mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-12 01:10:17 +00:00
abf4bdb9a9
* Convolution ND * Code unification across dimensions for generating tensor descriptors. * Example * Instances * Move convnd f32 instance file to comply with repo structure. * Conv 1D tensor layouts. * Formatting and use ReferenceConv * Reference ConvFwd supporting 1D and 2D convolution. * Debug printing TensorLayout name. * Conv fwd 1D instance f32 * Refactor conv ND example. Needed to support various conv dimensio. Needed to support various conv dimensions * Rename conv nd example director to prevent conflicts. * Refactor some common utility to single file. Plus some tests. * Refactor GetHostTensorDescriptor + UT. * Add 1D test case. * Test reference convolution 1d/2d * Remove some leftovers. * Fix convolution example error for 1D * Refactor test check errors utility function. * Test Conv2D Fwd XDL * More UT for 1D case. * Parameterize input & weight initializers. * Rename example to prevent conflicts. * Split convnd instance into separate files for 1d/2d * Address review comments. * Fix data type for flops/gbytes calculations. * Assign example number 11. * 3D cases for convolution utility functions. * 3D reference convolution. * Add support for 3D convolution. * Check for inputs bigger than 2GB. * Formatting * Support for bf16/f16/f32/i8 - conv instances + UT. * Use check_err from test_util.hpp. * Split convnd test into separate files for each dim. * Fix data generation and use proper instances. * Formatting * Skip tensor initialization if not necessary. * Fix CMakefiles. * Remove redundant conv2d_fwd test. * Lower problem size for conv3D UT. * 3D case for convnd example. * Remove leftovers after merge. * Add Conv Specialization string to GetTypeString * Skip instance causing numerical errors. * Small fixes. * Remove redundant includes. * Fix namespace name error. * Script for automatic testing and logging convolution fwd UTs * Comment out numactl cmd. * Refine weights initalization and relax rtol for fp16 * Move test_util.hpp to check_err.hpp * Refine weights initalization and relax rtol for fp16 * Refactor common part of test conv utils. * Move utility function to single common place. * Add additional common functions to utility. * Refactor convnd_fwd_xdl examples. * Remove redundant files. * Unify structure. * Add constructor to ConvParams. * And add input parameters validation. * Modify conv examples to use single utility file. * Remove check_error from host_tensor.hpp * Get rid of check_indices function. * Remove bf16_to_f32 function overload for scalars. * Fix namespace. * Add half_float::half for check_err. * Fix conv params size in UT. * Fix weights initialization for int8. * Fix weights initialization for int8. * Add type_convert when store output in ref conv 1D. * Get back old conv2d_fwd_xdl operation. * Silence conv debug print. * format * clean * clean * Fix merge. * Fix namespace for check_err * Formatting. * Fix merge artifacts. * Remove deleted header. * Fix some includes and use ck::utils::check_err. * Remove unused check_indices restored by previous merge. * Fix namespaces after merge. * Fix compilation error. * Small fixes. * Use common functions. * Fix filename * Fix namespaces. * Fix merge artifact - retrieve removed by accident fun. * Fix ConvForwardSpecialization. * Adhere to coding style rules. * Fix merge artifacts. Co-authored-by: Adam Osewski <aosewski@amd.com> Co-authored-by: Chao Liu <chao.liu2@amd.com>
535 lines
24 KiB
C++
535 lines
24 KiB
C++
#pragma once
|
|
#include <iomanip>
|
|
|
|
#include "check_err.hpp"
|
|
#include "config.hpp"
|
|
#include "device.hpp"
|
|
#include "host_tensor.hpp"
|
|
#include "host_tensor_generator.hpp"
|
|
#include "host_conv.hpp"
|
|
#include "tensor_layout.hpp"
|
|
#include "device_tensor.hpp"
|
|
#include "element_wise_operation.hpp"
|
|
#include "device_gemm.hpp"
|
|
#include "reference_gemm.hpp"
|
|
|
|
namespace ck {
|
|
namespace tensor_operation {
|
|
namespace device {
|
|
namespace device_gemm_instance {
|
|
|
|
using DeviceGemmNoOpPtr =
|
|
ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
|
|
ck::tensor_operation::element_wise::PassThrough,
|
|
ck::tensor_operation::element_wise::PassThrough>;
|
|
|
|
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
|
|
std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
|
|
|
|
} // namespace device_gemm_instance
|
|
} // namespace device
|
|
} // namespace tensor_operation
|
|
} // namespace ck
|
|
|
|
namespace ck {
|
|
namespace profiler {
|
|
|
|
template <typename ADataType,
|
|
typename BDataType,
|
|
typename CDataType,
|
|
typename ALayout,
|
|
typename BLayout,
|
|
typename CLayout>
|
|
void profile_gemm_impl(int do_verification,
|
|
int init_method,
|
|
bool do_log,
|
|
int nrepeat,
|
|
int M,
|
|
int N,
|
|
int K,
|
|
int StrideA,
|
|
int StrideB,
|
|
int StrideC,
|
|
int KBatch)
|
|
{
|
|
auto f_host_tensor_descriptor =
|
|
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
|
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
|
std::vector<std::size_t>({stride, 1}));
|
|
}
|
|
else
|
|
{
|
|
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
|
std::vector<std::size_t>({1, 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<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
|
|
|
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::size_t num_thread = 1;
|
|
switch(init_method)
|
|
{
|
|
case 0: break;
|
|
case 1:
|
|
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
|
|
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
|
|
break;
|
|
default:
|
|
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
|
|
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
|
|
}
|
|
|
|
// set zero to c_device_buf
|
|
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
|
|
|
|
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.GetElementSpace());
|
|
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
|
|
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
|
|
|
|
a_device_buf.ToDevice(a_m_k.mData.data());
|
|
b_device_buf.ToDevice(b_k_n.mData.data());
|
|
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
|
|
|
|
// add device GEMM instances
|
|
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmNoOpPtr> gemm_ptrs;
|
|
|
|
if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
|
|
is_same<CDataType, float>::value)
|
|
{
|
|
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
}
|
|
else if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
|
|
is_same<CDataType, half_t>::value)
|
|
{
|
|
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
if(KBatch > 1)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
else
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
|
|
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
}
|
|
else if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
|
|
is_same<BDataType, ck::bhalf_t>::value &&
|
|
is_same<CDataType, ck::bhalf_t>::value)
|
|
{
|
|
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
|
|
is_same<CDataType, int8_t>::value)
|
|
{
|
|
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(gemm_ptrs);
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
|
|
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
ck::tensor_operation::device::device_gemm_instance::
|
|
add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances(gemm_ptrs);
|
|
}
|
|
}
|
|
|
|
if(gemm_ptrs.size() <= 0)
|
|
{
|
|
throw std::runtime_error("wrong! no device GEMM instance found");
|
|
}
|
|
|
|
std::string best_gemm_name;
|
|
float best_ave_time = 0;
|
|
float best_tflops = 0;
|
|
float best_gb_per_sec = 0;
|
|
|
|
// profile device GEMM instances
|
|
for(auto& gemm_ptr : gemm_ptrs)
|
|
{
|
|
auto argument_ptr =
|
|
gemm_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,
|
|
ck::tensor_operation::element_wise::PassThrough{},
|
|
ck::tensor_operation::element_wise::PassThrough{},
|
|
ck::tensor_operation::element_wise::PassThrough{},
|
|
KBatch);
|
|
|
|
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
|
|
|
|
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
|
|
{
|
|
// re-init C to zero before profiling next kernel
|
|
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
|
|
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
|
|
|
|
std::string gemm_name = gemm_ptr->GetTypeString();
|
|
|
|
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
|
|
|
|
std::size_t flop = std::size_t(2) * M * N * K;
|
|
|
|
std::size_t num_btype =
|
|
sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + 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, " << gemm_name << std::endl;
|
|
|
|
if(tflops > best_tflops)
|
|
{
|
|
best_gemm_name = gemm_name;
|
|
best_tflops = tflops;
|
|
best_ave_time = ave_time;
|
|
best_gb_per_sec = gb_per_sec;
|
|
}
|
|
|
|
if(do_verification)
|
|
{
|
|
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
|
|
|
if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
|
|
is_same<BDataType, ck::bhalf_t>::value &&
|
|
is_same<CDataType, ck::bhalf_t>::value)
|
|
{
|
|
Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
|
Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
|
Tensor<float> c_m_n_host_result(
|
|
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
|
Tensor<float> c_m_n_device_f32_result(
|
|
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
|
|
|
bf16_to_f32_(a_m_k, a_f32_m_k);
|
|
bf16_to_f32_(b_k_n, b_f32_k_n);
|
|
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
|
|
|
|
using ReferenceGemmInstance = ck::tensor_operation::host::
|
|
ReferenceGemm<float, float, float, AElementOp, BElementOp, CElementOp>;
|
|
|
|
auto ref_gemm = ReferenceGemmInstance{};
|
|
auto ref_invoker = ref_gemm.MakeInvoker();
|
|
|
|
auto ref_argument = ref_gemm.MakeArgument(a_f32_m_k,
|
|
b_f32_k_n,
|
|
c_m_n_host_result,
|
|
a_element_op,
|
|
b_element_op,
|
|
c_element_op);
|
|
|
|
ref_invoker.Run(ref_argument);
|
|
|
|
ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData);
|
|
|
|
if(do_log)
|
|
{
|
|
LogRangeAsType<float>(
|
|
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
|
|
<< std::endl;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
Tensor<CDataType> c_m_n_host_result(
|
|
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
|
|
|
using ReferenceGemmInstance =
|
|
ck::tensor_operation::host::ReferenceGemm<ADataType,
|
|
BDataType,
|
|
CDataType,
|
|
AElementOp,
|
|
BElementOp,
|
|
CElementOp>;
|
|
|
|
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);
|
|
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
|
|
|
|
if(do_log)
|
|
{
|
|
LogRangeAsType<float>(
|
|
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
|
|
<< std::endl;
|
|
}
|
|
}
|
|
|
|
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_device: ", c_m_n_device_result.mData, ",")
|
|
<< std::endl;
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
std::cout << "does not support this GEMM problem" << std::endl;
|
|
}
|
|
}
|
|
|
|
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
|
|
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
|
|
}
|
|
|
|
} // namespace profiler
|
|
} // namespace ck
|