From 2dccd4a3ebd06b377afcecf5eaec058ec8dc9ddd Mon Sep 17 00:00:00 2001 From: Christopher Milan Date: Mon, 11 May 2026 16:22:12 -0700 Subject: [PATCH] am: autogen pmc (#16143) * am: autogen pmc * cleanup * fix * type --- tinygrad/runtime/autogen/am/__init__.py | 13 +- tinygrad/runtime/autogen/am/pmc.py | 3350 +++++++++++++++++++++++ tinygrad/runtime/support/amd.py | 20 +- 3 files changed, 3366 insertions(+), 17 deletions(-) create mode 100644 tinygrad/runtime/autogen/am/pmc.py diff --git a/tinygrad/runtime/autogen/am/__init__.py b/tinygrad/runtime/autogen/am/__init__.py index 6c5df51368..54c5f2cc92 100644 --- a/tinygrad/runtime/autogen/am/__init__.py +++ b/tinygrad/runtime/autogen/am/__init__.py @@ -2,13 +2,14 @@ import pathlib, hashlib, re, itertools from tinygrad.runtime.autogen import load, root __all__ = ["am", "pm4_soc15", "pm4_nv", "sdma_4_0_0", "sdma_5_0_0", "sdma_6_0_0", "smu_13_0_0", "smu_13_0_6", "smu_13_0_12", "smu_14_0_2", - "fw", "navi_offsets", "vega_offsets", "regs", "soc_9", "soc_11", "soc_12"] + "fw", "navi_offsets", "vega_offsets", "regs", "soc_9", "soc_11", "soc_12", "pmc"] am_src="https://github.com/ROCm/ROCK-Kernel-Driver/archive/33970e1351f5e511029602454979f3de7e22260f.tar.gz" rocm_src="https://github.com/ROCm/rocm-systems/archive/cccc350dc620e61ae2554978b62ab3532dc10bd9.tar.gz" AMD, AMDINC = "{}/drivers/gpu/drm/amd", "{}/drivers/gpu/drm/amd/include" inc, kern_rules = ["-include", "stdint.h"], [(r'le32_to_cpu', ''),] fw_src="https://gitlab.com/kernel-firmware/linux-firmware/-/archive/1e2c15348485939baf1b6d1f5a7a3b799d80703d/1e2c15348485939baf1b6d1f5a7a3b799d80703d.tar.gz" +pmc_src="https://raw.githubusercontent.com/ROCm/rocm-systems/cccc350dc620e61ae2554978b62ab3532dc10bd9/projects/rocprofiler-compute/src/rocprof_compute_soc/profile_configs/counter_defs.yaml" reg_files = { "gc": [(9,4,3), (11,0,0), (11,0,3), (11,5,0), (12,0,0)], @@ -88,4 +89,14 @@ def __getattr__(nm): case "soc_9" | "soc_11" | "soc_12": return load(f"am/{nm}", ["{}/projects/aqlprofile/linux/" + {9: "vega10", 11: "soc21", 12: "soc24"}[int(nm.split('_')[1])] + "_enum.h"], srcs=rocm_src, patterns=soc_patterns, macros=False) + case "pmc": + def genpmc(_, files, **kwargs): + from yaml import safe_load # type: ignore + with open(files[0], "r") as f: data = safe_load(f) + out = ["counters = {"] + for counter in [c for c in data['rocprofiler-sdk']['counters'] if any('block' in d for d in c['definitions'])]: + out.extend([f" {counter['name']!r}: {{", + *[f" {a!r}: ({d['block']!r}, {d['event']})," for d in counter['definitions'] for a in d['architectures']], " },"]) + return "\n".join(out + ["}"]) + return load("am/pmc", ["{}/counter_defs.yaml"], srcs=pmc_src, gen=genpmc) case _: raise AttributeError(f"no such autogen: {nm}") diff --git a/tinygrad/runtime/autogen/am/pmc.py b/tinygrad/runtime/autogen/am/pmc.py new file mode 100644 index 0000000000..ee8658e084 --- /dev/null +++ b/tinygrad/runtime/autogen/am/pmc.py @@ -0,0 +1,3350 @@ +counters = { + 'CPC_CPC_STAT_BUSY': { + 'gfx90a': ('CPC', 25), + 'gfx908': ('CPC', 25), + 'gfx940': ('CPC', 25), + 'gfx941': ('CPC', 25), + 'gfx942': ('CPC', 25), + 'gfx950': ('CPC', 25), + }, + 'CPC_CPC_STAT_IDLE': { + 'gfx90a': ('CPC', 26), + 'gfx908': ('CPC', 26), + 'gfx940': ('CPC', 26), + 'gfx941': ('CPC', 26), + 'gfx942': ('CPC', 26), + 'gfx950': ('CPC', 26), + }, + 'CPC_CPC_STAT_STALL': { + 'gfx90a': ('CPC', 27), + 'gfx908': ('CPC', 27), + 'gfx940': ('CPC', 27), + 'gfx941': ('CPC', 27), + 'gfx942': ('CPC', 27), + 'gfx950': ('CPC', 27), + }, + 'CPC_CPC_TCIU_BUSY': { + 'gfx90a': ('CPC', 28), + 'gfx908': ('CPC', 28), + 'gfx940': ('CPC', 28), + 'gfx941': ('CPC', 28), + 'gfx942': ('CPC', 28), + 'gfx950': ('CPC', 28), + }, + 'CPC_CPC_TCIU_IDLE': { + 'gfx90a': ('CPC', 29), + 'gfx908': ('CPC', 29), + 'gfx940': ('CPC', 29), + 'gfx941': ('CPC', 29), + 'gfx942': ('CPC', 29), + 'gfx950': ('CPC', 29), + }, + 'CPC_CPC_UTCL2IU_BUSY': { + 'gfx90a': ('CPC', 30), + 'gfx908': ('CPC', 30), + 'gfx940': ('CPC', 30), + 'gfx941': ('CPC', 30), + 'gfx942': ('CPC', 30), + 'gfx950': ('CPC', 30), + }, + 'CPC_CPC_UTCL2IU_IDLE': { + 'gfx90a': ('CPC', 31), + 'gfx908': ('CPC', 31), + 'gfx940': ('CPC', 31), + 'gfx941': ('CPC', 31), + 'gfx942': ('CPC', 31), + 'gfx950': ('CPC', 31), + }, + 'CPC_CPC_UTCL2IU_STALL': { + 'gfx90a': ('CPC', 32), + 'gfx908': ('CPC', 32), + 'gfx940': ('CPC', 32), + 'gfx941': ('CPC', 32), + 'gfx942': ('CPC', 32), + 'gfx950': ('CPC', 32), + }, + 'CPC_ME1_BUSY_FOR_PACKET_DECODE': { + 'gfx90a': ('CPC', 13), + 'gfx908': ('CPC', 13), + 'gfx940': ('CPC', 13), + 'gfx941': ('CPC', 13), + 'gfx942': ('CPC', 13), + 'gfx950': ('CPC', 13), + }, + 'CPC_ME1_DC0_SPI_BUSY': { + 'gfx90a': ('CPC', 33), + 'gfx908': ('CPC', 33), + 'gfx940': ('CPC', 33), + 'gfx941': ('CPC', 33), + 'gfx942': ('CPC', 33), + 'gfx950': ('CPC', 33), + }, + 'CPC_UTCL1_STALL_ON_TRANSLATION': { + 'gfx90a': ('CPC', 24), + 'gfx908': ('CPC', 24), + 'gfx940': ('CPC', 24), + 'gfx941': ('CPC', 24), + 'gfx942': ('CPC', 24), + 'gfx950': ('CPC', 24), + }, + 'CPC_ALWAYS_COUNT': { + 'gfx950': ('CPC', 0), + }, + 'CPC_ADC_VALID_CHUNK_NOT_AVAIL': { + 'gfx950': ('CPC', 3), + }, + 'CPC_ADC_DISPATCH_ALLOC_DONE': { + 'gfx950': ('CPC', 4), + }, + 'CPC_ADC_VALID_CHUNK_END': { + 'gfx950': ('CPC', 9), + }, + 'CPC_SYNC_FIFO_FULL_LEVEL': { + 'gfx950': ('CPC', 43), + }, + 'CPC_SYNC_FIFO_FULL': { + 'gfx950': ('CPC', 44), + }, + 'CPC_GD_BUSY': { + 'gfx950': ('CPC', 61), + }, + 'CPC_TG_SEND': { + 'gfx950': ('CPC', 62), + }, + 'CPC_WALK_NEXT_CHUNK': { + 'gfx950': ('CPC', 63), + }, + 'CPC_STALLED_BY_SE0_SPI': { + 'gfx950': ('CPC', 64), + }, + 'CPC_STALLED_BY_SE1_SPI': { + 'gfx950': ('CPC', 65), + }, + 'CPC_STALLED_BY_SE2_SPI': { + 'gfx950': ('CPC', 66), + }, + 'CPC_STALLED_BY_SE3_SPI': { + 'gfx950': ('CPC', 67), + }, + 'CPC_LTE_ALL': { + 'gfx950': ('CPC', 68), + }, + 'CPC_SYNC_WRREQ_FIFO_BUSY': { + 'gfx950': ('CPC', 69), + }, + 'CPC_CANE_BUSY': { + 'gfx950': ('CPC', 70), + }, + 'CPC_CANE_STALL': { + 'gfx950': ('CPC', 71), + }, + 'CPF_CMP_UTCL1_STALL_ON_TRANSLATION': { + 'gfx90a': ('CPF', 20), + 'gfx908': ('CPF', 20), + 'gfx940': ('CPF', 20), + 'gfx941': ('CPF', 20), + 'gfx942': ('CPF', 20), + 'gfx950': ('CPF', 20), + }, + 'CPF_CPF_STAT_BUSY': { + 'gfx90a': ('CPF', 23), + 'gfx908': ('CPF', 23), + 'gfx940': ('CPF', 23), + 'gfx941': ('CPF', 23), + 'gfx942': ('CPF', 23), + 'gfx950': ('CPF', 23), + }, + 'CPF_CPF_STAT_IDLE': { + 'gfx90a': ('CPF', 24), + 'gfx908': ('CPF', 24), + 'gfx940': ('CPF', 24), + 'gfx941': ('CPF', 24), + 'gfx942': ('CPF', 24), + 'gfx950': ('CPF', 24), + }, + 'CPF_CPF_STAT_STALL': { + 'gfx90a': ('CPF', 25), + 'gfx908': ('CPF', 25), + 'gfx940': ('CPF', 25), + 'gfx941': ('CPF', 25), + 'gfx942': ('CPF', 25), + 'gfx950': ('CPF', 25), + }, + 'CPF_CPF_TCIU_BUSY': { + 'gfx90a': ('CPF', 26), + 'gfx908': ('CPF', 26), + 'gfx940': ('CPF', 26), + 'gfx941': ('CPF', 26), + 'gfx942': ('CPF', 26), + 'gfx950': ('CPF', 26), + }, + 'CPF_CPF_TCIU_IDLE': { + 'gfx90a': ('CPF', 27), + 'gfx908': ('CPF', 27), + 'gfx940': ('CPF', 27), + 'gfx941': ('CPF', 27), + 'gfx942': ('CPF', 27), + 'gfx950': ('CPF', 27), + }, + 'CPF_CPF_TCIU_STALL': { + 'gfx90a': ('CPF', 28), + 'gfx908': ('CPF', 28), + 'gfx940': ('CPF', 28), + 'gfx941': ('CPF', 28), + 'gfx942': ('CPF', 28), + 'gfx950': ('CPF', 28), + }, + 'GL2C_EA_RDREQ': { + 'gfx12': ('GL2C', 140), + 'gfx1200': ('GL2C', 140), + 'gfx1201': ('GL2C', 140), + }, + 'GL2C_EA_RDREQ_128B': { + 'gfx10': ('GL2C', 102), + 'gfx1010': ('GL2C', 102), + 'gfx1030': ('GL2C', 102), + 'gfx1031': ('GL2C', 102), + 'gfx1032': ('GL2C', 102), + 'gfx11': ('GL2C', 102), + 'gfx1100': ('GL2C', 102), + 'gfx1101': ('GL2C', 102), + 'gfx1102': ('GL2C', 102), + 'gfx12': ('GL2C', 148), + 'gfx1200': ('GL2C', 148), + 'gfx1201': ('GL2C', 148), + }, + 'GL2C_EA_RDREQ_32B': { + 'gfx10': ('GL2C', 99), + 'gfx1010': ('GL2C', 99), + 'gfx1030': ('GL2C', 99), + 'gfx1031': ('GL2C', 99), + 'gfx1032': ('GL2C', 99), + 'gfx11': ('GL2C', 99), + 'gfx1100': ('GL2C', 99), + 'gfx1101': ('GL2C', 99), + 'gfx1102': ('GL2C', 99), + 'gfx12': ('GL2C', 146), + 'gfx1200': ('GL2C', 146), + 'gfx1201': ('GL2C', 146), + }, + 'GL2C_EA_RDREQ_64B': { + 'gfx10': ('GL2C', 100), + 'gfx1010': ('GL2C', 100), + 'gfx1030': ('GL2C', 100), + 'gfx1031': ('GL2C', 100), + 'gfx1032': ('GL2C', 100), + 'gfx11': ('GL2C', 100), + 'gfx1100': ('GL2C', 100), + 'gfx1101': ('GL2C', 100), + 'gfx1102': ('GL2C', 100), + 'gfx12': ('GL2C', 147), + 'gfx1200': ('GL2C', 147), + 'gfx1201': ('GL2C', 147), + }, + 'GL2C_EA_RDREQ_96B': { + 'gfx10': ('GL2C', 101), + 'gfx1010': ('GL2C', 101), + 'gfx1030': ('GL2C', 101), + 'gfx1031': ('GL2C', 101), + 'gfx1032': ('GL2C', 101), + 'gfx11': ('GL2C', 101), + 'gfx1100': ('GL2C', 101), + 'gfx1101': ('GL2C', 101), + 'gfx1102': ('GL2C', 101), + }, + 'GL2C_EA_WRREQ': { + 'gfx12': ('GL2C', 108), + 'gfx1200': ('GL2C', 108), + 'gfx1201': ('GL2C', 108), + }, + 'GL2C_EA_WRREQ_STALL': { + 'gfx12': ('GL2C', 122), + 'gfx1200': ('GL2C', 122), + 'gfx1201': ('GL2C', 122), + }, + 'GL2C_EA_WRREQ_64B': { + 'gfx10': ('GL2C', 85), + 'gfx1010': ('GL2C', 85), + 'gfx1030': ('GL2C', 85), + 'gfx1031': ('GL2C', 85), + 'gfx1032': ('GL2C', 85), + 'gfx11': ('GL2C', 85), + 'gfx1100': ('GL2C', 85), + 'gfx1101': ('GL2C', 85), + 'gfx1102': ('GL2C', 85), + 'gfx12': ('GL2C', 114), + 'gfx1200': ('GL2C', 114), + 'gfx1201': ('GL2C', 114), + }, + 'GL2C_HIT': { + 'gfx10': ('GL2C', 42), + 'gfx1010': ('GL2C', 42), + 'gfx1030': ('GL2C', 42), + 'gfx1031': ('GL2C', 42), + 'gfx1032': ('GL2C', 42), + 'gfx11': ('GL2C', 42), + 'gfx1100': ('GL2C', 42), + 'gfx1101': ('GL2C', 42), + 'gfx1102': ('GL2C', 42), + 'gfx12': ('GL2C', 41), + 'gfx1200': ('GL2C', 41), + 'gfx1201': ('GL2C', 41), + }, + 'GL2C_MC_RDREQ': { + 'gfx10': ('GL2C', 96), + 'gfx1010': ('GL2C', 96), + 'gfx1030': ('GL2C', 96), + 'gfx1031': ('GL2C', 96), + 'gfx1032': ('GL2C', 96), + 'gfx11': ('GL2C', 96), + 'gfx1100': ('GL2C', 96), + 'gfx1101': ('GL2C', 96), + 'gfx1102': ('GL2C', 96), + }, + 'GL2C_MC_WRREQ': { + 'gfx10': ('GL2C', 83), + 'gfx1010': ('GL2C', 83), + 'gfx1030': ('GL2C', 83), + 'gfx1031': ('GL2C', 83), + 'gfx1032': ('GL2C', 83), + 'gfx11': ('GL2C', 83), + 'gfx1100': ('GL2C', 83), + 'gfx1101': ('GL2C', 83), + 'gfx1102': ('GL2C', 83), + }, + 'GL2C_MC_WRREQ_STALL': { + 'gfx10': ('GL2C', 88), + 'gfx1010': ('GL2C', 88), + 'gfx1030': ('GL2C', 88), + 'gfx1031': ('GL2C', 88), + 'gfx1032': ('GL2C', 88), + 'gfx11': ('GL2C', 88), + 'gfx1100': ('GL2C', 88), + 'gfx1101': ('GL2C', 88), + 'gfx1102': ('GL2C', 88), + }, + 'GL2C_MISS': { + 'gfx10': ('GL2C', 43), + 'gfx1010': ('GL2C', 43), + 'gfx1030': ('GL2C', 43), + 'gfx1031': ('GL2C', 43), + 'gfx1032': ('GL2C', 43), + 'gfx11': ('GL2C', 43), + 'gfx1100': ('GL2C', 43), + 'gfx1101': ('GL2C', 43), + 'gfx1102': ('GL2C', 43), + 'gfx12': ('GL2C', 42), + 'gfx1200': ('GL2C', 42), + 'gfx1201': ('GL2C', 42), + }, + 'GRBM_COUNT': { + 'gfx10': ('GRBM', 0), + 'gfx1010': ('GRBM', 0), + 'gfx1030': ('GRBM', 0), + 'gfx1031': ('GRBM', 0), + 'gfx1032': ('GRBM', 0), + 'gfx11': ('GRBM', 0), + 'gfx1100': ('GRBM', 0), + 'gfx1101': ('GRBM', 0), + 'gfx1102': ('GRBM', 0), + 'gfx12': ('GRBM', 0), + 'gfx1200': ('GRBM', 0), + 'gfx1201': ('GRBM', 0), + 'gfx9': ('GRBM', 0), + 'gfx900': ('GRBM', 0), + 'gfx906': ('GRBM', 0), + 'gfx908': ('GRBM', 0), + 'gfx90a': ('GRBM', 0), + 'gfx940': ('GRBM', 0), + 'gfx941': ('GRBM', 0), + 'gfx942': ('GRBM', 0), + 'gfx950': ('GRBM', 0), + }, + 'GRBM_CPC_BUSY': { + 'gfx90a': ('GRBM', 30), + 'gfx908': ('GRBM', 30), + 'gfx940': ('GRBM', 30), + 'gfx941': ('GRBM', 30), + 'gfx942': ('GRBM', 30), + 'gfx950': ('GRBM', 30), + }, + 'GRBM_CPF_BUSY': { + 'gfx90a': ('GRBM', 31), + 'gfx908': ('GRBM', 31), + 'gfx940': ('GRBM', 31), + 'gfx941': ('GRBM', 31), + 'gfx942': ('GRBM', 31), + 'gfx950': ('GRBM', 31), + }, + 'GRBM_CP_BUSY': { + 'gfx10': ('GRBM', 3), + 'gfx1010': ('GRBM', 3), + 'gfx1030': ('GRBM', 3), + 'gfx1031': ('GRBM', 3), + 'gfx1032': ('GRBM', 3), + 'gfx90a': ('GRBM', 3), + 'gfx908': ('GRBM', 3), + 'gfx940': ('GRBM', 3), + 'gfx941': ('GRBM', 3), + 'gfx942': ('GRBM', 3), + 'gfx950': ('GRBM', 3), + }, + 'GRBM_EA_BUSY': { + 'gfx10': ('GRBM', 35), + 'gfx1010': ('GRBM', 35), + 'gfx1030': ('GRBM', 35), + 'gfx1031': ('GRBM', 35), + 'gfx1032': ('GRBM', 35), + 'gfx90a': ('GRBM', 35), + 'gfx908': ('GRBM', 35), + 'gfx940': ('GRBM', 35), + 'gfx941': ('GRBM', 35), + 'gfx942': ('GRBM', 35), + 'gfx950': ('GRBM', 35), + }, + 'GRBM_GDS_BUSY': { + 'gfx10': ('GRBM', 25), + 'gfx1010': ('GRBM', 25), + 'gfx1030': ('GRBM', 25), + 'gfx1031': ('GRBM', 25), + 'gfx1032': ('GRBM', 25), + }, + 'GRBM_GL2CC_BUSY': { + 'gfx10': ('GRBM', 40), + 'gfx1010': ('GRBM', 40), + 'gfx1030': ('GRBM', 40), + 'gfx1031': ('GRBM', 40), + 'gfx1032': ('GRBM', 40), + }, + 'GRBM_GUI_ACTIVE': { + 'gfx10': ('GRBM', 2), + 'gfx1010': ('GRBM', 2), + 'gfx1030': ('GRBM', 2), + 'gfx1031': ('GRBM', 2), + 'gfx1032': ('GRBM', 2), + 'gfx11': ('GRBM', 2), + 'gfx1100': ('GRBM', 2), + 'gfx1101': ('GRBM', 2), + 'gfx1102': ('GRBM', 2), + 'gfx12': ('GRBM', 2), + 'gfx1200': ('GRBM', 2), + 'gfx1201': ('GRBM', 2), + 'gfx9': ('GRBM', 2), + 'gfx900': ('GRBM', 2), + 'gfx906': ('GRBM', 2), + 'gfx908': ('GRBM', 2), + 'gfx90a': ('GRBM', 2), + 'gfx940': ('GRBM', 2), + 'gfx941': ('GRBM', 2), + 'gfx942': ('GRBM', 2), + 'gfx950': ('GRBM', 2), + }, + 'GRBM_SPI_BUSY': { + 'gfx10': ('GRBM', 11), + 'gfx1010': ('GRBM', 11), + 'gfx1030': ('GRBM', 11), + 'gfx1031': ('GRBM', 11), + 'gfx1032': ('GRBM', 11), + 'gfx90a': ('GRBM', 11), + 'gfx908': ('GRBM', 11), + 'gfx940': ('GRBM', 11), + 'gfx941': ('GRBM', 11), + 'gfx942': ('GRBM', 11), + 'gfx950': ('GRBM', 11), + }, + 'GRBM_TA_BUSY': { + 'gfx10': ('GRBM', 13), + 'gfx1010': ('GRBM', 13), + 'gfx1030': ('GRBM', 13), + 'gfx1031': ('GRBM', 13), + 'gfx1032': ('GRBM', 13), + 'gfx90a': ('GRBM', 13), + 'gfx908': ('GRBM', 13), + 'gfx940': ('GRBM', 13), + 'gfx941': ('GRBM', 13), + 'gfx942': ('GRBM', 13), + 'gfx950': ('GRBM', 13), + }, + 'GRBM_TC_BUSY': { + 'gfx90a': ('GRBM', 28), + 'gfx908': ('GRBM', 28), + 'gfx940': ('GRBM', 28), + 'gfx941': ('GRBM', 28), + 'gfx942': ('GRBM', 28), + 'gfx950': ('GRBM', 28), + }, + 'GRBM_UTCL2_BUSY': { + 'gfx90a': ('GRBM', 34), + 'gfx908': ('GRBM', 34), + 'gfx940': ('GRBM', 34), + 'gfx941': ('GRBM', 34), + 'gfx942': ('GRBM', 34), + 'gfx950': ('GRBM', 34), + }, + 'SPI_CSN_BUSY': { + 'gfx90a': ('SPI', 48), + 'gfx908': ('SPI', 48), + 'gfx940': ('SPI', 48), + 'gfx941': ('SPI', 48), + 'gfx942': ('SPI', 48), + 'gfx950': ('SPI', 48), + }, + 'SPI_CSN_NUM_THREADGROUPS': { + 'gfx90a': ('SPI', 49), + 'gfx908': ('SPI', 49), + 'gfx940': ('SPI', 49), + 'gfx941': ('SPI', 49), + 'gfx942': ('SPI', 49), + 'gfx950': ('SPI', 49), + }, + 'SPI_CSN_WAVE': { + 'gfx90a': ('SPI', 52), + 'gfx908': ('SPI', 52), + 'gfx940': ('SPI', 52), + 'gfx941': ('SPI', 52), + 'gfx942': ('SPI', 52), + 'gfx950': ('SPI', 52), + }, + 'SPI_CSN_WINDOW_VALID': { + 'gfx90a': ('SPI', 47), + 'gfx908': ('SPI', 47), + 'gfx940': ('SPI', 47), + 'gfx941': ('SPI', 47), + 'gfx942': ('SPI', 47), + 'gfx950': ('SPI', 47), + }, + 'SPI_RA_BAR_CU_FULL_CSN': { + 'gfx90a': ('SPI', 123), + 'gfx908': ('SPI', 123), + 'gfx940': ('SPI', 123), + 'gfx941': ('SPI', 123), + 'gfx942': ('SPI', 123), + 'gfx950': ('SPI', 123), + }, + 'SPI_RA_BULKY_CU_FULL_CSN': { + 'gfx90a': ('SPI', 125), + 'gfx908': ('SPI', 125), + 'gfx940': ('SPI', 125), + 'gfx941': ('SPI', 125), + 'gfx942': ('SPI', 125), + 'gfx950': ('SPI', 125), + }, + 'SPI_RA_LDS_CU_FULL_CSN': { + 'gfx90a': ('SPI', 120), + 'gfx908': ('SPI', 120), + 'gfx940': ('SPI', 120), + 'gfx941': ('SPI', 120), + 'gfx942': ('SPI', 120), + 'gfx950': ('SPI', 120), + }, + 'SPI_RA_REQ_NO_ALLOC': { + 'gfx90a': ('SPI', 79), + 'gfx908': ('SPI', 79), + 'gfx940': ('SPI', 79), + 'gfx941': ('SPI', 79), + 'gfx942': ('SPI', 79), + 'gfx950': ('SPI', 79), + }, + 'SPI_RA_REQ_NO_ALLOC_CSN': { + 'gfx90a': ('SPI', 85), + 'gfx908': ('SPI', 85), + 'gfx940': ('SPI', 85), + 'gfx941': ('SPI', 85), + 'gfx942': ('SPI', 85), + 'gfx950': ('SPI', 85), + }, + 'SPI_RA_RES_STALL_CSN': { + 'gfx90a': ('SPI', 91), + 'gfx908': ('SPI', 91), + 'gfx940': ('SPI', 91), + 'gfx941': ('SPI', 91), + 'gfx942': ('SPI', 91), + 'gfx950': ('SPI', 91), + }, + 'SPI_RA_SGPR_SIMD_FULL_CSN': { + 'gfx90a': ('SPI', 115), + 'gfx908': ('SPI', 115), + 'gfx940': ('SPI', 115), + 'gfx941': ('SPI', 115), + 'gfx942': ('SPI', 115), + 'gfx950': ('SPI', 115), + }, + 'SPI_RA_TGLIM_CU_FULL_CSN': { + 'gfx90a': ('SPI', 127), + 'gfx908': ('SPI', 127), + 'gfx940': ('SPI', 127), + 'gfx941': ('SPI', 127), + 'gfx942': ('SPI', 127), + 'gfx950': ('SPI', 127), + }, + 'SPI_RA_TMP_STALL_CSN': { + 'gfx90a': ('SPI', 97), + 'gfx908': ('SPI', 97), + 'gfx940': ('SPI', 97), + 'gfx941': ('SPI', 97), + 'gfx942': ('SPI', 97), + 'gfx950': ('SPI', 97), + }, + 'SPI_RA_VGPR_SIMD_FULL_CSN': { + 'gfx90a': ('SPI', 109), + 'gfx908': ('SPI', 109), + 'gfx940': ('SPI', 109), + 'gfx941': ('SPI', 109), + 'gfx942': ('SPI', 109), + 'gfx950': ('SPI', 109), + }, + 'SPI_RA_WAVE_SIMD_FULL_CSN': { + 'gfx90a': ('SPI', 103), + 'gfx908': ('SPI', 103), + 'gfx940': ('SPI', 103), + 'gfx941': ('SPI', 103), + 'gfx942': ('SPI', 103), + 'gfx950': ('SPI', 103), + }, + 'SPI_RA_WVLIM_STALL_CSN': { + 'gfx90a': ('SPI', 133), + 'gfx908': ('SPI', 133), + 'gfx940': ('SPI', 133), + 'gfx941': ('SPI', 133), + 'gfx942': ('SPI', 133), + 'gfx950': ('SPI', 133), + }, + 'SPI_SWC_CSC_WR': { + 'gfx90a': ('SPI', 189), + 'gfx908': ('SPI', 189), + 'gfx940': ('SPI', 189), + 'gfx941': ('SPI', 189), + 'gfx942': ('SPI', 189), + 'gfx950': ('SPI', 189), + }, + 'SPI_VWC_CSC_WR': { + 'gfx90a': ('SPI', 195), + 'gfx908': ('SPI', 195), + 'gfx940': ('SPI', 195), + 'gfx941': ('SPI', 195), + 'gfx942': ('SPI', 195), + 'gfx950': ('SPI', 195), + }, + 'SPI_CS0_WINDOW_VALID': { + 'gfx950': ('SPI', 0), + }, + 'SPI_CS0_BUSY': { + 'gfx950': ('SPI', 1), + }, + 'SPI_CS0_NUM_THREADGROUPS': { + 'gfx950': ('SPI', 2), + }, + 'SPI_CS0_CRAWLER_STALL': { + 'gfx950': ('SPI', 3), + }, + 'SPI_CS0_EVENT_WAVE': { + 'gfx950': ('SPI', 4), + }, + 'SPI_CS0_WAVE': { + 'gfx950': ('SPI', 5), + }, + 'SPI_CS1_WINDOW_VALID': { + 'gfx950': ('SPI', 6), + }, + 'SPI_CS1_BUSY': { + 'gfx950': ('SPI', 7), + }, + 'SPI_CS1_NUM_THREADGROUPS': { + 'gfx950': ('SPI', 8), + }, + 'SPI_CS1_CRAWLER_STALL': { + 'gfx950': ('SPI', 9), + }, + 'SPI_CS1_EVENT_WAVE': { + 'gfx950': ('SPI', 10), + }, + 'SPI_CS1_WAVE': { + 'gfx950': ('SPI', 11), + }, + 'SPI_CS2_WINDOW_VALID': { + 'gfx950': ('SPI', 12), + }, + 'SPI_CS2_BUSY': { + 'gfx950': ('SPI', 13), + }, + 'SPI_CS2_NUM_THREADGROUPS': { + 'gfx950': ('SPI', 14), + }, + 'SPI_CS2_CRAWLER_STALL': { + 'gfx950': ('SPI', 15), + }, + 'SPI_CS2_EVENT_WAVE': { + 'gfx950': ('SPI', 16), + }, + 'SPI_CS2_WAVE': { + 'gfx950': ('SPI', 17), + }, + 'SPI_CS3_WINDOW_VALID': { + 'gfx950': ('SPI', 18), + }, + 'SPI_CS3_BUSY': { + 'gfx950': ('SPI', 19), + }, + 'SPI_CS3_NUM_THREADGROUPS': { + 'gfx950': ('SPI', 20), + }, + 'SPI_CS3_CRAWLER_STALL': { + 'gfx950': ('SPI', 21), + }, + 'SPI_CS3_EVENT_WAVE': { + 'gfx950': ('SPI', 22), + }, + 'SPI_CS3_WAVE': { + 'gfx950': ('SPI', 23), + }, + 'SPI_CSQ_P0_Q0_OCCUPANCY': { + 'gfx950': ('SPI', 140), + }, + 'SPI_CSQ_P0_Q1_OCCUPANCY': { + 'gfx950': ('SPI', 141), + }, + 'SPI_CSQ_P0_Q2_OCCUPANCY': { + 'gfx950': ('SPI', 142), + }, + 'SPI_CSQ_P0_Q3_OCCUPANCY': { + 'gfx950': ('SPI', 143), + }, + 'SPI_CSQ_P0_Q4_OCCUPANCY': { + 'gfx950': ('SPI', 144), + }, + 'SPI_CSQ_P0_Q5_OCCUPANCY': { + 'gfx950': ('SPI', 145), + }, + 'SPI_CSQ_P0_Q6_OCCUPANCY': { + 'gfx950': ('SPI', 146), + }, + 'SPI_CSQ_P0_Q7_OCCUPANCY': { + 'gfx950': ('SPI', 147), + }, + 'SPI_CSQ_P1_Q0_OCCUPANCY': { + 'gfx950': ('SPI', 148), + }, + 'SPI_CSQ_P1_Q1_OCCUPANCY': { + 'gfx950': ('SPI', 149), + }, + 'SPI_CSQ_P1_Q2_OCCUPANCY': { + 'gfx950': ('SPI', 150), + }, + 'SPI_CSQ_P1_Q3_OCCUPANCY': { + 'gfx950': ('SPI', 151), + }, + 'SPI_CSQ_P1_Q4_OCCUPANCY': { + 'gfx950': ('SPI', 152), + }, + 'SPI_CSQ_P1_Q5_OCCUPANCY': { + 'gfx950': ('SPI', 153), + }, + 'SPI_CSQ_P1_Q6_OCCUPANCY': { + 'gfx950': ('SPI', 154), + }, + 'SPI_CSQ_P1_Q7_OCCUPANCY': { + 'gfx950': ('SPI', 155), + }, + 'SPI_CSQ_P2_Q0_OCCUPANCY': { + 'gfx950': ('SPI', 156), + }, + 'SPI_CSQ_P2_Q1_OCCUPANCY': { + 'gfx950': ('SPI', 157), + }, + 'SPI_CSQ_P2_Q2_OCCUPANCY': { + 'gfx950': ('SPI', 158), + }, + 'SPI_CSQ_P2_Q3_OCCUPANCY': { + 'gfx950': ('SPI', 159), + }, + 'SPI_CSQ_P2_Q4_OCCUPANCY': { + 'gfx950': ('SPI', 160), + }, + 'SPI_CSQ_P2_Q5_OCCUPANCY': { + 'gfx950': ('SPI', 161), + }, + 'SPI_CSQ_P2_Q6_OCCUPANCY': { + 'gfx950': ('SPI', 162), + }, + 'SPI_CSQ_P2_Q7_OCCUPANCY': { + 'gfx950': ('SPI', 163), + }, + 'SPI_CSQ_P3_Q0_OCCUPANCY': { + 'gfx950': ('SPI', 164), + }, + 'SPI_CSQ_P3_Q1_OCCUPANCY': { + 'gfx950': ('SPI', 165), + }, + 'SPI_CSQ_P3_Q2_OCCUPANCY': { + 'gfx950': ('SPI', 166), + }, + 'SPI_CSQ_P3_Q3_OCCUPANCY': { + 'gfx950': ('SPI', 167), + }, + 'SPI_CSQ_P3_Q4_OCCUPANCY': { + 'gfx950': ('SPI', 168), + }, + 'SPI_CSQ_P3_Q5_OCCUPANCY': { + 'gfx950': ('SPI', 169), + }, + 'SPI_CSQ_P3_Q6_OCCUPANCY': { + 'gfx950': ('SPI', 170), + }, + 'SPI_CSQ_P3_Q7_OCCUPANCY': { + 'gfx950': ('SPI', 171), + }, + 'SPI_CSQ_P0_OCCUPANCY': { + 'gfx950': ('SPI', 172), + }, + 'SPI_CSQ_P1_OCCUPANCY': { + 'gfx950': ('SPI', 173), + }, + 'SPI_CSQ_P2_OCCUPANCY': { + 'gfx950': ('SPI', 174), + }, + 'SPI_CSQ_P3_OCCUPANCY': { + 'gfx950': ('SPI', 175), + }, + 'SPI_VWC0_VDATA_VALID_WR': { + 'gfx950': ('SPI', 193), + }, + 'SPI_VWC1_VDATA_VALID_WR': { + 'gfx950': ('SPI', 194), + }, + 'SPI_CSC_WAVE_CNT_BUSY': { + 'gfx950': ('SPI', 225), + }, + 'SQC_DCACHE_ATOMIC': { + 'gfx90a': ('SQ', 298), + 'gfx908': ('SQ', 298), + 'gfx940': ('SQ', 298), + 'gfx941': ('SQ', 298), + 'gfx942': ('SQ', 298), + 'gfx950': ('SQ', 298), + }, + 'SQC_DCACHE_BUSY_CYCLES': { + 'gfx940': ('SQ', 289), + 'gfx941': ('SQ', 289), + 'gfx942': ('SQ', 289), + 'gfx950': ('SQ', 289), + }, + 'SQC_DCACHE_HITS': { + 'gfx90a': ('SQ', 291), + 'gfx908': ('SQ', 291), + 'gfx940': ('SQ', 291), + 'gfx941': ('SQ', 291), + 'gfx942': ('SQ', 291), + 'gfx950': ('SQ', 291), + }, + 'SQC_DCACHE_INPUT_VALID_READYB': { + 'gfx90a': ('SQ', 260), + 'gfx908': ('SQ', 260), + 'gfx940': ('SQ', 260), + 'gfx941': ('SQ', 260), + 'gfx942': ('SQ', 260), + 'gfx950': ('SQ', 260), + }, + 'SQC_DCACHE_MISSES': { + 'gfx90a': ('SQ', 292), + 'gfx908': ('SQ', 292), + 'gfx940': ('SQ', 292), + 'gfx941': ('SQ', 292), + 'gfx942': ('SQ', 292), + 'gfx950': ('SQ', 292), + }, + 'SQC_DCACHE_MISSES_DUPLICATE': { + 'gfx90a': ('SQ', 293), + 'gfx908': ('SQ', 293), + 'gfx940': ('SQ', 293), + 'gfx941': ('SQ', 293), + 'gfx942': ('SQ', 293), + 'gfx950': ('SQ', 293), + }, + 'SQC_DCACHE_REQ': { + 'gfx90a': ('SQ', 290), + 'gfx908': ('SQ', 290), + 'gfx940': ('SQ', 290), + 'gfx941': ('SQ', 290), + 'gfx942': ('SQ', 290), + 'gfx950': ('SQ', 290), + }, + 'SQC_DCACHE_REQ_READ_1': { + 'gfx90a': ('SQ', 323), + 'gfx908': ('SQ', 323), + 'gfx940': ('SQ', 323), + 'gfx941': ('SQ', 323), + 'gfx942': ('SQ', 323), + 'gfx950': ('SQ', 323), + }, + 'SQC_DCACHE_REQ_READ_16': { + 'gfx90a': ('SQ', 327), + 'gfx908': ('SQ', 327), + 'gfx940': ('SQ', 327), + 'gfx941': ('SQ', 327), + 'gfx942': ('SQ', 327), + 'gfx950': ('SQ', 327), + }, + 'SQC_DCACHE_REQ_READ_2': { + 'gfx90a': ('SQ', 324), + 'gfx908': ('SQ', 324), + 'gfx940': ('SQ', 324), + 'gfx941': ('SQ', 324), + 'gfx942': ('SQ', 324), + 'gfx950': ('SQ', 324), + }, + 'SQC_DCACHE_REQ_READ_4': { + 'gfx90a': ('SQ', 325), + 'gfx908': ('SQ', 325), + 'gfx940': ('SQ', 325), + 'gfx941': ('SQ', 325), + 'gfx942': ('SQ', 325), + 'gfx950': ('SQ', 325), + }, + 'SQC_DCACHE_REQ_READ_8': { + 'gfx90a': ('SQ', 326), + 'gfx908': ('SQ', 326), + 'gfx940': ('SQ', 326), + 'gfx941': ('SQ', 326), + 'gfx942': ('SQ', 326), + 'gfx950': ('SQ', 326), + }, + 'SQC_ICACHE_BUSY_CYCLES': { + 'gfx940': ('SQ', 269), + 'gfx941': ('SQ', 269), + 'gfx942': ('SQ', 269), + 'gfx950': ('SQ', 269), + }, + 'SQC_ICACHE_HITS': { + 'gfx90a': ('SQ', 271), + 'gfx908': ('SQ', 271), + 'gfx940': ('SQ', 271), + 'gfx941': ('SQ', 271), + 'gfx942': ('SQ', 271), + 'gfx950': ('SQ', 271), + 'gfx12': ('SQ', 302), + 'gfx1200': ('SQ', 302), + 'gfx1201': ('SQ', 302), + }, + 'SQC_ICACHE_INPUT_VALID_READYB': { + 'gfx940': ('SQ', 257), + 'gfx941': ('SQ', 257), + 'gfx942': ('SQ', 257), + 'gfx950': ('SQ', 257), + }, + 'SQC_ICACHE_MISSES': { + 'gfx90a': ('SQ', 272), + 'gfx908': ('SQ', 272), + 'gfx940': ('SQ', 272), + 'gfx941': ('SQ', 272), + 'gfx942': ('SQ', 272), + 'gfx950': ('SQ', 272), + 'gfx12': ('SQ', 303), + 'gfx1200': ('SQ', 303), + 'gfx1201': ('SQ', 303), + }, + 'SQC_ICACHE_MISSES_DUPLICATE': { + 'gfx90a': ('SQ', 273), + 'gfx908': ('SQ', 273), + 'gfx940': ('SQ', 273), + 'gfx941': ('SQ', 273), + 'gfx942': ('SQ', 273), + 'gfx950': ('SQ', 273), + }, + 'SQC_ICACHE_REQ': { + 'gfx90a': ('SQ', 270), + 'gfx908': ('SQ', 270), + 'gfx940': ('SQ', 270), + 'gfx941': ('SQ', 270), + 'gfx942': ('SQ', 270), + 'gfx950': ('SQ', 270), + 'gfx12': ('SQ', 301), + 'gfx1200': ('SQ', 301), + 'gfx1201': ('SQ', 301), + }, + 'SQC_LDS_BANK_CONFLICT': { + 'gfx10': ('SQ', 285), + 'gfx1010': ('SQ', 285), + 'gfx1030': ('SQ', 285), + 'gfx1031': ('SQ', 285), + 'gfx1032': ('SQ', 285), + 'gfx11': ('SQ', 256), + 'gfx1100': ('SQ', 256), + 'gfx1101': ('SQ', 256), + 'gfx1102': ('SQ', 256), + 'gfx12': ('SQ', 288), + 'gfx1200': ('SQ', 288), + 'gfx1201': ('SQ', 288), + }, + 'SQC_LDS_IDX_ACTIVE': { + 'gfx10': ('SQ', 290), + 'gfx1010': ('SQ', 290), + 'gfx1030': ('SQ', 290), + 'gfx1031': ('SQ', 290), + 'gfx1032': ('SQ', 290), + 'gfx11': ('SQ', 261), + 'gfx1100': ('SQ', 261), + 'gfx1101': ('SQ', 261), + 'gfx1102': ('SQ', 261), + 'gfx12': ('SQ', 293), + 'gfx1200': ('SQ', 293), + 'gfx1201': ('SQ', 293), + }, + 'SQC_TC_DATA_ATOMIC_REQ': { + 'gfx90a': ('SQ', 266), + 'gfx908': ('SQ', 266), + 'gfx940': ('SQ', 266), + 'gfx941': ('SQ', 266), + 'gfx942': ('SQ', 266), + 'gfx950': ('SQ', 266), + }, + 'SQC_TC_DATA_READ_REQ': { + 'gfx90a': ('SQ', 264), + 'gfx908': ('SQ', 264), + 'gfx940': ('SQ', 264), + 'gfx941': ('SQ', 264), + 'gfx942': ('SQ', 264), + 'gfx950': ('SQ', 264), + }, + 'SQC_TC_DATA_WRITE_REQ': { + 'gfx90a': ('SQ', 265), + 'gfx908': ('SQ', 265), + 'gfx940': ('SQ', 265), + 'gfx941': ('SQ', 265), + 'gfx942': ('SQ', 265), + 'gfx950': ('SQ', 265), + }, + 'SQC_TC_INST_REQ': { + 'gfx90a': ('SQ', 263), + 'gfx908': ('SQ', 263), + 'gfx940': ('SQ', 263), + 'gfx941': ('SQ', 263), + 'gfx942': ('SQ', 263), + 'gfx950': ('SQ', 263), + }, + 'SQC_TC_REQ': { + 'gfx90a': ('SQ', 262), + 'gfx908': ('SQ', 262), + 'gfx940': ('SQ', 262), + 'gfx941': ('SQ', 262), + 'gfx942': ('SQ', 262), + 'gfx950': ('SQ', 262), + }, + 'SQC_TC_STALL': { + 'gfx90a': ('SQ', 267), + 'gfx908': ('SQ', 267), + 'gfx940': ('SQ', 267), + 'gfx941': ('SQ', 267), + 'gfx942': ('SQ', 267), + 'gfx950': ('SQ', 267), + }, + 'SQ_ACCUM_PREV': { + 'gfx10': ('SQ', 1), + 'gfx1010': ('SQ', 1), + 'gfx1030': ('SQ', 1), + 'gfx1031': ('SQ', 1), + 'gfx1032': ('SQ', 1), + 'gfx11': ('SQ', 1), + 'gfx1100': ('SQ', 1), + 'gfx1101': ('SQ', 1), + 'gfx1102': ('SQ', 1), + 'gfx12': ('SQ', 1), + 'gfx1200': ('SQ', 1), + 'gfx1201': ('SQ', 1), + 'gfx90a': ('SQ', 1), + 'gfx908': ('SQ', 1), + 'gfx940': ('SQ', 1), + 'gfx941': ('SQ', 1), + 'gfx942': ('SQ', 1), + 'gfx950': ('SQ', 1), + }, + 'SQ_ACCUM_PREV_HIRES': { + 'gfx90a': ('SQ', 185), + 'gfx908': ('SQ', 158), + 'gfx940': ('SQ', 184), + 'gfx941': ('SQ', 184), + 'gfx942': ('SQ', 184), + 'gfx950': ('SQ', 200), + }, + 'SQ_ACTIVE_INST_ANY': { + 'gfx90a': ('SQ', 96), + 'gfx908': ('SQ', 69), + 'gfx940': ('SQ', 101), + 'gfx941': ('SQ', 101), + 'gfx942': ('SQ', 101), + 'gfx950': ('SQ', 117), + }, + 'SQ_ACTIVE_INST_EXP_GDS': { + 'gfx90a': ('SQ', 101), + 'gfx908': ('SQ', 74), + 'gfx940': ('SQ', 106), + 'gfx941': ('SQ', 106), + 'gfx942': ('SQ', 106), + 'gfx950': ('SQ', 122), + }, + 'SQ_ACTIVE_INST_FLAT': { + 'gfx90a': ('SQ', 103), + 'gfx908': ('SQ', 76), + 'gfx940': ('SQ', 108), + 'gfx941': ('SQ', 108), + 'gfx942': ('SQ', 108), + 'gfx950': ('SQ', 124), + }, + 'SQ_ACTIVE_INST_LDS': { + 'gfx90a': ('SQ', 98), + 'gfx908': ('SQ', 71), + 'gfx940': ('SQ', 103), + 'gfx941': ('SQ', 103), + 'gfx942': ('SQ', 103), + 'gfx950': ('SQ', 119), + }, + 'SQ_ACTIVE_INST_MISC': { + 'gfx90a': ('SQ', 102), + 'gfx908': ('SQ', 75), + 'gfx940': ('SQ', 107), + 'gfx941': ('SQ', 107), + 'gfx942': ('SQ', 107), + 'gfx950': ('SQ', 123), + }, + 'SQ_ACTIVE_INST_SCA': { + 'gfx90a': ('SQ', 100), + 'gfx908': ('SQ', 73), + 'gfx940': ('SQ', 105), + 'gfx941': ('SQ', 105), + 'gfx942': ('SQ', 105), + 'gfx950': ('SQ', 121), + }, + 'SQ_ACTIVE_INST_VALU': { + 'gfx9': ('SQ', 71), + 'gfx900': ('SQ', 71), + 'gfx906': ('SQ', 71), + 'gfx908': ('SQ', 72), + 'gfx90a': ('SQ', 99), + 'gfx940': ('SQ', 104), + 'gfx941': ('SQ', 104), + 'gfx942': ('SQ', 104), + 'gfx950': ('SQ', 120), + }, + 'SQ_ACTIVE_INST_VMEM': { + 'gfx90a': ('SQ', 97), + 'gfx908': ('SQ', 70), + 'gfx940': ('SQ', 102), + 'gfx941': ('SQ', 102), + 'gfx942': ('SQ', 102), + 'gfx950': ('SQ', 118), + }, + 'SQ_BUSY_CU_CYCLES': { + 'gfx90a': ('SQ', 13), + 'gfx908': ('SQ', 13), + 'gfx940': ('SQ', 13), + 'gfx941': ('SQ', 13), + 'gfx942': ('SQ', 13), + 'gfx950': ('SQ', 13), + }, + 'SQ_BUSY_CYCLES': { + 'gfx10': ('SQ', 3), + 'gfx1010': ('SQ', 3), + 'gfx1030': ('SQ', 3), + 'gfx1031': ('SQ', 3), + 'gfx1032': ('SQ', 3), + 'gfx11': ('SQ', 3), + 'gfx1100': ('SQ', 3), + 'gfx1101': ('SQ', 3), + 'gfx1102': ('SQ', 3), + 'gfx12': ('SQ', 3), + 'gfx1200': ('SQ', 3), + 'gfx1201': ('SQ', 3), + 'gfx90a': ('SQ', 3), + 'gfx908': ('SQ', 3), + 'gfx940': ('SQ', 3), + 'gfx941': ('SQ', 3), + 'gfx942': ('SQ', 3), + 'gfx950': ('SQ', 3), + }, + 'SQ_CYCLES': { + 'gfx90a': ('SQ', 2), + 'gfx908': ('SQ', 2), + 'gfx940': ('SQ', 2), + 'gfx941': ('SQ', 2), + 'gfx942': ('SQ', 2), + 'gfx950': ('SQ', 2), + }, + 'SQ_IFETCH': { + 'gfx90a': ('SQ', 115), + 'gfx908': ('SQ', 88), + 'gfx940': ('SQ', 120), + 'gfx941': ('SQ', 120), + 'gfx942': ('SQ', 120), + 'gfx950': ('SQ', 136), + }, + 'SQ_IFETCH_LEVEL': { + 'gfx90a': ('SQ', 116), + 'gfx908': ('SQ', 89), + 'gfx940': ('SQ', 121), + 'gfx941': ('SQ', 121), + 'gfx942': ('SQ', 121), + 'gfx950': ('SQ', 137), + }, + 'SQ_INSTS': { + 'gfx90a': ('SQ', 25), + 'gfx908': ('SQ', 25), + 'gfx940': ('SQ', 25), + 'gfx941': ('SQ', 25), + 'gfx942': ('SQ', 25), + 'gfx950': ('SQ', 25), + }, + 'SQ_INSTS_BRANCH': { + 'gfx90a': ('SQ', 64), + 'gfx908': ('SQ', 39), + 'gfx940': ('SQ', 69), + 'gfx941': ('SQ', 69), + 'gfx942': ('SQ', 69), + 'gfx950': ('SQ', 71), + }, + 'SQ_INSTS_EXP_GDS': { + 'gfx90a': ('SQ', 63), + 'gfx908': ('SQ', 38), + 'gfx940': ('SQ', 68), + 'gfx941': ('SQ', 68), + 'gfx942': ('SQ', 68), + 'gfx950': ('SQ', 70), + }, + 'SQ_INSTS_FLAT': { + 'gfx10': ('SQ', 57), + 'gfx1010': ('SQ', 57), + 'gfx1030': ('SQ', 57), + 'gfx1031': ('SQ', 57), + 'gfx1032': ('SQ', 57), + 'gfx9': ('SQ', 32), + 'gfx900': ('SQ', 32), + 'gfx906': ('SQ', 32), + 'gfx908': ('SQ', 33), + 'gfx90a': ('SQ', 58), + 'gfx940': ('SQ', 62), + 'gfx941': ('SQ', 62), + 'gfx942': ('SQ', 62), + 'gfx11': ('SQ', 56), + 'gfx1100': ('SQ', 56), + 'gfx1101': ('SQ', 56), + 'gfx1102': ('SQ', 56), + 'gfx12': ('SQ', 44), + 'gfx1200': ('SQ', 44), + 'gfx1201': ('SQ', 44), + 'gfx950': ('SQ', 64), + }, + 'SQ_INSTS_FLAT_LDS_ONLY': { + 'gfx9': ('SQ', 33), + 'gfx900': ('SQ', 33), + 'gfx906': ('SQ', 33), + 'gfx908': ('SQ', 34), + 'gfx90a': ('SQ', 59), + }, + 'SQ_INSTS_GDS': { + 'gfx10': ('SQ', 55), + 'gfx1010': ('SQ', 55), + 'gfx1030': ('SQ', 55), + 'gfx1031': ('SQ', 55), + 'gfx1032': ('SQ', 55), + 'gfx9': ('SQ', 35), + 'gfx900': ('SQ', 35), + 'gfx906': ('SQ', 35), + 'gfx908': ('SQ', 36), + 'gfx90a': ('SQ', 61), + 'gfx940': ('SQ', 66), + 'gfx941': ('SQ', 66), + 'gfx942': ('SQ', 66), + 'gfx11': ('SQ', 54), + 'gfx1100': ('SQ', 54), + 'gfx1101': ('SQ', 54), + 'gfx1102': ('SQ', 54), + 'gfx950': ('SQ', 68), + }, + 'SQ_INSTS_LDS': { + 'gfx10': ('SQ', 59), + 'gfx1010': ('SQ', 59), + 'gfx1030': ('SQ', 59), + 'gfx1031': ('SQ', 59), + 'gfx1032': ('SQ', 59), + 'gfx9': ('SQ', 34), + 'gfx900': ('SQ', 34), + 'gfx906': ('SQ', 34), + 'gfx908': ('SQ', 35), + 'gfx90a': ('SQ', 60), + 'gfx940': ('SQ', 65), + 'gfx941': ('SQ', 65), + 'gfx942': ('SQ', 65), + 'gfx11': ('SQ', 57), + 'gfx1100': ('SQ', 57), + 'gfx1101': ('SQ', 57), + 'gfx1102': ('SQ', 57), + 'gfx12': ('SQ', 45), + 'gfx1200': ('SQ', 45), + 'gfx1201': ('SQ', 45), + 'gfx950': ('SQ', 67), + }, + 'SQ_INSTS_MFMA': { + 'gfx90a': ('SQ', 52), + 'gfx908': ('SQ', 27), + 'gfx940': ('SQ', 56), + 'gfx941': ('SQ', 56), + 'gfx942': ('SQ', 56), + 'gfx950': ('SQ', 58), + }, + 'SQ_INSTS_SALU': { + 'gfx9': ('SQ', 30), + 'gfx900': ('SQ', 30), + 'gfx906': ('SQ', 30), + 'gfx908': ('SQ', 31), + 'gfx90a': ('SQ', 56), + 'gfx10': ('SQ', 60), + 'gfx1010': ('SQ', 60), + 'gfx1030': ('SQ', 60), + 'gfx1031': ('SQ', 60), + 'gfx1032': ('SQ', 60), + 'gfx940': ('SQ', 60), + 'gfx941': ('SQ', 60), + 'gfx942': ('SQ', 60), + 'gfx11': ('SQ', 58), + 'gfx1100': ('SQ', 58), + 'gfx1101': ('SQ', 58), + 'gfx1102': ('SQ', 58), + 'gfx12': ('SQ', 46), + 'gfx1200': ('SQ', 46), + 'gfx1201': ('SQ', 46), + 'gfx950': ('SQ', 62), + }, + 'SQ_INSTS_SENDMSG': { + 'gfx90a': ('SQ', 65), + 'gfx908': ('SQ', 40), + 'gfx940': ('SQ', 70), + 'gfx941': ('SQ', 70), + 'gfx942': ('SQ', 70), + 'gfx950': ('SQ', 72), + }, + 'SQ_INSTS_SMEM': { + 'gfx9': ('SQ', 31), + 'gfx900': ('SQ', 31), + 'gfx906': ('SQ', 31), + 'gfx908': ('SQ', 32), + 'gfx90a': ('SQ', 57), + 'gfx10': ('SQ', 61), + 'gfx1010': ('SQ', 61), + 'gfx1030': ('SQ', 61), + 'gfx1031': ('SQ', 61), + 'gfx1032': ('SQ', 61), + 'gfx940': ('SQ', 61), + 'gfx941': ('SQ', 61), + 'gfx942': ('SQ', 61), + 'gfx11': ('SQ', 59), + 'gfx1100': ('SQ', 59), + 'gfx1101': ('SQ', 59), + 'gfx1102': ('SQ', 59), + 'gfx12': ('SQ', 47), + 'gfx1200': ('SQ', 47), + 'gfx1201': ('SQ', 47), + 'gfx950': ('SQ', 63), + }, + 'SQ_INSTS_SMEM_NORM': { + 'gfx90a': ('SQ', 188), + 'gfx908': ('SQ', 161), + 'gfx940': ('SQ', 187), + 'gfx941': ('SQ', 187), + 'gfx942': ('SQ', 187), + 'gfx950': ('SQ', 203), + }, + 'SQ_INSTS_TEX_LOAD': { + 'gfx11': ('SQ', 66), + 'gfx1100': ('SQ', 66), + 'gfx1101': ('SQ', 66), + 'gfx1102': ('SQ', 66), + 'gfx12': ('SQ', 54), + 'gfx1200': ('SQ', 54), + 'gfx1201': ('SQ', 54), + }, + 'SQ_INSTS_TEX_STORE': { + 'gfx11': ('SQ', 67), + 'gfx1100': ('SQ', 67), + 'gfx1101': ('SQ', 67), + 'gfx1102': ('SQ', 67), + 'gfx12': ('SQ', 55), + 'gfx1200': ('SQ', 55), + 'gfx1201': ('SQ', 55), + }, + 'SQ_INSTS_VALU': { + 'gfx10': ('SQ', 64), + 'gfx1010': ('SQ', 64), + 'gfx1030': ('SQ', 64), + 'gfx1031': ('SQ', 64), + 'gfx1032': ('SQ', 64), + 'gfx9': ('SQ', 26), + 'gfx900': ('SQ', 26), + 'gfx906': ('SQ', 26), + 'gfx908': ('SQ', 26), + 'gfx90a': ('SQ', 26), + 'gfx940': ('SQ', 26), + 'gfx941': ('SQ', 26), + 'gfx942': ('SQ', 26), + 'gfx950': ('SQ', 26), + 'gfx11': ('SQ', 62), + 'gfx1100': ('SQ', 62), + 'gfx1101': ('SQ', 62), + 'gfx1102': ('SQ', 62), + 'gfx12': ('SQ', 50), + 'gfx1200': ('SQ', 50), + 'gfx1201': ('SQ', 50), + }, + 'SQ_INSTS_VALU_ADD_F16': { + 'gfx90a': ('SQ', 27), + 'gfx940': ('SQ', 27), + 'gfx941': ('SQ', 27), + 'gfx942': ('SQ', 27), + 'gfx950': ('SQ', 27), + }, + 'SQ_INSTS_VALU_ADD_F32': { + 'gfx90a': ('SQ', 31), + 'gfx940': ('SQ', 31), + 'gfx941': ('SQ', 31), + 'gfx942': ('SQ', 31), + 'gfx950': ('SQ', 31), + }, + 'SQ_INSTS_VALU_ADD_F64': { + 'gfx90a': ('SQ', 35), + 'gfx940': ('SQ', 35), + 'gfx941': ('SQ', 35), + 'gfx942': ('SQ', 35), + 'gfx950': ('SQ', 35), + }, + 'SQ_INSTS_VALU_CVT': { + 'gfx90a': ('SQ', 41), + 'gfx940': ('SQ', 41), + 'gfx941': ('SQ', 41), + 'gfx942': ('SQ', 41), + 'gfx950': ('SQ', 41), + }, + 'SQ_INSTS_VALU_FMA_F16': { + 'gfx90a': ('SQ', 29), + 'gfx940': ('SQ', 29), + 'gfx941': ('SQ', 29), + 'gfx942': ('SQ', 29), + 'gfx950': ('SQ', 29), + }, + 'SQ_INSTS_VALU_FMA_F32': { + 'gfx90a': ('SQ', 33), + 'gfx940': ('SQ', 33), + 'gfx941': ('SQ', 33), + 'gfx942': ('SQ', 33), + 'gfx950': ('SQ', 33), + }, + 'SQ_INSTS_VALU_FMA_F64': { + 'gfx90a': ('SQ', 37), + 'gfx940': ('SQ', 37), + 'gfx941': ('SQ', 37), + 'gfx942': ('SQ', 37), + 'gfx950': ('SQ', 37), + }, + 'SQ_INSTS_VALU_INT32': { + 'gfx90a': ('SQ', 39), + 'gfx940': ('SQ', 39), + 'gfx941': ('SQ', 39), + 'gfx942': ('SQ', 39), + 'gfx950': ('SQ', 39), + }, + 'SQ_INSTS_VALU_INT64': { + 'gfx90a': ('SQ', 40), + 'gfx940': ('SQ', 40), + 'gfx941': ('SQ', 40), + 'gfx942': ('SQ', 40), + 'gfx950': ('SQ', 40), + }, + 'SQ_INSTS_VALU_MFMA_BF16': { + 'gfx90a': ('SQ', 44), + 'gfx940': ('SQ', 44), + 'gfx941': ('SQ', 44), + 'gfx942': ('SQ', 44), + 'gfx950': ('SQ', 44), + }, + 'SQ_INSTS_VALU_MFMA_F16': { + 'gfx90a': ('SQ', 43), + 'gfx940': ('SQ', 43), + 'gfx941': ('SQ', 43), + 'gfx942': ('SQ', 43), + 'gfx950': ('SQ', 43), + }, + 'SQ_INSTS_VALU_MFMA_F32': { + 'gfx90a': ('SQ', 45), + 'gfx940': ('SQ', 45), + 'gfx941': ('SQ', 45), + 'gfx942': ('SQ', 45), + 'gfx950': ('SQ', 45), + }, + 'SQ_INSTS_VALU_MFMA_F64': { + 'gfx90a': ('SQ', 46), + 'gfx940': ('SQ', 46), + 'gfx941': ('SQ', 46), + 'gfx942': ('SQ', 46), + 'gfx950': ('SQ', 46), + }, + 'SQ_INSTS_VALU_MFMA_I8': { + 'gfx90a': ('SQ', 42), + 'gfx940': ('SQ', 42), + 'gfx941': ('SQ', 42), + 'gfx942': ('SQ', 42), + 'gfx950': ('SQ', 42), + }, + 'SQ_INSTS_VALU_MFMA_F8': { + 'gfx940': ('SQ', 48), + 'gfx941': ('SQ', 48), + 'gfx942': ('SQ', 48), + 'gfx950': ('SQ', 48), + }, + 'SQ_INSTS_VALU_MFMA_XF32': { + 'gfx950': ('SQ', 47), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_BF16': { + 'gfx90a': ('SQ', 49), + 'gfx940': ('SQ', 51), + 'gfx941': ('SQ', 51), + 'gfx942': ('SQ', 51), + 'gfx950': ('SQ', 52), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_F16': { + 'gfx90a': ('SQ', 48), + 'gfx940': ('SQ', 50), + 'gfx941': ('SQ', 50), + 'gfx942': ('SQ', 50), + 'gfx950': ('SQ', 51), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_F32': { + 'gfx90a': ('SQ', 50), + 'gfx940': ('SQ', 52), + 'gfx941': ('SQ', 52), + 'gfx942': ('SQ', 52), + 'gfx950': ('SQ', 53), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_F64': { + 'gfx90a': ('SQ', 51), + 'gfx940': ('SQ', 53), + 'gfx941': ('SQ', 53), + 'gfx942': ('SQ', 53), + 'gfx950': ('SQ', 54), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_I8': { + 'gfx90a': ('SQ', 47), + 'gfx940': ('SQ', 49), + 'gfx941': ('SQ', 49), + 'gfx942': ('SQ', 49), + 'gfx950': ('SQ', 50), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_F8': { + 'gfx940': ('SQ', 55), + 'gfx941': ('SQ', 55), + 'gfx942': ('SQ', 55), + 'gfx950': ('SQ', 56), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_XF32': { + 'gfx950': ('SQ', 55), + }, + 'SQ_VALU_MFMA_COEXEC_CYCLES': { + 'gfx950': ('SQ', 94), + }, + 'SQ_INSTS_VALU_MUL_F16': { + 'gfx90a': ('SQ', 28), + 'gfx940': ('SQ', 28), + 'gfx941': ('SQ', 28), + 'gfx942': ('SQ', 28), + 'gfx950': ('SQ', 28), + }, + 'SQ_INSTS_VALU_MUL_F32': { + 'gfx90a': ('SQ', 32), + 'gfx940': ('SQ', 32), + 'gfx941': ('SQ', 32), + 'gfx942': ('SQ', 32), + 'gfx950': ('SQ', 32), + }, + 'SQ_INSTS_VALU_MUL_F64': { + 'gfx90a': ('SQ', 36), + 'gfx940': ('SQ', 36), + 'gfx941': ('SQ', 36), + 'gfx942': ('SQ', 36), + 'gfx950': ('SQ', 36), + }, + 'SQ_INSTS_VALU_TRANS_F16': { + 'gfx90a': ('SQ', 30), + 'gfx940': ('SQ', 30), + 'gfx941': ('SQ', 30), + 'gfx942': ('SQ', 30), + 'gfx950': ('SQ', 30), + }, + 'SQ_INSTS_VALU_TRANS_F32': { + 'gfx90a': ('SQ', 34), + 'gfx940': ('SQ', 34), + 'gfx941': ('SQ', 34), + 'gfx942': ('SQ', 34), + 'gfx950': ('SQ', 34), + }, + 'SQ_INSTS_VALU_TRANS_F64': { + 'gfx90a': ('SQ', 38), + 'gfx940': ('SQ', 38), + 'gfx941': ('SQ', 38), + 'gfx942': ('SQ', 38), + 'gfx950': ('SQ', 38), + }, + 'SQ_INSTS_VMEM': { + 'gfx90a': ('SQ', 55), + 'gfx908': ('SQ', 30), + 'gfx940': ('SQ', 59), + 'gfx941': ('SQ', 59), + 'gfx942': ('SQ', 59), + 'gfx950': ('SQ', 61), + }, + 'SQ_INSTS_VMEM_RD': { + 'gfx9': ('SQ', 28), + 'gfx900': ('SQ', 28), + 'gfx906': ('SQ', 28), + 'gfx908': ('SQ', 29), + 'gfx90a': ('SQ', 54), + 'gfx940': ('SQ', 58), + 'gfx941': ('SQ', 58), + 'gfx942': ('SQ', 58), + 'gfx950': ('SQ', 60), + }, + 'SQ_INSTS_VMEM_WR': { + 'gfx9': ('SQ', 27), + 'gfx900': ('SQ', 27), + 'gfx906': ('SQ', 27), + 'gfx908': ('SQ', 28), + 'gfx90a': ('SQ', 53), + 'gfx940': ('SQ', 57), + 'gfx941': ('SQ', 57), + 'gfx942': ('SQ', 57), + 'gfx950': ('SQ', 59), + }, + 'SQ_INSTS_VSKIPPED': { + 'gfx90a': ('SQ', 66), + 'gfx908': ('SQ', 41), + 'gfx940': ('SQ', 71), + 'gfx941': ('SQ', 71), + 'gfx942': ('SQ', 71), + 'gfx950': ('SQ', 73), + }, + 'SQ_INSTS_WAVE32': { + 'gfx10': ('SQ', 71), + 'gfx1010': ('SQ', 71), + 'gfx1030': ('SQ', 71), + 'gfx1031': ('SQ', 71), + 'gfx1032': ('SQ', 71), + 'gfx11': ('SQ', 70), + 'gfx1100': ('SQ', 70), + 'gfx1101': ('SQ', 70), + 'gfx1102': ('SQ', 70), + 'gfx12': ('SQ', 58), + 'gfx1200': ('SQ', 58), + 'gfx1201': ('SQ', 58), + }, + 'SQ_INSTS_WAVE32_LDS': { + 'gfx10': ('SQ', 74), + 'gfx1010': ('SQ', 74), + 'gfx1030': ('SQ', 74), + 'gfx1031': ('SQ', 74), + 'gfx1032': ('SQ', 74), + 'gfx11': ('SQ', 72), + 'gfx1100': ('SQ', 72), + 'gfx1101': ('SQ', 72), + 'gfx1102': ('SQ', 72), + 'gfx12': ('SQ', 60), + 'gfx1200': ('SQ', 60), + 'gfx1201': ('SQ', 60), + }, + 'SQ_INSTS_WAVE32_VALU': { + 'gfx10': ('SQ', 75), + 'gfx1010': ('SQ', 75), + 'gfx1030': ('SQ', 75), + 'gfx1031': ('SQ', 75), + 'gfx1032': ('SQ', 75), + 'gfx11': ('SQ', 73), + 'gfx1100': ('SQ', 73), + 'gfx1101': ('SQ', 73), + 'gfx1102': ('SQ', 73), + 'gfx12': ('SQ', 61), + 'gfx1200': ('SQ', 61), + 'gfx1201': ('SQ', 61), + }, + 'SQ_INST_CYCLES_VALU': { + 'gfx12': ('SQ', 99), + 'gfx1200': ('SQ', 99), + 'gfx1201': ('SQ', 99), + }, + 'SQ_INST_CYCLES_SALU': { + 'gfx9': ('SQ', 84), + 'gfx900': ('SQ', 84), + 'gfx906': ('SQ', 84), + 'gfx908': ('SQ', 85), + 'gfx90a': ('SQ', 112), + 'gfx940': ('SQ', 117), + 'gfx941': ('SQ', 117), + 'gfx942': ('SQ', 117), + 'gfx950': ('SQ', 133), + }, + 'SQ_INST_CYCLES_SMEM': { + 'gfx90a': ('SQ', 111), + 'gfx908': ('SQ', 84), + 'gfx940': ('SQ', 116), + 'gfx941': ('SQ', 116), + 'gfx942': ('SQ', 116), + 'gfx950': ('SQ', 132), + }, + 'SQ_INST_CYCLES_VMEM': { + 'gfx10': ('SQ', 120), + 'gfx1010': ('SQ', 120), + 'gfx1030': ('SQ', 120), + 'gfx1031': ('SQ', 120), + 'gfx1032': ('SQ', 120), + 'gfx11': ('SQ', 106), + 'gfx1100': ('SQ', 106), + 'gfx1101': ('SQ', 106), + 'gfx1102': ('SQ', 106), + 'gfx12': ('SQ', 102), + 'gfx1200': ('SQ', 102), + 'gfx1201': ('SQ', 102), + }, + 'SQ_INST_CYCLES_VMEM_RD': { + 'gfx90a': ('SQ', 105), + 'gfx908': ('SQ', 78), + 'gfx940': ('SQ', 110), + 'gfx941': ('SQ', 110), + 'gfx942': ('SQ', 110), + 'gfx950': ('SQ', 126), + }, + 'SQ_INST_CYCLES_VMEM_WR': { + 'gfx90a': ('SQ', 104), + 'gfx908': ('SQ', 77), + 'gfx940': ('SQ', 109), + 'gfx941': ('SQ', 109), + 'gfx942': ('SQ', 109), + 'gfx950': ('SQ', 125), + }, + 'SQ_INST_LEVEL_GDS': { + 'gfx10': ('SQ', 98), + 'gfx1010': ('SQ', 98), + 'gfx1030': ('SQ', 98), + 'gfx1031': ('SQ', 98), + 'gfx1032': ('SQ', 98), + 'gfx11': ('SQ', 87), + 'gfx1100': ('SQ', 87), + 'gfx1101': ('SQ', 87), + 'gfx1102': ('SQ', 87), + }, + 'SQ_INST_LEVEL_LDS': { + 'gfx10': ('SQ', 99), + 'gfx1010': ('SQ', 99), + 'gfx1030': ('SQ', 99), + 'gfx1031': ('SQ', 99), + 'gfx1032': ('SQ', 99), + 'gfx90a': ('SQ', 69), + 'gfx908': ('SQ', 44), + 'gfx940': ('SQ', 74), + 'gfx941': ('SQ', 74), + 'gfx942': ('SQ', 74), + 'gfx11': ('SQ', 88), + 'gfx1100': ('SQ', 88), + 'gfx1101': ('SQ', 88), + 'gfx1102': ('SQ', 88), + 'gfx12': ('SQ', 75), + 'gfx1200': ('SQ', 75), + 'gfx1201': ('SQ', 75), + 'gfx950': ('SQ', 90), + }, + 'SQ_INST_LEVEL_SMEM': { + 'gfx90a': ('SQ', 68), + 'gfx908': ('SQ', 43), + 'gfx940': ('SQ', 73), + 'gfx941': ('SQ', 73), + 'gfx942': ('SQ', 73), + 'gfx950': ('SQ', 89), + }, + 'SQ_INST_LEVEL_VMEM': { + 'gfx90a': ('SQ', 67), + 'gfx908': ('SQ', 42), + 'gfx940': ('SQ', 72), + 'gfx941': ('SQ', 72), + 'gfx942': ('SQ', 72), + 'gfx950': ('SQ', 88), + }, + 'SQ_ITEMS': { + 'gfx90a': ('SQ', 14), + 'gfx908': ('SQ', 14), + 'gfx940': ('SQ', 14), + 'gfx941': ('SQ', 14), + 'gfx942': ('SQ', 14), + 'gfx950': ('SQ', 14), + }, + 'SQ_LDS_ADDR_CONFLICT': { + 'gfx90a': ('SQ', 122), + 'gfx908': ('SQ', 95), + 'gfx940': ('SQ', 127), + 'gfx941': ('SQ', 127), + 'gfx942': ('SQ', 127), + 'gfx950': ('SQ', 143), + }, + 'SQ_LDS_ATOMIC_RETURN': { + 'gfx90a': ('SQ', 125), + 'gfx908': ('SQ', 98), + 'gfx940': ('SQ', 130), + 'gfx941': ('SQ', 130), + 'gfx942': ('SQ', 130), + 'gfx950': ('SQ', 146), + }, + 'SQ_LDS_BANK_CONFLICT': { + 'gfx9': ('SQ', 93), + 'gfx900': ('SQ', 93), + 'gfx906': ('SQ', 93), + 'gfx908': ('SQ', 94), + 'gfx90a': ('SQ', 121), + 'gfx940': ('SQ', 126), + 'gfx941': ('SQ', 126), + 'gfx942': ('SQ', 126), + 'gfx950': ('SQ', 142), + }, + 'SQ_LDS_IDX_ACTIVE': { + 'gfx90a': ('SQ', 126), + 'gfx908': ('SQ', 99), + 'gfx940': ('SQ', 131), + 'gfx941': ('SQ', 131), + 'gfx942': ('SQ', 131), + 'gfx950': ('SQ', 147), + }, + 'SQ_LDS_MEM_VIOLATIONS': { + 'gfx90a': ('SQ', 124), + 'gfx908': ('SQ', 97), + 'gfx940': ('SQ', 129), + 'gfx941': ('SQ', 129), + 'gfx942': ('SQ', 129), + 'gfx950': ('SQ', 145), + }, + 'SQ_LDS_UNALIGNED_STALL': { + 'gfx90a': ('SQ', 123), + 'gfx908': ('SQ', 96), + 'gfx940': ('SQ', 128), + 'gfx941': ('SQ', 128), + 'gfx942': ('SQ', 128), + 'gfx950': ('SQ', 144), + }, + 'SQ_LEVEL_WAVES': { + 'gfx10': ('SQ', 7), + 'gfx1010': ('SQ', 7), + 'gfx1030': ('SQ', 7), + 'gfx1031': ('SQ', 7), + 'gfx1032': ('SQ', 7), + 'gfx90a': ('SQ', 5), + 'gfx908': ('SQ', 5), + 'gfx940': ('SQ', 5), + 'gfx941': ('SQ', 5), + 'gfx942': ('SQ', 5), + 'gfx950': ('SQ', 5), + }, + 'SQ_THREAD_CYCLES_VALU': { + 'gfx9': ('SQ', 85), + 'gfx900': ('SQ', 85), + 'gfx906': ('SQ', 85), + 'gfx908': ('SQ', 86), + 'gfx90a': ('SQ', 113), + 'gfx940': ('SQ', 118), + 'gfx941': ('SQ', 118), + 'gfx942': ('SQ', 118), + 'gfx950': ('SQ', 134), + }, + 'SQ_VALU_MFMA_BUSY_CYCLES': { + 'gfx90a': ('SQ', 72), + 'gfx940': ('SQ', 77), + 'gfx941': ('SQ', 77), + 'gfx942': ('SQ', 77), + 'gfx950': ('SQ', 93), + }, + 'SQ_WAIT_ANY': { + 'gfx10': ('SQ', 37), + 'gfx1010': ('SQ', 37), + 'gfx1030': ('SQ', 37), + 'gfx1031': ('SQ', 37), + 'gfx1032': ('SQ', 37), + 'gfx90a': ('SQ', 85), + 'gfx908': ('SQ', 58), + 'gfx940': ('SQ', 90), + 'gfx941': ('SQ', 90), + 'gfx942': ('SQ', 90), + 'gfx11': ('SQ', 35), + 'gfx1100': ('SQ', 35), + 'gfx1101': ('SQ', 35), + 'gfx1102': ('SQ', 35), + 'gfx12': ('SQ', 27), + 'gfx1200': ('SQ', 27), + 'gfx1201': ('SQ', 27), + 'gfx950': ('SQ', 106), + }, + 'SQ_WAIT_INST_ANY': { + 'gfx10': ('SQ', 28), + 'gfx1010': ('SQ', 28), + 'gfx1030': ('SQ', 28), + 'gfx1031': ('SQ', 28), + 'gfx1032': ('SQ', 28), + 'gfx90a': ('SQ', 88), + 'gfx908': ('SQ', 61), + 'gfx940': ('SQ', 93), + 'gfx941': ('SQ', 93), + 'gfx942': ('SQ', 93), + 'gfx11': ('SQ', 26), + 'gfx1100': ('SQ', 26), + 'gfx1101': ('SQ', 26), + 'gfx1102': ('SQ', 26), + 'gfx12': ('SQ', 26), + 'gfx1200': ('SQ', 26), + 'gfx1201': ('SQ', 26), + 'gfx950': ('SQ', 109), + }, + 'SQ_WAIT_INST_LDS': { + 'gfx10': ('SQ', 31), + 'gfx1010': ('SQ', 31), + 'gfx1030': ('SQ', 31), + 'gfx1031': ('SQ', 31), + 'gfx1032': ('SQ', 31), + 'gfx9': ('SQ', 63), + 'gfx900': ('SQ', 63), + 'gfx906': ('SQ', 63), + 'gfx908': ('SQ', 64), + 'gfx90a': ('SQ', 91), + 'gfx940': ('SQ', 96), + 'gfx941': ('SQ', 96), + 'gfx942': ('SQ', 96), + 'gfx11': ('SQ', 29), + 'gfx1100': ('SQ', 29), + 'gfx1101': ('SQ', 29), + 'gfx1102': ('SQ', 29), + 'gfx950': ('SQ', 112), + }, + 'SQ_WAVE32_INSTS': { + 'gfx10': ('SQ', 84), + 'gfx1010': ('SQ', 84), + 'gfx1030': ('SQ', 84), + 'gfx1031': ('SQ', 84), + 'gfx1032': ('SQ', 84), + 'gfx11': ('SQ', 82), + 'gfx1100': ('SQ', 82), + 'gfx1101': ('SQ', 82), + 'gfx1102': ('SQ', 82), + 'gfx12': ('SQ', 70), + 'gfx1200': ('SQ', 70), + 'gfx1201': ('SQ', 70), + }, + 'SQ_WAVE64_INSTS': { + 'gfx10': ('SQ', 85), + 'gfx1010': ('SQ', 85), + 'gfx1030': ('SQ', 85), + 'gfx1031': ('SQ', 85), + 'gfx1032': ('SQ', 85), + 'gfx11': ('SQ', 83), + 'gfx1100': ('SQ', 83), + 'gfx1101': ('SQ', 83), + 'gfx1102': ('SQ', 83), + 'gfx12': ('SQ', 71), + 'gfx1200': ('SQ', 71), + 'gfx1201': ('SQ', 71), + }, + 'SQ_WAVES': { + 'gfx10': ('SQ', 4), + 'gfx1010': ('SQ', 4), + 'gfx1030': ('SQ', 4), + 'gfx1031': ('SQ', 4), + 'gfx1032': ('SQ', 4), + 'gfx11': ('SQ', 4), + 'gfx1100': ('SQ', 4), + 'gfx1101': ('SQ', 4), + 'gfx1102': ('SQ', 4), + 'gfx12': ('SQ', 4), + 'gfx1200': ('SQ', 4), + 'gfx1201': ('SQ', 4), + 'gfx9': ('SQ', 4), + 'gfx900': ('SQ', 4), + 'gfx906': ('SQ', 4), + 'gfx908': ('SQ', 4), + 'gfx90a': ('SQ', 4), + 'gfx940': ('SQ', 4), + 'gfx941': ('SQ', 4), + 'gfx942': ('SQ', 4), + 'gfx950': ('SQ', 4), + }, + 'SQ_WAVES_EQ_64': { + 'gfx90a': ('SQ', 6), + 'gfx908': ('SQ', 6), + 'gfx940': ('SQ', 6), + 'gfx941': ('SQ', 6), + 'gfx942': ('SQ', 6), + 'gfx950': ('SQ', 6), + }, + 'SQ_WAVES_LT_16': { + 'gfx90a': ('SQ', 10), + 'gfx908': ('SQ', 10), + 'gfx940': ('SQ', 10), + 'gfx941': ('SQ', 10), + 'gfx942': ('SQ', 10), + 'gfx950': ('SQ', 10), + }, + 'SQ_WAVES_LT_32': { + 'gfx90a': ('SQ', 9), + 'gfx908': ('SQ', 9), + 'gfx940': ('SQ', 9), + 'gfx941': ('SQ', 9), + 'gfx942': ('SQ', 9), + 'gfx950': ('SQ', 9), + }, + 'SQ_WAVES_LT_48': { + 'gfx90a': ('SQ', 8), + 'gfx908': ('SQ', 8), + 'gfx940': ('SQ', 8), + 'gfx941': ('SQ', 8), + 'gfx942': ('SQ', 8), + 'gfx950': ('SQ', 8), + }, + 'SQ_WAVES_LT_64': { + 'gfx90a': ('SQ', 7), + 'gfx908': ('SQ', 7), + 'gfx940': ('SQ', 7), + 'gfx941': ('SQ', 7), + 'gfx942': ('SQ', 7), + 'gfx950': ('SQ', 7), + }, + 'SQ_WAVES_RESTORED': { + 'gfx90a': ('SQ', 186), + 'gfx908': ('SQ', 159), + 'gfx940': ('SQ', 185), + 'gfx941': ('SQ', 185), + 'gfx942': ('SQ', 185), + 'gfx950': ('SQ', 201), + }, + 'SQ_WAVES_SAVED': { + 'gfx90a': ('SQ', 187), + 'gfx908': ('SQ', 160), + 'gfx940': ('SQ', 186), + 'gfx941': ('SQ', 186), + 'gfx942': ('SQ', 186), + 'gfx950': ('SQ', 202), + }, + 'SQ_WAVE_CYCLES': { + 'gfx10': ('SQ', 26), + 'gfx1010': ('SQ', 26), + 'gfx1030': ('SQ', 26), + 'gfx1031': ('SQ', 26), + 'gfx1032': ('SQ', 26), + 'gfx90a': ('SQ', 74), + 'gfx908': ('SQ', 47), + 'gfx940': ('SQ', 79), + 'gfx941': ('SQ', 79), + 'gfx942': ('SQ', 79), + 'gfx11': ('SQ', 24), + 'gfx1100': ('SQ', 24), + 'gfx1101': ('SQ', 24), + 'gfx1102': ('SQ', 24), + 'gfx12': ('SQ', 24), + 'gfx1200': ('SQ', 24), + 'gfx1201': ('SQ', 24), + 'gfx950': ('SQ', 95), + }, + 'SQ_INSTS_VEC32_LEVEL_LDS': { + 'gfx12': ('SQ', 250), + 'gfx1200': ('SQ', 250), + 'gfx1201': ('SQ', 250), + }, + 'SQ_INSTS_VALU_FLOPS_FP16': { + 'gfx950': ('SQ', 81), + }, + 'SQ_INSTS_VALU_FLOPS_FP32': { + 'gfx950': ('SQ', 82), + }, + 'SQ_INSTS_VALU_FLOPS_FP64': { + 'gfx950': ('SQ', 83), + }, + 'SQ_INSTS_VALU_FLOPS_FP16_TRANS': { + 'gfx950': ('SQ', 84), + }, + 'SQ_INSTS_VALU_FLOPS_FP32_TRANS': { + 'gfx950': ('SQ', 85), + }, + 'SQ_INSTS_VALU_FLOPS_FP64_TRANS': { + 'gfx950': ('SQ', 86), + }, + 'SQ_INSTS_VALU_MFMA_F6F4': { + 'gfx950': ('SQ', 49), + }, + 'SQ_INSTS_VALU_MFMA_MOPS_F6F4': { + 'gfx950': ('SQ', 57), + }, + 'SQ_ACTIVE_INST_VALU2': { + 'gfx950': ('SQ', 74), + }, + 'SQ_INSTS_LDS_LOAD': { + 'gfx950': ('SQ', 75), + }, + 'SQ_INSTS_LDS_STORE': { + 'gfx950': ('SQ', 76), + }, + 'SQ_INSTS_LDS_ATOMIC': { + 'gfx950': ('SQ', 77), + }, + 'SQ_INSTS_LDS_LOAD_BANDWIDTH': { + 'gfx950': ('SQ', 78), + }, + 'SQ_INSTS_LDS_STORE_BANDWIDTH': { + 'gfx950': ('SQ', 79), + }, + 'SQ_INSTS_LDS_ATOMIC_BANDWIDTH': { + 'gfx950': ('SQ', 80), + }, + 'SQ_INSTS_VALU_IOPS': { + 'gfx950': ('SQ', 87), + }, + 'SQ_LDS_DATA_FIFO_FULL': { + 'gfx950': ('SQ', 152), + }, + 'SQ_LDS_CMD_FIFO_FULL': { + 'gfx950': ('SQ', 153), + }, + 'SQ_VMEM_TA_ADDR_FIFO_FULL': { + 'gfx90a': ('SQ', 133), + 'gfx942': ('SQ', 138), + 'gfx950': ('SQ', 154), + }, + 'SQ_VMEM_TA_CMD_FIFO_FULL': { + 'gfx90a': ('SQ', 134), + 'gfx942': ('SQ', 139), + 'gfx950': ('SQ', 155), + }, + 'SQ_VMEM_WR_TA_DATA_FIFO_FULL': { + 'gfx90a': ('SQ', 136), + 'gfx942': ('SQ', 141), + 'gfx950': ('SQ', 157), + }, + 'SQ_INSTS_FLAT_FLATSEG': { + 'gfx950': ('SQ', 65), + }, + 'SQ_INSTS_FLAT_NO_LDS': { + 'gfx950': ('SQ', 66), + }, + 'SQ_INSTS_EXP': { + 'gfx950': ('SQ', 69), + }, + 'SQ_EVENTS': { + 'gfx950': ('SQ', 16), + }, + 'TA_ADDR_STALLED_BY_TC_CYCLES': { + 'gfx90a': ('TA', 54), + 'gfx908': ('TA', 54), + 'gfx940': ('TA', 42), + 'gfx941': ('TA', 42), + 'gfx942': ('TA', 42), + 'gfx950': ('TA', 42), + }, + 'TA_ADDR_STALLED_BY_TD_CYCLES': { + 'gfx90a': ('TA', 55), + 'gfx908': ('TA', 55), + 'gfx940': ('TA', 43), + 'gfx941': ('TA', 43), + 'gfx942': ('TA', 43), + 'gfx950': ('TA', 43), + }, + 'TA_BUFFER_ATOMIC_WAVEFRONTS': { + 'gfx90a': ('TA', 47), + 'gfx908': ('TA', 47), + 'gfx940': ('TA', 35), + 'gfx941': ('TA', 35), + 'gfx942': ('TA', 35), + 'gfx950': ('TA', 35), + }, + 'TA_BUFFER_COALESCED_READ_CYCLES': { + 'gfx90a': ('TA', 52), + 'gfx908': ('TA', 52), + 'gfx940': ('TA', 40), + 'gfx941': ('TA', 40), + 'gfx942': ('TA', 40), + 'gfx950': ('TA', 40), + }, + 'TA_BUFFER_COALESCED_WRITE_CYCLES': { + 'gfx90a': ('TA', 53), + 'gfx908': ('TA', 53), + 'gfx940': ('TA', 41), + 'gfx941': ('TA', 41), + 'gfx942': ('TA', 41), + 'gfx950': ('TA', 41), + }, + 'TA_BUFFER_LOAD_WAVEFRONTS': { + 'gfx11': ('TA', 45), + 'gfx1100': ('TA', 45), + 'gfx1101': ('TA', 45), + 'gfx1102': ('TA', 45), + 'gfx12': ('TA', 45), + 'gfx1200': ('TA', 45), + 'gfx1201': ('TA', 45), + }, + 'TA_BUFFER_READ_WAVEFRONTS': { + 'gfx90a': ('TA', 45), + 'gfx908': ('TA', 45), + 'gfx940': ('TA', 33), + 'gfx941': ('TA', 33), + 'gfx942': ('TA', 33), + 'gfx950': ('TA', 33), + }, + 'TA_BUFFER_STORE_WAVEFRONTS': { + 'gfx11': ('TA', 46), + 'gfx1100': ('TA', 46), + 'gfx1101': ('TA', 46), + 'gfx1102': ('TA', 46), + 'gfx12': ('TA', 46), + 'gfx1200': ('TA', 46), + 'gfx1201': ('TA', 46), + }, + 'TA_BUFFER_TOTAL_CYCLES': { + 'gfx90a': ('TA', 49), + 'gfx908': ('TA', 49), + 'gfx940': ('TA', 37), + 'gfx941': ('TA', 37), + 'gfx942': ('TA', 37), + 'gfx950': ('TA', 37), + }, + 'TA_BUFFER_WAVEFRONTS': { + 'gfx90a': ('TA', 44), + 'gfx908': ('TA', 44), + 'gfx940': ('TA', 32), + 'gfx941': ('TA', 32), + 'gfx942': ('TA', 32), + 'gfx950': ('TA', 32), + }, + 'TA_BUFFER_WRITE_WAVEFRONTS': { + 'gfx90a': ('TA', 46), + 'gfx908': ('TA', 46), + 'gfx940': ('TA', 34), + 'gfx941': ('TA', 34), + 'gfx942': ('TA', 34), + 'gfx950': ('TA', 34), + }, + 'TA_DATA_STALLED_BY_TC_CYCLES': { + 'gfx90a': ('TA', 56), + 'gfx908': ('TA', 56), + 'gfx940': ('TA', 44), + 'gfx941': ('TA', 44), + 'gfx942': ('TA', 44), + 'gfx950': ('TA', 44), + }, + 'TA_FLAT_ATOMIC_WAVEFRONTS': { + 'gfx90a': ('TA', 103), + 'gfx908': ('TA', 103), + 'gfx940': ('TA', 54), + 'gfx941': ('TA', 54), + 'gfx942': ('TA', 54), + 'gfx950': ('TA', 54), + }, + 'TA_FLAT_LOAD_WAVEFRONTS': { + 'gfx10': ('TA', 101), + 'gfx1010': ('TA', 101), + 'gfx1030': ('TA', 101), + 'gfx1031': ('TA', 101), + 'gfx1032': ('TA', 101), + }, + 'TA_FLAT_READ_WAVEFRONTS': { + 'gfx9': ('TA', 101), + 'gfx900': ('TA', 101), + 'gfx906': ('TA', 101), + 'gfx908': ('TA', 101), + 'gfx90a': ('TA', 101), + 'gfx940': ('TA', 52), + 'gfx941': ('TA', 52), + 'gfx942': ('TA', 52), + 'gfx950': ('TA', 52), + }, + 'TA_FLAT_STORE_WAVEFRONTS': { + 'gfx10': ('TA', 102), + 'gfx1010': ('TA', 102), + 'gfx1030': ('TA', 102), + 'gfx1031': ('TA', 102), + 'gfx1032': ('TA', 102), + }, + 'TA_FLAT_WAVEFRONTS': { + 'gfx90a': ('TA', 100), + 'gfx908': ('TA', 100), + 'gfx940': ('TA', 51), + 'gfx941': ('TA', 51), + 'gfx942': ('TA', 51), + 'gfx950': ('TA', 51), + }, + 'TA_FLAT_WRITE_WAVEFRONTS': { + 'gfx9': ('TA', 102), + 'gfx900': ('TA', 102), + 'gfx906': ('TA', 102), + 'gfx908': ('TA', 102), + 'gfx90a': ('TA', 102), + 'gfx940': ('TA', 53), + 'gfx941': ('TA', 53), + 'gfx942': ('TA', 53), + 'gfx950': ('TA', 53), + }, + 'TA_TA_BUSY': { + 'gfx10': ('TA', 15), + 'gfx1010': ('TA', 15), + 'gfx1030': ('TA', 15), + 'gfx1031': ('TA', 15), + 'gfx1032': ('TA', 15), + 'gfx11': ('TA', 15), + 'gfx1100': ('TA', 15), + 'gfx1101': ('TA', 15), + 'gfx1102': ('TA', 15), + 'gfx12': ('TA', 15), + 'gfx1200': ('TA', 15), + 'gfx1201': ('TA', 15), + 'gfx9': ('TA', 15), + 'gfx900': ('TA', 15), + 'gfx906': ('TA', 15), + 'gfx908': ('TA', 15), + 'gfx90a': ('TA', 15), + 'gfx940': ('TA', 13), + 'gfx941': ('TA', 13), + 'gfx942': ('TA', 13), + 'gfx950': ('TA', 13), + }, + 'TA_TOTAL_WAVEFRONTS': { + 'gfx90a': ('TA', 32), + 'gfx908': ('TA', 32), + 'gfx940': ('TA', 29), + 'gfx941': ('TA', 29), + 'gfx942': ('TA', 29), + 'gfx950': ('TA', 29), + }, + 'TA_BUFFER_READ_LDS_WAVEFRONTS': { + 'gfx950': ('TA', 70), + }, + 'TA_FLAT_READ_LDS_WAVEFRONTS': { + 'gfx950': ('TA', 71), + }, + 'TA_BUFFER_COALESCEABLE_WAVEFRONTS': { + 'gfx950': ('TA', 36), + }, + 'TA_FLAT_COALESCEABLE_WAVEFRONTS': { + 'gfx950': ('TA', 55), + }, + 'TCA_BUSY': { + 'gfx90a': ('TCA', 2), + 'gfx908': ('TCA', 2), + 'gfx940': ('TCA', 2), + 'gfx941': ('TCA', 2), + 'gfx942': ('TCA', 2), + 'gfx950': ('TCA', 2), + }, + 'TCA_CYCLE': { + 'gfx90a': ('TCA', 1), + 'gfx908': ('TCA', 1), + 'gfx940': ('TCA', 1), + 'gfx941': ('TCA', 1), + 'gfx942': ('TCA', 1), + 'gfx950': ('TCA', 1), + }, + 'TCC_ALL_TC_OP_INV_EVICT': { + 'gfx90a': ('TCC', 80), + 'gfx908': ('TCC', 80), + 'gfx940': ('TCC', 80), + 'gfx941': ('TCC', 80), + 'gfx942': ('TCC', 80), + 'gfx950': ('TCC', 86), + }, + 'TCC_ALL_TC_OP_WB_WRITEBACK': { + 'gfx90a': ('TCC', 73), + 'gfx908': ('TCC', 73), + 'gfx940': ('TCC', 73), + 'gfx941': ('TCC', 73), + 'gfx942': ('TCC', 73), + 'gfx950': ('TCC', 79), + }, + 'TCC_ATOMIC': { + 'gfx90a': ('TCC', 14), + 'gfx908': ('TCC', 14), + 'gfx940': ('TCC', 14), + 'gfx941': ('TCC', 14), + 'gfx942': ('TCC', 14), + 'gfx950': ('TCC', 18), + }, + 'TCC_BUSY': { + 'gfx90a': ('TCC', 2), + 'gfx908': ('TCC', 2), + 'gfx940': ('TCC', 2), + 'gfx941': ('TCC', 2), + 'gfx942': ('TCC', 2), + 'gfx950': ('TCC', 2), + }, + 'TCC_CC_REQ': { + 'gfx90a': ('TCC', 7), + 'gfx908': ('TCC', 7), + 'gfx940': ('TCC', 7), + 'gfx941': ('TCC', 7), + 'gfx942': ('TCC', 7), + 'gfx950': ('TCC', 11), + }, + 'TCC_CYCLE': { + 'gfx90a': ('TCC', 1), + 'gfx908': ('TCC', 1), + 'gfx940': ('TCC', 1), + 'gfx941': ('TCC', 1), + 'gfx942': ('TCC', 1), + 'gfx950': ('TCC', 1), + }, + 'TCC_EA0_ATOMIC': { + 'gfx908': ('TCC', 36), + 'gfx940': ('TCC', 36), + 'gfx941': ('TCC', 36), + 'gfx942': ('TCC', 36), + 'gfx950': ('TCC', 40), + }, + 'TCC_EA0_ATOMIC_LEVEL': { + 'gfx908': ('TCC', 37), + 'gfx940': ('TCC', 37), + 'gfx941': ('TCC', 37), + 'gfx942': ('TCC', 37), + 'gfx950': ('TCC', 41), + }, + 'TCC_EA0_RDREQ': { + 'gfx908': ('TCC', 38), + 'gfx940': ('TCC', 38), + 'gfx941': ('TCC', 38), + 'gfx942': ('TCC', 38), + 'gfx950': ('TCC', 42), + }, + 'TCC_EA0_RDREQ_32B': { + 'gfx908': ('TCC', 39), + 'gfx940': ('TCC', 39), + 'gfx941': ('TCC', 39), + 'gfx942': ('TCC', 39), + 'gfx950': ('TCC', 43), + }, + 'TCC_EA0_RDREQ_DRAM': { + 'gfx908': ('TCC', 102), + 'gfx940': ('TCC', 102), + 'gfx941': ('TCC', 102), + 'gfx942': ('TCC', 102), + 'gfx950': ('TCC', 108), + }, + 'TCC_EA0_RDREQ_DRAM_CREDIT_STALL': { + 'gfx908': ('TCC', 43), + 'gfx940': ('TCC', 43), + 'gfx941': ('TCC', 43), + 'gfx942': ('TCC', 43), + 'gfx950': ('TCC', 49), + }, + 'TCC_EA0_RDREQ_GMI_CREDIT_STALL': { + 'gfx908': ('TCC', 42), + 'gfx940': ('TCC', 42), + 'gfx941': ('TCC', 42), + 'gfx942': ('TCC', 42), + 'gfx950': ('TCC', 48), + }, + 'TCC_EA0_RDREQ_IO_CREDIT_STALL': { + 'gfx908': ('TCC', 41), + 'gfx940': ('TCC', 41), + 'gfx941': ('TCC', 41), + 'gfx942': ('TCC', 41), + 'gfx950': ('TCC', 47), + }, + 'TCC_EA0_RDREQ_LEVEL': { + 'gfx908': ('TCC', 44), + 'gfx940': ('TCC', 44), + 'gfx941': ('TCC', 44), + 'gfx942': ('TCC', 44), + 'gfx950': ('TCC', 50), + }, + 'TCC_EA0_RD_UNCACHED_32B': { + 'gfx908': ('TCC', 40), + 'gfx940': ('TCC', 40), + 'gfx941': ('TCC', 40), + 'gfx942': ('TCC', 40), + 'gfx950': ('TCC', 46), + }, + 'TCC_EA0_WRREQ': { + 'gfx908': ('TCC', 26), + 'gfx90a': ('TCC', 26), + 'gfx940': ('TCC', 26), + 'gfx941': ('TCC', 26), + 'gfx942': ('TCC', 26), + 'gfx950': ('TCC', 30), + }, + 'TCC_EA0_WRREQ_64B': { + 'gfx908': ('TCC', 27), + 'gfx90a': ('TCC', 27), + 'gfx940': ('TCC', 27), + 'gfx941': ('TCC', 27), + 'gfx942': ('TCC', 27), + 'gfx950': ('TCC', 31), + }, + 'TCC_EA0_WRREQ_DRAM': { + 'gfx908': ('TCC', 103), + 'gfx940': ('TCC', 103), + 'gfx941': ('TCC', 103), + 'gfx942': ('TCC', 103), + 'gfx950': ('TCC', 109), + }, + 'TCC_EA0_WRREQ_DRAM_CREDIT_STALL': { + 'gfx908': ('TCC', 33), + 'gfx940': ('TCC', 33), + 'gfx941': ('TCC', 33), + 'gfx942': ('TCC', 33), + 'gfx950': ('TCC', 37), + }, + 'TCC_EA0_WRREQ_GMI_CREDIT_STALL': { + 'gfx908': ('TCC', 32), + 'gfx940': ('TCC', 32), + 'gfx941': ('TCC', 32), + 'gfx942': ('TCC', 32), + 'gfx950': ('TCC', 36), + }, + 'TCC_EA0_WRREQ_IO_CREDIT_STALL': { + 'gfx908': ('TCC', 31), + 'gfx940': ('TCC', 31), + 'gfx941': ('TCC', 31), + 'gfx942': ('TCC', 31), + 'gfx950': ('TCC', 35), + }, + 'TCC_EA0_WRREQ_LEVEL': { + 'gfx908': ('TCC', 35), + 'gfx940': ('TCC', 35), + 'gfx941': ('TCC', 35), + 'gfx942': ('TCC', 35), + 'gfx950': ('TCC', 39), + }, + 'TCC_EA0_WRREQ_PROBE_COMMAND': { + 'gfx940': ('TCC', 28), + 'gfx941': ('TCC', 28), + 'gfx942': ('TCC', 28), + 'gfx950': ('TCC', 32), + }, + 'TCC_EA0_WRREQ_STALL': { + 'gfx908': ('TCC', 30), + 'gfx940': ('TCC', 30), + 'gfx941': ('TCC', 30), + 'gfx942': ('TCC', 30), + 'gfx950': ('TCC', 34), + }, + 'TCC_EA0_WR_UNCACHED_32B': { + 'gfx908': ('TCC', 29), + 'gfx940': ('TCC', 29), + 'gfx941': ('TCC', 29), + 'gfx942': ('TCC', 29), + 'gfx950': ('TCC', 33), + }, + 'TCC_EA1_RDREQ': { + 'gfx906': ('TCC', 267), + }, + 'TCC_EA1_RDREQ_32B': { + 'gfx906': ('TCC', 268), + }, + 'TCC_EA1_WRREQ': { + 'gfx906': ('TCC', 256), + }, + 'TCC_EA1_WRREQ_64B': { + 'gfx906': ('TCC', 257), + }, + 'TCC_EA1_WRREQ_STALL': { + 'gfx906': ('TCC', 260), + }, + 'TCC_EA_ATOMIC': { + 'gfx90a': ('TCC', 36), + }, + 'TCC_EA_ATOMIC_LEVEL': { + 'gfx90a': ('TCC', 37), + }, + 'TCC_EA_RDREQ': { + 'gfx9': ('TCC', 41), + 'gfx900': ('TCC', 41), + 'gfx906': ('TCC', 41), + 'gfx908': ('TCC', 38), + 'gfx90a': ('TCC', 38), + }, + 'TCC_EA_RDREQ_32B': { + 'gfx9': ('TCC', 42), + 'gfx900': ('TCC', 42), + 'gfx906': ('TCC', 42), + 'gfx908': ('TCC', 39), + 'gfx90a': ('TCC', 39), + }, + 'TCC_EA_RDREQ_DRAM': { + 'gfx90a': ('TCC', 102), + }, + 'TCC_EA_RDREQ_DRAM_CREDIT_STALL': { + 'gfx90a': ('TCC', 43), + }, + 'TCC_EA_RDREQ_GMI_CREDIT_STALL': { + 'gfx90a': ('TCC', 42), + }, + 'TCC_EA_RDREQ_IO_CREDIT_STALL': { + 'gfx90a': ('TCC', 41), + }, + 'TCC_EA_RDREQ_LEVEL': { + 'gfx90a': ('TCC', 44), + }, + 'TCC_EA_RD_UNCACHED_32B': { + 'gfx90a': ('TCC', 40), + }, + 'TCC_EA_WRREQ': { + 'gfx9': ('TCC', 29), + 'gfx900': ('TCC', 29), + 'gfx906': ('TCC', 29), + 'gfx908': ('TCC', 26), + 'gfx90a': ('TCC', 26), + }, + 'TCC_EA_WRREQ_64B': { + 'gfx9': ('TCC', 30), + 'gfx900': ('TCC', 30), + 'gfx906': ('TCC', 30), + 'gfx908': ('TCC', 27), + 'gfx90a': ('TCC', 27), + }, + 'TCC_EA_WRREQ_DRAM': { + 'gfx90a': ('TCC', 103), + }, + 'TCC_EA_WRREQ_DRAM_CREDIT_STALL': { + 'gfx90a': ('TCC', 33), + }, + 'TCC_EA_WRREQ_GMI_CREDIT_STALL': { + 'gfx90a': ('TCC', 32), + }, + 'TCC_EA_WRREQ_IO_CREDIT_STALL': { + 'gfx90a': ('TCC', 31), + }, + 'TCC_EA_WRREQ_LEVEL': { + 'gfx90a': ('TCC', 35), + }, + 'TCC_EA_WRREQ_STALL': { + 'gfx9': ('TCC', 33), + 'gfx900': ('TCC', 33), + 'gfx906': ('TCC', 33), + 'gfx908': ('TCC', 30), + 'gfx90a': ('TCC', 30), + }, + 'TCC_EA_WR_UNCACHED_32B': { + 'gfx90a': ('TCC', 29), + }, + 'TCC_HIT': { + 'gfx9': ('TCC', 20), + 'gfx900': ('TCC', 20), + 'gfx906': ('TCC', 20), + 'gfx908': ('TCC', 17), + 'gfx90a': ('TCC', 17), + 'gfx940': ('TCC', 17), + 'gfx941': ('TCC', 17), + 'gfx942': ('TCC', 17), + 'gfx950': ('TCC', 21), + }, + 'TCC_INTERNAL_PROBE': { + 'gfx940': ('TCC', 11), + 'gfx941': ('TCC', 11), + 'gfx942': ('TCC', 11), + 'gfx950': ('TCC', 15), + }, + 'TCC_MISS': { + 'gfx9': ('TCC', 22), + 'gfx900': ('TCC', 22), + 'gfx906': ('TCC', 22), + 'gfx908': ('TCC', 19), + 'gfx90a': ('TCC', 19), + 'gfx940': ('TCC', 19), + 'gfx941': ('TCC', 19), + 'gfx942': ('TCC', 19), + 'gfx950': ('TCC', 23), + }, + 'TCC_NC_REQ': { + 'gfx908': ('TCC', 5), + 'gfx90a': ('TCC', 5), + 'gfx940': ('TCC', 5), + 'gfx941': ('TCC', 5), + 'gfx942': ('TCC', 5), + 'gfx950': ('TCC', 9), + }, + 'TCC_NORMAL_EVICT': { + 'gfx90a': ('TCC', 74), + 'gfx908': ('TCC', 74), + 'gfx940': ('TCC', 74), + 'gfx941': ('TCC', 74), + 'gfx942': ('TCC', 74), + 'gfx950': ('TCC', 80), + }, + 'TCC_NORMAL_WRITEBACK': { + 'gfx908': ('TCC', 68), + 'gfx90a': ('TCC', 68), + 'gfx940': ('TCC', 68), + 'gfx941': ('TCC', 68), + 'gfx942': ('TCC', 68), + 'gfx950': ('TCC', 74), + }, + 'TCC_PROBE': { + 'gfx90a': ('TCC', 9), + 'gfx908': ('TCC', 9), + 'gfx940': ('TCC', 9), + 'gfx941': ('TCC', 9), + 'gfx942': ('TCC', 9), + 'gfx950': ('TCC', 13), + }, + 'TCC_PROBE_ALL': { + 'gfx90a': ('TCC', 10), + 'gfx908': ('TCC', 10), + 'gfx940': ('TCC', 10), + 'gfx941': ('TCC', 10), + 'gfx942': ('TCC', 10), + 'gfx950': ('TCC', 14), + }, + 'TCC_PROBE_EVICT': { + 'gfx940': ('TCC', 81), + 'gfx941': ('TCC', 81), + 'gfx942': ('TCC', 81), + 'gfx950': ('TCC', 87), + }, + 'TCC_READ': { + 'gfx90a': ('TCC', 12), + 'gfx908': ('TCC', 12), + 'gfx940': ('TCC', 12), + 'gfx941': ('TCC', 12), + 'gfx942': ('TCC', 12), + 'gfx950': ('TCC', 16), + }, + 'TCC_REQ': { + 'gfx90a': ('TCC', 3), + 'gfx908': ('TCC', 3), + 'gfx940': ('TCC', 3), + 'gfx941': ('TCC', 3), + 'gfx942': ('TCC', 3), + 'gfx950': ('TCC', 6), + }, + 'TCC_RW_REQ': { + 'gfx90a': ('TCC', 8), + 'gfx908': ('TCC', 8), + 'gfx940': ('TCC', 8), + 'gfx941': ('TCC', 8), + 'gfx942': ('TCC', 8), + 'gfx950': ('TCC', 12), + }, + 'TCC_STREAMING_REQ': { + 'gfx90a': ('TCC', 4), + 'gfx908': ('TCC', 4), + 'gfx940': ('TCC', 4), + 'gfx941': ('TCC', 4), + 'gfx942': ('TCC', 4), + 'gfx950': ('TCC', 7), + }, + 'TCC_TAG_STALL': { + 'gfx90a': ('TCC', 45), + 'gfx908': ('TCC', 45), + 'gfx940': ('TCC', 45), + 'gfx941': ('TCC', 45), + 'gfx942': ('TCC', 45), + 'gfx950': ('TCC', 51), + }, + 'TCC_TOO_MANY_EA_WRREQS_STALL': { + 'gfx90a': ('TCC', 34), + 'gfx908': ('TCC', 34), + 'gfx940': ('TCC', 34), + 'gfx941': ('TCC', 34), + 'gfx942': ('TCC', 34), + 'gfx950': ('TCC', 38), + }, + 'TCC_UC_REQ': { + 'gfx90a': ('TCC', 6), + 'gfx908': ('TCC', 6), + 'gfx940': ('TCC', 6), + 'gfx941': ('TCC', 6), + 'gfx942': ('TCC', 6), + 'gfx950': ('TCC', 10), + }, + 'TCC_WRITE': { + 'gfx90a': ('TCC', 13), + 'gfx908': ('TCC', 13), + 'gfx940': ('TCC', 13), + 'gfx941': ('TCC', 13), + 'gfx942': ('TCC', 13), + 'gfx950': ('TCC', 17), + }, + 'TCC_WRITEBACK': { + 'gfx90a': ('TCC', 22), + 'gfx908': ('TCC', 22), + 'gfx940': ('TCC', 22), + 'gfx941': ('TCC', 22), + 'gfx942': ('TCC', 22), + 'gfx950': ('TCC', 26), + }, + 'TCC_BUBBLE': { + 'gfx90a': ('TCC', 56), + 'gfx908': ('TCC', 56), + 'gfx940': ('TCC', 56), + 'gfx941': ('TCC', 56), + 'gfx942': ('TCC', 56), + 'gfx950': ('TCC', 62), + }, + 'TCC_EA0_RDREQ_DRAM_32B': { + 'gfx950': ('TCC', 112), + }, + 'TCC_EA0_RDREQ_GMI_32B': { + 'gfx950': ('TCC', 113), + }, + 'TCC_EA0_RDREQ_IO_32B': { + 'gfx950': ('TCC', 114), + }, + 'TCC_EA0_WRREQ_WRITE_DRAM_32B': { + 'gfx950': ('TCC', 115), + }, + 'TCC_EA0_WRREQ_WRITE_ATOMIC_32B': { + 'gfx950': ('TCC', 116), + }, + 'TCC_EA0_WRREQ_WRITE_GMI_32B': { + 'gfx950': ('TCC', 117), + }, + 'TCC_EA0_WRREQ_ATOMIC_GMI_32B': { + 'gfx950': ('TCC', 118), + }, + 'TCC_EA0_WRREQ_WRITE_IO_32B': { + 'gfx950': ('TCC', 119), + }, + 'TCC_EA0_WRREQ_ATOMIC_IO_32B': { + 'gfx950': ('TCC', 120), + }, + 'TCC_READ_SECTORS': { + 'gfx950': ('TCC', 3), + }, + 'TCC_WRITE_SECTORS': { + 'gfx950': ('TCC', 4), + }, + 'TCC_ATOMIC_SECTORS': { + 'gfx950': ('TCC', 5), + }, + 'TCC_BYPASS_REQ': { + 'gfx950': ('TCC', 8), + }, + 'TCC_LATENCY_FIFO_FULL': { + 'gfx950': ('TCC', 27), + }, + 'TCC_SRC_FIFO_FULL': { + 'gfx950': ('TCC', 28), + }, + 'TCC_EA0_RDREQ_64B': { + 'gfx950': ('TCC', 44), + }, + 'TCC_EA0_RDREQ_128B': { + 'gfx950': ('TCC', 45), + }, + 'TCC_IB_REQ': { + 'gfx950': ('TCC', 67), + }, + 'TCC_IB_STALL': { + 'gfx950': ('TCC', 68), + }, + 'TCC_EA0_WRREQ_ATOMIC_DRAM': { + 'gfx950': ('TCC', 111), + }, + 'TCC_EA0_WRREQ_WRITE_DRAM': { + 'gfx950': ('TCC', 110), + }, + 'TCC_EA0_WRREQ_ATOMIC_DRAM_32B': { + 'gfx950': ('TCC', 116), + }, + 'TCC_CLIENT184_REQ': { + 'gfx908': ('TCC', 312), + }, + 'TCC_CLIENT185_REQ': { + 'gfx908': ('TCC', 313), + }, + 'TCC_CLIENT186_REQ': { + 'gfx908': ('TCC', 314), + }, + 'TCC_CLIENT187_REQ': { + 'gfx908': ('TCC', 315), + }, + 'TCC_CLIENT188_REQ': { + 'gfx908': ('TCC', 316), + }, + 'TCC_CLIENT189_REQ': { + 'gfx908': ('TCC', 317), + }, + 'TCC_CLIENT190_REQ': { + 'gfx908': ('TCC', 318), + }, + 'TCC_CLIENT191_REQ': { + 'gfx908': ('TCC', 319), + }, + 'TCP_REQ': { + 'gfx12': ('TCP', 9), + 'gfx1200': ('TCP', 9), + 'gfx1201': ('TCP', 9), + }, + 'TCP_REQ_MISS': { + 'gfx12': ('TCP', 17), + 'gfx1200': ('TCP', 17), + 'gfx1201': ('TCP', 17), + }, + 'TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES': { + 'gfx90a': ('TCP', 13), + 'gfx908': ('TCP', 13), + 'gfx940': ('TCP', 12), + 'gfx941': ('TCP', 12), + 'gfx942': ('TCP', 12), + 'gfx950': ('TCP', 12), + }, + 'TCP_GATE_EN1': { + 'gfx90a': ('TCP', 0), + 'gfx908': ('TCP', 0), + 'gfx940': ('TCP', 0), + 'gfx941': ('TCP', 0), + 'gfx942': ('TCP', 0), + 'gfx950': ('TCP', 0), + }, + 'TCP_GATE_EN2': { + 'gfx90a': ('TCP', 1), + 'gfx908': ('TCP', 1), + 'gfx940': ('TCP', 1), + 'gfx941': ('TCP', 1), + 'gfx942': ('TCP', 1), + 'gfx950': ('TCP', 1), + }, + 'TCP_PENDING_STALL_CYCLES': { + 'gfx90a': ('TCP', 22), + 'gfx908': ('TCP', 22), + 'gfx940': ('TCP', 21), + 'gfx941': ('TCP', 21), + 'gfx942': ('TCP', 21), + 'gfx950': ('TCP', 21), + }, + 'TCP_READ_TAGCONFLICT_STALL_CYCLES': { + 'gfx90a': ('TCP', 11), + 'gfx908': ('TCP', 11), + 'gfx940': ('TCP', 10), + 'gfx941': ('TCP', 10), + 'gfx942': ('TCP', 10), + 'gfx950': ('TCP', 10), + }, + 'TCP_TA_TCP_STATE_READ': { + 'gfx90a': ('TCP', 27), + 'gfx908': ('TCP', 27), + 'gfx940': ('TCP', 25), + 'gfx941': ('TCP', 25), + 'gfx942': ('TCP', 25), + 'gfx950': ('TCP', 25), + }, + 'TCP_TCC_ATOMIC_WITHOUT_RET_REQ': { + 'gfx90a': ('TCP', 72), + 'gfx908': ('TCP', 72), + 'gfx940': ('TCP', 68), + 'gfx941': ('TCP', 68), + 'gfx942': ('TCP', 68), + 'gfx950': ('TCP', 71), + }, + 'TCP_TCC_ATOMIC_WITH_RET_REQ': { + 'gfx90a': ('TCP', 71), + 'gfx908': ('TCP', 71), + 'gfx940': ('TCP', 67), + 'gfx941': ('TCP', 67), + 'gfx942': ('TCP', 67), + 'gfx950': ('TCP', 70), + }, + 'TCP_TCC_CC_ATOMIC_REQ': { + 'gfx90a': ('TCP', 83), + 'gfx908': ('TCP', 83), + 'gfx940': ('TCP', 79), + 'gfx941': ('TCP', 79), + 'gfx942': ('TCP', 79), + 'gfx950': ('TCP', 82), + }, + 'TCP_TCC_CC_READ_REQ': { + 'gfx90a': ('TCP', 81), + 'gfx908': ('TCP', 81), + 'gfx940': ('TCP', 77), + 'gfx941': ('TCP', 77), + 'gfx942': ('TCP', 77), + 'gfx950': ('TCP', 80), + }, + 'TCP_TCC_CC_WRITE_REQ': { + 'gfx90a': ('TCP', 82), + 'gfx908': ('TCP', 82), + 'gfx940': ('TCP', 78), + 'gfx941': ('TCP', 78), + 'gfx942': ('TCP', 78), + 'gfx950': ('TCP', 81), + }, + 'TCP_TCC_NC_ATOMIC_REQ': { + 'gfx90a': ('TCP', 77), + 'gfx908': ('TCP', 77), + 'gfx940': ('TCP', 73), + 'gfx941': ('TCP', 73), + 'gfx942': ('TCP', 73), + 'gfx950': ('TCP', 76), + }, + 'TCP_TCC_NC_READ_REQ': { + 'gfx90a': ('TCP', 75), + 'gfx908': ('TCP', 75), + 'gfx940': ('TCP', 71), + 'gfx941': ('TCP', 71), + 'gfx942': ('TCP', 71), + 'gfx950': ('TCP', 74), + }, + 'TCP_TCC_NC_WRITE_REQ': { + 'gfx90a': ('TCP', 76), + 'gfx908': ('TCP', 76), + 'gfx940': ('TCP', 72), + 'gfx941': ('TCP', 72), + 'gfx942': ('TCP', 72), + 'gfx950': ('TCP', 75), + }, + 'TCP_TCC_READ_REQ': { + 'gfx90a': ('TCP', 69), + 'gfx908': ('TCP', 69), + 'gfx940': ('TCP', 65), + 'gfx941': ('TCP', 65), + 'gfx942': ('TCP', 65), + 'gfx950': ('TCP', 68), + }, + 'TCP_TCC_READ_REQ_LATENCY': { + 'gfx90a': ('TCP', 66), + 'gfx908': ('TCP', 66), + 'gfx950': ('TCP', 65), + }, + 'TCP_TCC_RW_ATOMIC_REQ': { + 'gfx90a': ('TCP', 87), + 'gfx908': ('TCP', 87), + 'gfx940': ('TCP', 82), + 'gfx941': ('TCP', 82), + 'gfx942': ('TCP', 82), + 'gfx950': ('TCP', 85), + }, + 'TCP_TCC_RW_READ_REQ': { + 'gfx90a': ('TCP', 85), + 'gfx908': ('TCP', 85), + 'gfx940': ('TCP', 80), + 'gfx941': ('TCP', 80), + 'gfx942': ('TCP', 80), + 'gfx950': ('TCP', 83), + }, + 'TCP_TCC_RW_WRITE_REQ': { + 'gfx90a': ('TCP', 86), + 'gfx908': ('TCP', 86), + 'gfx940': ('TCP', 81), + 'gfx941': ('TCP', 81), + 'gfx942': ('TCP', 81), + 'gfx950': ('TCP', 84), + }, + 'TCP_TCC_UC_ATOMIC_REQ': { + 'gfx90a': ('TCP', 80), + 'gfx908': ('TCP', 80), + 'gfx940': ('TCP', 76), + 'gfx941': ('TCP', 76), + 'gfx942': ('TCP', 76), + 'gfx950': ('TCP', 79), + }, + 'TCP_TCC_UC_READ_REQ': { + 'gfx90a': ('TCP', 78), + 'gfx908': ('TCP', 78), + 'gfx940': ('TCP', 74), + 'gfx941': ('TCP', 74), + 'gfx942': ('TCP', 74), + 'gfx950': ('TCP', 77), + }, + 'TCP_TCC_UC_WRITE_REQ': { + 'gfx90a': ('TCP', 79), + 'gfx908': ('TCP', 79), + 'gfx940': ('TCP', 75), + 'gfx941': ('TCP', 75), + 'gfx942': ('TCP', 75), + 'gfx950': ('TCP', 78), + }, + 'TCP_TCC_WRITE_REQ': { + 'gfx90a': ('TCP', 70), + 'gfx908': ('TCP', 70), + 'gfx940': ('TCP', 66), + 'gfx941': ('TCP', 66), + 'gfx942': ('TCP', 66), + 'gfx950': ('TCP', 69), + }, + 'TCP_TCC_WRITE_REQ_LATENCY': { + 'gfx90a': ('TCP', 67), + 'gfx908': ('TCP', 67), + 'gfx950': ('TCP', 66), + }, + 'TCP_TCP_LATENCY': { + 'gfx90a': ('TCP', 65), + 'gfx908': ('TCP', 65), + 'gfx950': ('TCP', 64), + }, + 'TCP_TCP_TA_DATA_STALL_CYCLES': { + 'gfx9': ('TCP', 6), + 'gfx900': ('TCP', 6), + 'gfx906': ('TCP', 6), + 'gfx908': ('TCP', 6), + 'gfx90a': ('TCP', 6), + 'gfx940': ('TCP', 6), + 'gfx941': ('TCP', 6), + 'gfx942': ('TCP', 6), + 'gfx950': ('TCP', 6), + }, + 'TCP_TCR_TCP_STALL_CYCLES': { + 'gfx90a': ('TCP', 8), + 'gfx908': ('TCP', 8), + 'gfx940': ('TCP', 8), + 'gfx941': ('TCP', 8), + 'gfx942': ('TCP', 8), + 'gfx950': ('TCP', 8), + }, + 'TCP_TD_TCP_STALL_CYCLES': { + 'gfx90a': ('TCP', 7), + 'gfx908': ('TCP', 7), + 'gfx940': ('TCP', 7), + 'gfx941': ('TCP', 7), + 'gfx942': ('TCP', 7), + 'gfx950': ('TCP', 7), + }, + 'TCP_TOTAL_ACCESSES': { + 'gfx90a': ('TCP', 29), + 'gfx908': ('TCP', 29), + 'gfx940': ('TCP', 27), + 'gfx941': ('TCP', 27), + 'gfx942': ('TCP', 27), + 'gfx950': ('TCP', 27), + }, + 'TCP_TOTAL_ATOMIC_WITHOUT_RET': { + 'gfx90a': ('TCP', 39), + 'gfx908': ('TCP', 39), + 'gfx940': ('TCP', 37), + 'gfx941': ('TCP', 37), + 'gfx942': ('TCP', 37), + 'gfx950': ('TCP', 37), + }, + 'TCP_TOTAL_ATOMIC_WITH_RET': { + 'gfx90a': ('TCP', 38), + 'gfx908': ('TCP', 38), + 'gfx940': ('TCP', 36), + 'gfx941': ('TCP', 36), + 'gfx942': ('TCP', 36), + 'gfx950': ('TCP', 36), + }, + 'TCP_TOTAL_CACHE_ACCESSES': { + 'gfx90a': ('TCP', 60), + 'gfx908': ('TCP', 60), + 'gfx940': ('TCP', 60), + 'gfx941': ('TCP', 60), + 'gfx942': ('TCP', 60), + 'gfx950': ('TCP', 58), + }, + 'TCP_TOTAL_READ': { + 'gfx90a': ('TCP', 30), + 'gfx908': ('TCP', 30), + 'gfx940': ('TCP', 28), + 'gfx941': ('TCP', 28), + 'gfx942': ('TCP', 28), + 'gfx950': ('TCP', 28), + }, + 'TCP_TOTAL_WRITE': { + 'gfx90a': ('TCP', 32), + 'gfx908': ('TCP', 32), + 'gfx940': ('TCP', 30), + 'gfx941': ('TCP', 30), + 'gfx942': ('TCP', 30), + 'gfx950': ('TCP', 30), + }, + 'TCP_TOTAL_WRITEBACK_INVALIDATES': { + 'gfx90a': ('TCP', 45), + 'gfx908': ('TCP', 45), + 'gfx940': ('TCP', 43), + 'gfx941': ('TCP', 43), + 'gfx942': ('TCP', 43), + 'gfx950': ('TCP', 41), + }, + 'TCP_UTCL1_PERMISSION_MISS': { + 'gfx90a': ('TCP', 50), + 'gfx908': ('TCP', 50), + 'gfx940': ('TCP', 49), + 'gfx941': ('TCP', 49), + 'gfx942': ('TCP', 49), + 'gfx950': ('TCP', 47), + }, + 'TCP_UTCL1_REQUEST': { + 'gfx90a': ('TCP', 47), + 'gfx908': ('TCP', 47), + 'gfx940': ('TCP', 45), + 'gfx941': ('TCP', 45), + 'gfx942': ('TCP', 45), + 'gfx950': ('TCP', 43), + }, + 'TCP_UTCL1_TRANSLATION_HIT': { + 'gfx90a': ('TCP', 49), + 'gfx908': ('TCP', 49), + 'gfx940': ('TCP', 48), + 'gfx941': ('TCP', 48), + 'gfx942': ('TCP', 48), + 'gfx950': ('TCP', 46), + }, + 'TCP_UTCL1_TRANSLATION_MISS': { + 'gfx90a': ('TCP', 48), + 'gfx908': ('TCP', 48), + 'gfx940': ('TCP', 47), + 'gfx941': ('TCP', 47), + 'gfx942': ('TCP', 47), + 'gfx950': ('TCP', 45), + }, + 'TCP_VOLATILE': { + 'gfx90a': ('TCP', 28), + 'gfx908': ('TCP', 28), + 'gfx940': ('TCP', 26), + 'gfx941': ('TCP', 26), + 'gfx942': ('TCP', 26), + 'gfx950': ('TCP', 26), + }, + 'TCP_WRITE_TAGCONFLICT_STALL_CYCLES': { + 'gfx90a': ('TCP', 12), + 'gfx908': ('TCP', 12), + 'gfx940': ('TCP', 11), + 'gfx941': ('TCP', 11), + 'gfx942': ('TCP', 11), + 'gfx950': ('TCP', 11), + }, + 'TCP_CACHE_MISS': { + 'gfx950': ('TCP', 63), + }, + 'TCP_TCP_TA_ADDR_STALL_CYCLES': { + 'gfx950': ('TCP', 5), + }, + 'TCP_LFIFO_STALL_CYCLES': { + 'gfx950': ('TCP', 15), + }, + 'TCP_RFIFO_STALL_CYCLES': { + 'gfx950': ('TCP', 16), + }, + 'TCP_TCR_RDRET_STALL': { + 'gfx950': ('TCP', 17), + }, + 'TCP_UTCL1_SERIALIZATION_STALL': { + 'gfx950': ('TCP', 23), + }, + 'TCP_UTCL1_THRASHING_STALL': { + 'gfx950': ('TCP', 44), + }, + 'TCP_UTCL1_TRANSLATION_MISS_UNDER_MISS': { + 'gfx950': ('TCP', 48), + }, + 'TCP_UTCL1_STALL_INFLIGHT_MAX': { + 'gfx950': ('TCP', 49), + }, + 'TCP_UTCL1_STALL_LRU_INFLIGHT': { + 'gfx950': ('TCP', 50), + }, + 'TCP_UTCL1_STALL_MULTI_MISS': { + 'gfx950': ('TCP', 51), + }, + 'TCP_UTCL1_LFIFO_FULL': { + 'gfx950': ('TCP', 52), + }, + 'TCP_UTCL1_STALL_LFIFO_NOT_RES': { + 'gfx950': ('TCP', 53), + }, + 'TCP_UTCL1_STALL_UTCL2_REQ_OUT_OF_CREDITS': { + 'gfx950': ('TCP', 54), + }, + 'TCP_CLIENT_UTCL1_INFLIGHT': { + 'gfx950': ('TCP', 55), + }, + 'TCP_TAGRAM0_REQ': { + 'gfx950': ('TCP', 59), + }, + 'TCP_TAGRAM1_REQ': { + 'gfx950': ('TCP', 60), + }, + 'TCP_TAGRAM2_REQ': { + 'gfx950': ('TCP', 61), + }, + 'TCP_TAGRAM3_REQ': { + 'gfx950': ('TCP', 62), + }, + 'TCP_TCC_WRITE_REQ_HOLE_LATENCY': { + 'gfx950': ('TCP', 67), + }, + 'TCP_TOTAL_WBINVL1_VOL': { + 'gfx950': ('TCP', 38), + }, + 'TCP_SQ_TCP_INVALIDATE_VOL': { + 'gfx950': ('TCP', 39), + }, + 'TCP_CP_TCP_INVALIDATE_VOL': { + 'gfx950': ('TCP', 40), + }, + 'TCP_UTCL1_STALL_LFIFO_NO_RES': { + 'gfx950': ('TCP', 53), + }, + 'TD_ATOMIC_WAVEFRONT': { + 'gfx90a': ('TD', 26), + 'gfx908': ('TD', 26), + 'gfx940': ('TD', 17), + 'gfx941': ('TD', 17), + 'gfx942': ('TD', 17), + 'gfx950': ('TD', 17), + }, + 'TD_COALESCABLE_WAVEFRONT': { + 'gfx90a': ('TD', 32), + 'gfx908': ('TD', 32), + 'gfx940': ('TD', 21), + 'gfx941': ('TD', 21), + 'gfx942': ('TD', 21), + 'gfx950': ('TD', 21), + }, + 'TD_LOAD_WAVEFRONT': { + 'gfx90a': ('TD', 25), + 'gfx908': ('TD', 25), + 'gfx940': ('TD', 16), + 'gfx941': ('TD', 16), + 'gfx942': ('TD', 16), + 'gfx950': ('TD', 16), + }, + 'TD_SPI_STALL': { + 'gfx90a': ('TD', 18), + 'gfx940': ('TD', 15), + 'gfx941': ('TD', 15), + 'gfx942': ('TD', 15), + 'gfx950': ('TD', 15), + }, + 'TD_STORE_WAVEFRONT': { + 'gfx90a': ('TD', 27), + 'gfx908': ('TD', 27), + 'gfx940': ('TD', 18), + 'gfx941': ('TD', 18), + 'gfx942': ('TD', 18), + 'gfx950': ('TD', 18), + }, + 'TD_TC_STALL': { + 'gfx90a': ('TD', 15), + 'gfx908': ('TD', 15), + 'gfx940': ('TD', 12), + 'gfx941': ('TD', 12), + 'gfx942': ('TD', 12), + 'gfx950': ('TD', 12), + }, + 'TD_TD_BUSY': { + 'gfx90a': ('TD', 1), + 'gfx908': ('TD', 1), + 'gfx940': ('TD', 1), + 'gfx941': ('TD', 1), + 'gfx942': ('TD', 1), + 'gfx950': ('TD', 1), + }, + 'TD_WRITE_ACKT_WAVEFRONT': { + 'gfx950': ('TD', 27), + }, + 'TD_TD_SP_TRAFFIC': { + 'gfx950': ('TD', 29), + }, + 'SQC_DCACHE_INFLIGHT_LEVEL': { + 'gfx90a': ('SQ', 337), + 'gfx908': ('SQ', 337), + 'gfx940': ('SQ', 337), + 'gfx941': ('SQ', 337), + 'gfx942': ('SQ', 337), + 'gfx950': ('SQ', 337), + }, + 'SQC_ICACHE_INFLIGHT_LEVEL': { + 'gfx90a': ('SQ', 336), + 'gfx908': ('SQ', 336), + 'gfx940': ('SQ', 336), + 'gfx941': ('SQ', 336), + 'gfx942': ('SQ', 336), + 'gfx950': ('SQ', 336), + }, +} \ No newline at end of file diff --git a/tinygrad/runtime/support/amd.py b/tinygrad/runtime/support/amd.py index ae4522e423..27d3e1049f 100644 --- a/tinygrad/runtime/support/amd.py +++ b/tinygrad/runtime/support/amd.py @@ -1,8 +1,6 @@ -import functools, re, tinygrad.runtime.autogen.am +import functools, tinygrad.runtime.autogen.am from dataclasses import dataclass -from tinygrad.helpers import getbits, fetch - -ROCM_URL = "https://raw.githubusercontent.com/ROCm/rocm-systems/cccc350dc620e61ae2554978b62ab3532dc10bd9/projects" +from tinygrad.helpers import getbits @dataclass class AMDReg: @@ -36,22 +34,12 @@ def import_module(name:str, target:tuple[int, ...], submod=""): return getattr(mod, children[-1]) raise ImportError(f"Failed to import {submod+'.' if submod else ''}{name} {'.'.join(map(str, target))}") -def header_download(file, url) -> str: return fetch(f"{url}/{file}", subdir="defines").read_text() - def import_soc(ip): return getattr(tinygrad.runtime.autogen.am, f"soc_{ip[0]}") def import_pmc(ip) -> dict[str, tuple[str, int]]: - res:dict[str, tuple[str, int]] = {} - + from tinygrad.runtime.autogen.am import pmc # NOTE: precise arch for mi300+, generic for others, since rocm headers lack some archs - arch = f"gfx{ip[0]}{ip[1]:x}{ip[2]:x}" if ip[0] == 9 else f"gfx{ip[0]}" - - for sec in header_download("rocprofiler-compute/src/rocprof_compute_soc/profile_configs/counter_defs.yaml", ROCM_URL).split('- name: ')[1:]: - for arch_spec in sec.split('- architectures:')[1:]: - if arch in arch_spec and (block:=re.search(r'block:\s*([A-Za-z0-9_]+)', arch_spec)) and (ev:=re.search(r'event:\s*(\d+)', arch_spec)): - res[sec.splitlines()[0].strip()] = (block.group(1), int(ev.group(1))) - - return res + return {k:x for k,v in pmc.counters.items() if (x:=v.get(f"gfx{ip[0]}{ip[1]:x}{ip[2]:x}" if ip[0] == 9 else f"gfx{ip[0]}", None)) is not None} def import_asic_regs(prefix:str, version:tuple[int, ...], cls=AMDReg) -> dict[str, AMDReg]: return {reg:cls(name=reg, offset=off, segment=seg, fields=fields) for reg,(off,seg,fields) in import_module(prefix, version, submod="regs").items()}