From 97465e7448388e337e7b886bc7e8b9977e975bd0 Mon Sep 17 00:00:00 2001 From: vedithal-amd Date: Sat, 26 Jul 2025 15:33:23 -0400 Subject: [PATCH] add description for MI100 counters (#834) [ROCm/rocprofiler-compute commit: 6885cb068d61f5112f276dbfc7d90f3007726fbc] --- .../profile_configs/gfx908_counter_defs.yaml | 43 ++++++++++++------- 1 file changed, 28 insertions(+), 15 deletions(-) diff --git a/projects/rocprofiler-compute/src/rocprof_compute_soc/profile_configs/gfx908_counter_defs.yaml b/projects/rocprofiler-compute/src/rocprof_compute_soc/profile_configs/gfx908_counter_defs.yaml index 644a8843be..64bbcba4d2 100644 --- a/projects/rocprofiler-compute/src/rocprof_compute_soc/profile_configs/gfx908_counter_defs.yaml +++ b/projects/rocprofiler-compute/src/rocprof_compute_soc/profile_configs/gfx908_counter_defs.yaml @@ -2250,35 +2250,40 @@ rocprofiler-sdk: - gfx908 expression: reduce(TCC_WRITEBACK,sum) - name: TCC_EA_WR_UNCACHED_32B_sum - description: . + description: 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. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_WR_UNCACHED_32B,sum) - name: TCC_EA_WRREQ_STALL_sum - description: . + description: Number of cycles a write request was stalled. Sum over TCC instances. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_WRREQ_STALL,sum) - name: TCC_EA_WRREQ_IO_CREDIT_STALL_sum - description: . + description: Number of cycles a EA write request was stalled because the interface was out of IO credits. Sum over TCC + instances. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_WRREQ_IO_CREDIT_STALL,sum) - name: TCC_EA_WRREQ_GMI_CREDIT_STALL_sum - description: . + description: Number of cycles a EA write request was stalled because the interface was out of GMI credits. Sum over TCC + instances. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_WRREQ_GMI_CREDIT_STALL,sum) - name: TCC_EA_WRREQ_DRAM_CREDIT_STALL_sum - description: . + description: Number of cycles a EA write request was stalled because the interface was out of DRAM credits. Sum over TCC + instances. properties: [] definitions: - architectures: @@ -2292,56 +2297,64 @@ rocprofiler-sdk: - gfx908 expression: reduce(TCC_TOO_MANY_EA_WRREQS_STALL,sum) - name: TCC_EA_WRREQ_LEVEL_sum - description: . + description: 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. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_WRREQ_LEVEL,sum) - name: TCC_EA_RDREQ_LEVEL_sum - description: . + description: 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. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_RDREQ_LEVEL,sum) - name: TCC_EA_ATOMIC_sum - description: . + description: Number of transactions going over the TC_EA_wrreq interface that are actually atomic requests. Sum over TCC + instances. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_ATOMIC,sum) - name: TCC_EA_ATOMIC_LEVEL_sum - description: . + description: 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. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_ATOMIC_LEVEL,sum) - name: TCC_EA_RD_UNCACHED_32B_sum - description: . + description: Number of 32-byte TCC/EA read due to uncached traffic. A 64-byte request will be counted as 2 Sum over TCC + instances. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_RD_UNCACHED_32B,sum) - name: TCC_EA_RDREQ_IO_CREDIT_STALL_sum - description: . + description: 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. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_RDREQ_IO_CREDIT_STALL,sum) - name: TCC_EA_RDREQ_GMI_CREDIT_STALL_sum - description: . + description: Number of cycles a EA write request was stalled because the interface was out of GMI credits. Sum over TCC + instances. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_RDREQ_GMI_CREDIT_STALL,sum) - name: TCC_EA_RDREQ_DRAM_CREDIT_STALL_sum - description: . + description: 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. properties: [] definitions: - architectures: @@ -2376,14 +2389,14 @@ rocprofiler-sdk: - gfx908 expression: reduce(TCC_ALL_TC_OP_INV_EVICT,sum) - name: TCC_EA_RDREQ_DRAM_sum - description: . + description: Number of TCC/EA read requests (either 32-byte or 64-byte) destined for DRAM (MC). Sum over TCC instances. properties: [] definitions: - architectures: - gfx908 expression: reduce(TCC_EA_RDREQ_DRAM,sum) - name: TCC_EA_WRREQ_DRAM_sum - description: . + description: Number of TCC/EA write requests (either 32-byte of 64-byte) destined for DRAM (MC). Sum over TCC instances. properties: [] definitions: - architectures: