vl_idct.c revision 027704db75a61300e733b0f5a9efbb491189dce5
1508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König/************************************************************************** 2508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * 3508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * Copyright 2010 Christian König 4508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * All Rights Reserved. 5508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * 6508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * Permission is hereby granted, free of charge, to any person obtaining a 7508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * copy of this software and associated documentation files (the 8508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * "Software"), to deal in the Software without restriction, including 9508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * without limitation the rights to use, copy, modify, merge, publish, 10508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * distribute, sub license, and/or sell copies of the Software, and to 11508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * permit persons to whom the Software is furnished to do so, subject to 12508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * the following conditions: 13508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * 14508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * The above copyright notice and this permission notice (including the 15508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * next paragraph) shall be included in all copies or substantial portions 16508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * of the Software. 17508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * 18508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 19508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 21508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR 22508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 23508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 24508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * 26508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König **************************************************************************/ 27508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 28508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include "vl_idct.h" 29e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König#include "util/u_draw.h" 30508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include <assert.h> 31508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include <pipe/p_context.h> 32508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include <pipe/p_screen.h> 33508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include <util/u_inlines.h> 34508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include <util/u_sampler.h> 35508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include <util/u_format.h> 36508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include <tgsi/tgsi_ureg.h> 37508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#include "vl_types.h" 38508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 39508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#define BLOCK_WIDTH 8 40508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König#define BLOCK_HEIGHT 8 41ac1fd50163119a887487d748fab507b23e215c2bChristian König 4203c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König#define SCALE_FACTOR_16_TO_9 (32768.0f / 256.0f) 43508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 44ed49905944243863913bc2598f734ca038c85b94Christian König#define STAGE1_SCALE 4.0f 45ed49905944243863913bc2598f734ca038c85b94Christian König#define STAGE2_SCALE (SCALE_FACTOR_16_TO_9 / STAGE1_SCALE) 46ed49905944243863913bc2598f734ca038c85b94Christian König 47508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstruct vertex_shader_consts 48508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 49508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct vertex4f norm; 50508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König}; 51508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 52508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königenum VS_INPUT 53508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 54508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König VS_I_RECT, 55508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König VS_I_VPOS, 56508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 57508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König NUM_VS_INPUTS 58508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König}; 59508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 60508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königenum VS_OUTPUT 61508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 62508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König VS_O_VPOS, 63508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König VS_O_BLOCK, 64508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König VS_O_TEX, 657408a6ab89e0bc87209b50334604fae93277fdc6Christian König VS_O_START 66508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König}; 67508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 68e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königstatic const float const_matrix[8][8] = { 69508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.3535530f, 0.3535530f, 0.3535530f, 0.3535530f, 0.3535530f, 0.3535530f, 0.353553f, 0.3535530f }, 70508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.4903930f, 0.4157350f, 0.2777850f, 0.0975451f, -0.0975452f, -0.2777850f, -0.415735f, -0.4903930f }, 71508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.4619400f, 0.1913420f, -0.1913420f, -0.4619400f, -0.4619400f, -0.1913420f, 0.191342f, 0.4619400f }, 72508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.4157350f, -0.0975452f, -0.4903930f, -0.2777850f, 0.2777850f, 0.4903930f, 0.097545f, -0.4157350f }, 73508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.3535530f, -0.3535530f, -0.3535530f, 0.3535540f, 0.3535530f, -0.3535540f, -0.353553f, 0.3535530f }, 74508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.2777850f, -0.4903930f, 0.0975452f, 0.4157350f, -0.4157350f, -0.0975451f, 0.490393f, -0.2777850f }, 75508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.1913420f, -0.4619400f, 0.4619400f, -0.1913420f, -0.1913410f, 0.4619400f, -0.461940f, 0.1913420f }, 76508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 0.0975451f, -0.2777850f, 0.4157350f, -0.4903930f, 0.4903930f, -0.4157350f, 0.277786f, -0.0975458f } 77508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König}; 78508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 79e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König/* vertices for a quad covering a block */ 80e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königstatic const struct vertex2f const_quad[4] = { 81e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König {0.0f, 0.0f}, {1.0f, 0.0f}, {1.0f, 1.0f}, {0.0f, 1.0f} 82e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König}; 83e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 84508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void * 85508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königcreate_vert_shader(struct vl_idct *idct) 86508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 87508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct ureg_program *shader; 887408a6ab89e0bc87209b50334604fae93277fdc6Christian König struct ureg_src scale; 89508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct ureg_src vrect, vpos; 907408a6ab89e0bc87209b50334604fae93277fdc6Christian König struct ureg_dst t_vpos; 917408a6ab89e0bc87209b50334604fae93277fdc6Christian König struct ureg_dst o_vpos, o_block, o_tex, o_start; 92508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 93508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König shader = ureg_create(TGSI_PROCESSOR_VERTEX); 94508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König if (!shader) 95508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return NULL; 96508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 977408a6ab89e0bc87209b50334604fae93277fdc6Christian König scale = ureg_imm2f(shader, 987408a6ab89e0bc87209b50334604fae93277fdc6Christian König (float)BLOCK_WIDTH / idct->destination->width0, 997408a6ab89e0bc87209b50334604fae93277fdc6Christian König (float)BLOCK_HEIGHT / idct->destination->height0); 100508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 101508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König t_vpos = ureg_DECL_temporary(shader); 102508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 103508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vrect = ureg_DECL_vs_input(shader, VS_I_RECT); 104508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vpos = ureg_DECL_vs_input(shader, VS_I_VPOS); 105508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 106508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS); 107508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König o_block = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_BLOCK); 108e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König o_tex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_TEX); 109e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König o_start = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_START); 110508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 111508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König /* 112508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * t_vpos = vpos + vrect 113508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * o_vpos.xy = t_vpos * scale 114508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * o_vpos.zw = vpos 115508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * 116508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * o_block = vrect 117508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * o_tex = t_pos 118508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * o_start = vpos * scale 119508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König * 120508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König */ 121508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_ADD(shader, ureg_writemask(t_vpos, TGSI_WRITEMASK_XY), vpos, vrect); 1227408a6ab89e0bc87209b50334604fae93277fdc6Christian König ureg_MUL(shader, ureg_writemask(t_vpos, TGSI_WRITEMASK_XY), ureg_src(t_vpos), scale); 123508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_MOV(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_XY), ureg_src(t_vpos)); 124508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_MOV(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_ZW), vpos); 125508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 126508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_MOV(shader, ureg_writemask(o_block, TGSI_WRITEMASK_XY), vrect); 1277408a6ab89e0bc87209b50334604fae93277fdc6Christian König ureg_MOV(shader, ureg_writemask(o_tex, TGSI_WRITEMASK_XY), ureg_src(t_vpos)); 1287408a6ab89e0bc87209b50334604fae93277fdc6Christian König ureg_MUL(shader, ureg_writemask(o_start, TGSI_WRITEMASK_XY), vpos, scale); 129508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 130508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_release_temporary(shader, t_vpos); 131508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 132508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_END(shader); 133508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 134508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return ureg_create_shader_and_destroy(shader, idct->pipe); 135508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 136508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 137508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void 138cfe489b89723117e56674c2be7761c201f8d78ffChristian Königfetch_one(struct ureg_program *shader, struct ureg_dst m[2], 139cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_src tc, struct ureg_src sampler, 140cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_src start, bool right_side, float size) 141508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 142cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_dst t_tc, tmp; 143cfe489b89723117e56674c2be7761c201f8d78ffChristian König unsigned i, j; 144508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 145cfe489b89723117e56674c2be7761c201f8d78ffChristian König t_tc = ureg_DECL_temporary(shader); 146cfe489b89723117e56674c2be7761c201f8d78ffChristian König tmp = ureg_DECL_temporary(shader); 147cfe489b89723117e56674c2be7761c201f8d78ffChristian König 148cfe489b89723117e56674c2be7761c201f8d78ffChristian König m[0] = ureg_DECL_temporary(shader); 149cfe489b89723117e56674c2be7761c201f8d78ffChristian König m[1] = ureg_DECL_temporary(shader); 150cfe489b89723117e56674c2be7761c201f8d78ffChristian König 151cfe489b89723117e56674c2be7761c201f8d78ffChristian König /* 152cfe489b89723117e56674c2be7761c201f8d78ffChristian König * t_tc.x = right_side ? start.x : tc.x 153cfe489b89723117e56674c2be7761c201f8d78ffChristian König * t_tc.y = right_side ? tc.y : start.y 154cfe489b89723117e56674c2be7761c201f8d78ffChristian König * m[0..1].xyzw = tex(t_tc++, sampler) 155cfe489b89723117e56674c2be7761c201f8d78ffChristian König */ 156cfe489b89723117e56674c2be7761c201f8d78ffChristian König if(right_side) { 157cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_X), ureg_scalar(tc, TGSI_SWIZZLE_X)); 158cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_Y), ureg_scalar(start, TGSI_SWIZZLE_Y)); 159cfe489b89723117e56674c2be7761c201f8d78ffChristian König } else { 160cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_X), ureg_scalar(start, TGSI_SWIZZLE_X)); 161cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_Y), ureg_scalar(tc, TGSI_SWIZZLE_Y)); 162cfe489b89723117e56674c2be7761c201f8d78ffChristian König } 163508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König for(i = 0; i < 2; ++i) { 164cfe489b89723117e56674c2be7761c201f8d78ffChristian König for(j = 0; j < 4; ++j) { 165cfe489b89723117e56674c2be7761c201f8d78ffChristian König /* Nouveau and r600g can't writemask tex dst regs (yet?), do in two steps */ 166cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_TEX(shader, tmp, TGSI_TEXTURE_2D, ureg_src(t_tc), sampler); 167cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(m[i], TGSI_WRITEMASK_X << j), ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X)); 168cfe489b89723117e56674c2be7761c201f8d78ffChristian König 169cfe489b89723117e56674c2be7761c201f8d78ffChristian König if(i != 1 || j != 3) /* skip the last add */ 170cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_ADD(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_X << right_side), 171cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_src(t_tc), ureg_imm1f(shader, 1.0f / size)); 172cfe489b89723117e56674c2be7761c201f8d78ffChristian König } 173508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König } 174508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 175cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, t_tc); 176cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, tmp); 177cfe489b89723117e56674c2be7761c201f8d78ffChristian König} 178cfe489b89723117e56674c2be7761c201f8d78ffChristian König 179cfe489b89723117e56674c2be7761c201f8d78ffChristian Königstatic void 180cfe489b89723117e56674c2be7761c201f8d78ffChristian Königfetch_four(struct ureg_program *shader, struct ureg_dst m[2], 181cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_src tc, struct ureg_src sampler, 182cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_src start, bool right_side, float size) 183cfe489b89723117e56674c2be7761c201f8d78ffChristian König{ 184cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_dst t_tc; 185cfe489b89723117e56674c2be7761c201f8d78ffChristian König 186cfe489b89723117e56674c2be7761c201f8d78ffChristian König t_tc = ureg_DECL_temporary(shader); 187cfe489b89723117e56674c2be7761c201f8d78ffChristian König m[0] = ureg_DECL_temporary(shader); 188cfe489b89723117e56674c2be7761c201f8d78ffChristian König m[1] = ureg_DECL_temporary(shader); 189cfe489b89723117e56674c2be7761c201f8d78ffChristian König 190508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König /* 191cfe489b89723117e56674c2be7761c201f8d78ffChristian König * t_tc.x = right_side ? start.x : tc.x 192cfe489b89723117e56674c2be7761c201f8d78ffChristian König * t_tc.y = right_side ? tc.y : start.y 193cfe489b89723117e56674c2be7761c201f8d78ffChristian König * m[0..1] = tex(t_tc++, sampler) 194508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König */ 195cfe489b89723117e56674c2be7761c201f8d78ffChristian König if(right_side) { 196cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_X), ureg_scalar(start, TGSI_SWIZZLE_Y)); 197cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_Y), ureg_scalar(tc, TGSI_SWIZZLE_X)); 198ed8b767a8e09cff4d98a44cdc07b08f1b322c4d3Christian König } else { 199cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_X), ureg_scalar(start, TGSI_SWIZZLE_X)); 200cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_Y), ureg_scalar(tc, TGSI_SWIZZLE_Y)); 201ed8b767a8e09cff4d98a44cdc07b08f1b322c4d3Christian König } 202508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 203cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_TEX(shader, m[0], TGSI_TEXTURE_2D, ureg_src(t_tc), sampler); 204cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_MOV(shader, ureg_writemask(t_tc, TGSI_WRITEMASK_X), ureg_imm1f(shader, 4.0f / size)); 205cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_TEX(shader, m[1], TGSI_TEXTURE_2D, ureg_src(t_tc), sampler); 206508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 207cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, t_tc); 208cfe489b89723117e56674c2be7761c201f8d78ffChristian König} 209508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 210cfe489b89723117e56674c2be7761c201f8d78ffChristian Königstatic struct ureg_dst 211cfe489b89723117e56674c2be7761c201f8d78ffChristian Königmatrix_mul(struct ureg_program *shader, struct ureg_dst m[2][2]) 212cfe489b89723117e56674c2be7761c201f8d78ffChristian König{ 213cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_dst dst, tmp[2]; 214cfe489b89723117e56674c2be7761c201f8d78ffChristian König unsigned i; 215cfe489b89723117e56674c2be7761c201f8d78ffChristian König 216cfe489b89723117e56674c2be7761c201f8d78ffChristian König dst = ureg_DECL_temporary(shader); 217cfe489b89723117e56674c2be7761c201f8d78ffChristian König for(i = 0; i < 2; ++i) { 218cfe489b89723117e56674c2be7761c201f8d78ffChristian König tmp[i] = ureg_DECL_temporary(shader); 219508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König } 220508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 221cfe489b89723117e56674c2be7761c201f8d78ffChristian König /* 222cfe489b89723117e56674c2be7761c201f8d78ffChristian König * tmp[0..1] = dot4(m[0][0..1], m[1][0..1]) 223cfe489b89723117e56674c2be7761c201f8d78ffChristian König * dst = tmp[0] + tmp[1] 224cfe489b89723117e56674c2be7761c201f8d78ffChristian König */ 225cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_DP4(shader, ureg_writemask(tmp[0], TGSI_WRITEMASK_X), ureg_src(m[0][0]), ureg_src(m[1][0])); 226cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_DP4(shader, ureg_writemask(tmp[1], TGSI_WRITEMASK_X), ureg_src(m[0][1]), ureg_src(m[1][1])); 227cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_ADD(shader, ureg_writemask(dst, TGSI_WRITEMASK_X), ureg_src(tmp[0]), ureg_src(tmp[1])); 228508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 229508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König for(i = 0; i < 2; ++i) { 230508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_release_temporary(shader, tmp[i]); 231508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König } 232cfe489b89723117e56674c2be7761c201f8d78ffChristian König 233cfe489b89723117e56674c2be7761c201f8d78ffChristian König return dst; 234508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 235508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 236508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void * 237508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königcreate_transpose_frag_shader(struct vl_idct *idct) 238508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 239508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct ureg_program *shader; 240cfe489b89723117e56674c2be7761c201f8d78ffChristian König 241508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct ureg_src tc[2], sampler[2]; 242cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_src start[2]; 243cfe489b89723117e56674c2be7761c201f8d78ffChristian König 244cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_dst m[2][2]; 245cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_dst tmp, fragment; 246508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 247508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König shader = ureg_create(TGSI_PROCESSOR_FRAGMENT); 248508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König if (!shader) 249508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return NULL; 250508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 251508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König tc[0] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_BLOCK, TGSI_INTERPOLATE_LINEAR); 252508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König tc[1] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_TEX, TGSI_INTERPOLATE_LINEAR); 253508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 254508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König sampler[0] = ureg_DECL_sampler(shader, 0); 2552c9db2484b7c1cec7a3a629f70a5aa840e16268eChristian König sampler[1] = ureg_DECL_sampler(shader, 1); 256508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 257cfe489b89723117e56674c2be7761c201f8d78ffChristian König start[0] = ureg_imm1f(shader, 0.0f); 258cfe489b89723117e56674c2be7761c201f8d78ffChristian König start[1] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_START, TGSI_INTERPOLATE_CONSTANT); 259cfe489b89723117e56674c2be7761c201f8d78ffChristian König 260cfe489b89723117e56674c2be7761c201f8d78ffChristian König fetch_four(shader, m[0], tc[0], sampler[0], start[0], false, BLOCK_WIDTH); 261cfe489b89723117e56674c2be7761c201f8d78ffChristian König fetch_one(shader, m[1], tc[1], sampler[1], start[1], true, idct->destination->height0); 262c9e10c666adc64f6c5dfb04422560508f115aa54Christian König 263508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0); 264508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 265cfe489b89723117e56674c2be7761c201f8d78ffChristian König tmp = matrix_mul(shader, m); 26613e28cff7655adec0f89aed9c5ee74f8481133abChristian König ureg_MUL(shader, fragment, ureg_src(tmp), ureg_imm1f(shader, STAGE2_SCALE)); 267cfe489b89723117e56674c2be7761c201f8d78ffChristian König 268cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, tmp); 269cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[0][0]); 270cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[0][1]); 271cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[1][0]); 272cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[1][1]); 27303c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 274508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_END(shader); 275508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 276508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return ureg_create_shader_and_destroy(shader, idct->pipe); 277508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 278508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 279508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void * 280508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königcreate_matrix_frag_shader(struct vl_idct *idct) 281508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 282508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct ureg_program *shader; 283cfe489b89723117e56674c2be7761c201f8d78ffChristian König 284508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct ureg_src tc[2], sampler[2]; 285cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_src start[2]; 286cfe489b89723117e56674c2be7761c201f8d78ffChristian König 287cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_dst m[2][2]; 288cfe489b89723117e56674c2be7761c201f8d78ffChristian König struct ureg_dst tmp, fragment; 289508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 290508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König shader = ureg_create(TGSI_PROCESSOR_FRAGMENT); 291508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König if (!shader) 292508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return NULL; 293508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 294508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König tc[0] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_TEX, TGSI_INTERPOLATE_LINEAR); 295508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König tc[1] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_BLOCK, TGSI_INTERPOLATE_LINEAR); 296508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 2972c9db2484b7c1cec7a3a629f70a5aa840e16268eChristian König sampler[0] = ureg_DECL_sampler(shader, 1); 2982c9db2484b7c1cec7a3a629f70a5aa840e16268eChristian König sampler[1] = ureg_DECL_sampler(shader, 0); 299508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 300cfe489b89723117e56674c2be7761c201f8d78ffChristian König start[0] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_START, TGSI_INTERPOLATE_CONSTANT); 301cfe489b89723117e56674c2be7761c201f8d78ffChristian König start[1] = ureg_imm1f(shader, 0.0f); 302cfe489b89723117e56674c2be7761c201f8d78ffChristian König 30313e28cff7655adec0f89aed9c5ee74f8481133abChristian König fetch_four(shader, m[0], tc[0], sampler[0], start[0], false, idct->destination->width0); 304cfe489b89723117e56674c2be7761c201f8d78ffChristian König fetch_four(shader, m[1], tc[1], sampler[1], start[1], true, BLOCK_HEIGHT); 305c9e10c666adc64f6c5dfb04422560508f115aa54Christian König 306508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0); 307508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 308cfe489b89723117e56674c2be7761c201f8d78ffChristian König tmp = matrix_mul(shader, m); 30913e28cff7655adec0f89aed9c5ee74f8481133abChristian König ureg_MUL(shader, fragment, ureg_src(tmp), ureg_imm1f(shader, STAGE1_SCALE)); 310cfe489b89723117e56674c2be7761c201f8d78ffChristian König 311cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, tmp); 312cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[0][0]); 313cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[0][1]); 314cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[1][0]); 315cfe489b89723117e56674c2be7761c201f8d78ffChristian König ureg_release_temporary(shader, m[1][1]); 316508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 317508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ureg_END(shader); 318508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 319508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return ureg_create_shader_and_destroy(shader, idct->pipe); 320508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 321508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 32295febb69cc333dad75c0f2da19dd85f444281ad2Christian Königstatic void * 32395febb69cc333dad75c0f2da19dd85f444281ad2Christian Königcreate_empty_block_frag_shader(struct vl_idct *idct) 32495febb69cc333dad75c0f2da19dd85f444281ad2Christian König{ 32595febb69cc333dad75c0f2da19dd85f444281ad2Christian König struct ureg_program *shader; 32695febb69cc333dad75c0f2da19dd85f444281ad2Christian König struct ureg_dst fragment; 32795febb69cc333dad75c0f2da19dd85f444281ad2Christian König 32895febb69cc333dad75c0f2da19dd85f444281ad2Christian König shader = ureg_create(TGSI_PROCESSOR_FRAGMENT); 32995febb69cc333dad75c0f2da19dd85f444281ad2Christian König if (!shader) 33095febb69cc333dad75c0f2da19dd85f444281ad2Christian König return NULL; 33195febb69cc333dad75c0f2da19dd85f444281ad2Christian König 33295febb69cc333dad75c0f2da19dd85f444281ad2Christian König fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0); 33395febb69cc333dad75c0f2da19dd85f444281ad2Christian König 33495febb69cc333dad75c0f2da19dd85f444281ad2Christian König ureg_MOV(shader, fragment, ureg_imm1f(shader, 0.0f)); 33595febb69cc333dad75c0f2da19dd85f444281ad2Christian König 33695febb69cc333dad75c0f2da19dd85f444281ad2Christian König ureg_END(shader); 33795febb69cc333dad75c0f2da19dd85f444281ad2Christian König 33895febb69cc333dad75c0f2da19dd85f444281ad2Christian König return ureg_create_shader_and_destroy(shader, idct->pipe); 33995febb69cc333dad75c0f2da19dd85f444281ad2Christian König} 34095febb69cc333dad75c0f2da19dd85f444281ad2Christian König 341508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void 342508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königxfer_buffers_map(struct vl_idct *idct) 343508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 344508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct pipe_box rect = 345508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König { 346508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 0, 0, 0, 347508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->destination->width0, 348508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->destination->height0, 349508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 1 350508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König }; 351508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 352508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->tex_transfer = idct->pipe->get_transfer 353508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ( 35403c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->pipe, idct->textures.individual.source, 355508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König u_subresource(0, 0), 356508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD, 357508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König &rect 358508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ); 359508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 360508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->texels = idct->pipe->transfer_map(idct->pipe, idct->tex_transfer); 361e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 362e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vectors = pipe_buffer_map 363e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König ( 364e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe, 365e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.pos.buffer, 366e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD, 367e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König &idct->vec_transfer 368e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König ); 369508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 370508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 371508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void 372508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königxfer_buffers_unmap(struct vl_idct *idct) 373508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 374e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König pipe_buffer_unmap(idct->pipe, idct->vertex_bufs.individual.pos.buffer, idct->vec_transfer); 375e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 376508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->transfer_unmap(idct->pipe, idct->tex_transfer); 377508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->transfer_destroy(idct->pipe, idct->tex_transfer); 378508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 379508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 380508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic bool 381508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königinit_shaders(struct vl_idct *idct) 382508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 3835a8078486a013152d150a4524ebfab929eefe6c4Christian König idct->vs = create_vert_shader(idct); 3845a8078486a013152d150a4524ebfab929eefe6c4Christian König idct->transpose_fs = create_transpose_frag_shader(idct); 3855a8078486a013152d150a4524ebfab929eefe6c4Christian König idct->matrix_fs = create_matrix_frag_shader(idct); 3865a8078486a013152d150a4524ebfab929eefe6c4Christian König idct->eb_fs = create_empty_block_frag_shader(idct); 387508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 38821efda86875096333dc0412c0edab1e188f551d8Christian König return 38921efda86875096333dc0412c0edab1e188f551d8Christian König idct->vs != NULL && 39021efda86875096333dc0412c0edab1e188f551d8Christian König idct->transpose_fs != NULL && 39121efda86875096333dc0412c0edab1e188f551d8Christian König idct->matrix_fs != NULL && 39221efda86875096333dc0412c0edab1e188f551d8Christian König idct->eb_fs != NULL; 393508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 394508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 395508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void 396508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königcleanup_shaders(struct vl_idct *idct) 397508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 398508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->delete_vs_state(idct->pipe, idct->vs); 399508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->delete_fs_state(idct->pipe, idct->transpose_fs); 400508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->delete_fs_state(idct->pipe, idct->matrix_fs); 40195febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->pipe->delete_fs_state(idct->pipe, idct->eb_fs); 402508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 403508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 404508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic bool 405508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königinit_buffers(struct vl_idct *idct) 406508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 407508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct pipe_resource template; 408508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct pipe_sampler_view sampler_view; 409508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König struct pipe_vertex_element vertex_elems[2]; 41021efda86875096333dc0412c0edab1e188f551d8Christian König unsigned i; 411508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 41295febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->max_blocks = 413508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König align(idct->destination->width0, BLOCK_WIDTH) / BLOCK_WIDTH * 414508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König align(idct->destination->height0, BLOCK_HEIGHT) / BLOCK_HEIGHT * 415508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->destination->depth0; 416508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 417508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König memset(&template, 0, sizeof(struct pipe_resource)); 418508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.target = PIPE_TEXTURE_2D; 419c9e10c666adc64f6c5dfb04422560508f115aa54Christian König template.format = PIPE_FORMAT_R32G32B32A32_FLOAT; 420508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.last_level = 0; 421c9e10c666adc64f6c5dfb04422560508f115aa54Christian König template.width0 = 2; 422508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.height0 = 8; 423508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.depth0 = 1; 424508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.usage = PIPE_USAGE_IMMUTABLE; 425508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.bind = PIPE_BIND_SAMPLER_VIEW; 426508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.flags = 0; 427508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 42813e28cff7655adec0f89aed9c5ee74f8481133abChristian König template.format = PIPE_FORMAT_R16G16B16A16_SNORM; 429508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.width0 = idct->destination->width0; 430508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.height0 = idct->destination->height0; 431508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König template.depth0 = idct->destination->depth0; 43213e28cff7655adec0f89aed9c5ee74f8481133abChristian König template.usage = PIPE_USAGE_STREAM; 433508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->textures.individual.source = idct->pipe->screen->resource_create(idct->pipe->screen, &template); 434508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 43503c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König template.usage = PIPE_USAGE_STATIC; 436508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->textures.individual.intermediate = idct->pipe->screen->resource_create(idct->pipe->screen, &template); 437508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 438508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König for (i = 0; i < 4; ++i) { 43921efda86875096333dc0412c0edab1e188f551d8Christian König if(idct->textures.all[i] == NULL) 44021efda86875096333dc0412c0edab1e188f551d8Christian König return false; /* a texture failed to allocate */ 44121efda86875096333dc0412c0edab1e188f551d8Christian König 442508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König u_sampler_view_default_template(&sampler_view, idct->textures.all[i], idct->textures.all[i]->format); 443508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->sampler_views.all[i] = idct->pipe->create_sampler_view(idct->pipe, idct->textures.all[i], &sampler_view); 444508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König } 445508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 446e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.quad.stride = sizeof(struct vertex2f); 44795febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->vertex_bufs.individual.quad.max_index = 4 * idct->max_blocks - 1; 448e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.quad.buffer_offset = 0; 449e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.quad.buffer = pipe_buffer_create 450508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ( 451508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->screen, 452508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König PIPE_BIND_VERTEX_BUFFER, 45395febb69cc333dad75c0f2da19dd85f444281ad2Christian König sizeof(struct vertex2f) * 4 * idct->max_blocks 454508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ); 455508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 45621efda86875096333dc0412c0edab1e188f551d8Christian König if(idct->vertex_bufs.individual.quad.buffer == NULL) 45721efda86875096333dc0412c0edab1e188f551d8Christian König return false; 45821efda86875096333dc0412c0edab1e188f551d8Christian König 459e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.pos.stride = sizeof(struct vertex2f); 46095febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->vertex_bufs.individual.pos.max_index = 4 * idct->max_blocks - 1; 461e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.pos.buffer_offset = 0; 462e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.pos.buffer = pipe_buffer_create 463508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ( 464508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->screen, 465508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König PIPE_BIND_VERTEX_BUFFER, 46695febb69cc333dad75c0f2da19dd85f444281ad2Christian König sizeof(struct vertex2f) * 4 * idct->max_blocks 467508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König ); 468508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 46921efda86875096333dc0412c0edab1e188f551d8Christian König if(idct->vertex_bufs.individual.pos.buffer == NULL) 47021efda86875096333dc0412c0edab1e188f551d8Christian König return false; 47121efda86875096333dc0412c0edab1e188f551d8Christian König 472508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König /* Rect element */ 473508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[0].src_offset = 0; 474508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[0].instance_divisor = 0; 475508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[0].vertex_buffer_index = 0; 476508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[0].src_format = PIPE_FORMAT_R32G32_FLOAT; 477508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 478508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König /* Pos element */ 479508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[1].src_offset = 0; 480508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[1].instance_divisor = 0; 481508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[1].vertex_buffer_index = 1; 482508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König vertex_elems[1].src_format = PIPE_FORMAT_R32G32_FLOAT; 483508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 484508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->vertex_elems_state = idct->pipe->create_vertex_elements_state(idct->pipe, 2, vertex_elems); 485508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 486508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return true; 487508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 488508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 489508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königstatic void 490508a4a056c3140dc1f90b93acd46c06c30f7094eChristian Königcleanup_buffers(struct vl_idct *idct) 491508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 492508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König unsigned i; 493508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 494508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König assert(idct); 495508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 496508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König for (i = 0; i < 4; ++i) { 497508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König pipe_sampler_view_reference(&idct->sampler_views.all[i], NULL); 498508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König pipe_resource_reference(&idct->textures.all[i], NULL); 499508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König } 500508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 501508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König idct->pipe->delete_vertex_elements_state(idct->pipe, idct->vertex_elems_state); 502e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König pipe_resource_reference(&idct->vertex_bufs.individual.quad.buffer, NULL); 503e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König pipe_resource_reference(&idct->vertex_bufs.individual.pos.buffer, NULL); 504508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 505508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 506e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königstatic void 507e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königinit_constants(struct vl_idct *idct) 508508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 509e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König struct pipe_transfer *buf_transfer; 510e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König struct vertex2f *v; 511e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 5123dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König unsigned i; 513ac1fd50163119a887487d748fab507b23e215c2bChristian König 514ac1fd50163119a887487d748fab507b23e215c2bChristian König /* quad vectors */ 515e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König v = pipe_buffer_map 516e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König ( 517e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe, 518e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->vertex_bufs.individual.quad.buffer, 519e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD, 520e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König &buf_transfer 521e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König ); 52295febb69cc333dad75c0f2da19dd85f444281ad2Christian König for ( i = 0; i < idct->max_blocks; ++i) 52395febb69cc333dad75c0f2da19dd85f444281ad2Christian König memcpy(v + i * 4, &const_quad, sizeof(const_quad)); 524e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König pipe_buffer_unmap(idct->pipe, idct->vertex_bufs.individual.quad.buffer, buf_transfer); 525e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König} 526e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 527e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königstatic void 528e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königinit_state(struct vl_idct *idct) 529e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König{ 530e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König struct pipe_sampler_state sampler; 531e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König unsigned i; 532e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 533e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->num_blocks = 0; 53495febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->num_empty_blocks = 0; 535e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 536027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[0].scale[0] = idct->destination->width0; 537027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[1].scale[0] = idct->destination->width0; 538508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 539027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->fb_state[0].width = idct->destination->width0; 540027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->fb_state[1].width = idct->destination->width0; 541027704db75a61300e733b0f5a9efbb491189dce5Christian König 542027704db75a61300e733b0f5a9efbb491189dce5Christian König for(i = 0; i < 2; ++i) { 543027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[i].scale[1] = idct->destination->height0; 544027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[i].scale[2] = 1; 545027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[i].scale[3] = 1; 546027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[i].translate[0] = 0; 547027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[i].translate[1] = 0; 548027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[i].translate[2] = 0; 549027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->viewport[i].translate[3] = 0; 550027704db75a61300e733b0f5a9efbb491189dce5Christian König 551027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->fb_state[i].height = idct->destination->height0; 552027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->fb_state[i].nr_cbufs = 1; 553027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->fb_state[i].zsbuf = NULL; 554027704db75a61300e733b0f5a9efbb491189dce5Christian König } 555508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 556e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König for (i = 0; i < 4; ++i) { 557e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König memset(&sampler, 0, sizeof(sampler)); 558e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE; 559e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE; 560e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_EDGE; 561e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST; 562e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE; 563e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST; 564e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.compare_mode = PIPE_TEX_COMPARE_NONE; 565e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.compare_func = PIPE_FUNC_ALWAYS; 566e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.normalized_coords = 1; 567e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König /*sampler.shadow_ambient = ; */ 568e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König /*sampler.lod_bias = ; */ 569e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König sampler.min_lod = 0; 570e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König /*sampler.max_lod = ; */ 571e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König /*sampler.border_color[0] = ; */ 572e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König /*sampler.max_anisotropy = ; */ 573e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->samplers.all[i] = idct->pipe->create_sampler_state(idct->pipe, &sampler); 574e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König } 575e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König} 576e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 577e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königstatic void 578e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian Königcleanup_state(struct vl_idct *idct) 579e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König{ 580e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König unsigned i; 581e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 582e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König for (i = 0; i < 4; ++i) 583e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe->delete_sampler_state(idct->pipe, idct->samplers.all[i]); 584e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König} 585e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 5863dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian Königstruct pipe_resource * 5873dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian Königvl_idct_upload_matrix(struct pipe_context *pipe) 5883dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König{ 5893dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König struct pipe_resource template, *matrix; 5903dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König struct pipe_transfer *buf_transfer; 5913dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König unsigned i, j, pitch; 5923dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König float *f; 5933dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 5943dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König struct pipe_box rect = 5953dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König { 5963dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 0, 0, 0, 5973dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König BLOCK_WIDTH, 5983dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König BLOCK_HEIGHT, 5993dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 1 6003dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König }; 6013dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 6023dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König memset(&template, 0, sizeof(struct pipe_resource)); 6033dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.target = PIPE_TEXTURE_2D; 6043dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.format = PIPE_FORMAT_R32G32B32A32_FLOAT; 6053dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.last_level = 0; 6063dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.width0 = 2; 6073dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.height0 = 8; 6083dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.depth0 = 1; 6093dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.usage = PIPE_USAGE_IMMUTABLE; 6103dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.bind = PIPE_BIND_SAMPLER_VIEW; 6113dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König template.flags = 0; 6123dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 6133dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König matrix = pipe->screen->resource_create(pipe->screen, &template); 6143dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 6153dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König /* matrix */ 6163dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König buf_transfer = pipe->get_transfer 6173dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König ( 6183dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König pipe, matrix, 6193dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König u_subresource(0, 0), 6203dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD, 6213dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König &rect 6223dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König ); 6233dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König pitch = buf_transfer->stride / util_format_get_blocksize(buf_transfer->resource->format); 6243dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 6253dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König f = pipe->transfer_map(pipe, buf_transfer); 6263dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König for(i = 0; i < BLOCK_HEIGHT; ++i) 6273dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König for(j = 0; j < BLOCK_WIDTH; ++j) 6283dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König f[i * pitch * 4 + j] = const_matrix[j][i]; // transpose 6293dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 6303dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König pipe->transfer_unmap(pipe, buf_transfer); 6313dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König pipe->transfer_destroy(pipe, buf_transfer); 6323dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 6333dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König return matrix; 6343dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König} 6353dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König 63603c5a0ea5cd5b3e5931d6784749f87789a016b98Christian Königbool 6373dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian Königvl_idct_init(struct vl_idct *idct, struct pipe_context *pipe, struct pipe_resource *dst, struct pipe_resource *matrix) 638e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König{ 639e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König assert(idct && pipe && dst); 640e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 641e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe = pipe; 6423dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König pipe_resource_reference(&idct->textures.individual.matrix, matrix); 6433dd7bf7d39781f3ef4c0b53732945674c9924cdfChristian König pipe_resource_reference(&idct->textures.individual.transpose, matrix); 644508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König pipe_resource_reference(&idct->destination, dst); 645508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 646e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König init_state(idct); 647e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 648508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König if(!init_shaders(idct)) 649508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return false; 650508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 651508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König if(!init_buffers(idct)) { 652508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König cleanup_shaders(idct); 653508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return false; 654508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König } 655508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 656e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->surfaces.intermediate = idct->pipe->screen->get_tex_surface( 657e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe->screen, idct->textures.individual.intermediate, 0, 0, 0, 658e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET); 659027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->fb_state[0].cbufs[0] = idct->surfaces.intermediate; 660e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 661e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->surfaces.destination = idct->pipe->screen->get_tex_surface( 662e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe->screen, idct->destination, 0, 0, 0, 663e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET); 664027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->fb_state[1].cbufs[0] = idct->surfaces.destination; 665e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 666e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König init_constants(idct); 667508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König xfer_buffers_map(idct); 668508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 669508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König return true; 670508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 671508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 67203c5a0ea5cd5b3e5931d6784749f87789a016b98Christian Königvoid 67303c5a0ea5cd5b3e5931d6784749f87789a016b98Christian Königvl_idct_cleanup(struct vl_idct *idct) 674508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 675e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe->screen->tex_surface_destroy(idct->surfaces.destination); 676e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->pipe->screen->tex_surface_destroy(idct->surfaces.intermediate); 677e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 678508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König cleanup_shaders(idct); 679508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König cleanup_buffers(idct); 680e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 681e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König cleanup_state(idct); 682e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 683e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König pipe_resource_reference(&idct->destination, NULL); 684508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 685508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 68603c5a0ea5cd5b3e5931d6784749f87789a016b98Christian Königvoid 68703c5a0ea5cd5b3e5931d6784749f87789a016b98Christian Königvl_idct_add_block(struct vl_idct *idct, unsigned x, unsigned y, short *block) 688508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 68995febb69cc333dad75c0f2da19dd85f444281ad2Christian König struct vertex2f v, *v_dst; 690e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 691508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König unsigned tex_pitch; 692508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König short *texels; 693e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 694508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König unsigned i; 695508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 696508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König assert(idct); 697e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 69803c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König if(block) { 69995febb69cc333dad75c0f2da19dd85f444281ad2Christian König tex_pitch = idct->tex_transfer->stride / util_format_get_blocksize(idct->tex_transfer->resource->format); 70013e28cff7655adec0f89aed9c5ee74f8481133abChristian König texels = idct->texels + (y * tex_pitch * BLOCK_HEIGHT + x * BLOCK_WIDTH) * 4; 70103c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 70203c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König for (i = 0; i < BLOCK_HEIGHT; ++i) 70313e28cff7655adec0f89aed9c5ee74f8481133abChristian König memcpy(texels + i * tex_pitch * 4, block + i * BLOCK_WIDTH, BLOCK_WIDTH * 2); 704e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 70595febb69cc333dad75c0f2da19dd85f444281ad2Christian König /* non empty blocks fills the vector buffer from left to right */ 70695febb69cc333dad75c0f2da19dd85f444281ad2Christian König v_dst = idct->vectors + idct->num_blocks * 4; 70795febb69cc333dad75c0f2da19dd85f444281ad2Christian König 70803c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->num_blocks++; 70995febb69cc333dad75c0f2da19dd85f444281ad2Christian König 71003c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König } else { 71195febb69cc333dad75c0f2da19dd85f444281ad2Christian König 71295febb69cc333dad75c0f2da19dd85f444281ad2Christian König /* while empty blocks fills the vector buffer from right to left */ 71395febb69cc333dad75c0f2da19dd85f444281ad2Christian König v_dst = idct->vectors + (idct->max_blocks - idct->num_empty_blocks) * 4 - 4; 71495febb69cc333dad75c0f2da19dd85f444281ad2Christian König 71595febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->num_empty_blocks++; 71695febb69cc333dad75c0f2da19dd85f444281ad2Christian König } 71795febb69cc333dad75c0f2da19dd85f444281ad2Christian König 71895febb69cc333dad75c0f2da19dd85f444281ad2Christian König v.x = x; 71995febb69cc333dad75c0f2da19dd85f444281ad2Christian König v.y = y; 72095febb69cc333dad75c0f2da19dd85f444281ad2Christian König 72195febb69cc333dad75c0f2da19dd85f444281ad2Christian König for (i = 0; i < 4; ++i) { 72295febb69cc333dad75c0f2da19dd85f444281ad2Christian König v_dst[i] = v; 72303c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König } 724508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 725508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König 72603c5a0ea5cd5b3e5931d6784749f87789a016b98Christian Königvoid 72703c5a0ea5cd5b3e5931d6784749f87789a016b98Christian Königvl_idct_flush(struct vl_idct *idct) 728508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König{ 729508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König xfer_buffers_unmap(idct); 730e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 73103c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König if(idct->num_blocks > 0) { 73203c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 73303c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König /* first stage */ 734027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->pipe->set_framebuffer_state(idct->pipe, &idct->fb_state[0]); 735027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->pipe->set_viewport_state(idct->pipe, &idct->viewport[0]); 73603c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 73703c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->pipe->set_vertex_buffers(idct->pipe, 2, idct->vertex_bufs.all); 73803c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->pipe->bind_vertex_elements_state(idct->pipe, idct->vertex_elems_state); 7392c9db2484b7c1cec7a3a629f70a5aa840e16268eChristian König idct->pipe->set_fragment_sampler_views(idct->pipe, 2, idct->sampler_views.stage[0]); 7402c9db2484b7c1cec7a3a629f70a5aa840e16268eChristian König idct->pipe->bind_fragment_sampler_states(idct->pipe, 2, idct->samplers.stage[0]); 74103c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->pipe->bind_vs_state(idct->pipe, idct->vs); 74213e28cff7655adec0f89aed9c5ee74f8481133abChristian König idct->pipe->bind_fs_state(idct->pipe, idct->matrix_fs); 74303c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 74403c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, 0, idct->num_blocks * 4); 74503c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 74603c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König /* second stage */ 747027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->pipe->set_framebuffer_state(idct->pipe, &idct->fb_state[1]); 748027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->pipe->set_viewport_state(idct->pipe, &idct->viewport[1]); 74903c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 75003c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->pipe->set_vertex_buffers(idct->pipe, 2, idct->vertex_bufs.all); 75103c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->pipe->bind_vertex_elements_state(idct->pipe, idct->vertex_elems_state); 7522c9db2484b7c1cec7a3a629f70a5aa840e16268eChristian König idct->pipe->set_fragment_sampler_views(idct->pipe, 2, idct->sampler_views.stage[1]); 7532c9db2484b7c1cec7a3a629f70a5aa840e16268eChristian König idct->pipe->bind_fragment_sampler_states(idct->pipe, 2, idct->samplers.stage[1]); 75403c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König idct->pipe->bind_vs_state(idct->pipe, idct->vs); 75513e28cff7655adec0f89aed9c5ee74f8481133abChristian König idct->pipe->bind_fs_state(idct->pipe, idct->transpose_fs); 75603c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 75703c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, 0, idct->num_blocks * 4); 75895febb69cc333dad75c0f2da19dd85f444281ad2Christian König } 75903c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König 76095febb69cc333dad75c0f2da19dd85f444281ad2Christian König if(idct->num_empty_blocks > 0) { 76195febb69cc333dad75c0f2da19dd85f444281ad2Christian König 76295febb69cc333dad75c0f2da19dd85f444281ad2Christian König /* empty block handling */ 763027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->pipe->set_framebuffer_state(idct->pipe, &idct->fb_state[1]); 764027704db75a61300e733b0f5a9efbb491189dce5Christian König idct->pipe->set_viewport_state(idct->pipe, &idct->viewport[1]); 76595febb69cc333dad75c0f2da19dd85f444281ad2Christian König 76695febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->pipe->set_vertex_buffers(idct->pipe, 2, idct->vertex_bufs.all); 76795febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->pipe->bind_vertex_elements_state(idct->pipe, idct->vertex_elems_state); 76895febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->pipe->set_fragment_sampler_views(idct->pipe, 4, idct->sampler_views.all); 76995febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->pipe->bind_fragment_sampler_states(idct->pipe, 4, idct->samplers.all); 77095febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->pipe->bind_vs_state(idct->pipe, idct->vs); 77195febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->pipe->bind_fs_state(idct->pipe, idct->eb_fs); 77295febb69cc333dad75c0f2da19dd85f444281ad2Christian König 77395febb69cc333dad75c0f2da19dd85f444281ad2Christian König util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, 77495febb69cc333dad75c0f2da19dd85f444281ad2Christian König (idct->max_blocks - idct->num_empty_blocks) * 4, 77595febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->num_empty_blocks * 4); 77603c5a0ea5cd5b3e5931d6784749f87789a016b98Christian König } 777e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König 778e639e1b83ea65985cd84d12dc120d77cab80ba9eChristian König idct->num_blocks = 0; 77995febb69cc333dad75c0f2da19dd85f444281ad2Christian König idct->num_empty_blocks = 0; 780508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König xfer_buffers_map(idct); 781508a4a056c3140dc1f90b93acd46c06c30f7094eChristian König} 782