From 30dbeef2f46bdd6fe05d25dfa27cb4b2359dced3 Mon Sep 17 00:00:00 2001 From: Manuel Bottini Date: Wed, 26 Jun 2019 16:23:03 +0100 Subject: COMPMID-2411: Add (logistic and tanh) activation support for QSYMM16 for CL Change-Id: I8d72490b1cc58563ba7b94664135586bc40e6526 Signed-off-by: Manuel Bottini Reviewed-on: https://review.mlplatform.org/c/1466 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Reviewed-by: Georgios Pinitas --- src/core/CL/cl_kernels/activation_layer_qa8.cl | 196 --------------------- src/core/CL/cl_kernels/activation_layer_quant.cl | 158 +++++++++++++++++ src/core/CL/cl_kernels/activation_quant_helpers.h | 84 +++++++++ .../cl_kernels/depthwise_convolution_quantized.cl | 4 +- 4 files changed, 244 insertions(+), 198 deletions(-) delete mode 100644 src/core/CL/cl_kernels/activation_layer_qa8.cl create mode 100644 src/core/CL/cl_kernels/activation_layer_quant.cl create mode 100644 src/core/CL/cl_kernels/activation_quant_helpers.h (limited to 'src/core/CL/cl_kernels') diff --git a/src/core/CL/cl_kernels/activation_layer_qa8.cl b/src/core/CL/cl_kernels/activation_layer_qa8.cl deleted file mode 100644 index 41f23ca79b..0000000000 --- a/src/core/CL/cl_kernels/activation_layer_qa8.cl +++ /dev/null @@ -1,196 +0,0 @@ -/* - * Copyright (c) 2016-2019 ARM Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "helpers.h" - -#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) -#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE) - -#if defined(FLOAT_DOMAIN) -// Activations performed in the float domain - -#include "activation_float_helpers.h" - -#if defined(O2_VAL) && defined(S2_VAL) -#define OFFSET_OUT O2_VAL -#define SCALE_OUT S2_VAL -#else // defined(O2_VAL) && defined(S2_VAL) -#define OFFSET_OUT O1_VAL -#define SCALE_OUT S1_VAL -#endif // defined(O2_VAL) && defined(S2_VAL) - -/** This performs an activation function on QASYMM8 inputs with float transformations. - * - * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time - * - * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 - * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively. - * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively. - * @note Quantization offsets of the input/output tensors are passed in with -DO1_VAL= and -DO2_VAL= respectively. - * @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128. - * - * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8 - * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image - * @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr - * @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes) - * @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes) - * @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image - */ -__kernel void activation_layer_qa8_f32( - TENSOR3D_DECLARATION(input) -#ifndef IN_PLACE - , - TENSOR3D_DECLARATION(output) -#endif /* not IN_PLACE */ -) -{ - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); -#ifdef IN_PLACE - Tensor3D output = input; -#else /* IN_PLACE */ - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); -#endif /* IN_PLACE */ - - // Load data - TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr); - - VEC_FLOAT data_flt = CONVERT(data, VEC_FLOAT); - data_flt = round(data_flt - (float)O1_VAL) * ((float)S1_VAL); - data_flt = ACTIVATION(ACT, float, data_flt, A_VAL, B_VAL); - - data = CONVERT_SAT(round(data_flt / ((float)SCALE_OUT)) + (float)OFFSET_OUT, TYPE); - - // Store result - VSTORE(VEC_SIZE) - (data, 0, (__global DATA_TYPE *)output.ptr); -} - -#else // defined(FLOAT_DOMAIN) -// Activations performed in the quantized domain - -// RELU Activation -inline TYPE relu_op(TYPE x) -{ - return max((TYPE)CONST_0, x); -} -// Bounded RELU Activation -inline TYPE brelu_op(TYPE x) -{ - return min((TYPE)A_VAL, max(CONST_0, x)); -} -// Lower Upper Bounded RELU Activation -inline TYPE lu_brelu_op(TYPE x) -{ - return min(max(x, (TYPE)B_VAL), (TYPE)A_VAL); -} - -#define ACTIVATION_OP2(op, x) op##_op(x) -#define ACTIVATION_OP(op, x) ACTIVATION_OP2(op, x) - -#if defined(O1_VAL) && defined(O2_VAL) && defined(S1_VAL) && defined(S2_VAL) -#define PERFORM_ACTIVATION_QA8(act, data) \ - ({ \ - data = ACTIVATION_OP(act, data); \ - \ - VEC_DATA_TYPE(float, VEC_SIZE) \ - fdata = CONVERT(data, VEC_DATA_TYPE(float, VEC_SIZE)); \ - \ - fdata = round((fdata - (float)O1_VAL) * ((float)S1_VAL / (float)S2_VAL) + (float)O2_VAL); \ - data = CONVERT_SAT(fdata, VEC_DATA_TYPE(uchar, VEC_SIZE)); \ - }) -#else /* defined(O1_VAL) && defined(O2_VAL) && defined(S1_VAL) && defined(S2_VAL) */ -#define PERFORM_ACTIVATION_QA8(act, data) \ - ({ \ - data = ACTIVATION_OP(act, data); \ - }) -#endif /* defined(O1_VAL) && defined(O2_VAL) && defined(S1_VAL) && defined(S2_VAL) */ - -#if defined(ACT) -/** This performs an activation function on QASYMM8 inputs. - * - * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time - * - * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 - * @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH - * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively. - * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively. - * @note Quantization offsets of the input/output tensors are passed in with -DO1_VAL= and -DO2_VAL= respectively. - * @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128. - * - * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8 - * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image - * @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr - * @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes) - * @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes) - * @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image - */ -__kernel void activation_layer_qa8( - TENSOR3D_DECLARATION(input) -#ifndef IN_PLACE - , - TENSOR3D_DECLARATION(output) -#endif /* not IN_PLACE */ -) -{ - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); -#ifdef IN_PLACE - Tensor3D output = input; -#else /* IN_PLACE */ - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); -#endif /* IN_PLACE */ - - // Load data - TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr); - - data = PERFORM_ACTIVATION_QA8(ACT, data); - - // Store result - VSTORE(VEC_SIZE) - (data, 0, (__global DATA_TYPE *)output.ptr); -} -#endif // defined(ACT) -#endif // defined(FLOAT_DOMAIN) diff --git a/src/core/CL/cl_kernels/activation_layer_quant.cl b/src/core/CL/cl_kernels/activation_layer_quant.cl new file mode 100644 index 0000000000..ebd3408b23 --- /dev/null +++ b/src/core/CL/cl_kernels/activation_layer_quant.cl @@ -0,0 +1,158 @@ +/* + * Copyright (c) 2016-2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "activation_quant_helpers.h" + +#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE) + +#if defined(FLOAT_DOMAIN) +// Activations performed in the float domain + +#include "activation_float_helpers.h" + +/** This performs an activation function on quantized inputs with float transformations. + * + * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time + * + * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively. + * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively. + * @note Quantization offsets of the input/output tensors are passed in only if asymmetric with -DO1_VAL= and -DO2_VAL= respectively. + * @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128. + * + * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QSYMM16 + * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image + * @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr + * @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes) + * @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes) + * @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes) + * @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image + */ +__kernel void activation_layer_quant_f32( + TENSOR3D_DECLARATION(input) +#ifndef IN_PLACE + , + TENSOR3D_DECLARATION(output) +#endif /* not IN_PLACE */ +) +{ + // Get pixels pointer + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); +#ifdef IN_PLACE + Tensor3D output = input; +#else /* IN_PLACE */ + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); +#endif /* IN_PLACE */ + + // Load data + TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr); + + VEC_FLOAT data_flt = CONVERT(data, VEC_FLOAT); +#if defined(O1_VAL) + data_flt = round(data_flt - (float)O1_VAL) * ((float)S1_VAL); +#else // defined(O1_VAL) + data_flt = round(data_flt) * ((float)S1_VAL); +#endif // defined(O1_VAL) + data_flt = ACTIVATION(ACT, float, data_flt, A_VAL, B_VAL); + +#if defined(O2_VAL) + data = CONVERT_SAT(round(data_flt / ((float)S2_VAL)) + (float)O2_VAL, TYPE); +#else // defined(O2_VAL) + data = CONVERT_SAT(round(data_flt / ((float)S2_VAL)), TYPE); +#endif // defined(O2_VAL) + + // Store result + VSTORE(VEC_SIZE) + (data, 0, (__global DATA_TYPE *)output.ptr); +} + +#else // defined(FLOAT_DOMAIN) +// Activations performed in the quantized domain + +#if defined(ACT) +/** This performs an activation function on quantized inputs. + * + * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time + * + * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH + * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively. + * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively. + * @note Quantization offsets of the input/output tensors are passed in with -DO1_VAL= and -DO2_VAL= respectively. + * @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128. + * + * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QSYMM16 + * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image + * @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr + * @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes) + * @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes) + * @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes) + * @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image + */ +__kernel void activation_layer_quant( + TENSOR3D_DECLARATION(input) +#ifndef IN_PLACE + , + TENSOR3D_DECLARATION(output) +#endif /* not IN_PLACE */ +) +{ + // Get pixels pointer + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); +#ifdef IN_PLACE + Tensor3D output = input; +#else /* IN_PLACE */ + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); +#endif /* IN_PLACE */ + + // Load data + TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr); + + data = PERFORM_ACTIVATION_QUANT(ACT, data); + + // Store result + VSTORE(VEC_SIZE) + (data, 0, (__global DATA_TYPE *)output.ptr); +} +#endif // defined(ACT) +#endif // defined(FLOAT_DOMAIN) diff --git a/src/core/CL/cl_kernels/activation_quant_helpers.h b/src/core/CL/cl_kernels/activation_quant_helpers.h new file mode 100644 index 0000000000..402e7ac41f --- /dev/null +++ b/src/core/CL/cl_kernels/activation_quant_helpers.h @@ -0,0 +1,84 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "helpers.h" + +#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + +#if defined(S1_VAL) && !defined(S2_VAL) +#define S2_VAL S1_VAL +#endif // defined(S1_VAL) && !defined(S2_VAL) +#if defined(O1_VAL) && !defined(O2_VAL) +#define O2_VAL O1_VAL +#endif // defined(O1_VAL) && !defined(O2_VAL) + +// RELU Activation +inline TYPE relu_op(TYPE x) +{ + return max((TYPE)CONST_0, x); +} +// Bounded RELU Activation +inline TYPE brelu_op(TYPE x) +{ + return min((TYPE)A_VAL, max(CONST_0, x)); +} +// Lower Upper Bounded RELU Activation +inline TYPE lu_brelu_op(TYPE x) +{ + return min(max(x, (TYPE)B_VAL), (TYPE)A_VAL); +} + +#define ACTIVATION_OP2(op, x) op##_op(x) +#define ACTIVATION_OP(op, x) ACTIVATION_OP2(op, x) + +#if defined(S1_VAL) && defined(S2_VAL) +#if defined(O1_VAL) && defined(O2_VAL) +#define PERFORM_ACTIVATION_QUANT(act, data) \ + ({ \ + data = ACTIVATION_OP(act, data); \ + \ + VEC_DATA_TYPE(float, VEC_SIZE) \ + fdata = CONVERT(data, VEC_DATA_TYPE(float, VEC_SIZE)); \ + \ + fdata = round((fdata - (float)O1_VAL) * ((float)S1_VAL / (float)S2_VAL) + (float)O2_VAL); \ + data = CONVERT_SAT(fdata, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); \ + }) +#else // defined(O1_VAL) && defined(O2_VAL) +#define PERFORM_ACTIVATION_QUANT(act, data) \ + ({ \ + data = ACTIVATION_OP(act, data); \ + \ + VEC_DATA_TYPE(float, VEC_SIZE) \ + fdata = CONVERT(data, VEC_DATA_TYPE(float, VEC_SIZE)); \ + \ + fdata = round((fdata) * ((float)S1_VAL / (float)S2_VAL)); \ + data = CONVERT_SAT(fdata, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); \ + }) +#endif /* defined(O1_VAL) && defined(O2_VAL) */ +#else /* defined(S1_VAL) && defined(S2_VAL) */ +#define PERFORM_ACTIVATION_QUANT(act, data) \ + ({ \ + data = ACTIVATION_OP(act, data); \ + }) +#endif /* defined(S1_VAL) && defined(S2_VAL) */ \ No newline at end of file diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl index 13568b035d..8f2e441693 100644 --- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl +++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl @@ -31,8 +31,8 @@ #ifndef VEC_SIZE #define VEC_SIZE 8 #endif /* VEC_SIZE */ -#include "activation_layer_qa8.cl" -#define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QA8(ACTIVATION_TYPE, x) +#include "activation_layer_quant.cl" +#define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QUANT(ACTIVATION_TYPE, x) #else /* defined(ACTIVATION_TYPE) && defined(CONST_0) */ #define ACTIVATION_FUNC(x) (x) #endif /* defined(ACTIVATION_TYPE) && defined(CONST_0) */ -- cgit v1.2.1