ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 12:30:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add MoE & FP8 Blockscale WP Kernels for GFX950 (#2297)

Browse Source 
* [fix] align v3 gufusion pipeline

* fix device kernel selection.

* Add .co direct asm support by CK_USE_ASM_MOE_STAGE2_BLOCKSCALE

* experimental optimization for scale load in blkscale gemm

* Add asm for no-loop v3_128x128x128

* fix bugs

* tune fp8 example

* Update v1_128x128x128 to 2x2 instead of 4x1

* wip

* add warmup to asm launch

* wip2

* 16x16 function merged to moe

* temp save, a performant version.

* wip3

* Update .co binary to 16x16

* 16x16x128 correct; 64x64x128 failed

* update

* use mem_op::set when topk=1

* add mx fp8 b_preshuffle support, function not yet tested.

* Spilt the fp4 target. Fix the known bugs. 128x128x128 sanity checked; remove prints

* some fixes

* fix update

* remove some unnecessary hacky; enable 256x256x256 tilesize

* update for function debug

* Add pipeline v3. Have some runtime issue and register spill

* Fix pipe v3 correctness issue

* remove unnecessary hacky

* clang format

* fix a bug

* fix the bug, functional test passed

* tempsave; buggy at passed 4 e8m0 to scaled mfma

* added fp4_bpreshuffle example, build failures

* fixed some bugs

* implement shuffled scale mxfp4gemm, blocker: opsel not effect

* hotfix

* fix bugs, build passed

* (M, N, K)=(128, 128, 128) function failed.

* temp save for gemm1. Function not ready

* fix compile error. Gemm2 pass. Gemm1 WIP

* fix bug for a lds read

* update moe

* Compile pass. Gemm1 function WIP

* update moe

* fix fp8; fix even/odd

* tempsave

* update moe

* Revert "update"

This reverts commit c7d79dcb672616d9bc0fd9958f714fc80e7c84fd.

* Revert "use mem_op::set when topk=1"

This reverts commit 8c7772860735001a51421e7b6d0a28f6676d6c40.

* Add v3 128x128x128_4x4_16x16.co for gfx950

* temp cmake flag suppression  for aiter test

* add code for mxfp4 gemm, blockscale not supported yet

* gemm1 up-only pass. GU WIP

* function pass with inline asm hacky

* revert unexpected file change

* updated and build passed

* update CE elementOP

* added code for debug

* Gemm1 GUFusion function pass. Perf WIP

* Fix fp8/bf8; remove duplicated code

* disable the scheduler in v3; bring it back when compiler feature ready.

* update moe v1 pipeline

* Add gemm1 v1 32x128x128

* remove schedule barrier

* updated

* Fix fp8/bf8 B-row

* mfma using asm, device result correct, host result need to check

* gemm1 v3 64x128x128 debug

* fix cpu ref

* a/b thread_desc stride fix

* Use random scale for init1

* 16x16x128 input size blockscale function passed

* fix blockscale gemm bug

* tempsave. Almost all instances passed.

* v1 fix for mi350.

* temp save

* debug save

* update debug

* fix the bug, 128x128x256 tile function passed

* v3

* rename moe block selector and pipeline

* Add gemm1 v1

* Add gemm1 v1 to selector

* added mx moe block v3 support, function passed

* compile error fix

* Improve the pipeline

* Pack e8m0 as int32_t

* v1 compile pass. Function not ready

* debug synchronize issue over different GPU/ROCm

* minor fix

* Add profiler filter

* Add f4 ckProfiler

* Fix example compile error

* Add f4 profiler examples

* tempsave

* v1 function pass.

* v3 function pass

* align file and function name

* mx_moe_fp4 ready for aiter with clang-format.

* modify the way we represent fp4

* generalize the pipeline scheduling.

* init moe mx f4 scale shuffle

* Cmakelist diable compiler-bound flags

* mx_fp4 default parameter change

* Moe blockscale gemm1&gemm2 asm support for aiter. Suppression cmkae flag til new compler.

* update code

* tempsave; modify the way we represent fp4

* generalize the pipeline scheduling.

* Add gemm1 gfx942 .co support

* updated code, build passed.

* Update gemm2 asm with latest compiler flag

* Fix mx f4 ckProfiler

* Fix blockwise gemm mx v1

* lds conflict free + buffer load lds

* Add gemm2 v3 64x128x128

* fix a, b scale loading bugs, a, b scale loading now correctly

* Add gemm2 v3 64x128x128

* commit with debug info

* fix fp4 profiler

* Add mx fp4 pileline v1 instances

* Fix v2 topk_weight cal. Add silu asm.

* v2 tok_weight WIP

* init mx fp4 B no preshuffle version

* tempsave. compile pass, function wrong

* enable fp4 moe no weigth preshuffle, function pass

* update the TFlops calculation in the example

* Add gemm2 64x128x128 asm. Fix BF16 ref.

* fix 2 typos in fp4_preshuffle

* Better kernel selection in device classes

* correct preShuffleBuffer

we should used packed k to do shuffle.

* lds conflict free + buffer load lds

* optimize offset math in dma

* Fix fp4 ckProfiler

* Fix MX MFMA tests

* fix f4 pipeline issues

* gemm1 func pass

* update mx moe gemm1_bns tile size to 64x128x256

* update mx moe gemm1 gemm2 TF and BW calculation

* fix typo

* temp save

* Fix example_gemm_mx build

* rename the block pipeline

* correct a typo in tail

* Add rotating to mx examples

* fix the correctness issue

* Fix v1; use M padding

* Add NT flag to B/BScale buffer

* Merge gemm_mx_common.hpp

* temp save, 4.4~4.5

* Fix 'Merge gemm_mx_common.hpp'

* refactor the pipeline

* Pad the M for scale buffer unconditionaly

* update MX moe GEMM1 hotloopscheduling

* change the gemm1 tile from 64x128x128 to 128x64x128

* Unconditional Ascale padding

* Pad shuffled a scale only

* pad ascale

* add vmcnt guard for async copy

* Profiler add f4 wp

* Merge preshuffle device

* Add more fp4 wp instances

* Fix do_weight in gemm1. Fix cshuffle_datatype. Clang-format

* Clang-format after 2 merges

* Remove rocm6.3 workaround flags and macro

* Fix fp8 config

* Fix bf8 config

* flag and barrier fix for copmiler branch MainOpSelV3

* Add fp8 profiler instances

* Remove debug infos; Enable flags for blockscale f8

* No asm ver. for merging moe blocksale fp8 into mainline

* update the flag name for f8blockscale

* recover example

* fix performance bug of bpreshuffle f8 gemm

* clang format, remove  single rate mfma restriction for f8

* remove single rate mfma restriction for f8 blockscale gemm

* Fix moe blockscale gemm1 barrier 0x800 for new compiler

* add pipeline v1 for MOE Gemm2

* Use v1 pipeline for example_moe_gemm2_xdl_mx_fp4_bns

* Fix OOB; add MB96 instances

* remove unnecessary files

* fix the cmake issue

* Enable splitk for mxfp4; clang format;

* Generate random tensor values with multiple threads

* Use packed_size_v for A/BPackedSize

* Fix warning

* Fix target_compile_options for disabled target on gfx942

* fix moe pki4 on gfx950

* doc the kGroup definition

* Fix ThreadwiseTensorSliceTransfer_v4::Run (Fuse scale)

* Refactor thread_copy_lds_direct_load; fix gfx942 direct lds load example; fix f16_pki4 example

* Fix unknown compiler flag

* fix two failed examples.

* fix some failure tile size in gfx950 universal gemm. fix test_gemm_fp16

* workaround fix for test_gemm_f32; * We have very limited support for lds direct load if input matrix is not K major

* fix test_gemm_splitk;

* Fix compile for mx_mfma_op

* add mfma selection logic for multipled_v3

* Clean up

* Fix device gemm mx link error

* improve the global atomic pattern

* Revert unnecessary copyright updates

* restore minimum_occupancy logic

* Avoid data race in moe gemm2 ref

* Build fp8 gemm_multiply_multiply and moe only on gfx94/95

* update the instance in device_mx_gemm

* Resolve comments

* Copyright 2025

* Remove unused code

* fix library linking issue

---------

Co-authored-by: OscarXu <huaiguxu@amd.com>
Co-authored-by: lalala-sh <Jiaxing.Wen@amd.com>
Co-authored-by: mtgu0705 <mtgu@amd.com>
Co-authored-by: aska-0096 <haocwang@amd.com>
Co-authored-by: Your Name <you@example.com>
Co-authored-by: valarLip <340077269@qq.com>
Co-authored-by: feifei14119 <feiw@amd.com>
Co-authored-by: Lin, Qun <qlin@amd.com>
Co-authored-by: Andriy Roshchenko <andriy.roshchenko@amd.com>
Co-authored-by: joye <joye@amd.com>
Co-authored-by: asleepzzz <hanwen.chang@amd.com>

[ROCm/composable_kernel commit: 37554c31e8]
This commit is contained in:
Yi DING
2025-06-12 09:25:59 +08:00
committed by GitHub
parent e98451e42f
commit b40267f7a6
85 changed files with 32508 additions and 431 deletions
Download Patch File Download Diff File Expand all files Collapse all files

415
profiler/include/profiler/profile_gemm_blockscale_wp_impl.hpp Normal file
View File

@@ -0,0 +1,415 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_blockscale_wp.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"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename InOutDataType>
void preShuffleBuffer(const InOutDataType* src, InOutDataType* dst, int N, int K, int NXdl)
{
int KPack = 16;
int NLane = NXdl;
int KLane = 64 / NLane;
int K0 = K / (KLane * KPack);
// K -> K0 KLane KPack
// N -> N0 NLane
// N, K -> N0 K0 KLane NLane KPack
int tempk;
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K; ++k)
{
int n0 = n / NLane;
int n1 = n % NLane;
int k0 = k / (KLane * KPack);
tempk = k % (KLane * KPack);
int k1 = tempk / KPack;
int k2 = tempk % KPack;
int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
k1 * KPack * NLane + n1 * KPack + k2;
dst[outputIndex] = src[n * K + k];
}
}
}
template <typename A0DataType,
typename A1DataType,
typename B0DataType,
typename B1DataType,
typename ComputeDataType,
typename AccDataType,
typename EDataType,
index_t ScaleBlockM,
index_t ScaleBlockN,
index_t ScaleBlockK,
typename ALayout,
typename BLayout,
typename ELayout>
bool profile_gemm_blockscale_weighpreshuffle_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideE,
int n_warmup,
int n_iter,
uint64_t rotating = 0)
{
bool pass = true;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
ck::index_t Scale_Stride_AM = ((M + ScaleBlockM - 1) / ScaleBlockM);
ck::index_t Scale_Stride_BN = ck::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor>
? ((K + ScaleBlockK - 1) / ScaleBlockK)
: ((N + ScaleBlockN - 1) / ScaleBlockN);
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<A1DataType> a1_m_k(f_host_tensor_descriptor((M + ScaleBlockM - 1) / ScaleBlockM,
(K + ScaleBlockK - 1) / ScaleBlockK,
Scale_Stride_AM,
ck::tensor_layout::gemm::ColumnMajor{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<B0DataType> b_preshuffled_mfma16(
f_host_tensor_descriptor(K, N, StrideB, BLayout{})); // use layout only for size
Tensor<B0DataType> b_preshuffled_mfma32(
f_host_tensor_descriptor(K, N, StrideB, BLayout{})); // use layout only for size
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor((K + ScaleBlockK - 1) / ScaleBlockK,
(N + ScaleBlockN - 1) / ScaleBlockN,
Scale_Stride_BN,
BLayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
int total_gemm_needed =
a0_m_k.GetElementSpaceSizeInBytes() + b0_k_n.GetElementSpaceSizeInBytes() +
a1_m_k.GetElementSpaceSizeInBytes() + b1_k_n.GetElementSpaceSizeInBytes();
int rotating_count = std::max(
1,
std::min(n_iter,
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
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_device_result.mDesc << std::endl;
std::cout << "rotating count: " << rotating_count << 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;
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});
}
preShuffleBuffer(b0_k_n.mData.data(), b_preshuffled_mfma16.mData.data(), N, K, 16);
preShuffleBuffer(b0_k_n.mData.data(), b_preshuffled_mfma32.mData.data(), N, K, 32);
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf_mfma16(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf_mfma32(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
b_device_buf_mfma16.ToDevice(b_preshuffled_mfma16.mData.data());
b_device_buf_mfma32.ToDevice(b_preshuffled_mfma32.mData.data());
a1_device_buf.ToDevice(a1_m_k.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
using DeviceOp =
ck::tensor_operation::device::DeviceGemmMultipleD_BlockScale_BPreshuffle<ALayout,
BLayout,
ck::Tuple<>,
ELayout,
A0DataType,
A1DataType,
B0DataType,
B1DataType,
ck::Tuple<>,
EDataType,
ScaleBlockM,
ScaleBlockN,
ScaleBlockK,
AElementOp,
BElementOp,
CElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// Run reference GEMM
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
Tensor<float> a_m_k({M, K});
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 / ScaleBlockM, k / ScaleBlockK);
}
}
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 / ScaleBlockK, n / ScaleBlockN);
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<float,
float,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough,
float>;
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));
}
}
}
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
int NPerXdl = op_ptr->GetPreShuffleParameters();
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<A0DataType*>(a0_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(NPerXdl == 16 ? b_device_buf_mfma16.GetDeviceBuffer()
: b_device_buf_mfma32.GetDeviceBuffer()),
std::array<const void*, 0>{},
static_cast<EDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 0>{},
StrideE,
a1_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer(),
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_verification)
{
c_device_buf.FromDevice(e_m_n_device_result.mData.data());
#if defined CK_ENABLE_FP8
// set softer tolerances for fp8
if constexpr(is_same_v<A0DataType, f8_t> || is_same_v<B0DataType, f8_t> ||
is_same_v<EDataType, f8_t>)
{
std::string msg = "Error: Incorrect results!";
double rtol = 5e-2;
double atol = 5e-2;
bool current_pass = ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, msg, rtol, atol);
pass = pass & current_pass;
if(!current_pass)
{
std::cout << op_ptr->GetTypeString() << " failed" << std::endl;
}
}
else
{
#endif
pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
if(!pass)
{
std::cout << op_ptr->GetTypeString() << " failed" << std::endl;
}
#if defined CK_ENABLE_FP8
}
#endif
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a0_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b0_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", e_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",")
<< std::endl;
}
}
std::string op_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(
argument_ptr.get(),
StreamConfig{
nullptr, time_kernel, 0, n_warmup, n_iter, rotating_count > 1, rotating_count});
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 tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
if constexpr(is_same<EDataType, float>::value)
{
std::cout << "Best Perf for datatype = f32";
}
else if constexpr(is_same<EDataType, half_t>::value)
{
std::cout << "Best Perf for datatype = f16";
}
else if constexpr(is_same<EDataType, bhalf_t>::value)
{
std::cout << "Best Perf for datatype = bf16";
}
else if constexpr(is_same<EDataType, int8_t>::value)
{
std::cout << "Best Perf for datatype = int8";
}
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " ALayout = RowMajor";
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " ALayout = ColumnMajor";
}
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " BLayout = RowMajor";
}
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " BLayout = ColumnMajor";
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideE = " << StrideE << " : " << best_ave_time
<< " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck

20
profiler/include/profiler/profile_gemm_mx_impl.hpp
View File

@@ -226,6 +226,8 @@ bool profile_gemm_mx_impl(int do_verification,
return ck::type_convert<BDataType>(x);
};
using int_distr = std::uniform_int_distribution<int>;
using float_distr = std::uniform_real_distribution<float>;
switch(init_method)
{
case 0: // Initializations for development and debugging
@@ -245,21 +247,19 @@ bool profile_gemm_mx_impl(int do_verification,
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-4, 5}); // Z[-4,4]
b_k_n->GenerateTensorValue(GeneratorTensor_2<BDataType>{-4, 5}); // Z[-4,4]
a_m_k.GenerateTensorDistr(int_distr{-4, 5}); // Z[-4,4]
b_k_n->GenerateTensorDistr(int_distr{-4, 5}); // Z[-4,4]
a_m_k_scale.GenerateTensorValue(
GeneratorTensor_2<XDataType>{125, 129}); // scales: {0.25, 0.5, 1, 2}
b_k_n_scale.GenerateTensorValue(
GeneratorTensor_2<XDataType>{125, 129}); // scales: {0.25, 0.5, 1, 2}
a_m_k_scale.GenerateTensorDistr(int_distr{125, 129}); // scales: {0.25, 0.5, 1, 2}
b_k_n_scale.GenerateTensorDistr(int_distr{125, 129}); // scales: {0.25, 0.5, 1, 2}
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-2.0, 2.0});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<XDataType>{powf(2.0f, -125.0f), 1.0f});
a_m_k.GenerateTensorDistr(float_distr{-2.0, 2.0});
a_m_k_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
b_k_n->GenerateTensorValue(GeneratorTensor_3<BDataType>{-2.0, 2.0});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<XDataType>{powf(2.0f, -125.0f), 1.0f});
b_k_n->GenerateTensorDistr(float_distr{-2.0, 2.0});
b_k_n_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
break;
}