Skip to content

GitLab

Unverified Commit ec959387 authored by rocking's avatar rocking Committed by GitHub
Browse files

Merge branch 'develop' into ck_tile/fmha_receipt_aiter

parents c1e2fef7 0e5e29c4
Hide whitespace changes
Inline Side-by-side
Showing with 1242 additions and 41 deletions
example/24_batched_gemm/CMakeLists.txt
View file @ ec959387
...@@ -22,3 +22,6 @@ if(USE_BITINT_EXTENSION_INT4) ...@@ -22,3 +22,6 @@ if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp) add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
endif() endif()
add_example_executable(example_batched_gemm_xdl_fp16int4_b_scale_v3 batched_gemm_xdl_fp16int4_b_scale_v3.cpp)
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16int4_b_scale_v3)
example/24_batched_gemm/batched_gemm_xdl_fp16int4_b_scale_v3.cpp 0 → 100644
View file @ ec959387
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl_fpAintB_b_scale.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/check_err.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"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = ck::pk_i4_t;
using BScaleDataType = ck::half_t;
using AccDataType = F32;
using CShuffleDataType = F16;
using CDataType = F16;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto PermuteA = false;
static constexpr bool PermuteB = false;
static constexpr ck::index_t Scale_Block_N = 1;
static constexpr ck::index_t Scale_Block_K = 128;
static constexpr ck::index_t KPerBlock = 256;
// clang-format off
using DeviceBatchedGemmV2Instance =
ck::tensor_operation::device::DeviceBatchedGemm_Xdl_CShuffleV3_BScale<
ALayout, BLayout, CLayout,
ADataType, BDataType, BScaleDataType, CDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CElementOp, GemmDefault,
256, Scale_Block_N, Scale_Block_K,
16, 64,
KPerBlock, 8, 32,
16, 16,
1, 1,
S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 32, 32, 0,
1, 1, S<1, 16, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, CDataType, CDataType, PermuteA, PermuteB>;
// clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
AccDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_batched_gemm_example_fp16int4_b_scale.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_fp16_int4_b_scale_example(argc, argv); }
example/24_batched_gemm/run_batched_gemm_example_fp16int4_b_scale.inc 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <random>
#pragma once
struct ProblemSize final
{
ck::index_t M = 128;
ck::index_t N = 128;
ck::index_t K = 384;
ck::index_t stride_A = K;
ck::index_t stride_B = K;
ck::index_t stride_C = N;
ck::index_t batch_stride_A = M * K;
ck::index_t batch_stride_B = K * N;
ck::index_t batch_stride_C = M * N;
// Batched Gemm count
ck::index_t batch_count = 2;
// Split K count
ck::index_t KBatch = 1;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
};
template <typename DataType>
inline __host__ __device__ constexpr double get_rtol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 1e-1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 1.5e-1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
template <typename DataType>
inline __host__ __device__ constexpr double get_atol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 16.1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 8192.1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
using namespace ck::literals;
auto& [M,
N,
K,
stride_A,
stride_B,
stride_C,
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_count,
KBatch] = problem_size;
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, 1_uz, stride});
}
};
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
ck::index_t batch_BScale_Stride =
((K + Scale_Block_K - 1) / Scale_Block_K) * ((N + Scale_Block_N - 1) / Scale_Block_N);
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
Tensor<BDataType> b_g_k_n_permute(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
Tensor<BScaleDataType> b1_g_k_n(
f_host_tensor_descriptor(batch_count,
(K + Scale_Block_K - 1) / Scale_Block_K,
(N + Scale_Block_N - 1) / Scale_Block_N,
Scale_Stride_BN,
batch_BScale_Stride,
BLayout{}));
switch(config.init_method)
{
case 0:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_g_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
break;
case 3:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_g_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
break;
case 4:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_g_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
break;
case 5:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.5, 0.5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
b1_g_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
}
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, CLayout{}));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, CLayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "b1_g_k_n: " << b1_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_g_k_n_device_buf(sizeof(BDataType) * b_g_k_n_permute.mDesc.GetElementSpaceSize());
DeviceMem b1_g_scale_device_buf(sizeof(BScaleDataType) * b1_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_g_m_n_device_buf(sizeof(CDataType) *
c_g_m_n_device_result.mDesc.GetElementSpaceSize());
printf("a_g_m_k size: %zu, b_g_k_n size: %zu, b1_g_k_n size: %zu, c_g_m_n size: %zu\n",
a_g_m_k.mDesc.GetElementSpaceSize(),
b_g_k_n_permute.mDesc.GetElementSpaceSize(),
b1_g_k_n.mDesc.GetElementSpaceSize(),
c_g_m_n_device_result.mDesc.GetElementSpaceSize());
// weight permute
if constexpr(PermuteB)
{
printf("Permute B\n");
int K1 = KPerBlock;
int K0 = K / KPerBlock;
// int K0, N, K1
for(int bs = 0; bs < batch_count; bs++)
{
for(int j = 0; j < K0; j++)
{
for(int i = 0; i < N; i++)
{
for(int jj = 0; jj < K1; jj++)
{
b_g_k_n_permute(bs * batch_stride_B + j * N * K1 + i * K1 + jj) =
b_g_k_n(bs * batch_stride_B + i * K + (j * K1 + jj));
}
}
}
}
}
else
{
b_g_k_n_permute = b_g_k_n;
}
// vector pk_i4x4 permute
for(int bs = 0; bs < batch_count; bs++)
{
for(int i = 0; i < N; i++)
{
for(int j = 0; j < K; j += 8)
{
int input[8];
for(int k = 0; k < 4; k++)
{
int i4x2 = b_g_k_n_permute(bs, j + k * 2, i).data;
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
}
// permute 01234567->20643175
{
int hi = input[2];
int lo = input[0];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 0, i) = i4x2;
}
{
int hi = input[6];
int lo = input[4];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 2, i) = i4x2;
}
{
int hi = input[3];
int lo = input[1];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 4, i) = i4x2;
}
{
int hi = input[7];
int lo = input[5];
int i4x2 = (hi << 4) | lo;
b_g_k_n_permute(bs, j + 6, i) = i4x2;
}
}
}
}
a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
b_g_k_n_device_buf.ToDevice(b_g_k_n_permute.mData.data());
b1_g_scale_device_buf.ToDevice(b1_g_k_n.mData.data());
DeviceMem workspace;
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceBatchedGemmV2Instance{};
auto invoker = gemm.MakeInvoker();
float ave_time = 0;
auto argument =
gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
stride_A,
stride_B,
stride_C,
Scale_Stride_BN,
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_BScale_Stride,
static_cast<BScaleDataType*>(b1_g_scale_device_buf.GetDeviceBuffer()),
batch_count, // batch count
KBatch, // split K count
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
return true;
}
bool pass = true;
Tensor<float> b_g_k_n_dequant({batch_count, K, N});
if(config.do_verification)
{
float v_b = 0;
for(int bs = 0; bs < batch_count; bs++)
{
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
ck::pk_i4_t i4x2 = b_g_k_n(bs, k, n).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2.data >> 0) & 0xf;
else
i4 = (i4x2.data >> 4) & 0xf;
i4 = i4 - 8;
v_b = ck::type_convert<float>(i4);
b_g_k_n_dequant(bs, k, n) =
ck::type_convert<float>(v_b) *
ck::type_convert<float>(b1_g_k_n(bs, k / Scale_Block_K, n / Scale_Block_N));
}
}
}
auto ref_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_g_m_k,
b_g_k_n_dequant,
c_g_m_n_host_result,
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 0});
hip_check_error(hipDeviceSynchronize());
c_g_m_n_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
pass &= ck::utils::check_err(c_g_m_n_device_result,
c_g_m_n_host_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
}
if(config.time_kernel)
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = 2_uz * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K +
sizeof(BDataType) * K * N /
(ck::is_same_v<ck::remove_cvref_t<BDataType>, ck::pk_i4_t> ? 2 : 1) +
sizeof(CDataType) * M * N;
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, " << gemm.GetTypeString() << std::endl;
}
#if 0
// print A matrix
printf("A matrix:\n");
for(int bs = 0; bs < batch_count; bs++)
{
printf("batch %d -> Address: %p\n", bs, static_cast<void*>(&a_g_m_k(bs, 0, 0)));
for(int i = 0; i < M; i++)
{
for(int j = 0; j < K; j++)
{
printf("%.2f,", static_cast<float>(a_g_m_k(bs, i, j)));
}
printf("\n");
}
}
// print B matrix original
printf("B matrix original:\n");
for(int bs = 0; bs < batch_count; bs++)
{
printf("batch %d -> Address: %p\n", bs, static_cast<void*>(&b_g_k_n(bs, 0, 0)));
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
ck::pk_i4_t i4x2 = b_g_k_n(bs, k, n).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2.data >> 0) & 0xf;
else
i4 = (i4x2.data >> 4) & 0xf;
i4 = i4 - 8;
printf("%d,", static_cast<int>(i4));
}
printf("\n");
}
}
// print B matrix
printf("B matrix:\n");
for(int bs = 0; bs < batch_count; bs++)
{
printf("batch %d -> Address: %p\n", bs, static_cast<void*>(&b_g_k_n_dequant(bs, 0, 0)));
for(int i = 0; i < K; i++)
{
for(int j = 0; j < N; j++)
{
printf("%.2f, ", static_cast<float>(b_g_k_n_dequant(bs, i, j)));
}
printf("\n");
}
}
// print B scale matrix
printf("B Scale matrix:\n");
for(int bs = 0; bs < batch_count; bs++)
{
printf("batch %d -> Address: %p\n", bs, static_cast<void*>(&b1_g_k_n(bs, 0, 0)));
for(int i = 0; i < (K + Scale_Block_K - 1) / Scale_Block_K; i++)
{
for(int j = 0; j < (N + Scale_Block_N - 1) / Scale_Block_N; j++)
{
printf("%.2f, ", static_cast<float>(b1_g_k_n(bs, i, j)));
}
printf("\n");
}
}
// print C matrix
printf("C matrix:\n");
for(int bs = 0; bs < batch_count; bs++)
{
printf(
"batch %d -> Address: %p\n", bs, static_cast<void*>(&c_g_m_n_device_result(bs, 0, 0)));
for(int i = 0; i < M; i++)
{
for(int j = 0; j < N; j++)
{
printf("%.2f, ", static_cast<float>(c_g_m_n_device_result(bs, i, j)));
}
printf("\n");
}
}
printf("C reference matrix:\n");
for(int bs = 0; bs < batch_count; bs++)
{
printf("batch %d -> Address: %p\n", bs, static_cast<void*>(&c_g_m_n_host_result(bs, 0, 0)));
for(int i = 0; i < M; i++)
{
for(int j = 0; j < N; j++)
{
printf("%.2f, ", static_cast<float>(c_g_m_n_host_result(bs, i, j)));
}
printf("\n");
}
}
#endif
return pass;
}
bool run_batched_gemm_fp16_int4_b_scale_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
std::mt19937 gen(11939);
std::uniform_int_distribution<int> dis(0, 15);
problem_size.M = 128 * (dis(gen) + 1);
problem_size.N = 128 * (dis(gen) + 1);
problem_size.K = 256 * (dis(gen) + 2);
problem_size.batch_count = 2;
if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else if(argc >= 7)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.M = std::stoi(argv[4]);
problem_size.N = std::stoi(argv[5]);
problem_size.K = std::stoi(argv[6]);
if(argc >= 8)
{
problem_size.batch_count = std::stoi(argv[7]);
}
if(argc >= 9)
{
problem_size.KBatch = std::stoi(argv[8]);
}
}
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=n0, 1=yes)\n");
exit(0);
}
problem_size.stride_A = problem_size.K;
problem_size.stride_B = problem_size.K;
problem_size.stride_C = problem_size.N;
problem_size.batch_stride_A = problem_size.M * problem_size.K;
problem_size.batch_stride_B = problem_size.K * problem_size.N;
problem_size.batch_stride_C = problem_size.M * problem_size.N;
return run_batched_gemm(problem_size, config);
}
example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
View file @ ec959387
...@@ -32,6 +32,56 @@ using BiasLayout = typename LayoutSettingSelector<NDimSpatial>::BiasLayout; ...@@ -32,6 +32,56 @@ using BiasLayout = typename LayoutSettingSelector<NDimSpatial>::BiasLayout;
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
using ResidualLayout = typename LayoutSettingSelector<NDimSpatial>::ResidualLayout; using ResidualLayout = typename LayoutSettingSelector<NDimSpatial>::ResidualLayout;
#if defined(CK_USE_AMD_MFMA_GFX950)
template <ck::index_t NDimSpatial>
using DeviceConvFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>,
WeightLayout<NDimSpatial>,
ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>,
OutputLayout<NDimSpatial>,
InKernelDataType,
WeiKernelDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
4, // ABlockTransferSrcScalarPerVector
4, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
4, // BBlockTransferSrcScalarPerVector
4, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 16, 1, 16>,
4>;
#else // defined(CK_USE_AMD_MFMA_GFX950)
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
using DeviceConvFwdInstance = using DeviceConvFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
...@@ -80,6 +130,7 @@ using DeviceConvFwdInstance = ...@@ -80,6 +130,7 @@ using DeviceConvFwdInstance =
1, 1,
S<1, 16, 1, 16>, S<1, 16, 1, 16>,
4>; 4>;
#endif // defined(CK_USE_AMD_MFMA_GFX950)
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial, using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
......
example/31_batched_gemm_gemm/CMakeLists.txt
View file @ ec959387
...@@ -5,6 +5,6 @@ if(USE_BITINT_EXTENSION_INT4) ...@@ -5,6 +5,6 @@ if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp) add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1") if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx95" AND NOT GPU_TARGETS MATCHES "gfx1")
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp) add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
endif() endif()
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View file @ ec959387
...@@ -5,6 +5,6 @@ if(USE_BITINT_EXTENSION_INT4) ...@@ -5,6 +5,6 @@ if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1") if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx95" AND NOT GPU_TARGETS MATCHES "gfx1")
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
endif() endif()
example/62_convnd_activ/binary/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/convinvscale/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/convscale/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/convscale_add/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/convscale_reduce/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/convscale_relu/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/dynamic_unary/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/multi_AB/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/unary/CMakeLists.txt
View file @ ec959387
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/67_gemm_microscaling/CMakeLists.txt 0 → 100644
View file @ ec959387
add_custom_target(example_gemm_mx)
add_example_executable(example_gemm_mx_fp8 gemm_mx_fp8.cpp)
add_example_dependencies(example_gemm_mx example_gemm_mx_fp8)
example/67_gemm_microscaling/README.md 0 → 100644
View file @ ec959387
# GEMM Examples for Microscaling Formats
## example_gemm_mx_fp8
```bash
# arg1: verification (0=no, 1=CPU)
# arg2: initialization (0=no init, 1=integer value, 2=decimal value)
# arg3: time kernel (0=no, 1=yes)
# arg4: verbosity (0=no info, 1=verbose info)
# arg5 to 10: M (16x), N(16x), K(16x), StrideA, StrideB, StrideC
./bin/example_gemm_mx_fp8 1 1 0 1
```
```bash
# Implies: ./bin/example_gemm_mx_fp8 1 2 0 0
./bin/example_gemm_mx_fp8
```
\ No newline at end of file
example/67_gemm_microscaling/gemm_mx_common.hpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/utility/sequence.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_mx_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/host_tensor.hpp"
using ScaleDataType = ck::e8m0_bexp_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
struct ExecutionConfig final
{
int do_verification = 1; // (0=no, 1=CPU)
int init_method = 2; // (0=no init, 1=integer value, 2=decimal value)
bool time_kernel = false; // (0=no, 1=yes)
int verbosity = 0; // (0=no info, 1=verbose info)
};
struct ProblemSize final
{
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = -1;
ck::index_t StrideB = -1;
ck::index_t StrideC = -1;
};
bool parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig& config)
{
if(argc == 1)
{
// use default case
}
else if(argc == 5)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
config.verbosity = std::stoi(argv[4]);
}
else if(argc == 11)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
config.verbosity = std::stoi(argv[4]);
problem_size.M = std::stoi(argv[5]);
problem_size.N = std::stoi(argv[6]);
problem_size.K = std::stoi(argv[7]);
problem_size.StrideA = std::stoi(argv[8]);
problem_size.StrideB = std::stoi(argv[9]);
problem_size.StrideC = std::stoi(argv[10]);
}
else
{
std::cerr << "arg1: verification (0=no, 1=CPU)" << std::endl
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
<< std::endl
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
<< "arg4: verbosity (0=no info, 1=verbose info)" << std::endl
<< "arg5 to 10: M (16x), N(16x), K(16x), StrideA, StrideB, StrideC" << std::endl;
return false;
}
return true;
}
template <typename ADataType,
typename BDataType,
typename XDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename CElementWiseOp,
typename AccDataType,
typename CShuffleDataType,
ck::index_t MXVectorSize>
bool run_mx_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
using ELayout = CLayout;
using DsLayout = ck::Tuple<>;
using DsDataType = ck::Tuple<>;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = CElementWiseOp;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto BlkGemmPSched = ck::BlockGemmPipelineScheduler::Intrawave;
static constexpr auto BlkGemmPVer = ck::BlockGemmPipelineVersion::v3;
#if 1
// XXX: These parameters should not exist in MX-native GEMM kernel
static constexpr ck::index_t Scale_Block_M = 128;
static constexpr ck::index_t Scale_Block_N = 128;
#endif
static constexpr ck::index_t Scale_Block_K = MXVectorSize;
// XXX: DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3 is not designed to utilize MX-specific MFMA
// instructions.
//
// XXX: DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3 is not designed to utilize device-optimized
// scaled type convert functions.
//
// XXX: In DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3, KPerBlock is expected to be equal to
// ScaleBlockK (aka MXVectorSize).
// Additionally, the following is also expected:
// static_assert(ScaleBlockM % MPerBlock == 0);
// static_assert(ScaleBlockN % NPerBlock == 0);
// In MX-native GEMM kernel these requirements should be relaxed.
//
// XXX: It appears, by default we are using mfma_f32_16x16x4xf32
// MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>::selected_mfma.k_per_blk =
// MfmaSelector<float, 16, 16, float>::selected_mfma.k_per_blk = mfma_f32_16x16x4xf32
// XXX: GridwiseGemmMultiD_ABScale_xdl_cshuffle_v3 assumes scale type is float
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3
// ######| ALayout| BLayout| DsLayout| CLayout| ADataType| AScale| BDataType| BScale| DsDataType| CDataType| GemmAcc| CShuffleDataType|AElementwise|BElementwise| CElementwise| GemmSpec|Block| ScaleBlockM| ScaleBlockN| ScaleBlockK| M| N| K| AK1| BK1| M| N|MXdl|NXdl|ABlockTransfer|ABlockTransfer|ABlockTransfer|ABlockTransfer|ABlockTransfer|ABlockTransfer| ABlock|BBlockTransfer|BBlockTransfer|BBlockTransfer|BBlockTransfer|BBlockTransfer|BBlockTransfer| BBlock| CShuffle| CShuffle|CShuffleBlockTransfer|CDEShuffleBlockTransfer| BlkGemm| BlkGemm|ComputeTypeA|ComputeTypeB|LDSTypeA|LDSTypeB|
// ######| | | | | | DataType| | DataType| | | DataType| | Operation| Operation| Operation| | Size| | | | Per| Per| Per| | | Per| Per| Per| Per| ThreadCluster| ThreadCluster|SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar|LdsExtraM| ThreadCluster| ThreadCluster|SrcAccessOrder| SrcVector| SrcScalar| DstScalar|LdsExtraN| MXdl| NXdl| ClusterLengths| Scalar| PipeSched| PipelineVer| | | | |
// ######| | | | | | | | | | | | | | | | | | | | |Block|Block| Block| | | XDL| XDL|Wave|Wave| Lengths| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths| ArrangeOrder| | Dim| PerVector| PerVector_BK1| | PerWave| PerWave| MBlock_MPerBlock| PerVectors| | | | | | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | AK0_M_AK1| | | | | | | BK0_N_BK1| | | | | |PerShuffle|PerShuffle| NBlock_NPerBlock| | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, XDataType, BDataType, XDataType, DsDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, Scale_Block_M, Scale_Block_N, Scale_Block_K, 128, 128, 128, 16, 16, 16, 16, 4, 4, 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, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlkGemmPSched, BlkGemmPVer, float, float, float, float>;
// clang-format on
auto M = problem_size.M;
auto N = problem_size.N;
auto K = problem_size.K;
auto StrideA = problem_size.StrideA;
auto StrideB = problem_size.StrideB;
auto StrideC = problem_size.StrideC;
auto f_host_tensor_descriptor =
[](ck::index_t row, ck::index_t col, ck::index_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return HostTensorDescriptor({row, col}, {stride, 1});
}
else
{
return HostTensorDescriptor({row, col}, {1, stride});
}
};
auto f_get_default_stride =
[](ck::index_t row, ck::index_t col, ck::index_t stride, auto layout) {
if(stride == -1)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<ck::index_t>(col);
}
else
{
return static_cast<ck::index_t>(row);
}
}
else
return static_cast<ck::index_t>(stride);
};
StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
if(K % Scale_Block_K != 0)
{
throw std::runtime_error("wrong! K must be multiple of Scale_Block_K (16 or 32)");
};
auto Scale_Stride_AM = f_get_default_stride(M, K / Scale_Block_K, StrideA, ALayout{});
auto Scale_Stride_BN = f_get_default_stride(K / Scale_Block_K, N, StrideB, BLayout{});
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<XDataType> a_m_k_scale(
f_host_tensor_descriptor(M, K / Scale_Block_K, Scale_Stride_AM, ALayout{})); // scales for A
Tensor<XDataType> b_k_n_scale(
f_host_tensor_descriptor(K / Scale_Block_K, N, Scale_Stride_BN, BLayout{})); // scales for B
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{})); // host verification
Tensor<CDataType> c_m_n_device_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{})); // device result downloaded to host
if(config.verbosity >= 0)
{
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "a_m_k_scale: " << a_m_k_scale.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "b_k_n_scale: " << b_k_n_scale.mDesc << std::endl;
std::cout << "c_m_n_device_result: " << c_m_n_device_result.mDesc << std::endl;
}
switch(config.init_method)
{
case 0:
if(config.verbosity > 0)
{
std::cout << "NOTE: No input data initialization." << std::endl;
}
break;
case 1:
case 2:
ck::utils::FillConstant<ADataType>{ck::type_convert<ADataType>(1.0f)}(a_m_k);
ck::utils::FillConstant<XDataType>{ck::type_convert<XDataType>(0.5f)}(a_m_k_scale);
ck::utils::FillConstant<BDataType>{ck::type_convert<BDataType>(1.0f)}(b_k_n);
ck::utils::FillConstant<XDataType>{ck::type_convert<XDataType>(2.0f)}(b_k_n_scale);
if(config.verbosity > 0)
{
std::cout << "Init A = {1}" << std::endl;
std::cout << "Init A scale = {0.5}" << std::endl;
std::cout << "Init B = {1}" << std::endl;
std::cout << "Init B scale = {2.0}" << std::endl;
std::cout << "Expect C = {K}" << std::endl;
}
break;
default:
if(config.verbosity > 0)
{
std::cout << "NOTE: No input data initialization." << std::endl;
}
}
if(config.verbosity > 0)
std::cout << "Device memory allocation..." << std::endl;
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem a_scale_device_buf(sizeof(XDataType) * a_m_k_scale.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem b_scale_device_buf(sizeof(XDataType) * b_k_n_scale.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
if(config.verbosity > 0)
std::cout << "Upload data to device..." << std::endl;
a_device_buf.ToDevice(a_m_k.mData.data());
a_scale_device_buf.ToDevice(a_m_k_scale.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
b_scale_device_buf.ToDevice(b_k_n_scale.mData.data());
if(config.verbosity > 0)
std::cout << "Done." << std::endl;
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{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{},
c_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, NumDTensor>{},
StrideC,
a_scale_device_buf.GetDeviceBuffer(),
b_scale_device_buf.GetDeviceBuffer(),
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong!\n"
"Provided combination of compilation and runtime parameters is "
"not consistent with the supported device_gemm arguments.");
}
if(config.verbosity > 0)
std::cout << "Computing GEMM on device..." << std::endl;
float ave_time =
invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, config.verbosity, 20, 50});
bool res_verified = true;
if(config.do_verification > 0)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(config.verbosity > 0)
{
std::cout << "Done." << std::endl;
std::cout << "Computing GEMM on host..." << std::endl;
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceMXGemm<ADataType,
BDataType,
CDataType,
AccDataType,
float,
PassThrough,
PassThrough,
PassThrough,
float,
float>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
a_m_k_scale,
b_k_n,
b_k_n_scale,
c_m_n_host_result,
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
if(config.verbosity > 0)
{
std::cout << "Done." << std::endl;
std::cout << "Comparing results..." << std::endl;
}
if(config.init_method == 1)
{
res_verified =
res_verified && std::abs(static_cast<float>(K) - c_m_n_device_result(0, 0)) <= 0.0f;
std::cout << "Expected vs Computed: " << 1.0f * K << " vs " << c_m_n_device_result(0, 0)
<< ((res_verified) ? " (PASSED!)" : " (FAILED!)") << std::endl;
}
res_verified = res_verified && ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result,
"Error: Incorrect results!");
if(config.verbosity > 0 && res_verified)
std::cout << "Done." << std::endl;
}
else
{
if(config.verbosity > 0)
std::cout << "Done." << std::endl;
}
if(config.time_kernel)
{
std::size_t flop = std::size_t(2) * M * N * K + M * K + K * N; // GEMM + A scale + B scale
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N +
sizeof(XDataType) * (M * K + K * N) / Scale_Block_K;
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" << std::endl;
}
return res_verified;
}
template <typename ADataType,
typename BDataType,
typename XDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename CElementWiseOp,
typename AccDataType,
typename CShuffleDataType,
ck::index_t MXVectorSize>
bool run_mx_gemm_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
return parse_cmd_args(argc, argv, problem_size, config) &&
run_mx_gemm<ADataType,
BDataType,
XDataType,
CDataType,
ALayout,
BLayout,
CLayout,
CElementWiseOp,
AccDataType,
CShuffleDataType,
MXVectorSize>(problem_size, config);
}
example/67_gemm_microscaling/gemm_mx_fp8.cpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_mx_common.hpp"
using ADataType = ck::f8_t;
using BDataType = ck::f8_t;
#if 1
// XXX: MX-native GEMM kernel will work with e8m0_bexp_t scale type
using XDataType = float;
#else
using XDataType = ck::e8m0_bexp_t;
#endif
using AccDataType = float;
using CShuffleDataType = float;
using CDataType = float;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using CElementOp = PassThrough; // elementwise transformation for C matrix
constexpr ck::index_t mx_vector_size = 128; // scaling block size
int main(int argc, char* argv[])
{
return run_mx_gemm_example<ADataType,
BDataType,
XDataType,
CDataType,
ALayout,
BLayout,
CLayout,
CElementOp,
AccDataType,
CShuffleDataType,
mx_vector_size>(argc, argv)
? 0
: -1;
}
example/CMakeLists.txt
View file @ ec959387
...@@ -23,34 +23,34 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME) ...@@ -23,34 +23,34 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}") message("adding example ${EXAMPLE_NAME}")
set(result 1) set(result 1)
if(DEFINED DTYPES) if(DEFINED DTYPES)
foreach(source IN LISTS FILE_NAME) foreach(source IN LISTS FILE_NAME)
set(test 0) set(test 0)
if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES) if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES)
set(test 1) set(test 1)
endif() endif()
if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES) if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES)
set(test 1) set(test 1)
endif() endif()
if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES) if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES)
set(test 1) set(test 1)
endif() endif()
if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES) if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES)
set(test 1) set(test 1)
endif() endif()
if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES) if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES)
set(test 1) set(test 1)
endif() endif()
if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES) if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES)
set(test 1) set(test 1)
endif() endif()
if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES) if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES)
set(test 1) set(test 1)
endif() endif()
if(test EQUAL 1) if(test EQUAL 1)
message("removing example source file ${source} ") message("removing example source file ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}") list(REMOVE_ITEM FILE_NAME "${source}")
endif() endif()
endforeach() endforeach()
endif() endif()
set(EX_TARGETS ${SUPPORTED_GPU_TARGETS}) set(EX_TARGETS ${SUPPORTED_GPU_TARGETS})
...@@ -83,6 +83,13 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME) ...@@ -83,6 +83,13 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
list(REMOVE_ITEM FILE_NAME "${source}") list(REMOVE_ITEM FILE_NAME "${source}")
endif() endif()
endforeach() endforeach()
#Do not build any microscaling examples if gfx950 target is not on the list
foreach(source IN LISTS FILE_NAME)
if(NOT EX_TARGETS MATCHES "gfx950" AND source MATCHES "_mx")
message("removing microscaling example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any FP8 examples if CK_ENABLE_FP8 not set #Do not build any FP8 examples if CK_ENABLE_FP8 not set
foreach(source IN LISTS FILE_NAME) foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED CK_ENABLE_FP8 AND source MATCHES "_fp8") if(NOT DEFINED CK_ENABLE_FP8 AND source MATCHES "_fp8")
...@@ -102,7 +109,9 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME) ...@@ -102,7 +109,9 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
if(FILE_NAME MATCHES "_xdl") if(FILE_NAME MATCHES "_xdl")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic) list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(FILE_NAME MATCHES "_wmma") elseif(FILE_NAME MATCHES "_wmma")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030) list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
elseif(FILE_NAME MATCHES "_mx") #only build mx example for gfx950
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif() endif()
set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP) set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP)
add_executable(${EXAMPLE_NAME} ${FILE_NAME}) add_executable(${EXAMPLE_NAME} ${FILE_NAME})
...@@ -195,7 +204,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME) ...@@ -195,7 +204,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
if(FILE_NAME MATCHES "_xdl") if(FILE_NAME MATCHES "_xdl")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic) list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(FILE_NAME MATCHES "_wmma") elseif(FILE_NAME MATCHES "_wmma")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030) list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
endif() endif()
set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP) set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP)
add_executable(${EXAMPLE_NAME} ${FILE_NAME}) add_executable(${EXAMPLE_NAME} ${FILE_NAME})
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment