From aa1209a1bfc9fa24a24c1b47d309e27ba2cd90a7 Mon Sep 17 00:00:00 2001 From: zhenglin Date: Mon, 8 Jan 2018 13:54:52 +0800 Subject: APPBROWSER-375: Rewrite the transpose.cs with the new common code Change-Id: I373e349ac35ff52ebcc895723d8aa61b754519d4 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/115283 Tested-by: Jenkins Reviewed-by: Pablo Tello Reviewed-by: Joel Liang --- src/core/GLES_COMPUTE/cs_shaders/transpose.cs | 194 +++++++++----------------- 1 file changed, 68 insertions(+), 126 deletions(-) mode change 100644 => 100755 src/core/GLES_COMPUTE/cs_shaders/transpose.cs (limited to 'src/core/GLES_COMPUTE') diff --git a/src/core/GLES_COMPUTE/cs_shaders/transpose.cs b/src/core/GLES_COMPUTE/cs_shaders/transpose.cs old mode 100644 new mode 100755 index f8ad30312c..89bf9fbd89 --- a/src/core/GLES_COMPUTE/cs_shaders/transpose.cs +++ b/src/core/GLES_COMPUTE/cs_shaders/transpose.cs @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017, 2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -22,7 +22,12 @@ * SOFTWARE. */ layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in; -#include "helpers.h" + +#include "helpers_cs.h" + +#if defined(DATA_TYPE_FP16) +precision mediump float; +#endif // DATA_TYPE_FP16 #define SWAP_ROW_func(u0, l0) \ { \ @@ -59,144 +64,90 @@ layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = * @note The data type must be passed at compile time using "#define DATA_TYPE_NAME". e.g. "#define DATA_TYPE_FP32" * @note Optimization name must be passed using "#define OPTIMIZATION_NAME" for F16. e.g. "#define TRANSPOSE_8X8" * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32/F16 - * @param[in] src_stride_x Stride of the source matrix 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 matrix 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 matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix + * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32/F16 + * @param[in] src_attrs The attributes of the source matrix + * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as src_ptr + * @param[in] dst_attrs The attributes of the destination matrix */ - -layout(std140) uniform shader_params +SHADER_PARAMS_DECLARATION { - IMAGE_PARAM_DECLARATION(src); - IMAGE_PARAM_DECLARATION(dst); + ImageAttributes src_attrs; + ImageAttributes dst_attrs; }; #ifdef DATA_TYPE_FP32 -precision highp float; - -BUFFER_DECLARATION(src, 1, float, readonly); -BUFFER_DECLARATION(dst, 2, float, writeonly); - -#define LOAD16(r, name, offset) \ - { \ - r.x = LOAD4(name, offset); \ - r.y = LOAD4(name, offset + uint(1)); \ - r.z = LOAD4(name, offset + uint(2)); \ - r.w = LOAD4(name, offset + uint(3)); \ - } - -#define STORE16(name, offset, r) \ - { \ - STORE4(name, offset, r.x); \ - STORE4(name, offset + uint(1), r.y); \ - STORE4(name, offset + uint(2), r.z); \ - STORE4(name, offset + uint(3), r.w); \ - } +TENSOR_DECLARATION(1, srcBuffer, float, src_ptr, src_shift, 2, readonly); +TENSOR_DECLARATION(2, dstBuffer, float, dst_ptr, dst_shift, 2, writeonly); void main(void) { // compute source address - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); + ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); + ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(dst_attrs, dst_shift); // load the NxN block at (x, y) - vec4 u0; - vec4 u1; - vec4 u2; - vec4 u3; - LOAD16(u0, src, offset(src, 0, 0)); - LOAD16(u1, src, offset(src, 0, 1)); - LOAD16(u2, src, offset(src, 0, 2)); - LOAD16(u3, src, offset(src, 0, 3)); + vec4 u0 = VLOAD4(vec4, src_ptr, IMAGE_OFFSET(src_iter, 0, 0)); + vec4 u1 = VLOAD4(vec4, src_ptr, IMAGE_OFFSET(src_iter, 0, 1)); + vec4 u2 = VLOAD4(vec4, src_ptr, IMAGE_OFFSET(src_iter, 0, 2)); + vec4 u3 = VLOAD4(vec4, src_ptr, IMAGE_OFFSET(src_iter, 0, 3)); // transpose the block TRANSPOSE_4x4_func(u0, u1, u2, u3); // store the block at (y, x) - uint dst_offset_in_bytes = uint(16) * uint(gl_GlobalInvocationID.y) + uint(4) * uint(gl_GlobalInvocationID.x) * (dst.stride_y) + (dst.offset_first_element_in_bytes); + TENSOR_ITERATOR_ADVANCE_IN_BYTES(dst_iter, uint(16) * uint(gl_GlobalInvocationID.y) + uint(4) * uint(gl_GlobalInvocationID.x) * (dst_attrs.stride_y)); - STORE16(dst, uint((dst_offset_in_bytes + uint(0) * dst.stride_y) >> 2), u0); - STORE16(dst, uint((dst_offset_in_bytes + uint(1) * dst.stride_y) >> 2), u1); - STORE16(dst, uint((dst_offset_in_bytes + uint(2) * dst.stride_y) >> 2), u2); - STORE16(dst, uint((dst_offset_in_bytes + uint(3) * dst.stride_y) >> 2), u3); + VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 0), u0); + VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 1), u1); + VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 2), u2); + VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 3), u3); } -#elif defined(DATA_TYPE_FP16) /* DATA_TYPE_FP32 */ -precision mediump float; +#elif defined(DATA_TYPE_FP16) /* DATA_TYPE_FP16 */ #if defined(TRANSPOSE_4X4) - -BUFFER_DECLARATION(src, 1, uvec2, readonly); -BUFFER_DECLARATION(dst, 2, uvec2, writeonly); +TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly); +TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); void main(void) { // compute source address - Image src = GC_CONVERT_TO_IMAGE_STRUCT(src); - Image dst = GC_CONVERT_TO_IMAGE_STRUCT(dst); + ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); + ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(dst_attrs, dst_shift); // load the NxN block at (x, y) - vec4 u0; - vec4 u1; - vec4 u2; - vec4 u3; - uvec2 packed_s[4]; - - GC_LOAD1_2D_OFFSET(packed_s[0], src, 0, 0); - GC_LOAD1_2D_OFFSET(packed_s[1], src, 0, 1); - GC_LOAD1_2D_OFFSET(packed_s[2], src, 0, 2); - GC_LOAD1_2D_OFFSET(packed_s[3], src, 0, 3); - - u0 = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); - u1 = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); - u2 = vec4(unpackHalf2x16(packed_s[2].x), unpackHalf2x16(packed_s[2].y)); - u3 = vec4(unpackHalf2x16(packed_s[3].x), unpackHalf2x16(packed_s[3].y)); + vec4 u0 = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, 0)); + vec4 u1 = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, 1)); + vec4 u2 = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, 2)); + vec4 u3 = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, 3)); // transpose the block TRANSPOSE_4x4_func(u0, u1, u2, u3); // store the block at (y, x) - uint dst_offset_in_bytes = uint(8) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_step_y) + (dst.offset_first_element_in_bytes); - dst.current_offset = dst_offset_in_bytes; - - packed_s[0] = uvec2(packHalf2x16(u0.xy), packHalf2x16(u0.zw)); - packed_s[1] = uvec2(packHalf2x16(u1.xy), packHalf2x16(u1.zw)); - packed_s[2] = uvec2(packHalf2x16(u2.xy), packHalf2x16(u2.zw)); - packed_s[3] = uvec2(packHalf2x16(u3.xy), packHalf2x16(u3.zw)); - - GC_STORE1_2D_OFFSET(packed_s[0], dst, 0, 0); - GC_STORE1_2D_OFFSET(packed_s[1], dst, 0, 1); - GC_STORE1_2D_OFFSET(packed_s[2], dst, 0, 2); - GC_STORE1_2D_OFFSET(packed_s[3], dst, 0, 3); -} + TENSOR_ITERATOR_ADVANCE_IN_BYTES(dst_iter, uint(8) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_attrs.step_y)); -#elif defined(TRANSPOSE_8X8) /* TRANSPOSE_4X4 */ + STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 0), u0); + STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 1), u1); + STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 2), u2); + STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 3), u3); +} -BUFFER_DECLARATION(src, 1, uvec4, readonly); -BUFFER_DECLARATION(dst, 2, uvec4, writeonly); +#elif defined(TRANSPOSE_8X8) /* TRANSPOSE_8X8 */ +TENSOR_DECLARATION(1, srcBuffer, uvec4, src_ptr, src_shift, 4, readonly); +TENSOR_DECLARATION(2, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); void main(void) { // compute source address - Image src = GC_CONVERT_TO_IMAGE_STRUCT(src); - Image dst = GC_CONVERT_TO_IMAGE_STRUCT(dst); + ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); + ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(dst_attrs, dst_shift); - vec4 u[8][2]; - uvec4 packed_s[8]; + vec4 u[8][2]; for(int i = 0; i < 8; i++) { - GC_LOAD1_2D_OFFSET(packed_s[i], src, 0, i); - u[i][0] = vec4(unpackHalf2x16(packed_s[i].x), unpackHalf2x16(packed_s[i].y)); - u[i][1] = vec4(unpackHalf2x16(packed_s[i].z), unpackHalf2x16(packed_s[i].w)); + u[i] = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, i)); } // transpose the block @@ -207,41 +158,36 @@ void main(void) SWAP_4x4_func(u[0][1], u[1][1], u[2][1], u[3][1], u[4][0], u[5][0], u[6][0], u[7][0]); // store the block at (y, x) - uint dst_offset_in_bytes = uint(16) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_step_y) + (dst.offset_first_element_in_bytes); - dst.current_offset = dst_offset_in_bytes; + TENSOR_ITERATOR_ADVANCE_IN_BYTES(dst_iter, uint(16) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_attrs.step_y)); for(int i = 0; i < 8; i++) { - packed_s[i] = uvec4(packHalf2x16(u[i][0].xy), packHalf2x16(u[i][0].zw), packHalf2x16(u[i][1].xy), packHalf2x16(u[i][1].zw)); - GC_STORE1_2D_OFFSET(packed_s[i], dst, 0, i); + STORE_PACK8_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, i), u[i]); } } -#elif defined(TRANSPOSE_8X8_SQUARE) /* TRANSPOSE_4X4 */ - -BUFFER_DECLARATION(src, 1, uvec4, readonly); -BUFFER_DECLARATION(dst, 2, uvec4, writeonly); +#elif defined(TRANSPOSE_8X8_SQUARE) /* TRANSPOSE_8x8_SQUARE */ +TENSOR_DECLARATION(1, srcBuffer, uvec4, src_ptr, src_shift, 4, readonly); +TENSOR_DECLARATION(2, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); void main(void) { - Image src = GC_CONVERT_TO_IMAGE_STRUCT(src); - Image dst = GC_CONVERT_TO_IMAGE_STRUCT(dst); + ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); + ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(dst_attrs, dst_shift); if(gl_GlobalInvocationID.x <= gl_GlobalInvocationID.y) { - uint blk1_offset_in_bytes = src.current_offset; - uint blk2_offset_in_bytes = uint(16) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_step_y) + (dst.offset_first_element_in_bytes); + uint blk1_offset_in_bytes = CURRENT_ITEM_OFFSET_IN_BYTES(src_iter); + TENSOR_ITERATOR_ADVANCE_IN_BYTES(dst_iter, uint(16) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_attrs.step_y)); + uint blk2_offset_in_bytes = CURRENT_ITEM_OFFSET_IN_BYTES(dst_iter); // load block1 - vec4 u1[8][2]; - uvec4 packed_s[8]; + vec4 u1[8][2]; - src.current_offset = blk1_offset_in_bytes; + SET_TENSOR_ITERATOR_OFFSET_IN_BYTES(src_iter, blk1_offset_in_bytes); for(int i = 0; i < 8; i++) { - GC_LOAD1_2D_OFFSET(packed_s[i], src, 0, i); - u1[i][0] = vec4(unpackHalf2x16(packed_s[i].x), unpackHalf2x16(packed_s[i].y)); - u1[i][1] = vec4(unpackHalf2x16(packed_s[i].z), unpackHalf2x16(packed_s[i].w)); + u1[i] = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, i)); } // transpose block1 @@ -252,22 +198,19 @@ void main(void) SWAP_4x4_func(u1[0][1], u1[1][1], u1[2][1], u1[3][1], u1[4][0], u1[5][0], u1[6][0], u1[7][0]); // write to block2 - dst.current_offset = blk2_offset_in_bytes; + SET_TENSOR_ITERATOR_OFFSET_IN_BYTES(dst_iter, blk2_offset_in_bytes); for(int i = 0; i < 8; i++) { - packed_s[i] = uvec4(packHalf2x16(u1[i][0].xy), packHalf2x16(u1[i][0].zw), packHalf2x16(u1[i][1].xy), packHalf2x16(u1[i][1].zw)); - GC_STORE1_2D_OFFSET(packed_s[i], dst, 0, i); + STORE_PACK8_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, i), u1[i]); } // load block2 vec4 u2[8][2]; - src.current_offset = blk2_offset_in_bytes; + SET_TENSOR_ITERATOR_OFFSET_IN_BYTES(src_iter, blk2_offset_in_bytes); for(int i = 0; i < 8; i++) { - GC_LOAD1_2D_OFFSET(packed_s[i], src, 0, i); - u2[i][0] = vec4(unpackHalf2x16(packed_s[i].x), unpackHalf2x16(packed_s[i].y)); - u2[i][1] = vec4(unpackHalf2x16(packed_s[i].z), unpackHalf2x16(packed_s[i].w)); + u2[i] = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, i)); } // transpose block2 @@ -278,11 +221,10 @@ void main(void) SWAP_4x4_func(u2[0][1], u2[1][1], u2[2][1], u2[3][1], u2[4][0], u2[5][0], u2[6][0], u2[7][0]); // write to block1 - dst.current_offset = blk1_offset_in_bytes; + SET_TENSOR_ITERATOR_OFFSET_IN_BYTES(dst_iter, blk1_offset_in_bytes); for(int i = 0; i < 8; i++) { - packed_s[i] = uvec4(packHalf2x16(u2[i][0].xy), packHalf2x16(u2[i][0].zw), packHalf2x16(u2[i][1].xy), packHalf2x16(u2[i][1].zw)); - GC_STORE1_2D_OFFSET(packed_s[i], dst, 0, i); + STORE_PACK8_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, i), u2[i]); } } } -- cgit v1.2.1