lp_bld_sample_aos.c revision efc82aef35a2aac5d2ed9774f6d28f2626796416
1/************************************************************************** 2 * 3 * Copyright 2010 VMware, Inc. 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR 22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27 28/** 29 * @file 30 * Texture sampling -- SoA. 31 * 32 * @author Jose Fonseca <jfonseca@vmware.com> 33 * @author Brian Paul <brianp@vmware.com> 34 */ 35 36#include "pipe/p_defines.h" 37#include "pipe/p_state.h" 38#include "util/u_debug.h" 39#include "util/u_dump.h" 40#include "util/u_memory.h" 41#include "util/u_math.h" 42#include "util/u_format.h" 43#include "lp_bld_debug.h" 44#include "lp_bld_type.h" 45#include "lp_bld_const.h" 46#include "lp_bld_conv.h" 47#include "lp_bld_arit.h" 48#include "lp_bld_bitarit.h" 49#include "lp_bld_logic.h" 50#include "lp_bld_swizzle.h" 51#include "lp_bld_pack.h" 52#include "lp_bld_flow.h" 53#include "lp_bld_gather.h" 54#include "lp_bld_format.h" 55#include "lp_bld_init.h" 56#include "lp_bld_sample.h" 57#include "lp_bld_sample_aos.h" 58#include "lp_bld_quad.h" 59 60 61/** 62 * Build LLVM code for texture coord wrapping, for nearest filtering, 63 * for scaled integer texcoords. 64 * \param block_length is the length of the pixel block along the 65 * coordinate axis 66 * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size 67 * \param length the texture size along one dimension 68 * \param stride pixel stride along the coordinate axis (in bytes) 69 * \param is_pot if TRUE, length is a power of two 70 * \param wrap_mode one of PIPE_TEX_WRAP_x 71 * \param out_offset byte offset for the wrapped coordinate 72 * \param out_i resulting sub-block pixel coordinate for coord0 73 */ 74static void 75lp_build_sample_wrap_nearest_int(struct lp_build_sample_context *bld, 76 unsigned block_length, 77 LLVMValueRef coord, 78 LLVMValueRef length, 79 LLVMValueRef stride, 80 boolean is_pot, 81 unsigned wrap_mode, 82 LLVMValueRef *out_offset, 83 LLVMValueRef *out_i) 84{ 85 struct lp_build_context *int_coord_bld = &bld->int_coord_bld; 86 LLVMBuilderRef builder = bld->gallivm->builder; 87 LLVMValueRef length_minus_one; 88 89 length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one); 90 91 switch(wrap_mode) { 92 case PIPE_TEX_WRAP_REPEAT: 93 if(is_pot) 94 coord = LLVMBuildAnd(builder, coord, length_minus_one, ""); 95 else { 96 /* Add a bias to the texcoord to handle negative coords */ 97 LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024); 98 coord = LLVMBuildAdd(builder, coord, bias, ""); 99 coord = LLVMBuildURem(builder, coord, length, ""); 100 } 101 break; 102 103 case PIPE_TEX_WRAP_CLAMP_TO_EDGE: 104 coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero); 105 coord = lp_build_min(int_coord_bld, coord, length_minus_one); 106 break; 107 108 case PIPE_TEX_WRAP_CLAMP: 109 case PIPE_TEX_WRAP_CLAMP_TO_BORDER: 110 case PIPE_TEX_WRAP_MIRROR_REPEAT: 111 case PIPE_TEX_WRAP_MIRROR_CLAMP: 112 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: 113 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: 114 default: 115 assert(0); 116 } 117 118 lp_build_sample_partial_offset(int_coord_bld, block_length, coord, stride, 119 out_offset, out_i); 120} 121 122 123/** 124 * Build LLVM code for texture coord wrapping, for linear filtering, 125 * for scaled integer texcoords. 126 * \param block_length is the length of the pixel block along the 127 * coordinate axis 128 * \param coord0 the incoming texcoord (s,t,r or q) scaled to the texture size 129 * \param length the texture size along one dimension 130 * \param stride pixel stride along the coordinate axis (in bytes) 131 * \param is_pot if TRUE, length is a power of two 132 * \param wrap_mode one of PIPE_TEX_WRAP_x 133 * \param offset0 resulting relative offset for coord0 134 * \param offset1 resulting relative offset for coord0 + 1 135 * \param i0 resulting sub-block pixel coordinate for coord0 136 * \param i1 resulting sub-block pixel coordinate for coord0 + 1 137 */ 138static void 139lp_build_sample_wrap_linear_int(struct lp_build_sample_context *bld, 140 unsigned block_length, 141 LLVMValueRef coord0, 142 LLVMValueRef length, 143 LLVMValueRef stride, 144 boolean is_pot, 145 unsigned wrap_mode, 146 LLVMValueRef *offset0, 147 LLVMValueRef *offset1, 148 LLVMValueRef *i0, 149 LLVMValueRef *i1) 150{ 151 struct lp_build_context *int_coord_bld = &bld->int_coord_bld; 152 LLVMBuilderRef builder = bld->gallivm->builder; 153 LLVMValueRef length_minus_one; 154 LLVMValueRef lmask, umask, mask; 155 156 if (block_length != 1) { 157 /* 158 * If the pixel block covers more than one pixel then there is no easy 159 * way to calculate offset1 relative to offset0. Instead, compute them 160 * independently. 161 */ 162 163 LLVMValueRef coord1; 164 165 lp_build_sample_wrap_nearest_int(bld, 166 block_length, 167 coord0, 168 length, 169 stride, 170 is_pot, 171 wrap_mode, 172 offset0, i0); 173 174 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one); 175 176 lp_build_sample_wrap_nearest_int(bld, 177 block_length, 178 coord1, 179 length, 180 stride, 181 is_pot, 182 wrap_mode, 183 offset1, i1); 184 185 return; 186 } 187 188 /* 189 * Scalar pixels -- try to compute offset0 and offset1 with a single stride 190 * multiplication. 191 */ 192 193 *i0 = int_coord_bld->zero; 194 *i1 = int_coord_bld->zero; 195 196 length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one); 197 198 switch(wrap_mode) { 199 case PIPE_TEX_WRAP_REPEAT: 200 if (is_pot) { 201 coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, ""); 202 } 203 else { 204 /* Add a bias to the texcoord to handle negative coords */ 205 LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024); 206 coord0 = LLVMBuildAdd(builder, coord0, bias, ""); 207 coord0 = LLVMBuildURem(builder, coord0, length, ""); 208 } 209 210 mask = lp_build_compare(bld->gallivm, int_coord_bld->type, 211 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one); 212 213 *offset0 = lp_build_mul(int_coord_bld, coord0, stride); 214 *offset1 = LLVMBuildAnd(builder, 215 lp_build_add(int_coord_bld, *offset0, stride), 216 mask, ""); 217 break; 218 219 case PIPE_TEX_WRAP_CLAMP_TO_EDGE: 220 lmask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type, 221 PIPE_FUNC_GEQUAL, coord0, int_coord_bld->zero); 222 umask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type, 223 PIPE_FUNC_LESS, coord0, length_minus_one); 224 225 coord0 = lp_build_select(int_coord_bld, lmask, coord0, int_coord_bld->zero); 226 coord0 = lp_build_select(int_coord_bld, umask, coord0, length_minus_one); 227 228 mask = LLVMBuildAnd(builder, lmask, umask, ""); 229 230 *offset0 = lp_build_mul(int_coord_bld, coord0, stride); 231 *offset1 = lp_build_add(int_coord_bld, 232 *offset0, 233 LLVMBuildAnd(builder, stride, mask, "")); 234 break; 235 236 case PIPE_TEX_WRAP_CLAMP: 237 case PIPE_TEX_WRAP_CLAMP_TO_BORDER: 238 case PIPE_TEX_WRAP_MIRROR_REPEAT: 239 case PIPE_TEX_WRAP_MIRROR_CLAMP: 240 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: 241 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: 242 default: 243 assert(0); 244 *offset0 = int_coord_bld->zero; 245 *offset1 = int_coord_bld->zero; 246 break; 247 } 248} 249 250 251/** 252 * Sample a single texture image with nearest sampling. 253 * If sampling a cube texture, r = cube face in [0,5]. 254 * Return filtered color as two vectors of 16-bit fixed point values. 255 */ 256static void 257lp_build_sample_image_nearest(struct lp_build_sample_context *bld, 258 LLVMValueRef int_size, 259 LLVMValueRef row_stride_vec, 260 LLVMValueRef img_stride_vec, 261 LLVMValueRef data_ptr, 262 LLVMValueRef s, 263 LLVMValueRef t, 264 LLVMValueRef r, 265 LLVMValueRef *colors_lo, 266 LLVMValueRef *colors_hi) 267{ 268 const unsigned dims = bld->dims; 269 LLVMBuilderRef builder = bld->gallivm->builder; 270 struct lp_build_context i32, h16, u8n; 271 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type; 272 LLVMValueRef i32_c8; 273 LLVMValueRef width_vec, height_vec, depth_vec; 274 LLVMValueRef s_ipart, t_ipart = NULL, r_ipart = NULL; 275 LLVMValueRef x_stride; 276 LLVMValueRef x_offset, offset; 277 LLVMValueRef x_subcoord, y_subcoord, z_subcoord; 278 279 lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32)); 280 lp_build_context_init(&h16, bld->gallivm, lp_type_ufixed(16)); 281 lp_build_context_init(&u8n, bld->gallivm, lp_type_unorm(8)); 282 283 i32_vec_type = lp_build_vec_type(bld->gallivm, i32.type); 284 h16_vec_type = lp_build_vec_type(bld->gallivm, h16.type); 285 u8n_vec_type = lp_build_vec_type(bld->gallivm, u8n.type); 286 287 lp_build_extract_image_sizes(bld, 288 bld->int_size_type, 289 bld->int_coord_type, 290 int_size, 291 &width_vec, 292 &height_vec, 293 &depth_vec); 294 295 if (bld->static_state->normalized_coords) { 296 LLVMValueRef scaled_size; 297 LLVMValueRef flt_size; 298 299 /* scale size by 256 (8 fractional bits) */ 300 scaled_size = lp_build_shl_imm(&bld->int_size_bld, int_size, 8); 301 302 flt_size = lp_build_int_to_float(&bld->float_size_bld, scaled_size); 303 304 lp_build_unnormalized_coords(bld, flt_size, &s, &t, &r); 305 } 306 else { 307 /* scale coords by 256 (8 fractional bits) */ 308 s = lp_build_mul_imm(&bld->coord_bld, s, 256); 309 if (dims >= 2) 310 t = lp_build_mul_imm(&bld->coord_bld, t, 256); 311 if (dims >= 3) 312 r = lp_build_mul_imm(&bld->coord_bld, r, 256); 313 } 314 315 /* convert float to int */ 316 s = LLVMBuildFPToSI(builder, s, i32_vec_type, ""); 317 if (dims >= 2) 318 t = LLVMBuildFPToSI(builder, t, i32_vec_type, ""); 319 if (dims >= 3) 320 r = LLVMBuildFPToSI(builder, r, i32_vec_type, ""); 321 322 /* compute floor (shift right 8) */ 323 i32_c8 = lp_build_const_int_vec(bld->gallivm, i32.type, 8); 324 s_ipart = LLVMBuildAShr(builder, s, i32_c8, ""); 325 if (dims >= 2) 326 t_ipart = LLVMBuildAShr(builder, t, i32_c8, ""); 327 if (dims >= 3) 328 r_ipart = LLVMBuildAShr(builder, r, i32_c8, ""); 329 330 /* get pixel, row, image strides */ 331 x_stride = lp_build_const_vec(bld->gallivm, 332 bld->int_coord_bld.type, 333 bld->format_desc->block.bits/8); 334 335 /* Do texcoord wrapping, compute texel offset */ 336 lp_build_sample_wrap_nearest_int(bld, 337 bld->format_desc->block.width, 338 s_ipart, width_vec, x_stride, 339 bld->static_state->pot_width, 340 bld->static_state->wrap_s, 341 &x_offset, &x_subcoord); 342 offset = x_offset; 343 if (dims >= 2) { 344 LLVMValueRef y_offset; 345 lp_build_sample_wrap_nearest_int(bld, 346 bld->format_desc->block.height, 347 t_ipart, height_vec, row_stride_vec, 348 bld->static_state->pot_height, 349 bld->static_state->wrap_t, 350 &y_offset, &y_subcoord); 351 offset = lp_build_add(&bld->int_coord_bld, offset, y_offset); 352 if (dims >= 3) { 353 LLVMValueRef z_offset; 354 lp_build_sample_wrap_nearest_int(bld, 355 1, /* block length (depth) */ 356 r_ipart, depth_vec, img_stride_vec, 357 bld->static_state->pot_height, 358 bld->static_state->wrap_r, 359 &z_offset, &z_subcoord); 360 offset = lp_build_add(&bld->int_coord_bld, offset, z_offset); 361 } 362 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) { 363 LLVMValueRef z_offset; 364 /* The r coord is the cube face in [0,5] */ 365 z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec); 366 offset = lp_build_add(&bld->int_coord_bld, offset, z_offset); 367 } 368 } 369 370 /* 371 * Fetch the pixels as 4 x 32bit (rgba order might differ): 372 * 373 * rgba0 rgba1 rgba2 rgba3 374 * 375 * bit cast them into 16 x u8 376 * 377 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3 378 * 379 * unpack them into two 8 x i16: 380 * 381 * r0 g0 b0 a0 r1 g1 b1 a1 382 * r2 g2 b2 a2 r3 g3 b3 a3 383 * 384 * The higher 8 bits of the resulting elements will be zero. 385 */ 386 { 387 LLVMValueRef rgba8; 388 389 if (util_format_is_rgba8_variant(bld->format_desc)) { 390 /* 391 * Given the format is a rgba8, just read the pixels as is, 392 * without any swizzling. Swizzling will be done later. 393 */ 394 rgba8 = lp_build_gather(bld->gallivm, 395 bld->texel_type.length, 396 bld->format_desc->block.bits, 397 bld->texel_type.width, 398 data_ptr, offset); 399 400 rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, ""); 401 } 402 else { 403 rgba8 = lp_build_fetch_rgba_aos(bld->gallivm, 404 bld->format_desc, 405 u8n.type, 406 data_ptr, offset, 407 x_subcoord, 408 y_subcoord); 409 } 410 411 /* Expand one 4*rgba8 to two 2*rgba16 */ 412 lp_build_unpack2(bld->gallivm, u8n.type, h16.type, 413 rgba8, 414 colors_lo, colors_hi); 415 } 416} 417 418 419/** 420 * Sample a single texture image with (bi-)(tri-)linear sampling. 421 * Return filtered color as two vectors of 16-bit fixed point values. 422 */ 423static void 424lp_build_sample_image_linear(struct lp_build_sample_context *bld, 425 LLVMValueRef int_size, 426 LLVMValueRef row_stride_vec, 427 LLVMValueRef img_stride_vec, 428 LLVMValueRef data_ptr, 429 LLVMValueRef s, 430 LLVMValueRef t, 431 LLVMValueRef r, 432 LLVMValueRef *colors_lo, 433 LLVMValueRef *colors_hi) 434{ 435 const unsigned dims = bld->dims; 436 LLVMBuilderRef builder = bld->gallivm->builder; 437 struct lp_build_context i32, h16, u8n; 438 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type; 439 LLVMValueRef i32_c8, i32_c128, i32_c255; 440 LLVMValueRef width_vec, height_vec, depth_vec; 441 LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi; 442 LLVMValueRef t_ipart = NULL, t_fpart = NULL, t_fpart_lo = NULL, t_fpart_hi = NULL; 443 LLVMValueRef r_ipart = NULL, r_fpart = NULL, r_fpart_lo = NULL, r_fpart_hi = NULL; 444 LLVMValueRef x_stride, y_stride, z_stride; 445 LLVMValueRef x_offset0, x_offset1; 446 LLVMValueRef y_offset0, y_offset1; 447 LLVMValueRef z_offset0, z_offset1; 448 LLVMValueRef offset[2][2][2]; /* [z][y][x] */ 449 LLVMValueRef x_subcoord[2], y_subcoord[2], z_subcoord[2]; 450 LLVMValueRef neighbors_lo[2][2][2]; /* [z][y][x] */ 451 LLVMValueRef neighbors_hi[2][2][2]; /* [z][y][x] */ 452 LLVMValueRef packed_lo, packed_hi; 453 unsigned x, y, z; 454 unsigned i, j, k; 455 unsigned numj, numk; 456 457 lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32)); 458 lp_build_context_init(&h16, bld->gallivm, lp_type_ufixed(16)); 459 lp_build_context_init(&u8n, bld->gallivm, lp_type_unorm(8)); 460 461 i32_vec_type = lp_build_vec_type(bld->gallivm, i32.type); 462 h16_vec_type = lp_build_vec_type(bld->gallivm, h16.type); 463 u8n_vec_type = lp_build_vec_type(bld->gallivm, u8n.type); 464 465 lp_build_extract_image_sizes(bld, 466 bld->int_size_type, 467 bld->int_coord_type, 468 int_size, 469 &width_vec, 470 &height_vec, 471 &depth_vec); 472 473 if (bld->static_state->normalized_coords) { 474 LLVMValueRef scaled_size; 475 LLVMValueRef flt_size; 476 477 /* scale size by 256 (8 fractional bits) */ 478 scaled_size = lp_build_shl_imm(&bld->int_size_bld, int_size, 8); 479 480 flt_size = lp_build_int_to_float(&bld->float_size_bld, scaled_size); 481 482 lp_build_unnormalized_coords(bld, flt_size, &s, &t, &r); 483 } 484 else { 485 /* scale coords by 256 (8 fractional bits) */ 486 s = lp_build_mul_imm(&bld->coord_bld, s, 256); 487 if (dims >= 2) 488 t = lp_build_mul_imm(&bld->coord_bld, t, 256); 489 if (dims >= 3) 490 r = lp_build_mul_imm(&bld->coord_bld, r, 256); 491 } 492 493 /* convert float to int */ 494 s = LLVMBuildFPToSI(builder, s, i32_vec_type, ""); 495 if (dims >= 2) 496 t = LLVMBuildFPToSI(builder, t, i32_vec_type, ""); 497 if (dims >= 3) 498 r = LLVMBuildFPToSI(builder, r, i32_vec_type, ""); 499 500 /* subtract 0.5 (add -128) */ 501 i32_c128 = lp_build_const_int_vec(bld->gallivm, i32.type, -128); 502 s = LLVMBuildAdd(builder, s, i32_c128, ""); 503 if (dims >= 2) { 504 t = LLVMBuildAdd(builder, t, i32_c128, ""); 505 } 506 if (dims >= 3) { 507 r = LLVMBuildAdd(builder, r, i32_c128, ""); 508 } 509 510 /* compute floor (shift right 8) */ 511 i32_c8 = lp_build_const_int_vec(bld->gallivm, i32.type, 8); 512 s_ipart = LLVMBuildAShr(builder, s, i32_c8, ""); 513 if (dims >= 2) 514 t_ipart = LLVMBuildAShr(builder, t, i32_c8, ""); 515 if (dims >= 3) 516 r_ipart = LLVMBuildAShr(builder, r, i32_c8, ""); 517 518 /* compute fractional part (AND with 0xff) */ 519 i32_c255 = lp_build_const_int_vec(bld->gallivm, i32.type, 255); 520 s_fpart = LLVMBuildAnd(builder, s, i32_c255, ""); 521 if (dims >= 2) 522 t_fpart = LLVMBuildAnd(builder, t, i32_c255, ""); 523 if (dims >= 3) 524 r_fpart = LLVMBuildAnd(builder, r, i32_c255, ""); 525 526 /* get pixel, row and image strides */ 527 x_stride = lp_build_const_vec(bld->gallivm, bld->int_coord_bld.type, 528 bld->format_desc->block.bits/8); 529 y_stride = row_stride_vec; 530 z_stride = img_stride_vec; 531 532 /* do texcoord wrapping and compute texel offsets */ 533 lp_build_sample_wrap_linear_int(bld, 534 bld->format_desc->block.width, 535 s_ipart, width_vec, x_stride, 536 bld->static_state->pot_width, 537 bld->static_state->wrap_s, 538 &x_offset0, &x_offset1, 539 &x_subcoord[0], &x_subcoord[1]); 540 for (z = 0; z < 2; z++) { 541 for (y = 0; y < 2; y++) { 542 offset[z][y][0] = x_offset0; 543 offset[z][y][1] = x_offset1; 544 } 545 } 546 547 if (dims >= 2) { 548 lp_build_sample_wrap_linear_int(bld, 549 bld->format_desc->block.height, 550 t_ipart, height_vec, y_stride, 551 bld->static_state->pot_height, 552 bld->static_state->wrap_t, 553 &y_offset0, &y_offset1, 554 &y_subcoord[0], &y_subcoord[1]); 555 556 for (z = 0; z < 2; z++) { 557 for (x = 0; x < 2; x++) { 558 offset[z][0][x] = lp_build_add(&bld->int_coord_bld, 559 offset[z][0][x], y_offset0); 560 offset[z][1][x] = lp_build_add(&bld->int_coord_bld, 561 offset[z][1][x], y_offset1); 562 } 563 } 564 } 565 566 if (dims >= 3) { 567 lp_build_sample_wrap_linear_int(bld, 568 bld->format_desc->block.height, 569 r_ipart, depth_vec, z_stride, 570 bld->static_state->pot_depth, 571 bld->static_state->wrap_r, 572 &z_offset0, &z_offset1, 573 &z_subcoord[0], &z_subcoord[1]); 574 for (y = 0; y < 2; y++) { 575 for (x = 0; x < 2; x++) { 576 offset[0][y][x] = lp_build_add(&bld->int_coord_bld, 577 offset[0][y][x], z_offset0); 578 offset[1][y][x] = lp_build_add(&bld->int_coord_bld, 579 offset[1][y][x], z_offset1); 580 } 581 } 582 } 583 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) { 584 LLVMValueRef z_offset; 585 z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec); 586 for (y = 0; y < 2; y++) { 587 for (x = 0; x < 2; x++) { 588 /* The r coord is the cube face in [0,5] */ 589 offset[0][y][x] = lp_build_add(&bld->int_coord_bld, 590 offset[0][y][x], z_offset); 591 } 592 } 593 } 594 595 /* 596 * Transform 4 x i32 in 597 * 598 * s_fpart = {s0, s1, s2, s3} 599 * 600 * into 8 x i16 601 * 602 * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3} 603 * 604 * into two 8 x i16 605 * 606 * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1} 607 * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3} 608 * 609 * and likewise for t_fpart. There is no risk of loosing precision here 610 * since the fractional parts only use the lower 8bits. 611 */ 612 s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, ""); 613 if (dims >= 2) 614 t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, ""); 615 if (dims >= 3) 616 r_fpart = LLVMBuildBitCast(builder, r_fpart, h16_vec_type, ""); 617 618 { 619 LLVMTypeRef elem_type = LLVMInt32TypeInContext(bld->gallivm->context); 620 LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH]; 621 LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH]; 622 LLVMValueRef shuffle_lo; 623 LLVMValueRef shuffle_hi; 624 625 for (j = 0; j < h16.type.length; j += 4) { 626#ifdef PIPE_ARCH_LITTLE_ENDIAN 627 unsigned subindex = 0; 628#else 629 unsigned subindex = 1; 630#endif 631 LLVMValueRef index; 632 633 index = LLVMConstInt(elem_type, j/2 + subindex, 0); 634 for (i = 0; i < 4; ++i) 635 shuffles_lo[j + i] = index; 636 637 index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0); 638 for (i = 0; i < 4; ++i) 639 shuffles_hi[j + i] = index; 640 } 641 642 shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length); 643 shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length); 644 645 s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, 646 shuffle_lo, ""); 647 s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, 648 shuffle_hi, ""); 649 if (dims >= 2) { 650 t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, 651 shuffle_lo, ""); 652 t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, 653 shuffle_hi, ""); 654 } 655 if (dims >= 3) { 656 r_fpart_lo = LLVMBuildShuffleVector(builder, r_fpart, h16.undef, 657 shuffle_lo, ""); 658 r_fpart_hi = LLVMBuildShuffleVector(builder, r_fpart, h16.undef, 659 shuffle_hi, ""); 660 } 661 } 662 663 /* 664 * Fetch the pixels as 4 x 32bit (rgba order might differ): 665 * 666 * rgba0 rgba1 rgba2 rgba3 667 * 668 * bit cast them into 16 x u8 669 * 670 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3 671 * 672 * unpack them into two 8 x i16: 673 * 674 * r0 g0 b0 a0 r1 g1 b1 a1 675 * r2 g2 b2 a2 r3 g3 b3 a3 676 * 677 * The higher 8 bits of the resulting elements will be zero. 678 */ 679 numj = 1 + (dims >= 2); 680 numk = 1 + (dims >= 3); 681 682 for (k = 0; k < numk; k++) { 683 for (j = 0; j < numj; j++) { 684 for (i = 0; i < 2; i++) { 685 LLVMValueRef rgba8; 686 687 if (util_format_is_rgba8_variant(bld->format_desc)) { 688 /* 689 * Given the format is a rgba8, just read the pixels as is, 690 * without any swizzling. Swizzling will be done later. 691 */ 692 rgba8 = lp_build_gather(bld->gallivm, 693 bld->texel_type.length, 694 bld->format_desc->block.bits, 695 bld->texel_type.width, 696 data_ptr, offset[k][j][i]); 697 698 rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, ""); 699 } 700 else { 701 rgba8 = lp_build_fetch_rgba_aos(bld->gallivm, 702 bld->format_desc, 703 u8n.type, 704 data_ptr, offset[k][j][i], 705 x_subcoord[i], 706 y_subcoord[j]); 707 } 708 709 /* Expand one 4*rgba8 to two 2*rgba16 */ 710 lp_build_unpack2(bld->gallivm, u8n.type, h16.type, 711 rgba8, 712 &neighbors_lo[k][j][i], &neighbors_hi[k][j][i]); 713 } 714 } 715 } 716 717 /* 718 * Linear interpolation with 8.8 fixed point. 719 */ 720 if (dims == 1) { 721 /* 1-D lerp */ 722 packed_lo = lp_build_lerp(&h16, 723 s_fpart_lo, 724 neighbors_lo[0][0][0], 725 neighbors_lo[0][0][1]); 726 727 packed_hi = lp_build_lerp(&h16, 728 s_fpart_hi, 729 neighbors_hi[0][0][0], 730 neighbors_hi[0][0][1]); 731 } 732 else { 733 /* 2-D lerp */ 734 packed_lo = lp_build_lerp_2d(&h16, 735 s_fpart_lo, t_fpart_lo, 736 neighbors_lo[0][0][0], 737 neighbors_lo[0][0][1], 738 neighbors_lo[0][1][0], 739 neighbors_lo[0][1][1]); 740 741 packed_hi = lp_build_lerp_2d(&h16, 742 s_fpart_hi, t_fpart_hi, 743 neighbors_hi[0][0][0], 744 neighbors_hi[0][0][1], 745 neighbors_hi[0][1][0], 746 neighbors_hi[0][1][1]); 747 748 if (dims >= 3) { 749 LLVMValueRef packed_lo2, packed_hi2; 750 751 /* lerp in the second z slice */ 752 packed_lo2 = lp_build_lerp_2d(&h16, 753 s_fpart_lo, t_fpart_lo, 754 neighbors_lo[1][0][0], 755 neighbors_lo[1][0][1], 756 neighbors_lo[1][1][0], 757 neighbors_lo[1][1][1]); 758 759 packed_hi2 = lp_build_lerp_2d(&h16, 760 s_fpart_hi, t_fpart_hi, 761 neighbors_hi[1][0][0], 762 neighbors_hi[1][0][1], 763 neighbors_hi[1][1][0], 764 neighbors_hi[1][1][1]); 765 /* interp between two z slices */ 766 packed_lo = lp_build_lerp(&h16, r_fpart_lo, 767 packed_lo, packed_lo2); 768 packed_hi = lp_build_lerp(&h16, r_fpart_hi, 769 packed_hi, packed_hi2); 770 } 771 } 772 773 *colors_lo = packed_lo; 774 *colors_hi = packed_hi; 775} 776 777 778/** 779 * Sample the texture/mipmap using given image filter and mip filter. 780 * data0_ptr and data1_ptr point to the two mipmap levels to sample 781 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes. 782 * If we're using nearest miplevel sampling the '1' values will be null/unused. 783 */ 784static void 785lp_build_sample_mipmap(struct lp_build_sample_context *bld, 786 unsigned img_filter, 787 unsigned mip_filter, 788 LLVMValueRef s, 789 LLVMValueRef t, 790 LLVMValueRef r, 791 LLVMValueRef ilevel0, 792 LLVMValueRef ilevel1, 793 LLVMValueRef lod_fpart, 794 LLVMValueRef colors_lo_var, 795 LLVMValueRef colors_hi_var) 796{ 797 LLVMBuilderRef builder = bld->gallivm->builder; 798 LLVMValueRef size0; 799 LLVMValueRef size1; 800 LLVMValueRef row_stride0_vec; 801 LLVMValueRef row_stride1_vec; 802 LLVMValueRef img_stride0_vec; 803 LLVMValueRef img_stride1_vec; 804 LLVMValueRef data_ptr0; 805 LLVMValueRef data_ptr1; 806 LLVMValueRef colors0_lo, colors0_hi; 807 LLVMValueRef colors1_lo, colors1_hi; 808 809 /* sample the first mipmap level */ 810 lp_build_mipmap_level_sizes(bld, ilevel0, 811 &size0, 812 &row_stride0_vec, &img_stride0_vec); 813 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0); 814 if (img_filter == PIPE_TEX_FILTER_NEAREST) { 815 lp_build_sample_image_nearest(bld, 816 size0, 817 row_stride0_vec, img_stride0_vec, 818 data_ptr0, s, t, r, 819 &colors0_lo, &colors0_hi); 820 } 821 else { 822 assert(img_filter == PIPE_TEX_FILTER_LINEAR); 823 lp_build_sample_image_linear(bld, 824 size0, 825 row_stride0_vec, img_stride0_vec, 826 data_ptr0, s, t, r, 827 &colors0_lo, &colors0_hi); 828 } 829 830 /* Store the first level's colors in the output variables */ 831 LLVMBuildStore(builder, colors0_lo, colors_lo_var); 832 LLVMBuildStore(builder, colors0_hi, colors_hi_var); 833 834 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { 835 LLVMValueRef h16_scale = lp_build_const_float(bld->gallivm, 256.0); 836 LLVMTypeRef i32_type = LLVMIntTypeInContext(bld->gallivm->context, 32); 837 struct lp_build_if_state if_ctx; 838 LLVMValueRef need_lerp; 839 840 lod_fpart = LLVMBuildFMul(builder, lod_fpart, h16_scale, ""); 841 lod_fpart = LLVMBuildFPToSI(builder, lod_fpart, i32_type, "lod_fpart.fixed16"); 842 843 /* need_lerp = lod_fpart > 0 */ 844 need_lerp = LLVMBuildICmp(builder, LLVMIntSGT, 845 lod_fpart, LLVMConstNull(i32_type), 846 "need_lerp"); 847 848 lp_build_if(&if_ctx, bld->gallivm, need_lerp); 849 { 850 struct lp_build_context h16_bld; 851 852 lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16)); 853 854 /* sample the second mipmap level */ 855 lp_build_mipmap_level_sizes(bld, ilevel1, 856 &size1, 857 &row_stride1_vec, &img_stride1_vec); 858 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1); 859 if (img_filter == PIPE_TEX_FILTER_NEAREST) { 860 lp_build_sample_image_nearest(bld, 861 size1, 862 row_stride1_vec, img_stride1_vec, 863 data_ptr1, s, t, r, 864 &colors1_lo, &colors1_hi); 865 } 866 else { 867 lp_build_sample_image_linear(bld, 868 size1, 869 row_stride1_vec, img_stride1_vec, 870 data_ptr1, s, t, r, 871 &colors1_lo, &colors1_hi); 872 } 873 874 /* interpolate samples from the two mipmap levels */ 875 876 lod_fpart = LLVMBuildTrunc(builder, lod_fpart, h16_bld.elem_type, ""); 877 lod_fpart = lp_build_broadcast_scalar(&h16_bld, lod_fpart); 878 879#if HAVE_LLVM == 0x208 880 /* This is a work-around for a bug in LLVM 2.8. 881 * Evidently, something goes wrong in the construction of the 882 * lod_fpart short[8] vector. Adding this no-effect shuffle seems 883 * to force the vector to be properly constructed. 884 * Tested with mesa-demos/src/tests/mipmap_limits.c (press t, f). 885 */ 886 { 887 LLVMValueRef shuffles[8], shuffle; 888 int i; 889 assert(h16_bld.type.length <= Elements(shuffles)); 890 for (i = 0; i < h16_bld.type.length; i++) 891 shuffles[i] = lp_build_const_int32(bld->gallivm, 2 * (i & 1)); 892 shuffle = LLVMConstVector(shuffles, h16_bld.type.length); 893 lod_fpart = LLVMBuildShuffleVector(builder, 894 lod_fpart, lod_fpart, 895 shuffle, ""); 896 } 897#endif 898 899 colors0_lo = lp_build_lerp(&h16_bld, lod_fpart, 900 colors0_lo, colors1_lo); 901 colors0_hi = lp_build_lerp(&h16_bld, lod_fpart, 902 colors0_hi, colors1_hi); 903 904 LLVMBuildStore(builder, colors0_lo, colors_lo_var); 905 LLVMBuildStore(builder, colors0_hi, colors_hi_var); 906 } 907 lp_build_endif(&if_ctx); 908 } 909} 910 911 912 913/** 914 * Texture sampling in AoS format. Used when sampling common 32-bit/texel 915 * formats. 1D/2D/3D/cube texture supported. All mipmap sampling modes 916 * but only limited texture coord wrap modes. 917 */ 918void 919lp_build_sample_aos(struct lp_build_sample_context *bld, 920 unsigned unit, 921 LLVMValueRef s, 922 LLVMValueRef t, 923 LLVMValueRef r, 924 const LLVMValueRef *ddx, 925 const LLVMValueRef *ddy, 926 LLVMValueRef lod_bias, /* optional */ 927 LLVMValueRef explicit_lod, /* optional */ 928 LLVMValueRef texel_out[4]) 929{ 930 struct lp_build_context *int_bld = &bld->int_bld; 931 LLVMBuilderRef builder = bld->gallivm->builder; 932 const unsigned mip_filter = bld->static_state->min_mip_filter; 933 const unsigned min_filter = bld->static_state->min_img_filter; 934 const unsigned mag_filter = bld->static_state->mag_img_filter; 935 const unsigned dims = bld->dims; 936 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL; 937 LLVMValueRef ilevel0, ilevel1 = NULL; 938 LLVMValueRef packed, packed_lo, packed_hi; 939 LLVMValueRef unswizzled[4]; 940 LLVMValueRef face_ddx[4], face_ddy[4]; 941 struct lp_build_context h16_bld; 942 LLVMValueRef i32t_zero = lp_build_const_int32(bld->gallivm, 0); 943 944 /* we only support the common/simple wrap modes at this time */ 945 assert(lp_is_simple_wrap_mode(bld->static_state->wrap_s)); 946 if (dims >= 2) 947 assert(lp_is_simple_wrap_mode(bld->static_state->wrap_t)); 948 if (dims >= 3) 949 assert(lp_is_simple_wrap_mode(bld->static_state->wrap_r)); 950 951 952 /* make 16-bit fixed-pt builder context */ 953 lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16)); 954 955 /* cube face selection, compute pre-face coords, etc. */ 956 if (bld->static_state->target == PIPE_TEXTURE_CUBE) { 957 LLVMValueRef face, face_s, face_t; 958 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t); 959 s = face_s; /* vec */ 960 t = face_t; /* vec */ 961 /* use 'r' to indicate cube face */ 962 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */ 963 964 /* recompute ddx, ddy using the new (s,t) face texcoords */ 965 face_ddx[0] = lp_build_scalar_ddx(&bld->coord_bld, s); 966 face_ddx[1] = lp_build_scalar_ddx(&bld->coord_bld, t); 967 face_ddx[2] = NULL; 968 face_ddx[3] = NULL; 969 face_ddy[0] = lp_build_scalar_ddy(&bld->coord_bld, s); 970 face_ddy[1] = lp_build_scalar_ddy(&bld->coord_bld, t); 971 face_ddy[2] = NULL; 972 face_ddy[3] = NULL; 973 ddx = face_ddx; 974 ddy = face_ddy; 975 } 976 977 /* 978 * Compute the level of detail (float). 979 */ 980 if (min_filter != mag_filter || 981 mip_filter != PIPE_TEX_MIPFILTER_NONE) { 982 /* Need to compute lod either to choose mipmap levels or to 983 * distinguish between minification/magnification with one mipmap level. 984 */ 985 lp_build_lod_selector(bld, unit, ddx, ddy, 986 lod_bias, explicit_lod, 987 mip_filter, 988 &lod_ipart, &lod_fpart); 989 } else { 990 lod_ipart = i32t_zero; 991 } 992 993 /* 994 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1 995 */ 996 switch (mip_filter) { 997 default: 998 assert(0 && "bad mip_filter value in lp_build_sample_aos()"); 999 /* fall-through */ 1000 case PIPE_TEX_MIPFILTER_NONE: 1001 /* always use mip level 0 */ 1002 if (bld->static_state->target == PIPE_TEXTURE_CUBE) { 1003 /* XXX this is a work-around for an apparent bug in LLVM 2.7. 1004 * We should be able to set ilevel0 = const(0) but that causes 1005 * bad x86 code to be emitted. 1006 */ 1007 assert(lod_ipart); 1008 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0); 1009 } 1010 else { 1011 ilevel0 = i32t_zero; 1012 } 1013 break; 1014 case PIPE_TEX_MIPFILTER_NEAREST: 1015 assert(lod_ipart); 1016 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0); 1017 break; 1018 case PIPE_TEX_MIPFILTER_LINEAR: 1019 assert(lod_ipart); 1020 assert(lod_fpart); 1021 lp_build_linear_mip_levels(bld, unit, 1022 lod_ipart, &lod_fpart, 1023 &ilevel0, &ilevel1); 1024 break; 1025 } 1026 1027 /* 1028 * Get/interpolate texture colors. 1029 */ 1030 1031 packed_lo = lp_build_alloca(bld->gallivm, h16_bld.vec_type, "packed_lo"); 1032 packed_hi = lp_build_alloca(bld->gallivm, h16_bld.vec_type, "packed_hi"); 1033 1034 if (min_filter == mag_filter) { 1035 /* no need to distinquish between minification and magnification */ 1036 lp_build_sample_mipmap(bld, 1037 min_filter, mip_filter, 1038 s, t, r, 1039 ilevel0, ilevel1, lod_fpart, 1040 packed_lo, packed_hi); 1041 } 1042 else { 1043 /* Emit conditional to choose min image filter or mag image filter 1044 * depending on the lod being > 0 or <= 0, respectively. 1045 */ 1046 struct lp_build_if_state if_ctx; 1047 LLVMValueRef minify; 1048 1049 /* minify = lod >= 0.0 */ 1050 minify = LLVMBuildICmp(builder, LLVMIntSGE, 1051 lod_ipart, int_bld->zero, ""); 1052 1053 lp_build_if(&if_ctx, bld->gallivm, minify); 1054 { 1055 /* Use the minification filter */ 1056 lp_build_sample_mipmap(bld, 1057 min_filter, mip_filter, 1058 s, t, r, 1059 ilevel0, ilevel1, lod_fpart, 1060 packed_lo, packed_hi); 1061 } 1062 lp_build_else(&if_ctx); 1063 { 1064 /* Use the magnification filter */ 1065 lp_build_sample_mipmap(bld, 1066 mag_filter, PIPE_TEX_MIPFILTER_NONE, 1067 s, t, r, 1068 i32t_zero, NULL, NULL, 1069 packed_lo, packed_hi); 1070 } 1071 lp_build_endif(&if_ctx); 1072 } 1073 1074 /* 1075 * combine the values stored in 'packed_lo' and 'packed_hi' variables 1076 * into 'packed' 1077 */ 1078 packed = lp_build_pack2(bld->gallivm, 1079 h16_bld.type, lp_type_unorm(8), 1080 LLVMBuildLoad(builder, packed_lo, ""), 1081 LLVMBuildLoad(builder, packed_hi, "")); 1082 1083 /* 1084 * Convert to SoA and swizzle. 1085 */ 1086 lp_build_rgba8_to_f32_soa(bld->gallivm, 1087 bld->texel_type, 1088 packed, unswizzled); 1089 1090 if (util_format_is_rgba8_variant(bld->format_desc)) { 1091 lp_build_format_swizzle_soa(bld->format_desc, 1092 &bld->texel_bld, 1093 unswizzled, texel_out); 1094 } 1095 else { 1096 texel_out[0] = unswizzled[0]; 1097 texel_out[1] = unswizzled[1]; 1098 texel_out[2] = unswizzled[2]; 1099 texel_out[3] = unswizzled[3]; 1100 } 1101 1102 apply_sampler_swizzle(bld, texel_out); 1103} 1104