Skip to content

GitLab

Commit 6b9a4bd5 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop-staging-0423' into amd-master

parents 56de337f c5f1cdf7
Hide whitespace changes
Inline Side-by-side
Showing with 1228 additions and 348 deletions
example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
View file @ 6b9a4bd5
......@@ -8,6 +8,8 @@
#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/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -30,22 +32,6 @@ using DeviceElementwisePermuteInstance =
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;
......@@ -54,13 +40,16 @@ int main()
const int N = 120;
const int C = 128;
const int H = 32;
const int W = 1024;
const int W = 32;
std::vector<std::size_t> nchw = {N, C, H, W};
std::vector<std::size_t> nhwc = {N, H, W, C};
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};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
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});
......@@ -72,11 +61,6 @@ 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{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{});
......@@ -94,10 +78,11 @@ int main()
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 flop =
std::size_t(2) * ab_lengths[0] * ab_lengths[1] * ab_lengths[2] * ab_lengths[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 num_btype = (sizeof(ADataType) + sizeof(BDataType)) *
(ab_lengths[0] * ab_lengths[1] * ab_lengths[2] * ab_lengths[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......@@ -110,11 +95,16 @@ int main()
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
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();
Tensor<BDataType> host_b(nhwc);
host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
host_b, a, nchw, PassThrough{});
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_col.cpp
View file @ 6b9a4bd5
......@@ -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
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
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);
}
}
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
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
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,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
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 @ 6b9a4bd5
......@@ -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
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
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);
}
}
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
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
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()
{
......@@ -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,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
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 @ 6b9a4bd5
......@@ -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
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
1, // 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);
}
}
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
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
256, // BlockSize
128, // M0PerBlock
128, // M1PerBlock
8, // M0PerThread
8, // M1PerThread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
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,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
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 @ 6b9a4bd5
......@@ -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
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
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);
}
}
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
UnaryScaleSquare, // UnaryScaleSquare
4, // NumDim
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()
{
......@@ -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,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
auto argument =
broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
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 @ 6b9a4bd5
// 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/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
View file @ 6b9a4bd5
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
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 @ 6b9a4bd5
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
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/60_gemm_multi_ABD/CMakeLists.txt
View file @ 6b9a4bd5
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
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)
example/61_contraction_multi_ABD/CMakeLists.txt
View file @ 6b9a4bd5
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)
example/62_convnd_activ/CMakeLists.txt
View file @ 6b9a4bd5
......@@ -2,16 +2,9 @@ add_subdirectory(binary)
add_subdirectory(multi_AB)
add_subdirectory(unary)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_convnd_activ_xdl)
# ScaleAdd ScaleAdd Relu
add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp)
add_example_dependencies(example_convnd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16)
add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp)
add_example_dependencies(example_convnd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16)
set(target 1)
endif()
endforeach()
add_custom_target(example_convnd_activ_xdl)
# ScaleAdd ScaleAdd Relu
add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp)
add_example_dependencies(example_convnd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16)
add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp)
add_example_dependencies(example_convnd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16)
example/64_fpAintB_gemm/CMakeLists.txt
View file @ 6b9a4bd5
if(GPU_TARGETS MATCHES "gfx11")
add_custom_target(example_fpAintB_gemm_wmma)
add_example_executable(example_fp16int8_gemm_wmma fp16int8_gemm_wmma.cpp)
add_dependencies(example_fpAintB_gemm_wmma example_fp16int8_gemm_wmma)
endif()
add_custom_target(example_fpAintB_gemm_wmma)
add_example_executable(example_fp16int8_gemm_wmma fp16int8_gemm_wmma.cpp)
add_example_dependencies(example_fpAintB_gemm_wmma example_fp16int8_gemm_wmma)
example/CMakeLists.txt
View file @ 6b9a4bd5
......@@ -5,6 +5,12 @@ include_directories(BEFORE
add_custom_target(examples)
function(add_example_dependencies EXAMPLE_NAME FILE_NAME)
if(FILE_NAME)
add_dependencies(EXAMPLE_NAME FILE_NAME)
endif()
endfunction(add_example_dependencies EXAMPLE_NAME)
function(add_example_executable EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
set(result 1)
......@@ -38,12 +44,27 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
endif()
endforeach()
endif()
#Do not build any DL examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
message("removing dl example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any XDL examples if gfx9 targets are not on the list
foreach(source IN LISTS FILE_NAME)
if(NOT GPU_TARGETS MATCHES "gfx9" AND source MATCHES "_xdl")
message("removing xdl example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any WMMA examples if gfx11 targets are not on the list
foreach(source IN LISTS FILE_NAME)
if(NOT GPU_TARGETS MATCHES "gfx11" AND source MATCHES "_wmma")
message("removing wmma example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(FILE_NAME)
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
......@@ -97,12 +118,27 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
endif()
endforeach()
endif()
#Do not build any DL examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
message("removing dl example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any XDL examples if gfx9 targets are not on the list
foreach(source IN LISTS FILE_NAME)
if(NOT GPU_TARGETS MATCHES "gfx9" AND source MATCHES "_xdl")
message("removing xdl example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any WMMA examples if gfx11 targets are not on the list
foreach(source IN LISTS FILE_NAME)
if(NOT GPU_TARGETS MATCHES "gfx11" AND source MATCHES "_wmma")
message("removing wmma example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(FILE_NAME)
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
......
include/ck/ck.hpp
View file @ 6b9a4bd5
......@@ -45,6 +45,10 @@
#endif
// define general macros for various architectures
#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
defined(__gfx942__)
#define __gfx9__
#endif
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#define __gfx94__
#endif
......@@ -62,8 +66,7 @@
// buffer resource
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx94__)
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx9__)
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#elif defined(__gfx103__)
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
......@@ -75,8 +78,7 @@
#ifndef __HIP_DEVICE_COMPILE__ // for host code, define nothing
#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
#define CK_USE_AMD_V_MAC_F32
#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx103__) || \
defined(__gfx94__) // for GPU code
#elif defined(__gfx906__) || defined(__gfx9__) || defined(__gfx103__) // for GPU code
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8
......@@ -89,7 +91,7 @@
// MFMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_MFMA
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) // for GPU code
#elif defined(__gfx9__) // for GPU code
#define CK_USE_AMD_MFMA
#endif
......@@ -120,7 +122,7 @@
// buffer atomic add: floating point
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) // for GPU code
#elif defined(__gfx9__) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#else // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
......
include/ck/config.h.in
View file @ 6b9a4bd5
......@@ -104,6 +104,20 @@
#cmakedefine CK_ENABLE_INSTANCES_ONLY @CK_ENABLE_INSTANCES_ONLY@
#endif
//
// CK kernels which support XDL (MI series)
//
#ifndef CK_USE_XDL
#cmakedefine CK_USE_XDL @CK_USE_XDL@
#endif
//
// CK Kernels which support WMMA (recent Navi series)
//
#ifndef CK_USE_WMMA
#cmakedefine CK_USE_WMMA @CK_USE_WMMA@
#endif
// clang-format on
#endif // CK_CONFIG_H_IN
include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r2.hpp 0 → 100644
View file @ 6b9a4bd5
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r2.hpp"
namespace ck {
/**
* @brief Blockwise data transfer
*
* This version does following things to avoid scratch memory issue
* 1. Use StaticallyIndexedArray instead of C array for thread buffer
* 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
* 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
*
*/
template <typename ThreadGroup,
typename ElementwiseOperation,
typename DstInMemOps, // Sequence
typename BlockSliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,
typename SrcDatas,
typename DstDatas,
typename SrcDescs,
typename DstDescs,
typename SrcDimAccessOrder,
typename DstDimAccessOrder,
index_t SrcVectorDim,
index_t DstVectorDim,
typename SrcsScalarPerVector, // Sequence
typename DstsScalarPerVector, // Sequence
typename SrcsScalarStrideInVector, // Sequence
typename DstsScalarStrideInVector, // Sequence
typename ThreadTransferSrcsResetCoordinateAfterRun, // Sequence
typename ThreadTransferDstsResetCoordinateAfterRun, // Sequence
index_t NumThreadScratch = 1>
struct ThreadGroupTensorSliceTransfer_v4r2
{
static constexpr index_t nDim =
remove_reference_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
static constexpr index_t nSrc = SrcDescs::Size();
static constexpr index_t nDst = DstDescs::Size();
static constexpr auto thread_slice_lengths = BlockSliceLengths{} / ThreadClusterLengths{};
using Index = MultiIndex<nDim>;
__device__ constexpr ThreadGroupTensorSliceTransfer_v4r2(
const SrcDescs& src_descs,
const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
const DstDescs& dst_descs,
const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
const ElementwiseOperation& element_op)
: threadwise_transfer_(src_descs,
StaticallyIndexedArray<Index, nSrc>{},
dst_descs,
StaticallyIndexedArray<Index, nDst>{},
element_op)
{
static_assert(nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == SrcDimAccessOrder::Size() && nDim == SrcDimAccessOrder::Size(),
"wrong! nDim not consistent");
static_for<0, nSrc, 1>{}([&](auto src_i) {
static_assert(nDim ==
remove_cvref_t<tuple_element_t<src_i, SrcDescs>>::GetNumOfDimension(),
"wrong! nDim not consistent");
});
static_for<0, nDst, 1>{}([&](auto dst_i) {
static_assert(nDim ==
remove_cvref_t<tuple_element_t<dst_i, DstDescs>>::GetNumOfDimension(),
"wrong! nDim not consistent");
});
static_assert(
is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"wrong! ThreadGroup::GetNumOfThread() too small");
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(ThreadGroup::GetThreadId()));
const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
const auto src_thread_slice_origins = generate_tuple(
[&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
Number<nSrc>{});
const auto dst_thread_slice_origins = generate_tuple(
[&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
Number<nDst>{});
threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
}
}
template <typename SrcBuffers, index_t ThreadScratchId = 0>
__device__ void RunRead(const SrcDescs& src_descs,
const SrcBuffers& src_bufs,
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_descs, src_bufs, thread_scratch_id);
}
}
template <typename DstBuffers, index_t ThreadScratchId = 0>
__device__ void RunWrite(const DstDescs& dst_descs,
DstBuffers& dst_bufs,
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunWrite(dst_descs, dst_bufs, thread_scratch_id);
}
}
template <typename SrcBuffer, typename DstBuffer, index_t ThreadScratchId>
__device__ void Run(const SrcDescs& src_descs,
const SrcBuffer& src_bufs,
const DstDescs& dst_descs,
DstBuffer& dst_bufs,
Number<ThreadScratchId> thread_scratch_id)
{
RunRead(src_descs, src_bufs, thread_scratch_id);
RunWrite(dst_descs, dst_bufs, thread_scratch_id);
}
__device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, const Index& step)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(src_descs, step);
}
}
__device__ void MoveDstSliceWindow(const DstDescs& dst_descs, const Index& step)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_descs, step);
}
}
private:
static constexpr auto thread_cluster_desc_ =
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v3r2<decltype(thread_slice_lengths),
ElementwiseOperation,
DstInMemOps,
SrcDatas,
DstDatas,
SrcDescs,
DstDescs,
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorDim,
DstVectorDim,
SrcsScalarPerVector,
DstsScalarPerVector,
SrcsScalarStrideInVector,
DstsScalarStrideInVector,
ThreadTransferSrcsResetCoordinateAfterRun,
ThreadTransferDstsResetCoordinateAfterRun,
NumThreadScratch>;
ThreadwiseTransfer threadwise_transfer_;
};
} // namespace ck
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp
View file @ 6b9a4bd5
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -40,7 +40,8 @@ using is_tuple = decltype(std::declval<T&>().IsTuple());
* \tparam AElementwiseOperation A elementwise operation.
* \tparam BElementwiseOperation B elementwise operation.
* \tparam CDEElementwiseOperation CDE elementwise operation.
* \tparam ComputeType Compute data type (default: ADataType, first if tuple passed).
* \tparam AComputeType Compute data type for A tensor (default: ADataType, first if tuple passed).
* \tparam BComputeType Compute data type for B tensor (default: AComputeType).
*/
template <index_t NDimSpatial,
typename ALayout,
......@@ -54,12 +55,13 @@ template <index_t NDimSpatial,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputeType =
typename AComputeType =
decltype(UnpackDataType<is_detected<is_tuple, ADataType>::value,
Number<0>,
ADataType>())> // ComputeType is InputType by default (first
ADataType>()), // AComputeType is InputType by default (first
// in tuple for MultiAB), unpack if tuple was
// passed
typename BComputeType = AComputeType>
struct DeviceGroupedConvFwdMultipleABD : public BaseOperator
{
static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_multiple_d_splitk.hpp 0 → 100644
View file @ 6b9a4bd5
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include <iostream>
#include <vector>
#include <sstream>
#include "device_grouped_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
///
/// @brief Structure representing single GEMM problem arguments.
///
/// The pointer to the vector of those structures is passed to the GroupedGEMM entry
/// point kernel.
///
/// @tparam NumDTensor The number of D input tensors.
///
template <index_t NumDTensor = 0>
struct GroupedGemmMultipleDKernelArguments
{
__host__ __device__
GroupedGemmMultipleDKernelArguments(const void* p_a_grid_,
const void* p_b_grid_,
std::array<const void*, NumDTensor> p_ds_grid_,
void* p_e_grid_,
index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
std::array<index_t, NumDTensor> StrideDs_,
index_t StrideE_)
: p_a_grid{p_a_grid_},
p_b_grid{p_b_grid_},
p_ds_grid{p_ds_grid_},
p_e_grid{p_e_grid_},
M{M_},
N{N_},
K{K_},
StrideA{StrideA_},
StrideB{StrideB_},
StrideDs{StrideDs_},
StrideE{StrideE_}
{
}
const void* p_a_grid;
const void* p_b_grid;
std::array<const void*, NumDTensor> p_ds_grid;
void* p_e_grid;
index_t M;
index_t N;
index_t K;
index_t StrideA;
index_t StrideB;
std::array<index_t, NumDTensor> StrideDs;
index_t StrideE;
void Print() const
{
std::stringstream str;
for(auto sd : StrideDs)
str << sd << ",";
std::cout << "arg {"
<< "M:" << M << ", "
<< "N:" << N << ", "
<< "K:" << K << ", "
<< "SA:" << StrideA << ", "
<< "SB:" << StrideB << ", "
<< "SE:" << StrideE << ", "
<< "SDs: {" << str.str() << "}"
<< "}" << std::endl;
}
};
template <typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceGroupedGemmMultipleDSplitK : public DeviceGroupedGemm<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
//----------------------------------------------------------------------------------------------
/// @brief Sets the k batch size.
///
/// @param p_arg Pointer to the Argument we're going to change.
/// @param[in] kbatch The kbatch value.
///
virtual void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const = 0;
//----------------------------------------------------------------------------------------------
/// @brief Sets the device kernel arguments pointer.
///
/// @param p_arg The pointer to the Argument we're going to update.
/// @param[in] p_dev_kernel_args The pointer to the device memory which contains kernel
/// arguments.
///
virtual void SetDeviceKernelArgs(BaseArgument* p_arg, void* p_dev_kernel_args) const = 0;
//----------------------------------------------------------------------------------------------
/// @brief Gets the device kernel argument size.
///
/// @param[in] p_arg The pointer to the Device op Argument.
///
/// @return The device kernel argument size.
///
virtual size_t GetDeviceKernelArgSize(const BaseArgument* p_arg) const = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp 0 → 100644
View file @ 6b9a4bd5
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/math.hpp"
#include "ck/utility/sequence.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_dynamic_vector_dims.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/stream_utility.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InDataTypeTuple,
typename OutDataTypeTuple,
typename ElementwiseOperation,
index_t NumDim,
index_t BlockSize,
index_t M0PerBlock,
index_t M1PerBlock,
index_t M0PerThread,
index_t M1PerThread,
typename ThreadClusterArrangeOrder,
typename InScalarPerVectorSeq,
typename OutScalarPerVectorSeq>
struct DeviceElementwiseImpl
: public DeviceElementwise<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>
{
static constexpr int NumInput = InDataTypeTuple::Size();
static constexpr int NumOutput = OutDataTypeTuple::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static_assert(NumInput == InScalarPerVectorSeq::Size() &&
NumOutput == OutScalarPerVectorSeq::Size(),
"Tuple size is inconsistent with the number of in/out!");
static auto GenerateInDataTypePointerTuple()
{
return generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
return static_cast<const DataType*>(nullptr);
},
Number<NumInput>{});
};
static auto GenerateOutDataTypePointerTuple()
{
return generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
return static_cast<DataType*>(nullptr);
},
Number<NumOutput>{});
};
using InDataTypePointerTuple = decltype(GenerateInDataTypePointerTuple());
using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
static index_t GetLowestStrideDim(const std::array<index_t, NumDim>& strides)
{
index_t most_continous_dim = NumDim - 1;
index_t most_continous_dim_stride = strides[most_continous_dim];
for(index_t dim = 0; dim < NumDim; dim++)
{
if(strides[dim] < most_continous_dim_stride)
{
most_continous_dim_stride = strides[dim];
most_continous_dim = dim;
}
}
return most_continous_dim;
}
template <typename InOutDescriptor>
static auto PadInputOutputDescriptor(const InOutDescriptor& desc)
{
const auto M0 = desc.GetLength(I0);
const auto M1 = desc.GetLength(I1);
const auto pad_M0 = math::integer_divide_ceil(M0, M0PerThread) * M0PerThread - M0;
const auto pad_M1 = math::integer_divide_ceil(M1, M1PerThread) * M1PerThread - M1;
const auto padded_desc = transform_tensor_descriptor(
desc,
make_tuple(make_right_pad_transform(M0, pad_M0), make_right_pad_transform(M1, pad_M1)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return padded_desc;
}
static auto GenerateBatchDimsLenghtsTuple(const std::array<index_t, NumDim>& lengths,
const index_t M0_dim,
const index_t M1_dim)
{
// Generate batch dims, they will be merged to M0
// Add one more dim than needed in case that M0 is equal to M1
// If M0 is equal to M1, then will be one more batch dim
std::array<index_t, NumDim - 1> batch_dims;
index_t batch_dim = 0;
for(index_t i = 0; i < NumDim; i++)
{
if(i != M0_dim && i != M1_dim)
{
batch_dims[batch_dim] = lengths[i];
batch_dim++;
}
}
// Add dummy dim if M0_dim is not equal to M1_dim
if(M0_dim != M1_dim && NumDim >= 2)
batch_dims[NumDim - 2] = 1;
return generate_tuple([&](auto I) { return batch_dims[I]; }, Number<NumDim - 1>{});
}
static auto MakeDescriptor(const std::array<index_t, NumDim>& lengths,
const std::array<index_t, NumDim>& in_strides,
const std::array<index_t, NumDim>& out_strides,
const std::array<index_t, NumDim>& desc_strides)
{
const auto M0_dim = GetLowestStrideDim(out_strides);
const auto M1_dim = GetLowestStrideDim(in_strides);
// If M0_dim is equal to M1_dim, then make M0_dim dummy
const auto M0 = M0_dim == M1_dim ? I1 : lengths[M0_dim];
const auto M1 = lengths[M1_dim];
const auto M0_stride = M0_dim == M1_dim ? I1 : desc_strides[M0_dim];
const auto M1_stride = desc_strides[M1_dim];
const auto batch_dims_lenghts = GenerateBatchDimsLenghtsTuple(lengths, M0_dim, M1_dim);
const auto batch_dims_strides = GenerateBatchDimsLenghtsTuple(desc_strides, M0_dim, M1_dim);
const auto desc = make_naive_tensor_descriptor(
concat_tuple(batch_dims_lenghts, make_tuple(M0), make_tuple(M1)),
concat_tuple(batch_dims_strides, make_tuple(M0_stride), make_tuple(M1_stride)));
// Merged batch dims with M0
const auto transforms =
make_tuple(make_merge_transform(concat_tuple(batch_dims_lenghts, make_tuple(M0))),
make_pass_through_transform(M1));
using BatchElemsSequence =
typename arithmetic_sequence_gen<0, decltype(batch_dims_lenghts)::Size() + 1, 1>::type;
const auto lower_dims = make_tuple(BatchElemsSequence{}, Sequence<NumDim>{});
const auto upper_dims = make_tuple(Sequence<0>{}, Sequence<1>{});
// desc: (merged_dims + M0, M1)
auto merged_desc = transform_tensor_descriptor(desc, transforms, lower_dims, upper_dims);
return PadInputOutputDescriptor(merged_desc);
}
template <index_t NumTensors>
static auto GenerateInOutGridDescTuple()
{
std::array<index_t, NumDim> ones;
for(index_t d = 0; d < NumDim; d++)
{
ones[d] = 1;
}
return generate_tuple([&](auto) { return MakeDescriptor(ones, ones, ones, ones); },
Number<NumTensors>{});
};
using InGridDescTuple = decltype(GenerateInOutGridDescTuple<NumInput>());
using OutGridDescTuple = decltype(GenerateInOutGridDescTuple<NumOutput>());
using Block2TileMap = BlockToCTileMap_M00_N0_M01Adapt<M0PerBlock, M1PerBlock>;
using GridwiseElementwiseOp = GridwiseElementwise<InGridDescTuple,
OutGridDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
Block2TileMap,
ElementwiseOperation,
BlockSize,
M0PerBlock,
M1PerBlock,
M0PerThread,
M1PerThread,
ThreadClusterArrangeOrder,
InScalarPerVectorSeq,
OutScalarPerVectorSeq,
false>;
using GridwiseElementwiseOpSameInOutVectorDim = GridwiseElementwise<InGridDescTuple,
OutGridDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
Block2TileMap,
ElementwiseOperation,
BlockSize,
M0PerBlock,
M1PerBlock,
M0PerThread,
M1PerThread,
ThreadClusterArrangeOrder,
InScalarPerVectorSeq,
OutScalarPerVectorSeq,
true>;
struct Argument : public BaseArgument
{
Argument(const std::array<index_t, NumDim> lengths,
const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
const std::array<const void*, NumInput> in_dev_buffers,
const std::array<void*, NumOutput> out_dev_buffers,
ElementwiseOperation elementwise_op)
: lengths_(lengths),
inStridesArray_(inStridesArray),
outStridesArray_(outStridesArray),
elementwise_op_(elementwise_op)
{
in_dev_buffers_ = generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
return static_cast<const DataType*>(in_dev_buffers[I.value]);
},
Number<NumInput>{});
out_dev_buffers_ = generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
return static_cast<DataType*>(out_dev_buffers[I.value]);
},
Number<NumOutput>{});
}
InDataTypePointerTuple in_dev_buffers_;
OutDataTypePointerTuple out_dev_buffers_;
std::array<index_t, NumDim> lengths_;
std::array<std::array<index_t, NumDim>, NumInput> inStridesArray_;
std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray_;
ElementwiseOperation elementwise_op_;
};
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
auto in_grid_desc_tuple = generate_tuple(
[&](auto src_i) {
// Use Strides from first tensor to assert that M0 dim and
// M1 dim are the same for each tensor.
return MakeDescriptor(arg.lengths_,
arg.inStridesArray_[I0],
arg.outStridesArray_[I0],
arg.inStridesArray_[src_i]);
},
Number<NumInput>{});
auto out_grid_desc_tuple = generate_tuple(
[&](auto dst_i) {
return MakeDescriptor(arg.lengths_,
arg.inStridesArray_[I0],
arg.outStridesArray_[I0],
arg.outStridesArray_[dst_i]);
},
Number<NumOutput>{});
const index_t M0 = in_grid_desc_tuple.At(I0).GetLength(Number<I0>{});
const index_t M1 = in_grid_desc_tuple.At(I0).GetLength(Number<I1>{});
const auto block_2_tile_map = Block2TileMap(M0, M1);
const index_t grid_size = block_2_tile_map.CalculateGridSize(M0, M1);
const bool in_out_same_vector_dim = GetLowestStrideDim(arg.inStridesArray_[I0]) ==
GetLowestStrideDim(arg.outStridesArray_[I0]);
const auto kernel = in_out_same_vector_dim
? kernel_elementwise<GridwiseElementwiseOpSameInOutVectorDim,
InGridDescTuple,
OutGridDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
Block2TileMap,
ElementwiseOperation>
: kernel_elementwise<GridwiseElementwiseOp,
InGridDescTuple,
OutGridDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
Block2TileMap,
ElementwiseOperation>;
float elapsed_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
in_grid_desc_tuple,
out_grid_desc_tuple,
arg.in_dev_buffers_,
arg.out_dev_buffers_,
block_2_tile_map,
arg.elementwise_op_);
return elapsed_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static bool IsSupportedArgument(const Argument& arg)
{
const index_t M0_dim = GetLowestStrideDim(arg.inStridesArray_[I0]);
const index_t M1_dim = GetLowestStrideDim(arg.outStridesArray_[I0]);
auto IsScalarPerVectorValid = [&](const std::array<index_t, NumDim>& lengths,
const std::array<index_t, NumDim>& strides,
index_t scalarPerVector,
index_t M_dim) {
if(scalarPerVector == 1)
{
return true;
}
if(strides[M_dim] == 1 && lengths[M_dim] % scalarPerVector == 0)
{
return true;
}
return false;
};
bool is_valid = true;
static_for<0, NumInput, 1>{}([&](auto I) {
static_assert(M0PerThread % InScalarPerVectorSeq::At(I) == 0 &&
M1PerThread % InScalarPerVectorSeq::At(I) == 0);
is_valid &= IsScalarPerVectorValid(
arg.lengths_, arg.inStridesArray_[I.value], InScalarPerVectorSeq::At(I), M0_dim);
});
static_for<0, NumOutput, 1>{}([&](auto I) {
static_assert(M0PerThread % OutScalarPerVectorSeq::At(I) == 0 &&
M1PerThread % OutScalarPerVectorSeq::At(I) == 0);
is_valid &= IsScalarPerVectorValid(
arg.lengths_, arg.outStridesArray_[I.value], OutScalarPerVectorSeq::At(I), M1_dim);
});
return is_valid;
};
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto
MakeArgument(const std::array<index_t, NumDim> lengths,
const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
const std::array<const void*, NumInput> in_dev_buffers,
const std::array<void*, NumOutput> out_dev_buffers,
ElementwiseOperation elementwise_op)
{
return Argument{lengths,
inStridesArray,
outStridesArray,
in_dev_buffers,
out_dev_buffers,
elementwise_op};
}
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
const std::array<const void*, NumInput> in_dev_buffers,
const std::array<void*, NumOutput> out_dev_buffers,
ElementwiseOperation elementwise_op) override
{
return std::make_unique<Argument>(lengths,
inStridesArray,
outStridesArray,
in_dev_buffers,
out_dev_buffers,
elementwise_op);
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>();
};
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceElementwiseImpl<";
str << NumDim << ", ";
str << BlockSize << ", ";
str << M0PerBlock << ", ";
str << M1PerBlock << ", ";
str << M0PerThread << ", ";
str << M1PerThread << ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
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