Skip to content

GitLab

"nndet/arch/__init__.py" did not exist on "4560ce77c3d12bee24f6ac6a2a6b950b930c15d2"
Commit ac04f3cc authored by Khalique Ahmed's avatar Khalique Ahmed
Browse files

manual_merge

parents d39c3343 d8011adf
Hide whitespace changes
Inline Side-by-side
Showing with 686 additions and 595 deletions
src/targets/gpu/jit/ck_gemm.cpp
View file @ ac04f3cc
......@@ -27,6 +27,7 @@
#include <migraphx/make_op.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/ck.hpp>
#include <migraphx/env.hpp>
#include <migraphx/file_buffer.hpp>
#include <migraphx/gpu/compile_gen.hpp>
......@@ -37,8 +38,6 @@
#include <migraphx/reduce_dims.hpp>
#include <migraphx/stringutils.hpp>
#include "ck/host/device_gemm_multiple_d.hpp"
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -46,12 +45,6 @@ namespace gpu {
using namespace migraphx::gpu::gen; // NOLINT
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_LOG_CK_GEMM);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_TUNING);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_TUNING_VALUE);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_DEBUG);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TUNE_CK);
// NOLINTNEXTLINE
static const char* const ck_gemm_kernel = R"__migraphx__(
#include <args.hpp>
......@@ -79,220 +72,10 @@ MIGRAPHX_GLOBAL void ${kernel}(${params})
)__migraphx__";
// NOLINTNEXTLINE
static const char* const disable_warning_pragma = R"__migraphx__(
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Weverything"
${content}
#pragma clang diagnostic pop
)__migraphx__";
template <class P>
static std::string ck_disable_warnings(P p)
{
return interpolate_string(disable_warning_pragma,
{{"content", std::string{p.first, p.second}}});
}
static std::unordered_map<std::string, std::string> create_ck_header_strings()
{
std::unordered_map<std::string, std::string> result;
auto ck_headers = ck::host::GetHeaders();
std::transform(
ck_headers.begin(), ck_headers.end(), std::inserter(result, result.begin()), [&](auto&& p) {
return std::make_pair(p.first, ck_disable_warnings(p.second));
});
return result;
}
static std::vector<src_file> create_ck_headers()
{
static const auto& header_strings = create_ck_header_strings();
std::vector<src_file> srcs;
std::transform(
header_strings.begin(), header_strings.end(), std::back_inserter(srcs), [&](auto&& p) {
return src_file{fs::path{p.first},
{p.second.data(), p.second.data() + p.second.size()}};
});
return srcs;
}
static const std::vector<src_file>& ck_headers()
{
static const auto& headers = create_ck_headers();
return headers;
}
static bool transposed_matrix(const shape& s) { return s.strides().back() != 1; }
using tuning_entry = std::pair<std::vector<shape>, size_t>;
static std::vector<tuning_entry> read_tuning(const std::string& s)
{
if(not fs::exists(s))
return {};
return from_value<std::vector<tuning_entry>>(from_json_string(read_string(s)));
}
static float matrix_distance(const shape& x, const shape& y)
{
if(x.type() != y.type())
return std::numeric_limits<float>::max();
if(transposed_matrix(x) != transposed_matrix(y))
return std::numeric_limits<float>::max();
auto sum_squared = std::inner_product(x.lens().rbegin(),
x.lens().rbegin() + 2,
y.lens().rbegin(),
0,
std::plus<>{},
[](auto a, auto b) { return (a - b) * (a - b); });
return std::sqrt(sum_squared);
}
static std::size_t get_tuning_for(const std::vector<shape>& inputs)
{
static auto tuning = read_tuning(string_value_of(MIGRAPHX_CK_TUNING{}, ""));
if(tuning.empty())
{
std::cout << "*********** Warning: No CK tuning! for config:" << std::endl;
std::cout << " " << inputs[0] << std::endl;
std::cout << " " << inputs[1] << std::endl;
std::cout << " " << inputs[2] << std::endl;
}
auto it = std::find_if(
tuning.begin(), tuning.end(), [&](const auto& p) { return p.first == inputs; });
if(it == tuning.end())
{
std::cout << "*********** Warning: CK tuning missing for config!" << std::endl;
std::cout << " " << inputs[0] << std::endl;
std::cout << " " << inputs[1] << std::endl;
std::cout << " " << inputs[2] << std::endl;
std::vector<std::pair<float, std::size_t>> w;
std::transform(tuning.begin(), tuning.end(), std::back_inserter(w), [&](const auto& p) {
if(inputs.size() < 3 or p.first.size() < 3)
MIGRAPHX_THROW("Invalid CK config");
auto avg_distance = std::inner_product(
p.first.begin(),
p.first.begin() + 3,
inputs.begin(),
0.0f,
std::plus<>{},
[](const auto& x, const auto& y) { return matrix_distance(x, y) / 3.0f; });
return std::make_pair(avg_distance, p.second);
});
std::sort(w.begin(), w.end());
std::size_t default_value = 4;
if(not w.empty())
default_value = w.front().second;
auto tuning_val = value_of(MIGRAPHX_CK_TUNING_VALUE{}, default_value);
std::cout << "*********** Warning: CK try tuning: " << tuning_val << std::endl;
return tuning_val;
}
return it->second;
}
struct ck_gemm_compiler : compiler<ck_gemm_compiler>
{
static std::string get_layout(const shape& s)
{
return transposed_matrix(s) ? "ck::tensor_layout::gemm::ColumnMajor"
: "ck::tensor_layout::gemm::RowMajor";
}
static ck::host::DataType get_type(const shape& s)
{
if(s.type() == shape::half_type)
return ck::host::DataType::Half;
else if(s.type() == shape::float_type)
return ck::host::DataType::Float;
else if(s.type() == shape::int8_type)
return ck::host::DataType::Int8;
else if(s.type() == shape::int32_type)
return ck::host::DataType::Int32;
MIGRAPHX_THROW("Unsupported ck type");
}
template <class Iterator, class F>
static std::string ck_tuple(Iterator start, Iterator last, F f)
{
std::vector<std::string> s;
std::transform(start, last, std::back_inserter(s), f);
return "ck::Tuple<" + join_strings(s, ",") + ">";
}
static std::vector<shape> adjust_inputs(std::vector<shape> inputs, bool& swap_inputs)
{
swap_inputs = false;
auto c_shape = inputs.back();
if(not transposed_matrix(c_shape))
return inputs;
std::vector<int64_t> perm(c_shape.lens().size());
std::iota(perm.begin(), perm.end(), 0);
std::swap(perm[perm.size() - 1], perm[perm.size() - 2]);
std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](shape s) {
return reorder_shape(s, perm);
});
swap_inputs = true;
return inputs;
}
static std::size_t get_batch_count(const shape& s)
{
return std::accumulate(
s.lens().rbegin() + 2, s.lens().rend(), std::size_t{1}, std::multiplies<std::size_t>());
}
static void fold_batch_dims(shape& s)
{
auto lens = s.lens();
if(lens.size() <= 2)
return;
auto batch_count = get_batch_count(s);
auto m1 = lens.at(lens.size() - 2);
auto m2 = lens.at(lens.size() - 1);
if(transposed_matrix(s))
s = shape{s.type(), {m1, m2 * batch_count}};
else
s = shape{s.type(), {m1 * batch_count, m2}};
}
static void remove_batch_dims(shape& s)
{
auto lens = s.lens();
if(lens.size() <= 2)
return;
auto m1 = lens.at(lens.size() - 2);
auto m2 = lens.at(lens.size() - 1);
s = shape{s.type(), {m1, m2}};
}
std::vector<std::string> names() const { return {"ck_gemm", "gpu::ck_gemm"}; }
static bool standard_batch(const shape& s)
{
if(s.lens().size() < 3)
return true;
std::vector<std::size_t> lens(s.lens().begin(), s.lens().end() - 2);
std::vector<std::size_t> strides(s.strides().begin(), s.strides().end() - 2);
auto base = *(s.lens().end() - 2) * *(s.lens().end() - 1);
std::transform(strides.begin(), strides.end(), strides.begin(), [&](auto stride) {
return stride / base;
});
return shape{s.type(), lens, strides}.standard();
}
bool can_fold_batch(const std::vector<shape>& inputs) const
{
const auto& b_shape = inputs[1];
if(std::any_of(inputs.begin() + 2, inputs.end() - 1, [](auto input) {
return not standard_batch(input);
}))
return false;
const auto& b_strides = b_shape.strides();
return std::all_of(
b_strides.begin(), b_strides.end() - 2, [](auto stride) { return stride == 0; });
}
ck::host::device_gemm_multiple_d::Problem create_problem(const std::vector<shape>& inputs,
const value& v) const
{
......@@ -301,8 +84,7 @@ struct ck_gemm_compiler : compiler<ck_gemm_compiler>
const auto& c_shape = inputs.back();
// cppcheck-suppress unreadVariable
auto rank = a_shape.ndim();
auto rank = a_shape.ndim();
auto batch_count = get_batch_count(c_shape);
auto m = c_shape.lens()[rank - 2];
m = can_fold_batch(inputs) ? m * batch_count : m;
......@@ -352,12 +134,8 @@ struct ck_gemm_compiler : compiler<ck_gemm_compiler>
operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
{
const auto& a_shape = inputs[0];
const auto& b_shape = inputs[1];
const auto& c_shape = inputs.back();
auto tuning_value = v.get("tuning_value", 4);
if(not v.contains("tuning_value"))
tuning_value = get_tuning_for({a_shape, b_shape, c_shape});
auto tuning_value = v.get("tuning_value", 34);
auto batch_count = get_batch_count(c_shape);
auto problem = create_problem(inputs, v);
......
src/targets/gpu/jit/ck_gemm_softmax_gemm.cpp 0 → 100644
View file @ ac04f3cc
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <fstream>
#include <migraphx/filesystem.hpp>
#include <migraphx/gpu/compiler.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/env.hpp>
#include <migraphx/file_buffer.hpp>
#include <migraphx/gpu/ck.hpp>
#include <migraphx/gpu/compile_gen.hpp>
#include <migraphx/gpu/compile_hip.hpp>
#include <migraphx/gpu/compile_hip_code_object.hpp>
#include <migraphx/module.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/reduce_dims.hpp>
#include <migraphx/stringutils.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
using namespace migraphx::gpu::gen; // NOLINT
// NOLINTNEXTLINE
static const char* const ck_gemm_softmax_gemm_kernel = R"__migraphx__(
#include <args.hpp>
#include <migraphx/kernels/ck_gemm_softmax_gemm.hpp>
#include <migraphx/kernels/pointwise.hpp>
#include <migraphx/kernels/ops.hpp>
#include <migraphx/kernels/integral_constant.hpp>
#include <migraphx/kernels/generic_constant.hpp>
#include <${include}>
namespace migraphx {
${preamble}
extern "C" {
MIGRAPHX_GLOBAL void ${kernel}(${params})
{
transform_args(make_tensors(), rotate_last())(${args})([](auto... xs) {
auto settings = make_ck_gemm_softmax_gemm_settings(MIGRAPHX_MAKE_CONSTANT(float{SCALE}));
ck_gemm_softmax_gemm<${solution}, ${blocks_per_batch}>(settings, xs...);
});
}
}
} // namespace migraphx
)__migraphx__";
struct ck_gemm_softmax_gemm_compiler : compiler<ck_gemm_softmax_gemm_compiler>
{
std::vector<std::string> names() const
{
return {"ck_gemm_softmax_gemm", "gpu::ck_gemm_softmax_gemm"};
}
ck::host::device_batched_gemm_softmax_gemm::Problem
create_problem(const std::vector<shape>& inputs, const value&) const
{
const auto& a_shape = inputs[0];
const auto& b_shape = inputs[1];
const auto& b1_shape = inputs[2];
const auto& c_shape = inputs.back();
// cppcheck-suppress unreadVariable
auto rank = a_shape.ndim();
auto batch_count = get_batch_count(c_shape);
auto m = c_shape.lens()[rank - 2];
m = can_fold_batch(inputs) ? m * batch_count : m;
auto n = c_shape.lens().back();
auto k = a_shape.lens().back();
auto o = c_shape.lens().back();
const bool trans_a = transposed_matrix(a_shape);
const bool trans_b = transposed_matrix(b_shape);
const bool trans_b1 = transposed_matrix(b1_shape);
const bool trans_c = transposed_matrix(c_shape);
const auto a_type = get_type(a_shape);
const auto b_type = get_type(b_shape);
const auto b1_type = get_type(b1_shape);
const auto c_type = get_type(c_shape);
std::string ck_passthrough = "ck_passthrough";
return ck::host::device_batched_gemm_softmax_gemm::Problem{m,
n,
k,
o,
trans_a,
trans_b,
trans_b1,
trans_c,
a_type,
b_type,
b1_type,
c_type,
ck_passthrough,
ck_passthrough,
ck_passthrough,
ck_passthrough};
}
operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
{
const auto& c_shape = inputs.back();
auto tuning_value = v.get("tuning_value", 5);
auto batch_count = get_batch_count(c_shape);
auto problem = create_problem(inputs, v);
const auto include_header = problem.GetIncludeHeader();
const auto solutions = problem.GetSolutions(ctx.get_current_device().get_gfx_name());
const auto& solution = solutions.at(tuning_value);
const auto template_str = solution.template_str;
const auto blocks_per_batch = solution.grid_size;
const auto block_size = solution.block_size;
hip_compile_options options;
options.additional_src_files = ck_headers();
auto grid_size = can_fold_batch(inputs) ? blocks_per_batch : batch_count * blocks_per_batch;
options.set_launch_params(v, grid_size * block_size, block_size);
options.inputs = inputs;
options.output = c_shape;
options.kernel_name = v.get("kernel", "ck_gemm_softmax_gemm_kernel");
options.virtual_inputs = inputs;
if(can_fold_batch(inputs))
{
auto vinputs = inputs;
fold_batch_dims(vinputs[0]);
remove_batch_dims(vinputs[1]);
std::for_each(vinputs.begin() + 2, vinputs.end(), fold_batch_dims);
options.virtual_inputs = vinputs;
}
if(v.get("check", false) or enabled(MIGRAPHX_CK_DEBUG{}))
options.params += " -DMIGRAPHX_CK_CHECK=1";
// scale
assert(v.contains("scale"));
auto scale = v.at("scale").to<float>();
options.params += " -DSCALE=" + std::to_string(scale);
auto src = interpolate_string(ck_gemm_softmax_gemm_kernel,
{{"solution", template_str},
{"include", include_header},
{"params", enum_params(inputs.size(), "void * private_p")},
{"args", enum_params(inputs.size(), "private_p")},
{"blocks_per_batch", to_string(blocks_per_batch)},
{"preamble", v.get("preamble", std::string{})},
{"kernel", options.kernel_name}});
return compile_hip_code_object(src, options);
}
value create_settings(instruction_ref ins, const operation& op) const
{
auto v = op.to_value();
v["kernel"] = "ck_gemm_softmax_gemm_kernel";
if(not ins->module_inputs().empty())
{
auto* pm = ins->module_inputs().front();
v["preamble"] = generate_pointwise(*pm, "post_ck_gemm_softmax_gemm_function") +
"\nMIGRAPHX_LIFT_CLASS(post_ck_gemm_softmax_gemm, "
"post_ck_gemm_softmax_gemm_function);";
v["post"] = "ck_function_adaptor<post_ck_gemm_softmax_gemm>";
v["kernel"] = "ck_gemm_softmax_gemm_" + generate_name_from_ops(*pm) + "_kernel";
}
return v;
}
compiler_replace
compile(context& ctx, instruction_ref ins, const operation& op, const value& solution) const
{
auto shapes = to_shapes(ins->inputs());
auto v = create_settings(ins, op);
if(not solution.is_null())
v["tuning_value"] = solution;
return {compile_op(ctx, shapes, v),
[=](module& m, instruction_ref ins2, const operation& code_object) {
if(enabled(MIGRAPHX_LOG_CK_GEMM{}))
{
std::vector<shape> gemm_shapes{
shapes[0], shapes[1], shapes.back().with_type(shapes[0].type())};
std::cout << "gpu::ck_gemm_softmax_gemm: "
<< to_json_string(to_value(gemm_shapes)) << std::endl;
}
m.replace_instruction(ins2, code_object, ins2->inputs());
}};
}
optional<tuning_config>
get_tuning_config(context& ctx, instruction_ref ins, const operation& op, bool exhaustive) const
{
if(not exhaustive and not enabled(MIGRAPHX_TUNE_CK{}))
return nullopt;
tuning_config tc;
auto shapes = to_shapes(ins->inputs());
auto problem = create_problem(shapes, create_settings(ins, op));
auto solutions = problem.GetSolutions(ctx.get_current_device().get_gfx_name());
tc.solutions.resize(solutions.size());
std::iota(tc.solutions.begin(), tc.solutions.end(), 0);
std::vector<shape> gemm_shapes{shapes[0], shapes[1], shapes.back()};
tc.problem = to_value(gemm_shapes);
return tc;
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/jit/mlir.cpp
View file @ ac04f3cc
......@@ -57,11 +57,9 @@ struct mlir_compiler : compiler<mlir_compiler>
const operation&,
bool exhaustive) const
{
if(not exhaustive)
return nullopt;
auto shapes = to_shapes(ins->inputs());
auto* smod = ins->module_inputs().front();
return get_tuning_config_mlir(ctx, *smod, shapes);
return get_tuning_config_mlir(ctx, *smod, shapes, exhaustive);
}
};
......
src/targets/gpu/jit/roialign.cpp
View file @ ac04f3cc
......@@ -81,7 +81,7 @@ struct roialign_compiler : compiler<roialign_compiler>
// coord_trans_mode
auto ctm = v.at("coordinate_transformation_mode").to<std::string>();
float rois_offset = (ctm == "output_half_pixel") ? -0.5f : 0.0f;
float rois_offset = (ctm == "half_pixel") ? -0.5f : 0.0f;
options.params += " -DROIS_OFFSET=" + std::to_string(rois_offset);
// spatial_scale
......
src/targets/gpu/kernels/include/migraphx/kernels/ck.hpp
View file @ ac04f3cc
......@@ -154,6 +154,17 @@ struct ck_add
}
};
// In CK, the B matrix is ordered as N,K instead of K,N
template <class Dims>
constexpr auto ck_transposeb_dims(Dims dims)
{
return unpack(dims, [](auto k, auto n) { return make_const_array(n, k); });
}
template <class Tensor>
using ck_transposeb = decltype(make_shape(ck_transposeb_dims(get_shape_c<Tensor>{}.lens),
ck_transposeb_dims(get_shape_c<Tensor>{}.strides)));
#ifdef MIGRAPHX_CK_CHECK
#define MIGRAPHX_CK_STATIC_ASSERT static_assert
#else
......
src/targets/gpu/kernels/include/migraphx/kernels/ck_gemm.hpp
View file @ ac04f3cc
......@@ -33,17 +33,6 @@
namespace migraphx {
// In CK, the B matrix is ordered as N,K instead of K,N
template <class Dims>
constexpr auto ck_transposeb_dims(Dims dims)
{
return unpack(dims, [](auto k, auto n) { return make_const_array(n, k); });
}
template <class Tensor>
using ck_transposeb = decltype(make_shape(ck_transposeb_dims(get_shape_c<Tensor>{}.lens),
ck_transposeb_dims(get_shape_c<Tensor>{}.strides)));
template <class G, class E, class A, class B, class... Ds>
__device__ void ck_gemm_matrix(E e, A a, B b, Ds... ds)
{
......
src/targets/gpu/kernels/include/migraphx/kernels/ck_gemm_softmax_gemm.hpp 0 → 100644
View file @ ac04f3cc
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_KERNELS_CK_GEMM_SOFTMAX_GEMM_HPP
#define MIGRAPHX_GUARD_KERNELS_CK_GEMM_SOFTMAX_GEMM_HPP
#include <migraphx/kernels/index.hpp>
#include <migraphx/kernels/algorithm.hpp>
#include <migraphx/kernels/integral_constant.hpp>
#include <migraphx/kernels/tensor_view.hpp>
#include <migraphx/kernels/ck.hpp>
#include <migraphx/kernels/gemm_batcher.hpp>
namespace migraphx {
template <class T>
struct ck_gemm_softmax_gemm_settings
{
T scale{};
};
template <class... Ts>
constexpr ck_gemm_softmax_gemm_settings<Ts...> make_ck_gemm_softmax_gemm_settings(Ts... xs)
{
return {xs...};
}
template <class G, class C, class A, class B, class B1, class Settings>
__device__ void ck_gemm_softmax_gemm_matrix(C c, A a, B b, B1 b1, Settings s)
{
constexpr auto desc = G::make_descriptor(to_ck_tensor<A>(),
to_ck_tensor<ck_transposeb<B>>(),
to_ck_tensor<ck_transposeb<B1>>(),
to_ck_tensor<C>());
static_assert(desc.IsValid(), "Invalid ck gemm.");
G::Run(desc,
s.scale,
to_ck_const_pointer(a.data()),
to_ck_const_pointer(b.data()),
to_ck_const_pointer(b1.data()),
to_ck_pointer(c.data()));
}
template <class G, index_int BlocksPerBatch, class... Ts, class Settings>
__device__ void ck_gemm_softmax_gemm(Settings s, Ts... xs)
{
gemm_batch_args(make_index(), _c<BlocksPerBatch>, xs...)(
[&](auto... ys) { ck_gemm_softmax_gemm_matrix<G>(ys..., s); });
}
} // namespace migraphx
#endif
src/targets/gpu/kernels/include/migraphx/kernels/index.hpp
View file @ ac04f3cc
......@@ -31,6 +31,14 @@
#include <migraphx/kernels/debug.hpp>
#include <migraphx/kernels/functional.hpp>
#ifdef __clang__
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wreserved-identifier"
extern "C" __device__ size_t __ockl_get_enqueued_local_size(uint); // NOLINT
extern "C" __device__ size_t __ockl_get_local_size(uint); // NOLINT
#pragma clang diagnostic pop
#endif
namespace migraphx {
#if defined(MIGRAPHX_NGLOBAL) && defined(MIGRAPHX_NLOCAL)
......@@ -45,43 +53,37 @@ inline __device__ __attribute__((const)) index_int compute_global_size()
// This actualy works even when global is not divisible by local size.
// This doesnt actually do a multiplicatiosn. Instead it calls a device
// function to get the global size, which is why it works.
return blockDim.x * gridDim.x; // NOLINT
return blockDim.x * gridDim.x; // NOLINT
#endif
}
// We cant just use blockDim.x to get the local size since its broken on hip
// when global is not divisible by local size. In this case, we calulate the
// size for the last group.
#ifdef MIGRAPHX_NGROUP
// If global is divisible by local then local can be a const
#if(MIGRAPHX_NGLOBAL % MIGRAPHX_NLOCAL == 0) || (MIGRAPHX_NGROUP == 1)
#define MIGRAPHX_HAS_CONST_LOCAL 1
#endif
#endif
inline __device__ __attribute__((const)) index_int compute_local_size()
{
#ifdef MIGRAPHX_NLOCAL
const auto nlocal = MIGRAPHX_NLOCAL;
#else
const auto nlocal = blockDim.x; // NOLINT
#endif
#ifdef MIGRAPHX_NGROUP
const auto ngroup = MIGRAPHX_NGROUP;
#ifdef MIGRAPHX_HAS_CONST_LOCAL
return MIGRAPHX_NLOCAL;
#else
const auto ngroup = gridDim.x; // NOLINT
// Returns block size. For the non-uniform block it returns the size of the non-uniform block.
return __ockl_get_local_size(0); // NOLINT
#endif
const auto group_id = blockIdx.x; // NOLINT
const auto nglobal = compute_global_size();
if(group_id == ngroup - 1)
{
return 1 + (nglobal - 1) % nlocal;
}
else
{
return nlocal; // NOLINT
}
}
#ifdef MIGRAPHX_NGROUP
// If global is divisible by local then local can be a const
#if(MIGRAPHX_NGLOBAL % MIGRAPHX_NLOCAL == 0) || (MIGRAPHX_NGROUP == 1)
#define MIGRAPHX_HAS_CONST_LOCAL 1
#endif
inline __device__ __attribute__((const)) index_int compute_max_local_size()
{
#ifdef MIGRAPHX_LOCAL
return MIGRAPHX_NLOCAL;
#else
// Returns the block size. When workgrop has non-uniform block, this returns size of the uniform
// block.
return __ockl_get_enqueued_local_size(0); // NOLINT
#endif
}
struct index
{
......@@ -126,8 +128,8 @@ struct index
#else
__device__ index_int max_nlocal() const
{
MIGRAPHX_ASSERT(blockDim.x > 0);
return blockDim.x;
MIGRAPHX_ASSERT(compute_max_local_size() > 0);
return compute_max_local_size();
}
#endif
......@@ -249,7 +251,8 @@ struct index
#endif
inline __device__ __attribute__((const)) index make_index()
{
return index{blockIdx.x * blockDim.x + threadIdx.x, threadIdx.x, blockIdx.x}; // NOLINT
return index{
blockIdx.x * compute_max_local_size() + threadIdx.x, threadIdx.x, blockIdx.x}; // NOLINT
}
} // namespace migraphx
......
src/targets/gpu/kernels/include/migraphx/kernels/math.hpp
View file @ ac04f3cc
......@@ -101,7 +101,9 @@ MIGRAPHX_DEVICE_MATH(erf, ::erf)
MIGRAPHX_DEVICE_MATH(exp, ::exp)
MIGRAPHX_DEVICE_MATH(floor, ::floor)
MIGRAPHX_DEVICE_MATH(isnan, ::isnan)
MIGRAPHX_DEVICE_MATH(isinf, ::isinf)
MIGRAPHX_DEVICE_MATH(log, ::log)
MIGRAPHX_DEVICE_MATH(nearbyint, ::nearbyint)
MIGRAPHX_DEVICE_MATH(pow, ::pow)
MIGRAPHX_DEVICE_MATH(remainder, ::remainder)
MIGRAPHX_DEVICE_MATH(round, ::round)
......@@ -135,6 +137,7 @@ MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, ceil, ::hceil)
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, cos, ::hcos)
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, exp, ::hexp)
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, floor, ::hfloor)
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, isinf, ::__hisinf)
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, isnan, ::__hisnan)
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, log, ::hlog)
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, rsqrt, ::hrsqrt)
......@@ -150,6 +153,7 @@ MIGRAPHX_DEVICE_MATH_HALF(atan, ::atan)
MIGRAPHX_DEVICE_MATH_HALF(atanh, ::atanh)
MIGRAPHX_DEVICE_MATH_HALF(cosh, ::cosh)
MIGRAPHX_DEVICE_MATH_HALF(erf, ::erf)
MIGRAPHX_DEVICE_MATH_HALF(nearbyint, ::nearbyint)
MIGRAPHX_DEVICE_MATH_HALF(pow, ::pow)
MIGRAPHX_DEVICE_MATH_HALF(remainder, ::remainder)
MIGRAPHX_DEVICE_MATH_HALF(round, ::round)
......@@ -229,10 +233,12 @@ MIGRAPHX_DEVICE_MATH_VEC(erf)
MIGRAPHX_DEVICE_MATH_VEC(exp)
MIGRAPHX_DEVICE_MATH_VEC(floor)
MIGRAPHX_DEVICE_MATH_VEC(fmod)
MIGRAPHX_DEVICE_MATH_VEC(isinf)
MIGRAPHX_DEVICE_MATH_VEC(isnan)
MIGRAPHX_DEVICE_MATH_VEC(log)
MIGRAPHX_DEVICE_MATH_VEC(max)
MIGRAPHX_DEVICE_MATH_VEC(min)
MIGRAPHX_DEVICE_MATH_VEC(nearbyint)
MIGRAPHX_DEVICE_MATH_VEC(pow)
MIGRAPHX_DEVICE_MATH_VEC(remainder)
MIGRAPHX_DEVICE_MATH_VEC(round)
......
src/targets/gpu/lowering.cpp
View file @ ac04f3cc
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -40,6 +40,7 @@
#include <migraphx/op/if_op.hpp>
#include <migraphx/op/reshape.hpp>
#include <migraphx/op/quant_dot.hpp>
#include <migraphx/op/reshape_lazy.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/lowering.hpp>
......@@ -60,9 +61,8 @@ struct miopen_apply
const lowering* pass = nullptr;
std::unordered_map<std::string, std::function<instruction_ref(instruction_ref)>> apply_map{};
instruction_ref last{};
bool offload_copy = false;
bool int8_x4_format = true;
bool compute_fp32 = false;
bool offload_copy = false;
bool compute_fp32 = false;
context& get_context() const
{
......@@ -83,13 +83,10 @@ struct miopen_apply
assert(mod != nullptr);
assert(pass != nullptr);
auto& ctx = get_context();
int8_x4_format = get_int8_x4_format(ctx);
compute_fp32 = get_compute_fp32_flag();
offload_copy = (mod == mpm->get_root_module()) ? pass->offload_copy : false;
compute_fp32 = get_compute_fp32_flag();
offload_copy = (mod == mpm->get_root_module()) ? pass->offload_copy : false;
add_generic_op("contiguous");
add_extend_op("argmax");
add_extend_op("argmin");
add_extend_op("logsoftmax");
......@@ -115,6 +112,7 @@ struct miopen_apply
add_neg_op();
add_nms_op();
add_select_module_op();
add_reshape_lazy_op();
}
void copy_params() const
......@@ -230,18 +228,15 @@ struct miopen_apply
assert(refs.size() == 2);
auto output = insert_allocation(ins, ins->get_shape());
refs.push_back(output);
return mod->replace_instruction(
ins, rocblas_gemm<Op>{Op{}, 1, 0, int8_x4_format, compute_fp32}, refs);
return mod->replace_instruction(ins, rocblas_gemm<Op>{Op{}, 1, 0, compute_fp32}, refs);
});
}
void add_convolution_op(const std::string& name)
{
apply_map.emplace(name, [=](instruction_ref ins) {
operation conv = make_op(
"gpu::" + name,
{{"op", ins->get_operator().to_value()}, {"int8_x4_format", int8_x4_format}});
auto output = insert_allocation(ins, ins->get_shape());
operation conv = make_op("gpu::" + name, {{"op", ins->get_operator().to_value()}});
auto output = insert_allocation(ins, ins->get_shape());
return mod->replace_instruction(ins,
make_op("gpu::miopen_op", {{"op", to_value(conv)}}),
......@@ -376,6 +371,32 @@ struct miopen_apply
return mod->replace_instruction(ins, ins->get_operator(), inputs, ins->module_inputs());
});
}
/**
* Adds reshape lazy to reshape ops that can be aliased instead of copied.
* `gpu::contiguous` are added before and after the reshape; these contiguous
* instructions can be removed by the eliminate_contiguous pass.
*/
void add_reshape_lazy_op()
{
apply_map.emplace("reshape", [=](instruction_ref ins) {
std::vector<instruction_ref> before_contiguous_args = ins->inputs();
auto before_alloc = insert_allocation(ins, std::prev(ins)->get_shape());
before_contiguous_args.push_back(before_alloc);
auto before_contig =
mod->insert_instruction(ins, make_op("gpu::contiguous"), {before_contiguous_args});
auto new_lazy_reshape = mod->insert_instruction(
ins,
make_op("reshape_lazy", {{"dims", {ins->get_operator().to_value().at("dims")}}}),
before_contig);
std::vector<instruction_ref> after_contiguous_args = {new_lazy_reshape};
auto after_alloc = insert_allocation(new_lazy_reshape, new_lazy_reshape->get_shape());
after_contiguous_args.push_back(after_alloc);
return mod->replace_instruction(ins, make_op("gpu::contiguous"), after_contiguous_args);
});
}
};
void lowering::apply(module_pass_manager& mpm) const
......
src/targets/gpu/mlir.cpp
View file @ ac04f3cc
......@@ -22,7 +22,9 @@
* THE SOFTWARE.
*/
#include "migraphx/make_op.hpp"
#include <migraphx/stringutils.hpp>
#include <migraphx/gpu/mlir.hpp>
#include <ostream>
#ifdef MIGRAPHX_MLIR
#include <mlir-c/IR.h>
......@@ -33,6 +35,7 @@
#include <mlir-c/Dialect/Rock.h>
#include <mlir-c/IntegerSet.h>
#include <mlir-c/Pass.h>
#include <mlir-c/Support.h>
#include <mutex>
#if !defined(MLIR_MIGRAPHX_DIALECT_API_VERSION) || MLIR_MIGRAPHX_DIALECT_API_VERSION != 3
#warning "Incompatible version of rocMLIR library used, disabling"
......@@ -69,6 +72,7 @@ inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_MLIR);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNE_EXHAUSTIVE);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNING_DB);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNING_CFG);
......@@ -93,6 +97,8 @@ struct mlir_handle
friend bool operator==(ptr x, ptr y) { return x.get_value() == y.get_value(); }
friend bool operator!=(ptr x, ptr y) { return not(x == y); }
explicit operator bool() const noexcept { return obj != ptr(); }
T obj{};
};
......@@ -176,13 +182,85 @@ std::string mlir_print(F f, T x)
return ss.str();
}
struct mlir_logger
{
std::stringstream ss;
mlir_context* ctx;
std::optional<MlirDiagnosticHandlerID> id;
mlir_logger() : ctx(nullptr), id(std::nullopt) {}
mlir_logger(mlir_context* context) : ctx(context)
{
id =
mlirContextAttachDiagnosticHandler(ctx->get(), mlir_diagnostic_print_cb, this, nullptr);
}
~mlir_logger()
{
if(id.has_value())
mlirContextDetachDiagnosticHandler(ctx->get(), *id);
}
mlir_logger(const mlir_logger& other) = delete;
mlir_logger& operator=(const mlir_logger& other) = delete;
mlir_logger(mlir_logger&& other) noexcept
: ss(std::move(other.ss)), ctx(other.ctx), id(other.id)
{
other.ctx = nullptr;
other.id = std::nullopt;
}
mlir_logger& operator=(mlir_logger other) noexcept
{
std::swap(ss, other.ss);
std::swap(ctx, other.ctx);
std::swap(id, other.id);
return *this;
}
std::string str() const { return ss.str(); }
void clear() { ss = std::stringstream{}; }
static MlirLogicalResult mlir_diagnostic_print_cb(MlirDiagnostic diag, void* logger);
MlirLogicalResult handle(MlirDiagnostic diag);
};
MlirLogicalResult mlir_logger::mlir_diagnostic_print_cb(MlirDiagnostic diag, void* logger)
{
return reinterpret_cast<mlir_logger*>(logger)->handle(diag);
}
MlirLogicalResult mlir_logger::handle(MlirDiagnostic diag)
{
MlirDiagnosticSeverity sev = mlirDiagnosticGetSeverity(diag);
switch(sev)
{
case MlirDiagnosticSeverity::MlirDiagnosticError: ss << "Error: "; break;
case MlirDiagnosticSeverity::MlirDiagnosticWarning: ss << "Warning: "; break;
case MlirDiagnosticSeverity::MlirDiagnosticNote: ss << "Note: "; break;
case MlirDiagnosticSeverity::MlirDiagnosticRemark: ss << "Remark: "; break;
}
mlir_print(mlirDiagnosticPrint, diag, [&](auto s) { ss << s; });
ss << std::endl;
for(intptr_t i = 0, e = mlirDiagnosticGetNumNotes(diag); i < e; ++i)
{
(void)handle(mlirDiagnosticGetNote(diag, i));
}
return mlirLogicalResultSuccess();
}
struct mlir_program
{
mlir_program()
: ctx(mlirContextCreateWithRegistry(get_dialect_registry().get(),
/*threadingEnable=*/false)),
location(mlirLocationUnknownGet(ctx.get())),
mmodule(mlirModuleCreateEmpty(location))
mmodule(mlirModuleCreateEmpty(location)),
logger(&ctx)
{
mlirContextSetThreadPool(ctx.get(), get_thread_pool().get());
mlirContextLoadAllAvailableDialects(ctx.get());
......@@ -242,7 +320,10 @@ struct mlir_program
MlirType make_tensor(const shape& s) const
{
assert(s.standard());
if(not s.standard())
MIGRAPHX_THROW("MLIR expects all tensors to be in standard shape");
if(s.dynamic())
MIGRAPHX_THROW("MLIR does not support dynamic shapes");
std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
return mlirRankedTensorTypeGet(
lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());
......@@ -610,21 +691,49 @@ struct mlir_program
}
}
void run_high_level_pipeline() MIGRAPHX_TIDY_CONST
void run_high_level_pipeline()
{
mlir_pass_manager pm_front{mlirPassManagerCreate(ctx.get())};
mlirMIGraphXAddHighLevelPipeline(pm_front.get());
mlirPassManagerRunOnOp(pm_front.get(), mlirModuleGetOperation(mmodule.get()));
logger.clear();
if(mlirLogicalResultIsFailure(
mlirPassManagerRunOnOp(pm_front.get(), mlirModuleGetOperation(mmodule.get()))))
{
std::string error = "Invalid MLIR created: " + logger.str();
if(enabled(MIGRAPHX_TRACE_MLIR{}))
{
std::cout << error << std::endl;
}
MIGRAPHX_THROW(error);
}
}
void run_backend_pipeline() MIGRAPHX_TIDY_CONST
void run_backend_pipeline()
{
mlir_pass_manager pm_back{mlirPassManagerCreate(ctx.get())};
mlirMIGraphXAddBackendPipeline(pm_back.get(), target_arch.c_str());
mlirPassManagerRunOnOp(pm_back.get(), mlirModuleGetOperation(mmodule.get()));
logger.clear();
const size_t trace = value_of(MIGRAPHX_TRACE_MLIR{});
static std::mutex mutex;
auto mod_op = mlirModuleGetOperation(mmodule.get());
if(trace >= 2)
{
const std::lock_guard<std::mutex> lock(mutex);
std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
}
if(mlirLogicalResultIsFailure(mlirPassManagerRunOnOp(pm_back.get(), mod_op)))
{
std::string error = "MLIR backend compilation failed: " + logger.str();
if(enabled(MIGRAPHX_TRACE_MLIR{}))
{
std::cout << error << std::endl;
}
MIGRAPHX_THROW(error);
}
}
code_object_op compile(const value& solution) MIGRAPHX_TIDY_CONST
code_object_op compile(const value& solution)
{
// 1st pipeline to call
run_high_level_pipeline();
......@@ -645,8 +754,8 @@ struct mlir_program
void set_gpu_properties(const context& migraphx_ctx)
{
const auto& device = migraphx_ctx.get_current_device();
target_arch = device.get_device_name();
num_cu = device.get_cu_count();
target_arch = device.get_device_name();
num_cu = device.get_cu_count();
}
std::pair<std::size_t, std::size_t> get_launch_params() const
......@@ -678,12 +787,15 @@ struct mlir_program
MIGRAPHX_THROW("Failed setting tuning key: " + *str);
}
tuning_config get_tuning_config() MIGRAPHX_TIDY_CONST
tuning_config get_tuning_config(bool exhaustive)
{
tuning_config tc;
run_high_level_pipeline();
mlir_tuning_space params{
mlirRockTuningSpaceCreate(mmodule.get(), RocmlirTuningParamSetKindFull)};
auto tuning_mode =
exhaustive ? RocmlirTuningParamSetKindFull : RocmlirTuningParamSetKindQuick;
if(enabled(MIGRAPHX_MLIR_TUNE_EXHAUSTIVE{}))
tuning_mode = RocmlirTuningParamSetKindExhaustive;
mlir_tuning_space params{mlirRockTuningSpaceCreate(mmodule.get(), tuning_mode)};
for(auto i : range(mlirRockTuningGetNumParams(params.get())))
{
mlir_tuning_param param{mlirRockTuningParamCreate()};
......@@ -695,7 +807,8 @@ struct mlir_program
if(perf_key_bytes > perf_key.size())
MIGRAPHX_THROW("Tuning perf key was " + std::to_string(perf_key_bytes) +
" bytes and thus too long");
tc.solutions.emplace_back(perf_key.begin(), perf_key.begin() + perf_key_bytes);
tc.solutions.emplace_back(
std::string(perf_key.begin(), perf_key.begin() + perf_key_bytes));
}
std::array<char, ROCMLIR_TUNING_KEY_BUFSZ> tuning_key;
size_t tuning_key_bytes =
......@@ -717,7 +830,8 @@ struct mlir_program
if(not tuning_cfg_path.empty())
{
std::vector<std::string> tokens = split_string(prob_config, '\t');
std::string prob = tokens[1];
std::string prob = tokens[2];
if(starts_with(prob, "conv"))
{
tuning_cfg_path += ".conv";
......@@ -727,6 +841,8 @@ struct mlir_program
tuning_cfg_path += ".gemm";
}
std::ofstream tuning_cfg(tuning_cfg_path, std::ios::app);
prob =
trim(prob, [](unsigned char c) { return (c == '\0') or (std::isspace(c) != 0); });
tuning_cfg << prob << std::endl;
}
}
......@@ -799,6 +915,7 @@ struct mlir_program
mlir_context ctx;
MlirLocation location;
mlir_module mmodule;
mlir_logger logger;
problem_params pp;
std::deque<std::string> strings{};
std::string target_arch = "";
......@@ -867,15 +984,22 @@ code_object_op compile_mlir(const context& migraphx_ctx,
adjust_param_shapes(m, to_shapes(inputs));
const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
static std::mutex mutex;
if(trace)
{
const std::lock_guard<std::mutex> lock(mutex);
std::cout << m << std::endl;
}
mlir_program mp;
mp.set_gpu_properties(migraphx_ctx);
mp.parse(m);
auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
if(trace)
{
const std::lock_guard<std::mutex> lock(mutex);
std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
}
auto co = mp.compile(solution);
co.expected_inputs = to_shapes(inputs);
co.output = m.get_output_shapes().front();
......@@ -898,15 +1022,17 @@ instruction_ref insert_mlir(module& m,
return m.insert_instruction(ins, co, refs);
}
tuning_config
get_tuning_config_mlir(const context& migraphx_ctx, module m, const std::vector<shape>& inputs)
tuning_config get_tuning_config_mlir(const context& migraphx_ctx,
module m,
const std::vector<shape>& inputs,
bool exhaustive)
{
adjust_param_shapes(m, inputs);
mlir_program mp;
mp.set_gpu_properties(migraphx_ctx);
mp.parse(m);
return mp.get_tuning_config();
return mp.get_tuning_config(exhaustive);
}
#else
......@@ -935,7 +1061,7 @@ insert_mlir(module& m, instruction_ref, code_object_op co, const std::vector<ins
return m.end();
}
tuning_config get_tuning_config_mlir(const context&, module, const std::vector<shape>&)
tuning_config get_tuning_config_mlir(const context&, module, const std::vector<shape>&, bool)
{
return {};
}
......
src/targets/gpu/no_device.cpp 0 → 100644
View file @ ac04f3cc
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifdef __HIP_DEVICE_COMPILE__
#error \
"Device compilation not allowed for migraphx_gpu. Do not link with hip::device. Device code should go into migraphx_device or migraphx_kernels"
#endif
src/targets/gpu/pack_int8_args.cpp deleted 100644 → 0
View file @ d39c3343
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <iterator>
#include <migraphx/gpu/pack_int8_args.hpp>
#include <migraphx/gpu/int8_gemm_pack.hpp>
#include <migraphx/gpu/int8_conv_pack.hpp>
#include <migraphx/gpu/hip.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/program.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/permutation.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
static instruction_ref pad_ins(module& m, instruction_ref ins, int offset)
{
auto s = ins->get_shape();
auto lens = s.lens();
auto k = lens[lens.size() + offset];
auto pad_k = (k + 3) / 4 * 4;
auto pad_lens = lens;
pad_lens[lens.size() + offset] = pad_k;
auto ret_ins = ins;
if(pad_k != k)
{
std::vector<int64_t> pad_dims(lens.size() * 2, 0);
pad_dims[lens.size() + offset] = pad_k - k;
shape ps{s.type(), pad_lens};
auto ins_out =
m.insert_instruction(ins, make_op("hip::allocate", {{"shape", to_value(ps)}}));
auto pad = make_op("pad", {{"pads", pad_dims}});
ret_ins =
m.insert_instruction(std::next(ins), make_op("gpu::pad", pad.to_value()), ins, ins_out);
}
return ret_ins;
}
static std::vector<instruction_ref> pad_inputs(module& m, instruction_ref ins)
{
std::vector<instruction_ref> ret_inputs;
auto inputs = ins->inputs();
auto in0 = inputs.at(0);
auto sa = in0->get_shape();
bool transa = sa.transposed();
if(transa)
{
auto perm = find_permutation(sa);
auto val = in0->get_operator().to_value();
if(val.contains("dims"))
{
int offset = static_cast<int>(perm.back()) - static_cast<int>(perm.size());
auto t_in = in0->inputs().front();
auto p_in = pad_ins(m, t_in, offset);
auto dims = val.at("dims").to_vector<int64_t>();
auto r_in =
m.insert_instruction(ins, make_op("transpose", {{"permutation", dims}}), p_in);
ret_inputs.push_back(r_in);
}
else
{
shape cs{in0->get_shape().type(), in0->get_shape().lens()};
auto con_out =
m.insert_instruction(ins, make_op("hip::allocate", {{"shape", to_value(cs)}}));
auto cin0 = m.insert_instruction(ins, make_op("gpu::contiguous"), in0, con_out);
ret_inputs.push_back(pad_ins(m, cin0, -1));
}
}
else
{
ret_inputs.push_back(pad_ins(m, in0, -1));
}
auto in1 = inputs.at(1);
auto sb = in1->get_shape();
bool transb = sb.transposed();
if(transb)
{
auto perm = find_permutation(sb);
auto val = in1->get_operator().to_value();
if(val.contains("dims"))
{
int offset = static_cast<int>(perm[perm.size() - 2]) - static_cast<int>(perm.size());
auto t_in = in1->inputs().front();
auto p_in = pad_ins(m, t_in, offset);
auto dims = val.at("dims").to_vector<int64_t>();
auto r_in =
m.insert_instruction(ins, make_op("transpose", {{"permutation", dims}}), p_in);
ret_inputs.push_back(r_in);
}
else
{
shape cs{in1->get_shape().type(), in1->get_shape().lens()};
auto con_out =
m.insert_instruction(ins, make_op("hip::allocate", {{"shape", to_value(cs)}}));
auto cin1 = m.insert_instruction(ins, make_op("gpu::contiguous"), in1, con_out);
ret_inputs.push_back(pad_ins(m, cin1, -2));
}
}
else
{
ret_inputs.push_back(pad_ins(m, in1, -2));
}
std::copy(inputs.begin() + 2, inputs.end(), std::back_inserter(ret_inputs));
return ret_inputs;
}
void pack_int8_args::apply(module& m) const
{
for(auto ins : iterator_for(m))
{
if(ins->name() == "gpu::quant_gemm")
{
auto val = ins->get_operator().to_value();
assert(val.contains("int8_x4_format"));
if(not val.at("int8_x4_format").to<bool>())
{
continue;
}
auto inputs = ins->inputs();
auto lens = inputs.at(0)->get_shape().lens();
// gemm need the k to be multiple of 4, so need packing that dimension
auto old_inputs = inputs;
if((lens.back() % 4) != 0)
{
inputs = pad_inputs(m, ins);
}
bool transa = inputs[0]->get_shape().transposed();
bool transb = inputs[1]->get_shape().transposed();
if(not transb)
{
auto packed_b = m.insert_instruction(
ins, make_op("hip::allocate", {{"shape", to_value(inputs[1]->get_shape())}}));
auto output_b = m.insert_instruction(
ins, make_op("gpu::int8_gemm_pack_a"), {inputs[1], packed_b});
inputs[1] = output_b;
}
if(transa)
{
auto packed_a = m.insert_instruction(
ins, make_op("hip::allocate", {{"shape", to_value(inputs[0]->get_shape())}}));
auto output_a = m.insert_instruction(
ins, make_op("gpu::int8_gemm_pack_b"), {inputs[0], packed_a});
inputs[0] = output_a;
}
if(inputs != old_inputs)
{
m.replace_instruction(ins, ins->get_operator(), inputs);
}
}
else if(ins->name() == "gpu::quant_convolution")
{
auto val = ins->get_operator().to_value();
if(not val.at("int8_x4_format").to<bool>())
{
continue;
}
auto inputs = ins->inputs();
auto packed_x = m.insert_instruction(
ins,
make_op("hip::allocate",
{{"shape", to_value(pack_int8_shape(inputs[0]->get_shape()))}}));
auto output_x =
m.insert_instruction(ins, make_op("gpu::int8_conv_pack"), {inputs[0], packed_x});
instruction::replace_argument(ins, inputs[0], output_x);
auto packed_w = m.insert_instruction(
ins,
make_op("hip::allocate",
{{"shape", to_value(pack_int8_shape(inputs[1]->get_shape()))}}));
auto output_w =
m.insert_instruction(ins, make_op("gpu::int8_conv_pack"), {inputs[1], packed_w});
instruction::replace_argument(ins, inputs[1], output_w);
}
}
}
shape pack_int8_args::pack_int8_shape(const shape& s) const
{
if(s.type() != shape::int8_type)
{
MIGRAPHX_THROW("PACK_INT8_ARGS: only process int8_type");
}
auto lens = s.lens();
auto strides = s.strides();
lens[1] = (lens[1] + 3) / 4 * 4;
strides[0] = strides[1] * lens[1];
return {s.type(), lens, strides};
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/prefuse_ops.cpp
View file @ ac04f3cc
......@@ -21,17 +21,19 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <migraphx/permutation.hpp>
#include <migraphx/gpu/prefuse_ops.hpp>
#include <migraphx/gpu/gemm_softmax_gemm.hpp>
#include <migraphx/match/layernorm.hpp>
#include <migraphx/check_shapes.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/register_op.hpp>
#include <migraphx/pass_manager.hpp>
#include <migraphx/dead_code_elimination.hpp>
#include <migraphx/gpu/ck.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace {
template <class Derived, std::size_t N>
......@@ -45,40 +47,42 @@ struct layernorm_base
}
shape compute_shape(std::vector<shape> inputs, std::vector<module_ref> mods) const
{
std::size_t nargs = 1;
std::size_t nargs = N;
if(not mods.empty())
{
auto* pm = mods.front();
nargs = pm->get_parameter_names().size();
nargs += pm->get_parameter_names().size() - 1;
}
check_shapes{inputs, static_cast<const Derived&>(*this)}.has(nargs + N);
auto s = inputs.at(0);
check_shapes{inputs, static_cast<const Derived&>(*this)}.has(nargs);
auto s = inputs.front();
auto t = s.type();
if(not mods.empty())
t = mods.front()->get_output_shapes().front().type();
if(s.scalar())
{
return s;
}
else if(s.broadcasted())
{
return {t, s.lens()};
}
else
{
return s.with_lens(t, s.lens());
}
// Scalar output if all inputs are scalar
if(inputs.front().elements() == 1 and
all_of(inputs, [](const auto& ss) { return ss.scalar(); }))
return inputs.front();
auto l_s = shape::from_permutation(
t, s.lens(), find_permutation(std::vector<shape>(inputs.begin(), inputs.begin() + N)));
// just prelayernorm or preadd_layernorm
if(nargs <= N)
return l_s;
// else, layernorm + pointwise fusion, preserve layout of fused op
std::vector<shape> lp_s(inputs.begin() + N, inputs.end());
lp_s.insert(lp_s.begin(), l_s);
return shape::from_permutation(t, s.lens(), find_permutation(lp_s));
}
};
struct layernorm : layernorm_base<layernorm, 0>
struct layernorm : layernorm_base<layernorm, 1>
{
std::string name() const { return "gpu::prelayernorm"; }
};
MIGRAPHX_REGISTER_OP(layernorm);
struct add_layernorm : layernorm_base<add_layernorm, 1>
struct add_layernorm : layernorm_base<add_layernorm, 2>
{
std::string name() const { return "gpu::preadd_layernorm"; }
};
......@@ -117,6 +121,60 @@ struct find_add_layernorm
m.replace_instruction(ins, add_layernorm{op.epsilon}, add_ins->inputs());
}
};
struct pre_gemm_softmax_gemm : gemm_softmax_gemm
{
std::string name() const { return "gpu::pre_gemm_softmax_gemm"; }
};
MIGRAPHX_REGISTER_OP(pre_gemm_softmax_gemm);
MIGRAPHX_PRED_MATCHER(is_ck_gemm, instruction_ref ins)
{
if(ins->name() != "dot")
return false;
if(not pre_gemm_softmax_gemm::is_ck_supported_type(ins->get_shape().type()))
return false;
return true;
}
struct find_gemm_softmax_gemm
{
auto matcher() const
{
auto gemm1 =
match::skip(match::name("contiguous"))(match::name("dot")(is_ck_gemm().bind("gemm1")));
auto mul = match::name("mul")(
match::nargs(2), match::either_arg(0, 1)(match::is_constant().bind("scale"), gemm1));
auto softmax = match::name("softmax")(match::arg(0)(mul)).bind("softmax");
return match::name("dot")(is_ck_gemm().bind("gemm2"))(match::arg(0)(softmax));
}
void apply(module_pass_manager& mpm, const match::matcher_result& r) const
{
auto ins = r.result;
auto gemm2_ins = r.instructions["gemm2"];
auto gemm1_ins = r.instructions["gemm1"];
auto scale_lit = r.instructions["scale"];
float scale = 1.0;
scale_lit->eval().visit([&](const auto s) {
// CK only supports single-valued scale
if(std::all_of(
s.begin() + 1, s.end(), [&](auto v) { return float_equal(v, s.front()); }))
scale = s.front();
else
return;
});
auto inputs = gemm1_ins->inputs(); // A, B
inputs.push_back(gemm2_ins->inputs().back()); // B1
mpm.get_module().replace_instruction(
ins, pre_gemm_softmax_gemm{gemm2_ins->get_operator(), scale}, inputs);
}
};
} // namespace
void prefuse_ops::apply(module_pass_manager& mpm) const
......@@ -124,6 +182,8 @@ void prefuse_ops::apply(module_pass_manager& mpm) const
match::find_matches(mpm.get_module(), find_layernorm{});
mpm.run_pass(dead_code_elimination{});
match::find_matches(mpm.get_module(), find_add_layernorm{});
if(enabled(MIGRAPHX_ENABLE_CK{}))
match::find_matches(mpm, find_gemm_softmax_gemm{});
}
} // namespace gpu
......
src/targets/gpu/rocblas.cpp
View file @ ac04f3cc
......@@ -53,19 +53,6 @@ bool get_compute_fp32_flag()
return (starts_with(device_name, "gfx9") and device_name >= "gfx908");
}
bool get_int8_x4_format(context& ctx)
{
#if ROCBLAS_VERSION_MAJOR >= 3
(void)(ctx);
return false;
#else
// int8x4 packed format is only available starting from rocblas-v2.38 and it is deprecated in
// v3.0 and will be removed in v4.0
rocblas_gemm_flags flag;
rocblas_query_int8_layout_flag(ctx.get_stream().get_rocblas(), &flag);
return flag == rocblas_gemm_flags_pack_int8x4;
#endif
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/target.cpp
View file @ ac04f3cc
......@@ -49,6 +49,7 @@
#include <migraphx/rewrite_quantization.hpp>
#include <migraphx/rewrite_rnn.hpp>
#include <migraphx/schedule.hpp>
#include <migraphx/simplify_dyn_ops.hpp>
#include <migraphx/simplify_qdq.hpp>
#include <migraphx/simplify_reshapes.hpp>
#include <migraphx/split_single_dyn_dim.hpp>
......@@ -63,7 +64,6 @@
#include <migraphx/gpu/fuse_ops.hpp>
#include <migraphx/gpu/prefuse_ops.hpp>
#include <migraphx/gpu/lowering.hpp>
#include <migraphx/gpu/pack_int8_args.hpp>
#include <migraphx/gpu/schedule_model.hpp>
#include <migraphx/gpu/sync_device.hpp>
#include <migraphx/gpu/target.hpp>
......@@ -110,6 +110,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
{
split_single_dyn_dim{},
dead_code_elimination{},
simplify_dyn_ops{},
dead_code_elimination{},
normalize_ops{},
dead_code_elimination{},
simplify_qdq{},
......@@ -153,7 +155,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
dead_code_elimination{},
compile_miopen{&gctx},
dead_code_elimination{},
pack_int8_args{},
dead_code_elimination{},
fuse_ops{&ctx, options.fast_math},
dead_code_elimination{},
......
src/targets/gpu/time_op.cpp
View file @ ac04f3cc
......@@ -41,8 +41,7 @@ std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsig
}
using milliseconds = std::chrono::duration<double, std::milli>;
std::pair<double, double>
time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
double time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
{
// TODO: Use std::ref
......@@ -51,21 +50,19 @@ time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
auto output = op.compute_shape(inputs);
op.finalize(ctx, output, inputs);
auto args = generate_arguments(inputs);
auto run = [&] {
op.compute(ctx, output, args);
ctx.finish();
};
gctx.enable_perf_measurement();
auto start = context::create_event_for_timing();
auto stop = context::create_event_for_timing();
auto run = [&] { op.compute(ctx, output, args); };
run();
double host_time = 0.0;
double device_time = 0.0;
gctx.get_stream().record(start.get());
for(auto i : range(n))
{
(void)i;
host_time += time<milliseconds>(run);
device_time += gctx.get_elapsed_ms();
run();
}
return std::make_pair(host_time / n, device_time / n);
gctx.get_stream().record(stop.get());
gctx.finish();
return context::get_elapsed_ms(start.get(), stop.get()) / n;
}
} // namespace gpu
......
src/targets/ref/lowering.cpp
View file @ ac04f3cc
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......
src/tf/parse_reshape.cpp
View file @ ac04f3cc
......@@ -45,8 +45,7 @@ struct parse_reshape : op_parser<parse_reshape>
auto s = args[1]->eval();
std::vector<int64_t> dims;
s.visit([&](auto v) { copy(v, std::back_inserter(dims)); });
return info.add_instruction(make_op("reshape", {{"dims", dims}}),
info.make_contiguous(args[0]));
return info.add_instruction(make_op("reshape", {{"dims", dims}}), args[0]);
}
};
......
src/verify_args.cpp
View file @ ac04f3cc
......@@ -28,19 +28,20 @@ namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
bool verify_args(const std::string& name,
const argument& ref_arg,
const argument& target_arg,
double tolerance)
const verify::expected<argument>& ref_arg,
verify::tolerance tols)
{
bool passed = true;
visit_all(ref_arg, target_arg)([&](auto ref, auto target) {
double error;
passed = verify::verify_range(ref, target, tolerance, &error);
visit_all(ref_arg.data(), target_arg)([&](auto ref, auto target) {
double rms_error;
passed =
verify::verify_range_with_tolerance(target, verify::expected{ref}, tols, &rms_error);
if(not passed)
{
// TODO: Check for nans
std::cout << "FAILED: " << name << std::endl;
std::cout << "error: " << error << std::endl;
std::cout << "RMS Error: " << rms_error << std::endl;
if(ref.size() < 32)
std::cout << "ref:" << ref << std::endl;
if(target.size() < 32)
......@@ -78,16 +79,6 @@ bool verify_args(const std::string& name,
if(verify::range_zero(target))
std::cout << "Target data is all zeros" << std::endl;
// auto mxdiff = max_diff(ref, target);
// std::cout << "Max diff: " << mxdiff << std::endl;
// auto idx = mismatch_idx(ref, target, float_equal);
// if(idx < verify::range_distance(ref))
// {
// std::cout << "Mismatch at " << idx << ": " << ref[idx] << " != " << target[idx]
// << std::endl;
// }
auto ref_nan_idx = find_idx(ref, verify::not_finite);
if(ref_nan_idx >= 0)
std::cout << "Non finite number found in ref at " << ref_nan_idx << ": "
......@@ -97,11 +88,22 @@ bool verify_args(const std::string& name,
if(target_nan_idx >= 0)
std::cout << "Non finite number found in target at " << target_nan_idx << ": "
<< target[target_nan_idx] << std::endl;
// std::cout << std::endl;
std::cout << "MIGraphX verification passed successfully." << std::endl;
}
});
return passed;
}
bool verify_args_with_tolerance(const std::string& name,
const argument& target_arg,
const verify::expected<argument>& ref_arg,
std::size_t tolerance)
{
double rms_tol = 0.001;
target_arg.visit([&](auto ta) { rms_tol = verify::get_rms_tol(ta, tolerance); });
verify::tolerance tols{rms_tol};
return verify_args(name, target_arg, ref_arg, tols);
}
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
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