Skip to content

N
ngraph
Commit c75f7db3 authored by nishant.b.patel's avatar nishant.b.patel
Browse files
Options

Move quantized conv tests from test/builder to test/backend

parent baf1cb00
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 65 additions and 103 deletions
src/ngraph/runtime/interpreter/int_executable.hpp
View file @ c75f7db3
......@@ -51,6 +51,7 @@
#include "ngraph/op/passthrough.hpp"
#include "ngraph/op/product.hpp"
#include "ngraph/op/quantize.hpp"
#include "ngraph/op/quantized_convolution.hpp"
#include "ngraph/op/replace_slice.hpp"
#include "ngraph/op/reshape.hpp"
#include "ngraph/op/result.hpp"
......@@ -1134,12 +1135,75 @@ private:
break;
}
case OP_TYPEID::QuantizedConvolution:
{
const op::QuantizedConvolution* qc =
static_cast<const op::QuantizedConvolution*>(&node);
auto input_element_type = qc->get_input_element_type(0);
auto filter_element_type = qc->get_input_element_type(1);
auto output_element_type = qc->get_output_element_type(0);
if (input_element_type == element::u8 && filter_element_type == element::i8 &&
output_element_type == element::i8)
{
reference::convolution<uint8_t, int8_t, int8_t, int32_t>(
args[0]->get_data_ptr<const uint8_t>(),
args[1]->get_data_ptr<const int8_t>(),
out[0]->get_data_ptr<int8_t>(),
node.get_input_shape(0),
node.get_input_shape(1),
node.get_output_shape(0),
qc->get_window_movement_strides(),
qc->get_window_dilation_strides(),
qc->get_padding_below(),
qc->get_padding_above(),
qc->get_data_dilation_strides(),
args[2]->get_data_ptr<const float>(),
args[3]->get_data_ptr<const uint8_t>(),
args[4]->get_data_ptr<const float>(),
args[5]->get_data_ptr<const int8_t>(),
args[6]->get_data_ptr<const float>(),
args[7]->get_data_ptr<const int8_t>());
}
else if (input_element_type == element::u8 && filter_element_type == element::u8 &&
output_element_type == element::u8)
{
reference::convolution<uint8_t, uint8_t, uint8_t, int32_t>(
args[0]->get_data_ptr<const uint8_t>(),
args[1]->get_data_ptr<const uint8_t>(),
out[0]->get_data_ptr<uint8_t>(),
node.get_input_shape(0),
node.get_input_shape(1),
node.get_output_shape(0),
qc->get_window_movement_strides(),
qc->get_window_dilation_strides(),
qc->get_padding_below(),
qc->get_padding_above(),
qc->get_data_dilation_strides(),
args[2]->get_data_ptr<const float>(),
args[3]->get_data_ptr<const uint8_t>(),
args[4]->get_data_ptr<const float>(),
args[5]->get_data_ptr<const uint8_t>(),
args[6]->get_data_ptr<const float>(),
args[7]->get_data_ptr<const uint8_t>());
}
else
{
std::stringstream ss;
ss << "unsupported element type";
throw std::runtime_error(ss.str());
}
break;
}
case OP_TYPEID::QuantizedAvgPool:
case OP_TYPEID::QuantizedConvolutionBias:
case OP_TYPEID::QuantizedConvolutionBiasAdd:
case OP_TYPEID::QuantizedConvolutionBiasSignedAdd:
case OP_TYPEID::QuantizedConvolutionRelu:
case OP_TYPEID::QuantizedConvolution:
case OP_TYPEID::QuantizedMaxPool:
case OP_TYPEID::QuantizedDotBias:
case OP_TYPEID::QuantizedDot:
......
test/backend_test.in.cpp
View file @ c75f7db3
This diff is collapsed.
test/builder_quantization.cpp
View file @ c75f7db3
......@@ -146,108 +146,6 @@ static void constant_fold(std::shared_ptr<Function> f)
pass_manager.run_passes(f);
}
TEST(builder, scaled_QC)
{
Shape shape_a{1, 1, 3, 4}; // input shape
Shape shape_b{1, 1, 3, 3}; // filter shape
Shape shape_r{1, 1, 3, 4}; // output shape
vector<uint8_t> a_data = {1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4};
vector<int8_t> b_data = {1, 2, 3, 4, 5, 0, 0, 1, 2};
auto A = make_shared<op::Parameter>(element::u8, shape_a);
auto B = make_shared<op::Parameter>(element::i8, shape_b);
auto C = op::Constant::create(element::f32, Shape{1}, {0.0f});
auto D = op::Constant::create(element::f32, Shape{1}, {255.0f});
auto E = op::Constant::create(element::f32, Shape{1}, {-127.0f});
auto F = op::Constant::create(element::f32, Shape{1}, {127.0f});
auto G = op::Constant::create(element::f32, Shape{1}, {22.0f});
auto H = op::Constant::create(element::f32, Shape{1}, {90.0f});
auto CV = ngraph::builder::QuantizedConvolutionBuilder(A,
B,
Strides{1, 1}, // move_strides
Strides{1, 1}, // filter_dilation
CoordinateDiff{1, 1}, // below_pads
CoordinateDiff{1, 1}, // above_pads
Strides{1, 1}, // data_dilation
C,
D,
E,
F,
G,
H,
element::i8,
AxisSet{});
auto f = make_shared<Function>(NodeVector{CV}, ParameterVector{A, B});
constant_fold(f);
auto backend = runtime::Backend::create("CPU");
// Create some tensors for input/output
auto a = backend->create_tensor(element::u8, shape_a);
copy_data(a, a_data);
auto b = backend->create_tensor(element::i8, shape_b);
copy_data(b, b_data);
auto result = backend->create_tensor(element::i8, shape_r);
auto handle = backend->compile(f);
handle->call_with_validate({result}, {a, b});
EXPECT_EQ((vector<int8_t>{31, 48, 42, 45, 54, 102, 127, 61, 47, 74, 61, 55}),
read_vector<int8_t>(result));
}
TEST(builder, dynamic_scaled_QC)
{
Shape shape_a{1, 1, 3, 4}; // input shape
Shape shape_b{1, 1, 3, 3}; // filter shape
Shape shape_r{1, 1, 3, 4}; // output shape
vector<uint8_t> a_data = {1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4};
vector<int8_t> b_data = {1, 2, 3, 4, 5, 0, 0, 1, 2};
auto A = make_shared<op::Parameter>(element::u8, shape_a);
auto B = make_shared<op::Parameter>(element::i8, shape_b);
auto C = make_shared<op::Parameter>(element::f32, Shape{1});
auto D = make_shared<op::Parameter>(element::f32, Shape{1});
auto E = make_shared<op::Parameter>(element::f32, Shape{1});
auto F = make_shared<op::Parameter>(element::f32, Shape{1});
auto G = make_shared<op::Parameter>(element::f32, Shape{1});
auto H = make_shared<op::Parameter>(element::f32, Shape{1});
auto CV = ngraph::builder::QuantizedConvolutionBuilder(A,
B,
Strides{1, 1}, // move_strides
Strides{1, 1}, // filter_dilation
CoordinateDiff{1, 1}, // below_pads
CoordinateDiff{1, 1}, // above_pads
Strides{1, 1}, // data_dilation
C,
D,
E,
F,
G,
H,
element::i8,
AxisSet{});
auto f = make_shared<Function>(NodeVector{CV}, ParameterVector{A, B, C, D, E, F, G, H});
auto backend = runtime::Backend::create("CPU");
// Create some tensors for input/output
auto a = backend->create_tensor(element::u8, shape_a);
copy_data(a, a_data);
auto b = backend->create_tensor(element::i8, shape_b);
copy_data(b, b_data);
auto d = backend->create_tensor(element::f32, Shape{1});
copy_data(d, vector<float>{0.0f});
auto e = backend->create_tensor(element::f32, Shape{1});
copy_data(e, vector<float>{255.0f});
auto e_a = backend->create_tensor(element::f32, Shape{1});
copy_data(e_a, vector<float>{-127.0f});
auto g = backend->create_tensor(element::f32, Shape{1});
copy_data(g, vector<float>{127.0f});
auto h = backend->create_tensor(element::f32, Shape{1});
copy_data(h, vector<float>{22.0f});
auto i = backend->create_tensor(element::f32, Shape{1});
copy_data(i, vector<float>{90.0f});
auto result = backend->create_tensor(element::i8, shape_r);
auto handle = backend->compile(f);
handle->call_with_validate({result}, {a, b, d, e, e_a, g, h, i});
EXPECT_EQ((vector<int8_t>{31, 48, 42, 45, 54, 102, 127, 61, 47, 74, 61, 55}),
read_vector<int8_t>(result));
}
TEST(builder, scaled_QC_with_relu)
{
Shape shape_a{1, 1, 3, 3}; // input shape
......
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