IntenseWebs/openvino
3
0 Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Fix fused ops reshaping for OneDNN FC (#9217)

Browse Source
This commit is contained in:
Sergey Shlyapnikov authored and GitHub committed 2021-12-17 13:26:59 +03:00
1 parent e212a1ac83
commit 71f3e2da10
3 files changed
+68 -32
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

inference-engine/thirdparty/clDNN/src/graph_optimizer/add_onednn_optimization_attributes.cpp
Vendored
+22
View File
@@ -7,12 +7,34 @@
#include "pass_manager.h"
#include "program_node.h"
#ifdef ENABLE_ONEDNN_FOR_GPU
#include "fully_connected_inst.h"
#include <impls/onednn/utils.hpp>
#endif
using namespace cldnn;
void add_onednn_optimization_attributes::run(program& p) {
#ifdef ENABLE_ONEDNN_FOR_GPU
for (auto& node : p.get_processing_order()) {
if (node->get_preferred_impl_type() == impl_types::onednn) {
if (node->is_type<fully_connected>()) {
auto fc_prim = node->as<fully_connected>().get_primitive();
// Reshape fused ops tensors for OneDNN FC if needed
if (fc_prim->input_size == 3) {
for (auto& fused_prim : node->get_fused_primitives()) {
auto fused_node = fused_prim.node;
if (fused_node->is_type<eltwise>()) {
auto& dependency = node->get_dependency(fused_prim.dep_start_idx);
auto original_layout = dependency.get_output_layout();
onednn::combine_bf_with_first_spatial_dim(original_layout);
dependency.set_output_layout(original_layout, false);
}
}
}
}
node->init_onednn_primitive_attributes();
}
}
inference-engine/thirdparty/clDNN/src/impls/onednn/fully_connected_onednn.cpp
Vendored
-14
View File
@@ -128,20 +128,6 @@ public:
static primitive_impl* create(const fully_connected_node& arg) {
auto& engine = arg.get_program().get_engine();
auto desc = get_fully_connected_descriptor(arg);
auto prim = arg.get_primitive();
if (prim->input_size == 3) {
for (auto& fused_node : arg.get_fused_primitives()) {
auto node = fused_node.node;
if (node->is_type<eltwise>()) {
auto& dependency = arg.get_dependency(fused_node.dep_start_idx);
auto original_layout = dependency.get_output_layout();
onednn::combine_bf_with_first_spatial_dim(original_layout);
dependency.set_output_layout(original_layout, false);
}
}
}
auto attr = arg.get_onednn_primitive_attributes();
dnnl::primitive_desc prim_desc{&desc->data, attr.get(), engine.get_onednn_engine(), nullptr};
inference-engine/thirdparty/clDNN/tests/test_cases/fusings_gpu_test.cpp
Vendored
+46 -18
View File
@@ -608,6 +608,7 @@ public:
#define CASE_FC_U8S8_3D_1 {2, 32, 1, 3}, {2, 32, 1, 16}, {16, 3, 1, 1}, tensor{1}, tensor{0}, tensor{1}, 1, data_types::u8, format::bfyx, data_types::i8, format::oiyx, data_types::f32, format::bfyx
#define CASE_FC_U8S8_3D_2 {1, 1, 1, 3}, {1, 1, 1, 32}, {32, 3, 1, 1}, tensor{1}, tensor{0}, tensor{1}, 1, data_types::u8, format::bfyx, data_types::i8, format::oiyx, data_types::f32, format::bfyx
#define CASE_FC_U8S8_3D_3 {2, 3, 1, 1}, {2, 3, 1, 15}, {15, 1, 1, 1}, tensor{1}, tensor{0}, tensor{1}, 1, data_types::u8, format::bfyx, data_types::i8, format::oiyx, data_types::f32, format::bfyx
#define CASE_FC_U8S8_3D_4 {1, 512, 1, 1024}, {1, 384, 1, 1024}, {1024, 1024, 1, 1}, tensor{1}, tensor{0}, tensor{1}, 1, data_types::u8, format::bfyx, data_types::i8, format::oiyx, data_types::f32, format::bfyx
#define CASE_NORMALIZE_I8_1 {1, 2, 3, 3}, data_types::u8, format::bfyx, data_types::f32, format::bfyx
@@ -9258,7 +9259,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, gather_elements_activation_scale_eltwise,
}));
#ifdef ENABLE_ONEDNN_FOR_GPU
class ConvFusingTestOneDNN : public WeightsPrimitiveFusingTest<bc_test_params> {
class WeightsPrimitiveFusingTestOneDNN : public WeightsPrimitiveFusingTest<bc_test_params> {
public:
void execute(bc_test_params& p) {
// Onednn post operation has issue in a machine that does not support imad.
@@ -9299,7 +9300,7 @@ public:
}
};
class conv_int8_eltwise_onednn : public ConvFusingTestOneDNN {};
class conv_int8_eltwise_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_int8_eltwise_onednn, u8_eltwise_sum_out) {
auto p = GetParam();
@@ -9364,7 +9365,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_int8_eltwise_onednn,
bc_test_params{CASE_CONV3D_S8S8_5, 3, 4},
}));
class conv_fp32_activation_abs_onednn : public ConvFusingTestOneDNN {};
class conv_fp32_activation_abs_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_fp32_activation_abs_onednn, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9387,7 +9388,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp32_activation_abs_onednn,
bc_test_params{CASE_CONV_FP16_4, 2, 3},
}));
class conv_fp32_activation_mish_onednn : public ConvFusingTestOneDNN {};
class conv_fp32_activation_mish_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_fp32_activation_mish_onednn, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9410,7 +9411,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp32_activation_mish_onednn,
bc_test_params{CASE_CONV_FP16_4, 2, 3},
}));
class conv_fp32_activation_swish_onednn : public ConvFusingTestOneDNN {};
class conv_fp32_activation_swish_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_fp32_activation_swish_onednn, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9433,7 +9434,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp32_activation_swish_onednn,
bc_test_params{CASE_CONV_FP16_4, 2, 3},
}));
class conv_fp32_activation_hswish_onednn : public ConvFusingTestOneDNN {};
class conv_fp32_activation_hswish_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_fp32_activation_hswish_onednn, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9456,7 +9457,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp32_activation_hswish_onednn,
bc_test_params{CASE_CONV_FP16_4, 2, 3},
}));
class conv_fp32_activation_exp_onednn : public ConvFusingTestOneDNN {};
class conv_fp32_activation_exp_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_fp32_activation_exp_onednn, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9479,7 +9480,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp32_activation_exp_onednn,
bc_test_params{CASE_CONV_FP16_4, 2, 3},
}));
class conv_int8_quantize_u8_onednn : public ConvFusingTestOneDNN {};
class conv_int8_quantize_u8_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_int8_quantize_u8_onednn, per_channel) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9526,7 +9527,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_int8_quantize_u8_onednn,
bc_test_params{CASE_CONV_S8S8_3, 2, 3},
}));
class conv_int8_activation_eltwise_quantize_onednn : public ConvFusingTestOneDNN {};
class conv_int8_activation_eltwise_quantize_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_int8_activation_eltwise_quantize_onednn, bsv32_fsv32) {
auto p = GetParam();
layout eltwise_layout = get_output_layout(p);
@@ -9578,7 +9579,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_int8_activation_eltwise_quantize_oned
bc_test_params{CASE_CONV_S8S8_15, 2, 5},
}));
class conv_int8_scale_shift_swish_onednn : public ConvFusingTestOneDNN {};
class conv_int8_scale_shift_swish_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_int8_scale_shift_swish_onednn, bsv32_fsv32) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9617,7 +9618,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_int8_scale_shift_swish_onednn,
bc_test_params{CASE_CONV_S8S8_15, 2, 7},
}));
class conv_int8_eltwise_scale_onednn : public ConvFusingTestOneDNN {};
class conv_int8_eltwise_scale_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_int8_eltwise_scale_onednn, u8_eltwise_prod_out_reuse) {
auto p = GetParam();
@@ -9667,7 +9668,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_int8_eltwise_scale_onednn,
// Limitations: no
// DNNL_VERBOSE log without optimization: attr-post-ops:eltwise_linear:12.75:127.5+eltwise_linear:1:-128
// DNNL_VERBOSE log with optimization: attr-post-ops:eltwise_linear:12.75:-0.5
class post_ops_optimizations_onednn_eltw_linear_eltw_linear : public ConvFusingTestOneDNN {};
class post_ops_optimizations_onednn_eltw_linear_eltw_linear : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(post_ops_optimizations_onednn_eltw_linear_eltw_linear, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9718,7 +9719,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_onednn_eltw_linear_
// Limitations: beta = 0 in eltw_linear
// DNNL_VERBOSE log without optimization: attr-post-ops:eltwise_linear:12.75:127.5+eltwise_round+eltwise_linear:2.00784+eltwise_clip:0:512
// DNNL_VERBOSE log with optimization: attr-post-ops:eltwise_linear:12.75:127.5+eltwise_round:0:0:2.00784+eltwise_clip:0:512
class post_ops_optimizations_onednn_eltw_non_linear_eltw_linear : public ConvFusingTestOneDNN {};
class post_ops_optimizations_onednn_eltw_non_linear_eltw_linear : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(post_ops_optimizations_onednn_eltw_non_linear_eltw_linear, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9769,7 +9770,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_onednn_eltw_non_lin
// Limitations: alpha = 1 and scale = 1 in eltw_linear; binary_add is a constant compile-time buffer
// DNNL_VERBOSE log without optimization: attr-oscale:2 attr-post-ops:binary_add:f32:2+eltwise_linear:1:-127+eltwise_clip:-127:127
// DNNL_VERBOSE log with optimization: attr-oscale:2 attr-post-ops:binary_add:f32:2+eltwise_clip:-127:127
class post_ops_optimizations_onednn_binary_add_eltw_linear : public ConvFusingTestOneDNN {};
class post_ops_optimizations_onednn_binary_add_eltw_linear : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(post_ops_optimizations_onednn_binary_add_eltw_linear, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9820,7 +9821,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_onednn_binary_add_e
// Limitations: beta = 0 in eltw_linear; binary_mul is a constant compile-time buffer
// DNNL_VERBOSE log without optimization: attr-oscale:2 attr-post-ops:binary_mul:f32:2+eltwise_linear:2.01575+eltwise_clip:0:512
// DNNL_VERBOSE log with optimization: attr-oscale:2 attr-post-ops:binary_mul:f32:2+eltwise_clip:0:512
class post_ops_optimizations_onednn_binary_mul_eltw_linear : public ConvFusingTestOneDNN {};
class post_ops_optimizations_onednn_binary_mul_eltw_linear : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(post_ops_optimizations_onednn_binary_mul_eltw_linear, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9871,7 +9872,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_onednn_binary_mul_e
// Limitations: beta = 0 in eltw_linear
// DNNL_VERBOSE log without optimization: attr-oscale:2 attr-post-ops:eltwise_linear:2.01575+eltwise_clip:0:512
// DNNL_VERBOSE log with optimization: attr-oscale:2 attr-post-ops:eltwise_clip:0:512
class post_ops_optimizations_onednn_oscale_eltw_linear : public ConvFusingTestOneDNN {};
class post_ops_optimizations_onednn_oscale_eltw_linear : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(post_ops_optimizations_onednn_oscale_eltw_linear, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9920,7 +9921,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_onednn_oscale_eltw_
// Limitations: beta = 0 in eltw_linear
// DNNL_VERBOSE log without optimization: attr-post-ops:eltwise_relu+sum:1:0:u8+eltwise_linear:12.7+eltwise_clip:0:127
// DNNL_VERBOSE log with optimization: attr-post-ops:eltwise_relu:0:0:12.7+sum:12.7:0:u8+eltwise_clip:0:127
class post_ops_optimizations_onednn_eltw_any_sum_eltw_linear : public ConvFusingTestOneDNN {};
class post_ops_optimizations_onednn_eltw_any_sum_eltw_linear : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(post_ops_optimizations_onednn_eltw_any_sum_eltw_linear, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -9970,7 +9971,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_onednn_eltw_any_sum
// Input range uses in 2 cases: not per-tensor output range or out_lo > out_hi
// Here's out_lo > out_hi and no optimizations
// DNNL_VERBOSE log: attr-post-ops:eltwise_linear:12.75:127.5+eltwise_round+eltwise_linear:-1:127
class post_ops_optimizations_input_range : public ConvFusingTestOneDNN {};
class post_ops_optimizations_input_range : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(post_ops_optimizations_input_range, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
@@ -10015,6 +10016,33 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_input_range,
bc_test_params{CASE_CONV_S8S8_14, 2, 3},
bc_test_params{CASE_CONV_S8S8_15, 2, 3},
}));
class fc_int8_inputs_fused_fp32_sum : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(fc_int8_inputs_fused_fp32_sum, basic) {
auto p = GetParam();
auto shift_layout = layout{ p.default_type, p.default_format, tensor{1, 1, 1, p.kernel.batch[0]} };
create_topologies(input_layout("input", get_input_layout(p)),
data("weights", get_mem(get_fc_weights_layout(p))),
data("bias", get_mem(get_fc_bias_layout(p))),
data("shift_data", get_mem(shift_layout, 1)),
fully_connected("fc_prim", "input", "weights", "bias", cldnn::data_types::f32, "", padding(), get_fc_output_dim_size(p)),
eltwise("shift", {"fc_prim", "shift_data"}, eltwise_mode::sum, cldnn::data_types::f32),
crop("crop", "shift", get_output_layout(p).size, {0, 0, 0, 0}),
reorder("reorder_bfyx", "crop", p.default_format, data_types::f32)
);
tolerance = 1e-5f;
execute(p);
}
INSTANTIATE_TEST_SUITE_P(fusings_gpu, fc_int8_inputs_fused_fp32_sum, ::testing::ValuesIn(std::vector<bc_test_params>{
// OneDNN has issue with small shapes - ticket 7064
// bc_test_params{ CASE_FC_U8S8_3D_1, 2, 4 },
// bc_test_params{ CASE_FC_U8S8_3D_2, 2, 4 },
bc_test_params{ CASE_FC_U8S8_3D_4, 2, 4 },
}));
#endif