AI/rocm-systems
2
0
Fork 0
程式碼 問題 合併請求 Actions 套件 專案 版本發布 Wiki 動態

[rocprofiler-sdk] Add derived metrics for Navi4 (#238)

瀏覽代碼 
* add more derived metrics for navi4.

* addr comments

* addr comments, and add more derived counters.

* EOF.

* misc.

* remove duplicate counter.

* misc.

* Remove gfx12 architecture definition for ldslatency

* remove extra architectures for gfx12.

* use wgp for normalization

* move these changes to another PR.

---------

Co-authored-by: Venkateshwar Reddy Kandula <venkateshwar.kandula1306@gmail.com>
此提交包含在：
Venkateshwar Reddy Kandula
2025-09-12 02:30:10 -05:00
提交者 GitHub
父節點 7d847dde3f
當前提交 26e7c4231e
共有 1 個檔案被更改，包括 140 行新增 和 0 行删除
下載 Patch 檔 下載差異檔 展開所有檔案 折疊所有檔案
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/counter_defs.yaml
+140
查看文件
@@ -467,6 +467,9 @@ rocprofiler-sdk:
- gfx941
- gfx942
- gfx950
- gfx12
- gfx1200
- gfx1201
expression: simd_count
- name: CpUtil
description: 'Unit: percent'
@@ -627,6 +630,9 @@ rocprofiler-sdk:
- gfx941
- gfx942
- gfx950
- gfx12
- gfx1200
- gfx1201
expression: FETCH_SIZE
- name: FlatLDSInsts
description: The average number of FLAT instructions that read or write to LDS executed per work item (affected by flow
@@ -1037,6 +1043,12 @@ rocprofiler-sdk:
- gfx1102
block: GL2C
event: 88
- architectures:
- gfx12
- gfx1200
- gfx1201
block: GL2C
event: 122
- name: GL2C_MC_WRREQ_sum
description: Number of transactions (either 32-byte or 64-byte) going over the GL2C_MC_wrreq interface. Sum over GL2C
instances.
@@ -1107,6 +1119,9 @@ rocprofiler-sdk:
- gfx1100
- gfx1101
- gfx1102
- gfx12
- gfx1200
- gfx1201
expression: reduce(GL2C_MC_WRREQ_STALL,max)
- name: GPUBusy
description: The percentage of time GPU was busy.
@@ -1127,6 +1142,9 @@ rocprofiler-sdk:
- gfx906
- gfx908
- gfx90a
- gfx12
- gfx1200
- gfx1201
expression: 100*reduce(GRBM_GUI_ACTIVE,max)/reduce(GRBM_COUNT,max)
- name: GPU_UTIL
description: Percentage of the time that GUI is active
@@ -1498,7 +1516,16 @@ rocprofiler-sdk:
definitions:
- architectures:
- gfx90a
- gfx940
- gfx941
- gfx942
- gfx950
expression: 100*reduce(SQ_LDS_IDX_ACTIVE,sum)/(reduce(GRBM_GUI_ACTIVE,max)*CU_NUM)
- architectures:
- gfx12
- gfx1200
- gfx1201
expression: 100*reduce(SQC_LDS_IDX_ACTIVE,sum)/(reduce(GRBM_GUI_ACTIVE,max)*(SIMD_NUM/4))
- name: MAX_WAVE_SIZE
description: Max wave size constant
properties: []
@@ -1532,6 +1559,9 @@ rocprofiler-sdk:
- gfx1100
- gfx1101
- gfx1102
- gfx12
- gfx1200
- gfx1201
expression: reduce(SQ_WAVE_CYCLES,sum)/reduce(SQ_BUSY_CYCLES,sum)
- architectures:
- gfx90a
@@ -1549,6 +1579,9 @@ rocprofiler-sdk:
- gfx1100
- gfx1101
- gfx1102
- gfx12
- gfx1200
- gfx1201
expression: reduce(SQ_WAVE_CYCLES,sum)/reduce(GRBM_GUI_ACTIVE,max)/CU_NUM
- architectures:
- gfx10
@@ -1576,6 +1609,9 @@ rocprofiler-sdk:
- gfx1100
- gfx1101
- gfx1102
- gfx12
- gfx1200
- gfx1201
expression: 100*reduce(SQ_WAVE_CYCLES,sum)/reduce(GRBM_GUI_ACTIVE,max)/CU_NUM/32
- architectures:
- gfx90a
@@ -1604,6 +1640,9 @@ rocprofiler-sdk:
- gfx906
- gfx908
- gfx90a
- gfx12
- gfx1200
- gfx1201
expression: 100*reduce(TA_TA_BUSY,max)/reduce(GRBM_GUI_ACTIVE,max)
- name: MemUnitStalled
description: 'The percentage of GPUTime the memory unit is stalled. Try reducing the number or size of fetches and writes
@@ -1742,6 +1781,9 @@ rocprofiler-sdk:
- gfx906
- gfx908
- gfx90a
- gfx12
- gfx1200
- gfx1201
expression: reduce(SQ_INSTS_SALU,sum)/reduce(SQ_WAVES,sum)
- name: SE_NUM
description: SE_NUM
@@ -1766,6 +1808,9 @@ rocprofiler-sdk:
- gfx941
- gfx942
- gfx950
- gfx12
- gfx1200
- gfx1201
expression: array_count/simd_arrays_per_engine
- name: SFetchInsts
description: The average number of scalar fetch instructions from the video memory executed per work-item (affected by
@@ -1786,6 +1831,9 @@ rocprofiler-sdk:
- gfx906
- gfx908
- gfx90a
- gfx12
- gfx1200
- gfx1201
expression: reduce(SQ_INSTS_SMEM,sum)/reduce(SQ_WAVES,sum)
- name: SPI_CSN_BUSY
description: Number of clocks with outstanding waves (SPI or SH). Requires SPI_DEBUG_CNTL.DEBUG_PIPE_SEL to select source,
@@ -2725,6 +2773,22 @@ rocprofiler-sdk:
- gfx950
block: SQ
event: 271
- architectures:
- gfx12
- gfx1200
- gfx1201
block: SQ
event: 302
- name: SQC_ICACHE_MISSES
description: Number of cache misses, includes uncached requests. {per-Bank, nondeterministic, C2}
properties: []
definitions:
- architectures:
- gfx12
- gfx1200
- gfx1201
block: SQ
event: 303
- name: SQC_ICACHE_INPUT_VALID_READYB
description: ' Input stalled by SQC (per-SQ, nondeterministic, unwindowed)'
properties: []
@@ -2775,6 +2839,12 @@ rocprofiler-sdk:
- gfx950
block: SQ
event: 270
- architectures:
- gfx12
- gfx1200
- gfx1201
block: SQ
event: 301
- name: SQC_LDS_BANK_CONFLICT
description: Number of cycles LDS is stalled by bank conflicts. (emulated, C1)
properties: []
@@ -5434,6 +5504,27 @@ rocprofiler-sdk:
- gfx950
block: SQ
event: 16
- name: SQ_INST_CYCLES_VALU
description: Number of cycles needed to execute VALU operations (SIMD cycles), where there is overlapping V_OP32_1
and V_OP32_T instruction, count them separately.
properties: []
definitions:
- architectures:
- gfx1201
- gfx12
- gfx1200
block: SQ
event: 99
- name: SQ_INSTS_VEC32_LEVEL_LDS
description: Number of in-flight wave32 LDS (indexed, flat) instructions issued.{level, nondeterministic}
properties: []
definitions:
- architectures:
- gfx1201
- gfx12
- gfx1200
block: SQ
event: 250
- name: ScaPipeIssueUtil
description: 'Unit: percent'
properties: []
@@ -9850,6 +9941,26 @@ rocprofiler-sdk:
- architectures:
- gfx950
expression: reduce(TCP_UTCL1_STALL_UTCL2_REQ_OUT_OF_CREDITS,sum)
- name: TCP_REQ
description: Total cache line accesses
properties: []
definitions:
- architectures:
- gfx12
- gfx1200
- gfx1201
block: TCP
event: 9
- name: TCP_REQ_MISS
description: Total cache requests that missed
properties: []
definitions:
- architectures:
- gfx12
- gfx1200
- gfx1201
block: TCP
event: 17
- name: TD_ATOMIC_WAVEFRONT
description: Count the wavefronts with opcode = atomic.
properties: []
@@ -10159,6 +10270,11 @@ rocprofiler-sdk:
- gfx942
- gfx950
expression: 100*reduce(SQ_ACTIVE_INST_VALU,sum)/CU_NUM/reduce(GRBM_GUI_ACTIVE,max)
- architectures:
- gfx12
- gfx1200
- gfx1201
expression: 100*reduce(SQ_INST_CYCLES_VALU,sum)/(SIMD_NUM/4)/reduce(GRBM_GUI_ACTIVE,max)
- name: VALUInsts
description: The average number of vector ALU instructions executed per work-item (affected by flow control).
properties: []
@@ -10177,6 +10293,9 @@ rocprofiler-sdk:
- gfx906
- gfx908
- gfx90a
- gfx12
- gfx1200
- gfx1201
expression: reduce(SQ_INSTS_VALU,sum)/reduce(SQ_WAVES,sum)
- name: VALUUtilization
description: 'The percentage of active vector ALU threads in a wave. A lower number can mean either more thread divergence
@@ -10255,6 +10374,11 @@ rocprofiler-sdk:
- gfx941
- gfx942
expression: 100*reduce(SQ_ACTIVE_INST_VALU,sum)/(reduce(GRBM_GUI_ACTIVE,max)*CU_NUM)
- architectures:
- gfx12
- gfx1200
- gfx1201
expression: 100*reduce(SQ_INST_CYCLES_VALU,sum)/(reduce(GRBM_GUI_ACTIVE,max)*(SIMD_NUM/4))
- name: VmemLatency
description: 'Unit: cycles'
properties: []
@@ -10412,6 +10536,9 @@ rocprofiler-sdk:
- gfx906
- gfx908
- gfx90a
- gfx12
- gfx1200
- gfx1201
expression: reduce(SQ_WAVES,sum)
- name: WriteSize
description: The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or
@@ -10450,6 +10577,9 @@ rocprofiler-sdk:
- gfx1100
- gfx1101
- gfx1102
- gfx12
- gfx1200
- gfx1201
expression: 100*GL2C_WRREQ_STALL_max/reduce(GRBM_GUI_ACTIVE,max)
- name: sL1dCacheHitRate
description: 'Unit: percent'
@@ -10555,3 +10685,13 @@ rocprofiler-sdk:
- gfx942
- gfx950
expression: TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES_sum/TCP_GATE_EN1_sum
- name: L0CacheHit
description: The percentage of read requests that hit the data in the L0 cache. The L0 cache contains vector data, which
is data that may vary in each thread across the wavefront. Value range 0% (no hit) to 100% (optimal).
properties: []
definitions:
- architectures:
- gfx12
- gfx1200
- gfx1201
expression: (1-(reduce(TCP_REQ_MISS,sum)/reduce(TCP_REQ,sum)))*100