/* * Copyright (c) 2017-2018 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" #if defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) /** This kernel applies dot product to each plane on the input tensor and the corrispective column of the reshaped weight tensor. * * @note Datatype and source width and height should be given as a preprocessor argument using -DDATA_TYPE=type, -DSRC_WIDTH=width and -DSRC_HEIGHT=height. e.g. -DDATA_TYPE=short * * @param[in] src_ptr Pointer to the source tensor. Supported data types: 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 Y 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] weights_ptr Pointer to the weights tensor. Same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[out] dst_ptr Pointer to the destination tensor. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor */ __kernel void gemm_mv(TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(weights), VECTOR_DECLARATION(dst)) { Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); int y = get_global_id(1) * 4; int z = get_global_id(2); __global uchar *current_weights = weights_ptr + weights_offset_first_element_in_bytes + z * weights_stride_y; __global uchar *input_ptr = src.ptr; DATA_TYPE acc0 = (DATA_TYPE)0; DATA_TYPE acc1 = (DATA_TYPE)0; DATA_TYPE acc2 = (DATA_TYPE)0; DATA_TYPE acc3 = (DATA_TYPE)0; // This kernel handle 4 rows in per thread so that it can reuse the weights for(int i = 0; i < SRC_WIDTH; i += 4) { VEC_DATA_TYPE(DATA_TYPE, 4) weights = vload4(0, (__global DATA_TYPE *)(current_weights + i * weights_stride_x)); int4 offset = (int4)i * (int4)src_stride_x + (int4)(0, 1, 2, 3) * (int4)src_stride_y; VEC_DATA_TYPE(DATA_TYPE, 4) tmp0 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s0)); VEC_DATA_TYPE(DATA_TYPE, 4) tmp1 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s1)); VEC_DATA_TYPE(DATA_TYPE, 4) tmp2 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s2)); VEC_DATA_TYPE(DATA_TYPE, 4) tmp3 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s3)); acc0 += dot(weights, tmp0); acc1 += dot(weights, tmp1); acc2 += dot(weights, tmp2); acc3 += dot(weights, tmp3); } __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y + z * SRC_HEIGHT) * dst_stride_x; int rows_left = SRC_HEIGHT - (y + 4); // This if check is used to handle the last few rows when it can't be divided by the four if(rows_left >= 0) { VEC_DATA_TYPE(DATA_TYPE, 4) out = (VEC_DATA_TYPE(DATA_TYPE, 4))(acc0, acc1, acc2, acc3); vstore4(out, 0, (__global DATA_TYPE *)output_ptr); } else { switch(rows_left) { case -1: // three rows left; one is padding *((__global DATA_TYPE *)(output_ptr + 2 * dst_stride_x)) = acc2; case -2: // two rows left; two are padding *((__global DATA_TYPE *)(output_ptr + 1 * dst_stride_x)) = acc1; case -3: // one row left; three are padding *((__global DATA_TYPE *)(output_ptr + 0 * dst_stride_x)) = acc0; break; } } } #endif /* defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) */ #if defined(SRC_WIDTH) && defined(SRC_HEIGHT) /** This kernel applies dot product to each plane on the input tensor and the corresponding column of the reshaped weight tensor. * * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8 * @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 Y 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] weights_ptr Pointer to the weights tensor. Same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[out] dst_ptr Pointer to the destination tensor. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[in] input_offset Input's quantization offset * @param[in] weights_offset Weights's quantization offset */ __kernel void gemm_mv_quantized(TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(weights), VECTOR_DECLARATION(dst), const int input_offset, const int weights_offset) { Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); int y = get_global_id(1) * 4; int z = get_global_id(2); __global uchar *current_weights = weights_ptr + weights_offset_first_element_in_bytes + z * weights_stride_y; __global uchar *input_ptr = src.ptr; int acc0 = 0; int acc1 = 0; int acc2 = 0; int acc3 = 0; // This kernel handle 4 rows in per thread so that it can reuse the weights for(int i = 0; i < SRC_WIDTH; i += 4) { int4 w = convert_int4(vload4(0, (__global uchar *)(current_weights + i * weights_stride_x))) + (int4)weights_offset; int4 offset = (int4)i * (int4)src_stride_x + (int4)(0, 1, 2, 3) * (int4)src_stride_y; int4 tmp0 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s0))) + (int4)input_offset; int4 tmp1 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s1))) + (int4)input_offset; int4 tmp2 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s2))) + (int4)input_offset; int4 tmp3 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s3))) + (int4)input_offset; // Accumulate acc0 += tmp0.s0 * w.s0 + tmp0.s1 * w.s1 + tmp0.s2 * w.s2 + tmp0.s3 * w.s3; acc1 += tmp1.s0 * w.s0 + tmp1.s1 * w.s1 + tmp1.s2 * w.s2 + tmp1.s3 * w.s3; acc2 += tmp2.s0 * w.s0 + tmp2.s1 * w.s1 + tmp2.s2 * w.s2 + tmp2.s3 * w.s3; acc3 += tmp3.s0 * w.s0 + tmp3.s1 * w.s1 + tmp3.s2 * w.s2 + tmp3.s3 * w.s3; } __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y + z * SRC_HEIGHT) * dst_stride_x; int rows_left = SRC_HEIGHT - (y + 4); // This if check is used to handle the last few rows when it can't be divided by the four if(rows_left >= 0) { vstore4((int4)(acc0, acc1, acc2, acc3), 0, (__global int *)output_ptr); } else { switch(rows_left) { case -1: // three rows left; one is padding *((__global int *)(output_ptr + 2 * dst_stride_x)) = acc2; case -2: // two rows left; two are padding *((__global int *)(output_ptr + 1 * dst_stride_x)) = acc1; case -3: // one row left; three are padding *((__global int *)(output_ptr + 0 * dst_stride_x)) = acc0; break; } } } #endif /* defined(SRC_WIDTH) && defined(SRC_HEIGHT) */