ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-30 19:57:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'develop' into ck_tile/gemm_blockscale_abquant

Browse Source
This commit is contained in:
kensclin
2025-12-12 01:06:46 +08:00
committed by GitHub
parent c2bb4d261f 4dcc3e59c1
commit 2debb6ca08
179 changed files with 9598 additions and 2091 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
example/65_gemm_multiply_multiply/CMakeLists.txt
View File

@@ -77,3 +77,5 @@ example_compile_options(example_moe_gemm1_xdl_fp8_blockscale PRIVATE ${BLOCKSCAL
add_example_executable(example_gemm_add_add_wmma_fp16 gemm_add_add_wmma_fp16.cpp)
add_example_executable(example_gemm_multiply_multiply_wmma_fp16_bpreshuffle gemm_multiply_multiply_wmma_fp16_bpreshuffle.cpp)
add_example_executable(example_gemm_multiply_multiply_wmma_fp8_bpreshuffle gemm_multiply_multiply_wmma_fp8_bpreshuffle.cpp)
add_example_executable(example_gemm_multiply_multiply_wmma_fp8_ab_scale gemm_multiply_multiply_wmma_fp8_ab_scale.cpp)
add_example_executable(example_gemm_multiply_multiply_wmma_fp8_blockscale_bpreshuffle gemm_multiply_multiply_wmma_fp8_blockscale_bpreshuffle.cpp)

345
example/65_gemm_multiply_multiply/gemm_multiply_multiply_wmma_fp8_ab_scale.cpp Normal file
View File

@@ -0,0 +1,345 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3_ab_scale.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using FP8 = ck::f8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = FP8;
using A1DataType = F32;
using B0DataType = FP8;
using B1DataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using B0Layout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t Scale_Block_M = 1;
static constexpr ck::index_t Scale_Block_N = 128;
static constexpr ck::index_t Scale_Block_K = 128;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_Wmma_CShuffle_V3
// clang-format off
<Row, Col, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType,
AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
128, 128, 128,
16, 16,
16, 16,
4, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 1, S<1, 32, 1, 8>, S<8>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool flush_cache = true;
// GEMM shape
ck::index_t M = 128;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
ck::index_t KBatch = 1;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 8 || argc == 9)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
flush_cache = std::stoi(argv[7]);
if(argc == 9)
{
KBatch = std::stoi(argv[8]);
}
StrideA = K;
StrideB = K;
StrideE = N;
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: M, N, K\n");
printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n");
printf("arg8: KBatch (default: 1)\n");
exit(0);
}
ck::index_t Scale_Stride_AM = (K + Scale_Block_K - 1) / Scale_Block_K;
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return ck::HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return ck::HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
ck::Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
ck::Tensor<A1DataType> a1_m_k(f_host_tensor_descriptor((M + Scale_Block_M - 1) / Scale_Block_M,
(K + Scale_Block_K - 1) / Scale_Block_K,
Scale_Stride_AM,
A0Layout{}));
ck::Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
ck::Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K,
(N + Scale_Block_N - 1) / Scale_Block_N,
Scale_Stride_BN,
B0Layout{}));
ck::Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
ck::Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-1, 1});
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-1, 1});
break;
case 2:
a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 3:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 4:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
}
ck::DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
ck::DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize());
ck::DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
ck::DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
ck::DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
a1_device_buf.ToDevice(a1_m_k.mData.data());
b0_device_buf.ToDevice(b0_k_n.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumDTensor = DsDataType::Size();
// do GEMM
auto device_op = DeviceOpInstance{};
std::string op_name = device_op.GetTypeString();
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(static_cast<A0DataType*>(a0_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
std::array<const void*, NumDTensor>{},
static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, NumDTensor>{},
StrideE,
static_cast<const A1DataType*>(a1_device_buf.GetDeviceBuffer()),
static_cast<const B1DataType*>(b1_device_buf.GetDeviceBuffer()),
a_element_op,
b_element_op,
cde_element_op,
KBatch);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float ave_time = .0;
ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 0, 50, 100});
int pass = 0;
if(do_verification)
{
ck::Tensor<AccDataType> c_m_n({M, N});
ck::Tensor<float> a_m_k({M, K});
ck::Tensor<float> b_k_n({K, N});
for(int m = 0; m < M; m++)
{
for(int k = 0; k < K; k++)
{
a_m_k(m, k) = ck::type_convert<float>(a0_m_k(m, k)) *
a1_m_k(m / Scale_Block_M, k / Scale_Block_K);
}
}
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
b_k_n(k, n) = ck::type_convert<float>(b0_k_n(k, n)) *
b1_k_n(k / Scale_Block_K, n / Scale_Block_N);
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<float,
float,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
e_m_n_host_result(m, n) = ck::type_convert<EDataType>(c_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, "Error: Incorrect results!", 5e-2, 5e-2)
? 0
: 1;
}
if(flush_cache)
{
int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype;
ave_time = invoker.Run(argument,
StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf});
}
else
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100});
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< op_name << ", KBatch " << KBatch << std::endl;
return pass;
}

357
example/65_gemm_multiply_multiply/gemm_multiply_multiply_wmma_fp8_blockscale_bpreshuffle.cpp Normal file
View File

@@ -0,0 +1,357 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3_blockscale_bpreshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
#include "common.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using FP8 = ck::f8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = FP8;
using A1DataType = F32;
using B0DataType = FP8;
using B1DataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using A1Layout = Col;
using B0Layout = Col;
using D0Layout = Row;
using D1Layout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
static constexpr int KPack = 16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t Scale_Block_M = 1;
static constexpr ck::index_t Scale_Block_N = 128;
static constexpr ck::index_t Scale_Block_K = 128;
using DeviceOpInstance =
ck::tensor_operation::device::DeviceGemmMultiD_BlockScale_Wmma_CShuffle_V3_BPreshuffle
// clang-format off
<Row, Col, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
128, 128, 128,
16, 16,
16, 16,
4, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 16, 16, 0,
1, 1,
S<1, 32, 1, 8>, S<8>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool flush_cache = true;
// GEMM shape
ck::index_t M = 128;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
ck::index_t KBatch = 1;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 8 || argc == 9)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
flush_cache = std::stoi(argv[7]);
if(argc == 9)
{
KBatch = std::stoi(argv[8]);
}
StrideA = K;
StrideB = K;
StrideE = N;
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: M, N, K\n");
printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n");
printf("arg8: KBatch (default: 1)\n");
exit(0);
}
// Transpose the AScale tensor for better performance
ck::index_t Scale_Stride_AK = (M + Scale_Block_M - 1) / Scale_Block_M;
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return ck::HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return ck::HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
ck::Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
ck::Tensor<A1DataType> a1_m_k(f_host_tensor_descriptor((M + Scale_Block_M - 1) / Scale_Block_M,
(K + Scale_Block_K - 1) / Scale_Block_K,
Scale_Stride_AK,
A1Layout{}));
ck::Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
ck::Tensor<B0DataType> b0_preshuffled(
f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); // use laout only for size
ck::Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K,
(N + Scale_Block_N - 1) / Scale_Block_N,
Scale_Stride_BN,
B0Layout{}));
ck::Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
ck::Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
case 2:
a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 3:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 4:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
}
ck::DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
ck::DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize());
ck::DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
ck::DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
ck::DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
a1_device_buf.ToDevice(a1_m_k.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumDTensor = DsDataType::Size();
// do GEMM
auto device_op = DeviceOpInstance{};
std::string op_name = device_op.GetTypeString();
int NPerWmma = device_op.GetPreShuffleParameters();
preShuffleBuffer<KPack>(b0_k_n.mData.data(), b0_preshuffled.mData.data(), N, K, NPerWmma);
b0_device_buf.ToDevice(b0_preshuffled.mData.data());
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
b0_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, NumDTensor>{},
StrideE,
a1_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer(),
a_element_op,
b_element_op,
cde_element_op,
KBatch);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float ave_time = 0.0f;
if(flush_cache)
{
int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype;
ave_time = invoker.Run(argument,
StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf});
}
else
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100});
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< op_name << ", KBatch " << KBatch << std::endl;
if(do_verification)
{
ck::Tensor<AccDataType> c_m_n({M, N});
ck::Tensor<float> a_m_k({M, K});
ck::Tensor<float> b_k_n({K, N});
for(int m = 0; m < M; m++)
{
for(int k = 0; k < K; k++)
{
a_m_k(m, k) = ck::type_convert<float>(a0_m_k(m, k)) *
a1_m_k(m / Scale_Block_M, k / Scale_Block_K);
}
}
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
b_k_n(k, n) = ck::type_convert<float>(b0_k_n(k, n)) *
b1_k_n(k / Scale_Block_K, n / Scale_Block_N);
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<float,
float,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
e_m_n_host_result(m, n) = ck::type_convert<EDataType>(c_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, "Error: Incorrect results!", 5e-2, 5e-2)
? 0
: 1;
}
return 0;
}

6
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View File

@@ -770,7 +770,7 @@ def create_kernel(
class CompatibilityRuleFactory:
@staticmethod
def get_rules() -> list[CompatibilityRule]:
def get_rules() -> List[CompatibilityRule]:
# in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
def check_mode(problem_ctx: ProblemContext, kernel_ctx: KernelContext) -> bool:
if problem_ctx.mode == "group":
@@ -812,7 +812,7 @@ class CompatibilityRuleFactoryGfx9(CompatibilityRuleFactory):
_AVAILABLE_PIPELINES = frozenset({"qr", "qr_async", "qs"})
@classmethod
def get_rules(cls) -> list[CompatibilityRule]:
def get_rules(cls) -> List[CompatibilityRule]:
rules = CompatibilityRuleFactory.get_rules()
def check_hdim_tile(
@@ -846,7 +846,7 @@ class CompatibilityRuleFactoryGfx950(CompatibilityRuleFactoryGfx9):
)
@classmethod
def get_rules(cls) -> list[CompatibilityRule]:
def get_rules(cls) -> List[CompatibilityRule]:
rules = CompatibilityRuleFactoryGfx9.get_rules()
def check_tile_pipeline(

1
example/ck_tile/38_block_scale_gemm/CMakeLists.txt
View File

@@ -17,6 +17,7 @@ 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_bf8.cpp
gemm_bquant_quantgrouped_fp8.cpp
gemm_bquant_quantgrouped_preshuffleb.cpp

4
example/ck_tile/38_block_scale_gemm/README.md
View File

@@ -23,7 +23,7 @@ This folder contains examples of quant GEMMs using the ck_tile tile-programming
- **Preshuffled GEMM**: Shuffle the GEMM of B (weight) matrix in the warp layout and bypass the shared memory to do the GEMM calculation. Best performance solution for GEMM.
- **TransposeC**: Transpose the C Matrix Output layout to have the best coalesced scale reading
- **Preshuffled Quant**: Preshuffle the input matrix to load multiple Quant warp blocks along the selected dimension.
- **Precision**: Supports fp16, bf16, fp8, bf8, int4 (for B Matrix).
- **Precision**: Supports fp16, bf16, fp8, bf8, int4 (for B Matrix), uint8 (split into two fp4 in the pipeline (for B Matrix)).
- **Validation**: CPU/GPU validation and error tolerance options.
## build
@@ -53,7 +53,7 @@ args:
-stride_b Tensor B stride (default:0)
-stride_c Tensor C stride (default:0)
-v 0: No validation, 1: Validation on CPU, 2: Validation on GPU (default:1)
-prec Data type. For AQuant: fp8, bf8, i4fp8, or i4bf8; for Bquant: fp8, bf8, fp8i4, or bf8i4 (default for both AQuant and Bquant: fp8)
-prec Data type. For AQuant: fp8, bf8, i4fp8, or i4bf8; for Bquant: fp8, bf8, fp8i4, bf8i4, or bf16fp4 (default for both AQuant and Bquant: fp8)
-warmup Number of iterations before benchmarking the kernel (default:50)
-repeat Number of iterations to benchmark the kernel (default:1000)
-timer gpu:gpu timer, cpu:cpu timer (default:gpu)

41
example/ck_tile/38_block_scale_gemm/gemm_bquant_quantgrouped_bf16mxfp4.cpp Normal file
View File

@@ -0,0 +1,41 @@
// SPDX-License-Identifier: MIT
// Copyright (c) , Advanced Micro Devices, Inc. All rights reserved.
#include "run_gemm_quant_example.inc"
template <typename T>
using GemmConfig = GemmConfigQuantPrefill<T>;
#define RUN_GEMM_EXAMPLE_PREC_TYPE \
run_gemm_example_prec_type<GemmConfig<ck_tile::pk_fp4_raw_t>, \
TypeConfig, \
QuantGroupSize, \
ck_tile::QuantType::BQuantGrouped>(arg_parser);
void bquant_quantgrouped_bf16fp4_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut)
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::bf16_t,
ck_tile::pk_fp4_raw_t,
ck_tile::bf16_t,
ck_tile::pk_fp4_raw_t>{});
lut[hash_multiple_strings(
{"bf16fp4", "bquant", "non-preshuffleb", "non-preshufflequant", "1x1x32"})] =
[](const ck_tile::ArgParser& arg_parser) {
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 32>>;
return RUN_GEMM_EXAMPLE_PREC_TYPE;
};
lut[hash_multiple_strings(
{"bf16fp4", "bquant", "non-preshuffleb", "non-preshufflequant", "1x1x64"})] =
[](const ck_tile::ArgParser& arg_parser) {
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 64>>;
return RUN_GEMM_EXAMPLE_PREC_TYPE;
};
lut[hash_multiple_strings(
{"bf16fp4", "bquant", "non-preshuffleb", "non-preshufflequant", "1x1x128"})] =
[](const ck_tile::ArgParser& arg_parser) {
using QuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
return RUN_GEMM_EXAMPLE_PREC_TYPE;
};
}

5
example/ck_tile/38_block_scale_gemm/gemm_quant.cpp
View File

@@ -32,7 +32,7 @@ auto create_args(int argc, char* argv[])
.insert("prec",
"fp8",
"Data type. For AQuant: fp8, bf8, i4fp8, or i4bf8; for Bquant: fp8, bf8, fp8i4, "
"or bf8i4; for ABQuant: fp8, bf8")
"bf8i4 or bf16fp4; for ABQuant: fp8, bf8")
.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")
@@ -109,6 +109,8 @@ void bquant_quantgrouped_fp8i4_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut);
void bquant_quantgrouped_bf8i4_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut);
void bquant_quantgrouped_bf16fp4_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut);
void bquant_quantgrouped_preshuffleb_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut);
void bquant_quantgrouped_preshufflequant_instance_factory(
@@ -141,6 +143,7 @@ int main(int argc, char* argv[])
bquant_quantgrouped_bf8_instance_factory(lut);
bquant_quantgrouped_fp8i4_instance_factory(lut);
bquant_quantgrouped_bf8i4_instance_factory(lut);
bquant_quantgrouped_bf16fp4_instance_factory(lut);
bquant_quantgrouped_preshuffleb_instance_factory(lut);
bquant_quantgrouped_preshufflequant_instance_factory(lut);
bquant_quantgrouped_preshuffleb_preshufflequant_instance_factory(lut);

6
example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
View File

@@ -69,8 +69,10 @@ auto calculate_rtol_atol(const ck_tile::index_t K,
const ck_tile::index_t kbatch,
const float max_accumulated_value)
{
using ComputeType =
std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>;
using ComputeType = std::conditional_t<
std::is_same_v<BDataType, ck_tile::pk_fp4_raw_t>,
ADataType,
std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>>;
// Calculate thresholds
const auto rtol = ck_tile::get_relative_threshold<ComputeType, CDataType, AccDataType>(
ck_tile::integer_divide_ceil(K, kbatch));

145
example/ck_tile/38_block_scale_gemm/run_gemm_quant_example.inc
View File

@@ -145,21 +145,25 @@ float gemm_calc_quant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::str
has_hot_loop_v,
tail_number_v>>>>;
using GemmPipeline = std::conditional_t<
QuantMode == ck_tile::QuantType::RowColQuant ||
QuantMode == ck_tile::QuantType::TensorQuant,
ck_tile::GemmPipelineAgBgCrCompV3<PipelineProblem>,
std::conditional_t<
QuantMode == ck_tile::QuantType::AQuantGrouped,
std::conditional_t<GemmConfig::PreshuffleQuant == true,
ck_tile::AQuantGemmPipelineAgBgCrCompV3<PipelineProblem>,
ck_tile::AQuantGemmPipelineAgBgCrMem<PipelineProblem>>,
std::conditional_t<
QuantMode == ck_tile::QuantType::BQuantGrouped,
std::conditional_t<GemmConfig::PreshuffleB == true,
ck_tile::WPQuantBPipelineAgBgCrV2<PipelineProblem>,
ck_tile::BQuantGemmPipelineAgBgCrCompV3<PipelineProblem>>,
ck_tile::ABQuantGemmPipelineAgBgCrCompV3<PipelineProblem>>>>;
using GemmPipeline = std::conditional_t < QuantMode == ck_tile::QuantType::RowColQuant ||
QuantMode == ck_tile::QuantType::TensorQuant,
ck_tile::GemmPipelineAgBgCrCompV3<PipelineProblem>,
std::conditional_t<
QuantMode == ck_tile::QuantType::AQuantGrouped,
std::conditional_t<GemmConfig::PreshuffleQuant == true,
ck_tile::AQuantGemmPipelineAgBgCrCompV3<PipelineProblem>,
ck_tile::AQuantGemmPipelineAgBgCrMem<PipelineProblem>>,
std::conditional_t<
GemmConfig::PreshuffleB == true,
ck_tile::WPQuantBPipelineAgBgCrV2<PipelineProblem>,
std::conditional_t<
std::is_same_v<typename TypeConfig::BDataType, ck_tile::pk_fp4_raw_t>,
ck_tile::MxFp4GemmPipelineAgBgCrCompV3<PipelineProblem>,
std::conditional_t<
QuantMode == ck_tile::QuantType::ABQuantGrouped,
ck_tile::ABQuantGemmPipelineAgBgCrCompV3<
PipelineProblem,
ck_tile::BQuantGemmPipelineAgBgCrCompV3<PipelineProblem>>>>>>;
constexpr bool TiledPermuteN =
(BQuantGroupSize::kN > 1) ? false : GemmConfig::TiledMMAPermuteN;
@@ -168,28 +172,31 @@ float gemm_calc_quant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::str
printf(
"TiledPermuteN: %d (QuantGroupSize::kN=%d)\n", TiledPermuteN, BQuantGroupSize::kN);
}
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<typename TypeConfig::ADataType,
typename TypeConfig::BDataType,
ck_tile::tuple<>,
typename TypeConfig::AccDataType,
typename TypeConfig::CDataType,
ck_tile::tuple<>,
CLayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
GemmConfig::N_Warp,
GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile,
transpose_c,
ck_tile::memory_operation_enum::set,
1,
false,
1,
TiledPermuteN>>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<
typename TypeConfig::ADataType,
std::conditional_t<
std::is_same_v<typename TypeConfig::BDataType, ck_tile::pk_fp4_raw_t>,
typename TypeConfig::ADataType,
typename TypeConfig::BDataType>,
ck_tile::tuple<>,
typename TypeConfig::AccDataType,
typename TypeConfig::CDataType,
ck_tile::tuple<>,
CLayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
GemmConfig::N_Warp,
GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile,
transpose_c,
ck_tile::memory_operation_enum::set,
1,
false,
1,
TiledPermuteN>>;
using Kernel =
ck_tile::QuantGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue, QuantMode>;
@@ -226,7 +233,11 @@ float gemm_calc_quant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::str
ck_tile::HostTensor<typename TypeConfig::ADataType> a_m(ck_tile::host_tensor_descriptor(
args.M, args.K, args.stride_A, is_row_major(ALayout{})));
ck_tile::HostTensor<typename TypeConfig::BDataType> b_n(ck_tile::host_tensor_descriptor(
args.K, args.N, args.stride_B, is_row_major(BLayout{})));
std::is_same_v<typename TypeConfig::BDataType, ck_tile::pk_fp4_raw_t> ? args.K / 2
: args.K,
args.N,
args.stride_B,
is_row_major(BLayout{})));
auto size_a_buffer = a_m.get_element_space_size_in_bytes();
auto size_b_buffer = b_n.get_element_space_size_in_bytes();
@@ -484,7 +495,11 @@ int run_gemm_example_with_layouts(const ck_tile::ArgParser& arg_parser,
int rotating_count = arg_parser.get_int("rotating_count");
stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
stride_B = ck_tile::get_default_stride(
(std::is_same_v<BDataType, ck_tile::pk_fp4_raw_t>) ? (K / 2) : K,
N,
stride_B,
is_row_major(b_layout));
stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
// Conditional stride calculation based on QuantMode
@@ -516,8 +531,11 @@ int run_gemm_example_with_layouts(const ck_tile::ArgParser& arg_parser,
ck_tile::HostTensor<ADataType> a_m_k(
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
ck_tile::HostTensor<BDataType> b_k_n(
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
ck_tile::HostTensor<BDataType> b_k_n(ck_tile::host_tensor_descriptor(
(std::is_same_v<BDataType, ck_tile::pk_fp4_raw_t>) ? (K / 2) : K,
N,
stride_B,
is_row_major(b_layout)));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
@@ -563,13 +581,22 @@ int run_gemm_example_with_layouts(const ck_tile::ArgParser& arg_parser,
{
ck_tile::FillUniformDistribution<ck_tile::pk_int4_t>{-5.0f, 5.0f, fill_seed(gen)}(
b_k_n);
ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
*bq_tensor_ptr);
}
else if constexpr(std::is_same_v<BDataType, ck_tile::pk_fp4_raw_t>)
{
ck_tile::FillUniformDistribution<BDataType>{-5.0f, 5.0f, fill_seed(gen)}(b_k_n);
ck_tile::FillUniformDistribution<BQDataType>{125.f, 130.f, fill_seed(gen)}(
*bq_tensor_ptr);
}
else
{
ck_tile::FillUniformDistribution<BDataType>{-2.0f, 3.0f, fill_seed(gen)}(b_k_n);
ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
*bq_tensor_ptr);
}
ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
*bq_tensor_ptr);
ck_tile::FillUniformDistribution<ADataType>{-5.0f, 5.0f, fill_seed(gen)}(a_m_k);
}
else if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped)
@@ -817,13 +844,23 @@ int run_gemm_example_with_layouts(const ck_tile::ArgParser& arg_parser,
}
else if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
{
ck_tile::reference_gemm_quant<ADataType,
AQDataType,
BDataType,
AccDataType,
CDataType,
BQuantGroupSize,
false>(a_m_k, *bq_tensor_ptr, b_k_n, c_m_n_host_ref);
if constexpr(std::is_same_v<BDataType, ck_tile::pk_fp4_raw_t>)
ck_tile::reference_mxfp4gemm_quant<ADataType,
BQDataType,
BDataType,
AccDataType,
CDataType,
BQuantGroupSize,
false>(
a_m_k, *bq_tensor_ptr, b_k_n, c_m_n_host_ref);
else
ck_tile::reference_gemm_quant<ADataType,
AQDataType,
BDataType,
AccDataType,
CDataType,
BQuantGroupSize,
false>(a_m_k, *bq_tensor_ptr, b_k_n, c_m_n_host_ref);
}
else if constexpr(QuantMode == ck_tile::QuantType::ABQuantGrouped)
{
@@ -896,16 +933,18 @@ int run_gemm_example_prec_type(const ck_tile::ArgParser& arg_parser)
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
if((QuantMode == ck_tile::QuantType::AQuantGrouped ||
QuantMode == ck_tile::QuantType::RowColQuant) &&
QuantMode == ck_tile::QuantType::RowColQuant ||
std::is_same_v<typename TypeConfig::BDataType, ck_tile::pk_fp4_raw_t>) &&
GemmConfig::PreshuffleB)
{
throw std::runtime_error(
"Preshuffling weight matrix is not supported for AQuant or RowColQuant");
"Preshuffling weight matrix is not supported for AQuant, RowColQuant or bf16_fp4_gemm");
}
if constexpr(std::is_same_v<typename TypeConfig::ADataType, ck_tile::pk_int4_t> ||
std::is_same_v<typename TypeConfig::ADataType, ck_tile::fp8_t> ||
std::is_same_v<typename TypeConfig::ADataType, ck_tile::bf8_t>)
std::is_same_v<typename TypeConfig::ADataType, ck_tile::bf8_t> ||
std::is_same_v<typename TypeConfig::ADataType, ck_tile::bf16_t>)
{
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");

2
example/test_old_ck_gpu_reference.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
// Standalone test program for Old CK GPU references
// Tests naive_conv_fwd (existing) and future backward ops