projects / mesa.git / blob
? search:


d1592c22f4399baea1d9ed53f4f3407cddd66d9d
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_sample_soa.c
1 /**************************************************************************
2 *
3 * Copyright 2009 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 */
34
35 #include "pipe/p_defines.h"
36 #include "pipe/p_state.h"
37 #include "util/u_debug.h"
38 #include "util/u_dump.h"
39 #include "util/u_memory.h"
40 #include "util/u_math.h"
41 #include "util/u_format.h"
42 #include "util/u_cpu_detect.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_logic.h"
49 #include "lp_bld_swizzle.h"
50 #include "lp_bld_pack.h"
51 #include "lp_bld_flow.h"
52 #include "lp_bld_format.h"
53 #include "lp_bld_sample.h"
54
55
56 /**
57 * Keep all information for sampling code generation in a single place.
58 */
59 struct lp_build_sample_context
60 {
61 LLVMBuilderRef builder;
62
63 const struct lp_sampler_static_state *static_state;
64
65 struct lp_sampler_dynamic_state *dynamic_state;
66
67 const struct util_format_description *format_desc;
68
69 /** regular scalar float type */
70 struct lp_type float_type;
71 struct lp_build_context float_bld;
72
73 /** regular scalar float type */
74 struct lp_type int_type;
75 struct lp_build_context int_bld;
76
77 /** Incoming coordinates type and build context */
78 struct lp_type coord_type;
79 struct lp_build_context coord_bld;
80
81 /** Unsigned integer coordinates */
82 struct lp_type uint_coord_type;
83 struct lp_build_context uint_coord_bld;
84
85 /** Signed integer coordinates */
86 struct lp_type int_coord_type;
87 struct lp_build_context int_coord_bld;
88
89 /** Output texels type and build context */
90 struct lp_type texel_type;
91 struct lp_build_context texel_bld;
92 };
93
94
95 /**
96 * Does the given texture wrap mode allow sampling the texture border color?
97 * XXX maybe move this into gallium util code.
98 */
99 static boolean
100 wrap_mode_uses_border_color(unsigned mode)
101 {
102 switch (mode) {
103 case PIPE_TEX_WRAP_REPEAT:
104 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
105 case PIPE_TEX_WRAP_MIRROR_REPEAT:
106 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
107 return FALSE;
108 case PIPE_TEX_WRAP_CLAMP:
109 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
110 case PIPE_TEX_WRAP_MIRROR_CLAMP:
111 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
112 return TRUE;
113 default:
114 assert(0 && "unexpected wrap mode");
115 return FALSE;
116 }
117 }
118
119
120 static LLVMValueRef
121 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
122 LLVMValueRef data_array, LLVMValueRef level)
123 {
124 LLVMValueRef indexes[2], data_ptr;
125 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
126 indexes[1] = level;
127 data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
128 data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
129 return data_ptr;
130 }
131
132
133 static LLVMValueRef
134 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
135 LLVMValueRef data_array, int level)
136 {
137 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
138 return lp_build_get_mipmap_level(bld, data_array, lvl);
139 }
140
141
142 /**
143 * Dereference stride_array[mipmap_level] array to get a stride.
144 * Return stride as a vector.
145 */
146 static LLVMValueRef
147 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
148 LLVMValueRef stride_array, LLVMValueRef level)
149 {
150 LLVMValueRef indexes[2], stride;
151 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
152 indexes[1] = level;
153 stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
154 stride = LLVMBuildLoad(bld->builder, stride, "");
155 stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
156 return stride;
157 }
158
159
160 /** Dereference stride_array[0] array to get a stride (as vector). */
161 static LLVMValueRef
162 lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld,
163 LLVMValueRef stride_array, int level)
164 {
165 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
166 return lp_build_get_level_stride_vec(bld, stride_array, lvl);
167 }
168
169
170 static int
171 texture_dims(enum pipe_texture_target tex)
172 {
173 switch (tex) {
174 case PIPE_TEXTURE_1D:
175 return 1;
176 case PIPE_TEXTURE_2D:
177 case PIPE_TEXTURE_CUBE:
178 return 2;
179 case PIPE_TEXTURE_3D:
180 return 3;
181 default:
182 assert(0 && "bad texture target in texture_dims()");
183 return 2;
184 }
185 }
186
187
188 static LLVMValueRef
189 lp_build_swizzle_chan_soa(struct lp_type type,
190 const LLVMValueRef *unswizzled,
191 enum util_format_swizzle swizzle)
192 {
193 switch (swizzle) {
194 case PIPE_SWIZZLE_RED:
195 case PIPE_SWIZZLE_GREEN:
196 case PIPE_SWIZZLE_BLUE:
197 case PIPE_SWIZZLE_ALPHA:
198 return unswizzled[swizzle];
199 case PIPE_SWIZZLE_ZERO:
200 return lp_build_zero(type);
201 case PIPE_SWIZZLE_ONE:
202 return lp_build_one(type);
203 default:
204 assert(0);
205 return lp_build_undef(type);
206 }
207 }
208
209
210 static void
211 lp_build_swizzle_soa(struct lp_build_sample_context *bld,
212 LLVMValueRef *texel)
213 {
214 LLVMValueRef unswizzled[4];
215 unsigned char swizzles[4];
216 unsigned chan;
217
218 for (chan = 0; chan < 4; ++chan) {
219 unswizzled[chan] = texel[chan];
220 }
221
222 swizzles[0] = bld->static_state->swizzle_r;
223 swizzles[1] = bld->static_state->swizzle_g;
224 swizzles[2] = bld->static_state->swizzle_b;
225 swizzles[3] = bld->static_state->swizzle_a;
226
227 for (chan = 0; chan < 4; ++chan) {
228 unsigned swizzle = swizzles[chan];
229 texel[chan] = lp_build_swizzle_chan_soa(bld->texel_type,
230 unswizzled, swizzle);
231 }
232 }
233
234
235
236 /**
237 * Generate code to fetch a texel from a texture at int coords (x, y, z).
238 * The computation depends on whether the texture is 1D, 2D or 3D.
239 * The result, texel, will be:
240 * texel[0] = red values
241 * texel[1] = green values
242 * texel[2] = blue values
243 * texel[3] = alpha values
244 */
245 static void
246 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
247 LLVMValueRef width,
248 LLVMValueRef height,
249 LLVMValueRef depth,
250 LLVMValueRef x,
251 LLVMValueRef y,
252 LLVMValueRef z,
253 LLVMValueRef y_stride,
254 LLVMValueRef z_stride,
255 LLVMValueRef data_ptr,
256 LLVMValueRef *texel)
257 {
258 const int dims = texture_dims(bld->static_state->target);
259 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
260 LLVMValueRef offset;
261 LLVMValueRef i, j;
262 LLVMValueRef use_border = NULL;
263
264 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
265 if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
266 LLVMValueRef b1, b2;
267 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
268 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
269 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
270 }
271
272 if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
273 LLVMValueRef b1, b2;
274 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
275 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
276 if (use_border) {
277 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
278 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
279 }
280 else {
281 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
282 }
283 }
284
285 if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
286 LLVMValueRef b1, b2;
287 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
288 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
289 if (use_border) {
290 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
291 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
292 }
293 else {
294 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
295 }
296 }
297
298 /*
299 * Describe the coordinates in terms of pixel blocks.
300 *
301 * TODO: pixel blocks are power of two. LLVM should convert rem/div to
302 * bit arithmetic. Verify this.
303 */
304
305 if (bld->format_desc->block.width == 1) {
306 i = bld->uint_coord_bld.zero;
307 }
308 else {
309 LLVMValueRef block_width = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.width);
310 i = LLVMBuildURem(bld->builder, x, block_width, "");
311 x = LLVMBuildUDiv(bld->builder, x, block_width, "");
312 }
313
314 if (bld->format_desc->block.height == 1) {
315 j = bld->uint_coord_bld.zero;
316 }
317 else {
318 LLVMValueRef block_height = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.height);
319 j = LLVMBuildURem(bld->builder, y, block_height, "");
320 y = LLVMBuildUDiv(bld->builder, y, block_height, "");
321 }
322
323 /* convert x,y,z coords to linear offset from start of texture, in bytes */
324 offset = lp_build_sample_offset(&bld->uint_coord_bld,
325 bld->format_desc,
326 x, y, z, y_stride, z_stride);
327
328 if (use_border) {
329 /* If we can sample the border color, it means that texcoords may
330 * lie outside the bounds of the texture image. We need to do
331 * something to prevent reading out of bounds and causing a segfault.
332 *
333 * Simply AND the texture coords with !use_border. This will cause
334 * coords which are out of bounds to become zero. Zero's guaranteed
335 * to be inside the texture image.
336 */
337 offset = lp_build_andc(&bld->uint_coord_bld, offset, use_border);
338 }
339
340 lp_build_fetch_rgba_soa(bld->builder,
341 bld->format_desc,
342 bld->texel_type,
343 data_ptr, offset,
344 i, j,
345 texel);
346
347 lp_build_swizzle_soa(bld, texel);
348
349 /*
350 * Note: if we find an app which frequently samples the texture border
351 * we might want to implement a true conditional here to avoid sampling
352 * the texture whenever possible (since that's quite a bit of code).
353 * Ex:
354 * if (use_border) {
355 * texel = border_color;
356 * }
357 * else {
358 * texel = sample_texture(coord);
359 * }
360 * As it is now, we always sample the texture, then selectively replace
361 * the texel color results with the border color.
362 */
363
364 if (use_border) {
365 /* select texel color or border color depending on use_border */
366 int chan;
367 for (chan = 0; chan < 4; chan++) {
368 LLVMValueRef border_chan =
369 lp_build_const_vec(bld->texel_type,
370 bld->static_state->border_color[chan]);
371 texel[chan] = lp_build_select(&bld->texel_bld, use_border,
372 border_chan, texel[chan]);
373 }
374 }
375 }
376
377
378 static LLVMValueRef
379 lp_build_sample_packed(struct lp_build_sample_context *bld,
380 LLVMValueRef x,
381 LLVMValueRef y,
382 LLVMValueRef y_stride,
383 LLVMValueRef data_array)
384 {
385 LLVMValueRef offset;
386 LLVMValueRef data_ptr;
387
388 offset = lp_build_sample_offset(&bld->uint_coord_bld,
389 bld->format_desc,
390 x, y, NULL, y_stride, NULL);
391
392 assert(bld->format_desc->block.width == 1);
393 assert(bld->format_desc->block.height == 1);
394 assert(bld->format_desc->block.bits <= bld->texel_type.width);
395
396 /* get pointer to mipmap level 0 data */
397 data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
398
399 return lp_build_gather(bld->builder,
400 bld->texel_type.length,
401 bld->format_desc->block.bits,
402 bld->texel_type.width,
403 data_ptr, offset);
404 }
405
406
407 /**
408 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
409 */
410 static LLVMValueRef
411 lp_build_coord_mirror(struct lp_build_sample_context *bld,
412 LLVMValueRef coord)
413 {
414 struct lp_build_context *coord_bld = &bld->coord_bld;
415 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
416 LLVMValueRef fract, flr, isOdd;
417
418 /* fract = coord - floor(coord) */
419 fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
420
421 /* flr = ifloor(coord); */
422 flr = lp_build_ifloor(coord_bld, coord);
423
424 /* isOdd = flr & 1 */
425 isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
426
427 /* make coord positive or negative depending on isOdd */
428 coord = lp_build_set_sign(coord_bld, fract, isOdd);
429
430 /* convert isOdd to float */
431 isOdd = lp_build_int_to_float(coord_bld, isOdd);
432
433 /* add isOdd to coord */
434 coord = lp_build_add(coord_bld, coord, isOdd);
435
436 return coord;
437 }
438
439
440 /**
441 * We only support a few wrap modes in lp_build_sample_wrap_int() at this time.
442 * Return whether the given mode is supported by that function.
443 */
444 static boolean
445 is_simple_wrap_mode(unsigned mode)
446 {
447 switch (mode) {
448 case PIPE_TEX_WRAP_REPEAT:
449 case PIPE_TEX_WRAP_CLAMP:
450 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
451 return TRUE;
452 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
453 default:
454 return FALSE;
455 }
456 }
457
458
459 /**
460 * Build LLVM code for texture wrap mode, for scaled integer texcoords.
461 * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
462 * \param length the texture size along one dimension
463 * \param is_pot if TRUE, length is a power of two
464 * \param wrap_mode one of PIPE_TEX_WRAP_x
465 */
466 static LLVMValueRef
467 lp_build_sample_wrap_int(struct lp_build_sample_context *bld,
468 LLVMValueRef coord,
469 LLVMValueRef length,
470 boolean is_pot,
471 unsigned wrap_mode)
472 {
473 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
474 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
475 LLVMValueRef length_minus_one;
476
477 length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
478
479 switch(wrap_mode) {
480 case PIPE_TEX_WRAP_REPEAT:
481 if(is_pot)
482 coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
483 else
484 /* Signed remainder won't give the right results for negative
485 * dividends but unsigned remainder does.*/
486 coord = LLVMBuildURem(bld->builder, coord, length, "");
487 break;
488
489 case PIPE_TEX_WRAP_CLAMP:
490 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
491 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
492 coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
493 coord = lp_build_min(int_coord_bld, coord, length_minus_one);
494 break;
495
496 case PIPE_TEX_WRAP_MIRROR_REPEAT:
497 case PIPE_TEX_WRAP_MIRROR_CLAMP:
498 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
499 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
500 /* FIXME */
501 _debug_printf("llvmpipe: failed to translate texture wrap mode %s\n",
502 util_dump_tex_wrap(wrap_mode, TRUE));
503 coord = lp_build_max(uint_coord_bld, coord, uint_coord_bld->zero);
504 coord = lp_build_min(uint_coord_bld, coord, length_minus_one);
505 break;
506
507 default:
508 assert(0);
509 }
510
511 return coord;
512 }
513
514
515 /**
516 * Build LLVM code for texture wrap mode for linear filtering.
517 * \param x0_out returns first integer texcoord
518 * \param x1_out returns second integer texcoord
519 * \param weight_out returns linear interpolation weight
520 */
521 static void
522 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
523 LLVMValueRef coord,
524 LLVMValueRef length,
525 boolean is_pot,
526 unsigned wrap_mode,
527 LLVMValueRef *x0_out,
528 LLVMValueRef *x1_out,
529 LLVMValueRef *weight_out)
530 {
531 struct lp_build_context *coord_bld = &bld->coord_bld;
532 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
533 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
534 LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0);
535 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
536 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
537 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
538 LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
539 LLVMValueRef coord0, coord1, weight;
540
541 switch(wrap_mode) {
542 case PIPE_TEX_WRAP_REPEAT:
543 /* mul by size and subtract 0.5 */
544 coord = lp_build_mul(coord_bld, coord, length_f);
545 coord = lp_build_sub(coord_bld, coord, half);
546 /* convert to int */
547 coord0 = lp_build_ifloor(coord_bld, coord);
548 coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
549 /* compute lerp weight */
550 weight = lp_build_fract(coord_bld, coord);
551 /* repeat wrap */
552 if (is_pot) {
553 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
554 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
555 }
556 else {
557 /* Signed remainder won't give the right results for negative
558 * dividends but unsigned remainder does.*/
559 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
560 coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
561 }
562 break;
563
564 case PIPE_TEX_WRAP_CLAMP:
565 if (bld->static_state->normalized_coords) {
566 coord = lp_build_mul(coord_bld, coord, length_f);
567 }
568 weight = lp_build_fract(coord_bld, coord);
569 coord0 = lp_build_clamp(coord_bld, coord, coord_bld->zero,
570 length_f_minus_one);
571 coord1 = lp_build_add(coord_bld, coord, coord_bld->one);
572 coord1 = lp_build_clamp(coord_bld, coord1, coord_bld->zero,
573 length_f_minus_one);
574 coord0 = lp_build_ifloor(coord_bld, coord0);
575 coord1 = lp_build_ifloor(coord_bld, coord1);
576 break;
577
578 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
579 if (bld->static_state->normalized_coords) {
580 /* clamp to [0,1] */
581 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
582 /* mul by tex size and subtract 0.5 */
583 coord = lp_build_mul(coord_bld, coord, length_f);
584 coord = lp_build_sub(coord_bld, coord, half);
585 }
586 else {
587 LLVMValueRef min, max;
588 /* clamp to [0.5, length - 0.5] */
589 min = lp_build_const_vec(coord_bld->type, 0.5F);
590 max = lp_build_sub(coord_bld, length_f, min);
591 coord = lp_build_clamp(coord_bld, coord, min, max);
592 }
593 /* compute lerp weight */
594 weight = lp_build_fract(coord_bld, coord);
595 /* coord0 = floor(coord); */
596 coord0 = lp_build_ifloor(coord_bld, coord);
597 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
598 /* coord0 = max(coord0, 0) */
599 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
600 /* coord1 = min(coord1, length-1) */
601 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
602 break;
603
604 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
605 {
606 LLVMValueRef min, max;
607 if (bld->static_state->normalized_coords) {
608 /* min = -1.0 / (2 * length) = -0.5 / length */
609 min = lp_build_mul(coord_bld,
610 lp_build_const_vec(coord_bld->type, -0.5F),
611 lp_build_rcp(coord_bld, length_f));
612 /* max = 1.0 - min */
613 max = lp_build_sub(coord_bld, coord_bld->one, min);
614 /* coord = clamp(coord, min, max) */
615 coord = lp_build_clamp(coord_bld, coord, min, max);
616 /* scale coord to length (and sub 0.5?) */
617 coord = lp_build_mul(coord_bld, coord, length_f);
618 coord = lp_build_sub(coord_bld, coord, half);
619 }
620 else {
621 /* clamp to [-0.5, length + 0.5] */
622 min = lp_build_const_vec(coord_bld->type, -0.5F);
623 max = lp_build_sub(coord_bld, length_f, min);
624 coord = lp_build_clamp(coord_bld, coord, min, max);
625 coord = lp_build_sub(coord_bld, coord, half);
626 }
627 /* compute lerp weight */
628 weight = lp_build_fract(coord_bld, coord);
629 /* convert to int */
630 coord0 = lp_build_ifloor(coord_bld, coord);
631 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
632 }
633 break;
634
635 case PIPE_TEX_WRAP_MIRROR_REPEAT:
636 /* compute mirror function */
637 coord = lp_build_coord_mirror(bld, coord);
638
639 /* scale coord to length */
640 coord = lp_build_mul(coord_bld, coord, length_f);
641 coord = lp_build_sub(coord_bld, coord, half);
642
643 /* compute lerp weight */
644 weight = lp_build_fract(coord_bld, coord);
645
646 /* convert to int coords */
647 coord0 = lp_build_ifloor(coord_bld, coord);
648 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
649
650 /* coord0 = max(coord0, 0) */
651 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
652 /* coord1 = min(coord1, length-1) */
653 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
654 break;
655
656 case PIPE_TEX_WRAP_MIRROR_CLAMP:
657 {
658 LLVMValueRef min, max;
659 /* min = 1.0 / (2 * length) */
660 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
661 /* max = 1.0 - min */
662 max = lp_build_sub(coord_bld, coord_bld->one, min);
663
664 coord = lp_build_abs(coord_bld, coord);
665 coord = lp_build_clamp(coord_bld, coord, min, max);
666 coord = lp_build_mul(coord_bld, coord, length_f);
667 if(0)coord = lp_build_sub(coord_bld, coord, half);
668 weight = lp_build_fract(coord_bld, coord);
669 coord0 = lp_build_ifloor(coord_bld, coord);
670 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
671 }
672 break;
673
674 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
675 {
676 LLVMValueRef min, max;
677 /* min = 1.0 / (2 * length) */
678 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
679 /* max = 1.0 - min */
680 max = lp_build_sub(coord_bld, coord_bld->one, min);
681
682 coord = lp_build_abs(coord_bld, coord);
683 coord = lp_build_clamp(coord_bld, coord, min, max);
684 coord = lp_build_mul(coord_bld, coord, length_f);
685 coord = lp_build_sub(coord_bld, coord, half);
686 weight = lp_build_fract(coord_bld, coord);
687 coord0 = lp_build_ifloor(coord_bld, coord);
688 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
689 }
690 break;
691
692 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
693 {
694 LLVMValueRef min, max;
695 /* min = -1.0 / (2 * length) = -0.5 / length */
696 min = lp_build_mul(coord_bld,
697 lp_build_const_vec(coord_bld->type, -0.5F),
698 lp_build_rcp(coord_bld, length_f));
699 /* max = 1.0 - min */
700 max = lp_build_sub(coord_bld, coord_bld->one, min);
701
702 coord = lp_build_abs(coord_bld, coord);
703 coord = lp_build_clamp(coord_bld, coord, min, max);
704 coord = lp_build_mul(coord_bld, coord, length_f);
705 coord = lp_build_sub(coord_bld, coord, half);
706 weight = lp_build_fract(coord_bld, coord);
707 coord0 = lp_build_ifloor(coord_bld, coord);
708 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
709 }
710 break;
711
712 default:
713 assert(0);
714 coord0 = NULL;
715 coord1 = NULL;
716 weight = NULL;
717 }
718
719 *x0_out = coord0;
720 *x1_out = coord1;
721 *weight_out = weight;
722 }
723
724
725 /**
726 * Build LLVM code for texture wrap mode for nearest filtering.
727 * \param coord the incoming texcoord (nominally in [0,1])
728 * \param length the texture size along one dimension, as int
729 * \param is_pot if TRUE, length is a power of two
730 * \param wrap_mode one of PIPE_TEX_WRAP_x
731 */
732 static LLVMValueRef
733 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
734 LLVMValueRef coord,
735 LLVMValueRef length,
736 boolean is_pot,
737 unsigned wrap_mode)
738 {
739 struct lp_build_context *coord_bld = &bld->coord_bld;
740 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
741 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
742 LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0);
743 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
744 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
745 LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
746 LLVMValueRef icoord;
747
748 switch(wrap_mode) {
749 case PIPE_TEX_WRAP_REPEAT:
750 coord = lp_build_mul(coord_bld, coord, length_f);
751 icoord = lp_build_ifloor(coord_bld, coord);
752 if (is_pot)
753 icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
754 else
755 /* Signed remainder won't give the right results for negative
756 * dividends but unsigned remainder does.*/
757 icoord = LLVMBuildURem(bld->builder, icoord, length, "");
758 break;
759
760 case PIPE_TEX_WRAP_CLAMP:
761 /* mul by size */
762 if (bld->static_state->normalized_coords) {
763 coord = lp_build_mul(coord_bld, coord, length_f);
764 }
765 /* floor */
766 icoord = lp_build_ifloor(coord_bld, coord);
767 /* clamp to [0, size-1]. Note: int coord builder type */
768 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
769 length_minus_one);
770 break;
771
772 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
773 {
774 LLVMValueRef min, max;
775 if (bld->static_state->normalized_coords) {
776 /* min = 1.0 / (2 * length) */
777 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
778 /* max = length - min */
779 max = lp_build_sub(coord_bld, length_f, min);
780 /* scale coord to length */
781 coord = lp_build_mul(coord_bld, coord, length_f);
782 }
783 else {
784 /* clamp to [0.5, length - 0.5] */
785 min = lp_build_const_vec(coord_bld->type, 0.5F);
786 max = lp_build_sub(coord_bld, length_f, min);
787 }
788 /* coord = clamp(coord, min, max) */
789 coord = lp_build_clamp(coord_bld, coord, min, max);
790 icoord = lp_build_ifloor(coord_bld, coord);
791 }
792 break;
793
794 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
795 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
796 {
797 LLVMValueRef min, max;
798 if (bld->static_state->normalized_coords) {
799 /* min = -1.0 / (2 * length) = -0.5 / length */
800 min = lp_build_mul(coord_bld,
801 lp_build_const_vec(coord_bld->type, -0.5F),
802 lp_build_rcp(coord_bld, length_f));
803 /* max = length - min */
804 max = lp_build_sub(coord_bld, length_f, min);
805 /* scale coord to length */
806 coord = lp_build_mul(coord_bld, coord, length_f);
807 }
808 else {
809 /* clamp to [-0.5, length + 0.5] */
810 min = lp_build_const_vec(coord_bld->type, -0.5F);
811 max = lp_build_sub(coord_bld, length_f, min);
812 }
813 /* coord = clamp(coord, min, max) */
814 coord = lp_build_clamp(coord_bld, coord, min, max);
815 icoord = lp_build_ifloor(coord_bld, coord);
816 }
817 break;
818
819 case PIPE_TEX_WRAP_MIRROR_REPEAT:
820 {
821 LLVMValueRef min, max;
822 /* min = 1.0 / (2 * length) */
823 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
824 /* max = length - min */
825 max = lp_build_sub(coord_bld, length_f, min);
826
827 /* compute mirror function */
828 coord = lp_build_coord_mirror(bld, coord);
829
830 /* scale coord to length */
831 coord = lp_build_mul(coord_bld, coord, length_f);
832
833 /* coord = clamp(coord, min, max) */
834 coord = lp_build_clamp(coord_bld, coord, min, max);
835 icoord = lp_build_ifloor(coord_bld, coord);
836 }
837 break;
838
839 case PIPE_TEX_WRAP_MIRROR_CLAMP:
840 coord = lp_build_abs(coord_bld, coord);
841 coord = lp_build_mul(coord_bld, coord, length_f);
842 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f_minus_one);
843 icoord = lp_build_ifloor(coord_bld, coord);
844 break;
845
846 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
847 {
848 LLVMValueRef min, max;
849 /* min = 1.0 / (2 * length) */
850 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
851 /* max = length - min */
852 max = lp_build_sub(coord_bld, length_f, min);
853
854 coord = lp_build_abs(coord_bld, coord);
855 coord = lp_build_mul(coord_bld, coord, length_f);
856 coord = lp_build_clamp(coord_bld, coord, min, max);
857 icoord = lp_build_ifloor(coord_bld, coord);
858 }
859 break;
860
861 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
862 {
863 LLVMValueRef min, max;
864 /* min = 1.0 / (2 * length) */
865 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
866 min = lp_build_negate(coord_bld, min);
867 /* max = length - min */
868 max = lp_build_sub(coord_bld, length_f, min);
869
870 coord = lp_build_abs(coord_bld, coord);
871 coord = lp_build_mul(coord_bld, coord, length_f);
872 coord = lp_build_clamp(coord_bld, coord, min, max);
873 icoord = lp_build_ifloor(coord_bld, coord);
874 }
875 break;
876
877 default:
878 assert(0);
879 icoord = NULL;
880 }
881
882 return icoord;
883 }
884
885
886 /**
887 * Codegen equivalent for u_minify().
888 * Return max(1, base_size >> level);
889 */
890 static LLVMValueRef
891 lp_build_minify(struct lp_build_sample_context *bld,
892 LLVMValueRef base_size,
893 LLVMValueRef level)
894 {
895 LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify");
896 size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
897 return size;
898 }
899
900
901 /**
902 * Generate code to compute texture level of detail (lambda).
903 * \param s vector of texcoord s values
904 * \param t vector of texcoord t values
905 * \param r vector of texcoord r values
906 * \param shader_lod_bias vector float with the shader lod bias,
907 * \param width scalar int texture width
908 * \param height scalar int texture height
909 * \param depth scalar int texture depth
910 */
911 static LLVMValueRef
912 lp_build_lod_selector(struct lp_build_sample_context *bld,
913 LLVMValueRef s,
914 LLVMValueRef t,
915 LLVMValueRef r,
916 const LLVMValueRef *ddx,
917 const LLVMValueRef *ddy,
918 LLVMValueRef shader_lod_bias,
919 LLVMValueRef width,
920 LLVMValueRef height,
921 LLVMValueRef depth)
922
923 {
924 if (bld->static_state->min_lod == bld->static_state->max_lod) {
925 /* User is forcing sampling from a particular mipmap level.
926 * This is hit during mipmap generation.
927 */
928 return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
929 }
930 else {
931 const int dims = texture_dims(bld->static_state->target);
932 struct lp_build_context *float_bld = &bld->float_bld;
933 LLVMValueRef sampler_lod_bias = LLVMConstReal(LLVMFloatType(),
934 bld->static_state->lod_bias);
935 LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(),
936 bld->static_state->min_lod);
937 LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(),
938 bld->static_state->max_lod);
939
940 LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
941 LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
942 LLVMValueRef rho, lod;
943
944 /*
945 * dsdx = abs(s[1] - s[0]);
946 * dsdy = abs(s[2] - s[0]);
947 * dtdx = abs(t[1] - t[0]);
948 * dtdy = abs(t[2] - t[0]);
949 * drdx = abs(r[1] - r[0]);
950 * drdy = abs(r[2] - r[0]);
951 */
952 dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
953 dsdx = lp_build_abs(float_bld, dsdx);
954 dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
955 dsdy = lp_build_abs(float_bld, dsdy);
956 if (dims > 1) {
957 dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
958 dtdx = lp_build_abs(float_bld, dtdx);
959 dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
960 dtdy = lp_build_abs(float_bld, dtdy);
961 if (dims > 2) {
962 drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
963 drdx = lp_build_abs(float_bld, drdx);
964 drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
965 drdy = lp_build_abs(float_bld, drdy);
966 }
967 }
968
969 /* Compute rho = max of all partial derivatives scaled by texture size.
970 * XXX this could be vectorized somewhat
971 */
972 rho = LLVMBuildMul(bld->builder,
973 lp_build_max(float_bld, dsdx, dsdy),
974 lp_build_int_to_float(float_bld, width), "");
975 if (dims > 1) {
976 LLVMValueRef max;
977 max = LLVMBuildMul(bld->builder,
978 lp_build_max(float_bld, dtdx, dtdy),
979 lp_build_int_to_float(float_bld, height), "");
980 rho = lp_build_max(float_bld, rho, max);
981 if (dims > 2) {
982 max = LLVMBuildMul(bld->builder,
983 lp_build_max(float_bld, drdx, drdy),
984 lp_build_int_to_float(float_bld, depth), "");
985 rho = lp_build_max(float_bld, rho, max);
986 }
987 }
988
989 /* compute lod = log2(rho) */
990 lod = lp_build_log2(float_bld, rho);
991
992 /* add sampler lod bias */
993 lod = LLVMBuildAdd(bld->builder, lod, sampler_lod_bias, "sampler LOD bias");
994
995 /* add shader lod bias */
996 /* XXX for now we take only the first element since our lod is scalar */
997 shader_lod_bias = LLVMBuildExtractElement(bld->builder, shader_lod_bias,
998 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
999 lod = LLVMBuildAdd(bld->builder, lod, shader_lod_bias, "shader LOD bias");
1000
1001 /* clamp lod */
1002 lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
1003
1004 return lod;
1005 }
1006 }
1007
1008
1009 /**
1010 * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
1011 * mipmap level index.
1012 * Note: this is all scalar code.
1013 * \param lod scalar float texture level of detail
1014 * \param level_out returns integer
1015 */
1016 static void
1017 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
1018 unsigned unit,
1019 LLVMValueRef lod,
1020 LLVMValueRef *level_out)
1021 {
1022 struct lp_build_context *float_bld = &bld->float_bld;
1023 struct lp_build_context *int_bld = &bld->int_bld;
1024 LLVMValueRef last_level, level;
1025
1026 LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
1027
1028 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
1029 bld->builder, unit);
1030
1031 /* convert float lod to integer */
1032 level = lp_build_iround(float_bld, lod);
1033
1034 /* clamp level to legal range of levels */
1035 *level_out = lp_build_clamp(int_bld, level, zero, last_level);
1036 }
1037
1038
1039 /**
1040 * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
1041 * two (adjacent) mipmap level indexes. Later, we'll sample from those
1042 * two mipmap levels and interpolate between them.
1043 */
1044 static void
1045 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
1046 unsigned unit,
1047 LLVMValueRef lod,
1048 LLVMValueRef *level0_out,
1049 LLVMValueRef *level1_out,
1050 LLVMValueRef *weight_out)
1051 {
1052 struct lp_build_context *float_bld = &bld->float_bld;
1053 struct lp_build_context *int_bld = &bld->int_bld;
1054 LLVMValueRef last_level, level;
1055
1056 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
1057 bld->builder, unit);
1058
1059 /* convert float lod to integer */
1060 level = lp_build_ifloor(float_bld, lod);
1061
1062 /* compute level 0 and clamp to legal range of levels */
1063 *level0_out = lp_build_clamp(int_bld, level,
1064 int_bld->zero,
1065 last_level);
1066 /* compute level 1 and clamp to legal range of levels */
1067 *level1_out = lp_build_add(int_bld, *level0_out, int_bld->one);
1068 *level1_out = lp_build_min(int_bld, *level1_out, last_level);
1069
1070 *weight_out = lp_build_fract(float_bld, lod);
1071 }
1072
1073
1074 /**
1075 * Generate code to sample a mipmap level with nearest filtering.
1076 * If sampling a cube texture, r = cube face in [0,5].
1077 */
1078 static void
1079 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
1080 LLVMValueRef width_vec,
1081 LLVMValueRef height_vec,
1082 LLVMValueRef depth_vec,
1083 LLVMValueRef row_stride_vec,
1084 LLVMValueRef img_stride_vec,
1085 LLVMValueRef data_ptr,
1086 LLVMValueRef s,
1087 LLVMValueRef t,
1088 LLVMValueRef r,
1089 LLVMValueRef colors_out[4])
1090 {
1091 const int dims = texture_dims(bld->static_state->target);
1092 LLVMValueRef x, y, z;
1093
1094 /*
1095 * Compute integer texcoords.
1096 */
1097 x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1098 bld->static_state->pot_width,
1099 bld->static_state->wrap_s);
1100 lp_build_name(x, "tex.x.wrapped");
1101
1102 if (dims >= 2) {
1103 y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1104 bld->static_state->pot_height,
1105 bld->static_state->wrap_t);
1106 lp_build_name(y, "tex.y.wrapped");
1107
1108 if (dims == 3) {
1109 z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1110 bld->static_state->pot_height,
1111 bld->static_state->wrap_r);
1112 lp_build_name(z, "tex.z.wrapped");
1113 }
1114 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1115 z = r;
1116 }
1117 else {
1118 z = NULL;
1119 }
1120 }
1121 else {
1122 y = z = NULL;
1123 }
1124
1125 /*
1126 * Get texture colors.
1127 */
1128 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1129 x, y, z,
1130 row_stride_vec, img_stride_vec,
1131 data_ptr, colors_out);
1132 }
1133
1134
1135 /**
1136 * Generate code to sample a mipmap level with linear filtering.
1137 * If sampling a cube texture, r = cube face in [0,5].
1138 */
1139 static void
1140 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1141 LLVMValueRef width_vec,
1142 LLVMValueRef height_vec,
1143 LLVMValueRef depth_vec,
1144 LLVMValueRef row_stride_vec,
1145 LLVMValueRef img_stride_vec,
1146 LLVMValueRef data_ptr,
1147 LLVMValueRef s,
1148 LLVMValueRef t,
1149 LLVMValueRef r,
1150 LLVMValueRef colors_out[4])
1151 {
1152 const int dims = texture_dims(bld->static_state->target);
1153 LLVMValueRef x0, y0, z0, x1, y1, z1;
1154 LLVMValueRef s_fpart, t_fpart, r_fpart;
1155 LLVMValueRef neighbors[2][2][4];
1156 int chan;
1157
1158 /*
1159 * Compute integer texcoords.
1160 */
1161 lp_build_sample_wrap_linear(bld, s, width_vec,
1162 bld->static_state->pot_width,
1163 bld->static_state->wrap_s,
1164 &x0, &x1, &s_fpart);
1165 lp_build_name(x0, "tex.x0.wrapped");
1166 lp_build_name(x1, "tex.x1.wrapped");
1167
1168 if (dims >= 2) {
1169 lp_build_sample_wrap_linear(bld, t, height_vec,
1170 bld->static_state->pot_height,
1171 bld->static_state->wrap_t,
1172 &y0, &y1, &t_fpart);
1173 lp_build_name(y0, "tex.y0.wrapped");
1174 lp_build_name(y1, "tex.y1.wrapped");
1175
1176 if (dims == 3) {
1177 lp_build_sample_wrap_linear(bld, r, depth_vec,
1178 bld->static_state->pot_depth,
1179 bld->static_state->wrap_r,
1180 &z0, &z1, &r_fpart);
1181 lp_build_name(z0, "tex.z0.wrapped");
1182 lp_build_name(z1, "tex.z1.wrapped");
1183 }
1184 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1185 z0 = z1 = r; /* cube face */
1186 r_fpart = NULL;
1187 }
1188 else {
1189 z0 = z1 = NULL;
1190 r_fpart = NULL;
1191 }
1192 }
1193 else {
1194 y0 = y1 = t_fpart = NULL;
1195 z0 = z1 = r_fpart = NULL;
1196 }
1197
1198 /*
1199 * Get texture colors.
1200 */
1201 /* get x0/x1 texels */
1202 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1203 x0, y0, z0,
1204 row_stride_vec, img_stride_vec,
1205 data_ptr, neighbors[0][0]);
1206 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1207 x1, y0, z0,
1208 row_stride_vec, img_stride_vec,
1209 data_ptr, neighbors[0][1]);
1210
1211 if (dims == 1) {
1212 /* Interpolate two samples from 1D image to produce one color */
1213 for (chan = 0; chan < 4; chan++) {
1214 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1215 neighbors[0][0][chan],
1216 neighbors[0][1][chan]);
1217 }
1218 }
1219 else {
1220 /* 2D/3D texture */
1221 LLVMValueRef colors0[4];
1222
1223 /* get x0/x1 texels at y1 */
1224 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1225 x0, y1, z0,
1226 row_stride_vec, img_stride_vec,
1227 data_ptr, neighbors[1][0]);
1228 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1229 x1, y1, z0,
1230 row_stride_vec, img_stride_vec,
1231 data_ptr, neighbors[1][1]);
1232
1233 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1234 for (chan = 0; chan < 4; chan++) {
1235 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1236 s_fpart, t_fpart,
1237 neighbors[0][0][chan],
1238 neighbors[0][1][chan],
1239 neighbors[1][0][chan],
1240 neighbors[1][1][chan]);
1241 }
1242
1243 if (dims == 3) {
1244 LLVMValueRef neighbors1[2][2][4];
1245 LLVMValueRef colors1[4];
1246
1247 /* get x0/x1/y0/y1 texels at z1 */
1248 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1249 x0, y0, z1,
1250 row_stride_vec, img_stride_vec,
1251 data_ptr, neighbors1[0][0]);
1252 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1253 x1, y0, z1,
1254 row_stride_vec, img_stride_vec,
1255 data_ptr, neighbors1[0][1]);
1256 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1257 x0, y1, z1,
1258 row_stride_vec, img_stride_vec,
1259 data_ptr, neighbors1[1][0]);
1260 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1261 x1, y1, z1,
1262 row_stride_vec, img_stride_vec,
1263 data_ptr, neighbors1[1][1]);
1264
1265 /* Bilinear interpolate the four samples from the second Z slice */
1266 for (chan = 0; chan < 4; chan++) {
1267 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1268 s_fpart, t_fpart,
1269 neighbors1[0][0][chan],
1270 neighbors1[0][1][chan],
1271 neighbors1[1][0][chan],
1272 neighbors1[1][1][chan]);
1273 }
1274
1275 /* Linearly interpolate the two samples from the two 3D slices */
1276 for (chan = 0; chan < 4; chan++) {
1277 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1278 r_fpart,
1279 colors0[chan], colors1[chan]);
1280 }
1281 }
1282 else {
1283 /* 2D tex */
1284 for (chan = 0; chan < 4; chan++) {
1285 colors_out[chan] = colors0[chan];
1286 }
1287 }
1288 }
1289 }
1290
1291
1292 /** Helper used by lp_build_cube_lookup() */
1293 static LLVMValueRef
1294 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
1295 {
1296 /* ima = -0.5 / abs(coord); */
1297 LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
1298 LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1299 LLVMValueRef ima = lp_build_mul(coord_bld, negHalf,
1300 lp_build_rcp(coord_bld, absCoord));
1301 return ima;
1302 }
1303
1304
1305 /**
1306 * Helper used by lp_build_cube_lookup()
1307 * \param sign scalar +1 or -1
1308 * \param coord float vector
1309 * \param ima float vector
1310 */
1311 static LLVMValueRef
1312 lp_build_cube_coord(struct lp_build_context *coord_bld,
1313 LLVMValueRef sign, int negate_coord,
1314 LLVMValueRef coord, LLVMValueRef ima)
1315 {
1316 /* return negate(coord) * ima * sign + 0.5; */
1317 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
1318 LLVMValueRef res;
1319
1320 assert(negate_coord == +1 || negate_coord == -1);
1321
1322 if (negate_coord == -1) {
1323 coord = lp_build_negate(coord_bld, coord);
1324 }
1325
1326 res = lp_build_mul(coord_bld, coord, ima);
1327 if (sign) {
1328 sign = lp_build_broadcast_scalar(coord_bld, sign);
1329 res = lp_build_mul(coord_bld, res, sign);
1330 }
1331 res = lp_build_add(coord_bld, res, half);
1332
1333 return res;
1334 }
1335
1336
1337 /** Helper used by lp_build_cube_lookup()
1338 * Return (major_coord >= 0) ? pos_face : neg_face;
1339 */
1340 static LLVMValueRef
1341 lp_build_cube_face(struct lp_build_sample_context *bld,
1342 LLVMValueRef major_coord,
1343 unsigned pos_face, unsigned neg_face)
1344 {
1345 LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1346 major_coord,
1347 bld->float_bld.zero, "");
1348 LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
1349 LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
1350 LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
1351 return res;
1352 }
1353
1354
1355
1356 /**
1357 * Generate code to do cube face selection and per-face texcoords.
1358 */
1359 static void
1360 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1361 LLVMValueRef s,
1362 LLVMValueRef t,
1363 LLVMValueRef r,
1364 LLVMValueRef *face,
1365 LLVMValueRef *face_s,
1366 LLVMValueRef *face_t)
1367 {
1368 struct lp_build_context *float_bld = &bld->float_bld;
1369 struct lp_build_context *coord_bld = &bld->coord_bld;
1370 LLVMValueRef rx, ry, rz;
1371 LLVMValueRef arx, ary, arz;
1372 LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
1373 LLVMValueRef arx_ge_ary, arx_ge_arz;
1374 LLVMValueRef ary_ge_arx, ary_ge_arz;
1375 LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
1376 LLVMValueRef rx_pos, ry_pos, rz_pos;
1377
1378 assert(bld->coord_bld.type.length == 4);
1379
1380 /*
1381 * Use the average of the four pixel's texcoords to choose the face.
1382 */
1383 rx = lp_build_mul(float_bld, c25,
1384 lp_build_sum_vector(&bld->coord_bld, s));
1385 ry = lp_build_mul(float_bld, c25,
1386 lp_build_sum_vector(&bld->coord_bld, t));
1387 rz = lp_build_mul(float_bld, c25,
1388 lp_build_sum_vector(&bld->coord_bld, r));
1389
1390 arx = lp_build_abs(float_bld, rx);
1391 ary = lp_build_abs(float_bld, ry);
1392 arz = lp_build_abs(float_bld, rz);
1393
1394 /*
1395 * Compare sign/magnitude of rx,ry,rz to determine face
1396 */
1397 arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
1398 arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
1399 ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
1400 ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
1401
1402 arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
1403 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1404
1405 rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
1406 ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
1407 rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
1408
1409 {
1410 struct lp_build_flow_context *flow_ctx;
1411 struct lp_build_if_state if_ctx;
1412
1413 flow_ctx = lp_build_flow_create(bld->builder);
1414 lp_build_flow_scope_begin(flow_ctx);
1415
1416 *face_s = bld->coord_bld.undef;
1417 *face_t = bld->coord_bld.undef;
1418 *face = bld->int_bld.undef;
1419
1420 lp_build_name(*face_s, "face_s");
1421 lp_build_name(*face_t, "face_t");
1422 lp_build_name(*face, "face");
1423
1424 lp_build_flow_scope_declare(flow_ctx, face_s);
1425 lp_build_flow_scope_declare(flow_ctx, face_t);
1426 lp_build_flow_scope_declare(flow_ctx, face);
1427
1428 lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
1429 {
1430 /* +/- X face */
1431 LLVMValueRef sign = lp_build_sgn(float_bld, rx);
1432 LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
1433 *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
1434 *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1435 *face = lp_build_cube_face(bld, rx,
1436 PIPE_TEX_FACE_POS_X,
1437 PIPE_TEX_FACE_NEG_X);
1438 }
1439 lp_build_else(&if_ctx);
1440 {
1441 struct lp_build_flow_context *flow_ctx2;
1442 struct lp_build_if_state if_ctx2;
1443
1444 LLVMValueRef face_s2 = bld->coord_bld.undef;
1445 LLVMValueRef face_t2 = bld->coord_bld.undef;
1446 LLVMValueRef face2 = bld->int_bld.undef;
1447
1448 flow_ctx2 = lp_build_flow_create(bld->builder);
1449 lp_build_flow_scope_begin(flow_ctx2);
1450 lp_build_flow_scope_declare(flow_ctx2, &face_s2);
1451 lp_build_flow_scope_declare(flow_ctx2, &face_t2);
1452 lp_build_flow_scope_declare(flow_ctx2, &face2);
1453
1454 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1455
1456 lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
1457 {
1458 /* +/- Y face */
1459 LLVMValueRef sign = lp_build_sgn(float_bld, ry);
1460 LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
1461 face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
1462 face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
1463 face2 = lp_build_cube_face(bld, ry,
1464 PIPE_TEX_FACE_POS_Y,
1465 PIPE_TEX_FACE_NEG_Y);
1466 }
1467 lp_build_else(&if_ctx2);
1468 {
1469 /* +/- Z face */
1470 LLVMValueRef sign = lp_build_sgn(float_bld, rz);
1471 LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
1472 face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
1473 face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1474 face2 = lp_build_cube_face(bld, rz,
1475 PIPE_TEX_FACE_POS_Z,
1476 PIPE_TEX_FACE_NEG_Z);
1477 }
1478 lp_build_endif(&if_ctx2);
1479 lp_build_flow_scope_end(flow_ctx2);
1480 lp_build_flow_destroy(flow_ctx2);
1481
1482 *face_s = face_s2;
1483 *face_t = face_t2;
1484 *face = face2;
1485 }
1486
1487 lp_build_endif(&if_ctx);
1488 lp_build_flow_scope_end(flow_ctx);
1489 lp_build_flow_destroy(flow_ctx);
1490 }
1491 }
1492
1493
1494
1495 /**
1496 * Sample the texture/mipmap using given image filter and mip filter.
1497 * data0_ptr and data1_ptr point to the two mipmap levels to sample
1498 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
1499 * If we're using nearest miplevel sampling the '1' values will be null/unused.
1500 */
1501 static void
1502 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
1503 unsigned img_filter,
1504 unsigned mip_filter,
1505 LLVMValueRef s,
1506 LLVMValueRef t,
1507 LLVMValueRef r,
1508 LLVMValueRef lod_fpart,
1509 LLVMValueRef width0_vec,
1510 LLVMValueRef width1_vec,
1511 LLVMValueRef height0_vec,
1512 LLVMValueRef height1_vec,
1513 LLVMValueRef depth0_vec,
1514 LLVMValueRef depth1_vec,
1515 LLVMValueRef row_stride0_vec,
1516 LLVMValueRef row_stride1_vec,
1517 LLVMValueRef img_stride0_vec,
1518 LLVMValueRef img_stride1_vec,
1519 LLVMValueRef data_ptr0,
1520 LLVMValueRef data_ptr1,
1521 LLVMValueRef *colors_out)
1522 {
1523 LLVMValueRef colors0[4], colors1[4];
1524 int chan;
1525
1526 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1527 lp_build_sample_image_nearest(bld,
1528 width0_vec, height0_vec, depth0_vec,
1529 row_stride0_vec, img_stride0_vec,
1530 data_ptr0, s, t, r, colors0);
1531
1532 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1533 /* sample the second mipmap level, and interp */
1534 lp_build_sample_image_nearest(bld,
1535 width1_vec, height1_vec, depth1_vec,
1536 row_stride1_vec, img_stride1_vec,
1537 data_ptr1, s, t, r, colors1);
1538 }
1539 }
1540 else {
1541 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1542
1543 lp_build_sample_image_linear(bld,
1544 width0_vec, height0_vec, depth0_vec,
1545 row_stride0_vec, img_stride0_vec,
1546 data_ptr0, s, t, r, colors0);
1547
1548 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1549 /* sample the second mipmap level, and interp */
1550 lp_build_sample_image_linear(bld,
1551 width1_vec, height1_vec, depth1_vec,
1552 row_stride1_vec, img_stride1_vec,
1553 data_ptr1, s, t, r, colors1);
1554 }
1555 }
1556
1557 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1558 /* interpolate samples from the two mipmap levels */
1559 for (chan = 0; chan < 4; chan++) {
1560 colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1561 colors0[chan], colors1[chan]);
1562 }
1563 }
1564 else {
1565 /* use first/only level's colors */
1566 for (chan = 0; chan < 4; chan++) {
1567 colors_out[chan] = colors0[chan];
1568 }
1569 }
1570 }
1571
1572
1573
1574 /**
1575 * General texture sampling codegen.
1576 * This function handles texture sampling for all texture targets (1D,
1577 * 2D, 3D, cube) and all filtering modes.
1578 */
1579 static void
1580 lp_build_sample_general(struct lp_build_sample_context *bld,
1581 unsigned unit,
1582 LLVMValueRef s,
1583 LLVMValueRef t,
1584 LLVMValueRef r,
1585 const LLVMValueRef *ddx,
1586 const LLVMValueRef *ddy,
1587 LLVMValueRef lodbias,
1588 LLVMValueRef width,
1589 LLVMValueRef height,
1590 LLVMValueRef depth,
1591 LLVMValueRef width_vec,
1592 LLVMValueRef height_vec,
1593 LLVMValueRef depth_vec,
1594 LLVMValueRef row_stride_array,
1595 LLVMValueRef img_stride_array,
1596 LLVMValueRef data_array,
1597 LLVMValueRef *colors_out)
1598 {
1599 struct lp_build_context *float_bld = &bld->float_bld;
1600 const unsigned mip_filter = bld->static_state->min_mip_filter;
1601 const unsigned min_filter = bld->static_state->min_img_filter;
1602 const unsigned mag_filter = bld->static_state->mag_img_filter;
1603 const int dims = texture_dims(bld->static_state->target);
1604 LLVMValueRef lod = NULL, lod_fpart = NULL;
1605 LLVMValueRef ilevel0, ilevel1 = NULL, ilevel0_vec, ilevel1_vec = NULL;
1606 LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1607 LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1608 LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1609 LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1610 LLVMValueRef data_ptr0, data_ptr1 = NULL;
1611
1612 /*
1613 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1614 mip_filter, min_filter, mag_filter);
1615 */
1616
1617 /*
1618 * Compute the level of detail (float).
1619 */
1620 if (min_filter != mag_filter ||
1621 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1622 /* Need to compute lod either to choose mipmap levels or to
1623 * distinguish between minification/magnification with one mipmap level.
1624 */
1625 lod = lp_build_lod_selector(bld, s, t, r, ddx, ddy, lodbias,
1626 width, height, depth);
1627 }
1628
1629 /*
1630 * Compute integer mipmap level(s) to fetch texels from.
1631 */
1632 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1633 /* always use mip level 0 */
1634 ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1635 }
1636 else {
1637 if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1638 lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1639 }
1640 else {
1641 assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1642 lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1643 &lod_fpart);
1644 lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1645 }
1646 }
1647
1648 /*
1649 * Convert scalar integer mipmap levels into vectors.
1650 */
1651 ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1652 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1653 ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1654
1655 /*
1656 * Compute width, height at mipmap level 'ilevel0'
1657 */
1658 width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1659 if (dims >= 2) {
1660 height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1661 row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1662 ilevel0);
1663 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1664 img_stride0_vec = lp_build_get_level_stride_vec(bld,
1665 img_stride_array,
1666 ilevel0);
1667 if (dims == 3) {
1668 depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1669 }
1670 }
1671 }
1672 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1673 /* compute width, height, depth for second mipmap level at 'ilevel1' */
1674 width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1675 if (dims >= 2) {
1676 height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1677 row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1678 ilevel1);
1679 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1680 img_stride1_vec = lp_build_get_level_stride_vec(bld,
1681 img_stride_array,
1682 ilevel1);
1683 if (dims ==3) {
1684 depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1685 }
1686 }
1687 }
1688 }
1689
1690 /*
1691 * Choose cube face, recompute per-face texcoords.
1692 */
1693 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1694 LLVMValueRef face, face_s, face_t;
1695 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
1696 s = face_s; /* vec */
1697 t = face_t; /* vec */
1698 /* use 'r' to indicate cube face */
1699 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
1700 }
1701
1702 /*
1703 * Get pointer(s) to image data for mipmap level(s).
1704 */
1705 data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1706 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1707 data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1708 }
1709
1710 /*
1711 * Get/interpolate texture colors.
1712 */
1713 if (min_filter == mag_filter) {
1714 /* no need to distinquish between minification and magnification */
1715 lp_build_sample_mipmap(bld, min_filter, mip_filter, s, t, r, lod_fpart,
1716 width0_vec, width1_vec,
1717 height0_vec, height1_vec,
1718 depth0_vec, depth1_vec,
1719 row_stride0_vec, row_stride1_vec,
1720 img_stride0_vec, img_stride1_vec,
1721 data_ptr0, data_ptr1,
1722 colors_out);
1723 }
1724 else {
1725 /* Emit conditional to choose min image filter or mag image filter
1726 * depending on the lod being >0 or <= 0, respectively.
1727 */
1728 struct lp_build_flow_context *flow_ctx;
1729 struct lp_build_if_state if_ctx;
1730 LLVMValueRef minify;
1731
1732 flow_ctx = lp_build_flow_create(bld->builder);
1733 lp_build_flow_scope_begin(flow_ctx);
1734
1735 lp_build_flow_scope_declare(flow_ctx, &colors_out[0]);
1736 lp_build_flow_scope_declare(flow_ctx, &colors_out[1]);
1737 lp_build_flow_scope_declare(flow_ctx, &colors_out[2]);
1738 lp_build_flow_scope_declare(flow_ctx, &colors_out[3]);
1739
1740 /* minify = lod > 0.0 */
1741 minify = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1742 lod, float_bld->zero, "");
1743
1744 lp_build_if(&if_ctx, flow_ctx, bld->builder, minify);
1745 {
1746 /* Use the minification filter */
1747 lp_build_sample_mipmap(bld, min_filter, mip_filter,
1748 s, t, r, lod_fpart,
1749 width0_vec, width1_vec,
1750 height0_vec, height1_vec,
1751 depth0_vec, depth1_vec,
1752 row_stride0_vec, row_stride1_vec,
1753 img_stride0_vec, img_stride1_vec,
1754 data_ptr0, data_ptr1,
1755 colors_out);
1756 }
1757 lp_build_else(&if_ctx);
1758 {
1759 /* Use the magnification filter */
1760 lp_build_sample_mipmap(bld, mag_filter, mip_filter,
1761 s, t, r, lod_fpart,
1762 width0_vec, width1_vec,
1763 height0_vec, height1_vec,
1764 depth0_vec, depth1_vec,
1765 row_stride0_vec, row_stride1_vec,
1766 img_stride0_vec, img_stride1_vec,
1767 data_ptr0, data_ptr1,
1768 colors_out);
1769 }
1770 lp_build_endif(&if_ctx);
1771
1772 lp_build_flow_scope_end(flow_ctx);
1773 lp_build_flow_destroy(flow_ctx);
1774 }
1775 }
1776
1777
1778
1779 static void
1780 lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
1781 struct lp_type dst_type,
1782 LLVMValueRef packed,
1783 LLVMValueRef *rgba)
1784 {
1785 LLVMValueRef mask = lp_build_const_int_vec(dst_type, 0xff);
1786 unsigned chan;
1787
1788 /* Decode the input vector components */
1789 for (chan = 0; chan < 4; ++chan) {
1790 unsigned start = chan*8;
1791 unsigned stop = start + 8;
1792 LLVMValueRef input;
1793
1794 input = packed;
1795
1796 if(start)
1797 input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(dst_type, start), "");
1798
1799 if(stop < 32)
1800 input = LLVMBuildAnd(builder, input, mask, "");
1801
1802 input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
1803
1804 rgba[chan] = input;
1805 }
1806 }
1807
1808
1809 static void
1810 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1811 LLVMValueRef s,
1812 LLVMValueRef t,
1813 LLVMValueRef width,
1814 LLVMValueRef height,
1815 LLVMValueRef stride_array,
1816 LLVMValueRef data_array,
1817 LLVMValueRef *texel)
1818 {
1819 LLVMBuilderRef builder = bld->builder;
1820 struct lp_build_context i32, h16, u8n;
1821 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1822 LLVMValueRef i32_c8, i32_c128, i32_c255;
1823 LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1824 LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1825 LLVMValueRef x0, x1;
1826 LLVMValueRef y0, y1;
1827 LLVMValueRef neighbors[2][2];
1828 LLVMValueRef neighbors_lo[2][2];
1829 LLVMValueRef neighbors_hi[2][2];
1830 LLVMValueRef packed, packed_lo, packed_hi;
1831 LLVMValueRef unswizzled[4];
1832 LLVMValueRef stride;
1833
1834 assert(bld->static_state->target == PIPE_TEXTURE_2D);
1835 assert(bld->static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR);
1836 assert(bld->static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR);
1837 assert(bld->static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE);
1838
1839 lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1840 lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1841 lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1842
1843 i32_vec_type = lp_build_vec_type(i32.type);
1844 h16_vec_type = lp_build_vec_type(h16.type);
1845 u8n_vec_type = lp_build_vec_type(u8n.type);
1846
1847 if (bld->static_state->normalized_coords) {
1848 LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1849 LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1850 LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1851 s = lp_build_mul(&bld->coord_bld, s, fp_width);
1852 t = lp_build_mul(&bld->coord_bld, t, fp_height);
1853 }
1854
1855 /* scale coords by 256 (8 fractional bits) */
1856 s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1857 t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1858
1859 /* convert float to int */
1860 s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1861 t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1862
1863 /* subtract 0.5 (add -128) */
1864 i32_c128 = lp_build_const_int_vec(i32.type, -128);
1865 s = LLVMBuildAdd(builder, s, i32_c128, "");
1866 t = LLVMBuildAdd(builder, t, i32_c128, "");
1867
1868 /* compute floor (shift right 8) */
1869 i32_c8 = lp_build_const_int_vec(i32.type, 8);
1870 s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1871 t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1872
1873 /* compute fractional part (AND with 0xff) */
1874 i32_c255 = lp_build_const_int_vec(i32.type, 255);
1875 s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1876 t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1877
1878 x0 = s_ipart;
1879 y0 = t_ipart;
1880
1881 x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
1882 y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
1883
1884 x0 = lp_build_sample_wrap_int(bld, x0, width, bld->static_state->pot_width,
1885 bld->static_state->wrap_s);
1886 y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
1887 bld->static_state->wrap_t);
1888
1889 x1 = lp_build_sample_wrap_int(bld, x1, width, bld->static_state->pot_width,
1890 bld->static_state->wrap_s);
1891 y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
1892 bld->static_state->wrap_t);
1893
1894 /*
1895 * Transform 4 x i32 in
1896 *
1897 * s_fpart = {s0, s1, s2, s3}
1898 *
1899 * into 8 x i16
1900 *
1901 * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1902 *
1903 * into two 8 x i16
1904 *
1905 * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1906 * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1907 *
1908 * and likewise for t_fpart. There is no risk of loosing precision here
1909 * since the fractional parts only use the lower 8bits.
1910 */
1911
1912 s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1913 t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1914
1915 {
1916 LLVMTypeRef elem_type = LLVMInt32Type();
1917 LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1918 LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1919 LLVMValueRef shuffle_lo;
1920 LLVMValueRef shuffle_hi;
1921 unsigned i, j;
1922
1923 for(j = 0; j < h16.type.length; j += 4) {
1924 unsigned subindex = util_cpu_caps.little_endian ? 0 : 1;
1925 LLVMValueRef index;
1926
1927 index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1928 for(i = 0; i < 4; ++i)
1929 shuffles_lo[j + i] = index;
1930
1931 index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1932 for(i = 0; i < 4; ++i)
1933 shuffles_hi[j + i] = index;
1934 }
1935
1936 shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1937 shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1938
1939 s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1940 t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1941 s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1942 t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1943 }
1944
1945 stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
1946
1947 /*
1948 * Fetch the pixels as 4 x 32bit (rgba order might differ):
1949 *
1950 * rgba0 rgba1 rgba2 rgba3
1951 *
1952 * bit cast them into 16 x u8
1953 *
1954 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1955 *
1956 * unpack them into two 8 x i16:
1957 *
1958 * r0 g0 b0 a0 r1 g1 b1 a1
1959 * r2 g2 b2 a2 r3 g3 b3 a3
1960 *
1961 * The higher 8 bits of the resulting elements will be zero.
1962 */
1963
1964 neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array);
1965 neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array);
1966 neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array);
1967 neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array);
1968
1969 neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, "");
1970 neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, "");
1971 neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, "");
1972 neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, "");
1973
1974 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]);
1975 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]);
1976 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]);
1977 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]);
1978
1979 /*
1980 * Linear interpolate with 8.8 fixed point.
1981 */
1982
1983 packed_lo = lp_build_lerp_2d(&h16,
1984 s_fpart_lo, t_fpart_lo,
1985 neighbors_lo[0][0],
1986 neighbors_lo[0][1],
1987 neighbors_lo[1][0],
1988 neighbors_lo[1][1]);
1989
1990 packed_hi = lp_build_lerp_2d(&h16,
1991 s_fpart_hi, t_fpart_hi,
1992 neighbors_hi[0][0],
1993 neighbors_hi[0][1],
1994 neighbors_hi[1][0],
1995 neighbors_hi[1][1]);
1996
1997 packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
1998
1999 /*
2000 * Convert to SoA and swizzle.
2001 */
2002
2003 packed = LLVMBuildBitCast(builder, packed, i32_vec_type, "");
2004
2005 lp_build_rgba8_to_f32_soa(bld->builder,
2006 bld->texel_type,
2007 packed, unswizzled);
2008
2009 lp_build_format_swizzle_soa(bld->format_desc,
2010 bld->texel_type, unswizzled,
2011 texel);
2012
2013 lp_build_swizzle_soa(bld, texel);
2014 }
2015
2016
2017 static void
2018 lp_build_sample_compare(struct lp_build_sample_context *bld,
2019 LLVMValueRef p,
2020 LLVMValueRef *texel)
2021 {
2022 struct lp_build_context *texel_bld = &bld->texel_bld;
2023 LLVMValueRef res;
2024 unsigned chan;
2025
2026 if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
2027 return;
2028
2029 /* TODO: Compare before swizzling, to avoid redundant computations */
2030 res = NULL;
2031 for(chan = 0; chan < 4; ++chan) {
2032 LLVMValueRef cmp;
2033 cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
2034 cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
2035
2036 if(res)
2037 res = lp_build_add(texel_bld, res, cmp);
2038 else
2039 res = cmp;
2040 }
2041
2042 assert(res);
2043 res = lp_build_mul(texel_bld, res, lp_build_const_vec(texel_bld->type, 0.25));
2044
2045 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
2046 for(chan = 0; chan < 3; ++chan)
2047 texel[chan] = res;
2048 texel[3] = texel_bld->one;
2049 }
2050
2051
2052 /**
2053 * Just set texels to white instead of actually sampling the texture.
2054 * For debugging.
2055 */
2056 static void
2057 lp_build_sample_nop(struct lp_build_sample_context *bld,
2058 LLVMValueRef *texel)
2059 {
2060 struct lp_build_context *texel_bld = &bld->texel_bld;
2061 unsigned chan;
2062
2063 for (chan = 0; chan < 4; chan++) {
2064 /*lp_bld_mov(texel_bld, texel, texel_bld->one);*/
2065 texel[chan] = texel_bld->one;
2066 }
2067 }
2068
2069
2070 /**
2071 * Build texture sampling code.
2072 * 'texel' will return a vector of four LLVMValueRefs corresponding to
2073 * R, G, B, A.
2074 * \param type vector float type to use for coords, etc.
2075 */
2076 void
2077 lp_build_sample_soa(LLVMBuilderRef builder,
2078 const struct lp_sampler_static_state *static_state,
2079 struct lp_sampler_dynamic_state *dynamic_state,
2080 struct lp_type type,
2081 unsigned unit,
2082 unsigned num_coords,
2083 const LLVMValueRef *coords,
2084 const LLVMValueRef *ddx,
2085 const LLVMValueRef *ddy,
2086 LLVMValueRef lodbias,
2087 LLVMValueRef *texel)
2088 {
2089 struct lp_build_sample_context bld;
2090 LLVMValueRef width, width_vec;
2091 LLVMValueRef height, height_vec;
2092 LLVMValueRef depth, depth_vec;
2093 LLVMValueRef row_stride_array, img_stride_array;
2094 LLVMValueRef data_array;
2095 LLVMValueRef s;
2096 LLVMValueRef t;
2097 LLVMValueRef r;
2098
2099 /* Setup our build context */
2100 memset(&bld, 0, sizeof bld);
2101 bld.builder = builder;
2102 bld.static_state = static_state;
2103 bld.dynamic_state = dynamic_state;
2104 bld.format_desc = util_format_description(static_state->format);
2105
2106 bld.float_type = lp_type_float(32);
2107 bld.int_type = lp_type_int(32);
2108 bld.coord_type = type;
2109 bld.uint_coord_type = lp_uint_type(type);
2110 bld.int_coord_type = lp_int_type(type);
2111 bld.texel_type = type;
2112
2113 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
2114 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
2115 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
2116 lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
2117 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
2118 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
2119
2120 /* Get the dynamic state */
2121 width = dynamic_state->width(dynamic_state, builder, unit);
2122 height = dynamic_state->height(dynamic_state, builder, unit);
2123 depth = dynamic_state->depth(dynamic_state, builder, unit);
2124 row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
2125 img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit);
2126 data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
2127 /* Note that data_array is an array[level] of pointers to texture images */
2128
2129 s = coords[0];
2130 t = coords[1];
2131 r = coords[2];
2132
2133 width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
2134 height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
2135 depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
2136
2137 if (0) {
2138 /* For debug: no-op texture sampling */
2139 lp_build_sample_nop(&bld, texel);
2140 }
2141 else if (util_format_is_rgba8_variant(bld.format_desc) &&
2142 static_state->target == PIPE_TEXTURE_2D &&
2143 static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
2144 static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
2145 static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
2146 is_simple_wrap_mode(static_state->wrap_s) &&
2147 is_simple_wrap_mode(static_state->wrap_t)) {
2148 /* special case */
2149 lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
2150 row_stride_array, data_array, texel);
2151 }
2152 else {
2153 lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy, lodbias,
2154 width, height, depth,
2155 width_vec, height_vec, depth_vec,
2156 row_stride_array, img_stride_array,
2157 data_array,
2158 texel);
2159 }
2160
2161 lp_build_sample_compare(&bld, r, texel);
2162 }