From 7317e393de848cb20dd326a45a9efb18a3ee7701 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Fri, 17 Jan 2020 11:27:49 +0000 Subject: COMPMID-2941 Add QAsymm8 and S32 support for ArgMinMax to CL Select operation does not work with any combination of datatypes. Added a new OpenCL build option, DATA_TYPE_SELECT, in oder to specify the select result. Change-Id: I5c73c2a6b1d121288b7f6edba3e0cb1104636844 Signed-off-by: Michalis Spyrou Reviewed-on: https://review.mlplatform.org/c/2601 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio --- src/core/CL/cl_kernels/arg_min_max.cl | 65 +++++++++++++++++++++-------------- 1 file changed, 39 insertions(+), 26 deletions(-) (limited to 'src/core/CL/cl_kernels/arg_min_max.cl') diff --git a/src/core/CL/cl_kernels/arg_min_max.cl b/src/core/CL/cl_kernels/arg_min_max.cl index 3f75377636..06dcc8ddde 100644 --- a/src/core/CL/cl_kernels/arg_min_max.cl +++ b/src/core/CL/cl_kernels/arg_min_max.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 ARM Limited. + * Copyright (c) 2019-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,6 +23,19 @@ */ #include "helpers.h" +#if defined(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((VEC_DATA_TYPE(DATA_TYPE_SELECT, 16))0, (VEC_DATA_TYPE(DATA_TYPE_SELECT, 16)) - 1, x > y) +#define ISLESS(x, y) select((VEC_DATA_TYPE(DATA_TYPE_SELECT, 16))0, (VEC_DATA_TYPE(DATA_TYPE_SELECT, 16)) - 1, x < y) +#endif // defined(WIDTH) +#endif // defined(FLOAT_DATA_TYPE) + #if defined(ARG_MAX) #define CONDITION_TO_USE(x, y) ISGREATER(x, y) #elif defined(ARG_MIN) @@ -31,7 +44,7 @@ #error "Unsupported reduction operation!" #endif // defined(ARG_MAX) -#if defined(DATA_TYPE_OUTPUT) +#if defined(DATA_TYPE_OUTPUT) && defined(DATA_TYPE_SELECT) #if defined(WIDTH) #if defined(ARG_MIN) #if defined(PREV_OUTPUT) @@ -81,20 +94,20 @@ inline DATA_TYPE_OUTPUT arg_idx_min(__global const DATA_TYPE *input, const int x 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) + VEC_DATA_TYPE(DATA_TYPE_SELECT, 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))); + idx_sel.s0123 = (in.s0123 < in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), VEC_DATA_TYPE(DATA_TYPE_SELECT, 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))); + idx_sel.s01 = (in.s01 < in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), VEC_DATA_TYPE(DATA_TYPE_SELECT, 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)); + idx_sel.s0 = (in.s0 < in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), DATA_TYPE_SELECT)); res.s0 = select(res.s1, res.s0, CONVERT(idx_sel.s0, int)); return res.s0 + x_elem; @@ -150,20 +163,20 @@ inline DATA_TYPE_OUTPUT arg_idx_max(__global const DATA_TYPE *input, const int x 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) + VEC_DATA_TYPE(DATA_TYPE_SELECT, 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))); + idx_sel.s0123 = (in.s0123 > in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), VEC_DATA_TYPE(DATA_TYPE_SELECT, 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))); + idx_sel.s01 = (in.s01 > in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), VEC_DATA_TYPE(DATA_TYPE_SELECT, 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)); + idx_sel.s0 = (in.s0 > in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), DATA_TYPE_SELECT)); res.s0 = select(res.s1, res.s0, CONVERT(idx_sel.s0, int)); return res.s0 + x_elem; @@ -276,7 +289,7 @@ __kernel void arg_min_max_x( * * @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 data type of the select results must be passed at compile time using -DDATA_TYPE_SELECT: e.g. -DDATA_TYPE_SELECT=int * @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 @@ -299,15 +312,15 @@ __kernel void arg_min_max_y( 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, 16) + res = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, 0)), VEC_DATA_TYPE(DATA_TYPE, 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, 16) + in = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, y)), VEC_DATA_TYPE(DATA_TYPE, 16)); VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)); @@ -324,7 +337,7 @@ __kernel void arg_min_max_y( /** 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 data type of the select results must be passed at compile time using -DDATA_TYPE_SELECT: e.g. -DDATA_TYPE_SELECT=int * @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 @@ -351,15 +364,15 @@ __kernel void arg_min_max_z( 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, 16) + res = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)), VEC_DATA_TYPE(DATA_TYPE, 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, 16) + in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, z)), VEC_DATA_TYPE(DATA_TYPE, 16)); VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)); @@ -376,7 +389,7 @@ __kernel void arg_min_max_z( /** 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 data type of the select results must be passed at compile time using -DDATA_TYPE_SELECT: e.g. -DDATA_TYPE_SELECT=int * @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 * @@ -408,15 +421,15 @@ __kernel void arg_min_max_w( 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, 16) + res = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, 0)), VEC_DATA_TYPE(DATA_TYPE, 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, 16) + in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, w)), VEC_DATA_TYPE(DATA_TYPE, 16)); VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) cond_conv = CONVERT(CONDITION_TO_USE(in, res), VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)); @@ -428,4 +441,4 @@ __kernel void arg_min_max_w( 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 +#endif /* defined(DATA_TYPE_OUTPUT) && defined(DATA_TYPE_SELECT) */ \ No newline at end of file -- cgit v1.2.1