ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK][EXAMPLES] (#2826)

Browse Source 
-Added parameter to enable/disable verification and timing of kernel in various examples that missed it.
-Added parameter to change number of groups to execute in grouped_gemm_examples.

Signed-off-by: Michal Kulikowski <Michal.Kulikowski@amd.com>
This commit is contained in:
Michał Kulikowski
2025-09-11 21:33:00 +02:00
committed by GitHub
parent f3239395dc
commit ffe9775e70
33 changed files with 578 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
View File

@@ -236,7 +236,7 @@ void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M,
<< " GB/s, " << std::endl;
}
int main()
int main(int argc, char* argv[])
{
// GEMM shape
ck::index_t M = 1024;
@@ -249,6 +249,25 @@ int main()
ck::index_t StrideD1 = 1024;
ck::index_t StrideE = 1024;
bool do_verification = true;
bool time_kernel = false;
if(argc == 1)
{
// use default
}
else if(argc == 3)
{
do_verification = std::stoi(argv[1]);
time_kernel = static_cast<bool>(std::stoi(argv[2]));
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: time kernel (0=no, 1=yes)\n");
exit(0);
}
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<D0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
@@ -357,6 +376,7 @@ int main()
normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, false});
bool pass = true;
if(do_verification)
{
// verification
Tensor<LayerNormOutDataType> host_layerNorm_m_n(
@@ -383,27 +403,25 @@ int main()
1e-2);
}
if(time_kernel)
{
// evaluate kernel perf
bool time_kernel = true;
float gemm_reduce_mean_reduce_square_mean_ave_time =
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, true});
float normalize_ave_time =
normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, true});
if(time_kernel)
DumpGemmLayerNormPerf<ADataType,
BDataType,
EDataType,
D0DataType,
D1DataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
DumpGemmLayerNormPerf<ADataType,
BDataType,
EDataType,
D0DataType,
D1DataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
}
return pass ? 0 : 1;

48
example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
View File

@@ -221,7 +221,7 @@ void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M,
<< " GB/s, " << std::endl;
}
int main()
int main(int argc, char* argv[])
{
// GEMM shape
ck::index_t M = 1024;
@@ -232,6 +232,25 @@ int main()
ck::index_t StrideB = 1024;
ck::index_t StrideE = 1024;
bool do_verification = true;
bool time_kernel = false;
if(argc == 1)
{
// use default
}
else if(argc == 3)
{
do_verification = std::stoi(argv[1]);
time_kernel = static_cast<bool>(std::stoi(argv[2]));
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: time kernel (0=no, 1=yes)\n");
exit(0);
}
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
@@ -333,6 +352,7 @@ int main()
normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, false});
bool pass = true;
if(do_verification)
{
// verification
Tensor<LayerNormOutDataType> host_layerNorm_m_n(
@@ -354,25 +374,23 @@ int main()
layerNorm_m_n, host_layerNorm_m_n, "Error: Incorrect results d1", 1e-3, 1e-3);
}
if(time_kernel)
{
// evaluate kernel perf
bool time_kernel = true;
float gemm_reduce_mean_reduce_square_mean_ave_time =
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, true});
float normalize_ave_time =
normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, true});
if(time_kernel)
DumpGemmLayerNormPerf<ADataType,
BDataType,
EDataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
DumpGemmLayerNormPerf<ADataType,
BDataType,
EDataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
}
return pass ? 0 : 1;