diff options
author | Manuel Bottini <manuel.bottini@arm.com> | 2019-10-21 17:59:07 +0100 |
---|---|---|
committer | Manuel Bottini <manuel.bottini@arm.com> | 2019-12-03 13:58:56 +0000 |
commit | 7b9998d0fe1f98768b690ead10ebfa166d1b873d (patch) | |
tree | d3f6b81fb2e414a9e0f8ed9597eab27ef970d725 /src/core/CL/cl_kernels | |
parent | f9179d393a07eb9eed753e315df79d22391906c6 (diff) | |
download | ComputeLibrary-7b9998d0fe1f98768b690ead10ebfa166d1b873d.tar.gz |
COMPMID-1816: Use parallel reduction on 0 axis in CL ARG_MIN/ARG_MAX
Introducing new CLArgMinMax kernel
Change-Id: I0b8254207cc3859d19ceef9b6429cf5c1c586db0
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2202
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels')
-rw-r--r-- | src/core/CL/cl_kernels/arg_min_max.cl | 431 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/helpers.h | 13 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/reduction_operation.cl | 111 |
3 files changed, 462 insertions, 93 deletions
diff --git a/src/core/CL/cl_kernels/arg_min_max.cl b/src/core/CL/cl_kernels/arg_min_max.cl new file mode 100644 index 0000000000..3f75377636 --- /dev/null +++ b/src/core/CL/cl_kernels/arg_min_max.cl @@ -0,0 +1,431 @@ +/* + * 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" + +#if defined(ARG_MAX) +#define CONDITION_TO_USE(x, y) ISGREATER(x, y) +#elif defined(ARG_MIN) +#define CONDITION_TO_USE(x, y) ISLESS(x, y) +#else // !(defined(ARG_MAX) || defined(ARG_MIN)) +#error "Unsupported reduction operation!" +#endif // defined(ARG_MAX) + +#if defined(DATA_TYPE_OUTPUT) +#if defined(WIDTH) +#if defined(ARG_MIN) +#if defined(PREV_OUTPUT) +/** Find index minimum value of a vector + * + * @param[in] input Pointer to the first value. + * + * @return index of the vector. + */ +inline DATA_TYPE_OUTPUT arg_idx_min_prev_out(__global const DATA_TYPE *input, __global const DATA_TYPE_OUTPUT *prev_res, const int x_idx) +{ + int end_elem = (x_idx + 1) * 16; + if(end_elem > WIDTH) + { + end_elem = WIDTH - x_idx * 16; + } + DATA_TYPE_OUTPUT res = prev_res[0]; + for(int x_v = 1; x_v < end_elem; ++x_v) + { + res = select(res, prev_res[x_v], *(input + prev_res[x_v]) < * (input + res)); + } + return res; +} +#else // !defined(PREV_OUTPUT) +/** Find index minimum value of a vector + * + * @param[in] input Pointer to the first value. + * + * @return index of the vector. + */ +inline DATA_TYPE_OUTPUT arg_idx_min(__global const DATA_TYPE *input, const int x_idx) +{ +#if WIDTH < 16 + DATA_TYPE_OUTPUT res = 0; + for(DATA_TYPE_OUTPUT x_v = res + 1; x_v < WIDTH; ++x_v) + { + res = select(res, x_v, *(input + x_v) < * (input + res)); + } + return res; +#else // WIDTH >= 16 + int x_elem = x_idx * 16; + const int x_goback = select(0, 16 - WIDTH % 16, x_elem + 16 > WIDTH); + x_elem -= x_goback; + + VEC_DATA_TYPE(DATA_TYPE, 16) + in = vload16(0, input - x_goback); + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + res = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 }; + + VEC_DATA_TYPE(COND_DATA_TYPE, 8) + idx_sel = (in.s01234567 <= in.s89abcdef); + in.s01234567 = select(in.s89abcdef, in.s01234567, idx_sel); + res.s01234567 = select(res.s89abcdef, res.s01234567, CONVERT(idx_sel, int8)); + + idx_sel.s0123 = (in.s0123 < in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), VEC_DATA_TYPE(COND_DATA_TYPE, 4))); + in.s0123 = select(in.s4567, in.s0123, idx_sel.s0123); + res.s0123 = select(res.s4567, res.s0123, CONVERT(idx_sel.s0123, int4)); + + idx_sel.s01 = (in.s01 < in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), VEC_DATA_TYPE(COND_DATA_TYPE, 2))); + in.s01 = select(in.s23, in.s01, idx_sel.s01); + res.s01 = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2)); + + idx_sel.s0 = (in.s0 < in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), COND_DATA_TYPE)); + res.s0 = select(res.s1, res.s0, CONVERT(idx_sel.s0, int)); + + return res.s0 + x_elem; +#endif // WIDTH < 16 +} +#endif // defined(PREV_OUTPUT) +#endif // defined(ARG_MIN) +#if defined(ARG_MAX) +#if defined(PREV_OUTPUT) +/** Find index maximum value of a vector + * + * @param[in] input Pointer to the first value. + * + * @return index of the vector. + */ +inline DATA_TYPE_OUTPUT arg_idx_max_prev_out(__global const DATA_TYPE *input, __global const DATA_TYPE_OUTPUT *prev_res, const int x_idx) +{ + int end_elem = (x_idx + 1) * 16; + if(end_elem > WIDTH) + { + end_elem = WIDTH - x_idx * 16; + } + DATA_TYPE_OUTPUT res = prev_res[0]; + for(int x_v = 1; x_v < end_elem; ++x_v) + { + res = select(res, prev_res[x_v], *(input + prev_res[x_v]) > *(input + res)); + } + return res; +} +#else // !defined(PREV_OUTPUT) +/** Find index maximum value of a vector + * + * @param[in] input Pointer to the first value. + * + * @return index of the vector. + */ +inline DATA_TYPE_OUTPUT arg_idx_max(__global const DATA_TYPE *input, const int x_idx) +{ +#if WIDTH < 16 + DATA_TYPE_OUTPUT res = 0; + for(DATA_TYPE_OUTPUT x_v = res + 1; x_v < WIDTH; ++x_v) + { + res = select(res, x_v, *(input + x_v) > *(input + res)); + } + return res; +#else // WIDTH >= 16 + int x_elem = x_idx * 16; + const int x_goback = select(0, 16 - WIDTH % 16, x_elem + 16 > WIDTH); + x_elem -= x_goback; + + VEC_DATA_TYPE(DATA_TYPE, 16) + in = vload16(0, input - x_goback); + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + res = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 }; + + VEC_DATA_TYPE(COND_DATA_TYPE, 8) + idx_sel = (in.s01234567 >= in.s89abcdef); + in.s01234567 = select(in.s89abcdef, in.s01234567, idx_sel); + res.s01234567 = select(res.s89abcdef, res.s01234567, CONVERT(idx_sel, int8)); + + idx_sel.s0123 = (in.s0123 > in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), VEC_DATA_TYPE(COND_DATA_TYPE, 4))); + in.s0123 = select(in.s4567, in.s0123, idx_sel.s0123); + res.s0123 = select(res.s4567, res.s0123, CONVERT(idx_sel.s0123, int4)); + + idx_sel.s01 = (in.s01 > in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), VEC_DATA_TYPE(COND_DATA_TYPE, 2))); + in.s01 = select(in.s23, in.s01, idx_sel.s01); + res.s01 = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2)); + + idx_sel.s0 = (in.s0 > in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), COND_DATA_TYPE)); + res.s0 = select(res.s1, res.s0, CONVERT(idx_sel.s0, int)); + + return res.s0 + x_elem; +#endif // WIDTH < 16 +} +#endif // defined(PREV_OUTPUT) +#endif // defined(ARG_MAX) + +/** This kernel performs parallel reduction given an operation on x-axis. + * + * @note In case the results of previous stages are passed the flag PREV_OUTPUT has to be passed using -DPREV_OUTPUT + * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The data type of the output must be passed at compile time using -DDATA_TYPE_OUTPUT: e.g. -DDATA_TYPE_OUTPUT=uint + * @note The arg_max flag must be passed at compile time using -DARG_MAX if we want to compute the ArgMax + * @note The arg_min flag must be passed at compile time using -DARG_MIN if we want to compute the ArgMin + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: S32/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_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] prev_res_ptr (Optional) Pointer to previous results tensor. Supported data types: U32/S32 + * @param[in] prev_res_stride_x (Optional) Stride of the output tensor in X dimension (in bytes) + * @param[in] prev_res_step_x (Optional) prev_res_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] prev_res_stride_y (Optional) Stride of the output tensor in Y dimension (in bytes) + * @param[in] prev_res_step_y (Optional) prev_res_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] prev_res_offset_first_element_in_bytes (Optional) The offset of the first element in the previous results tensor + * @param[in] partial_res_ptr The local buffer to hold partial result values. Supported data types: U32/S32 + * @param[in] partial_res_stride_x Stride of the output tensor in X dimension (in bytes) + * @param[in] partial_res_step_x partial_res_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] partial_res_stride_y Stride of the output tensor in Y dimension (in bytes) + * @param[in] partial_res_step_y partial_res_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] partial_res_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] local_results Local buffer for storing the partial result + */ +__kernel void arg_min_max_x( + IMAGE_DECLARATION(src), +#if defined(PREV_OUTPUT) + IMAGE_DECLARATION(prev_res), +#endif // defined(PREV_OUTPUT) + IMAGE_DECLARATION(partial_res), + __local DATA_TYPE_OUTPUT *local_results) +{ +#if defined(PREV_OUTPUT) + Image src = CONVERT_TO_IMAGE_STRUCT_NO_STEP(src); + Image prev_res = CONVERT_TO_IMAGE_STRUCT(prev_res); +#else // !defined(PREV_OUTPUT) + Image src = CONVERT_TO_IMAGE_STRUCT(src); +#endif // defined(PREV_OUTPUT) + Image partial_res = CONVERT_TO_IMAGE_STRUCT(partial_res); + + unsigned int lsize = get_local_size(0); + unsigned int lid = get_local_id(0); + + const uint x_idx = get_global_id(0); + const uint y_idx = get_global_id(1); + const __global DATA_TYPE *src_in_row = (const __global DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + y_idx * src_step_y); + + for(unsigned int y = 0; y < get_local_size(1); ++y) + { +#if defined(ARG_MAX) +#if defined(PREV_OUTPUT) + local_results[lid] = arg_idx_max_prev_out(src_in_row, (__global DATA_TYPE_OUTPUT *)offset(&prev_res, 0, y), x_idx); +#else // !defined(PREV_OUTPUT) + local_results[lid] = arg_idx_max((__global DATA_TYPE *)offset(&src, 0, y), x_idx); +#endif // defined(PREV_OUTPUT) +#else // defined(ARG_MIN) +#if defined(PREV_OUTPUT) + local_results[lid] = arg_idx_min_prev_out(src_in_row, (__global DATA_TYPE_OUTPUT *)offset(&prev_res, 0, y), x_idx); +#else // !defined(PREV_OUTPUT) + local_results[lid] = arg_idx_min((__global DATA_TYPE *)offset(&src, 0, y), x_idx); +#endif // defined(PREV_OUTPUT) +#endif // defined(ARG_MAX) || defined(ARG_MIN) + + barrier(CLK_LOCAL_MEM_FENCE); + + // Perform parallel reduction + for(unsigned int i = lsize >> 1; i > 0; i >>= 1) + { + if(lid < i) + { + DATA_TYPE tmp0 = *(src_in_row + local_results[lid]); + DATA_TYPE tmp1 = *(src_in_row + local_results[lid + i]); +#if defined(ARG_MAX) + local_results[lid] = select( + local_results[lid], + local_results[lid + i], + ((tmp0 == tmp1) && (local_results[lid + i] < local_results[lid])) || (tmp0 < tmp1)); +#else // defined(ARG_MIN) + local_results[lid] = select( + local_results[lid], + local_results[lid + i], + ((tmp0 == tmp1) && (local_results[lid + i] < local_results[lid])) || (tmp0 > tmp1)); +#endif // defined(ARG_MAX) || defined(ARG_MIN) + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + if(lid == 0) + { + ((__global DATA_TYPE_OUTPUT *)offset(&partial_res, get_group_id(0), y))[0] = local_results[0]; + } + } +} +#endif // defined(WIDTH) + +#if defined(HEIGHT) +/** This kernel performs reduction on y-axis. + * + * @note The input data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The data type of the output must be passed at compile time using -DDATA_TYPE_OUTPUT: e.g. -DDATA_TYPE_OUTPUT=uint + * @note The data type of the intermediate results must be passed at compile time using -DDATA_TYPE_PROMOTED: e.g. -DDATA_TYPE_PROMOTED=uint + * @note The height size must be passed at compile time using -DHEIGHT e.g. -DHEIGHT=128 + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: S32/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_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] output_ptr The local buffer to hold sumed values. Supported data types: U32/S32 + * @param[in] output_stride_x Stride of the output tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the output tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor + */ +__kernel void arg_min_max_y( + IMAGE_DECLARATION(src), + IMAGE_DECLARATION(output)) +{ + Image src = CONVERT_TO_IMAGE_STRUCT(src); + Image output = CONVERT_TO_IMAGE_STRUCT(output); + + VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) + res = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); + + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + indx = 0; + for(unsigned int y = 1; y < HEIGHT; ++y) + { + VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) + in = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, y)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); + + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)); + indx = select(indx, y, cond_conv); + res = select(res, in, CONDITION_TO_USE(in, res)); + } + + // Store result + vstore16(indx, 0, (__global DATA_TYPE_OUTPUT *)output.ptr); +} +#endif // defined(HEIGHT) + +#if defined(DEPTH) +/** This kernel performs reduction on z-axis. + * + * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The data type of the intermediate results must be passed at compile time using -DDATA_TYPE_PROMOTED: e.g. -DDATA_TYPE_PROMOTED=uint + * @note The depth size must be passed at compile time using -DDEPTH e.g. -DDEPTH=128 + * + * @param[in] input_ptr Pointer to the source tensor. Supported data types: S32/F16/F32 + * @param[in] input_stride_x Stride of the source tensor 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 tensor 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 tensor + * @param[in] output_ptr The local buffer to hold sumed values. Supported data types: U32/S32 + * @param[in] output_stride_x Stride of the output tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the output tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the output tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor + */ +__kernel void arg_min_max_z( + TENSOR3D_DECLARATION(input), + TENSOR3D_DECLARATION(output)) +{ + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); + + VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) + res = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); + + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + indx = 0; + for(DATA_TYPE_OUTPUT z = 1; z < DEPTH; ++z) + { + VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) + in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, z)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); + + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)); + indx = select(indx, z, cond_conv); + res = select(res, in, CONDITION_TO_USE(in, res)); + } + + // Store result + vstore16(indx, 0, (__global DATA_TYPE_OUTPUT *)output.ptr); +} +#endif /* defined(DEPTH) */ + +#if defined(BATCH) && defined(DEPTH) +/** This kernel performs reduction on w-axis. + * + * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The data type of the intermediate results must be passed at compile time using -DDATA_TYPE_PROMOTED: e.g. -DDATA_TYPE_PROMOTED=uint + * @note The batch size must be passed at compile time using -DBATCH e.g. -DBATCH=128 + * @note The depth size must be passed at compile time using -DBATCH e.g. -DDEPTH=128 + * + * @param[in] input_ptr Pointer to the source tensor. Supported data types: S32/F16/F32 + * @param[in] input_stride_x Stride of the source tensor 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 tensor 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_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] output_ptr The local buffer to hold sumed values. Supported data types: U32/S32 + * @param[in] output_stride_x Stride of the output tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the output tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the output tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_stride_w Stride of the output tensor in W dimension (in bytes) + * @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor + */ +__kernel void arg_min_max_w( + TENSOR4D_DECLARATION(input), + TENSOR4D_DECLARATION(output)) +{ + Tensor4D input = CONVERT_TO_TENSOR4D_STRUCT(input, DEPTH); + Tensor4D output = CONVERT_TO_TENSOR4D_STRUCT(output, DEPTH); + + VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) + res = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); + + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + indx = 0; + for(DATA_TYPE_OUTPUT w = 1; w < BATCH; ++w) + { + VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) + in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, w)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); + + VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) + cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)); + indx = select(indx, w, cond_conv); + res = select(res, in, CONDITION_TO_USE(in, res)); + } + + // Store result + vstore16(indx, 0, (__global DATA_TYPE_OUTPUT *)output.ptr); +} +#endif /* defined(BATCH) && defined(DEPTH) */ +#endif // defined(DATA_TYPE_OUTPUT)
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h index eaeaa6034d..ec5701dc69 100644 --- a/src/core/CL/cl_kernels/helpers.h +++ b/src/core/CL/cl_kernels/helpers.h @@ -266,6 +266,19 @@ #define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x))) #define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round) +#if FLOAT_DATA_TYPE +#define ISGREATER(x, y) isgreater(x, y) +#define ISLESS(x, y) isless(x, y) +#else // !FLOAT_DATA_TYPE +#if defined(WIDTH) +#define ISGREATER(x, y) (x > y) ? 1 : 0 +#define ISLESS(x, y) (x < y) ? 1 : 0 +#else // !defined(WIDTH) +#define ISGREATER(x, y) select((int16)0, (int16)-1, x > y) +#define ISLESS(x, y) select((int16)0, (int16)-1, x < y) +#endif // defined(WIDTH) +#endif // FLOAT_DATA_TYPE + #define VECTOR_DECLARATION(name) \ __global uchar *name##_ptr, \ uint name##_stride_x, \ diff --git a/src/core/CL/cl_kernels/reduction_operation.cl b/src/core/CL/cl_kernels/reduction_operation.cl index 5a4bb9ff4c..451b962b01 100644 --- a/src/core/CL/cl_kernels/reduction_operation.cl +++ b/src/core/CL/cl_kernels/reduction_operation.cl @@ -23,19 +23,6 @@ */ #include "helpers.h" -#if FLOAT_DATA_TYPE -#define ISGREATER(x, y) isgreater(x, y) -#define ISLESS(x, y) isless(x, y) -#else // !FLOAT_DATA_TYPE -#if defined(WIDTH) -#define ISGREATER(x, y) (x > y) ? 1 : 0 -#define ISLESS(x, y) (x < y) ? 1 : 0 -#else // !defined(WIDTH) -#define ISGREATER(x, y) select((int16)0, (int16)-1, x > y) -#define ISLESS(x, y) select((int16)0, (int16)-1, x < y) -#endif // defined(WIDTH) -#endif // FLOAT_DATA_TYPE - /** Calculate square sum of a vector * * @param[in] input Pointer to the first pixel. @@ -164,7 +151,7 @@ __kernel void reduction_operation_x( * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float * @note The width size must be passed at compile time using -DWIDTH e.g. -DWIDTH=128 * @note The product flag must be passed at compile time using -DPROD if we want to compute the product, otherwise sum will be used - * @note In case of ARG_MIN and ARG_MAX the condition data type must be passed at compile time using -DCOND_DATA_TYPE e.g. -DCOND_DATA_TYPE=short + * @note In case of MIN and MAX the condition data type must be passed at compile time using -DCOND_DATA_TYPE e.g. -DCOND_DATA_TYPE=short * * @param[in] src_ptr Pointer to the source tensor. Supported data types: S32/F16/F32 and QASYMM8 for operation MEAN * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) @@ -184,32 +171,19 @@ __kernel void reduction_operation_non_parallel_x( DATA_TYPE_PROMOTED res = *((__global DATA_TYPE *)vector_offset(&src, 0)); -#if defined(ARG_MAX) || defined(ARG_MIN) - uint indx = 0; -#endif // defined(ARG_MAX) || defined(ARG_MIN) - for(unsigned int x = 1; x < WIDTH; ++x) { DATA_TYPE_PROMOTED in = *((__global DATA_TYPE *)vector_offset(&src, x)); -#if defined(ARG_MAX) - indx = select(indx, x, ISGREATER(in, res)); - res = select(res, in, CONVERT(ISGREATER(in, res), COND_DATA_TYPE)); -#elif defined(ARG_MIN) - indx = select(indx, x, ISLESS(in, res)); - res = select(res, in, CONVERT(ISLESS(in, res), COND_DATA_TYPE)); -#elif defined(MIN) +#if defined(MIN) res = select(res, in, CONVERT(ISLESS(in, res), COND_DATA_TYPE)); #elif defined(MAX) - res = select(res, in, CONVERT(ISGREATER(in, res), COND_DATA_TYPE)); -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) + res = select(res, in, CONVERT(ISGREATER(in, res), COND_DATA_TYPE)); +#else // !(defined(MAX) || defined(MIN)) res += in; -#endif // defined(ARG_MAX) || defined(ARG_MIN) +#endif // defined(MAX) || defined(MIN) } // Store result -#if defined(ARG_MAX) || defined(ARG_MIN) - *((__global uint *)output.ptr) = indx; -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) #if defined(MEAN) res /= WIDTH; #endif // defined(MEAN) @@ -218,7 +192,6 @@ __kernel void reduction_operation_non_parallel_x( #else // defined(MIN) || defined(MAX) *((__global uchar *)output.ptr) = convert_uchar(res); #endif // defined(MIN) || defined(MAX) -#endif // defined(ARG_MAX) || defined(ARG_MIN) } #endif // defined(WIDTH) @@ -255,27 +228,15 @@ __kernel void reduction_operation_y( res *= res; #endif // defined(SUM_SQUARE) -#if defined(ARG_MAX) || defined(ARG_MIN) - uint16 indx = 0; -#endif // defined(ARG_MAX) || defined(ARG_MIN) - for(unsigned int y = 1; y < HEIGHT; ++y) { VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) in = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, y)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); -#if defined(ARG_MAX) - uint16 cond_conv = CONVERT(ISGREATER(in, res), uint16); - indx = select(indx, y, cond_conv); - res = select(res, in, ISGREATER(in, res)); -#elif defined(ARG_MIN) - uint16 cond_conv = CONVERT(ISLESS(in, res), uint16); - indx = select(indx, y, cond_conv); - res = select(res, in, ISLESS(in, res)); -#elif defined(MIN) +#if defined(MIN) res = select(res, in, ISLESS(in, res)); #elif defined(MAX) - res = select(res, in, ISGREATER(in, res)); -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) + res = select(res, in, ISGREATER(in, res)); +#else // !(defined(MAX) || defined(MIN)) #if defined(SUM_SQUARE) in *= in; #endif // defined(SUM_SQUARE) @@ -284,18 +245,14 @@ __kernel void reduction_operation_y( #else // !defined(PROD) res += in; #endif // defined(PROD) -#endif // defined(ARG_MAX) || defined(ARG_MIN) +#endif // defined(MAX) || defined(MIN) } // Store result -#if defined(ARG_MAX) || defined(ARG_MIN) - vstore16(indx, 0, (__global uint *)output.ptr); -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) #if defined(MEAN) res /= HEIGHT; #endif // defined(MEAN) vstore16(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr); -#endif // defined(ARG_MAX) || defined(ARG_MIN) } #endif // defined(HEIGHT) @@ -340,10 +297,6 @@ __kernel void reduction_operation_z( res *= res; #endif // defined(SUM_SQUARE) -#if defined(ARG_MAX) || defined(ARG_MIN) - uint16 indx = 0; -#endif // defined(ARG_MAX) || defined(ARG_MIN) - for(unsigned int z = 1; z < DEPTH; ++z) { VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) @@ -354,19 +307,11 @@ __kernel void reduction_operation_z( in1 = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 8, 0, z)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); #endif // defined(COMPLEX) -#if defined(ARG_MAX) - uint16 cond_conv = CONVERT(ISGREATER(in, res), uint16); - indx = select(indx, z, cond_conv); - res = select(res, in, ISGREATER(in, res)); -#elif defined(ARG_MIN) - uint16 cond_conv = CONVERT(ISLESS(in, res), uint16); - indx = select(indx, z, cond_conv); - res = select(res, in, ISLESS(in, res)); -#elif defined(MIN) +#if defined(MIN) res = select(res, in, ISLESS(in, res)); #elif defined(MAX) - res = select(res, in, ISGREATER(in, res)); -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) + res = select(res, in, ISGREATER(in, res)); +#else // !(defined(MAX) || defined(MIN)) #if defined(SUM_SQUARE) in *= in; #endif // defined(SUM_SQUARE) @@ -377,14 +322,11 @@ __kernel void reduction_operation_z( #if defined(COMPLEX) res1 += in1; #endif // defined(COMPLEX) -#endif //defined(PROD) -#endif // defined(ARG_MAX) || defined(ARG_MIN) +#endif // defined(PROD) +#endif // defined(MAX) || defined(MIN) } // Store result -#if defined(ARG_MAX) || defined(ARG_MIN) - vstore16(indx, 0, (__global uint *)output.ptr); -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) #if defined(MEAN) res /= DEPTH; #endif // defined(MEAN) @@ -392,7 +334,6 @@ __kernel void reduction_operation_z( #if defined(COMPLEX) vstore16(CONVERT(res1, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)tensor3D_offset(&output, 8, 0, 0)); #endif // defined(COMPLEX) -#endif // defined(ARG_MAX) || defined(ARG_MIN) } #endif /* defined(DEPTH) */ @@ -438,28 +379,16 @@ __kernel void reduction_operation_w( res *= res; #endif // defined(SUM_SQUARE) -#if defined(ARG_MAX) || defined(ARG_MIN) - uint16 indx = 0; -#endif // defined(ARG_MAX) || defined(ARG_MIN) - for(unsigned int w = 1; w < BATCH; ++w) { VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, w)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); -#if defined(ARG_MAX) - uint16 cond_conv = CONVERT(ISGREATER(in, res), uint16); - indx = select(indx, w, cond_conv); - res = select(res, in, ISGREATER(in, res)); -#elif defined(ARG_MIN) - uint16 cond_conv = CONVERT(ISLESS(in, res), uint16); - indx = select(indx, w, cond_conv); - res = select(res, in, ISLESS(in, res)); -#elif defined(MIN) +#if defined(MIN) res = select(res, in, ISLESS(in, res)); #elif defined(MAX) - res = select(res, in, ISGREATER(in, res)); -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) + res = select(res, in, ISGREATER(in, res)); +#else // !(defined(MAX) || defined(MIN)) #if defined(SUM_SQUARE) in *= in; #endif // defined(SUM_SQUARE) @@ -468,17 +397,13 @@ __kernel void reduction_operation_w( #else //!defined(PROD) res += in; #endif //defined(PROD) -#endif // defined(ARG_MAX) || defined(ARG_MIN) +#endif // defined(MAX) || defined(MIN) } // Store result -#if defined(ARG_MAX) || defined(ARG_MIN) - vstore16(indx, 0, (__global uint *)output.ptr); -#else // !(defined(ARG_MAX) || defined(ARG_MIN)) #if defined(MEAN) res /= BATCH; #endif // defined(MEAN) vstore16(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr); -#endif // defined(ARG_MAX) || defined(ARG_MIN) } #endif /* defined(BATCH) && defined(DEPTH) */ |