Skip to content

GitLab

"scripts/git@developer.sourcefind.cn:zhaoyu6/sglang.git" did not exist on "2c7f01bc899a9d772d77f0477116707924013c6b"
Commit 019d4b7c authored by illsilin's avatar illsilin
Browse files

merge from public repo

parents 07307ea1 5063a39f
Hide whitespace changes
Inline Side-by-side
Showing with 1080 additions and 80 deletions
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_irregular_instance.cpp 0 → 100644
View file @ 019d4b7c
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_irregular_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v2>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_irregular_instance.cpp 0 → 100644
View file @ 019d4b7c
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_irregular_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v5>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profiler/profile_gemm_b_scale_impl.hpp 0 → 100644
View file @ 019d4b7c
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, 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_xdl_cshuffle_v3_b_scale.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_b_scale.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 ADataType,
typename BDataType,
typename BScaleDataType,
typename ComputeDataType,
typename AccDataType,
typename CDataType,
index_t ScaleBlockK,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_gemm_b_scale_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});
}
};
ck::index_t Scale_Stride_BN = ck::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor>
? ((K + ScaleBlockK - 1) / ScaleBlockK)
: N;
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<BScaleDataType> b1_k_n(f_host_tensor_descriptor(
(K + ScaleBlockK - 1) / ScaleBlockK, // K direction group size is ScaleBlockK
N, // N direction group size is 1
Scale_Stride_BN,
BLayout{}));
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{}));
int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() +
b_k_n.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 << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_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});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
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});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
}
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 b1_device_buf(sizeof(BScaleDataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
using DeviceOp = ck::tensor_operation::device::DeviceGemmV2BScale<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
BScaleDataType,
CDataType,
1,
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<float> b_k_n_dequant({K, N});
float v_b = 0;
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
ck::pk_i4_t i4x2 = b_k_n(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_k_n_dequant(k, n) = ck::type_convert<float>(v_b) *
ck::type_convert<float>(b1_k_n(k / ScaleBlockK, n));
}
}
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_dequant, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
std::string best_op_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(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;
}
b_device_buf.ToDevice(b_k_n_permute.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,
Scale_Stride_BN,
static_cast<BScaleDataType*>(b1_device_buf.GetDeviceBuffer()),
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<int8_t>(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();
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_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;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_gemm_universal_impl.hpp
View file @ 019d4b7c
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -65,11 +65,13 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -65,11 +65,13 @@ bool profile_gemm_universal_impl(int do_verification,
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); 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(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); 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{})); Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes(); std::size_t total_gemm_needed =
int rotating_count = std::max( a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes();
int rotating_count = std::max(
1, 1,
std::min(n_iter, std::min(n_iter,
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed)))); static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
...@@ -86,9 +88,13 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -86,9 +88,13 @@ bool profile_gemm_universal_impl(int do_verification,
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2}); a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2}); b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
break; break;
default: case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); 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 AElementOp = ck::tensor_operation::element_wise::PassThrough;
...@@ -100,11 +106,10 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -100,11 +106,10 @@ bool profile_gemm_universal_impl(int do_verification,
const auto c_element_op = CElementOp{}; const auto c_element_op = CElementOp{};
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize()); DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.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()); DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data()); a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
using DeviceOp = ck::tensor_operation::device::DeviceGemmV2<ALayout, using DeviceOp = ck::tensor_operation::device::DeviceGemmV2<ALayout,
BLayout, BLayout,
...@@ -153,6 +158,84 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -153,6 +158,84 @@ bool profile_gemm_universal_impl(int do_verification,
// profile device GEMM instances // profile device GEMM instances
for(auto& op_ptr : op_ptrs) 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;
}
b_device_buf.ToDevice(b_k_n_permute.mData.data());
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38}; std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
if(KBatch > 0) if(KBatch > 0)
...@@ -240,7 +323,15 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -240,7 +323,15 @@ bool profile_gemm_universal_impl(int do_verification,
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + 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; sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......
profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
View file @ 019d4b7c
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -89,6 +89,7 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification, ...@@ -89,6 +89,7 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
in_device_buf.ToDevice(input.mData.data()); in_device_buf.ToDevice(input.mData.data());
out_device_buf.ToDevice(output.mData.data()); out_device_buf.ToDevice(output.mData.data());
float max_accumulated_value = 0;
if(do_verification) if(do_verification)
{ {
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial, auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
...@@ -114,6 +115,8 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification, ...@@ -114,6 +115,8 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
{}); {});
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
max_accumulated_value =
*std::max_element(weight_host_result.mData.begin(), weight_host_result.mData.end());
} }
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial,
...@@ -237,7 +240,39 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification, ...@@ -237,7 +240,39 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
{ {
wei_device_buf.FromDevice(weight_device_result.mData.data()); wei_device_buf.FromDevice(weight_device_result.mData.data());
bool pass = ck::utils::check_err(weight_device_result, weight_host_result); using ComputeType =
std::conditional_t<sizeof(ComputeTypeA) < sizeof(ComputeTypeB),
ComputeTypeA,
ComputeTypeB>;
using AccDataType =
std::conditional_t<std::is_same_v<ComputeType, int8_t>, int32_t, float>;
const index_t num_accums = output.GetElementSize() / conv_param.K_;
const index_t num_accums_split_k = split_k_list[split_k_id];
// Calculate thresholds
auto rtol =
ck::utils::get_relative_threshold<ComputeType, WeiDataType, AccDataType>(
num_accums / num_accums_split_k);
auto atol =
ck::utils::get_absolute_threshold<ComputeType, WeiDataType, AccDataType>(
max_accumulated_value / num_accums_split_k,
num_accums / num_accums_split_k);
// Calculate error due to split_k accumulation
auto rtol_split_k =
ck::utils::get_relative_threshold<WeiDataType, WeiDataType, WeiDataType>(
num_accums_split_k);
auto atol_split_k =
ck::utils::get_absolute_threshold<WeiDataType, WeiDataType, WeiDataType>(
max_accumulated_value, num_accums_split_k);
// Use higher threshold
rtol = std::max(rtol, rtol_split_k);
atol = std::max(atol, atol_split_k);
bool pass = ck::utils::check_err(weight_device_result,
weight_host_result,
"Error: Incorrect results!",
rtol,
atol);
std::cout << "Relative error threshold: " << rtol
<< " Absolute error threshold: " << atol << std::endl;
if(!pass) if(!pass)
{ {
......
profiler/src/CMakeLists.txt
View file @ 019d4b7c
...@@ -58,6 +58,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") ...@@ -58,6 +58,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
...@@ -141,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") ...@@ -141,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
endif() endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
...@@ -177,5 +179,4 @@ if(DL_KERNELS) ...@@ -177,5 +179,4 @@ if(DL_KERNELS)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif() endif()
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
profiler/src/profile_gemm_b_scale.cpp 0 → 100644
View file @ 019d4b7c
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include "profiler/profile_gemm_b_scale_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
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
};
enum struct BScaleBlockTile
{
K_64, // 0
K_128, // 1
};
#define OP_NAME "gemm_b_scale"
#define OP_DESC "Int4-dequant GEMM"
int profile_gemm_b_scale(int argc, char* argv[])
{
if(argc != 16 && argc != 19)
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
"f16->f8; 7: f8->bf16, "
"comp f8; 8: f16@i4)\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: B scale block tile (0: 64, 1: 128):\n");
printf("arg5: verification (0: no; 1: yes)\n");
printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg7: print tensor value (0: no; 1: yes)\n");
printf("arg8: time kernel (0=no, 1=yes)\n");
printf("arg9 to 14: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg15: split k into mulitiple batch\n");
printf("optional:\n");
printf("arg16: number of warm-up cycles (default 1)\n");
printf("arg17: number of iterations (default 10)\n");
printf("arg18: memory for rotating buffer (default 0, size in MB)\n");
exit(1);
}
printf("Start profiling\n");
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const auto B_scale_block = static_cast<BScaleBlockTile>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const int init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const int M = std::stoi(argv[9]);
const int N = std::stoi(argv[10]);
const int K = std::stoi(argv[11]);
const int StrideA = std::stoi(argv[12]);
const int StrideB = std::stoi(argv[13]);
const int StrideC = std::stoi(argv[14]);
const int KBatch = std::stoi(argv[15]);
printf("M:%d, N:%d, K:%d, StrideA:%d, StrideB:%d, StrideC:%d, KBatch:%d\n",
M,
N,
K,
StrideA,
StrideB,
StrideC,
KBatch);
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
if(argc == 19)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
rotating = std::stoull(argv[18]) * 1024 * 1024;
printf("n_warmup:%d, n_iter:%d, rotating:%lu\n", n_warmup, n_iter, rotating);
}
using F32 = float;
using F16 = ck::half_t;
using I4 = ck::pk_i4_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto b_scale_type,
auto comp_type,
auto acc_type,
auto c_type,
auto scale_block_k,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using BScaleDataType = decltype(b_scale_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_b_scale_impl<ADataType,
BDataType,
BScaleDataType,
ComputeDataType,
AccDataType,
CDataType,
scale_block_k,
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::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN &&
B_scale_block == BScaleBlockTile::K_128)
{
printf("F16_I4_F16 MK_NK_MN K_128\n");
return profile(
F16{}, I4{}, F16{}, F16{}, F32{}, F16{}, ck::Number<128>{}, Row{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_b_scale);
profiler/src/profile_gemm_multiply_multiply.cpp
View file @ 019d4b7c
...@@ -28,6 +28,7 @@ enum struct GemmDataType ...@@ -28,6 +28,7 @@ enum struct GemmDataType
F16_F16_F16_F8, // 6 F16_F16_F16_F8, // 6
F8_F8_BF16, // 7 F8_F8_BF16, // 7
INT8_INT8_BF16, // 8 INT8_INT8_BF16, // 8
F8_F8_F16, // 9
}; };
#define OP_NAME "gemm_multiply_multiply" #define OP_NAME "gemm_multiply_multiply"
...@@ -40,7 +41,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[]) ...@@ -40,7 +41,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: " printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
"f16->f8; 7: f8->bf16, " "f16->f8; 7: f8->bf16, "
"comp f8; 8: int8->bf16)\n"); "comp f8; 8: int8->bf16; 9: f8->f16, comp f8;)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\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(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n"); printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
...@@ -89,6 +90,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[]) ...@@ -89,6 +90,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
using F32 = float; using F32 = float;
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F8 = ck::f8_t; using F8 = ck::f8_t;
using I8 = int8_t; using I8 = int8_t;
using I32 = int; using I32 = int;
...@@ -165,6 +167,11 @@ int profile_gemm_multiply_multiply(int argc, char* argv[]) ...@@ -165,6 +167,11 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
return profile( return profile(
F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{}); F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
} }
else if(data_type == GemmDataType::F8_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(
F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, F16{}, Row{}, Col{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::INT8_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
return profile( return profile(
......
profiler/src/profile_gemm_universal.cpp
View file @ 019d4b7c
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_universal_impl.hpp" #include "profiler/profile_gemm_universal_impl.hpp"
#include "profiler_operation_registry.hpp" #include "profiler_operation_registry.hpp"
...@@ -27,6 +27,8 @@ enum struct GemmDataType ...@@ -27,6 +27,8 @@ enum struct GemmDataType
F16_F8_F16, // 5 F16_F8_F16, // 5
F16_F16_F16_F8, // 6 F16_F16_F16_F8, // 6
F8_F8_BF16, // 7 F8_F8_BF16, // 7
F16_I4_F16, // 8
BF16_I4_BF16, // 9
}; };
#define OP_NAME "gemm_universal" #define OP_NAME "gemm_universal"
...@@ -39,7 +41,7 @@ int profile_gemm_universal(int argc, char* argv[]) ...@@ -39,7 +41,7 @@ int profile_gemm_universal(int argc, char* argv[])
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: " printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
"f16->f8; 7: f8->bf16, " "f16->f8; 7: f8->bf16, "
"comp f8)\n"); "comp f8; 8: f16@i4; 9: bf16@i4\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\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(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n"); printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
...@@ -103,6 +105,7 @@ int profile_gemm_universal(int argc, char* argv[]) ...@@ -103,6 +105,7 @@ int profile_gemm_universal(int argc, char* argv[])
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94) #if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t; using F8 = ck::f8_t;
using I4 = ck::pk_i4_t;
#endif #endif
using Row = ck::tensor_layout::gemm::RowMajor; using Row = ck::tensor_layout::gemm::RowMajor;
...@@ -207,6 +210,14 @@ int profile_gemm_universal(int argc, char* argv[]) ...@@ -207,6 +210,14 @@ int profile_gemm_universal(int argc, char* argv[])
{ {
return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{}); return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{});
} }
else if(data_type == GemmDataType::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, I4{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::BF16_I4_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(BF16{}, I4{}, BF16{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
#endif #endif
else else
{ {
......
profiler/src/profile_gemm_universal_streamk.cpp 100755 → 100644
View file @ 019d4b7c
...@@ -83,8 +83,9 @@ int profile_gemm_universal_streamk(int argc, char* argv[]) ...@@ -83,8 +83,9 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
rotating = std::stoull(argv[18]) * 1024 * 1024; rotating = std::stoull(argv[18]) * 1024 * 1024;
} }
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94) #if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t; using F8 = ck::f8_t;
...@@ -165,6 +166,22 @@ int profile_gemm_universal_streamk(int argc, char* argv[]) ...@@ -165,6 +166,22 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}); return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
} }
#endif #endif
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, Col{}, Row{});
}
else else
{ {
std::cout << "this data_type & layout is not implemented" << std::endl; std::cout << "this data_type & layout is not implemented" << std::endl;
......
pyproject.toml
View file @ 019d4b7c
...@@ -21,16 +21,19 @@ dependencies = [] ...@@ -21,16 +21,19 @@ dependencies = []
"Bug Tracker" = "https://github.com/rocm/composable_kernel/issues" "Bug Tracker" = "https://github.com/rocm/composable_kernel/issues"
[tool.setuptools] [tool.setuptools]
packages = ["ck4inductor", "ck4inductor.include", "ck4inductor.library"] packages = ["ck4inductor", "ck4inductor.include", "ck4inductor.library", "ck4inductor.universal_gemm", "ck4inductor.batched_universal_gemm", "ck4inductor.grouped_conv_fwd"]
[tool.setuptools.package-dir] [tool.setuptools.package-dir]
ck4inductor = "python/ck4inductor" ck4inductor = "python/ck4inductor"
"ck4inductor.universal_gemm" = "python/ck4inductor/universal_gemm"
"ck4inductor.batched_universal_gemm" = "python/ck4inductor/batched_universal_gemm"
"ck4inductor.grouped_conv_fwd" = "python/ck4inductor/grouped_conv_fwd"
"ck4inductor.include" = "include" "ck4inductor.include" = "include"
"ck4inductor.library" = "library" "ck4inductor.library" = "library"
[tool.setuptools.package-data] [tool.setuptools.package-data]
"ck4inductor.include" = ["ck/**/*.hpp"] "ck4inductor.include" = ["ck/**/*.hpp"]
"ck4inductor.library" = ["src/tensor_operation_instance/gpu/gemm_universal/**/*.hpp"] "ck4inductor.library" = ["src/tensor_operation_instance/gpu/gemm_universal/**/*.hpp", "src/tensor_operation_instance/gpu/gemm_universal_batched/**/*.hpp", "include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/**/*.hpp"]
[tool.setuptools.dynamic] [tool.setuptools.dynamic]
version = { attr = "setuptools_scm.get_version" } version = { attr = "setuptools_scm.get_version" }
python/ck4inductor/universal_gemm/gen_instances.py
View file @ 019d4b7c
...@@ -68,12 +68,13 @@ def parse_instances(str_instances: List[str]) -> List[CKGemmOperation]: ...@@ -68,12 +68,13 @@ def parse_instances(str_instances: List[str]) -> List[CKGemmOperation]:
template_args.insert(2, tuple()) # ds layout template_args.insert(2, tuple()) # ds layout
template_args.insert(6, tuple()) # ds dtype template_args.insert(6, tuple()) # ds dtype
try:
new_instance = CKGemmOperation( new_instance = CKGemmOperation(
*template_args, # type: ignore[arg-type] *template_args, # type: ignore[arg-type]
) )
op_instances.append(new_instance)
op_instances.append(new_instance) except TypeError as e:
log.debug(f"{e} when parsing {line}")
return op_instances return op_instances
......
python/test/test_gen_instances.py 0 → 100644
View file @ 019d4b7c
# SPDX-License-Identifier: MIT
# Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
import logging
import unittest
from ck4inductor.universal_gemm.gen_instances import (
gen_ops_library as gen_gemm_ops_library,
)
from ck4inductor.universal_gemm.gen_instances import (
gen_ops_preselected as gen_gemm_ops_preselected,
)
from ck4inductor.grouped_conv_fwd.gen_instances import (
gen_conv_ops_library as gen_conv_ops_library,
)
from ck4inductor.batched_universal_gemm.gen_instances import (
gen_ops_library as gen_batched_gemm_ops_library,
)
log = logging.getLogger(__name__)
class TestGenInstances(unittest.TestCase):
def test_gen_gemm_instances(self):
instances = gen_gemm_ops_library()
log.debug("%d gemm instances from library" % len(instances))
self.assertTrue(instances)
def test_preselected_gemm_instances(self):
instances = gen_gemm_ops_preselected()
log.debug("%d preselected gemm instances" % len(instances))
self.assertTrue(instances)
def test_gen_conv_instances(self):
instances = gen_conv_ops_library()
log.debug("%d gemm instances from library" % len(instances))
self.assertTrue(instances)
def test_gen_batched_gemm_instances(self):
instances = gen_batched_gemm_ops_library()
log.debug("%d gemm instances from library" % len(instances))
self.assertTrue(instances)
script/cmake-ck-dev.sh
View file @ 019d4b7c
...@@ -15,9 +15,9 @@ else ...@@ -15,9 +15,9 @@ else
fi fi
cmake \ cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \ -D CMAKE_PREFIX_PATH=/opt/rocm/ \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \ -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \ -D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \
-D CMAKE_BUILD_TYPE=Release \ -D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=ON \ -D BUILD_DEV=ON \
-D GPU_TARGETS=$GPU_TARGETS \ -D GPU_TARGETS=$GPU_TARGETS \
......
script/process_perf_data.py
View file @ 019d4b7c
...@@ -149,6 +149,12 @@ def parse_logfile(logfile): ...@@ -149,6 +149,12 @@ def parse_logfile(logfile):
lst=line.split() lst=line.split()
line_dict=dict(zip(lst[1:],lst)) line_dict=dict(zip(lst[1:],lst))
res.append(line_dict['TFlops,']) res.append(line_dict['TFlops,'])
elif 'perf_tile_gemm_basic' in logfile or 'perf_tile_gemm_mem_pipeline' in logfile:
for line in open(logfile):
if 'TFlops' in line:
lst=line.split()
line_dict=dict(zip(lst[1:],lst))
res.append(line_dict['TFlops,'])
return res return res
...@@ -330,6 +336,14 @@ def main(): ...@@ -330,6 +336,14 @@ def main():
for i in range(1,len(results)+1): for i in range(1,len(results)+1):
testlist.append("Test%i"%i) testlist.append("Test%i"%i)
table_name="ck_fmha_bwd_tflops" table_name="ck_fmha_bwd_tflops"
if 'gemm_basic_fp16' in filename:
for i in range(1, len(results)+1):
testlist.append("Test%i"%i)
table_name="ck_tile_gemm_basic_fp16_tflops"
if 'gemm_mem_pipeline_fp16' in filename:
for i in range(1, len(results)+1):
testlist.append("Test%i"%i)
table_name="ck_tile_gemm_mem_pipeline_fp16_tflops"
tflops_base = get_baseline(table_name,conn) tflops_base = get_baseline(table_name,conn)
store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, sqlEngine) store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, sqlEngine)
......
script/process_perf_data.sh
View file @ 019d4b7c
...@@ -43,3 +43,19 @@ file=./perf_fmha_bwd_gfx90a.log ...@@ -43,3 +43,19 @@ file=./perf_fmha_bwd_gfx90a.log
if [ -e "$file" ]; then if [ -e "$file" ]; then
python3 process_perf_data.py perf_fmha_bwd_gfx90a.log python3 process_perf_data.py perf_fmha_bwd_gfx90a.log
fi fi
file=./perf_tile_gemm_basic_fp16_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_tile_gemm_basic_fp16_gfx942.log
fi
file=./perf_tile_gemm_basic_fp16_gfx90a.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_tile_gemm_basic_fp16_gfx90a.log
fi
file=./perf_tile_gemm_mem_pipeline_fp16_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_tile_gemm_mem_pipeline_fp16_gfx942.log
fi
file=./perf_tile_gemm_mem_pipeline_fp16_gfx90a.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_tile_gemm_mem_pipeline_fp16_gfx90a.log
fi
script/process_qa_data.sh
View file @ 019d4b7c
...@@ -52,3 +52,19 @@ file=./perf_fmha_bwd_gfx90a.log ...@@ -52,3 +52,19 @@ file=./perf_fmha_bwd_gfx90a.log
if [ -e "$file" ]; then if [ -e "$file" ]; then
python3 process_perf_data.py perf_fmha_bwd_gfx90a.log python3 process_perf_data.py perf_fmha_bwd_gfx90a.log
fi fi
file=./perf_gemm_basic_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_gemm_basic_gfx942.log
fi
file=./perf_gemm_basic_gfx90a.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_gemm_basic_gfx90a.log
fi
file=./perf_gemm_mem_pipeline_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_gemm_mem_pipeline_gfx942.log
fi
file=./perf_gemm_mem_pipeline_gfx90a.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_gemm_mem_pipeline_gfx90a.log
fi
test/CMakeLists.txt
View file @ 019d4b7c
...@@ -7,6 +7,34 @@ include(gtest) ...@@ -7,6 +7,34 @@ include(gtest)
add_custom_target(tests) add_custom_target(tests)
# list of tests that are labelled as REGRESSION_TEST for make regression (runtime more than 30 seconds)
# all other tests are labelled as SMOKE_TEST
set(REGRESSION_TESTS
test_gemm_standalone_xdl_fp16
test_gemm_fp16
test_gemm_splitk
test_batched_gemm
test_gemm_universal
test_batched_gemm_softmax_gemm_fp16
test_batched_gemm_softmax_gemm_permute_fp16
test_batched_gemm_bias_softmax_gemm_permute_fp16
test_batched_gemm_softmax_gemm_permute_bf16
test_batched_gemm_bias_softmax_gemm_permute_bf16
test_grouped_gemm_splitk
test_reduce_no_index
test_reduce_with_index
test_convnd_fwd
test_convnd_bwd_data
test_grouped_convnd_fwd
test_grouped_convnd_bwd_weight
test_softmax_rank3
test_softmax_rank4
test_batchnorm_fwd_rank_4
test_batchnorm_bwd_rank_4
test_grouped_convnd_bwd_data_xdl
test_conv_tensor_rearrange
)
function(add_test_executable TEST_NAME) function(add_test_executable TEST_NAME)
message("adding test ${TEST_NAME}") message("adding test ${TEST_NAME}")
set(result 1) set(result 1)
...@@ -43,6 +71,12 @@ function(add_test_executable TEST_NAME) ...@@ -43,6 +71,12 @@ function(add_test_executable TEST_NAME)
set(TEST_TARGETS ${SUPPORTED_GPU_TARGETS}) set(TEST_TARGETS ${SUPPORTED_GPU_TARGETS})
foreach(source IN LISTS ARGN)
if(NOT DEFINED DPP_KERNELS AND source MATCHES "_dpp")
message("removing dpp test ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
foreach(source IN LISTS ARGN) foreach(source IN LISTS ARGN)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl") if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
message("removing dl test ${source} ") message("removing dl test ${source} ")
...@@ -82,6 +116,15 @@ function(add_test_executable TEST_NAME) ...@@ -82,6 +116,15 @@ function(add_test_executable TEST_NAME)
endif() endif()
#message("add_test returns ${result}") #message("add_test returns ${result}")
set(result ${result} PARENT_SCOPE) set(result ${result} PARENT_SCOPE)
if(result EQUAL 0 AND NOT "${TEST_NAME}" IN_LIST REGRESSION_TESTS)
message("adding to SMOKE TEST FILTER ${TEST_NAME}")
set_tests_properties(${TEST_NAME} PROPERTIES LABELS "SMOKE_TEST")
add_dependencies(smoke ${TEST_NAME})
elseif(result EQUAL 0 AND "${TEST_NAME}" IN_LIST REGRESSION_TESTS)
message("Adding to REGRESSION TEST FILTER ${TEST_NAME}")
set_tests_properties(${TEST_NAME} PROPERTIES LABELS "REGRESSION_TEST")
add_dependencies(regression ${TEST_NAME})
endif()
endfunction() endfunction()
function(add_gtest_executable TEST_NAME) function(add_gtest_executable TEST_NAME)
...@@ -162,6 +205,15 @@ function(add_gtest_executable TEST_NAME) ...@@ -162,6 +205,15 @@ function(add_gtest_executable TEST_NAME)
endif() endif()
#message("add_gtest returns ${result}") #message("add_gtest returns ${result}")
set(result ${result} PARENT_SCOPE) set(result ${result} PARENT_SCOPE)
if(result EQUAL 0 AND NOT "${TEST_NAME}" IN_LIST REGRESSION_TESTS)
#message("adding to smoke test FILTER ${TEST_NAME}")
set_tests_properties(${TEST_NAME} PROPERTIES LABELS "SMOKE_TEST")
add_dependencies(smoke ${TEST_NAME})
elseif(result EQUAL 0 AND "${TEST_NAME}" IN_LIST REGRESSION_TESTS)
#message("Adding to REGRESSION TEST FILTER ${TEST_NAME}")
set_tests_properties(${TEST_NAME} PROPERTIES LABELS "REGRESSION_TEST")
add_dependencies(regression ${TEST_NAME})
endif()
endfunction() endfunction()
add_compile_options(-Wno-c++20-extensions) add_compile_options(-Wno-c++20-extensions)
......
test/ck_tile/batched_gemm/test_batched_gemm_util.hpp
View file @ 019d4b7c
...@@ -24,12 +24,9 @@ class TestCkTileBatchedGemm : public ::testing::Test ...@@ -24,12 +24,9 @@ class TestCkTileBatchedGemm : public ::testing::Test
using AccDataType = std::tuple_element_t<5, Tuple>; using AccDataType = std::tuple_element_t<5, Tuple>;
using CDataType = std::tuple_element_t<6, Tuple>; using CDataType = std::tuple_element_t<6, Tuple>;
struct batched_gemm_kargs : public ck_tile::BatchedGemmHostArgs
{
};
template <typename ALayout, typename BLayout, typename CLayout> template <typename ALayout, typename BLayout, typename CLayout>
void invoke_batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config& s) void invoke_batched_gemm(const ck_tile::BatchedGemmHostArgs& args,
const ck_tile::stream_config& s)
{ {
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part. // The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false; constexpr bool kPadM = false;
...@@ -94,9 +91,9 @@ class TestCkTileBatchedGemm : public ::testing::Test ...@@ -94,9 +91,9 @@ class TestCkTileBatchedGemm : public ::testing::Test
using Kernel = using Kernel =
ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>; ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args); auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args); const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch, args.batch_count);
constexpr dim3 blocks = Kernel::BlockSize(); constexpr dim3 blocks = Kernel::BlockSize();
if(s.log_level_ > 0) if(s.log_level_ > 0)
...@@ -185,21 +182,23 @@ class TestCkTileBatchedGemm : public ::testing::Test ...@@ -185,21 +182,23 @@ class TestCkTileBatchedGemm : public ::testing::Test
c_m_n_dev_buf.SetZero(); c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero(); c_m_n_dev_result.SetZero();
batched_gemm_kargs kargs{a_m_k_dev_buf.GetDeviceBuffer(), ck_tile::BatchedGemmHostArgs args;
b_k_n_dev_buf.GetDeviceBuffer(), args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
c_m_n_dev_buf.GetDeviceBuffer(), args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
M, args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
N, args.k_batch = 1;
K, args.M = M;
StrideA, args.N = N;
StrideB, args.K = K;
StrideC, args.stride_A = StrideA;
BatchStrideA, args.stride_B = StrideB;
BatchStrideB, args.stride_C = StrideC;
BatchStrideC, args.batch_stride_A = BatchStrideA;
BatchCount}; args.batch_stride_B = BatchStrideB;
args.batch_stride_C = BatchStrideC;
invoke_batched_gemm<ALayout, BLayout, CLayout>(kargs, args.batch_count = BatchCount;
invoke_batched_gemm<ALayout, BLayout, CLayout>(args,
ck_tile::stream_config{nullptr, false}); ck_tile::stream_config{nullptr, false});
std::cout << "Run kernel with M =" << M << " N =" << N << " K =" << K std::cout << "Run kernel with M =" << M << " N =" << N << " K =" << K
......
test/ck_tile/gemm/test_gemm_pipeline_util.hpp
View file @ 019d4b7c
...@@ -31,22 +31,8 @@ class TestCkTileGemmPipeline : public ::testing::Test ...@@ -31,22 +31,8 @@ class TestCkTileGemmPipeline : public ::testing::Test
static constexpr auto PipelineType = std::tuple_element_t<8, Tuple>::value; static constexpr auto PipelineType = std::tuple_element_t<8, Tuple>::value;
// TODO: expose tile size through test t-param ? // TODO: expose tile size through test t-param ?
struct gemm_args
{
const void* p_a;
const void* p_b;
void* p_c;
ck_tile::index_t kbatch;
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;
};
template <bool PadM, bool PadN, bool PadK> template <bool PadM, bool PadN, bool PadK>
void invoke_gemm(const gemm_args& args, const ck_tile::stream_config& s) void invoke_gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{ {
// TODO: This should be parameterized in tests // TODO: This should be parameterized in tests
constexpr ck_tile::index_t M_Tile = 128; constexpr ck_tile::index_t M_Tile = 128;
...@@ -88,7 +74,9 @@ class TestCkTileGemmPipeline : public ::testing::Test ...@@ -88,7 +74,9 @@ class TestCkTileGemmPipeline : public ::testing::Test
ck_tile:: ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>>; GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>>;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(args.K); const ck_tile::index_t k_grain = args.k_batch * K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop); const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop); const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
...@@ -117,17 +105,9 @@ class TestCkTileGemmPipeline : public ::testing::Test ...@@ -117,17 +105,9 @@ class TestCkTileGemmPipeline : public ::testing::Test
has_hot_loop_v, has_hot_loop_v,
tail_number_v>>>; tail_number_v>>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>; using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args.p_a, auto kargs = Kernel::MakeKernelArgs(args);
args.p_b,
args.p_c, const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
args.M,
args.N,
args.K,
args.stride_A,
args.stride_B,
args.stride_C);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize(); constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs)) if(!Kernel::IsSupportedArgument(kargs))
...@@ -319,11 +299,11 @@ class TestCkTileGemmPipeline : public ::testing::Test ...@@ -319,11 +299,11 @@ class TestCkTileGemmPipeline : public ::testing::Test
c_m_n_dev_buf.SetZero(); c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero(); c_m_n_dev_result.SetZero();
gemm_args args; ck_tile::GemmHostArgs args;
args.p_a = a_m_k_dev_buf.GetDeviceBuffer(); args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
args.p_b = b_k_n_dev_buf.GetDeviceBuffer(); args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
args.p_c = c_m_n_dev_buf.GetDeviceBuffer(); args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
args.kbatch = kbatch; args.k_batch = kbatch;
args.M = M; args.M = M;
args.N = N; args.N = N;
args.K = K; args.K = K;
......
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