AI/rocm-systems
2
0
Forka 0
Codice Problemi Pull Request Actions Pacchetti Progetti Rilasci Wiki Attività

Add 'projects/rocprofiler-sdk/' from commit 'bf0fad1d5406fbc51403ba1aa9621a9d4a9bce2b'

Sfoglia il codice sorgente 
git-subtree-dir: projects/rocprofiler-sdk
git-subtree-mainline: 50a90550e9
git-subtree-split: bf0fad1d54
This commit is contained in:
systems-assistant[bot]
2025-07-22 22:52:46 +00:00
parent 50a90550e9 bf0fad1d54
commit ad0fb25ed5
1272 ha cambiato i file con 230117 aggiunte e 0 eliminazioni
Scarica il file Patch Scarica il file Diff Expand all files Collapse all files
projects/rocprofiler-sdk/source/share/CMakeLists.txt
+6
Vedi File
@@ -0,0 +1,6 @@
#
#
#
add_subdirectory(rocprofiler-sdk)
add_subdirectory(rocprofiler-sdk-rocpd)
projects/rocprofiler-sdk/source/share/rocprofiler-sdk-rocpd/CMakeLists.txt
+15
Vedi File
@@ -0,0 +1,15 @@
#
#
#
set(rocpd_schemas data_views.sql marker_views.sql rocpd_indexes.sql rocpd_tables.sql
rocpd_views.sql summary_views.sql)
foreach(_FILE ${rocpd_schemas})
configure_file(${_FILE} ${PROJECT_BINARY_DIR}/share/rocprofiler-sdk-rocpd/${_FILE}
COPYONLY)
install(
FILES ${PROJECT_BINARY_DIR}/share/rocprofiler-sdk-rocpd/${_FILE}
DESTINATION share/rocprofiler-sdk-rocpd
COMPONENT rocpd)
endforeach()
projects/rocprofiler-sdk/source/share/rocprofiler-sdk-rocpd/data_views.sql
+722
Vedi File
@@ -0,0 +1,722 @@
--
-- Useful views
--
-- Code objects
CREATE VIEW IF NOT EXISTS
`code_objects` AS
SELECT
CO.id,
CO.guid,
CO.nid,
P.pid,
A.absolute_index AS agent_abs_index,
CO.uri,
CO.load_base,
CO.load_size,
CO.load_delta,
CO.storage_type AS storage_type_str,
JSON_EXTRACT(CO.extdata, '$.size') AS code_object_size,
JSON_EXTRACT(CO.extdata, '$.storage_type') AS storage_type,
JSON_EXTRACT(CO.extdata, '$.memory_base') AS memory_base,
JSON_EXTRACT(CO.extdata, '$.memory_size') AS memory_size
FROM
`rocpd_info_code_object` CO
INNER JOIN `rocpd_info_agent` A ON CO.agent_id = A.id
AND CO.guid = A.guid
INNER JOIN `rocpd_info_process` P ON CO.pid = P.id
AND CO.guid = P.guid;
CREATE VIEW IF NOT EXISTS
`kernel_symbols` AS
SELECT
KS.id,
KS.guid,
KS.nid,
P.pid,
KS.code_object_id,
KS.kernel_name,
KS.display_name,
KS.kernel_object,
KS.kernarg_segment_size,
KS.kernarg_segment_alignment,
KS.group_segment_size,
KS.private_segment_size,
KS.sgpr_count,
KS.arch_vgpr_count,
KS.accum_vgpr_count,
JSON_EXTRACT(KS.extdata, '$.size') AS kernel_symbol_size,
JSON_EXTRACT(KS.extdata, '$.kernel_id') AS kernel_id,
JSON_EXTRACT(KS.extdata, '$.kernel_code_entry_byte_offset') AS kernel_code_entry_byte_offset,
JSON_EXTRACT(KS.extdata, '$.formatted_kernel_name') AS formatted_kernel_name,
JSON_EXTRACT(KS.extdata, '$.demangled_kernel_name') AS demangled_kernel_name,
JSON_EXTRACT(KS.extdata, '$.truncated_kernel_name') AS truncated_kernel_name,
JSON_EXTRACT(KS.extdata, '$.kernel_address.handle') AS kernel_address
FROM
`rocpd_info_kernel_symbol` KS
INNER JOIN `rocpd_info_process` P ON KS.pid = P.id
AND KS.guid = P.guid;
-- Processes
CREATE VIEW IF NOT EXISTS
`processes` AS
SELECT
N.id AS nid,
N.machine_id,
N.system_name,
N.hostname,
N.release AS system_release,
N.version AS system_version,
P.guid,
P.ppid,
P.pid,
P.init,
P.start,
P.end,
P.fini,
P.command
FROM
`rocpd_info_process` P
INNER JOIN `rocpd_info_node` N ON N.id = P.nid
AND N.guid = P.guid;
-- Threads
CREATE VIEW IF NOT EXISTS
`threads` AS
SELECT
N.id AS nid,
N.machine_id,
N.system_name,
N.hostname,
N.release AS system_release,
N.version AS system_version,
P.guid,
P.ppid,
P.pid,
T.tid,
T.start,
T.end,
T.name
FROM
`rocpd_info_thread` T
INNER JOIN `rocpd_info_process` P ON P.id = T.pid
AND N.guid = T.guid
INNER JOIN `rocpd_info_node` N ON N.id = T.nid
AND N.guid = T.guid;
-- CPU regions
CREATE VIEW IF NOT EXISTS
`regions` AS
SELECT
R.id,
R.guid,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
S.string AS name,
R.nid,
P.pid,
T.tid,
R.start,
R.end,
(R.end - R.start) AS duration,
R.event_id,
E.stack_id,
E.parent_stack_id,
E.correlation_id AS corr_id,
E.extdata,
E.call_stack,
E.line_info
FROM
`rocpd_region` R
INNER JOIN `rocpd_event` E ON E.id = R.event_id
AND E.guid = R.guid
INNER JOIN `rocpd_string` S ON S.id = R.name_id
AND S.guid = R.guid
INNER JOIN `rocpd_info_process` P ON P.id = R.pid
AND P.guid = R.guid
INNER JOIN `rocpd_info_thread` T ON T.id = R.tid
AND T.guid = R.guid;
CREATE VIEW IF NOT EXISTS
`region_args` AS
SELECT
R.id,
R.guid,
R.nid,
P.pid,
A.type,
A.name,
A.value
FROM
`rocpd_region` R
INNER JOIN `rocpd_event` E ON E.id = R.event_id
AND E.guid = R.guid
INNER JOIN `rocpd_arg` A ON A.event_id = E.id
AND A.guid = R.guid
INNER JOIN `rocpd_info_process` P ON P.id = R.pid
AND P.guid = R.guid;
--
-- Samples
CREATE VIEW IF NOT EXISTS
`samples` AS
SELECT
R.id,
R.guid,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = T.name_id
AND RS.guid = T.guid
) AS name,
T.nid,
P.pid,
TH.tid,
R.timestamp,
R.event_id,
E.stack_id AS stack_id,
E.parent_stack_id AS parent_stack_id,
E.correlation_id AS corr_id,
E.extdata AS extdata,
E.call_stack AS call_stack,
E.line_info AS line_info
FROM
`rocpd_sample` R
INNER JOIN `rocpd_track` T ON T.id = R.track_id
AND T.guid = R.guid
INNER JOIN `rocpd_event` E ON E.id = R.event_id
AND E.guid = R.guid
INNER JOIN `rocpd_info_process` P ON P.id = T.pid
AND P.guid = T.guid
INNER JOIN `rocpd_info_thread` TH ON TH.id = T.tid
AND TH.guid = T.guid;
--
-- Provides samples view with the same columns as regions view
CREATE VIEW IF NOT EXISTS
`sample_regions` AS
SELECT
R.id,
R.guid,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = T.name_id
AND RS.guid = T.guid
) AS name,
T.nid,
P.pid,
TH.tid,
R.timestamp AS start,
R.timestamp AS END,
(R.timestamp - R.timestamp) AS duration,
R.event_id,
E.stack_id AS stack_id,
E.parent_stack_id AS parent_stack_id,
E.correlation_id AS corr_id,
E.extdata AS extdata,
E.call_stack AS call_stack,
E.line_info AS line_info
FROM
`rocpd_sample` R
INNER JOIN `rocpd_track` T ON T.id = R.track_id
AND T.guid = R.guid
INNER JOIN `rocpd_event` E ON E.id = R.event_id
AND E.guid = R.guid
INNER JOIN `rocpd_info_process` P ON P.id = T.pid
AND P.guid = T.guid
INNER JOIN `rocpd_info_thread` TH ON TH.id = T.tid
AND TH.guid = T.guid;
--
-- Provides a unified view of the regions and samples
CREATE VIEW IF NOT EXISTS
`regions_and_samples` AS
SELECT
*
FROM
`regions`
UNION ALL
SELECT
*
FROM
`sample_regions`;
--
-- Kernel information
CREATE VIEW
`kernels` AS
SELECT
K.id,
K.guid,
T.tid,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
R.string AS region,
S.display_name AS name,
K.nid,
P.pid,
A.absolute_index AS agent_abs_index,
A.logical_index AS agent_log_index,
A.type_index AS agent_type_index,
A.type AS agent_type,
S.code_object_id AS code_object_id,
K.kernel_id,
K.dispatch_id,
K.stream_id,
K.queue_id,
Q.name AS queue,
ST.name AS stream,
K.start,
K.end,
(K.end - K.start) AS duration,
K.grid_size_x AS grid_x,
K.grid_size_y AS grid_y,
K.grid_size_z AS grid_z,
K.workgroup_size_x AS workgroup_x,
K.workgroup_size_y AS workgroup_y,
K.workgroup_size_z AS workgroup_z,
K.group_segment_size AS lds_size,
K.private_segment_size AS scratch_size,
S.group_segment_size AS static_lds_size,
S.private_segment_size AS static_scratch_size,
E.stack_id,
E.parent_stack_id,
E.correlation_id AS corr_id
FROM
`rocpd_kernel_dispatch` K
INNER JOIN `rocpd_info_agent` A ON A.id = K.agent_id
AND A.guid = K.guid
INNER JOIN `rocpd_event` E ON E.id = K.event_id
AND E.guid = K.guid
INNER JOIN `rocpd_string` R ON R.id = K.region_name_id
AND R.guid = K.guid
INNER JOIN `rocpd_info_kernel_symbol` S ON S.id = K.kernel_id
AND S.guid = K.guid
LEFT JOIN `rocpd_info_stream` ST ON ST.id = K.stream_id
AND ST.guid = K.guid
LEFT JOIN `rocpd_info_queue` Q ON Q.id = K.queue_id
AND Q.guid = K.guid
INNER JOIN `rocpd_info_process` P ON P.id = Q.pid
AND P.guid = Q.guid
INNER JOIN `rocpd_info_thread` T ON T.id = K.tid
AND T.guid = K.guid;
--
-- Performance Monitoring Counters (PMC)
CREATE VIEW IF NOT EXISTS
`pmc_info` AS
SELECT
PMC_I.id,
PMC_I.guid,
PMC_I.nid,
P.pid,
A.absolute_index AS agent_abs_index,
PMC_I.is_constant,
PMC_I.is_derived,
PMC_I.name,
PMC_I.description,
PMC_I.block,
PMC_I.expression
FROM
`rocpd_info_pmc` PMC_I
INNER JOIN `rocpd_info_agent` A ON PMC_I.agent_id = A.id
AND PMC_I.guid = A.guid
INNER JOIN `rocpd_info_process` P ON P.id = PMC_I.pid
AND PMC_I.guid = P.guid;
CREATE VIEW IF NOT EXISTS
`pmc_events` AS
SELECT
PMC_E.id,
PMC_E.guid,
PMC_E.pmc_id,
E.id AS event_id,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
(
SELECT
display_name
FROM
`rocpd_info_kernel_symbol` KS
WHERE
KS.id = K.kernel_id
AND KS.guid = K.guid
) AS name,
K.nid,
P.pid,
K.dispatch_id,
K.start,
K.end,
(K.end - K.start) AS duration,
PMC_I.name AS counter_name,
PMC_E.value AS counter_value
FROM
`rocpd_pmc_event` PMC_E
INNER JOIN `rocpd_info_pmc` PMC_I ON PMC_I.id = PMC_E.pmc_id
AND PMC_I.guid = PMC_E.guid
INNER JOIN `rocpd_event` E ON E.id = PMC_E.event_id
AND E.guid = PMC_E.guid
INNER JOIN `rocpd_kernel_dispatch` K ON K.event_id = PMC_E.event_id
AND K.guid = PMC_E.guid
INNER JOIN `rocpd_info_process` P ON P.id = K.pid
AND P.guid = K.guid;
-- events with arguments ---
CREATE VIEW IF NOT EXISTS
`events_args` AS
SELECT
E.id AS event_id,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
E.stack_id,
E.parent_stack_id,
E.correlation_id,
A.position AS arg_position,
A.type AS arg_type,
A.name AS arg_name,
A.value AS arg_value,
E.call_stack,
E.line_info,
A.extdata
FROM
`rocpd_event` E
INNER JOIN `rocpd_arg` A ON A.event_id = E.id
AND A.guid = E.guid;
-- list of astream arguments enriched by the corresponding stream descriptions
CREATE VIEW IF NOT EXISTS
`stream_args` AS
SELECT
A.id AS argument_id,
A.event_id AS event_id,
A.position AS arg_position,
A.type AS arg_type,
A.value AS arg_value,
JSON_EXTRACT(A.extdata, '$.stream_id') AS stream_id,
S.nid,
P.pid,
S.name AS stream_name,
S.extdata AS extdata
FROM
`rocpd_arg` A
INNER JOIN `rocpd_info_stream` S ON JSON_EXTRACT(A.extdata, '$.stream_id') = S.id
AND A.guid = S.guid
INNER JOIN `rocpd_info_process` P ON P.id = S.pid
AND P.guid = S.guid
WHERE
A.name = 'stream';
--
--
CREATE VIEW IF NOT EXISTS
`memory_copies` AS
SELECT
M.id,
M.guid,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
M.nid,
P.pid,
T.tid,
M.start,
M.end,
(M.end - M.start) AS duration,
S.string AS name,
R.string AS region_name,
M.stream_id,
M.queue_id,
ST.name AS stream_name,
Q.name AS queue_name,
M.size,
dst_agent.name AS dst_device,
dst_agent.absolute_index AS dst_agent_abs_index,
dst_agent.logical_index AS dst_agent_log_index,
dst_agent.type_index AS dst_agent_type_index,
dst_agent.type AS dst_agent_type,
M.dst_address,
src_agent.name AS src_device,
src_agent.absolute_index AS src_agent_abs_index,
src_agent.logical_index AS src_agent_log_index,
src_agent.type_index AS src_agent_type_index,
src_agent.type AS src_agent_type,
M.src_address,
E.stack_id,
E.parent_stack_id,
E.correlation_id AS corr_id
FROM
`rocpd_memory_copy` M
INNER JOIN `rocpd_string` S ON S.id = M.name_id
AND S.guid = M.guid
LEFT JOIN `rocpd_string` R ON R.id = M.region_name_id
AND R.guid = M.guid
INNER JOIN `rocpd_info_agent` dst_agent ON dst_agent.id = M.dst_agent_id
AND dst_agent.guid = M.guid
INNER JOIN `rocpd_info_agent` src_agent ON src_agent.id = M.src_agent_id
AND src_agent.guid = M.guid
LEFT JOIN `rocpd_info_queue` Q ON Q.id = M.queue_id
AND Q.guid = M.guid
LEFT JOIN `rocpd_info_stream` ST ON ST.id = M.stream_id
AND ST.guid = M.guid
INNER JOIN `rocpd_event` E ON E.id = M.event_id
AND E.guid = M.guid
INNER JOIN `rocpd_info_process` P ON P.id = M.pid
AND P.guid = M.guid
INNER JOIN `rocpd_info_thread` T ON T.id = M.tid
AND T.guid = M.guid;
--
--
CREATE VIEW IF NOT EXISTS
`memory_allocations` AS
SELECT
M.id,
M.guid,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
M.nid,
P.pid,
T.tid,
M.start,
M.end,
(M.end - M.start) AS duration,
M.type,
M.level,
A.name AS agent_name,
A.absolute_index AS agent_abs_index,
A.logical_index AS agent_log_index,
A.type_index AS agent_type_index,
A.type AS agent_type,
M.address,
M.size,
M.queue_id,
Q.name AS queue_name,
M.stream_id,
ST.name AS stream_name,
E.stack_id,
E.parent_stack_id,
E.correlation_id AS corr_id
FROM
`rocpd_memory_allocate` M
LEFT JOIN `rocpd_info_agent` A ON M.agent_id = A.id
AND M.guid = A.guid
LEFT JOIN `rocpd_info_queue` Q ON Q.id = M.queue_id
AND Q.guid = M.guid
LEFT JOIN `rocpd_info_stream` ST ON ST.id = M.stream_id
AND ST.guid = M.guid
INNER JOIN `rocpd_event` E ON E.id = M.event_id
AND E.guid = M.guid
INNER JOIN `rocpd_info_process` P ON P.id = M.pid
AND P.guid = M.guid
INNER JOIN `rocpd_info_thread` T ON T.id = M.tid
AND P.guid = M.guid;
--
--
CREATE VIEW IF NOT EXISTS
`scratch_memory` AS
SELECT
M.id,
M.guid,
M.nid,
P.pid,
M.type AS operation,
A.name AS agent_name,
A.absolute_index AS agent_abs_index,
A.logical_index AS agent_log_index,
A.type_index AS agent_type_index,
A.type AS agent_type,
M.queue_id,
T.tid,
JSON_EXTRACT(M.extdata, '$.flags') AS alloc_flags,
M.start,
M.end,
M.size,
M.address,
E.correlation_id,
E.stack_id,
E.parent_stack_id,
E.correlation_id AS corr_id,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
E.extdata AS event_extdata
FROM
`rocpd_memory_allocate` M
LEFT JOIN `rocpd_info_agent` A ON M.agent_id = A.id
AND M.guid = A.guid
LEFT JOIN `rocpd_info_queue` Q ON Q.id = M.queue_id
AND Q.guid = M.guid
INNER JOIN `rocpd_event` E ON E.id = M.event_id
AND E.guid = M.guid
INNER JOIN `rocpd_info_process` P ON P.id = M.pid
AND P.guid = M.guid
INNER JOIN `rocpd_info_thread` T ON T.id = M.tid
AND T.guid = M.guid
WHERE
M.level = 'SCRATCH'
ORDER BY
M.start ASC;
--
--
CREATE VIEW IF NOT EXISTS
`counters_collection` AS
SELECT
MIN(PMC_E.id) AS id,
PMC_E.guid,
K.dispatch_id,
K.kernel_id,
E.id AS event_id,
E.correlation_id,
E.stack_id,
E.parent_stack_id,
P.pid,
T.tid,
K.agent_id,
A.absolute_index AS agent_abs_index,
A.logical_index AS agent_log_index,
A.type_index AS agent_type_index,
A.type AS agent_type,
K.queue_id,
k.grid_size_x AS grid_size_x,
k.grid_size_y AS grid_size_y,
k.grid_size_z AS grid_size_z,
(K.grid_size_x * K.grid_size_y * K.grid_size_z) AS grid_size,
S.display_name AS kernel_name,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = K.region_name_id
AND RS.guid = K.guid
) AS kernel_region,
K.workgroup_size_x AS workgroup_size_x,
K.workgroup_size_y AS workgroup_size_y,
K.workgroup_size_z AS workgroup_size_z,
(K.workgroup_size_x * K.workgroup_size_y * K.workgroup_size_z) AS workgroup_size,
K.group_segment_size AS lds_block_size,
K.private_segment_size AS scratch_size,
S.arch_vgpr_count AS vgpr_count,
S.accum_vgpr_count,
S.sgpr_count,
PMC_I.name AS counter_name,
PMC_I.symbol AS counter_symbol,
PMC_I.component,
PMC_I.description,
PMC_I.block,
PMC_I.expression,
PMC_I.value_type,
PMC_I.id AS counter_id,
SUM(PMC_E.value) AS value,
K.start,
K.end,
PMC_I.is_constant,
PMC_I.is_derived,
(K.end - K.start) AS duration,
(
SELECT
string
FROM
`rocpd_string` RS
WHERE
RS.id = E.category_id
AND RS.guid = E.guid
) AS category,
K.nid,
E.extdata,
S.code_object_id
FROM
`rocpd_pmc_event` PMC_E
INNER JOIN `rocpd_info_pmc` PMC_I ON PMC_I.id = PMC_E.pmc_id
AND PMC_I.guid = PMC_E.guid
INNER JOIN `rocpd_event` E ON E.id = PMC_E.event_id
AND E.guid = PMC_E.guid
INNER JOIN `rocpd_kernel_dispatch` K ON K.event_id = PMC_E.event_id
AND K.guid = PMC_E.guid
INNER JOIN `rocpd_info_agent` A ON A.id = K.agent_id
AND A.guid = K.guid
INNER JOIN `rocpd_info_kernel_symbol` S ON S.id = K.kernel_id
AND S.guid = K.guid
INNER JOIN `rocpd_info_process` P ON P.id = K.pid
AND P.guid = K.guid
INNER JOIN `rocpd_info_thread` T ON T.id = K.tid
AND T.guid = K.guid
GROUP BY
PMC_E.guid,
K.dispatch_id,
PMC_I.name,
K.agent_id;
projects/rocprofiler-sdk/source/share/rocprofiler-sdk-rocpd/marker_views.sql
+3
Vedi File
@@ -0,0 +1,3 @@
--
-- Views related to markers
--
projects/rocprofiler-sdk/source/share/rocprofiler-sdk-rocpd/rocpd_indexes.sql
+45
Vedi File
@@ -0,0 +1,45 @@
--
-- Indexes for the various fields
--
-- string field
-- CREATE INDEX `rocpd_string{{uuid}}_string_idx` ON `rocpd_string{{uuid}}` ("string");
-- guid field
-- CREATE INDEX `rocpd_string{{uuid}}_guid_idx` ON `rocpd_string{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_node{{uuid}}_guid_idx` ON `rocpd_info_node{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_process{{uuid}}_guid_idx` ON `rocpd_info_process{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_thread{{uuid}}_guid_idx` ON `rocpd_info_thread{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_agent{{uuid}}_guid_idx` ON `rocpd_info_agent{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_queue{{uuid}}_guid_idx` ON `rocpd_info_queue{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_stream{{uuid}}_guid_idx` ON `rocpd_info_stream{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_pmc{{uuid}}_guid_idx` ON `rocpd_info_pmc{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_code_object{{uuid}}_guid_idx` ON `rocpd_info_code_object{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_info_kernel_symbol{{uuid}}_guid_idx` ON `rocpd_info_kernel_symbol{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_track{{uuid}}_guid_idx` ON `rocpd_track{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_event{{uuid}}_guid_idx` ON `rocpd_event{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_arg{{uuid}}_guid_idx` ON `rocpd_arg{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_pmc_event{{uuid}}_guid_idx` ON `rocpd_pmc_event{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_region{{uuid}}_guid_idx` ON `rocpd_region{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_sample{{uuid}}_guid_idx` ON `rocpd_sample{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_kernel_dispatch{{uuid}}_guid_idx` ON `rocpd_kernel_dispatch{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_memory_copy{{uuid}}_guid_idx` ON `rocpd_memory_copy{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_memory_allocate{{uuid}}_guid_idx` ON `rocpd_memory_allocate{{uuid}}` ("id", "guid");
-- CREATE INDEX `rocpd_event{{uuid}}_category_idx` ON `rocpd_event{{uuid}}` ("id", "guid", "category_id");
-- CREATE INDEX `rocpd_region{{uuid}}_event_idx` ON `rocpd_region{{uuid}}` ("id", "guid", "event_id");
-- CREATE INDEX `rocpd_region{{uuid}}_name_idx` ON `rocpd_region{{uuid}}` ("id", "guid", "name_id");
-- CREATE INDEX `rocpd_sample{{uuid}}_event_idx` ON `rocpd_sample{{uuid}}` ("id", "guid", "event_id");
-- CREATE INDEX `rocpd_sample{{uuid}}_track_idx` ON `rocpd_sample{{uuid}}` ("id", "guid", "track_id");
-- CREATE INDEX `rocpd_track{{uuid}}_name_idx` ON `rocpd_track{{uuid}}` ("id", "guid", "name_id");
-- CREATE INDEX `rocpd_memory_copy{{uuid}}_guid_nid_pid_idx` ON `rocpd_memory_copy{{uuid}}` ("guid", "nid", "pid");
-- CREATE INDEX `rocpd_kernel_dispatch{{uuid}}_guid_nid_pid_idx` ON `rocpd_kernel_dispatch{{uuid}}` ("guid", "nid", "pid");
-- CREATE INDEX `rocpd_region{{uuid}}_guid_idx` ON `rocpd_region{{uuid}}` ("guid", "nid", "pid");
-- CREATE INDEX `rocpd_sample{{uuid}}_guid_nid_pid_idx` ON `rocpd_sample{{uuid}}` ("guid", "nid", "pid");
-- CREATE INDEX `rocpd_region{{uuid}}_guid_idx` ON `rocpd_region{{uuid}}` ("guid");
-- CREATE INDEX `rocpd_region{{uuid}}_nid_idx` ON `rocpd_region{{uuid}}` ("nid");
-- CREATE INDEX `rocpd_region{{uuid}}_pid_idx` ON `rocpd_region{{uuid}}` ("pid");
-- CREATE INDEX `rocpd_region{{uuid}}_start_idx` ON `rocpd_region{{uuid}}` ("start");
-- CREATE INDEX `rocpd_region{{uuid}}_end_idx` ON `rocpd_region{{uuid}}` ("end");
projects/rocprofiler-sdk/source/share/rocprofiler-sdk-rocpd/rocpd_tables.sql
+373
Vedi File
@@ -0,0 +1,373 @@
-- Enable foreign key support for cascading
PRAGMA foreign_keys = ON;
CREATE TABLE IF NOT EXISTS
"rocpd_metadata{{uuid}}" (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"tag" TEXT NOT NULL,
"value" TEXT NOT NULL
);
CREATE TABLE IF NOT EXISTS
`rocpd_string{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"string" TEXT NOT NULL UNIQUE ON CONFLICT ABORT
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_node{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"hash" BIGINT NOT NULL UNIQUE,
"machine_id" TEXT NOT NULL UNIQUE,
"system_name" TEXT,
"hostname" TEXT,
"release" TEXT,
"version" TEXT,
"hardware_name" TEXT,
"domain_name" TEXT
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_process{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"ppid" INTEGER,
"pid" INTEGER NOT NULL,
"init" BIGINT,
"fini" BIGINT,
"start" BIGINT,
"end" BIGINT,
"command" TEXT,
"environment" JSONB DEFAULT "{}" NOT NULL,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_thread{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"ppid" INTEGER,
"pid" INTEGER NOT NULL,
"tid" INTEGER NOT NULL,
"name" TEXT,
"start" BIGINT,
"end" BIGINT,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_agent{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"type" TEXT CHECK ("type" IN ('CPU', 'GPU')),
"absolute_index" INTEGER,
"logical_index" INTEGER,
"type_index" INTEGER,
"uuid" INTEGER,
"name" TEXT,
"model_name" TEXT,
"vendor_name" TEXT,
"product_name" TEXT,
"user_name" TEXT,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_queue{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"name" TEXT,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_stream{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"name" TEXT,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE
);
-- 2993533, 2269219937, 2993533
-- 2993533, 2269219937, 2993533
-- Performance monitoring counters (PMC) descriptions
CREATE TABLE IF NOT EXISTS
`rocpd_info_pmc{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"agent_id" INTEGER,
"target_arch" TEXT CHECK ("target_arch" IN ('CPU', 'GPU')),
"event_code" INT,
"instance_id" INTEGER,
"name" TEXT NOT NULL,
"symbol" TEXT NOT NULL,
"description" TEXT,
"long_description" TEXT DEFAULT "",
"component" TEXT,
"units" TEXT DEFAULT "",
"value_type" TEXT CHECK ("value_type" IN ('ABS', 'ACCUM', 'RELATIVE')),
"block" TEXT,
"expression" TEXT,
"is_constant" INTEGER,
"is_derived" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (agent_id) REFERENCES `rocpd_info_agent{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_code_object{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"agent_id" INTEGER,
"uri" TEXT,
"load_base" BIGINT,
"load_size" BIGINT,
"load_delta" BIGINT,
"storage_type" TEXT CHECK ("storage_type" IN ('FILE', 'MEMORY')),
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (agent_id) REFERENCES `rocpd_info_agent{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_info_kernel_symbol{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"code_object_id" INTEGER NOT NULL,
"kernel_name" TEXT,
"display_name" TEXT,
"kernel_object" INTEGER,
"kernarg_segment_size" INTEGER,
"kernarg_segment_alignment" INTEGER,
"group_segment_size" INTEGER,
"private_segment_size" INTEGER,
"sgpr_count" INTEGER,
"arch_vgpr_count" INTEGER,
"accum_vgpr_count" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (code_object_id) REFERENCES `rocpd_info_code_object{{uuid}}` (id) ON UPDATE CASCADE
);
-- Stores repetitive info for samples
CREATE TABLE IF NOT EXISTS
`rocpd_track{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER,
"tid" INTEGER,
"name_id" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (tid) REFERENCES `rocpd_info_thread{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (name_id) REFERENCES `rocpd_string{{uuid}}` (id) ON UPDATE CASCADE
);
-- Storage for a region, instant, and counter
CREATE TABLE IF NOT EXISTS
`rocpd_event{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"category_id" INTEGER,
"stack_id" INTEGER,
"parent_stack_id" INTEGER,
"correlation_id" INTEGER,
"call_stack" JSONB DEFAULT "{}" NOT NULL,
"line_info" JSONB DEFAULT "{}" NOT NULL,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (category_id) REFERENCES `rocpd_string{{uuid}}` (id) ON UPDATE CASCADE
);
-- stores arguments for events
CREATE TABLE IF NOT EXISTS
`rocpd_arg{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"event_id" INTEGER NOT NULL,
"position" INTEGER NOT NULL,
"type" TEXT NOT NULL,
"name" TEXT NOT NULL,
"value" TEXT, -- TODO: discuss make it value_id and integer, refer to string table --
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (event_id) REFERENCES `rocpd_event{{uuid}}` (id) ON UPDATE CASCADE
);
-- Region with a start/stop on the same thread (CPU)
CREATE TABLE IF NOT EXISTS
`rocpd_pmc_event{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"event_id" INTEGER,
"pmc_id" INTEGER NOT NULL,
"value" REAL DEFAULT 0.0,
"extdata" JSONB DEFAULT "{}",
FOREIGN KEY (pmc_id) REFERENCES `rocpd_info_pmc{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (event_id) REFERENCES `rocpd_event{{uuid}}` (id) ON UPDATE CASCADE
);
-- Region with a start/stop on the same thread (CPU)
CREATE TABLE IF NOT EXISTS
`rocpd_region{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"tid" INTEGER NOT NULL,
"start" BIGINT NOT NULL,
"end" BIGINT NOT NULL,
"name_id" INTEGER NOT NULL,
"event_id" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (tid) REFERENCES `rocpd_info_thread{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (name_id) REFERENCES `rocpd_string{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (event_id) REFERENCES `rocpd_event{{uuid}}` (id) ON UPDATE CASCADE
);
-- Instantaneous sample
CREATE TABLE IF NOT EXISTS
`rocpd_sample{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"track_id" INTEGER NOT NULL,
"timestamp" BIGINT NOT NULL,
"event_id" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (track_id) REFERENCES `rocpd_track{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (event_id) REFERENCES `rocpd_event{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_kernel_dispatch{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"tid" INTEGER,
"agent_id" INTEGER NOT NULL,
"kernel_id" INTEGER NOT NULL,
"dispatch_id" INTEGER NOT NULL,
"queue_id" INTEGER NOT NULL,
"stream_id" INTEGER NOT NULL,
"start" BIGINT NOT NULL,
"end" BIGINT NOT NULL,
"private_segment_size" INTEGER,
"group_segment_size" INTEGER,
"workgroup_size_x" INTEGER NOT NULL,
"workgroup_size_y" INTEGER NOT NULL,
"workgroup_size_z" INTEGER NOT NULL,
"grid_size_x" INTEGER NOT NULL,
"grid_size_y" INTEGER NOT NULL,
"grid_size_z" INTEGER NOT NULL,
"region_name_id" INTEGER,
"event_id" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (tid) REFERENCES `rocpd_info_thread{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (agent_id) REFERENCES `rocpd_info_agent{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (kernel_id) REFERENCES `rocpd_info_kernel_symbol{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (queue_id) REFERENCES `rocpd_info_queue{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (stream_id) REFERENCES `rocpd_info_stream{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (region_name_id) REFERENCES `rocpd_string{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (event_id) REFERENCES `rocpd_event{{uuid}}` (id) ON UPDATE CASCADE
);
CREATE TABLE IF NOT EXISTS
`rocpd_memory_copy{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"tid" INTEGER,
"start" BIGINT NOT NULL,
"end" BIGINT NOT NULL,
"name_id" INTEGER NOT NULL,
"dst_agent_id" INTEGER,
"dst_address" INTEGER,
"src_agent_id" INTEGER,
"src_address" INTEGER,
"size" INTEGER NOT NULL,
"queue_id" INTEGER,
"stream_id" INTEGER,
"region_name_id" INTEGER,
"event_id" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (tid) REFERENCES `rocpd_info_thread{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (name_id) REFERENCES `rocpd_string{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (dst_agent_id) REFERENCES `rocpd_info_agent{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (src_agent_id) REFERENCES `rocpd_info_agent{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (stream_id) REFERENCES `rocpd_info_stream{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (queue_id) REFERENCES `rocpd_info_queue{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (region_name_id) REFERENCES `rocpd_string{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (event_id) REFERENCES `rocpd_event{{uuid}}` (id) ON UPDATE CASCADE
);
-- Memory allocations (real memory, virtual memory, and scratch memory)
CREATE TABLE IF NOT EXISTS
`rocpd_memory_allocate{{uuid}}` (
"id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
"guid" TEXT DEFAULT "{{guid}}" NOT NULL,
"nid" INTEGER NOT NULL,
"pid" INTEGER NOT NULL,
"tid" INTEGER,
"agent_id" INTEGER,
"type" TEXT CHECK ("type" IN ('ALLOC', 'FREE', 'REALLOC', 'RECLAIM')),
"level" TEXT CHECK ("level" IN ('REAL', 'VIRTUAL', 'SCRATCH')),
"start" BIGINT NOT NULL,
"end" BIGINT NOT NULL,
"address" INTEGER,
"size" INTEGER NOT NULL,
"queue_id" INTEGER,
"stream_id" INTEGER,
"event_id" INTEGER,
"extdata" JSONB DEFAULT "{}" NOT NULL,
FOREIGN KEY (nid) REFERENCES `rocpd_info_node{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (pid) REFERENCES `rocpd_info_process{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (tid) REFERENCES `rocpd_info_thread{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (agent_id) REFERENCES `rocpd_info_agent{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (stream_id) REFERENCES `rocpd_info_stream{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (queue_id) REFERENCES `rocpd_info_queue{{uuid}}` (id) ON UPDATE CASCADE,
FOREIGN KEY (event_id) REFERENCES `rocpd_event{{uuid}}` (id) ON UPDATE CASCADE
);
INSERT INTO
`rocpd_metadata{{uuid}}` ("tag", "value")
VALUES
("schema_version", "3"),
("uuid", "{{uuid}}"),
("guid", "{{guid}}");
projects/rocprofiler-sdk/source/share/rocprofiler-sdk-rocpd/rocpd_views.sql
+139
Vedi File
@@ -0,0 +1,139 @@
CREATE VIEW IF NOT EXISTS
`rocpd_metadata` AS
SELECT
*
FROM
`rocpd_metadata{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_string` AS
SELECT
*
FROM
`rocpd_string{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_node` AS
SELECT
*
FROM
`rocpd_info_node{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_process` AS
SELECT
*
FROM
`rocpd_info_process{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_thread` AS
SELECT
*
FROM
`rocpd_info_thread{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_agent` AS
SELECT
*
FROM
`rocpd_info_agent{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_queue` AS
SELECT
*
FROM
`rocpd_info_queue{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_stream` AS
SELECT
*
FROM
`rocpd_info_stream{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_pmc` AS
SELECT
*
FROM
`rocpd_info_pmc{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_code_object` AS
SELECT
*
FROM
`rocpd_info_code_object{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_info_kernel_symbol` AS
SELECT
*
FROM
`rocpd_info_kernel_symbol{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_track` AS
SELECT
*
FROM
`rocpd_track{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_event` AS
SELECT
*
FROM
`rocpd_event{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_arg` AS
SELECT
*
FROM
`rocpd_arg{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_pmc_event` AS
SELECT
*
FROM
`rocpd_pmc_event{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_region` AS
SELECT
*
FROM
`rocpd_region{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_sample` AS
SELECT
*
FROM
`rocpd_sample{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_kernel_dispatch` AS
SELECT
*
FROM
`rocpd_kernel_dispatch{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_memory_copy` AS
SELECT
*
FROM
`rocpd_memory_copy{{uuid}}`;
CREATE VIEW IF NOT EXISTS
`rocpd_memory_allocate` AS
SELECT
*
FROM
`rocpd_memory_allocate{{uuid}}`;
projects/rocprofiler-sdk/source/share/rocprofiler-sdk-rocpd/summary_views.sql
+376
Vedi File
@@ -0,0 +1,376 @@
--
-- Useful summary views
--
--
-- Sorted list of kernels which consume the most overall time
CREATE VIEW IF NOT EXISTS
`top_kernels` AS
SELECT
S.display_name AS name,
COUNT(K.kernel_id) AS total_calls,
SUM(K.end - K.start) / 1000.0 AS total_duration,
(SUM(K.end - K.start) / COUNT(K.kernel_id)) / 1000.0 AS average,
SUM(K.end - K.start) * 100.0 / (
SELECT
SUM(A.end - A.start)
FROM
`rocpd_kernel_dispatch` A
) AS percentage
FROM
`rocpd_kernel_dispatch` K
INNER JOIN `rocpd_info_kernel_symbol` S ON S.id = K.kernel_id
AND S.guid = K.guid
GROUP BY
name
ORDER BY
total_duration DESC;
--
-- GPU utilization metrics including kernels and memory copy operations
CREATE VIEW IF NOT EXISTS
`busy` AS
SELECT
A.agent_id,
AG.type,
GpuTime,
WallTime,
GpuTime * 1.0 / WallTime AS Busy
FROM
(
SELECT
agent_id,
guid,
SUM(END - start) AS GpuTime
FROM
(
SELECT
agent_id,
guid,
END,
start
FROM
`rocpd_kernel_dispatch`
UNION ALL
SELECT
dst_agent_id AS agent_id,
guid,
END,
start
FROM
`rocpd_memory_copy`
)
GROUP BY
agent_id,
guid
) A
INNER JOIN (
SELECT
MAX(END) - MIN(start) AS WallTime
FROM
(
SELECT
END,
start
FROM
`rocpd_kernel_dispatch`
UNION ALL
SELECT
END,
start
FROM
`rocpd_memory_copy`
)
) W ON 1 = 1
INNER JOIN `rocpd_info_agent` AG ON AG.id = A.agent_id
AND AG.guid = A.guid;
--
-- Overall performance summary including kernels and memory copy operations
CREATE VIEW
`top` AS
SELECT
name,
COUNT(*) AS total_calls,
SUM(duration) / 1000.0 AS total_duration,
(SUM(duration) / COUNT(*)) / 1000.0 AS average,
SUM(duration) * 100.0 / total_time AS percentage
FROM
(
-- Kernel operations
SELECT
ks.display_name AS name,
(kd.end - kd.start) AS duration
FROM
`rocpd_kernel_dispatch` kd
INNER JOIN `rocpd_info_kernel_symbol` ks ON kd.kernel_id = ks.id
AND kd.guid = ks.guid
UNION ALL
-- Memory operations
SELECT
rs.string AS name,
(END - start) AS duration
FROM
`rocpd_memory_copy` mc
INNER JOIN `rocpd_string` rs ON rs.id = mc.name_id
AND rs.guid = mc.guid
UNION ALL
-- Regions
SELECT
rs.string AS name,
(END - start) AS duration
FROM
`rocpd_region` rr
INNER JOIN `rocpd_string` rs ON rs.id = rr.name_id
AND rs.guid = rr.guid
) operations
CROSS JOIN (
SELECT
SUM(END - start) AS total_time
FROM
(
SELECT
END,
start
FROM
`rocpd_kernel_dispatch`
UNION ALL
SELECT
END,
start
FROM
`rocpd_memory_copy`
UNION ALL
SELECT
END,
start
FROM
`rocpd_region`
)
) TOTAL
GROUP BY
name
ORDER BY
total_duration DESC;
-- Kernel summary by name
CREATE VIEW
`kernel_summary` AS
WITH
avg_data AS (
SELECT
name,
AVG(duration) AS avg_duration
FROM
`kernels`
GROUP BY
name
),
aggregated_data AS (
SELECT
K.name,
COUNT(*) AS calls,
SUM(K.duration) AS total_duration,
SUM(CAST(K.duration AS REAL) * CAST(K.duration AS REAL)) AS sqr_duration,
A.avg_duration AS average_duration,
MIN(K.duration) AS min_duration,
MAX(K.duration) AS max_duration,
SUM(CAST((K.duration - A.avg_duration) AS REAL) * CAST((K.duration - A.avg_duration) AS REAL)) / (COUNT(*) - 1) AS variance_duration,
SQRT(
SUM(CAST((K.duration - A.avg_duration) AS REAL) * CAST((K.duration - A.avg_duration) AS REAL)) / (COUNT(*) - 1)
) AS std_dev_duration
FROM
`kernels` K
JOIN avg_data A ON K.name = A.name
GROUP BY
K.name
),
total_duration AS (
SELECT
SUM(total_duration) AS grand_total_duration
FROM
aggregated_data
)
SELECT
AD.name AS name,
AD.calls,
AD.total_duration AS "DURATION (nsec)",
AD.sqr_duration AS "SQR (nsec)",
AD.average_duration AS "AVERAGE (nsec)",
(CAST(AD.total_duration AS REAL) / TD.grand_total_duration) * 100 AS "PERCENT (INC)",
AD.min_duration AS "MIN (nsec)",
AD.max_duration AS "MAX (nsec)",
AD.variance_duration AS "VARIANCE",
AD.std_dev_duration AS "STD_DEV"
FROM
aggregated_data AD
CROSS JOIN total_duration TD;
--
-- Kernel summary by region name
CREATE VIEW
`kernel_summary_region` AS
WITH
avg_data AS (
SELECT
region,
AVG(duration) AS avg_duration
FROM
`kernels`
GROUP BY
region
),
aggregated_data AS (
SELECT
K.region AS name,
COUNT(*) AS calls,
SUM(K.duration) AS total_duration,
SUM(CAST(K.duration AS REAL) * CAST(K.duration AS REAL)) AS sqr_duration,
A.avg_duration AS average_duration,
MIN(K.duration) AS min_duration,
MAX(K.duration) AS max_duration,
SUM(CAST((K.duration - A.avg_duration) AS REAL) * CAST((K.duration - A.avg_duration) AS REAL)) / (COUNT(*) - 1) AS variance_duration,
SQRT(
SUM(CAST((K.duration - A.avg_duration) AS REAL) * CAST((K.duration - A.avg_duration) AS REAL)) / (COUNT(*) - 1)
) AS std_dev_duration
FROM
`kernels` K
JOIN avg_data A ON K.region = A.region
GROUP BY
K.region
),
total_duration AS (
SELECT
SUM(total_duration) AS grand_total_duration
FROM
aggregated_data
)
SELECT
AD.name AS name,
AD.calls,
AD.total_duration AS "DURATION (nsec)",
AD.sqr_duration AS "SQR (nsec)",
AD.average_duration AS "AVERAGE (nsec)",
(CAST(AD.total_duration AS REAL) / TD.grand_total_duration) * 100 AS "PERCENT (INC)",
AD.min_duration AS "MIN (nsec)",
AD.max_duration AS "MAX (nsec)",
AD.variance_duration AS "VARIANCE",
AD.std_dev_duration AS "STD_DEV"
FROM
aggregated_data AD
CROSS JOIN total_duration TD;
--
-- Memory copy summary
CREATE VIEW
`memory_copy_summary` AS
WITH
avg_data AS (
SELECT
name,
AVG(duration) AS avg_duration
FROM
`memory_copies`
GROUP BY
name
),
aggregated_data AS (
SELECT
MC.name,
COUNT(*) AS calls,
SUM(MC.duration) AS total_duration,
SUM(CAST(MC.duration AS REAL) * CAST(MC.duration AS REAL)) AS sqr_duration,
A.avg_duration AS average_duration,
MIN(MC.duration) AS min_duration,
MAX(MC.duration) AS max_duration,
SUM(
CAST((MC.duration - A.avg_duration) AS REAL) * CAST((MC.duration - A.avg_duration) AS REAL)
) / (COUNT(*) - 1) AS variance_duration,
SQRT(
SUM(
CAST((MC.duration - A.avg_duration) AS REAL) * CAST((MC.duration - A.avg_duration) AS REAL)
) / (COUNT(*) - 1)
) AS std_dev_duration
FROM
`memory_copies` MC
JOIN avg_data A ON MC.name = A.name
GROUP BY
MC.name
),
total_duration AS (
SELECT
SUM(total_duration) AS grand_total_duration
FROM
aggregated_data
)
SELECT
AD.name AS name,
AD.calls,
AD.total_duration AS "DURATION (nsec)",
AD.sqr_duration AS "SQR (nsec)",
AD.average_duration AS "AVERAGE (nsec)",
(CAST(AD.total_duration AS REAL) / TD.grand_total_duration) * 100 AS "PERCENT (INC)",
AD.min_duration AS "MIN (nsec)",
AD.max_duration AS "MAX (nsec)",
AD.variance_duration AS "VARIANCE",
AD.std_dev_duration AS "STD_DEV"
FROM
aggregated_data AD
CROSS JOIN total_duration TD;
--
-- Memory allocation summary
CREATE VIEW
`memory_allocation_summary` AS
WITH
avg_data AS (
SELECT
type AS name,
AVG(duration) AS avg_duration
FROM
`memory_allocations`
GROUP BY
type
),
aggregated_data AS (
SELECT
MA.type AS name,
COUNT(*) AS calls,
SUM(MA.duration) AS total_duration,
SUM(CAST(MA.duration AS REAL) * CAST(MA.duration AS REAL)) AS sqr_duration,
A.avg_duration AS average_duration,
MIN(MA.duration) AS min_duration,
MAX(MA.duration) AS max_duration,
SUM(
CAST((MA.duration - A.avg_duration) AS REAL) * CAST((MA.duration - A.avg_duration) AS REAL)
) / (COUNT(*) - 1) AS variance_duration,
SQRT(
SUM(
CAST((MA.duration - A.avg_duration) AS REAL) * CAST((MA.duration - A.avg_duration) AS REAL)
) / (COUNT(*) - 1)
) AS std_dev_duration
FROM
`memory_allocations` MA
JOIN avg_data A ON MA.type = A.name
GROUP BY
MA.type
),
total_duration AS (
SELECT
SUM(total_duration) AS grand_total_duration
FROM
aggregated_data
)
SELECT
'MEMORY_ALLOCATION_' || AD.name AS name,
AD.calls,
AD.total_duration AS "DURATION (nsec)",
AD.sqr_duration AS "SQR (nsec)",
AD.average_duration AS "AVERAGE (nsec)",
(CAST(AD.total_duration AS REAL) / TD.grand_total_duration) * 100 AS "PERCENT (INC)",
AD.min_duration AS "MIN (nsec)",
AD.max_duration AS "MAX (nsec)",
AD.variance_duration AS "VARIANCE",
AD.std_dev_duration AS "STD_DEV"
FROM
aggregated_data AD
CROSS JOIN total_duration TD;
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/CMakeLists.txt
+51
Vedi File
@@ -0,0 +1,51 @@
# MIT License
#
# Copyright (c) 2023-2025 Advanced Micro Devices, Inc. All rights reserved.
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
# non-executable files
set(share_files basic_counters.xml counter_defs.yaml derived_counters.xml)
# executables
set(share_progs convert-counters-collection-format.py)
foreach(_FILE ${share_files})
configure_file(${_FILE} ${PROJECT_BINARY_DIR}/share/rocprofiler-sdk/${_FILE} COPYONLY)
install(
FILES ${PROJECT_BINARY_DIR}/share/rocprofiler-sdk/${_FILE}
DESTINATION share/rocprofiler-sdk
COMPONENT core)
endforeach()
foreach(_FILE ${share_progs})
configure_file(${_FILE} ${PROJECT_BINARY_DIR}/share/rocprofiler-sdk/${_FILE}
USE_SOURCE_PERMISSIONS COPYONLY)
install(
PROGRAMS ${PROJECT_BINARY_DIR}/share/rocprofiler-sdk/${_FILE}
DESTINATION share/rocprofiler-sdk
COMPONENT core)
endforeach()
set(CONVERT_CC_FORMAT_PATH
${PROJECT_BINARY_DIR}/share/rocprofiler-sdk/convert-counters-collection-format.py)
add_executable(rocprofiler-sdk::convert-counters-collection-format IMPORTED)
set_property(TARGET rocprofiler-sdk::convert-counters-collection-format
PROPERTY IMPORTED_LOCATION ${CONVERT_CC_FORMAT_PATH})
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/basic_counters.xml
+780
Vedi File
@@ -0,0 +1,780 @@
<gfx8 base="gfx8">
<metric name="MAX_WAVE_SIZE" expr=wave_front_size descr="Max wave size constant"></metric>
<metric name="SE_NUM" expr=array_count/simd_arrays_per_engine descr="SE_NUM"></metric>
<metric name="SIMD_NUM" expr=simd_per_cu/CU_NUM descr="SIMD Number"></metric>
<metric name="CU_NUM" expr=cu_per_simd_array*array_count descr="CU_NUM"></metric>
<metric name="GRBM_COUNT" block=GRBM event=0 descr="Tie High - Count Number of Clocks"></metric>
<metric name="GRBM_GUI_ACTIVE" block=GRBM event=2 descr="The GUI is Active"></metric>
<metric name="SQ_WAVES" block=SQ event=4 descr="Count number of waves sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_INSTS_VALU" block=SQ event=26 descr="Number of VALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_WR" block=SQ event=27 descr="Number of VMEM write instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_RD" block=SQ event=28 descr="Number of VMEM read instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SALU" block=SQ event=30 descr="Number of SALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SMEM" block=SQ event=31 descr="Number of SMEM instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT" block=SQ event=32 descr="Number of FLAT instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT_LDS_ONLY" block=SQ event=33 descr="Number of FLAT instructions issued that read/wrote only from/to LDS (only works if EARLY_TA_DONE is enabled). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_LDS" block=SQ event=34 descr="Number of LDS instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_GDS" block=SQ event=35 descr="Number of GDS instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_WAIT_INST_LDS" block=SQ event=61 descr="Number of wave-cycles spent waiting for LDS instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"></metric>
<metric name="SQ_ACTIVE_INST_VALU" block=SQ event=69 descr="Number of cycles the SQ instruction arbiter is working on a VALU instruction. (per-simd, nondeterministic). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_SALU" block=SQ event=86 descr="Number of cycles needed to execute non-memory read scalar operations. (per-simd, emulated)"></metric>
<metric name="SQ_THREAD_CYCLES_VALU" block=SQ event=89 descr="Number of thread-cycles used to execute VALU operations (similar to INST_CYCLES_VALU but multiplied by # of active threads). (per-simd)"></metric>
<metric name="SQ_LDS_BANK_CONFLICT" block=SQ event=97 descr="Number of cycles LDS is stalled by bank conflicts. (emulated)"></metric>
<metric name="TA_TA_BUSY" block=TA event=15 descr="TA block is busy. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS" block=TA event=101 descr="Number of flat opcode reads processed by the TA."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS" block=TA event=102 descr="Number of flat opcode writes processed by the TA."></metric>
<metric name="TCC_HIT" block=TCC event=18 descr="Number of cache hits."></metric>
<metric name="TCC_MISS" block=TCC event=19 descr="Number of cache misses. UC reads count as misses."></metric>
<metric name="TCC_MC_RDREQ" block=TCC event=35 descr="Number of 32-byte reads. The hardware actually does 64-byte reads but the number is adjusted to provide uniformity."></metric>
<metric name="TCC_MC_WRREQ" block=TCC event=26 descr="Number of 32-byte transactions going over the TC_MC_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests."></metric>
<metric name="TCC_MC_WRREQ_STALL" block=TCC event=28 descr="Number of cycles a write request was stalled."></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES" block=TCP event=3 descr="TCP stalls TA data interface. Now Windowed."></metric>
</gfx8>
<gfx9>
<metric name="MAX_WAVE_SIZE" expr=wave_front_size descr="Max wave size constant"></metric>
<metric name="SE_NUM" expr=array_count/simd_arrays_per_engine descr="SE_NUM"></metric>
<metric name="SIMD_NUM" expr=simd_per_cu/CU_NUM descr="SIMD Number"></metric>
<metric name="CU_NUM" expr=cu_per_simd_array*array_count descr="CU_NUM"></metric>
<metric name="GRBM_COUNT" block=GRBM event=0 descr="Tie High - Count Number of Clocks"></metric>
<metric name="GRBM_GUI_ACTIVE" block=GRBM event=2 descr="The GUI is Active"></metric>
<metric name="SQ_WAVES" block=SQ event=4 descr="Count number of waves sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_INSTS_VALU" block=SQ event=26 descr="Number of VALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_WR" block=SQ event=27 descr="Number of VMEM write instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_RD" block=SQ event=28 descr="Number of VMEM read instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SALU" block=SQ event=30 descr="Number of SALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SMEM" block=SQ event=31 descr="Number of SMEM instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT" block=SQ event=32 descr="Number of FLAT instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT_LDS_ONLY" block=SQ event=33 descr="Number of FLAT instructions issued that read/wrote only from/to LDS (only works if EARLY_TA_DONE is enabled). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_LDS" block=SQ event=34 descr="Number of LDS instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_GDS" block=SQ event=35 descr="Number of GDS instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_WAIT_INST_LDS" block=SQ event=63 descr="Number of wave-cycles spent waiting for LDS instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"></metric>
<metric name="SQ_ACTIVE_INST_VALU" block=SQ event=71 descr="regspec 71? Number of cycles the SQ instruction arbiter is working on a VALU instruction. (per-simd, nondeterministic). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_SALU" block=SQ event=84 descr="Number of cycles needed to execute non-memory read scalar operations. (per-simd, emulated)"></metric>
<metric name="SQ_THREAD_CYCLES_VALU" block=SQ event=85 descr="Number of thread-cycles used to execute VALU operations (similar to INST_CYCLES_VALU but multiplied by # of active threads). (per-simd)"></metric>
<metric name="SQ_LDS_BANK_CONFLICT" block=SQ event=93 descr="Number of cycles LDS is stalled by bank conflicts. (emulated)"></metric>
<metric name="TA_TA_BUSY" block=TA event=15 descr="TA block is busy. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS" block=TA event=101 descr="Number of flat opcode reads processed by the TA."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS" block=TA event=102 descr="Number of flat opcode writes processed by the TA."></metric>
<metric name="TCC_HIT" block=TCC event=20 descr="Number of cache hits."></metric>
<metric name="TCC_MISS" block=TCC event=22 descr="Number of cache misses. UC reads count as misses."></metric>
<metric name="TCC_EA_WRREQ" block=TCC event=29 descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands."></metric>
<metric name="TCC_EA_WRREQ_64B" block=TCC event=30 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."></metric>
<metric name="TCC_EA_WRREQ_STALL" block=TCC event=33 descr="Number of cycles a write request was stalled."></metric>
<metric name="TCC_EA_RDREQ" block=TCC event=41 descr="Number of TCC/EA read requests (either 32-byte or 64-byte)"></metric>
<metric name="TCC_EA_RDREQ_32B" block=TCC event=42 descr="Number of 32-byte TCC/EA read requests"></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES" block=TCP event=6 descr="TCP stalls TA data interface. Now Windowed."></metric>
</gfx9>
<gfx900 base="gfx9">
</gfx900>
<gfx906 base="gfx9">
# EA1
<metric name="MAX_WAVE_SIZE" expr=wave_front_size descr="Max wave size constant"></metric>
<metric name="SE_NUM" expr=array_count/simd_arrays_per_engine descr="SE_NUM"></metric>
<metric name="SIMD_NUM" expr=simd_per_cu/CU_NUM descr="SIMD Number"></metric>
<metric name="CU_NUM" expr=cu_per_simd_array*array_count descr="CU_NUM"></metric>
<metric name="TCC_EA1_WRREQ" block=TCC event=256 descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands."></metric>
<metric name="TCC_EA1_WRREQ_64B" block=TCC event=257 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."></metric>
<metric name="TCC_EA1_WRREQ_STALL" block=TCC event=260 descr="Number of cycles a write request was stalled."></metric>
<metric name="TCC_EA1_RDREQ" block=TCC event=267 descr="Number of TCC/EA read requests (either 32-byte or 64-byte)"></metric>
<metric name="TCC_EA1_RDREQ_32B" block=TCC event=268 descr="Number of 32-byte TCC/EA read requests"></metric>
</gfx906>
<gfx908 base="gfx9">
<metric name="SQ_INSTS_VMEM_WR" block=SQ event=28 descr="Number of VMEM write instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_RD" block=SQ event=29 descr="Number of VMEM read instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SALU" block=SQ event=31 descr="Number of SALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SMEM" block=SQ event=32 descr="Number of SMEM instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT" block=SQ event=33 descr="Number of FLAT instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT_LDS_ONLY" block=SQ event=34 descr="Number of FLAT instructions issued that read/wrote only from/to LDS (only works if EARLY_TA_DONE is enabled). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_LDS" block=SQ event=35 descr="Number of LDS instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_GDS" block=SQ event=36 descr="Number of GDS instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_WAIT_INST_LDS" block=SQ event=64 descr="Number of wave-cycles spent waiting for LDS instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"></metric>
<metric name="SQ_ACTIVE_INST_VALU" block=SQ event=72 descr="regspec 71? Number of cycles the SQ instruction arbiter is working on a VALU instruction. (per-simd, nondeterministic). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_SALU" block=SQ event=85 descr="Number of cycles needed to execute non-memory read scalar operations. (per-simd, emulated)"></metric>
<metric name="SQ_THREAD_CYCLES_VALU" block=SQ event=86 descr="Number of thread-cycles used to execute VALU operations (similar to INST_CYCLES_VALU but multiplied by # of active threads). (per-simd)"></metric>
<metric name="SQ_LDS_BANK_CONFLICT" block=SQ event=94 descr="Number of cycles LDS is stalled by bank conflicts. (emulated)"></metric>
<metric name="TCC_HIT" block=TCC event=17 descr="Number of cache hits."></metric>
<metric name="TCC_MISS" block=TCC event=19 descr="Number of cache misses. UC reads count as misses."></metric>
<metric name="TCC_EA_WRREQ" block=TCC event=26 descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands."></metric>
<metric name="TCC_EA_WRREQ_64B" block=TCC event=27 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."></metric>
<metric name="TCC_EA_WRREQ_STALL" block=TCC event=30 descr="Number of cycles a write request was stalled."></metric>
<metric name="TCC_EA_RDREQ" block=TCC event=38 descr="Number of TCC/EA read requests (either 32-byte or 64-byte)"></metric>
<metric name="TCC_EA_RDREQ_32B" block=TCC event=39 descr="Number of 32-byte TCC/EA read requests"></metric>
</gfx908>
<gfx90a>
<metric name="MAX_WAVE_SIZE" expr=wave_front_size descr="Max wave size constant"></metric>
<metric name="SE_NUM" expr=array_count/simd_arrays_per_engine descr="SE_NUM"></metric>
<metric name="SIMD_NUM" expr=simd_per_cu/CU_NUM descr="SIMD Number"></metric>
<metric name="CU_NUM" expr=cu_per_simd_array*array_count descr="CU_NUM"></metric>
<metric name="SQ_WAIT_INST_LDS" block=SQ event=91 descr="Number of wave-cycles spent waiting for LDS instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES" block=TCP event=6 descr="TCP stalls TA data interface. Now Windowed."></metric>
<metric name="GRBM_COUNT" block=GRBM event=0 descr="Tie High - Count Number of Clocks"></metric>
<metric name="GRBM_GUI_ACTIVE" block=GRBM event=2 descr="The GUI is Active"></metric>
<metric name="GRBM_CP_BUSY" block=GRBM event=3 descr="Any of the Command Processor (CPG/CPC/CPF) blocks are busy."></metric>
<metric name="GRBM_SPI_BUSY" block=GRBM event=11 descr="Any of the Shader Pipe Interpolators (SPI) are busy in the shader engine(s)."></metric>
<metric name="GRBM_TA_BUSY" block=GRBM event=13 descr="Any of the Texture Pipes (TA) are busy in the shader engine(s)."></metric>
<metric name="GRBM_TC_BUSY" block=GRBM event=28 descr="Any of the Texture Cache Blocks (TCP/TCI/TCA/TCC) are busy."></metric>
<metric name="GRBM_CPC_BUSY" block=GRBM event=30 descr="The Command Processor Compute (CPC) is busy."></metric>
<metric name="GRBM_CPF_BUSY" block=GRBM event=31 descr="The Command Processor Fetchers (CPF) is busy."></metric>
<metric name="GRBM_UTCL2_BUSY" block=GRBM event=34 descr="The Unified Translation Cache Level-2 (UTCL2) block is busy."></metric>
<metric name="GRBM_EA_BUSY" block=GRBM event=35 descr="The Efficiency Arbiter (EA) block is busy."></metric>
<metric name="CPC_ME1_BUSY_FOR_PACKET_DECODE" block=CPC event=13 descr="Me1 busy for packet decode."></metric>
<metric name="CPC_UTCL1_STALL_ON_TRANSLATION" block=CPC event=24 descr="One of the UTCL1s is stalled waiting on translation, XNACK or PENDING response."></metric>
<metric name="CPC_CPC_STAT_BUSY" block=CPC event=25 descr="CPC Busy."></metric>
<metric name="CPC_CPC_STAT_IDLE" block=CPC event=26 descr="CPC Idle."></metric>
<metric name="CPC_CPC_STAT_STALL" block=CPC event=27 descr="CPC Stalled."></metric>
<metric name="CPC_CPC_TCIU_BUSY" block=CPC event=28 descr="CPC TCIU interface Busy."></metric>
<metric name="CPC_CPC_TCIU_IDLE" block=CPC event=29 descr="CPC TCIU interface Idle."></metric>
<metric name="CPC_CPC_UTCL2IU_BUSY" block=CPC event=30 descr="CPC UTCL2 interface Busy."></metric>
<metric name="CPC_CPC_UTCL2IU_IDLE" block=CPC event=31 descr="CPC UTCL2 interface Idle."></metric>
<metric name="CPC_CPC_UTCL2IU_STALL" block=CPC event=32 descr="CPC UTCL2 interface Stalled waiting on Free, Tags or Translation."></metric>
<metric name="CPC_ME1_DC0_SPI_BUSY" block=CPC event=33 descr="CPC Me1 Processor Busy."></metric>
<metric name="CPF_CMP_UTCL1_STALL_ON_TRANSLATION" block=CPF event=20 descr="One of the Compute UTCL1s is stalled waiting on translation, XNACK or PENDING response."></metric>
<metric name="CPF_CPF_STAT_BUSY" block=CPF event=23 descr="CPF Busy."></metric>
<metric name="CPF_CPF_STAT_IDLE" block=CPF event=24 descr="CPF Idle."></metric>
<metric name="CPF_CPF_STAT_STALL" block=CPF event=25 descr="CPF Stalled."></metric>
<metric name="CPF_CPF_TCIU_BUSY" block=CPF event=26 descr="CPF TCIU interface Busy."></metric>
<metric name="CPF_CPF_TCIU_IDLE" block=CPF event=27 descr="CPF TCIU interface Idle."></metric>
<metric name="CPF_CPF_TCIU_STALL" block=CPF event=28 descr="CPF TCIU interface Stalled waiting on Free, Tags."></metric>
<metric name="SPI_CSN_WINDOW_VALID" block=SPI event=47 descr="Clock count enabled by perfcounter_start event. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_CSN_BUSY" block=SPI event=48 descr="Number of clocks with outstanding waves (SPI or SH). Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_CSN_NUM_THREADGROUPS" block=SPI event=49 descr="Number of threadgroups launched. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_CSN_WAVE" block=SPI event=52 descr="Number of waves. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_RA_REQ_NO_ALLOC" block=SPI event=79 descr="Arb cycles with requests but no allocation. Source is RA0"></metric>
<metric name="SPI_RA_REQ_NO_ALLOC_CSN" block=SPI event=85 descr="Arb cycles with CSn req and no CSn alloc. Source is RA0"></metric>
<metric name="SPI_RA_RES_STALL_CSN" block=SPI event=91 descr="Arb cycles with CSn req and no CSn fits. Source is RA0"></metric>
<metric name="SPI_RA_TMP_STALL_CSN" block=SPI event=97 descr="Cycles where csn wants to req but does not fit in temp space."></metric>
<metric name="SPI_RA_WAVE_SIMD_FULL_CSN" block=SPI event=103 descr="Sum of SIMD where WAVE can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_VGPR_SIMD_FULL_CSN" block=SPI event=109 descr="Sum of SIMD where VGPR can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_SGPR_SIMD_FULL_CSN" block=SPI event=115 descr="Sum of SIMD where SGPR can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_LDS_CU_FULL_CSN" block=SPI event=120 descr="Sum of CU where LDS can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_BAR_CU_FULL_CSN" block=SPI event=123 descr="Sum of CU where BARRIER can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_BULKY_CU_FULL_CSN" block=SPI event=125 descr="Sum of CU where BULKY can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_TGLIM_CU_FULL_CSN" block=SPI event=127 descr="Cycles where csn wants to req but all CU are at tg_limit"></metric>
<metric name="SPI_RA_WVLIM_STALL_CSN" block=SPI event=133 descr="Number of clocks csn is stalled due to WAVE LIMIT."></metric>
<metric name="SPI_SWC_CSC_WR" block=SPI event=189 descr="Number of clocks to write CSC waves to SGPRs (need to multiply this value by 4) Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_VWC_CSC_WR" block=SPI event=195 descr="Number of clocks to write CSC waves to VGPRs (need to multiply this value by 4) Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SQ_ACCUM_PREV" block=SQ event=1 descr="For counter N, increment by the value of counter N-1. Only accumulates once every 4 cycles."></metric>
<metric name="SQ_CYCLES" block=SQ event=2 descr="Clock cycles. (nondeterministic, per-simd, global)"></metric>
<metric name="SQ_BUSY_CYCLES" block=SQ event=3 descr="Clock cycles while SQ is reporting that it is busy. (nondeterministic, per-simd, global)"></metric>
<metric name="SQ_WAVES" block=SQ event=4 descr="Count number of waves sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_LEVEL_WAVES" block=SQ event=5 descr="Track the number of waves. Set ACCUM_PREV for the next counter to use this. (level, per-simd, global)"></metric>
<metric name="SQ_WAVES_EQ_64" block=SQ event=6 descr="Count number of waves with exactly 64 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_64" block=SQ event=7 descr="Count number of waves with <64 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_48" block=SQ event=8 descr="Count number of waves with <48 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_32" block=SQ event=9 descr="Count number of waves sent <32 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_16" block=SQ event=10 descr="Count number of waves sent <16 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_BUSY_CU_CYCLES" block=SQ event=13 descr="Count quad-cycles each CU is busy. (nondeterministic, per-simd)"></metric>
<metric name="SQ_ITEMS" block=SQ event=14 descr="Number of valid items per wave. (per-simd, global)"></metric>
<metric name="SQ_INSTS" block=SQ event=25 descr="Number of instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU" block=SQ event=26 descr="Number of VALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_ADD_F16" block=SQ event=27 descr="Number of VALU ADD/SUB instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MUL_F16" block=SQ event=28 descr="Number of VALU MUL instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_FMA_F16" block=SQ event=29 descr="Number of VALU FMA/MAD instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_TRANS_F16" block=SQ event=30 descr="Number of VALU transcendental instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_ADD_F32" block=SQ event=31 descr="Number of VALU ADD/SUB instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MUL_F32" block=SQ event=32 descr="Number of VALU MUL instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_FMA_F32" block=SQ event=33 descr="Number of VALU FMA/MAD instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_TRANS_F32" block=SQ event=34 descr="Number of VALU transcendental instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_ADD_F64" block=SQ event=35 descr="Number of VALU ADD/SUB instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MUL_F64" block=SQ event=36 descr="Number of VALU MUL instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_FMA_F64" block=SQ event=37 descr="Number of VALU FMA/MAD instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_TRANS_F64" block=SQ event=38 descr="Number of VALU transcendental instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_INT32" block=SQ event=39 descr="Number of VALU 32-bit integer (signed or unsigned) instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_INT64" block=SQ event=40 descr="Number of VALU 64-bit integer (signed or unsigned) instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_CVT" block=SQ event=41 descr="Number of VALU data conversion instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_I8" block=SQ event=42 descr="Number of VALU V_MFMA_*_I8 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_F16" block=SQ event=43 descr="Number of VALU V_MFMA_*_F16 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_BF16" block=SQ event=44 descr="Number of VALU V_MFMA_*_BF16 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_F32" block=SQ event=45 descr="Number of VALU V_MFMA_*_F32 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_F64" block=SQ event=46 descr="Number of VALU V_MFMA_*_F64 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_I8" block=SQ event=47 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type I8. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_F16" block=SQ event=48 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type F16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_BF16" block=SQ event=49 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type BF16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_F32" block=SQ event=50 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type F32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_F64" block=SQ event=51 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type F64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_MFMA" block=SQ event=52 descr="Number of MFMA instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_WR" block=SQ event=53 descr="Number of VMEM write instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_RD" block=SQ event=54 descr="Number of VMEM read instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM" block=SQ event=55 descr="Number of VMEM instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SALU" block=SQ event=56 descr="Number of SALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SMEM" block=SQ event=57 descr="Number of SMEM instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT" block=SQ event=58 descr="Number of FLAT instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT_LDS_ONLY" block=SQ event=59 descr="Number of FLAT instructions issued that read/wrote only from/to LDS (only works if EARLY_TA_DONE is enabled). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_LDS" block=SQ event=60 descr="Number of LDS instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_GDS" block=SQ event=61 descr="Number of GDS instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_EXP_GDS" block=SQ event=63 descr="Number of EXP and GDS instructions issued, excluding skipped export instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_BRANCH" block=SQ event=64 descr="Number of Branch instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SENDMSG" block=SQ event=65 descr="Number of Sendmsg instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VSKIPPED" block=SQ event=66 descr="Number of vector instructions skipped. (per-simd, emulated)"></metric>
<metric name="SQ_INST_LEVEL_VMEM" block=SQ event=67 descr="Number of in-flight VMEM instructions. Set next counter to ACCUM_PREV and divide by INSTS_VMEM for average latency. Includes FLAT instructions. (per-simd, level, nondeterministic)"></metric>
<metric name="SQ_INST_LEVEL_SMEM" block=SQ event=68 descr="Number of in-flight SMEM instructions (*2 load/store; *2 atomic; *2 memtime; *4 wb/inv). Set next counter to ACCUM_PREV and divide by INSTS_SMEM for average latency per smem request. Falls slightly short of total request latency because some fetches are divided into two requests that may finish at different times and this counter collects the average latency of the two. (per-simd, level, nondeterministic)"></metric>
<metric name="SQ_INST_LEVEL_LDS" block=SQ event=69 descr="Number of in-flight LDS instructions. Set next counter to ACCUM_PREV and divide by INSTS_LDS for average latency. Includes FLAT instructions. (per-simd, level, nondeterministic)"></metric>
<metric name="SQ_VALU_MFMA_BUSY_CYCLES" block=SQ event=72 descr="Number of cycles the MFMA ALU is busy (per-simd, emulated)"></metric>
<metric name="SQ_WAVE_CYCLES" block=SQ event=74 descr="Number of wave-cycles spent by waves in the CUs (per-simd, nondeterministic). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_WAIT_ANY" block=SQ event=85 descr="Number of wave-cycles spent waiting for anything (per-simd, nondeterministic). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_WAIT_INST_ANY" block=SQ event=88 descr="Number of wave-cycles spent waiting for any instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"></metric>
<metric name="SQ_ACTIVE_INST_ANY" block=SQ event=96 descr="Number of cycles each wave is working on an instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_VMEM" block=SQ event=97 descr="Number of cycles the SQ instruction arbiter is working on a VMEM instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_LDS" block=SQ event=98 descr="Number of cycles the SQ instruction arbiter is working on a LDS instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_VALU" block=SQ event=99 descr="Number of cycles the SQ instruction arbiter is working on a VALU instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_SCA" block=SQ event=100 descr="Number of cycles the SQ instruction arbiter is working on a SALU or SMEM instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_EXP_GDS" block=SQ event=101 descr="Number of cycles the SQ instruction arbiter is working on an EXPORT or GDS instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_MISC" block=SQ event=102 descr="Number of cycles the SQ instruction aribter is working on a BRANCH or SENDMSG instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_FLAT" block=SQ event=103 descr="Number of cycles the SQ instruction arbiter is working on a FLAT instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_VMEM_WR" block=SQ event=104 descr="Number of cycles needed to send addr and cmd data for VMEM write instructions. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_VMEM_RD" block=SQ event=105 descr="Number of cycles needed to send addr and cmd data for VMEM read instructions. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_SMEM" block=SQ event=111 descr="Number of cycles needed to execute scalar memory reads. (per-simd, emulated)"></metric>
<metric name="SQ_INST_CYCLES_SALU" block=SQ event=112 descr="Number of cycles needed to execute non-memory read scalar operations. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_THREAD_CYCLES_VALU" block=SQ event=113 descr="Number of thread-cycles used to execute VALU operations (similar to INST_CYCLES_VALU but multiplied by # of active threads). (per-simd)"></metric>
<metric name="SQ_IFETCH" block=SQ event=115 descr="Number of instruction fetch requests from cache. (per-simd, emulated)"></metric>
<metric name="SQ_IFETCH_LEVEL" block=SQ event=116 descr="Number of instruction fetch requests from cache. (per-simd, level)"></metric>
<metric name="SQ_LDS_BANK_CONFLICT" block=SQ event=121 descr="Number of cycles LDS is stalled by bank conflicts. (emulated)"></metric>
<metric name="SQ_LDS_ADDR_CONFLICT" block=SQ event=122 descr="Number of cycles LDS is stalled by address conflicts. (emulated,nondeterministic)"></metric>
<metric name="SQ_LDS_UNALIGNED_STALL" block=SQ event=123 descr="Number of cycles LDS is stalled processing flat unaligned load/store ops. (emulated)"></metric>
<metric name="SQ_LDS_MEM_VIOLATIONS" block=SQ event=124 descr="Number of threads that have a memory violation in the LDS.(emulated)"></metric>
<metric name="SQ_LDS_ATOMIC_RETURN" block=SQ event=125 descr="Number of atomic return cycles in LDS. (per-simd, emulated)"></metric>
<metric name="SQ_LDS_IDX_ACTIVE" block=SQ event=126 descr="Number of cycles LDS is used for indexed (non-direct,non-interpolation) operations. (per-simd, emulated)"></metric>
<metric name="SQ_ACCUM_PREV_HIRES" block=SQ event=185 descr="For counter N, increment by the value of counter N-1."></metric>
<metric name="SQ_WAVES_RESTORED" block=SQ event=186 descr="Count number of context-restored waves sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_SAVED" block=SQ event=187 descr="Count number of context-saved waves. (per-simd, emulated, global)"></metric>
<metric name="SQ_INSTS_SMEM_NORM" block=SQ event=188 descr="Number of SMEM instructions issued normalized to match smem_level (*2 load/store; *2 atomic; *2 memtime; *4 wb/inv). (per-simd, emulated)"></metric>
<metric name="SQC_DCACHE_INPUT_VALID_READYB" block=SQ event=260 descr="Input stalled by SQC (per-SQ, nondeterministic, unwindowed)"></metric>
<metric name="SQC_TC_REQ" block=SQ event=262 descr="Total number of TC requests that were issued by instruction and constant caches. (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_INST_REQ" block=SQ event=263 descr="Number of insruction requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_DATA_READ_REQ" block=SQ event=264 descr="Number of data read requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_DATA_WRITE_REQ" block=SQ event=265 descr="Number of data write requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_DATA_ATOMIC_REQ" block=SQ event=266 descr="Number of data atomic requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_STALL" block=SQ event=267 descr="Valid request stalled TC request interface (no-credits). (No-Masking, nondeterministic, unwindowed)"></metric>
<metric name="SQC_ICACHE_REQ" block=SQ event=270 descr="Number of requests. (per-SQ, per-Bank)"></metric>
<metric name="SQC_ICACHE_HITS" block=SQ event=271 descr="Number of cache hits. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_ICACHE_MISSES" block=SQ event=272 descr="Number of cache misses, includes uncached requests. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_ICACHE_MISSES_DUPLICATE" block=SQ event=273 descr="Number of misses that were duplicates (access to a non-resident, miss pending CL). (per-SQ, per-Bank, nondeterministic)" ></metric>
<metric name="SQC_DCACHE_REQ" block=SQ event=290 descr="Number of requests (post-bank-serialization). (per-SQ, per-Bank)"></metric>
<metric name="SQC_DCACHE_HITS" block=SQ event=291 descr="Number of cache hits. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_DCACHE_MISSES" block=SQ event=292 descr="Number of cache misses, includes uncached requests. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_DCACHE_MISSES_DUPLICATE" block=SQ event=293 descr="Number of misses that were duplicates (access to a non-resident, miss pending CL). (per-SQ, per-Bank, nondeterministic)" ></metric>
<metric name="SQC_DCACHE_ATOMIC" block=SQ event=298 descr="Number of atomic requests. (per-SQ, per-Bank)"></metric>
<metric name="SQC_DCACHE_REQ_READ_1" block=SQ event=323 descr="Number of constant cache 1 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_2" block=SQ event=324 descr="Number of constant cache 2 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_4" block=SQ event=325 descr="Number of constant cache 4 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_8" block=SQ event=326 descr="Number of constant cache 8 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_16" block=SQ event=327 descr="Number of constant cache 16 dw read requests. (per-SQ)"></metric>
<metric name="TA_TA_BUSY" block=TA event=15 descr="TA block is busy. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_TOTAL_WAVEFRONTS" block=TA event=32 descr="Total number of wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_WAVEFRONTS" block=TA event=44 descr="Number of buffer wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_READ_WAVEFRONTS" block=TA event=45 descr="Number of buffer read wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_WRITE_WAVEFRONTS" block=TA event=46 descr="Number of buffer write wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_ATOMIC_WAVEFRONTS" block=TA event=47 descr="Number of buffer atomic wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_TOTAL_CYCLES" block=TA event=49 descr="Number of buffer cycles issued to TC."></metric>
<metric name="TA_BUFFER_COALESCED_READ_CYCLES" block=TA event=52 descr="Number of buffer coalesced read cycles issued to TC."></metric>
<metric name="TA_BUFFER_COALESCED_WRITE_CYCLES" block=TA event=53 descr="Number of buffer coalesced write cycles issued to TC."></metric>
<metric name="TA_ADDR_STALLED_BY_TC_CYCLES" block=TA event=54 descr="Number of cycles addr path stalled by TC. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_ADDR_STALLED_BY_TD_CYCLES" block=TA event=55 descr="Number of cycles addr path stalled by TD. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_DATA_STALLED_BY_TC_CYCLES" block=TA event=56 descr="Number of cycles data path stalled by TC. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_FLAT_WAVEFRONTS" block=TA event=100 descr="Number of flat opcode wavfronts processed by the TA."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS" block=TA event=101 descr="Number of flat opcode reads processed by the TA."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS" block=TA event=102 descr="Number of flat opcode writes processed by the TA."></metric>
<metric name="TA_FLAT_ATOMIC_WAVEFRONTS" block=TA event=103 descr="Number of flat opcode atomics processed by the TA."></metric>
<metric name="TD_TD_BUSY" block=TD event=1 descr="TD is processing or waiting for data. Perf_Windowing not supported for this counter."></metric>
<metric name="TD_TC_STALL" block=TD event=15 descr="TD is stalled waiting for TC data."></metric>
<metric name="TD_SPI_STALL" block=TD event=18 descr="TD is stalled SPI vinit"></metric>
<metric name="TD_LOAD_WAVEFRONT" block=TD event=25 descr="Count the wavefronts with opcode = load, include atomics and store."></metric>
<metric name="TD_ATOMIC_WAVEFRONT" block=TD event=26 descr="Count the wavefronts with opcode = atomic."></metric>
<metric name="TD_STORE_WAVEFRONT" block=TD event=27 descr="Count the wavefronts with opcode = store."></metric>
<metric name="TD_COALESCABLE_WAVEFRONT" block=TD event=32 descr="Count wavefronts that TA finds coalescable."></metric>
<metric name="TCP_GATE_EN1" block=TCP event=0 descr="TCP interface clocks are turned on. Not Windowed."></metric>
<metric name="TCP_GATE_EN2" block=TCP event=1 descr="TCP core clocks are turned on. Not Windowed."></metric>
<metric name="TCP_TD_TCP_STALL_CYCLES" block=TCP event=7 descr="TD stalls TCP"></metric>
<metric name="TCP_TCR_TCP_STALL_CYCLES" block=TCP event=8 descr="TCR stalls TCP_TCR_req interface"></metric>
<metric name="TCP_READ_TAGCONFLICT_STALL_CYCLES" block=TCP event=11 descr="Tagram conflict stall on a read"></metric>
<metric name="TCP_WRITE_TAGCONFLICT_STALL_CYCLES" block=TCP event=12 descr="Tagram conflict stall on a write"></metric>
<metric name="TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES" block=TCP event=13 descr="Tagram conflict stall on an atomic"></metric>
<metric name="TCP_PENDING_STALL_CYCLES" block=TCP event=22 descr="Stall due to data pending from L2"></metric>
<metric name="TCP_TA_TCP_STATE_READ" block=TCP event=27 descr="Number of state reads"></metric>
<metric name="TCP_VOLATILE" block=TCP event=28 descr="Total number of L1 volatile pixels/buffers from TA"></metric>
<metric name="TCP_TOTAL_ACCESSES" block=TCP event=29 descr="Total number of pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_READ+TCP_PERF_SEL_TOTAL_NONREAD"></metric>
<metric name="TCP_TOTAL_READ" block=TCP event=30 descr="Total number of read pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_HIT_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_EVICT_READ"></metric>
<metric name="TCP_TOTAL_WRITE" block=TCP event=32 descr="Total number of local write pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_MISS_LRU_WRITE+ TCP_PERF_SEL_TOTAL_MISS_EVICT_WRITE"></metric>
<metric name="TCP_TOTAL_ATOMIC_WITH_RET" block=TCP event=38 descr="Total number of atomic with return pixels/buffers from TA"></metric>
<metric name="TCP_TOTAL_ATOMIC_WITHOUT_RET" block=TCP event=39 descr="Total number of atomic without return pixels/buffers from TA"></metric>
<metric name="TCP_TOTAL_WRITEBACK_INVALIDATES" block=TCP event=45 descr="Total number of cache invalidates. Equals TCP_PERF_SEL_TOTAL_WBINVL1+ TCP_PERF_SEL_TOTAL_WBINVL1_VOL+ TCP_PERF_SEL_CP_TCP_INVALIDATE+ TCP_PERF_SEL_SQ_TCP_INVALIDATE_VOL. Not Windowed."></metric>
<metric name="TCP_UTCL1_REQUEST" block=TCP event=47 descr="Total CLIENT_UTCL1 NORMAL requests"></metric>
<metric name="TCP_UTCL1_TRANSLATION_MISS" block=TCP event=48 descr="Total utcl1 translation misses"></metric>
<metric name="TCP_UTCL1_TRANSLATION_HIT" block=TCP event=49 descr="Total utcl1 translation hits"></metric>
<metric name="TCP_UTCL1_PERMISSION_MISS" block=TCP event=50 descr="Total utcl1 permission misses"></metric>
<metric name="TCP_TOTAL_CACHE_ACCESSES" block=TCP event=60 descr="Count of total cache line (tag) accesses (includes hits and misses)."></metric>
<metric name="TCP_TCP_LATENCY" block=TCP event=65 descr="Total TCP wave latency (from first clock of wave entering to first clock of wave leaving), divide by TA_TCP_STATE_READ to avg wave latency"></metric>
<metric name="TCP_TCC_READ_REQ_LATENCY" block=TCP event=66 descr="Total TCP->TCC request latency for reads and atomics with return. Not Windowed."></metric>
<metric name="TCP_TCC_WRITE_REQ_LATENCY" block=TCP event=67 descr="Total TCP->TCC request latency for writes and atomics without return. Not Windowed."></metric>
<metric name="TCP_TCC_READ_REQ" block=TCP event=69 descr="Total read requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_WRITE_REQ" block=TCP event=70 descr="Total write requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_ATOMIC_WITH_RET_REQ" block=TCP event=71 descr="Total atomic with return requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_ATOMIC_WITHOUT_RET_REQ" block=TCP event=72 descr="Total atomic without return requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_NC_READ_REQ" block=TCP event=75 descr="Total read requests with NC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_NC_WRITE_REQ" block=TCP event=76 descr="Total write requests with NC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_NC_ATOMIC_REQ" block=TCP event=77 descr="Total atomic requests with NC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_UC_READ_REQ" block=TCP event=78 descr="Total read requests with UC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_UC_WRITE_REQ" block=TCP event=79 descr="Total write requests with UC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_UC_ATOMIC_REQ" block=TCP event=80 descr="Total atomic requests with UC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_CC_READ_REQ" block=TCP event=81 descr="Total write requests with CC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_CC_WRITE_REQ" block=TCP event=82 descr="Total write requests with CC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_CC_ATOMIC_REQ" block=TCP event=83 descr="Total atomic requests with CC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_RW_READ_REQ" block=TCP event=85 descr="Total write requests with RW mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_RW_WRITE_REQ" block=TCP event=86 descr="Total write requests with RW mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_RW_ATOMIC_REQ" block=TCP event=87 descr="Total atomic requests with RW mtype from this TCP to all TCCs"></metric>
<metric name="TCA_CYCLE" block=TCA event=1 descr="Number of cycles. Not windowable."></metric>
<metric name="TCA_BUSY" block=TCA event=2 descr="Number of cycles we have a request pending. Not windowable."></metric>
<metric name="TCC_CYCLE" block=TCC event=1 descr="Number of cycles. Not windowable."></metric>
<metric name="TCC_BUSY" block=TCC event=2 descr="Number of cycles we have a request pending. Not windowable."></metric>
<metric name="TCC_REQ" block=TCC event=3 descr="Number of requests of all types. This is measured at the tag block. This may be more than the number of requests arriving at the TCC, but it is a good indication of the total amount of work that needs to be performed."></metric>
<metric name="TCC_STREAMING_REQ" block=TCC event=4 descr="Number of streaming requests. This is measured at the tag block."></metric>
<metric name="TCC_NC_REQ" block=TCC event=5 descr="The number of noncoherently cached requests. This is measured at the tag block."></metric>
<metric name="TCC_UC_REQ" block=TCC event=6 descr="The number of uncached requests. This is measured at the tag block."></metric>
<metric name="TCC_CC_REQ" block=TCC event=7 descr="The number of coherently cached requests. This is measured at the tag block."></metric>
<metric name="TCC_RW_REQ" block=TCC event=8 descr="The number of RW requests. This is measured at the tag block."></metric>
<metric name="TCC_PROBE" block=TCC event=9 descr="Number of probe requests. Not windowable."></metric>
<metric name="TCC_PROBE_ALL" block=TCC event=10 descr="Number of external probe requests with with EA_TCC_preq_all== 1. Not windowable."></metric>
<metric name="TCC_READ" block=TCC event=12 descr="Number of read requests. Compressed reads are included in this, but metadata reads are not included."></metric>
<metric name="TCC_WRITE" block=TCC event=13 descr="Number of write requests."></metric>
<metric name="TCC_ATOMIC" block=TCC event=14 descr="Number of atomic requests of all types."></metric>
<metric name="TCC_HIT" block=TCC event=17 descr="Number of cache hits."></metric>
<metric name="TCC_MISS" block=TCC event=19 descr="Number of cache misses. UC reads count as misses."></metric>
<metric name="TCC_WRITEBACK" block=TCC event=22 descr="Number of lines written back to main memory. This includes writebacks of dirty lines and uncached write/atomic requests."></metric>
<metric name="TCC_EA_WRREQ" block=TCC event=26 descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands."></metric>
<metric name="TCC_EA_WRREQ_64B" block=TCC event=27 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."></metric>
<metric name="TCC_EA_WR_UNCACHED_32B" block=TCC event=29 descr="Number of 32-byte write/atomic going over the TC_EA_wrreq interface due to uncached traffic. Note that CC mtypes can produce uncached requests, and those are included in this. A 64-byte request will be counted as 2"></metric>
<metric name="TCC_EA_WRREQ_STALL" block=TCC event=30 descr="Number of cycles a write request was stalled."></metric>
<metric name="TCC_EA_WRREQ_IO_CREDIT_STALL" block=TCC event=31 descr="Number of cycles a EA write request was stalled because the interface was out of IO credits."></metric>
<metric name="TCC_EA_WRREQ_GMI_CREDIT_STALL" block=TCC event=32 descr="Number of cycles a EA write request was stalled because the interface was out of GMI credits."></metric>
<metric name="TCC_EA_WRREQ_DRAM_CREDIT_STALL" block=TCC event=33 descr="Number of cycles a EA write request was stalled because the interface was out of DRAM credits."></metric>
<metric name="TCC_TOO_MANY_EA_WRREQS_STALL" block=TCC event=34 descr="Number of cycles the TCC could not send a EA write request because it already reached its maximum number of pending EA write requests."></metric>
<metric name="TCC_EA_WRREQ_LEVEL" block=TCC event=35 descr="The sum of the number of EA write requests in flight. This is primarily meant for measure average EA write latency. Average write latency = TCC_PERF_SEL_EA_WRREQ_LEVEL/TCC_PERF_SEL_EA_WRREQ."></metric>
<metric name="TCC_EA_ATOMIC" block=TCC event=36 descr="Number of transactions going over the TC_EA_wrreq interface that are actually atomic requests."></metric>
<metric name="TCC_EA_ATOMIC_LEVEL" block=TCC event=37 descr="The sum of the number of EA atomics in flight. This is primarily meant for measure average EA atomic latency. Average atomic latency = TCC_PERF_SEL_EA_WRREQ_ATOMIC_LEVEL/TCC_PERF_SEL_EA_WRREQ_ATOMIC."></metric>
<metric name="TCC_EA_RDREQ" block=TCC event=38 descr="Number of TCC/EA read requests (either 32-byte or 64-byte)"></metric>
<metric name="TCC_EA_RDREQ_32B" block=TCC event=39 descr="Number of 32-byte TCC/EA read requests"></metric>
<metric name="TCC_EA_RD_UNCACHED_32B" block=TCC event=40 descr="Number of 32-byte TCC/EA read due to uncached traffic. A 64-byte request will be counted as 2"></metric>
<metric name="TCC_EA_RDREQ_IO_CREDIT_STALL" block=TCC event=41 descr="Number of cycles there was a stall because the read request interface was out of IO credits. Stalls occur regardless of whether a read needed to be performed or not."></metric>
<metric name="TCC_EA_RDREQ_GMI_CREDIT_STALL" block=TCC event=42 descr="Number of cycles there was a stall because the read request interface was out of GMI credits. Stalls occur regardless of whether a read needed to be performed or not."></metric>
<metric name="TCC_EA_RDREQ_DRAM_CREDIT_STALL" block=TCC event=43 descr="Number of cycles there was a stall because the read request interface was out of DRAM credits. Stalls occur regardless of whether a read needed to be performed or not."></metric>
<metric name="TCC_EA_RDREQ_LEVEL" block=TCC event=44 descr="The sum of the number of TCC/EA read requests in flight. This is primarily meant for measure average EA read latency. Average read latency = TCC_PERF_SEL_EA_RDREQ_LEVEL/TCC_PERF_SEL_EA_RDREQ."></metric>
<metric name="TCC_TAG_STALL" block=TCC event=45 descr="Number of cycles the normal request pipeline in the tag was stalled for any reason. Normally, stalls of this nature are measured exactly from one point the pipeline, but that is not the case for this counter. Probes can stall the pipeline at a variety of places, and there is no single point that can reasonably measure the total stalls accurately."></metric>
<metric name="TCC_NORMAL_WRITEBACK" block=TCC event=68 descr="Number of writebacks due to requests that are not writeback requests."></metric>
<metric name="TCC_ALL_TC_OP_WB_WRITEBACK" block=TCC event=73 descr="Number of writebacks due to all TC_OP writeback requests."></metric>
<metric name="TCC_NORMAL_EVICT" block=TCC event=74 descr="Number of evictions due to requests that are not invalidate or probe requests."></metric>
<metric name="TCC_ALL_TC_OP_INV_EVICT" block=TCC event=80 descr="Number of evictions due to all TC_OP invalidate requests."></metric>
<metric name="TCC_EA_RDREQ_DRAM" block=TCC event=102 descr="Number of TCC/EA read requests (either 32-byte or 64-byte) destined for DRAM (MC)."></metric>
<metric name="TCC_EA_WRREQ_DRAM" block=TCC event=103 descr="Number of TCC/EA write requests (either 32-byte of 64-byte) destined for DRAM (MC)."></metric>
</gfx90a>
<gfx940>
<metric name="MAX_WAVE_SIZE" expr=wave_front_size descr="Max wave size constant"></metric>
<metric name="SE_NUM" expr=array_count/simd_arrays_per_engine descr="SE_NUM"></metric>
<metric name="SIMD_NUM" expr=simd_per_cu/CU_NUM descr="SIMD Number"></metric>
<metric name="CU_NUM" expr=cu_per_simd_array*array_count descr="CU_NUM"></metric>
<metric name="SQ_WAIT_INST_LDS" block=SQ event=96 descr="Number of wave-cycles spent waiting for LDS instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES" block=TCP event=6 descr="TCP stalls TA data interface. Now Windowed."></metric>
<metric name="GRBM_COUNT" block=GRBM event=0 descr="Tie High - Count Number of Clocks"></metric>
<metric name="GRBM_GUI_ACTIVE" block=GRBM event=2 descr="The GUI is Active"></metric>
<metric name="GRBM_CP_BUSY" block=GRBM event=3 descr="Any of the Command Processor (CPG/CPC/CPF) blocks are busy."></metric>
<metric name="GRBM_SPI_BUSY" block=GRBM event=11 descr="Any of the Shader Pipe Interpolators (SPI) are busy in the shader engine(s)."></metric>
<metric name="GRBM_TA_BUSY" block=GRBM event=13 descr="Any of the Texture Pipes (TA) are busy in the shader engine(s)."></metric>
<metric name="GRBM_TC_BUSY" block=GRBM event=28 descr="Any of the Texture Cache Blocks (TCP/TCI/TCA/TCC) are busy."></metric>
<metric name="GRBM_CPC_BUSY" block=GRBM event=30 descr="The Command Processor Compute (CPC) is busy."></metric>
<metric name="GRBM_CPF_BUSY" block=GRBM event=31 descr="The Command Processor Fetchers (CPF) is busy."></metric>
<metric name="GRBM_UTCL2_BUSY" block=GRBM event=34 descr="The Unified Translation Cache Level-2 (UTCL2) block is busy."></metric>
<metric name="GRBM_EA_BUSY" block=GRBM event=35 descr="The Efficiency Arbiter (EA) block is busy."></metric>
<metric name="CPC_ME1_BUSY_FOR_PACKET_DECODE" block=CPC event=13 descr="Me1 busy for packet decode."></metric>
<metric name="CPC_UTCL1_STALL_ON_TRANSLATION" block=CPC event=24 descr="One of the UTCL1s is stalled waiting on translation, XNACK or PENDING response."></metric>
<metric name="CPC_CPC_STAT_BUSY" block=CPC event=25 descr="CPC Busy."></metric>
<metric name="CPC_CPC_STAT_IDLE" block=CPC event=26 descr="CPC Idle."></metric>
<metric name="CPC_CPC_STAT_STALL" block=CPC event=27 descr="CPC Stalled."></metric>
<metric name="CPC_CPC_TCIU_BUSY" block=CPC event=28 descr="CPC TCIU interface Busy."></metric>
<metric name="CPC_CPC_TCIU_IDLE" block=CPC event=29 descr="CPC TCIU interface Idle."></metric>
<metric name="CPC_CPC_UTCL2IU_BUSY" block=CPC event=30 descr="CPC UTCL2 interface Busy."></metric>
<metric name="CPC_CPC_UTCL2IU_IDLE" block=CPC event=31 descr="CPC UTCL2 interface Idle."></metric>
<metric name="CPC_CPC_UTCL2IU_STALL" block=CPC event=32 descr="CPC UTCL2 interface Stalled waiting on Free, Tags or Translation."></metric>
<metric name="CPC_ME1_DC0_SPI_BUSY" block=CPC event=33 descr="CPC Me1 Processor Busy."></metric>
<metric name="CPF_CMP_UTCL1_STALL_ON_TRANSLATION" block=CPF event=20 descr="One of the Compute UTCL1s is stalled waiting on translation, XNACK or PENDING response."></metric>
<metric name="CPF_CPF_STAT_BUSY" block=CPF event=23 descr="CPF Busy."></metric>
<metric name="CPF_CPF_STAT_IDLE" block=CPF event=24 descr="CPF Idle."></metric>
<metric name="CPF_CPF_STAT_STALL" block=CPF event=25 descr="CPF Stalled."></metric>
<metric name="CPF_CPF_TCIU_BUSY" block=CPF event=26 descr="CPF TCIU interface Busy."></metric>
<metric name="CPF_CPF_TCIU_IDLE" block=CPF event=27 descr="CPF TCIU interface Idle."></metric>
<metric name="CPF_CPF_TCIU_STALL" block=CPF event=28 descr="CPF TCIU interface Stalled waiting on Free, Tags."></metric>
<metric name="SPI_CSN_WINDOW_VALID" block=SPI event=47 descr="Clock count enabled by perfcounter_start event. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_CSN_BUSY" block=SPI event=48 descr="Number of clocks with outstanding waves (SPI or SH). Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_CSN_NUM_THREADGROUPS" block=SPI event=49 descr="Number of threadgroups launched. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_CSN_WAVE" block=SPI event=52 descr="Number of waves. Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_RA_REQ_NO_ALLOC" block=SPI event=79 descr="Arb cycles with requests but no allocation. Source is RA0"></metric>
<metric name="SPI_RA_REQ_NO_ALLOC_CSN" block=SPI event=85 descr="Arb cycles with CSn req and no CSn alloc. Source is RA0"></metric>
<metric name="SPI_RA_RES_STALL_CSN" block=SPI event=91 descr="Arb cycles with CSn req and no CSn fits. Source is RA0"></metric>
<metric name="SPI_RA_TMP_STALL_CSN" block=SPI event=97 descr="Cycles where csn wants to req but does not fit in temp space."></metric>
<metric name="SPI_RA_WAVE_SIMD_FULL_CSN" block=SPI event=103 descr="Sum of SIMD where WAVE can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_VGPR_SIMD_FULL_CSN" block=SPI event=109 descr="Sum of SIMD where VGPR can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_SGPR_SIMD_FULL_CSN" block=SPI event=115 descr="Sum of SIMD where SGPR can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_LDS_CU_FULL_CSN" block=SPI event=120 descr="Sum of CU where LDS can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_BAR_CU_FULL_CSN" block=SPI event=123 descr="Sum of CU where BARRIER can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_BULKY_CU_FULL_CSN" block=SPI event=125 descr="Sum of CU where BULKY can't take csn wave when !fits. Source is RA0"></metric>
<metric name="SPI_RA_TGLIM_CU_FULL_CSN" block=SPI event=127 descr="Cycles where csn wants to req but all CU are at tg_limit"></metric>
<metric name="SPI_RA_WVLIM_STALL_CSN" block=SPI event=133 descr="Number of clocks csn is stalled due to WAVE LIMIT."></metric>
<metric name="SPI_SWC_CSC_WR" block=SPI event=189 descr="Number of clocks to write CSC waves to SGPRs (need to multiply this value by 4) Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SPI_VWC_CSC_WR" block=SPI event=195 descr="Number of clocks to write CSC waves to VGPRs (need to multiply this value by 4) Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source, DEBUG_PIPE_SEL = 1, source is CS1; DEBUG_PIPE_SEL = 2, source is CS2; DEBUG_PIPE_SEL = 3, source is CS3; default, source is CS0;"></metric>
<metric name="SQ_ACCUM_PREV" block=SQ event=1 descr="For counter N, increment by the value of counter N-1. Only accumulates once every 4 cycles."></metric>
<metric name="SQ_CYCLES" block=SQ event=2 descr="Clock cycles. (nondeterministic, per-simd, global)"></metric>
<metric name="SQ_BUSY_CYCLES" block=SQ event=3 descr="Clock cycles while SQ is reporting that it is busy. (nondeterministic, per-simd, global)"></metric>
<metric name="SQ_WAVES" block=SQ event=4 descr="Count number of waves sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_LEVEL_WAVES" block=SQ event=5 descr="Track the number of waves. Set ACCUM_PREV for the next counter to use this. (level, per-simd, global)"></metric>
<metric name="SQ_WAVES_EQ_64" block=SQ event=6 descr="Count number of waves with exactly 64 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_64" block=SQ event=7 descr="Count number of waves with <64 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_48" block=SQ event=8 descr="Count number of waves with <48 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_32" block=SQ event=9 descr="Count number of waves sent <32 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_LT_16" block=SQ event=10 descr="Count number of waves sent <16 active threads sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_BUSY_CU_CYCLES" block=SQ event=13 descr="Count quad-cycles each CU is busy. (nondeterministic, per-simd)"></metric>
<metric name="SQ_ITEMS" block=SQ event=14 descr="Number of valid items per wave. (per-simd, global)"></metric>
<metric name="SQ_INSTS" block=SQ event=25 descr="Number of instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU" block=SQ event=26 descr="Number of VALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_ADD_F16" block=SQ event=27 descr="Number of VALU ADD/SUB instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MUL_F16" block=SQ event=28 descr="Number of VALU MUL instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_FMA_F16" block=SQ event=29 descr="Number of VALU FMA/MAD instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_TRANS_F16" block=SQ event=30 descr="Number of VALU transcendental instructions on float16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_ADD_F32" block=SQ event=31 descr="Number of VALU ADD/SUB instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MUL_F32" block=SQ event=32 descr="Number of VALU MUL instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_FMA_F32" block=SQ event=33 descr="Number of VALU FMA/MAD instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_TRANS_F32" block=SQ event=34 descr="Number of VALU transcendental instructions on float32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_ADD_F64" block=SQ event=35 descr="Number of VALU ADD/SUB instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MUL_F64" block=SQ event=36 descr="Number of VALU MUL instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_FMA_F64" block=SQ event=37 descr="Number of VALU FMA/MAD instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_TRANS_F64" block=SQ event=38 descr="Number of VALU transcendental instructions on float64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_INT32" block=SQ event=39 descr="Number of VALU 32-bit integer (signed or unsigned) instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_INT64" block=SQ event=40 descr="Number of VALU 64-bit integer (signed or unsigned) instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_CVT" block=SQ event=41 descr="Number of VALU data conversion instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_I8" block=SQ event=42 descr="Number of VALU V_MFMA_*_I8 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_F16" block=SQ event=43 descr="Number of VALU V_MFMA_*_F16 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_BF16" block=SQ event=44 descr="Number of VALU V_MFMA_*_BF16 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_F32" block=SQ event=45 descr="Number of VALU V_MFMA_*_F32 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_F64" block=SQ event=46 descr="Number of VALU V_MFMA_*_F64 instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_I8" block=SQ event=49 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type I8. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_F16" block=SQ event=50 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type F16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_BF16" block=SQ event=51 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type BF16. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_F32" block=SQ event=52 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type F32. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VALU_MFMA_MOPS_F64" block=SQ event=53 descr="Number of VALU matrix math operations (add or mul) performed dividied by 512, assuming a full EXEC mask, of data type F64. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_MFMA" block=SQ event=56 descr="Number of MFMA instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_WR" block=SQ event=57 descr="Number of VMEM write instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM_RD" block=SQ event=58 descr="Number of VMEM read instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VMEM" block=SQ event=59 descr="Number of VMEM instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SALU" block=SQ event=60 descr="Number of SALU instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SMEM" block=SQ event=61 descr="Number of SMEM instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_FLAT" block=SQ event=62 descr="Number of FLAT instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_LDS" block=SQ event=65 descr="Number of LDS instructions issued (including FLAT). (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_GDS" block=SQ event=66 descr="Number of GDS instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_EXP_GDS" block=SQ event=68 descr="Number of EXP and GDS instructions issued, excluding skipped export instructions. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_BRANCH" block=SQ event=69 descr="Number of Branch instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_SENDMSG" block=SQ event=70 descr="Number of Sendmsg instructions issued. (per-simd, emulated)"></metric>
<metric name="SQ_INSTS_VSKIPPED" block=SQ event=71 descr="Number of vector instructions skipped. (per-simd, emulated)"></metric>
<metric name="SQ_INST_LEVEL_VMEM" block=SQ event=72 descr="Number of in-flight VMEM instructions. Set next counter to ACCUM_PREV and divide by INSTS_VMEM for average latency. Includes FLAT instructions. (per-simd, level, nondeterministic)"></metric>
<metric name="SQ_INST_LEVEL_SMEM" block=SQ event=73 descr="Number of in-flight SMEM instructions (*2 load/store; *2 atomic; *2 memtime; *4 wb/inv). Set next counter to ACCUM_PREV and divide by INSTS_SMEM for average latency per smem request. Falls slightly short of total request latency because some fetches are divided into two requests that may finish at different times and this counter collects the average latency of the two. (per-simd, level, nondeterministic)"></metric>
<metric name="SQ_INST_LEVEL_LDS" block=SQ event=74 descr="Number of in-flight LDS instructions. Set next counter to ACCUM_PREV and divide by INSTS_LDS for average latency. Includes FLAT instructions. (per-simd, level, nondeterministic)"></metric>
<metric name="SQ_VALU_MFMA_BUSY_CYCLES" block=SQ event=77 descr="Number of cycles the MFMA ALU is busy (per-simd, emulated)"></metric>
<metric name="SQ_WAVE_CYCLES" block=SQ event=79 descr="Number of wave-cycles spent by waves in the CUs (per-simd, nondeterministic). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_WAIT_ANY" block=SQ event=90 descr="Number of wave-cycles spent waiting for anything (per-simd, nondeterministic). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_WAIT_INST_ANY" block=SQ event=93 descr="Number of wave-cycles spent waiting for any instruction issue. In units of 4 cycles. (per-simd, nondeterministic)"></metric>
<metric name="SQ_ACTIVE_INST_ANY" block=SQ event=101 descr="Number of cycles each wave is working on an instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_VMEM" block=SQ event=102 descr="Number of cycles the SQ instruction arbiter is working on a VMEM instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_LDS" block=SQ event=103 descr="Number of cycles the SQ instruction arbiter is working on a LDS instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_VALU" block=SQ event=104 descr="Number of cycles the SQ instruction arbiter is working on a VALU instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_SCA" block=SQ event=105 descr="Number of cycles the SQ instruction arbiter is working on a SALU or SMEM instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_EXP_GDS" block=SQ event=106 descr="Number of cycles the SQ instruction arbiter is working on an EXPORT or GDS instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_MISC" block=SQ event=107 descr="Number of cycles the SQ instruction aribter is working on a BRANCH or SENDMSG instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_ACTIVE_INST_FLAT" block=SQ event=108 descr="Number of cycles the SQ instruction arbiter is working on a FLAT instruction. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_VMEM_WR" block=SQ event=109 descr="Number of cycles needed to send addr and cmd data for VMEM write instructions. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_VMEM_RD" block=SQ event=110 descr="Number of cycles needed to send addr and cmd data for VMEM read instructions. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_INST_CYCLES_SMEM" block=SQ event=116 descr="Number of cycles needed to execute scalar memory reads. (per-simd, emulated)"></metric>
<metric name="SQ_INST_CYCLES_SALU" block=SQ event=117 descr="Number of cycles needed to execute non-memory read scalar operations. (per-simd, emulated). Units in quad-cycles(4 cycles)"></metric>
<metric name="SQ_THREAD_CYCLES_VALU" block=SQ event=118 descr="Number of thread-cycles used to execute VALU operations (similar to INST_CYCLES_VALU but multiplied by # of active threads). (per-simd)"></metric>
<metric name="SQ_IFETCH" block=SQ event=120 descr="Number of instruction fetch requests from cache. (per-simd, emulated)"></metric>
<metric name="SQ_IFETCH_LEVEL" block=SQ event=121 descr="Number of instruction fetch requests from cache. (per-simd, level)"></metric>
<metric name="SQ_LDS_BANK_CONFLICT" block=SQ event=126 descr="Number of cycles LDS is stalled by bank conflicts. (emulated)"></metric>
<metric name="SQ_LDS_ADDR_CONFLICT" block=SQ event=127 descr="Number of cycles LDS is stalled by address conflicts. (emulated,nondeterministic)"></metric>
<metric name="SQ_LDS_UNALIGNED_STALL" block=SQ event=128 descr="Number of cycles LDS is stalled processing flat unaligned load/store ops. (emulated)"></metric>
<metric name="SQ_LDS_MEM_VIOLATIONS" block=SQ event=129 descr="Number of threads that have a memory violation in the LDS.(emulated)"></metric>
<metric name="SQ_LDS_ATOMIC_RETURN" block=SQ event=130 descr="Number of atomic return cycles in LDS. (per-simd, emulated)"></metric>
<metric name="SQ_LDS_IDX_ACTIVE" block=SQ event=131 descr="Number of cycles LDS is used for indexed (non-direct,non-interpolation) operations. (per-simd, emulated)"></metric>
<metric name="SQ_ACCUM_PREV_HIRES" block=SQ event=184 descr="For counter N, increment by the value of counter N-1."></metric>
<metric name="SQ_WAVES_RESTORED" block=SQ event=185 descr="Count number of context-restored waves sent to SQs. (per-simd, emulated, global)"></metric>
<metric name="SQ_WAVES_SAVED" block=SQ event=186 descr="Count number of context-saved waves. (per-simd, emulated, global)"></metric>
<metric name="SQ_INSTS_SMEM_NORM" block=SQ event=187 descr="Number of SMEM instructions issued normalized to match smem_level (*2 load/store; *2 atomic; *2 memtime; *4 wb/inv). (per-simd, emulated)"></metric>
<metric name="SQC_ICACHE_INPUT_VALID_READYB" block=SQ event=257 descr=" Input stalled by SQC (per-SQ, nondeterministic, unwindowed)"></metric>
<metric name="SQC_DCACHE_INPUT_VALID_READYB" block=SQ event=260 descr="Input stalled by SQC (per-SQ, nondeterministic, unwindowed)"></metric>
<metric name="SQC_TC_REQ" block=SQ event=262 descr="Total number of TC requests that were issued by instruction and constant caches. (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_INST_REQ" block=SQ event=263 descr="Number of insruction requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_DATA_READ_REQ" block=SQ event=264 descr="Number of data read requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_DATA_WRITE_REQ" block=SQ event=265 descr="Number of data write requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_DATA_ATOMIC_REQ" block=SQ event=266 descr="Number of data atomic requests to the TC (No-Masking, nondeterministic)"></metric>
<metric name="SQC_TC_STALL" block=SQ event=267 descr="Valid request stalled TC request interface (no-credits). (No-Masking, nondeterministic, unwindowed)"></metric>
<metric name="SQC_ICACHE_BUSY_CYCLES" block=SQ event=269 descr="Clock cycles while cache is reporting that it is busy. (No-Masking, nondeterministic, unwindowed)"></metric>
<metric name="SQC_ICACHE_REQ" block=SQ event=270 descr="Number of requests. (per-SQ, per-Bank)"></metric>
<metric name="SQC_ICACHE_HITS" block=SQ event=271 descr="Number of cache hits. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_ICACHE_MISSES" block=SQ event=272 descr="Number of cache misses, includes uncached requests. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_ICACHE_MISSES_DUPLICATE" block=SQ event=273 descr="Number of misses that were duplicates (access to a non-resident, miss pending CL). (per-SQ, per-Bank, nondeterministic)" ></metric>
<metric name="SQC_DCACHE_BUSY_CYCLES" block=SQ event=289 descr=" Clock cycles while cache is reporting that it is busy. (No-Masking, nondeterministic, unwindowed)"></metric>
<metric name="SQC_DCACHE_REQ" block=SQ event=290 descr="Number of requests (post-bank-serialization). (per-SQ, per-Bank)"></metric>
<metric name="SQC_DCACHE_HITS" block=SQ event=291 descr="Number of cache hits. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_DCACHE_MISSES" block=SQ event=292 descr="Number of cache misses, includes uncached requests. (per-SQ, per-Bank, nondeterministic)"></metric>
<metric name="SQC_DCACHE_MISSES_DUPLICATE" block=SQ event=293 descr="Number of misses that were duplicates (access to a non-resident, miss pending CL). (per-SQ, per-Bank, nondeterministic)" ></metric>
<metric name="SQC_DCACHE_ATOMIC" block=SQ event=298 descr="Number of atomic requests. (per-SQ, per-Bank)"></metric>
<metric name="SQC_DCACHE_REQ_READ_1" block=SQ event=323 descr="Number of constant cache 1 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_2" block=SQ event=324 descr="Number of constant cache 2 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_4" block=SQ event=325 descr="Number of constant cache 4 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_8" block=SQ event=326 descr="Number of constant cache 8 dw read requests. (per-SQ)"></metric>
<metric name="SQC_DCACHE_REQ_READ_16" block=SQ event=327 descr="Number of constant cache 16 dw read requests. (per-SQ)"></metric>
<metric name="TA_TA_BUSY" block=TA event=13 descr="TA block is busy. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_TOTAL_WAVEFRONTS" block=TA event=29 descr="Total number of wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_WAVEFRONTS" block=TA event=32 descr="Number of buffer wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_READ_WAVEFRONTS" block=TA event=33 descr="Number of buffer read wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_WRITE_WAVEFRONTS" block=TA event=34 descr="Number of buffer write wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_ATOMIC_WAVEFRONTS" block=TA event=35 descr="Number of buffer atomic wavefronts processed by TA."></metric>
<metric name="TA_BUFFER_TOTAL_CYCLES" block=TA event=37 descr="Number of buffer cycles issued to TC."></metric>
<metric name="TA_BUFFER_COALESCED_READ_CYCLES" block=TA event=40 descr="Number of buffer coalesced read cycles issued to TC."></metric>
<metric name="TA_BUFFER_COALESCED_WRITE_CYCLES" block=TA event=41 descr="Number of buffer coalesced write cycles issued to TC."></metric>
<metric name="TA_ADDR_STALLED_BY_TC_CYCLES" block=TA event=42 descr="Number of cycles addr path stalled by TC. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_ADDR_STALLED_BY_TD_CYCLES" block=TA event=43 descr="Number of cycles addr path stalled by TD. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_DATA_STALLED_BY_TC_CYCLES" block=TA event=44 descr="Number of cycles data path stalled by TC. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_FLAT_WAVEFRONTS" block=TA event=51 descr="Number of flat opcode wavfronts processed by the TA."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS" block=TA event=52 descr="Number of flat opcode reads processed by the TA."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS" block=TA event=53 descr="Number of flat opcode writes processed by the TA."></metric>
<metric name="TA_FLAT_ATOMIC_WAVEFRONTS" block=TA event=54 descr="Number of flat opcode atomics processed by the TA."></metric>
<metric name="TD_TD_BUSY" block=TD event=1 descr="TD is processing or waiting for data. Perf_Windowing not supported for this counter."></metric>
<metric name="TD_TC_STALL" block=TD event=12 descr="TD is stalled waiting for TC data."></metric>
<metric name="TD_SPI_STALL" block=TD event=15 descr="TD is stalled SPI vinit"></metric>
<metric name="TD_LOAD_WAVEFRONT" block=TD event=16 descr="Count the wavefronts with opcode = load, include atomics and store."></metric>
<metric name="TD_ATOMIC_WAVEFRONT" block=TD event=17 descr="Count the wavefronts with opcode = atomic."></metric>
<metric name="TD_STORE_WAVEFRONT" block=TD event=18 descr="Count the wavefronts with opcode = store."></metric>
<metric name="TD_COALESCABLE_WAVEFRONT" block=TD event=21 descr="Count wavefronts that TA finds coalescable."></metric>
<metric name="TCP_GATE_EN1" block=TCP event=0 descr="TCP interface clocks are turned on. Not Windowed."></metric>
<metric name="TCP_GATE_EN2" block=TCP event=1 descr="TCP core clocks are turned on. Not Windowed."></metric>
<metric name="TCP_TD_TCP_STALL_CYCLES" block=TCP event=7 descr="TD stalls TCP"></metric>
<metric name="TCP_TCR_TCP_STALL_CYCLES" block=TCP event=8 descr="TCR stalls TCP_TCR_req interface"></metric>
<metric name="TCP_READ_TAGCONFLICT_STALL_CYCLES" block=TCP event=10 descr="Tagram conflict stall on a read"></metric>
<metric name="TCP_WRITE_TAGCONFLICT_STALL_CYCLES" block=TCP event=11 descr="Tagram conflict stall on a write"></metric>
<metric name="TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES" block=TCP event=12 descr="Tagram conflict stall on an atomic"></metric>
<metric name="TCP_PENDING_STALL_CYCLES" block=TCP event=21 descr="Stall due to data pending from L2"></metric>
<metric name="TCP_TA_TCP_STATE_READ" block=TCP event=25 descr="Number of state reads"></metric>
<metric name="TCP_VOLATILE" block=TCP event=26 descr="Total number of L1 volatile pixels/buffers from TA"></metric>
<metric name="TCP_TOTAL_ACCESSES" block=TCP event=27 descr="Total number of pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_READ+TCP_PERF_SEL_TOTAL_NONREAD"></metric>
<metric name="TCP_TOTAL_READ" block=TCP event=28 descr="Total number of read pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_HIT_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_EVICT_READ"></metric>
<metric name="TCP_TOTAL_WRITE" block=TCP event=30 descr="Total number of local write pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_MISS_LRU_WRITE+ TCP_PERF_SEL_TOTAL_MISS_EVICT_WRITE"></metric>
<metric name="TCP_TOTAL_ATOMIC_WITH_RET" block=TCP event=36 descr="Total number of atomic with return pixels/buffers from TA"></metric>
<metric name="TCP_TOTAL_ATOMIC_WITHOUT_RET" block=TCP event=37 descr="Total number of atomic without return pixels/buffers from TA"></metric>
<metric name="TCP_TOTAL_WRITEBACK_INVALIDATES" block=TCP event=43 descr="Total number of cache invalidates. Equals TCP_PERF_SEL_TOTAL_WBINVL1+ TCP_PERF_SEL_TOTAL_WBINVL1_VOL+ TCP_PERF_SEL_CP_TCP_INVALIDATE+ TCP_PERF_SEL_SQ_TCP_INVALIDATE_VOL. Not Windowed."></metric>
<metric name="TCP_UTCL1_REQUEST" block=TCP event=45 descr="Total CLIENT_UTCL1 NORMAL requests"></metric>
<metric name="TCP_UTCL1_TRANSLATION_MISS" block=TCP event=47 descr="Total utcl1 translation misses"></metric>
<metric name="TCP_UTCL1_TRANSLATION_HIT" block=TCP event=48 descr="Total utcl1 translation hits"></metric>
<metric name="TCP_UTCL1_PERMISSION_MISS" block=TCP event=49 descr="Total utcl1 permission misses"></metric>
<metric name="TCP_TOTAL_CACHE_ACCESSES" block=TCP event=60 descr="Count of total cache line (tag) accesses (includes hits and misses)."></metric>
<metric name="TCP_TCC_READ_REQ" block=TCP event=65 descr="Total read requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_WRITE_REQ" block=TCP event=66 descr="Total write requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_ATOMIC_WITH_RET_REQ" block=TCP event=67 descr="Total atomic with return requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_ATOMIC_WITHOUT_RET_REQ" block=TCP event=68 descr="Total atomic without return requests from TCP to all TCCs"></metric>
<metric name="TCP_TCC_NC_READ_REQ" block=TCP event=71 descr="Total read requests with NC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_NC_WRITE_REQ" block=TCP event=72 descr="Total write requests with NC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_NC_ATOMIC_REQ" block=TCP event=73 descr="Total atomic requests with NC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_UC_READ_REQ" block=TCP event=74 descr="Total read requests with UC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_UC_WRITE_REQ" block=TCP event=75 descr="Total write requests with UC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_UC_ATOMIC_REQ" block=TCP event=76 descr="Total atomic requests with UC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_CC_READ_REQ" block=TCP event=77 descr="Total write requests with CC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_CC_WRITE_REQ" block=TCP event=78 descr="Total write requests with CC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_CC_ATOMIC_REQ" block=TCP event=79 descr="Total atomic requests with CC mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_RW_READ_REQ" block=TCP event=80 descr="Total write requests with RW mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_RW_WRITE_REQ" block=TCP event=81 descr="Total write requests with RW mtype from this TCP to all TCCs"></metric>
<metric name="TCP_TCC_RW_ATOMIC_REQ" block=TCP event=82 descr="Total atomic requests with RW mtype from this TCP to all TCCs"></metric>
<metric name="TCA_CYCLE" block=TCA event=1 descr="Number of cycles. Not windowable."></metric>
<metric name="TCA_BUSY" block=TCA event=2 descr="Number of cycles we have a request pending. Not windowable."></metric>
<metric name="TCC_CYCLE" block=TCC event=1 descr="Number of cycles. Not windowable."></metric>
<metric name="TCC_BUSY" block=TCC event=2 descr="Number of cycles we have a request pending. Not windowable."></metric>
<metric name="TCC_REQ" block=TCC event=3 descr="Number of requests of all types. This is measured at the tag block. This may be more than the number of requests arriving at the TCC, but it is a good indication of the total amount of work that needs to be performed."></metric>
<metric name="TCC_STREAMING_REQ" block=TCC event=4 descr="Number of streaming requests. This is measured at the tag block."></metric>
<metric name="TCC_NC_REQ" block=TCC event=5 descr="The number of noncoherently cached requests. This is measured at the tag block."></metric>
<metric name="TCC_UC_REQ" block=TCC event=6 descr="The number of uncached requests. This is measured at the tag block."></metric>
<metric name="TCC_CC_REQ" block=TCC event=7 descr="The number of coherently cached requests. This is measured at the tag block."></metric>
<metric name="TCC_RW_REQ" block=TCC event=8 descr="The number of RW requests. This is measured at the tag block."></metric>
<metric name="TCC_PROBE" block=TCC event=9 descr="Number of probe requests. Not windowable."></metric>
<metric name="TCC_PROBE_ALL" block=TCC event=10 descr="Number of external probe requests with with EA_TCC_preq_all== 1. Not windowable."></metric>
<metric name="TCC_INTERNAL_PROBE" block=TCC event=11 descr="Number of self-probes spawned by TCC for CC writes/atomic operations. Not windowable."></metric>
<metric name="TCC_READ" block=TCC event=12 descr="Number of read requests. Compressed reads are included in this, but metadata reads are not included."></metric>
<metric name="TCC_WRITE" block=TCC event=13 descr="Number of write requests."></metric>
<metric name="TCC_ATOMIC" block=TCC event=14 descr="Number of atomic requests of all types."></metric>
<metric name="TCC_HIT" block=TCC event=17 descr="Number of cache hits."></metric>
<metric name="TCC_MISS" block=TCC event=19 descr="Number of cache misses. UC reads count as misses."></metric>
<metric name="TCC_WRITEBACK" block=TCC event=22 descr="Number of lines written back to main memory. This includes writebacks of dirty lines and uncached write/atomic requests."></metric>
<metric name="TCC_EA0_WRREQ" block=TCC event=26 descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands."></metric>
<metric name="TCC_EA0_WRREQ_64B" block=TCC event=27 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."></metric>
<metric name="TCC_EA0_WRREQ_PROBE_COMMAND" block=TCC event=28 descr="Number of probe commands going over the TC_EA_wrreq interface."></metric>
<metric name="TCC_EA0_WR_UNCACHED_32B" block=TCC event=29 descr="Number of 32-byte write/atomic going over the TC_EA_wrreq interface due to uncached traffic. Note that CC mtypes can produce uncached requests, and those are included in this. A 64-byte request will be counted as 2"></metric>
<metric name="TCC_EA0_WRREQ_STALL" block=TCC event=30 descr="Number of cycles a write request was stalled."></metric>
<metric name="TCC_EA0_WRREQ_IO_CREDIT_STALL" block=TCC event=31 descr="Number of cycles a EA write request was stalled because the interface was out of IO credits."></metric>
<metric name="TCC_EA0_WRREQ_GMI_CREDIT_STALL" block=TCC event=32 descr="Number of cycles a EA write request was stalled because the interface was out of GMI credits."></metric>
<metric name="TCC_EA0_WRREQ_DRAM_CREDIT_STALL" block=TCC event=33 descr="Number of cycles a EA write request was stalled because the interface was out of DRAM credits."></metric>
<metric name="TCC_TOO_MANY_EA_WRREQS_STALL" block=TCC event=34 descr="Number of cycles the TCC could not send a EA write request because it already reached its maximum number of pending EA write requests."></metric>
<metric name="TCC_EA0_WRREQ_LEVEL" block=TCC event=35 descr="The sum of the number of EA write requests in flight. This is primarily meant for measure average EA write latency. Average write latency = TCC_PERF_SEL_EA_WRREQ_LEVEL/TCC_PERF_SEL_EA_WRREQ."></metric>
<metric name="TCC_EA0_ATOMIC" block=TCC event=36 descr="Number of transactions going over the TC_EA_wrreq interface that are actually atomic requests."></metric>
<metric name="TCC_EA0_ATOMIC_LEVEL" block=TCC event=37 descr="The sum of the number of EA atomics in flight. This is primarily meant for measure average EA atomic latency. Average atomic latency = TCC_PERF_SEL_EA_WRREQ_ATOMIC_LEVEL/TCC_PERF_SEL_EA_WRREQ_ATOMIC."></metric>
<metric name="TCC_EA0_RDREQ" block=TCC event=38 descr="Number of TCC/EA read requests (either 32-byte or 64-byte)"></metric>
<metric name="TCC_EA0_RDREQ_32B" block=TCC event=39 descr="Number of 32-byte TCC/EA read requests"></metric>
<metric name="TCC_EA0_RD_UNCACHED_32B" block=TCC event=40 descr="Number of 32-byte TCC/EA read due to uncached traffic. A 64-byte request will be counted as 2"></metric>
<metric name="TCC_EA0_RDREQ_IO_CREDIT_STALL" block=TCC event=41 descr="Number of cycles there was a stall because the read request interface was out of IO credits. Stalls occur regardless of whether a read needed to be performed or not."></metric>
<metric name="TCC_EA0_RDREQ_GMI_CREDIT_STALL" block=TCC event=42 descr="Number of cycles there was a stall because the read request interface was out of GMI credits. Stalls occur regardless of whether a read needed to be performed or not."></metric>
<metric name="TCC_EA0_RDREQ_DRAM_CREDIT_STALL" block=TCC event=43 descr="Number of cycles there was a stall because the read request interface was out of DRAM credits. Stalls occur regardless of whether a read needed to be performed or not."></metric>
<metric name="TCC_EA0_RDREQ_LEVEL" block=TCC event=44 descr="The sum of the number of TCC/EA read requests in flight. This is primarily meant for measure average EA read latency. Average read latency = TCC_PERF_SEL_EA_RDREQ_LEVEL/TCC_PERF_SEL_EA_RDREQ."></metric>
<metric name="TCC_TAG_STALL" block=TCC event=45 descr="Number of cycles the normal request pipeline in the tag was stalled for any reason. Normally, stalls of this nature are measured exactly from one point the pipeline, but that is not the case for this counter. Probes can stall the pipeline at a variety of places, and there is no single point that can reasonably measure the total stalls accurately."></metric>
<metric name="TCC_NORMAL_WRITEBACK" block=TCC event=68 descr="Number of writebacks due to requests that are not writeback requests."></metric>
<metric name="TCC_ALL_TC_OP_WB_WRITEBACK" block=TCC event=73 descr="Number of writebacks due to all TC_OP writeback requests."></metric>
<metric name="TCC_NORMAL_EVICT" block=TCC event=74 descr="Number of evictions due to requests that are not invalidate or probe requests."></metric>
<metric name="TCC_ALL_TC_OP_INV_EVICT" block=TCC event=80 descr="Number of evictions due to all TC_OP invalidate requests."></metric>
<metric name="TCC_PROBE_EVICT" block=TCC event=81 descr="Number of evictions/invalidations due to probes. Not windowable."></metric>
<metric name="TCC_EA0_RDREQ_DRAM" block=TCC event=102 descr="Number of TCC/EA read requests (either 32-byte or 64-byte) destined for DRAM (MC)."></metric>
<metric name="TCC_EA0_WRREQ_DRAM" block=TCC event=103 descr="Number of TCC/EA write requests (either 32-byte of 64-byte) destined for DRAM (MC)."></metric>
</gfx940>
<gfx941 base="gfx940"></gfx941>
<gfx942 base="gfx940"></gfx942>
<gfx10>
<metric name="MAX_WAVE_SIZE" expr=wave_front_size descr="Max wave size constant"></metric>
<metric name="SE_NUM" expr=array_count/simd_arrays_per_engine descr="SE_NUM"></metric>
<metric name="SIMD_NUM" expr=simd_per_cu/CU_NUM descr="SIMD Number"></metric>
<metric name="CU_NUM" expr=cu_per_simd_array*array_count descr="CU_NUM"></metric>
<metric name="GRBM_COUNT" block=GRBM event=0 descr="Tie High - Count Number of Clocks"></metric>
<metric name="GRBM_GUI_ACTIVE" block=GRBM event=2 descr="The GUI is Active"></metric>
<metric name="GRBM_CP_BUSY" block=GRBM event=3 descr="Any of the Command Processor (CPG/CPC/CPF) blocks are busy."></metric>
<metric name="GRBM_SPI_BUSY" block=GRBM event=11 descr="Any of the Shader Pipe Interpolators (SPI) are busy in the shader engine(s)."></metric>
<metric name="GRBM_TA_BUSY" block=GRBM event=13 descr="Any of the Texture Pipes (TA) are busy in the shader engine(s)."></metric>
<metric name="GRBM_GDS_BUSY" block=GRBM event=25 descr="The Global Data Share (GDS) is busy."></metric>
<metric name="GRBM_EA_BUSY" block=GRBM event=35 descr="The Efficiency Arbiter (EA) block is busy."></metric>
<metric name="GRBM_GL2CC_BUSY" block=GRBM event=40 descr="The GL2CC block is busy."></metric>
<metric name="GL2C_HIT" block=GL2C event=42 descr="Number of cache hits"></metric>
<metric name="GL2C_MISS" block=GL2C event=43 descr="Number of cache misses. UC reads count as misses."></metric>
<metric name="GL2C_MC_WRREQ" block=GL2C event=83 descr="Number of transactions (either 32-byte or 64-byte) going over the GL2C_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands"></metric>
<metric name="GL2C_EA_WRREQ_64B" block=GL2C event=85 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."></metric>
<metric name="GL2C_MC_WRREQ_STALL" block=GL2C event=88 descr="Number of cycles a write request was stalled."></metric>
<metric name="GL2C_MC_RDREQ" block=GL2C event=96 descr="Number of GL2C/EA read requests (either 32-byte or 64-byte or 128-byte)."></metric>
<metric name="GL2C_EA_RDREQ_32B" block=GL2C event=99 descr="Number of 32-byte GL2C/EA read requests"></metric>
<metric name="GL2C_EA_RDREQ_64B" block=GL2C event=100 descr="Number of 64-byte GL2C/EA read requests"></metric>
<metric name="GL2C_EA_RDREQ_96B" block=GL2C event=101 descr="Number of 96-byte GL2C/EA read requests"></metric>
<metric name="GL2C_EA_RDREQ_128B" block=GL2C event=102 descr="Number of 128-byte GL2C/EA read requests"></metric>
<metric name="SQ_ACCUM_PREV" block=SQ event=1 descr="For counter N, increment by the value of counter N-1."></metric>
<metric name="SQ_BUSY_CYCLES" block=SQ event=3 descr="Clock cycles while SQ is reporting that it is busy. {nondeterministic, global, C2}"></metric>
<metric name="SQ_WAVES" block=SQ event=4 descr="Count number of waves sent to SQs. {emulated, global, C1}"></metric>
<metric name="SQ_LEVEL_WAVES" block=SQ event=7 descr="Track the aggregated number of waves over certain period of time, Set next counter to ACCUM_PREV and divide by SQ_PERF_SEL_WAVES for average wave life."></metric>
<metric name="SQ_WAVE_CYCLES" block=SQ event=26 descr="Number of clock cycles spent by waves in the SQs. Incremented by # of living (valid) waves each cycle. {nondeterministic, C1}"></metric>
<metric name="SQ_WAIT_INST_ANY" block=SQ event=28 descr="Number of clock cycles spent waiting for any instruction issue. In units of cycles. {nondeterministic}"></metric>
<metric name="SQ_WAIT_ANY" block=SQ event=37 descr="Number of clock cycles spent waiting for anything. {nondeterministic, C1}"></metric>
<metric name="SQ_INSTS_WAVE32" block=SQ event=71 descr="Number of wave32 instructions issued, for flat, lds, valu, tex. {emulated, C1}"></metric>
<metric name="SQ_INSTS_WAVE32_LDS" block=SQ event=74 descr="Number of wave32 LDS indexed instructions issued. Wave64 may count 1 or 2, depending on what gets issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_WAVE32_VALU" block=SQ event=75 descr="Number of wave32 valu instructions issued. Wave64 may count 1 or 2, depending on what gets issued. {emulated, C1}"></metric>
<metric name="SQ_WAVE32_INSTS" block=SQ event=84 descr="Number of instructions issued by wave32 waves. Skipped instructions are not counted. {emulated}"></metric>
<metric name="SQ_WAVE64_INSTS" block=SQ event=85 descr="Number of instructions issued by wave64 waves. Skipped instructions are not counted. {emulated}"></metric>
<metric name="SQ_INST_LEVEL_GDS" block=SQ event=98 descr="Number of in-flight GDS instructions. Set next counter to ACCUM_PREV and divide by INSTS_GDS for average latency. {level, nondeterministic, C1}"></metric>
<metric name="SQ_INST_LEVEL_LDS" block=SQ event=99 descr="Number of in-flight LDS instructions. Set next counter to ACCUM_PREV and divide by INSTS_LDS for average latency. Includes FLAT instructions. {level, nondeterministic, C1}"></metric>
<metric name="SQ_INST_CYCLES_VMEM" block=SQ event=120 descr="Number of cycles needed to send addr and data for VMEM (lds, buffer, image, flat, scratch, global) instructions, windowed by perf_en. {emulated, C1}"></metric>
<metric name="SQC_LDS_BANK_CONFLICT" block=SQ event=285 descr="Number of cycles LDS is stalled by bank conflicts. (emulated, C1)"></metric>
<metric name="SQC_LDS_IDX_ACTIVE" block=SQ event=290 descr="Number of cycles LDS is used for indexed (non-direct,non-interpolation) operations. {per-simd, emulated, C1}"></metric>
<metric name="SQ_INSTS_VALU" block=SQ event=64 descr="Number of VALU instructions issued excluding skipped instructions. {emulated, C1}"></metric>
<metric name="SQ_INSTS_SALU" block=SQ event=60 descr="Number of SALU instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_SMEM" block=SQ event=61 descr="Number of SMEM instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_FLAT" block=SQ event=57 descr="Number of FLAT instructions issued. {emulated, C2}"></metric>
<metric name="SQ_INSTS_LDS" block=SQ event=59 descr="Number of LDS indexed instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_GDS" block=SQ event=55 descr="Number of GDS instructions issued. {emulated, C1}"></metric>
<metric name="SQ_WAIT_INST_LDS" block=SQ event=31 descr="Number of clock cycles spent waiting for LDS (indexed) instruction issue. In units of cycles. {nondeterministic, C1}"></metric>
<metric name="TA_TA_BUSY" block=TA event=15 descr="TA block is busy. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_FLAT_LOAD_WAVEFRONTS" block=TA event=101 descr=" Number of flat load vec32 packets processed by TA, same as flat_read_wavefronts in earlier IP"></metric>
<metric name="TA_FLAT_STORE_WAVEFRONTS" block=TA event=102 descr="Number of flat store vec32 packets processed by TA, same as flat_write_wavefronts in earlier IP"></metric>
</gfx10>
<gfx1010 base="gfx10">
</gfx1010>
<gfx1030 base="gfx10">
</gfx1030>
<gfx1031 base="gfx10">
</gfx1031>
<gfx1032 base="gfx10">
</gfx1032>
<gfx11>
<metric name="MAX_WAVE_SIZE" expr=wave_front_size descr="Max wave size constant"></metric>
<metric name="SE_NUM" expr=array_count/simd_arrays_per_engine descr="SE_NUM"></metric>
<metric name="SIMD_NUM" expr=simd_per_cu/CU_NUM descr="SIMD Number"></metric>
<metric name="CU_NUM" expr=cu_per_simd_array*array_count descr="CU_NUM"></metric>
<metric name="GRBM_COUNT" block=GRBM event=0 descr="Tie High - Count Number of Clocks"></metric>
<metric name="GRBM_GUI_ACTIVE" block=GRBM event=2 descr="The GUI is Active"></metric>
<metric name="GL2C_HIT" block=GL2C event=42 descr="Number of cache hits"></metric>
<metric name="GL2C_MISS" block=GL2C event=43 descr="Number of cache misses. UC reads count as misses."></metric>
<metric name="GL2C_MC_WRREQ" block=GL2C event=83 descr="Number of transactions (either 32-byte or 64-byte) going over the GL2C_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands"></metric>
<metric name="GL2C_EA_WRREQ_64B" block=GL2C event=85 descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface."></metric>
<metric name="GL2C_MC_WRREQ_STALL" block=GL2C event=88 descr="Number of cycles a write request was stalled."></metric>
<metric name="GL2C_MC_RDREQ" block=GL2C event=96 descr="Number of GL2C/EA read requests (either 32-byte or 64-byte or 128-byte)."></metric>
<metric name="GL2C_EA_RDREQ_32B" block=GL2C event=99 descr="Number of 32-byte GL2C/EA read requests"></metric>
<metric name="GL2C_EA_RDREQ_64B" block=GL2C event=100 descr="Number of 64-byte GL2C/EA read requests"></metric>
<metric name="GL2C_EA_RDREQ_96B" block=GL2C event=101 descr="Number of 96-byte GL2C/EA read requests"></metric>
<metric name="GL2C_EA_RDREQ_128B" block=GL2C event=102 descr="Number of 128-byte GL2C/EA read requests"></metric>
<metric name="SQ_ACCUM_PREV" block=SQ event=1 descr="For counter N, increment by the value of counter N-1."></metric>
<metric name="SQ_BUSY_CYCLES" block=SQ event=3 descr="Clock cycles while SQ is reporting that it is busy. {nondeterministic, global, C2}"></metric>
<metric name="SQ_WAVES" block=SQ event=4 descr="Count number of waves sent to SQs. {emulated, global, C1}"></metric>
<metric name="SQ_WAVE_CYCLES" block=SQ event=24 descr="Number of clock cycles spent by waves in the SQs. Incremented by number of living (valid) waves each cycle. {nondeterministic, C1}"></metric>
<metric name="SQ_WAIT_INST_ANY" block=SQ event=26 descr="Number of clock-cycles spent waiting for any instruction issue. In units of cycles. (nondeterministic)"></metric>
<metric name="SQ_WAIT_ANY" block=SQ event=35 descr="Number of wave-cycles spent waiting for anything (nondeterministic, C1)"></metric>
<metric name="SQ_INSTS_WAVE32" block=SQ event=70 descr="Number of wave32 instructions issued, for flat, lds, valu, tex. {emulated, C1}"></metric>
<metric name="SQ_INSTS_WAVE32_LDS" block=SQ event=72 descr="Number of wave32 LDS indexed instructions issued. Wave64 may count 1 or 2, depending on what gets issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_WAVE32_VALU" block=SQ event=73 descr="Number of wave32 valu instructions issued. Wave64 may count 1 or 2, depending on what gets issued. {emulated, C1}"></metric>
<metric name="SQ_WAVE32_INSTS" block=SQ event=82 descr="Number of instructions issued by wave32 waves. Skipped instructions are not counted. {emulated}"></metric>
<metric name="SQ_WAVE64_INSTS" block=SQ event=83 descr="Number of instructions issued by wave64 waves. Skipped instructions are not counted. {emulated}"></metric>
<metric name="SQ_INST_LEVEL_GDS" block=SQ event=87 descr="Number of in-flight GDS instructions. Set next counter to ACCUM_PREV and divide by INSTS_GDS for average latency. {level, nondeterministic, C1}"></metric>
<metric name="SQ_INST_LEVEL_LDS" block=SQ event=88 descr="Number of in-flight LDS instructions. Set next counter to ACCUM_PREV and divide by INSTS_LDS for average latency. Includes FLAT instructions. {level, nondeterministic, C1}"></metric>
<metric name="SQ_INST_CYCLES_VMEM" block=SQ event=106 descr="Number of cycles needed to send addr and data for VMEM (lds, buffer, image, flat, scratch, global) instructions, windowed by perf_en. {emulated, C1}"></metric>
<metric name="SQC_LDS_BANK_CONFLICT" block=SQ event=256 descr="Number of cycles LDS is stalled by bank conflicts. (emulated, C1)"></metric>
<metric name="SQC_LDS_IDX_ACTIVE" block=SQ event=261 descr="Number of cycles LDS is used for indexed (non-direct,non-interpolation) operations. {per-simd, emulated, C1}"></metric>
<metric name="SQ_INSTS_VALU" block=SQ event=62 descr="Number of VALU instructions issued excluding skipped instructions. {emulated, C1}"></metric>
<metric name="SQ_INSTS_SALU" block=SQ event=58 descr="Number of SALU instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_SMEM" block=SQ event=59 descr="Number of SMEM instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_FLAT" block=SQ event=56 descr="Number of FLAT instructions issued. {emulated, C2}"></metric>
<metric name="SQ_INSTS_LDS" block=SQ event=57 descr="Number of LDS indexed instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_GDS" block=SQ event=54 descr="Number of GDS instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_TEX_LOAD" block=SQ event=66 descr="Number of buffer load, image load, sample, or atomic (with return) instructions issued. {emulated, C1}"></metric>
<metric name="SQ_INSTS_TEX_STORE" block=SQ event=67 descr="Number of buffer store, image store, or atomic (without return) instructions issued. {emulated, C1}"></metric>
<metric name="SQ_WAIT_INST_LDS" block=SQ event=29 descr="Number of clock cycles spent waiting for LDS (indexed) instruction issue. In units of cycles. {nondeterministic, C1}"></metric>
<metric name="TA_TA_BUSY" block=TA event=15 descr="TA block is busy. Perf_Windowing not supported for this counter."></metric>
<metric name="TA_BUFFER_LOAD_WAVEFRONTS" block=TA event=45 descr="Number of buffer load vec32 packets processed by TA"></metric>
<metric name="TA_BUFFER_STORE_WAVEFRONTS" block=TA event=46 descr="Number of buffer store vec32 packets processed by TA"></metric>
</gfx11>
<gfx1100 base="gfx11">
</gfx1100>
<gfx1101 base="gfx11">
</gfx1101>
<gfx1102 base="gfx11">
</gfx1102>
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/convert-counters-collection-format.py
Executable
+224
Vedi File
@@ -0,0 +1,224 @@
#!/usr/bin/env python3
# MIT License
#
# Copyright (c) 2023-2025 Advanced Micro Devices, Inc. All rights reserved.
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import os
import pandas as pd
import argparse
import logging
def read_csv(file_path):
df = pd.DataFrame()
try:
df = pd.read_csv(file_path)
except Exception as e:
logging.info(f"Error reading {file_path}: {e}")
raise
return df
def get_counter_collection_files(root_path):
file_paths = []
for root, _, files in os.walk(root_path):
if "pmc_" in root:
for file in files:
if file.endswith("counter_collection.csv"):
file_path = os.path.join(root, file)
file_paths.append(file_path)
return file_paths
def get_combined_df(args):
files_list = []
for input in args.input:
if os.path.isfile(input):
files_list.append(input)
elif os.path.isdir(input):
files_list.extend(get_counter_collection_files(input))
if not files_list:
raise ValueError("Valid Input files not found")
logging.info(f"Processing files: {files_list}")
combined_df = pd.DataFrame()
for file in files_list:
combined_df = pd.concat([combined_df, read_csv(file)], ignore_index=True)
return combined_df
def write_to_file(df, args):
logging.info(f"Saving output file to : {args.output}")
directory, file_path = os.path.split(args.output)
if directory:
os.makedirs(directory, exist_ok=True)
if not args.retain_agent_prefix:
df["Agent_Id"] = df["Agent_Id"].apply(lambda x: x.split(" ")[-1])
df.to_csv(args.output, index=False)
def main(args):
logging.basicConfig(level=args.loglevel)
input_df = get_combined_df(args)
# Validate
columns = [
"Correlation_Id",
"Dispatch_Id",
"Agent_Id",
"Queue_Id",
"Process_Id",
"Thread_Id",
"Grid_Size",
"Kernel_Id",
"Kernel_Name",
"Workgroup_Size",
"LDS_Block_Size",
"Scratch_Size",
"VGPR_Count",
"SGPR_Count",
"Counter_Name",
"Counter_Value",
"Start_Timestamp",
"End_Timestamp",
]
for col in input_df.columns:
if col not in columns:
logging.debug(f"Unexpected column {col} found in rocprofv3 input file")
# Convert
indexes = [
"Dispatch_Id",
"Agent_Id",
"Grid_Size",
"Kernel_Name",
"LDS_Block_Size",
"Queue_Id",
"SGPR_Count",
"Scratch_Size",
"VGPR_Count",
"Workgroup_Size",
]
# Drop duplicate counters in multiple PMC lines
input_df.drop_duplicates(
subset=indexes + ["Counter_Name"], keep="first", inplace=True
)
pivoted_data = input_df.pivot_table(
index=indexes, columns="Counter_Name", values="Counter_Value", aggfunc="sum"
).reset_index()
# Save
write_to_file(pivoted_data, args)
def strtobool(val):
"""Convert a string representation of truth to true or false.
True values are 'y', 'yes', 't', 'true', 'on', and '1'; false values
are 'n', 'no', 'f', 'false', 'off', and '0'. Raises ValueError if
'val' is anything else.
"""
if isinstance(val, (list, tuple)):
if len(val) > 1:
val_type = type(val).__name__
raise ValueError(f"invalid truth value {val} (type={val_type})")
else:
val = val[0]
if isinstance(val, bool):
return val
elif isinstance(val, str) and val.lower() in ("y", "yes", "t", "true", "on", "1"):
return True
elif isinstance(val, str) and val.lower() in ("n", "no", "f", "false", "off", "0"):
return False
else:
val_type = type(val).__name__
raise ValueError(f"invalid truth value {val} (type={val_type})")
class booleanArgAction(argparse.Action):
def __call__(self, parser, args, value, option_string=None):
setattr(args, self.dest, strtobool(value))
def add_parser_bool_argument(gparser, *args, **kwargs):
gparser.add_argument(
*args,
**kwargs,
action=booleanArgAction,
nargs="?",
const=True,
type=str,
required=False,
metavar="BOOL",
default=False,
)
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"-i",
"--input",
help="Rocprofv3 Counter Collection input files and/or directories containing `*counter_collection.csv` files",
nargs="+",
default=[],
required=True,
)
parser.add_argument(
"-o",
"--output",
help="Rocprofv1 formatted output file name",
default=None,
type=str,
required=True,
)
parser.add_argument(
"-d",
"--debug",
help="Debug Logs",
action="store_const",
dest="loglevel",
const=logging.DEBUG,
default=logging.WARNING,
)
parser.add_argument(
"-v",
"--verbose",
help="Verbose Logs",
action="store_const",
dest="loglevel",
const=logging.INFO,
)
advanced_options = parser.add_argument_group("Advanced options")
add_parser_bool_argument(
advanced_options,
"--retain-agent-prefix",
help="retains the agent prefix",
)
return parser.parse_args()
if __name__ == "__main__":
main(parse_args())
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/counter_defs.yaml
+10085
Vedi File
File diff soppresso perché troppo grande Carica Diff
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/derived_counters.xml
+581
Vedi File
@@ -0,0 +1,581 @@
<common_derived>
# GPUBusy The percentage of time GPU was busy.
<metric
name="GPUBusy"
descr="The percentage of time GPU was busy."
expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT
></metric>
# Wavefronts Total wavefronts.
<metric
name="Wavefronts"
descr="Total wavefronts."
expr=SQ_WAVES
></metric>
# VALUInsts The average number of vector ALU instructions executed per work-item (affected by flow control).
<metric
name="VALUInsts"
descr="The average number of vector ALU instructions executed per work-item (affected by flow control)."
expr=SQ_INSTS_VALU/SQ_WAVES
></metric>
# SALUInsts The average number of scalar ALU instructions executed per work-item (affected by flow control).
<metric
name="SALUInsts"
descr="The average number of scalar ALU instructions executed per work-item (affected by flow control)."
expr=SQ_INSTS_SALU/SQ_WAVES
></metric>
# SFetchInsts The average number of scalar fetch instructions from the video memory executed per work-item (affected by flow control).
<metric
name="SFetchInsts"
descr="The average number of scalar fetch instructions from the video memory executed per work-item (affected by flow control)."
expr=SQ_INSTS_SMEM/SQ_WAVES
></metric>
# GDSInsts The average number of GDS read or GDS write instructions executed per work item (affected by flow control).
<metric
name="GDSInsts"
descr="The average number of GDS read or GDS write instructions executed per work item (affected by flow control)."
expr=SQ_INSTS_GDS/SQ_WAVES
></metric>
# MemUnitBusy The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound).
<metric
name="MemUnitBusy"
descr="The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound)."
expr=100*reduce(TA_TA_BUSY,max)/GRBM_GUI_ACTIVE/SE_NUM
></metric>
# ALUStalledByLDS The percentage of GPUTime ALU units are stalled by the LDS input queue being full or the output queue being not ready. If there are LDS bank conflicts, reduce them. Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% (optimal) to 100% (bad).
<metric
name="ALUStalledByLDS"
descr="The percentage of GPUTime ALU units are stalled by the LDS input queue being full or the output queue being not ready. If there are LDS bank conflicts, reduce them. Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% (optimal) to 100% (bad)."
expr=400*SQ_WAIT_INST_LDS/SQ_WAVES/GRBM_GUI_ACTIVE
></metric>
</common_derived>
<gfx8 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
<metric name="TA_BUSY_min" expr=reduce(TA_TA_BUSY,min) descr="TA block is busy. Min over TA instances."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS_sum" expr=reduce(TA_FLAT_READ_WAVEFRONTS,sum) descr="Number of flat opcode reads processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS_sum" expr=reduce(TA_FLAT_WRITE_WAVEFRONTS,sum) descr="Number of flat opcode writes processed by the TA. Sum over TA instances."></metric>
<metric name="TCC_HIT_sum" expr=reduce(TCC_HIT,sum) descr="Number of cache hits. Sum over TCC instances."></metric>
<metric name="TCC_MISS_sum" expr=reduce(TCC_MISS,sum) descr="Number of cache misses. Sum over TCC instances."></metric>
<metric name="TCC_MC_RDREQ_sum" expr=reduce(TCC_MC_RDREQ,sum) descr="Number of 32-byte reads. Sum over TCC instaces."></metric>
<metric name="TCC_MC_WRREQ_sum" expr=reduce(TCC_MC_WRREQ,sum) descr="Number of 32-byte transactions going over the TC_MC_wrreq interface. Sum over TCC instaces."></metric>
<metric name="TCC_WRREQ_STALL_max" expr=reduce(TCC_MC_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
<metric name="FETCH_SIZE" expr=(TCC_MC_RDREQ_sum*32)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_SIZE" expr=(TCC_MC_WRREQ_sum*32)/1024 descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_REQ_32B" expr=TCC_MC_WRREQ_sum descr="The total number of 32-byte effective memory writes."></metric>
<metric name="VFetchInsts" expr=(SQ_INSTS_VMEM_RD-TA_FLAT_READ_WAVEFRONTS_sum)/SQ_WAVES descr="The average number of vector fetch instructions from the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that fetch from video memory."></metric>
<metric name="VWriteInsts" expr=(SQ_INSTS_VMEM_WR-TA_FLAT_WRITE_WAVEFRONTS_sum)/SQ_WAVES descr="The average number of vector write instructions to the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that write to video memory."></metric>
<metric name="FlatVMemInsts" expr=(SQ_INSTS_FLAT-SQ_INSTS_FLAT_LDS_ONLY)/SQ_WAVES descr="The average number of FLAT instructions that read from or write to the video memory executed per work item (affected by flow control). Includes FLAT instructions that read from or write to scratch."></metric>
<metric name="LDSInsts" expr=(SQ_INSTS_LDS-SQ_INSTS_FLAT_LDS_ONLY)/SQ_WAVES descr="The average number of LDS read or LDS write instructions executed per work item (affected by flow control). Excludes FLAT instructions that read from or write to LDS."></metric>
<metric name="FlatLDSInsts" expr=SQ_INSTS_FLAT_LDS_ONLY/SQ_WAVES descr="The average number of FLAT instructions that read or write to LDS executed per work item (affected by flow control)."></metric>
<metric name="VALUUtilization" expr=100*SQ_THREAD_CYCLES_VALU/(SQ_ACTIVE_INST_VALU*MAX_WAVE_SIZE) descr="The percentage of active vector ALU threads in a wave. A lower number can mean either more thread divergence in a wave or that the work-group size is not a multiple of 64. Value range: 0% (bad), 100% (ideal - no thread divergence)."></metric>
<metric name="VALUBusy" expr=100*SQ_ACTIVE_INST_VALU*4/SIMD_NUM/GRBM_GUI_ACTIVE descr="The percentage of GPUTime vector ALU instructions are processed. Value range: 0% (bad) to 100% (optimal)."></metric>
<metric name="SALUBusy" expr=100*SQ_INST_CYCLES_SALU*4/SIMD_NUM/GRBM_GUI_ACTIVE descr="The percentage of GPUTime scalar ALU instructions are processed. Value range: 0% (bad) to 100% (optimal)."></metric>
<metric name="FetchSize" expr=FETCH_SIZE descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WriteSize" expr=WRITE_SIZE descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="MemWrites32B" expr=WRITE_REQ_32B descr="The total number of effective 32B write transactions to the memory"></metric>
<metric name="L2CacheHit" expr=100*reduce(TCC_HIT,sum)/(reduce(TCC_HIT,sum)+reduce(TCC_MISS,sum)) descr="The percentage of fetch, write, atomic, and other instructions that hit the data in L2 cache. Value range: 0% (no hit) to 100% (optimal)."></metric>
<metric name="MemUnitStalled" expr=100*reduce(TCP_TCP_TA_DATA_STALL_CYCLES,max)/GRBM_GUI_ACTIVE/SE_NUM descr="The percentage of GPUTime the memory unit is stalled. Try reducing the number or size of fetches and writes if possible. Value range: 0% (optimal) to 100% (bad)."></metric>
<metric name="WriteUnitStalled" expr=100*TCC_WRREQ_STALL_max/GRBM_GUI_ACTIVE descr="The percentage of GPUTime the Write unit is stalled. Value range: 0% to 100% (bad)."></metric>
# LDSBankConflict The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad).
<metric name="LDSBankConflict" expr=100*SQ_LDS_BANK_CONFLICT/GRBM_GUI_ACTIVE/CU_NUM descr="The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad)."></metric>
</gfx8>
<gfx9 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
<metric name="TA_BUSY_min" expr=reduce(TA_TA_BUSY,min) descr="TA block is busy. Min over TA instances."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS_sum" expr=reduce(TA_FLAT_READ_WAVEFRONTS,sum) descr="Number of flat opcode reads processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS_sum" expr=reduce(TA_FLAT_WRITE_WAVEFRONTS,sum) descr="Number of flat opcode writes processed by the TA. Sum over TA instances."></metric>
<metric name="TCC_HIT_sum" expr=reduce(TCC_HIT,sum) descr="Number of cache hits. Sum over TCC instances."></metric>
<metric name="TCC_MISS_sum" expr=reduce(TCC_MISS,sum) descr="Number of cache misses. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_32B_sum" expr=reduce(TCC_EA_RDREQ_32B,sum) descr="Number of 32-byte TCC/EA read requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_sum" expr=reduce(TCC_EA_RDREQ,sum) descr="Number of TCC/EA read requests (either 32-byte or 64-byte). Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_sum" expr=reduce(TCC_EA_WRREQ,sum) descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_64B_sum" expr=reduce(TCC_EA_WRREQ_64B,sum) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_WRREQ_STALL_max" expr=reduce(TCC_EA_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
<metric name="GPU_UTIL" expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Percentage of the time that GUI is active"></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES_sum" expr=reduce(TCP_TCP_TA_DATA_STALL_CYCLES,sum) descr="Total number of TCP stalls TA data interface."></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES_max" expr=reduce(TCP_TCP_TA_DATA_STALL_CYCLES,max) descr="Maximum number of TCP stalls TA data interface."></metric>
<metric name="FETCH_SIZE" expr=(TCC_EA_RDREQ_32B_sum*32+(TCC_EA_RDREQ_sum-TCC_EA_RDREQ_32B_sum)*64)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_SIZE" expr=((TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum)*32+TCC_EA_WRREQ_64B_sum*64)/1024 descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_REQ_32B" expr=TCC_EA_WRREQ_64B_sum*2+(TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum) descr="The total number of 32-byte effective memory writes."></metric>
<metric name="VFetchInsts" expr=(SQ_INSTS_VMEM_RD-TA_FLAT_READ_WAVEFRONTS_sum)/SQ_WAVES descr="The average number of vector fetch instructions from the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that fetch from video memory."></metric>
<metric name="VWriteInsts" expr=(SQ_INSTS_VMEM_WR-TA_FLAT_WRITE_WAVEFRONTS_sum)/SQ_WAVES descr="The average number of vector write instructions to the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that write to video memory."></metric>
<metric name="FlatVMemInsts" expr=(SQ_INSTS_FLAT-SQ_INSTS_FLAT_LDS_ONLY)/SQ_WAVES descr="The average number of FLAT instructions that read from or write to the video memory executed per work item (affected by flow control). Includes FLAT instructions that read from or write to scratch."></metric>
<metric name="LDSInsts" expr=(SQ_INSTS_LDS-SQ_INSTS_FLAT_LDS_ONLY)/SQ_WAVES descr="The average number of LDS read or LDS write instructions executed per work item (affected by flow control). Excludes FLAT instructions that read from or write to LDS."></metric>
<metric name="FlatLDSInsts" expr=SQ_INSTS_FLAT_LDS_ONLY/SQ_WAVES descr="The average number of FLAT instructions that read or write to LDS executed per work item (affected by flow control)."></metric>
<metric name="VALUUtilization" expr=100*SQ_THREAD_CYCLES_VALU/(SQ_ACTIVE_INST_VALU*MAX_WAVE_SIZE) descr="The percentage of active vector ALU threads in a wave. A lower number can mean either more thread divergence in a wave or that the work-group size is not a multiple of 64. Value range: 0% (bad), 100% (ideal - no thread divergence)."></metric>
<metric name="VALUBusy" expr=100*SQ_ACTIVE_INST_VALU*4/SIMD_NUM/GRBM_GUI_ACTIVE descr="The percentage of GPUTime vector ALU instructions are processed. Value range: 0% (bad) to 100% (optimal)."></metric>
<metric name="SALUBusy" expr=100*SQ_INST_CYCLES_SALU*4/SIMD_NUM/GRBM_GUI_ACTIVE descr="The percentage of GPUTime scalar ALU instructions are processed. Value range: 0% (bad) to 100% (optimal)."></metric>
<metric name="FetchSize" expr=FETCH_SIZE descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WriteSize" expr=WRITE_SIZE descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="MemWrites32B" expr=WRITE_REQ_32B descr="The total number of effective 32B write transactions to the memory"></metric>
<metric name="L2CacheHit" expr=100*reduce(TCC_HIT,sum)/(reduce(TCC_HIT,sum)+reduce(TCC_MISS,sum)) descr="The percentage of fetch, write, atomic, and other instructions that hit the data in L2 cache. Value range: 0% (no hit) to 100% (optimal)."></metric>
<metric name="MemUnitStalled" expr=100*TCP_TCP_TA_DATA_STALL_CYCLES_max/GRBM_GUI_ACTIVE/SE_NUM descr="The percentage of GPUTime the memory unit is stalled. Try reducing the number or size of fetches and writes if possible. Value range: 0% (optimal) to 100% (bad)."></metric>
<metric name="WriteUnitStalled" expr=100*TCC_WRREQ_STALL_max/GRBM_GUI_ACTIVE descr="The percentage of GPUTime the Write unit is stalled. Value range: 0% to 100% (bad)."></metric>
# LDSBankConflict The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad).
<metric name="LDSBankConflict" expr=100*SQ_LDS_BANK_CONFLICT/GRBM_GUI_ACTIVE/CU_NUM descr="The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad)."></metric>
</gfx9>
<gfx900 base="gfx9">
</gfx900>
<gfx906 base="gfx9">
# EA1
<metric name="TCC_EA1_RDREQ_32B_sum" expr=reduce(TCC_EA1_RDREQ_32B,sum) descr="Number of 32-byte TCC/EA read requests. Sum over TCC EA1s."></metric>
<metric name="TCC_EA1_RDREQ_sum" expr=reduce(TCC_EA1_RDREQ,sum) descr="Number of TCC/EA read requests (either 32-byte or 64-byte). Sum over TCC EA1s."></metric>
<metric name="TCC_EA1_WRREQ_sum" expr=reduce(TCC_EA1_WRREQ,sum) descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Sum over TCC EA1s."></metric>
<metric name="TCC_EA1_WRREQ_64B_sum" expr=reduce(TCC_EA1_WRREQ_64B,sum) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC EA1s."></metric>
<metric name="TCC_WRREQ1_STALL_max" expr=reduce(TCC_EA1_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
<metric name="RDATA1_SIZE" expr=(TCC_EA1_RDREQ_32B_sum*32+(TCC_EA1_RDREQ_sum-TCC_EA1_RDREQ_32B_sum)*64) descr="The total kilobytes fetched from the video memory. This is measured on EA1s."></metric>
<metric name="WDATA1_SIZE" expr=((TCC_EA1_WRREQ_sum-TCC_EA1_WRREQ_64B_sum)*32+TCC_EA1_WRREQ_64B_sum*64) descr="The total kilobytes written to the video memory. This is measured on EA1s."></metric>
# both EA0 and EA1 should be included
<metric name="FETCH_SIZE" expr=(TCC_EA_RDREQ_32B_sum*32+(TCC_EA_RDREQ_sum-TCC_EA_RDREQ_32B_sum)*64+RDATA1_SIZE)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_SIZE" expr=((TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum)*32+TCC_EA_WRREQ_64B_sum*64+WDATA1_SIZE)/1024 descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_REQ_32B" expr=(TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum)+(TCC_EA1_WRREQ_sum-TCC_EA1_WRREQ_64B_sum)+(TCC_EA_WRREQ_64B_sum+TCC_EA1_WRREQ_64B_sum)*2 descr="The total number of 32-byte effective memory writes."></metric>
</gfx906>
<gfx908 base="gfx9">
<metric name="TCC_HIT_sum" expr=reduce(TCC_HIT,sum) descr="Number of cache hits. Sum over TCC instances."></metric>
<metric name="TCC_MISS_sum" expr=reduce(TCC_MISS,sum) descr="Number of cache misses. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_32B_sum" expr=reduce(TCC_EA_RDREQ_32B,sum) descr="Number of 32-byte TCC/EA read requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_sum" expr=reduce(TCC_EA_RDREQ,sum) descr="Number of TCC/EA read requests (either 32-byte or 64-byte). Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_sum" expr=reduce(TCC_EA_WRREQ,sum) descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_64B_sum" expr=reduce(TCC_EA_WRREQ_64B,sum) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_WRREQ_STALL_max" expr=reduce(TCC_EA_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
<metric name="CU_UTILIZATION" expr=GRBM_GUI_ACTIVE/GRBM_COUNT descr="The total number of active cycles divided by total number of elapsed cycles"></metric>
</gfx908>
<gfx90a base="gfx9">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="MeanOccupancyPerCU" expr=SQ_LEVEL_WAVES*0+SQ_ACCUM_PREV_HIRES/GRBM_GUI_ACTIVE/CU_NUM descr="Mean occupancy per compute unit."></metric>
<metric name="MeanOccupancyPerActiveCU" expr=SQ_LEVEL_WAVES*0+SQ_ACCUM_PREV_HIRES*4/SQ_BUSY_CYCLES/CU_NUM descr="Mean occupancy per active compute unit."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
<metric name="TA_BUSY_min" expr=reduce(TA_TA_BUSY,min) descr="TA block is busy. Min over TA instances."></metric>
<metric name="TA_TA_BUSY_sum" expr=reduce(TA_TA_BUSY,sum) descr="TA block is busy. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_TOTAL_WAVEFRONTS_sum" expr=reduce(TA_TOTAL_WAVEFRONTS,sum) descr="Total number of wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_ADDR_STALLED_BY_TC_CYCLES_sum" expr=reduce(TA_ADDR_STALLED_BY_TC_CYCLES,sum) descr="Number of cycles addr path stalled by TC. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_ADDR_STALLED_BY_TD_CYCLES_sum" expr=reduce(TA_ADDR_STALLED_BY_TD_CYCLES,sum) descr="Number of cycles addr path stalled by TD. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_DATA_STALLED_BY_TC_CYCLES_sum" expr=reduce(TA_DATA_STALLED_BY_TC_CYCLES,sum) descr="Number of cycles data path stalled by TC. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_FLAT_WAVEFRONTS_sum" expr=reduce(TA_FLAT_WAVEFRONTS,sum) descr="Number of flat opcode wavfronts processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS_sum" expr=reduce(TA_FLAT_READ_WAVEFRONTS,sum) descr="Number of flat opcode reads processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS_sum" expr=reduce(TA_FLAT_WRITE_WAVEFRONTS,sum) descr="Number of flat opcode writes processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_ATOMIC_WAVEFRONTS_sum" expr=reduce(TA_FLAT_ATOMIC_WAVEFRONTS,sum) descr="Number of flat opcode atomics processed by the TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_WAVEFRONTS,sum) descr="Number of buffer wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_READ_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_READ_WAVEFRONTS,sum) descr="Number of buffer read wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_WRITE_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_WRITE_WAVEFRONTS,sum) descr="Number of buffer write wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_ATOMIC_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_ATOMIC_WAVEFRONTS,sum) descr="Number of buffer atomic wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_TOTAL_CYCLES_sum" expr=reduce(TA_BUFFER_TOTAL_CYCLES,sum) descr="Number of buffer cycles issued to TC. Sum over TA instances."></metric>
<metric name="TA_BUFFER_COALESCED_READ_CYCLES_sum" expr=reduce(TA_BUFFER_COALESCED_READ_CYCLES,sum) descr="Number of buffer coalesced read cycles issued to TC. Sum over TA instances."></metric>
<metric name="TA_BUFFER_COALESCED_WRITE_CYCLES_sum" expr=reduce(TA_BUFFER_COALESCED_WRITE_CYCLES,sum) descr="Number of buffer coalesced write cycles issued to TC. Sum over TA instances."></metric>
<metric name="TD_TD_BUSY_sum" expr=reduce(TD_TD_BUSY,sum) descr="TD is processing or waiting for data. Perf_Windowing not supported for this counter. Sum over TD instances."></metric>
<metric name="TD_TC_STALL_sum" expr=reduce(TD_TC_STALL,sum) descr="TD is stalled waiting for TC data. Sum over TD instances."></metric>
<metric name="TD_LOAD_WAVEFRONT_sum" expr=reduce(TD_LOAD_WAVEFRONT,sum) descr="Count the wavefronts with opcode = load, include atomics and store. Sum over TD instances."></metric>
<metric name="TD_ATOMIC_WAVEFRONT_sum" expr=reduce(TD_ATOMIC_WAVEFRONT,sum) descr="Count the wavefronts with opcode = atomic. Sum over TD instances."></metric>
<metric name="TD_STORE_WAVEFRONT_sum" expr=reduce(TD_STORE_WAVEFRONT,sum) descr="Count the wavefronts with opcode = store. Sum over TD instances."></metric>
<metric name="TD_COALESCABLE_WAVEFRONT_sum" expr=reduce(TD_COALESCABLE_WAVEFRONT,sum) descr="Count wavefronts that TA finds coalescable. Sum over TD instances."></metric>
<metric name="TD_SPI_STALL_sum" expr=reduce(TD_SPI_STALL,sum) descr="TD is stalled SPI vinit, sum of TCP instances"></metric>
<metric name="TCP_GATE_EN1_sum" expr=reduce(TCP_GATE_EN1,sum) descr="TCP interface clocks are turned on. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_GATE_EN2_sum" expr=reduce(TCP_GATE_EN2,sum) descr="TCP core clocks are turned on. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_TD_TCP_STALL_CYCLES_sum" expr=reduce(TCP_TD_TCP_STALL_CYCLES,sum) descr="TD stalls TCP. Sum over TCP instances."></metric>
<metric name="TCP_TCR_TCP_STALL_CYCLES_sum" expr=reduce(TCP_TCR_TCP_STALL_CYCLES,sum) descr="TCR stalls TCP_TCR_req interface. Sum over TCP instances."></metric>
<metric name="TCP_READ_TAGCONFLICT_STALL_CYCLES_sum" expr=reduce(TCP_READ_TAGCONFLICT_STALL_CYCLES,sum) descr="Tagram conflict stall on a read. Sum over TCP instances."></metric>
<metric name="TCP_WRITE_TAGCONFLICT_STALL_CYCLES_sum" expr=reduce(TCP_WRITE_TAGCONFLICT_STALL_CYCLES,sum) descr="Tagram conflict stall on a write. Sum over TCP instances."></metric>
<metric name="TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES_sum" expr=reduce(TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES,sum) descr="Tagram conflict stall on an atomic. Sum over TCP instances."></metric>
<metric name="TCP_VOLATILE_sum" expr=reduce(TCP_VOLATILE,sum) descr="Total number of L1 volatile pixels/buffers from TA. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_ACCESSES_sum" expr=reduce(TCP_TOTAL_ACCESSES,sum) descr="Total number of pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_READ+TCP_PERF_SEL_TOTAL_NONREAD. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_READ_sum" expr=reduce(TCP_TOTAL_READ,sum) descr="Total number of read pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_HIT_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_EVICT_READ. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_WRITE_sum" expr=reduce(TCP_TOTAL_WRITE,sum) descr="Total number of local write pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_MISS_LRU_WRITE+ TCP_PERF_SEL_TOTAL_MISS_EVICT_WRITE. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_ATOMIC_WITH_RET_sum" expr=reduce(TCP_TOTAL_ATOMIC_WITH_RET,sum) descr="Total number of atomic with return pixels/buffers from TA. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_ATOMIC_WITHOUT_RET_sum" expr=reduce(TCP_TOTAL_ATOMIC_WITHOUT_RET,sum) descr="Total number of atomic without return pixels/buffers from TA Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_WRITEBACK_INVALIDATES_sum" expr=reduce(TCP_TOTAL_WRITEBACK_INVALIDATES,sum) descr="Total number of cache invalidates. Equals TCP_PERF_SEL_TOTAL_WBINVL1+ TCP_PERF_SEL_TOTAL_WBINVL1_VOL+ TCP_PERF_SEL_CP_TCP_INVALIDATE+ TCP_PERF_SEL_SQ_TCP_INVALIDATE_VOL. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_REQUEST_sum" expr=reduce(TCP_UTCL1_REQUEST,sum) descr="Total CLIENT_UTCL1 NORMAL requests Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_TRANSLATION_MISS_sum" expr=reduce(TCP_UTCL1_TRANSLATION_MISS,sum) descr="Total utcl1 translation misses Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_TRANSLATION_HIT_sum" expr=reduce(TCP_UTCL1_TRANSLATION_HIT,sum) descr="Total utcl1 translation hits Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_PERMISSION_MISS_sum" expr=reduce(TCP_UTCL1_PERMISSION_MISS,sum) descr="Total utcl1 permission misses Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_CACHE_ACCESSES_sum" expr=reduce(TCP_TOTAL_CACHE_ACCESSES,sum) descr="Count of total cache line (tag) accesses (includes hits and misses). Sum over TCP instances."></metric>
<metric name="TCP_TCP_LATENCY_sum" expr=reduce(TCP_TCP_LATENCY,sum) descr="Total TCP wave latency (from first clock of wave entering to first clock of wave leaving), divide by TA_TCP_STATE_READ to avg wave latency Sum over TCP instances."></metric>
<metric name="TCP_TA_TCP_STATE_READ_sum" expr=reduce(TCP_TA_TCP_STATE_READ,sum) descr="Number of state reads Sum over TCP instances."></metric>
<metric name="TCP_TCC_READ_REQ_LATENCY_sum" expr=reduce(TCP_TCC_READ_REQ_LATENCY,sum) descr="Total TCP->TCC request latency for reads and atomics with return. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_TCC_WRITE_REQ_LATENCY_sum" expr=reduce(TCP_TCC_WRITE_REQ_LATENCY,sum) descr="Total TCP->TCC request latency for writes and atomics without return. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_TCC_READ_REQ_sum" expr=reduce(TCP_TCC_READ_REQ,sum) descr="Total read requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_WRITE_REQ_sum" expr=reduce(TCP_TCC_WRITE_REQ,sum) descr="Total write requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_ATOMIC_WITH_RET_REQ_sum" expr=reduce(TCP_TCC_ATOMIC_WITH_RET_REQ,sum) descr="Total atomic with return requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum" expr=reduce(TCP_TCC_ATOMIC_WITHOUT_RET_REQ,sum) descr="Total atomic without return requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_NC_READ_REQ_sum" expr=reduce(TCP_TCC_NC_READ_REQ,sum) descr="Total read requests with NC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_NC_WRITE_REQ_sum" expr=reduce(TCP_TCC_NC_WRITE_REQ,sum) descr="Total write requests with NC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_NC_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_NC_ATOMIC_REQ,sum) descr="Total atomic requests with NC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_UC_READ_REQ_sum" expr=reduce(TCP_TCC_UC_READ_REQ,sum) descr="Total read requests with UC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_UC_WRITE_REQ_sum" expr=reduce(TCP_TCC_UC_WRITE_REQ,sum) descr="Total write requests with UC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_UC_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_UC_ATOMIC_REQ,sum) descr="Total atomic requests with UC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_CC_READ_REQ_sum" expr=reduce(TCP_TCC_CC_READ_REQ,sum) descr="Total write requests with CC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_CC_WRITE_REQ_sum" expr=reduce(TCP_TCC_CC_WRITE_REQ,sum) descr="Total write requests with CC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_CC_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_CC_ATOMIC_REQ,sum) descr="Total atomic requests with CC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_RW_READ_REQ_sum" expr=reduce(TCP_TCC_RW_READ_REQ,sum) descr="Total write requests with RW mtype from this TCP to all TCCs. Sum over TCP instances."></metric>
<metric name="TCP_TCC_RW_WRITE_REQ_sum" expr=reduce(TCP_TCC_RW_WRITE_REQ,sum) descr="Total write requests with RW mtype from this TCP to all TCCs. Sum over TCP instances."></metric>
<metric name="TCP_TCC_RW_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_RW_ATOMIC_REQ,sum) descr="Total atomic requests with RW mtype from this TCP to all TCCs. Sum over TCP instances."></metric>
<metric name="TCP_PENDING_STALL_CYCLES_sum" expr=reduce(TCP_PENDING_STALL_CYCLES,sum) descr="Stall due to data pending from L2. Sum over TCP instances."></metric>
<metric name="TCA_CYCLE_sum" expr=reduce(TCA_CYCLE,sum) descr="Number of cycles. Sum over all TCA instances "></metric>
<metric name="TCA_BUSY_sum" expr=reduce(TCA_BUSY,sum) descr="Number of cycles we have a request pending. Sum over all TCA instances."></metric>
<metric name="TCC_BUSY_avr" expr=reduce(TCC_BUSY,avr) descr="TCC_BUSY avr over all memory channels."></metric>
<metric name="TCC_WRREQ_STALL_max" expr=reduce(TCC_EA_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
<metric name="TCC_CYCLE_sum" expr=reduce(TCC_CYCLE,sum) descr="Number of cycles. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_BUSY_sum" expr=reduce(TCC_BUSY,sum) descr="Number of cycles we have a request pending. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_REQ_sum" expr=reduce(TCC_REQ,sum) descr="Number of requests of all types. This is measured at the tag block. This may be more than the number of requests arriving at the TCC, but it is a good indication of the total amount of work that needs to be performed. Sum over TCC instances."></metric>
<metric name="TCC_STREAMING_REQ_sum" expr=reduce(TCC_STREAMING_REQ,sum) descr="Number of streaming requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_NC_REQ_sum" expr=reduce(TCC_NC_REQ,sum) descr="The number of noncoherently cached requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_UC_REQ_sum" expr=reduce(TCC_UC_REQ,sum) descr="The number of uncached requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_CC_REQ_sum" expr=reduce(TCC_CC_REQ,sum) descr="The number of coherently cached requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_RW_REQ_sum" expr=reduce(TCC_RW_REQ,sum) descr="The number of RW requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_PROBE_sum" expr=reduce(TCC_PROBE,sum) descr="Number of probe requests. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_PROBE_ALL_sum" expr=reduce(TCC_PROBE_ALL,sum) descr="Number of external probe requests with with EA_TCC_preq_all== 1. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_READ_sum" expr=reduce(TCC_READ,sum) descr="Number of read requests. Compressed reads are included in this, but metadata reads are not included. Sum over TCC instances."></metric>
<metric name="TCC_WRITE_sum" expr=reduce(TCC_WRITE,sum) descr="Number of write requests. Sum over TCC instances."></metric>
<metric name="TCC_ATOMIC_sum" expr=reduce(TCC_ATOMIC,sum) descr="Number of atomic requests of all types. Sum over TCC instances."></metric>
<metric name="TCC_HIT_sum" expr=reduce(TCC_HIT,sum) descr="Number of cache hits. Sum over TCC instances."></metric>
<metric name="TCC_MISS_sum" expr=reduce(TCC_MISS,sum) descr="Number of cache misses. UC reads count as misses. Sum over TCC instances."></metric>
<metric name="TCC_WRITEBACK_sum" expr=reduce(TCC_WRITEBACK,sum) descr="Number of lines written back to main memory. This includes writebacks of dirty lines and uncached write/atomic requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_sum" expr=reduce(TCC_EA_WRREQ,sum) descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_64B_sum" expr=reduce(TCC_EA_WRREQ_64B,sum) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_EA_WR_UNCACHED_32B_sum" expr=reduce(TCC_EA_WR_UNCACHED_32B,sum) descr="Number of 32-byte write/atomic going over the TC_EA_wrreq interface due to uncached traffic. Note that CC mtypes can produce uncached requests, and those are included in this. A 64-byte request will be counted as 2. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_STALL_sum" expr=reduce(TCC_EA_WRREQ_STALL,sum) descr="Number of cycles a write request was stalled. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_IO_CREDIT_STALL_sum" expr=reduce(TCC_EA_WRREQ_IO_CREDIT_STALL,sum) descr="Number of cycles a EA write request was stalled because the interface was out of IO credits. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_GMI_CREDIT_STALL_sum" expr=reduce(TCC_EA_WRREQ_GMI_CREDIT_STALL,sum) descr="Number of cycles a EA write request was stalled because the interface was out of GMI credits. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_DRAM_CREDIT_STALL_sum" expr=reduce(TCC_EA_WRREQ_DRAM_CREDIT_STALL,sum) descr="Number of cycles a EA write request was stalled because the interface was out of DRAM credits. Sum over TCC instances."></metric>
<metric name="TCC_TOO_MANY_EA_WRREQS_STALL_sum" expr=reduce(TCC_TOO_MANY_EA_WRREQS_STALL,sum) descr="Number of cycles the TCC could not send a EA write request because it already reached its maximum number of pending EA write requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_LEVEL_sum" expr=reduce(TCC_EA_WRREQ_LEVEL,sum) descr="The sum of the number of EA write requests in flight. This is primarily meant for measure average EA write latency. Average write latency = TCC_PERF_SEL_EA_WRREQ_LEVEL/TCC_PERF_SEL_EA_WRREQ. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_LEVEL_sum" expr=reduce(TCC_EA_RDREQ_LEVEL,sum) descr="The sum of the number of TCC/EA read requests in flight. This is primarily meant for measure average EA read latency. Average read latency = TCC_PERF_SEL_EA_RDREQ_LEVEL/TCC_PERF_SEL_EA_RDREQ. Sum over TCC instances."></metric>
<metric name="TCC_EA_ATOMIC_sum" expr=reduce(TCC_EA_ATOMIC,sum) descr="Number of transactions going over the TC_EA_wrreq interface that are actually atomic requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_ATOMIC_LEVEL_sum" expr=reduce(TCC_EA_ATOMIC_LEVEL,sum) descr="The sum of the number of EA atomics in flight. This is primarily meant for measure average EA atomic latency. Average atomic latency = TCC_PERF_SEL_EA_WRREQ_ATOMIC_LEVEL/TCC_PERF_SEL_EA_WRREQ_ATOMIC. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_sum" expr=reduce(TCC_EA_RDREQ,sum) descr="Number of TCC/EA read requests (either 32-byte or 64-byte) Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_32B_sum" expr=reduce(TCC_EA_RDREQ_32B,sum) descr="Number of 32-byte TCC/EA read requests Sum over TCC instances."></metric>
<metric name="TCC_EA_RD_UNCACHED_32B_sum" expr=reduce(TCC_EA_RD_UNCACHED_32B,sum) descr="Number of 32-byte TCC/EA read due to uncached traffic. A 64-byte request will be counted as 2 Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_IO_CREDIT_STALL_sum" expr=reduce(TCC_EA_RDREQ_IO_CREDIT_STALL,sum) descr="Number of cycles there was a stall because the read request interface was out of IO credits. Stalls occur regardless of whether a read needed to be performed or not. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_GMI_CREDIT_STALL_sum" expr=reduce(TCC_EA_RDREQ_GMI_CREDIT_STALL,sum) descr="Number of cycles there was a stall because the read request interface was out of GMI credits. Stalls occur regardless of whether a read needed to be performed or not. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_DRAM_CREDIT_STALL_sum" expr=reduce(TCC_EA_RDREQ_DRAM_CREDIT_STALL,sum) descr="Number of cycles there was a stall because the read request interface was out of DRAM credits. Stalls occur regardless of whether a read needed to be performed or not. Sum over TCC instances."></metric>
<metric name="TCC_TAG_STALL_sum" expr=reduce(TCC_TAG_STALL,sum) descr="Total number of cycles the normal request pipeline in the tag is stalled for any reason."></metric>
<metric name="TCC_NORMAL_WRITEBACK_sum" expr=reduce(TCC_NORMAL_WRITEBACK,sum) descr="Number of writebacks due to requests that are not writeback requests. Sum over TCC instances."></metric>
<metric name="TCC_ALL_TC_OP_WB_WRITEBACK_sum" expr=reduce(TCC_ALL_TC_OP_WB_WRITEBACK,sum) descr="Number of writebacks due to all TC_OP writeback requests. Sum over TCC instances."></metric>
<metric name="TCC_NORMAL_EVICT_sum" expr=reduce(TCC_NORMAL_EVICT,sum) descr="Number of evictions due to requests that are not invalidate or probe requests. Sum over TCC instances."></metric>
<metric name="TCC_ALL_TC_OP_INV_EVICT_sum" expr=reduce(TCC_ALL_TC_OP_INV_EVICT,sum) descr="Number of evictions due to all TC_OP invalidate requests. Sum over TCC instances."></metric>
<metric name="TCC_EA_RDREQ_DRAM_sum" expr=reduce(TCC_EA_RDREQ_DRAM,sum) descr="Number of TCC/EA read requests (either 32-byte or 64-byte) destined for DRAM (MC). Sum over TCC instances."></metric>
<metric name="TCC_EA_WRREQ_DRAM_sum" expr=reduce(TCC_EA_WRREQ_DRAM,sum) descr="Number of TCC/EA write requests (either 32-byte of 64-byte) destined for DRAM (MC). Sum over TCC instances."></metric>
<metric name="FETCH_SIZE" expr=(TCC_EA_RDREQ_32B_sum*32+(TCC_EA_RDREQ_sum-TCC_EA_RDREQ_32B_sum)*64)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_SIZE" expr=((TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum)*32+TCC_EA_WRREQ_64B_sum*64)/1024 descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_REQ_32B" expr=TCC_EA_WRREQ_64B_sum*2+(TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum) descr="The total number of 32-byte effective memory writes."></metric>
<metric name="CU_OCCUPANCY" expr=(SQ_CYCLES/(SQ_WAVE_CYCLES*4))/MAX_WAVE_SIZE descr="The ratio of active waves on a CU to the maximum number of active waves supported by the CU"></metric>
<metric name="CU_UTILIZATION" expr=GRBM_GUI_ACTIVE/GRBM_COUNT descr="The total number of active cycles divided by total number of elapsed cycles"></metric>
<metric name="TOTAL_16_OPS" expr=(SQ_INSTS_VALU_FMA_F16*2+SQ_INSTS_VALU_ADD_F16+SQ_INSTS_VALU_MUL_F16+SQ_INSTS_VALU_TRANS_F16)*64+((SQ_INSTS_VALU_MFMA_MOPS_F16+SQ_INSTS_VALU_MFMA_MOPS_BF16)*512) descr="The number of 16 bits OPS executed"></metric>
<metric name="TOTAL_32_OPS" expr=(SQ_INSTS_VALU_FMA_F32*2+SQ_INSTS_VALU_INT32+SQ_INSTS_VALU_ADD_F32+SQ_INSTS_VALU_MUL_F32+SQ_INSTS_VALU_TRANS_F32)*64+(SQ_INSTS_VALU_MFMA_MOPS_F32*512) descr="The number of 32 bits OPS executed"></metric>
<metric name="TOTAL_64_OPS" expr=(SQ_INSTS_VALU_FMA_F64*2+SQ_INSTS_VALU_INT64+SQ_INSTS_VALU_ADD_F64+SQ_INSTS_VALU_MUL_F64)*64+(SQ_INSTS_VALU_MFMA_MOPS_F64*512) descr="The number of 64 bits OPS executed"></metric>
<metric name="AggSysCycles" expr=GRBM_GUI_ACTIVE*CU_NUM descr="Unit: cycles"></metric>
## IP Block Utilization Metrics
<metric name="GpuUtil" expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Unit: percent"></metric>
<metric name="CpUtil" expr=100*GRBM_CP_BUSY/GRBM_GUI_ACTIVE descr="Unit: percent"></metric>
<metric name="SpiUtil" expr=100*GRBM_SPI_BUSY/GRBM_GUI_ACTIVE descr="Unit: percent"></metric>
<metric name="TaUtil" expr=100*GRBM_TA_BUSY/GRBM_GUI_ACTIVE descr="Unit: percent"></metric>
<metric name="TcUtil" expr=100*GRBM_TC_BUSY/GRBM_GUI_ACTIVE descr="Unit: percent"></metric>
<metric name="EaUtil" expr=100*GRBM_EA_BUSY/GRBM_GUI_ACTIVE descr="Unit: percent"></metric>
## Instruction Fetch Metrics
<metric name="InstrFetchLatency" expr=SQ_ACCUM_PREV_HIRES/SQ_IFETCH descr="Unit: cycles"></metric>
## Wavefront Metrics
<metric name="WaveOccupancy" expr=SQ_ACCUM_PREV_HIRES/GRBM_GUI_ACTIVE descr="Unit: wavefronts"></metric>
<metric name="WaveDuration" expr=4*SQ_WAVE_CYCLES/SQ_WAVES descr="Unit: cycles"></metric>
<metric name="WaveDepWait" expr=100*SQ_WAIT_ANY/SQ_WAVE_CYCLES descr="Unit: percent"></metric>
<metric name="WaveIssueWait" expr=100*SQ_WAIT_INST_ANY/SQ_WAVE_CYCLES descr="Unit: percent"></metric>
<metric name="WaveExec" expr=100*SQ_ACTIVE_INST_ANY/SQ_WAVE_CYCLES descr="Unit: percent"></metric>
## Compute Unit Metrics
<metric name="ValuIops" expr=(SQ_INSTS_VALU_INT32+SQ_INSTS_VALU_INT64)*64 descr="Unit: IOP"></metric>
<metric name="MfmaFlops" expr=(SQ_INSTS_VALU_MFMA_MOPS_F16+SQ_INSTS_VALU_MFMA_MOPS_BF16+SQ_INSTS_VALU_MFMA_MOPS_F32+SQ_INSTS_VALU_MFMA_MOPS_F64)*512 descr="Unit: FLOP"></metric>
<metric name="MfmaFlopsF16" expr=SQ_INSTS_VALU_MFMA_MOPS_F16*512 descr="Unit: FLOP"></metric>
<metric name="MfmaFlopsBF16" expr=SQ_INSTS_VALU_MFMA_MOPS_BF16*512 descr="Unit: FLOP"></metric>
<metric name="MfmaFlopsF32" expr=SQ_INSTS_VALU_MFMA_MOPS_F32*512 descr="Unit: FLOP"></metric>
<metric name="MfmaFlopsF64" expr=SQ_INSTS_VALU_MFMA_MOPS_F64*512 descr="Unit: IOP"></metric>
<metric name="ScaPipeIssueUtil" expr=100*SQ_ACTIVE_INST_SCA/(GRBM_GUI_ACTIVE*CU_NUM) descr="Unit: percent"></metric>
<metric name="ValuPipeIssueUtil" expr=100*SQ_ACTIVE_INST_VALU/(GRBM_GUI_ACTIVE*CU_NUM) descr="Unit: percent"></metric>
<metric name="VmemPipeIssueUtil" expr=400*(SQ_ACTIVE_INST_VMEM+SQ_ACTIVE_INST_FLAT)/(GRBM_GUI_ACTIVE*CU_NUM) descr="Unit: percent"></metric>
<metric name="MfmaUtil" expr=100*SQ_VALU_MFMA_BUSY_CYCLES/(GRBM_GUI_ACTIVE*CU_NUM*4) descr="Unit: percent"></metric>
<metric name="AvgNumActiveThreads" expr=SQ_THREAD_CYCLES_VALU/SQ_ACTIVE_INST_VALU descr="Unit: percent"></metric>
<metric name="VmemLatency" expr=SQ_ACCUM_PREV_HIRES/SQ_INSTS_VMEM descr="Unit: cycles"></metric>
<metric name="SmemLatency" expr=SQ_ACCUM_PREV_HIRES/SQ_INSTS_SMEM_NORM descr="Unit: cycles"></metric>
## Local Data Share (LDS) Metrics
<metric name="LdsUtil" expr=100*SQ_LDS_IDX_ACTIVE/(GRBM_GUI_ACTIVE*CU_NUM) descr="Unit: percent"></metric>
<metric name="LdsPipeIssueUtil" expr=400*SQ_ACTIVE_INST_LDS/(GRBM_GUI_ACTIVE*CU_NUM*2) descr="Unit: percent"></metric>
<metric name="LdsLatency" expr=SQ_ACCUM_PREV_HIRES/SQ_INSTS_LDS descr="Unit: cycles"></metric>
<metric name="LdsBankConflict" expr=SQ_LDS_BANK_CONFLICT/(SQ_LDS_IDX_ACTIVE-SQ_LDS_BANK_CONFLICT) descr="Unit: conflicts/access"></metric>
## L1I and sL1D Cache Metrics
<metric name="L1iCacheHitRate" expr=100*SQC_ICACHE_HITS/SQC_ICACHE_REQ descr="Unit: percent"></metric>
<metric name="sL1dCacheHitRate" expr=100*SQC_DCACHE_HITS/SQC_DCACHE_REQ descr="Unit: percent"></metric>
## vL1D Cache Metrics
<metric name="vL1dBufCoalesceRate" expr=6400*TA_TOTAL_WAVEFRONTS_sum/(TCP_TOTAL_ACCESSES_sum*4) descr="Unit: percent"></metric>
<metric name="vL1dCacheUtil" expr=100*TCP_GATE_EN2_sum/TCP_GATE_EN1_sum descr="Unit: percent"></metric>
<metric name="vL1dCacheTcbHitRate" expr=100*TCP_UTCL1_TRANSLATION_HIT_sum/TCP_UTCL1_REQUEST_sum descr="Unit: percent"></metric>
<metric name="vL1dCacheWaveLatency" expr=TCP_TCP_LATENCY_sum/TCP_TA_TCP_STATE_READ_sum descr="Unit: cycles"></metric>
<metric name="vL1dReadFromL2Latency" expr=TCP_TCC_READ_REQ_LATENCY_sum/(TCP_TCC_READ_REQ_sum+TCP_TCC_ATOMIC_WITH_RET_REQ_sum) descr="Unit: cycles"></metric>
<metric name="vL1dWriteToL2Latency" expr=TCP_TCC_WRITE_REQ_LATENCY_sum/(TCP_TCC_WRITE_REQ_sum+TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum) descr="Unit: cycles"></metric>
<metric name="vL1dRdTagConfStallRate" expr=100*TCP_READ_TAGCONFLICT_STALL_CYCLES_sum/TCP_GATE_EN2_sum descr="Unit: percent"></metric>
<metric name="vL1dWrTagConfStallRate" expr=100*TCP_WRITE_TAGCONFLICT_STALL_CYCLES_sum/TCP_GATE_EN2_sum descr="Unit: percent"></metric>
<metric name="vL1dAtomicTagConfStallRate" expr=100*TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES_sum/TCP_GATE_EN2_sum descr="Unit: percent"></metric>
<metric name="vL1dMissReqStallRate" expr=100*TCP_TCR_TCP_STALL_CYCLES_sum/TCP_GATE_EN2_sum descr="Unit: percent"></metric>
<metric name="vL1dDataPendRate" expr=100*TCP_PENDING_STALL_CYCLES_sum/TCP_GATE_EN2_sum descr="Unit: percent"></metric>
<metric name="vL1dDataRetStallRate" expr=100*TD_TC_STALL_sum/TD_TD_BUSY_sum descr="Unit: percent"></metric>
## L2 Cache Metrics
<metric name="L2CacheHitRate" expr=100*TCC_HIT_sum/(TCC_HIT_sum+TCC_MISS_sum) descr="Unit: percent"></metric>
<metric name="L2CacheTagRamStallRate" expr=100*TCC_TAG_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaRdLatency" expr=TCC_EA_RDREQ_LEVEL_sum/TCC_EA_RDREQ_sum descr="Unit: cycles"></metric>
<metric name="EaRdIoStallRate" expr=100*TCC_EA_RDREQ_IO_CREDIT_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaRdGmiStallRate" expr=100*TCC_EA_RDREQ_GMI_CREDIT_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaRdDramStallRate" expr=100*TCC_EA_RDREQ_DRAM_CREDIT_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaWrLatency" expr=TCC_EA_WRREQ_LEVEL_sum/TCC_EA_WRREQ_sum descr="Unit: cycles"></metric>
<metric name="EaWrIoStallRate" expr=100*TCC_EA_WRREQ_IO_CREDIT_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaWrGmiStallRate" expr=100*TCC_EA_WRREQ_GMI_CREDIT_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaWrDramStallRate" expr=100*TCC_EA_WRREQ_DRAM_CREDIT_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaWrStarveRate" expr=100*TCC_TOO_MANY_EA_WRREQS_STALL_sum/TCC_BUSY_sum descr="Unit: percent"></metric>
<metric name="EaAtomicLatency" expr=TCC_EA_ATOMIC_LEVEL_sum/TCC_EA_ATOMIC_sum descr="Unit: cycles"></metric>
</gfx90a>
<gfx940>
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES_sum" expr=reduce(TCP_TCP_TA_DATA_STALL_CYCLES,sum) descr="Total number of TCP stalls TA data interface."></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES_max" expr=reduce(TCP_TCP_TA_DATA_STALL_CYCLES,max) descr="Maximum number of TCP stalls TA data interface."></metric>
<metric name="MeanOccupancyPerCU" expr=reduce(SQ_LEVEL_WAVES,sum)*0+reduce(SQ_ACCUM_PREV_HIRES,sum)/reduce(GRBM_GUI_ACTIVE,sum)/CU_NUM descr="Mean occupancy per compute unit."></metric>
<metric name="MeanOccupancyPerActiveCU" expr=SQ_LEVEL_WAVES*0+SQ_ACCUM_PREV_HIRES*4/SQ_BUSY_CYCLES/CU_NUM descr="Mean occupancy per active compute unit."></metric>
<metric name="VFetchInsts" expr=(SQ_INSTS_VMEM_RD-TA_FLAT_READ_WAVEFRONTS_sum)/SQ_WAVES descr="The average number of vector fetch instructions from the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that fetch from video memory."></metric>
<metric name="VWriteInsts" expr=(SQ_INSTS_VMEM_WR-TA_FLAT_WRITE_WAVEFRONTS_sum)/SQ_WAVES descr="The average number of vector write instructions to the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that write to video memory."></metric>
<metric name="VALUUtilization" expr=100*SQ_THREAD_CYCLES_VALU/(SQ_ACTIVE_INST_VALU*MAX_WAVE_SIZE) descr="The percentage of active vector ALU threads in a wave. A lower number can mean either more thread divergence in a wave or that the work-group size is not a multiple of 64. Value range: 0% (bad), 100% (ideal - no thread divergence)."></metric>
<metric name="VALUBusy" expr=100*reduce(SQ_ACTIVE_INST_VALU,sum)*4/SIMD_NUM/reduce(GRBM_GUI_ACTIVE,sum) descr="The percentage of GPUTime vector ALU instructions are processed. Value range: 0% (bad) to 100% (optimal)."></metric>
<metric name="SALUBusy" expr=100*reduce(SQ_INST_CYCLES_SALU,sum)*4/SIMD_NUM/reduce(GRBM_GUI_ACTIVE,sum) descr="The percentage of GPUTime scalar ALU instructions are processed. Value range: 0% (bad) to 100% (optimal)."></metric>
<metric name="FetchSize" expr=FETCH_SIZE descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WriteSize" expr=WRITE_SIZE descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="MemWrites32B" expr=WRITE_REQ_32B descr="The total number of effective 32B write transactions to the memory"></metric>
<metric name="MemUnitStalled" expr=100*TCP_TCP_TA_DATA_STALL_CYCLES_max/GRBM_GUI_ACTIVE/SE_NUM descr="The percentage of GPUTime the memory unit is stalled. Try reducing the number or size of fetches and writes if possible. Value range: 0% (optimal) to 100% (bad)."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
<metric name="TA_BUSY_min" expr=reduce(TA_TA_BUSY,min) descr="TA block is busy. Min over TA instances."></metric>
<metric name="TA_TA_BUSY_sum" expr=reduce(TA_TA_BUSY,sum) descr="TA block is busy. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_TOTAL_WAVEFRONTS_sum" expr=reduce(TA_TOTAL_WAVEFRONTS,sum) descr="Total number of wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_ADDR_STALLED_BY_TC_CYCLES_sum" expr=reduce(TA_ADDR_STALLED_BY_TC_CYCLES,sum) descr="Number of cycles addr path stalled by TC. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_ADDR_STALLED_BY_TD_CYCLES_sum" expr=reduce(TA_ADDR_STALLED_BY_TD_CYCLES,sum) descr="Number of cycles addr path stalled by TD. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_DATA_STALLED_BY_TC_CYCLES_sum" expr=reduce(TA_DATA_STALLED_BY_TC_CYCLES,sum) descr="Number of cycles data path stalled by TC. Perf_Windowing not supported for this counter. Sum over TA instances."></metric>
<metric name="TA_FLAT_WAVEFRONTS_sum" expr=reduce(TA_FLAT_WAVEFRONTS,sum) descr="Number of flat opcode wavfronts processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_READ_WAVEFRONTS_sum" expr=reduce(TA_FLAT_READ_WAVEFRONTS,sum) descr="Number of flat opcode reads processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_WRITE_WAVEFRONTS_sum" expr=reduce(TA_FLAT_WRITE_WAVEFRONTS,sum) descr="Number of flat opcode writes processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_ATOMIC_WAVEFRONTS_sum" expr=reduce(TA_FLAT_ATOMIC_WAVEFRONTS,sum) descr="Number of flat opcode atomics processed by the TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_WAVEFRONTS,sum) descr="Number of buffer wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_READ_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_READ_WAVEFRONTS,sum) descr="Number of buffer read wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_WRITE_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_WRITE_WAVEFRONTS,sum) descr="Number of buffer write wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_ATOMIC_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_ATOMIC_WAVEFRONTS,sum) descr="Number of buffer atomic wavefronts processed by TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_TOTAL_CYCLES_sum" expr=reduce(TA_BUFFER_TOTAL_CYCLES,sum) descr="Number of buffer cycles issued to TC. Sum over TA instances."></metric>
<metric name="TA_BUFFER_COALESCED_READ_CYCLES_sum" expr=reduce(TA_BUFFER_COALESCED_READ_CYCLES,sum) descr="Number of buffer coalesced read cycles issued to TC. Sum over TA instances."></metric>
<metric name="TA_BUFFER_COALESCED_WRITE_CYCLES_sum" expr=reduce(TA_BUFFER_COALESCED_WRITE_CYCLES,sum) descr="Number of buffer coalesced write cycles issued to TC. Sum over TA instances."></metric>
<metric name="TD_TD_BUSY_sum" expr=reduce(TD_TD_BUSY,sum) descr="TD is processing or waiting for data. Perf_Windowing not supported for this counter. Sum over TD instances."></metric>
<metric name="TD_TC_STALL_sum" expr=reduce(TD_TC_STALL,sum) descr="TD is stalled waiting for TC data. Sum over TD instances."></metric>
<metric name="TD_LOAD_WAVEFRONT_sum" expr=reduce(TD_LOAD_WAVEFRONT,sum) descr="Count the wavefronts with opcode = load, include atomics and store. Sum over TD instances."></metric>
<metric name="TD_ATOMIC_WAVEFRONT_sum" expr=reduce(TD_ATOMIC_WAVEFRONT,sum) descr="Count the wavefronts with opcode = atomic. Sum over TD instances."></metric>
<metric name="TD_STORE_WAVEFRONT_sum" expr=reduce(TD_STORE_WAVEFRONT,sum) descr="Count the wavefronts with opcode = store. Sum over TD instances."></metric>
<metric name="TD_COALESCABLE_WAVEFRONT_sum" expr=reduce(TD_COALESCABLE_WAVEFRONT,sum) descr="Count wavefronts that TA finds coalescable. Sum over TD instances."></metric>
<metric name="TD_SPI_STALL_sum" expr=reduce(TD_SPI_STALL,sum) descr="TD is stalled SPI vinit, sum of TCP instances"></metric>
<metric name="TCP_GATE_EN1_sum" expr=reduce(TCP_GATE_EN1,sum) descr="TCP interface clocks are turned on. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_GATE_EN2_sum" expr=reduce(TCP_GATE_EN2,sum) descr="TCP core clocks are turned on. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_TD_TCP_STALL_CYCLES_sum" expr=reduce(TCP_TD_TCP_STALL_CYCLES,sum) descr="TD stalls TCP. Sum over TCP instances."></metric>
<metric name="TCP_TCR_TCP_STALL_CYCLES_sum" expr=reduce(TCP_TCR_TCP_STALL_CYCLES,sum) descr="TCR stalls TCP_TCR_req interface. Sum over TCP instances."></metric>
<metric name="TCP_READ_TAGCONFLICT_STALL_CYCLES_sum" expr=reduce(TCP_READ_TAGCONFLICT_STALL_CYCLES,sum) descr="Tagram conflict stall on a read. Sum over TCP instances."></metric>
<metric name="TCP_WRITE_TAGCONFLICT_STALL_CYCLES_sum" expr=reduce(TCP_WRITE_TAGCONFLICT_STALL_CYCLES,sum) descr="Tagram conflict stall on a write. Sum over TCP instances."></metric>
<metric name="TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES_sum" expr=reduce(TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES,sum) descr="Tagram conflict stall on an atomic. Sum over TCP instances."></metric>
<metric name="TCP_VOLATILE_sum" expr=reduce(TCP_VOLATILE,sum) descr="Total number of L1 volatile pixels/buffers from TA. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_ACCESSES_sum" expr=reduce(TCP_TOTAL_ACCESSES,sum) descr="Total number of pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_READ+TCP_PERF_SEL_TOTAL_NONREAD. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_READ_sum" expr=reduce(TCP_TOTAL_READ,sum) descr="Total number of read pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_HIT_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_LRU_READ + TCP_PERF_SEL_TOTAL_MISS_EVICT_READ. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_WRITE_sum" expr=reduce(TCP_TOTAL_WRITE,sum) descr="Total number of local write pixels/buffers from TA. Equals TCP_PERF_SEL_TOTAL_MISS_LRU_WRITE+ TCP_PERF_SEL_TOTAL_MISS_EVICT_WRITE. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_ATOMIC_WITH_RET_sum" expr=reduce(TCP_TOTAL_ATOMIC_WITH_RET,sum) descr="Total number of atomic with return pixels/buffers from TA. Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_ATOMIC_WITHOUT_RET_sum" expr=reduce(TCP_TOTAL_ATOMIC_WITHOUT_RET,sum) descr="Total number of atomic without return pixels/buffers from TA Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_WRITEBACK_INVALIDATES_sum" expr=reduce(TCP_TOTAL_WRITEBACK_INVALIDATES,sum) descr="Total number of cache invalidates. Equals TCP_PERF_SEL_TOTAL_WBINVL1+ TCP_PERF_SEL_TOTAL_WBINVL1_VOL+ TCP_PERF_SEL_CP_TCP_INVALIDATE+ TCP_PERF_SEL_SQ_TCP_INVALIDATE_VOL. Not Windowed. Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_REQUEST_sum" expr=reduce(TCP_UTCL1_REQUEST,sum) descr="Total CLIENT_UTCL1 NORMAL requests Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_TRANSLATION_MISS_sum" expr=reduce(TCP_UTCL1_TRANSLATION_MISS,sum) descr="Total utcl1 translation misses Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_TRANSLATION_HIT_sum" expr=reduce(TCP_UTCL1_TRANSLATION_HIT,sum) descr="Total utcl1 translation hits Sum over TCP instances."></metric>
<metric name="TCP_UTCL1_PERMISSION_MISS_sum" expr=reduce(TCP_UTCL1_PERMISSION_MISS,sum) descr="Total utcl1 permission misses Sum over TCP instances."></metric>
<metric name="TCP_TOTAL_CACHE_ACCESSES_sum" expr=reduce(TCP_TOTAL_CACHE_ACCESSES,sum) descr="Count of total cache line (tag) accesses (includes hits and misses). Sum over TCP instances."></metric>
<metric name="TCP_TA_TCP_STATE_READ_sum" expr=reduce(TCP_TA_TCP_STATE_READ,sum) descr="Number of state reads Sum over TCP instances."></metric>
<metric name="TCP_TCC_READ_REQ_sum" expr=reduce(TCP_TCC_READ_REQ,sum) descr="Total read requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_WRITE_REQ_sum" expr=reduce(TCP_TCC_WRITE_REQ,sum) descr="Total write requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_ATOMIC_WITH_RET_REQ_sum" expr=reduce(TCP_TCC_ATOMIC_WITH_RET_REQ,sum) descr="Total atomic with return requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum" expr=reduce(TCP_TCC_ATOMIC_WITHOUT_RET_REQ,sum) descr="Total atomic without return requests from TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_NC_READ_REQ_sum" expr=reduce(TCP_TCC_NC_READ_REQ,sum) descr="Total read requests with NC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_NC_WRITE_REQ_sum" expr=reduce(TCP_TCC_NC_WRITE_REQ,sum) descr="Total write requests with NC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_NC_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_NC_ATOMIC_REQ,sum) descr="Total atomic requests with NC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_UC_READ_REQ_sum" expr=reduce(TCP_TCC_UC_READ_REQ,sum) descr="Total read requests with UC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_UC_WRITE_REQ_sum" expr=reduce(TCP_TCC_UC_WRITE_REQ,sum) descr="Total write requests with UC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_UC_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_UC_ATOMIC_REQ,sum) descr="Total atomic requests with UC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_CC_READ_REQ_sum" expr=reduce(TCP_TCC_CC_READ_REQ,sum) descr="Total write requests with CC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_CC_WRITE_REQ_sum" expr=reduce(TCP_TCC_CC_WRITE_REQ,sum) descr="Total write requests with CC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_CC_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_CC_ATOMIC_REQ,sum) descr="Total atomic requests with CC mtype from this TCP to all TCCs Sum over TCP instances."></metric>
<metric name="TCP_TCC_RW_READ_REQ_sum" expr=reduce(TCP_TCC_RW_READ_REQ,sum) descr="Total write requests with RW mtype from this TCP to all TCCs. Sum over TCP instances."></metric>
<metric name="TCP_TCC_RW_WRITE_REQ_sum" expr=reduce(TCP_TCC_RW_WRITE_REQ,sum) descr="Total write requests with RW mtype from this TCP to all TCCs. Sum over TCP instances."></metric>
<metric name="TCP_TCC_RW_ATOMIC_REQ_sum" expr=reduce(TCP_TCC_RW_ATOMIC_REQ,sum) descr="Total atomic requests with RW mtype from this TCP to all TCCs. Sum over TCP instances."></metric>
<metric name="TCP_PENDING_STALL_CYCLES_sum" expr=reduce(TCP_PENDING_STALL_CYCLES,sum) descr="Stall due to data pending from L2. Sum over TCP instances."></metric>
<metric name="TCA_CYCLE_sum" expr=reduce(TCA_CYCLE,sum) descr="Number of cycles. Sum over all TCA instances "></metric>
<metric name="TCA_BUSY_sum" expr=reduce(TCA_BUSY,sum) descr="Number of cycles we have a request pending. Sum over all TCA instances."></metric>
<metric name="TCC_BUSY_avr" expr=reduce(TCC_BUSY,avr) descr="TCC_BUSY avr over all memory channels."></metric>
<metric name="TCC_WRREQ_STALL_max" expr=reduce(TCC_EA0_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over TCC instances."></metric>
<metric name="TCC_CYCLE_sum" expr=reduce(TCC_CYCLE,sum) descr="Number of cycles. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_BUSY_sum" expr=reduce(TCC_BUSY,sum) descr="Number of cycles we have a request pending. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_REQ_sum" expr=reduce(TCC_REQ,sum) descr="Number of requests of all types. This is measured at the tag block. This may be more than the number of requests arriving at the TCC, but it is a good indication of the total amount of work that needs to be performed. Sum over TCC instances."></metric>
<metric name="TCC_STREAMING_REQ_sum" expr=reduce(TCC_STREAMING_REQ,sum) descr="Number of streaming requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_NC_REQ_sum" expr=reduce(TCC_NC_REQ,sum) descr="The number of noncoherently cached requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_UC_REQ_sum" expr=reduce(TCC_UC_REQ,sum) descr="The number of uncached requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_CC_REQ_sum" expr=reduce(TCC_CC_REQ,sum) descr="The number of coherently cached requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_RW_REQ_sum" expr=reduce(TCC_RW_REQ,sum) descr="The number of RW requests. This is measured at the tag block. Sum over TCC instances."></metric>
<metric name="TCC_PROBE_sum" expr=reduce(TCC_PROBE,sum) descr="Number of probe requests. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_PROBE_ALL_sum" expr=reduce(TCC_PROBE_ALL,sum) descr="Number of external probe requests with with EA_TCC_preq_all== 1. Not windowable. Sum over TCC instances."></metric>
<metric name="TCC_READ_sum" expr=reduce(TCC_READ,sum) descr="Number of read requests. Compressed reads are included in this, but metadata reads are not included. Sum over TCC instances."></metric>
<metric name="TCC_WRITE_sum" expr=reduce(TCC_WRITE,sum) descr="Number of write requests. Sum over TCC instances."></metric>
<metric name="TCC_ATOMIC_sum" expr=reduce(TCC_ATOMIC,sum) descr="Number of atomic requests of all types. Sum over TCC instances."></metric>
<metric name="TCC_HIT_sum" expr=reduce(TCC_HIT,sum) descr="Number of cache hits. Sum over TCC instances."></metric>
<metric name="TCC_MISS_sum" expr=reduce(TCC_MISS,sum) descr="Number of cache misses. UC reads count as misses. Sum over TCC instances."></metric>
<metric name="TCC_WRITEBACK_sum" expr=reduce(TCC_WRITEBACK,sum) descr="Number of lines written back to main memory. This includes writebacks of dirty lines and uncached write/atomic requests. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_sum" expr=reduce(TCC_EA0_WRREQ,sum) descr="Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_64B_sum" expr=reduce(TCC_EA0_WRREQ_64B,sum) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WR_UNCACHED_32B_sum" expr=reduce(TCC_EA0_WR_UNCACHED_32B,sum) descr="Number of 32-byte write/atomic going over the TC_EA_wrreq interface due to uncached traffic. Note that CC mtypes can produce uncached requests, and those are included in this. A 64-byte request will be counted as 2. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_STALL_sum" expr=reduce(TCC_EA0_WRREQ_STALL,sum) descr="Number of cycles a write request was stalled. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_IO_CREDIT_STALL_sum" expr=reduce(TCC_EA0_WRREQ_IO_CREDIT_STALL,sum) descr="Number of cycles a EA write request was stalled because the interface was out of IO credits. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_GMI_CREDIT_STALL_sum" expr=reduce(TCC_EA0_WRREQ_GMI_CREDIT_STALL,sum) descr="Number of cycles a EA write request was stalled because the interface was out of GMI credits. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_DRAM_CREDIT_STALL_sum" expr=reduce(TCC_EA0_WRREQ_DRAM_CREDIT_STALL,sum) descr="Number of cycles a EA write request was stalled because the interface was out of DRAM credits. Sum over TCC instances."></metric>
<metric name="TCC_TOO_MANY_EA_WRREQS_STALL_sum" expr=reduce(TCC_TOO_MANY_EA_WRREQS_STALL,sum) descr="Number of cycles the TCC could not send a EA write request because it already reached its maximum number of pending EA write requests. Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_LEVEL_sum" expr=reduce(TCC_EA0_WRREQ_LEVEL,sum) descr="The sum of the number of EA write requests in flight. This is primarily meant for measure average EA write latency. Average write latency = TCC_PERF_SEL_EA_WRREQ_LEVEL/TCC_PERF_SEL_EA_WRREQ. Sum over TCC instances."></metric>
<metric name="TCC_EA0_RDREQ_LEVEL_sum" expr=reduce(TCC_EA0_RDREQ_LEVEL,sum) descr="The sum of the number of TCC/EA read requests in flight. This is primarily meant for measure average EA read latency. Average read latency = TCC_PERF_SEL_EA_RDREQ_LEVEL/TCC_PERF_SEL_EA_RDREQ. Sum over TCC instances."></metric>
<metric name="TCC_EA0_ATOMIC_sum" expr=reduce(TCC_EA0_ATOMIC,sum) descr="Number of transactions going over the TC_EA_wrreq interface that are actually atomic requests. Sum over TCC instances."></metric>
<metric name="TCC_EA0_ATOMIC_LEVEL_sum" expr=reduce(TCC_EA0_ATOMIC_LEVEL,sum) descr="The sum of the number of EA atomics in flight. This is primarily meant for measure average EA atomic latency. Average atomic latency = TCC_PERF_SEL_EA_WRREQ_ATOMIC_LEVEL/TCC_PERF_SEL_EA_WRREQ_ATOMIC. Sum over TCC instances."></metric>
<metric name="TCC_EA0_RDREQ_sum" expr=reduce(TCC_EA0_RDREQ,sum) descr="Number of TCC/EA read requests (either 32-byte or 64-byte) Sum over TCC instances."></metric>
<metric name="TCC_EA0_RDREQ_32B_sum" expr=reduce(TCC_EA0_RDREQ_32B,sum) descr="Number of 32-byte TCC/EA read requests Sum over TCC instances."></metric>
<metric name="TCC_EA0_RD_UNCACHED_32B_sum" expr=reduce(TCC_EA0_RD_UNCACHED_32B,sum) descr="Number of 32-byte TCC/EA read due to uncached traffic. A 64-byte request will be counted as 2 Sum over TCC instances."></metric>
<metric name="TCC_EA0_RDREQ_IO_CREDIT_STALL_sum" expr=reduce(TCC_EA0_RDREQ_IO_CREDIT_STALL,sum) descr="Number of cycles there was a stall because the read request interface was out of IO credits. Stalls occur regardless of whether a read needed to be performed or not. Sum over TCC instances."></metric>
<metric name="TCC_EA0_RDREQ_GMI_CREDIT_STALL_sum" expr=reduce(TCC_EA0_RDREQ_GMI_CREDIT_STALL,sum) descr="Number of cycles there was a stall because the read request interface was out of GMI credits. Stalls occur regardless of whether a read needed to be performed or not. Sum over TCC instances."></metric>
<metric name="TCC_EA0_RDREQ_DRAM_CREDIT_STALL_sum" expr=reduce(TCC_EA0_RDREQ_DRAM_CREDIT_STALL,sum) descr="Number of cycles there was a stall because the read request interface was out of DRAM credits. Stalls occur regardless of whether a read needed to be performed or not. Sum over TCC instances."></metric>
<metric name="TCC_TAG_STALL_sum" expr=reduce(TCC_TAG_STALL,sum) descr="Total number of cycles the normal request pipeline in the tag is stalled for any reason."></metric>
<metric name="TCC_NORMAL_WRITEBACK_sum" expr=reduce(TCC_NORMAL_WRITEBACK,sum) descr="Number of writebacks due to requests that are not writeback requests. Sum over TCC instances."></metric>
<metric name="TCC_ALL_TC_OP_WB_WRITEBACK_sum" expr=reduce(TCC_ALL_TC_OP_WB_WRITEBACK,sum) descr="Number of writebacks due to all TC_OP writeback requests. Sum over TCC instances."></metric>
<metric name="TCC_NORMAL_EVICT_sum" expr=reduce(TCC_NORMAL_EVICT,sum) descr="Number of evictions due to requests that are not invalidate or probe requests. Sum over TCC instances."></metric>
<metric name="TCC_ALL_TC_OP_INV_EVICT_sum" expr=reduce(TCC_ALL_TC_OP_INV_EVICT,sum) descr="Number of evictions due to all TC_OP invalidate requests. Sum over TCC instances."></metric>
<metric name="TCC_EA0_RDREQ_DRAM_sum" expr=reduce(TCC_EA0_RDREQ_DRAM,sum) descr="Number of TCC/EA read requests (either 32-byte or 64-byte) destined for DRAM (MC). Sum over TCC instances."></metric>
<metric name="TCC_EA0_WRREQ_DRAM_sum" expr=reduce(TCC_EA0_WRREQ_DRAM,sum) descr="Number of TCC/EA write requests (either 32-byte of 64-byte) destined for DRAM (MC). Sum over TCC instances."></metric>
<metric name="FETCH_SIZE" expr=(TCC_EA0_RDREQ_32B_sum*32+(TCC_EA0_RDREQ_sum-TCC_EA0_RDREQ_32B_sum)*64)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_SIZE" expr=((TCC_EA0_WRREQ_sum-TCC_EA0_WRREQ_64B_sum)*32+TCC_EA0_WRREQ_64B_sum*64)/1024 descr="The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WRITE_REQ_32B" expr=TCC_EA0_WRREQ_64B_sum*2+(TCC_EA0_WRREQ_sum-TCC_EA0_WRREQ_64B_sum) descr="The total number of 32-byte effective memory writes."></metric>
<metric name="CU_OCCUPANCY" expr=(SQ_CYCLES/(SQ_WAVE_CYCLES*4))/MAX_WAVE_SIZE descr="The ratio of active waves on a CU to the maximum number of active waves supported by the CU"></metric>
<metric name="CU_UTILIZATION" expr=GRBM_GUI_ACTIVE/GRBM_COUNT descr="The total number of active cycles divided by total number of elapsed cycles"></metric>
<metric name="TOTAL_16_OPS" expr=(SQ_INSTS_VALU_FMA_F16*2+SQ_INSTS_VALU_ADD_F16+SQ_INSTS_VALU_MUL_F16+SQ_INSTS_VALU_TRANS_F16)*64+((SQ_INSTS_VALU_MFMA_MOPS_F16+SQ_INSTS_VALU_MFMA_MOPS_BF16)*512) descr="The number of 16 bits OPS executed"></metric>
<metric name="TOTAL_32_OPS" expr=(SQ_INSTS_VALU_FMA_F32*2+SQ_INSTS_VALU_INT32+SQ_INSTS_VALU_ADD_F32+SQ_INSTS_VALU_MUL_F32+SQ_INSTS_VALU_TRANS_F32)*64+(SQ_INSTS_VALU_MFMA_MOPS_F32*512) descr="The number of 32 bits OPS executed"></metric>
<metric name="TOTAL_64_OPS" expr=(SQ_INSTS_VALU_FMA_F64*2+SQ_INSTS_VALU_INT64+SQ_INSTS_VALU_ADD_F64+SQ_INSTS_VALU_MUL_F64)*64+(SQ_INSTS_VALU_MFMA_MOPS_F64*512) descr="The number of 64 bits OPS executed"></metric>
<metric name="GPU_UTIL" expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Percentage of the time that GUI is active"></metric>
</gfx940>
<gfx10 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="MeanOccupancyPerCU" expr=GRBM_COUNT*0+SQ_LEVEL_WAVES*0+SQ_ACCUM_PREV/GRBM_GUI_ACTIVE/CU_NUM descr="Mean occupancy per compute unit."></metric>
<metric name="MeanOccupancyPerActiveCU" expr=GRBM_COUNT*0+SQ_LEVEL_WAVES*0+SQ_ACCUM_PREV*4/SQ_BUSY_CYCLES/CU_NUM descr="Mean occupancy per active compute unit."></metric>
<metric name="GPU_UTIL" expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Percentage of the time that GUI is active"></metric>
<metric name="CP_UTIL" expr=100*GRBM_CP_BUSY/GRBM_GUI_ACTIVE descr="Percentage of the GRBM_GUI_ACTIVE time that any of the Command Processor (CPG/CPC/CPF) blocks are busy"></metric>
<metric name="SPI_UTIL" expr=100*GRBM_SPI_BUSY/GRBM_GUI_ACTIVE descr="Percentage of the GRBM_GUI_ACTIVE time that any of the Shader Pipe Interpolators (SPI) are busy in the shader engine(s)"></metric>
<metric name="TA_UTIL" expr=100*GRBM_TA_BUSY/GRBM_GUI_ACTIVE descr="Percentage of the GRBM_GUI_ACTIVE time that any of the Texture Pipes (TA) are busy in the shader engine(s)."></metric>
<metric name="GDS_UTIL" expr=100*GRBM_GDS_BUSY/GRBM_GUI_ACTIVE descr="Percentage of the GRBM_GUI_ACTIVE time that the Global Data Share (GDS) is busy."></metric>
<metric name="EA_UTIL" expr=100*GRBM_EA_BUSY/GRBM_GUI_ACTIVE descr="Percentage of the GRBM_GUI_ACTIVE time that the Efficiency Arbiter (EA) block is busy."></metric>
<metric name="WAVE_DEP_WAIT" expr=100*SQ_WAIT_ANY/SQ_WAVE_CYCLES descr="Percentage of the SQ_WAVE_CYCLE time spent waiting for anything."></metric>
<metric name="WAVE_ISSUE_WAIT" expr=100*SQ_WAIT_INST_ANY/SQ_WAVE_CYCLES descr="Percentage of the SQ_WAVE_CYCLE time spent waiting for any instruction issue."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
<metric name="TA_BUSY_min" expr=reduce(TA_TA_BUSY,min) descr="TA block is busy. Min over TA instances."></metric>
<metric name="TA_FLAT_LOAD_WAVEFRONTS_sum" expr=reduce(TA_FLAT_LOAD_WAVEFRONTS,sum) descr="Number of flat load vec32 packets processed by the TA. Sum over TA instances."></metric>
<metric name="TA_FLAT_STORE_WAVEFRONTS_sum" expr=reduce(TA_FLAT_STORE_WAVEFRONTS,sum) descr="Number of flat store vec32 packets processed by the TA. Sum over TA instances."></metric>
<metric name="GL2C_HIT_sum" expr=reduce(GL2C_HIT,sum) descr="Number of cache hits. Sum over GL2C instances."></metric>
<metric name="GL2C_MISS_sum" expr=reduce(GL2C_MISS,sum) descr="Number of cache misses. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_32B_sum" expr=reduce(GL2C_EA_RDREQ_32B,sum) descr="Number of 32-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_64B_sum" expr=reduce(GL2C_EA_RDREQ_64B,sum) descr="Number of 64-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_96B_sum" expr=reduce(GL2C_EA_RDREQ_96B,sum) descr="Number of 96-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_128B_sum" expr=reduce(GL2C_EA_RDREQ_128B,sum) descr="Number of 128-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_MC_RDREQ_sum" expr=reduce(GL2C_MC_RDREQ,sum) descr="Number of GL2C/EA read requests (either 32-byte or 64-byte or 128-byte). Sum over GL2C instances."></metric>
<metric name="GL2C_MC_WRREQ_sum" expr=reduce(GL2C_MC_WRREQ,sum) descr="Number of transactions (either 32-byte or 64-byte) going over the GL2C_MC_wrreq interface. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_WRREQ_64B_sum" expr=reduce(GL2C_EA_WRREQ_64B,sum) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the GL2C_EA_wrreq interface. Sum over GL2C instances."></metric>
<metric name="GL2C_WRREQ_STALL_max" expr=reduce(GL2C_MC_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over GL2C instances."></metric>
<metric name="L2CacheHit" expr=100*reduce(GL2C_HIT,sum)/(reduce(GL2C_HIT,sum)+reduce(GL2C_MISS,sum)) descr="The percentage of fetch, write, atomic, and other instructions that hit the data in L2 cache. Value range: 0% (no hit) to 100% (optimal)."></metric>
<metric name="FETCH_SIZE" expr=(GL2C_EA_RDREQ_32B_sum*32+GL2C_EA_RDREQ_64B_sum*64+GL2C_EA_RDREQ_96B_sum*96+GL2C_EA_RDREQ_128B_sum*128)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WriteUnitStalled" expr=100*GL2C_WRREQ_STALL_max/GRBM_GUI_ACTIVE descr="The percentage of GPUTime the Write unit is stalled. Value range: 0% to 100% (bad)."></metric>
<metric name="LDSBankConflict" expr=100*SQC_LDS_BANK_CONFLICT/SQC_LDS_IDX_ACTIVE descr="The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad)."></metric>
</gfx10>
<gfx1030 base="gfx10">
</gfx1030>
<gfx1031 base="gfx10">
</gfx1031>
<gfx1010 base="gfx10">
</gfx1010>
<gfx1032 base="gfx10">
</gfx1032>
<gfx11 base="common_derived">
<metric name="SQ_WAVES_sum" expr=reduce(SQ_WAVES,sum) descr="Count number of waves sent to SQs. (per-simd, emulated, global). Sum over SQ instances."></metric>
<metric name="GPU_UTIL" expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT descr="Percentage of the time that GUI is active"></metric>
<metric name="WAVE_DEP_WAIT" expr=100*SQ_WAIT_ANY/SQ_WAVE_CYCLES descr="Percentage of the SQ_WAVE_CYCLE time spent waiting for anything."></metric>
<metric name="WAVE_ISSUE_WAIT" expr=100*SQ_WAIT_INST_ANY/SQ_WAVE_CYCLES descr="Percentage of the SQ_WAVE_CYCLE time spent waiting for any instruction issue."></metric>
<metric name="TA_BUSY_avr" expr=reduce(TA_TA_BUSY,avr) descr="TA block is busy. Average over TA instances."></metric>
<metric name="TA_BUSY_max" expr=reduce(TA_TA_BUSY,max) descr="TA block is busy. Max over TA instances."></metric>
<metric name="TA_BUSY_min" expr=reduce(TA_TA_BUSY,min) descr="TA block is busy. Min over TA instances."></metric>
<metric name="TA_BUFFER_LOAD_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_LOAD_WAVEFRONTS,sum) descr="Number of buffer load vec32 packets processed by the TA. Sum over TA instances."></metric>
<metric name="TA_BUFFER_STORE_WAVEFRONTS_sum" expr=reduce(TA_BUFFER_STORE_WAVEFRONTS,sum) descr="Number of buffer store vec32 packets processed by the TA. Sum over TA instances."></metric>
<metric name="GL2C_HIT_sum" expr=reduce(GL2C_HIT,sum) descr="Number of cache hits. Sum over GL2C instances."></metric>
<metric name="GL2C_MISS_sum" expr=reduce(GL2C_MISS,sum) descr="Number of cache misses. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_32B_sum" expr=reduce(GL2C_EA_RDREQ_32B,sum) descr="Number of 32-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_64B_sum" expr=reduce(GL2C_EA_RDREQ_64B,sum) descr="Number of 64-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_96B_sum" expr=reduce(GL2C_EA_RDREQ_96B,sum) descr="Number of 96-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_RDREQ_128B_sum" expr=reduce(GL2C_EA_RDREQ_128B,sum) descr="Number of 128-byte GL2C/EA read requests. Sum over GL2C instances."></metric>
<metric name="GL2C_MC_RDREQ_sum" expr=reduce(GL2C_MC_RDREQ,sum) descr="Number of GL2C/EA read requests (either 32-byte or 64-byte or 128-byte). Sum over GL2C instances."></metric>
<metric name="GL2C_MC_WRREQ_sum" expr=reduce(GL2C_MC_WRREQ,sum) descr="Number of transactions (either 32-byte or 64-byte) going over the GL2C_MC_wrreq interface. Sum over GL2C instances."></metric>
<metric name="GL2C_EA_WRREQ_64B_sum" expr=reduce(GL2C_EA_WRREQ_64B,sum) descr="Number of 64-byte transactions going (64-byte write or CMPSWAP) over the GL2C_EA_wrreq interface. Sum over GL2C instances."></metric>
<metric name="GL2C_WRREQ_STALL_max" expr=reduce(GL2C_MC_WRREQ_STALL,max) descr="Number of cycles a write request was stalled. Max over GL2C instances."></metric>
<metric name="L2CacheHit" expr=100*reduce(GL2C_HIT,sum)/(reduce(GL2C_HIT,sum)+reduce(GL2C_MISS,sum)) descr="The percentage of fetch, write, atomic, and other instructions that hit the data in L2 cache. Value range: 0% (no hit) to 100% (optimal)."></metric>
<metric name="FETCH_SIZE" expr=(GL2C_EA_RDREQ_32B_sum*32+GL2C_EA_RDREQ_64B_sum*64+GL2C_EA_RDREQ_96B_sum*96+GL2C_EA_RDREQ_128B_sum*128)/1024 descr="The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account."></metric>
<metric name="WriteUnitStalled" expr=100*GL2C_WRREQ_STALL_max/GRBM_GUI_ACTIVE descr="The percentage of GPUTime the Write unit is stalled. Value range: 0% to 100% (bad)."></metric>
<metric name="LDSBankConflict" expr=100*SQC_LDS_BANK_CONFLICT/SQC_LDS_IDX_ACTIVE descr="The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad)."></metric>
</gfx11>
<gfx1100 base="gfx11">
</gfx1100>
<gfx1101 base="gfx11">
</gfx1101>
<gfx1102 base="gfx11">
</gfx1102>
<gfx11 base="gfx11"></gfx11>
#Mi300
<gfx941 base="gfx940"></gfx941>
<gfx942 base="gfx940"></gfx942>
#Navi21
<gfx1032 base="gfx1032"></gfx1032>