1 /* 2 * Copyright 2016 Red Hat 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 * SOFTWARE. 22 */ 23 24 #include <stdbool.h> 25 26 #include "st_tgsi_lower_yuv.h" 27 #include "tgsi/tgsi_transform.h" 28 #include "tgsi/tgsi_scan.h" 29 #include "tgsi/tgsi_dump.h" 30 #include "util/u_debug.h" 31 32 #include "util/bitscan.h" 33 34 struct tgsi_yuv_transform { 35 struct tgsi_transform_context base; 36 struct tgsi_shader_info info; 37 struct tgsi_full_src_register imm[4]; 38 struct { 39 struct tgsi_full_src_register src; 40 struct tgsi_full_dst_register dst; 41 } tmp[2]; 42 #define A 0 43 #define B 1 44 45 /* Maps a primary sampler (used for Y) to the U or UV sampler. In 46 * case of 3-plane YUV format, the V plane is next sampler after U. 47 */ 48 unsigned char sampler_map[PIPE_MAX_SAMPLERS][2]; 49 50 bool first_instruction_emitted; 51 unsigned free_slots; 52 unsigned lower_nv12; 53 unsigned lower_iyuv; 54 }; 55 56 static inline struct tgsi_yuv_transform * 57 tgsi_yuv_transform(struct tgsi_transform_context *tctx) 58 { 59 return (struct tgsi_yuv_transform *)tctx; 60 } 61 62 static void 63 reg_dst(struct tgsi_full_dst_register *dst, 64 const struct tgsi_full_dst_register *orig_dst, unsigned wrmask) 65 { 66 *dst = *orig_dst; 67 dst->Register.WriteMask &= wrmask; 68 assert(dst->Register.WriteMask); 69 } 70 71 static inline void 72 get_swiz(unsigned *swiz, const struct tgsi_src_register *src) 73 { 74 swiz[0] = src->SwizzleX; 75 swiz[1] = src->SwizzleY; 76 swiz[2] = src->SwizzleZ; 77 swiz[3] = src->SwizzleW; 78 } 79 80 static void 81 reg_src(struct tgsi_full_src_register *src, 82 const struct tgsi_full_src_register *orig_src, 83 unsigned sx, unsigned sy, unsigned sz, unsigned sw) 84 { 85 unsigned swiz[4]; 86 get_swiz(swiz, &orig_src->Register); 87 *src = *orig_src; 88 src->Register.SwizzleX = swiz[sx]; 89 src->Register.SwizzleY = swiz[sy]; 90 src->Register.SwizzleZ = swiz[sz]; 91 src->Register.SwizzleW = swiz[sw]; 92 } 93 94 #define TGSI_SWIZZLE__ TGSI_SWIZZLE_X /* don't-care value! */ 95 #define SWIZ(x,y,z,w) TGSI_SWIZZLE_ ## x, TGSI_SWIZZLE_ ## y, \ 96 TGSI_SWIZZLE_ ## z, TGSI_SWIZZLE_ ## w 97 98 static inline struct tgsi_full_instruction 99 tex_instruction(unsigned samp) 100 { 101 struct tgsi_full_instruction inst; 102 103 inst = tgsi_default_full_instruction(); 104 inst.Instruction.Opcode = TGSI_OPCODE_TEX; 105 inst.Instruction.Texture = 1; 106 inst.Texture.Texture = TGSI_TEXTURE_2D; 107 inst.Instruction.NumDstRegs = 1; 108 inst.Instruction.NumSrcRegs = 2; 109 inst.Src[1].Register.File = TGSI_FILE_SAMPLER; 110 inst.Src[1].Register.Index = samp; 111 112 return inst; 113 } 114 115 static inline struct tgsi_full_instruction 116 mov_instruction(void) 117 { 118 struct tgsi_full_instruction inst; 119 120 inst = tgsi_default_full_instruction(); 121 inst.Instruction.Opcode = TGSI_OPCODE_MOV; 122 inst.Instruction.Saturate = 0; 123 inst.Instruction.NumDstRegs = 1; 124 inst.Instruction.NumSrcRegs = 1; 125 126 return inst; 127 } 128 129 static inline struct tgsi_full_instruction 130 dp3_instruction(void) 131 { 132 struct tgsi_full_instruction inst; 133 134 inst = tgsi_default_full_instruction(); 135 inst.Instruction.Opcode = TGSI_OPCODE_DP3; 136 inst.Instruction.NumDstRegs = 1; 137 inst.Instruction.NumSrcRegs = 2; 138 139 return inst; 140 } 141 142 143 144 static void 145 emit_immed(struct tgsi_transform_context *tctx, int idx, 146 float x, float y, float z, float w) 147 { 148 struct tgsi_yuv_transform *ctx = tgsi_yuv_transform(tctx); 149 struct tgsi_shader_info *info = &ctx->info; 150 struct tgsi_full_immediate immed; 151 152 immed = tgsi_default_full_immediate(); 153 immed.Immediate.NrTokens = 1 + 4; /* one for the token itself */ 154 immed.u[0].Float = x; 155 immed.u[1].Float = y; 156 immed.u[2].Float = z; 157 immed.u[3].Float = w; 158 tctx->emit_immediate(tctx, &immed); 159 160 ctx->imm[idx].Register.File = TGSI_FILE_IMMEDIATE; 161 ctx->imm[idx].Register.Index = info->immediate_count + idx; 162 ctx->imm[idx].Register.SwizzleX = TGSI_SWIZZLE_X; 163 ctx->imm[idx].Register.SwizzleY = TGSI_SWIZZLE_Y; 164 ctx->imm[idx].Register.SwizzleZ = TGSI_SWIZZLE_Z; 165 ctx->imm[idx].Register.SwizzleW = TGSI_SWIZZLE_W; 166 } 167 168 static void 169 emit_samp(struct tgsi_transform_context *tctx, unsigned samp) 170 { 171 tgsi_transform_sampler_decl(tctx, samp); 172 tgsi_transform_sampler_view_decl(tctx, samp, PIPE_TEXTURE_2D, 173 TGSI_RETURN_TYPE_FLOAT); 174 } 175 176 /* Emit extra declarations we need: 177 * + 2 TEMP to hold intermediate results 178 * + 1 (for 2-plane YUV) or 2 (for 3-plane YUV) extra samplers per 179 * lowered YUV sampler 180 * + extra immediates for doing CSC 181 */ 182 static void 183 emit_decls(struct tgsi_transform_context *tctx) 184 { 185 struct tgsi_yuv_transform *ctx = tgsi_yuv_transform(tctx); 186 struct tgsi_shader_info *info = &ctx->info; 187 unsigned mask, tempbase, i; 188 struct tgsi_full_declaration decl; 189 190 /* 191 * Declare immediates for CSC conversion: 192 */ 193 194 /* ITU-R BT.601 conversion */ 195 emit_immed(tctx, 0, 1.164, 0.000, 1.596, 0.0); 196 emit_immed(tctx, 1, 1.164, -0.392, -0.813, 0.0); 197 emit_immed(tctx, 2, 1.164, 2.017, 0.000, 0.0); 198 emit_immed(tctx, 3, 0.0625, 0.500, 0.500, 1.0); 199 200 /* 201 * Declare extra samplers / sampler-views: 202 */ 203 204 mask = ctx->lower_nv12 | ctx->lower_iyuv; 205 while (mask) { 206 unsigned extra, y_samp = u_bit_scan(&mask); 207 208 extra = u_bit_scan(&ctx->free_slots); 209 ctx->sampler_map[y_samp][0] = extra; 210 emit_samp(tctx, extra); 211 212 if (ctx->lower_iyuv & (1 << y_samp)) { 213 extra = u_bit_scan(&ctx->free_slots); 214 ctx->sampler_map[y_samp][1] = extra; 215 emit_samp(tctx, extra); 216 } 217 } 218 219 /* 220 * Declare extra temp: 221 */ 222 223 tempbase = info->file_max[TGSI_FILE_TEMPORARY] + 1; 224 225 for (i = 0; i < 2; i++) { 226 decl = tgsi_default_full_declaration(); 227 decl.Declaration.File = TGSI_FILE_TEMPORARY; 228 decl.Range.First = decl.Range.Last = tempbase + i; 229 tctx->emit_declaration(tctx, &decl); 230 231 ctx->tmp[i].src.Register.File = TGSI_FILE_TEMPORARY; 232 ctx->tmp[i].src.Register.Index = tempbase + i; 233 ctx->tmp[i].src.Register.SwizzleX = TGSI_SWIZZLE_X; 234 ctx->tmp[i].src.Register.SwizzleY = TGSI_SWIZZLE_Y; 235 ctx->tmp[i].src.Register.SwizzleZ = TGSI_SWIZZLE_Z; 236 ctx->tmp[i].src.Register.SwizzleW = TGSI_SWIZZLE_W; 237 238 ctx->tmp[i].dst.Register.File = TGSI_FILE_TEMPORARY; 239 ctx->tmp[i].dst.Register.Index = tempbase + i; 240 ctx->tmp[i].dst.Register.WriteMask = TGSI_WRITEMASK_XYZW; 241 } 242 } 243 244 /* call with YUV in tmpA.xyz */ 245 static void 246 yuv_to_rgb(struct tgsi_transform_context *tctx, 247 struct tgsi_full_dst_register *dst) 248 { 249 struct tgsi_yuv_transform *ctx = tgsi_yuv_transform(tctx); 250 struct tgsi_full_instruction inst; 251 252 /* 253 * IMM[0] FLT32 { 1.164, 0.000, 1.596, 0.0 } 254 * IMM[1] FLT32 { 1.164, -0.392, -0.813, 0.0 } 255 * IMM[2] FLT32 { 1.164, 2.017, 0.000, 0.0 } 256 * IMM[3] FLT32 { 0.0625, 0.500, 0.500, 1.0 } 257 */ 258 259 /* SUB tmpA.xyz, tmpA, imm[3] */ 260 inst = tgsi_default_full_instruction(); 261 inst.Instruction.Opcode = TGSI_OPCODE_ADD; 262 inst.Instruction.Saturate = 0; 263 inst.Instruction.NumDstRegs = 1; 264 inst.Instruction.NumSrcRegs = 2; 265 reg_dst(&inst.Dst[0], &ctx->tmp[A].dst, TGSI_WRITEMASK_XYZ); 266 reg_src(&inst.Src[0], &ctx->tmp[A].src, SWIZ(X, Y, Z, _)); 267 reg_src(&inst.Src[1], &ctx->imm[3], SWIZ(X, Y, Z, _)); 268 inst.Src[1].Register.Negate = 1; 269 tctx->emit_instruction(tctx, &inst); 270 271 /* DP3 dst.x, tmpA, imm[0] */ 272 inst = dp3_instruction(); 273 reg_dst(&inst.Dst[0], dst, TGSI_WRITEMASK_X); 274 reg_src(&inst.Src[0], &ctx->tmp[A].src, SWIZ(X, Y, Z, W)); 275 reg_src(&inst.Src[1], &ctx->imm[0], SWIZ(X, Y, Z, W)); 276 tctx->emit_instruction(tctx, &inst); 277 278 /* DP3 dst.y, tmpA, imm[1] */ 279 inst = dp3_instruction(); 280 reg_dst(&inst.Dst[0], dst, TGSI_WRITEMASK_Y); 281 reg_src(&inst.Src[0], &ctx->tmp[A].src, SWIZ(X, Y, Z, W)); 282 reg_src(&inst.Src[1], &ctx->imm[1], SWIZ(X, Y, Z, W)); 283 tctx->emit_instruction(tctx, &inst); 284 285 /* DP3 dst.z, tmpA, imm[2] */ 286 inst = dp3_instruction(); 287 reg_dst(&inst.Dst[0], dst, TGSI_WRITEMASK_Z); 288 reg_src(&inst.Src[0], &ctx->tmp[A].src, SWIZ(X, Y, Z, W)); 289 reg_src(&inst.Src[1], &ctx->imm[2], SWIZ(X, Y, Z, W)); 290 tctx->emit_instruction(tctx, &inst); 291 292 /* MOV dst.w, imm[0].x */ 293 inst = mov_instruction(); 294 reg_dst(&inst.Dst[0], dst, TGSI_WRITEMASK_W); 295 reg_src(&inst.Src[0], &ctx->imm[3], SWIZ(_, _, _, W)); 296 tctx->emit_instruction(tctx, &inst); 297 } 298 299 static void 300 lower_nv12(struct tgsi_transform_context *tctx, 301 struct tgsi_full_instruction *originst) 302 { 303 struct tgsi_yuv_transform *ctx = tgsi_yuv_transform(tctx); 304 struct tgsi_full_instruction inst; 305 struct tgsi_full_src_register *coord = &originst->Src[0]; 306 unsigned samp = originst->Src[1].Register.Index; 307 308 /* sample Y: 309 * TEX tempA.x, coord, texture[samp], 2D; 310 */ 311 inst = tex_instruction(samp); 312 reg_dst(&inst.Dst[0], &ctx->tmp[A].dst, TGSI_WRITEMASK_X); 313 reg_src(&inst.Src[0], coord, SWIZ(X, Y, Z, W)); 314 tctx->emit_instruction(tctx, &inst); 315 316 /* sample UV: 317 * TEX tempB.xy, coord, texture[sampler_map[samp][0]], 2D; 318 * MOV tempA.yz, tempB._xy_ 319 */ 320 inst = tex_instruction(ctx->sampler_map[samp][0]); 321 reg_dst(&inst.Dst[0], &ctx->tmp[B].dst, TGSI_WRITEMASK_XY); 322 reg_src(&inst.Src[0], coord, SWIZ(X, Y, Z, W)); 323 tctx->emit_instruction(tctx, &inst); 324 325 inst = mov_instruction(); 326 reg_dst(&inst.Dst[0], &ctx->tmp[A].dst, TGSI_WRITEMASK_YZ); 327 reg_src(&inst.Src[0], &ctx->tmp[B].src, SWIZ(_, X, Y, _)); 328 tctx->emit_instruction(tctx, &inst); 329 330 /* At this point, we have YUV in tempA.xyz, rest is common: */ 331 yuv_to_rgb(tctx, &originst->Dst[0]); 332 } 333 334 static void 335 lower_iyuv(struct tgsi_transform_context *tctx, 336 struct tgsi_full_instruction *originst) 337 { 338 struct tgsi_yuv_transform *ctx = tgsi_yuv_transform(tctx); 339 struct tgsi_full_instruction inst; 340 struct tgsi_full_src_register *coord = &originst->Src[0]; 341 unsigned samp = originst->Src[1].Register.Index; 342 343 /* sample Y: 344 * TEX tempA.x, coord, texture[samp], 2D; 345 */ 346 inst = tex_instruction(samp); 347 reg_dst(&inst.Dst[0], &ctx->tmp[A].dst, TGSI_WRITEMASK_X); 348 reg_src(&inst.Src[0], coord, SWIZ(X, Y, Z, W)); 349 tctx->emit_instruction(tctx, &inst); 350 351 /* sample U: 352 * TEX tempB.x, coord, texture[sampler_map[samp][0]], 2D; 353 * MOV tempA.y, tempB._x__ 354 */ 355 inst = tex_instruction(ctx->sampler_map[samp][0]); 356 reg_dst(&inst.Dst[0], &ctx->tmp[B].dst, TGSI_WRITEMASK_X); 357 reg_src(&inst.Src[0], coord, SWIZ(X, Y, Z, W)); 358 tctx->emit_instruction(tctx, &inst); 359 360 inst = mov_instruction(); 361 reg_dst(&inst.Dst[0], &ctx->tmp[A].dst, TGSI_WRITEMASK_Y); 362 reg_src(&inst.Src[0], &ctx->tmp[B].src, SWIZ(_, X, _, _)); 363 tctx->emit_instruction(tctx, &inst); 364 365 /* sample V: 366 * TEX tempB.x, coord, texture[sampler_map[samp][1]], 2D; 367 * MOV tempA.z, tempB.__x_ 368 */ 369 inst = tex_instruction(ctx->sampler_map[samp][1]); 370 reg_dst(&inst.Dst[0], &ctx->tmp[B].dst, TGSI_WRITEMASK_X); 371 reg_src(&inst.Src[0], coord, SWIZ(X, Y, Z, W)); 372 tctx->emit_instruction(tctx, &inst); 373 374 inst = mov_instruction(); 375 reg_dst(&inst.Dst[0], &ctx->tmp[A].dst, TGSI_WRITEMASK_Z); 376 reg_src(&inst.Src[0], &ctx->tmp[B].src, SWIZ(_, _, X, _)); 377 tctx->emit_instruction(tctx, &inst); 378 379 /* At this point, we have YUV in tempA.xyz, rest is common: */ 380 yuv_to_rgb(tctx, &originst->Dst[0]); 381 } 382 383 static void 384 transform_instr(struct tgsi_transform_context *tctx, 385 struct tgsi_full_instruction *inst) 386 { 387 struct tgsi_yuv_transform *ctx = tgsi_yuv_transform(tctx); 388 389 if (!ctx->first_instruction_emitted) { 390 emit_decls(tctx); 391 ctx->first_instruction_emitted = true; 392 } 393 394 switch (inst->Instruction.Opcode) { 395 /* TODO what other tex opcode's can be used w/ external eglimgs? */ 396 case TGSI_OPCODE_TEX: { 397 unsigned samp = inst->Src[1].Register.Index; 398 if (ctx->lower_nv12 & (1 << samp)) { 399 lower_nv12(tctx, inst); 400 } else if (ctx->lower_iyuv & (1 << samp)) { 401 lower_iyuv(tctx, inst); 402 } else { 403 goto skip; 404 } 405 break; 406 } 407 default: 408 skip: 409 tctx->emit_instruction(tctx, inst); 410 return; 411 } 412 } 413 414 extern const struct tgsi_token * 415 st_tgsi_lower_yuv(const struct tgsi_token *tokens, unsigned free_slots, 416 unsigned lower_nv12, unsigned lower_iyuv) 417 { 418 struct tgsi_yuv_transform ctx; 419 struct tgsi_token *newtoks; 420 int newlen; 421 422 assert(!(lower_nv12 & lower_iyuv)); /* bitmasks should be mutually exclusive */ 423 424 // tgsi_dump(tokens, 0); 425 // debug_printf("\n"); 426 427 memset(&ctx, 0, sizeof(ctx)); 428 ctx.base.transform_instruction = transform_instr; 429 ctx.free_slots = free_slots; 430 ctx.lower_nv12 = lower_nv12; 431 ctx.lower_iyuv = lower_iyuv; 432 tgsi_scan_shader(tokens, &ctx.info); 433 434 /* TODO better job of figuring out how many extra tokens we need.. 435 * this is a pain about tgsi_transform :-/ 436 */ 437 newlen = tgsi_num_tokens(tokens) + 120; 438 newtoks = tgsi_alloc_tokens(newlen); 439 if (!newtoks) 440 return NULL; 441 442 tgsi_transform_shader(tokens, newtoks, newlen, &ctx.base); 443 444 // tgsi_dump(newtoks, 0); 445 // debug_printf("\n"); 446 447 return newtoks; 448 } 449
