Skip to content

GitLab

Unverified Commit 12d70ff3 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge pull request #64 from ROCm/merge_from_public 

Merge from public
parents 5dbbf5d6 6422c23b
Show whitespace changes
Inline Side-by-side
Showing with 2054 additions and 586 deletions
example/44_elementwise_permute/elementwise_permute_3d.cpp deleted 100644 → 0
View file @ 5dbbf5d6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
#include "ck/library/utility/algorithm.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwise3dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
2, // NumDim_m, {N, C}
2, // NumDim_n, {H, W}
1, // NumDim_k, {D}
4, // MPerThread
4, // NPerThread
4, // KPerThread
ck::Sequence<4>, // InScalarPerVectorSeq
ck::Sequence<4>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
{
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
{
auto a_val = A_ncdhw(n, c, d, h, w);
functor(B_ndhwc(n, d, h, w, c), a_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
const int N = 4;
const int C = 16;
const int H = 32;
const int W = 5;
const int D = 16;
std::vector<std::size_t> ncdhw = {N, C, D, H, W};
std::vector<std::size_t> ndhwc = {N, D, H, W, C};
Tensor<ADataType> a(ncdhw);
Tensor<BDataType> b(ndhwc);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (ncdhw): " << a.mDesc << std::endl;
std::cout << "B (ndhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
std::size_t num_btype =
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(ndhwc);
host_elementwise4D(host_b, a, PassThrough{});
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
View file @ 12d70ff3
......@@ -6,7 +6,9 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -21,27 +23,19 @@ using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // Elementwise op
PassThrough, // Elementwise
4, // NumDim
8, // MPerThread
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_nhwc, const HostTensorA& A_nchw, Functor functor)
{
for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
{
auto a_val = A_nchw(n, c, h, w);
functor(B_nhwc(n, h, w, c), a_val);
}
}
ck::Sequence<8>>; // OutScalarPerVectorSeq
int main()
{
......@@ -50,18 +44,6 @@ int main()
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 128};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
......@@ -72,9 +54,22 @@ int main()
1,
static_cast<int>(nhwc[2] * nhwc[3]),
static_cast<int>(nhwc[3])};
ck::ranges::copy(nchw, ab_lengths.begin());
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
......@@ -106,10 +101,16 @@ int main()
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{});
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance =
ck::tensor_operation::host::ReferenceElementwise<1, ADataType, BDataType, PassThrough>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(as, host_b, PassThrough{});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
......
example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp deleted 100644 → 0
View file @ 5dbbf5d6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.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"
using F16 = ck::half_t;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // Elementwise op
3, // NumDim_M
1, // NumDim_N
1, // MPerThread
1, // NPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
const std::vector<std::size_t>& shape_nchw,
Functor functor)
{
for(std::size_t n = 0; n < shape_nchw[0]; ++n)
for(std::size_t c = 0; c < shape_nchw[1]; ++c)
for(std::size_t h = 0; h < shape_nchw[2]; ++h)
for(std::size_t w = 0; w < shape_nchw[3]; ++w)
{
auto a_val = A_nchw(n, c, h, w);
functor(B_nhwc(n, h, w, c), a_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
const int N = 120;
const int C = 128;
const int H = 32;
const int W = 1024;
std::vector<std::size_t> nchw = {N, C, H, W};
std::vector<std::size_t> nhwc = {N, H, W, C};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths{N, H, W, C};
std::array<ck::index_t, 4> a_strides = {C * H * W, W, 1, H * W};
std::array<ck::index_t, 4> b_strides = {H * W * C, W * C, C, 1};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
host_b, a, nchw, PassThrough{});
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp
View file @ 12d70ff3
......@@ -6,8 +6,10 @@
#include <random>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -21,43 +23,23 @@ using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
using UnaryScale = ck::tensor_operation::element_wise::Scale;
using UnarySquare = ck::tensor_operation::element_wise::UnarySquare;
using UnaryScaleSquare =
ck::tensor_operation::element_wise::UnaryCombinedOp<UnarySquare, UnaryScale>;
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
8, // MPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
std::size_t N = A_nchw.mDesc.GetLengths()[0];
std::size_t C = A_nchw.mDesc.GetLengths()[1];
std::size_t H = A_nchw.mDesc.GetLengths()[2];
std::size_t W = A_nchw.mDesc.GetLengths()[3];
for(std::size_t w = 0; w < W; ++w)
for(std::size_t h = 0; h < H; ++h)
for(std::size_t c = 0; c < C; ++c)
for(std::size_t n = 0; n < N; ++n)
{
ADataType tmp_val;
auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
functor_b(tmp_val, a_val);
functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
scale * tmp_val);
}
}
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<8>>; // OutScalarPerVectorSeq
int main()
{
......@@ -66,8 +48,21 @@ int main()
std::vector<std::size_t> nchw = {16, 8, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 8};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
float scale = 1.f;
auto i = 0;
std::mt19937 gen(11939);
......@@ -90,28 +85,14 @@ int main()
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
......@@ -125,11 +106,10 @@ int main()
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
std::size_t flop = std::size_t(5) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype =
(2 * sizeof(ADataType) + sizeof(BDataType)) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
......@@ -141,10 +121,17 @@ int main()
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance = ck::tensor_operation::host::
ReferenceElementwise<1, ADataType, BDataType, UnaryScaleSquare>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(
as, host_b, UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
......
example/44_elementwise_permute/elementwise_permute_4D_fp16_row.cpp
View file @ 12d70ff3
......@@ -5,8 +5,10 @@
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -20,38 +22,23 @@ using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
using UnaryScale = ck::tensor_operation::element_wise::Scale;
using UnarySquare = ck::tensor_operation::element_wise::UnarySquare;
using UnaryScaleSquare =
ck::tensor_operation::element_wise::UnaryCombinedOp<UnarySquare, UnaryScale>;
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
8, // MPerThread
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
{
ADataType tmp_val;
auto a_val = A_nchw(n, c, h, w);
functor_b(tmp_val, a_val);
functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
}
}
ck::Sequence<8>>; // OutScalarPerVectorSeq
int main()
{
......@@ -60,18 +47,6 @@ int main()
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 128};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
......@@ -85,15 +60,29 @@ int main()
ck::ranges::copy(nchw, ab_lengths.begin());
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
......@@ -123,10 +112,17 @@ int main()
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance = ck::tensor_operation::host::
ReferenceElementwise<1, ADataType, BDataType, UnaryScaleSquare>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(
as, host_b, UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
......
example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp
View file @ 12d70ff3
......@@ -5,8 +5,10 @@
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -20,53 +22,47 @@ using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
using UnaryScale = ck::tensor_operation::element_wise::Scale;
using UnarySquare = ck::tensor_operation::element_wise::UnarySquare;
using UnaryScaleSquare =
ck::tensor_operation::element_wise::UnaryCombinedOp<UnarySquare, UnaryScale>;
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
1, // MPerThread
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
std::size_t N = A_nchw.mDesc.GetLengths()[0];
std::size_t C = A_nchw.mDesc.GetLengths()[1];
std::size_t H = A_nchw.mDesc.GetLengths()[2];
std::size_t W = A_nchw.mDesc.GetLengths()[3];
for(std::size_t w = 0; w < W; ++w)
for(std::size_t h = 0; h < H; ++h)
for(std::size_t c = 0; c < C; ++c)
for(std::size_t n = 0; n < N; ++n)
{
ADataType tmp_val;
auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
functor_b(tmp_val, a_val);
functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {5, 4, 2, 3};
std::vector<std::size_t> nhwc = {5, 2, 3, 4};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
std::vector<std::size_t> nchw = {16, 8, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 8};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
float scale = 1.f;
auto i = 0;
......@@ -90,28 +86,14 @@ int main()
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
......@@ -141,10 +123,17 @@ int main()
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance = ck::tensor_operation::host::
ReferenceElementwise<1, ADataType, BDataType, UnaryScaleSquare>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(
as, host_b, UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
......
example/44_elementwise_permute/elementwise_permute_4D_fp32_row.cpp
View file @ 12d70ff3
......@@ -5,8 +5,10 @@
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -20,38 +22,23 @@ using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
using UnaryScale = ck::tensor_operation::element_wise::Scale;
using UnarySquare = ck::tensor_operation::element_wise::UnarySquare;
using UnaryScaleSquare =
ck::tensor_operation::element_wise::UnaryCombinedOp<UnarySquare, UnaryScale>;
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
8, // MPerThread
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
{
ADataType tmp_val;
auto a_val = A_nchw(n, c, h, w);
functor_b(tmp_val, a_val);
functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
}
}
ck::Sequence<8>>; // OutScalarPerVectorSeq
int main()
{
......@@ -60,18 +47,6 @@ int main()
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 128};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
......@@ -85,15 +60,28 @@ int main()
ck::ranges::copy(nchw, ab_lengths.begin());
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
......@@ -123,10 +111,17 @@ int main()
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance = ck::tensor_operation::host::
ReferenceElementwise<1, ADataType, BDataType, UnaryScaleSquare>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(
as, host_b, UnaryScaleSquare{UnarySquare{}, UnaryScale{scale}});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
......
example/44_elementwise_permute/elementwise_trinary_4D_fp16.cpp 0 → 100644
View file @ 12d70ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using UnaryScale = ck::tensor_operation::element_wise::Scale;
using UnarySquare = ck::tensor_operation::element_wise::UnarySquare;
using UnaryScaleSquare =
ck::tensor_operation::element_wise::UnaryCombinedOp<UnarySquare, UnaryScale>;
using BinaryAdd = ck::tensor_operation::element_wise::Add;
// B = alpha * A0 * A0 + beta * A1 * A1 + gamma * A2 * A2
using TrinaryAddUnaryScaleSquare =
ck::tensor_operation::element_wise::TrinaryWithUnaryCombinedOp<BinaryAdd,
BinaryAdd,
UnaryScaleSquare,
UnaryScaleSquare,
UnaryScaleSquare>;
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<ADataType, ADataType, ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
TrinaryAddUnaryScaleSquare, // ElementwiseOp
4, // NumDim
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<8, 8, 8>, // InScalarPerVectorSeq
ck::Sequence<8>>; // OutScalarPerVectorSeq
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> ab_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
static_cast<int>(nchw[2] * nchw[3]),
static_cast<int>(nchw[3]),
1};
ck::ranges::copy(nchw, ab_lengths.begin());
std::array<Tensor<ADataType>, 3> as = {Tensor<ADataType>(ab_lengths, ab_strides),
Tensor<ADataType>(ab_lengths, ab_strides),
Tensor<ADataType>(ab_lengths, ab_strides)};
Tensor<ADataType>& a0 = as[0];
Tensor<ADataType>& a1 = as[1];
Tensor<ADataType>& a2 = as[2];
Tensor<BDataType> b(ab_lengths, ab_strides);
float alpha = 3.f;
float beta = 2.f;
float gamma = 4.f;
a0.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
a1.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
a2.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a0_device_buf(sizeof(ADataType) * a0.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(ADataType) * a1.mDesc.GetElementSpaceSize());
DeviceMem a2_device_buf(sizeof(ADataType) * a2.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0.mData.data());
a1_device_buf.ToDevice(a1.mData.data());
a2_device_buf.ToDevice(a2.mData.data());
std::array<const void*, 3> inputs = {a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer(),
a2_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto unary_scale_op_a0 = UnaryScaleSquare{UnarySquare{}, UnaryScale{alpha}};
auto unary_scale_op_a1 = UnaryScaleSquare{UnarySquare{}, UnaryScale{beta}};
auto unary_scale_op_a2 = UnaryScaleSquare{UnarySquare{}, UnaryScale{gamma}};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths,
{ab_strides, ab_strides, ab_strides},
{ab_strides},
inputs,
output,
TrinaryAddUnaryScaleSquare{
BinaryAdd{}, BinaryAdd{}, unary_scale_op_a0, unary_scale_op_a1, unary_scale_op_a2});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A0 (nchw): " << a0.mDesc << std::endl;
std::cout << "A1 (nchw): " << a1.mDesc << std::endl;
std::cout << "A2 (nchw): " << a2.mDesc << std::endl;
std::cout << "B (nchw): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(5) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
Tensor<BDataType> host_b(ab_lengths, ab_strides);
using ReferenceElementwiseInstance = ck::tensor_operation::host::
ReferenceElementwise<3, ADataType, BDataType, TrinaryAddUnaryScaleSquare>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(
as,
host_b,
TrinaryAddUnaryScaleSquare{
BinaryAdd{}, BinaryAdd{}, unary_scale_op_a0, unary_scale_op_a1, unary_scale_op_a2});
ref_invoker.Run(ref_argument);
const double threshold = std::pow(2, -10) * 2;
b_device_buf.FromDevice(b.mData.data());
pass &= ck::utils::check_err(
b.mData, host_b.mData, "Error: Incorrect results b", threshold, threshold);
}
return pass ? 0 : 1;
}
example/46_gemm_add_multiply/README.md
View file @ 12d70ff3
......@@ -8,19 +8,3 @@
#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
./bin/example_gemm_add_multiply_dl_fp16 1 1 1
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
arg.a_grid_desc_k0_m0_m1_k1_{2048, 3840, 2}
arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
arg.e_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {960, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 3.99904 ms, 32.22 TFlops, 31.9913 GB/s, DeviceGemmMultipleD_Dl<256, 128, 128, 16, 2, 4, 4, 1>
```
example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
View file @ 12d70ff3
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp)
set(target 1)
endif()
endforeach()
add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute_xdl.cpp)
example/52_im2col_col2im/CMakeLists.txt
View file @ 12d70ff3
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_im2col_col2im)
add_custom_target(example_im2col_col2im)
add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
add_example_dependencies(example_im2col_col2im example_image_to_column_f32)
add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
add_example_dependencies(example_im2col_col2im example_image_to_column_f32)
add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
add_example_dependencies(example_im2col_col2im example_column_to_image_f32)
set(target 1)
endif()
endforeach()
add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
add_example_dependencies(example_im2col_col2im example_column_to_image_f32)
example/59_grouped_gemm_multi_ABD/CMakeLists.txt 0 → 100644
View file @ 12d70ff3
add_custom_target(example_grouped_gemm_xdl_multi_abd)
add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16 grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp)
add_example_dependencies(example_grouped_gemm_xdl_multi_abd example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16)
add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8 grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp)
add_example_dependencies(example_grouped_gemm_xdl_multi_abd example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8)
example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp 0 → 100644
View file @ 12d70ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multi_abd_xdl_fixed_nk.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.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"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<D0DataType>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Col;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using DsLayout = ck::Tuple<Row>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = AddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Multi_ABD_Fixed_NK
// clang-format off
///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
///######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
///######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 128, 16, 128, 32, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1>;
// clang-format on
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
std::vector<Tensor<A0DataType>> a0_tensors;
std::vector<Tensor<B1DataType>> b_tensors;
std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<D0DataType>> d0_tensors;
std::vector<Tensor<EDataType>> c_host_tensors;
std::vector<Tensor<EDataType>> c_device_tensors;
a0_tensors.reserve(group_count);
b_tensors.reserve(group_count);
b0_tensors.reserve(group_count);
b1_tensors.reserve(group_count);
d0_tensors.reserve(group_count);
c_host_tensors.reserve(group_count);
c_device_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
d0_tensors_device, c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
d0_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
a0_tensors.push_back(Tensor<A0DataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A0Layout{})));
b_tensors.push_back(Tensor<B1DataType>(f_host_tensor_descriptor(
problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
b0_tensors.push_back(Tensor<B0DataType>(f_host_tensor_descriptor(
problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
b1_tensors.push_back(Tensor<B1DataType>(
f_host_tensor_descriptor(problem_size.Ks[i], problem_size.Ns[i], 0, B1Layout{})));
d0_tensors.push_back(Tensor<D0DataType>(
f_host_tensor_descriptor(problem_size.Ms[i], problem_size.Ns[i], 0, ELayout{})));
c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
std::cout << "gemm[" << i << "] a_m_k: " << a0_tensors[i].mDesc
<< " b_k_n: " << b0_tensors[i].mDesc << " d_m_n: " << d0_tensors[i].mDesc
<< " c_m_n: " << c_device_tensors[i].mDesc << std::endl;
flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
num_btype += sizeof(A0DataType) * a0_tensors[i].mDesc.GetElementSize() +
sizeof(B0DataType) * b0_tensors[i].mDesc.GetElementSize() +
sizeof(B1DataType) * b1_tensors[i].mDesc.GetElementSize() +
sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSize() +
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
switch(config.init_method)
{
case 0: break;
case 1:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
b0_tensors[i].GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_tensors[i].GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
break;
case 2:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_tensors[i].GenerateTensorValue(GeneratorTensor_3<B0DataType>{-5, 5});
b1_tensors[i].GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
default:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
}
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 1;
using GroupedGemmKernelArgument = ck::tensor_operation::device::
GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
d0_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
a0_tensors_device[i]->ToDevice(a0_tensors[i].mData.data(),
a0_tensors[i].mDesc.GetElementSpaceSize() *
sizeof(A0DataType));
b0_tensors_device[i]->ToDevice(b0_tensors[i].mData.data(),
b0_tensors[i].mDesc.GetElementSpaceSize() *
sizeof(B0DataType));
b1_tensors_device[i]->ToDevice(b1_tensors[i].mData.data(),
b1_tensors[i].mDesc.GetElementSpaceSize() *
sizeof(B1DataType));
d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
c_tensors_device[i]->SetZero();
gemm_descs.push_back(
{sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {0}, 1});
grouped_gemm_kernel_args_.push_back(
{std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
b1_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
std::array<ck::index_t, NumDTensor>{0},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
std::vector<std::array<const void*, NumATensor>> p_As = {};
std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
// do GEMM
auto argument = gemm.MakeArgument(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
DeviceMem gemm_kernel_args_dev(gemm.GetDeviceKernelArgSize(&argument));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
gemm.GetDeviceKernelArgSize(&argument),
hipMemcpyHostToDevice));
gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
gemm.SetKBatch(argument, config.k_batch);
gemm.SetElementwiseOps(argument, a_element_op, b_element_op, cde_element_op);
invoker.Run(argument, StreamConfig{nullptr, false});
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
}
bool pass = true;
if(config.do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B1DataType,
EDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
for(int n = 0; n < problem_size.Ns[i]; ++n)
{
for(int k = 0; k < problem_size.Ks[i]; ++k)
{
b_element_op(b_tensors[i](k, n), b0_tensors[i](k, n), b1_tensors[i](k, n));
}
}
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
c_device_tensors[i].mDesc.GetElementSize() *
sizeof(EDataType));
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a0_tensors[i],
b_tensors[i],
c_host_tensors[i],
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < problem_size.Ms[i]; ++m)
{
for(int n = 0; n < problem_size.Ns[i]; ++n)
{
cde_element_op(
c_host_tensors[i](m, n), c_host_tensors[i](m, n), d0_tensors[i](m, n));
}
}
pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
}
}
return pass;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(32 + rand() % 32);
problem_size.Ns.push_back(1024);
problem_size.Ks.push_back(512);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
}
if(argc == 5)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
config.k_batch = std::stoi(argv[4]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4: k_batch (>0)\n");
exit(0);
}
return !run_grouped_gemm(problem_size, config);
}
example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp 0 → 100644
View file @ 12d70ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multi_abd_xdl_fixed_nk.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.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"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using Scale = ck::tensor_operation::element_wise::Scale;
using AddScale = ck::tensor_operation::element_wise::BinaryWithUnaryCombinedOp<Add, Scale, Scale>;
using A0DataType = F16;
using A1DataType = F32;
using AsDataType = ck::Tuple<A0DataType, A1DataType>;
using B0DataType = F16;
using BsDataType = ck::Tuple<B0DataType>;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F32;
using DsDataType = ck::Tuple<D0DataType>;
using EDataType = F32;
using A0Layout = Row;
using A1Layout = Row;
using AsLayout = ck::Tuple<A0Layout, A1Layout>;
using B0Layout = Col;
using BsLayout = ck::Tuple<B0Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using AElementOp = AddScale;
using BElementOp = PassThrough;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Multi_ABD_Fixed_NK
// clang-format off
///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
///######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
///######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 128, 16, 128, 32, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ck::half_t>;
// clang-format on
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
std::vector<Tensor<A0DataType>> a0_tensors;
std::vector<Tensor<A1DataType>> a1_tensors;
std::vector<Tensor<B0DataType>> b_tensors;
std::vector<Tensor<D0DataType>> d0_tensors;
std::vector<Tensor<EDataType>> e_host_tensors;
std::vector<Tensor<EDataType>> e_device_tensors;
a0_tensors.reserve(group_count);
a1_tensors.reserve(group_count);
b_tensors.reserve(group_count);
d0_tensors.reserve(group_count);
e_host_tensors.reserve(group_count);
e_device_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, a1_tensors_device, b_tensors_device,
d0_tensors_device, c_tensors_device;
a0_tensors_device.reserve(group_count);
a1_tensors_device.reserve(group_count);
b_tensors_device.reserve(group_count);
d0_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
a0_tensors.push_back(Tensor<A0DataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A0Layout{})));
a1_tensors.push_back(Tensor<A1DataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A1Layout{})));
b_tensors.push_back(Tensor<B0DataType>(f_host_tensor_descriptor(
problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
d0_tensors.push_back(Tensor<D0DataType>(
f_host_tensor_descriptor(problem_size.Ms[i], problem_size.Ns[i], 0, ELayout{})));
e_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
e_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
std::cout << "gemm[" << i << "] a_m_k: " << a0_tensors[i].mDesc
<< " b_k_n: " << b_tensors[i].mDesc << " d_m_n: " << d0_tensors[i].mDesc
<< " c_m_n: " << e_device_tensors[i].mDesc << std::endl;
flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
num_btype += sizeof(A0DataType) * a0_tensors[i].mDesc.GetElementSize() +
sizeof(A1DataType) * a1_tensors[i].mDesc.GetElementSize() +
sizeof(B0DataType) * b_tensors[i].mDesc.GetElementSize() +
sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSize() +
sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSize();
switch(config.init_method)
{
case 0: break;
case 1:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
a1_tensors[i].GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
break;
case 2:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
a1_tensors[i].GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
break;
default:
a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
a1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
}
constexpr ck::index_t NumATensor = 2;
constexpr ck::index_t NumBTensor = 1;
constexpr ck::index_t NumDTensor = 1;
using GroupedGemmKernelArgument = ck::tensor_operation::device::
GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
a1_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(A1DataType) * sum_of_m * problem_size.Ks[i]));
b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
d0_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
a0_tensors_device[i]->ToDevice(a0_tensors[i].mData.data(),
a0_tensors[i].mDesc.GetElementSpaceSize() *
sizeof(A0DataType));
a1_tensors_device[i]->ToDevice(a1_tensors[i].mData.data(),
a1_tensors[i].mDesc.GetElementSpaceSize() *
sizeof(A1DataType));
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
b_tensors[i].mDesc.GetElementSpaceSize() *
sizeof(B0DataType));
d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
c_tensors_device[i]->SetZero();
gemm_descs.push_back({sum_of_m,
problem_size.Ns[i],
problem_size.Ks[i],
{1, 1},
{problem_size.stride_Bs[i]},
{0},
1});
grouped_gemm_kernel_args_.push_back(
{std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer(),
a1_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
std::array<ck::index_t, NumATensor>{problem_size.stride_As[i],
problem_size.stride_As[i]},
std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i]},
std::array<ck::index_t, NumDTensor>{0},
problem_size.stride_Cs[i]});
}
constexpr float scale = 1.f;
auto a_element_op = AElementOp{Add{}, Scale{scale}, Scale{scale}};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
std::vector<std::array<const void*, NumATensor>> p_As = {};
std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
// do GEMM
auto argument = gemm.MakeArgument(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
DeviceMem gemm_kernel_args_dev(gemm.GetDeviceKernelArgSize(&argument));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
gemm.GetDeviceKernelArgSize(&argument),
hipMemcpyHostToDevice));
gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
gemm.SetKBatch(argument, config.k_batch);
gemm.SetElementwiseOps(argument, a_element_op, b_element_op, cde_element_op);
invoker.Run(argument, StreamConfig{nullptr, false});
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
}
bool pass = true;
if(config.do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B0DataType,
EDataType,
AccDataType,
PassThrough,
BElementOp,
PassThrough>;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
for(int m = 0; m < problem_size.Ms[i]; ++m)
{
for(int k = 0; k < problem_size.Ks[i]; ++k)
{
a_element_op(a0_tensors[i](m, k), a0_tensors[i](m, k), a1_tensors[i](m, k));
}
}
c_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data(),
e_device_tensors[i].mDesc.GetElementSize() *
sizeof(EDataType));
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a0_tensors[i],
b_tensors[i],
e_host_tensors[i],
PassThrough{},
b_element_op,
PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < problem_size.Ms[i]; ++m)
{
for(int n = 0; n < problem_size.Ns[i]; ++n)
{
cde_element_op(
e_host_tensors[i](m, n), e_host_tensors[i](m, n), d0_tensors[i](m, n));
}
}
pass &= ck::utils::check_err(e_device_tensors[i], e_host_tensors[i]);
}
}
return pass;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(32 + rand() % 32);
problem_size.Ns.push_back(1024);
problem_size.Ks.push_back(512);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
}
if(argc == 5)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
config.k_batch = std::stoi(argv[4]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4: k_batch (>0)\n");
exit(0);
}
return !run_grouped_gemm(problem_size, config);
}
example/60_gemm_multi_ABD/CMakeLists.txt
View file @ 12d70ff3
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
add_example_executable(example_gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp)
add_example_executable(example_gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp)
add_example_executable(example_gemm_multi_ABD_xdl_fastgelu_bf16_i8 gemm_multi_ABD_xdl_fastgelu_bf16_i8.cpp)
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp 0 → 100644
View file @ 12d70ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<D0DataType>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = AddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
// clang-format off
///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
///######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
///######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
b0_device_buf.ToDevice(b0_k_n.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 1;
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer()},
std::array<const void*, NumDTensor>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, NumATensor>{StrideA},
std::array<ck::index_t, NumBTensor>{StrideB, 0},
std::array<ck::index_t, NumDTensor>{StrideD},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
Tensor<A0DataType> a_m_k({M, K});
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K; ++k)
{
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B1DataType,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fastgelu_bf16_i8.cpp 0 → 100644
View file @ 12d70ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = FastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
// clang-format off
///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
///######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
///######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
b0_device_buf.ToDevice(b0_k_n.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 0;
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer()},
std::array<const void*, NumDTensor>{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, NumATensor>{StrideA},
std::array<ck::index_t, NumBTensor>{StrideB, 0},
std::array<ck::index_t, NumDTensor>{},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
Tensor<A0DataType> a_m_k({M, K});
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K; ++k)
{
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B1DataType,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp
View file @ 12d70ff3
......@@ -37,7 +37,7 @@ using DDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using BLayout = Row;
using DLayout = Row;
using ELayout = Row;
......@@ -141,9 +141,9 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
1,
2,
8,
8,
1,
1,
1,
......@@ -161,10 +161,10 @@ int main(int argc, char* argv[])
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD = 4096;
ck::index_t StrideE = 4096;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideD = N;
ck::index_t StrideE = N;
float alpha = 1.0f;
float beta = 1.0f;
......
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp 0 → 100644
View file @ 12d70ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType>;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = BF16;
using DsDataType = ck::Tuple<B1DataType, D0DataType>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<B1Layout, D0Layout>;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MultiplyAddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
// clang-format off
///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
///######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
///######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
b0_device_buf.ToDevice(b0_k_n.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 1;
constexpr ck::index_t NumDTensor = 2;
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer(),
d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, NumATensor>{StrideA},
std::array<ck::index_t, NumBTensor>{StrideB},
std::array<ck::index_t, NumDTensor>{0, StrideD},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
Tensor<A0DataType> a_m_k({M, K});
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
#if 0
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K; ++k)
{
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
}
}
#endif
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B0DataType,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), b1_k_n(0, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
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