mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 21:21:22 +00:00
592d73ad73
* initial work on adding support of gfx12 in tile_engine for GEMM benchmarking
* add stage("Run TILE_ENGINE_GEMM Tests on gfx1201") to Jenkins config
* make tile_[m/n/k] validation arch dependent
243 lines
8.0 KiB
Python
243 lines
8.0 KiB
Python
# SPDX-License-Identifier: MIT
|
|
# Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
# -*- coding: utf-8 -*-
|
|
|
|
"""
|
|
Mappings and utility functions for kernel code generation.
|
|
"""
|
|
|
|
import subprocess
|
|
import re
|
|
from functools import lru_cache
|
|
|
|
DATA_TYPE_MAP = {
|
|
"fp32": "float",
|
|
"fp16": "ck_tile::half_t",
|
|
"bf16": "ck_tile::bf16_t",
|
|
"int8": "ck_tile::int8_t",
|
|
"fp8": "ck_tile::fp8_t",
|
|
"bf8": "ck_tile::bf8_t",
|
|
"int4": "ck_tile::pk_int4_t",
|
|
"int32": "ck_tile::int32_t",
|
|
}
|
|
|
|
LAYOUT_MAP = {
|
|
"r": "ck_tile::tensor_layout::gemm::RowMajor",
|
|
"c": "ck_tile::tensor_layout::gemm::ColumnMajor",
|
|
}
|
|
|
|
DEFAULT_EPILOGUE = """
|
|
using GemmEpilogue = ck_tile::DefaultGemm2DEpilogue<
|
|
ck_tile::DefaultGemm2DEpilogueProblem<ADataType,
|
|
BDataType,
|
|
ck_tile::tuple<>,
|
|
AccDataType,
|
|
CDataType,
|
|
ck_tile::tuple<>,
|
|
CLayout,
|
|
ck_tile::element_wise::PassThrough,
|
|
TilePartitioner::MPerBlock,
|
|
TilePartitioner::NPerBlock,
|
|
kPadM,
|
|
kPadN,
|
|
WarpTileM,
|
|
WarpTileN,
|
|
WarpTileK,
|
|
UniversalGemmProblem::TransposeC,
|
|
true,
|
|
memory_operation>>;
|
|
"""
|
|
|
|
CSHUFFLE_EPILOGUE = """
|
|
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
|
ck_tile::CShuffleEpilogueProblem<ADataType,
|
|
BDataType,
|
|
ck_tile::tuple<>,
|
|
AccDataType,
|
|
CDataType,
|
|
ck_tile::tuple<>,
|
|
CLayout,
|
|
ck_tile::element_wise::PassThrough,
|
|
TilePartitioner::MPerBlock,
|
|
TilePartitioner::NPerBlock,
|
|
WarpM,
|
|
WarpN,
|
|
WarpTileM,
|
|
WarpTileN,
|
|
WarpTileK,
|
|
UniversalGemmProblem::TransposeC,
|
|
memory_operation>>;
|
|
"""
|
|
|
|
PIPELINE_MAP = {
|
|
"mem": ["ck_tile::BaseGemmPipelineAgBgCrMem", "ck_tile::GemmPipelineAgBgCrMem"],
|
|
"compv3": [
|
|
"ck_tile::BaseGemmPipelineAgBgCrCompV3",
|
|
"ck_tile::GemmPipelineAgBgCrCompV3",
|
|
],
|
|
"compv4": [
|
|
"ck_tile::BaseGemmPipelineAgBgCrCompV4",
|
|
"ck_tile::GemmPipelineAgBgCrCompV4",
|
|
],
|
|
}
|
|
|
|
SCHEDULER_MAP = {
|
|
"interwave": "ck_tile::GemmPipelineScheduler::Interwave",
|
|
"intrawave": "ck_tile::GemmPipelineScheduler::Intrawave",
|
|
}
|
|
|
|
EPILOGUE_MAP = {"default": DEFAULT_EPILOGUE, "cshuffle": CSHUFFLE_EPILOGUE}
|
|
|
|
|
|
def BOOL_MAP(b_):
|
|
return {True: "true", False: "false"}[bool(b_)]
|
|
|
|
|
|
# To Do: add some more supported combinations
|
|
warp_tile_supported_combinations = {
|
|
"gfx90a": {
|
|
"fp16_fp16_fp16": [
|
|
[32, 32, 8],
|
|
[16, 16, 16],
|
|
[32, 32, 16],
|
|
[16, 16, 32],
|
|
[4, 64, 16],
|
|
[64, 4, 16],
|
|
],
|
|
"bf16_bf16_bf16": [
|
|
[32, 32, 8],
|
|
[16, 16, 16],
|
|
[32, 32, 16],
|
|
[16, 16, 32],
|
|
[4, 64, 16],
|
|
[64, 4, 16],
|
|
],
|
|
"fp8_fp8_fp16": [[32, 32, 16], [32, 32, 32]],
|
|
"bf8_bf8_fp16": [[32, 32, 16], [32, 32, 32]],
|
|
},
|
|
"gfx942": {
|
|
"fp16_fp16_fp16": [
|
|
[32, 32, 8],
|
|
[16, 16, 16],
|
|
[32, 32, 16],
|
|
[16, 16, 32],
|
|
[4, 64, 16],
|
|
[64, 4, 16],
|
|
],
|
|
"bf16_bf16_bf16": [
|
|
[32, 32, 8],
|
|
[16, 16, 16],
|
|
[32, 32, 16],
|
|
[16, 16, 32],
|
|
[4, 64, 16],
|
|
[64, 4, 16],
|
|
],
|
|
"fp8_fp8_fp16": [[32, 32, 16], [32, 32, 32], [16, 16, 32], [16, 16, 64]],
|
|
"bf8_bf8_fp16": [[32, 32, 16], [32, 32, 32], [16, 16, 64], [16, 16, 32]],
|
|
"int8_int8_int32": [[16, 16, 32], [32, 32, 16]],
|
|
},
|
|
"gfx950": {
|
|
"fp16_fp16_fp16": [
|
|
[32, 32, 8],
|
|
[16, 16, 16],
|
|
[32, 32, 16],
|
|
[16, 16, 32],
|
|
[4, 64, 16],
|
|
[64, 4, 16],
|
|
],
|
|
"bf16_bf16_bf16": [
|
|
[32, 32, 8],
|
|
[16, 16, 16],
|
|
[32, 32, 16],
|
|
[16, 16, 32],
|
|
[4, 64, 16],
|
|
[64, 4, 16],
|
|
],
|
|
"fp8_fp8_fp16": [
|
|
[32, 32, 16],
|
|
[32, 32, 32],
|
|
[16, 16, 32],
|
|
[16, 16, 64],
|
|
[16, 16, 128],
|
|
[32, 32, 64],
|
|
],
|
|
"bf8_bf8_fp16": [
|
|
[32, 32, 16],
|
|
[32, 32, 32],
|
|
[16, 16, 64],
|
|
[16, 16, 32],
|
|
[16, 16, 128],
|
|
[32, 32, 64],
|
|
],
|
|
"fp8_bf8_fp16": [
|
|
[16, 16, 128],
|
|
[32, 32, 64],
|
|
],
|
|
"bf8_fp8_fp16": [
|
|
[16, 16, 128],
|
|
[32, 32, 64],
|
|
],
|
|
},
|
|
"gfx1201": {
|
|
"fp16_fp16_fp16": [
|
|
[16, 16, 16],
|
|
],
|
|
},
|
|
}
|
|
|
|
# To Do: remove some unsupported combinations
|
|
trait_unsupported_combinations = {
|
|
("compv3", "cshuffle", "interwave"),
|
|
("compv3", "default", "interwave"),
|
|
("compv4", "cshuffle", "interwave"),
|
|
("compv4", "default", "interwave"),
|
|
}
|
|
|
|
|
|
ELEMENT_SIZE_MAP = {
|
|
"fp16": 2,
|
|
"bf16": 2,
|
|
"int8": 1,
|
|
"fp8": 1,
|
|
"bf8": 1,
|
|
"int4": 0.5,
|
|
"int32": 4,
|
|
}
|
|
|
|
|
|
def element_size(data_type: str) -> float:
|
|
"""Calculate the size (in bytes) of a single element for given data type."""
|
|
data_type = data_type.lower()
|
|
if data_type not in ELEMENT_SIZE_MAP:
|
|
raise ValueError(f"Unsupported data type: {data_type}")
|
|
return ELEMENT_SIZE_MAP[data_type]
|
|
|
|
|
|
GPU_NAME_PATTERN = re.compile(r"Name:\s*(gfx\d+\w*)")
|
|
|
|
|
|
@lru_cache(maxsize=1)
|
|
def get_gpu_name_by_id(gpu_id: int = 0) -> str:
|
|
"""Retrieve GPU name (e.g. gfx90a) by device ID"""
|
|
try:
|
|
output = subprocess.check_output(
|
|
["rocminfo"], text=True, stderr=subprocess.PIPE, timeout=5
|
|
)
|
|
if matches := GPU_NAME_PATTERN.finditer(output):
|
|
gpu_list = [m.group(1) for m in matches]
|
|
return gpu_list[gpu_id] if gpu_id < len(gpu_list) else ""
|
|
|
|
return ""
|
|
|
|
except subprocess.CalledProcessError as e:
|
|
print(f"GPU query failed (exit {e.returncode}): {e.stderr.strip()}")
|
|
except FileNotFoundError:
|
|
print("ROCm tools not installed (requires rocminfo)")
|
|
except subprocess.TimeoutExpired:
|
|
print("GPU query timeout (5s)")
|
|
except Exception as e:
|
|
print(f"GPU detection error: {str(e)}")
|
|
|
|
return ""
|