Skip to content

GitLab

Commit e08b425f authored by charlie's avatar charlie
Browse files

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into refactor_dynamic_compute

parents fbe13c96 5fa42993
Show whitespace changes
Inline Side-by-side
Showing with 479 additions and 642 deletions
src/targets/gpu/fuse_mlir.cpp
View file @ e08b425f
......@@ -49,7 +49,7 @@ struct mlir_conv
std::string name() const { return "gpu::mlir_conv"; }
shape compute_shape(std::vector<shape> inputs, const std::vector<module_ref>& mods) const
{
check_shapes{inputs, *this}.standard();
check_shapes{inputs, *this}.packed_or_broadcasted();
if(mods.size() != 1)
MIGRAPHX_THROW("should have one submodule.");
if(inputs.size() < 2)
......@@ -70,6 +70,9 @@ MIGRAPHX_PRED_MATCHER(is_mlir_conv, instruction_ref ins)
auto group = v.at("group").to<int>();
if(group != 1)
return false;
// Avoid MLIR assertion: Index < Length && "Invalid index!"
if(ins->get_shape().lens().size() != 4)
return false;
return true;
}
......@@ -96,9 +99,10 @@ struct find_conv_pointwise
i.name());
}))
return;
// Only fuse with fp32 for now
// Only fuse with fp32/fp16
if(std::any_of(ins->inputs().begin(), ins->inputs().end(), [&](auto i) {
return i->get_shape().type() != shape::type_t::float_type;
return not contains({shape::type_t::float_type, shape::type_t::half_type},
i->get_shape().type());
}))
return;
std::sort(names.begin(), names.end());
......
src/targets/gpu/fuse_ops.cpp
View file @ e08b425f
......@@ -26,7 +26,6 @@
#include <migraphx/gpu/fuse_ops.hpp>
#include <migraphx/matcher.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/gpu/convolution.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/gemm.hpp>
......@@ -190,10 +189,12 @@ MIGRAPHX_PRED_MATCHER(fusable_conv, instruction_ref ins)
return false;
auto wei = ins->inputs().at(1)->get_shape();
assert(wei.lens().size() == 4);
auto conv = any_cast<miopen_convolution>(ins->get_operator());
if(conv.op.group > 1)
auto miopen_conv_op = ins->get_operator().to_value();
auto algo = miopen_conv_op.at("algo").to<miopenConvFwdAlgorithm_t>();
auto conv_op = from_value<op::convolution>(miopen_conv_op["op"]);
if(conv_op.group > 1)
return false;
if(wei.lens()[1] > 512 and conv.algo != miopenConvolutionFwdAlgoWinograd)
if(wei.lens()[1] > 512 and algo != miopenConvolutionFwdAlgoWinograd)
return false;
// Do not fuse non-symmetric input
......@@ -201,13 +202,12 @@ MIGRAPHX_PRED_MATCHER(fusable_conv, instruction_ref ins)
if(input_lens[2] != input_lens[3] or wei.lens()[2] != wei.lens()[3])
return false;
auto op = conv.op;
// Dont fuse winograd for non-3x3s since there is no fused windograd for those configs
if(conv.algo == miopenConvolutionFwdAlgoWinograd and wei.lens()[2] != 3 and
wei.lens()[3] != 3 and contains({{1, 1}}, op.stride))
if(algo == miopenConvolutionFwdAlgoWinograd and wei.lens()[2] != 3 and wei.lens()[3] != 3 and
contains({{1, 1}}, conv_op.stride))
return false;
return contains({{0, 0, 0, 0}, {1, 1, 1, 1}, {2, 2, 2, 2}}, op.padding) and
contains({{0, 0}, {1, 1}}, op.stride) and contains({{1, 1}}, op.dilation);
return contains({{0, 0, 0, 0}, {1, 1, 1, 1}, {2, 2, 2, 2}}, conv_op.padding) and
contains({{0, 0}, {1, 1}}, conv_op.stride) and contains({{1, 1}}, conv_op.dilation);
}
void move_broadcasted_back(std::vector<instruction_ref>& args)
......@@ -462,7 +462,7 @@ void apply_conv_bias(context& ctx, module& m, const match::matcher_result& r)
auto ins = r.result;
auto input_ins = conv_ins->inputs().at(0);
auto weights_ins = conv_ins->inputs().at(1);
auto conv_op = any_cast<miopen_convolution>(conv_ins->get_operator()).op;
auto conv_op = from_value<op::convolution>((conv_ins->get_operator()).to_value()["op"]);
auto alloc_ins = ins->inputs().back();
auto old_ws_ins = conv_ins->inputs().at(2);
......@@ -528,7 +528,7 @@ struct find_conv_pointwise
auto ins = r.result;
auto input_ins = conv_ins->inputs().at(0);
auto weights_ins = conv_ins->inputs().at(1);
auto conv_op = any_cast<miopen_convolution>(conv_ins->get_operator()).op;
auto conv_op = from_value<op::convolution>(conv_ins->get_operator().to_value()["op"]);
auto alloc_ins = ins->inputs().back();
module_ref pm = ins->module_inputs().front();
......
src/targets/gpu/include/migraphx/gpu/batch_norm_inference.hpp deleted 100644 → 0
View file @ fbe13c96
/*
* 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_RTGLIB_BATCHNORM_HPP
#define MIGRAPHX_GUARD_RTGLIB_BATCHNORM_HPP
#include <migraphx/argument.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/reflect.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct miopen_batch_norm_inference
{
op::batch_norm_inference op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
std::string name() const { return "gpu::batch_norm_inference"; }
shape compute_shape(const std::vector<shape>& inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/convolution.hpp
View file @ e08b425f
......@@ -25,18 +25,40 @@
#define MIGRAPHX_GUARD_RTGLIB_CONVOLUTION_HPP
#include <migraphx/shape.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/generate.hpp>
#include <migraphx/operation.hpp>
#include <migraphx/register_op.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/op/identity.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/op/deconvolution.hpp>
#include <unordered_map>
#include <migraphx/reflect.hpp>
#include <migraphx/gpu/context.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
inline shape reshape_if_1d(const shape& input)
{
shape new_shape{input};
auto dims = new_shape.lens();
if(dims.size() == 3)
{
std::vector<size_t> new_dims = dims;
new_dims.insert(new_dims.begin() + 2, 1);
new_shape = shape{input.type(), new_dims};
}
return new_shape;
}
template <class Op>
struct miopen_convolution
{
op::convolution op;
Op op;
bool int8_x4_format = false;
shared<convolution_descriptor> cd = nullptr;
miopenConvFwdAlgorithm_t algo{};
#ifdef MIGRAPHX_HAS_FIND_2_API
......@@ -48,29 +70,276 @@ struct miopen_convolution
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack(f(self.op.padding, "padding"),
f(self.op.stride, "stride"),
f(self.op.dilation, "dilation"),
f(self.op.group, "group"),
f(self.op.padding_mode, "padding_mode"),
return pack(f(self.op, "op"),
#ifdef MIGRAPHX_HAS_FIND_2_API
f(self.solution_object, "solution_object"),
#endif
f(self.algo, "algo"),
f(self.int8_x4_format, "int8_x4_format"),
f(self.solution_id, "solution_id"));
}
std::string name() const { return "gpu::convolution"; }
shape compute_shape(const std::vector<shape>& inputs) const;
std::string name() const { return "gpu::" + op.name(); }
inline shape compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, op}.has(4).standard();
std::vector<shape> conv_inputs(inputs.begin(), inputs.begin() + 2);
check_shapes{conv_inputs, op}.max_ndims(5);
return migraphx::compute_shape<Op>(op, conv_inputs);
}
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
shape find(context& ctx, const shape& output_shape, std::vector<shape> inputs);
void finalize(context& ctx, const shape& output_shape, const std::vector<shape>& inputs);
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
{
auto x_desc = make_tensor(reshape_if_1d(args[0].get_shape()), int8_x4_format);
auto w_desc = make_tensor(reshape_if_1d(args[1].get_shape()), int8_x4_format);
auto y_desc = make_tensor(reshape_if_1d(output_shape));
auto* miopen_stream_handle = ctx.get_stream().get_miopen();
auto workspace_size = args[2].get_shape().bytes();
#ifdef MIGRAPHX_HAS_FIND_2_API
{
const miopenTensorArgument_t tensor_args[3] = {
{miopenTensorConvolutionX, nullptr, args[0].implicit()},
{miopenTensorConvolutionW, nullptr, args[1].implicit()},
{miopenTensorConvolutionY, nullptr, args[3].implicit()},
};
if(solution_ptr.get() == nullptr)
MIGRAPHX_THROW("MIOpen " + op.name() + " : Load MIOpen Solution before running it");
auto status = miopenRunSolution(miopen_stream_handle,
solution_ptr.get(),
3,
tensor_args,
args[2].implicit(),
workspace_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() +
" : running convolution using find_2.0 failed");
return args[3];
}
#else
// else use immediate mode
if(solution_id == 0)
MIGRAPHX_THROW("MIOpen " + op.name() + " : invalid solution ID");
auto status = miopenConvolutionForwardImmediate(miopen_stream_handle,
w_desc.get(),
args[1].implicit(),
x_desc.get(),
args[0].implicit(),
cd.get(),
y_desc.get(),
args[3].implicit(),
args[2].implicit(),
workspace_size,
solution_id);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": running convolution failed");
return args[3];
#endif
}
inline void set_conv_descriptor()
{
if(cd == nullptr)
{
cd = (op.name() == "deconvolution") ? make_deconv(op) : make_conv(op);
}
}
value compile(migraphx::context& ctx, const shape& output, const std::vector<shape>& input)
{
set_conv_descriptor();
auto ws = find(any_cast<migraphx::gpu::context>(ctx), output, input);
return {{"workspace", ws.bytes()}};
}
shape find(context& ctx, const shape& output_shape, const std::vector<shape>& inputs)
{
shape workspace_shape{};
auto x_desc = make_tensor(reshape_if_1d(inputs[0]), int8_x4_format);
auto w_desc = make_tensor(reshape_if_1d(inputs[1]), int8_x4_format);
auto y_desc = make_tensor(reshape_if_1d(output_shape));
std::size_t workspace_size = 0;
#ifdef MIGRAPHX_HAS_FIND_2_API
{
auto conv_problem = make_obj<miopen_problem>(
&miopenCreateConvProblem, cd.get(), miopenProblemDirectionForward);
set_tensor_descriptor(miopenTensorConvolutionX, x_desc, conv_problem);
set_tensor_descriptor(miopenTensorConvolutionW, w_desc, conv_problem);
set_tensor_descriptor(miopenTensorConvolutionY, y_desc, conv_problem);
auto* miopen_stream_handle = ctx.get_stream().get_miopen();
solution_ptr = find_solution(miopen_stream_handle, conv_problem.get());
auto status = miopenGetSolutionWorkspaceSize(solution_ptr.get(), &workspace_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + " : failed to get solution's workspace size");
std::size_t solution_size;
status = miopenGetSolutionSize(solution_ptr.get(), &solution_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + ": Failed to fetch solution size");
auto solution_binary = std::vector<char>{};
solution_binary.resize(solution_size);
status = miopenSaveSolution(solution_ptr.get(), solution_binary.data());
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + ": Saving solution failed");
solution_object = value::binary{solution_binary.data(), solution_size};
return shape{shape::int8_type, {workspace_size}};
}
#else
auto status = miopenConvolutionForwardGetWorkSpaceSize(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
&workspace_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + " : Failed to get forward workspace size");
workspace_shape = shape{shape::int8_type, {workspace_size}};
auto x_shape = inputs[0];
auto w_shape = inputs[1];
if(int8_x4_format)
{
x_shape = pack_int8_shape(x_shape);
w_shape = pack_int8_shape(w_shape);
}
auto x = to_gpu(generate_argument(x_shape));
auto w = to_gpu(generate_argument(w_shape));
auto y = allocate_gpu(output_shape);
auto workspace = allocate_gpu(workspace_shape);
int algo_count = 1;
miopenConvAlgoPerf_t perf;
status = miopenFindConvolutionForwardAlgorithm(ctx.get_stream().get_miopen(),
x_desc.get(),
x.implicit(),
w_desc.get(),
w.implicit(),
cd.get(),
y_desc.get(),
y.implicit(),
1,
&algo_count,
&perf,
workspace.implicit(),
workspace_size,
false);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + " : find convolution failed");
algo = perf.fwd_algo;
size_t solution_count;
status = miopenConvolutionForwardGetSolutionCount(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
&solution_count);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": get solution count failed");
std::vector<miopenConvSolution_t> solutions(solution_count);
status = miopenConvolutionForwardGetSolution(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
solution_count,
&solution_count,
solutions.data());
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": get solution failed");
solution_id = solutions.front().solution_id;
return shape{shape::int8_type, {perf.memory}};
#endif
}
void finalize(context& ctx, const shape& output_shape, const std::vector<shape>& inputs)
{
#ifdef MIGRAPHX_HAS_FIND_2_API
{
(void)(ctx); // avoid warnings
(void)(output_shape);
(void)(inputs);
// load solution
if(solution_ptr == nullptr)
{
miopenSolution_t ptr;
auto status =
miopenLoadSolution(&ptr,
reinterpret_cast<const char*>(solution_object.data()),
solution_object.size());
solution_ptr = miopen_solution{ptr};
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": loading convolution solution failed");
}
}
#else
// Use immediate mode API
{
set_conv_descriptor();
if(solution_id == 0)
{
// Check that workspace hasn't changed
auto size = inputs.at(2).bytes();
auto ws = find(ctx, output_shape, inputs);
if(ws.bytes() > size)
MIGRAPHX_THROW("MIOpen " + op.name() +
": workspace has changed during finalization.");
}
auto x_desc = make_tensor(reshape_if_1d(inputs[0]), int8_x4_format);
auto w_desc = make_tensor(reshape_if_1d(inputs[1]), int8_x4_format);
auto y_desc = make_tensor(reshape_if_1d(output_shape));
auto status = miopenConvolutionForwardCompileSolution(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
solution_id);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen Convolution: compile solution failed");
}
#endif
}
inline std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
};
inline shape pack_int8_shape(const shape& s) const
{
if(s.type() != shape::int8_type)
{
return s;
}
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/include/migraphx/gpu/deconvolution.hpp deleted 100644 → 0
View file @ fbe13c96
/*
* 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_RTGLIB_DECONVOLUTION_HPP
#define MIGRAPHX_GUARD_RTGLIB_DECONVOLUTION_HPP
#include <migraphx/shape.hpp>
#include <migraphx/op/deconvolution.hpp>
#include <migraphx/gpu/miopen.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct miopen_deconvolution
{
op::deconvolution op;
shared<convolution_descriptor> cd;
miopenConvFwdAlgorithm_t algo{};
uint64_t solution_id = 0;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack_join(op::deconvolution::reflect(self.op, f),
pack(f(self.solution_id, "solution_id")));
}
std::string name() const { return "gpu::deconv"; }
shape compute_shape(const std::vector<shape>& inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
shape find(context& ctx, const shape& output_shape, std::vector<shape> inputs);
void finalize(context& ctx, const shape& output_shape, std::vector<shape> inputs);
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/mlir.hpp
View file @ e08b425f
......@@ -36,7 +36,8 @@ struct module;
namespace gpu {
std::string dump_mlir(const module& m);
code_object_op compile_mlir(const context& ctx, const module& m);
code_object_op
compile_mlir(const context& ctx, module m, const std::vector<instruction_ref>& inputs);
instruction_ref insert_mlir(module& m,
instruction_ref ins,
......
src/targets/gpu/include/migraphx/gpu/perfdb.hpp
View file @ e08b425f
......@@ -41,7 +41,7 @@ struct problem_params
shape output;
};
std::string get_mlir_perf_for_conv(const problem_params& pp);
std::string get_mlir_perf_for_conv(const problem_params& pp, bool xdlops);
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
......
src/targets/gpu/include/migraphx/gpu/quant_convolution.hpp deleted 100644 → 0
View file @ fbe13c96
/*
* 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_RTGLIB_QUANT_CONVOLUTION_HPP
#define MIGRAPHX_GUARD_RTGLIB_QUANT_CONVOLUTION_HPP
#include <migraphx/shape.hpp>
#include <migraphx/reflect.hpp>
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/gpu/miopen.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct miopen_quant_convolution
{
op::quant_convolution op;
bool int8_x4_format = false;
shared<convolution_descriptor> cd;
miopenConvFwdAlgorithm_t algo{};
uint64_t solution_id = 0;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
// TODO: Add algo
return pack_join(migraphx::reflect(self.op, f),
pack(f(self.int8_x4_format, "int8_x4_format")));
}
std::string name() const { return "gpu::quant_convolution"; }
shape compute_shape(const std::vector<shape>& inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
shape find(context& ctx, const shape& output_shape, std::vector<shape> inputs);
void finalize(context& ctx, const shape& output_shape, std::vector<shape> inputs);
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
private:
shape pack_int8_shape(const shape& s) const;
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/jit/mlir.cpp
View file @ e08b425f
......@@ -41,7 +41,7 @@ struct mlir_compiler : compiler<mlir_compiler>
{
auto* smod = ins->module_inputs().front();
assert(smod->get_parameter_names().size() == ins->inputs().size() - 1);
return insert(compile_mlir(ctx, *smod));
return insert(compile_mlir(ctx, *smod, ins->inputs()));
}
compiler_replace insert(code_object_op co) const
......
src/targets/gpu/lowering.cpp
View file @ e08b425f
......@@ -37,15 +37,12 @@
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/op/quant_dot.hpp>
#include <migraphx/gpu/batch_norm_inference.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/convolution.hpp>
#include <migraphx/gpu/deconvolution.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/gpu/gemm.hpp>
#include <migraphx/gpu/int8_conv_pack.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/gpu/quant_convolution.hpp>
#include <migraphx/gpu/rocblas.hpp>
#include <migraphx/gpu/compiler.hpp>
#include <migraphx/iterator_for.hpp>
......@@ -113,16 +110,15 @@ struct miopen_apply
add_extend_op("scatter_none");
add_extend_op("topk");
add_batch_norm_inference_op();
add_convolution_op();
add_deconvolution_op();
add_convolution_op<op::convolution>("convolution");
add_convolution_op<op::deconvolution>("deconvolution");
add_convolution_op<op::quant_convolution>("quant_convolution");
add_gemm_op<op::dot>("dot");
add_gemm_op<op::quant_dot>("quant_dot");
add_if_op();
add_loop_op();
add_neg_op();
add_nms_op();
add_quant_convolution_op();
}
void copy_params() const
......@@ -230,38 +226,6 @@ struct miopen_apply
return mod->insert_instruction(ins, make_op("allocate", {{"shape", to_value(s)}}));
}
void add_convolution_op()
{
apply_map.emplace("convolution", [=](instruction_ref ins) {
auto&& op = any_cast<op::convolution>(ins->get_operator());
auto conv = miopen_convolution{op, make_conv(op)};
auto ws = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
auto workspace = insert_allocation(ins, ws);
auto output = insert_allocation(ins, ins->get_shape());
return mod->replace_instruction(
ins, conv, ins->inputs().at(0), ins->inputs().at(1), workspace, output);
});
}
void add_deconvolution_op()
{
apply_map.emplace("deconvolution", [=](instruction_ref ins) {
auto&& op = any_cast<op::deconvolution>(ins->get_operator());
auto conv = miopen_deconvolution{op, make_deconv(op)};
auto ws = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
auto workspace = insert_allocation(ins, ws);
auto output = insert_allocation(ins, ins->get_shape());
return mod->replace_instruction(
ins, conv, ins->inputs().at(0), ins->inputs().at(1), workspace, output);
});
}
template <typename Op>
void add_gemm_op(const std::string& name)
{
......@@ -275,31 +239,33 @@ struct miopen_apply
});
}
void add_quant_convolution_op()
template <typename Op>
void add_convolution_op(const std::string& name)
{
apply_map.emplace("quant_convolution", [=](instruction_ref ins) {
auto&& op = any_cast<op::quant_convolution>(ins->get_operator());
shape ws;
miopen_quant_convolution conv;
auto compile_quant_conv_with_format = [&](bool format) {
conv = miopen_quant_convolution{op, format, make_conv(op)};
ws = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
apply_map.emplace(name, [=](instruction_ref ins) {
operation conv =
miopen_convolution<Op>{any_cast<Op>(ins->get_operator()), int8_x4_format};
migraphx::context ctx = get_context();
size_t ws_bytes = 0;
auto compile_conv_with_format = [&](bool format) {
conv = miopen_convolution<Op>{any_cast<Op>(ins->get_operator()), format};
auto ws = conv.compile(ctx, ins->get_shape(), to_shapes(ins->inputs()));
ws_bytes = ws.get("workspace", 0);
};
try
{
compile_quant_conv_with_format(int8_x4_format);
{ // for the regular convolution and deconvolution, this try would always succeed
compile_conv_with_format(int8_x4_format);
}
catch(migraphx::exception&)
{
// In case no solver supports the default format, retry using the other format.
compile_quant_conv_with_format(not int8_x4_format);
compile_conv_with_format(not int8_x4_format);
}
auto args = ins->inputs();
auto workspace = insert_allocation(ins, ws);
auto output = insert_allocation(ins, ins->get_shape());
auto workspace = insert_allocation(ins, shape{shape::int8_type, {ws_bytes}});
return mod->replace_instruction(ins, conv, args[0], args[1], workspace, output);
});
}
......@@ -334,43 +300,6 @@ struct miopen_apply
});
}
void add_batch_norm_inference_op()
{
apply_map.emplace("batch_norm_inference", [=](instruction_ref ins) {
auto&& op = any_cast<op::batch_norm_inference>(ins->get_operator());
auto output = insert_allocation(ins, ins->get_shape());
shape old_shape = ins->inputs().at(1)->get_shape();
auto input = ins->inputs()[0];
auto input_lens = input->get_shape().lens();
std::vector<int64_t> rsp_lens(input_lens.size(), 1);
// for per_activation case, also need to reshape input
if(op.bn_mode == op::batch_norm_inference::per_activation)
{
std::copy(input_lens.begin() + 1, input_lens.end(), rsp_lens.begin() + 1);
}
else
{
rsp_lens[1] = static_cast<int64_t>(old_shape.elements());
}
auto reshape_op = op::reshape{rsp_lens};
std::vector<instruction_ref> reshapes;
std::transform(ins->inputs().begin() + 1,
ins->inputs().end(),
std::back_inserter(reshapes),
[&](auto i) { return mod->insert_instruction(ins, reshape_op, i); });
return mod->replace_instruction(ins,
miopen_batch_norm_inference{op},
input,
reshapes[0],
reshapes[1],
reshapes[2],
reshapes[3],
output);
});
}
// use 0 - input to represent neg
void add_neg_op()
{
......
src/targets/gpu/mlir.cpp
View file @ e08b425f
......@@ -21,6 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include "migraphx/make_op.hpp"
#include <migraphx/gpu/mlir.hpp>
#ifdef MIGRAPHX_MLIR
......@@ -43,8 +44,9 @@
#include <migraphx/gpu/code_object_op.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/gpu/perfdb.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/permutation.hpp>
#include <deque>
#include <variant>
......@@ -194,7 +196,6 @@ struct mlir_program
MlirType make_tensor(const shape& s) const
{
assert(s.standard());
std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
return mlirRankedTensorTypeGet(
lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());
......@@ -502,11 +503,12 @@ struct mlir_program
{
pp =
problem_params{ins->get_operator(), to_shapes(ins->inputs()), ins->get_shape()};
std::string tuned = get_tune_params();
// check if HW supports xdlops
bool xdlops = contains(get_xdlops_archs(), target_name);
std::string tuned = get_tune_params(xdlops);
if(not tuned.empty())
ops.add_attributes({{"perf_config", tuned}});
// check if HW supports xdlops
if(contains(get_xdlops_archs(), target_name))
if(xdlops)
ops.add_attributes({{"xdlopsV2", true}});
}
......@@ -571,7 +573,7 @@ struct mlir_program
MIGRAPHX_THROW("Failed to compile mlir program");
}
std::string get_tune_params() { return get_mlir_perf_for_conv(pp); }
std::string get_tune_params(bool xdlops) { return get_mlir_perf_for_conv(pp, xdlops); }
mlir_context ctx;
MlirLocation location;
......@@ -589,8 +591,54 @@ std::string dump_mlir(const module& m)
return mlir_print(&mlirOperationPrint, mod_op);
}
code_object_op compile_mlir(const context&, const module& m)
void adjust_param_shapes(module& m, const std::vector<instruction_ref>& 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();
auto param = m.get_parameter(name);
if(input.standard())
continue;
auto lens = input.lens();
auto strides = input.strides();
std::vector<operation> ops;
if(input.transposed())
{
auto perm = find_permutation(input);
auto iperm = invert_permutation(perm);
lens = reorder_dims(lens, iperm);
strides = reorder_dims(strides, iperm);
ops.push_back(make_op("transpose", {{"permutation", perm}}));
}
if(input.broadcasted())
{
std::transform(lens.begin(),
lens.end(),
strides.begin(),
lens.begin(),
[](auto len, auto stride) -> std::size_t {
if(stride == 0)
return 1;
return len;
});
ops.push_back(make_op("multibroadcast", {{"out_lens", input.lens()}}));
}
auto new_param =
std::accumulate(ops.begin(),
ops.end(),
m.add_parameter(name + ".0", shape{input.type(), lens}),
[&](auto x, auto op) { return m.insert_instruction(param, op, x); });
m.replace_instruction(param, new_param);
m.remove_instruction(param);
}
}
code_object_op compile_mlir(const context&, module m, const std::vector<instruction_ref>& inputs)
{
adjust_param_shapes(m, inputs);
const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
if(trace)
std::cout << m << std::endl;
......@@ -662,13 +710,19 @@ instruction_ref insert_mlir(module& m,
std::string dump_mlir(const module&) { return {}; }
code_object_op compile_mlir(const context&, const module&) { return {}; }
template <class T>
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>&)
{
return {};
}
// NOLINTEND(performance-unnecessary-value-param)
instruction_ref
// cppcheck-suppress funcArgNamesDifferent
insert_mlir(module& m, instruction_ref, code_object_op co, const std::vector<instruction_ref>&)
......
src/targets/gpu/perfdb.cpp
View file @ e08b425f
......@@ -108,16 +108,17 @@ auto query_miopen_db(const std::string& query)
} // namespace
std::string get_mlir_perf_for_conv(const problem_params& pp)
std::string get_mlir_perf_for_conv(const problem_params& pp, bool xdlops)
{
std::string solver = xdlops ? "ConvMlirIgemmFwdXdlops" : "ConvMlirIgemmFwd";
std::string query = "select P.* \
from perf_db P, config C \
where P.config = C.id AND \
P.solver = 'ConvMlirIgemmFwdXdlops' AND \
P.solver = '${solver}' AND \
${config}";
auto results =
query_miopen_db(interpolate_string(query, {{"config", generate_miopen_config(pp)}}));
auto results = query_miopen_db(
interpolate_string(query, {{"config", generate_miopen_config(pp)}, {"solver", solver}}));
if(results.empty())
return "";
return results.front().at("params");
......
src/targets/gpu/quant_convolution.cpp deleted 100644 → 0
View file @ fbe13c96
/*
* 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 <migraphx/gpu/quant_convolution.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/generate.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
shape miopen_quant_convolution::compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(4).standard();
return op.normalize_compute_shape({inputs.at(0), inputs.at(1)});
}
argument miopen_quant_convolution::compute(context& ctx,
const shape& output_shape,
const std::vector<argument>& args) const
{
auto x_desc = make_tensor(args[0].get_shape(), int8_x4_format);
auto w_desc = make_tensor(args[1].get_shape(), int8_x4_format);
auto y_desc = make_tensor(output_shape);
float alpha = 1;
float beta = 0;
auto status = miopenConvolutionForward(ctx.get_stream().get_miopen(),
&alpha,
x_desc.get(),
args[0].implicit(),
w_desc.get(),
args[1].implicit(),
cd.get(),
algo,
&beta,
y_desc.get(),
args[3].implicit(),
args[2].implicit(),
args[2].get_shape().bytes());
if(status != miopenStatusSuccess)
{
MIGRAPHX_THROW("QUANT_CONVOLUTION: run convolution forward failed");
}
return args[3];
}
shape miopen_quant_convolution::find(context& ctx,
const shape& output_shape,
std::vector<shape> inputs)
{
shape workspace_shape{};
auto x_desc = make_tensor(inputs[0], int8_x4_format);
auto w_desc = make_tensor(inputs[1], int8_x4_format);
auto y_desc = make_tensor(output_shape);
std::size_t workspace_size = 0;
miopenConvolutionForwardGetWorkSpaceSize(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
&workspace_size);
workspace_shape = shape{shape::int8_type, {workspace_size}};
auto x_shape = inputs[0];
auto w_shape = inputs[1];
if(int8_x4_format)
{
x_shape = pack_int8_shape(x_shape);
w_shape = pack_int8_shape(w_shape);
}
auto x = to_gpu(generate_argument(x_shape));
auto w = to_gpu(generate_argument(w_shape));
auto y = allocate_gpu(output_shape);
auto workspace = allocate_gpu(workspace_shape);
int algo_count = 1;
miopenConvAlgoPerf_t perf;
auto status = miopenFindConvolutionForwardAlgorithm(ctx.get_stream().get_miopen(),
x_desc.get(),
x.implicit(),
w_desc.get(),
w.implicit(),
cd.get(),
y_desc.get(),
y.implicit(),
1,
&algo_count,
&perf,
workspace.implicit(),
workspace_size,
false);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen Quant Convolution: find convolution failed");
algo = perf.fwd_algo;
size_t solution_count;
status = miopenConvolutionForwardGetSolutionCount(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
&solution_count);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen Quant Convolution: get solution count failed");
std::vector<miopenConvSolution_t> solutions(solution_count);
status = miopenConvolutionForwardGetSolution(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
solution_count,
&solution_count,
solutions.data());
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen Quant Convolution: get solution failed");
solution_id = solutions.front().solution_id;
return shape{shape::int8_type, {perf.memory}};
}
void miopen_quant_convolution::finalize(context& ctx,
const shape& output_shape,
std::vector<shape> inputs)
{
if(cd == nullptr)
cd = make_conv(op);
if(solution_id == 0)
{
// Check that workspace hasn't changed
auto size = inputs.at(2).bytes();
auto ws = find(ctx, output_shape, inputs);
if(ws.bytes() > size)
MIGRAPHX_THROW("MIOpen Quant Convolution: workspace has changed during finalization.");
}
auto x_desc = make_tensor(inputs[0], int8_x4_format);
auto w_desc = make_tensor(inputs[1], int8_x4_format);
auto y_desc = make_tensor(output_shape);
auto status = miopenConvolutionForwardCompileSolution(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
solution_id);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen Quant Convolution: compile solution failed");
}
shape miopen_quant_convolution::pack_int8_shape(const shape& s) const
{
if(s.type() != shape::int8_type)
{
MIGRAPHX_THROW("PACK_INT8_SHAPE: 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/target.cpp
View file @ e08b425f
......@@ -41,7 +41,6 @@
#include <migraphx/propagate_constant.hpp>
#include <migraphx/register_target.hpp>
#include <migraphx/replace_allocate.hpp>
#include <migraphx/rewrite_batchnorm.hpp>
#include <migraphx/rewrite_gelu.hpp>
#include <migraphx/rewrite_pooling.hpp>
#include <migraphx/rewrite_quantization.hpp>
......@@ -110,8 +109,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
dead_code_elimination{},
insert_pad{},
dead_code_elimination{},
rewrite_batchnorm{},
dead_code_elimination{},
rewrite_rnn{},
dead_code_elimination{},
inline_module{},
......
src/targets/ref/lowering.cpp
View file @ e08b425f
......@@ -26,7 +26,6 @@
#include <migraphx/instruction.hpp>
#include <migraphx/dfor.hpp>
#include <migraphx/op/identity.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/op/deconvolution.hpp>
#include <migraphx/op/quant_convolution.hpp>
......@@ -73,84 +72,6 @@ typename std::conditional_t<std::is_integral<T>{}, std::make_signed<T>, std::ena
return x;
}
//
// ref implemenataion of batch norm for inference
//
// inputs are:
// args[0] -> input data buffer
// args[1] -> mini batch mean
// args[2] -> mini batch variance
// args[3] -> gamma
// args[4] -> bias
//
// The equation to compute batch norm for inference is:
//
// output[i] = bias + gamma * (input[i] + mean) / sqrt(variance + epsilon)
//
// the input data format should be nchw
//
struct ref_batch_norm_inference
{
op::batch_norm_inference op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
std::string name() const { return "ref::batch_norm_inference"; }
shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
argument compute(context&, const shape& output_shape, std::vector<argument> args) const
{
argument output{output_shape};
double epsilon = op.epsilon;
auto input = args[0];
auto arg_gamma = args[1];
auto arg_bias = args[2];
auto mini_batch_mean = args[3];
auto mini_batch_variance = args[4];
if(op.bn_mode == op::batch_norm_inference::spatial)
{
visit_all(output, input, mini_batch_mean, mini_batch_variance, arg_gamma, arg_bias)(
[&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {
par_for(output_shape.elements(), [&](auto i) {
auto idx = output_shape.multi(i);
auto c = idx[1];
assert((variance[c] + epsilon) > 0);
result[i] =
gamma[c] * (buffer[i] - mean[c]) / std::sqrt(variance[c] + epsilon) +
bias[c];
});
});
}
if(op.bn_mode == op::batch_norm_inference::per_activation)
{
visit_all(output, input, mini_batch_mean, mini_batch_variance, arg_gamma, arg_bias)(
[&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {
par_for(output_shape.elements(), [&](auto i) {
auto idx = output_shape.multi(i);
idx[0] = 0;
auto index = output_shape.index(idx);
assert((variance[index] + epsilon) > 0);
result[i] = gamma[index] * (buffer[i] - mean[index]) /
std::sqrt(variance[index] + epsilon) +
bias[index];
});
});
}
return output;
}
};
MIGRAPHX_REGISTER_OP(ref_batch_norm_inference)
struct ref_lrn
{
op::lrn op;
......@@ -643,8 +564,6 @@ struct ref_apply
void init()
{
apply_map["batch_norm_inference"] =
extend_op<ref_batch_norm_inference, op::batch_norm_inference>();
apply_map["convolution"] = extend_op<ref_convolution<op::convolution>, op::convolution>();
apply_map["dot"] = extend_op<ref_gemm, op::dot>();
apply_map["quant_dot"] = extend_op<ref_quant_gemm, op::quant_dot>();
......
src/tf/parse_batchnorm.cpp
View file @ e08b425f
......@@ -23,6 +23,7 @@
*/
#include <migraphx/tf/op_parser.hpp>
#include <migraphx/tf/tf_parser.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/make_op.hpp>
......@@ -38,16 +39,37 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
instruction_ref parse(const op_desc& /*opd*/,
const tf_parser& /*parser*/,
tf_parser::node_info info,
const std::vector<instruction_ref>& args) const
std::vector<instruction_ref> args) const
{
float epsilon = 1e-5f;
float momentum = 0.9f;
// different default epsilon than from ONNX
float epsilon = 1e-4f;
if(contains(info.attributes, "epsilon"))
{
epsilon = info.attributes.at("epsilon").f();
}
auto op = make_op("batch_norm_inference", {{"epsilon", epsilon}, {"momentum", momentum}});
return info.add_instruction(op, args);
auto x_lens = args[0]->get_shape().lens();
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 =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[1]);
auto bias_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[2]);
auto mean_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[3]);
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 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);
return info.add_broadcastable_binary_op("add", r0, bias_unsqueeze);
}
};
......
test/api/test_custom_op_gpu.cpp
View file @ e08b425f
......@@ -55,7 +55,8 @@ struct half_copy_host final : migraphx::experimental_custom_op_base
hipMemcpyHostToHost,
ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
MIGRAPHX_HIP_ASSERT(hipMemset(output_buffer_ptr, 0, copy_bytes));
MIGRAPHX_HIP_ASSERT(
hipMemsetAsync(output_buffer_ptr, 0, copy_bytes, ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
return inputs[1];
}
......@@ -97,7 +98,8 @@ struct half_copy_device final : migraphx::experimental_custom_op_base
hipMemcpyDeviceToDevice,
ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
MIGRAPHX_HIP_ASSERT(hipMemset(output_buffer_ptr, 0, copy_bytes));
MIGRAPHX_HIP_ASSERT(
hipMemsetAsync(output_buffer_ptr, 0, copy_bytes, ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
return inputs[1];
}
......@@ -124,7 +126,7 @@ struct half_copy_device_same_buffer final : migraphx::experimental_custom_op_bas
virtual bool runs_on_offload_target() const override { return true; }
virtual migraphx::argument
compute(migraphx::context, migraphx::shape, migraphx::arguments inputs) const override
compute(migraphx::context ctx, migraphx::shape, migraphx::arguments inputs) const override
{
// This custom op simply sets first half size_bytes of the input 0, and rest of the half
// bytes are copied. for this custom_op, it does its computation on the "device". Therefore,
......@@ -133,7 +135,8 @@ struct half_copy_device_same_buffer final : migraphx::experimental_custom_op_bas
auto input_bytes = inputs[0].get_shape().bytes();
auto copy_bytes = input_bytes / 2;
MIGRAPHX_HIP_ASSERT(hipSetDevice(0));
MIGRAPHX_HIP_ASSERT(hipMemset(buffer_ptr, 0, copy_bytes));
MIGRAPHX_HIP_ASSERT(
hipMemsetAsync(buffer_ptr, 0, copy_bytes, ctx.get_queue<hipStream_t>()));
MIGRAPHX_HIP_ASSERT(hipDeviceSynchronize());
return inputs[0];
}
......
test/gpu/mlir.cpp
View file @ e08b425f
......@@ -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);
migraphx::gpu::insert_mlir(*mm, mm->end(), compile_mlir(ctx, mmlir, inputs), inputs);
return p;
}
......
test/onnx/batch_norm_invalid_rank_test.onnx → test/onnx/batch_norm_rank_2_test.onnx
View file @ e08b425f
batch_norm_invalid_rank_test:
7
batch_norm_rank_2_test:
J
x
scale
bias
mean
variancey"BatchNormalizationbatch_norm_invalid_rank_testZ
variancey"BatchNormalization*
epsilon75batch_norm_rank_2_testZ
x


Z

Z
scale

Z
Z
bias

Z
Z
mean

Z
Z
variance

b
b
y


B
\ No newline at end of file

B
\ No newline at end of file
test/onnx/gen_onnx.py
View file @ e08b425f
......@@ -331,6 +331,24 @@ def batch_norm_flat_test():
return ([node], [x, scale, bias, mean, var], [out])
@onnx_test
def batch_norm_rank_2_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [2, 5])
scale = helper.make_tensor_value_info('scale', TensorProto.FLOAT, [5])
bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [5])
mean = helper.make_tensor_value_info('mean', TensorProto.FLOAT, [5])
var = helper.make_tensor_value_info('variance', TensorProto.FLOAT, [5])
out = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 5])
node = onnx.helper.make_node(
'BatchNormalization',
inputs=['x', 'scale', 'bias', 'mean', 'variance'],
outputs=['y'],
epsilon=1e-6)
return ([node], [x, scale, bias, mean, var], [out])
@onnx_test
def batch_norm_1d_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT16, [2, 3, 4])
......@@ -385,23 +403,6 @@ def batch_norm_3d_test():
return ([node], [x, scale, bias, mean, var], [out])
@onnx_test
def batch_norm_invalid_rank_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [8, 8])
scale = helper.make_tensor_value_info('scale', TensorProto.FLOAT, [8])
bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [8])
mean = helper.make_tensor_value_info('mean', TensorProto.FLOAT, [8])
var = helper.make_tensor_value_info('variance', TensorProto.FLOAT, [8])
out = helper.make_tensor_value_info('y', TensorProto.FLOAT, [8, 8])
node = onnx.helper.make_node(
'BatchNormalization',
inputs=['x', 'scale', 'bias', 'mean', 'variance'],
outputs=['y'])
return ([node], [x, scale, bias, mean, var], [out])
@onnx_test
def batch_norm_invalid_bias_rank_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [2, 3, 4, 4])
......
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