diff options
author | Giorgio Arena <giorgio.arena@arm.com> | 2021-08-19 13:04:56 +0100 |
---|---|---|
committer | Giorgio Arena <giorgio.arena@arm.com> | 2021-08-23 13:33:31 +0000 |
commit | 511771fbe0a74e6d9dfd37ba9b4926a8315ec7aa (patch) | |
tree | e2296560f37935232d482982d025e148f1c6c61e /src/core/CL/cl_kernels/nhwc | |
parent | 19884630c37ae9de2f65a88ea2cda5630a551bad (diff) | |
download | ComputeLibrary-511771fbe0a74e6d9dfd37ba9b4926a8315ec7aa.tar.gz |
Remove padding from ClScaleKernel
- Merge quantized kernels with fp for bilinear interpolation (both NCHW and NHWC)
- Pass dimensions at compile time rather than at run time
- Use tile-based approach to rework the NCHW kernels
- Remove unused functions/files
Resolve COMPMID-4723
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Change-Id: Ifcdf02beb9daa9f318395751b3c85eb2fe874082
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6138
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels/nhwc')
-rw-r--r-- | src/core/CL/cl_kernels/nhwc/scale.cl | 109 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/nhwc/scale_quantized.cl | 124 |
2 files changed, 55 insertions, 178 deletions
diff --git a/src/core/CL/cl_kernels/nhwc/scale.cl b/src/core/CL/cl_kernels/nhwc/scale.cl index 1ea5e73df1..69cbbcd5f3 100644 --- a/src/core/CL/cl_kernels/nhwc/scale.cl +++ b/src/core/CL/cl_kernels/nhwc/scale.cl @@ -22,7 +22,6 @@ * SOFTWARE. */ #include "helpers.h" -#include "warp_helpers.h" #if defined(DEPTH_OUT) /** Performs scale on an image interpolating with the NEAREAST NEIGHBOUR method. Input and output are single channel F32. (NHWC) @@ -46,28 +45,20 @@ * @param[in] out_stride_z Stride of the destination image in Z dimension (in bytes) * @param[in] out_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image - * @param[in] input_width Input image width - * @param[in] input_height Input image height - * @param[in] scale_x The scale factor along x dimension - * @param[in] scale_y The scale factor along y dimension */ __kernel void scale_nearest_neighbour_nhwc( TENSOR4D_DECLARATION(in), - TENSOR4D_DECLARATION(out), - const float input_width, - const float input_height, - const float scale_x, - const float scale_y) + TENSOR4D_DECLARATION(out)) { Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); #ifdef SAMPLING_POLICY_TOP_LEFT - float new_x = get_global_id(1) * scale_x; - float new_y = (get_global_id(2) % DEPTH_OUT) * scale_y; + float new_x = get_global_id(1) * SCALE_X; + float new_y = (get_global_id(2) % DEPTH_OUT) * SCALE_Y; #elif SAMPLING_POLICY_CENTER - float new_x = (get_global_id(1) + 0.5f) * scale_x; - float new_y = ((get_global_id(2) % DEPTH_OUT) + 0.5f) * scale_y; + float new_x = (get_global_id(1) + 0.5f) * SCALE_X; + float new_y = ((get_global_id(2) % DEPTH_OUT) + 0.5f) * SCALE_Y; #else /* SAMPLING_POLICY */ #error("Unsupported sampling policy"); #endif /* SAMPLING_POLICY */ @@ -75,8 +66,8 @@ __kernel void scale_nearest_neighbour_nhwc( new_x = round(new_x); new_y = round(new_y); #endif /* ALIGN_CORNERS */ - const float clamped_x = clamp(new_x, 0.0f, input_width - 1); - const float clamped_y = clamp(new_y, 0.0f, input_height - 1); + const float clamped_x = clamp(new_x, 0.0f, (float)SRC_WIDTH - 1); + const float clamped_y = clamp(new_y, 0.0f, (float)SRC_HEIGHT - 1); *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))); } @@ -104,71 +95,81 @@ __kernel void scale_nearest_neighbour_nhwc( * @param[in] out_stride_z Stride of the destination image in Z dimension (in bytes) * @param[in] out_step_z dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image - * @param[in] input_width Input image width - * @param[in] input_height Input image height - * @param[in] scale_x The scale factor along x dimension - * @param[in] scale_y The scale factor along y dimension * */ __kernel void scale_bilinear_nhwc( TENSOR4D_DECLARATION(in), - TENSOR4D_DECLARATION(out), - const float input_width, - const float input_height, - const float scale_x, - const float scale_y) + TENSOR4D_DECLARATION(out)) { Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); #ifdef SAMPLING_POLICY_TOP_LEFT - const float new_x = get_global_id(1) * scale_x; - const float new_y = (get_global_id(2) % DEPTH_OUT) * scale_y; + const float new_x = get_global_id(1) * SCALE_X; + const float new_y = (get_global_id(2) % DEPTH_OUT) * SCALE_Y; #elif SAMPLING_POLICY_CENTER - const float new_x = (get_global_id(1) + 0.5f) * scale_x - 0.5f; - const float new_y = ((get_global_id(2) % DEPTH_OUT) + 0.5f) * scale_y - 0.5f; + const float new_x = (get_global_id(1) + 0.5f) * SCALE_X - 0.5f; + const float new_y = ((get_global_id(2) % DEPTH_OUT) + 0.5f) * SCALE_Y - 0.5f; #else /* SAMPLING_POLICY */ #error("Unsupported sampling policy"); #endif /* SAMPLING_POLICY */ const float new_xf = floor(new_x); const float new_yf = floor(new_y); - const float clamped_x = clamp(new_xf, 0.0f, input_width - 1); - const float clamped_x1 = clamp(new_xf + 1, 0.0f, input_width - 1); - const float clamped_y = clamp(new_yf, 0.0f, input_height - 1); - const float clamped_y1 = clamp(new_yf + 1, 0.0f, input_height - 1); + const float clamped_x = clamp(new_xf, 0.0f, SRC_WIDTH - 1.f); + const float clamped_x1 = clamp(new_xf + 1, 0.0f, SRC_WIDTH - 1.f); + const float clamped_y = clamp(new_yf, 0.0f, SRC_HEIGHT - 1.f); + const float clamped_y1 = clamp(new_yf + 1, 0.0f, SRC_HEIGHT - 1.f); + +#if defined(OFFSET) && defined(SCALE) +#define IN_DATA_TYPE int +#else // defined(OFFSET) && defined(SCALE) +#define IN_DATA_TYPE float +#endif // defined(OFFSET) && defined(SCALE) #ifndef BORDER_MODE_REPLICATE - const bool check_x = (0.f <= new_xf && new_xf < input_width); - const bool check_x1 = (-1.f <= new_xf && new_xf < input_width - 1); - const bool check_y = (0.f <= new_yf && new_yf < input_height); - const bool check_y1 = (-1.f <= new_yf && new_yf < input_height - 1); - const float ins_0 = select((float)(CONSTANT_VALUE), (float)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), - (get_global_id(2) / DEPTH_OUT)))), - check_x && check_y); - const float ins_1 = select((float)(CONSTANT_VALUE), (float)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), - (get_global_id(2) / DEPTH_OUT)))), - check_x1 && check_y); - const float ins_2 = select((float)(CONSTANT_VALUE), (float)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), - (get_global_id(2) / DEPTH_OUT)))), - check_x && check_y1); - const float ins_3 = select((float)(CONSTANT_VALUE), (float)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), - (get_global_id(2) / DEPTH_OUT)))), - check_x1 && check_y1); - float4 ins = (float4)(ins_0, ins_1, ins_2, ins_3); + const bool check_x = (0.f <= new_xf && new_xf < (float)SRC_WIDTH); + const bool check_x1 = (-1.f <= new_xf && new_xf < SRC_WIDTH - 1.f); + const bool check_y = (0.f <= new_yf && new_yf < (float)SRC_HEIGHT); + const bool check_y1 = (-1.f <= new_yf && new_yf < SRC_HEIGHT - 1.f); + + const IN_DATA_TYPE ins_0 = select((IN_DATA_TYPE)(CONSTANT_VALUE), (IN_DATA_TYPE)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), + (get_global_id(2) / DEPTH_OUT)))), + check_x && check_y); + const IN_DATA_TYPE ins_1 = select((IN_DATA_TYPE)(CONSTANT_VALUE), (IN_DATA_TYPE)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), + (get_global_id(2) / DEPTH_OUT)))), + check_x1 && check_y); + const IN_DATA_TYPE ins_2 = select((IN_DATA_TYPE)(CONSTANT_VALUE), (IN_DATA_TYPE)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), + (get_global_id(2) / DEPTH_OUT)))), + check_x && check_y1); + const IN_DATA_TYPE ins_3 = select((IN_DATA_TYPE)(CONSTANT_VALUE), (IN_DATA_TYPE)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), + (get_global_id(2) / DEPTH_OUT)))), + check_x1 && check_y1); + VEC_DATA_TYPE(IN_DATA_TYPE, 4) + ins = (VEC_DATA_TYPE(IN_DATA_TYPE, 4))(ins_0, ins_1, ins_2, ins_3); #else /* BORDER_MODE_REPLICATE */ - float4 ins = (float4)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT)))); + VEC_DATA_TYPE(IN_DATA_TYPE, 4) + ins = (VEC_DATA_TYPE(IN_DATA_TYPE, 4))(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), + *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), + *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT))), + *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT)))); #endif /* BORDER_MODE_REPLICATE */ const float a = new_x - new_xf; const float b = 1.f - a; const float a1 = new_y - new_yf; const float b1 = 1.f - a1; + +#if defined(OFFSET) && defined(SCALE) + const float4 insf32 = convert_float4(ins - (int4)OFFSET) * (float4)SCALE; + const float fr = ((insf32.s0 * b * b1) + (insf32.s1 * a * b1) + (insf32.s2 * b * a1) + (insf32.s3 * a * a1)); + DATA_TYPE res = CONVERT_SAT(convert_int_sat_rtp(fr / SCALE) + OFFSET, DATA_TYPE); + + *((__global DATA_TYPE *)out.ptr) = res; +#else // defined(OFFSET) && defined(SCALE) const float fr = ((ins.s0 * b * b1) + (ins.s1 * a * b1) + (ins.s2 * b * a1) + (ins.s3 * a * a1)); *((__global DATA_TYPE *)out.ptr) = CONVERT(fr, DATA_TYPE); +#endif // defined(OFFSET) && defined(SCALE) } #endif /* defined(DEPTH_OUT) */
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nhwc/scale_quantized.cl b/src/core/CL/cl_kernels/nhwc/scale_quantized.cl deleted file mode 100644 index de9bb607b0..0000000000 --- a/src/core/CL/cl_kernels/nhwc/scale_quantized.cl +++ /dev/null @@ -1,124 +0,0 @@ -/* - * Copyright (c) 2018-2021 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_asymm.h" -#include "warp_helpers_quantized.h" - -#if defined(DEPTH_OUT) -/** Performs scale on an image interpolating with the BILINEAR method. (NHWC) - * - * @note Sampling policy to be used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT - * @note Scale value for QASYMM8 data type to used is passed as -DSCALE=<VALUE> e.g. -DSCALE=0.5 - * @note Offset value for QASYMM8 data type to used is passed as -DOFFSET=<VALUE> e.g. -DOFFSET=1 - * @note If border mode replicate is used, is should be passed as -DBORDER_MODE_REPLICATE - * @note Output tensor's depth should be given as a preprocessor argument using -DDEPTH_OUT=size. e.g. -DDEPTH=16 - * @note The value to be used at the edges of the images shoud be given as a preprocessor argument using -DCONSTANT_VALUE=value. - * - * @param[in] in_ptr Pointer to the source image. Supported data types: QASYMM8. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] in_stride_z Stride of the source image in Z dimension (in bytes) - * @param[in] in_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: same as @p in_ptr - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] out_stride_z Stride of the destination image in Z dimension (in bytes) - * @param[in] out_step_z dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image - * @param[in] input_width Input image width - * @param[in] input_height Input image height - * @param[in] scale_x The scale factor along x dimension - * @param[in] scale_y The scale factor along y dimension - * @param[in] constant_border_value Constant border value to use - */ -__kernel void scale_bilinear_quantized_nhwc( - TENSOR4D_DECLARATION(in), - TENSOR4D_DECLARATION(out), - const float input_width, - const float input_height, - const float scale_x, - const float scale_y) -{ - Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); - Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); - -#ifdef SAMPLING_POLICY_TOP_LEFT - const float new_x = get_global_id(1) * scale_x; - const float new_y = (get_global_id(2) % DEPTH_OUT) * scale_y; -#elif SAMPLING_POLICY_CENTER - const float new_x = (get_global_id(1) + 0.5f) * scale_x - 0.5f; - const float new_y = ((get_global_id(2) % DEPTH_OUT) + 0.5f) * scale_y - 0.5f; -#else /* SAMPLING_POLICY */ -#error("Unsupported sampling policy"); -#endif /* SAMPLING_POLICY */ - - const float new_xf = floor(new_x); - const float new_yf = floor(new_y); - const float clamped_x = clamp(new_xf, 0.0f, input_width - 1); - const float clamped_x1 = clamp(new_xf + 1, 0.0f, input_width - 1); - const float clamped_y = clamp(new_yf, 0.0f, input_height - 1); - const float clamped_y1 = clamp(new_yf + 1, 0.0f, input_height - 1); - -#ifndef BORDER_MODE_REPLICATE - const bool check_x = (0.f <= new_xf && new_xf < input_width); - const bool check_x1 = (-1.f <= new_xf && new_xf < input_width - 1); - const bool check_y = (0.f <= new_yf && new_yf < input_height); - const bool check_y1 = (-1.f <= new_yf && new_yf < input_height - 1); - const int ins_0 = select((int)(CONSTANT_VALUE), (int)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), - (get_global_id(2) / DEPTH_OUT)))), - check_x && check_y); - const int ins_1 = select((int)(CONSTANT_VALUE), (int)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), - (get_global_id(2) / DEPTH_OUT)))), - check_x1 && check_y); - const int ins_2 = select((int)(CONSTANT_VALUE), (int)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), - (get_global_id(2) / DEPTH_OUT)))), - check_x && check_y1); - const int ins_3 = select((int)(CONSTANT_VALUE), (int)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), - (get_global_id(2) / DEPTH_OUT)))), - check_x1 && check_y1); - int4 ins = (int4)(ins_0, ins_1, ins_2, ins_3); -#else /* BORDER_MODE_REPLICATE */ - int4 ins = (int4)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT)))); -#endif /* BORDER_MODE_REPLICATE */ - - const float a = new_x - new_xf; - const float b = 1.f - a; - const float a1 = new_y - new_yf; - const float b1 = 1.f - a1; - const float4 insf32 = convert_float4(ins - (int4)OFFSET) * (float4)SCALE; - - const float fr = ((insf32.s0 * b * b1) + (insf32.s1 * a * b1) + (insf32.s2 * b * a1) + (insf32.s3 * a * a1)); - - DATA_TYPE res = CONVERT_SAT(convert_int_sat_rtp(fr / SCALE) + OFFSET, DATA_TYPE); - - *((__global DATA_TYPE *)out.ptr) = res; -} -#endif /* defined(DEPTH_OUT) */
\ No newline at end of file |