Skip to content

GitLab

Commit 0c5e5fc5 authored by turneram's avatar turneram
Browse files

Merge branch 'ck-elementwise2' into ck-poc

parents 65b6a759 0e237605
Hide whitespace changes
Inline Side-by-side
Showing with 215 additions and 67 deletions
src/targets/gpu/jit/ck_elementwise.cpp
View file @ 0c5e5fc5
......@@ -25,6 +25,8 @@
#include <migraphx/make_op.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/compile_gen.hpp>
#include <migraphx/gpu/compile_hip_code_object.hpp>
#include <migraphx/gpu/compile_hip.hpp>
#include <migraphx/ranges.hpp>
......@@ -39,6 +41,113 @@ namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
using namespace migraphx::gpu::gen; // NOLINT
// static const char* const ck_elementwise_kernel = R"__migraphx__(
// //#include <migraphx/kernels/ck_elementwise.hpp>
// #include <migraphx/kernels/ops.hpp>
// #include <migraphx/kernels/integral_constant.hpp>
// #include <migraphx/kernels/generic_constant.hpp>
// #include <args.hpp>
// #include <migraphx/kernels/index.hpp>
// #include <migraphx/kernels/algorithm.hpp>
// #include <migraphx/kernels/integral_constant.hpp>
// #include <migraphx/kernels/tensor_view.hpp>
// #include "ck/device_utility/device_prop.hpp"
// #include "ck/device_utility/kernel_launch.hpp"
// #include "ck/tensor_operation/gpu/device/device_base.hpp"
// #include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
// #include "ck/tensor_operation/gpu/grid/gridwise_binary_elementwise_1d.hpp"
// namespace migraphx {
// using ADataType = float;
// using BDataType = float;
// using CDataType = float;
// using ElementwiseFunctor = float;
// static constexpr auto I0 = ck::Number<0>{};
// template <class L, class S, class N>
// constexpr auto MakeDescriptor_M(const L& lengths, const S& strides, const N& ndim)
// {
// auto gridSize = 72;
// auto blockSize = 1024;
// //constexpr auto ndim = 1;
// // auto idx = make_index();
// auto tupleOfShape = generate_tuple([&](auto I) { return static_cast<ck::index_t>(lengths[I]);
// },
// ck::Number<ndim>{});
// auto tupleOfStride = generate_tuple(
// [&](auto I) { return static_cast<ck::index_t>(strides[I]); }, ck::Number<1>{});
// const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
// auto desc_m = desc;
// // merge nd to 1d desc - [s0 * s1 * ...]
// if constexpr(ndim > 1)
// {
// desc_m = transform_tensor_descriptor(
// desc,
// make_tuple(make_merge_transform(tupleOfShape)),
// make_tuple(generate_sequence_v2([&](auto I) { return I; }, ck::Number<ndim>{})),
// make_tuple(ck::Sequence<0>{}));
// }
// const auto M = desc_m.GetLength(I0);
// const ck::index_t loop_step = /* idx.nglobal(); // */ gridSize * blockSize /* * MPerThread
// */; const auto pad = ck::math::integer_least_multiple(M, loop_step) - M; const
// auto desc_m_pad =
// transform_tensor_descriptor(desc_m,
// make_tuple(ck::make_right_pad_transform(M, pad)),
// make_tuple(ck::Sequence<0>{}),
// make_tuple(ck::Sequence<0>{}));
// return desc_m_pad;
// }
// struct Add
// {
// template <typename Y, typename X0, typename X1>
// __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const
// {
// y = x0 + x1;
// };
// };
// extern "C" {
// __global__ void ck_elementwise_kernel(void* a_p, void* b_p, void* c_p)
// {
// make_tensors()(a_p, b_p, c_p)([](auto a_t, auto b_t, auto c_t) {
// constexpr auto lengths = get_shape_c<decltype(a_t)>{}.lens;
// constexpr auto strides = get_shape_c<decltype(a_t)>{}.strides;
// constexpr auto ndim = _c<decltype(lengths.size()){}>[1];
// constexpr auto a_desc = MakeDescriptor_M(lengths, strides, ndim);
// using AGridDesc_M = decltype(a_desc);
// using GridwiseBinEltwise = ck::GridwiseBinaryElementwise_1D<ADataType,
// BDataType,
// CDataType,
// CDataType,
// AGridDesc_M,
// AGridDesc_M,
// AGridDesc_M,
// Add,
// 1,
// 1,
// 1,
// 1>;
// auto op = Add{};
// GridwiseBinEltwise::Run(a_t.data(), b_t.data(), c_t.data(), a_desc, a_desc, a_desc, op);
// });
// }
// }
// } // namespace migraphx
// )__migraphx__";
// NOLINTNEXTLINE
static const char* const ck_elementwise_kernel = R"__migraphx__(
#include <migraphx/kernels/ck_elementwise.hpp>
......@@ -51,10 +160,10 @@ namespace migraphx {
extern "C" {
__global__ void ck_elementwise_kernel(void* a_p, void* b_p, void* c_p)
__global__ void ck_elementwise_kernel(void* a_p, void* b_p, void* c_p)
{
make_tensors()(a_p, b_p, c_p)([](auto&&... xs) {
ck_elementwise(xs...);
make_tensors()(a_p, b_p, c_p)([](auto&&... xs) {
ck_elementwise(xs...);
});
}
......@@ -68,16 +177,30 @@ struct ck_elementwise_compiler : compiler<ck_elementwise_compiler>
{
std::vector<std::string> names() const { return {"ck_elementwise"}; }
static std::size_t oversubscribe_if(bool b)
{
if(b)
return 256;
else
return 1;
}
operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
{
hip_compile_options options;
auto out_s = inputs.back();
options.set_launch_params(v, compute_global_for(ctx, out_s.elements()));
options.inputs = inputs;
options.output = out_s;
options.output = inputs.back();
options.virtual_inputs = reduce_dims(inputs);
options.params = "-Wno-float-equal";
auto axis = find_fast_axis(options.virtual_inputs);
auto vec = vectorize::elements(axis, options.virtual_inputs);
auto preloads = preload::broadcasts(axis, options.virtual_inputs);
options.kernel_name = "ck_elementwise_kernel";
options.virtual_inputs = inputs;
options.set_launch_params(
v,
compute_global_for(ctx,
options.output.elements() / (vec.size * 4),
oversubscribe_if(not preloads.is_preloading())));
return compile_hip_code_object(ck_elementwise_kernel, options);
}
......
src/targets/gpu/kernels/include/migraphx/kernels/ck_elementwise.hpp
View file @ 0c5e5fc5
......@@ -37,46 +37,57 @@
namespace migraphx {
using ADataType = float;
using BDataType = float;
using CDataType = float;
using ElementwiseFunctor = float;
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using ElementwiseFunctor = ck::half_t;
static constexpr auto I0 = ck::Number<0>{};
template <class L, class S, class N>
constexpr auto MakeDescriptor_M(const L& lengths, const S& strides, const N& /* ndim */)
template <ck::index_t ndim>
struct CKBinaryElementwise
{
auto gridSize = 72;
auto blockSize = 1024;
constexpr auto ndim = 1;
// auto idx = make_index();
auto tupleOfShape = generate_tuple([&](auto I) { return static_cast<ck::index_t>(lengths[I]); },
ck::Number<ndim>{});
auto tupleOfStride = generate_tuple(
[&](auto I) { return static_cast<ck::index_t>(strides[I]); }, ck::Number<1>{});
const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
auto desc_m = desc;
// merge nd to 1d desc - [s0 * s1 * ...]
if constexpr(ndim > 1)
template <class Desc_M>
constexpr auto PadDescriptor_M_1d(Desc_M desc_m)
{
desc_m = transform_tensor_descriptor(
desc,
make_tuple(make_merge_transform(tupleOfShape)),
make_tuple(generate_sequence_v2([&](auto I) { return I; }, ck::Number<ndim>{})),
make_tuple(ck::Sequence<0>{}));
auto gridSize = 72;
auto blockSize = 1024;
auto MPerThread = 8;
const auto M = desc_m.GetLength(I0);
const ck::index_t loop_step = gridSize * blockSize * MPerThread;
const auto pad = ck::math::integer_least_multiple(M, loop_step) - M;
const auto desc_m_pad =
transform_tensor_descriptor(desc_m,
make_tuple(ck::make_right_pad_transform(M, pad)),
make_tuple(ck::Sequence<0>{}),
make_tuple(ck::Sequence<0>{}));
return desc_m_pad;
}
const auto M = desc_m.GetLength(I0);
const ck::index_t loop_step = /* idx.nglobal(); // */ gridSize * blockSize /* * MPerThread */;
const auto pad = ck::math::integer_least_multiple(M, loop_step) - M;
const auto desc_m_pad =
transform_tensor_descriptor(desc_m,
make_tuple(ck::make_right_pad_transform(M, pad)),
make_tuple(ck::Sequence<0>{}),
make_tuple(ck::Sequence<0>{}));
return desc_m_pad;
}
template <class L, class S>
constexpr auto MakeDescriptor_M(const L& lengths, const S& strides)
{
auto tupleOfShape = generate_tuple(
[&](auto I) { return static_cast<ck::index_t>(lengths[I]); }, ck::Number<ndim>{});
auto tupleOfStride = generate_tuple(
[&](auto I) { return static_cast<ck::index_t>(strides[I]); }, ck::Number<ndim>{});
const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
// merge nd to 1d desc - [s0 * s1 * ...]
if constexpr(ndim > 1)
{
const auto desc_m = transform_tensor_descriptor(
desc,
make_tuple(make_merge_transform(tupleOfShape)),
make_tuple(generate_sequence_v2([&](auto I) { return I; }, ck::Number<ndim>{})),
make_tuple(ck::Sequence<0>{}));
return PadDescriptor_M_1d(desc_m);
}
else
{
return PadDescriptor_M_1d(desc);
}
}
};
struct Add
{
......@@ -90,29 +101,41 @@ struct Add
template <class T, class U, class V>
__device__ void ck_elementwise(const T& a_t, const U& b_t, const V& c_t)
{
auto idx = make_index();
if(idx.global == 0)
{
constexpr auto lengths = get_shape_c<T>{}.lens;
constexpr auto strides = get_shape_c<T>{}.strides;
constexpr auto a_desc = MakeDescriptor_M(lengths, strides, 1);
constexpr auto a_lens = get_shape_c<T>{}.lens;
constexpr auto a_strides = get_shape_c<T>{}.strides;
constexpr ck::index_t a_ndim = decltype(a_lens.size()){};
auto a_bin_op = CKBinaryElementwise<a_ndim>{};
constexpr auto a_desc = a_bin_op.MakeDescriptor_M(a_lens, a_strides);
using AGridDesc_M = decltype(a_desc);
using GridwiseBinEltwise = ck::GridwiseBinaryElementwise_1D<ADataType,
BDataType,
CDataType,
CDataType,
AGridDesc_M,
AGridDesc_M,
AGridDesc_M,
Add,
1,
1,
1,
1>;
auto op = Add{};
GridwiseBinEltwise::Run(a_t.data(), b_t.data(), c_t.data(), a_desc, a_desc, a_desc, op);
}
constexpr auto b_lens = get_shape_c<U>{}.lens;
constexpr auto b_strides = get_shape_c<U>{}.strides;
constexpr ck::index_t b_ndim = decltype(b_lens.size()){};
auto b_bin_op = CKBinaryElementwise<b_ndim>{};
constexpr auto b_desc = b_bin_op.MakeDescriptor_M(b_lens, b_strides);
constexpr auto c_lens = get_shape_c<V>{}.lens;
constexpr auto c_strides = get_shape_c<V>{}.strides;
constexpr ck::index_t c_ndim = decltype(c_lens.size()){};
auto c_bin_op = CKBinaryElementwise<c_ndim>{};
constexpr auto c_desc = c_bin_op.MakeDescriptor_M(c_lens, c_strides);
using AGridDesc_M = decltype(a_desc);
using BGridDesc_M = decltype(b_desc);
using CGridDesc_M = decltype(c_desc);
using GridwiseBinEltwise = ck::GridwiseBinaryElementwise_1D<ADataType,
BDataType,
CDataType,
CDataType,
AGridDesc_M,
BGridDesc_M,
CGridDesc_M,
Add,
8,
8,
8,
8>;
auto op = Add{};
GridwiseBinEltwise::Run(a_t.data(), b_t.data(), c_t.data(), a_desc, b_desc, c_desc, op);
}
} // namespace migraphx
......
test/verify/0ck_gemm_test.cpp → test/verify/0ck_elementwise_half_test.cpp
View file @ 0c5e5fc5
......@@ -27,18 +27,20 @@
#include <migraphx/generate.hpp>
#include <migraphx/make_op.hpp>
struct ck_gemm : verify_program<ck_gemm>
struct ck_elementwise_half : verify_program<ck_elementwise_half>
{
migraphx::program create_program() const
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape m1_shape{migraphx::shape::float_type, {128, 256}};
migraphx::shape m2_shape{migraphx::shape::float_type, {256, 256}};
migraphx::shape m1_shape{migraphx::shape::half_type, {2, 384, 3072}};
migraphx::shape m2_shape{migraphx::shape::half_type, {3072}};
auto l1 = mm->add_parameter("1", m1_shape);
auto l2 = mm->add_parameter("2", m2_shape);
l2 = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {2, 384, 3072}}}), l2);
mm->add_instruction(migraphx::make_op("ck_gemm"), l1, l2);
mm->add_instruction(migraphx::make_op("ck_elementwise"), l1, l2);
return p;
}
......
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