Skip to content

Commit

Permalink
[IE CLDNN] Added int8 output suppirt into bfyx_to_fsv16 fp kernel (op…
Browse files Browse the repository at this point in the history 
…envinotoolkit#2906)
  • Loading branch information
@vladimir-paramuzov @sdurawa
vladimir-paramuzov authored and sdurawa committed Nov 6, 2020
1 parent afc98ca commit f6b35e8
Show file tree
Hide file tree
Showing 3 changed files with 51 additions and 25 deletions.
5 changes: 4 additions & 1 deletion ...nel_selector/core/actual_kernels/convolution/convolution_kernel_bfyx_to_b_fs_yx_fsv16.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,8 @@ ParamsKey ConvolutionKernel_bfyx_to_bfyx_f16::GetSupportedKey() const {
k.EnableInputDataType(Datatype::F32);
k.EnableOutputDataType(Datatype::F16);
k.EnableOutputDataType(Datatype::F32);
k.EnableOutputDataType(Datatype::UINT8);
k.EnableOutputDataType(Datatype::INT8);
k.EnableInputWeightsType(WeightsType::F16);
k.EnableInputWeightsType(WeightsType::F32);
k.EnableInputLayout(DataLayout::bfyx);
Expand All @@ -67,6 +69,7 @@ ParamsKey ConvolutionKernel_bfyx_to_bfyx_f16::GetSupportedKey() const {
k.EnableBatching();
k.EnableSubGroup();
k.EnableSubGroupShort();
k.EnableDifferentTypes();
return k;
}

Expand Down Expand Up @@ -132,7 +135,7 @@ JitConstants ConvolutionKernel_bfyx_to_bfyx_f16::GetJitConstants(const convoluti
auto blockWidth = dispatchData.cldnnStyle.blockWidth;

if (!params.fused_ops.empty()) {
auto input_dt = GetUnitType(params);
auto input_dt = GetActivationType(params);
FusedOpsConfiguration conf_vec = { "_VEC",
{"b", "(f_block*16)", "y", "x"},
"dst",
Expand Down
45 changes: 21 additions & 24 deletions ...gine/thirdparty/clDNN/kernel_selector/core/cl_kernels/convolution_gpu_bfyx_to_bfyx_f16.cl
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
// Copyright (c) 2016-2019 Intel Corporation
// Copyright (c) 2016-2020 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
Expand All @@ -13,13 +13,11 @@
// limitations under the License.

#include "include/include_all.cl"
#include "include/unit_type.cl"
#include "include/mmad.cl"

#define FEATURE_SLICE_SIZE 16

// OUTPUT_X_BLOCK_SIZE is one of 2, 4, 8
#define UNIT_BLOCK_WRITEN(ptr, offset, val) CAT(UNIT_BLOCK_WRITE, OUTPUT_X_BLOCK_SIZE)(ptr, offset, val)
#define DT_OUTPUT_BLOCK_WRITEN(ptr, offset, val) BLOCK_WRITEN(OUTPUT_TYPE, OUTPUT_X_BLOCK_SIZE, ptr, offset, val)

__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE)))
__attribute__((reqd_work_group_size(1, SUB_GROUP_SIZE, 1)))
Expand All @@ -43,9 +41,6 @@ KERNEL(convolution_bfyx_to_bfyx_f16)(
const int x = (xy % X_BLOCKS) * OUTPUT_X_BLOCK_SIZE;
const int y = (xy / X_BLOCKS);

typedef MAKE_VECTOR_TYPE(UNIT_TYPE, OUTPUT_X_BLOCK_SIZE) vec_t;
typedef MAKE_VECTOR_TYPE(UNIT_TYPE, 8) wei_t;

const int input_x = x * STRIDE_SIZE_X - PADDING_SIZE_X;
const int input_y = y * STRIDE_SIZE_Y - PADDING_SIZE_Y;

Expand Down Expand Up @@ -104,12 +99,12 @@ KERNEL(convolution_bfyx_to_bfyx_f16)(
bias_offset += split_idx * BIAS_LENGTH;
# endif

vec_t dst = (vec_t)(UNIT_BLOCK_READ(biases, bias_offset));
MAKE_VECTOR_TYPE(INPUT0_TYPE, OUTPUT_X_BLOCK_SIZE) dst = (MAKE_VECTOR_TYPE(INPUT0_TYPE, OUTPUT_X_BLOCK_SIZE))(DT_INPUT_BLOCK_READ(biases, bias_offset));
#else
vec_t dst = UNIT_VAL_ZERO;
MAKE_VECTOR_TYPE(INPUT0_TYPE, OUTPUT_X_BLOCK_SIZE) dst = INPUT0_VAL_ZERO;
#endif

UNIT_TYPE line_cache[INPUT0_FEATURE_NUM * INPUT_BLOCK_SIZE];
INPUT0_TYPE line_cache[INPUT0_FEATURE_NUM * INPUT_BLOCK_SIZE];
for (int ic = 0; ic < INPUT0_FEATURE_NUM; ic++)
{
__attribute__((opencl_unroll_hint(INPUT_BLOCK_SIZE)))
Expand All @@ -125,11 +120,10 @@ KERNEL(convolution_bfyx_to_bfyx_f16)(
xb * input_x_pitch +
yb * input_y_pitch];
else
line_cache[ic * INPUT_BLOCK_SIZE + i] = UNIT_VAL_ZERO;
line_cache[ic * INPUT_BLOCK_SIZE + i] = INPUT0_VAL_ZERO;
}
}


__attribute__((opencl_unroll_hint(FILTER_SIZE_Y)))
for (int kh = 0; kh < FILTER_SIZE_Y; kh++)
{
Expand All @@ -138,18 +132,18 @@ KERNEL(convolution_bfyx_to_bfyx_f16)(
{
uint offset = filter_offset + kh * filter_y_pitch + kw * filter_x_pitch;

UNIT_TYPE wei[INPUT0_FEATURE_NUM];
FILTER_TYPE wei[INPUT0_FEATURE_NUM];
__attribute__((opencl_unroll_hint(INPUT0_FEATURE_NUM)))
for (int ic = 0; ic < INPUT0_FEATURE_NUM; ic++)
wei[ic] = UNIT_BLOCK_READ(weights, offset + ic * filter_isv_pitch);
wei[ic] = DT_FILTER_BLOCK_READ(weights, offset + ic * filter_isv_pitch);

__attribute__((opencl_unroll_hint(OUTPUT_X_BLOCK_SIZE)))
for (int i = 0; i < OUTPUT_X_BLOCK_SIZE; i++)
{
const uint buf_offset = (kw*DILATION_SIZE_X + STRIDE_SIZE_X * i + (kh) * INPUT_LINE_SIZE) / SUB_GROUP_SIZE;
const uint buf_group = (kw*DILATION_SIZE_X + STRIDE_SIZE_X * i + (kh) * INPUT_LINE_SIZE) % SUB_GROUP_SIZE;

UNIT_TYPE src[INPUT0_FEATURE_NUM];
INPUT0_TYPE src[INPUT0_FEATURE_NUM];
__attribute__((opencl_unroll_hint(INPUT0_FEATURE_NUM)))
for (int ic = 0; ic < INPUT0_FEATURE_NUM; ic++) {
src[ic] = intel_sub_group_shuffle(line_cache[ic * INPUT_BLOCK_SIZE + buf_offset], buf_group);
Expand All @@ -159,17 +153,20 @@ KERNEL(convolution_bfyx_to_bfyx_f16)(
}
}

dst = ACTIVATION(dst, ACTIVATION_PARAMS);
MAKE_VECTOR_TYPE(OUTPUT_TYPE, OUTPUT_X_BLOCK_SIZE) res;
#ifndef HAS_FUSED_OPS
res = ACTIVATION(dst, ACTIVATION_PARAMS);
#endif

#if OUTPUT_LEFTOVERS
if ((f_block+1)*FEATURE_SLICE_SIZE >= OUTPUT_FEATURE_NUM) {
for (int i = 0; i < OUTPUT_X_BLOCK_SIZE; i++) {
#if HAS_FUSED_OPS
FUSED_OPS_SCALAR;
dst[i] = FUSED_OPS_RESULT_SCALAR;
res[i] = FUSED_OPS_RESULT_SCALAR;
#endif
if ((f_block*FEATURE_SLICE_SIZE + lid < OUTPUT_FEATURE_NUM) && (x + i) < OUTPUT_SIZE_X)
output[output_offset + i * output_x_pitch + lid] = dst[i];
output[output_offset + i * output_x_pitch + lid] = res[i];
}
}
else
Expand All @@ -178,17 +175,17 @@ KERNEL(convolution_bfyx_to_bfyx_f16)(
if (x + OUTPUT_X_BLOCK_SIZE <= OUTPUT_SIZE_X) {
#if HAS_FUSED_OPS
FUSED_OPS_VEC;
dst = FUSED_OPS_RESULT_VEC;
res = FUSED_OPS_RESULT_VEC;
#endif
UNIT_BLOCK_WRITEN(output, output_offset, dst);
DT_OUTPUT_BLOCK_WRITEN(output, output_offset, res);
} else {
const int x_tail = OUTPUT_SIZE_X - x;
const int x_tail = OUTPUT_SIZE_X % OUTPUT_X_BLOCK_SIZE;
for (int i = 0; i < x_tail; i++) {
#if HAS_FUSED_OPS
FUSED_OPS_SCALAR;
dst[i] = FUSED_OPS_RESULT_SCALAR;
FUSED_OPS_SCALAR;
res[i] = FUSED_OPS_RESULT_SCALAR;
#endif
UNIT_BLOCK_WRITE(output, output_offset + i * output_x_pitch, dst[i]);
DT_OUTPUT_BLOCK_WRITE(output, output_offset + i * output_x_pitch, res[i]);
}
}
}
Expand Down
26 changes: 26 additions & 0 deletions inference-engine/thirdparty/clDNN/tests/test_cases/fusings_gpu_test.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -472,6 +472,7 @@ class ConvEltwTest : public ::BaseFusingTest<conv_eltw_test_params> {
#define CASE_CONV_FP32_11 {1, 32, 4, 5, 4}, {1, 16, 2, 3, 2}, {1, 1, 3, 3, 3}, tensor{1}, tensor{0}, tensor{1}, 2, data_types::f32, format::b_fs_zyx_fsv16, data_types::f32, format::os_is_zyx_isv16_osv16, data_types::f32, format::bfzyx
#define CASE_CONV_FP32_12 {1, 16, 4, 5, 4}, {1, 16, 2, 3, 2}, {1, 1, 3, 3, 3}, tensor{1}, tensor{0}, tensor{1}, 2, data_types::f32, format::b_fs_zyx_fsv16, data_types::f32, format::os_is_zyx_isv16_osv16, data_types::f32, format::bfzyx
#define CASE_CONV_FP32_13 {1, 16, 18, 5, 4}, {1, 16, 16, 3, 2}, {1, 1, 3, 3, 3}, tensor{1}, tensor{0}, tensor{1}, 2, data_types::f32, format::b_fs_zyx_fsv16, data_types::f32, format::os_is_zyx_isv16_osv16, data_types::f32, format::bfzyx
#define CASE_CONV_FP32_14 {1, 3, 4, 5}, {1, 30, 2, 3}, {1, 1, 3, 3}, tensor{1}, tensor{0}, tensor{1}, 1, data_types::f32, format::bfyx, data_types::f32, format::bfyx, data_types::f32, format::bfyx

#define CASE_CONV_FP16_1 {1, 15, 4, 5}, {1, 30, 2, 3}, {1, 1, 3, 3}, tensor{1}, tensor{0}, tensor{1}, 1, data_types::f16, format::bfyx, data_types::f16, format::bfyx, data_types::f16, format::bfyx
#define CASE_CONV_FP16_2 {1, 16, 4, 5}, {1, 32, 2, 3}, {1, 1, 3, 3}, tensor{1}, tensor{0}, tensor{1}, 1, data_types::f16, format::b_fs_yx_fsv16, data_types::f16, format::os_is_yx_isv16_osv16, data_types::f16, format::bfyx
Expand Down Expand Up @@ -853,6 +854,31 @@ INSTANTIATE_TEST_CASE_P(fusings_gpu, conv_fp32_eltwise_b_fs_zyx_fsv16,
bc_test_params{CASE_CONV_FP16_12, 2, 3},
}), );

class conv_fp32_quantize_u8_first_conv : public ConvFusingTest {};
TEST_P(conv_fp32_quantize_u8_first_conv, basic) {
auto p = GetParam();
create_topologies(input_layout("input", get_input_layout(p)),
data("weights", get_mem(get_weights_layout(p))),
data("bias", get_mem(get_bias_layout(p))),
data("in_lo", get_mem(get_per_channel_layout(p), min_random, 0)),
data("in_hi", get_mem(get_per_channel_layout(p), 1, max_random)),
data("out_lo", get_mem(get_single_element_layout(p), 0)),
data("out_hi", get_mem(get_single_element_layout(p), 255)),
reorder("reordered_input", "input", format::b_fs_yx_fsv16, p.data_type),
convolution("conv_prim", "reordered_input", {"weights"}, {"bias"}, p.groups, p.stride, p.pad, p.dilation),
quantize("quantize", "conv_prim", "in_lo", "in_hi", "out_lo", "out_hi", 256, data_types::u8),
reorder("reorder_bfyx", "quantize", p.default_format, data_types::f32)
);

tolerance = 1.0f;
execute(p);
}

INSTANTIATE_TEST_CASE_P(fusings_gpu, conv_fp32_quantize_u8_first_conv,
::testing::ValuesIn(std::vector<bc_test_params>{
bc_test_params{CASE_CONV_FP32_14, 2, 3},
}), );

class conv_fp32_quantize_u8 : public ConvFusingTest {};
TEST_P(conv_fp32_quantize_u8, basic) {
auto p = GetParam();
Expand Down

0 comments on commit f6b35e8

Please sign in to comment.