projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm: fix incorrect floor(), itrunc()
[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 * If sampling a cube texture, r = cube face in [0,5].
998 */
999 static void
1000 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
1001 LLVMValueRef width_vec,
1002 LLVMValueRef height_vec,
1003 LLVMValueRef depth_vec,
1004 LLVMValueRef row_stride_vec,
1005 LLVMValueRef img_stride_vec,
1006 LLVMValueRef data_ptr,
1007 LLVMValueRef s,
1008 LLVMValueRef t,
1009 LLVMValueRef r,
1010 LLVMValueRef colors_out[4])
1011 {
1012 const int dims = texture_dims(bld->static_state->target);
1013 LLVMValueRef x, y, z;
1014
1015 /*
1016 * Compute integer texcoords.
1017 */
1018 x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1019 bld->static_state->pot_width,
1020 bld->static_state->wrap_s);
1021 lp_build_name(x, "tex.x.wrapped");
1022
1023 if (dims >= 2) {
1024 y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1025 bld->static_state->pot_height,
1026 bld->static_state->wrap_t);
1027 lp_build_name(y, "tex.y.wrapped");
1028
1029 if (dims == 3) {
1030 z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1031 bld->static_state->pot_height,
1032 bld->static_state->wrap_r);
1033 lp_build_name(z, "tex.z.wrapped");
1034 }
1035 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1036 z = r;
1037 }
1038 else {
1039 z = NULL;
1040 }
1041 }
1042 else {
1043 y = z = NULL;
1044 }
1045
1046 /*
1047 * Get texture colors.
1048 */
1049 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1050 x, y, z,
1051 row_stride_vec, img_stride_vec,
1052 data_ptr, colors_out);
1053 }
1054
1055
1056 /**
1057 * Generate code to sample a mipmap level with linear filtering.
1058 * If sampling a cube texture, r = cube face in [0,5].
1059 */
1060 static void
1061 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1062 LLVMValueRef width_vec,
1063 LLVMValueRef height_vec,
1064 LLVMValueRef depth_vec,
1065 LLVMValueRef row_stride_vec,
1066 LLVMValueRef img_stride_vec,
1067 LLVMValueRef data_ptr,
1068 LLVMValueRef s,
1069 LLVMValueRef t,
1070 LLVMValueRef r,
1071 LLVMValueRef colors_out[4])
1072 {
1073 const int dims = texture_dims(bld->static_state->target);
1074 LLVMValueRef x0, y0, z0, x1, y1, z1;
1075 LLVMValueRef s_fpart, t_fpart, r_fpart;
1076 LLVMValueRef neighbors[2][2][4];
1077 int chan;
1078
1079 /*
1080 * Compute integer texcoords.
1081 */
1082 lp_build_sample_wrap_linear(bld, s, width_vec,
1083 bld->static_state->pot_width,
1084 bld->static_state->wrap_s,
1085 &x0, &x1, &s_fpart);
1086 lp_build_name(x0, "tex.x0.wrapped");
1087 lp_build_name(x1, "tex.x1.wrapped");
1088
1089 if (dims >= 2) {
1090 lp_build_sample_wrap_linear(bld, t, height_vec,
1091 bld->static_state->pot_height,
1092 bld->static_state->wrap_t,
1093 &y0, &y1, &t_fpart);
1094 lp_build_name(y0, "tex.y0.wrapped");
1095 lp_build_name(y1, "tex.y1.wrapped");
1096
1097 if (dims == 3) {
1098 lp_build_sample_wrap_linear(bld, r, depth_vec,
1099 bld->static_state->pot_depth,
1100 bld->static_state->wrap_r,
1101 &z0, &z1, &r_fpart);
1102 lp_build_name(z0, "tex.z0.wrapped");
1103 lp_build_name(z1, "tex.z1.wrapped");
1104 }
1105 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1106 z0 = z1 = r; /* cube face */
1107 r_fpart = NULL;
1108 }
1109 else {
1110 z0 = z1 = NULL;
1111 r_fpart = NULL;
1112 }
1113 }
1114 else {
1115 y0 = y1 = t_fpart = NULL;
1116 z0 = z1 = r_fpart = NULL;
1117 }
1118
1119 /*
1120 * Get texture colors.
1121 */
1122 /* get x0/x1 texels */
1123 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1124 x0, y0, z0,
1125 row_stride_vec, img_stride_vec,
1126 data_ptr, neighbors[0][0]);
1127 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1128 x1, y0, z0,
1129 row_stride_vec, img_stride_vec,
1130 data_ptr, neighbors[0][1]);
1131
1132 if (dims == 1) {
1133 /* Interpolate two samples from 1D image to produce one color */
1134 for (chan = 0; chan < 4; chan++) {
1135 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1136 neighbors[0][0][chan],
1137 neighbors[0][1][chan]);
1138 }
1139 }
1140 else {
1141 /* 2D/3D texture */
1142 LLVMValueRef colors0[4];
1143
1144 /* get x0/x1 texels at y1 */
1145 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1146 x0, y1, z0,
1147 row_stride_vec, img_stride_vec,
1148 data_ptr, neighbors[1][0]);
1149 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1150 x1, y1, z0,
1151 row_stride_vec, img_stride_vec,
1152 data_ptr, neighbors[1][1]);
1153
1154 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1155 for (chan = 0; chan < 4; chan++) {
1156 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1157 s_fpart, t_fpart,
1158 neighbors[0][0][chan],
1159 neighbors[0][1][chan],
1160 neighbors[1][0][chan],
1161 neighbors[1][1][chan]);
1162 }
1163
1164 if (dims == 3) {
1165 LLVMValueRef neighbors1[2][2][4];
1166 LLVMValueRef colors1[4];
1167
1168 /* get x0/x1/y0/y1 texels at z1 */
1169 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1170 x0, y0, z1,
1171 row_stride_vec, img_stride_vec,
1172 data_ptr, neighbors1[0][0]);
1173 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1174 x1, y0, z1,
1175 row_stride_vec, img_stride_vec,
1176 data_ptr, neighbors1[0][1]);
1177 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1178 x0, y1, z1,
1179 row_stride_vec, img_stride_vec,
1180 data_ptr, neighbors1[1][0]);
1181 lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1182 x1, y1, z1,
1183 row_stride_vec, img_stride_vec,
1184 data_ptr, neighbors1[1][1]);
1185
1186 /* Bilinear interpolate the four samples from the second Z slice */
1187 for (chan = 0; chan < 4; chan++) {
1188 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1189 s_fpart, t_fpart,
1190 neighbors1[0][0][chan],
1191 neighbors1[0][1][chan],
1192 neighbors1[1][0][chan],
1193 neighbors1[1][1][chan]);
1194 }
1195
1196 /* Linearly interpolate the two samples from the two 3D slices */
1197 for (chan = 0; chan < 4; chan++) {
1198 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1199 r_fpart,
1200 colors0[chan], colors1[chan]);
1201 }
1202 }
1203 else {
1204 /* 2D tex */
1205 for (chan = 0; chan < 4; chan++) {
1206 colors_out[chan] = colors0[chan];
1207 }
1208 }
1209 }
1210 }
1211
1212
1213 /** Helper used by lp_build_cube_lookup() */
1214 static LLVMValueRef
1215 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
1216 {
1217 /* ima = -0.5 / abs(coord); */
1218 LLVMValueRef negHalf = lp_build_const_scalar(coord_bld->type, -0.5);
1219 LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1220 LLVMValueRef ima = lp_build_mul(coord_bld, negHalf,
1221 lp_build_rcp(coord_bld, absCoord));
1222 return ima;
1223 }
1224
1225
1226 /**
1227 * Helper used by lp_build_cube_lookup()
1228 * \param sign scalar +1 or -1
1229 * \param coord float vector
1230 * \param ima float vector
1231 */
1232 static LLVMValueRef
1233 lp_build_cube_coord(struct lp_build_context *coord_bld,
1234 LLVMValueRef sign, int negate_coord,
1235 LLVMValueRef coord, LLVMValueRef ima)
1236 {
1237 /* return negate(coord) * ima * sign + 0.5; */
1238 LLVMValueRef half = lp_build_const_scalar(coord_bld->type, 0.5);
1239 LLVMValueRef res;
1240
1241 assert(negate_coord == +1 || negate_coord == -1);
1242
1243 if (negate_coord == -1) {
1244 coord = lp_build_negate(coord_bld, coord);
1245 }
1246
1247 res = lp_build_mul(coord_bld, coord, ima);
1248 if (sign) {
1249 sign = lp_build_broadcast_scalar(coord_bld, sign);
1250 res = lp_build_mul(coord_bld, res, sign);
1251 }
1252 res = lp_build_add(coord_bld, res, half);
1253
1254 return res;
1255 }
1256
1257
1258 /** Helper used by lp_build_cube_lookup()
1259 * Return (major_coord >= 0) ? pos_face : neg_face;
1260 */
1261 static LLVMValueRef
1262 lp_build_cube_face(struct lp_build_sample_context *bld,
1263 LLVMValueRef major_coord,
1264 unsigned pos_face, unsigned neg_face)
1265 {
1266 LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1267 major_coord,
1268 bld->float_bld.zero, "");
1269 LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
1270 LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
1271 LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
1272 return res;
1273 }
1274
1275
1276
1277 /**
1278 * Generate code to do cube face selection and per-face texcoords.
1279 */
1280 static void
1281 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1282 LLVMValueRef s,
1283 LLVMValueRef t,
1284 LLVMValueRef r,
1285 LLVMValueRef *face,
1286 LLVMValueRef *face_s,
1287 LLVMValueRef *face_t)
1288 {
1289 struct lp_build_context *float_bld = &bld->float_bld;
1290 struct lp_build_context *coord_bld = &bld->coord_bld;
1291 LLVMValueRef rx, ry, rz;
1292 LLVMValueRef arx, ary, arz;
1293 LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
1294 LLVMValueRef arx_ge_ary, arx_ge_arz;
1295 LLVMValueRef ary_ge_arx, ary_ge_arz;
1296 LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
1297 LLVMValueRef rx_pos, ry_pos, rz_pos;
1298
1299 assert(bld->coord_bld.type.length == 4);
1300
1301 /*
1302 * Use the average of the four pixel's texcoords to choose the face.
1303 */
1304 rx = lp_build_mul(float_bld, c25,
1305 lp_build_sum_vector(&bld->coord_bld, s));
1306 ry = lp_build_mul(float_bld, c25,
1307 lp_build_sum_vector(&bld->coord_bld, t));
1308 rz = lp_build_mul(float_bld, c25,
1309 lp_build_sum_vector(&bld->coord_bld, r));
1310
1311 arx = lp_build_abs(float_bld, rx);
1312 ary = lp_build_abs(float_bld, ry);
1313 arz = lp_build_abs(float_bld, rz);
1314
1315 /*
1316 * Compare sign/magnitude of rx,ry,rz to determine face
1317 */
1318 arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
1319 arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
1320 ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
1321 ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
1322
1323 arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
1324 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1325
1326 rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
1327 ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
1328 rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
1329
1330 {
1331 struct lp_build_flow_context *flow_ctx;
1332 struct lp_build_if_state if_ctx;
1333
1334 flow_ctx = lp_build_flow_create(bld->builder);
1335 lp_build_flow_scope_begin(flow_ctx);
1336
1337 *face_s = bld->coord_bld.undef;
1338 *face_t = bld->coord_bld.undef;
1339 *face = bld->int_bld.undef;
1340
1341 lp_build_name(*face_s, "face_s");
1342 lp_build_name(*face_t, "face_t");
1343 lp_build_name(*face, "face");
1344
1345 lp_build_flow_scope_declare(flow_ctx, face_s);
1346 lp_build_flow_scope_declare(flow_ctx, face_t);
1347 lp_build_flow_scope_declare(flow_ctx, face);
1348
1349 lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
1350 {
1351 /* +/- X face */
1352 LLVMValueRef sign = lp_build_sgn(float_bld, rx);
1353 LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
1354 *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
1355 *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1356 *face = lp_build_cube_face(bld, rx,
1357 PIPE_TEX_FACE_POS_X,
1358 PIPE_TEX_FACE_NEG_X);
1359 }
1360 lp_build_else(&if_ctx);
1361 {
1362 struct lp_build_flow_context *flow_ctx2;
1363 struct lp_build_if_state if_ctx2;
1364
1365 LLVMValueRef face_s2 = bld->coord_bld.undef;
1366 LLVMValueRef face_t2 = bld->coord_bld.undef;
1367 LLVMValueRef face2 = bld->int_bld.undef;
1368
1369 flow_ctx2 = lp_build_flow_create(bld->builder);
1370 lp_build_flow_scope_begin(flow_ctx2);
1371 lp_build_flow_scope_declare(flow_ctx2, &face_s2);
1372 lp_build_flow_scope_declare(flow_ctx2, &face_t2);
1373 lp_build_flow_scope_declare(flow_ctx2, &face2);
1374
1375 ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1376
1377 lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
1378 {
1379 /* +/- Y face */
1380 LLVMValueRef sign = lp_build_sgn(float_bld, ry);
1381 LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
1382 face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
1383 face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
1384 face2 = lp_build_cube_face(bld, ry,
1385 PIPE_TEX_FACE_POS_Y,
1386 PIPE_TEX_FACE_NEG_Y);
1387 }
1388 lp_build_else(&if_ctx2);
1389 {
1390 /* +/- Z face */
1391 LLVMValueRef sign = lp_build_sgn(float_bld, rz);
1392 LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
1393 face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
1394 face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1395 face2 = lp_build_cube_face(bld, rz,
1396 PIPE_TEX_FACE_POS_Z,
1397 PIPE_TEX_FACE_NEG_Z);
1398 }
1399 lp_build_endif(&if_ctx2);
1400 lp_build_flow_scope_end(flow_ctx2);
1401 lp_build_flow_destroy(flow_ctx2);
1402
1403 *face_s = face_s2;
1404 *face_t = face_t2;
1405 *face = face2;
1406 }
1407
1408 lp_build_endif(&if_ctx);
1409 lp_build_flow_scope_end(flow_ctx);
1410 lp_build_flow_destroy(flow_ctx);
1411 }
1412 }
1413
1414
1415
1416 /**
1417 * General texture sampling codegen.
1418 * This function handles texture sampling for all texture targets (1D,
1419 * 2D, 3D, cube) and all filtering modes.
1420 */
1421 static void
1422 lp_build_sample_general(struct lp_build_sample_context *bld,
1423 unsigned unit,
1424 LLVMValueRef s,
1425 LLVMValueRef t,
1426 LLVMValueRef r,
1427 LLVMValueRef width,
1428 LLVMValueRef height,
1429 LLVMValueRef depth,
1430 LLVMValueRef width_vec,
1431 LLVMValueRef height_vec,
1432 LLVMValueRef depth_vec,
1433 LLVMValueRef row_stride_array,
1434 LLVMValueRef img_stride_vec,
1435 LLVMValueRef data_array,
1436 LLVMValueRef *colors_out)
1437 {
1438 const unsigned mip_filter = bld->static_state->min_mip_filter;
1439 const unsigned min_filter = bld->static_state->min_img_filter;
1440 const unsigned mag_filter = bld->static_state->mag_img_filter;
1441 const int dims = texture_dims(bld->static_state->target);
1442 LLVMValueRef lod, lod_fpart;
1443 LLVMValueRef ilevel0, ilevel1, ilevel0_vec, ilevel1_vec;
1444 LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1445 LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1446 LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1447 LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1448 LLVMValueRef data_ptr0, data_ptr1;
1449 int chan;
1450
1451 /*
1452 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
1453 mip_filter, min_filter, mag_filter);
1454 */
1455
1456 /*
1457 * Compute the level of detail (mipmap level index(es)).
1458 */
1459 if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1460 /* always use mip level 0 */
1461 ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1462 }
1463 else {
1464 /* compute float LOD */
1465 lod = lp_build_lod_selector(bld, s, t, r, width, height, depth);
1466
1467 if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1468 lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1469 }
1470 else {
1471 assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1472 lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1473 &lod_fpart);
1474 lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1475 }
1476 }
1477
1478 /*
1479 * Convert scalar integer mipmap levels into vectors.
1480 */
1481 ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1482 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1483 ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1484
1485 /*
1486 * Compute width, height at mipmap level 'ilevel0'
1487 */
1488 width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1489 if (dims >= 2) {
1490 height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1491 row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1492 ilevel0);
1493 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1494 img_stride0_vec = lp_build_mul(&bld->int_coord_bld,
1495 row_stride0_vec, height0_vec);
1496 if (dims == 3) {
1497 depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1498 }
1499 }
1500 }
1501 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1502 /* compute width, height, depth for second mipmap level at ilevel1 */
1503 width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1504 if (dims >= 2) {
1505 height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1506 row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1507 ilevel1);
1508 if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1509 img_stride1_vec = lp_build_mul(&bld->int_coord_bld,
1510 row_stride1_vec, height1_vec);
1511 if (dims ==3) {
1512 depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1513 }
1514 }
1515 }
1516 }
1517
1518 /*
1519 * Choose cube face, recompute texcoords.
1520 */
1521 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1522 LLVMValueRef face, face_s, face_t;
1523 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
1524 s = face_s; /* vec */
1525 t = face_t; /* vec */
1526 /* use 'r' to indicate cube face */
1527 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
1528 }
1529
1530 /*
1531 * Get pointer(s) to image data for mipmap level(s).
1532 */
1533 data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1534 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1535 data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1536 }
1537
1538 /*
1539 * Get/interpolate texture colors.
1540 */
1541 /* XXX temporarily force this path: */
1542 if (1 /*min_filter == mag_filter*/) {
1543 /* same filter for minification or magnification */
1544 LLVMValueRef colors0[4], colors1[4];
1545
1546 if (min_filter == PIPE_TEX_FILTER_NEAREST) {
1547 lp_build_sample_image_nearest(bld,
1548 width0_vec, height0_vec, depth0_vec,
1549 row_stride0_vec, img_stride0_vec,
1550 data_ptr0, s, t, r, colors0);
1551
1552 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1553 /* sample the second mipmap level, and interp */
1554 lp_build_sample_image_nearest(bld,
1555 width1_vec, height1_vec, depth1_vec,
1556 row_stride1_vec, img_stride1_vec,
1557 data_ptr1, s, t, r, colors1);
1558 }
1559 }
1560 else {
1561 assert(min_filter == PIPE_TEX_FILTER_LINEAR);
1562
1563 lp_build_sample_image_linear(bld,
1564 width0_vec, height0_vec, depth0_vec,
1565 row_stride0_vec, img_stride0_vec,
1566 data_ptr0, s, t, r, colors0);
1567
1568
1569 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1570 /* sample the second mipmap level, and interp */
1571 lp_build_sample_image_linear(bld,
1572 width1_vec, height1_vec, depth1_vec,
1573 row_stride1_vec, img_stride1_vec,
1574 data_ptr1, s, t, r, colors1);
1575 }
1576 }
1577
1578 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1579 /* interpolate samples from the two mipmap levels */
1580 for (chan = 0; chan < 4; chan++) {
1581 colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1582 colors0[chan], colors1[chan]);
1583 }
1584 }
1585 else {
1586 /* use first/only level's colors */
1587 for (chan = 0; chan < 4; chan++) {
1588 colors_out[chan] = colors0[chan];
1589 }
1590 }
1591 }
1592 else {
1593 /* emit conditional to choose min image filter or mag image filter
1594 * depending on the lod being >0 or <= 0, respectively.
1595 */
1596 abort();
1597 }
1598 }
1599
1600
1601
1602 static void
1603 lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
1604 struct lp_type dst_type,
1605 LLVMValueRef packed,
1606 LLVMValueRef *rgba)
1607 {
1608 LLVMValueRef mask = lp_build_int_const_scalar(dst_type, 0xff);
1609 unsigned chan;
1610
1611 /* Decode the input vector components */
1612 for (chan = 0; chan < 4; ++chan) {
1613 unsigned start = chan*8;
1614 unsigned stop = start + 8;
1615 LLVMValueRef input;
1616
1617 input = packed;
1618
1619 if(start)
1620 input = LLVMBuildLShr(builder, input, lp_build_int_const_scalar(dst_type, start), "");
1621
1622 if(stop < 32)
1623 input = LLVMBuildAnd(builder, input, mask, "");
1624
1625 input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
1626
1627 rgba[chan] = input;
1628 }
1629 }
1630
1631
1632 static void
1633 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1634 LLVMValueRef s,
1635 LLVMValueRef t,
1636 LLVMValueRef width,
1637 LLVMValueRef height,
1638 LLVMValueRef stride_array,
1639 LLVMValueRef data_array,
1640 LLVMValueRef *texel)
1641 {
1642 LLVMBuilderRef builder = bld->builder;
1643 struct lp_build_context i32, h16, u8n;
1644 LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1645 LLVMValueRef i32_c8, i32_c128, i32_c255;
1646 LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1647 LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1648 LLVMValueRef x0, x1;
1649 LLVMValueRef y0, y1;
1650 LLVMValueRef neighbors[2][2];
1651 LLVMValueRef neighbors_lo[2][2];
1652 LLVMValueRef neighbors_hi[2][2];
1653 LLVMValueRef packed, packed_lo, packed_hi;
1654 LLVMValueRef unswizzled[4];
1655 LLVMValueRef stride;
1656
1657 lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1658 lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1659 lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1660
1661 i32_vec_type = lp_build_vec_type(i32.type);
1662 h16_vec_type = lp_build_vec_type(h16.type);
1663 u8n_vec_type = lp_build_vec_type(u8n.type);
1664
1665 if (bld->static_state->normalized_coords) {
1666 LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1667 LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1668 LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1669 s = lp_build_mul(&bld->coord_bld, s, fp_width);
1670 t = lp_build_mul(&bld->coord_bld, t, fp_height);
1671 }
1672
1673 /* scale coords by 256 (8 fractional bits) */
1674 s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1675 t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1676
1677 /* convert float to int */
1678 s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1679 t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1680
1681 /* subtract 0.5 (add -128) */
1682 i32_c128 = lp_build_int_const_scalar(i32.type, -128);
1683 s = LLVMBuildAdd(builder, s, i32_c128, "");
1684 t = LLVMBuildAdd(builder, t, i32_c128, "");
1685
1686 /* compute floor (shift right 8) */
1687 i32_c8 = lp_build_int_const_scalar(i32.type, 8);
1688 s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1689 t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1690
1691 /* compute fractional part (AND with 0xff) */
1692 i32_c255 = lp_build_int_const_scalar(i32.type, 255);
1693 s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1694 t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1695
1696 x0 = s_ipart;
1697 y0 = t_ipart;
1698
1699 x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
1700 y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
1701
1702 x0 = lp_build_sample_wrap_int(bld, x0, width, bld->static_state->pot_width,
1703 bld->static_state->wrap_s);
1704 y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
1705 bld->static_state->wrap_t);
1706
1707 x1 = lp_build_sample_wrap_int(bld, x1, width, bld->static_state->pot_width,
1708 bld->static_state->wrap_s);
1709 y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
1710 bld->static_state->wrap_t);
1711
1712 /*
1713 * Transform 4 x i32 in
1714 *
1715 * s_fpart = {s0, s1, s2, s3}
1716 *
1717 * into 8 x i16
1718 *
1719 * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1720 *
1721 * into two 8 x i16
1722 *
1723 * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1724 * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1725 *
1726 * and likewise for t_fpart. There is no risk of loosing precision here
1727 * since the fractional parts only use the lower 8bits.
1728 */
1729
1730 s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1731 t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1732
1733 {
1734 LLVMTypeRef elem_type = LLVMInt32Type();
1735 LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1736 LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1737 LLVMValueRef shuffle_lo;
1738 LLVMValueRef shuffle_hi;
1739 unsigned i, j;
1740
1741 for(j = 0; j < h16.type.length; j += 4) {
1742 unsigned subindex = util_cpu_caps.little_endian ? 0 : 1;
1743 LLVMValueRef index;
1744
1745 index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1746 for(i = 0; i < 4; ++i)
1747 shuffles_lo[j + i] = index;
1748
1749 index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1750 for(i = 0; i < 4; ++i)
1751 shuffles_hi[j + i] = index;
1752 }
1753
1754 shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1755 shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1756
1757 s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1758 t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1759 s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1760 t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1761 }
1762
1763 stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
1764
1765 /*
1766 * Fetch the pixels as 4 x 32bit (rgba order might differ):
1767 *
1768 * rgba0 rgba1 rgba2 rgba3
1769 *
1770 * bit cast them into 16 x u8
1771 *
1772 * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1773 *
1774 * unpack them into two 8 x i16:
1775 *
1776 * r0 g0 b0 a0 r1 g1 b1 a1
1777 * r2 g2 b2 a2 r3 g3 b3 a3
1778 *
1779 * The higher 8 bits of the resulting elements will be zero.
1780 */
1781
1782 neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array);
1783 neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array);
1784 neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array);
1785 neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array);
1786
1787 neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, "");
1788 neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, "");
1789 neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, "");
1790 neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, "");
1791
1792 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]);
1793 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]);
1794 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]);
1795 lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]);
1796
1797 /*
1798 * Linear interpolate with 8.8 fixed point.
1799 */
1800
1801 packed_lo = lp_build_lerp_2d(&h16,
1802 s_fpart_lo, t_fpart_lo,
1803 neighbors_lo[0][0],
1804 neighbors_lo[0][1],
1805 neighbors_lo[1][0],
1806 neighbors_lo[1][1]);
1807
1808 packed_hi = lp_build_lerp_2d(&h16,
1809 s_fpart_hi, t_fpart_hi,
1810 neighbors_hi[0][0],
1811 neighbors_hi[0][1],
1812 neighbors_hi[1][0],
1813 neighbors_hi[1][1]);
1814
1815 packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
1816
1817 /*
1818 * Convert to SoA and swizzle.
1819 */
1820
1821 packed = LLVMBuildBitCast(builder, packed, i32_vec_type, "");
1822
1823 lp_build_rgba8_to_f32_soa(bld->builder,
1824 bld->texel_type,
1825 packed, unswizzled);
1826
1827 lp_build_format_swizzle_soa(bld->format_desc,
1828 bld->texel_type, unswizzled,
1829 texel);
1830 }
1831
1832
1833 static void
1834 lp_build_sample_compare(struct lp_build_sample_context *bld,
1835 LLVMValueRef p,
1836 LLVMValueRef *texel)
1837 {
1838 struct lp_build_context *texel_bld = &bld->texel_bld;
1839 LLVMValueRef res;
1840 unsigned chan;
1841
1842 if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1843 return;
1844
1845 /* TODO: Compare before swizzling, to avoid redundant computations */
1846 res = NULL;
1847 for(chan = 0; chan < 4; ++chan) {
1848 LLVMValueRef cmp;
1849 cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
1850 cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
1851
1852 if(res)
1853 res = lp_build_add(texel_bld, res, cmp);
1854 else
1855 res = cmp;
1856 }
1857
1858 assert(res);
1859 res = lp_build_mul(texel_bld, res, lp_build_const_scalar(texel_bld->type, 0.25));
1860
1861 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1862 for(chan = 0; chan < 3; ++chan)
1863 texel[chan] = res;
1864 texel[3] = texel_bld->one;
1865 }
1866
1867
1868 /**
1869 * Build texture sampling code.
1870 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1871 * R, G, B, A.
1872 * \param type vector float type to use for coords, etc.
1873 */
1874 void
1875 lp_build_sample_soa(LLVMBuilderRef builder,
1876 const struct lp_sampler_static_state *static_state,
1877 struct lp_sampler_dynamic_state *dynamic_state,
1878 struct lp_type type,
1879 unsigned unit,
1880 unsigned num_coords,
1881 const LLVMValueRef *coords,
1882 LLVMValueRef lodbias,
1883 LLVMValueRef *texel)
1884 {
1885 struct lp_build_sample_context bld;
1886 LLVMValueRef width, width_vec;
1887 LLVMValueRef height, height_vec;
1888 LLVMValueRef depth, depth_vec;
1889 LLVMValueRef stride_array;
1890 LLVMValueRef data_array;
1891 LLVMValueRef s;
1892 LLVMValueRef t;
1893 LLVMValueRef r;
1894
1895 (void) lp_build_lod_selector; /* temporary to silence warning */
1896 (void) lp_build_nearest_mip_level;
1897 (void) lp_build_linear_mip_levels;
1898 (void) lp_build_minify;
1899
1900 /* Setup our build context */
1901 memset(&bld, 0, sizeof bld);
1902 bld.builder = builder;
1903 bld.static_state = static_state;
1904 bld.dynamic_state = dynamic_state;
1905 bld.format_desc = util_format_description(static_state->format);
1906
1907 bld.float_type = lp_type_float(32);
1908 bld.int_type = lp_type_int(32);
1909 bld.coord_type = type;
1910 bld.uint_coord_type = lp_uint_type(type);
1911 bld.int_coord_type = lp_int_type(type);
1912 bld.texel_type = type;
1913
1914 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
1915 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
1916 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
1917 lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
1918 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
1919 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
1920
1921 /* Get the dynamic state */
1922 width = dynamic_state->width(dynamic_state, builder, unit);
1923 height = dynamic_state->height(dynamic_state, builder, unit);
1924 depth = dynamic_state->depth(dynamic_state, builder, unit);
1925 stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
1926 data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
1927 /* Note that data_array is an array[level] of pointers to texture images */
1928
1929 s = coords[0];
1930 t = coords[1];
1931 r = coords[2];
1932
1933 width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
1934 height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
1935 depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
1936
1937 if (lp_format_is_rgba8(bld.format_desc) &&
1938 static_state->target == PIPE_TEXTURE_2D &&
1939 static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
1940 static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
1941 static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
1942 is_simple_wrap_mode(static_state->wrap_s) &&
1943 is_simple_wrap_mode(static_state->wrap_t)) {
1944 /* special case */
1945 lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
1946 stride_array, data_array, texel);
1947 }
1948 else {
1949 lp_build_sample_general(&bld, unit, s, t, r,
1950 width, height, depth,
1951 width_vec, height_vec, depth_vec,
1952 stride_array, NULL, data_array,
1953 texel);
1954 }
1955
1956 lp_build_sample_compare(&bld, r, texel);
1957 }