Skip to content

N
ngraph
Unverified Commit 21525b8e authored by Robert Kimball's avatar Robert Kimball Committed by GitHub
Browse files
Options

Merge branch 'master' into master

parents e165a460 5713b34d
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 113 additions and 7 deletions
src/ngraph/runtime/cpu/builder/convolution.cpp
View file @ 21525b8e
......@@ -186,7 +186,9 @@ namespace ngraph
auto& arg0_tensor = external_function->get_tensor_data(args[0].get_name());
auto& arg1_tensor = external_function->get_tensor_data(args[1].get_name());
auto& arg2_tensor = external_function->get_tensor_data(args[2].get_name());
auto& arg3_tensor = external_function->get_tensor_data(args[3].get_name());
auto& out_tensor = external_function->get_tensor_data(out[0].get_name());
size_t arg3_size = args[3].get_size();
if (runtime::cpu::mkldnn_utils::use_mkldnn_kernel(node))
{
......@@ -196,8 +198,14 @@ namespace ngraph
node, args, out);
auto& deps = mkldnn_emitter->get_primitive_deps(conv_index);
auto functor = [&, conv_index](CPURuntimeContext* ctx,
CPUExecutionContext* ectx) {
auto functor = [&, conv_index, arg3_size](CPURuntimeContext* ctx,
CPUExecutionContext* ectx) {
if (out_tensor != arg3_tensor)
{
memcpy(static_cast<char*>(out_tensor),
static_cast<char*>(arg3_tensor),
arg3_size);
}
cpu::mkldnn_utils::set_memory_ptr(ctx, deps[0], arg0_tensor);
cpu::mkldnn_utils::set_memory_ptr(ctx, deps[1], arg1_tensor);
cpu::mkldnn_utils::set_memory_ptr(ctx, deps[2], arg2_tensor);
......@@ -219,7 +227,9 @@ namespace ngraph
auto& arg0_tensor = external_function->get_tensor_data(args[0].get_name());
auto& arg1_tensor = external_function->get_tensor_data(args[1].get_name());
auto& arg2_tensor = external_function->get_tensor_data(args[2].get_name());
auto& out_tensor = external_function->get_tensor_data(out[0].get_name());
size_t arg2_size = args[2].get_size();
if (runtime::cpu::mkldnn_utils::use_mkldnn_kernel(node))
{
......@@ -228,8 +238,14 @@ namespace ngraph
node, args, out);
auto& deps = mkldnn_emitter->get_primitive_deps(conv_index);
auto functor = [&, conv_index](CPURuntimeContext* ctx,
CPUExecutionContext* ectx) {
auto functor = [&, conv_index, arg2_size](CPURuntimeContext* ctx,
CPUExecutionContext* ectx) {
if (out_tensor != arg2_tensor)
{
memcpy(static_cast<char*>(out_tensor),
static_cast<char*>(arg2_tensor),
arg2_size);
}
cpu::mkldnn_utils::set_memory_ptr(ctx, deps[0], arg0_tensor);
cpu::mkldnn_utils::set_memory_ptr(ctx, deps[1], arg1_tensor);
cpu::mkldnn_utils::set_memory_ptr(ctx, deps[2], out_tensor);
......
src/ngraph/runtime/cpu/cpu_emitter.cpp
View file @ 21525b8e
......@@ -2410,6 +2410,11 @@ namespace ngraph
node, args, out);
auto& deps = mkldnn_emitter->get_primitive_deps(qconv_index);
writer << "if (" << out[0].get_name() << " != " << args[3].get_name() << ")\n";
writer.block_begin();
writer << "memcpy(" << out[0].get_name() << ", " << args[3].get_name() << ", "
<< args[3].get_size() * args[3].get_element_type().size() << ");\n";
writer.block_end();
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[0])
<< ", " << args[0].get_name() << ");\n";
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[1])
......@@ -2441,6 +2446,11 @@ namespace ngraph
node, args, out);
auto& deps = mkldnn_emitter->get_primitive_deps(qconv_index);
writer << "if (" << out[0].get_name() << " != " << args[3].get_name() << ")\n";
writer.block_begin();
writer << "memcpy(" << out[0].get_name() << ", " << args[3].get_name() << ", "
<< args[3].get_size() * args[3].get_element_type().size() << ");\n";
writer.block_end();
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[0])
<< ", " << args[0].get_name() << ");\n";
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[1])
......@@ -2500,6 +2510,11 @@ namespace ngraph
node, args, out);
auto& deps = mkldnn_emitter->get_primitive_deps(conv_index);
writer << "if (" << out[0].get_name() << " != " << args[3].get_name() << ")\n";
writer.block_begin();
writer << "memcpy(" << out[0].get_name() << ", " << args[3].get_name() << ", "
<< args[3].get_size() * args[3].get_element_type().size() << ");\n";
writer.block_end();
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[0])
<< ", " << args[0].get_name() << ");\n";
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[1])
......@@ -2527,6 +2542,11 @@ namespace ngraph
node, args, out);
auto& deps = mkldnn_emitter->get_primitive_deps(conv_index);
writer << "if (" << out[0].get_name() << " != " << args[2].get_name() << ")\n";
writer.block_begin();
writer << "memcpy(" << out[0].get_name() << ", " << args[2].get_name() << ", "
<< args[2].get_size() * args[2].get_element_type().size() << ");\n";
writer.block_end();
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[0])
<< ", " << args[0].get_name() << ");\n";
writer << "cpu::mkldnn_utils::set_memory_ptr(ctx, " << to_string(deps[1])
......
src/ngraph/runtime/cpu/mkldnn_emitter.cpp
View file @ 21525b8e
......@@ -772,10 +772,18 @@ size_t MKLDNNEmitter::build_reorder(const mkldnn::memory::desc& input_desc,
size_t input_index = build_memory_primitive(input_desc);
size_t result_index = build_memory_primitive(result_desc);
size_t primitive_index = insert_primitive(
new mkldnn::reorder(*m_mkldnn_primitives[input_index], *m_mkldnn_primitives[result_index]));
size_t primitive_index = 0;
try
{
primitive_index = insert_primitive(new mkldnn::reorder(*m_mkldnn_primitives[input_index],
*m_mkldnn_primitives[result_index]));
m_primitive_deps[primitive_index] = {input_index, result_index};
}
catch (const mkldnn::error& e)
{
throw ngraph_error("Could not create mkldnn primitive " + e.message);
}
m_primitive_deps[primitive_index] = {input_index, result_index};
return primitive_index;
}
......
test/builder_quantization.cpp
View file @ 21525b8e
......@@ -679,6 +679,68 @@ TEST(builder, scaled_QC_with_bias_signed_add_and_relu)
read_vector<uint8_t>(result));
}
TEST(builder, scaled_QC_with_bias_signed_add_and_relu_nhwc)
{
Shape shape_a{1, 3, 4, 1}; // input shape
Shape shape_b{1, 3, 3, 1}; // 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};
vector<int32_t> c_data = {5};
vector<int8_t> conv_2_data = {-1, -2, -3, -4, -5, -6, -10, 0, 1, 2, 3, 4};
auto A = make_shared<op::Parameter>(element::u8, shape_a);
auto A_reshape = make_shared<op::Reshape>(A, AxisVector{0, 3, 1, 2}, Shape{1, 1, 3, 4});
auto B = make_shared<op::Parameter>(element::i8, shape_b);
auto B_reshape = make_shared<op::Reshape>(B, AxisVector{0, 3, 1, 2}, Shape{1, 1, 3, 3});
auto Add = make_shared<op::Parameter>(element::i8, shape_a);
auto Add_reshape = make_shared<op::Reshape>(Add, AxisVector{0, 3, 1, 2}, Shape{1, 1, 3, 4});
auto Bias = make_shared<op::Parameter>(element::i32, Shape{1});
auto C = op::Constant::create(element::f32, Shape{}, {0.0f});
auto D = op::Constant::create(element::f32, Shape{}, {255.0f});
auto E = op::Constant::create(element::f32, Shape{}, {-127.0f});
auto F = op::Constant::create(element::f32, Shape{}, {127.0f});
auto G = op::Constant::create(element::f32, Shape{}, {22.0f});
auto H = op::Constant::create(element::f32, Shape{}, {90.0f});
auto I = op::Constant::create(element::f32, Shape{}, {22.0f});
auto J = op::Constant::create(element::f32, Shape{}, {90.0f});
auto CV =
ngraph::builder::ScaledQuantizedConvolutionBiasSignedAdd(A_reshape,
B_reshape,
Bias,
Add_reshape,
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,
I,
J,
true);
auto f = make_shared<Function>(NodeVector{CV}, ParameterVector{A, B, Bias, Add});
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 c = backend->create_tensor(element::i32, Shape{1});
copy_data(c, c_data);
auto d = backend->create_tensor(element::i8, shape_a);
copy_data(d, conv_2_data);
auto result = backend->create_tensor(element::u8, shape_r);
auto handle = backend->compile(f);
backend->call_with_validate(handle, {result}, {a, b, c, d});
EXPECT_EQ((vector<uint8_t>{74, 106, 93, 97, 112, 127, 127, 127, 110, 127, 127, 127}),
read_vector<uint8_t>(result));
}
TEST(builder, dynamic_scaled_QC_with_bias_signed_add_and_relu)
{
Shape shape_a{1, 1, 3, 4}; // 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