Skip to content

GitLab

Commit e7f7ea10 authored by Alan Turner's avatar Alan Turner
Browse files

Merge remote-tracking branch 'origin/optimize' into ck-gsg

parents 9c6ba1ed ae13eb93
Hide whitespace changes
Inline Side-by-side
Showing with 633 additions and 236 deletions
src/targets/ref/lowering.cpp
View file @ e7f7ea10
......@@ -26,15 +26,12 @@
#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>
#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>
......@@ -75,84 +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;
......@@ -237,15 +156,16 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
argument compute(context&, shape output_shape, std::vector<argument> args) const
{
std::vector<std::size_t> padding;
if(op.use_dynamic_same_auto_pad)
if(op.padding_mode != op::padding_mode_t::default_)
{
auto input_lens = args[0].get_shape().lens();
std::vector<std::size_t> img_lens{input_lens.begin() + 2, input_lens.end()};
auto input_lens = args[0].get_shape().lens();
auto weights_lens = args[1].get_shape().lens();
std::vector<std::size_t> k_lens{weights_lens.begin() + 2, weights_lens.end()};
padding = calc_dyn_auto_pad(img_lens, k_lens, op.stride, op.dilation);
output_shape =
compute_padded_shape({args.at(0).get_shape(), args.at(1).get_shape()}, padding);
padding =
op.padding_mode == op::same_upper
? calc_dyn_auto_pad(input_lens, weights_lens, op.stride, op.dilation, true)
: calc_dyn_auto_pad(input_lens, weights_lens, op.stride, op.dilation, false);
output_shape = compute_padded_shape(
args[0].get_shape(), args[1].get_shape(), padding, op.stride, op.dilation);
}
else
{
......@@ -313,34 +233,6 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
});
return result;
}
private:
/*!
* Used for dynamic auto padding since padding needs to be computed at evaulation time.
* \param inputs two fixed shape inputs [input_tensor, weights]
* \param padding from auto_pad calculation
*/
shape compute_padded_shape(const std::vector<shape>& inputs,
const std::vector<std::size_t>& padding) const
{
const shape& input = inputs.at(0);
const shape& weights = inputs.at(1);
const size_t num_spatial_dims = input.lens().size() - 2;
std::vector<size_t> output_lens{input.lens()[0], weights.lens()[0]};
// calculate the output shape of the convolution: ((W - K + 2P) / S) + 1
for(size_t i = 0; i < num_spatial_dims; i++)
{
auto padding_factor = padding[i] + padding[i + num_spatial_dims];
output_lens.push_back(std::size_t(std::max<std::ptrdiff_t>(
1,
(input.lens()[i + 2] - (1 + op.dilation[i] * (weights.lens()[i + 2] - 1)) +
padding_factor) /
op.stride[i] +
1)));
}
return inputs[0].with_lens(output_lens);
}
};
struct ref_im2col
......@@ -537,65 +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>>
{
......@@ -731,16 +564,12 @@ 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>();
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>();
......
src/tf/parse_batchnorm.cpp
View file @ e7f7ea10
......@@ -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);
}
};
......
src/tf/parse_conv.cpp
View file @ e7f7ea10
......@@ -75,7 +75,6 @@ struct parse_conv : op_parser<parse_conv>
const std::string& pad_mode = info.attributes.at("padding").s();
if(pad_mode.find("SAME") != std::string::npos)
{
op.padding_mode = op::padding_mode_t::same;
std::vector<size_t> weight_dims = weights->get_shape().lens();
size_t weight_h = weight_dims[2];
size_t weight_w = weight_dims[3];
......@@ -87,10 +86,6 @@ struct parse_conv : op_parser<parse_conv>
op.padding = std::vector<size_t>(pads.begin(), pads.end());
}
else if(pad_mode.find("VALID") != std::string::npos)
{
op.padding_mode = op::padding_mode_t::valid;
}
else if(pad_mode.find("EXPLICIT") != std::string::npos)
{
std::vector<size_t> padding;
......
src/tf/parse_depthwiseconv.cpp
View file @ e7f7ea10
......@@ -80,7 +80,6 @@ struct parse_depthwiseconv : op_parser<parse_depthwiseconv>
if(pad_mode.find("SAME") != std::string::npos)
{
op.padding_mode = op::padding_mode_t::same;
std::vector<size_t> weight_dims = weights->get_shape().lens();
size_t weight_h = weight_dims[2];
size_t weight_w = weight_dims[3];
......@@ -101,10 +100,6 @@ struct parse_depthwiseconv : op_parser<parse_depthwiseconv>
op.padding[1] = pads[1];
}
}
else if(pad_mode.find("VALID") != std::string::npos)
{
op.padding_mode = op::padding_mode_t::valid;
}
}
std::vector<int64_t> new_weights_shape;
......
test/api/test_custom_op_gpu.cpp
View file @ e7f7ea10
......@@ -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/fuse_pointwise.cpp
View file @ e7f7ea10
......@@ -272,6 +272,35 @@ TEST_CASE(contiguous_input)
EXPECT(p1 == p2);
}
TEST_CASE(contiguous_boolean_input)
{
migraphx::shape s{migraphx::shape::bool_type, {2, 3}};
migraphx::shape s_lit{migraphx::shape::bool_type, {1}, {0}};
migraphx::program p1;
{
auto* mm = p1.get_main_module();
auto x = mm->add_parameter("x", s);
auto one = mm->add_literal(migraphx::literal(s_lit, {1.0}));
auto yb =
mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", s.lens()}}), one);
auto y = mm->add_instruction(migraphx::make_op("contiguous"), yb);
auto xor1 = mm->add_instruction(migraphx::make_op("logical_xor"), x, y);
mm->add_return({xor1});
}
run_pass(p1);
migraphx::program p2;
{
auto* mm = p2.get_main_module();
auto x = mm->add_parameter("x", s);
auto xor1 = add_pointwise(p2, "main:pointwise0", {x}, [=](auto* pm, const auto& inputs) {
auto y = pm->add_literal(migraphx::literal(s_lit, {1}));
return pm->add_instruction(migraphx::make_op("logical_xor"), inputs[0], y);
});
mm->add_return({xor1});
}
}
TEST_CASE(all_scalar_input)
{
migraphx::shape s{migraphx::shape::float_type};
......
test/gpu/literal.cpp
View file @ e7f7ea10
......@@ -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 @ e7f7ea10
......@@ -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;
}
......@@ -141,7 +141,7 @@ TEST_CASE(conv)
const std::string mlir_output = R"__migraphx__(
module {
func.func @main(%arg0: tensor<2x8x3x3xf32>, %arg1: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {kernel = "mixr"} {
%0 = migraphx.convolution(%arg1, %arg0) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1], use_dynamic_same_auto_pad = 0 : i64} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
%0 = migraphx.convolution(%arg1, %arg0) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
return %0 : tensor<1x2x2x2xf32>
}
}
......@@ -164,7 +164,7 @@ TEST_CASE(conv_add_relu)
const std::string mlir_output = R"__migraphx__(
module {
func.func @main(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x8x3x3xf32>, %arg2: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {kernel = "mixr"} {
%0 = migraphx.convolution(%arg2, %arg1) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1], use_dynamic_same_auto_pad = 0 : i64} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
%0 = migraphx.convolution(%arg2, %arg1) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
%1 = migraphx.add(%0, %arg0) : (tensor<1x2x2x2xf32>, tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
%2 = migraphx.relu(%1) : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
return %2 : tensor<1x2x2x2xf32>
......
test/gpu/quantization.cpp
View file @ e7f7ea10
......@@ -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/layout_nhwc.cpp 0 → 100644
View file @ e7f7ea10
/*
* 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/layout_nhwc.hpp>
#include <migraphx/dead_code_elimination.hpp>
#include <migraphx/pass_manager.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/generate.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/instruction.hpp>
#include <basic_ops.hpp>
#include <migraphx/make_op.hpp>
#include <test.hpp>
void run_pass(migraphx::module& m)
{
migraphx::run_passes(m, {migraphx::layout_nhwc{}, migraphx::dead_code_elimination{}});
}
migraphx::operation layout(std::vector<int64_t> permutation = {0, 1, 2, 3})
{
return migraphx::make_op("layout", {{"permutation", permutation}});
}
migraphx::instruction_ref add_layout_nhwc(migraphx::module& m, migraphx::instruction_ref ins)
{
return m.add_instruction(layout({0, 2, 3, 1}), ins);
}
TEST_CASE(conv_relu)
{
migraphx::module m1;
{
auto x = m1.add_parameter("x", {migraphx::shape::float_type, {1, 8, 16, 16}});
auto w = m1.add_literal(
migraphx::generate_literal({migraphx::shape::float_type, {16, 8, 3, 3}}));
auto conv = m1.add_instruction(
migraphx::make_op("convolution",
{{"padding", {1, 1}}, {"stride", {2, 2}}, {"dilation", {1, 1}}}),
x,
w);
m1.add_instruction(migraphx::make_op("relu"), conv);
}
run_pass(m1);
migraphx::module m2;
{
auto x = add_layout_nhwc(
m2, m2.add_parameter("x", {migraphx::shape::float_type, {1, 8, 16, 16}}));
auto w = add_layout_nhwc(m2,
m2.add_literal(migraphx::generate_literal(
{migraphx::shape::float_type, {16, 8, 3, 3}})));
auto conv = m2.add_instruction(
migraphx::make_op("convolution",
{{"padding", {1, 1}}, {"stride", {2, 2}}, {"dilation", {1, 1}}}),
x,
w);
auto conv_layout = m2.add_instruction(layout(), conv);
m2.add_instruction(migraphx::make_op("relu"), conv_layout);
}
EXPECT(m1.sort() == m2.sort());
}
TEST_CASE(conv_add)
{
migraphx::module m1;
{
auto x = m1.add_parameter("x", {migraphx::shape::float_type, {1, 8, 16, 16}});
auto w = m1.add_literal(
migraphx::generate_literal({migraphx::shape::float_type, {16, 8, 3, 3}}));
auto y = m1.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {16}}));
auto conv = m1.add_instruction(
migraphx::make_op("convolution",
{{"padding", {1, 1}}, {"stride", {2, 2}}, {"dilation", {1, 1}}}),
x,
w);
auto b = m1.add_instruction(
migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", conv->get_shape().lens()}}),
y);
m1.add_instruction(migraphx::make_op("add"), conv, b);
}
run_pass(m1);
migraphx::module m2;
{
auto x = add_layout_nhwc(
m2, m2.add_parameter("x", {migraphx::shape::float_type, {1, 8, 16, 16}}));
auto w = add_layout_nhwc(m2,
m2.add_literal(migraphx::generate_literal(
{migraphx::shape::float_type, {16, 8, 3, 3}})));
auto y = m2.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {16}}));
auto conv = m2.add_instruction(
migraphx::make_op("convolution",
{{"padding", {1, 1}}, {"stride", {2, 2}}, {"dilation", {1, 1}}}),
x,
w);
auto conv_layout = m2.add_instruction(layout(), conv);
auto b = m2.add_instruction(
migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", conv->get_shape().lens()}}),
y);
m2.add_instruction(migraphx::make_op("add"), conv_layout, b);
}
EXPECT(m1.sort() == m2.sort());
}
int main(int argc, const char* argv[]) { test::run(argc, argv); }
test/onnx/batch_norm_invalid_rank_test.onnx → test/onnx/batch_norm_rank_2_test.onnx
View file @ e7f7ea10
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 @ e7f7ea10
......@@ -332,6 +332,24 @@ 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])
......@@ -386,23 +404,6 @@ 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])
......@@ -420,6 +421,74 @@ def batch_norm_invalid_bias_rank_test():
return ([node], [x, scale, bias, mean, var], [out])
@onnx_test
def binary_dyn_brcst_prelu_test():
arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT,
[None, 3, 4, 5])
arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [4, 5])
arg_out = helper.make_tensor_value_info('out', TensorProto.FLOAT,
[None, 3, 4, 5])
node = onnx.helper.make_node(
'PRelu',
inputs=['0', '1'],
outputs=['out'],
)
return ([node], [arg0, arg1], [arg_out])
@onnx_test
def binary_dyn_brcst_add_test():
arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT16, [4, 5])
arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT,
[None, 3, 4, 5])
arg_out = helper.make_tensor_value_info('out', TensorProto.FLOAT,
[None, 3, 4, 5])
node = onnx.helper.make_node(
'Add',
inputs=['0', '1'],
outputs=['out'],
)
return ([node], [arg0, arg1], [arg_out])
@onnx_test
def binary_dyn_brcst_attr_error_test():
arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT16, [4, 5])
arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT,
[None, 3, 4, 5])
arg_out = helper.make_tensor_value_info('out', TensorProto.FLOAT,
[None, 3, 4, 5])
node = onnx.helper.make_node('Add',
inputs=['0', '1'],
outputs=['out'],
broadcast=1,
axis=1)
return ([node], [arg0, arg1], [arg_out])
@onnx_test
def binary_dyn_brcst_mul_test():
arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT,
[None, 3, 4, 5])
arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [4, 1])
arg_out = helper.make_tensor_value_info('out', TensorProto.FLOAT,
[None, 3, 4, 5])
node = onnx.helper.make_node(
'Mul',
inputs=['0', '1'],
outputs=['out'],
)
return ([node], [arg0, arg1], [arg_out])
@onnx_test
def cast_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT16, [10])
......@@ -3685,6 +3754,16 @@ 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])
......@@ -5318,6 +5397,20 @@ 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])
......@@ -5700,6 +5793,92 @@ def split_test_default():
return ([node], [x], [y1, y2])
@onnx_test
def split_test_no_attribute():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [300, 15])
y1 = helper.make_tensor_value_info('y1', TensorProto.FLOAT, [75, 15])
y2 = helper.make_tensor_value_info('y2', TensorProto.FLOAT, [75, 15])
y3 = helper.make_tensor_value_info('y3', TensorProto.FLOAT, [75, 15])
y4 = helper.make_tensor_value_info('y4', TensorProto.FLOAT, [75, 15])
split = np.ones(4) * 75
split_tensor = helper.make_tensor(name="split",
data_type=TensorProto.INT64,
dims=split.shape,
vals=split.astype(np.int64))
const_node = helper.make_node("Constant",
inputs=[],
outputs=['split'],
value=split_tensor)
node = onnx.helper.make_node(
'Split',
inputs=['x', 'split'],
outputs=['y1', 'y2', 'y3', 'y4'],
)
return ([const_node, node], [x], [y1, y2, y3, y4])
@onnx_test
def split_test_no_attribute_invalid_split():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [300, 15])
y1 = helper.make_tensor_value_info('y1', TensorProto.FLOAT, [75, 15])
y2 = helper.make_tensor_value_info('y2', TensorProto.FLOAT, [75, 15])
y3 = helper.make_tensor_value_info('y3', TensorProto.FLOAT, [75, 15])
y4 = helper.make_tensor_value_info('y4', TensorProto.FLOAT, [75, 15])
split = np.ones(4)
split_tensor = helper.make_tensor(name="split",
data_type=TensorProto.INT64,
dims=split.shape,
vals=split.astype(np.int64))
const_node = helper.make_node("Constant",
inputs=[],
outputs=['split'],
value=split_tensor)
node = onnx.helper.make_node(
'Split',
inputs=['x', 'split'],
outputs=['y1', 'y2', 'y3', 'y4'],
)
return ([const_node, node], [x], [y1, y2, y3, y4])
@onnx_test
def split_test_invalid_split():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10, 15])
y1 = helper.make_tensor_value_info('y1', TensorProto.FLOAT, [10, 7])
y2 = helper.make_tensor_value_info('y2', TensorProto.FLOAT, [10, 4])
y3 = helper.make_tensor_value_info('y3', TensorProto.FLOAT, [10, 4])
node = onnx.helper.make_node('Split',
inputs=['x'],
outputs=['y1', 'y2', 'y3'],
axis=1,
split=[1, 1, 1])
return ([node], [x], [y1, y2, y3])
@onnx_test
def split_test_no_attribute_invalid_input_split():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10, 15])
y1 = helper.make_tensor_value_info('y1', TensorProto.FLOAT, [10, 7])
y2 = helper.make_tensor_value_info('y2', TensorProto.FLOAT, [10, 4])
y3 = helper.make_tensor_value_info('y3', TensorProto.FLOAT, [10, 4])
node = onnx.helper.make_node('Split',
inputs=['x'],
outputs=['y1', 'y2', 'y3'],
axis=1,
split=[])
return ([node], [x], [y1, y2, y3])
@onnx_test
def sqrt_test():
x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10, 15])
......
test/onnx/onnx_test.cpp
View file @ e7f7ea10
......@@ -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>
......@@ -394,6 +393,31 @@ TEST_CASE(batch_norm_flat_test)
EXPECT(p == prog);
}
TEST_CASE(batch_norm_rank_2_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
auto x = mm->add_parameter("x", {migraphx::shape::float_type, {2, 5}});
auto scale = mm->add_parameter("scale", {migraphx::shape::float_type, {5}});
auto bias = mm->add_parameter("bias", {migraphx::shape::float_type, {5}});
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 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});
add_common_op(*mm, migraphx::make_op("add"), {r0, bias});
auto prog = optimize_onnx("batch_norm_rank_2_test.onnx");
EXPECT(p == prog);
}
TEST_CASE(batch_norm_1d_test)
{
migraphx::program p;
......@@ -497,6 +521,76 @@ TEST_CASE(batch_norm_invalid_bias_rank)
EXPECT(test::throws([&] { migraphx::parse_onnx("batch_norm_invalid_bias_rank.onnx"); }));
}
TEST_CASE(binary_dyn_brcst_prelu_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
auto l0 = mm->add_parameter(
"0",
migraphx::shape{migraphx::shape::float_type, {{1, 4, 0}, {3, 3, 0}, {4, 4, 0}, {5, 5, 0}}});
auto l1 = mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {4, 5}});
auto ret = add_common_op(*mm, migraphx::make_op("prelu"), {l0, l1});
mm->add_return({ret});
migraphx::onnx_options options;
options.default_dyn_dim_value = {1, 4, 0};
auto prog = migraphx::parse_onnx("binary_dyn_brcst_prelu_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(binary_dyn_brcst_add_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
auto l0 = mm->add_parameter("0", migraphx::shape{migraphx::shape::half_type, {4, 5}});
auto l1 = mm->add_parameter(
"1",
migraphx::shape{migraphx::shape::float_type, {{1, 4, 0}, {3, 3, 0}, {4, 4, 0}, {5, 5, 0}}});
auto ret = add_common_op(*mm, migraphx::make_op("add"), {l0, l1});
mm->add_return({ret});
migraphx::onnx_options options;
options.default_dyn_dim_value = {1, 4, 0};
auto prog = migraphx::parse_onnx("binary_dyn_brcst_add_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(binary_dyn_brcst_attr_error_test)
{
migraphx::onnx_options options;
options.default_dyn_dim_value = {1, 4, 0};
EXPECT(test::throws(
[&] { migraphx::parse_onnx("binary_dyn_brcst_attr_error_test.onnx", options); }));
}
TEST_CASE(binary_dyn_brcst_mul_test)
{
migraphx::program p;
auto* mm = p.get_main_module();
auto l0 = mm->add_parameter(
"0",
migraphx::shape{migraphx::shape::float_type, {{1, 4, 0}, {3, 3, 0}, {4, 4, 0}, {5, 5, 0}}});
auto l1 = mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {4, 1}});
auto bl1 = mm->add_instruction(
migraphx::make_op("multibroadcast",
{{"out_dyn_dims", to_value(l0->get_shape().dyn_dims())}}),
l1,
l0);
auto ret = mm->add_instruction(migraphx::make_op("mul"), l0, bl1);
mm->add_return({ret});
migraphx::onnx_options options;
options.default_dyn_dim_value = {1, 4, 0};
auto prog = migraphx::parse_onnx("binary_dyn_brcst_mul_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(cast_test)
{
migraphx::program p;
......@@ -856,8 +950,7 @@ TEST_CASE(conv_autopad_same_test)
auto l0 = mm->add_parameter("0", {migraphx::shape::float_type, {1, 3, 32, 32}});
auto l1 = mm->add_parameter("1", {migraphx::shape::float_type, {1, 3, 3, 3}});
migraphx::op::convolution op;
op.padding = {1, 1, 1, 1};
op.padding_mode = migraphx::op::padding_mode_t::same;
op.padding = {1, 1, 1, 1};
mm->add_instruction(op, l0, l1);
auto prog = optimize_onnx("conv_autopad_same_test.onnx");
......@@ -1034,15 +1127,11 @@ TEST_CASE(conv_dynamic_batch_same_upper)
auto l0 = mm->add_parameter(
"0", {migraphx::shape::float_type, {{1, 10, 0}, {3, 3, 0}, {5, 5, 0}, {5, 5, 0}}});
auto l1 = mm->add_parameter("1", {migraphx::shape::float_type, {1, 3, 3, 3}});
auto c0 =
mm->add_instruction(migraphx::make_op("convolution",
{{"padding", {1, 1, 1, 1}},
{"stride", {1, 1}},
{"dilation", {1, 1}},
{"padding_mode", migraphx::op::padding_mode_t::same},
{"use_dynamic_same_auto_pad", false}}),
l0,
l1);
auto c0 = mm->add_instruction(
migraphx::make_op("convolution",
{{"padding", {1, 1, 1, 1}}, {"stride", {1, 1}}, {"dilation", {1, 1}}}),
l0,
l1);
mm->add_return({c0});
migraphx::onnx_options options;
......@@ -1064,8 +1153,7 @@ TEST_CASE(conv_dynamic_img_same_upper)
{{"padding", {0, 0}},
{"stride", {1, 1}},
{"dilation", {1, 1}},
{"padding_mode", migraphx::op::padding_mode_t::same_upper},
{"use_dynamic_same_auto_pad", true}}),
{"padding_mode", migraphx::op::padding_mode_t::same_upper}}),
l0,
l1);
mm->add_return({c0});
......@@ -1089,8 +1177,7 @@ TEST_CASE(conv_dynamic_kernel_same_lower)
{{"padding", {0, 0}},
{"stride", {1, 1}},
{"dilation", {1, 1}},
{"padding_mode", migraphx::op::padding_mode_t::same_lower},
{"use_dynamic_same_auto_pad", true}}),
{"padding_mode", migraphx::op::padding_mode_t::same_lower}}),
l0,
l1);
mm->add_return({c0});
......@@ -3483,6 +3570,21 @@ 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;
......@@ -5206,6 +5308,24 @@ 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;
......@@ -5487,6 +5607,31 @@ TEST_CASE(split_test)
EXPECT(p == prog);
}
TEST_CASE(split_test_no_attribute)
{
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape si{migraphx::shape::int64_type, {4}, {1}};
std::vector<int> ind = {75, 75, 75, 75};
auto input = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {300, 15}});
mm->add_literal(migraphx::literal(si, ind));
auto r1 = mm->add_instruction(
migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {75}}}), input);
auto r2 = mm->add_instruction(
migraphx::make_op("slice", {{"axes", {0}}, {"starts", {75}}, {"ends", {150}}}), input);
auto r3 = mm->add_instruction(
migraphx::make_op("slice", {{"axes", {0}}, {"starts", {150}}, {"ends", {225}}}), input);
auto r4 = mm->add_instruction(
migraphx::make_op("slice", {{"axes", {0}}, {"starts", {225}}, {"ends", {300}}}), input);
mm->add_return({r1, r2, r3, r4});
auto prog = migraphx::parse_onnx("split_test_no_attribute.onnx");
EXPECT(p == prog);
}
TEST_CASE(split_test_default)
{
migraphx::program p;
......@@ -5502,6 +5647,23 @@ TEST_CASE(split_test_default)
EXPECT(p == prog);
}
TEST_CASE(split_test_no_attribute_invalid_split)
{
EXPECT(
test::throws([&] { migraphx::parse_onnx("split_test_no_attribute_invalid_split.onnx"); }));
}
TEST_CASE(split_test_invalid_split)
{
EXPECT(test::throws([&] { migraphx::parse_onnx("split_test_invalid_split.onnx"); }));
}
TEST_CASE(split_test_no_attribute_invalid_input_split)
{
EXPECT(test::throws(
[&] { migraphx::parse_onnx("split_test_no_attribute_invalid_input_split.onnx"); }));
}
TEST_CASE(sqrt_test)
{
migraphx::program p;
......
test/onnx/split_test_invalid_split.onnx 0 → 100644
View file @ e7f7ea10
split_test_invalid_split:
5
xy1y2y3"Split*
axis*
split@@@split_test_invalid_splitZ
x


b
y1


b
y2


b
y3


B
\ No newline at end of file
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