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 "pipe/p_shader_tokens.h"
39 #include "util/u_debug.h"
40 #include "util/u_dump.h"
41 #include "util/u_memory.h"
42 #include "util/u_math.h"
43 #include "util/u_format.h"
44 #include "util/u_cpu_detect.h"
45 #include "util/format_rgb9e5.h"
46 #include "lp_bld_debug.h"
47 #include "lp_bld_type.h"
48 #include "lp_bld_const.h"
49 #include "lp_bld_conv.h"
50 #include "lp_bld_arit.h"
51 #include "lp_bld_bitarit.h"
52 #include "lp_bld_logic.h"
53 #include "lp_bld_printf.h"
54 #include "lp_bld_swizzle.h"
55 #include "lp_bld_flow.h"
56 #include "lp_bld_gather.h"
57 #include "lp_bld_format.h"
58 #include "lp_bld_sample.h"
59 #include "lp_bld_sample_aos.h"
60 #include "lp_bld_struct.h"
61 #include "lp_bld_quad.h"
62 #include "lp_bld_pack.h"
63 #include "lp_bld_intr.h"
64
65
66 /**
67 * Generate code to fetch a texel from a texture at int coords (x, y, z).
68 * The computation depends on whether the texture is 1D, 2D or 3D.
69 * The result, texel, will be float vectors:
70 * texel[0] = red values
71 * texel[1] = green values
72 * texel[2] = blue values
73 * texel[3] = alpha values
74 */
75 static void
lp_build_sample_texel_soa(struct lp_build_sample_context * bld,LLVMValueRef width,LLVMValueRef height,LLVMValueRef depth,LLVMValueRef x,LLVMValueRef y,LLVMValueRef z,LLVMValueRef y_stride,LLVMValueRef z_stride,LLVMValueRef data_ptr,LLVMValueRef mipoffsets,LLVMValueRef texel_out[4])76 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
77 LLVMValueRef width,
78 LLVMValueRef height,
79 LLVMValueRef depth,
80 LLVMValueRef x,
81 LLVMValueRef y,
82 LLVMValueRef z,
83 LLVMValueRef y_stride,
84 LLVMValueRef z_stride,
85 LLVMValueRef data_ptr,
86 LLVMValueRef mipoffsets,
87 LLVMValueRef texel_out[4])
88 {
89 const struct lp_static_sampler_state *static_state = bld->static_sampler_state;
90 const unsigned dims = bld->dims;
91 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
92 LLVMBuilderRef builder = bld->gallivm->builder;
93 LLVMValueRef offset;
94 LLVMValueRef i, j;
95 LLVMValueRef use_border = NULL;
96
97 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
98 if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
99 static_state->min_img_filter,
100 static_state->mag_img_filter)) {
101 LLVMValueRef b1, b2;
102 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
103 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
104 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
105 }
106
107 if (dims >= 2 &&
108 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
109 static_state->min_img_filter,
110 static_state->mag_img_filter)) {
111 LLVMValueRef b1, b2;
112 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
113 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
114 if (use_border) {
115 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
116 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
117 }
118 else {
119 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
120 }
121 }
122
123 if (dims == 3 &&
124 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
125 static_state->min_img_filter,
126 static_state->mag_img_filter)) {
127 LLVMValueRef b1, b2;
128 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
129 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
130 if (use_border) {
131 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
132 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
133 }
134 else {
135 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
136 }
137 }
138
139 /* convert x,y,z coords to linear offset from start of texture, in bytes */
140 lp_build_sample_offset(&bld->int_coord_bld,
141 bld->format_desc,
142 x, y, z, y_stride, z_stride,
143 &offset, &i, &j);
144 if (mipoffsets) {
145 offset = lp_build_add(&bld->int_coord_bld, offset, mipoffsets);
146 }
147
148 if (use_border) {
149 /* If we can sample the border color, it means that texcoords may
150 * lie outside the bounds of the texture image. We need to do
151 * something to prevent reading out of bounds and causing a segfault.
152 *
153 * Simply AND the texture coords with !use_border. This will cause
154 * coords which are out of bounds to become zero. Zero's guaranteed
155 * to be inside the texture image.
156 */
157 offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
158 }
159
160 lp_build_fetch_rgba_soa(bld->gallivm,
161 bld->format_desc,
162 bld->texel_type, TRUE,
163 data_ptr, offset,
164 i, j,
165 bld->cache,
166 texel_out);
167
168 /*
169 * Note: if we find an app which frequently samples the texture border
170 * we might want to implement a true conditional here to avoid sampling
171 * the texture whenever possible (since that's quite a bit of code).
172 * Ex:
173 * if (use_border) {
174 * texel = border_color;
175 * }
176 * else {
177 * texel = sample_texture(coord);
178 * }
179 * As it is now, we always sample the texture, then selectively replace
180 * the texel color results with the border color.
181 */
182
183 if (use_border) {
184 /* select texel color or border color depending on use_border. */
185 const struct util_format_description *format_desc = bld->format_desc;
186 int chan;
187 struct lp_type border_type = bld->texel_type;
188 border_type.length = 4;
189 /*
190 * Only replace channels which are actually present. The others should
191 * get optimized away eventually by sampler_view swizzle anyway but it's
192 * easier too.
193 */
194 for (chan = 0; chan < 4; chan++) {
195 unsigned chan_s;
196 /* reverse-map channel... */
197 for (chan_s = 0; chan_s < 4; chan_s++) {
198 if (chan_s == format_desc->swizzle[chan]) {
199 break;
200 }
201 }
202 if (chan_s <= 3) {
203 /* use the already clamped color */
204 LLVMValueRef idx = lp_build_const_int32(bld->gallivm, chan);
205 LLVMValueRef border_chan;
206
207 border_chan = lp_build_extract_broadcast(bld->gallivm,
208 border_type,
209 bld->texel_type,
210 bld->border_color_clamped,
211 idx);
212 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
213 border_chan, texel_out[chan]);
214 }
215 }
216 }
217 }
218
219
220 /**
221 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR_REPEAT mode.
222 * (Note that with pot sizes could do this much more easily post-scale
223 * with some bit arithmetic.)
224 */
225 static LLVMValueRef
lp_build_coord_mirror(struct lp_build_sample_context * bld,LLVMValueRef coord,boolean posOnly)226 lp_build_coord_mirror(struct lp_build_sample_context *bld,
227 LLVMValueRef coord, boolean posOnly)
228 {
229 struct lp_build_context *coord_bld = &bld->coord_bld;
230 LLVMValueRef fract;
231 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
232
233 /*
234 * We can just use 2*(x - round(0.5*x)) to do all the mirroring,
235 * it all works out. (The result is in range [-1, 1.0], negative if
236 * the coord is in the "odd" section, otherwise positive.)
237 */
238
239 coord = lp_build_mul(coord_bld, coord, half);
240 fract = lp_build_round(coord_bld, coord);
241 fract = lp_build_sub(coord_bld, coord, fract);
242 coord = lp_build_add(coord_bld, fract, fract);
243
244 if (posOnly) {
245 /*
246 * Theoretically it's not quite 100% accurate because the spec says
247 * that ultimately a scaled coord of -x.0 should map to int coord
248 * -x + 1 with mirroring, not -x (this does not matter for bilinear
249 * filtering).
250 */
251 coord = lp_build_abs(coord_bld, coord);
252 /* kill off NaNs */
253 /* XXX: not safe without arch rounding, fract can be anything. */
254 coord = lp_build_max_ext(coord_bld, coord, coord_bld->zero,
255 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
256 }
257
258 return coord;
259 }
260
261
262 /**
263 * Helper to compute the first coord and the weight for
264 * linear wrap repeat npot textures
265 */
266 void
lp_build_coord_repeat_npot_linear(struct lp_build_sample_context * bld,LLVMValueRef coord_f,LLVMValueRef length_i,LLVMValueRef length_f,LLVMValueRef * coord0_i,LLVMValueRef * weight_f)267 lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld,
268 LLVMValueRef coord_f,
269 LLVMValueRef length_i,
270 LLVMValueRef length_f,
271 LLVMValueRef *coord0_i,
272 LLVMValueRef *weight_f)
273 {
274 struct lp_build_context *coord_bld = &bld->coord_bld;
275 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
276 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
277 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
278 int_coord_bld->one);
279 LLVMValueRef mask;
280 /* wrap with normalized floats is just fract */
281 coord_f = lp_build_fract(coord_bld, coord_f);
282 /* mul by size and subtract 0.5 */
283 coord_f = lp_build_mul(coord_bld, coord_f, length_f);
284 coord_f = lp_build_sub(coord_bld, coord_f, half);
285 /*
286 * we avoided the 0.5/length division before the repeat wrap,
287 * now need to fix up edge cases with selects
288 */
289 /*
290 * Note we do a float (unordered) compare so we can eliminate NaNs.
291 * (Otherwise would need fract_safe above).
292 */
293 mask = lp_build_compare(coord_bld->gallivm, coord_bld->type,
294 PIPE_FUNC_LESS, coord_f, coord_bld->zero);
295
296 /* convert to int, compute lerp weight */
297 lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f);
298 *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
299 }
300
301
302 /**
303 * Build LLVM code for texture wrap mode for linear filtering.
304 * \param x0_out returns first integer texcoord
305 * \param x1_out returns second integer texcoord
306 * \param weight_out returns linear interpolation weight
307 */
308 static void
lp_build_sample_wrap_linear(struct lp_build_sample_context * bld,boolean is_gather,LLVMValueRef coord,LLVMValueRef length,LLVMValueRef length_f,LLVMValueRef offset,boolean is_pot,unsigned wrap_mode,LLVMValueRef * x0_out,LLVMValueRef * x1_out,LLVMValueRef * weight_out)309 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
310 boolean is_gather,
311 LLVMValueRef coord,
312 LLVMValueRef length,
313 LLVMValueRef length_f,
314 LLVMValueRef offset,
315 boolean is_pot,
316 unsigned wrap_mode,
317 LLVMValueRef *x0_out,
318 LLVMValueRef *x1_out,
319 LLVMValueRef *weight_out)
320 {
321 struct lp_build_context *coord_bld = &bld->coord_bld;
322 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
323 LLVMBuilderRef builder = bld->gallivm->builder;
324 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
325 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
326 LLVMValueRef coord0, coord1, weight;
327
328 switch(wrap_mode) {
329 case PIPE_TEX_WRAP_REPEAT:
330 if (is_pot) {
331 /* mul by size and subtract 0.5 */
332 coord = lp_build_mul(coord_bld, coord, length_f);
333 coord = lp_build_sub(coord_bld, coord, half);
334 if (offset) {
335 offset = lp_build_int_to_float(coord_bld, offset);
336 coord = lp_build_add(coord_bld, coord, offset);
337 }
338 /* convert to int, compute lerp weight */
339 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
340 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
341 /* repeat wrap */
342 coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
343 coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, "");
344 }
345 else {
346 LLVMValueRef mask;
347 if (offset) {
348 offset = lp_build_int_to_float(coord_bld, offset);
349 offset = lp_build_div(coord_bld, offset, length_f);
350 coord = lp_build_add(coord_bld, coord, offset);
351 }
352 lp_build_coord_repeat_npot_linear(bld, coord,
353 length, length_f,
354 &coord0, &weight);
355 mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
356 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
357 coord1 = LLVMBuildAnd(builder,
358 lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
359 mask, "");
360 }
361 break;
362
363 case PIPE_TEX_WRAP_CLAMP:
364 if (bld->static_sampler_state->normalized_coords) {
365 /* scale coord to length */
366 coord = lp_build_mul(coord_bld, coord, length_f);
367 }
368 if (offset) {
369 offset = lp_build_int_to_float(coord_bld, offset);
370 coord = lp_build_add(coord_bld, coord, offset);
371 }
372
373 /*
374 * clamp to [0, length]
375 *
376 * Unlike some other wrap modes, this should be correct for gather
377 * too. GL_CLAMP explicitly does this clamp on the coord prior to
378 * actual wrapping (which is per sample).
379 */
380 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
381
382 coord = lp_build_sub(coord_bld, coord, half);
383
384 /* convert to int, compute lerp weight */
385 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
386 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
387 break;
388
389 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
390 {
391 struct lp_build_context abs_coord_bld = bld->coord_bld;
392 abs_coord_bld.type.sign = FALSE;
393
394 if (bld->static_sampler_state->normalized_coords) {
395 /* mul by tex size */
396 coord = lp_build_mul(coord_bld, coord, length_f);
397 }
398 if (offset) {
399 offset = lp_build_int_to_float(coord_bld, offset);
400 coord = lp_build_add(coord_bld, coord, offset);
401 }
402
403 /* clamp to length max */
404 coord = lp_build_min_ext(coord_bld, coord, length_f,
405 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
406 if (!is_gather) {
407 /* subtract 0.5 */
408 coord = lp_build_sub(coord_bld, coord, half);
409 /* clamp to [0, length - 0.5] */
410 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
411 /* convert to int, compute lerp weight */
412 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
413 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
414 } else {
415 /*
416 * The non-gather path will end up with coords 0, 1 if coord was
417 * smaller than 0.5 (with corresponding weight 0.0 so it doesn't
418 * really matter what the second coord is). But for gather, we
419 * really need to end up with coords 0, 0.
420 */
421 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
422 coord0 = lp_build_sub(coord_bld, coord, half);
423 coord1 = lp_build_add(coord_bld, coord, half);
424 /* Values range ([-0.5, length_f - 0.5], [0.5, length_f + 0.5] */
425 coord0 = lp_build_itrunc(coord_bld, coord0);
426 coord1 = lp_build_itrunc(coord_bld, coord1);
427 weight = coord_bld->undef;
428 }
429 /* coord1 = min(coord1, length-1) */
430 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
431 break;
432 }
433
434 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
435 if (bld->static_sampler_state->normalized_coords) {
436 /* scale coord to length */
437 coord = lp_build_mul(coord_bld, coord, length_f);
438 }
439 if (offset) {
440 offset = lp_build_int_to_float(coord_bld, offset);
441 coord = lp_build_add(coord_bld, coord, offset);
442 }
443 /*
444 * We don't need any clamp. Technically, for very large (pos or neg)
445 * (or infinite) values, clamp against [-length, length] would be
446 * correct, but we don't need to guarantee any specific
447 * result for such coords (the ifloor will be undefined, but for modes
448 * requiring border all resulting coords are safe).
449 */
450 coord = lp_build_sub(coord_bld, coord, half);
451 /* convert to int, compute lerp weight */
452 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
453 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
454 break;
455
456 case PIPE_TEX_WRAP_MIRROR_REPEAT:
457 if (offset) {
458 offset = lp_build_int_to_float(coord_bld, offset);
459 offset = lp_build_div(coord_bld, offset, length_f);
460 coord = lp_build_add(coord_bld, coord, offset);
461 }
462 if (!is_gather) {
463 /* compute mirror function */
464 coord = lp_build_coord_mirror(bld, coord, TRUE);
465
466 /* scale coord to length */
467 coord = lp_build_mul(coord_bld, coord, length_f);
468 coord = lp_build_sub(coord_bld, coord, half);
469
470 /* convert to int, compute lerp weight */
471 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
472 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
473
474 /* coord0 = max(coord0, 0) */
475 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
476 /* coord1 = min(coord1, length-1) */
477 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
478 } else {
479 /*
480 * This is pretty reasonable in the end, all what the tests care
481 * about is nasty edge cases (scaled coords x.5, so the individual
482 * coords are actually integers, which is REALLY tricky to get right
483 * due to this working differently both for negative numbers as well
484 * as for even/odd cases). But with enough magic it's not too complex
485 * after all.
486 * Maybe should try a bit arithmetic one though for POT textures...
487 */
488 LLVMValueRef isNeg;
489 /*
490 * Wrapping just once still works, even though it means we can
491 * get "wrong" sign due to performing mirror in the middle of the
492 * two coords (because this can only happen very near the odd/even
493 * edges, so both coords will actually end up as 0 or length - 1
494 * in the end).
495 * For GL4 gather with per-sample offsets we'd need to the mirroring
496 * per coord too.
497 */
498 coord = lp_build_coord_mirror(bld, coord, FALSE);
499 coord = lp_build_mul(coord_bld, coord, length_f);
500
501 /*
502 * NaNs should be safe here, we'll do away with them with
503 * the ones' complement plus min.
504 */
505 coord0 = lp_build_sub(coord_bld, coord, half);
506 coord0 = lp_build_ifloor(coord_bld, coord0);
507 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
508 /* ones complement for neg numbers (mirror(negX) = X - 1) */
509 isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS,
510 coord0, int_coord_bld->zero);
511 coord0 = lp_build_xor(int_coord_bld, coord0, isNeg);
512 isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS,
513 coord1, int_coord_bld->zero);
514 coord1 = lp_build_xor(int_coord_bld, coord1, isNeg);
515 coord0 = lp_build_min(int_coord_bld, coord0, length_minus_one);
516 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
517
518 weight = coord_bld->undef;
519 }
520 break;
521
522 case PIPE_TEX_WRAP_MIRROR_CLAMP:
523 if (bld->static_sampler_state->normalized_coords) {
524 /* scale coord to length */
525 coord = lp_build_mul(coord_bld, coord, length_f);
526 }
527 if (offset) {
528 offset = lp_build_int_to_float(coord_bld, offset);
529 coord = lp_build_add(coord_bld, coord, offset);
530 }
531 /*
532 * XXX: probably not correct for gather, albeit I'm not
533 * entirely sure as it's poorly specified. The wrapping looks
534 * correct according to the spec which is against gl 1.2.1,
535 * however negative values will be swapped - gl re-specified
536 * wrapping with newer versions (no more pre-clamp except with
537 * GL_CLAMP).
538 */
539 coord = lp_build_abs(coord_bld, coord);
540
541 /* clamp to [0, length] */
542 coord = lp_build_min_ext(coord_bld, coord, length_f,
543 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
544
545 coord = lp_build_sub(coord_bld, coord, half);
546
547 /* convert to int, compute lerp weight */
548 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
549 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
550 break;
551
552 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
553 {
554 struct lp_build_context abs_coord_bld = bld->coord_bld;
555 abs_coord_bld.type.sign = FALSE;
556
557 if (bld->static_sampler_state->normalized_coords) {
558 /* scale coord to length */
559 coord = lp_build_mul(coord_bld, coord, length_f);
560 }
561 if (offset) {
562 offset = lp_build_int_to_float(coord_bld, offset);
563 coord = lp_build_add(coord_bld, coord, offset);
564 }
565 if (!is_gather) {
566 coord = lp_build_abs(coord_bld, coord);
567
568 /* clamp to length max */
569 coord = lp_build_min_ext(coord_bld, coord, length_f,
570 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
571 /* subtract 0.5 */
572 coord = lp_build_sub(coord_bld, coord, half);
573 /* clamp to [0, length - 0.5] */
574 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
575
576 /* convert to int, compute lerp weight */
577 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
578 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
579 /* coord1 = min(coord1, length-1) */
580 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
581 } else {
582 /*
583 * The non-gather path will swap coord0/1 if coord was negative,
584 * which is ok for filtering since the filter weight matches
585 * accordingly. Also, if coord is close to zero, coord0/1 will
586 * be 0 and 1, instead of 0 and 0 (again ok due to filter
587 * weight being 0.0). Both issues need to be fixed for gather.
588 */
589 LLVMValueRef isNeg;
590
591 /*
592 * Actually wanted to cheat here and use:
593 * coord1 = lp_build_iround(coord_bld, coord);
594 * but it's not good enough for some tests (even piglit
595 * textureGather is set up in a way so the coords area always
596 * .5, that is right at the crossover points).
597 * So do ordinary sub/floor, then do ones' complement
598 * for negative numbers.
599 * (Note can't just do sub|add/abs/itrunc per coord neither -
600 * because the spec demands that mirror(3.0) = 3 but
601 * mirror(-3.0) = 2.)
602 */
603 coord = lp_build_sub(coord_bld, coord, half);
604 coord0 = lp_build_ifloor(coord_bld, coord);
605 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
606 isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, coord0,
607 int_coord_bld->zero);
608 coord0 = lp_build_xor(int_coord_bld, isNeg, coord0);
609 coord0 = lp_build_min(int_coord_bld, coord0, length_minus_one);
610
611 isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, coord1,
612 int_coord_bld->zero);
613 coord1 = lp_build_xor(int_coord_bld, isNeg, coord1);
614 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
615
616 weight = coord_bld->undef;
617 }
618 }
619 break;
620
621 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
622 {
623 if (bld->static_sampler_state->normalized_coords) {
624 /* scale coord to length */
625 coord = lp_build_mul(coord_bld, coord, length_f);
626 }
627 if (offset) {
628 offset = lp_build_int_to_float(coord_bld, offset);
629 coord = lp_build_add(coord_bld, coord, offset);
630 }
631 /*
632 * XXX: probably not correct for gather due to swapped
633 * order if coord is negative (same rationale as for
634 * MIRROR_CLAMP).
635 */
636 coord = lp_build_abs(coord_bld, coord);
637
638 /*
639 * We don't need any clamp. Technically, for very large
640 * (or infinite) values, clamp against length would be
641 * correct, but we don't need to guarantee any specific
642 * result for such coords (the ifloor will be undefined, but
643 * for modes requiring border all resulting coords are safe).
644 */
645 coord = lp_build_sub(coord_bld, coord, half);
646
647 /* convert to int, compute lerp weight */
648 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
649 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
650 }
651 break;
652
653 default:
654 assert(0);
655 coord0 = NULL;
656 coord1 = NULL;
657 weight = NULL;
658 }
659
660 *x0_out = coord0;
661 *x1_out = coord1;
662 *weight_out = weight;
663 }
664
665
666 /**
667 * Build LLVM code for texture wrap mode for nearest filtering.
668 * \param coord the incoming texcoord (nominally in [0,1])
669 * \param length the texture size along one dimension, as int vector
670 * \param length_f the texture size along one dimension, as float vector
671 * \param offset texel offset along one dimension (as int vector)
672 * \param is_pot if TRUE, length is a power of two
673 * \param wrap_mode one of PIPE_TEX_WRAP_x
674 */
675 static LLVMValueRef
lp_build_sample_wrap_nearest(struct lp_build_sample_context * bld,LLVMValueRef coord,LLVMValueRef length,LLVMValueRef length_f,LLVMValueRef offset,boolean is_pot,unsigned wrap_mode)676 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
677 LLVMValueRef coord,
678 LLVMValueRef length,
679 LLVMValueRef length_f,
680 LLVMValueRef offset,
681 boolean is_pot,
682 unsigned wrap_mode)
683 {
684 struct lp_build_context *coord_bld = &bld->coord_bld;
685 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
686 LLVMBuilderRef builder = bld->gallivm->builder;
687 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
688 LLVMValueRef icoord;
689
690 switch(wrap_mode) {
691 case PIPE_TEX_WRAP_REPEAT:
692 if (is_pot) {
693 coord = lp_build_mul(coord_bld, coord, length_f);
694 icoord = lp_build_ifloor(coord_bld, coord);
695 if (offset) {
696 icoord = lp_build_add(int_coord_bld, icoord, offset);
697 }
698 icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
699 }
700 else {
701 if (offset) {
702 offset = lp_build_int_to_float(coord_bld, offset);
703 offset = lp_build_div(coord_bld, offset, length_f);
704 coord = lp_build_add(coord_bld, coord, offset);
705 }
706 /* take fraction, unnormalize */
707 coord = lp_build_fract_safe(coord_bld, coord);
708 coord = lp_build_mul(coord_bld, coord, length_f);
709 icoord = lp_build_itrunc(coord_bld, coord);
710 }
711 break;
712
713 case PIPE_TEX_WRAP_CLAMP:
714 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
715 if (bld->static_sampler_state->normalized_coords) {
716 /* scale coord to length */
717 coord = lp_build_mul(coord_bld, coord, length_f);
718 }
719
720 if (offset) {
721 offset = lp_build_int_to_float(coord_bld, offset);
722 coord = lp_build_add(coord_bld, coord, offset);
723 }
724 /* floor */
725 /* use itrunc instead since we clamp to 0 anyway */
726 icoord = lp_build_itrunc(coord_bld, coord);
727
728 /* clamp to [0, length - 1]. */
729 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
730 length_minus_one);
731 break;
732
733 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
734 if (bld->static_sampler_state->normalized_coords) {
735 /* scale coord to length */
736 coord = lp_build_mul(coord_bld, coord, length_f);
737 }
738 /* no clamp necessary, border masking will handle this */
739 icoord = lp_build_ifloor(coord_bld, coord);
740 if (offset) {
741 icoord = lp_build_add(int_coord_bld, icoord, offset);
742 }
743 break;
744
745 case PIPE_TEX_WRAP_MIRROR_REPEAT:
746 if (offset) {
747 offset = lp_build_int_to_float(coord_bld, offset);
748 offset = lp_build_div(coord_bld, offset, length_f);
749 coord = lp_build_add(coord_bld, coord, offset);
750 }
751 /* compute mirror function */
752 coord = lp_build_coord_mirror(bld, coord, TRUE);
753
754 /* scale coord to length */
755 assert(bld->static_sampler_state->normalized_coords);
756 coord = lp_build_mul(coord_bld, coord, length_f);
757
758 /* itrunc == ifloor here */
759 icoord = lp_build_itrunc(coord_bld, coord);
760
761 /* clamp to [0, length - 1] */
762 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
763 break;
764
765 case PIPE_TEX_WRAP_MIRROR_CLAMP:
766 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
767 if (bld->static_sampler_state->normalized_coords) {
768 /* scale coord to length */
769 coord = lp_build_mul(coord_bld, coord, length_f);
770 }
771 if (offset) {
772 offset = lp_build_int_to_float(coord_bld, offset);
773 coord = lp_build_add(coord_bld, coord, offset);
774 }
775 coord = lp_build_abs(coord_bld, coord);
776
777 /* itrunc == ifloor here */
778 icoord = lp_build_itrunc(coord_bld, coord);
779 /*
780 * Use unsigned min due to possible undef values (NaNs, overflow)
781 */
782 {
783 struct lp_build_context abs_coord_bld = *int_coord_bld;
784 abs_coord_bld.type.sign = FALSE;
785 /* clamp to [0, length - 1] */
786 icoord = lp_build_min(&abs_coord_bld, icoord, length_minus_one);
787 }
788 break;
789
790 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
791 if (bld->static_sampler_state->normalized_coords) {
792 /* scale coord to length */
793 coord = lp_build_mul(coord_bld, coord, length_f);
794 }
795 if (offset) {
796 offset = lp_build_int_to_float(coord_bld, offset);
797 coord = lp_build_add(coord_bld, coord, offset);
798 }
799 coord = lp_build_abs(coord_bld, coord);
800
801 /* itrunc == ifloor here */
802 icoord = lp_build_itrunc(coord_bld, coord);
803 break;
804
805 default:
806 assert(0);
807 icoord = NULL;
808 }
809
810 return icoord;
811 }
812
813
814 /**
815 * Do shadow test/comparison.
816 * \param p shadow ref value
817 * \param texel the texel to compare against
818 */
819 static LLVMValueRef
lp_build_sample_comparefunc(struct lp_build_sample_context * bld,LLVMValueRef p,LLVMValueRef texel)820 lp_build_sample_comparefunc(struct lp_build_sample_context *bld,
821 LLVMValueRef p,
822 LLVMValueRef texel)
823 {
824 struct lp_build_context *texel_bld = &bld->texel_bld;
825 LLVMValueRef res;
826
827 if (0) {
828 //lp_build_print_value(bld->gallivm, "shadow cmp coord", p);
829 lp_build_print_value(bld->gallivm, "shadow cmp texel", texel);
830 }
831
832 /* result = (p FUNC texel) ? 1 : 0 */
833 /*
834 * honor d3d10 floating point rules here, which state that comparisons
835 * are ordered except NOT_EQUAL which is unordered.
836 */
837 if (bld->static_sampler_state->compare_func != PIPE_FUNC_NOTEQUAL) {
838 res = lp_build_cmp_ordered(texel_bld, bld->static_sampler_state->compare_func,
839 p, texel);
840 }
841 else {
842 res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func,
843 p, texel);
844 }
845 return res;
846 }
847
848
849 /**
850 * Generate code to sample a mipmap level with nearest filtering.
851 * If sampling a cube texture, r = cube face in [0,5].
852 */
853 static void
lp_build_sample_image_nearest(struct lp_build_sample_context * bld,LLVMValueRef size,LLVMValueRef row_stride_vec,LLVMValueRef img_stride_vec,LLVMValueRef data_ptr,LLVMValueRef mipoffsets,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef colors_out[4])854 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
855 LLVMValueRef size,
856 LLVMValueRef row_stride_vec,
857 LLVMValueRef img_stride_vec,
858 LLVMValueRef data_ptr,
859 LLVMValueRef mipoffsets,
860 const LLVMValueRef *coords,
861 const LLVMValueRef *offsets,
862 LLVMValueRef colors_out[4])
863 {
864 const unsigned dims = bld->dims;
865 LLVMValueRef width_vec;
866 LLVMValueRef height_vec;
867 LLVMValueRef depth_vec;
868 LLVMValueRef flt_size;
869 LLVMValueRef flt_width_vec;
870 LLVMValueRef flt_height_vec;
871 LLVMValueRef flt_depth_vec;
872 LLVMValueRef x, y = NULL, z = NULL;
873
874 lp_build_extract_image_sizes(bld,
875 &bld->int_size_bld,
876 bld->int_coord_type,
877 size,
878 &width_vec, &height_vec, &depth_vec);
879
880 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
881
882 lp_build_extract_image_sizes(bld,
883 &bld->float_size_bld,
884 bld->coord_type,
885 flt_size,
886 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
887
888 /*
889 * Compute integer texcoords.
890 */
891 x = lp_build_sample_wrap_nearest(bld, coords[0], width_vec,
892 flt_width_vec, offsets[0],
893 bld->static_texture_state->pot_width,
894 bld->static_sampler_state->wrap_s);
895 lp_build_name(x, "tex.x.wrapped");
896
897 if (dims >= 2) {
898 y = lp_build_sample_wrap_nearest(bld, coords[1], height_vec,
899 flt_height_vec, offsets[1],
900 bld->static_texture_state->pot_height,
901 bld->static_sampler_state->wrap_t);
902 lp_build_name(y, "tex.y.wrapped");
903
904 if (dims == 3) {
905 z = lp_build_sample_wrap_nearest(bld, coords[2], depth_vec,
906 flt_depth_vec, offsets[2],
907 bld->static_texture_state->pot_depth,
908 bld->static_sampler_state->wrap_r);
909 lp_build_name(z, "tex.z.wrapped");
910 }
911 }
912 if (has_layer_coord(bld->static_texture_state->target)) {
913 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
914 /* add cube layer to face */
915 z = lp_build_add(&bld->int_coord_bld, coords[2], coords[3]);
916 }
917 else {
918 z = coords[2];
919 }
920 lp_build_name(z, "tex.z.layer");
921 }
922
923 /*
924 * Get texture colors.
925 */
926 lp_build_sample_texel_soa(bld,
927 width_vec, height_vec, depth_vec,
928 x, y, z,
929 row_stride_vec, img_stride_vec,
930 data_ptr, mipoffsets, colors_out);
931
932 if (bld->static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) {
933 LLVMValueRef cmpval;
934 cmpval = lp_build_sample_comparefunc(bld, coords[4], colors_out[0]);
935 /* this is really just a AND 1.0, cmpval but llvm is clever enough */
936 colors_out[0] = lp_build_select(&bld->texel_bld, cmpval,
937 bld->texel_bld.one, bld->texel_bld.zero);
938 colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
939 }
940
941 }
942
943
944 /**
945 * Like a lerp, but inputs are 0/~0 masks, so can simplify slightly.
946 */
947 static LLVMValueRef
lp_build_masklerp(struct lp_build_context * bld,LLVMValueRef weight,LLVMValueRef mask0,LLVMValueRef mask1)948 lp_build_masklerp(struct lp_build_context *bld,
949 LLVMValueRef weight,
950 LLVMValueRef mask0,
951 LLVMValueRef mask1)
952 {
953 struct gallivm_state *gallivm = bld->gallivm;
954 LLVMBuilderRef builder = gallivm->builder;
955 LLVMValueRef weight2;
956
957 weight2 = lp_build_sub(bld, bld->one, weight);
958 weight = LLVMBuildBitCast(builder, weight,
959 lp_build_int_vec_type(gallivm, bld->type), "");
960 weight2 = LLVMBuildBitCast(builder, weight2,
961 lp_build_int_vec_type(gallivm, bld->type), "");
962 weight = LLVMBuildAnd(builder, weight, mask1, "");
963 weight2 = LLVMBuildAnd(builder, weight2, mask0, "");
964 weight = LLVMBuildBitCast(builder, weight, bld->vec_type, "");
965 weight2 = LLVMBuildBitCast(builder, weight2, bld->vec_type, "");
966 return lp_build_add(bld, weight, weight2);
967 }
968
969 /**
970 * Like a 2d lerp, but inputs are 0/~0 masks, so can simplify slightly.
971 */
972 static LLVMValueRef
lp_build_masklerp2d(struct lp_build_context * bld,LLVMValueRef weight0,LLVMValueRef weight1,LLVMValueRef mask00,LLVMValueRef mask01,LLVMValueRef mask10,LLVMValueRef mask11)973 lp_build_masklerp2d(struct lp_build_context *bld,
974 LLVMValueRef weight0,
975 LLVMValueRef weight1,
976 LLVMValueRef mask00,
977 LLVMValueRef mask01,
978 LLVMValueRef mask10,
979 LLVMValueRef mask11)
980 {
981 LLVMValueRef val0 = lp_build_masklerp(bld, weight0, mask00, mask01);
982 LLVMValueRef val1 = lp_build_masklerp(bld, weight0, mask10, mask11);
983 return lp_build_lerp(bld, weight1, val0, val1, 0);
984 }
985
986 /*
987 * this is a bit excessive code for something OpenGL just recommends
988 * but does not require.
989 */
990 #define ACCURATE_CUBE_CORNERS 1
991
992 /**
993 * Generate code to sample a mipmap level with linear filtering.
994 * If sampling a cube texture, r = cube face in [0,5].
995 * If linear_mask is present, only pixels having their mask set
996 * will receive linear filtering, the rest will use nearest.
997 */
998 static void
lp_build_sample_image_linear(struct lp_build_sample_context * bld,boolean is_gather,LLVMValueRef size,LLVMValueRef linear_mask,LLVMValueRef row_stride_vec,LLVMValueRef img_stride_vec,LLVMValueRef data_ptr,LLVMValueRef mipoffsets,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef colors_out[4])999 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1000 boolean is_gather,
1001 LLVMValueRef size,
1002 LLVMValueRef linear_mask,
1003 LLVMValueRef row_stride_vec,
1004 LLVMValueRef img_stride_vec,
1005 LLVMValueRef data_ptr,
1006 LLVMValueRef mipoffsets,
1007 const LLVMValueRef *coords,
1008 const LLVMValueRef *offsets,
1009 LLVMValueRef colors_out[4])
1010 {
1011 LLVMBuilderRef builder = bld->gallivm->builder;
1012 struct lp_build_context *ivec_bld = &bld->int_coord_bld;
1013 struct lp_build_context *coord_bld = &bld->coord_bld;
1014 struct lp_build_context *texel_bld = &bld->texel_bld;
1015 const unsigned dims = bld->dims;
1016 LLVMValueRef width_vec;
1017 LLVMValueRef height_vec;
1018 LLVMValueRef depth_vec;
1019 LLVMValueRef flt_size;
1020 LLVMValueRef flt_width_vec;
1021 LLVMValueRef flt_height_vec;
1022 LLVMValueRef flt_depth_vec;
1023 LLVMValueRef fall_off[4], have_corners;
1024 LLVMValueRef z1 = NULL;
1025 LLVMValueRef z00 = NULL, z01 = NULL, z10 = NULL, z11 = NULL;
1026 LLVMValueRef x00 = NULL, x01 = NULL, x10 = NULL, x11 = NULL;
1027 LLVMValueRef y00 = NULL, y01 = NULL, y10 = NULL, y11 = NULL;
1028 LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
1029 LLVMValueRef xs[4], ys[4], zs[4];
1030 LLVMValueRef neighbors[2][2][4];
1031 int chan, texel_index;
1032 boolean seamless_cube_filter, accurate_cube_corners;
1033 unsigned chan_swiz = bld->static_texture_state->swizzle_r;
1034
1035 seamless_cube_filter = (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
1036 bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) &&
1037 bld->static_sampler_state->seamless_cube_map;
1038
1039 accurate_cube_corners = ACCURATE_CUBE_CORNERS && seamless_cube_filter;
1040
1041 lp_build_extract_image_sizes(bld,
1042 &bld->int_size_bld,
1043 bld->int_coord_type,
1044 size,
1045 &width_vec, &height_vec, &depth_vec);
1046
1047 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
1048
1049 lp_build_extract_image_sizes(bld,
1050 &bld->float_size_bld,
1051 bld->coord_type,
1052 flt_size,
1053 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
1054
1055 /*
1056 * Compute integer texcoords.
1057 */
1058
1059 if (!seamless_cube_filter) {
1060 lp_build_sample_wrap_linear(bld, is_gather, coords[0], width_vec,
1061 flt_width_vec, offsets[0],
1062 bld->static_texture_state->pot_width,
1063 bld->static_sampler_state->wrap_s,
1064 &x00, &x01, &s_fpart);
1065 lp_build_name(x00, "tex.x0.wrapped");
1066 lp_build_name(x01, "tex.x1.wrapped");
1067 x10 = x00;
1068 x11 = x01;
1069
1070 if (dims >= 2) {
1071 lp_build_sample_wrap_linear(bld, is_gather, coords[1], height_vec,
1072 flt_height_vec, offsets[1],
1073 bld->static_texture_state->pot_height,
1074 bld->static_sampler_state->wrap_t,
1075 &y00, &y10, &t_fpart);
1076 lp_build_name(y00, "tex.y0.wrapped");
1077 lp_build_name(y10, "tex.y1.wrapped");
1078 y01 = y00;
1079 y11 = y10;
1080
1081 if (dims == 3) {
1082 lp_build_sample_wrap_linear(bld, is_gather, coords[2], depth_vec,
1083 flt_depth_vec, offsets[2],
1084 bld->static_texture_state->pot_depth,
1085 bld->static_sampler_state->wrap_r,
1086 &z00, &z1, &r_fpart);
1087 z01 = z10 = z11 = z00;
1088 lp_build_name(z00, "tex.z0.wrapped");
1089 lp_build_name(z1, "tex.z1.wrapped");
1090 }
1091 }
1092 if (has_layer_coord(bld->static_texture_state->target)) {
1093 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
1094 /* add cube layer to face */
1095 z00 = z01 = z10 = z11 = z1 =
1096 lp_build_add(&bld->int_coord_bld, coords[2], coords[3]);
1097 }
1098 else {
1099 z00 = z01 = z10 = z11 = z1 = coords[2]; /* cube face or layer */
1100 }
1101 lp_build_name(z00, "tex.z0.layer");
1102 lp_build_name(z1, "tex.z1.layer");
1103 }
1104 }
1105 else {
1106 struct lp_build_if_state edge_if;
1107 LLVMTypeRef int1t;
1108 LLVMValueRef new_faces[4], new_xcoords[4][2], new_ycoords[4][2];
1109 LLVMValueRef coord0, coord1, have_edge, have_corner;
1110 LLVMValueRef fall_off_ym_notxm, fall_off_ym_notxp, fall_off_x, fall_off_y;
1111 LLVMValueRef fall_off_yp_notxm, fall_off_yp_notxp;
1112 LLVMValueRef x0, x1, y0, y1, y0_clamped, y1_clamped;
1113 LLVMValueRef face = coords[2];
1114 LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5f);
1115 LLVMValueRef length_minus_one = lp_build_sub(ivec_bld, width_vec, ivec_bld->one);
1116 /* XXX drop height calcs. Could (should) do this without seamless filtering too */
1117 height_vec = width_vec;
1118 flt_height_vec = flt_width_vec;
1119
1120 /* XXX the overflow logic is actually sort of duplicated with trilinear,
1121 * since an overflow in one mip should also have a corresponding overflow
1122 * in another.
1123 */
1124 /* should always have normalized coords, and offsets are undefined */
1125 assert(bld->static_sampler_state->normalized_coords);
1126 /*
1127 * The coords should all be between [0,1] however we can have NaNs,
1128 * which will wreak havoc. In particular the y1_clamped value below
1129 * can be -INT_MAX (on x86) and be propagated right through (probably
1130 * other values might be bogus in the end too).
1131 * So kill off the NaNs here.
1132 */
1133 coord0 = lp_build_max_ext(coord_bld, coords[0], coord_bld->zero,
1134 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
1135 coord0 = lp_build_mul(coord_bld, coord0, flt_width_vec);
1136 /* instead of clamp, build mask if overflowed */
1137 coord0 = lp_build_sub(coord_bld, coord0, half);
1138 /* convert to int, compute lerp weight */
1139 /* not ideal with AVX (and no AVX2) */
1140 lp_build_ifloor_fract(coord_bld, coord0, &x0, &s_fpart);
1141 x1 = lp_build_add(ivec_bld, x0, ivec_bld->one);
1142 coord1 = lp_build_max_ext(coord_bld, coords[1], coord_bld->zero,
1143 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
1144 coord1 = lp_build_mul(coord_bld, coord1, flt_height_vec);
1145 coord1 = lp_build_sub(coord_bld, coord1, half);
1146 lp_build_ifloor_fract(coord_bld, coord1, &y0, &t_fpart);
1147 y1 = lp_build_add(ivec_bld, y0, ivec_bld->one);
1148
1149 fall_off[0] = lp_build_cmp(ivec_bld, PIPE_FUNC_LESS, x0, ivec_bld->zero);
1150 fall_off[1] = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, x1, length_minus_one);
1151 fall_off[2] = lp_build_cmp(ivec_bld, PIPE_FUNC_LESS, y0, ivec_bld->zero);
1152 fall_off[3] = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, y1, length_minus_one);
1153
1154 fall_off_x = lp_build_or(ivec_bld, fall_off[0], fall_off[1]);
1155 fall_off_y = lp_build_or(ivec_bld, fall_off[2], fall_off[3]);
1156 have_edge = lp_build_or(ivec_bld, fall_off_x, fall_off_y);
1157 have_edge = lp_build_any_true_range(ivec_bld, ivec_bld->type.length, have_edge);
1158
1159 /* needed for accurate corner filtering branch later, rely on 0 init */
1160 int1t = LLVMInt1TypeInContext(bld->gallivm->context);
1161 have_corners = lp_build_alloca(bld->gallivm, int1t, "have_corner");
1162
1163 for (texel_index = 0; texel_index < 4; texel_index++) {
1164 xs[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "xs");
1165 ys[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "ys");
1166 zs[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "zs");
1167 }
1168
1169 lp_build_if(&edge_if, bld->gallivm, have_edge);
1170
1171 have_corner = lp_build_and(ivec_bld, fall_off_x, fall_off_y);
1172 have_corner = lp_build_any_true_range(ivec_bld, ivec_bld->type.length, have_corner);
1173 LLVMBuildStore(builder, have_corner, have_corners);
1174
1175 /*
1176 * Need to feed clamped values here for cheap corner handling,
1177 * but only for y coord (as when falling off both edges we only
1178 * fall off the x one) - this should be sufficient.
1179 */
1180 y0_clamped = lp_build_max(ivec_bld, y0, ivec_bld->zero);
1181 y1_clamped = lp_build_min(ivec_bld, y1, length_minus_one);
1182
1183 /*
1184 * Get all possible new coords.
1185 */
1186 lp_build_cube_new_coords(ivec_bld, face,
1187 x0, x1, y0_clamped, y1_clamped,
1188 length_minus_one,
1189 new_faces, new_xcoords, new_ycoords);
1190
1191 /* handle fall off x-, x+ direction */
1192 /* determine new coords, face (not both fall_off vars can be true at same time) */
1193 x00 = lp_build_select(ivec_bld, fall_off[0], new_xcoords[0][0], x0);
1194 y00 = lp_build_select(ivec_bld, fall_off[0], new_ycoords[0][0], y0_clamped);
1195 x10 = lp_build_select(ivec_bld, fall_off[0], new_xcoords[0][1], x0);
1196 y10 = lp_build_select(ivec_bld, fall_off[0], new_ycoords[0][1], y1_clamped);
1197 x01 = lp_build_select(ivec_bld, fall_off[1], new_xcoords[1][0], x1);
1198 y01 = lp_build_select(ivec_bld, fall_off[1], new_ycoords[1][0], y0_clamped);
1199 x11 = lp_build_select(ivec_bld, fall_off[1], new_xcoords[1][1], x1);
1200 y11 = lp_build_select(ivec_bld, fall_off[1], new_ycoords[1][1], y1_clamped);
1201
1202 z00 = z10 = lp_build_select(ivec_bld, fall_off[0], new_faces[0], face);
1203 z01 = z11 = lp_build_select(ivec_bld, fall_off[1], new_faces[1], face);
1204
1205 /* handle fall off y-, y+ direction */
1206 /*
1207 * Cheap corner logic: just hack up things so a texel doesn't fall
1208 * off both sides (which means filter weights will be wrong but we'll only
1209 * use valid texels in the filter).
1210 * This means however (y) coords must additionally be clamped (see above).
1211 * This corner handling should be fully OpenGL (but not d3d10) compliant.
1212 */
1213 fall_off_ym_notxm = lp_build_andnot(ivec_bld, fall_off[2], fall_off[0]);
1214 fall_off_ym_notxp = lp_build_andnot(ivec_bld, fall_off[2], fall_off[1]);
1215 fall_off_yp_notxm = lp_build_andnot(ivec_bld, fall_off[3], fall_off[0]);
1216 fall_off_yp_notxp = lp_build_andnot(ivec_bld, fall_off[3], fall_off[1]);
1217
1218 x00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_xcoords[2][0], x00);
1219 y00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_ycoords[2][0], y00);
1220 x01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_xcoords[2][1], x01);
1221 y01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_ycoords[2][1], y01);
1222 x10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_xcoords[3][0], x10);
1223 y10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_ycoords[3][0], y10);
1224 x11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_xcoords[3][1], x11);
1225 y11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_ycoords[3][1], y11);
1226
1227 z00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_faces[2], z00);
1228 z01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_faces[2], z01);
1229 z10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_faces[3], z10);
1230 z11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_faces[3], z11);
1231
1232 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
1233 /* now can add cube layer to face (per sample) */
1234 z00 = lp_build_add(ivec_bld, z00, coords[3]);
1235 z01 = lp_build_add(ivec_bld, z01, coords[3]);
1236 z10 = lp_build_add(ivec_bld, z10, coords[3]);
1237 z11 = lp_build_add(ivec_bld, z11, coords[3]);
1238 }
1239
1240 LLVMBuildStore(builder, x00, xs[0]);
1241 LLVMBuildStore(builder, x01, xs[1]);
1242 LLVMBuildStore(builder, x10, xs[2]);
1243 LLVMBuildStore(builder, x11, xs[3]);
1244 LLVMBuildStore(builder, y00, ys[0]);
1245 LLVMBuildStore(builder, y01, ys[1]);
1246 LLVMBuildStore(builder, y10, ys[2]);
1247 LLVMBuildStore(builder, y11, ys[3]);
1248 LLVMBuildStore(builder, z00, zs[0]);
1249 LLVMBuildStore(builder, z01, zs[1]);
1250 LLVMBuildStore(builder, z10, zs[2]);
1251 LLVMBuildStore(builder, z11, zs[3]);
1252
1253 lp_build_else(&edge_if);
1254
1255 LLVMBuildStore(builder, x0, xs[0]);
1256 LLVMBuildStore(builder, x1, xs[1]);
1257 LLVMBuildStore(builder, x0, xs[2]);
1258 LLVMBuildStore(builder, x1, xs[3]);
1259 LLVMBuildStore(builder, y0, ys[0]);
1260 LLVMBuildStore(builder, y0, ys[1]);
1261 LLVMBuildStore(builder, y1, ys[2]);
1262 LLVMBuildStore(builder, y1, ys[3]);
1263 if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
1264 LLVMValueRef cube_layer = lp_build_add(ivec_bld, face, coords[3]);
1265 LLVMBuildStore(builder, cube_layer, zs[0]);
1266 LLVMBuildStore(builder, cube_layer, zs[1]);
1267 LLVMBuildStore(builder, cube_layer, zs[2]);
1268 LLVMBuildStore(builder, cube_layer, zs[3]);
1269 }
1270 else {
1271 LLVMBuildStore(builder, face, zs[0]);
1272 LLVMBuildStore(builder, face, zs[1]);
1273 LLVMBuildStore(builder, face, zs[2]);
1274 LLVMBuildStore(builder, face, zs[3]);
1275 }
1276
1277 lp_build_endif(&edge_if);
1278
1279 x00 = LLVMBuildLoad(builder, xs[0], "");
1280 x01 = LLVMBuildLoad(builder, xs[1], "");
1281 x10 = LLVMBuildLoad(builder, xs[2], "");
1282 x11 = LLVMBuildLoad(builder, xs[3], "");
1283 y00 = LLVMBuildLoad(builder, ys[0], "");
1284 y01 = LLVMBuildLoad(builder, ys[1], "");
1285 y10 = LLVMBuildLoad(builder, ys[2], "");
1286 y11 = LLVMBuildLoad(builder, ys[3], "");
1287 z00 = LLVMBuildLoad(builder, zs[0], "");
1288 z01 = LLVMBuildLoad(builder, zs[1], "");
1289 z10 = LLVMBuildLoad(builder, zs[2], "");
1290 z11 = LLVMBuildLoad(builder, zs[3], "");
1291 }
1292
1293 if (linear_mask) {
1294 /*
1295 * Whack filter weights into place. Whatever texel had more weight is
1296 * the one which should have been selected by nearest filtering hence
1297 * just use 100% weight for it.
1298 */
1299 struct lp_build_context *c_bld = &bld->coord_bld;
1300 LLVMValueRef w1_mask, w1_weight;
1301 LLVMValueRef half = lp_build_const_vec(bld->gallivm, c_bld->type, 0.5f);
1302
1303 w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, s_fpart, half);
1304 /* this select is really just a "and" */
1305 w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
1306 s_fpart = lp_build_select(c_bld, linear_mask, s_fpart, w1_weight);
1307 if (dims >= 2) {
1308 w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, t_fpart, half);
1309 w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
1310 t_fpart = lp_build_select(c_bld, linear_mask, t_fpart, w1_weight);
1311 if (dims == 3) {
1312 w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, r_fpart, half);
1313 w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
1314 r_fpart = lp_build_select(c_bld, linear_mask, r_fpart, w1_weight);
1315 }
1316 }
1317 }
1318
1319 /*
1320 * Get texture colors.
1321 */
1322 /* get x0/x1 texels */
1323 lp_build_sample_texel_soa(bld,
1324 width_vec, height_vec, depth_vec,
1325 x00, y00, z00,
1326 row_stride_vec, img_stride_vec,
1327 data_ptr, mipoffsets, neighbors[0][0]);
1328 lp_build_sample_texel_soa(bld,
1329 width_vec, height_vec, depth_vec,
1330 x01, y01, z01,
1331 row_stride_vec, img_stride_vec,
1332 data_ptr, mipoffsets, neighbors[0][1]);
1333
1334 if (dims == 1) {
1335 assert(!is_gather);
1336 if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
1337 /* Interpolate two samples from 1D image to produce one color */
1338 for (chan = 0; chan < 4; chan++) {
1339 colors_out[chan] = lp_build_lerp(texel_bld, s_fpart,
1340 neighbors[0][0][chan],
1341 neighbors[0][1][chan],
1342 0);
1343 }
1344 }
1345 else {
1346 LLVMValueRef cmpval0, cmpval1;
1347 cmpval0 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1348 cmpval1 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1349 /* simplified lerp, AND mask with weight and add */
1350 colors_out[0] = lp_build_masklerp(texel_bld, s_fpart,
1351 cmpval0, cmpval1);
1352 colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
1353 }
1354 }
1355 else {
1356 /* 2D/3D texture */
1357 struct lp_build_if_state corner_if;
1358 LLVMValueRef colors0[4], colorss[4];
1359
1360 /* get x0/x1 texels at y1 */
1361 lp_build_sample_texel_soa(bld,
1362 width_vec, height_vec, depth_vec,
1363 x10, y10, z10,
1364 row_stride_vec, img_stride_vec,
1365 data_ptr, mipoffsets, neighbors[1][0]);
1366 lp_build_sample_texel_soa(bld,
1367 width_vec, height_vec, depth_vec,
1368 x11, y11, z11,
1369 row_stride_vec, img_stride_vec,
1370 data_ptr, mipoffsets, neighbors[1][1]);
1371
1372 /*
1373 * To avoid having to duplicate linear_mask / fetch code use
1374 * another branch (with corner condition though edge would work
1375 * as well) here.
1376 */
1377 if (accurate_cube_corners) {
1378 LLVMValueRef c00, c01, c10, c11, c00f, c01f, c10f, c11f;
1379 LLVMValueRef have_corner, one_third;
1380
1381 colorss[0] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs0");
1382 colorss[1] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs1");
1383 colorss[2] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs2");
1384 colorss[3] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs3");
1385
1386 have_corner = LLVMBuildLoad(builder, have_corners, "");
1387
1388 lp_build_if(&corner_if, bld->gallivm, have_corner);
1389
1390 one_third = lp_build_const_vec(bld->gallivm, coord_bld->type,
1391 1.0f/3.0f);
1392
1393 /* find corner */
1394 c00 = lp_build_and(ivec_bld, fall_off[0], fall_off[2]);
1395 c00f = LLVMBuildBitCast(builder, c00, coord_bld->vec_type, "");
1396 c01 = lp_build_and(ivec_bld, fall_off[1], fall_off[2]);
1397 c01f = LLVMBuildBitCast(builder, c01, coord_bld->vec_type, "");
1398 c10 = lp_build_and(ivec_bld, fall_off[0], fall_off[3]);
1399 c10f = LLVMBuildBitCast(builder, c10, coord_bld->vec_type, "");
1400 c11 = lp_build_and(ivec_bld, fall_off[1], fall_off[3]);
1401 c11f = LLVMBuildBitCast(builder, c11, coord_bld->vec_type, "");
1402
1403 if (!is_gather) {
1404 /*
1405 * we can't use standard 2d lerp as we need per-element weight
1406 * in case of corners, so just calculate bilinear result as
1407 * w00*s00 + w01*s01 + w10*s10 + w11*s11.
1408 * (This is actually less work than using 2d lerp, 7 vs. 9
1409 * instructions, however calculating the weights needs another 6,
1410 * so actually probably not slower than 2d lerp only for 4 channels
1411 * as weights only need to be calculated once - of course fixing
1412 * the weights has additional cost.)
1413 */
1414 LLVMValueRef w00, w01, w10, w11, wx0, wy0, c_weight, tmp;
1415 wx0 = lp_build_sub(coord_bld, coord_bld->one, s_fpart);
1416 wy0 = lp_build_sub(coord_bld, coord_bld->one, t_fpart);
1417 w00 = lp_build_mul(coord_bld, wx0, wy0);
1418 w01 = lp_build_mul(coord_bld, s_fpart, wy0);
1419 w10 = lp_build_mul(coord_bld, wx0, t_fpart);
1420 w11 = lp_build_mul(coord_bld, s_fpart, t_fpart);
1421
1422 /* find corner weight */
1423 c_weight = lp_build_select(coord_bld, c00, w00, coord_bld->zero);
1424 c_weight = lp_build_select(coord_bld, c01, w01, c_weight);
1425 c_weight = lp_build_select(coord_bld, c10, w10, c_weight);
1426 c_weight = lp_build_select(coord_bld, c11, w11, c_weight);
1427
1428 /*
1429 * add 1/3 of the corner weight to the weight of the 3 other
1430 * samples and null out corner weight.
1431 */
1432 c_weight = lp_build_mul(coord_bld, c_weight, one_third);
1433 w00 = lp_build_add(coord_bld, w00, c_weight);
1434 w00 = lp_build_andnot(coord_bld, w00, c00f);
1435 w01 = lp_build_add(coord_bld, w01, c_weight);
1436 w01 = lp_build_andnot(coord_bld, w01, c01f);
1437 w10 = lp_build_add(coord_bld, w10, c_weight);
1438 w10 = lp_build_andnot(coord_bld, w10, c10f);
1439 w11 = lp_build_add(coord_bld, w11, c_weight);
1440 w11 = lp_build_andnot(coord_bld, w11, c11f);
1441
1442 if (bld->static_sampler_state->compare_mode ==
1443 PIPE_TEX_COMPARE_NONE) {
1444 for (chan = 0; chan < 4; chan++) {
1445 colors0[chan] = lp_build_mul(coord_bld, w00,
1446 neighbors[0][0][chan]);
1447 tmp = lp_build_mul(coord_bld, w01, neighbors[0][1][chan]);
1448 colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
1449 tmp = lp_build_mul(coord_bld, w10, neighbors[1][0][chan]);
1450 colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
1451 tmp = lp_build_mul(coord_bld, w11, neighbors[1][1][chan]);
1452 colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
1453 }
1454 }
1455 else {
1456 LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1457 cmpval00 = lp_build_sample_comparefunc(bld, coords[4],
1458 neighbors[0][0][0]);
1459 cmpval01 = lp_build_sample_comparefunc(bld, coords[4],
1460 neighbors[0][1][0]);
1461 cmpval10 = lp_build_sample_comparefunc(bld, coords[4],
1462 neighbors[1][0][0]);
1463 cmpval11 = lp_build_sample_comparefunc(bld, coords[4],
1464 neighbors[1][1][0]);
1465 /*
1466 * inputs to interpolation are just masks so just add
1467 * masked weights together
1468 */
1469 cmpval00 = LLVMBuildBitCast(builder, cmpval00,
1470 coord_bld->vec_type, "");
1471 cmpval01 = LLVMBuildBitCast(builder, cmpval01,
1472 coord_bld->vec_type, "");
1473 cmpval10 = LLVMBuildBitCast(builder, cmpval10,
1474 coord_bld->vec_type, "");
1475 cmpval11 = LLVMBuildBitCast(builder, cmpval11,
1476 coord_bld->vec_type, "");
1477 colors0[0] = lp_build_and(coord_bld, w00, cmpval00);
1478 tmp = lp_build_and(coord_bld, w01, cmpval01);
1479 colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
1480 tmp = lp_build_and(coord_bld, w10, cmpval10);
1481 colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
1482 tmp = lp_build_and(coord_bld, w11, cmpval11);
1483 colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
1484 colors0[1] = colors0[2] = colors0[3] = colors0[0];
1485 }
1486 }
1487 else {
1488 /*
1489 * We don't have any weights to adjust, so instead calculate
1490 * the fourth texel as simply the average of the other 3.
1491 * (This would work for non-gather too, however we'd have
1492 * a boatload more of the select stuff due to there being
1493 * 4 times as many colors as weights.)
1494 */
1495 LLVMValueRef col00, col01, col10, col11;
1496 LLVMValueRef colc, colc0, colc1;
1497 col10 = lp_build_swizzle_soa_channel(texel_bld,
1498 neighbors[1][0], chan_swiz);
1499 col11 = lp_build_swizzle_soa_channel(texel_bld,
1500 neighbors[1][1], chan_swiz);
1501 col01 = lp_build_swizzle_soa_channel(texel_bld,
1502 neighbors[0][1], chan_swiz);
1503 col00 = lp_build_swizzle_soa_channel(texel_bld,
1504 neighbors[0][0], chan_swiz);
1505
1506 /*
1507 * The spec says for comparison filtering, the comparison
1508 * must happen before synthesizing the new value.
1509 * This means all gathered values are always 0 or 1,
1510 * except for the non-existing texel, which can be 0,1/3,2/3,1...
1511 * Seems like we'd be allowed to just return 0 or 1 too, so we
1512 * could simplify and pass down the compare mask values to the
1513 * end (using int arithmetic/compare on the mask values to
1514 * construct the fourth texel) and only there convert to floats
1515 * but it's probably not worth it (it might be easier for the cpu
1516 * but not for the code)...
1517 */
1518 if (bld->static_sampler_state->compare_mode !=
1519 PIPE_TEX_COMPARE_NONE) {
1520 LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1521 cmpval00 = lp_build_sample_comparefunc(bld, coords[4], col00);
1522 cmpval01 = lp_build_sample_comparefunc(bld, coords[4], col01);
1523 cmpval10 = lp_build_sample_comparefunc(bld, coords[4], col10);
1524 cmpval11 = lp_build_sample_comparefunc(bld, coords[4], col11);
1525 col00 = lp_build_select(texel_bld, cmpval00,
1526 texel_bld->one, texel_bld->zero);
1527 col01 = lp_build_select(texel_bld, cmpval01,
1528 texel_bld->one, texel_bld->zero);
1529 col10 = lp_build_select(texel_bld, cmpval10,
1530 texel_bld->one, texel_bld->zero);
1531 col11 = lp_build_select(texel_bld, cmpval11,
1532 texel_bld->one, texel_bld->zero);
1533 }
1534
1535 /*
1536 * Null out corner color.
1537 */
1538 col00 = lp_build_andnot(coord_bld, col00, c00f);
1539 col01 = lp_build_andnot(coord_bld, col01, c01f);
1540 col10 = lp_build_andnot(coord_bld, col10, c10f);
1541 col11 = lp_build_andnot(coord_bld, col11, c11f);
1542
1543 /*
1544 * New corner texel color is all colors added / 3.
1545 */
1546 colc0 = lp_build_add(coord_bld, col00, col01);
1547 colc1 = lp_build_add(coord_bld, col10, col11);
1548 colc = lp_build_add(coord_bld, colc0, colc1);
1549 colc = lp_build_mul(coord_bld, one_third, colc);
1550
1551 /*
1552 * Replace the corner texel color with the new value.
1553 */
1554 col00 = lp_build_select(coord_bld, c00, colc, col00);
1555 col01 = lp_build_select(coord_bld, c01, colc, col01);
1556 col10 = lp_build_select(coord_bld, c10, colc, col10);
1557 col11 = lp_build_select(coord_bld, c11, colc, col11);
1558
1559 colors0[0] = col10;
1560 colors0[1] = col11;
1561 colors0[2] = col01;
1562 colors0[3] = col00;
1563 }
1564
1565 LLVMBuildStore(builder, colors0[0], colorss[0]);
1566 LLVMBuildStore(builder, colors0[1], colorss[1]);
1567 LLVMBuildStore(builder, colors0[2], colorss[2]);
1568 LLVMBuildStore(builder, colors0[3], colorss[3]);
1569
1570 lp_build_else(&corner_if);
1571 }
1572
1573 if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
1574 if (is_gather) {
1575 /*
1576 * Just assign the red channel (no component selection yet).
1577 * This is a bit hackish, we usually do the swizzle at the
1578 * end of sampling (much less values to swizzle), but this
1579 * obviously cannot work when using gather.
1580 */
1581 colors0[0] = lp_build_swizzle_soa_channel(texel_bld,
1582 neighbors[1][0],
1583 chan_swiz);
1584 colors0[1] = lp_build_swizzle_soa_channel(texel_bld,
1585 neighbors[1][1],
1586 chan_swiz);
1587 colors0[2] = lp_build_swizzle_soa_channel(texel_bld,
1588 neighbors[0][1],
1589 chan_swiz);
1590 colors0[3] = lp_build_swizzle_soa_channel(texel_bld,
1591 neighbors[0][0],
1592 chan_swiz);
1593 }
1594 else {
1595 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1596 for (chan = 0; chan < 4; chan++) {
1597 colors0[chan] = lp_build_lerp_2d(texel_bld,
1598 s_fpart, t_fpart,
1599 neighbors[0][0][chan],
1600 neighbors[0][1][chan],
1601 neighbors[1][0][chan],
1602 neighbors[1][1][chan],
1603 0);
1604 }
1605 }
1606 }
1607 else {
1608 LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1609 cmpval00 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1610 cmpval01 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1611 cmpval10 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][0][0]);
1612 cmpval11 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][1][0]);
1613
1614 if (is_gather) {
1615 /* more hacks for swizzling, should be X, ONE or ZERO... */
1616 colors0[0] = lp_build_select(texel_bld, cmpval10,
1617 texel_bld->one, texel_bld->zero);
1618 colors0[1] = lp_build_select(texel_bld, cmpval11,
1619 texel_bld->one, texel_bld->zero);
1620 colors0[2] = lp_build_select(texel_bld, cmpval01,
1621 texel_bld->one, texel_bld->zero);
1622 colors0[3] = lp_build_select(texel_bld, cmpval00,
1623 texel_bld->one, texel_bld->zero);
1624 }
1625 else {
1626 colors0[0] = lp_build_masklerp2d(texel_bld, s_fpart, t_fpart,
1627 cmpval00, cmpval01, cmpval10, cmpval11);
1628 colors0[1] = colors0[2] = colors0[3] = colors0[0];
1629 }
1630 }
1631
1632 if (accurate_cube_corners) {
1633 LLVMBuildStore(builder, colors0[0], colorss[0]);
1634 LLVMBuildStore(builder, colors0[1], colorss[1]);
1635 LLVMBuildStore(builder, colors0[2], colorss[2]);
1636 LLVMBuildStore(builder, colors0[3], colorss[3]);
1637
1638 lp_build_endif(&corner_if);
1639
1640 colors0[0] = LLVMBuildLoad(builder, colorss[0], "");
1641 colors0[1] = LLVMBuildLoad(builder, colorss[1], "");
1642 colors0[2] = LLVMBuildLoad(builder, colorss[2], "");
1643 colors0[3] = LLVMBuildLoad(builder, colorss[3], "");
1644 }
1645
1646 if (dims == 3) {
1647 LLVMValueRef neighbors1[2][2][4];
1648 LLVMValueRef colors1[4];
1649
1650 assert(!is_gather);
1651
1652 /* get x0/x1/y0/y1 texels at z1 */
1653 lp_build_sample_texel_soa(bld,
1654 width_vec, height_vec, depth_vec,
1655 x00, y00, z1,
1656 row_stride_vec, img_stride_vec,
1657 data_ptr, mipoffsets, neighbors1[0][0]);
1658 lp_build_sample_texel_soa(bld,
1659 width_vec, height_vec, depth_vec,
1660 x01, y01, z1,
1661 row_stride_vec, img_stride_vec,
1662 data_ptr, mipoffsets, neighbors1[0][1]);
1663 lp_build_sample_texel_soa(bld,
1664 width_vec, height_vec, depth_vec,
1665 x10, y10, z1,
1666 row_stride_vec, img_stride_vec,
1667 data_ptr, mipoffsets, neighbors1[1][0]);
1668 lp_build_sample_texel_soa(bld,
1669 width_vec, height_vec, depth_vec,
1670 x11, y11, z1,
1671 row_stride_vec, img_stride_vec,
1672 data_ptr, mipoffsets, neighbors1[1][1]);
1673
1674 if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
1675 /* Bilinear interpolate the four samples from the second Z slice */
1676 for (chan = 0; chan < 4; chan++) {
1677 colors1[chan] = lp_build_lerp_2d(texel_bld,
1678 s_fpart, t_fpart,
1679 neighbors1[0][0][chan],
1680 neighbors1[0][1][chan],
1681 neighbors1[1][0][chan],
1682 neighbors1[1][1][chan],
1683 0);
1684 }
1685 /* Linearly interpolate the two samples from the two 3D slices */
1686 for (chan = 0; chan < 4; chan++) {
1687 colors_out[chan] = lp_build_lerp(texel_bld,
1688 r_fpart,
1689 colors0[chan], colors1[chan],
1690 0);
1691 }
1692 }
1693 else {
1694 LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1695 cmpval00 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1696 cmpval01 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1697 cmpval10 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][0][0]);
1698 cmpval11 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][1][0]);
1699 colors1[0] = lp_build_masklerp2d(texel_bld, s_fpart, t_fpart,
1700 cmpval00, cmpval01, cmpval10, cmpval11);
1701 /* Linearly interpolate the two samples from the two 3D slices */
1702 colors_out[0] = lp_build_lerp(texel_bld,
1703 r_fpart,
1704 colors0[0], colors1[0],
1705 0);
1706 colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
1707 }
1708 }
1709 else {
1710 /* 2D tex */
1711 for (chan = 0; chan < 4; chan++) {
1712 colors_out[chan] = colors0[chan];
1713 }
1714 }
1715 }
1716 if (is_gather) {
1717 /*
1718 * For gather, we can't do our usual channel swizzling done later,
1719 * so do it here. It only really matters for 0/1 swizzles in case
1720 * of comparison filtering, since in this case the results would be
1721 * wrong, without comparison it should all work out alright but it
1722 * can't hurt to do that here, since it will instantly drop all
1723 * calculations above, though it's a rather stupid idea to do
1724 * gather on a channel which will always return 0 or 1 in any case...
1725 */
1726 if (chan_swiz == PIPE_SWIZZLE_1) {
1727 for (chan = 0; chan < 4; chan++) {
1728 colors_out[chan] = texel_bld->one;
1729 }
1730 } else if (chan_swiz == PIPE_SWIZZLE_0) {
1731 for (chan = 0; chan < 4; chan++) {
1732 colors_out[chan] = texel_bld->zero;
1733 }
1734 }
1735 }
1736 }
1737
1738
1739 /**
1740 * Sample the texture/mipmap using given image filter and mip filter.
1741 * ilevel0 and ilevel1 indicate the two mipmap levels to sample
1742 * from (vectors or scalars).
1743 * If we're using nearest miplevel sampling the '1' values will be null/unused.
1744 */
1745 static void
lp_build_sample_mipmap(struct lp_build_sample_context * bld,unsigned img_filter,unsigned mip_filter,boolean is_gather,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef ilevel0,LLVMValueRef ilevel1,LLVMValueRef lod_fpart,LLVMValueRef * colors_out)1746 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
1747 unsigned img_filter,
1748 unsigned mip_filter,
1749 boolean is_gather,
1750 const LLVMValueRef *coords,
1751 const LLVMValueRef *offsets,
1752 LLVMValueRef ilevel0,
1753 LLVMValueRef ilevel1,
1754 LLVMValueRef lod_fpart,
1755 LLVMValueRef *colors_out)
1756 {
1757 LLVMBuilderRef builder = bld->gallivm->builder;
1758 LLVMValueRef size0 = NULL;
1759 LLVMValueRef size1 = NULL;
1760 LLVMValueRef row_stride0_vec = NULL;
1761 LLVMValueRef row_stride1_vec = NULL;
1762 LLVMValueRef img_stride0_vec = NULL;
1763 LLVMValueRef img_stride1_vec = NULL;
1764 LLVMValueRef data_ptr0 = NULL;
1765 LLVMValueRef data_ptr1 = NULL;
1766 LLVMValueRef mipoff0 = NULL;
1767 LLVMValueRef mipoff1 = NULL;
1768 LLVMValueRef colors0[4], colors1[4];
1769 unsigned chan;
1770
1771 /* sample the first mipmap level */
1772 lp_build_mipmap_level_sizes(bld, ilevel0,
1773 &size0,
1774 &row_stride0_vec, &img_stride0_vec);
1775 if (bld->num_mips == 1) {
1776 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
1777 }
1778 else {
1779 /* This path should work for num_lods 1 too but slightly less efficient */
1780 data_ptr0 = bld->base_ptr;
1781 mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
1782 }
1783 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1784 lp_build_sample_image_nearest(bld, size0,
1785 row_stride0_vec, img_stride0_vec,
1786 data_ptr0, mipoff0, coords, offsets,
1787 colors0);
1788 }
1789 else {
1790 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1791 lp_build_sample_image_linear(bld, is_gather, size0, NULL,
1792 row_stride0_vec, img_stride0_vec,
1793 data_ptr0, mipoff0, coords, offsets,
1794 colors0);
1795 }
1796
1797 /* Store the first level's colors in the output variables */
1798 for (chan = 0; chan < 4; chan++) {
1799 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1800 }
1801
1802 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1803 struct lp_build_if_state if_ctx;
1804 LLVMValueRef need_lerp;
1805
1806 /* need_lerp = lod_fpart > 0 */
1807 if (bld->num_lods == 1) {
1808 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
1809 lod_fpart, bld->lodf_bld.zero,
1810 "need_lerp");
1811 }
1812 else {
1813 /*
1814 * We'll do mip filtering if any of the quads (or individual
1815 * pixel in case of per-pixel lod) need it.
1816 * It might be better to split the vectors here and only fetch/filter
1817 * quads which need it (if there's one lod per quad).
1818 */
1819 need_lerp = lp_build_compare(bld->gallivm, bld->lodf_bld.type,
1820 PIPE_FUNC_GREATER,
1821 lod_fpart, bld->lodf_bld.zero);
1822 need_lerp = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods, need_lerp);
1823 lp_build_name(need_lerp, "need_lerp");
1824 }
1825
1826 lp_build_if(&if_ctx, bld->gallivm, need_lerp);
1827 {
1828 /*
1829 * We unfortunately need to clamp lod_fpart here since we can get
1830 * negative values which would screw up filtering if not all
1831 * lod_fpart values have same sign.
1832 */
1833 lod_fpart = lp_build_max(&bld->lodf_bld, lod_fpart,
1834 bld->lodf_bld.zero);
1835 /* sample the second mipmap level */
1836 lp_build_mipmap_level_sizes(bld, ilevel1,
1837 &size1,
1838 &row_stride1_vec, &img_stride1_vec);
1839 if (bld->num_mips == 1) {
1840 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
1841 }
1842 else {
1843 data_ptr1 = bld->base_ptr;
1844 mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
1845 }
1846 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1847 lp_build_sample_image_nearest(bld, size1,
1848 row_stride1_vec, img_stride1_vec,
1849 data_ptr1, mipoff1, coords, offsets,
1850 colors1);
1851 }
1852 else {
1853 lp_build_sample_image_linear(bld, FALSE, size1, NULL,
1854 row_stride1_vec, img_stride1_vec,
1855 data_ptr1, mipoff1, coords, offsets,
1856 colors1);
1857 }
1858
1859 /* interpolate samples from the two mipmap levels */
1860
1861 if (bld->num_lods != bld->coord_type.length)
1862 lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
1863 bld->lodf_bld.type,
1864 bld->texel_bld.type,
1865 lod_fpart);
1866
1867 for (chan = 0; chan < 4; chan++) {
1868 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1869 colors0[chan], colors1[chan],
1870 0);
1871 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1872 }
1873 }
1874 lp_build_endif(&if_ctx);
1875 }
1876 }
1877
1878
1879 /**
1880 * Sample the texture/mipmap using given mip filter, and using
1881 * both nearest and linear filtering at the same time depending
1882 * on linear_mask.
1883 * lod can be per quad but linear_mask is always per pixel.
1884 * ilevel0 and ilevel1 indicate the two mipmap levels to sample
1885 * from (vectors or scalars).
1886 * If we're using nearest miplevel sampling the '1' values will be null/unused.
1887 */
1888 static void
lp_build_sample_mipmap_both(struct lp_build_sample_context * bld,LLVMValueRef linear_mask,unsigned mip_filter,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef ilevel0,LLVMValueRef ilevel1,LLVMValueRef lod_fpart,LLVMValueRef lod_positive,LLVMValueRef * colors_out)1889 lp_build_sample_mipmap_both(struct lp_build_sample_context *bld,
1890 LLVMValueRef linear_mask,
1891 unsigned mip_filter,
1892 const LLVMValueRef *coords,
1893 const LLVMValueRef *offsets,
1894 LLVMValueRef ilevel0,
1895 LLVMValueRef ilevel1,
1896 LLVMValueRef lod_fpart,
1897 LLVMValueRef lod_positive,
1898 LLVMValueRef *colors_out)
1899 {
1900 LLVMBuilderRef builder = bld->gallivm->builder;
1901 LLVMValueRef size0 = NULL;
1902 LLVMValueRef size1 = NULL;
1903 LLVMValueRef row_stride0_vec = NULL;
1904 LLVMValueRef row_stride1_vec = NULL;
1905 LLVMValueRef img_stride0_vec = NULL;
1906 LLVMValueRef img_stride1_vec = NULL;
1907 LLVMValueRef data_ptr0 = NULL;
1908 LLVMValueRef data_ptr1 = NULL;
1909 LLVMValueRef mipoff0 = NULL;
1910 LLVMValueRef mipoff1 = NULL;
1911 LLVMValueRef colors0[4], colors1[4];
1912 unsigned chan;
1913
1914 /* sample the first mipmap level */
1915 lp_build_mipmap_level_sizes(bld, ilevel0,
1916 &size0,
1917 &row_stride0_vec, &img_stride0_vec);
1918 if (bld->num_mips == 1) {
1919 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
1920 }
1921 else {
1922 /* This path should work for num_lods 1 too but slightly less efficient */
1923 data_ptr0 = bld->base_ptr;
1924 mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
1925 }
1926
1927 lp_build_sample_image_linear(bld, FALSE, size0, linear_mask,
1928 row_stride0_vec, img_stride0_vec,
1929 data_ptr0, mipoff0, coords, offsets,
1930 colors0);
1931
1932 /* Store the first level's colors in the output variables */
1933 for (chan = 0; chan < 4; chan++) {
1934 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1935 }
1936
1937 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1938 struct lp_build_if_state if_ctx;
1939 LLVMValueRef need_lerp;
1940
1941 /*
1942 * We'll do mip filtering if any of the quads (or individual
1943 * pixel in case of per-pixel lod) need it.
1944 * Note using lod_positive here not lod_fpart since it may be the same
1945 * condition as that used in the outer "if" in the caller hence llvm
1946 * should be able to merge the branches in this case.
1947 */
1948 need_lerp = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods, lod_positive);
1949 lp_build_name(need_lerp, "need_lerp");
1950
1951 lp_build_if(&if_ctx, bld->gallivm, need_lerp);
1952 {
1953 /*
1954 * We unfortunately need to clamp lod_fpart here since we can get
1955 * negative values which would screw up filtering if not all
1956 * lod_fpart values have same sign.
1957 */
1958 lod_fpart = lp_build_max(&bld->lodf_bld, lod_fpart,
1959 bld->lodf_bld.zero);
1960 /* sample the second mipmap level */
1961 lp_build_mipmap_level_sizes(bld, ilevel1,
1962 &size1,
1963 &row_stride1_vec, &img_stride1_vec);
1964 if (bld->num_mips == 1) {
1965 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
1966 }
1967 else {
1968 data_ptr1 = bld->base_ptr;
1969 mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
1970 }
1971
1972 lp_build_sample_image_linear(bld, FALSE, size1, linear_mask,
1973 row_stride1_vec, img_stride1_vec,
1974 data_ptr1, mipoff1, coords, offsets,
1975 colors1);
1976
1977 /* interpolate samples from the two mipmap levels */
1978
1979 if (bld->num_lods != bld->coord_type.length)
1980 lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
1981 bld->lodf_bld.type,
1982 bld->texel_bld.type,
1983 lod_fpart);
1984
1985 for (chan = 0; chan < 4; chan++) {
1986 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1987 colors0[chan], colors1[chan],
1988 0);
1989 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1990 }
1991 }
1992 lp_build_endif(&if_ctx);
1993 }
1994 }
1995
1996
1997 /**
1998 * Build (per-coord) layer value.
1999 * Either clamp layer to valid values or fill in optional out_of_bounds
2000 * value and just return value unclamped.
2001 */
2002 static LLVMValueRef
lp_build_layer_coord(struct lp_build_sample_context * bld,unsigned texture_unit,boolean is_cube_array,LLVMValueRef layer,LLVMValueRef * out_of_bounds)2003 lp_build_layer_coord(struct lp_build_sample_context *bld,
2004 unsigned texture_unit,
2005 boolean is_cube_array,
2006 LLVMValueRef layer,
2007 LLVMValueRef *out_of_bounds)
2008 {
2009 LLVMValueRef num_layers;
2010 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
2011
2012 num_layers = bld->dynamic_state->depth(bld->dynamic_state, bld->gallivm,
2013 bld->context_ptr, texture_unit);
2014
2015 if (out_of_bounds) {
2016 LLVMValueRef out1, out;
2017 assert(!is_cube_array);
2018 num_layers = lp_build_broadcast_scalar(int_coord_bld, num_layers);
2019 out = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, layer, int_coord_bld->zero);
2020 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, layer, num_layers);
2021 *out_of_bounds = lp_build_or(int_coord_bld, out, out1);
2022 return layer;
2023 }
2024 else {
2025 LLVMValueRef maxlayer;
2026 LLVMValueRef s = is_cube_array ? lp_build_const_int32(bld->gallivm, 6) :
2027 bld->int_bld.one;
2028 maxlayer = lp_build_sub(&bld->int_bld, num_layers, s);
2029 maxlayer = lp_build_broadcast_scalar(int_coord_bld, maxlayer);
2030 return lp_build_clamp(int_coord_bld, layer, int_coord_bld->zero, maxlayer);
2031 }
2032 }
2033
2034
2035 /**
2036 * Calculate cube face, lod, mip levels.
2037 */
2038 static void
lp_build_sample_common(struct lp_build_sample_context * bld,boolean is_lodq,unsigned texture_index,unsigned sampler_index,LLVMValueRef * coords,const struct lp_derivatives * derivs,LLVMValueRef lod_bias,LLVMValueRef explicit_lod,LLVMValueRef * lod_pos_or_zero,LLVMValueRef * lod,LLVMValueRef * lod_fpart,LLVMValueRef * ilevel0,LLVMValueRef * ilevel1)2039 lp_build_sample_common(struct lp_build_sample_context *bld,
2040 boolean is_lodq,
2041 unsigned texture_index,
2042 unsigned sampler_index,
2043 LLVMValueRef *coords,
2044 const struct lp_derivatives *derivs, /* optional */
2045 LLVMValueRef lod_bias, /* optional */
2046 LLVMValueRef explicit_lod, /* optional */
2047 LLVMValueRef *lod_pos_or_zero,
2048 LLVMValueRef *lod,
2049 LLVMValueRef *lod_fpart,
2050 LLVMValueRef *ilevel0,
2051 LLVMValueRef *ilevel1)
2052 {
2053 const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
2054 const unsigned min_filter = bld->static_sampler_state->min_img_filter;
2055 const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
2056 const unsigned target = bld->static_texture_state->target;
2057 LLVMValueRef first_level, cube_rho = NULL;
2058 LLVMValueRef lod_ipart = NULL;
2059 struct lp_derivatives cube_derivs;
2060
2061 /*
2062 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
2063 mip_filter, min_filter, mag_filter);
2064 */
2065
2066 /*
2067 * Choose cube face, recompute texcoords for the chosen face and
2068 * compute rho here too (as it requires transform of derivatives).
2069 */
2070 if (target == PIPE_TEXTURE_CUBE || target == PIPE_TEXTURE_CUBE_ARRAY) {
2071 boolean need_derivs;
2072 need_derivs = ((min_filter != mag_filter ||
2073 mip_filter != PIPE_TEX_MIPFILTER_NONE) &&
2074 !bld->static_sampler_state->min_max_lod_equal &&
2075 !explicit_lod);
2076 lp_build_cube_lookup(bld, coords, derivs, &cube_rho, &cube_derivs, need_derivs);
2077 derivs = &cube_derivs;
2078 if (target == PIPE_TEXTURE_CUBE_ARRAY) {
2079 /* calculate cube layer coord now */
2080 LLVMValueRef layer = lp_build_iround(&bld->coord_bld, coords[3]);
2081 LLVMValueRef six = lp_build_const_int_vec(bld->gallivm, bld->int_coord_type, 6);
2082 layer = lp_build_mul(&bld->int_coord_bld, layer, six);
2083 coords[3] = lp_build_layer_coord(bld, texture_index, TRUE, layer, NULL);
2084 /* because of seamless filtering can't add it to face (coords[2]) here. */
2085 }
2086 }
2087 else if (target == PIPE_TEXTURE_1D_ARRAY ||
2088 target == PIPE_TEXTURE_2D_ARRAY) {
2089 coords[2] = lp_build_iround(&bld->coord_bld, coords[2]);
2090 coords[2] = lp_build_layer_coord(bld, texture_index, FALSE, coords[2], NULL);
2091 }
2092
2093 if (bld->static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) {
2094 /*
2095 * Clamp p coords to [0,1] for fixed function depth texture format here.
2096 * Technically this is not entirely correct for unorm depth as the ref value
2097 * should be converted to the depth format (quantization!) and comparison
2098 * then done in texture format. This would actually help performance (since
2099 * only need to do it once and could save the per-sample conversion of texels
2100 * to floats instead), but it would need more messy code (would need to push
2101 * at least some bits down to actual fetch so conversion could be skipped,
2102 * and would have ugly interaction with border color, would need to convert
2103 * border color to that format too or do some other tricks to make it work).
2104 */
2105 const struct util_format_description *format_desc = bld->format_desc;
2106 unsigned chan_type;
2107 /* not entirely sure we couldn't end up with non-valid swizzle here */
2108 chan_type = format_desc->swizzle[0] <= PIPE_SWIZZLE_W ?
2109 format_desc->channel[format_desc->swizzle[0]].type :
2110 UTIL_FORMAT_TYPE_FLOAT;
2111 if (chan_type != UTIL_FORMAT_TYPE_FLOAT) {
2112 coords[4] = lp_build_clamp(&bld->coord_bld, coords[4],
2113 bld->coord_bld.zero, bld->coord_bld.one);
2114 }
2115 }
2116
2117 /*
2118 * Compute the level of detail (float).
2119 */
2120 if (min_filter != mag_filter ||
2121 mip_filter != PIPE_TEX_MIPFILTER_NONE || is_lodq) {
2122 /* Need to compute lod either to choose mipmap levels or to
2123 * distinguish between minification/magnification with one mipmap level.
2124 */
2125 lp_build_lod_selector(bld, is_lodq, texture_index, sampler_index,
2126 coords[0], coords[1], coords[2], cube_rho,
2127 derivs, lod_bias, explicit_lod,
2128 mip_filter, lod,
2129 &lod_ipart, lod_fpart, lod_pos_or_zero);
2130 if (is_lodq) {
2131 LLVMValueRef last_level;
2132 last_level = bld->dynamic_state->last_level(bld->dynamic_state,
2133 bld->gallivm,
2134 bld->context_ptr,
2135 texture_index);
2136 first_level = bld->dynamic_state->first_level(bld->dynamic_state,
2137 bld->gallivm,
2138 bld->context_ptr,
2139 texture_index);
2140 last_level = lp_build_sub(&bld->int_bld, last_level, first_level);
2141 last_level = lp_build_int_to_float(&bld->float_bld, last_level);
2142 last_level = lp_build_broadcast_scalar(&bld->lodf_bld, last_level);
2143
2144 switch (mip_filter) {
2145 case PIPE_TEX_MIPFILTER_NONE:
2146 *lod_fpart = bld->lodf_bld.zero;
2147 break;
2148 case PIPE_TEX_MIPFILTER_NEAREST:
2149 *lod_fpart = lp_build_round(&bld->lodf_bld, *lod_fpart);
2150 /* fallthrough */
2151 case PIPE_TEX_MIPFILTER_LINEAR:
2152 *lod_fpart = lp_build_clamp(&bld->lodf_bld, *lod_fpart,
2153 bld->lodf_bld.zero, last_level);
2154 break;
2155 }
2156 return;
2157 }
2158
2159 } else {
2160 lod_ipart = bld->lodi_bld.zero;
2161 *lod_pos_or_zero = bld->lodi_bld.zero;
2162 }
2163
2164 if (bld->num_lods != bld->num_mips) {
2165 /* only makes sense if there's just a single mip level */
2166 assert(bld->num_mips == 1);
2167 lod_ipart = lp_build_extract_range(bld->gallivm, lod_ipart, 0, 1);
2168 }
2169
2170 /*
2171 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
2172 */
2173 switch (mip_filter) {
2174 default:
2175 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
2176 /* fall-through */
2177 case PIPE_TEX_MIPFILTER_NONE:
2178 /* always use mip level 0 */
2179 first_level = bld->dynamic_state->first_level(bld->dynamic_state,
2180 bld->gallivm, bld->context_ptr,
2181 texture_index);
2182 first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level);
2183 *ilevel0 = first_level;
2184 break;
2185 case PIPE_TEX_MIPFILTER_NEAREST:
2186 assert(lod_ipart);
2187 lp_build_nearest_mip_level(bld, texture_index, lod_ipart, ilevel0, NULL);
2188 break;
2189 case PIPE_TEX_MIPFILTER_LINEAR:
2190 assert(lod_ipart);
2191 assert(*lod_fpart);
2192 lp_build_linear_mip_levels(bld, texture_index,
2193 lod_ipart, lod_fpart,
2194 ilevel0, ilevel1);
2195 break;
2196 }
2197 }
2198
2199 static void
lp_build_clamp_border_color(struct lp_build_sample_context * bld,unsigned sampler_unit)2200 lp_build_clamp_border_color(struct lp_build_sample_context *bld,
2201 unsigned sampler_unit)
2202 {
2203 struct gallivm_state *gallivm = bld->gallivm;
2204 LLVMBuilderRef builder = gallivm->builder;
2205 LLVMValueRef border_color_ptr =
2206 bld->dynamic_state->border_color(bld->dynamic_state, gallivm,
2207 bld->context_ptr, sampler_unit);
2208 LLVMValueRef border_color;
2209 const struct util_format_description *format_desc = bld->format_desc;
2210 struct lp_type vec4_type = bld->texel_type;
2211 struct lp_build_context vec4_bld;
2212 LLVMValueRef min_clamp = NULL;
2213 LLVMValueRef max_clamp = NULL;
2214
2215 /*
2216 * For normalized format need to clamp border color (technically
2217 * probably should also quantize the data). Really sucks doing this
2218 * here but can't avoid at least for now since this is part of
2219 * sampler state and texture format is part of sampler_view state.
2220 * GL expects also expects clamping for uint/sint formats too so
2221 * do that as well (d3d10 can't end up here with uint/sint since it
2222 * only supports them with ld).
2223 */
2224 vec4_type.length = 4;
2225 lp_build_context_init(&vec4_bld, gallivm, vec4_type);
2226
2227 /*
2228 * Vectorized clamping of border color. Loading is a bit of a hack since
2229 * we just cast the pointer to float array to pointer to vec4
2230 * (int or float).
2231 */
2232 border_color_ptr = lp_build_array_get_ptr(gallivm, border_color_ptr,
2233 lp_build_const_int32(gallivm, 0));
2234 border_color_ptr = LLVMBuildBitCast(builder, border_color_ptr,
2235 LLVMPointerType(vec4_bld.vec_type, 0), "");
2236 border_color = LLVMBuildLoad(builder, border_color_ptr, "");
2237 /* we don't have aligned type in the dynamic state unfortunately */
2238 LLVMSetAlignment(border_color, 4);
2239
2240 /*
2241 * Instead of having some incredibly complex logic which will try to figure out
2242 * clamping necessary for each channel, simply use the first channel, and treat
2243 * mixed signed/unsigned normalized formats specially.
2244 * (Mixed non-normalized, which wouldn't work at all here, do not exist for a
2245 * good reason.)
2246 */
2247 if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN) {
2248 int chan;
2249 /* d/s needs special handling because both present means just sampling depth */
2250 if (util_format_is_depth_and_stencil(format_desc->format)) {
2251 chan = format_desc->swizzle[0];
2252 }
2253 else {
2254 chan = util_format_get_first_non_void_channel(format_desc->format);
2255 }
2256 if (chan >= 0 && chan <= PIPE_SWIZZLE_W) {
2257 unsigned chan_type = format_desc->channel[chan].type;
2258 unsigned chan_norm = format_desc->channel[chan].normalized;
2259 unsigned chan_pure = format_desc->channel[chan].pure_integer;
2260 if (chan_type == UTIL_FORMAT_TYPE_SIGNED) {
2261 if (chan_norm) {
2262 min_clamp = lp_build_const_vec(gallivm, vec4_type, -1.0F);
2263 max_clamp = vec4_bld.one;
2264 }
2265 else if (chan_pure) {
2266 /*
2267 * Border color was stored as int, hence need min/max clamp
2268 * only if chan has less than 32 bits..
2269 */
2270 unsigned chan_size = format_desc->channel[chan].size;
2271 if (chan_size < 32) {
2272 min_clamp = lp_build_const_int_vec(gallivm, vec4_type,
2273 0 - (1 << (chan_size - 1)));
2274 max_clamp = lp_build_const_int_vec(gallivm, vec4_type,
2275 (1 << (chan_size - 1)) - 1);
2276 }
2277 }
2278 /* TODO: no idea about non-pure, non-normalized! */
2279 }
2280 else if (chan_type == UTIL_FORMAT_TYPE_UNSIGNED) {
2281 if (chan_norm) {
2282 min_clamp = vec4_bld.zero;
2283 max_clamp = vec4_bld.one;
2284 }
2285 /*
2286 * Need a ugly hack here, because we don't have Z32_FLOAT_X8X24
2287 * we use Z32_FLOAT_S8X24 to imply sampling depth component
2288 * and ignoring stencil, which will blow up here if we try to
2289 * do a uint clamp in a float texel build...
2290 * And even if we had that format, mesa st also thinks using z24s8
2291 * means depth sampling ignoring stencil.
2292 */
2293 else if (chan_pure) {
2294 /*
2295 * Border color was stored as uint, hence never need min
2296 * clamp, and only need max clamp if chan has less than 32 bits.
2297 */
2298 unsigned chan_size = format_desc->channel[chan].size;
2299 if (chan_size < 32) {
2300 max_clamp = lp_build_const_int_vec(gallivm, vec4_type,
2301 (1 << chan_size) - 1);
2302 }
2303 /* TODO: no idea about non-pure, non-normalized! */
2304 }
2305 }
2306 else if (chan_type == UTIL_FORMAT_TYPE_FIXED) {
2307 /* TODO: I have no idea what clamp this would need if any! */
2308 }
2309 }
2310 /* mixed plain formats (or different pure size) */
2311 switch (format_desc->format) {
2312 case PIPE_FORMAT_B10G10R10A2_UINT:
2313 case PIPE_FORMAT_R10G10B10A2_UINT:
2314 {
2315 unsigned max10 = (1 << 10) - 1;
2316 max_clamp = lp_build_const_aos(gallivm, vec4_type, max10, max10,
2317 max10, (1 << 2) - 1, NULL);
2318 }
2319 break;
2320 case PIPE_FORMAT_R10SG10SB10SA2U_NORM:
2321 min_clamp = lp_build_const_aos(gallivm, vec4_type, -1.0F, -1.0F,
2322 -1.0F, 0.0F, NULL);
2323 max_clamp = vec4_bld.one;
2324 break;
2325 case PIPE_FORMAT_R8SG8SB8UX8U_NORM:
2326 case PIPE_FORMAT_R5SG5SB6U_NORM:
2327 min_clamp = lp_build_const_aos(gallivm, vec4_type, -1.0F, -1.0F,
2328 0.0F, 0.0F, NULL);
2329 max_clamp = vec4_bld.one;
2330 break;
2331 default:
2332 break;
2333 }
2334 }
2335 else {
2336 /* cannot figure this out from format description */
2337 if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
2338 /* s3tc formats are always unorm */
2339 min_clamp = vec4_bld.zero;
2340 max_clamp = vec4_bld.one;
2341 }
2342 else if (format_desc->layout == UTIL_FORMAT_LAYOUT_RGTC ||
2343 format_desc->layout == UTIL_FORMAT_LAYOUT_ETC) {
2344 switch (format_desc->format) {
2345 case PIPE_FORMAT_RGTC1_UNORM:
2346 case PIPE_FORMAT_RGTC2_UNORM:
2347 case PIPE_FORMAT_LATC1_UNORM:
2348 case PIPE_FORMAT_LATC2_UNORM:
2349 case PIPE_FORMAT_ETC1_RGB8:
2350 min_clamp = vec4_bld.zero;
2351 max_clamp = vec4_bld.one;
2352 break;
2353 case PIPE_FORMAT_RGTC1_SNORM:
2354 case PIPE_FORMAT_RGTC2_SNORM:
2355 case PIPE_FORMAT_LATC1_SNORM:
2356 case PIPE_FORMAT_LATC2_SNORM:
2357 min_clamp = lp_build_const_vec(gallivm, vec4_type, -1.0F);
2358 max_clamp = vec4_bld.one;
2359 break;
2360 default:
2361 assert(0);
2362 break;
2363 }
2364 }
2365 /*
2366 * all others from subsampled/other group, though we don't care
2367 * about yuv (and should not have any from zs here)
2368 */
2369 else if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_YUV){
2370 switch (format_desc->format) {
2371 case PIPE_FORMAT_R8G8_B8G8_UNORM:
2372 case PIPE_FORMAT_G8R8_G8B8_UNORM:
2373 case PIPE_FORMAT_G8R8_B8R8_UNORM:
2374 case PIPE_FORMAT_R8G8_R8B8_UNORM:
2375 case PIPE_FORMAT_R1_UNORM: /* doesn't make sense but ah well */
2376 min_clamp = vec4_bld.zero;
2377 max_clamp = vec4_bld.one;
2378 break;
2379 case PIPE_FORMAT_R8G8Bx_SNORM:
2380 min_clamp = lp_build_const_vec(gallivm, vec4_type, -1.0F);
2381 max_clamp = vec4_bld.one;
2382 break;
2383 /*
2384 * Note smallfloat formats usually don't need clamping
2385 * (they still have infinite range) however this is not
2386 * true for r11g11b10 and r9g9b9e5, which can't represent
2387 * negative numbers (and additionally r9g9b9e5 can't represent
2388 * very large numbers). d3d10 seems happy without clamping in
2389 * this case, but gl spec is pretty clear: "for floating
2390 * point and integer formats, border values are clamped to
2391 * the representable range of the format" so do that here.
2392 */
2393 case PIPE_FORMAT_R11G11B10_FLOAT:
2394 min_clamp = vec4_bld.zero;
2395 break;
2396 case PIPE_FORMAT_R9G9B9E5_FLOAT:
2397 min_clamp = vec4_bld.zero;
2398 max_clamp = lp_build_const_vec(gallivm, vec4_type, MAX_RGB9E5);
2399 break;
2400 default:
2401 assert(0);
2402 break;
2403 }
2404 }
2405 }
2406
2407 if (min_clamp) {
2408 border_color = lp_build_max(&vec4_bld, border_color, min_clamp);
2409 }
2410 if (max_clamp) {
2411 border_color = lp_build_min(&vec4_bld, border_color, max_clamp);
2412 }
2413
2414 bld->border_color_clamped = border_color;
2415 }
2416
2417
2418 /**
2419 * General texture sampling codegen.
2420 * This function handles texture sampling for all texture targets (1D,
2421 * 2D, 3D, cube) and all filtering modes.
2422 */
2423 static void
lp_build_sample_general(struct lp_build_sample_context * bld,unsigned sampler_unit,boolean is_gather,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef lod_positive,LLVMValueRef lod_fpart,LLVMValueRef ilevel0,LLVMValueRef ilevel1,LLVMValueRef * colors_out)2424 lp_build_sample_general(struct lp_build_sample_context *bld,
2425 unsigned sampler_unit,
2426 boolean is_gather,
2427 const LLVMValueRef *coords,
2428 const LLVMValueRef *offsets,
2429 LLVMValueRef lod_positive,
2430 LLVMValueRef lod_fpart,
2431 LLVMValueRef ilevel0,
2432 LLVMValueRef ilevel1,
2433 LLVMValueRef *colors_out)
2434 {
2435 LLVMBuilderRef builder = bld->gallivm->builder;
2436 const struct lp_static_sampler_state *sampler_state = bld->static_sampler_state;
2437 const unsigned mip_filter = sampler_state->min_mip_filter;
2438 const unsigned min_filter = sampler_state->min_img_filter;
2439 const unsigned mag_filter = sampler_state->mag_img_filter;
2440 LLVMValueRef texels[4];
2441 unsigned chan;
2442
2443 /* if we need border color, (potentially) clamp it now */
2444 if (lp_sampler_wrap_mode_uses_border_color(sampler_state->wrap_s,
2445 min_filter,
2446 mag_filter) ||
2447 (bld->dims > 1 &&
2448 lp_sampler_wrap_mode_uses_border_color(sampler_state->wrap_t,
2449 min_filter,
2450 mag_filter)) ||
2451 (bld->dims > 2 &&
2452 lp_sampler_wrap_mode_uses_border_color(sampler_state->wrap_r,
2453 min_filter,
2454 mag_filter))) {
2455 lp_build_clamp_border_color(bld, sampler_unit);
2456 }
2457
2458
2459 /*
2460 * Get/interpolate texture colors.
2461 */
2462
2463 for (chan = 0; chan < 4; ++chan) {
2464 texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
2465 lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]);
2466 }
2467
2468 if (min_filter == mag_filter) {
2469 /* no need to distinguish between minification and magnification */
2470 lp_build_sample_mipmap(bld, min_filter, mip_filter,
2471 is_gather,
2472 coords, offsets,
2473 ilevel0, ilevel1, lod_fpart,
2474 texels);
2475 }
2476 else {
2477 /*
2478 * Could also get rid of the if-logic and always use mipmap_both, both
2479 * for the single lod and multi-lod case if nothing really uses this.
2480 */
2481 if (bld->num_lods == 1) {
2482 /* Emit conditional to choose min image filter or mag image filter
2483 * depending on the lod being > 0 or <= 0, respectively.
2484 */
2485 struct lp_build_if_state if_ctx;
2486
2487 lod_positive = LLVMBuildTrunc(builder, lod_positive,
2488 LLVMInt1TypeInContext(bld->gallivm->context),
2489 "lod_pos");
2490
2491 lp_build_if(&if_ctx, bld->gallivm, lod_positive);
2492 {
2493 /* Use the minification filter */
2494 lp_build_sample_mipmap(bld, min_filter, mip_filter, FALSE,
2495 coords, offsets,
2496 ilevel0, ilevel1, lod_fpart,
2497 texels);
2498 }
2499 lp_build_else(&if_ctx);
2500 {
2501 /* Use the magnification filter */
2502 lp_build_sample_mipmap(bld, mag_filter, PIPE_TEX_MIPFILTER_NONE,
2503 FALSE,
2504 coords, offsets,
2505 ilevel0, NULL, NULL,
2506 texels);
2507 }
2508 lp_build_endif(&if_ctx);
2509 }
2510 else {
2511 LLVMValueRef need_linear, linear_mask;
2512 unsigned mip_filter_for_nearest;
2513 struct lp_build_if_state if_ctx;
2514
2515 if (min_filter == PIPE_TEX_FILTER_LINEAR) {
2516 linear_mask = lod_positive;
2517 mip_filter_for_nearest = PIPE_TEX_MIPFILTER_NONE;
2518 }
2519 else {
2520 linear_mask = lp_build_not(&bld->lodi_bld, lod_positive);
2521 mip_filter_for_nearest = mip_filter;
2522 }
2523 need_linear = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods,
2524 linear_mask);
2525 lp_build_name(need_linear, "need_linear");
2526
2527 if (bld->num_lods != bld->coord_type.length) {
2528 linear_mask = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
2529 bld->lodi_type,
2530 bld->int_coord_type,
2531 linear_mask);
2532 }
2533
2534 lp_build_if(&if_ctx, bld->gallivm, need_linear);
2535 {
2536 /*
2537 * Do sampling with both filters simultaneously. This means using
2538 * a linear filter and doing some tricks (with weights) for the pixels
2539 * which need nearest filter.
2540 * Note that it's probably rare some pixels need nearest and some
2541 * linear filter but the fixups required for the nearest pixels
2542 * aren't all that complicated so just always run a combined path
2543 * if at least some pixels require linear.
2544 */
2545 lp_build_sample_mipmap_both(bld, linear_mask, mip_filter,
2546 coords, offsets,
2547 ilevel0, ilevel1,
2548 lod_fpart, lod_positive,
2549 texels);
2550 }
2551 lp_build_else(&if_ctx);
2552 {
2553 /*
2554 * All pixels require just nearest filtering, which is way
2555 * cheaper than linear, hence do a separate path for that.
2556 */
2557 lp_build_sample_mipmap(bld, PIPE_TEX_FILTER_NEAREST,
2558 mip_filter_for_nearest, FALSE,
2559 coords, offsets,
2560 ilevel0, ilevel1, lod_fpart,
2561 texels);
2562 }
2563 lp_build_endif(&if_ctx);
2564 }
2565 }
2566
2567 for (chan = 0; chan < 4; ++chan) {
2568 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
2569 lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]);
2570 }
2571 }
2572
2573
2574 /**
2575 * Texel fetch function.
2576 * In contrast to general sampling there is no filtering, no coord minification,
2577 * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
2578 * directly to be applied to the selected mip level (after adding texel offsets).
2579 * This function handles texel fetch for all targets where texel fetch is supported
2580 * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
2581 */
2582 static void
lp_build_fetch_texel(struct lp_build_sample_context * bld,unsigned texture_unit,const LLVMValueRef * coords,LLVMValueRef explicit_lod,const LLVMValueRef * offsets,LLVMValueRef * colors_out)2583 lp_build_fetch_texel(struct lp_build_sample_context *bld,
2584 unsigned texture_unit,
2585 const LLVMValueRef *coords,
2586 LLVMValueRef explicit_lod,
2587 const LLVMValueRef *offsets,
2588 LLVMValueRef *colors_out)
2589 {
2590 struct lp_build_context *perquadi_bld = &bld->lodi_bld;
2591 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
2592 unsigned dims = bld->dims, chan;
2593 unsigned target = bld->static_texture_state->target;
2594 boolean out_of_bound_ret_zero = TRUE;
2595 LLVMValueRef size, ilevel;
2596 LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
2597 LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
2598 LLVMValueRef width, height, depth, i, j;
2599 LLVMValueRef offset, out_of_bounds, out1;
2600
2601 out_of_bounds = int_coord_bld->zero;
2602
2603 if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) {
2604 if (bld->num_mips != int_coord_bld->type.length) {
2605 ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
2606 perquadi_bld->type, explicit_lod, 0);
2607 }
2608 else {
2609 ilevel = explicit_lod;
2610 }
2611 lp_build_nearest_mip_level(bld, texture_unit, ilevel, &ilevel,
2612 out_of_bound_ret_zero ? &out_of_bounds : NULL);
2613 }
2614 else {
2615 assert(bld->num_mips == 1);
2616 if (bld->static_texture_state->target != PIPE_BUFFER) {
2617 ilevel = bld->dynamic_state->first_level(bld->dynamic_state, bld->gallivm,
2618 bld->context_ptr, texture_unit);
2619 }
2620 else {
2621 ilevel = lp_build_const_int32(bld->gallivm, 0);
2622 }
2623 }
2624 lp_build_mipmap_level_sizes(bld, ilevel,
2625 &size,
2626 &row_stride_vec, &img_stride_vec);
2627 lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
2628 size, &width, &height, &depth);
2629
2630 if (target == PIPE_TEXTURE_1D_ARRAY ||
2631 target == PIPE_TEXTURE_2D_ARRAY) {
2632 if (out_of_bound_ret_zero) {
2633 z = lp_build_layer_coord(bld, texture_unit, FALSE, z, &out1);
2634 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
2635 }
2636 else {
2637 z = lp_build_layer_coord(bld, texture_unit, FALSE, z, NULL);
2638 }
2639 }
2640
2641 /* This is a lot like border sampling */
2642 if (offsets[0]) {
2643 /*
2644 * coords are really unsigned, offsets are signed, but I don't think
2645 * exceeding 31 bits is possible
2646 */
2647 x = lp_build_add(int_coord_bld, x, offsets[0]);
2648 }
2649 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
2650 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
2651 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
2652 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
2653
2654 if (dims >= 2) {
2655 if (offsets[1]) {
2656 y = lp_build_add(int_coord_bld, y, offsets[1]);
2657 }
2658 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
2659 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
2660 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
2661 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
2662
2663 if (dims >= 3) {
2664 if (offsets[2]) {
2665 z = lp_build_add(int_coord_bld, z, offsets[2]);
2666 }
2667 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
2668 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
2669 out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
2670 out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
2671 }
2672 }
2673
2674 lp_build_sample_offset(int_coord_bld,
2675 bld->format_desc,
2676 x, y, z, row_stride_vec, img_stride_vec,
2677 &offset, &i, &j);
2678
2679 if (bld->static_texture_state->target != PIPE_BUFFER) {
2680 offset = lp_build_add(int_coord_bld, offset,
2681 lp_build_get_mip_offsets(bld, ilevel));
2682 }
2683
2684 offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
2685
2686 lp_build_fetch_rgba_soa(bld->gallivm,
2687 bld->format_desc,
2688 bld->texel_type, TRUE,
2689 bld->base_ptr, offset,
2690 i, j,
2691 bld->cache,
2692 colors_out);
2693
2694 if (out_of_bound_ret_zero) {
2695 /*
2696 * Only needed for ARB_robust_buffer_access_behavior and d3d10.
2697 * Could use min/max above instead of out-of-bounds comparisons
2698 * if we don't care about the result returned for out-of-bounds.
2699 */
2700 for (chan = 0; chan < 4; chan++) {
2701 colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
2702 bld->texel_bld.zero, colors_out[chan]);
2703 }
2704 }
2705 }
2706
2707
2708 /**
2709 * Just set texels to white instead of actually sampling the texture.
2710 * For debugging.
2711 */
2712 void
lp_build_sample_nop(struct gallivm_state * gallivm,struct lp_type type,const LLVMValueRef * coords,LLVMValueRef texel_out[4])2713 lp_build_sample_nop(struct gallivm_state *gallivm,
2714 struct lp_type type,
2715 const LLVMValueRef *coords,
2716 LLVMValueRef texel_out[4])
2717 {
2718 LLVMValueRef one = lp_build_one(gallivm, type);
2719 unsigned chan;
2720
2721 for (chan = 0; chan < 4; chan++) {
2722 texel_out[chan] = one;
2723 }
2724 }
2725
2726
2727 /**
2728 * Build the actual texture sampling code.
2729 * 'texel' will return a vector of four LLVMValueRefs corresponding to
2730 * R, G, B, A.
2731 * \param type vector float type to use for coords, etc.
2732 * \param sample_key
2733 * \param derivs partial derivatives of (s,t,r,q) with respect to x and y
2734 */
2735 static void
lp_build_sample_soa_code(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,struct lp_type type,unsigned sample_key,unsigned texture_index,unsigned sampler_index,LLVMValueRef context_ptr,LLVMValueRef thread_data_ptr,const LLVMValueRef * coords,const LLVMValueRef * offsets,const struct lp_derivatives * derivs,LLVMValueRef lod,LLVMValueRef texel_out[4])2736 lp_build_sample_soa_code(struct gallivm_state *gallivm,
2737 const struct lp_static_texture_state *static_texture_state,
2738 const struct lp_static_sampler_state *static_sampler_state,
2739 struct lp_sampler_dynamic_state *dynamic_state,
2740 struct lp_type type,
2741 unsigned sample_key,
2742 unsigned texture_index,
2743 unsigned sampler_index,
2744 LLVMValueRef context_ptr,
2745 LLVMValueRef thread_data_ptr,
2746 const LLVMValueRef *coords,
2747 const LLVMValueRef *offsets,
2748 const struct lp_derivatives *derivs, /* optional */
2749 LLVMValueRef lod, /* optional */
2750 LLVMValueRef texel_out[4])
2751 {
2752 unsigned target = static_texture_state->target;
2753 unsigned dims = texture_dims(target);
2754 unsigned num_quads = type.length / 4;
2755 unsigned mip_filter, min_img_filter, mag_img_filter, i;
2756 struct lp_build_sample_context bld;
2757 struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
2758 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
2759 LLVMBuilderRef builder = gallivm->builder;
2760 LLVMValueRef tex_width, newcoords[5];
2761 enum lp_sampler_lod_property lod_property;
2762 enum lp_sampler_lod_control lod_control;
2763 enum lp_sampler_op_type op_type;
2764 LLVMValueRef lod_bias = NULL;
2765 LLVMValueRef explicit_lod = NULL;
2766 boolean op_is_tex, op_is_lodq, op_is_gather;
2767
2768 if (0) {
2769 enum pipe_format fmt = static_texture_state->format;
2770 debug_printf("Sample from %s\n", util_format_name(fmt));
2771 }
2772
2773 lod_property = (sample_key & LP_SAMPLER_LOD_PROPERTY_MASK) >>
2774 LP_SAMPLER_LOD_PROPERTY_SHIFT;
2775 lod_control = (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >>
2776 LP_SAMPLER_LOD_CONTROL_SHIFT;
2777 op_type = (sample_key & LP_SAMPLER_OP_TYPE_MASK) >>
2778 LP_SAMPLER_OP_TYPE_SHIFT;
2779
2780 op_is_tex = op_type == LP_SAMPLER_OP_TEXTURE;
2781 op_is_lodq = op_type == LP_SAMPLER_OP_LODQ;
2782 op_is_gather = op_type == LP_SAMPLER_OP_GATHER;
2783
2784 if (lod_control == LP_SAMPLER_LOD_BIAS) {
2785 lod_bias = lod;
2786 assert(lod);
2787 assert(derivs == NULL);
2788 }
2789 else if (lod_control == LP_SAMPLER_LOD_EXPLICIT) {
2790 explicit_lod = lod;
2791 assert(lod);
2792 assert(derivs == NULL);
2793 }
2794 else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
2795 assert(derivs);
2796 assert(lod == NULL);
2797 }
2798 else {
2799 assert(derivs == NULL);
2800 assert(lod == NULL);
2801 }
2802
2803 if (static_texture_state->format == PIPE_FORMAT_NONE) {
2804 /*
2805 * If there's nothing bound, format is NONE, and we must return
2806 * all zero as mandated by d3d10 in this case.
2807 */
2808 unsigned chan;
2809 LLVMValueRef zero = lp_build_zero(gallivm, type);
2810 for (chan = 0; chan < 4; chan++) {
2811 texel_out[chan] = zero;
2812 }
2813 return;
2814 }
2815
2816 assert(type.floating);
2817
2818 /* Setup our build context */
2819 memset(&bld, 0, sizeof bld);
2820 bld.gallivm = gallivm;
2821 bld.context_ptr = context_ptr;
2822 bld.static_sampler_state = &derived_sampler_state;
2823 bld.static_texture_state = static_texture_state;
2824 bld.dynamic_state = dynamic_state;
2825 bld.format_desc = util_format_description(static_texture_state->format);
2826 bld.dims = dims;
2827
2828 if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD || op_is_lodq) {
2829 bld.no_quad_lod = TRUE;
2830 }
2831 if (gallivm_debug & GALLIVM_DEBUG_NO_RHO_APPROX || op_is_lodq) {
2832 bld.no_rho_approx = TRUE;
2833 }
2834 if (gallivm_debug & GALLIVM_DEBUG_NO_BRILINEAR || op_is_lodq) {
2835 bld.no_brilinear = TRUE;
2836 }
2837
2838 bld.vector_width = lp_type_width(type);
2839
2840 bld.float_type = lp_type_float(32);
2841 bld.int_type = lp_type_int(32);
2842 bld.coord_type = type;
2843 bld.int_coord_type = lp_int_type(type);
2844 bld.float_size_in_type = lp_type_float(32);
2845 bld.float_size_in_type.length = dims > 1 ? 4 : 1;
2846 bld.int_size_in_type = lp_int_type(bld.float_size_in_type);
2847 bld.texel_type = type;
2848
2849 /* always using the first channel hopefully should be safe,
2850 * if not things WILL break in other places anyway.
2851 */
2852 if (bld.format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
2853 bld.format_desc->channel[0].pure_integer) {
2854 if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
2855 bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
2856 }
2857 else if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
2858 bld.texel_type = lp_type_uint_vec(type.width, type.width * type.length);
2859 }
2860 }
2861 else if (util_format_has_stencil(bld.format_desc) &&
2862 !util_format_has_depth(bld.format_desc)) {
2863 /* for stencil only formats, sample stencil (uint) */
2864 bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
2865 }
2866
2867 if (!static_texture_state->level_zero_only ||
2868 !static_sampler_state->max_lod_pos || op_is_lodq) {
2869 derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter;
2870 } else {
2871 derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
2872 }
2873 if (op_is_gather) {
2874 /*
2875 * gather4 is exactly like GL_LINEAR filtering but in the end skipping
2876 * the actual filtering. Using mostly the same paths, so cube face
2877 * selection, coord wrapping etc. all naturally uses the same code.
2878 */
2879 derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
2880 derived_sampler_state.min_img_filter = PIPE_TEX_FILTER_LINEAR;
2881 derived_sampler_state.mag_img_filter = PIPE_TEX_FILTER_LINEAR;
2882 }
2883 mip_filter = derived_sampler_state.min_mip_filter;
2884
2885 if (0) {
2886 debug_printf(" .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter);
2887 }
2888
2889 if (static_texture_state->target == PIPE_TEXTURE_CUBE ||
2890 static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY)
2891 {
2892 /*
2893 * Seamless filtering ignores wrap modes.
2894 * Setting to CLAMP_TO_EDGE is correct for nearest filtering, for
2895 * bilinear it's not correct but way better than using for instance repeat.
2896 * Note we even set this for non-seamless. Technically GL allows any wrap
2897 * mode, which made sense when supporting true borders (can get seamless
2898 * effect with border and CLAMP_TO_BORDER), but gallium doesn't support
2899 * borders and d3d9 requires wrap modes to be ignored and it's a pain to fix
2900 * up the sampler state (as it makes it texture dependent).
2901 */
2902 derived_sampler_state.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
2903 derived_sampler_state.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
2904 }
2905 /*
2906 * We could force CLAMP to CLAMP_TO_EDGE here if min/mag filter is nearest,
2907 * so AoS path could be used. Not sure it's worth the trouble...
2908 */
2909
2910 min_img_filter = derived_sampler_state.min_img_filter;
2911 mag_img_filter = derived_sampler_state.mag_img_filter;
2912
2913
2914 /*
2915 * This is all a bit complicated different paths are chosen for performance
2916 * reasons.
2917 * Essentially, there can be 1 lod per element, 1 lod per quad or 1 lod for
2918 * everything (the last two options are equivalent for 4-wide case).
2919 * If there's per-quad lod but we split to 4-wide so we can use AoS, per-quad
2920 * lod is calculated then the lod value extracted afterwards so making this
2921 * case basically the same as far as lod handling is concerned for the
2922 * further sample/filter code as the 1 lod for everything case.
2923 * Different lod handling mostly shows up when building mipmap sizes
2924 * (lp_build_mipmap_level_sizes() and friends) and also in filtering
2925 * (getting the fractional part of the lod to the right texels).
2926 */
2927
2928 /*
2929 * There are other situations where at least the multiple int lods could be
2930 * avoided like min and max lod being equal.
2931 */
2932 bld.num_mips = bld.num_lods = 1;
2933
2934 if (bld.no_quad_lod && bld.no_rho_approx &&
2935 ((mip_filter != PIPE_TEX_MIPFILTER_NONE && op_is_tex &&
2936 (static_texture_state->target == PIPE_TEXTURE_CUBE ||
2937 static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY)) ||
2938 op_is_lodq)) {
2939 /*
2940 * special case for using per-pixel lod even for implicit lod,
2941 * which is generally never required (ok by APIs) except to please
2942 * some (somewhat broken imho) tests (because per-pixel face selection
2943 * can cause derivatives to be different for pixels outside the primitive
2944 * due to the major axis division even if pre-project derivatives are
2945 * looking normal).
2946 * For lodq, we do it to simply avoid scalar pack / unpack (albeit for
2947 * cube maps we do indeed get per-pixel lod values).
2948 */
2949 bld.num_mips = type.length;
2950 bld.num_lods = type.length;
2951 }
2952 else if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT ||
2953 (explicit_lod || lod_bias || derivs)) {
2954 if ((!op_is_tex && target != PIPE_BUFFER) ||
2955 (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
2956 bld.num_mips = type.length;
2957 bld.num_lods = type.length;
2958 }
2959 else if (op_is_tex && min_img_filter != mag_img_filter) {
2960 bld.num_mips = 1;
2961 bld.num_lods = type.length;
2962 }
2963 }
2964 /* TODO: for true scalar_lod should only use 1 lod value */
2965 else if ((!op_is_tex && explicit_lod && target != PIPE_BUFFER) ||
2966 (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
2967 bld.num_mips = num_quads;
2968 bld.num_lods = num_quads;
2969 }
2970 else if (op_is_tex && min_img_filter != mag_img_filter) {
2971 bld.num_mips = 1;
2972 bld.num_lods = num_quads;
2973 }
2974
2975
2976 bld.lodf_type = type;
2977 /* we want native vector size to be able to use our intrinsics */
2978 if (bld.num_lods != type.length) {
2979 /* TODO: this currently always has to be per-quad or per-element */
2980 bld.lodf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
2981 }
2982 bld.lodi_type = lp_int_type(bld.lodf_type);
2983 bld.levelf_type = bld.lodf_type;
2984 if (bld.num_mips == 1) {
2985 bld.levelf_type.length = 1;
2986 }
2987 bld.leveli_type = lp_int_type(bld.levelf_type);
2988 bld.float_size_type = bld.float_size_in_type;
2989 /* Note: size vectors may not be native. They contain minified w/h/d/_ values,
2990 * with per-element lod that is w0/h0/d0/_/w1/h1/d1_/... so up to 8x4f32 */
2991 if (bld.num_mips > 1) {
2992 bld.float_size_type.length = bld.num_mips == type.length ?
2993 bld.num_mips * bld.float_size_in_type.length :
2994 type.length;
2995 }
2996 bld.int_size_type = lp_int_type(bld.float_size_type);
2997
2998 lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
2999 lp_build_context_init(&bld.float_vec_bld, gallivm, type);
3000 lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
3001 lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
3002 lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
3003 lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
3004 lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
3005 lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
3006 lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
3007 lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
3008 lp_build_context_init(&bld.levelf_bld, gallivm, bld.levelf_type);
3009 lp_build_context_init(&bld.leveli_bld, gallivm, bld.leveli_type);
3010 lp_build_context_init(&bld.lodf_bld, gallivm, bld.lodf_type);
3011 lp_build_context_init(&bld.lodi_bld, gallivm, bld.lodi_type);
3012
3013 /* Get the dynamic state */
3014 tex_width = dynamic_state->width(dynamic_state, gallivm,
3015 context_ptr, texture_index);
3016 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm,
3017 context_ptr, texture_index);
3018 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm,
3019 context_ptr, texture_index);
3020 bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm,
3021 context_ptr, texture_index);
3022 bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm,
3023 context_ptr, texture_index);
3024 /* Note that mip_offsets is an array[level] of offsets to texture images */
3025
3026 if (dynamic_state->cache_ptr && thread_data_ptr) {
3027 bld.cache = dynamic_state->cache_ptr(dynamic_state, gallivm,
3028 thread_data_ptr, texture_index);
3029 }
3030
3031 /* width, height, depth as single int vector */
3032 if (dims <= 1) {
3033 bld.int_size = tex_width;
3034 }
3035 else {
3036 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
3037 tex_width,
3038 LLVMConstInt(i32t, 0, 0), "");
3039 if (dims >= 2) {
3040 LLVMValueRef tex_height =
3041 dynamic_state->height(dynamic_state, gallivm,
3042 context_ptr, texture_index);
3043 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
3044 tex_height,
3045 LLVMConstInt(i32t, 1, 0), "");
3046 if (dims >= 3) {
3047 LLVMValueRef tex_depth =
3048 dynamic_state->depth(dynamic_state, gallivm, context_ptr,
3049 texture_index);
3050 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
3051 tex_depth,
3052 LLVMConstInt(i32t, 2, 0), "");
3053 }
3054 }
3055 }
3056
3057 for (i = 0; i < 5; i++) {
3058 newcoords[i] = coords[i];
3059 }
3060
3061 if (util_format_is_pure_integer(static_texture_state->format) &&
3062 !util_format_has_depth(bld.format_desc) && op_is_tex &&
3063 (static_sampler_state->min_mip_filter == PIPE_TEX_MIPFILTER_LINEAR ||
3064 static_sampler_state->min_img_filter == PIPE_TEX_FILTER_LINEAR ||
3065 static_sampler_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR)) {
3066 /*
3067 * Bail if impossible filtering is specified (the awkard additional
3068 * depth check is because it is legal in gallium to have things like S8Z24
3069 * here which would say it's pure int despite such formats should sample
3070 * the depth component).
3071 * In GL such filters make the texture incomplete, this makes it robust
3072 * against state trackers which set this up regardless (we'd crash in the
3073 * lerp later otherwise).
3074 * At least in some apis it may be legal to use such filters with lod
3075 * queries and/or gather (at least for gather d3d10 says only the wrap
3076 * bits are really used hence filter bits are likely simply ignored).
3077 * For fetch, we don't get valid samplers either way here.
3078 */
3079 unsigned chan;
3080 LLVMValueRef zero = lp_build_zero(gallivm, type);
3081 for (chan = 0; chan < 4; chan++) {
3082 texel_out[chan] = zero;
3083 }
3084 return;
3085 }
3086
3087 if (0) {
3088 /* For debug: no-op texture sampling */
3089 lp_build_sample_nop(gallivm,
3090 bld.texel_type,
3091 newcoords,
3092 texel_out);
3093 }
3094
3095 else if (op_type == LP_SAMPLER_OP_FETCH) {
3096 lp_build_fetch_texel(&bld, texture_index, newcoords,
3097 lod, offsets,
3098 texel_out);
3099 }
3100
3101 else {
3102 LLVMValueRef lod_fpart = NULL, lod_positive = NULL;
3103 LLVMValueRef ilevel0 = NULL, ilevel1 = NULL, lod = NULL;
3104 boolean use_aos;
3105
3106 use_aos = util_format_fits_8unorm(bld.format_desc) &&
3107 op_is_tex &&
3108 /* not sure this is strictly needed or simply impossible */
3109 derived_sampler_state.compare_mode == PIPE_TEX_COMPARE_NONE &&
3110 lp_is_simple_wrap_mode(derived_sampler_state.wrap_s);
3111
3112 use_aos &= bld.num_lods <= num_quads ||
3113 derived_sampler_state.min_img_filter ==
3114 derived_sampler_state.mag_img_filter;
3115 if (dims > 1) {
3116 use_aos &= lp_is_simple_wrap_mode(derived_sampler_state.wrap_t);
3117 if (dims > 2) {
3118 use_aos &= lp_is_simple_wrap_mode(derived_sampler_state.wrap_r);
3119 }
3120 }
3121 if ((static_texture_state->target == PIPE_TEXTURE_CUBE ||
3122 static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) &&
3123 derived_sampler_state.seamless_cube_map &&
3124 (derived_sampler_state.min_img_filter == PIPE_TEX_FILTER_LINEAR ||
3125 derived_sampler_state.mag_img_filter == PIPE_TEX_FILTER_LINEAR)) {
3126 /* theoretically possible with AoS filtering but not implemented (complex!) */
3127 use_aos = 0;
3128 }
3129
3130 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
3131 !use_aos && util_format_fits_8unorm(bld.format_desc)) {
3132 debug_printf("%s: using floating point linear filtering for %s\n",
3133 __FUNCTION__, bld.format_desc->short_name);
3134 debug_printf(" min_img %d mag_img %d mip %d target %d seamless %d"
3135 " wraps %d wrapt %d wrapr %d\n",
3136 derived_sampler_state.min_img_filter,
3137 derived_sampler_state.mag_img_filter,
3138 derived_sampler_state.min_mip_filter,
3139 static_texture_state->target,
3140 derived_sampler_state.seamless_cube_map,
3141 derived_sampler_state.wrap_s,
3142 derived_sampler_state.wrap_t,
3143 derived_sampler_state.wrap_r);
3144 }
3145
3146 lp_build_sample_common(&bld, op_is_lodq, texture_index, sampler_index,
3147 newcoords,
3148 derivs, lod_bias, explicit_lod,
3149 &lod_positive, &lod, &lod_fpart,
3150 &ilevel0, &ilevel1);
3151
3152 if (op_is_lodq) {
3153 texel_out[0] = lod_fpart;
3154 texel_out[1] = lod;
3155 texel_out[2] = texel_out[3] = bld.coord_bld.zero;
3156 return;
3157 }
3158
3159 if (use_aos && static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
3160 /* The aos path doesn't do seamless filtering so simply add cube layer
3161 * to face now.
3162 */
3163 newcoords[2] = lp_build_add(&bld.int_coord_bld, newcoords[2], newcoords[3]);
3164 }
3165
3166 /*
3167 * we only try 8-wide sampling with soa or if we have AVX2
3168 * as it appears to be a loss with just AVX)
3169 */
3170 if (num_quads == 1 || !use_aos ||
3171 (util_cpu_caps.has_avx2 &&
3172 (bld.num_lods == 1 ||
3173 derived_sampler_state.min_img_filter == derived_sampler_state.mag_img_filter))) {
3174 if (use_aos) {
3175 /* do sampling/filtering with fixed pt arithmetic */
3176 lp_build_sample_aos(&bld, sampler_index,
3177 newcoords[0], newcoords[1],
3178 newcoords[2],
3179 offsets, lod_positive, lod_fpart,
3180 ilevel0, ilevel1,
3181 texel_out);
3182 }
3183
3184 else {
3185 lp_build_sample_general(&bld, sampler_index,
3186 op_type == LP_SAMPLER_OP_GATHER,
3187 newcoords, offsets,
3188 lod_positive, lod_fpart,
3189 ilevel0, ilevel1,
3190 texel_out);
3191 }
3192 }
3193 else {
3194 unsigned j;
3195 struct lp_build_sample_context bld4;
3196 struct lp_type type4 = type;
3197 unsigned i;
3198 LLVMValueRef texelout4[4];
3199 LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
3200
3201 type4.length = 4;
3202
3203 /* Setup our build context */
3204 memset(&bld4, 0, sizeof bld4);
3205 bld4.no_quad_lod = bld.no_quad_lod;
3206 bld4.no_rho_approx = bld.no_rho_approx;
3207 bld4.no_brilinear = bld.no_brilinear;
3208 bld4.gallivm = bld.gallivm;
3209 bld4.context_ptr = bld.context_ptr;
3210 bld4.static_texture_state = bld.static_texture_state;
3211 bld4.static_sampler_state = bld.static_sampler_state;
3212 bld4.dynamic_state = bld.dynamic_state;
3213 bld4.format_desc = bld.format_desc;
3214 bld4.dims = bld.dims;
3215 bld4.row_stride_array = bld.row_stride_array;
3216 bld4.img_stride_array = bld.img_stride_array;
3217 bld4.base_ptr = bld.base_ptr;
3218 bld4.mip_offsets = bld.mip_offsets;
3219 bld4.int_size = bld.int_size;
3220 bld4.cache = bld.cache;
3221
3222 bld4.vector_width = lp_type_width(type4);
3223
3224 bld4.float_type = lp_type_float(32);
3225 bld4.int_type = lp_type_int(32);
3226 bld4.coord_type = type4;
3227 bld4.int_coord_type = lp_int_type(type4);
3228 bld4.float_size_in_type = lp_type_float(32);
3229 bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
3230 bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
3231 bld4.texel_type = bld.texel_type;
3232 bld4.texel_type.length = 4;
3233
3234 bld4.num_mips = bld4.num_lods = 1;
3235 if (bld4.no_quad_lod && bld4.no_rho_approx &&
3236 (static_texture_state->target == PIPE_TEXTURE_CUBE ||
3237 static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) &&
3238 (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
3239 bld4.num_mips = type4.length;
3240 bld4.num_lods = type4.length;
3241 }
3242 if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT &&
3243 (explicit_lod || lod_bias || derivs)) {
3244 if ((!op_is_tex && target != PIPE_BUFFER) ||
3245 (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
3246 bld4.num_mips = type4.length;
3247 bld4.num_lods = type4.length;
3248 }
3249 else if (op_is_tex && min_img_filter != mag_img_filter) {
3250 bld4.num_mips = 1;
3251 bld4.num_lods = type4.length;
3252 }
3253 }
3254
3255 /* we want native vector size to be able to use our intrinsics */
3256 bld4.lodf_type = type4;
3257 if (bld4.num_lods != type4.length) {
3258 bld4.lodf_type.length = 1;
3259 }
3260 bld4.lodi_type = lp_int_type(bld4.lodf_type);
3261 bld4.levelf_type = type4;
3262 if (bld4.num_mips != type4.length) {
3263 bld4.levelf_type.length = 1;
3264 }
3265 bld4.leveli_type = lp_int_type(bld4.levelf_type);
3266 bld4.float_size_type = bld4.float_size_in_type;
3267 if (bld4.num_mips > 1) {
3268 bld4.float_size_type.length = bld4.num_mips == type4.length ?
3269 bld4.num_mips * bld4.float_size_in_type.length :
3270 type4.length;
3271 }
3272 bld4.int_size_type = lp_int_type(bld4.float_size_type);
3273
3274 lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
3275 lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
3276 lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
3277 lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
3278 lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
3279 lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
3280 lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
3281 lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
3282 lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
3283 lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
3284 lp_build_context_init(&bld4.levelf_bld, gallivm, bld4.levelf_type);
3285 lp_build_context_init(&bld4.leveli_bld, gallivm, bld4.leveli_type);
3286 lp_build_context_init(&bld4.lodf_bld, gallivm, bld4.lodf_type);
3287 lp_build_context_init(&bld4.lodi_bld, gallivm, bld4.lodi_type);
3288
3289 for (i = 0; i < num_quads; i++) {
3290 LLVMValueRef s4, t4, r4;
3291 LLVMValueRef lod_positive4, lod_fpart4 = NULL;
3292 LLVMValueRef ilevel04, ilevel14 = NULL;
3293 LLVMValueRef offsets4[4] = { NULL };
3294 unsigned num_lods = bld4.num_lods;
3295
3296 s4 = lp_build_extract_range(gallivm, newcoords[0], 4*i, 4);
3297 t4 = lp_build_extract_range(gallivm, newcoords[1], 4*i, 4);
3298 r4 = lp_build_extract_range(gallivm, newcoords[2], 4*i, 4);
3299
3300 if (offsets[0]) {
3301 offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4);
3302 if (dims > 1) {
3303 offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4);
3304 if (dims > 2) {
3305 offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4);
3306 }
3307 }
3308 }
3309 lod_positive4 = lp_build_extract_range(gallivm, lod_positive, num_lods * i, num_lods);
3310 ilevel04 = bld.num_mips == 1 ? ilevel0 :
3311 lp_build_extract_range(gallivm, ilevel0, num_lods * i, num_lods);
3312 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
3313 ilevel14 = lp_build_extract_range(gallivm, ilevel1, num_lods * i, num_lods);
3314 lod_fpart4 = lp_build_extract_range(gallivm, lod_fpart, num_lods * i, num_lods);
3315 }
3316
3317 if (use_aos) {
3318 /* do sampling/filtering with fixed pt arithmetic */
3319 lp_build_sample_aos(&bld4, sampler_index,
3320 s4, t4, r4, offsets4,
3321 lod_positive4, lod_fpart4,
3322 ilevel04, ilevel14,
3323 texelout4);
3324 }
3325
3326 else {
3327 /* this path is currently unreachable and hence might break easily... */
3328 LLVMValueRef newcoords4[5];
3329 newcoords4[0] = s4;
3330 newcoords4[1] = t4;
3331 newcoords4[2] = r4;
3332 newcoords4[3] = lp_build_extract_range(gallivm, newcoords[3], 4*i, 4);
3333 newcoords4[4] = lp_build_extract_range(gallivm, newcoords[4], 4*i, 4);
3334
3335 lp_build_sample_general(&bld4, sampler_index,
3336 op_type == LP_SAMPLER_OP_GATHER,
3337 newcoords4, offsets4,
3338 lod_positive4, lod_fpart4,
3339 ilevel04, ilevel14,
3340 texelout4);
3341 }
3342 for (j = 0; j < 4; j++) {
3343 texelouttmp[j][i] = texelout4[j];
3344 }
3345 }
3346
3347 for (j = 0; j < 4; j++) {
3348 texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
3349 }
3350 }
3351 }
3352
3353 if (target != PIPE_BUFFER && op_type != LP_SAMPLER_OP_GATHER) {
3354 apply_sampler_swizzle(&bld, texel_out);
3355 }
3356
3357 /*
3358 * texel type can be a (32bit) int/uint (for pure int formats only),
3359 * however we are expected to always return floats (storage is untyped).
3360 */
3361 if (!bld.texel_type.floating) {
3362 unsigned chan;
3363 for (chan = 0; chan < 4; chan++) {
3364 texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan],
3365 lp_build_vec_type(gallivm, type), "");
3366 }
3367 }
3368 }
3369
3370
3371 #define USE_TEX_FUNC_CALL 1
3372
3373 #define LP_MAX_TEX_FUNC_ARGS 32
3374
3375 static inline void
get_target_info(enum pipe_texture_target target,unsigned * num_coords,unsigned * num_derivs,unsigned * num_offsets,unsigned * layer)3376 get_target_info(enum pipe_texture_target target,
3377 unsigned *num_coords, unsigned *num_derivs,
3378 unsigned *num_offsets, unsigned *layer)
3379 {
3380 unsigned dims = texture_dims(target);
3381 *num_coords = dims;
3382 *num_offsets = dims;
3383 *num_derivs = (target == PIPE_TEXTURE_CUBE ||
3384 target == PIPE_TEXTURE_CUBE_ARRAY) ? 3 : dims;
3385 *layer = has_layer_coord(target) ? 2: 0;
3386 if (target == PIPE_TEXTURE_CUBE_ARRAY) {
3387 /*
3388 * dims doesn't include r coord for cubes - this is handled
3389 * by layer instead, but need to fix up for cube arrays...
3390 */
3391 *layer = 3;
3392 *num_coords = 3;
3393 }
3394 }
3395
3396
3397 /**
3398 * Generate the function body for a texture sampling function.
3399 */
3400 static void
lp_build_sample_gen_func(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,struct lp_type type,unsigned texture_index,unsigned sampler_index,LLVMValueRef function,unsigned num_args,unsigned sample_key)3401 lp_build_sample_gen_func(struct gallivm_state *gallivm,
3402 const struct lp_static_texture_state *static_texture_state,
3403 const struct lp_static_sampler_state *static_sampler_state,
3404 struct lp_sampler_dynamic_state *dynamic_state,
3405 struct lp_type type,
3406 unsigned texture_index,
3407 unsigned sampler_index,
3408 LLVMValueRef function,
3409 unsigned num_args,
3410 unsigned sample_key)
3411 {
3412 LLVMBuilderRef old_builder;
3413 LLVMBasicBlockRef block;
3414 LLVMValueRef coords[5];
3415 LLVMValueRef offsets[3] = { NULL };
3416 LLVMValueRef lod = NULL;
3417 LLVMValueRef context_ptr;
3418 LLVMValueRef thread_data_ptr = NULL;
3419 LLVMValueRef texel_out[4];
3420 struct lp_derivatives derivs;
3421 struct lp_derivatives *deriv_ptr = NULL;
3422 unsigned num_param = 0;
3423 unsigned i, num_coords, num_derivs, num_offsets, layer;
3424 enum lp_sampler_lod_control lod_control;
3425 boolean need_cache = FALSE;
3426
3427 lod_control = (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >>
3428 LP_SAMPLER_LOD_CONTROL_SHIFT;
3429
3430 get_target_info(static_texture_state->target,
3431 &num_coords, &num_derivs, &num_offsets, &layer);
3432
3433 if (dynamic_state->cache_ptr) {
3434 const struct util_format_description *format_desc;
3435 format_desc = util_format_description(static_texture_state->format);
3436 if (format_desc && format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
3437 need_cache = TRUE;
3438 }
3439 }
3440
3441 /* "unpack" arguments */
3442 context_ptr = LLVMGetParam(function, num_param++);
3443 if (need_cache) {
3444 thread_data_ptr = LLVMGetParam(function, num_param++);
3445 }
3446 for (i = 0; i < num_coords; i++) {
3447 coords[i] = LLVMGetParam(function, num_param++);
3448 }
3449 for (i = num_coords; i < 5; i++) {
3450 /* This is rather unfortunate... */
3451 coords[i] = lp_build_undef(gallivm, type);
3452 }
3453 if (layer) {
3454 coords[layer] = LLVMGetParam(function, num_param++);
3455 }
3456 if (sample_key & LP_SAMPLER_SHADOW) {
3457 coords[4] = LLVMGetParam(function, num_param++);
3458 }
3459 if (sample_key & LP_SAMPLER_OFFSETS) {
3460 for (i = 0; i < num_offsets; i++) {
3461 offsets[i] = LLVMGetParam(function, num_param++);
3462 }
3463 }
3464 if (lod_control == LP_SAMPLER_LOD_BIAS ||
3465 lod_control == LP_SAMPLER_LOD_EXPLICIT) {
3466 lod = LLVMGetParam(function, num_param++);
3467 }
3468 else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
3469 for (i = 0; i < num_derivs; i++) {
3470 derivs.ddx[i] = LLVMGetParam(function, num_param++);
3471 derivs.ddy[i] = LLVMGetParam(function, num_param++);
3472 }
3473 deriv_ptr = &derivs;
3474 }
3475
3476 assert(num_args == num_param);
3477
3478 /*
3479 * Function body
3480 */
3481
3482 old_builder = gallivm->builder;
3483 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
3484 gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
3485 LLVMPositionBuilderAtEnd(gallivm->builder, block);
3486
3487 lp_build_sample_soa_code(gallivm,
3488 static_texture_state,
3489 static_sampler_state,
3490 dynamic_state,
3491 type,
3492 sample_key,
3493 texture_index,
3494 sampler_index,
3495 context_ptr,
3496 thread_data_ptr,
3497 coords,
3498 offsets,
3499 deriv_ptr,
3500 lod,
3501 texel_out);
3502
3503 LLVMBuildAggregateRet(gallivm->builder, texel_out, 4);
3504
3505 LLVMDisposeBuilder(gallivm->builder);
3506 gallivm->builder = old_builder;
3507
3508 gallivm_verify_function(gallivm, function);
3509 }
3510
3511
3512 /**
3513 * Call the matching function for texture sampling.
3514 * If there's no match, generate a new one.
3515 */
3516 static void
lp_build_sample_soa_func(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,const struct lp_sampler_params * params)3517 lp_build_sample_soa_func(struct gallivm_state *gallivm,
3518 const struct lp_static_texture_state *static_texture_state,
3519 const struct lp_static_sampler_state *static_sampler_state,
3520 struct lp_sampler_dynamic_state *dynamic_state,
3521 const struct lp_sampler_params *params)
3522 {
3523 LLVMBuilderRef builder = gallivm->builder;
3524 LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(
3525 LLVMGetInsertBlock(builder)));
3526 LLVMValueRef function, inst;
3527 LLVMValueRef args[LP_MAX_TEX_FUNC_ARGS];
3528 LLVMBasicBlockRef bb;
3529 LLVMValueRef tex_ret;
3530 unsigned num_args = 0;
3531 char func_name[64];
3532 unsigned i, num_coords, num_derivs, num_offsets, layer;
3533 unsigned texture_index = params->texture_index;
3534 unsigned sampler_index = params->sampler_index;
3535 unsigned sample_key = params->sample_key;
3536 const LLVMValueRef *coords = params->coords;
3537 const LLVMValueRef *offsets = params->offsets;
3538 const struct lp_derivatives *derivs = params->derivs;
3539 enum lp_sampler_lod_control lod_control;
3540 boolean need_cache = FALSE;
3541
3542 lod_control = (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >>
3543 LP_SAMPLER_LOD_CONTROL_SHIFT;
3544
3545 get_target_info(static_texture_state->target,
3546 &num_coords, &num_derivs, &num_offsets, &layer);
3547
3548 if (dynamic_state->cache_ptr) {
3549 const struct util_format_description *format_desc;
3550 format_desc = util_format_description(static_texture_state->format);
3551 if (format_desc && format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
3552 /*
3553 * This is not 100% correct, if we have cache but the
3554 * util_format_s3tc_prefer is true the cache won't get used
3555 * regardless (could hook up the block decode there...) */
3556 need_cache = TRUE;
3557 }
3558 }
3559 /*
3560 * texture function matches are found by name.
3561 * Thus the name has to include both the texture and sampler unit
3562 * (which covers all static state) plus the actual texture function
3563 * (including things like offsets, shadow coord, lod control).
3564 * Additionally lod_property has to be included too.
3565 */
3566
3567 util_snprintf(func_name, sizeof(func_name), "texfunc_res_%d_sam_%d_%x",
3568 texture_index, sampler_index, sample_key);
3569
3570 function = LLVMGetNamedFunction(module, func_name);
3571
3572 if(!function) {
3573 LLVMTypeRef arg_types[LP_MAX_TEX_FUNC_ARGS];
3574 LLVMTypeRef ret_type;
3575 LLVMTypeRef function_type;
3576 LLVMTypeRef val_type[4];
3577 unsigned num_param = 0;
3578
3579 /*
3580 * Generate the function prototype.
3581 */
3582
3583 arg_types[num_param++] = LLVMTypeOf(params->context_ptr);
3584 if (need_cache) {
3585 arg_types[num_param++] = LLVMTypeOf(params->thread_data_ptr);
3586 }
3587 for (i = 0; i < num_coords; i++) {
3588 arg_types[num_param++] = LLVMTypeOf(coords[0]);
3589 assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[i]));
3590 }
3591 if (layer) {
3592 arg_types[num_param++] = LLVMTypeOf(coords[layer]);
3593 assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[layer]));
3594 }
3595 if (sample_key & LP_SAMPLER_SHADOW) {
3596 arg_types[num_param++] = LLVMTypeOf(coords[0]);
3597 }
3598 if (sample_key & LP_SAMPLER_OFFSETS) {
3599 for (i = 0; i < num_offsets; i++) {
3600 arg_types[num_param++] = LLVMTypeOf(offsets[0]);
3601 assert(LLVMTypeOf(offsets[0]) == LLVMTypeOf(offsets[i]));
3602 }
3603 }
3604 if (lod_control == LP_SAMPLER_LOD_BIAS ||
3605 lod_control == LP_SAMPLER_LOD_EXPLICIT) {
3606 arg_types[num_param++] = LLVMTypeOf(params->lod);
3607 }
3608 else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
3609 for (i = 0; i < num_derivs; i++) {
3610 arg_types[num_param++] = LLVMTypeOf(derivs->ddx[i]);
3611 arg_types[num_param++] = LLVMTypeOf(derivs->ddy[i]);
3612 assert(LLVMTypeOf(derivs->ddx[0]) == LLVMTypeOf(derivs->ddx[i]));
3613 assert(LLVMTypeOf(derivs->ddy[0]) == LLVMTypeOf(derivs->ddy[i]));
3614 }
3615 }
3616
3617 val_type[0] = val_type[1] = val_type[2] = val_type[3] =
3618 lp_build_vec_type(gallivm, params->type);
3619 ret_type = LLVMStructTypeInContext(gallivm->context, val_type, 4, 0);
3620 function_type = LLVMFunctionType(ret_type, arg_types, num_param, 0);
3621 function = LLVMAddFunction(module, func_name, function_type);
3622
3623 for (i = 0; i < num_param; ++i) {
3624 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
3625
3626 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
3627 }
3628 }
3629
3630 LLVMSetFunctionCallConv(function, LLVMFastCallConv);
3631 LLVMSetLinkage(function, LLVMInternalLinkage);
3632
3633 lp_build_sample_gen_func(gallivm,
3634 static_texture_state,
3635 static_sampler_state,
3636 dynamic_state,
3637 params->type,
3638 texture_index,
3639 sampler_index,
3640 function,
3641 num_param,
3642 sample_key);
3643 }
3644
3645 num_args = 0;
3646 args[num_args++] = params->context_ptr;
3647 if (need_cache) {
3648 args[num_args++] = params->thread_data_ptr;
3649 }
3650 for (i = 0; i < num_coords; i++) {
3651 args[num_args++] = coords[i];
3652 }
3653 if (layer) {
3654 args[num_args++] = coords[layer];
3655 }
3656 if (sample_key & LP_SAMPLER_SHADOW) {
3657 args[num_args++] = coords[4];
3658 }
3659 if (sample_key & LP_SAMPLER_OFFSETS) {
3660 for (i = 0; i < num_offsets; i++) {
3661 args[num_args++] = offsets[i];
3662 }
3663 }
3664 if (lod_control == LP_SAMPLER_LOD_BIAS ||
3665 lod_control == LP_SAMPLER_LOD_EXPLICIT) {
3666 args[num_args++] = params->lod;
3667 }
3668 else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
3669 for (i = 0; i < num_derivs; i++) {
3670 args[num_args++] = derivs->ddx[i];
3671 args[num_args++] = derivs->ddy[i];
3672 }
3673 }
3674
3675 assert(num_args <= LP_MAX_TEX_FUNC_ARGS);
3676
3677 tex_ret = LLVMBuildCall(builder, function, args, num_args, "");
3678 bb = LLVMGetInsertBlock(builder);
3679 inst = LLVMGetLastInstruction(bb);
3680 LLVMSetInstructionCallConv(inst, LLVMFastCallConv);
3681
3682 for (i = 0; i < 4; i++) {
3683 params->texel[i] = LLVMBuildExtractValue(gallivm->builder, tex_ret, i, "");
3684 }
3685 }
3686
3687
3688 /**
3689 * Build texture sampling code.
3690 * Either via a function call or inline it directly.
3691 */
3692 void
lp_build_sample_soa(const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,struct gallivm_state * gallivm,const struct lp_sampler_params * params)3693 lp_build_sample_soa(const struct lp_static_texture_state *static_texture_state,
3694 const struct lp_static_sampler_state *static_sampler_state,
3695 struct lp_sampler_dynamic_state *dynamic_state,
3696 struct gallivm_state *gallivm,
3697 const struct lp_sampler_params *params)
3698 {
3699 boolean use_tex_func = FALSE;
3700
3701 /*
3702 * Do not use a function call if the sampling is "simple enough".
3703 * We define this by
3704 * a) format
3705 * b) no mips (either one level only or no mip filter)
3706 * No mips will definitely make the code smaller, though
3707 * the format requirement is a bit iffy - there's some (SoA) formats
3708 * which definitely generate less code. This does happen to catch
3709 * some important cases though which are hurt quite a bit by using
3710 * a call (though not really because of the call overhead but because
3711 * they are reusing the same texture unit with some of the same
3712 * parameters).
3713 * Ideally we'd let llvm recognize this stuff by doing IPO passes.
3714 */
3715
3716 if (USE_TEX_FUNC_CALL) {
3717 const struct util_format_description *format_desc;
3718 boolean simple_format;
3719 boolean simple_tex;
3720 enum lp_sampler_op_type op_type;
3721 format_desc = util_format_description(static_texture_state->format);
3722 simple_format = !format_desc ||
3723 (util_format_is_rgba8_variant(format_desc) &&
3724 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB);
3725
3726 op_type = (params->sample_key & LP_SAMPLER_OP_TYPE_MASK) >>
3727 LP_SAMPLER_OP_TYPE_SHIFT;
3728 simple_tex =
3729 op_type != LP_SAMPLER_OP_TEXTURE ||
3730 ((static_sampler_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE ||
3731 static_texture_state->level_zero_only == TRUE) &&
3732 static_sampler_state->min_img_filter == static_sampler_state->mag_img_filter);
3733
3734 use_tex_func = format_desc && !(simple_format && simple_tex);
3735 }
3736
3737 if (use_tex_func) {
3738 lp_build_sample_soa_func(gallivm,
3739 static_texture_state,
3740 static_sampler_state,
3741 dynamic_state,
3742 params);
3743 }
3744 else {
3745 lp_build_sample_soa_code(gallivm,
3746 static_texture_state,
3747 static_sampler_state,
3748 dynamic_state,
3749 params->type,
3750 params->sample_key,
3751 params->texture_index,
3752 params->sampler_index,
3753 params->context_ptr,
3754 params->thread_data_ptr,
3755 params->coords,
3756 params->offsets,
3757 params->derivs,
3758 params->lod,
3759 params->texel);
3760 }
3761 }
3762
3763
3764 void
lp_build_size_query_soa(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_state,struct lp_sampler_dynamic_state * dynamic_state,const struct lp_sampler_size_query_params * params)3765 lp_build_size_query_soa(struct gallivm_state *gallivm,
3766 const struct lp_static_texture_state *static_state,
3767 struct lp_sampler_dynamic_state *dynamic_state,
3768 const struct lp_sampler_size_query_params *params)
3769 {
3770 LLVMValueRef lod, level = 0, size;
3771 LLVMValueRef first_level = NULL;
3772 int dims, i;
3773 boolean has_array;
3774 unsigned num_lods = 1;
3775 struct lp_build_context bld_int_vec4;
3776 LLVMValueRef context_ptr = params->context_ptr;
3777 unsigned texture_unit = params->texture_unit;
3778 unsigned target = params->target;
3779
3780 if (static_state->format == PIPE_FORMAT_NONE) {
3781 /*
3782 * If there's nothing bound, format is NONE, and we must return
3783 * all zero as mandated by d3d10 in this case.
3784 */
3785 unsigned chan;
3786 LLVMValueRef zero = lp_build_const_vec(gallivm, params->int_type, 0.0F);
3787 for (chan = 0; chan < 4; chan++) {
3788 params->sizes_out[chan] = zero;
3789 }
3790 return;
3791 }
3792
3793 /*
3794 * Do some sanity verification about bound texture and shader dcl target.
3795 * Not entirely sure what's possible but assume array/non-array
3796 * always compatible (probably not ok for OpenGL but d3d10 has no
3797 * distinction of arrays at the resource level).
3798 * Everything else looks bogus (though not entirely sure about rect/2d).
3799 * Currently disabled because it causes assertion failures if there's
3800 * nothing bound (or rather a dummy texture, not that this case would
3801 * return the right values).
3802 */
3803 if (0 && static_state->target != target) {
3804 if (static_state->target == PIPE_TEXTURE_1D)
3805 assert(target == PIPE_TEXTURE_1D_ARRAY);
3806 else if (static_state->target == PIPE_TEXTURE_1D_ARRAY)
3807 assert(target == PIPE_TEXTURE_1D);
3808 else if (static_state->target == PIPE_TEXTURE_2D)
3809 assert(target == PIPE_TEXTURE_2D_ARRAY);
3810 else if (static_state->target == PIPE_TEXTURE_2D_ARRAY)
3811 assert(target == PIPE_TEXTURE_2D);
3812 else if (static_state->target == PIPE_TEXTURE_CUBE)
3813 assert(target == PIPE_TEXTURE_CUBE_ARRAY);
3814 else if (static_state->target == PIPE_TEXTURE_CUBE_ARRAY)
3815 assert(target == PIPE_TEXTURE_CUBE);
3816 else
3817 assert(0);
3818 }
3819
3820 dims = texture_dims(target);
3821
3822 switch (target) {
3823 case PIPE_TEXTURE_1D_ARRAY:
3824 case PIPE_TEXTURE_2D_ARRAY:
3825 case PIPE_TEXTURE_CUBE_ARRAY:
3826 has_array = TRUE;
3827 break;
3828 default:
3829 has_array = FALSE;
3830 break;
3831 }
3832
3833 assert(!params->int_type.floating);
3834
3835 lp_build_context_init(&bld_int_vec4, gallivm, lp_type_int_vec(32, 128));
3836
3837 if (params->explicit_lod) {
3838 /* FIXME: this needs to honor per-element lod */
3839 lod = LLVMBuildExtractElement(gallivm->builder, params->explicit_lod,
3840 lp_build_const_int32(gallivm, 0), "");
3841 first_level = dynamic_state->first_level(dynamic_state, gallivm,
3842 context_ptr, texture_unit);
3843 level = LLVMBuildAdd(gallivm->builder, lod, first_level, "level");
3844 lod = lp_build_broadcast_scalar(&bld_int_vec4, level);
3845 } else {
3846 lod = bld_int_vec4.zero;
3847 }
3848
3849 size = bld_int_vec4.undef;
3850
3851 size = LLVMBuildInsertElement(gallivm->builder, size,
3852 dynamic_state->width(dynamic_state, gallivm,
3853 context_ptr, texture_unit),
3854 lp_build_const_int32(gallivm, 0), "");
3855
3856 if (dims >= 2) {
3857 size = LLVMBuildInsertElement(gallivm->builder, size,
3858 dynamic_state->height(dynamic_state, gallivm,
3859 context_ptr, texture_unit),
3860 lp_build_const_int32(gallivm, 1), "");
3861 }
3862
3863 if (dims >= 3) {
3864 size = LLVMBuildInsertElement(gallivm->builder, size,
3865 dynamic_state->depth(dynamic_state, gallivm,
3866 context_ptr, texture_unit),
3867 lp_build_const_int32(gallivm, 2), "");
3868 }
3869
3870 size = lp_build_minify(&bld_int_vec4, size, lod, TRUE);
3871
3872 if (has_array) {
3873 LLVMValueRef layers = dynamic_state->depth(dynamic_state, gallivm,
3874 context_ptr, texture_unit);
3875 if (target == PIPE_TEXTURE_CUBE_ARRAY) {
3876 /*
3877 * It looks like GL wants number of cubes, d3d10.1 has it undefined?
3878 * Could avoid this by passing in number of cubes instead of total
3879 * number of layers (might make things easier elsewhere too).
3880 */
3881 LLVMValueRef six = lp_build_const_int32(gallivm, 6);
3882 layers = LLVMBuildSDiv(gallivm->builder, layers, six, "");
3883 }
3884 size = LLVMBuildInsertElement(gallivm->builder, size, layers,
3885 lp_build_const_int32(gallivm, dims), "");
3886 }
3887
3888 /*
3889 * d3d10 requires zero for x/y/z values (but not w, i.e. mip levels)
3890 * if level is out of bounds (note this can't cover unbound texture
3891 * here, which also requires returning zero).
3892 */
3893 if (params->explicit_lod && params->is_sviewinfo) {
3894 LLVMValueRef last_level, out, out1;
3895 struct lp_build_context leveli_bld;
3896
3897 /* everything is scalar for now */
3898 lp_build_context_init(&leveli_bld, gallivm, lp_type_int_vec(32, 32));
3899 last_level = dynamic_state->last_level(dynamic_state, gallivm,
3900 context_ptr, texture_unit);
3901
3902 out = lp_build_cmp(&leveli_bld, PIPE_FUNC_LESS, level, first_level);
3903 out1 = lp_build_cmp(&leveli_bld, PIPE_FUNC_GREATER, level, last_level);
3904 out = lp_build_or(&leveli_bld, out, out1);
3905 if (num_lods == 1) {
3906 out = lp_build_broadcast_scalar(&bld_int_vec4, out);
3907 }
3908 else {
3909 /* TODO */
3910 assert(0);
3911 }
3912 size = lp_build_andnot(&bld_int_vec4, size, out);
3913 }
3914 for (i = 0; i < dims + (has_array ? 1 : 0); i++) {
3915 params->sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec4.type, params->int_type,
3916 size,
3917 lp_build_const_int32(gallivm, i));
3918 }
3919 if (params->is_sviewinfo) {
3920 for (; i < 4; i++) {
3921 params->sizes_out[i] = lp_build_const_vec(gallivm, params->int_type, 0.0);
3922 }
3923 }
3924
3925 /*
3926 * if there's no explicit_lod (buffers, rects) queries requiring nr of
3927 * mips would be illegal.
3928 */
3929 if (params->is_sviewinfo && params->explicit_lod) {
3930 struct lp_build_context bld_int_scalar;
3931 LLVMValueRef num_levels;
3932 lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32));
3933
3934 if (static_state->level_zero_only) {
3935 num_levels = bld_int_scalar.one;
3936 }
3937 else {
3938 LLVMValueRef last_level;
3939
3940 last_level = dynamic_state->last_level(dynamic_state, gallivm,
3941 context_ptr, texture_unit);
3942 num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level);
3943 num_levels = lp_build_add(&bld_int_scalar, num_levels, bld_int_scalar.one);
3944 }
3945 params->sizes_out[3] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, params->int_type),
3946 num_levels);
3947 }
3948 }
3949