From 72f39be2f372b9a810cb27320dba5d0722407549 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Mon, 19 Feb 2018 15:33:41 +0000 Subject: COMPMID-939 Fix mismatches and finalize CLSoftmaxLayer optimization Change-Id: I4404f91a270e0ba7bbb7451c4c43a485fd4a3f6c Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/121105 Tested-by: Jenkins Reviewed-by: Anthony Barbier --- src/core/CL/cl_kernels/softmax_layer.cl | 174 +------------------------------- 1 file changed, 1 insertion(+), 173 deletions(-) (limited to 'src/core/CL/cl_kernels/softmax_layer.cl') diff --git a/src/core/CL/cl_kernels/softmax_layer.cl b/src/core/CL/cl_kernels/softmax_layer.cl index fc45ae6ff6..aa1fa01c53 100644 --- a/src/core/CL/cl_kernels/softmax_layer.cl +++ b/src/core/CL/cl_kernels/softmax_layer.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -88,178 +88,6 @@ __constant VEC_DATA_TYPE(DATA_TYPE, 16) type_min = (VEC_DATA_TYPE(DATA_TYPE, 16) __constant uint16 idx16 = (uint16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); __constant uint4 idx4 = (uint4)(0, 1, 2, 3); -/** Identifies the maximum value across the 1st dimension. - * - * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Fixed point position must be given as a preprocessor argument using -DFIXED_POINT_POSITION=pos. e.g. DFIXED_POINT_POSITION=4 - * @note In case the input is not multiple of 16 -DNON_MULTIPLE_OF_16 must be passed. - * - * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: QS8/QS16/F16/F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor slice. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] width Input image width - */ -__kernel void softmax_layer_max( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - uint width) -{ - Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); - Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); - - // Initialize local maximum - VEC_DATA_TYPE(DATA_TYPE, 16) - max_val = (VEC_DATA_TYPE(DATA_TYPE, 16))type_min; - - // Calculate max of row - const uint width4 = width >> 4; - for(uint i = 0; i < width4; i++) - { - VEC_DATA_TYPE(DATA_TYPE, 16) - data = vload16(0, (__global DATA_TYPE *)offset(&src, i << 4, 0)); - max_val = MAX_OP(data, max_val, DATA_TYPE, 16); - } - -#ifdef NON_MULTIPLE_OF_16 - // Handle non multiple of 16 - VEC_DATA_TYPE(DATA_TYPE, 16) - data = vload16(0, (__global DATA_TYPE *)offset(&src, width4 << 4, 0)); - VEC_DATA_TYPE(SELECT_DATA_TYPE, 16) - widx = CONVERT(((uint16)(width4 << 4) + idx16) < width, VEC_DATA_TYPE(SELECT_DATA_TYPE, 16)); - max_val = MAX_OP(max_val, select(type_min, data, widx), DATA_TYPE, 16); -#endif /* NON_MULTIPLE_OF_16 */ - - // Perform max reduction - max_val.s01234567 = MAX_OP(max_val.s01234567, max_val.s89ABCDEF, DATA_TYPE, 8); - max_val.s0123 = MAX_OP(max_val.s0123, max_val.s4567, DATA_TYPE, 4); - max_val.s01 = MAX_OP(max_val.s01, max_val.s23, DATA_TYPE, 2); - max_val.s0 = MAX_OP(max_val.s0, max_val.s1, DATA_TYPE, 1); - - // Store result - *((__global DATA_TYPE *)dst.ptr) = max_val.s0; -} - -/** Shifts the values of the input tensor by the max calculated in softmax_layer_max kernel, - * then gets the exponent of each element as sums all elements across each row. - * - * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * @note Fixed point position must be given as a preprocessor argument using -DFIXED_POINT_POSITION=pos. e.g. DFIXED_POINT_POSITION=4 - * @note In case the input is not multiple of 16 -DNON_MULTIPLE_OF_16 must be passed. - * @note Beta can be optionally passed at compile time using -DBETA (if undefined, assume it equals 1.0) - * - * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: QS8/QS16/F16/F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[in] max_ptr Pointer to the max values tensor slice. Supported data types: same as @p src_ptr - * @param[in] max_stride_x Stride of the max values tensor in X dimension (in bytes) - * @param[in] max_step_x max_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] max_stride_y Stride of the max values tensor in Y dimension (in bytes) - * @param[in] max_step_y max_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] max_stride_z Stride of the max values tensor in Z dimension (in bytes) - * @param[in] max_step_z max_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] max_offset_first_element_in_bytes The offset of the first element in the max values tensor - * @param[out] dst_ptr Pointer to the destination tensor slice. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[out] sum_ptr Pointer to the sum values tensor slice. Supported data types: same as @p src_ptr - * @param[in] sum_stride_x Stride of the sum values tensor in X dimension (in bytes) - * @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] sum_stride_y Stride of the sum values tensor in Y dimension (in bytes) - * @param[in] sum_step_y sum_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] sum_stride_z Stride of the sum values tensor in Z dimension (in bytes) - * @param[in] sum_step_z sum_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor - * @param[in] width Input image width - */ -__kernel void softmax_layer_shift_exp_sum( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(max), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(sum), - uint width) -{ - Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); - Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); - Image max = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(max); - Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(sum); - -#ifdef BETA - // Initialize beta - VEC_DATA_TYPE(DATA_TYPE, 16) - beta = (VEC_DATA_TYPE(DATA_TYPE, 16))BETA; -#endif /* BETA */ - - // Load max value of 1D logits vector (row) - DATA_TYPE max_val = *((__global DATA_TYPE *)offset(&max, 0, 0)); - - // Set sum vector - VEC_DATA_TYPE(DATA_TYPE, 16) - sum1D = 0; - - // Shift values, exp and sum - const uint width4 = width >> 4; - for(uint i = 0; i < width4; i++) - { - VEC_DATA_TYPE(DATA_TYPE, 16) - data = vload16(0, (__global DATA_TYPE *)offset(&src, i << 4, 0)); - data = SUB_OP(data, max_val, DATA_TYPE, 16); -#ifdef BETA - data = MUL_OP(data, beta, DATA_TYPE, 16); -#endif /* BETA */ - data = EXP_OP(data, DATA_TYPE, 16); - vstore16(data, 0, (__global DATA_TYPE *)offset(&dst, i << 4, 0)); - sum1D = ADD_OP(sum1D, data, DATA_TYPE, 16); - } - -#ifdef NON_MULTIPLE_OF_16 - // Handle non multiple of 16 - VEC_DATA_TYPE(DATA_TYPE, 16) - data = vload16(0, (__global DATA_TYPE *)offset(&src, width4 << 4, 0)); - data = SUB_OP(data, max_val, DATA_TYPE, 16); -#ifdef BETA - data = MUL_OP(data, beta, DATA_TYPE, 16); -#endif /* BETA */ - data = EXP_OP(data, DATA_TYPE, 16); - VEC_DATA_TYPE(SELECT_DATA_TYPE, 16) - widx = CONVERT(((uint16)(width4 << 4) + idx16) < width, VEC_DATA_TYPE(SELECT_DATA_TYPE, 16)); - data = select(0, data, widx); - vstore16(data, 0, (__global DATA_TYPE *)offset(&dst, width4 << 4, 0)); - sum1D = ADD_OP(sum1D, data, DATA_TYPE, 16); -#endif /* NON_MULTIPLE_OF_16 */ - - // Perform min/max reduction - sum1D.s01234567 = ADD_OP(sum1D.s01234567, sum1D.s89ABCDEF, DATA_TYPE, 8); - sum1D.s0123 = ADD_OP(sum1D.s0123, sum1D.s4567, DATA_TYPE, 4); - sum1D.s01 = ADD_OP(sum1D.s01, sum1D.s23, DATA_TYPE, 2); - sum1D.s0 = ADD_OP(sum1D.s0, sum1D.s1, DATA_TYPE, 1); - - // Calculate and store result - *((__global DATA_TYPE *)sum.ptr) = sum1D.s0; -} - /** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel. * * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short -- cgit v1.2.1