Skip to content

GitLab

Unverified Commit 881a4bd4 authored by Umang Yadav's avatar Umang Yadav Committed by GitHub
Browse files

Merge branch 'develop' into fix_parse_if

parents a2d710e3 4b1c1c41
Hide whitespace changes
Inline Side-by-side
Showing with 347 additions and 371 deletions
src/targets/gpu/include/migraphx/gpu/mlir.hpp
View file @ 881a4bd4
......@@ -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 @ 881a4bd4
......@@ -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/jit/mlir.cpp
View file @ 881a4bd4
......@@ -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/leaky_relu.cpp → src/targets/gpu/jit/pad.cpp
View file @ 881a4bd4
......@@ -21,45 +21,80 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <migraphx/gpu/leaky_relu.hpp>
#include <migraphx/gpu/compiler.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/gpu/compile_hip_code_object.hpp>
#include <migraphx/gpu/compile_hip.hpp>
#include <migraphx/gpu/compile_gen.hpp>
#include <migraphx/reduce_dims.hpp>
#include <migraphx/float_equal.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
shape miopen_leaky_relu::compute_shape(const std::vector<shape>& inputs) const
using namespace migraphx::gpu::gen; // NOLINT
static const char* const pointwise_kernel = R"__migraphx__(
#include <migraphx/kernels/pad.hpp>
#include <migraphx/kernels/index.hpp>
#include <migraphx/kernels/ops.hpp>
#include <args.hpp>
namespace migraphx {
extern "C" {
__global__ void pad_kernel(void* input_p, void* output_p)
{
check_shapes{inputs, *this}.has(2).not_broadcasted();
return inputs.at(1);
auto offsets = index_ints<${offsets}>{};
auto idx = make_index();
make_tensors()(input_p, output_p)([&](auto input, auto output) {
pad(idx, offsets, input, output, ${pad_val});
});
}
}
argument miopen_leaky_relu::compute(context& ctx,
const shape& output_shape,
const std::vector<argument>& args) const
{
float alpha = 1;
float beta = 0;
auto x_desc = make_tensor(args[0].get_shape());
auto y_desc = make_tensor(output_shape);
miopenActivationForward(ctx.get_stream().get_miopen(),
ad.get(),
&alpha,
x_desc.get(),
args[0].implicit(),
&beta,
y_desc.get(),
args[1].implicit());
} // namespace migraphx
return args[1];
}
)__migraphx__";
void miopen_leaky_relu::finalize(context&, const shape&, const std::vector<shape>&)
struct pad_compiler : compiler<pad_compiler>
{
ad = make_leaky_relu(op.alpha);
}
std::vector<std::string> names() const { return {"pad"}; }
operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
{
hip_compile_options options;
options.inputs = inputs;
options.output = inputs.back();
options.virtual_inputs = reduce_dims(inputs);
options.kernel_name = "pad_kernel";
options.set_launch_params(v, compute_global_for(ctx, inputs.at(1).elements()));
auto pad_val = v.get("value", 0.f);
auto pad_val_string = to_string(pad_val);
if(float_equal(pad_val, std::numeric_limits<float>::lowest()))
pad_val_string = "lowest{}";
if(float_equal(pad_val, std::numeric_limits<float>::max()))
pad_val_string = "highest{}";
auto padding = v.at("pads").to_vector<int64_t>();
auto input_lens = inputs.front().lens();
std::vector<size_t> offsets(input_lens.size());
std::copy(padding.begin(), padding.begin() + offsets.size(), offsets.begin());
auto src = interpolate_string(
pointwise_kernel,
{{"pad_val", to_string(pad_val_string)}, {"offsets", to_string_range(offsets)}});
return compile_hip_code_object(src, options);
}
compiler_replace compile(context& ctx, instruction_ref ins, const operation& op) const
{
return replace(compile_op(ctx, to_shapes(ins->inputs()), op.to_value()));
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/jit/scatternd.cpp
View file @ 881a4bd4
......@@ -79,9 +79,10 @@ struct scatternd_compiler : compiler<scatternd_compiler>
{
assert(starts_with(op.name(), "scatternd_"));
auto reduction = op.name().substr(10);
return insert(compile_op(ctx,
to_shapes({ins->inputs().begin() + 1, ins->inputs().end()}),
{{"reduction", reduction}}));
return insert(compile_op(
ctx,
to_shapes(std::vector<instruction_ref>{ins->inputs().begin() + 1, ins->inputs().end()}),
{{"reduction", reduction}}));
}
compiler_replace insert(const operation& op) const
......
src/targets/gpu/include/migraphx/gpu/quant_convolution.hpp → src/targets/gpu/kernels/include/migraphx/kernels/pad.hpp
View file @ 881a4bd4
......@@ -21,53 +21,43 @@
* 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
#ifndef MIGRAPHX_GUARD_KERNELS_PAD_HPP
#define MIGRAPHX_GUARD_KERNELS_PAD_HPP
#include <migraphx/shape.hpp>
#include <migraphx/reflect.hpp>
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/kernels/shape.hpp>
#include <migraphx/kernels/index.hpp>
#include <migraphx/kernels/algorithm.hpp>
#include <migraphx/kernels/ranges.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct miopen_quant_convolution
template <class Offsets, class Input, class Output, class PadVal>
__device__ void pad(const index& idx,
const Offsets& offsets,
const Input& input,
Output& output,
const PadVal& pad_val)
{
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;
}
auto output_shape = output.get_shape();
idx.global_stride(output_shape.elements(), [&](auto i) {
// 1. get current multi-index for output
// 2. get the size of the input to determine input boundaries
// 3. compute the corresponding multi-index for input by accounting for offsets
// 4. if current multi-index is within offsets or input's new multi-index is out of bounds,
// use pad value instead of input's value
auto multi = output_shape.multi(i);
auto input_bounds = input.get_shape().lens;
auto input_idx = multi - offsets;
auto range_multi = range(multi.size());
if(any_of(range_multi.begin(), range_multi.end(), [&](auto j) {
return multi[j] < offsets[j] or input_idx[j] >= input_bounds[j];
}))
output[multi] = pad_val;
else
output[multi] = input[input_idx];
});
}
private:
shape pack_int8_shape(const shape& s) const;
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/elu.hpp → src/targets/gpu/kernels/include/migraphx/kernels/ranges.hpp
View file @ 881a4bd4
......@@ -21,44 +21,29 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_RTGLIB_ELU_HPP
#define MIGRAPHX_GUARD_RTGLIB_ELU_HPP
#ifndef MIGRAPHX_GUARD_KERNELS_RANGES_HPP
#define MIGRAPHX_GUARD_KERNELS_RANGES_HPP
#include <migraphx/op/elu.hpp>
#include <migraphx/shape.hpp>
#include <migraphx/reflect.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/kernels/iota_iterator.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct miopen_elu
template <class Iterator>
struct iterator_range
{
op::elu op;
shared<activation_descriptor> ad;
Iterator start;
Iterator last;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
constexpr Iterator begin() const { return start; }
std::string name() const { return "gpu::elu"; }
shape compute_shape(const std::vector<shape>& inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
void finalize(context&, const shape&, const std::vector<shape>&);
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
constexpr Iterator end() const { return last; }
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
constexpr iterator_range<iota_iterator> range(diff_int start, diff_int last)
{
return {{start, {}}, {last, {}}};
}
constexpr iterator_range<iota_iterator> range(diff_int last) { return range(0, last); }
#endif
} // namespace migraphx
#endif // MIGRAPHX_GUARD_KERNELS_RANGES_HPP
src/targets/gpu/lowering.cpp
View file @ 881a4bd4
......@@ -39,12 +39,10 @@
#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>
......@@ -97,14 +95,11 @@ struct miopen_apply
add_extend_op("argmax");
add_extend_op("argmin");
add_extend_op("elu");
add_extend_op("gather");
add_extend_op("leaky_relu");
add_extend_op("logsoftmax");
add_extend_op("lrn");
add_extend_op("multinomial");
add_extend_op("nonzero");
add_extend_op("pad");
add_extend_op("pooling");
add_extend_op("prefix_scan_sum");
add_extend_op("reverse");
......@@ -114,15 +109,15 @@ struct miopen_apply
add_extend_op("scatter_none");
add_extend_op("topk");
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 +225,6 @@ 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 +238,33 @@ 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);
});
}
......
src/targets/gpu/mlir.cpp
View file @ 881a4bd4
......@@ -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>
......@@ -370,7 +372,11 @@ struct mlir_program
mlir_operation_state& add_results(const std::vector<shape>& outputs)
{
auto x = prog->make_tensors(outputs);
std::vector<shape> reshaped(outputs.size());
std::transform(outputs.begin(), outputs.end(), reshaped.begin(), [](const shape& r) {
return shape{r.type(), r.lens()};
});
auto x = prog->make_tensors(reshaped);
mlirOperationStateAddResults(&op_state, x.size(), x.data());
return *this;
}
......@@ -502,11 +508,12 @@ 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 +578,7 @@ 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 +596,54 @@ 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 +715,19 @@ 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 @ 881a4bd4
......@@ -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 query = "select P.* \
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 @ a2d710e3
/*
* 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 @ 881a4bd4
......@@ -138,12 +138,12 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
dead_code_elimination{},
pack_int8_args{},
dead_code_elimination{},
adjust_allocation{gpu_allocation_model{}},
dead_code_elimination{},
fuse_ops{&ctx, options.fast_math},
dead_code_elimination{},
replace_allocate{gpu_allocation_model{}, options.offload_copy},
dead_code_elimination{},
adjust_allocation{gpu_allocation_model{}},
dead_code_elimination{},
compile_ops{&ctx},
dead_code_elimination{},
write_literals{&ctx},
......
src/targets/ref/lowering.cpp
View file @ 881a4bd4
......@@ -31,9 +31,7 @@
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/op/dot.hpp>
#include <migraphx/op/quant_dot.hpp>
#include <migraphx/op/elu.hpp>
#include <migraphx/op/im2col.hpp>
#include <migraphx/op/leaky_relu.hpp>
#include <migraphx/op/logsoftmax.hpp>
#include <migraphx/op/loop.hpp>
#include <migraphx/op/lrn.hpp>
......@@ -431,65 +429,6 @@ struct ref_quant_gemm
};
MIGRAPHX_REGISTER_OP(ref_gemm)
struct leaky_relu_op
{
op::leaky_relu op;
std::string name() const { return "ref::leaky_relu"; }
auto fcn() const
{
auto a = op.alpha;
return [a](auto x) { return x > 0 ? x : x * a; };
}
};
struct elu_op
{
op::elu op;
std::string name() const { return "ref::elu"; }
auto fcn() const
{
auto a = op.alpha;
return [a](auto x) { return x > 0 ? x : a * std::expm1(x); };
}
};
template <typename Op>
struct ref_unary : auto_register_op<ref_unary<Op>>
{
ref_unary() = default;
template <class T>
ref_unary(T pop) : op(Op{std::move(pop)})
{
}
Op op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op.op, f);
}
std::string name() const { return op.name(); }
shape compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(1);
const auto& s = inputs.at(0);
return {s.type(), s.lens()};
}
argument compute(context&, const shape& output_shape, std::vector<argument> args) const
{
argument result{output_shape};
visit_all(result, args[0])([&](auto output, auto input) {
assert(input.get_shape().standard());
std::transform(input.begin(), input.end(), output.begin(), op.fcn());
});
return result;
}
};
template <class Op>
struct ref_softmax : auto_register_op<ref_softmax<Op>>
{
......@@ -630,9 +569,7 @@ struct ref_apply
apply_map["quant_dot"] = extend_op<ref_quant_gemm, op::quant_dot>();
apply_map["quant_convolution"] =
extend_op<ref_convolution<op::quant_convolution>, op::quant_convolution>();
apply_map["elu"] = extend_op<ref_unary<elu_op>, op::elu>();
apply_map["im2col"] = extend_op<ref_im2col, op::im2col>();
apply_map["leaky_relu"] = extend_op<ref_unary<leaky_relu_op>, op::leaky_relu>();
apply_map["logsoftmax"] = extend_op<ref_softmax<op::logsoftmax>, op::logsoftmax>();
apply_map["lrn"] = extend_op<ref_lrn, op::lrn>();
apply_map["pad"] = extend_op<ref_pad, op::pad>();
......
test/api/test_custom_op_gpu.cpp
View file @ 881a4bd4
......@@ -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/literal.cpp
View file @ 881a4bd4
......@@ -48,4 +48,4 @@ void gpu_literal_test()
}
}
int main() { gpu_literal_test(); }
int main() { gpu_literal_test(); } // NOLINT (bugprone-exception-escape)
test/gpu/mlir.cpp
View file @ 881a4bd4
......@@ -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/gpu/quantization.cpp
View file @ 881a4bd4
......@@ -30,7 +30,6 @@
#include <migraphx/ref/target.hpp>
#include <migraphx/gpu/target.hpp>
#include <migraphx/verify.hpp>
#include <migraphx/quantization.hpp>
#include <migraphx/dead_code_elimination.hpp>
#include <migraphx/propagate_constant.hpp>
#include <migraphx/pass_manager.hpp>
......
test/onnx/gen_onnx.py
View file @ 881a4bd4
......@@ -4590,6 +4590,16 @@ def neg_test():
return ([node], [x], [y])
@onnx_test
def neg_dynamic_test():
x = helper.make_tensor_value_info('0', TensorProto.INT64, [None, 3])
y = helper.make_tensor_value_info('1', TensorProto.INT64, [None, 3])
node = onnx.helper.make_node('Neg', inputs=['0'], outputs=['1'])
return ([node], [x], [y])
@onnx_test
def nms_test():
b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [1, 6, 4])
......@@ -6223,6 +6233,20 @@ def sinh_test():
return ([node], [x], [y])
@onnx_test
def sinh_dynamic_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [None])
y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [None])
node = onnx.helper.make_node(
'Sinh',
inputs=['x'],
outputs=['y'],
)
return ([node], [x], [y])
@onnx_test
def size_float_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [2, 3, 4])
......
test/onnx/onnx_test.cpp
View file @ 881a4bd4
......@@ -42,7 +42,6 @@
#include <migraphx/op/lrn.hpp>
#include <migraphx/op/reshape.hpp>
#include <migraphx/op/unknown.hpp>
#include <random>
#include <migraphx/serialize.hpp>
......@@ -3944,6 +3943,21 @@ TEST_CASE(neg_test)
EXPECT(p == prog);
}
TEST_CASE(neg_dynamic_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::int64_type, {{1, 10, 0}, {3, 3, 0}}};
auto input = mm->add_parameter("0", s);
auto ret = mm->add_instruction(migraphx::make_op("neg"), input);
mm->add_return({ret});
migraphx::onnx_options options;
options.default_dyn_dim_value = {1, 10, 0};
auto prog = migraphx::parse_onnx("neg_dynamic_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(nms_test)
{
migraphx::program p;
......@@ -5667,6 +5681,24 @@ TEST_CASE(sinh_test)
EXPECT(p == prog);
}
TEST_CASE(sinh_dynamic_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape::dynamic_dimension dd{1, 10, 0};
std::vector<migraphx::shape::dynamic_dimension> dyn_dims;
dyn_dims.push_back(dd);
auto input = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, dyn_dims});
auto ret = mm->add_instruction(migraphx::make_op("sinh"), input);
mm->add_return({ret});
migraphx::onnx_options options;
options.default_dyn_dim_value = dd;
auto prog = parse_onnx("sinh_dynamic_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(size_float_test)
{
migraphx::program p;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment