mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-05 22:22:27 +00:00
This commit is contained in:
Ding, Yi
21
example/ck_tile/42_mx_gemm/CMakeLists.txt
Normal file
21
example/ck_tile/42_mx_gemm/CMakeLists.txt
Normal file
@@ -0,0 +1,21 @@
|
||||
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
# SPDX-License-Identifier: MIT
|
||||
|
||||
set(SUPPORTED_GPUS gfx950)
|
||||
|
||||
set(has_supported_gpu FALSE)
|
||||
foreach(gpu IN LISTS GPU_TARGETS)
|
||||
if(gpu IN_LIST SUPPORTED_GPUS)
|
||||
set(has_supported_gpu TRUE)
|
||||
break()
|
||||
endif()
|
||||
endforeach()
|
||||
|
||||
if(has_supported_gpu)
|
||||
add_executable(tile_example_mx_gemm mx_gemm.cpp)
|
||||
set(EXAMPLE_MX_GEMM_COMPILE_OPTIONS -Wno-undefined-func-template)
|
||||
if(CK_USE_OCP_FP8)
|
||||
list(APPEND EXAMPLE_MX_GEMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
|
||||
endif()
|
||||
target_compile_options(tile_example_mx_gemm PRIVATE ${EXAMPLE_MX_GEMM_COMPILE_OPTIONS})
|
||||
endif()
|
||||
128
example/ck_tile/42_mx_gemm/mx_gemm.cpp
Normal file
128
example/ck_tile/42_mx_gemm/mx_gemm.cpp
Normal file
@@ -0,0 +1,128 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#include <hip/hip_runtime.h>
|
||||
|
||||
#include <cstring>
|
||||
#include <iostream>
|
||||
#include <ostream>
|
||||
#include <string>
|
||||
#include <tuple>
|
||||
#include <type_traits>
|
||||
|
||||
#include "ck_tile/host.hpp"
|
||||
#include "mx_gemm.hpp"
|
||||
#include "mx_gemm_instance.hpp"
|
||||
|
||||
template <typename Layout>
|
||||
static constexpr inline auto is_row_major(Layout layout_)
|
||||
{
|
||||
return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
|
||||
ck_tile::tensor_layout::gemm::RowMajor>>{};
|
||||
}
|
||||
|
||||
template <typename GemmConfig,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout,
|
||||
typename ScaleM,
|
||||
typename ScaleN,
|
||||
bool UsePersistentKernel = false>
|
||||
float invoke_mx_gemm(ck_tile::DeviceMem& a_dev_buf,
|
||||
ck_tile::DeviceMem& b_dev_buf,
|
||||
ck_tile::DeviceMem& c_dev_buf,
|
||||
ck_tile::index_t M,
|
||||
ck_tile::index_t N,
|
||||
ck_tile::index_t K,
|
||||
ck_tile::index_t stride_A,
|
||||
ck_tile::index_t stride_B,
|
||||
ck_tile::index_t stride_C,
|
||||
ck_tile::index_t kbatch,
|
||||
ScaleM scale_m,
|
||||
ScaleN scale_n,
|
||||
int n_warmup,
|
||||
int n_repeat)
|
||||
{
|
||||
MXGemmHostArgs<ScaleM, ScaleN> args(a_dev_buf.GetDeviceBuffer(),
|
||||
b_dev_buf.GetDeviceBuffer(),
|
||||
c_dev_buf.GetDeviceBuffer(),
|
||||
kbatch,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
stride_C,
|
||||
scale_m,
|
||||
scale_n);
|
||||
|
||||
// Simplified invocation - comp_async handles hot loop and tail internally
|
||||
auto invoke_splitk_path = [&](auto split_k_) {
|
||||
return mx_gemm_calc<GemmConfig,
|
||||
ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ScaleM,
|
||||
ScaleN,
|
||||
UsePersistentKernel,
|
||||
split_k_.value>(
|
||||
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
|
||||
};
|
||||
|
||||
float ave_time = (args.k_batch == 1) ? invoke_splitk_path(std::false_type{})
|
||||
: invoke_splitk_path(std::true_type{});
|
||||
|
||||
constexpr int APackedSize = ck_tile::numeric_traits<ADataType>::PackedSize;
|
||||
constexpr int BPackedSize = ck_tile::numeric_traits<BDataType>::PackedSize;
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N * K / 32;
|
||||
std::size_t num_byte = sizeof(ADataType) * M * K / APackedSize +
|
||||
sizeof(BDataType) * N * K / BPackedSize + sizeof(CDataType) * M * N +
|
||||
sizeof(ck_tile::e8m0_t) * M * K / 32 +
|
||||
sizeof(ck_tile::e8m0_t) * N * K / 32;
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
float gb_per_sec = num_byte / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Run " << ck_tile::gemm_prec_str<ADataType, BDataType>() << " MX GEMM kernel " //
|
||||
<< " M = " << M << " N = " << N << " K = " << K << " StrideA = " << stride_A
|
||||
<< " StrideB = " << stride_B << " StrideC = " << stride_C << " : " << ave_time
|
||||
<< " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
|
||||
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
auto create_args(int argc, char* argv[])
|
||||
{
|
||||
ck_tile::ArgParser arg_parser;
|
||||
arg_parser.insert("m", "4096", "m dimension")
|
||||
.insert("n", "4096", "n dimension")
|
||||
.insert("k", "4096", "k dimension")
|
||||
.insert("a_layout", "R", "A tensor data layout - Row by default")
|
||||
.insert("b_layout", "C", "B tensor data layout - Row by default")
|
||||
.insert("c_layout", "R", "C tensor data layout - Row by default")
|
||||
.insert("stride_a", "0", "Tensor A stride")
|
||||
.insert("stride_b", "0", "Tensor B stride")
|
||||
.insert("stride_c", "0", "Tensor C stride")
|
||||
.insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
|
||||
.insert(
|
||||
"mx_prec", "fp4xfp4", "data type for activation and weight, support: fp4xfp4, fp8xfp8")
|
||||
.insert("warmup", "50", "number of iterations before benchmark the kernel")
|
||||
.insert("repeat", "100", "number of iterations to benchmark the kernel")
|
||||
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
|
||||
.insert("split_k", "1", "splitK value")
|
||||
.insert("init", "0", "0:random, 1:constant(1)");
|
||||
bool result = arg_parser.parse(argc, argv);
|
||||
return std::make_tuple(result, arg_parser);
|
||||
}
|
||||
|
||||
#include "run_mx_gemm.inc"
|
||||
|
||||
int main(int argc, char* argv[]) { return run_mx_gemm_example(argc, argv); }
|
||||
99
example/ck_tile/42_mx_gemm/mx_gemm.hpp
Normal file
99
example/ck_tile/42_mx_gemm/mx_gemm.hpp
Normal file
@@ -0,0 +1,99 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <string>
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/kernel_launch.hpp"
|
||||
#include "ck_tile/ops/epilogue.hpp"
|
||||
#include "ck_tile/ops/gemm.hpp"
|
||||
#include "ck_tile/ops/gemm_mx/kernel/scale_pointer.hpp"
|
||||
|
||||
template <typename ScaleM, typename ScaleN>
|
||||
struct MXGemmHostArgs : ck_tile::UniversalGemmHostArgs<1, 1, 0>
|
||||
{
|
||||
using Base = ck_tile::UniversalGemmHostArgs<1, 1, 0>;
|
||||
|
||||
MXGemmHostArgs(const void* a_ptr,
|
||||
const void* b_ptr,
|
||||
void* c_ptr_,
|
||||
ck_tile::index_t k_batch_,
|
||||
ck_tile::index_t M_,
|
||||
ck_tile::index_t N_,
|
||||
ck_tile::index_t K_,
|
||||
ck_tile::index_t stride_A_,
|
||||
ck_tile::index_t stride_B_,
|
||||
ck_tile::index_t stride_C_,
|
||||
ScaleM scale_m_,
|
||||
ScaleN scale_n_)
|
||||
: Base({a_ptr},
|
||||
{b_ptr},
|
||||
{},
|
||||
c_ptr_,
|
||||
k_batch_,
|
||||
M_,
|
||||
N_,
|
||||
K_,
|
||||
{stride_A_},
|
||||
{stride_B_},
|
||||
{},
|
||||
stride_C_),
|
||||
scale_m(scale_m_),
|
||||
scale_n(scale_n_)
|
||||
{
|
||||
}
|
||||
|
||||
ScaleM scale_m;
|
||||
ScaleN scale_n;
|
||||
};
|
||||
|
||||
// GEMM config with 16x16 warp tile
|
||||
|
||||
struct MxGemmConfig
|
||||
{
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 128;
|
||||
static constexpr ck_tile::index_t K_Tile = 512;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 128;
|
||||
|
||||
static constexpr bool kPadM = false;
|
||||
static constexpr bool kPadN = false;
|
||||
static constexpr bool kPadK = false;
|
||||
|
||||
static constexpr bool TransposeC = false;
|
||||
static constexpr bool UseStructuredSparsity = false;
|
||||
|
||||
static constexpr int kBlockPerCu = 1;
|
||||
static constexpr int TileParitionerGroupNum = 8;
|
||||
static constexpr int TileParitionerM01 = 4;
|
||||
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Intrawave;
|
||||
static constexpr ck_tile::index_t NumWaveGroups = 1;
|
||||
static constexpr bool DoubleSmemBuffer = false; // comp_async uses double buffer
|
||||
static constexpr bool Preshuffle = false;
|
||||
|
||||
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
|
||||
static constexpr bool TiledMMAPermuteN = false;
|
||||
};
|
||||
struct MXfp4_GemmConfig16 : MxGemmConfig
|
||||
{
|
||||
static constexpr ck_tile::index_t M_Tile = 64;
|
||||
static constexpr ck_tile::index_t N_Tile = 64;
|
||||
static constexpr ck_tile::index_t K_Tile = 256;
|
||||
};
|
||||
|
||||
// GEMM config with 16x16 warp tile
|
||||
struct MXfp8_GemmConfig16 : MxGemmConfig
|
||||
{
|
||||
static constexpr ck_tile::index_t M_Tile = 64;
|
||||
static constexpr ck_tile::index_t N_Tile = 64;
|
||||
static constexpr ck_tile::index_t K_Tile = 256;
|
||||
};
|
||||
106
example/ck_tile/42_mx_gemm/mx_gemm_instance.hpp
Normal file
106
example/ck_tile/42_mx_gemm/mx_gemm_instance.hpp
Normal file
@@ -0,0 +1,106 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/host.hpp"
|
||||
#include "mx_gemm.hpp"
|
||||
#include "ck_tile/ops/gemm_mx/pipeline/gemm_pipeline_ag_bg_cr_comp_async.hpp"
|
||||
#include "ck_tile/ops/gemm_mx/kernel/gemm_mx_kernel.hpp"
|
||||
|
||||
template <typename Layout>
|
||||
using is_row_major_t = ck_tile::bool_constant<
|
||||
std::is_same_v<ck_tile::remove_cvref_t<Layout>, ck_tile::tensor_layout::gemm::RowMajor>>;
|
||||
|
||||
template <typename GemmConfig,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout,
|
||||
typename ScaleM,
|
||||
typename ScaleN,
|
||||
bool persistent,
|
||||
bool Splitk>
|
||||
float mx_gemm_calc(const MXGemmHostArgs<ScaleM, ScaleN>& args, const ck_tile::stream_config& s)
|
||||
{
|
||||
using GemmShape = ck_tile::TileGemmShape<
|
||||
ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
|
||||
ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
|
||||
ck_tile::
|
||||
sequence<GemmConfig::M_Warp_Tile, GemmConfig::N_Warp_Tile, GemmConfig::K_Warp_Tile>>;
|
||||
|
||||
using MXGemmTraits = ck_tile::TileGemmUniversalTraits<GemmConfig::kPadM,
|
||||
GemmConfig::kPadN,
|
||||
GemmConfig::kPadK,
|
||||
GemmConfig::DoubleSmemBuffer,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
GemmConfig::TransposeC,
|
||||
GemmConfig::UseStructuredSparsity,
|
||||
persistent,
|
||||
GemmConfig::NumWaveGroups,
|
||||
GemmConfig::Preshuffle>;
|
||||
|
||||
using ComputeDataType = ADataType;
|
||||
static_assert(sizeof(ComputeDataType) >= sizeof(BDataType),
|
||||
"mixed_prec_gemm requires ADataType is a wider type than BDataType");
|
||||
|
||||
using MXPipelineProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
MXGemmTraits,
|
||||
GemmConfig::Scheduler>;
|
||||
|
||||
// Use the new MX comp_async pipeline with MX scaling support
|
||||
using MXGemmPipeline = ck_tile::MXGemmPipelineAgBgCrCompAsync<MXPipelineProblem>;
|
||||
|
||||
using TilePartitioner =
|
||||
ck_tile::GemmSpatiallyLocalTilePartitioner<GemmShape,
|
||||
GemmConfig::TileParitionerGroupNum,
|
||||
GemmConfig::TileParitionerM01>;
|
||||
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ComputeDataType,
|
||||
ComputeDataType,
|
||||
ck_tile::tuple<>, // DsDataType
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ck_tile::tuple<>, // DsLayout
|
||||
CLayout,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
GemmConfig::M_Warp,
|
||||
GemmConfig::N_Warp,
|
||||
GemmConfig::M_Warp_Tile,
|
||||
GemmConfig::N_Warp_Tile,
|
||||
GemmConfig::K_Warp_Tile,
|
||||
MXPipelineProblem::TransposeC>>;
|
||||
|
||||
using Kernel = ck_tile::MXGemmKernel<TilePartitioner, MXGemmPipeline, GemmEpilogue>;
|
||||
|
||||
auto kargs = Kernel::MakeKernelArgs(std::array<const void*, 1>{args.as_ptr},
|
||||
std::array<const void*, 1>{args.bs_ptr},
|
||||
std::array<const void*, 0>{},
|
||||
args.e_ptr,
|
||||
args.k_batch,
|
||||
args.M,
|
||||
args.N,
|
||||
args.K,
|
||||
std::array<ck_tile::index_t, 1>{args.stride_As},
|
||||
std::array<ck_tile::index_t, 1>{args.stride_Bs},
|
||||
std::array<ck_tile::index_t, 0>{},
|
||||
args.stride_E,
|
||||
args.scale_m,
|
||||
args.scale_n);
|
||||
|
||||
const auto kernel = ck_tile::make_kernel<Kernel::kBlockPerCu>(
|
||||
Kernel{}, Kernel::GridSize(kargs), Kernel::BlockSize(), 0, kargs);
|
||||
|
||||
return ck_tile::launch_kernel(s, kernel);
|
||||
}
|
||||
220
example/ck_tile/42_mx_gemm/run_mx_gemm.inc
Normal file
220
example/ck_tile/42_mx_gemm/run_mx_gemm.inc
Normal file
@@ -0,0 +1,220 @@
|
||||
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
||||
// SPDX-License-Identifier: MIT
|
||||
|
||||
// Calculate relative and absolute error thresholds for MX GEMM
|
||||
template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
|
||||
auto calculate_rtol_atol(const ck_tile::index_t K, const float max_accumulated_value)
|
||||
{
|
||||
using ComputeType =
|
||||
std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>;
|
||||
// Calculate thresholds
|
||||
const auto rtol = ck_tile::get_relative_threshold<ComputeType, CDataType, AccDataType>(K);
|
||||
const auto atol = ck_tile::get_absolute_threshold<ComputeType, CDataType, AccDataType>(
|
||||
max_accumulated_value, K);
|
||||
return ck_tile::make_tuple(rtol, atol);
|
||||
}
|
||||
|
||||
// Use e8m0_t directly without packing - simpler and cleaner approach
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename GemmConfig,
|
||||
bool UsePersistentKernel,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
int run_mx_gemm_with_layouts(int argc, char* argv[], ALayout, BLayout, CLayout)
|
||||
{
|
||||
auto [result, arg_parser] = create_args(argc, argv);
|
||||
if(!result)
|
||||
return -1;
|
||||
|
||||
ck_tile::index_t M = arg_parser.get_int("m");
|
||||
ck_tile::index_t N = arg_parser.get_int("n");
|
||||
ck_tile::index_t K = arg_parser.get_int("k");
|
||||
ck_tile::index_t stride_A = arg_parser.get_int("stride_a");
|
||||
ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
|
||||
ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
|
||||
int validation = arg_parser.get_int("v");
|
||||
int n_warmup = arg_parser.get_int("warmup");
|
||||
int n_repeat = arg_parser.get_int("repeat");
|
||||
int kbatch = arg_parser.get_int("split_k");
|
||||
int init_method = arg_parser.get_int("init");
|
||||
|
||||
using CDataType = ck_tile::fp16_t;
|
||||
|
||||
// Use get_default_stride helper for automatic leading dimension calculation (only if not
|
||||
// explicitly provided)
|
||||
if(stride_A == 0)
|
||||
stride_A = ck_tile::get_default_stride(M, K, 0, is_row_major(ALayout{}));
|
||||
if(stride_B == 0)
|
||||
stride_B = ck_tile::get_default_stride(K, N, 0, is_row_major(BLayout{}));
|
||||
if(stride_C == 0)
|
||||
stride_C = ck_tile::get_default_stride(M, N, 0, is_row_major(CLayout{}));
|
||||
|
||||
ck_tile::HostTensor<ADataType> a_host(
|
||||
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(ALayout{})));
|
||||
ck_tile::HostTensor<BDataType> b_host(
|
||||
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(BLayout{})));
|
||||
ck_tile::HostTensor<CDataType> c_host(
|
||||
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
||||
|
||||
// Scale tensors - follow parent matrix layouts for optimal memory access
|
||||
// A scales: [M, K/32] with A's layout
|
||||
// B scales: [K/32, N] with B's layout
|
||||
using ScaleType = ck_tile::e8m0_t;
|
||||
ck_tile::index_t scale_k_size = K / 32;
|
||||
|
||||
// Follow A/BLayout to get the layouts for the scale tensors
|
||||
ck_tile::index_t stride_scale_a =
|
||||
ck_tile::get_default_stride(M, scale_k_size, 0, is_row_major(ALayout{}));
|
||||
ck_tile::index_t stride_scale_b =
|
||||
ck_tile::get_default_stride(scale_k_size, N, 0, is_row_major(BLayout{}));
|
||||
|
||||
ck_tile::HostTensor<ScaleType> scale_a_host(
|
||||
ck_tile::host_tensor_descriptor(M, scale_k_size, stride_scale_a, is_row_major(ALayout{})));
|
||||
ck_tile::HostTensor<ScaleType> scale_b_host(
|
||||
ck_tile::host_tensor_descriptor(scale_k_size, N, stride_scale_b, is_row_major(BLayout{})));
|
||||
int seed = 1234;
|
||||
switch(init_method)
|
||||
{
|
||||
case 0:
|
||||
// Initialize A, B, and scales to random values
|
||||
ck_tile::FillUniformDistribution<ADataType>{-2.f, 2.f, seed++}(a_host);
|
||||
ck_tile::FillUniformDistribution<BDataType>{-2.f, 2.f, seed++}(b_host);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{0.001f, 10.f, seed++}(scale_a_host);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{0.001f, 10.f, seed++}(scale_b_host);
|
||||
break;
|
||||
case 1:
|
||||
// Initialize A, B, and scales to 1.0
|
||||
ck_tile::FillConstant<ADataType>{ADataType(1.f)}(a_host);
|
||||
ck_tile::FillConstant<BDataType>{BDataType(1.f)}(b_host);
|
||||
ck_tile::FillConstant<ScaleType>{ScaleType(1.f)}(scale_a_host);
|
||||
ck_tile::FillConstant<ScaleType>{ScaleType(1.f)}(scale_b_host);
|
||||
break;
|
||||
case 2:
|
||||
// Initialize A and B with random values but with constant 1.0 scales
|
||||
ck_tile::FillUniformDistribution<ADataType>{-2.f, 2.f, seed++}(a_host);
|
||||
ck_tile::FillUniformDistribution<BDataType>{-2.f, 2.f, seed++}(b_host);
|
||||
ck_tile::FillConstant<ScaleType>{ScaleType(0.1f)}(scale_a_host);
|
||||
ck_tile::FillConstant<ScaleType>{ScaleType(0.1f)}(scale_b_host);
|
||||
break;
|
||||
}
|
||||
|
||||
// Device buffers for A, B, C, and scale tensors
|
||||
ck_tile::DeviceMem a_dev_buf(a_host.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem b_dev_buf(b_host.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem c_dev_buf(c_host.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem scale_a_dev_buf(scale_a_host.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem scale_b_dev_buf(scale_b_host.get_element_space_size_in_bytes());
|
||||
|
||||
a_dev_buf.ToDevice(a_host.data());
|
||||
b_dev_buf.ToDevice(b_host.data());
|
||||
c_dev_buf.SetZero(); // Initialize C buffer to zero
|
||||
scale_a_dev_buf.ToDevice(scale_a_host.data());
|
||||
scale_b_dev_buf.ToDevice(scale_b_host.data());
|
||||
|
||||
// Scale pointers - use e8m0_t* directly
|
||||
using ScaleM = ck_tile::MXScalePointer<ScaleType, 1, 32>; // in blocks of 32 in K
|
||||
using ScaleN = ck_tile::MXScalePointer<ScaleType, 1, 32>;
|
||||
ScaleM scale_m(reinterpret_cast<ScaleType*>(scale_a_dev_buf.GetDeviceBuffer()));
|
||||
ScaleN scale_n(reinterpret_cast<ScaleType*>(scale_b_dev_buf.GetDeviceBuffer()));
|
||||
|
||||
float ave_time = invoke_mx_gemm<GemmConfig,
|
||||
ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ScaleM,
|
||||
ScaleN,
|
||||
UsePersistentKernel>(a_dev_buf,
|
||||
b_dev_buf,
|
||||
c_dev_buf,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
stride_C,
|
||||
kbatch,
|
||||
scale_m,
|
||||
scale_n,
|
||||
n_warmup,
|
||||
n_repeat);
|
||||
|
||||
(void)ave_time;
|
||||
|
||||
bool pass = true;
|
||||
if(validation > 0)
|
||||
{
|
||||
// get output data from device
|
||||
c_dev_buf.FromDevice(c_host.data());
|
||||
|
||||
// compute reference
|
||||
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
|
||||
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
||||
c_m_n_host_ref.SetZero();
|
||||
|
||||
ck_tile::reference_mx_gemm<ADataType, BDataType, ScaleType, AccDataType, CDataType>(
|
||||
a_host, b_host, c_m_n_host_ref, scale_a_host, scale_b_host);
|
||||
|
||||
const float max_accumulated_value =
|
||||
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
|
||||
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
|
||||
K, max_accumulated_value);
|
||||
const double rtol = rtol_atol.at(ck_tile::number<0>{});
|
||||
const double atol = rtol_atol.at(ck_tile::number<1>{});
|
||||
pass = ck_tile::check_err(c_host, c_m_n_host_ref, "Error: Incorrect results!", rtol, atol);
|
||||
|
||||
std::cout << "Relative error threshold: " << rtol << " Absolute error threshold: " << atol
|
||||
<< std::endl;
|
||||
std::cout << "The CPU verification result is: " << (pass ? "correct" : "fail") << std::endl;
|
||||
}
|
||||
return pass ? 0 : -1;
|
||||
}
|
||||
|
||||
int run_mx_gemm_example(int argc, char* argv[])
|
||||
{
|
||||
auto [result, arg_parser] = create_args(argc, argv);
|
||||
if(!result)
|
||||
return -1;
|
||||
|
||||
using Row = ck_tile::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
std::string mx_prec = arg_parser.get_str("mx_prec");
|
||||
std::string a_layout = arg_parser.get_str("a_layout");
|
||||
std::string b_layout = arg_parser.get_str("b_layout");
|
||||
|
||||
if(a_layout == "R" && b_layout == "C")
|
||||
{
|
||||
if(mx_prec == "fp4" || mx_prec == "fp4xfp4")
|
||||
{
|
||||
return run_mx_gemm_with_layouts<ck_tile::pk_fp4_t,
|
||||
ck_tile::pk_fp4_t,
|
||||
float,
|
||||
MXfp4_GemmConfig16,
|
||||
true>(argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(mx_prec == "fp8" || mx_prec == "fp8xfp8")
|
||||
{
|
||||
return run_mx_gemm_with_layouts<ck_tile::fp8_t,
|
||||
ck_tile::fp8_t,
|
||||
float,
|
||||
MXfp8_GemmConfig16,
|
||||
true>(argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Only fp4/8 is supported currently!");
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Only A=Row, B=Col layout is supported currently!");
|
||||
}
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user