rocprofiler_kernel_dispatch_info_t + header record for buffered counter collection (#758)

* Update include/rocprofiler-sdk

- defines.h
  - ROCPROFILER_VERSION_10_0 -> ROCPROFILER_SDK_VERSION_0_0
- fwd.h
  - rocprofiler_counter_record_kind_t
  - rocprofiler_kernel_dispatch_info_t
  - rocprofiler_record_counter_t
    - has dispatch id instead of correlation id
  - rocprofiler_counter_info_v0_t
    - added rocprofiler_counter_id_t field
    - added is_constant field
    - reordered better packing
- dispatch_profile.h
  - added rocprofiler_profile_counting_dispatch_record_t for use as a header record for rocprofiler_profile_counting_dispatch_data_t
- callback_tracing.h
  - rocprofiler_callback_tracing_kernel_dispatch_data_t uses rocprofiler_kernel_dispatch_info_t
- buffer_tracing.h
  - rocprofiler_buffer_tracing_kernel_dispatch_record_t uses rocprofiler_kernel_dispatch_info_t

* Update lib/rocprofiler-sdk/*

- transition to rocprofiler_kernel_dispatch_info_t
- set id and is_constant values for rocprofiler_counter_info_v0_t in rocprofiler_query_counter_info

* Update lib/rocprofiler-sdk-tool

- transition to rocprofiler_kernel_dispatch_info_t

* Update lib/rocprofiler-sdk/counters/tests/core.cpp

- transition to rocprofiler_kernel_dispatch_info_t

* Update samples

- transition to rocprofiler_kernel_dispatch_info_t
- transition to rocprofiler_counter_record_kind_t

* Update tests

- transition to rocprofiler_kernel_dispatch_info_t
- transition to rocprofiler_counter_record_kind_t
- improve integration test validation for counter-collection
- update serialization for new/additional types

* Fix tests/counter-collection/validate.py

- loosen restrictions on the length of counter description

* Update include/rocprofiler-sdk/buffer_tracing.h

- remove accidental packed attribute

* Update lib/rocprofiler-sdk/counters/xml/derived_counters.xml

- Add description for TCC_TAG_STALL_sum (reference: https://rocm.docs.amd.com/en/develop/conceptual/gpu-arch/mi300-mi200-performance-counters.html)

* Update tests/page-migration/validate.py
This commit is contained in:
Jonathan R. Madsen
2024-04-12 17:30:34 -05:00
committato da GitHub
parent fd3d97287c
commit 07537b6231
23 ha cambiato i file con 489 aggiunte e 273 eliminazioni
+26 -17
Vedi File
@@ -275,23 +275,32 @@ WriteInterceptor(const void* packets,
corr_id->add_kern_count();
}
auto dispatch_id = ++sequence_counter;
auto callback_record =
callback_record_t{sizeof(callback_record_t),
rocprofiler_timestamp_t{0},
rocprofiler_timestamp_t{0},
queue.get_agent().get_rocp_agent()->id,
queue.get_id(),
kernel_id,
dispatch_id,
kernel_pkt.kernel_dispatch.private_segment_size,
kernel_pkt.kernel_dispatch.group_segment_size,
rocprofiler_dim3_t{kernel_pkt.kernel_dispatch.workgroup_size_x,
kernel_pkt.kernel_dispatch.workgroup_size_y,
kernel_pkt.kernel_dispatch.workgroup_size_z},
rocprofiler_dim3_t{kernel_pkt.kernel_dispatch.grid_size_x,
kernel_pkt.kernel_dispatch.grid_size_y,
kernel_pkt.kernel_dispatch.grid_size_z}};
// computes the "size" based on the offset of reserved_padding field
constexpr auto kernel_dispatch_info_rt_size =
common::compute_runtime_sizeof<rocprofiler_kernel_dispatch_info_t>();
static_assert(kernel_dispatch_info_rt_size < sizeof(rocprofiler_kernel_dispatch_info_t),
"failed to compute size field based on offset of reserved_padding field");
auto dispatch_id = ++sequence_counter;
auto callback_record = callback_record_t{
sizeof(callback_record_t),
rocprofiler_timestamp_t{0},
rocprofiler_timestamp_t{0},
rocprofiler_kernel_dispatch_info_t{
.size = kernel_dispatch_info_rt_size,
.agent_id = queue.get_agent().get_rocp_agent()->id,
.queue_id = queue.get_id(),
.kernel_id = kernel_id,
.dispatch_id = dispatch_id,
.private_segment_size = kernel_pkt.kernel_dispatch.private_segment_size,
.group_segment_size = kernel_pkt.kernel_dispatch.group_segment_size,
.workgroup_size = rocprofiler_dim3_t{kernel_pkt.kernel_dispatch.workgroup_size_x,
kernel_pkt.kernel_dispatch.workgroup_size_y,
kernel_pkt.kernel_dispatch.workgroup_size_z},
.grid_size = rocprofiler_dim3_t{kernel_pkt.kernel_dispatch.grid_size_x,
kernel_pkt.kernel_dispatch.grid_size_y,
kernel_pkt.kernel_dispatch.grid_size_z}}};
{
auto tracer_data = callback_record;