Skip to content

N
ngraph
Commit 4fd8228d authored by Sergey Shalnov's avatar Sergey Shalnov Committed by Scott Cyphers
Browse files
Options

IntelGPU backend: Reshape operation 5D support implemented (#2433)

parent b78321c6
Hide whitespace changes
Inline Side-by-side
Showing with 168 additions and 73 deletions
src/ngraph/runtime/intelgpu/intelgpu_backend.cpp
View file @ 4fd8228d
......@@ -931,35 +931,51 @@ shared_ptr<runtime::Executable>
arguments_check(op, 1, 1);
const shared_ptr<op::Reshape> op_reshape = static_pointer_cast<op::Reshape>(op);
const AxisVector& reshape_axes = op_reshape->get_input_order();
if (op_reshape->get_is_transpose())
if ((get_input_type(op) != element::f32) || (get_input_shape(op).size() > 4) ||
(get_output_shape(op).size() > 4))
{
vector<uint16_t> permute_order({0, 1, 2, 3}); // No action by default
const AxisVector& reshape_axes = op_reshape->get_input_order();
const size_t max_dim = 4;
const size_t scale =
reshape_axes.size() < max_dim ? max_dim - reshape_axes.size() : 0;
// Need to scale indexes up according on array rank.
// For example, in 2D array, indexes are 0,1 but in 4D array it should be 2,3
// because cldnn::tensor is always 4D assuming cldnn::bfyx model
size_t rindex = max_dim;
for (auto i = reshape_axes.crbegin(); i != reshape_axes.crend() && rindex > 0;
++i, --rindex)
{
permute_order.at(rindex - 1) = *i + scale;
}
const cldnn::permute cldnn_permute(
get_output_name(op), get_input_name(op), permute_order);
topology.add(cldnn_permute);
do_reshape_operation(topology,
get_input_name(op),
get_input_shape(op),
get_input_type(op),
get_output_name(op),
get_output_shape(op),
get_output_type(op),
reshape_axes);
}
else
{
const cldnn::tensor new_shape =
intelgpu_space::create_cldnn_tensor(get_output_shape(op));
const cldnn::reshape reshape_op(get_output_name(op), get_input_name(op), new_shape);
topology.add(reshape_op);
if (op_reshape->get_is_transpose())
{
vector<uint16_t> permute_order({0, 1, 2, 3}); // No action by default
const size_t max_dim = 4;
const size_t scale =
reshape_axes.size() < max_dim ? max_dim - reshape_axes.size() : 0;
// Need to scale indexes up according on array rank.
// For example, in 2D array, indexes are 0,1 but in 4D array it should be 2,3
// because cldnn::tensor is always 4D assuming cldnn::bfyx model
size_t rindex = max_dim;
for (auto i = reshape_axes.crbegin(); i != reshape_axes.crend() && rindex > 0;
++i, --rindex)
{
permute_order.at(rindex - 1) = *i + scale;
}
const cldnn::permute cldnn_permute(
get_output_name(op), get_input_name(op), permute_order);
topology.add(cldnn_permute);
}
else
{
const cldnn::tensor new_shape =
intelgpu_space::create_cldnn_tensor(get_output_shape(op));
const cldnn::reshape reshape_op(
get_output_name(op), get_input_name(op), new_shape);
topology.add(reshape_op);
}
}
break;
}
......
src/ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.cpp
View file @ 4fd8228d
......@@ -1453,3 +1453,63 @@ void runtime::intelgpu::do_negative_operation(cldnn::topology& topology,
gws);
topology.add(op_negative);
}
void runtime::intelgpu::do_reshape_operation(cldnn::topology& topology,
const string& input_name,
const Shape& input_shape,
const element::Type& input_type,
const string& output_name,
const Shape& output_shape,
const element::Type& output_type,
const AxisVector& reshape_axes)
{
const cldnn::layout layout = IntelGPULayout::create_cldnn_layout(output_type, output_shape);
const string entry_point_name = "reshape_" + output_name;
const string& input_type_name = get_opencl_type_name(input_type);
const string& output_type_name = get_opencl_type_name(output_type);
const size_t dst_shape_size = shape_size(output_shape);
codegen::CodeWriter writer;
gen_func_def(writer,
entry_point_name,
{input_type_name},
{input_shape},
output_type_name,
{dst_shape_size});
writer.block_begin();
{
writer << "// input: " << input_shape << "\n";
writer << "//output: " << output_shape << "\n";
writer << "//axes: " << reshape_axes << "\n\n";
writer << "uint output_it = 0;\n";
// Main operation loop
for (auto const i : reshape_axes)
{
writer << "for (uint i" << i << " = 0; i" << i << " < " << input_shape.at(i) << "; ++i"
<< i << ")\n";
writer.block_begin();
}
writer << "output[output_it] = input0" << access_dims(input_shape) << ";\n"
<< "++output_it;\n";
// Closing brackets for loop
for (auto const i : reshape_axes)
{
writer.block_end();
}
}
writer.block_end();
const cldnn::custom_gpu_primitive op_reshape(output_name,
{input_name},
{writer.get_code()},
entry_point_name,
get_kernel_args(1, 1),
"",
layout,
{1});
topology.add(op_reshape);
}
src/ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.hpp
View file @ 4fd8228d
......@@ -21,6 +21,7 @@
#include "ngraph/runtime/intelgpu/code_writer.hpp"
#include "ngraph/axis_set.hpp"
#include "ngraph/axis_vector.hpp"
#include "ngraph/coordinate.hpp"
#include "ngraph/shape.hpp"
#include "ngraph/strides.hpp"
......@@ -195,6 +196,15 @@ namespace ngraph
const Shape& output_shape,
const element::Type& output_type);
void do_reshape_operation(cldnn::topology& topology,
const std::string& input_name,
const Shape& input_shape,
const element::Type& input_type,
const std::string& output_name,
const Shape& output_shape,
const element::Type& output_type,
const AxisVector& reshape_axes);
// Helper functions used in cldnn::custom_gpu_primitive kernels
std::string get_opencl_type_name(const element::Type& ngraph_type);
std::string get_opencl_type_min_max_value(const element::Type& ngraph_type,
......
src/ngraph/runtime/intelgpu/unit_test.manifest
View file @ 4fd8228d
......@@ -61,7 +61,6 @@ replace_slice_matrix
replace_slice_matrix_inplace
replace_slice_scalar
replace_slice_vector
reshape_6d
reverse_sequence_n2c3h4w2
reverse_sequence_n4c3h2w2
reverse_sequence_n4d2c3h2w2
......
test/backend_reshape.in.cpp
View file @ 4fd8228d
......@@ -256,18 +256,15 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_m2m_dim_change_transpose)
NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_021)
{
vector<float> a_data(2 * 3 * 4);
for (int i = 0; i < 2 * 3 * 4; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 3, 4};
Shape shape_r{2, 4, 3};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto r = make_shared<op::Reshape>(A, AxisVector{0, 2, 1}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto backend = runtime::Backend::create("${BACKEND_NAME}");
// Create some tensors for input/output
......@@ -284,18 +281,15 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_021)
NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_210)
{
vector<float> a_data(2 * 3 * 4);
for (int i = 0; i < 2 * 3 * 4; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 3, 4};
Shape shape_r{4, 3, 2};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto r = make_shared<op::Reshape>(A, AxisVector{2, 1, 0}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto backend = runtime::Backend::create("${BACKEND_NAME}");
// Create some tensors for input/output
......@@ -312,18 +306,15 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_210)
NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_201)
{
vector<float> a_data(2 * 3 * 4);
for (int i = 0; i < 2 * 3 * 4; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 3, 4};
Shape shape_r{4, 2, 3};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto r = make_shared<op::Reshape>(A, AxisVector{2, 0, 1}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto backend = runtime::Backend::create("${BACKEND_NAME}");
// Create some tensors for input/output
......@@ -340,18 +331,15 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_201)
NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_102)
{
vector<float> a_data(2 * 3 * 4);
for (int i = 0; i < 2 * 3 * 4; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 3, 4};
Shape shape_r{3, 2, 4};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto r = make_shared<op::Reshape>(A, AxisVector{1, 0, 2}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto backend = runtime::Backend::create("${BACKEND_NAME}");
// Create some tensors for input/output
......@@ -368,18 +356,15 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_102)
NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_120)
{
vector<float> a_data(2 * 3 * 4);
for (int i = 0; i < 2 * 3 * 4; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 3, 4};
Shape shape_r{3, 4, 2};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto r = make_shared<op::Reshape>(A, AxisVector{1, 2, 0}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto backend = runtime::Backend::create("${BACKEND_NAME}");
// Create some tensors for input/output
......@@ -396,18 +381,15 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_3d_transpose_120)
NGRAPH_TEST(${BACKEND_NAME}, reshape_4d_transpose)
{
vector<float> a_data(2 * 2 * 5 * 5);
for (int i = 0; i < 2 * 2 * 5 * 5; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 2, 5, 5};
Shape shape_r{2, 5, 5, 2};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto r = make_shared<op::Reshape>(A, AxisVector{0, 2, 3, 1}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto backend = runtime::Backend::create("${BACKEND_NAME}");
// Create some tensors for input/output
......@@ -430,18 +412,15 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_4d_transpose)
NGRAPH_TEST(${BACKEND_NAME}, reshape_4d_no_transpose)
{
vector<float> a_data(2 * 2 * 5 * 5);
for (int i = 0; i < 2 * 2 * 5 * 5; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 2, 5, 5};
Shape shape_r{2, 5, 5, 2};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto r = make_shared<op::Reshape>(A, AxisVector{0, 1, 2, 3}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto backend = runtime::Backend::create("${BACKEND_NAME}");
// Create some tensors for input/output
......@@ -517,16 +496,13 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_transposed_shape_change)
//
NGRAPH_TEST(${BACKEND_NAME}, reshape_6d)
{
vector<float> a_data(2 * 2 * 3 * 3 * 2 * 4);
for (int i = 0; i < 2 * 2 * 3 * 3 * 2 * 4; i++)
{
a_data[i] = float(i + 1);
}
Shape shape_a{2, 2, 3, 3, 2, 4};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
Shape shape_r{3, 2, 2, 4, 3, 2};
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto r = make_shared<op::Reshape>(A, AxisVector{2, 4, 0, 5, 3, 1}, shape_r);
auto f = make_shared<Function>(r, ParameterVector{A});
......@@ -565,3 +541,37 @@ NGRAPH_TEST(${BACKEND_NAME}, reshape_6d)
215., 287., 200., 272., 208., 280., 216., 288.}),
read_vector<float>(result));
}
#if NGRAPH_INTERPRETER_ENABLE
NGRAPH_TEST(${BACKEND_NAME}, reshape_shufflenet_5d)
{
Shape shape_a{1, 112, 56, 56};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
Shape shape_b{1, 4, 28, 56, 56};
auto B = make_shared<op::Parameter>(element::f32, shape_b);
Shape shape_c{1, 28, 4, 56, 56};
auto C = make_shared<op::Parameter>(element::f32, shape_c);
Shape shape_r{1, 112, 56, 56};
vector<float> a_data(shape_size(shape_a));
iota(a_data.begin(), a_data.end(), 1);
auto r0 = make_shared<op::Reshape>(A, AxisVector{0, 1, 2, 3}, shape_b);
auto r1 = make_shared<op::Reshape>(r0, AxisVector{0, 2, 1, 3, 4}, shape_c);
auto r2 = make_shared<op::Reshape>(r1, AxisVector{0, 1, 2, 3, 4}, shape_r);
auto f = make_shared<Function>(r2, ParameterVector{A});
auto ref_func = clone_function(*f);
auto bk_func = clone_function(*f);
vector<vector<float>> args;
args.push_back(a_data);
auto ref_results = execute(ref_func, args, "INTERPRETER");
auto bk_results = execute(bk_func, args, "${BACKEND_NAME}");
EXPECT_TRUE(test::all_close_f(ref_results.at(0), bk_results.at(0)));
}
#endif //NGRAPH_INTERPRETER_ENABLE
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