ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-13 02:27:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

formatting

Browse Source
This commit is contained in:
khushbu
2025-12-11 22:29:23 -05:00
parent 995d1a5cf6
commit 44aaaacbec
8 changed files with 184 additions and 124 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
example/ck_tile/38_block_scale_gemm/CMakeLists.txt
View File

@@ -16,7 +16,7 @@ if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx12")
# gemm_aquant_quantgrouped_preshufflequant.cpp
# gemm_bquant_quantgrouped_bf8i4.cpp
# gemm_bquant_quantgrouped_fp8i4.cpp
gemm_bquant_quantgrouped_bf16mxfp4.cpp
# gemm_bquant_quantgrouped_bf16mxfp4.cpp
# gemm_bquant_quantgrouped_bf8.cpp
# gemm_bquant_quantgrouped_fp8.cpp
# gemm_bquant_quantgrouped_preshuffleb.cpp

63
example/ck_tile/38_block_scale_gemm/gemm_bquant_quantgrouped_preshufflequant.cpp
View File

@@ -35,30 +35,45 @@ void bquant_quantgrouped_preshufflequant_instance_factory(
QuantGroupSize,
ck_tile::QuantType::BQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8", "bquant", "non-preshuffleb", "preshufflequant", "1x16x128"})] =
[](const ck_tile::ArgParser& arg_parser) {
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::fp8_t,
ck_tile::fp8_t,
ck_tile::half_t,
float>{});
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 16, 128>>;
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
QuantGroupSize,
ck_tile::QuantType::BQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8", "bquant", "non-preshuffleb", "preshufflequant", "1x32x128"})] =
[](const ck_tile::ArgParser& arg_parser) {
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::fp8_t,
ck_tile::fp8_t,
ck_tile::half_t,
float>{});
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 32, 128>>;
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
QuantGroupSize,
ck_tile::QuantType::BQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"bquant",
"non-preshuffleb",
"preshufflequant",
"1x16x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 16, 128>>;
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
QuantGroupSize,
ck_tile::QuantType::BQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"bquant",
"non-preshuffleb",
"preshufflequant",
"1x32x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 32, 128>>;
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
QuantGroupSize,
ck_tile::QuantType::BQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"bquant",
"non-preshuffleb",
"preshufflequant",
"1x64x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 64, 128>>;
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
QuantGroupSize,
ck_tile::QuantType::BQuantGrouped>(arg_parser);
};
// lut[hash_multiple_strings({"bf8", "bquant", "non-preshuffleb", "preshufflequant",
// "1x1x128"})] =
// [](const ck_tile::ArgParser& arg_parser) {

6
example/ck_tile/38_block_scale_gemm/gemm_quant.cpp
View File

@@ -33,14 +33,14 @@ auto create_args(int argc, char* argv[])
"fp8",
"Data type. For AQuant: fp8, bf8, i4fp8, or i4bf8; for Bquant: fp8, bf8, fp8i4, "
"bf8i4 or bf16fp4")
.insert("warmup", "1", "Number of iterations before benchmarking the kernel")
.insert("repeat", "0", "Number of iterations to benchmark the kernel")
.insert("warmup", "50", "Number of iterations before benchmarking the kernel")
.insert("repeat", "1000", "Number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("split_k", "1", "SplitK value")
.insert("device", "0", "Device id that will be used to run the kernel")
.insert("init", "0", "0:random, 1:linear, 2:constant(1)")
.insert("flush_cache", "true", "Flush cache before running the kernel")
.insert("rotating_count", "0", "Rotating count")
.insert("rotating_count", "1000", "Rotating count")
.insert("quant_mode", "bquant", "Choose aquant, bquant, tensor or rowcol")
.insert("preshuffleb", "false", "Enable preshuffle of tensor B")
.insert("preshufflequant", "false", "Enable preshuffle of quant tensor")

6
example/ck_tile/38_block_scale_gemm/run_gemm_quant_example.inc
View File

@@ -357,9 +357,9 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
<< " C_Type = " << ck_tile::DataTypeTraits<typename TypeConfig::CDataType>::name
<< " QuantMode = " << quant_type_to_string(QuantMode)
<< " PreshuffleQuant = " << (GemmConfig::PreshuffleQuant ? "true" : "false") << " : "
<< " PreshuffleB = " << (GemmConfig::PreshuffleB ? "true" : "false") << " : "
<< ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
<< " PreshuffleB = \n"
<< (GemmConfig::PreshuffleB ? "true" : "false") << " : " << ave_time << " ms, "
<< tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
return ave_time;
}