projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Merge commit 'origin/gallium-context-transfers-2'
[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
189 /**
190 * Generate code to fetch a texel from a texture at int coords (x, y, z).
191 * The computation depends on whether the texture is 1D, 2D or 3D.
192 * The result, texel, will be:
193 * texel[0] = red values
194 * texel[1] = green values
195 * texel[2] = blue values
196 * texel[3] = alpha values
197 */
198 static void
199 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
200 LLVMValueRef width,
201 LLVMValueRef height,
202 LLVMValueRef depth,
203 LLVMValueRef x,
204 LLVMValueRef y,
205 LLVMValueRef z,
206 LLVMValueRef y_stride,
207 LLVMValueRef z_stride,
208 LLVMValueRef data_ptr,
209 LLVMValueRef *texel)
210 {
211 const int dims = texture_dims(bld->static_state->target);
212 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
213 LLVMValueRef offset;
214 LLVMValueRef packed;
215 LLVMValueRef use_border = NULL;
216
217 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
218 if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
219 LLVMValueRef b1, b2;
220 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
221 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
222 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
223 }
224
225 if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
226 LLVMValueRef b1, b2;
227 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
228 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
229 if (use_border) {
230 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
231 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
232 }
233 else {
234 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
235 }
236 }
237
238 if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
239 LLVMValueRef b1, b2;
240 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
241 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
242 if (use_border) {
243 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
244 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
245 }
246 else {
247 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
248 }
249 }
250
251 /*
252 * Note: if we find an app which frequently samples the texture border
253 * we might want to implement a true conditional here to avoid sampling
254 * the texture whenever possible (since that's quite a bit of code).
255 * Ex:
256 * if (use_border) {
257 * texel = border_color;
258 * }
259 * else {
260 * texel = sample_texture(coord);
261 * }
262 * As it is now, we always sample the texture, then selectively replace
263 * the texel color results with the border color.
264 */
265
266 /* convert x,y,z coords to linear offset from start of texture, in bytes */
267 offset = lp_build_sample_offset(&bld->uint_coord_bld,
268 bld->format_desc,
269 x, y, z, y_stride, z_stride);
270
271 assert(bld->format_desc->block.width == 1);
272 assert(bld->format_desc->block.height == 1);
273 assert(bld->format_desc->block.bits <= bld->texel_type.width);
274
275 /* gather the texels from the texture */
276 packed = lp_build_gather(bld->builder,
277 bld->texel_type.length,
278 bld->format_desc->block.bits,
279 bld->texel_type.width,
280 data_ptr, offset);
281
282 texel[0] = texel[1] = texel[2] = texel[3] = NULL;
283
284 /* convert texels to float rgba */
285 lp_build_unpack_rgba_soa(bld->builder,
286 bld->format_desc,
287 bld->texel_type,
288 packed, texel);
289
290 if (use_border) {
291 /* select texel color or border color depending on use_border */
292 int chan;
293 for (chan = 0; chan < 4; chan++) {
294 LLVMValueRef border_chan =
295 lp_build_const_scalar(bld->texel_type,
296 bld->static_state->border_color[chan]);
297 texel[chan] = lp_build_select(&bld->texel_bld, use_border,
298 border_chan, texel[chan]);
299 }
300 }
301 }
302
303
304 static LLVMValueRef
305 lp_build_sample_packed(struct lp_build_sample_context *bld,
306 LLVMValueRef x,
307 LLVMValueRef y,
308 LLVMValueRef y_stride,
309 LLVMValueRef data_array)
310 {
311 LLVMValueRef offset;
312 LLVMValueRef data_ptr;
313
314 offset = lp_build_sample_offset(&bld->uint_coord_bld,
315 bld->format_desc,
316 x, y, NULL, y_stride, NULL);
317
318 assert(bld->format_desc->block.width == 1);
319 assert(bld->format_desc->block.height == 1);
320 assert(bld->format_desc->block.bits <= bld->texel_type.width);
321
322 /* get pointer to mipmap level 0 data */
323 data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
324
325 return lp_build_gather(bld->builder,
326 bld->texel_type.length,
327 bld->format_desc->block.bits,
328 bld->texel_type.width,
329 data_ptr, offset);
330 }
331
332
333 /**
334 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
335 */
336 static LLVMValueRef
337 lp_build_coord_mirror(struct lp_build_sample_context *bld,
338 LLVMValueRef coord)
339 {
340 struct lp_build_context *coord_bld = &bld->coord_bld;
341 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
342 LLVMValueRef fract, flr, isOdd;
343
344 /* fract = coord - floor(coord) */
345 fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
346
347 /* flr = ifloor(coord); */
348 flr = lp_build_ifloor(coord_bld, coord);
349
350 /* isOdd = flr & 1 */
351 isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
352
353 /* make coord positive or negative depending on isOdd */
354 coord = lp_build_set_sign(coord_bld, fract, isOdd);
355
356 /* convert isOdd to float */
357 isOdd = lp_build_int_to_float(coord_bld, isOdd);
358
359 /* add isOdd to coord */
360 coord = lp_build_add(coord_bld, coord, isOdd);
361
362 return coord;
363 }
364
365
366 /**
367 * We only support a few wrap modes in lp_build_sample_wrap_int() at this time.
368 * Return whether the given mode is supported by that function.
369 */
370 static boolean
371 is_simple_wrap_mode(unsigned mode)
372 {
373 switch (mode) {
374 case PIPE_TEX_WRAP_REPEAT:
375 case PIPE_TEX_WRAP_CLAMP:
376 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
377 return TRUE;
378 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
379 default:
380 return FALSE;
381 }
382 }
383
384
385 /**
386 * Build LLVM code for texture wrap mode, for scaled integer texcoords.
387 * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
388 * \param length the texture size along one dimension
389 * \param is_pot if TRUE, length is a power of two
390 * \param wrap_mode one of PIPE_TEX_WRAP_x
391 */
392 static LLVMValueRef
393 lp_build_sample_wrap_int(struct lp_build_sample_context *bld,
394 LLVMValueRef coord,
395 LLVMValueRef length,
396 boolean is_pot,
397 unsigned wrap_mode)
398 {
399 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
400 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
401 LLVMValueRef length_minus_one;
402
403 length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
404
405 switch(wrap_mode) {
406 case PIPE_TEX_WRAP_REPEAT:
407 if(is_pot)
408 coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
409 else
410 /* Signed remainder won't give the right results for negative
411 * dividends but unsigned remainder does.*/
412 coord = LLVMBuildURem(bld->builder, coord, length, "");
413 break;
414
415 case PIPE_TEX_WRAP_CLAMP:
416 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
417 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
418 coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
419 coord = lp_build_min(int_coord_bld, coord, length_minus_one);
420 break;
421
422 case PIPE_TEX_WRAP_MIRROR_REPEAT:
423 case PIPE_TEX_WRAP_MIRROR_CLAMP:
424 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
425 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
426 /* FIXME */
427 _debug_printf("llvmpipe: failed to translate texture wrap mode %s\n",
428 util_dump_tex_wrap(wrap_mode, TRUE));
429 coord = lp_build_max(uint_coord_bld, coord, uint_coord_bld->zero);
430 coord = lp_build_min(uint_coord_bld, coord, length_minus_one);
431 break;
432
433 default:
434 assert(0);
435 }
436
437 return coord;
438 }
439
440
441 /**
442 * Build LLVM code for texture wrap mode for linear filtering.
443 * \param x0_out returns first integer texcoord
444 * \param x1_out returns second integer texcoord
445 * \param weight_out returns linear interpolation weight
446 */
447 static void
448 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
449 LLVMValueRef coord,
450 LLVMValueRef length,
451 boolean is_pot,
452 unsigned wrap_mode,
453 LLVMValueRef *x0_out,
454 LLVMValueRef *x1_out,
455 LLVMValueRef *weight_out)
456 {
457 struct lp_build_context *coord_bld = &bld->coord_bld;
458 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
459 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
460 LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0);
461 LLVMValueRef half = lp_build_const_scalar(coord_bld->type, 0.5);
462 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
463 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
464 LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
465 LLVMValueRef coord0, coord1, weight;
466
467 switch(wrap_mode) {
468 case PIPE_TEX_WRAP_REPEAT:
469 /* mul by size and subtract 0.5 */
470 coord = lp_build_mul(coord_bld, coord, length_f);
471 coord = lp_build_sub(coord_bld, coord, half);
472 /* convert to int */
473 coord0 = lp_build_ifloor(coord_bld, coord);
474 coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
475 /* compute lerp weight */
476 weight = lp_build_fract(coord_bld, coord);
477 /* repeat wrap */
478 if (is_pot) {
479 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
480 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
481 }
482 else {
483 /* Signed remainder won't give the right results for negative
484 * dividends but unsigned remainder does.*/
485 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
486 coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
487 }
488 break;
489
490 case PIPE_TEX_WRAP_CLAMP:
491 if (bld->static_state->normalized_coords) {
492 coord = lp_build_mul(coord_bld, coord, length_f);
493 }
494 weight = lp_build_fract(coord_bld, coord);
495 coord0 = lp_build_clamp(coord_bld, coord, coord_bld->zero,
496 length_f_minus_one);
497 coord1 = lp_build_add(coord_bld, coord, coord_bld->one);
498 coord1 = lp_build_clamp(coord_bld, coord1, coord_bld->zero,
499 length_f_minus_one);
500 coord0 = lp_build_ifloor(coord_bld, coord0);
501 coord1 = lp_build_ifloor(coord_bld, coord1);
502 break;
503
504 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
505 if (bld->static_state->normalized_coords) {
506 /* clamp to [0,1] */
507 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
508 /* mul by tex size and subtract 0.5 */
509 coord = lp_build_mul(coord_bld, coord, length_f);
510 coord = lp_build_sub(coord_bld, coord, half);
511 }
512 else {
513 LLVMValueRef min, max;
514 /* clamp to [0.5, length - 0.5] */
515 min = lp_build_const_scalar(coord_bld->type, 0.5F);
516 max = lp_build_sub(coord_bld, length_f, min);
517 coord = lp_build_clamp(coord_bld, coord, min, max);
518 }
519 /* compute lerp weight */
520 weight = lp_build_fract(coord_bld, coord);
521 /* coord0 = floor(coord); */
522 coord0 = lp_build_ifloor(coord_bld, coord);
523 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
524 /* coord0 = max(coord0, 0) */
525 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
526 /* coord1 = min(coord1, length-1) */
527 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
528 break;
529
530 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
531 {
532 LLVMValueRef min, max;
533 if (bld->static_state->normalized_coords) {
534 /* min = -1.0 / (2 * length) = -0.5 / length */
535 min = lp_build_mul(coord_bld,
536 lp_build_const_scalar(coord_bld->type, -0.5F),
537 lp_build_rcp(coord_bld, length_f));
538 /* max = 1.0 - min */
539 max = lp_build_sub(coord_bld, coord_bld->one, min);
540 /* coord = clamp(coord, min, max) */
541 coord = lp_build_clamp(coord_bld, coord, min, max);
542 /* scale coord to length (and sub 0.5?) */
543 coord = lp_build_mul(coord_bld, coord, length_f);
544 coord = lp_build_sub(coord_bld, coord, half);
545 }
546 else {
547 /* clamp to [-0.5, length + 0.5] */
548 min = lp_build_const_scalar(coord_bld->type, -0.5F);
549 max = lp_build_sub(coord_bld, length_f, min);
550 coord = lp_build_clamp(coord_bld, coord, min, max);
551 coord = lp_build_sub(coord_bld, coord, half);
552 }
553 /* compute lerp weight */
554 weight = lp_build_fract(coord_bld, coord);
555 /* convert to int */
556 coord0 = lp_build_ifloor(coord_bld, coord);
557 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
558 }
559 break;
560
561 case PIPE_TEX_WRAP_MIRROR_REPEAT:
562 /* compute mirror function */
563 coord = lp_build_coord_mirror(bld, coord);
564
565 /* scale coord to length */
566 coord = lp_build_mul(coord_bld, coord, length_f);
567 coord = lp_build_sub(coord_bld, coord, half);
568
569 /* compute lerp weight */
570 weight = lp_build_fract(coord_bld, coord);
571
572 /* convert to int coords */
573 coord0 = lp_build_ifloor(coord_bld, coord);
574 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
575
576 /* coord0 = max(coord0, 0) */
577 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
578 /* coord1 = min(coord1, length-1) */
579 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
580 break;
581
582 case PIPE_TEX_WRAP_MIRROR_CLAMP:
583 {
584 LLVMValueRef min, max;
585 /* min = 1.0 / (2 * length) */
586 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
587 /* max = 1.0 - min */
588 max = lp_build_sub(coord_bld, coord_bld->one, min);
589
590 coord = lp_build_abs(coord_bld, coord);
591 coord = lp_build_clamp(coord_bld, coord, min, max);
592 coord = lp_build_mul(coord_bld, coord, length_f);
593 if(0)coord = lp_build_sub(coord_bld, coord, half);
594 weight = lp_build_fract(coord_bld, coord);
595 coord0 = lp_build_ifloor(coord_bld, coord);
596 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
597 }
598 break;
599
600 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
601 {
602 LLVMValueRef min, max;
603 /* min = 1.0 / (2 * length) */
604 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
605 /* max = 1.0 - min */
606 max = lp_build_sub(coord_bld, coord_bld->one, min);
607
608 coord = lp_build_abs(coord_bld, coord);
609 coord = lp_build_clamp(coord_bld, coord, min, max);
610 coord = lp_build_mul(coord_bld, coord, length_f);
611 coord = lp_build_sub(coord_bld, coord, half);
612 weight = lp_build_fract(coord_bld, coord);
613 coord0 = lp_build_ifloor(coord_bld, coord);
614 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
615 }
616 break;
617
618 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
619 {
620 LLVMValueRef min, max;
621 /* min = -1.0 / (2 * length) = -0.5 / length */
622 min = lp_build_mul(coord_bld,
623 lp_build_const_scalar(coord_bld->type, -0.5F),
624 lp_build_rcp(coord_bld, length_f));
625 /* max = 1.0 - min */
626 max = lp_build_sub(coord_bld, coord_bld->one, min);
627
628 coord = lp_build_abs(coord_bld, coord);
629 coord = lp_build_clamp(coord_bld, coord, min, max);
630 coord = lp_build_mul(coord_bld, coord, length_f);
631 coord = lp_build_sub(coord_bld, coord, half);
632 weight = lp_build_fract(coord_bld, coord);
633 coord0 = lp_build_ifloor(coord_bld, coord);
634 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
635 }
636 break;
637
638 default:
639 assert(0);
640 coord0 = NULL;
641 coord1 = NULL;
642 weight = NULL;
643 }
644
645 *x0_out = coord0;
646 *x1_out = coord1;
647 *weight_out = weight;
648 }
649
650
651 /**
652 * Build LLVM code for texture wrap mode for nearest filtering.
653 * \param coord the incoming texcoord (nominally in [0,1])
654 * \param length the texture size along one dimension, as int
655 * \param is_pot if TRUE, length is a power of two
656 * \param wrap_mode one of PIPE_TEX_WRAP_x
657 */
658 static LLVMValueRef
659 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
660 LLVMValueRef coord,
661 LLVMValueRef length,
662 boolean is_pot,
663 unsigned wrap_mode)
664 {
665 struct lp_build_context *coord_bld = &bld->coord_bld;
666 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
667 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
668 LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0);
669 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
670 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
671 LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
672 LLVMValueRef icoord;
673
674 switch(wrap_mode) {
675 case PIPE_TEX_WRAP_REPEAT:
676 coord = lp_build_mul(coord_bld, coord, length_f);
677 icoord = lp_build_ifloor(coord_bld, coord);
678 if (is_pot)
679 icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
680 else
681 /* Signed remainder won't give the right results for negative
682 * dividends but unsigned remainder does.*/
683 icoord = LLVMBuildURem(bld->builder, icoord, length, "");
684 break;
685
686 case PIPE_TEX_WRAP_CLAMP:
687 /* mul by size */
688 if (bld->static_state->normalized_coords) {
689 coord = lp_build_mul(coord_bld, coord, length_f);
690 }
691 /* floor */
692 icoord = lp_build_ifloor(coord_bld, coord);
693 /* clamp to [0, size-1]. Note: int coord builder type */
694 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
695 length_minus_one);
696 break;
697
698 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
699 {
700 LLVMValueRef min, max;
701 if (bld->static_state->normalized_coords) {
702 /* min = 1.0 / (2 * length) */
703 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
704 /* max = length - min */
705 max = lp_build_sub(coord_bld, length_f, min);
706 /* scale coord to length */
707 coord = lp_build_mul(coord_bld, coord, length_f);
708 }
709 else {
710 /* clamp to [0.5, length - 0.5] */
711 min = lp_build_const_scalar(coord_bld->type, 0.5F);
712 max = lp_build_sub(coord_bld, length_f, min);
713 }
714 /* coord = clamp(coord, min, max) */
715 coord = lp_build_clamp(coord_bld, coord, min, max);
716 icoord = lp_build_ifloor(coord_bld, coord);
717 }
718 break;
719
720 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
721 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
722 {
723 LLVMValueRef min, max;
724 if (bld->static_state->normalized_coords) {
725 /* min = -1.0 / (2 * length) = -0.5 / length */
726 min = lp_build_mul(coord_bld,
727 lp_build_const_scalar(coord_bld->type, -0.5F),
728 lp_build_rcp(coord_bld, length_f));
729 /* max = length - min */
730 max = lp_build_sub(coord_bld, length_f, min);
731 /* scale coord to length */
732 coord = lp_build_mul(coord_bld, coord, length_f);
733 }
734 else {
735 /* clamp to [-0.5, length + 0.5] */
736 min = lp_build_const_scalar(coord_bld->type, -0.5F);
737 max = lp_build_sub(coord_bld, length_f, min);
738 }
739 /* coord = clamp(coord, min, max) */
740 coord = lp_build_clamp(coord_bld, coord, min, max);
741 icoord = lp_build_ifloor(coord_bld, coord);
742 }
743 break;
744
745 case PIPE_TEX_WRAP_MIRROR_REPEAT:
746 {
747 LLVMValueRef min, max;
748 /* min = 1.0 / (2 * length) */
749 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
750 /* max = length - min */
751 max = lp_build_sub(coord_bld, length_f, min);
752
753 /* compute mirror function */
754 coord = lp_build_coord_mirror(bld, coord);
755
756 /* scale coord to length */
757 coord = lp_build_mul(coord_bld, coord, length_f);
758
759 /* coord = clamp(coord, min, max) */
760 coord = lp_build_clamp(coord_bld, coord, min, max);
761 icoord = lp_build_ifloor(coord_bld, coord);
762 }
763 break;
764
765 case PIPE_TEX_WRAP_MIRROR_CLAMP:
766 coord = lp_build_abs(coord_bld, coord);
767 coord = lp_build_mul(coord_bld, coord, length_f);
768 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f_minus_one);
769 icoord = lp_build_ifloor(coord_bld, coord);
770 break;
771
772 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
773 {
774 LLVMValueRef min, max;
775 /* min = 1.0 / (2 * length) */
776 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
777 /* max = length - min */
778 max = lp_build_sub(coord_bld, length_f, min);
779
780 coord = lp_build_abs(coord_bld, coord);
781 coord = lp_build_mul(coord_bld, coord, length_f);
782 coord = lp_build_clamp(coord_bld, coord, min, max);
783 icoord = lp_build_ifloor(coord_bld, coord);
784 }
785 break;
786
787 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
788 {
789 LLVMValueRef min, max;
790 /* min = 1.0 / (2 * length) */
791 min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
792 min = lp_build_negate(coord_bld, min);
793 /* max = length - min */
794 max = lp_build_sub(coord_bld, length_f, min);
795
796 coord = lp_build_abs(coord_bld, coord);
797 coord = lp_build_mul(coord_bld, coord, length_f);
798 coord = lp_build_clamp(coord_bld, coord, min, max);
799 icoord = lp_build_ifloor(coord_bld, coord);
800 }
801 break;
802
803 default:
804 assert(0);
805 icoord = NULL;
806 }
807
808 return icoord;
809 }
810
811
812 /**
813 * Codegen equivalent for u_minify().
814 * Return max(1, base_size >> level);
815 */
816 static LLVMValueRef
817 lp_build_minify(struct lp_build_sample_context *bld,
818 LLVMValueRef base_size,
819 LLVMValueRef level)
820 {
821 LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify");
822 size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
823 return size;
824 }
825
826
827 /**
828 * Generate code to compute texture level of detail (lambda).
829 * \param s vector of texcoord s values
830 * \param t vector of texcoord t values
831 * \param r vector of texcoord r values
832 * \param width scalar int texture width
833 * \param height scalar int texture height
834 * \param depth scalar int texture depth
835 */
836 static LLVMValueRef
837 lp_build_lod_selector(struct lp_build_sample_context *bld,
838 LLVMValueRef s,
839 LLVMValueRef t,
840 LLVMValueRef r,
841 LLVMValueRef width,
842 LLVMValueRef height,
843 LLVMValueRef depth)
844
845 {
846 const int dims = texture_dims(bld->static_state->target);
847 struct lp_build_context *float_bld = &bld->float_bld;
848 LLVMValueRef lod_bias = LLVMConstReal(LLVMFloatType(), bld->static_state->lod_bias);
849 LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
850 LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(), bld->static_state->max_lod);
851
852 LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
853 LLVMValueRef index1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
854 LLVMValueRef index2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
855
856 LLVMValueRef s0, s1, s2;
857 LLVMValueRef t0, t1, t2;
858 LLVMValueRef r0, r1, r2;
859 LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
860 LLVMValueRef rho, lod;
861
862 /*
863 * dsdx = abs(s[1] - s[0]);
864 * dsdy = abs(s[2] - s[0]);
865 * dtdx = abs(t[1] - t[0]);
866 * dtdy = abs(t[2] - t[0]);
867 * drdx = abs(r[1] - r[0]);
868 * drdy = abs(r[2] - r[0]);
869 * XXX we're assuming a four-element quad in 2x2 layout here.
870 */
871 s0 = LLVMBuildExtractElement(bld->builder, s, index0, "s0");
872 s1 = LLVMBuildExtractElement(bld->builder, s, index1, "s1");
873 s2 = LLVMBuildExtractElement(bld->builder, s, index2, "s2");
874 dsdx = LLVMBuildSub(bld->builder, s1, s0, "");
875 dsdx = lp_build_abs(float_bld, dsdx);
876 dsdy = LLVMBuildSub(bld->builder, s2, s0, "");
877 dsdy = lp_build_abs(float_bld, dsdy);
878 if (dims > 1) {
879 t0 = LLVMBuildExtractElement(bld->builder, t, index0, "t0");
880 t1 = LLVMBuildExtractElement(bld->builder, t, index1, "t1");
881 t2 = LLVMBuildExtractElement(bld->builder, t, index2, "t2");
882 dtdx = LLVMBuildSub(bld->builder, t1, t0, "");
883 dtdx = lp_build_abs(float_bld, dtdx);
884 dtdy = LLVMBuildSub(bld->builder, t2, t0, "");
885 dtdy = lp_build_abs(float_bld, dtdy);
886 if (dims > 2) {
887 r0 = LLVMBuildExtractElement(bld->builder, r, index0, "r0");
888 r1 = LLVMBuildExtractElement(bld->builder, r, index1, "r1");
889 r2 = LLVMBuildExtractElement(bld->builder, r, index2, "r2");
890 drdx = LLVMBuildSub(bld->builder, r1, r0, "");
891 drdx = lp_build_abs(float_bld, drdx);
892 drdy = LLVMBuildSub(bld->builder, r2, r0, "");
893 drdy = lp_build_abs(float_bld, drdy);
894 }
895 }
896
897 /* Compute rho = max of all partial derivatives scaled by texture size.
898 * XXX this could be vectorized somewhat
899 */
900 rho = LLVMBuildMul(bld->builder,
901 lp_build_max(float_bld, dsdx, dsdy),
902 lp_build_int_to_float(float_bld, width), "");
903 if (dims > 1) {
904 LLVMValueRef max;
905 max = LLVMBuildMul(bld->builder,
906 lp_build_max(float_bld, dtdx, dtdy),
907 lp_build_int_to_float(float_bld, height), "");
908 rho = lp_build_max(float_bld, rho, max);
909 if (dims > 2) {
910 max = LLVMBuildMul(bld->builder,
911 lp_build_max(float_bld, drdx, drdy),
912 lp_build_int_to_float(float_bld, depth), "");
913 rho = lp_build_max(float_bld, rho, max);
914 }
915 }
916
917 /* compute lod = log2(rho) */
918 lod = lp_build_log2(float_bld, rho);
919
920 /* add lod bias */
921 lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "LOD bias");
922
923 /* clamp lod */
924 lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
925
926 return lod;
927 }
928
929
930 /**
931 * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
932 * mipmap level index.
933 * Note: this is all scalar code.
934 * \param lod scalar float texture level of detail
935 * \param level_out returns integer
936 */
937 static void
938 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
939 unsigned unit,
940 LLVMValueRef lod,
941 LLVMValueRef *level_out)
942 {
943 struct lp_build_context *float_bld = &bld->float_bld;
944 struct lp_build_context *int_bld = &bld->int_bld;
945 LLVMValueRef last_level, level;
946
947 LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
948
949 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
950 bld->builder, unit);
951
952 /* convert float lod to integer */
953 level = lp_build_iround(float_bld, lod);
954
955 /* clamp level to legal range of levels */
956 *level_out = lp_build_clamp(int_bld, level, zero, last_level);
957 }
958
959
960 /**
961 * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
962 * two (adjacent) mipmap level indexes. Later, we'll sample from those
963 * two mipmap levels and interpolate between them.
964 */
965 static void
966 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
967 unsigned unit,
968 LLVMValueRef lod,
969 LLVMValueRef *level0_out,
970 LLVMValueRef *level1_out,
971 LLVMValueRef *weight_out)
972 {
973 struct lp_build_context *float_bld = &bld->float_bld;
974 struct lp_build_context *int_bld = &bld->int_bld;
975 LLVMValueRef last_level, level;
976
977 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
978 bld->builder, unit);
979
980 /* convert float lod to integer */
981 level = lp_build_ifloor(float_bld, lod);
982
983 /* compute level 0 and clamp to legal range of levels */
984 *level0_out = lp_build_clamp(int_bld, level,
985 int_bld->zero,
986 last_level);
987 /* compute level 1 and clamp to legal range of levels */
988 *level1_out = lp_build_add(int_bld, *level0_out, int_bld->one);
989 *level1_out = lp_build_min(int_bld, *level1_out, int_bld->zero);
990
991 *weight_out = lp_build_fract(float_bld, lod);
992 }
993
994
995 /**
996 * Generate code to sample a mipmap level with nearest filtering.
997 */
998 static void
999 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
1000 LLVMValueRef width_vec,
1001 LLVMValueRef height_vec,
1002 LLVMValueRef depth_vec,
1003 LLVMValueRef row_stride_vec,
1004 LLVMValueRef img_stride_vec,
1005 LLVMValueRef data_ptr,
1006 LLVMValueRef s,
1007 LLVMValueRef t,
1008 LLVMValueRef r,
1009 LLVMValueRef colors_out[4])
1010 {
1011 const int dims = texture_dims(bld->static_state->target);
1012 LLVMValueRef x, y, z;
1013
1014 /*
1015 * Compute integer texcoords.
1016 */
1017 x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1018 bld->static_state->pot_width,
1019 bld->static_state->wrap_s);
1020 lp_build_name(x, "tex.x.wrapped");
1021
1022 if (dims >= 2) {
1023 y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1024 bld->static_state->pot_height,
1025 bld->static_state->wrap_t);
1026 lp_build_name(y, "tex.y.wrapped");
1027
1028 if (dims == 3) {
1029 z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1030 bld->static_state->pot_height,
1031 bld->static_state->wrap_r);
1032 lp_build_name(z, "tex.z.wrapped");
1033 }
1034 else {
1035 z = NULL;
1036 }
1037 }
1038 else {
1039 y = z = NULL;
1040 }
1041
1042 /*
1043 * Get texture colors.
1044 */
1045 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1046 x, y, z,
1047 row_stride_vec, img_stride_vec,
1048 data_ptr, colors_out);
1049 }
1050
1051
1052 /**
1053 * Generate code to sample a mipmap level with linear filtering.
1054 * 1D, 2D and 3D images are suppored.
1055 */
1056 static void
1057 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1058 LLVMValueRef width_vec,
1059 LLVMValueRef height_vec,
1060 LLVMValueRef depth_vec,
1061 LLVMValueRef row_stride_vec,
1062 LLVMValueRef img_stride_vec,
1063 LLVMValueRef data_ptr,
1064 LLVMValueRef s,
1065 LLVMValueRef t,
1066 LLVMValueRef r,
1067 LLVMValueRef colors_out[4])
1068 {
1069 const int dims = texture_dims(bld->static_state->target);
1070 LLVMValueRef x0, y0, z0, x1, y1, z1;
1071 LLVMValueRef s_fpart, t_fpart, r_fpart;
1072 LLVMValueRef neighbors[2][2][4];
1073 int chan;
1074
1075 /*
1076 * Compute integer texcoords.
1077 */
1078 lp_build_sample_wrap_linear(bld, s, width_vec,
1079 bld->static_state->pot_width,
1080 bld->static_state->wrap_s,
1081 &x0, &x1, &s_fpart);
1082 lp_build_name(x0, "tex.x0.wrapped");
1083 lp_build_name(x1, "tex.x1.wrapped");
1084
1085 if (dims >= 2) {
1086 lp_build_sample_wrap_linear(bld, t, height_vec,
1087 bld->static_state->pot_height,
1088 bld->static_state->wrap_t,
1089 &y0, &y1, &t_fpart);
1090 lp_build_name(y0, "tex.y0.wrapped");
1091 lp_build_name(y1, "tex.y1.wrapped");
1092
1093 if (dims == 3) {
1094 lp_build_sample_wrap_linear(bld, r, depth_vec,
1095 bld->static_state->pot_depth,
1096 bld->static_state->wrap_r,
1097 &z0, &z1, &r_fpart);
1098 lp_build_name(z0, "tex.z0.wrapped");
1099 lp_build_name(z1, "tex.z1.wrapped");
1100 }
1101 else {
1102 z0 = z1 = r_fpart = NULL;
1103 }
1104 }
1105 else {
1106 y0 = y1 = t_fpart = NULL;
1107 z0 = z1 = r_fpart = NULL;
1108 }
1109
1110 /*
1111 * Get texture colors.
1112 */
1113 /* get x0/x1 texels */
1114 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1115 x0, y0, z0,
1116 row_stride_vec, img_stride_vec,
1117 data_ptr, neighbors[0][0]);
1118 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1119 x1, y0, z0,
1120 row_stride_vec, img_stride_vec,
1121 data_ptr, neighbors[0][1]);
1122
1123 if (dims == 1) {
1124 /* Interpolate two samples from 1D image to produce one color */
1125 for (chan = 0; chan < 4; chan++) {
1126 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1127 neighbors[0][0][chan],
1128 neighbors[0][1][chan]);
1129 }
1130 }
1131 else {
1132 /* 2D/3D texture */
1133 LLVMValueRef colors0[4];
1134
1135 /* get x0/x1 texels at y1 */
1136 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1137 x0, y1, z0,
1138 row_stride_vec, img_stride_vec,
1139 data_ptr, neighbors[1][0]);
1140 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1141 x1, y1, z0,
1142 row_stride_vec, img_stride_vec,
1143 data_ptr, neighbors[1][1]);
1144
1145 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1146 for (chan = 0; chan < 4; chan++) {
1147 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1148 s_fpart, t_fpart,
1149 neighbors[0][0][chan],
1150 neighbors[0][1][chan],
1151 neighbors[1][0][chan],
1152 neighbors[1][1][chan]);
1153 }
1154
1155 if (dims == 3) {
1156 LLVMValueRef neighbors1[2][2][4];
1157 LLVMValueRef colors1[4];
1158
1159 /* get x0/x1/y0/y1 texels at z1 */
1160 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1161 x0, y0, z1,
1162 row_stride_vec, img_stride_vec,
1163 data_ptr, neighbors1[0][0]);
1164 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1165 x1, y0, z1,
1166 row_stride_vec, img_stride_vec,
1167 data_ptr, neighbors1[0][1]);
1168 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1169 x0, y1, z1,
1170 row_stride_vec, img_stride_vec,
1171 data_ptr, neighbors1[1][0]);
1172 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1173 x1, y1, z1,
1174 row_stride_vec, img_stride_vec,
1175 data_ptr, neighbors1[1][1]);
1176
1177 /* Bilinear interpolate the four samples from the second Z slice */
1178 for (chan = 0; chan < 4; chan++) {
1179 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1180 s_fpart, t_fpart,
1181 neighbors1[0][0][chan],
1182 neighbors1[0][1][chan],
1183 neighbors1[1][0][chan],
1184 neighbors1[1][1][chan]);
1185 }
1186
1187 /* Linearly interpolate the two samples from the two 3D slices */
1188 for (chan = 0; chan < 4; chan++) {
1189 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1190 r_fpart,
1191 colors0[chan], colors1[chan]);
1192 }
1193 }
1194 else {
1195 /* 2D tex */
1196 for (chan = 0; chan < 4; chan++) {
1197 colors_out[chan] = colors0[chan];
1198 }
1199 }
1200 }
1201 }
1202
1203
1204 /**
1205 * Generate code to do cube face selection and per-face texcoords.
1206 */
1207 static void
1208 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1209 LLVMValueRef s,
1210 LLVMValueRef t,
1211 LLVMValueRef r,
1212 LLVMValueRef *face,
1213 LLVMValueRef *face_s,
1214 LLVMValueRef *face_t)
1215 {
1216 #if 0
1217 struct lp_build_context *float_bld = &bld->float_bld;
1218 LLVMValueRef rx, ry, rz;
1219 LLVMValueRef arx, ary, arz;
1220 LLVMValueRef sc, tc, ma;
1221 LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
1222 LLVMValueRef arx_ge_ary, arx_ge_arz;
1223 LLVMValueRef ary_ge_arx, ary_ge_arz;
1224 LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
1225 LLVMValueRef rx_pos, ry_pos, rz_pos;
1226
1227 assert(bld->coord_bld.type.length == 4);
1228
1229 /*
1230 * Use the average of the four pixel's texcoords to choose the face.
1231 */
1232 rx = lp_build_mul(float_bld, c25,
1233 lp_build_sum_vector(&bld->coord_bld, s));
1234 ry = lp_build_mul(float_bld, c25,
1235 lp_build_sum_vector(&bld->coord_bld, t));
1236 rz = lp_build_mul(float_bld, c25,
1237 lp_build_sum_vector(&bld->coord_bld, r));
1238
1239 arx = lp_build_abs(float_bld, rx);
1240 ary = lp_build_abs(float_bld, ry);
1241 arz = lp_build_abs(float_bld, rz);
1242
1243 /*
1244 * Compare sign/magnitude of rx,ry,rz to determine face
1245 */
1246 arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
1247 arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
1248 ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
1249 ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
1250
1251 arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
1252 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1253
1254 rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
1255 ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
1256 rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
1257
1258 {
1259 struct lp_build_flow_context *flow_ctx;
1260 struct lp_build_if_state if_ctx, if2_ctx;
1261
1262 flow_ctx = lp_build_flow_create(bld->builder);
1263
1264 lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
1265 {
1266 #if 0
1267 lp_build_if(&if2_ctx, flow_ctx, bld->builder, rx_pos);
1268 {
1269 /* Positive X face */
1270 }
1271 lp_build_else(&if2_ctx);
1272 {
1273 /* Negative X face */
1274 }
1275 lp_build_endif(&if2_ctx);
1276 #endif
1277 }
1278 lp_build_else(&if_ctx);
1279 {
1280
1281
1282 }
1283 lp_build_endif(&if_ctx);
1284
1285 lp_build_flow_destroy(flow_ctx);
1286 }
1287 #endif
1288 }
1289
1290
1291
1292 /**
1293 * General texture sampling codegen.
1294 * This function handles texture sampling for all texture targets (1D,
1295 * 2D, 3D, cube) and all filtering modes.
1296 */
1297 static void
1298 lp_build_sample_general(struct lp_build_sample_context *bld,
1299 unsigned unit,
1300 LLVMValueRef s,
1301 LLVMValueRef t,
1302 LLVMValueRef r,
1303 LLVMValueRef width,
1304 LLVMValueRef height,
1305 LLVMValueRef depth,
1306 LLVMValueRef width_vec,
1307 LLVMValueRef height_vec,
1308 LLVMValueRef depth_vec,
1309 LLVMValueRef row_stride_array,
1310 LLVMValueRef img_stride_vec,
1311 LLVMValueRef data_array,
1312 LLVMValueRef *colors_out)
1313 {
1314 const unsigned mip_filter = bld->static_state->min_mip_filter;
1315 const unsigned min_filter = bld->static_state->min_img_filter;
1316 const unsigned mag_filter = bld->static_state->mag_img_filter;
1317 const int dims = texture_dims(bld->static_state->target);
1318 LLVMValueRef lod, lod_fpart;
1319 LLVMValueRef ilevel0, ilevel1, ilevel0_vec, ilevel1_vec;
1320 LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1321 LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1322 LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1323 LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1324 LLVMValueRef data_ptr0, data_ptr1;
1325 int chan;
1326
1327 /*
1328 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1329 mip_filter, min_filter, mag_filter);
1330 */
1331
1332 /*
1333 * Compute the level of detail (mipmap level index(es)).
1334 */
1335 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1336 /* always use mip level 0 */
1337 ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1338 }
1339 else {
1340 /* compute float LOD */
1341 lod = lp_build_lod_selector(bld, s, t, r, width, height, depth);
1342
1343 if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1344 lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1345 }
1346 else {
1347 assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1348 lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1349 &lod_fpart);
1350 lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1351 }
1352 }
1353
1354 /*
1355 * Convert scalar integer mipmap levels into vectors.
1356 */
1357 ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1358 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1359 ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1360
1361 /*
1362 * Compute width, height at mipmap level 'ilevel0'
1363 */
1364 width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1365 if (dims >= 2) {
1366 height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1367 row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1368 ilevel0);
1369 if (dims == 3) {
1370 depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1371 img_stride0_vec = lp_build_mul(&bld->int_coord_bld,
1372 row_stride0_vec, height0_vec);
1373 }
1374 }
1375 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1376 /* compute width, height, depth for second mipmap level at ilevel1 */
1377 width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1378 if (dims >= 2) {
1379 height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1380 row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1381 ilevel1);
1382 if (dims == 3) {
1383 depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1384 img_stride1_vec = lp_build_mul(&bld->int_coord_bld,
1385 row_stride1_vec, height1_vec);
1386 }
1387 }
1388 }
1389
1390 /*
1391 * Choose cube face, recompute texcoords.
1392 */
1393 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1394 LLVMValueRef face, face_s, face_t;
1395 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
1396 }
1397
1398 /*
1399 * Get pointer(s) to image data for mipmap level(s).
1400 */
1401 data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1402 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1403 data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1404 }
1405
1406 /*
1407 * Get/interpolate texture colors.
1408 */
1409 /* XXX temporarily force this path: */
1410 if (1 /*min_filter == mag_filter*/) {
1411 /* same filter for minification or magnification */
1412 LLVMValueRef colors0[4], colors1[4];
1413
1414 if (min_filter == PIPE_TEX_FILTER_NEAREST) {
1415 lp_build_sample_image_nearest(bld,
1416 width0_vec, height0_vec, depth0_vec,
1417 row_stride0_vec, img_stride0_vec,
1418 data_ptr0, s, t, r, colors0);
1419
1420 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1421 /* sample the second mipmap level, and interp */
1422 lp_build_sample_image_nearest(bld,
1423 width1_vec, height1_vec, depth1_vec,
1424 row_stride1_vec, img_stride1_vec,
1425 data_ptr1, s, t, r, colors1);
1426 }
1427 }
1428 else {
1429 assert(min_filter == PIPE_TEX_FILTER_LINEAR);
1430
1431 lp_build_sample_image_linear(bld,
1432 width0_vec, height0_vec, depth0_vec,
1433 row_stride0_vec, img_stride0_vec,
1434 data_ptr0, s, t, r, colors0);
1435
1436
1437 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1438 /* sample the second mipmap level, and interp */
1439 lp_build_sample_image_linear(bld,
1440 width1_vec, height1_vec, depth1_vec,
1441 row_stride1_vec, img_stride1_vec,
1442 data_ptr1, s, t, r, colors1);
1443 }
1444 }
1445
1446 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1447 /* interpolate samples from the two mipmap levels */
1448 for (chan = 0; chan < 4; chan++) {
1449 colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1450 colors0[chan], colors1[chan]);
1451 }
1452 }
1453 else {
1454 /* use first/only level's colors */
1455 for (chan = 0; chan < 4; chan++) {
1456 colors_out[chan] = colors0[chan];
1457 }
1458 }
1459 }
1460 else {
1461 /* emit conditional to choose min image filter or mag image filter
1462 * depending on the lod being >0 or <= 0, respectively.
1463 */
1464 abort();
1465 }
1466 }
1467
1468
1469
1470 static void
1471 lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
1472 struct lp_type dst_type,
1473 LLVMValueRef packed,
1474 LLVMValueRef *rgba)
1475 {
1476 LLVMValueRef mask = lp_build_int_const_scalar(dst_type, 0xff);
1477 unsigned chan;
1478
1479 /* Decode the input vector components */
1480 for (chan = 0; chan < 4; ++chan) {
1481 unsigned start = chan*8;
1482 unsigned stop = start + 8;
1483 LLVMValueRef input;
1484
1485 input = packed;
1486
1487 if(start)
1488 input = LLVMBuildLShr(builder, input, lp_build_int_const_scalar(dst_type, start), "");
1489
1490 if(stop < 32)
1491 input = LLVMBuildAnd(builder, input, mask, "");
1492
1493 input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
1494
1495 rgba[chan] = input;
1496 }
1497 }
1498
1499
1500 static void
1501 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1502 LLVMValueRef s,
1503 LLVMValueRef t,
1504 LLVMValueRef width,
1505 LLVMValueRef height,
1506 LLVMValueRef stride_array,
1507 LLVMValueRef data_array,
1508 LLVMValueRef *texel)
1509 {
1510 LLVMBuilderRef builder = bld->builder;
1511 struct lp_build_context i32, h16, u8n;
1512 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1513 LLVMValueRef i32_c8, i32_c128, i32_c255;
1514 LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1515 LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1516 LLVMValueRef x0, x1;
1517 LLVMValueRef y0, y1;
1518 LLVMValueRef neighbors[2][2];
1519 LLVMValueRef neighbors_lo[2][2];
1520 LLVMValueRef neighbors_hi[2][2];
1521 LLVMValueRef packed, packed_lo, packed_hi;
1522 LLVMValueRef unswizzled[4];
1523 LLVMValueRef stride;
1524
1525 lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1526 lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1527 lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1528
1529 i32_vec_type = lp_build_vec_type(i32.type);
1530 h16_vec_type = lp_build_vec_type(h16.type);
1531 u8n_vec_type = lp_build_vec_type(u8n.type);
1532
1533 if (bld->static_state->normalized_coords) {
1534 LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1535 LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1536 LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1537 s = lp_build_mul(&bld->coord_bld, s, fp_width);
1538 t = lp_build_mul(&bld->coord_bld, t, fp_height);
1539 }
1540
1541 /* scale coords by 256 (8 fractional bits) */
1542 s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1543 t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1544
1545 /* convert float to int */
1546 s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1547 t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1548
1549 /* subtract 0.5 (add -128) */
1550 i32_c128 = lp_build_int_const_scalar(i32.type, -128);
1551 s = LLVMBuildAdd(builder, s, i32_c128, "");
1552 t = LLVMBuildAdd(builder, t, i32_c128, "");
1553
1554 /* compute floor (shift right 8) */
1555 i32_c8 = lp_build_int_const_scalar(i32.type, 8);
1556 s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1557 t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1558
1559 /* compute fractional part (AND with 0xff) */
1560 i32_c255 = lp_build_int_const_scalar(i32.type, 255);
1561 s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1562 t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1563
1564 x0 = s_ipart;
1565 y0 = t_ipart;
1566
1567 x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
1568 y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
1569
1570 x0 = lp_build_sample_wrap_int(bld, x0, width, bld->static_state->pot_width,
1571 bld->static_state->wrap_s);
1572 y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
1573 bld->static_state->wrap_t);
1574
1575 x1 = lp_build_sample_wrap_int(bld, x1, width, bld->static_state->pot_width,
1576 bld->static_state->wrap_s);
1577 y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
1578 bld->static_state->wrap_t);
1579
1580 /*
1581 * Transform 4 x i32 in
1582 *
1583 * s_fpart = {s0, s1, s2, s3}
1584 *
1585 * into 8 x i16
1586 *
1587 * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1588 *
1589 * into two 8 x i16
1590 *
1591 * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1592 * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1593 *
1594 * and likewise for t_fpart. There is no risk of loosing precision here
1595 * since the fractional parts only use the lower 8bits.
1596 */
1597
1598 s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1599 t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1600
1601 {
1602 LLVMTypeRef elem_type = LLVMInt32Type();
1603 LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1604 LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1605 LLVMValueRef shuffle_lo;
1606 LLVMValueRef shuffle_hi;
1607 unsigned i, j;
1608
1609 for(j = 0; j < h16.type.length; j += 4) {
1610 unsigned subindex = util_cpu_caps.little_endian ? 0 : 1;
1611 LLVMValueRef index;
1612
1613 index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1614 for(i = 0; i < 4; ++i)
1615 shuffles_lo[j + i] = index;
1616
1617 index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1618 for(i = 0; i < 4; ++i)
1619 shuffles_hi[j + i] = index;
1620 }
1621
1622 shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1623 shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1624
1625 s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1626 t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1627 s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1628 t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1629 }
1630
1631 stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
1632
1633 /*
1634 * Fetch the pixels as 4 x 32bit (rgba order might differ):
1635 *
1636 * rgba0 rgba1 rgba2 rgba3
1637 *
1638 * bit cast them into 16 x u8
1639 *
1640 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1641 *
1642 * unpack them into two 8 x i16:
1643 *
1644 * r0 g0 b0 a0 r1 g1 b1 a1
1645 * r2 g2 b2 a2 r3 g3 b3 a3
1646 *
1647 * The higher 8 bits of the resulting elements will be zero.
1648 */
1649
1650 neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array);
1651 neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array);
1652 neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array);
1653 neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array);
1654
1655 neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, "");
1656 neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, "");
1657 neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, "");
1658 neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, "");
1659
1660 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]);
1661 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]);
1662 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]);
1663 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]);
1664
1665 /*
1666 * Linear interpolate with 8.8 fixed point.
1667 */
1668
1669 packed_lo = lp_build_lerp_2d(&h16,
1670 s_fpart_lo, t_fpart_lo,
1671 neighbors_lo[0][0],
1672 neighbors_lo[0][1],
1673 neighbors_lo[1][0],
1674 neighbors_lo[1][1]);
1675
1676 packed_hi = lp_build_lerp_2d(&h16,
1677 s_fpart_hi, t_fpart_hi,
1678 neighbors_hi[0][0],
1679 neighbors_hi[0][1],
1680 neighbors_hi[1][0],
1681 neighbors_hi[1][1]);
1682
1683 packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
1684
1685 /*
1686 * Convert to SoA and swizzle.
1687 */
1688
1689 packed = LLVMBuildBitCast(builder, packed, i32_vec_type, "");
1690
1691 lp_build_rgba8_to_f32_soa(bld->builder,
1692 bld->texel_type,
1693 packed, unswizzled);
1694
1695 lp_build_format_swizzle_soa(bld->format_desc,
1696 bld->texel_type, unswizzled,
1697 texel);
1698 }
1699
1700
1701 static void
1702 lp_build_sample_compare(struct lp_build_sample_context *bld,
1703 LLVMValueRef p,
1704 LLVMValueRef *texel)
1705 {
1706 struct lp_build_context *texel_bld = &bld->texel_bld;
1707 LLVMValueRef res;
1708 unsigned chan;
1709
1710 if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1711 return;
1712
1713 /* TODO: Compare before swizzling, to avoid redundant computations */
1714 res = NULL;
1715 for(chan = 0; chan < 4; ++chan) {
1716 LLVMValueRef cmp;
1717 cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
1718 cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
1719
1720 if(res)
1721 res = lp_build_add(texel_bld, res, cmp);
1722 else
1723 res = cmp;
1724 }
1725
1726 assert(res);
1727 res = lp_build_mul(texel_bld, res, lp_build_const_scalar(texel_bld->type, 0.25));
1728
1729 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1730 for(chan = 0; chan < 3; ++chan)
1731 texel[chan] = res;
1732 texel[3] = texel_bld->one;
1733 }
1734
1735
1736 /**
1737 * Build texture sampling code.
1738 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1739 * R, G, B, A.
1740 * \param type vector float type to use for coords, etc.
1741 */
1742 void
1743 lp_build_sample_soa(LLVMBuilderRef builder,
1744 const struct lp_sampler_static_state *static_state,
1745 struct lp_sampler_dynamic_state *dynamic_state,
1746 struct lp_type type,
1747 unsigned unit,
1748 unsigned num_coords,
1749 const LLVMValueRef *coords,
1750 LLVMValueRef lodbias,
1751 LLVMValueRef *texel)
1752 {
1753 struct lp_build_sample_context bld;
1754 LLVMValueRef width, width_vec;
1755 LLVMValueRef height, height_vec;
1756 LLVMValueRef depth, depth_vec;
1757 LLVMValueRef stride_array;
1758 LLVMValueRef data_array;
1759 LLVMValueRef s;
1760 LLVMValueRef t;
1761 LLVMValueRef r;
1762
1763 (void) lp_build_lod_selector; /* temporary to silence warning */
1764 (void) lp_build_nearest_mip_level;
1765 (void) lp_build_linear_mip_levels;
1766 (void) lp_build_minify;
1767
1768 /* Setup our build context */
1769 memset(&bld, 0, sizeof bld);
1770 bld.builder = builder;
1771 bld.static_state = static_state;
1772 bld.dynamic_state = dynamic_state;
1773 bld.format_desc = util_format_description(static_state->format);
1774
1775 bld.float_type = lp_type_float(32);
1776 bld.int_type = lp_type_int(32);
1777 bld.coord_type = type;
1778 bld.uint_coord_type = lp_uint_type(type);
1779 bld.int_coord_type = lp_int_type(type);
1780 bld.texel_type = type;
1781
1782 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
1783 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
1784 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
1785 lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
1786 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
1787 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
1788
1789 /* Get the dynamic state */
1790 width = dynamic_state->width(dynamic_state, builder, unit);
1791 height = dynamic_state->height(dynamic_state, builder, unit);
1792 depth = dynamic_state->depth(dynamic_state, builder, unit);
1793 stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
1794 data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
1795 /* Note that data_array is an array[level] of pointers to texture images */
1796
1797 s = coords[0];
1798 t = coords[1];
1799 r = coords[2];
1800
1801 width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
1802 height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
1803 depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
1804
1805 if (lp_format_is_rgba8(bld.format_desc) &&
1806 static_state->target == PIPE_TEXTURE_2D &&
1807 static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
1808 static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
1809 static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
1810 is_simple_wrap_mode(static_state->wrap_s) &&
1811 is_simple_wrap_mode(static_state->wrap_t)) {
1812 /* special case */
1813 lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
1814 stride_array, data_array, texel);
1815 }
1816 else {
1817 lp_build_sample_general(&bld, unit, s, t, r,
1818 width, height, depth,
1819 width_vec, height_vec, depth_vec,
1820 stride_array, NULL, data_array,
1821 texel);
1822 }
1823
1824 lp_build_sample_compare(&bld, r, texel);
1825 }