mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
[ckProfiler] Add infrastructure and instances to profile gemm_universal with B preshuffle (#2427)
* works on mi300 * fix(profiler): add error message for unsupported type/layout * refactor(preshuffle.inc): add type aliases for code readability
This commit is contained in:
Aviral Goel
@@ -0,0 +1,444 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2023-2025, 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/device_gemm_v2.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/gemm_universal_preshuffle.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 T>
|
||||
void preShuffleBuffer(const T* src, T* 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 ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
bool profile_gemm_universal_preshuffle_impl(int do_verification,
|
||||
int init_method,
|
||||
bool do_log,
|
||||
bool time_kernel,
|
||||
int M,
|
||||
int N,
|
||||
int K,
|
||||
int StrideA,
|
||||
int StrideB,
|
||||
int StrideC,
|
||||
int KBatch,
|
||||
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});
|
||||
}
|
||||
};
|
||||
|
||||
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<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<BDataType> b_preshuffled(
|
||||
f_host_tensor_descriptor(K, N, StrideB, BLayout{})); // for preshuffle
|
||||
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
|
||||
std::size_t total_gemm_needed =
|
||||
a_m_k.GetElementSpaceSizeInBytes() + b_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 << "a_m_k: " << a_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
|
||||
std::cout << "rotating count: " << rotating_count << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
|
||||
break;
|
||||
case 2:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
}
|
||||
|
||||
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
const auto a_element_op = AElementOp{};
|
||||
const auto b_element_op = BElementOp{};
|
||||
const auto c_element_op = CElementOp{};
|
||||
|
||||
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize());
|
||||
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_device_buf.ToDevice(a_m_k.mData.data());
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGemmV2BPreshuffle<ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
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)
|
||||
{
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
AccDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CElementOp,
|
||||
ComputeDataType>;
|
||||
|
||||
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_host_result, a_element_op, b_element_op, c_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
}
|
||||
|
||||
std::string best_op_name;
|
||||
std::optional<std::string> best_op_object_name;
|
||||
float best_ave_time = 0;
|
||||
float best_tflops = 0;
|
||||
float best_gb_per_sec = 0;
|
||||
float best_kbatch = 0;
|
||||
|
||||
// profile device GEMM instances
|
||||
for(auto& op_ptr : op_ptrs)
|
||||
{
|
||||
const int KPerBlock = op_ptr->GetKPerBlock();
|
||||
|
||||
if(op_ptr->GetPermuteB())
|
||||
{
|
||||
int K1 = KPerBlock;
|
||||
int K0 = K / KPerBlock;
|
||||
|
||||
// int K0, N, K1
|
||||
for(int j = 0; j < K0; j++)
|
||||
{
|
||||
for(int i = 0; i < N; i++)
|
||||
{
|
||||
for(int jj = 0; jj < K1; jj++)
|
||||
{
|
||||
b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(is_same_v<BDataType, pk_i4_t> && is_same_v<ADataType, half_t>)
|
||||
{
|
||||
// vector pk_i4x4 permute
|
||||
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_k_n_permute(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_k_n_permute(j + 0, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[6];
|
||||
int lo = input[4];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 2, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[3];
|
||||
int lo = input[1];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 4, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[7];
|
||||
int lo = input[5];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 6, i) = i4x2;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
b_k_n_permute = b_k_n;
|
||||
}
|
||||
int NPerXdl = op_ptr->GetPreShuffleParameters();
|
||||
|
||||
preShuffleBuffer<BDataType>(
|
||||
b_k_n_permute.mData.data(), b_preshuffled.mData.data(), N, K, NPerXdl);
|
||||
|
||||
b_device_buf.ToDevice(b_preshuffled.mData.data());
|
||||
|
||||
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
|
||||
|
||||
if(KBatch > 0)
|
||||
{
|
||||
kbatch_list = {KBatch};
|
||||
}
|
||||
|
||||
for(std::size_t i = 0; i < kbatch_list.size(); i++)
|
||||
{
|
||||
auto kbatch_curr = kbatch_list[i];
|
||||
|
||||
auto argument_ptr =
|
||||
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
kbatch_curr,
|
||||
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(c_m_n_device_result.mData.data());
|
||||
|
||||
#if defined CK_ENABLE_FP8
|
||||
// set softer tolerances for fp8
|
||||
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
|
||||
is_same_v<CDataType, f8_t>)
|
||||
{
|
||||
std::string msg = "Error: Incorrect results!";
|
||||
double rtol = 1e-1;
|
||||
double atol = 1e-1;
|
||||
pass = pass & ck::utils::check_err(
|
||||
c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
|
||||
}
|
||||
else
|
||||
{
|
||||
#endif
|
||||
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
|
||||
#if defined CK_ENABLE_FP8
|
||||
}
|
||||
#endif
|
||||
|
||||
if(do_log)
|
||||
{
|
||||
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
|
||||
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
|
||||
<< std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "c_device: ", c_m_n_device_result.mData, ",")
|
||||
<< std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
std::optional<std::string> op_obj_name = op_ptr->GetObjectName();
|
||||
|
||||
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;
|
||||
|
||||
static constexpr index_t BPackedSize = []() {
|
||||
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
|
||||
return 2;
|
||||
else
|
||||
return 1;
|
||||
}();
|
||||
|
||||
std::size_t num_btype = sizeof(ADataType) * M * K +
|
||||
sizeof(BDataType) * K * N / BPackedSize +
|
||||
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: " << std::setw(10) << ave_time << " ms, " << tflops
|
||||
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
|
||||
<< kbatch_curr << std::endl;
|
||||
|
||||
if(tflops > best_tflops && ave_time > 1e-10)
|
||||
{
|
||||
best_op_name = op_name;
|
||||
best_op_object_name = op_obj_name;
|
||||
best_tflops = tflops;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
best_kbatch = kbatch_curr;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_ptr->GetTypeString() << " does not support this problem"
|
||||
<< std::endl;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(is_same<CDataType, float>::value)
|
||||
{
|
||||
std::cout << "Best Perf for datatype = f32";
|
||||
}
|
||||
else if constexpr(is_same<CDataType, half_t>::value)
|
||||
{
|
||||
std::cout << "Best Perf for datatype = f16";
|
||||
}
|
||||
else if constexpr(is_same<CDataType, bhalf_t>::value)
|
||||
{
|
||||
std::cout << "Best Perf for datatype = bf16";
|
||||
}
|
||||
else if constexpr(is_same<CDataType, 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 << " StrideC = " << StrideC << " KBatch = " << best_kbatch
|
||||
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
|
||||
<< " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
if(best_op_object_name)
|
||||
std::cout << best_op_object_name.value() << std::endl;
|
||||
|
||||
return pass;
|
||||
}
|
||||
|
||||
} // namespace profiler
|
||||
} // namespace ck
|
||||
@@ -63,6 +63,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
|
||||
list(APPEND PROFILER_OPS profile_gemm_multiply_multiply_wp.cpp)
|
||||
list(APPEND PROFILER_OPS profile_gemm_ab_scale.cpp)
|
||||
list(APPEND PROFILER_OPS profile_gemm_blockscale_wp.cpp)
|
||||
list(APPEND PROFILER_OPS profile_gemm_universal_preshuffle.cpp)
|
||||
endif()
|
||||
if(SUPPORTED_GPU_TARGETS MATCHES "gfx95")
|
||||
list(APPEND PROFILER_OPS profile_gemm_mx.cpp)
|
||||
@@ -171,6 +172,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
|
||||
list(APPEND DEVICE_INSTANCES device_gemm_multiply_multiply_wp_instance)
|
||||
list(APPEND DEVICE_INSTANCES device_gemm_ab_scale_instance)
|
||||
list(APPEND DEVICE_INSTANCES device_gemm_blockscale_wp_instance)
|
||||
list(APPEND DEVICE_INSTANCES device_gemm_universal_preshuffle_instance)
|
||||
endif()
|
||||
if(SUPPORTED_GPU_TARGETS MATCHES "gfx95")
|
||||
list(APPEND DEVICE_INSTANCES device_gemm_mx_instance)
|
||||
|
||||
180
profiler/src/profile_gemm_universal_preshuffle.cpp
Normal file
180
profiler/src/profile_gemm_universal_preshuffle.cpp
Normal file
@@ -0,0 +1,180 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
#include <initializer_list>
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
|
||||
#include "profiler/profile_gemm_universal_preshuffle_impl.hpp"
|
||||
#include "profiler_operation_registry.hpp"
|
||||
|
||||
enum struct GemmMatrixLayout
|
||||
{
|
||||
// MK_KN_MN, // 0
|
||||
MK_NK_MN = 1, // 1
|
||||
// KM_KN_MN, // 2
|
||||
// KM_NK_MN, // 3
|
||||
};
|
||||
|
||||
enum struct GemmDataType
|
||||
{
|
||||
F8_F8_F16 = 0, // 0
|
||||
F8_F8_BF16 = 1, // 1
|
||||
// F32_F32_F32, // 0
|
||||
// F16_F16_F16, // 1
|
||||
// BF16_BF16_BF16, // 2
|
||||
// INT8_INT8_INT8, // 3
|
||||
// F8_F16_F16, // 4
|
||||
// F16_F8_F16, // 5
|
||||
// F16_F16_F16_F8, // 6
|
||||
// F8_F8_BF16, // 7
|
||||
// F16_I4_F16, // 8
|
||||
// BF16_I4_BF16, // 9
|
||||
|
||||
};
|
||||
|
||||
#define OP_NAME "gemm_universal_preshuffle"
|
||||
#define OP_DESC "Universal GEMM Preshuffle"
|
||||
|
||||
int profile_gemm_universal_preshuffle(int argc, char* argv[])
|
||||
{
|
||||
if(argc != 15 && argc != 18)
|
||||
{
|
||||
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
|
||||
printf("arg2: data type (0: f8->bf16, 1: f8->f16)\n");
|
||||
// printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
|
||||
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
|
||||
// printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
|
||||
// printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
|
||||
printf("arg4: verification (0: no; 1: yes)\n");
|
||||
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
|
||||
printf("arg6: print tensor value (0: no; 1: yes)\n");
|
||||
printf("arg7: time kernel (0=no, 1=yes)\n");
|
||||
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
|
||||
printf("arg14: split k into mulitiple batch\n");
|
||||
printf("optional:\n");
|
||||
printf("arg15: number of warm-up cycles (default 1)\n");
|
||||
printf("arg16: number of iterations (default 10)\n");
|
||||
printf("arg17: memory for rotating buffer (default 0, size in MB)\n");
|
||||
exit(1);
|
||||
}
|
||||
|
||||
int M;
|
||||
int N;
|
||||
int StrideA;
|
||||
int StrideB;
|
||||
// Analyze the unsupported matrix shapes, switch the M and N number
|
||||
if(std::stoi(argv[9]) % 8 != 0 && std::stoi(argv[8]) % 8 == 0)
|
||||
{
|
||||
M = std::stoi(argv[9]);
|
||||
StrideA = std::stoi(argv[12]);
|
||||
N = std::stoi(argv[8]);
|
||||
StrideB = std::stoi(argv[11]);
|
||||
}
|
||||
else
|
||||
{
|
||||
M = std::stoi(argv[8]);
|
||||
StrideA = std::stoi(argv[11]);
|
||||
N = std::stoi(argv[9]);
|
||||
StrideB = std::stoi(argv[12]);
|
||||
}
|
||||
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
|
||||
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
|
||||
const bool do_verification = std::stoi(argv[4]);
|
||||
const int init_method = std::stoi(argv[5]);
|
||||
const bool do_log = std::stoi(argv[6]);
|
||||
const bool time_kernel = std::stoi(argv[7]);
|
||||
|
||||
const int K = std::stoi(argv[10]);
|
||||
|
||||
const int StrideC = std::stoi(argv[13]);
|
||||
const int KBatch = std::stoi(argv[14]);
|
||||
|
||||
int n_warmup = 1;
|
||||
int n_iter = 10;
|
||||
uint64_t rotating = 0;
|
||||
if(argc == 18)
|
||||
{
|
||||
n_warmup = std::stoi(argv[15]);
|
||||
n_iter = std::stoi(argv[16]);
|
||||
rotating = std::stoull(argv[17]) * 1024 * 1024;
|
||||
}
|
||||
|
||||
using F32 = float;
|
||||
using F16 = ck::half_t;
|
||||
using BF16 = ck::bhalf_t;
|
||||
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94) || defined(CK_USE_WMMA_FP8)
|
||||
using F8 = ck::f8_t;
|
||||
#endif
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
auto profile = [&](auto a_type,
|
||||
auto b_type,
|
||||
auto comp_type,
|
||||
auto acc_type,
|
||||
auto c_type,
|
||||
auto a_layout,
|
||||
auto b_layout,
|
||||
auto c_layout) {
|
||||
using ADataType = decltype(a_type);
|
||||
using BDataType = decltype(b_type);
|
||||
using ComputeDataType = decltype(comp_type);
|
||||
using AccDataType = decltype(acc_type);
|
||||
using CDataType = decltype(c_type);
|
||||
|
||||
using ALayout = decltype(a_layout);
|
||||
using BLayout = decltype(b_layout);
|
||||
using CLayout = decltype(c_layout);
|
||||
|
||||
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
|
||||
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
|
||||
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
|
||||
|
||||
bool pass = ck::profiler::profile_gemm_universal_preshuffle_impl<ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
(StrideA < 0) ? DefaultStrideA : StrideA,
|
||||
(StrideB < 0) ? DefaultStrideB : StrideB,
|
||||
(StrideC < 0) ? DefaultStrideC : StrideC,
|
||||
KBatch,
|
||||
n_warmup,
|
||||
n_iter,
|
||||
rotating);
|
||||
|
||||
return pass ? 0 : 1;
|
||||
};
|
||||
|
||||
if(data_type == GemmDataType::F8_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
return profile(F8{}, F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
|
||||
}
|
||||
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94) || defined(CK_USE_WMMA_FP8)
|
||||
if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{});
|
||||
}
|
||||
#endif
|
||||
else
|
||||
{
|
||||
std::cout << "this data_type & layout is not implemented" << std::endl;
|
||||
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_universal_preshuffle);
|
||||
Reference in New Issue
Block a user