projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Merge remote branch 'origin/master' into nv50-compiler
[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 * @author Brian Paul <brianp@vmware.com>
34 */
35
36 #include "pipe/p_defines.h"
37 #include "pipe/p_state.h"
38 #include "util/u_debug.h"
39 #include "util/u_dump.h"
40 #include "util/u_memory.h"
41 #include "util/u_math.h"
42 #include "util/u_format.h"
43 #include "lp_bld_debug.h"
44 #include "lp_bld_type.h"
45 #include "lp_bld_const.h"
46 #include "lp_bld_conv.h"
47 #include "lp_bld_arit.h"
48 #include "lp_bld_logic.h"
49 #include "lp_bld_swizzle.h"
50 #include "lp_bld_pack.h"
51 #include "lp_bld_flow.h"
52 #include "lp_bld_gather.h"
53 #include "lp_bld_format.h"
54 #include "lp_bld_sample.h"
55 #include "lp_bld_quad.h"
56
57
58 /**
59 * Keep all information for sampling code generation in a single place.
60 */
61 struct lp_build_sample_context
62 {
63 LLVMBuilderRef builder;
64
65 const struct lp_sampler_static_state *static_state;
66
67 struct lp_sampler_dynamic_state *dynamic_state;
68
69 const struct util_format_description *format_desc;
70
71 /** regular scalar float type */
72 struct lp_type float_type;
73 struct lp_build_context float_bld;
74
75 /** regular scalar float type */
76 struct lp_type int_type;
77 struct lp_build_context int_bld;
78
79 /** Incoming coordinates type and build context */
80 struct lp_type coord_type;
81 struct lp_build_context coord_bld;
82
83 /** Unsigned integer coordinates */
84 struct lp_type uint_coord_type;
85 struct lp_build_context uint_coord_bld;
86
87 /** Signed integer coordinates */
88 struct lp_type int_coord_type;
89 struct lp_build_context int_coord_bld;
90
91 /** Output texels type and build context */
92 struct lp_type texel_type;
93 struct lp_build_context texel_bld;
94 };
95
96
97 /**
98 * Does the given texture wrap mode allow sampling the texture border color?
99 * XXX maybe move this into gallium util code.
100 */
101 static boolean
102 wrap_mode_uses_border_color(unsigned mode)
103 {
104 switch (mode) {
105 case PIPE_TEX_WRAP_REPEAT:
106 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
107 case PIPE_TEX_WRAP_MIRROR_REPEAT:
108 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
109 return FALSE;
110 case PIPE_TEX_WRAP_CLAMP:
111 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
112 case PIPE_TEX_WRAP_MIRROR_CLAMP:
113 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
114 return TRUE;
115 default:
116 assert(0 && "unexpected wrap mode");
117 return FALSE;
118 }
119 }
120
121
122 static LLVMValueRef
123 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
124 LLVMValueRef data_array, LLVMValueRef level)
125 {
126 LLVMValueRef indexes[2], data_ptr;
127 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
128 indexes[1] = level;
129 data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
130 data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
131 return data_ptr;
132 }
133
134
135 static LLVMValueRef
136 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
137 LLVMValueRef data_array, int level)
138 {
139 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
140 return lp_build_get_mipmap_level(bld, data_array, lvl);
141 }
142
143
144 /**
145 * Dereference stride_array[mipmap_level] array to get a stride.
146 * Return stride as a vector.
147 */
148 static LLVMValueRef
149 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
150 LLVMValueRef stride_array, LLVMValueRef level)
151 {
152 LLVMValueRef indexes[2], stride;
153 indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
154 indexes[1] = level;
155 stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
156 stride = LLVMBuildLoad(bld->builder, stride, "");
157 stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
158 return stride;
159 }
160
161
162 /** Dereference stride_array[0] array to get a stride (as vector). */
163 static LLVMValueRef
164 lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld,
165 LLVMValueRef stride_array, int level)
166 {
167 LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
168 return lp_build_get_level_stride_vec(bld, stride_array, lvl);
169 }
170
171
172 static int
173 texture_dims(enum pipe_texture_target tex)
174 {
175 switch (tex) {
176 case PIPE_TEXTURE_1D:
177 return 1;
178 case PIPE_TEXTURE_2D:
179 case PIPE_TEXTURE_RECT:
180 case PIPE_TEXTURE_CUBE:
181 return 2;
182 case PIPE_TEXTURE_3D:
183 return 3;
184 default:
185 assert(0 && "bad texture target in texture_dims()");
186 return 2;
187 }
188 }
189
190
191 static void
192 apply_sampler_swizzle(struct lp_build_sample_context *bld,
193 LLVMValueRef *texel)
194 {
195 unsigned char swizzles[4];
196
197 swizzles[0] = bld->static_state->swizzle_r;
198 swizzles[1] = bld->static_state->swizzle_g;
199 swizzles[2] = bld->static_state->swizzle_b;
200 swizzles[3] = bld->static_state->swizzle_a;
201
202 lp_build_swizzle_soa_inplace(&bld->texel_bld, texel, swizzles);
203 }
204
205
206
207 /**
208 * Generate code to fetch a texel from a texture at int coords (x, y, z).
209 * The computation depends on whether the texture is 1D, 2D or 3D.
210 * The result, texel, will be:
211 * texel[0] = red values
212 * texel[1] = green values
213 * texel[2] = blue values
214 * texel[3] = alpha values
215 */
216 static void
217 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
218 LLVMValueRef width,
219 LLVMValueRef height,
220 LLVMValueRef depth,
221 LLVMValueRef x,
222 LLVMValueRef y,
223 LLVMValueRef z,
224 LLVMValueRef y_stride,
225 LLVMValueRef z_stride,
226 LLVMValueRef data_ptr,
227 LLVMValueRef texel_out[4])
228 {
229 const int dims = texture_dims(bld->static_state->target);
230 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
231 LLVMValueRef offset;
232 LLVMValueRef i, j;
233 LLVMValueRef use_border = NULL;
234
235 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
236 if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
237 LLVMValueRef b1, b2;
238 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
239 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
240 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
241 }
242
243 if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
244 LLVMValueRef b1, b2;
245 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
246 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
247 if (use_border) {
248 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
249 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
250 }
251 else {
252 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
253 }
254 }
255
256 if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
257 LLVMValueRef b1, b2;
258 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
259 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
260 if (use_border) {
261 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
262 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
263 }
264 else {
265 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
266 }
267 }
268
269 /* convert x,y,z coords to linear offset from start of texture, in bytes */
270 lp_build_sample_offset(&bld->uint_coord_bld,
271 bld->format_desc,
272 x, y, z, y_stride, z_stride,
273 &offset, &i, &j);
274
275 if (use_border) {
276 /* If we can sample the border color, it means that texcoords may
277 * lie outside the bounds of the texture image. We need to do
278 * something to prevent reading out of bounds and causing a segfault.
279 *
280 * Simply AND the texture coords with !use_border. This will cause
281 * coords which are out of bounds to become zero. Zero's guaranteed
282 * to be inside the texture image.
283 */
284 offset = lp_build_andc(&bld->uint_coord_bld, offset, use_border);
285 }
286
287 lp_build_fetch_rgba_soa(bld->builder,
288 bld->format_desc,
289 bld->texel_type,
290 data_ptr, offset,
291 i, j,
292 texel_out);
293
294 apply_sampler_swizzle(bld, texel_out);
295
296 /*
297 * Note: if we find an app which frequently samples the texture border
298 * we might want to implement a true conditional here to avoid sampling
299 * the texture whenever possible (since that's quite a bit of code).
300 * Ex:
301 * if (use_border) {
302 * texel = border_color;
303 * }
304 * else {
305 * texel = sample_texture(coord);
306 * }
307 * As it is now, we always sample the texture, then selectively replace
308 * the texel color results with the border color.
309 */
310
311 if (use_border) {
312 /* select texel color or border color depending on use_border */
313 int chan;
314 for (chan = 0; chan < 4; chan++) {
315 LLVMValueRef border_chan =
316 lp_build_const_vec(bld->texel_type,
317 bld->static_state->border_color[chan]);
318 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
319 border_chan, texel_out[chan]);
320 }
321 }
322 }
323
324
325 /**
326 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
327 */
328 static LLVMValueRef
329 lp_build_coord_mirror(struct lp_build_sample_context *bld,
330 LLVMValueRef coord)
331 {
332 struct lp_build_context *coord_bld = &bld->coord_bld;
333 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
334 LLVMValueRef fract, flr, isOdd;
335
336 /* fract = coord - floor(coord) */
337 fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
338
339 /* flr = ifloor(coord); */
340 flr = lp_build_ifloor(coord_bld, coord);
341
342 /* isOdd = flr & 1 */
343 isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
344
345 /* make coord positive or negative depending on isOdd */
346 coord = lp_build_set_sign(coord_bld, fract, isOdd);
347
348 /* convert isOdd to float */
349 isOdd = lp_build_int_to_float(coord_bld, isOdd);
350
351 /* add isOdd to coord */
352 coord = lp_build_add(coord_bld, coord, isOdd);
353
354 return coord;
355 }
356
357
358 /**
359 * We only support a few wrap modes in lp_build_sample_wrap_linear_int() at this time.
360 * Return whether the given mode is supported by that function.
361 */
362 static boolean
363 is_simple_wrap_mode(unsigned mode)
364 {
365 switch (mode) {
366 case PIPE_TEX_WRAP_REPEAT:
367 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
368 return TRUE;
369 default:
370 return FALSE;
371 }
372 }
373
374
375 /**
376 * Build LLVM code for texture wrap mode, for scaled integer texcoords.
377 * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
378 * \param length the texture size along one dimension
379 * \param is_pot if TRUE, length is a power of two
380 * \param wrap_mode one of PIPE_TEX_WRAP_x
381 * \param i0 resulting sub-block pixel coordinate for coord0
382 */
383 static void
384 lp_build_sample_wrap_nearest_int(struct lp_build_sample_context *bld,
385 unsigned block_length,
386 LLVMValueRef coord,
387 LLVMValueRef length,
388 LLVMValueRef stride,
389 boolean is_pot,
390 unsigned wrap_mode,
391 LLVMValueRef *out_offset,
392 LLVMValueRef *out_i)
393 {
394 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
395 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
396 LLVMValueRef length_minus_one;
397
398 length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
399
400 switch(wrap_mode) {
401 case PIPE_TEX_WRAP_REPEAT:
402 if(is_pot)
403 coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
404 else
405 /* Signed remainder won't give the right results for negative
406 * dividends but unsigned remainder does.*/
407 coord = LLVMBuildURem(bld->builder, coord, length, "");
408 break;
409
410 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
411 coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
412 coord = lp_build_min(int_coord_bld, coord, length_minus_one);
413 break;
414
415 case PIPE_TEX_WRAP_CLAMP:
416 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
417 case PIPE_TEX_WRAP_MIRROR_REPEAT:
418 case PIPE_TEX_WRAP_MIRROR_CLAMP:
419 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
420 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
421 default:
422 assert(0);
423 }
424
425 lp_build_sample_partial_offset(uint_coord_bld, block_length, coord, stride,
426 out_offset, out_i);
427 }
428
429
430 /**
431 * Build LLVM code for texture wrap mode, for scaled integer texcoords.
432 * \param coord0 the incoming texcoord (s,t,r or q) scaled to the texture size
433 * \param length the texture size along one dimension
434 * \param stride pixel stride along the coordinate axis
435 * \param block_length is the length of the pixel block along the
436 * coordinate axis
437 * \param is_pot if TRUE, length is a power of two
438 * \param wrap_mode one of PIPE_TEX_WRAP_x
439 * \param offset0 resulting relative offset for coord0
440 * \param offset1 resulting relative offset for coord0 + 1
441 * \param i0 resulting sub-block pixel coordinate for coord0
442 * \param i1 resulting sub-block pixel coordinate for coord0 + 1
443 */
444 static void
445 lp_build_sample_wrap_linear_int(struct lp_build_sample_context *bld,
446 unsigned block_length,
447 LLVMValueRef coord0,
448 LLVMValueRef length,
449 LLVMValueRef stride,
450 boolean is_pot,
451 unsigned wrap_mode,
452 LLVMValueRef *offset0,
453 LLVMValueRef *offset1,
454 LLVMValueRef *i0,
455 LLVMValueRef *i1)
456 {
457 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
458 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
459 LLVMValueRef length_minus_one;
460 LLVMValueRef lmask, umask, mask;
461
462 if (block_length != 1) {
463 /*
464 * If the pixel block covers more than one pixel then there is no easy
465 * way to calculate offset1 relative to offset0. Instead, compute them
466 * independently.
467 */
468
469 LLVMValueRef coord1;
470
471 lp_build_sample_wrap_nearest_int(bld,
472 block_length,
473 coord0,
474 length,
475 stride,
476 is_pot,
477 wrap_mode,
478 offset0, i0);
479
480 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
481
482 lp_build_sample_wrap_nearest_int(bld,
483 block_length,
484 coord1,
485 length,
486 stride,
487 is_pot,
488 wrap_mode,
489 offset1, i1);
490
491 return;
492 }
493
494 /*
495 * Scalar pixels -- try to compute offset0 and offset1 with a single stride
496 * multiplication.
497 */
498
499 *i0 = uint_coord_bld->zero;
500 *i1 = uint_coord_bld->zero;
501
502 length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
503
504 switch(wrap_mode) {
505 case PIPE_TEX_WRAP_REPEAT:
506 if (is_pot) {
507 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
508 }
509 else {
510 /* Signed remainder won't give the right results for negative
511 * dividends but unsigned remainder does.*/
512 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
513 }
514
515 mask = lp_build_compare(bld->builder, int_coord_bld->type,
516 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
517
518 *offset0 = lp_build_mul(uint_coord_bld, coord0, stride);
519 *offset1 = LLVMBuildAnd(bld->builder,
520 lp_build_add(uint_coord_bld, *offset0, stride),
521 mask, "");
522 break;
523
524 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
525 lmask = lp_build_compare(int_coord_bld->builder, int_coord_bld->type,
526 PIPE_FUNC_GEQUAL, coord0, int_coord_bld->zero);
527 umask = lp_build_compare(int_coord_bld->builder, int_coord_bld->type,
528 PIPE_FUNC_LESS, coord0, length_minus_one);
529
530 coord0 = lp_build_select(int_coord_bld, lmask, coord0, int_coord_bld->zero);
531 coord0 = lp_build_select(int_coord_bld, umask, coord0, length_minus_one);
532
533 mask = LLVMBuildAnd(bld->builder, lmask, umask, "");
534
535 *offset0 = lp_build_mul(uint_coord_bld, coord0, stride);
536 *offset1 = lp_build_add(uint_coord_bld,
537 *offset0,
538 LLVMBuildAnd(bld->builder, stride, mask, ""));
539 break;
540
541 case PIPE_TEX_WRAP_CLAMP:
542 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
543 case PIPE_TEX_WRAP_MIRROR_REPEAT:
544 case PIPE_TEX_WRAP_MIRROR_CLAMP:
545 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
546 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
547 default:
548 assert(0);
549 *offset0 = uint_coord_bld->zero;
550 *offset1 = uint_coord_bld->zero;
551 break;
552 }
553 }
554
555
556 /**
557 * Build LLVM code for texture wrap mode for linear filtering.
558 * \param x0_out returns first integer texcoord
559 * \param x1_out returns second integer texcoord
560 * \param weight_out returns linear interpolation weight
561 */
562 static void
563 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
564 LLVMValueRef coord,
565 LLVMValueRef length,
566 boolean is_pot,
567 unsigned wrap_mode,
568 LLVMValueRef *x0_out,
569 LLVMValueRef *x1_out,
570 LLVMValueRef *weight_out)
571 {
572 struct lp_build_context *coord_bld = &bld->coord_bld;
573 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
574 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
575 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
576 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
577 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
578 LLVMValueRef coord0, coord1, weight;
579
580 switch(wrap_mode) {
581 case PIPE_TEX_WRAP_REPEAT:
582 /* mul by 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 /* convert to int */
586 coord0 = lp_build_ifloor(coord_bld, coord);
587 coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
588 /* compute lerp weight */
589 weight = lp_build_fract(coord_bld, coord);
590 /* repeat wrap */
591 if (is_pot) {
592 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
593 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
594 }
595 else {
596 /* Signed remainder won't give the right results for negative
597 * dividends but unsigned remainder does.*/
598 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
599 coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
600 }
601 break;
602
603 case PIPE_TEX_WRAP_CLAMP:
604 if (bld->static_state->normalized_coords) {
605 /* scale coord to length */
606 coord = lp_build_mul(coord_bld, coord, length_f);
607 }
608
609 /* clamp to [0, length] */
610 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
611
612 coord = lp_build_sub(coord_bld, coord, half);
613
614 weight = lp_build_fract(coord_bld, coord);
615 coord0 = lp_build_ifloor(coord_bld, coord);
616 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
617 break;
618
619 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
620 if (bld->static_state->normalized_coords) {
621 /* clamp to [0,1] */
622 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
623 /* mul by tex size and subtract 0.5 */
624 coord = lp_build_mul(coord_bld, coord, length_f);
625 coord = lp_build_sub(coord_bld, coord, half);
626 }
627 else {
628 LLVMValueRef min, max;
629 /* clamp to [0.5, length - 0.5] */
630 min = half;
631 max = lp_build_sub(coord_bld, length_f, min);
632 coord = lp_build_clamp(coord_bld, coord, min, max);
633 }
634 /* compute lerp weight */
635 weight = lp_build_fract(coord_bld, coord);
636 /* coord0 = floor(coord); */
637 coord0 = lp_build_ifloor(coord_bld, coord);
638 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
639 /* coord0 = max(coord0, 0) */
640 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
641 /* coord1 = min(coord1, length-1) */
642 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
643 break;
644
645 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
646 {
647 LLVMValueRef min, max;
648 if (bld->static_state->normalized_coords) {
649 /* scale coord to length */
650 coord = lp_build_mul(coord_bld, coord, length_f);
651 }
652 /* clamp to [-0.5, length + 0.5] */
653 min = lp_build_const_vec(coord_bld->type, -0.5F);
654 max = lp_build_sub(coord_bld, length_f, min);
655 coord = lp_build_clamp(coord_bld, coord, min, max);
656 coord = lp_build_sub(coord_bld, coord, half);
657 /* compute lerp weight */
658 weight = lp_build_fract(coord_bld, coord);
659 /* convert to int */
660 coord0 = lp_build_ifloor(coord_bld, coord);
661 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
662 }
663 break;
664
665 case PIPE_TEX_WRAP_MIRROR_REPEAT:
666 /* compute mirror function */
667 coord = lp_build_coord_mirror(bld, coord);
668
669 /* scale coord to length */
670 coord = lp_build_mul(coord_bld, coord, length_f);
671 coord = lp_build_sub(coord_bld, coord, half);
672
673 /* compute lerp weight */
674 weight = lp_build_fract(coord_bld, coord);
675
676 /* convert to int coords */
677 coord0 = lp_build_ifloor(coord_bld, coord);
678 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
679
680 /* coord0 = max(coord0, 0) */
681 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
682 /* coord1 = min(coord1, length-1) */
683 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
684 break;
685
686 case PIPE_TEX_WRAP_MIRROR_CLAMP:
687 coord = lp_build_abs(coord_bld, coord);
688
689 if (bld->static_state->normalized_coords) {
690 /* scale coord to length */
691 coord = lp_build_mul(coord_bld, coord, length_f);
692 }
693
694 /* clamp to [0, length] */
695 coord = lp_build_min(coord_bld, coord, length_f);
696
697 coord = lp_build_sub(coord_bld, coord, half);
698
699 weight = lp_build_fract(coord_bld, coord);
700 coord0 = lp_build_ifloor(coord_bld, coord);
701 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
702 break;
703
704 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
705 {
706 LLVMValueRef min, max;
707
708 coord = lp_build_abs(coord_bld, coord);
709
710 if (bld->static_state->normalized_coords) {
711 /* scale coord to length */
712 coord = lp_build_mul(coord_bld, coord, length_f);
713 }
714
715 /* clamp to [0.5, length - 0.5] */
716 min = half;
717 max = lp_build_sub(coord_bld, length_f, min);
718 coord = lp_build_clamp(coord_bld, coord, min, max);
719
720 coord = lp_build_sub(coord_bld, coord, half);
721
722 weight = lp_build_fract(coord_bld, coord);
723 coord0 = lp_build_ifloor(coord_bld, coord);
724 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
725 }
726 break;
727
728 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
729 {
730 LLVMValueRef min, max;
731
732 coord = lp_build_abs(coord_bld, coord);
733
734 if (bld->static_state->normalized_coords) {
735 /* scale coord to length */
736 coord = lp_build_mul(coord_bld, coord, length_f);
737 }
738
739 /* clamp to [-0.5, length + 0.5] */
740 min = lp_build_negate(coord_bld, half);
741 max = lp_build_sub(coord_bld, length_f, min);
742 coord = lp_build_clamp(coord_bld, coord, min, max);
743
744 coord = lp_build_sub(coord_bld, coord, half);
745
746 weight = lp_build_fract(coord_bld, coord);
747 coord0 = lp_build_ifloor(coord_bld, coord);
748 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
749 }
750 break;
751
752 default:
753 assert(0);
754 coord0 = NULL;
755 coord1 = NULL;
756 weight = NULL;
757 }
758
759 *x0_out = coord0;
760 *x1_out = coord1;
761 *weight_out = weight;
762 }
763
764
765 /**
766 * Build LLVM code for texture wrap mode for nearest filtering.
767 * \param coord the incoming texcoord (nominally in [0,1])
768 * \param length the texture size along one dimension, as int
769 * \param is_pot if TRUE, length is a power of two
770 * \param wrap_mode one of PIPE_TEX_WRAP_x
771 */
772 static LLVMValueRef
773 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
774 LLVMValueRef coord,
775 LLVMValueRef length,
776 boolean is_pot,
777 unsigned wrap_mode)
778 {
779 struct lp_build_context *coord_bld = &bld->coord_bld;
780 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
781 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
782 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
783 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
784 LLVMValueRef icoord;
785
786 switch(wrap_mode) {
787 case PIPE_TEX_WRAP_REPEAT:
788 coord = lp_build_mul(coord_bld, coord, length_f);
789 icoord = lp_build_ifloor(coord_bld, coord);
790 if (is_pot)
791 icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
792 else
793 /* Signed remainder won't give the right results for negative
794 * dividends but unsigned remainder does.*/
795 icoord = LLVMBuildURem(bld->builder, icoord, length, "");
796 break;
797
798 case PIPE_TEX_WRAP_CLAMP:
799 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
800 if (bld->static_state->normalized_coords) {
801 /* scale coord to length */
802 coord = lp_build_mul(coord_bld, coord, length_f);
803 }
804
805 /* floor */
806 icoord = lp_build_ifloor(coord_bld, coord);
807
808 /* clamp to [0, length - 1]. */
809 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
810 length_minus_one);
811 break;
812
813 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
814 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
815 {
816 LLVMValueRef min, max;
817
818 if (bld->static_state->normalized_coords) {
819 /* scale coord to length */
820 coord = lp_build_mul(coord_bld, coord, length_f);
821 }
822
823 icoord = lp_build_ifloor(coord_bld, coord);
824
825 /* clamp to [-1, length] */
826 min = lp_build_negate(int_coord_bld, int_coord_bld->one);
827 max = length;
828 icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
829 }
830 break;
831
832 case PIPE_TEX_WRAP_MIRROR_REPEAT:
833 /* compute mirror function */
834 coord = lp_build_coord_mirror(bld, coord);
835
836 /* scale coord to length */
837 assert(bld->static_state->normalized_coords);
838 coord = lp_build_mul(coord_bld, coord, length_f);
839
840 icoord = lp_build_ifloor(coord_bld, coord);
841
842 /* clamp to [0, length - 1] */
843 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
844 break;
845
846 case PIPE_TEX_WRAP_MIRROR_CLAMP:
847 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
848 coord = lp_build_abs(coord_bld, coord);
849
850 if (bld->static_state->normalized_coords) {
851 /* scale coord to length */
852 coord = lp_build_mul(coord_bld, coord, length_f);
853 }
854
855 icoord = lp_build_ifloor(coord_bld, coord);
856
857 /* clamp to [0, length - 1] */
858 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
859 break;
860
861 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
862 coord = lp_build_abs(coord_bld, coord);
863
864 if (bld->static_state->normalized_coords) {
865 /* scale coord to length */
866 coord = lp_build_mul(coord_bld, coord, length_f);
867 }
868
869 icoord = lp_build_ifloor(coord_bld, coord);
870
871 /* clamp to [0, length] */
872 icoord = lp_build_min(int_coord_bld, icoord, length);
873 break;
874
875 default:
876 assert(0);
877 icoord = NULL;
878 }
879
880 return icoord;
881 }
882
883
884 /**
885 * Codegen equivalent for u_minify().
886 * Return max(1, base_size >> level);
887 */
888 static LLVMValueRef
889 lp_build_minify(struct lp_build_sample_context *bld,
890 LLVMValueRef base_size,
891 LLVMValueRef level)
892 {
893 LLVMValueRef size = LLVMBuildLShr(bld->builder, base_size, level, "minify");
894 size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
895 return size;
896 }
897
898
899 /**
900 * Generate code to compute texture level of detail (lambda).
901 * \param ddx partial derivatives of (s, t, r, q) with respect to X
902 * \param ddy partial derivatives of (s, t, r, q) with respect to Y
903 * \param lod_bias optional float vector with the shader lod bias
904 * \param explicit_lod optional float vector with the explicit lod
905 * \param width scalar int texture width
906 * \param height scalar int texture height
907 * \param depth scalar int texture depth
908 *
909 * XXX: The resulting lod is scalar, so ignore all but the first element of
910 * derivatives, lod_bias, etc that are passed by the shader.
911 */
912 static LLVMValueRef
913 lp_build_lod_selector(struct lp_build_sample_context *bld,
914 const LLVMValueRef ddx[4],
915 const LLVMValueRef ddy[4],
916 LLVMValueRef lod_bias, /* optional */
917 LLVMValueRef explicit_lod, /* optional */
918 LLVMValueRef width,
919 LLVMValueRef height,
920 LLVMValueRef depth)
921
922 {
923 if (bld->static_state->min_lod == bld->static_state->max_lod) {
924 /* User is forcing sampling from a particular mipmap level.
925 * This is hit during mipmap generation.
926 */
927 return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
928 }
929 else {
930 struct lp_build_context *float_bld = &bld->float_bld;
931 LLVMValueRef sampler_lod_bias = LLVMConstReal(LLVMFloatType(),
932 bld->static_state->lod_bias);
933 LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(),
934 bld->static_state->min_lod);
935 LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(),
936 bld->static_state->max_lod);
937 LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
938 LLVMValueRef lod;
939
940 if (explicit_lod) {
941 lod = LLVMBuildExtractElement(bld->builder, explicit_lod,
942 index0, "");
943 }
944 else {
945 const int dims = texture_dims(bld->static_state->target);
946 LLVMValueRef dsdx, dsdy;
947 LLVMValueRef dtdx = NULL, dtdy = NULL, drdx = NULL, drdy = NULL;
948 LLVMValueRef rho;
949
950 dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
951 dsdx = lp_build_abs(float_bld, dsdx);
952 dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
953 dsdy = lp_build_abs(float_bld, dsdy);
954 if (dims > 1) {
955 dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
956 dtdx = lp_build_abs(float_bld, dtdx);
957 dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
958 dtdy = lp_build_abs(float_bld, dtdy);
959 if (dims > 2) {
960 drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
961 drdx = lp_build_abs(float_bld, drdx);
962 drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
963 drdy = lp_build_abs(float_bld, drdy);
964 }
965 }
966
967 /* Compute rho = max of all partial derivatives scaled by texture size.
968 * XXX this could be vectorized somewhat
969 */
970 rho = LLVMBuildFMul(bld->builder,
971 lp_build_max(float_bld, dsdx, dsdy),
972 lp_build_int_to_float(float_bld, width), "");
973 if (dims > 1) {
974 LLVMValueRef max;
975 max = LLVMBuildFMul(bld->builder,
976 lp_build_max(float_bld, dtdx, dtdy),
977 lp_build_int_to_float(float_bld, height), "");
978 rho = lp_build_max(float_bld, rho, max);
979 if (dims > 2) {
980 max = LLVMBuildFMul(bld->builder,
981 lp_build_max(float_bld, drdx, drdy),
982 lp_build_int_to_float(float_bld, depth), "");
983 rho = lp_build_max(float_bld, rho, max);
984 }
985 }
986
987 /* compute lod = log2(rho) */
988 lod = lp_build_log2(float_bld, rho);
989
990 /* add shader lod bias */
991 if (lod_bias) {
992 lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
993 index0, "");
994 lod = LLVMBuildFAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
995 }
996 }
997
998 /* add sampler lod bias */
999 lod = LLVMBuildFAdd(bld->builder, lod, sampler_lod_bias, "sampler_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 level = lp_build_add(int_bld, level, int_bld->one);
1068 *level1_out = lp_build_clamp(int_bld, level,
1069 int_bld->zero,
1070 last_level);
1071
1072 *weight_out = lp_build_fract(float_bld, lod);
1073 }
1074
1075
1076 /**
1077 * Generate code to sample a mipmap level with nearest filtering.
1078 * If sampling a cube texture, r = cube face in [0,5].
1079 */
1080 static void
1081 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
1082 LLVMValueRef width_vec,
1083 LLVMValueRef height_vec,
1084 LLVMValueRef depth_vec,
1085 LLVMValueRef row_stride_vec,
1086 LLVMValueRef img_stride_vec,
1087 LLVMValueRef data_ptr,
1088 LLVMValueRef s,
1089 LLVMValueRef t,
1090 LLVMValueRef r,
1091 LLVMValueRef colors_out[4])
1092 {
1093 const int dims = texture_dims(bld->static_state->target);
1094 LLVMValueRef x, y, z;
1095
1096 /*
1097 * Compute integer texcoords.
1098 */
1099 x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1100 bld->static_state->pot_width,
1101 bld->static_state->wrap_s);
1102 lp_build_name(x, "tex.x.wrapped");
1103
1104 if (dims >= 2) {
1105 y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1106 bld->static_state->pot_height,
1107 bld->static_state->wrap_t);
1108 lp_build_name(y, "tex.y.wrapped");
1109
1110 if (dims == 3) {
1111 z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1112 bld->static_state->pot_height,
1113 bld->static_state->wrap_r);
1114 lp_build_name(z, "tex.z.wrapped");
1115 }
1116 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1117 z = r;
1118 }
1119 else {
1120 z = NULL;
1121 }
1122 }
1123 else {
1124 y = z = NULL;
1125 }
1126
1127 /*
1128 * Get texture colors.
1129 */
1130 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1131 x, y, z,
1132 row_stride_vec, img_stride_vec,
1133 data_ptr, colors_out);
1134 }
1135
1136
1137 /**
1138 * Generate code to sample a mipmap level with linear filtering.
1139 * If sampling a cube texture, r = cube face in [0,5].
1140 */
1141 static void
1142 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1143 LLVMValueRef width_vec,
1144 LLVMValueRef height_vec,
1145 LLVMValueRef depth_vec,
1146 LLVMValueRef row_stride_vec,
1147 LLVMValueRef img_stride_vec,
1148 LLVMValueRef data_ptr,
1149 LLVMValueRef s,
1150 LLVMValueRef t,
1151 LLVMValueRef r,
1152 LLVMValueRef colors_out[4])
1153 {
1154 const int dims = texture_dims(bld->static_state->target);
1155 LLVMValueRef x0, y0, z0, x1, y1, z1;
1156 LLVMValueRef s_fpart, t_fpart, r_fpart;
1157 LLVMValueRef neighbors[2][2][4];
1158 int chan;
1159
1160 /*
1161 * Compute integer texcoords.
1162 */
1163 lp_build_sample_wrap_linear(bld, s, width_vec,
1164 bld->static_state->pot_width,
1165 bld->static_state->wrap_s,
1166 &x0, &x1, &s_fpart);
1167 lp_build_name(x0, "tex.x0.wrapped");
1168 lp_build_name(x1, "tex.x1.wrapped");
1169
1170 if (dims >= 2) {
1171 lp_build_sample_wrap_linear(bld, t, height_vec,
1172 bld->static_state->pot_height,
1173 bld->static_state->wrap_t,
1174 &y0, &y1, &t_fpart);
1175 lp_build_name(y0, "tex.y0.wrapped");
1176 lp_build_name(y1, "tex.y1.wrapped");
1177
1178 if (dims == 3) {
1179 lp_build_sample_wrap_linear(bld, r, depth_vec,
1180 bld->static_state->pot_depth,
1181 bld->static_state->wrap_r,
1182 &z0, &z1, &r_fpart);
1183 lp_build_name(z0, "tex.z0.wrapped");
1184 lp_build_name(z1, "tex.z1.wrapped");
1185 }
1186 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1187 z0 = z1 = r; /* cube face */
1188 r_fpart = NULL;
1189 }
1190 else {
1191 z0 = z1 = NULL;
1192 r_fpart = NULL;
1193 }
1194 }
1195 else {
1196 y0 = y1 = t_fpart = NULL;
1197 z0 = z1 = r_fpart = NULL;
1198 }
1199
1200 /*
1201 * Get texture colors.
1202 */
1203 /* get x0/x1 texels */
1204 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1205 x0, y0, z0,
1206 row_stride_vec, img_stride_vec,
1207 data_ptr, neighbors[0][0]);
1208 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1209 x1, y0, z0,
1210 row_stride_vec, img_stride_vec,
1211 data_ptr, neighbors[0][1]);
1212
1213 if (dims == 1) {
1214 /* Interpolate two samples from 1D image to produce one color */
1215 for (chan = 0; chan < 4; chan++) {
1216 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1217 neighbors[0][0][chan],
1218 neighbors[0][1][chan]);
1219 }
1220 }
1221 else {
1222 /* 2D/3D texture */
1223 LLVMValueRef colors0[4];
1224
1225 /* get x0/x1 texels at y1 */
1226 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1227 x0, y1, z0,
1228 row_stride_vec, img_stride_vec,
1229 data_ptr, neighbors[1][0]);
1230 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1231 x1, y1, z0,
1232 row_stride_vec, img_stride_vec,
1233 data_ptr, neighbors[1][1]);
1234
1235 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1236 for (chan = 0; chan < 4; chan++) {
1237 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1238 s_fpart, t_fpart,
1239 neighbors[0][0][chan],
1240 neighbors[0][1][chan],
1241 neighbors[1][0][chan],
1242 neighbors[1][1][chan]);
1243 }
1244
1245 if (dims == 3) {
1246 LLVMValueRef neighbors1[2][2][4];
1247 LLVMValueRef colors1[4];
1248
1249 /* get x0/x1/y0/y1 texels at z1 */
1250 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1251 x0, y0, z1,
1252 row_stride_vec, img_stride_vec,
1253 data_ptr, neighbors1[0][0]);
1254 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1255 x1, y0, z1,
1256 row_stride_vec, img_stride_vec,
1257 data_ptr, neighbors1[0][1]);
1258 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1259 x0, y1, z1,
1260 row_stride_vec, img_stride_vec,
1261 data_ptr, neighbors1[1][0]);
1262 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1263 x1, y1, z1,
1264 row_stride_vec, img_stride_vec,
1265 data_ptr, neighbors1[1][1]);
1266
1267 /* Bilinear interpolate the four samples from the second Z slice */
1268 for (chan = 0; chan < 4; chan++) {
1269 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1270 s_fpart, t_fpart,
1271 neighbors1[0][0][chan],
1272 neighbors1[0][1][chan],
1273 neighbors1[1][0][chan],
1274 neighbors1[1][1][chan]);
1275 }
1276
1277 /* Linearly interpolate the two samples from the two 3D slices */
1278 for (chan = 0; chan < 4; chan++) {
1279 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1280 r_fpart,
1281 colors0[chan], colors1[chan]);
1282 }
1283 }
1284 else {
1285 /* 2D tex */
1286 for (chan = 0; chan < 4; chan++) {
1287 colors_out[chan] = colors0[chan];
1288 }
1289 }
1290 }
1291 }
1292
1293
1294 /** Helper used by lp_build_cube_lookup() */
1295 static LLVMValueRef
1296 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
1297 {
1298 /* ima = -0.5 / abs(coord); */
1299 LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
1300 LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1301 LLVMValueRef ima = lp_build_div(coord_bld, negHalf, absCoord);
1302 return ima;
1303 }
1304
1305
1306 /**
1307 * Helper used by lp_build_cube_lookup()
1308 * \param sign scalar +1 or -1
1309 * \param coord float vector
1310 * \param ima float vector
1311 */
1312 static LLVMValueRef
1313 lp_build_cube_coord(struct lp_build_context *coord_bld,
1314 LLVMValueRef sign, int negate_coord,
1315 LLVMValueRef coord, LLVMValueRef ima)
1316 {
1317 /* return negate(coord) * ima * sign + 0.5; */
1318 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
1319 LLVMValueRef res;
1320
1321 assert(negate_coord == +1 || negate_coord == -1);
1322
1323 if (negate_coord == -1) {
1324 coord = lp_build_negate(coord_bld, coord);
1325 }
1326
1327 res = lp_build_mul(coord_bld, coord, ima);
1328 if (sign) {
1329 sign = lp_build_broadcast_scalar(coord_bld, sign);
1330 res = lp_build_mul(coord_bld, res, sign);
1331 }
1332 res = lp_build_add(coord_bld, res, half);
1333
1334 return res;
1335 }
1336
1337
1338 /** Helper used by lp_build_cube_lookup()
1339 * Return (major_coord >= 0) ? pos_face : neg_face;
1340 */
1341 static LLVMValueRef
1342 lp_build_cube_face(struct lp_build_sample_context *bld,
1343 LLVMValueRef major_coord,
1344 unsigned pos_face, unsigned neg_face)
1345 {
1346 LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1347 major_coord,
1348 bld->float_bld.zero, "");
1349 LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
1350 LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
1351 LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
1352 return res;
1353 }
1354
1355
1356
1357 /**
1358 * Generate code to do cube face selection and compute per-face texcoords.
1359 */
1360 static void
1361 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1362 LLVMValueRef s,
1363 LLVMValueRef t,
1364 LLVMValueRef r,
1365 LLVMValueRef *face,
1366 LLVMValueRef *face_s,
1367 LLVMValueRef *face_t)
1368 {
1369 struct lp_build_context *float_bld = &bld->float_bld;
1370 struct lp_build_context *coord_bld = &bld->coord_bld;
1371 LLVMValueRef rx, ry, rz;
1372 LLVMValueRef arx, ary, arz;
1373 LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
1374 LLVMValueRef arx_ge_ary, arx_ge_arz;
1375 LLVMValueRef ary_ge_arx, ary_ge_arz;
1376 LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
1377 LLVMValueRef rx_pos, ry_pos, rz_pos;
1378
1379 assert(bld->coord_bld.type.length == 4);
1380
1381 /*
1382 * Use the average of the four pixel's texcoords to choose the face.
1383 */
1384 rx = lp_build_mul(float_bld, c25,
1385 lp_build_sum_vector(&bld->coord_bld, s));
1386 ry = lp_build_mul(float_bld, c25,
1387 lp_build_sum_vector(&bld->coord_bld, t));
1388 rz = lp_build_mul(float_bld, c25,
1389 lp_build_sum_vector(&bld->coord_bld, r));
1390
1391 arx = lp_build_abs(float_bld, rx);
1392 ary = lp_build_abs(float_bld, ry);
1393 arz = lp_build_abs(float_bld, rz);
1394
1395 /*
1396 * Compare sign/magnitude of rx,ry,rz to determine face
1397 */
1398 arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
1399 arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
1400 ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
1401 ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
1402
1403 arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
1404 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1405
1406 rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
1407 ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
1408 rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
1409
1410 {
1411 struct lp_build_flow_context *flow_ctx;
1412 struct lp_build_if_state if_ctx;
1413
1414 flow_ctx = lp_build_flow_create(bld->builder);
1415 lp_build_flow_scope_begin(flow_ctx);
1416
1417 *face_s = bld->coord_bld.undef;
1418 *face_t = bld->coord_bld.undef;
1419 *face = bld->int_bld.undef;
1420
1421 lp_build_name(*face_s, "face_s");
1422 lp_build_name(*face_t, "face_t");
1423 lp_build_name(*face, "face");
1424
1425 lp_build_flow_scope_declare(flow_ctx, face_s);
1426 lp_build_flow_scope_declare(flow_ctx, face_t);
1427 lp_build_flow_scope_declare(flow_ctx, face);
1428
1429 lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
1430 {
1431 /* +/- X face */
1432 LLVMValueRef sign = lp_build_sgn(float_bld, rx);
1433 LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
1434 *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
1435 *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1436 *face = lp_build_cube_face(bld, rx,
1437 PIPE_TEX_FACE_POS_X,
1438 PIPE_TEX_FACE_NEG_X);
1439 }
1440 lp_build_else(&if_ctx);
1441 {
1442 struct lp_build_flow_context *flow_ctx2;
1443 struct lp_build_if_state if_ctx2;
1444
1445 LLVMValueRef face_s2 = bld->coord_bld.undef;
1446 LLVMValueRef face_t2 = bld->coord_bld.undef;
1447 LLVMValueRef face2 = bld->int_bld.undef;
1448
1449 flow_ctx2 = lp_build_flow_create(bld->builder);
1450 lp_build_flow_scope_begin(flow_ctx2);
1451 lp_build_flow_scope_declare(flow_ctx2, &face_s2);
1452 lp_build_flow_scope_declare(flow_ctx2, &face_t2);
1453 lp_build_flow_scope_declare(flow_ctx2, &face2);
1454
1455 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1456
1457 lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
1458 {
1459 /* +/- Y face */
1460 LLVMValueRef sign = lp_build_sgn(float_bld, ry);
1461 LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
1462 face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
1463 face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
1464 face2 = lp_build_cube_face(bld, ry,
1465 PIPE_TEX_FACE_POS_Y,
1466 PIPE_TEX_FACE_NEG_Y);
1467 }
1468 lp_build_else(&if_ctx2);
1469 {
1470 /* +/- Z face */
1471 LLVMValueRef sign = lp_build_sgn(float_bld, rz);
1472 LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
1473 face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
1474 face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1475 face2 = lp_build_cube_face(bld, rz,
1476 PIPE_TEX_FACE_POS_Z,
1477 PIPE_TEX_FACE_NEG_Z);
1478 }
1479 lp_build_endif(&if_ctx2);
1480 lp_build_flow_scope_end(flow_ctx2);
1481 lp_build_flow_destroy(flow_ctx2);
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 /* sample the first mipmap level */
1528 lp_build_sample_image_nearest(bld,
1529 width0_vec, height0_vec, depth0_vec,
1530 row_stride0_vec, img_stride0_vec,
1531 data_ptr0, s, t, r, colors0);
1532
1533 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1534 /* sample the second mipmap level */
1535 lp_build_sample_image_nearest(bld,
1536 width1_vec, height1_vec, depth1_vec,
1537 row_stride1_vec, img_stride1_vec,
1538 data_ptr1, s, t, r, colors1);
1539 }
1540 }
1541 else {
1542 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1543
1544 /* sample the first mipmap level */
1545 lp_build_sample_image_linear(bld,
1546 width0_vec, height0_vec, depth0_vec,
1547 row_stride0_vec, img_stride0_vec,
1548 data_ptr0, s, t, r, colors0);
1549
1550 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1551 /* sample the second mipmap level */
1552 lp_build_sample_image_linear(bld,
1553 width1_vec, height1_vec, depth1_vec,
1554 row_stride1_vec, img_stride1_vec,
1555 data_ptr1, s, t, r, colors1);
1556 }
1557 }
1558
1559 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1560 /* interpolate samples from the two mipmap levels */
1561 for (chan = 0; chan < 4; chan++) {
1562 colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1563 colors0[chan], colors1[chan]);
1564 }
1565 }
1566 else {
1567 /* use first/only level's colors */
1568 for (chan = 0; chan < 4; chan++) {
1569 colors_out[chan] = colors0[chan];
1570 }
1571 }
1572 }
1573
1574
1575
1576 /**
1577 * General texture sampling codegen.
1578 * This function handles texture sampling for all texture targets (1D,
1579 * 2D, 3D, cube) and all filtering modes.
1580 */
1581 static void
1582 lp_build_sample_general(struct lp_build_sample_context *bld,
1583 unsigned unit,
1584 LLVMValueRef s,
1585 LLVMValueRef t,
1586 LLVMValueRef r,
1587 const LLVMValueRef *ddx,
1588 const LLVMValueRef *ddy,
1589 LLVMValueRef lod_bias, /* optional */
1590 LLVMValueRef explicit_lod, /* optional */
1591 LLVMValueRef width,
1592 LLVMValueRef height,
1593 LLVMValueRef depth,
1594 LLVMValueRef width_vec,
1595 LLVMValueRef height_vec,
1596 LLVMValueRef depth_vec,
1597 LLVMValueRef row_stride_array,
1598 LLVMValueRef img_stride_array,
1599 LLVMValueRef data_array,
1600 LLVMValueRef *colors_out)
1601 {
1602 struct lp_build_context *float_bld = &bld->float_bld;
1603 const unsigned mip_filter = bld->static_state->min_mip_filter;
1604 const unsigned min_filter = bld->static_state->min_img_filter;
1605 const unsigned mag_filter = bld->static_state->mag_img_filter;
1606 const int dims = texture_dims(bld->static_state->target);
1607 LLVMValueRef lod = NULL, lod_fpart = NULL;
1608 LLVMValueRef ilevel0, ilevel1 = NULL, ilevel0_vec, ilevel1_vec = NULL;
1609 LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1610 LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1611 LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1612 LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1613 LLVMValueRef data_ptr0, data_ptr1 = NULL;
1614 LLVMValueRef face_ddx[4], face_ddy[4];
1615
1616 /*
1617 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1618 mip_filter, min_filter, mag_filter);
1619 */
1620
1621 /*
1622 * Choose cube face, recompute texcoords and derivatives for the chosen face.
1623 */
1624 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1625 LLVMValueRef face, face_s, face_t;
1626 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
1627 s = face_s; /* vec */
1628 t = face_t; /* vec */
1629 /* use 'r' to indicate cube face */
1630 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
1631
1632 /* recompute ddx, ddy using the new (s,t) face texcoords */
1633 face_ddx[0] = lp_build_ddx(&bld->coord_bld, s);
1634 face_ddx[1] = lp_build_ddx(&bld->coord_bld, t);
1635 face_ddx[2] = NULL;
1636 face_ddx[3] = NULL;
1637 face_ddy[0] = lp_build_ddy(&bld->coord_bld, s);
1638 face_ddy[1] = lp_build_ddy(&bld->coord_bld, t);
1639 face_ddy[2] = NULL;
1640 face_ddy[3] = NULL;
1641 ddx = face_ddx;
1642 ddy = face_ddy;
1643 }
1644
1645 /*
1646 * Compute the level of detail (float).
1647 */
1648 if (min_filter != mag_filter ||
1649 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1650 /* Need to compute lod either to choose mipmap levels or to
1651 * distinguish between minification/magnification with one mipmap level.
1652 */
1653 lod = lp_build_lod_selector(bld, ddx, ddy,
1654 lod_bias, explicit_lod,
1655 width, height, depth);
1656 }
1657
1658 /*
1659 * Compute integer mipmap level(s) to fetch texels from.
1660 */
1661 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1662 /* always use mip level 0 */
1663 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1664 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1665 * We should be able to set ilevel0 = const(0) but that causes
1666 * bad x86 code to be emitted.
1667 */
1668 lod = lp_build_const_elem(bld->coord_bld.type, 0.0);
1669 lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1670 }
1671 else {
1672 ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1673 }
1674 }
1675 else {
1676 assert(lod);
1677 if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1678 lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1679 }
1680 else {
1681 assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1682 lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1683 &lod_fpart);
1684 lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1685 }
1686 }
1687
1688 /*
1689 * Convert scalar integer mipmap levels into vectors.
1690 */
1691 ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1692 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1693 ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1694
1695 /*
1696 * Compute width, height at mipmap level 'ilevel0'
1697 */
1698 width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1699 if (dims >= 2) {
1700 height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1701 row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1702 ilevel0);
1703 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1704 img_stride0_vec = lp_build_get_level_stride_vec(bld,
1705 img_stride_array,
1706 ilevel0);
1707 if (dims == 3) {
1708 depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1709 }
1710 }
1711 }
1712 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1713 /* compute width, height, depth for second mipmap level at 'ilevel1' */
1714 width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1715 if (dims >= 2) {
1716 height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1717 row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1718 ilevel1);
1719 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1720 img_stride1_vec = lp_build_get_level_stride_vec(bld,
1721 img_stride_array,
1722 ilevel1);
1723 if (dims ==3) {
1724 depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1725 }
1726 }
1727 }
1728 }
1729
1730 /*
1731 * Get pointer(s) to image data for mipmap level(s).
1732 */
1733 data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1734 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1735 data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1736 }
1737
1738 /*
1739 * Get/interpolate texture colors.
1740 */
1741 if (min_filter == mag_filter) {
1742 /* no need to distinquish between minification and magnification */
1743 lp_build_sample_mipmap(bld, min_filter, mip_filter, s, t, r, lod_fpart,
1744 width0_vec, width1_vec,
1745 height0_vec, height1_vec,
1746 depth0_vec, depth1_vec,
1747 row_stride0_vec, row_stride1_vec,
1748 img_stride0_vec, img_stride1_vec,
1749 data_ptr0, data_ptr1,
1750 colors_out);
1751 }
1752 else {
1753 /* Emit conditional to choose min image filter or mag image filter
1754 * depending on the lod being >0 or <= 0, respectively.
1755 */
1756 struct lp_build_flow_context *flow_ctx;
1757 struct lp_build_if_state if_ctx;
1758 LLVMValueRef minify;
1759
1760 flow_ctx = lp_build_flow_create(bld->builder);
1761 lp_build_flow_scope_begin(flow_ctx);
1762
1763 lp_build_flow_scope_declare(flow_ctx, &colors_out[0]);
1764 lp_build_flow_scope_declare(flow_ctx, &colors_out[1]);
1765 lp_build_flow_scope_declare(flow_ctx, &colors_out[2]);
1766 lp_build_flow_scope_declare(flow_ctx, &colors_out[3]);
1767
1768 /* minify = lod > 0.0 */
1769 minify = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1770 lod, float_bld->zero, "");
1771
1772 lp_build_if(&if_ctx, flow_ctx, bld->builder, minify);
1773 {
1774 /* Use the minification filter */
1775 lp_build_sample_mipmap(bld, min_filter, mip_filter,
1776 s, t, r, lod_fpart,
1777 width0_vec, width1_vec,
1778 height0_vec, height1_vec,
1779 depth0_vec, depth1_vec,
1780 row_stride0_vec, row_stride1_vec,
1781 img_stride0_vec, img_stride1_vec,
1782 data_ptr0, data_ptr1,
1783 colors_out);
1784 }
1785 lp_build_else(&if_ctx);
1786 {
1787 /* Use the magnification filter */
1788 lp_build_sample_mipmap(bld, mag_filter, mip_filter,
1789 s, t, r, lod_fpart,
1790 width0_vec, width1_vec,
1791 height0_vec, height1_vec,
1792 depth0_vec, depth1_vec,
1793 row_stride0_vec, row_stride1_vec,
1794 img_stride0_vec, img_stride1_vec,
1795 data_ptr0, data_ptr1,
1796 colors_out);
1797 }
1798 lp_build_endif(&if_ctx);
1799
1800 lp_build_flow_scope_end(flow_ctx);
1801 lp_build_flow_destroy(flow_ctx);
1802 }
1803 }
1804
1805
1806
1807 static void
1808 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1809 LLVMValueRef s,
1810 LLVMValueRef t,
1811 LLVMValueRef width,
1812 LLVMValueRef height,
1813 LLVMValueRef stride_array,
1814 LLVMValueRef data_array,
1815 LLVMValueRef texel_out[4])
1816 {
1817 LLVMBuilderRef builder = bld->builder;
1818 struct lp_build_context i32, h16, u8n;
1819 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1820 LLVMValueRef i32_c8, i32_c128, i32_c255;
1821 LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1822 LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1823 LLVMValueRef data_ptr;
1824 LLVMValueRef x_stride, y_stride;
1825 LLVMValueRef x_offset0, x_offset1;
1826 LLVMValueRef y_offset0, y_offset1;
1827 LLVMValueRef offset[2][2];
1828 LLVMValueRef x_subcoord[2], y_subcoord[2];
1829 LLVMValueRef neighbors_lo[2][2];
1830 LLVMValueRef neighbors_hi[2][2];
1831 LLVMValueRef packed, packed_lo, packed_hi;
1832 LLVMValueRef unswizzled[4];
1833 const unsigned level = 0;
1834 unsigned i, j;
1835
1836 assert(bld->static_state->target == PIPE_TEXTURE_2D
1837 || bld->static_state->target == PIPE_TEXTURE_RECT);
1838 assert(bld->static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR);
1839 assert(bld->static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR);
1840 assert(bld->static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE);
1841
1842 lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1843 lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1844 lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1845
1846 i32_vec_type = lp_build_vec_type(i32.type);
1847 h16_vec_type = lp_build_vec_type(h16.type);
1848 u8n_vec_type = lp_build_vec_type(u8n.type);
1849
1850 if (bld->static_state->normalized_coords) {
1851 LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1852 LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1853 LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1854 s = lp_build_mul(&bld->coord_bld, s, fp_width);
1855 t = lp_build_mul(&bld->coord_bld, t, fp_height);
1856 }
1857
1858 /* scale coords by 256 (8 fractional bits) */
1859 s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1860 t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1861
1862 /* convert float to int */
1863 s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1864 t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1865
1866 /* subtract 0.5 (add -128) */
1867 i32_c128 = lp_build_const_int_vec(i32.type, -128);
1868 s = LLVMBuildAdd(builder, s, i32_c128, "");
1869 t = LLVMBuildAdd(builder, t, i32_c128, "");
1870
1871 /* compute floor (shift right 8) */
1872 i32_c8 = lp_build_const_int_vec(i32.type, 8);
1873 s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1874 t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1875
1876 /* compute fractional part (AND with 0xff) */
1877 i32_c255 = lp_build_const_int_vec(i32.type, 255);
1878 s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1879 t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1880
1881 x_stride = lp_build_const_vec(bld->uint_coord_bld.type,
1882 bld->format_desc->block.bits/8);
1883
1884 y_stride = lp_build_get_const_level_stride_vec(bld, stride_array, level);
1885
1886 lp_build_sample_wrap_linear_int(bld,
1887 bld->format_desc->block.width,
1888 s_ipart, width, x_stride,
1889 bld->static_state->pot_width,
1890 bld->static_state->wrap_s,
1891 &x_offset0, &x_offset1,
1892 &x_subcoord[0], &x_subcoord[1]);
1893 lp_build_sample_wrap_linear_int(bld,
1894 bld->format_desc->block.height,
1895 t_ipart, height, y_stride,
1896 bld->static_state->pot_height,
1897 bld->static_state->wrap_t,
1898 &y_offset0, &y_offset1,
1899 &y_subcoord[0], &y_subcoord[1]);
1900
1901 offset[0][0] = lp_build_add(&bld->uint_coord_bld, x_offset0, y_offset0);
1902 offset[0][1] = lp_build_add(&bld->uint_coord_bld, x_offset1, y_offset0);
1903 offset[1][0] = lp_build_add(&bld->uint_coord_bld, x_offset0, y_offset1);
1904 offset[1][1] = lp_build_add(&bld->uint_coord_bld, x_offset1, y_offset1);
1905
1906 /*
1907 * Transform 4 x i32 in
1908 *
1909 * s_fpart = {s0, s1, s2, s3}
1910 *
1911 * into 8 x i16
1912 *
1913 * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1914 *
1915 * into two 8 x i16
1916 *
1917 * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1918 * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1919 *
1920 * and likewise for t_fpart. There is no risk of loosing precision here
1921 * since the fractional parts only use the lower 8bits.
1922 */
1923
1924 s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1925 t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1926
1927 {
1928 LLVMTypeRef elem_type = LLVMInt32Type();
1929 LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1930 LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1931 LLVMValueRef shuffle_lo;
1932 LLVMValueRef shuffle_hi;
1933
1934 for(j = 0; j < h16.type.length; j += 4) {
1935 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1936 unsigned subindex = 0;
1937 #else
1938 unsigned subindex = 1;
1939 #endif
1940 LLVMValueRef index;
1941
1942 index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1943 for(i = 0; i < 4; ++i)
1944 shuffles_lo[j + i] = index;
1945
1946 index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1947 for(i = 0; i < 4; ++i)
1948 shuffles_hi[j + i] = index;
1949 }
1950
1951 shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1952 shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1953
1954 s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1955 t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1956 s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1957 t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1958 }
1959
1960 /*
1961 * get pointer to mipmap level 0 data
1962 */
1963 data_ptr = lp_build_get_const_mipmap_level(bld, data_array, level);
1964
1965 /*
1966 * Fetch the pixels as 4 x 32bit (rgba order might differ):
1967 *
1968 * rgba0 rgba1 rgba2 rgba3
1969 *
1970 * bit cast them into 16 x u8
1971 *
1972 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1973 *
1974 * unpack them into two 8 x i16:
1975 *
1976 * r0 g0 b0 a0 r1 g1 b1 a1
1977 * r2 g2 b2 a2 r3 g3 b3 a3
1978 *
1979 * The higher 8 bits of the resulting elements will be zero.
1980 */
1981
1982 for (j = 0; j < 2; ++j) {
1983 for (i = 0; i < 2; ++i) {
1984 LLVMValueRef rgba8;
1985
1986 if (util_format_is_rgba8_variant(bld->format_desc)) {
1987 /*
1988 * Given the format is a rgba8, just read the pixels as is,
1989 * without any swizzling. Swizzling will be done later.
1990 */
1991 rgba8 = lp_build_gather(bld->builder,
1992 bld->texel_type.length,
1993 bld->format_desc->block.bits,
1994 bld->texel_type.width,
1995 data_ptr, offset[j][i]);
1996
1997 rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
1998
1999 }
2000 else {
2001 rgba8 = lp_build_fetch_rgba_aos(bld->builder,
2002 bld->format_desc,
2003 u8n.type,
2004 data_ptr, offset[j][i],
2005 x_subcoord[i],
2006 y_subcoord[j]);
2007 }
2008
2009 lp_build_unpack2(builder, u8n.type, h16.type,
2010 rgba8,
2011 &neighbors_lo[j][i], &neighbors_hi[j][i]);
2012 }
2013 }
2014
2015 /*
2016 * Linear interpolate with 8.8 fixed point.
2017 */
2018
2019 packed_lo = lp_build_lerp_2d(&h16,
2020 s_fpart_lo, t_fpart_lo,
2021 neighbors_lo[0][0],
2022 neighbors_lo[0][1],
2023 neighbors_lo[1][0],
2024 neighbors_lo[1][1]);
2025
2026 packed_hi = lp_build_lerp_2d(&h16,
2027 s_fpart_hi, t_fpart_hi,
2028 neighbors_hi[0][0],
2029 neighbors_hi[0][1],
2030 neighbors_hi[1][0],
2031 neighbors_hi[1][1]);
2032
2033 packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
2034
2035 /*
2036 * Convert to SoA and swizzle.
2037 */
2038
2039 lp_build_rgba8_to_f32_soa(bld->builder,
2040 bld->texel_type,
2041 packed, unswizzled);
2042
2043 if (util_format_is_rgba8_variant(bld->format_desc)) {
2044 lp_build_format_swizzle_soa(bld->format_desc,
2045 &bld->texel_bld,
2046 unswizzled, texel_out);
2047 } else {
2048 texel_out[0] = unswizzled[0];
2049 texel_out[1] = unswizzled[1];
2050 texel_out[2] = unswizzled[2];
2051 texel_out[3] = unswizzled[3];
2052 }
2053
2054 apply_sampler_swizzle(bld, texel_out);
2055 }
2056
2057
2058 static void
2059 lp_build_sample_compare(struct lp_build_sample_context *bld,
2060 LLVMValueRef p,
2061 LLVMValueRef texel[4])
2062 {
2063 struct lp_build_context *texel_bld = &bld->texel_bld;
2064 LLVMValueRef res;
2065 unsigned chan;
2066
2067 if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
2068 return;
2069
2070 /* TODO: Compare before swizzling, to avoid redundant computations */
2071 res = NULL;
2072 for(chan = 0; chan < 4; ++chan) {
2073 LLVMValueRef cmp;
2074 cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
2075 cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
2076
2077 if(res)
2078 res = lp_build_add(texel_bld, res, cmp);
2079 else
2080 res = cmp;
2081 }
2082
2083 assert(res);
2084 res = lp_build_mul(texel_bld, res, lp_build_const_vec(texel_bld->type, 0.25));
2085
2086 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
2087 for(chan = 0; chan < 3; ++chan)
2088 texel[chan] = res;
2089 texel[3] = texel_bld->one;
2090 }
2091
2092
2093 /**
2094 * Just set texels to white instead of actually sampling the texture.
2095 * For debugging.
2096 */
2097 static void
2098 lp_build_sample_nop(struct lp_build_sample_context *bld,
2099 LLVMValueRef texel_out[4])
2100 {
2101 struct lp_build_context *texel_bld = &bld->texel_bld;
2102 unsigned chan;
2103
2104 for (chan = 0; chan < 4; chan++) {
2105 /*lp_bld_mov(texel_bld, texel, texel_bld->one);*/
2106 texel_out[chan] = texel_bld->one;
2107 }
2108 }
2109
2110
2111 /**
2112 * Build texture sampling code.
2113 * 'texel' will return a vector of four LLVMValueRefs corresponding to
2114 * R, G, B, A.
2115 * \param type vector float type to use for coords, etc.
2116 * \param ddx partial derivatives of (s,t,r,q) with respect to x
2117 * \param ddy partial derivatives of (s,t,r,q) with respect to y
2118 */
2119 void
2120 lp_build_sample_soa(LLVMBuilderRef builder,
2121 const struct lp_sampler_static_state *static_state,
2122 struct lp_sampler_dynamic_state *dynamic_state,
2123 struct lp_type type,
2124 unsigned unit,
2125 unsigned num_coords,
2126 const LLVMValueRef *coords,
2127 const LLVMValueRef ddx[4],
2128 const LLVMValueRef ddy[4],
2129 LLVMValueRef lod_bias, /* optional */
2130 LLVMValueRef explicit_lod, /* optional */
2131 LLVMValueRef texel_out[4])
2132 {
2133 struct lp_build_sample_context bld;
2134 LLVMValueRef width, width_vec;
2135 LLVMValueRef height, height_vec;
2136 LLVMValueRef depth, depth_vec;
2137 LLVMValueRef row_stride_array, img_stride_array;
2138 LLVMValueRef data_array;
2139 LLVMValueRef s;
2140 LLVMValueRef t;
2141 LLVMValueRef r;
2142
2143 if (0) {
2144 enum pipe_format fmt = static_state->format;
2145 debug_printf("Sample from %s\n", util_format_name(fmt));
2146 }
2147
2148 assert(type.floating);
2149
2150 /* Setup our build context */
2151 memset(&bld, 0, sizeof bld);
2152 bld.builder = builder;
2153 bld.static_state = static_state;
2154 bld.dynamic_state = dynamic_state;
2155 bld.format_desc = util_format_description(static_state->format);
2156
2157 bld.float_type = lp_type_float(32);
2158 bld.int_type = lp_type_int(32);
2159 bld.coord_type = type;
2160 bld.uint_coord_type = lp_uint_type(type);
2161 bld.int_coord_type = lp_int_type(type);
2162 bld.texel_type = type;
2163
2164 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
2165 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
2166 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
2167 lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
2168 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
2169 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
2170
2171 /* Get the dynamic state */
2172 width = dynamic_state->width(dynamic_state, builder, unit);
2173 height = dynamic_state->height(dynamic_state, builder, unit);
2174 depth = dynamic_state->depth(dynamic_state, builder, unit);
2175 row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
2176 img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit);
2177 data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
2178 /* Note that data_array is an array[level] of pointers to texture images */
2179
2180 s = coords[0];
2181 t = coords[1];
2182 r = coords[2];
2183
2184 width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
2185 height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
2186 depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
2187
2188 if (0) {
2189 /* For debug: no-op texture sampling */
2190 lp_build_sample_nop(&bld, texel_out);
2191 }
2192 else if (util_format_fits_8unorm(bld.format_desc) &&
2193 bld.format_desc->nr_channels > 1 &&
2194 (static_state->target == PIPE_TEXTURE_2D ||
2195 static_state->target == PIPE_TEXTURE_RECT) &&
2196 static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
2197 static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
2198 static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
2199 is_simple_wrap_mode(static_state->wrap_s) &&
2200 is_simple_wrap_mode(static_state->wrap_t)) {
2201 /* special case */
2202 lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
2203 row_stride_array, data_array, texel_out);
2204 }
2205 else {
2206 lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
2207 lod_bias, explicit_lod,
2208 width, height, depth,
2209 width_vec, height_vec, depth_vec,
2210 row_stride_array, img_stride_array,
2211 data_array,
2212 texel_out);
2213 }
2214
2215 lp_build_sample_compare(&bld, r, texel_out);
2216 }