Skip to content

GitLab

Commit 64998903 authored by Chris Austen's avatar Chris Austen
Browse files

Merge remote-tracking branch 'origin/develop' into rel57_workitems

parents 2190957c e4dc75ea
Show whitespace changes
Inline Side-by-side
Showing with 643 additions and 192 deletions
src/targets/gpu/include/migraphx/gpu/mlir.hpp
View file @ 64998903
......@@ -29,6 +29,7 @@
#include <migraphx/gpu/config.hpp>
#include <migraphx/gpu/code_object_op.hpp>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/gpu/tuning_config.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -36,16 +37,19 @@ struct module;
namespace gpu {
MIGRAPHX_GPU_EXPORT std::string dump_mlir(const module& m);
MIGRAPHX_GPU_EXPORT code_object_op compile_mlir(const context& ctx,
module m,
const std::vector<instruction_ref>& inputs);
const std::vector<instruction_ref>& inputs,
const value& solution);
MIGRAPHX_GPU_EXPORT instruction_ref insert_mlir(module& m,
instruction_ref ins,
code_object_op co,
const std::vector<instruction_ref>& inputs);
MIGRAPHX_GPU_EXPORT tuning_config get_tuning_config_mlir(module m,
const std::vector<shape>& inputs);
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
......
src/targets/gpu/include/migraphx/gpu/tuning_config.hpp 0 → 100644
View file @ 64998903
/*
* 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_GPU_TUNING_CONFIG_HPP
#define MIGRAPHX_GUARD_GPU_TUNING_CONFIG_HPP
#include <migraphx/config.hpp>
#include <migraphx/value.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct tuning_config
{
value problem;
std::vector<value> solutions;
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif // MIGRAPHX_GUARD_GPU_TUNING_CONFIG_HPP
src/targets/gpu/jit/mlir.cpp
View file @ 64998903
......@@ -36,11 +36,12 @@ struct mlir_compiler : compiler<mlir_compiler>
operation compile_op(context&, const std::vector<shape>&, const value&) const { return {}; }
compiler_replace compile(context& ctx, instruction_ref ins, const operation&) const
compiler_replace
compile(context& ctx, instruction_ref ins, const operation&, const value& solution) const
{
auto* smod = ins->module_inputs().front();
assert(smod->get_parameter_names().size() == ins->inputs().size() - 1);
return insert(compile_mlir(ctx, *smod, ins->inputs()));
return insert(compile_mlir(ctx, *smod, ins->inputs(), solution));
}
compiler_replace insert(code_object_op co) const
......@@ -50,6 +51,16 @@ struct mlir_compiler : compiler<mlir_compiler>
m.replace_instruction(ins, mlir);
}};
}
optional<tuning_config>
get_tuning_config(context&, instruction_ref ins, 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(*smod, shapes);
}
};
} // namespace gpu
......
src/targets/gpu/jit/pointwise.cpp
View file @ 64998903
......@@ -72,7 +72,7 @@ struct pointwise_compiler : compiler<pointwise_compiler>
hip_compile_options options;
options.inputs = inputs;
options.output = inputs.back();
options.virtual_inputs = reduce_dims(inputs);
options.virtual_inputs = reduce_dims(normalize_permutation(inputs));
options.params = "-Wno-float-equal";
auto axis = find_fast_axis(options.virtual_inputs);
auto vec = vectorize::elements(ctx, axis, options.virtual_inputs);
......
src/targets/gpu/jit/reduce.cpp
View file @ 64998903
......@@ -84,7 +84,7 @@ static shape get_reduced_shape(const shape& s, const std::vector<T>& axes)
std::fill(lens.begin(), lens.end(), 1);
for(const auto& axis : axes)
lens[axis] = s.lens()[axis];
return shape{s.type(), lens};
return s.with_lens(lens);
}
template <class T>
......@@ -93,7 +93,7 @@ static shape get_output_shape(const shape& s, const std::vector<T>& axes)
auto lens = s.lens();
for(const auto& axis : axes)
lens[axis] = 1;
return shape{s.type(), lens};
return s.with_lens(lens);
}
template <class ReduceLens>
......@@ -228,7 +228,7 @@ struct fused_reduce_compiler : compiler<fused_reduce_compiler>
auto virtual_inputs = inputs;
virtual_inputs.push_back(get_reduced_shape(inputs.front(), axes));
virtual_inputs.push_back(get_output_shape(inputs.front(), axes));
virtual_inputs = reduce_dims(virtual_inputs);
virtual_inputs = reduce_dims(normalize_permutation(virtual_inputs));
auto reduce_output_shape = virtual_inputs.back();
virtual_inputs.pop_back();
auto reduction_shape = virtual_inputs.back();
......
src/targets/gpu/mlir.cpp
View file @ 64998903
......@@ -52,6 +52,7 @@
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/gpu/perfdb.hpp>
#include <migraphx/gpu/tuning_config.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/permutation.hpp>
#include <deque>
......@@ -134,6 +135,10 @@ using mlir_block = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirBlock, mlirBlockD
using mlir_pass_manager = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirPassManager, mlirPassManagerDestroy);
using mlir_tuning_table = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirRockTuningTable,
mlirRockTuningTableDestroy);
using mlir_tuning_space = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirRockTuningSpace,
mlirRockTuningSpaceDestroy);
using mlir_tuning_param = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirRockTuningParam,
mlirRockTuningParamDestroy);
std::string_view to_string_view(MlirStringRef s) { return {s.data, s.length}; }
......@@ -616,18 +621,30 @@ struct mlir_program
}
}
code_object_op compile() MIGRAPHX_TIDY_CONST
void run_high_level_pipeline() MIGRAPHX_TIDY_CONST
{
mlir_pass_manager pm_front{mlirPassManagerCreate(ctx.get())};
mlir_pass_manager pm_back{mlirPassManagerCreate(ctx.get())};
// 1st pipeline to call
mlirMIGraphXAddHighLevelPipeline(pm_front.get());
mlirPassManagerRunOnOp(pm_front.get(), mlirModuleGetOperation(mmodule.get()));
}
// 2nd pipeline to call
get_module_tuned();
void run_backend_pipeline() MIGRAPHX_TIDY_CONST
{
mlir_pass_manager pm_back{mlirPassManagerCreate(ctx.get())};
mlirMIGraphXAddBackendPipeline(pm_back.get(), target_arch.c_str());
mlirPassManagerRunOnOp(pm_back.get(), mlirModuleGetOperation(mmodule.get()));
}
code_object_op compile(const value& solution) MIGRAPHX_TIDY_CONST
{
// 1st pipeline to call
run_high_level_pipeline();
if(solution.is_null())
get_module_tuned();
else
set_tuning(solution);
// 2nd pipeline to call
run_backend_pipeline();
code_object_op op{};
op.symbol_name = sym_name;
......@@ -658,6 +675,33 @@ struct mlir_program
MIGRAPHX_THROW("Failed to compile mlir program");
}
void set_tuning(const value& v)
{
auto str = v.to<std::string>();
// We need to make a copy of the buffer since mlirRockTuningSetFromStr may modify the string
std::vector<char> buffer(str.begin(), str.end());
buffer.push_back(0);
if(not mlirRockTuningSetFromStr(mmodule.get(), buffer.data()))
MIGRAPHX_THROW("Failed setting tuning key: " + str);
}
tuning_config get_tuning_config() MIGRAPHX_TIDY_CONST
{
tuning_config tc;
run_high_level_pipeline();
mlir_tuning_space params{mlirRockTuningSpaceCreate(mmodule.get())};
for(auto i : range(mlirRockTuningGetNumParamsFull(params.get())))
{
mlir_tuning_param param{mlirRockTuningParamCreate()};
if(not mlirRockTuningParamGet(params.get(), i, param.get()))
MIGRAPHX_THROW("Incorrect mlir tuning parameter: " + std::to_string(i));
tc.solutions.push_back(std::string{mlirRockTuningGetParamStr(param.get())});
}
mlir_tuning_table tuning_table{mlirRockTuningTableCreate()};
tc.problem = std::string{mlirRockTuningGetKey(tuning_table.get(), mmodule.get())};
return tc;
}
std::string get_tune_params(bool xdlops) const { return get_mlir_perf_for_conv(pp, xdlops); }
// This function appends to tuning cfg file that could be
......@@ -749,14 +793,14 @@ std::string dump_mlir(const module& m)
return mlir_print(&mlirOperationPrint, mod_op);
}
void adjust_param_shapes(module& m, const std::vector<instruction_ref>& inputs)
void adjust_param_shapes(module& m, const std::vector<shape>& inputs)
{
auto names = m.get_parameter_names();
std::sort(names.begin(), names.end());
for(auto i : range(names.size()))
{
const auto& name = names[i];
const auto& input = inputs[i]->get_shape();
const auto& input = inputs[i];
auto param = m.get_parameter(name);
if(input.standard())
continue;
......@@ -794,9 +838,12 @@ void adjust_param_shapes(module& m, const std::vector<instruction_ref>& inputs)
}
}
code_object_op compile_mlir(const context&, module m, const std::vector<instruction_ref>& inputs)
code_object_op compile_mlir(const context&,
module m,
const std::vector<instruction_ref>& inputs,
const value& solution)
{
adjust_param_shapes(m, inputs);
adjust_param_shapes(m, to_shapes(inputs));
const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
if(trace)
......@@ -808,7 +855,8 @@ code_object_op compile_mlir(const context&, module m, const std::vector<instruct
auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
if(trace)
std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
auto co = mp.compile();
auto co = mp.compile(solution);
co.expected_inputs = to_shapes(inputs);
co.output = m.get_output_shapes().front();
return co;
}
......@@ -829,6 +877,16 @@ instruction_ref insert_mlir(module& m,
return m.insert_instruction(ins, co, refs);
}
tuning_config get_tuning_config_mlir(module m, const std::vector<shape>& inputs)
{
adjust_param_shapes(m, inputs);
mlir_program mp;
mp.find_target();
mp.parse(m);
return mp.get_tuning_config();
}
#else
std::string dump_mlir(const module&) { return {}; }
......@@ -840,11 +898,11 @@ void use(T&)
// Disabling clang-tidy warning on non-real useage.
// NOLINTBEGIN(performance-unnecessary-value-param)
code_object_op compile_mlir(const context&, module, const std::vector<instruction_ref>&)
code_object_op
compile_mlir(const context&, module, const std::vector<instruction_ref>&, const value&)
{
return {};
}
// NOLINTEND(performance-unnecessary-value-param)
instruction_ref
// cppcheck-suppress funcArgNamesDifferent
......@@ -854,6 +912,9 @@ insert_mlir(module& m, instruction_ref, code_object_op co, const std::vector<ins
return m.end();
}
tuning_config get_tuning_config_mlir(module, const std::vector<shape>&) { return {}; }
// NOLINTEND(performance-unnecessary-value-param)
#endif
} // namespace gpu
......
src/targets/gpu/target.cpp
View file @ 64998903
......@@ -75,7 +75,9 @@ namespace gpu {
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_SCHEDULE_PASS)
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_REDUCE_FUSION)
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_ENABLE_NHWC)
#ifndef _WIN32
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_ENABLE_CK)
#endif
struct id_pass
{
......@@ -136,7 +138,9 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
dead_code_elimination{},
enable_pass(not enabled(MIGRAPHX_DISABLE_REDUCE_FUSION{}), fuse_reduce{}),
dead_code_elimination{},
#ifndef _WIN32
enable_pass(enabled(MIGRAPHX_ENABLE_CK{}), fuse_ck{}),
#endif
dead_code_elimination{},
enable_pass(mlir_enabled(), fuse_mlir{&ctx}),
dead_code_elimination{},
......
src/tf/parse_batchnorm.cpp
View file @ 64998903
......@@ -52,7 +52,6 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
auto x_type = args[0]->get_shape().type();
// unsqueeze tensors of shape (C) to broadcast correctly
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto scale_unsqueeze =
......@@ -64,11 +63,11 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
auto var_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[4]);
auto numer = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto x_sub_mean = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
auto denom = info.add_broadcastable_binary_op("pow", var_eps, rt);
auto div0 = info.add_broadcastable_binary_op("div", numer, denom);
auto r0 = info.add_broadcastable_binary_op("mul", div0, scale_unsqueeze);
auto rsqrt = info.add_instruction(make_op("rsqrt"), var_eps);
auto mul0 = info.add_broadcastable_binary_op("mul", scale_unsqueeze, rsqrt);
auto r0 = info.add_broadcastable_binary_op("mul", x_sub_mean, mul0);
return info.add_broadcastable_binary_op("add", r0, bias_unsqueeze);
}
};
......
test/api/CMakeLists.txt
View file @ 64998903
......@@ -36,7 +36,7 @@ endfunction()
function(add_c_api_test TEST_NAME TEST_SRC TEST_DIR)
set(NAME test_api_${TEST_NAME})
add_executable(${NAME} EXCLUDE_FROM_ALL ${TEST_SRC})
target_link_libraries(${NAME} migraphx_c migraphx)
target_link_libraries(${NAME} migraphx_c)
target_include_directories(${NAME} PUBLIC ../include)
add_test(NAME ${NAME} COMMAND $<TARGET_FILE:${NAME}> WORKING_DIRECTORY ${TEST_DIR})
add_dependencies(tests ${NAME})
......
test/api/test_custom_op.cpp
View file @ 64998903
......@@ -99,7 +99,7 @@ TEST_CASE(run_sigmoid_custom_op)
EXPECT(bool{result == migraphx::argument(s, expected_result.data())});
}
extern "C" void migraphx_test_private_disable_exception_catch(bool b);
extern "C" MIGRAPHX_C_EXPORT void migraphx_test_private_disable_exception_catch(bool);
TEST_CASE(run_sigmoid_with_incorrect_shape)
{
......
test/api/test_gpu.cpp
View file @ 64998903
......@@ -34,7 +34,6 @@ TEST_CASE(load_and_run)
auto shapes_before = p.get_output_shapes();
migraphx::compile_options options;
options.set_offload_copy();
options.set_exhaustive_tune_flag();
p.compile(migraphx::target("gpu"), options);
auto shapes_after = p.get_output_shapes();
CHECK(shapes_before.size() == 1);
......
test/gpu/mlir.cpp
View file @ 64998903
......@@ -84,7 +84,7 @@ migraphx::program create_program_from_mlir(const migraphx::module& mmlir)
inputs.push_back(mm->add_parameter("output", mmlir.get_output_shapes().front()));
migraphx::gpu::context ctx;
migraphx::gpu::insert_mlir(*mm, mm->end(), compile_mlir(ctx, mmlir, inputs), inputs);
migraphx::gpu::insert_mlir(*mm, mm->end(), compile_mlir(ctx, mmlir, inputs, {}), inputs);
return p;
}
......
test/include/test.hpp
View file @ 64998903
......@@ -384,7 +384,7 @@ bool throws(F f, const std::string& msg = "")
}
template <class T, class U>
auto near(T px, U py, double ptol = 1e-6f)
auto within_abs(T px, U py, double ptol = 1e-6f)
{
return make_function("near", [](auto x, auto y, auto tol) { return std::abs(x - y) < tol; })(
px, py, ptol);
......
test/jit.cpp
View file @ 64998903
......@@ -82,9 +82,9 @@ TEST_CASE(generate_module)
auto f = compile_module<float(float, float)>(m);
EXPECT(test::near(f(2, 2), 2));
EXPECT(test::near(f(10, 6), 4));
EXPECT(test::near(f(1, 2), std::sqrt(3)));
EXPECT(test::within_abs(f(2, 2), 2));
EXPECT(test::within_abs(f(10, 6), 4));
EXPECT(test::within_abs(f(1, 2), std::sqrt(3)));
}
TEST_CASE(generate_module_with_literals)
......@@ -99,9 +99,9 @@ TEST_CASE(generate_module_with_literals)
auto f = compile_module<float(float, float)>(m);
EXPECT(test::near(f(1, 2), 2));
EXPECT(test::near(f(9, 6), 4));
EXPECT(test::near(f(0, 2), std::sqrt(3)));
EXPECT(test::within_abs(f(1, 2), 2));
EXPECT(test::within_abs(f(9, 6), 4));
EXPECT(test::within_abs(f(0, 2), std::sqrt(3)));
}
int main(int argc, const char* argv[]) { test::run(argc, argv); }
test/multi_target/multitarget_test.cpp
View file @ 64998903
......@@ -34,13 +34,13 @@
#include <migraphx/literal.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/shape.hpp>
#include <migraphx/verify.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/check_shapes.hpp>
#include <migraphx/functional.hpp>
#include <basic_ops.hpp>
#include <migraphx/compile_options.hpp>
#include <migraphx/register_target.hpp>
#include <migraphx/generate.hpp>
#include "test.hpp"
// check if it is custom_op or run_on_module operator
......@@ -180,38 +180,74 @@ TEST_CASE(multitarget_compile_cpu_gpu)
auto z_param = mm->add_parameter("z", s);
auto cpu_ins = mm->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 1}}), {x_param, y_param}, {cpu_mod});
auto cpu_ins_0 =
mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), cpu_ins);
auto gpu_ins = mm->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 0}}), {cpu_ins, z_param}, {gpu_mod});
mm->add_return({gpu_ins});
p.compile({migraphx::make_target("gpu"), migraphx::make_target("cpu")});
migraphx::make_op("run_on_target", {{"target_id", 0}}), {cpu_ins_0, z_param}, {gpu_mod});
auto gpu_ins_0 =
mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), gpu_ins);
mm->add_return({gpu_ins_0});
migraphx::compile_options gpu_opts;
gpu_opts.offload_copy = true;
p.compile({migraphx::make_target("gpu"), migraphx::make_target("cpu")}, {gpu_opts});
EXPECT(check_compiled_program(p, {migraphx::make_target("gpu"), migraphx::make_target("cpu")}));
migraphx::parameter_map params;
params["x"] = migraphx::fill_argument(s, 1);
params["y"] = migraphx::fill_argument(s, 2);
params["z"] = migraphx::fill_argument(s, 3);
auto result = p.eval(params).back();
auto gold = migraphx::fill_argument(s, 6);
EXPECT(gold == result);
}
TEST_CASE(single_target_compile)
TEST_CASE(single_target_multi_compile)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape boxes_s{migraphx::shape::float_type, {1, 6, 4}};
auto* mm = p.get_main_module();
auto boxes_param = mm->add_parameter("boxes", boxes_s);
auto* gpu_mod = p.create_module("gpu_mod");
auto boxes_param_gpu = gpu_mod->add_parameter("boxes_param_gpu", boxes_s);
migraphx::shape scores_s{migraphx::shape::float_type, {1, 1, 6}};
std::vector<float> scores_vec = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
auto boxes_l = mm->add_parameter("boxes", boxes_s);
auto scores_l = mm->add_literal(migraphx::literal(scores_s, scores_vec));
auto max_out_l = mm->add_literal(int64_t{4});
auto iou_threshold = mm->add_literal(0.5f);
auto score_threshold = mm->add_literal(0.0f);
auto r = mm->add_instruction(migraphx::make_op("nonmaxsuppression", {{"center_point_box", 1}}),
boxes_l,
auto scores_l = gpu_mod->add_literal(migraphx::literal(scores_s, scores_vec));
auto max_out_l = gpu_mod->add_literal(int64_t{4});
auto iou_threshold = gpu_mod->add_literal(0.5f);
auto score_threshold = gpu_mod->add_literal(0.0f);
auto r = gpu_mod->add_instruction(
migraphx::make_op("nonmaxsuppression",
{{"center_point_box", true}, {"use_dyn_output", true}}),
boxes_param_gpu,
scores_l,
max_out_l,
iou_threshold,
score_threshold);
mm->add_return({r});
p.compile(migraphx::make_target("gpu"));
EXPECT(is_compiled_gpu_module(*p.get_main_module()));
gpu_mod->add_return({r});
auto run_on_gpu = mm->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 0}}), {boxes_param}, {gpu_mod});
auto run_on_gpu_0 =
mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_on_gpu);
mm->add_return({run_on_gpu_0});
// compile using multi-target compilation path
migraphx::compile_options gpu_opts;
gpu_opts.offload_copy = true;
// need to add "ref" to avoid ambigious call to "compile()"
p.compile({migraphx::make_target("gpu"), migraphx::make_target("ref")}, {gpu_opts});
EXPECT(check_compiled_program(p, {migraphx::make_target("gpu"), migraphx::make_target("ref")}));
// eval
migraphx::parameter_map params;
std::vector<float> boxes_vec = {0.5, 0.5, 1.0, 1.0, 0.5, 0.6, 1.0, 1.0, 0.5, 0.4, 1.0, 1.0,
0.5, 10.5, 1.0, 1.0, 0.5, 10.6, 1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
params["boxes"] = migraphx::argument(boxes_s, boxes_vec.data());
auto output = p.eval(params).back();
std::vector<int64_t> gold_vec = {0, 0, 3, 0, 0, 0, 0, 0, 5};
auto gold =
migraphx::argument(migraphx::shape{migraphx::shape::int64_type, {3, 3}}, gold_vec.data());
EXPECT(output == gold);
}
TEST_CASE(multitarget_compile_if_then_else)
......@@ -225,53 +261,64 @@ TEST_CASE(multitarget_compile_if_then_else)
auto y = mm->add_parameter("y", ds);
auto* then_mod = p.create_module("if_gpu_mod");
std::vector<float> data1 = {0.384804, -1.77948, -0.453775, 0.477438, -1.06333, -1.12893};
std::vector<float> data1(ds.elements(), 1);
auto l1 = then_mod->add_literal(migraphx::literal(ds, data1));
auto a1 = then_mod->add_instruction(migraphx::make_op("add"), x, l1);
auto gpu_x = then_mod->add_parameter("gpu_x", ds);
auto a1 = then_mod->add_instruction(migraphx::make_op("add"), gpu_x, l1);
then_mod->add_return({a1});
auto* else_mod = p.create_module("else_cpu_mod");
std::vector<float> data2 = {-0.258047, 0.360394, 0.536804, -0.577762, 1.0217, 1.02442};
std::vector<float> data2(ds.elements(), 2);
auto l2 = else_mod->add_literal(migraphx::literal(ds, data2));
auto a2 = else_mod->add_instruction(migraphx::make_op("mul"), y, l2);
auto cpu_y = else_mod->add_parameter("cpu_y", ds);
auto a2 = else_mod->add_instruction(migraphx::make_op("mul"), cpu_y, l2);
else_mod->add_return({a2});
auto* run_on_cpu_mod = p.create_module("run_on_cpu");
auto run_cpu_ins = run_on_cpu_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 1}}), {}, {else_mod});
run_on_cpu_mod->add_return({run_cpu_ins});
migraphx::make_op("run_on_target", {{"target_id", 1}}), {y}, {else_mod});
auto run_cpu_ins_0 = run_on_cpu_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_cpu_ins);
run_on_cpu_mod->add_return({run_cpu_ins_0});
auto* run_on_gpu_mod = p.create_module("run_on_gpu");
auto run_gpu_ins = run_on_gpu_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 0}}), {}, {then_mod});
run_on_gpu_mod->add_return({run_gpu_ins});
migraphx::make_op("run_on_target", {{"target_id", 0}}), {x}, {then_mod});
auto run_gpu_ins_0 = run_on_gpu_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_gpu_ins);
run_on_gpu_mod->add_return({run_gpu_ins_0});
auto ret =
mm->add_instruction(migraphx::make_op("if"), {cond}, {run_on_gpu_mod, run_on_cpu_mod});
auto r = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), ret);
mm->add_return({r});
// compile
p.compile({migraphx::make_target("gpu"), migraphx::make_target("cpu")});
migraphx::compile_options gpu_opts;
gpu_opts.offload_copy = true;
p.compile({migraphx::make_target("gpu"), migraphx::make_target("cpu")}, {gpu_opts});
EXPECT(check_compiled_program(p, {migraphx::make_target("gpu"), migraphx::make_target("cpu")}));
migraphx::parameter_map params;
params["x"] = migraphx::fill_argument(ds, 2);
params["y"] = migraphx::fill_argument(ds, 3);
for(bool cond_val : {true, false})
{
params["cond"] = migraphx::argument(cond_s, &cond_val);
auto result = p.eval(params).back();
auto gold = migraphx::fill_argument(ds, (cond_val ? 3 : 6));
EXPECT(gold == result);
}
}
// TODO : FPGA compilation is broken right now, below test mentions fpga but doesn't compile for it
TEST_CASE(multitarget_compile_nested_if_then_else)
{
float seed = 0.0f;
std::mt19937 gen(seed);
std::uniform_real_distribution<> dis(0.0, 1.0);
auto get_random_values = [&](size_t elements) {
std::vector<float> rand_samples(elements);
std::generate(rand_samples.begin(), rand_samples.end(), [&]() { return dis(gen); });
return rand_samples;
};
std::unordered_map<std::size_t, std::size_t> counter_map = {{0, 0}, {1, 0}};
migraphx::shape ds{migraphx::shape::float_type, {2, 3}};
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape cond_s{migraphx::shape::bool_type};
auto cond = mm->add_parameter("cond", cond_s);
auto cond_0 = mm->add_parameter("cond_0", cond_s);
auto cond_1 = mm->add_parameter("cond_1", cond_s);
auto x = mm->add_parameter("x", ds);
auto y = mm->add_parameter("y", ds);
auto z = mm->add_parameter("z", ds);
......@@ -281,19 +328,21 @@ TEST_CASE(multitarget_compile_nested_if_then_else)
std::string mod_name =
"target_" + std::to_string(tid) + "_" + std::to_string(counter_map[tid]++);
auto* test_mod = prog.create_module(mod_name);
std::vector<float> data = get_random_values(ds.elements());
std::vector<float> data(ds.elements(), -1);
auto l1 = test_mod->add_literal(migraphx::literal(ds, data));
auto test_mod_param = test_mod->add_parameter(mod_name, ds);
// instruction with local literal and main_mod param as inputs
auto ins1 = test_mod->add_instruction(migraphx::make_op("add"), x, l1);
// instructinon with local param and local ins as inputs
auto ins2 = test_mod->add_instruction(migraphx::make_op("mul"), ins1, test_mod_param);
// instruction with local ins and parent ins as inputs
auto ins3 = test_mod->add_instruction(migraphx::make_op("sub"), ins2, inputs.front());
auto test_mod_param_0 = test_mod->add_parameter(mod_name + "_param_0", ds);
auto test_mod_param_1 = test_mod->add_parameter(mod_name + "_param_1", ds);
auto test_mod_param_2 = test_mod->add_parameter(mod_name + "_param_2", ds);
auto ins1 = test_mod->add_instruction(migraphx::make_op("add"), test_mod_param_0, l1);
auto ins2 = test_mod->add_instruction(migraphx::make_op("mul"), ins1, test_mod_param_1);
auto ins3 = test_mod->add_instruction(migraphx::make_op("sub"), ins2, test_mod_param_2);
test_mod->add_return({ins3});
auto* run_on_target_mod = prog.create_module("run_on_" + mod_name);
run_on_target_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", tid}}), {inputs.front()}, {test_mod});
auto run_ins = run_on_target_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", tid}}), inputs, {test_mod});
auto run_ins_0 = run_on_target_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_ins);
run_on_target_mod->add_return({run_ins_0});
return run_on_target_mod;
};
......@@ -307,15 +356,30 @@ TEST_CASE(multitarget_compile_nested_if_then_else)
ref_mod->add_return({ref_add});
auto* then_mod = p.create_module("then_mod");
auto then_mod_param = then_mod->add_parameter("then_mod_param", ds);
auto then_mod_ref_ins = then_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 3}}), {then_mod_param, y}, {ref_mod});
auto then_mod_cond = then_mod->add_parameter("then_mod_cond", cond_s);
auto then_mod_param_0 = then_mod->add_parameter("then_mod_param_0", ds);
auto then_mod_param_1 = then_mod->add_parameter("then_mod_param_1", ds);
auto then_mod_param_2 = then_mod->add_parameter("then_mod_param_2", ds);
auto then_mod_ref_ins =
then_mod->add_instruction(migraphx::make_op("run_on_target", {{"target_id", 3}}),
{then_mod_param_0, then_mod_param_1},
{ref_mod});
auto then_mod_ref_ins_0 = then_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), then_mod_ref_ins);
then_mod->add_instruction(
auto then_mod_if = then_mod->add_instruction(
migraphx::make_op("if"),
{cond},
{create_test_module(p, {z}, 1), create_test_module(p, {then_mod_ref_ins_0}, 0)});
{then_mod_cond,
then_mod_param_0,
then_mod_param_1,
then_mod_param_2,
then_mod_ref_ins_0,
then_mod_param_1,
then_mod_param_2},
{create_test_module(p, {then_mod_param_0, then_mod_param_1, then_mod_param_2}, 1),
create_test_module(p, {then_mod_ref_ins_0, then_mod_param_1, then_mod_param_2}, 0)});
auto then_mod_if_0 =
then_mod->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), then_mod_if);
then_mod->add_return({then_mod_if_0});
// create nested else_mod with multiple targets.
// else_mod has one instruction that runs a module on "fpga" and another instruction that
......@@ -327,52 +391,104 @@ TEST_CASE(multitarget_compile_nested_if_then_else)
fpga_mod->add_return({fpga_add});
auto* else_mod = p.create_module("else_mod");
auto else_mod_param = else_mod->add_parameter("else_mod_param", ds);
auto else_mod_fpga_ins = else_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 2}}), {else_mod_param, y}, {fpga_mod});
auto else_mod_cond = else_mod->add_parameter("else_mod_cond", cond_s);
auto else_mod_param_0 = else_mod->add_parameter("else_mod_param_0", ds);
auto else_mod_param_1 = else_mod->add_parameter("else_mod_param_1", ds);
auto else_mod_param_2 = else_mod->add_parameter("else_mod_param_2", ds);
auto else_mod_fpga_ins =
else_mod->add_instruction(migraphx::make_op("run_on_target", {{"target_id", 2}}),
{else_mod_param_0, else_mod_param_2},
{fpga_mod});
auto else_mod_fpga_ins_0 = else_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), else_mod_fpga_ins);
else_mod->add_instruction(migraphx::make_op("if"),
{cond},
{create_test_module(p, {else_mod_fpga_ins_0}, 0),
create_test_module(p, {else_mod_param}, 1)});
auto else_mod_if = else_mod->add_instruction(
migraphx::make_op("if"),
{else_mod_cond,
else_mod_fpga_ins_0,
else_mod_param_0,
else_mod_param_1,
else_mod_param_2,
else_mod_param_1,
else_mod_param_0},
{create_test_module(p, {else_mod_fpga_ins_0, else_mod_param_0, else_mod_param_1}, 0),
create_test_module(p, {else_mod_param_2, else_mod_param_1, else_mod_param_0}, 1)});
auto else_mod_if_0 =
else_mod->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), else_mod_if);
else_mod->add_return({else_mod_if_0});
// Create nested and multi-target main module using "If"
auto main_if_ins =
mm->add_instruction(migraphx::make_op("if"), {cond, x}, {then_mod, else_mod});
auto main_if_ins = mm->add_instruction(
migraphx::make_op("if"), {cond_0, cond_1, x, y, z, cond_1, x, y, z}, {then_mod, else_mod});
auto r = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), main_if_ins);
mm->add_return({r});
// compile
migraphx::compile_options gpu_opts;
gpu_opts.offload_copy = true;
p.compile({migraphx::make_target("gpu"),
migraphx::make_target("cpu"),
migraphx::make_target("fpga"),
migraphx::make_target("ref")});
migraphx::make_target("ref"),
migraphx::make_target("ref")},
{gpu_opts});
EXPECT(check_compiled_program(p,
{migraphx::make_target("gpu"),
migraphx::make_target("cpu"),
migraphx::make_target("fpga"),
migraphx::make_target("ref"),
migraphx::make_target("ref")}));
// do evaluation using different conditions
// TODO: make two conditional to cover all the paths
migraphx::parameter_map params;
float x_i = 2.0;
float y_i = 3.0;
float z_i = 4.0;
params["x"] = migraphx::fill_argument(ds, x_i);
params["y"] = migraphx::fill_argument(ds, y_i);
params["z"] = migraphx::fill_argument(ds, z_i);
// cover all paths with different combination of conditions
std::vector<std::pair<bool, bool>> test_conds = {
{true, true}, {true, false}, {false, true}, {false, false}};
for(auto [cond_val_0, cond_val_1] : test_conds)
{
params["cond_0"] = migraphx::argument(cond_s, &cond_val_0);
params["cond_1"] = migraphx::argument(cond_s, &cond_val_1);
auto result = p.eval(params).back();
// main has one instruction that is : if_then_else
// then mod is doing : {tmp = x+y; (cond) ? (((x-1)*y)-z) : (((tmp-1)*y)-z);}
// else mod is doing : {tmp = x+z; (cond) ? (((tmp-1)*x)-y) : (((z-1)*y)-x);}
float gold_i = -1.0;
if(cond_val_0)
{
float tmp_i = x_i + y_i;
gold_i = (cond_val_1) ? (((x_i - 1) * y_i) - z_i) : (((tmp_i - 1) * y_i) - z_i);
}
else
{
float tmp_i = x_i + z_i;
gold_i = (cond_val_1) ? (((tmp_i - 1) * x_i) - y_i) : (((z_i - 1) * y_i) - x_i);
}
auto gold = migraphx::fill_argument(ds, gold_i);
EXPECT(gold == result);
}
}
// TODO : FPGA compilation is broken right now, below test mentions fpga but doesn't compile for it
TEST_CASE(multitarget_select_module)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape lit_s{migraphx::shape{migraphx::shape::float_type, {1}}};
auto literal_ins = mm->add_literal(migraphx::literal{lit_s, {6}});
// create batch submodules
auto create_submodule = [&](std::size_t batch_size, const std::string& module_name) {
auto* submod = p.create_module(module_name);
migraphx::shape sm_shape{migraphx::shape::float_type, {batch_size, 4}};
auto sm_input = submod->add_parameter("data", sm_shape);
migraphx::shape lit_s{migraphx::shape{migraphx::shape::float_type, {1}}};
auto literal_ins = submod->add_literal(migraphx::literal{lit_s, {6}});
auto broadcast_lit =
submod->add_instruction(migraphx::make_op("multibroadcast"), literal_ins, sm_input);
auto add_ins0 = submod->add_instruction(migraphx::make_op("add"), sm_input, broadcast_lit);
auto add_ins1 = submod->add_instruction(migraphx::make_op("add"), add_ins0, broadcast_lit);
submod->add_return({add_ins0, add_ins1});
submod->add_return({add_ins1});
return submod;
};
auto* batch1 = create_submodule(1, "batch_1");
......@@ -380,36 +496,45 @@ TEST_CASE(multitarget_select_module)
auto* batch3 = create_submodule(3, "batch_3");
auto* batch4 = create_submodule(4, "batch_4");
migraphx::shape s{migraphx::shape::float_type, {{1, 4}, {4, 4}}};
auto input = mm->add_parameter("data", s);
auto* run_cpu_mod = p.create_module("cpu_mod");
auto cpu_param = run_cpu_mod->add_parameter(
"cpu_data", migraphx::shape{migraphx::shape::float_type, {1, 4}});
auto cpu_param =
run_cpu_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {1, 4}});
auto run_cpu_ins = run_cpu_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 1}}), {cpu_param}, {batch1});
run_cpu_mod->add_return({run_cpu_ins});
auto run_cpu_ins_0 = run_cpu_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_cpu_ins);
run_cpu_mod->add_return({run_cpu_ins_0});
auto* run_gpu_mod = p.create_module("gpu_mod");
auto gpu_param = run_gpu_mod->add_parameter(
"gpu_data", migraphx::shape{migraphx::shape::float_type, {2, 4}});
auto gpu_param =
run_gpu_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {2, 4}});
auto run_gpu_ins = run_gpu_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 0}}), {gpu_param}, {batch2});
run_gpu_mod->add_return({run_gpu_ins});
auto run_gpu_ins_0 = run_gpu_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_gpu_ins);
run_gpu_mod->add_return({run_gpu_ins_0});
auto* run_fpga_mod = p.create_module("fpga_mod");
auto fpga_param = run_fpga_mod->add_parameter(
"fpga_data", migraphx::shape{migraphx::shape::float_type, {3, 4}});
auto fpga_param =
run_fpga_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {3, 4}});
auto run_fpga_ins = run_fpga_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 2}}), {fpga_param}, {batch3});
run_fpga_mod->add_return({run_fpga_ins});
auto run_fpga_ins_0 = run_fpga_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_fpga_ins);
run_fpga_mod->add_return({run_fpga_ins_0});
auto* run_ref_mod = p.create_module("ref_mod");
auto ref_param = run_fpga_mod->add_parameter(
"ref_data", migraphx::shape{migraphx::shape::float_type, {4, 4}});
auto ref_param =
run_ref_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {4, 4}});
auto run_ref_ins = run_ref_mod->add_instruction(
migraphx::make_op("run_on_target", {{"target_id", 3}}), {ref_param}, {batch4});
run_ref_mod->add_return({run_ref_ins});
auto run_ref_ins_0 = run_ref_mod->add_instruction(
migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_ref_ins);
run_ref_mod->add_return({run_ref_ins_0});
auto* mm = p.get_main_module();
migraphx::shape dyn_s{migraphx::shape::float_type, {{1, 4}, {4, 4}}};
auto input = mm->add_parameter("data", dyn_s);
std::vector<migraphx::shape> sub_shapes = {};
sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {4, 4}}});
sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {4, 4}}});
......@@ -419,18 +544,34 @@ TEST_CASE(multitarget_select_module)
{input},
{run_cpu_mod, run_gpu_mod, run_fpga_mod, run_ref_mod});
auto ret0 = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), sm_ins);
auto ret1 = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 1}}), sm_ins);
mm->add_return({ret0, ret1});
mm->add_return({ret0});
// compile
migraphx::compile_options gpu_opts;
gpu_opts.offload_copy = true;
p.compile({migraphx::make_target("gpu"),
migraphx::make_target("cpu"),
migraphx::make_target("fpga"),
migraphx::make_target("ref")});
migraphx::make_target("ref"),
migraphx::make_target("ref")},
{gpu_opts});
EXPECT(check_compiled_program(p,
{migraphx::make_target("gpu"),
migraphx::make_target("cpu"),
migraphx::make_target("fpga"),
migraphx::make_target("ref"),
migraphx::make_target("ref")}));
// program does the 12+x where x has dynamic shape {{1, 4}, {4, 4}}
for(const size_t bs : {1, 2, 3, 4})
{
migraphx::shape arg_shape{migraphx::shape::float_type, {bs, 4}};
migraphx::parameter_map params;
params["data"] = migraphx::generate_argument(arg_shape, arg_shape.elements());
std::vector<float> input_data;
params["data"].visit([&](const auto& vec) { input_data.assign(vec.begin(), vec.end()); });
std::transform(input_data.begin(), input_data.end(), input_data.begin(), [](const auto& i) {
return i + 12.0;
});
auto result = p.eval(params).back();
EXPECT(migraphx::argument(arg_shape, input_data.data()) == result);
}
}
int main(int argc, const char* argv[]) { test::run(argc, argv); }
test/onnx/.onnxrt-commit
View file @ 64998903
3be6eb53c8b359703cb645ed2cb1cdf106924b7c
21a71d52bd2074b770807b209939ec11e2c64fa7
test/onnx/gen_onnx.py
View file @ 64998903
......@@ -6165,6 +6165,101 @@ def shape_test():
return ([node], [x], [y])
@onnx_test()
def shape_dyn_test0():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT,
[None, 4, None, None])
y = helper.make_tensor_value_info('y', TensorProto.INT64, [4])
node = onnx.helper.make_node(
'Shape',
inputs=['x'],
outputs=['y'],
)
return ([node], [x], [y])
@onnx_test()
def shape_dyn_test1():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT,
[None, 4, None, None])
y = helper.make_tensor_value_info('y', TensorProto.INT64, [2])
node = onnx.helper.make_node('Shape', inputs=['x'], outputs=['y'], start=2)
return ([node], [x], [y])
@onnx_test()
def shape_dyn_test2():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT,
[None, 4, None, None])
y = helper.make_tensor_value_info('y', TensorProto.INT64, [2])
node = onnx.helper.make_node('Shape',
inputs=['x'],
outputs=['y'],
start=-2)
return ([node], [x], [y])
@onnx_test()
def shape_dyn_test3():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT,
[None, 4, None, None])
y = helper.make_tensor_value_info('y', TensorProto.INT64, [2])
node = onnx.helper.make_node('Shape',
inputs=['x'],
outputs=['y'],
start=1,
end=2)
return ([node], [x], [y])
@onnx_test()
def shape_end_oob_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT,
[None, 4, None, None])
y = helper.make_tensor_value_info('y', TensorProto.INT64, [2])
node = onnx.helper.make_node('Shape', inputs=['x'], outputs=['y'], end=5)
return ([node], [x], [y])
@onnx_test()
def shape_start_oob_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT,
[None, 4, None, None])
y = helper.make_tensor_value_info('y', TensorProto.INT64, [2])
node = onnx.helper.make_node('Shape',
inputs=['x'],
outputs=['y'],
start=-6)
return ([node], [x], [y])
@onnx_test()
def shape_end_less_start_error():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT,
[None, 4, None, None])
y = helper.make_tensor_value_info('y', TensorProto.INT64, [2])
node = onnx.helper.make_node('Shape',
inputs=['x'],
outputs=['y'],
start=3,
end=1)
return ([node], [x], [y])
@onnx_test()
def shape_gather_test():
values = np.array([1])
......
test/onnx/onnx_test.cpp
View file @ 64998903
......@@ -440,14 +440,13 @@ TEST_CASE(batch_norm_flat_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {1}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {1}});
auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-6f}});
auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, scale});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), {var_eps});
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, bias});
auto prog = optimize_onnx("batch_norm_flat_test.onnx");
......@@ -465,14 +464,13 @@ TEST_CASE(batch_norm_rank_2_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {5}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {5}});
auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-6f}});
auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, scale});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), {var_eps});
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, bias});
auto prog = optimize_onnx("batch_norm_rank_2_test.onnx");
......@@ -490,7 +488,6 @@ TEST_CASE(batch_norm_1d_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {3}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {3}});
auto rt = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {1e-5f}});
auto usq_scale = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), scale);
......@@ -498,11 +495,11 @@ TEST_CASE(batch_norm_1d_test)
auto usq_mean = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), mean);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), var);
auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, usq_scale});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});
auto prog = optimize_onnx("batch_norm_1d_test.onnx");
......@@ -520,7 +517,6 @@ TEST_CASE(batch_norm_2d_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {3}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {3}});
auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-5f}});
auto usq_scale = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), scale);
......@@ -528,11 +524,11 @@ TEST_CASE(batch_norm_2d_test)
auto usq_mean = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), mean);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), var);
auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, usq_scale});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});
auto prog = optimize_onnx("batch_norm_2d_test.onnx");
......@@ -550,7 +546,6 @@ TEST_CASE(batch_norm_3d_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::half_type, {2}});
auto var = mm->add_parameter("variance", {migraphx::shape::half_type, {2}});
auto rt = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {1e-6f}});
auto usq_scale =
......@@ -561,12 +556,13 @@ TEST_CASE(batch_norm_3d_test)
mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2, 3}}}), mean);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2, 3}}}), var);
auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, usq_scale});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});
auto prog = optimize_onnx("batch_norm_3d_test.onnx");
EXPECT(p == prog);
......@@ -908,7 +904,6 @@ TEST_CASE(constant_test)
TEST_CASE(constant_fill_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {2, 3}};
......@@ -1105,7 +1100,6 @@ TEST_CASE(conv_bn_relu_maxpool_test)
auto p5 = mm->add_parameter("5", {migraphx::shape::float_type, {1}});
auto p6 = mm->add_parameter("6", {migraphx::shape::float_type, {1}});
auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-5f}});
uint64_t axis = 1;
......@@ -1120,25 +1114,12 @@ TEST_CASE(conv_bn_relu_maxpool_test)
auto usq_mean = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), p5);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), p6);
auto mb_mean = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), usq_mean);
auto numer = mm->add_instruction(migraphx::make_op("sub"), l5, mb_mean);
auto mb_eps =
mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 1}}}), eps);
auto var_eps = mm->add_instruction(migraphx::make_op("add"), usq_var, mb_eps);
auto mb_rt =
mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 1}}}), rt);
auto denom = mm->add_instruction(migraphx::make_op("pow"), var_eps, mb_rt);
auto mb_denom = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), denom);
auto div0 = mm->add_instruction(migraphx::make_op("div"), numer, mb_denom);
auto mb_scale = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), usq_scale);
auto r0 = mm->add_instruction(migraphx::make_op("mul"), div0, mb_scale);
auto mb_bias = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), usq_bias);
auto l6 = mm->add_instruction(migraphx::make_op("add"), r0, mb_bias);
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {l5, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
auto l6 = add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});
auto l7 = mm->add_instruction(migraphx::make_op("relu"), l6);
mm->add_instruction(migraphx::make_op("pooling",
......@@ -6079,6 +6060,118 @@ TEST_CASE(shape_test)
EXPECT(p == prog);
}
TEST_CASE(shape_dyn_test0)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}}};
auto p0 = mm->add_parameter("x", s);
migraphx::shape s_shape{migraphx::shape::int64_type, {4}};
auto ret = mm->add_instruction(migraphx::make_op("dimensions_of", {{"end", 4}}), p0);
mm->add_return({ret});
migraphx::onnx_options options;
options.map_dyn_input_dims["x"] = {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}};
auto prog = parse_onnx("shape_dyn_test0.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(shape_dyn_test1)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}}};
auto p0 = mm->add_parameter("x", s);
migraphx::shape s_shape{migraphx::shape::int64_type, {4}};
auto ret =
mm->add_instruction(migraphx::make_op("dimensions_of", {{"start", 2}, {"end", 4}}), p0);
mm->add_return({ret});
migraphx::onnx_options options;
options.map_dyn_input_dims["x"] = {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}};
auto prog = parse_onnx("shape_dyn_test1.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(shape_dyn_test2)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}}};
auto p0 = mm->add_parameter("x", s);
migraphx::shape s_shape{migraphx::shape::int64_type, {4}};
auto ret =
mm->add_instruction(migraphx::make_op("dimensions_of", {{"start", 2}, {"end", 4}}), p0);
mm->add_return({ret});
migraphx::onnx_options options;
options.map_dyn_input_dims["x"] = {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}};
auto prog = parse_onnx("shape_dyn_test2.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(shape_dyn_test3)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}}};
auto p0 = mm->add_parameter("x", s);
migraphx::shape s_shape{migraphx::shape::int64_type, {4}};
auto ret =
mm->add_instruction(migraphx::make_op("dimensions_of", {{"start", 1}, {"end", 2}}), p0);
mm->add_return({ret});
migraphx::onnx_options options;
options.map_dyn_input_dims["x"] = {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}};
auto prog = parse_onnx("shape_dyn_test3.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(shape_end_oob_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}}};
auto p0 = mm->add_parameter("x", s);
migraphx::shape s_shape{migraphx::shape::int64_type, {4}};
auto ret = mm->add_instruction(migraphx::make_op("dimensions_of", {{"end", 4}}), p0);
mm->add_return({ret});
migraphx::onnx_options options;
options.map_dyn_input_dims["x"] = {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}};
auto prog = migraphx::parse_onnx("shape_end_oob_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(shape_start_oob_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}}};
auto p0 = mm->add_parameter("x", s);
migraphx::shape s_shape{migraphx::shape::int64_type, {4}};
auto ret = mm->add_instruction(migraphx::make_op("dimensions_of", {{"end", 4}}), p0);
mm->add_return({ret});
migraphx::onnx_options options;
options.map_dyn_input_dims["x"] = {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}};
auto prog = migraphx::parse_onnx("shape_start_oob_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(shape_end_less_start_error)
{
migraphx::onnx_options options;
options.map_dyn_input_dims["x"] = {{1, 4, {1, 4}}, {4, 4}, {2, 4}, {2, 4}};
EXPECT(test::throws([&] { migraphx::parse_onnx("shape_end_less_start_error.onnx", options); }));
}
TEST_CASE(shape_gather_test)
{
migraphx::program p;
......@@ -7150,7 +7243,8 @@ TEST_CASE(variable_batch_user_input_test6)
TEST_CASE(variable_batch_user_input_test7)
{
// if entry in map_dyn_input_dims is all fixed dynamic_dimensions, convert it to a static shape
// if entry in map_dyn_input_dims is all fixed dynamic_dimensions, convert it to a static
// shape
migraphx::program p;
auto* mm = p.get_main_module();
auto l0 = mm->add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 16, 16}});
......
test/onnx/selu_test.onnx
View file @ 64998903
No preview for this file type
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