Skip to content

GitLab

Unverified Commit f5de8b57 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Merge branch 'develop' into modified_grouped_gemm_addressing_method

parents e83c7061 fa9a0a5c
Hide whitespace changes
Inline Side-by-side
Showing with 850 additions and 268 deletions
library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt 0 → 100644
View file @ f5de8b57
# device_normalization_instance
set(DEVICE_NORMALIZATION_INSTANCE_SOURCE
device_softmax_f32_f32_instance.cpp
device_softmax_f16_f16_instance.cpp
)
add_library(device_normalization_instance OBJECT ${DEVICE_NORMALIZATION_INSTANCE_SOURCE})
set_target_properties(device_normalization_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(device_normalization_instance)
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f16_f16_instance.cpp 0 → 100644
View file @ f5de8b57
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_normalization_instance {
using F16 = ck::half_t;
using F32 = float;
template <index_t Rank, index_t Reduce>
using device_softmax_f16_f16_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8>
// clang-format on
>;
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 1>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 2>{});
}
void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 3>{});
}
} // namespace device_normalization_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f32_f32_instance.cpp 0 → 100644
View file @ f5de8b57
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_normalization_instance {
using F32 = float;
template <index_t Rank, index_t Reduce>
using device_softmax_f32_f32_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4>
// clang-format on
>;
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 2>{});
}
void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 2>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 3>{});
}
} // namespace device_normalization_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ f5de8b57
......@@ -22,6 +22,7 @@ set(PROFILER_SOURCE
src/profile_conv_bwd_weight.cpp
src/profile_batched_gemm_reduce.cpp
src/profile_gemm_add_add_fastgelu.cpp
src/profile_normalization.cpp
)
add_executable(ckProfiler ${PROFILER_SOURCE})
......@@ -46,4 +47,5 @@ target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_normalization_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ f5de8b57
......@@ -6,7 +6,7 @@
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -21,32 +21,28 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DeviceBatchedGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceBatchedGemmReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
DInElementOps,
DOutElementOps>;
using F32 = float;
using F16 = ck::half_t;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
using DeviceGemmReduceNoOpPtr =
ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
std::vector<DeviceGemmReduceNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
......@@ -59,7 +55,7 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename ReduceDataType,
typename ALayout,
typename BLayout,
typename CLayout>
......@@ -99,16 +95,16 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
Tensor<ReduceDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
......@@ -135,20 +131,23 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto dxs_in_element_op = DxsInElementOps{};
const auto dxs_out_element_op = DxsOutElementOps{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
const auto reduce0_op = ReduceOp0{};
const auto reduce1_op = ReduceOp1{};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
if(do_verification)
{
......@@ -160,6 +159,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
BElementOp,
CElementOp>;
using ReduceAccDataType = ReduceDataType;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
......@@ -172,21 +173,22 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetIdentityValue<float>();
float d1_acc = d1_reduce_op.GetIdentityValue<float>();
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d1_val;
ReduceAccDataType d0_val =
ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
ReduceAccDataType d1_val;
UnarySquareElementOp{}(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
square(d1_val, d0_val);
reduce0_op(reduce0_acc, d0_val);
reduce1_op(reduce1_acc, d1_val);
}
d0_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d0_acc);
d1_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d1_acc);
d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
}
}
......@@ -194,17 +196,19 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
d1_g_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceBatchedGemmReduceNoOpPtr>
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmReduceNoOpPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
......@@ -257,31 +261,32 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
// 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()),
&dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op,
BatchCount);
auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{},
gemm_element_ops,
{},
reduce_in_element_ops,
reduce_out_element_ops,
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......@@ -311,8 +316,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
float c_error = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
float d0_error = check_error(d0_g_m_host_result, d0_g_m_device_result);
......
profiler/include/profile_gemm_bias_add_reduce_impl.hpp
View file @ f5de8b57
......@@ -21,33 +21,28 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmBiasAddReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmBiasAddReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
DInElementOps,
DOutElementOps>;
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
using F32 = float;
using F16 = ck::half_t;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmBiasAddReduceNoOpPtr =
ck::tensor_operation::device::DeviceGemmReducePtr<1, ReducePtrsGlobal::Size()>;
void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
} // namespace device_gemm_instance
......@@ -61,9 +56,9 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename C0DataType,
typename C1DataType,
typename DDataType,
typename BiasDataType,
typename D0DataType,
typename ReduceDataType,
typename ALayout,
typename BLayout,
typename CLayout>
......@@ -77,7 +72,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
int StrideA,
int StrideB,
int StrideC,
int StrideC1)
int StrideD0)
{
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
......@@ -102,24 +97,24 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<C0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<C1DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_host_result(
Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> reduce0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_host_result(
Tensor<ReduceDataType> reduce1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_device_result(
Tensor<ReduceDataType> reduce0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_device_result(
Tensor<ReduceDataType> reduce1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
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_host_result.mDesc << std::endl;
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
......@@ -130,50 +125,53 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
bias_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c1_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
d0_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
std::srand(0);
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);
bias_n.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5}, num_thread);
c1_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
d0_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
using C1ElementOp = PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using D0ElementOp = PassThrough;
using ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto c1_element_op = C1ElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
auto d0_element_op = D0ElementOp{};
const auto reduce0_op = ReduceOp0{};
const auto reduce1_op = ReduceOp1{};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
auto div = UnaryDivElementOp{N};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&div, &div};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
ReduceDataType,
AElementOp,
BElementOp,
CElementOp>;
using ReduceAccDataType = DDataType;
using ReduceAccDataType = ReduceDataType;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -189,53 +187,53 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
ReduceAccDataType acc = static_cast<ReduceAccDataType>(c_m_n_host_result(m, n)) +
static_cast<ReduceAccDataType>(bias_n(n));
ReduceAccDataType c1 = static_cast<ReduceAccDataType>(c1_m_n(m, n));
ReduceAccDataType d0 = static_cast<ReduceAccDataType>(d0_m_n(m, n));
c_element_op(acc, acc);
c1_element_op(c1, c1);
acc += c1;
d0_element_op(d0, d0);
acc += d0;
c_m_n_host_result(m, n) = static_cast<CDataType>(acc);
}
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType c_val =
ReduceAccDataType d0_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
square(d1_val, d0_val);
reduce0_op(reduce0_acc, d0_val);
reduce1_op(reduce1_acc, d1_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
div(reduce0_acc, reduce0_acc);
div(reduce1_acc, reduce1_acc);
reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
}
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());
DeviceMem bias_device_buf(sizeof(C0DataType) * bias_n.mDesc.GetElementSpace());
DeviceMem c1_device_buf(sizeof(C1DataType) * c1_m_n.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
reduce0_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
reduce1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
bias_device_buf.ToDevice(bias_n.mData.data());
c1_device_buf.ToDevice(c1_m_n.mData.data());
d0_device_buf.ToDevice(d0_m_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasAddReduceNoOpPtr>
......@@ -249,7 +247,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
......@@ -257,7 +255,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
......@@ -265,7 +263,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
......@@ -273,7 +271,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
gemm_ptrs);
}
}
......@@ -291,34 +289,31 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
// 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()),
static_cast<C0DataType*>(bias_device_buf.GetDeviceBuffer()),
static_cast<C1DataType*>(c1_device_buf.GetDeviceBuffer()),
&dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op,
c1_element_op,
dxs_in_element_op,
dxs_out_element_op);
auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
bias_device_buf.GetDeviceBuffer(),
{d0_device_buf.GetDeviceBuffer()},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{StrideD0},
gemm_element_ops,
{&d0_element_op},
reduce_in_element_ops,
reduce_out_element_ops);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......@@ -328,9 +323,9 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(C0DataType) * M * N +
sizeof(C1DataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
sizeof(CDataType) * M * N + sizeof(BiasDataType) * M * N +
sizeof(D0DataType) * M * N + sizeof(ReduceDataType) * M +
sizeof(ReduceDataType) * M;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......@@ -350,12 +345,12 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
ck::utils::check_err(reduce0_m_device_result.mData, reduce0_m_host_result.mData);
ck::utils::check_err(reduce1_m_device_result.mData, reduce1_m_host_result.mData);
if(do_log)
{
......@@ -365,13 +360,17 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
<< std::endl;
}
}
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ f5de8b57
......@@ -21,21 +21,17 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
DInElementOps,
DOutElementOps>;
using F32 = float;
using F16 = ck::half_t;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr =
ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -60,7 +56,7 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename ReduceDataType,
typename ALayout,
typename BLayout,
typename CLayout>
......@@ -95,22 +91,22 @@ bool profile_gemm_reduce_impl(int do_verification,
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_host_result(
Tensor<ReduceDataType> reduce0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_host_result(
Tensor<ReduceDataType> reduce1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_device_result(
Tensor<ReduceDataType> reduce0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_device_result(
Tensor<ReduceDataType> reduce1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
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_host_result.mDesc << std::endl;
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
......@@ -130,34 +126,37 @@ bool profile_gemm_reduce_impl(int do_verification,
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
const auto reduce0_op = ReduceOp0{};
const auto reduce1_op = ReduceOp1{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
auto div = UnaryDivElementOp{N};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&div, &div};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
ReduceDataType,
AElementOp,
BElementOp,
CElementOp>;
using ReduceAccDataType = DDataType;
using ReduceAccDataType = ReduceDataType;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -169,37 +168,37 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType c_val =
ReduceAccDataType d0_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
square(d1_val, d0_val);
reduce0_op(reduce0_acc, d0_val);
reduce1_op(reduce1_acc, d1_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
div(reduce0_acc, reduce0_acc);
div(reduce1_acc, reduce1_acc);
reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
}
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());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
reduce0_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
reduce1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
......@@ -258,30 +257,31 @@ bool profile_gemm_reduce_impl(int do_verification,
// 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()),
&dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op);
auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{},
gemm_element_ops,
{},
reduce_in_element_ops,
reduce_out_element_ops);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......@@ -311,12 +311,12 @@ bool profile_gemm_reduce_impl(int do_verification,
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
ck::utils::check_err(reduce0_m_device_result.mData, reduce0_m_host_result.mData);
ck::utils::check_err(reduce1_m_device_result.mData, reduce1_m_host_result.mData);
if(do_log)
{
......@@ -326,13 +326,17 @@ bool profile_gemm_reduce_impl(int do_verification,
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
<< std::endl;
}
}
......
profiler/include/profile_grouped_gemm_impl.hpp
View file @ f5de8b57
......@@ -232,6 +232,10 @@ void profile_grouped_gemm_impl(int do_verification,
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
......
profiler/include/profile_normalization_impl.hpp 0 → 100644
View file @ f5de8b57
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_normalization_instance {
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>&);
} // namespace device_normalization_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
enum struct NormType
{
LAYERNORM,
BATCHNORM,
SOFTMAX,
};
enum struct NormDataType
{
F32_F32, // in, out
F16_F16,
BF16_BF16,
INT8_INT8,
};
// clang-format off
template <typename NormDataType> std::string type_to_string();
template <> std::string type_to_string<float>() { return "f32"; }
template <> std::string type_to_string<half_t>() { return "f16"; }
template <> std::string type_to_string<bhalf_t>() { return "bf16"; }
template <> std::string type_to_string<int8_t>() { return "int8"; }
template <> std::string type_to_string<int32_t>() { return "int32"; }
// clang-format on
template <typename InDataType, typename AccDataType, typename OutDataType>
void profile_normalization_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> in_length,
std::vector<index_t> in_strides,
std::vector<index_t> reduce_dims,
AccDataType alpha,
AccDataType beta,
NormType norm_type)
{
Tensor<InDataType> in = in_strides.empty() ? Tensor<InDataType>(in_length)
: Tensor<InDataType>(in_length, in_strides);
Tensor<OutDataType> out(in.mDesc);
switch(init_method)
{
// case 0: break;
case 0:
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{});
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
}
Tensor<OutDataType> out_ref(out);
DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
in_dev.ToDevice(in.mData.data());
out_dev.ToDevice(out.mData.data());
std::vector<index_t> i_in_lengths(in.mDesc.GetLengths().begin(), in.mDesc.GetLengths().end());
std::vector<index_t> i_in_strides(in.mDesc.GetStrides().begin(), in.mDesc.GetStrides().end());
// add device normalization instances
std::vector<tensor_operation::device::DeviceNormalizationPtr> instances;
if(norm_type == NormType::SOFTMAX)
{
if constexpr(is_same<InDataType, half_t>::value && is_same<OutDataType, half_t>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
tensor_operation::device::device_normalization_instance::
add_device_softmax_f16_f16_rank3_instances(instances);
if(in_length.size() == 4)
tensor_operation::device::device_normalization_instance::
add_device_softmax_f16_f16_rank4_instances(instances);
}
else if constexpr(is_same<InDataType, float>::value && is_same<OutDataType, float>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
tensor_operation::device::device_normalization_instance::
add_device_softmax_f32_f32_rank3_instances(instances);
if(in_length.size() == 4)
tensor_operation::device::device_normalization_instance::
add_device_softmax_f32_f32_rank4_instances(instances);
}
}
if(instances.size() <= 0)
{
throw std::runtime_error("wrong! no device normalization instance found");
}
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
for(auto& inst_ptr : instances)
{
// Is this user's responsibility to check if problem mismatches kernel instance (ie. rank 3
// problem to rank 4 kernel) other than invoking IsSupportedArgument()?
if(!(inst_ptr->GetRank() == static_cast<index_t>(i_in_lengths.size()) &&
inst_ptr->GetNumReduceDim() == static_cast<index_t>(reduce_dims.size())))
{
continue;
}
auto argument_ptr = inst_ptr->MakeArgumentPointer(i_in_lengths,
i_in_strides,
reduce_dims,
&alpha,
&beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer());
if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = [", in_length, ", ")
<< "], "
<< "scaler = [" << alpha << ", " << beta << "]." << std::endl;
return;
}
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes =
in.mDesc.GetElementSize() * sizeof(InDataType) +
(beta == 0.0f ? 1 : 2) * out.mDesc.GetElementSize() * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
// TODO: factory method to dynamically switch between different reference normalizations
using ReferenceFactory =
tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
ReferenceFactory{}.MakeInvoker().Run({in, out_ref, alpha, beta, reduce_dims});
out_dev.FromDevice(out.mData.data());
bool pass;
if(std::is_same<InDataType, int8_t>::value)
{
pass = ck::utils::check_err(
out.mData, out_ref.mData, "Error: Incorrect results!", 0, 1);
if(do_log)
{
LogRangeAsType<int>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<int>(std::cout << "out_ref : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<int>(std::cout << "out : ", out.mData, ",") << std::endl;
}
}
else
{
pass = ck::utils::check_err(out.mData, out_ref.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "out_ref : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out : ", out.mData, ",") << std::endl;
}
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "input lengths = [", in_length, ", ")
<< "], "
<< "scaler = [" << alpha << ", " << beta << "]." << std::endl;
}
}
}
std::cout << "Best Perf for datatype = " << type_to_string<InDataType>() << "_"
<< type_to_string<OutDataType>() << ", ";
LogRange(std::cout << "length = ", i_in_lengths, ",") << ", ";
LogRange(std::cout << "stride = ", i_in_strides, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dims, ",") << ", ";
std::cout << "alpha = " << alpha << ", "
<< "beta = " << beta << ", " << best_avg_time << " ms, " << best_gb_per_sec
<< " GB/s, " << best_instance_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/src/profile_normalization.cpp 0 → 100644
View file @ f5de8b57
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/include/profile_normalization_impl.hpp"
using ck::index_t;
using ck::profiler::NormDataType;
using ck::profiler::NormType;
struct ArgParser
{
std::unordered_map<std::string, NormType> norm_dict = {{"layernorm", NormType::LAYERNORM},
{"batchnorm", NormType::BATCHNORM},
{"softmax", NormType::SOFTMAX}};
std::unordered_map<std::string, std::vector<int>> long_opts = {
{"length", {}}, {"stride", {}}, {"reduce", {}}, {"alpha", {}}, {"beta", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help()
{
std::cout << "arg1: tensor operation (layernorm/batchnorm/softmax)\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg3: verification (0: no; 1: yes)\n"
<< "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg5: print tensor value (0: no; 1: yes)\n"
<< "arg6: time kernel (0=n0, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 8 4 256) \n"
<< "--stride: tensor strides (e.g, --stride 1024 256 1)\n"
<< "--reduce: to-reduce dimensions (e.g, --reduce 2)\n"
<< "--alpha: alpha scaling value\n"
<< "--beta: beta scaling value\n"
<< std::endl;
}
int profile_normalization(int argc, char* argv[])
{
if(argc <= 2)
{
print_help();
return 0;
}
ArgParser arg_parser;
// short unnamed options
const NormType norm_type = arg_parser.norm_dict[argv[1]];
const NormDataType data_type = static_cast<NormDataType>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
const std::vector<index_t> stride = arg_parser.long_opts["stride"];
const std::vector<index_t> reduce = arg_parser.long_opts["reduce"];
const index_t alpha =
arg_parser.long_opts["alpha"].empty() ? 1 : arg_parser.long_opts["alpha"][0];
const index_t beta = arg_parser.long_opts["beta"].empty() ? 0 : arg_parser.long_opts["beta"][0];
if(data_type == NormDataType::F16_F16)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else if(data_type == NormDataType::F32_F32)
{
ck::profiler::profile_normalization_impl<float, float, float>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
// hijack main() for quick debugging
// int main(int argc, char* argv[])
// {
// profile_normalization(argc, argv);
// return 0;
// }
profiler/src/profiler.cpp
View file @ f5de8b57
......@@ -20,6 +20,7 @@ int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_convnd_fwd(int argc, char* argv[]);
int profile_convnd_bwd_data(int, char*[], int);
int profile_conv_bwd_weight(int, char*[]);
int profile_normalization(int, char*[]);
int profile_reduce(int, char*[]);
static void print_helper_message()
......@@ -130,6 +131,11 @@ int main(int argc, char* argv[])
{
return profile_gemm_add_add_fastgelu(argc, argv);
}
else if(strcmp(argv[1], "batchnorm") == 0 || strcmp(argv[1], "layernorm") == 0 ||
strcmp(argv[1], "softmax") == 0)
{
return profile_normalization(argc, argv);
}
else
{
print_helper_message();
......
test/softmax/CMakeLists.txt
View file @ f5de8b57
......@@ -2,7 +2,10 @@ add_custom_target(test_softmax)
add_gtest_executable(test_softmax_fp32 test_softmax_fp32.cpp)
add_gtest_executable(test_softmax_fp16 test_softmax_fp16.cpp)
add_gtest_executable(test_softmax_int8 test_softmax_int8.cpp)
target_link_libraries(test_softmax_fp32 PRIVATE host_tensor)
target_link_libraries(test_softmax_fp16 PRIVATE host_tensor)
target_link_libraries(test_softmax_int8 PRIVATE host_tensor)
add_dependencies(test_softmax test_softmax_fp32)
add_dependencies(test_softmax test_softmax_fp16)
\ No newline at end of file
add_dependencies(test_softmax test_softmax_fp16)
add_dependencies(test_softmax test_softmax_int8)
\ No newline at end of file
test/softmax/test_softmax_fp16.cpp
View file @ f5de8b57
......@@ -15,14 +15,19 @@ class TestSoftmaxFP16 : public ck::TestSoftmax<Tuple>
// clang-format off
using KernelTypes = ::testing::Types<
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
std::tuple<ck::half_t, float, float, I<3>, I<1>, I<256>, I<8>, I<32>, I<1>, I<8>, I<1>, I<8>, I<4>>, // mixed precision
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<1>, I<256>, I<8>, I<32>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<1>, I<256>, I<4>, I<64>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<1>, I<256>, I<2>, I<128>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<16>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<32>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<2>, I<256>, I<8>, I<32>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<2>, I<256>, I<4>, I<64>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<2>, I<256>, I<2>, I<128>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<8>, I<8>>
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<16>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, float, ck::half_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<32>, I<1>, I<8>, I<8>>
>;
// clang-format on
TYPED_TEST_SUITE(TestSoftmaxFP16, KernelTypes);
......
test/softmax/test_softmax_fp32.cpp
View file @ f5de8b57
......@@ -15,14 +15,19 @@ class TestSoftmaxFP32 : public ck::TestSoftmax<Tuple>
// clang-format off
using KernelTypes = ::testing::Types<
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
std::tuple<float, float, ck::half_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<4>, I<8>>, // mixed precision
std::tuple<float, float, float, I<3>, I<1>, I<256>, I<8>, I<32>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<1>, I<256>, I<4>, I<64>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<1>, I<256>, I<2>, I<128>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<16>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<2>, I<256>, I<8>, I<32>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<2>, I<256>, I<4>, I<64>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<2>, I<256>, I<2>, I<128>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<4>, I<1>, I<4>, I<4>>
std::tuple<float, float, float, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<16>, I<1>, I<4>, I<4>>
>;
// clang-format on
TYPED_TEST_SUITE(TestSoftmaxFP32, KernelTypes);
......
test/softmax/test_softmax_int8.cpp 0 → 100644
View file @ f5de8b57
#include "gtest/gtest.h"
#include "test_softmax_util.hpp"
template <ck::index_t N>
using I = ck::Number<N>;
template <typename Tuple>
class TestSoftmaxINT8 : public ck::TestSoftmax<Tuple>
{
};
// clang-format off
using KernelTypes = ::testing::Types<
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
std::tuple<int8_t, float, int8_t, I<3>, I<1>, I<256>, I<8>, I<32>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<1>, I<256>, I<4>, I<64>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<1>, I<256>, I<2>, I<128>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<32>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<1>, I<256>, I<1>, I<256>, I<1>, I<64>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<2>, I<256>, I<8>, I<32>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<2>, I<256>, I<4>, I<64>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<2>, I<256>, I<2>, I<128>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<16>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<32>, I<1>, I<16>, I<16>>,
std::tuple<int8_t, float, int8_t, I<3>, I<2>, I<256>, I<1>, I<256>, I<1>, I<64>, I<1>, I<16>, I<16>>
>;
// clang-format on
TYPED_TEST_SUITE(TestSoftmaxINT8, KernelTypes);
TYPED_TEST(TestSoftmaxINT8, Test_INT8) { this->Run(); }
test/softmax/test_softmax_util.hpp
View file @ f5de8b57
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <iostream>
#include <gtest/gtest.h>
......@@ -16,6 +18,18 @@
namespace ck {
template <typename Range>
std::string serialize_range(const Range& range)
{
std::stringstream ss;
for(auto& r : range)
{
ss << r << ", ";
}
std::string str = ss.str();
return std::string(str.begin(), str.end() - 2);
}
template <typename Tuple>
class TestSoftmax : public ::testing::Test
{
......@@ -80,23 +94,43 @@ class TestSoftmax : public ::testing::Test
auto argument_ptr = device_instance.MakeArgumentPointer(i_in_lengths,
i_in_strides,
reduce_dims,
alpha,
beta,
&alpha,
&beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer());
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
{
FAIL() << "Unsupported argument";
// std::cout << "Skipped due to unsupported argument: "
// << "input lengths = [" << serialize_range(in_length) << "], "
// << "scaler = [" << alpha << ", " << beta << "]." << std::endl;
return;
}
auto invoker_ptr = device_instance.MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get());
ref_instance_invoker_.Run({in, out_ref, alpha, beta, Rank, reduce_dims});
ref_instance_invoker_.Run({in, out_ref, alpha, beta, reduce_dims});
out_dev.FromDevice(out.mData.data());
EXPECT_TRUE(ck::utils::check_err(out.mData, out_ref.mData));
bool pass;
if(std::is_same<InDataType, int8_t>::value)
{
EXPECT_TRUE(pass = ck::utils::check_err(
out.mData, out_ref.mData, "Error: Incorrect results!", 0, 1));
}
else
{
EXPECT_TRUE(pass = ck::utils::check_err(out.mData, out_ref.mData));
}
if(!pass)
{
FAIL() << "Failure in input lengths = [" << serialize_range(in_length) << "], "
<< "scaler = [" << alpha << ", " << beta << "].";
}
}
void Run()
......@@ -105,13 +139,14 @@ class TestSoftmax : public ::testing::Test
{
for(auto scale : this->scales_)
{
this->RunSingle(in_length, std::get<0>(scale), std::get<1>(scale));
this->RunSingle(in_length, scale[0], scale[1]);
}
}
}
std::vector<std::vector<index_t>> in_lengths_ = {{1, 8, 128}, {2, 128, 1024}, {3, 9, 1032}};
std::vector<std::tuple<AccDataType, AccDataType>> scales_ = {{1, 0}, {2, 2}, {0, 1}};
std::vector<std::vector<index_t>> in_lengths_ = {
{1, 8, 128}, {2, 128, 1024}, {3, 9, 1032}, {4, 4, 2048}, {8, 1, 8192}};
std::vector<std::vector<AccDataType>> scales_ = {{1, 0}, {1, 1}, {0, 1}, {2, 2}};
typename ReferenceInstance::Invoker ref_instance_invoker_;
};
......
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