1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
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.
26 **************************************************************************/
30 * Texture sampling -- SoA.
32 * @author Jose Fonseca <jfonseca@vmware.com>
33 * @author Brian Paul <brianp@vmware.com>
36 #include "pipe/p_defines.h"
37 #include "pipe/p_state.h"
38 #include "util/u_debug.h"
39 #include "util/u_dump.h"
40 #include "util/u_memory.h"
41 #include "util/u_math.h"
42 #include "util/u_format.h"
43 #include "lp_bld_debug.h"
44 #include "lp_bld_type.h"
45 #include "lp_bld_const.h"
46 #include "lp_bld_conv.h"
47 #include "lp_bld_arit.h"
48 #include "lp_bld_bitarit.h"
49 #include "lp_bld_logic.h"
50 #include "lp_bld_printf.h"
51 #include "lp_bld_swizzle.h"
52 #include "lp_bld_flow.h"
53 #include "lp_bld_gather.h"
54 #include "lp_bld_format.h"
55 #include "lp_bld_sample.h"
56 #include "lp_bld_sample_aos.h"
57 #include "lp_bld_struct.h"
58 #include "lp_bld_quad.h"
62 * Generate code to fetch a texel from a texture at int coords (x, y, z).
63 * The computation depends on whether the texture is 1D, 2D or 3D.
64 * The result, texel, will be float vectors:
65 * texel[0] = red values
66 * texel[1] = green values
67 * texel[2] = blue values
68 * texel[3] = alpha values
71 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
79 LLVMValueRef y_stride,
80 LLVMValueRef z_stride,
81 LLVMValueRef data_ptr,
82 LLVMValueRef texel_out[4])
84 const struct lp_sampler_static_state *static_state = bld->static_state;
85 const unsigned dims = bld->dims;
86 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
89 LLVMValueRef use_border = NULL;
91 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
92 if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
93 static_state->min_img_filter,
94 static_state->mag_img_filter)) {
96 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
97 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
98 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
102 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
103 static_state->min_img_filter,
104 static_state->mag_img_filter)) {
106 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
107 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
109 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
110 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
113 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
118 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
119 static_state->min_img_filter,
120 static_state->mag_img_filter)) {
122 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
123 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
125 use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
126 use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
129 use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
133 /* convert x,y,z coords to linear offset from start of texture, in bytes */
134 lp_build_sample_offset(&bld->int_coord_bld,
136 x, y, z, y_stride, z_stride,
140 /* If we can sample the border color, it means that texcoords may
141 * lie outside the bounds of the texture image. We need to do
142 * something to prevent reading out of bounds and causing a segfault.
144 * Simply AND the texture coords with !use_border. This will cause
145 * coords which are out of bounds to become zero. Zero's guaranteed
146 * to be inside the texture image.
148 offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
151 lp_build_fetch_rgba_soa(bld->builder,
159 * Note: if we find an app which frequently samples the texture border
160 * we might want to implement a true conditional here to avoid sampling
161 * the texture whenever possible (since that's quite a bit of code).
164 * texel = border_color;
167 * texel = sample_texture(coord);
169 * As it is now, we always sample the texture, then selectively replace
170 * the texel color results with the border color.
174 /* select texel color or border color depending on use_border */
175 LLVMValueRef border_color_ptr =
176 bld->dynamic_state->border_color(bld->dynamic_state,
179 for (chan = 0; chan < 4; chan++) {
180 LLVMValueRef border_chan =
181 lp_build_array_get(bld->builder, border_color_ptr,
182 lp_build_const_int32(chan));
183 LLVMValueRef border_chan_vec =
184 lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan);
185 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
186 border_chan_vec, texel_out[chan]);
190 apply_sampler_swizzle(bld, texel_out);
195 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
198 lp_build_coord_mirror(struct lp_build_sample_context *bld,
201 struct lp_build_context *coord_bld = &bld->coord_bld;
202 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
203 LLVMValueRef fract, flr, isOdd;
205 lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
207 /* isOdd = flr & 1 */
208 isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
210 /* make coord positive or negative depending on isOdd */
211 coord = lp_build_set_sign(coord_bld, fract, isOdd);
213 /* convert isOdd to float */
214 isOdd = lp_build_int_to_float(coord_bld, isOdd);
216 /* add isOdd to coord */
217 coord = lp_build_add(coord_bld, coord, isOdd);
224 * Build LLVM code for texture wrap mode for linear filtering.
225 * \param x0_out returns first integer texcoord
226 * \param x1_out returns second integer texcoord
227 * \param weight_out returns linear interpolation weight
230 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
233 LLVMValueRef length_f,
236 LLVMValueRef *x0_out,
237 LLVMValueRef *x1_out,
238 LLVMValueRef *weight_out)
240 struct lp_build_context *coord_bld = &bld->coord_bld;
241 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
242 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
243 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
244 LLVMValueRef coord0, coord1, weight;
247 case PIPE_TEX_WRAP_REPEAT:
248 /* mul by size and subtract 0.5 */
249 coord = lp_build_mul(coord_bld, coord, length_f);
250 coord = lp_build_sub(coord_bld, coord, half);
251 /* convert to int, compute lerp weight */
252 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
255 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
256 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
257 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
260 /* Add a bias to the texcoord to handle negative coords */
261 LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024);
263 coord0 = LLVMBuildAdd(bld->builder, coord0, bias, "");
264 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
265 mask = lp_build_compare(bld->builder, int_coord_bld->type,
266 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
267 coord1 = LLVMBuildAnd(bld->builder,
268 lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
273 case PIPE_TEX_WRAP_CLAMP:
274 if (bld->static_state->normalized_coords) {
275 /* scale coord to length */
276 coord = lp_build_mul(coord_bld, coord, length_f);
279 /* clamp to [0, length] */
280 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
282 coord = lp_build_sub(coord_bld, coord, half);
284 /* convert to int, compute lerp weight */
285 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
286 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
289 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
291 struct lp_build_context abs_coord_bld = bld->coord_bld;
292 abs_coord_bld.type.sign = FALSE;
294 if (bld->static_state->normalized_coords) {
295 /* mul by tex size */
296 coord = lp_build_mul(coord_bld, coord, length_f);
297 /* clamp to length max */
298 coord = lp_build_min(coord_bld, coord, length_f);
300 coord = lp_build_sub(coord_bld, coord, half);
301 /* clamp to [0, length - 0.5] */
302 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
304 /* XXX this is odd normalized ranges from 0 to length-0.5 after denorm
305 but non-normalized ranges from to 0.5 to length-0.5 after clamp.
306 Is this missing the sub 0.5? */
308 LLVMValueRef min, max;
309 /* clamp to [0.5, length - 0.5] */
311 max = lp_build_sub(coord_bld, length_f, min);
312 coord = lp_build_clamp(coord_bld, coord, min, max);
314 /* convert to int, compute lerp weight */
315 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
316 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
317 /* coord1 = min(coord1, length-1) */
318 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
322 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
325 if (bld->static_state->normalized_coords) {
326 /* scale coord to length */
327 coord = lp_build_mul(coord_bld, coord, length_f);
329 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
330 coord = lp_build_sub(coord_bld, coord, half);
331 min = lp_build_const_vec(coord_bld->type, -1.0F);
332 coord = lp_build_clamp(coord_bld, coord, min, length_f);
333 /* convert to int, compute lerp weight */
334 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
335 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
339 case PIPE_TEX_WRAP_MIRROR_REPEAT:
340 /* compute mirror function */
341 coord = lp_build_coord_mirror(bld, coord);
343 /* scale coord to length */
344 coord = lp_build_mul(coord_bld, coord, length_f);
345 coord = lp_build_sub(coord_bld, coord, half);
347 /* convert to int, compute lerp weight */
348 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
349 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
351 /* coord0 = max(coord0, 0) */
352 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
353 /* coord1 = min(coord1, length-1) */
354 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
357 case PIPE_TEX_WRAP_MIRROR_CLAMP:
358 coord = lp_build_abs(coord_bld, coord);
360 if (bld->static_state->normalized_coords) {
361 /* scale coord to length */
362 coord = lp_build_mul(coord_bld, coord, length_f);
365 /* clamp to [0, length] */
366 coord = lp_build_min(coord_bld, coord, length_f);
368 coord = lp_build_sub(coord_bld, coord, half);
370 /* convert to int, compute lerp weight */
371 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
372 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
375 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
377 LLVMValueRef min, max;
378 struct lp_build_context abs_coord_bld = bld->coord_bld;
379 abs_coord_bld.type.sign = FALSE;
380 coord = lp_build_abs(coord_bld, coord);
382 if (bld->static_state->normalized_coords) {
383 /* scale coord to length */
384 coord = lp_build_mul(coord_bld, coord, length_f);
387 /* clamp to [0.5, length - 0.5] */
389 max = lp_build_sub(coord_bld, length_f, min);
390 coord = lp_build_clamp(coord_bld, coord, min, max);
392 coord = lp_build_sub(coord_bld, coord, half);
394 /* convert to int, compute lerp weight */
395 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
396 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
400 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
402 coord = lp_build_abs(coord_bld, coord);
404 if (bld->static_state->normalized_coords) {
405 /* scale coord to length */
406 coord = lp_build_mul(coord_bld, coord, length_f);
409 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
410 /* skip -0.5 clamp (always positive), do sub first */
411 coord = lp_build_sub(coord_bld, coord, half);
412 coord = lp_build_min(coord_bld, coord, length_f);
414 /* convert to int, compute lerp weight */
415 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
416 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
429 *weight_out = weight;
434 * Build LLVM code for texture wrap mode for nearest filtering.
435 * \param coord the incoming texcoord (nominally in [0,1])
436 * \param length the texture size along one dimension, as int vector
437 * \param is_pot if TRUE, length is a power of two
438 * \param wrap_mode one of PIPE_TEX_WRAP_x
441 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
444 LLVMValueRef length_f,
448 struct lp_build_context *coord_bld = &bld->coord_bld;
449 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
450 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
454 case PIPE_TEX_WRAP_REPEAT:
455 coord = lp_build_mul(coord_bld, coord, length_f);
456 icoord = lp_build_ifloor(coord_bld, coord);
458 icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
460 /* Add a bias to the texcoord to handle negative coords */
461 LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024);
462 icoord = LLVMBuildAdd(bld->builder, icoord, bias, "");
463 icoord = LLVMBuildURem(bld->builder, icoord, length, "");
467 case PIPE_TEX_WRAP_CLAMP:
468 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
469 if (bld->static_state->normalized_coords) {
470 /* scale coord to length */
471 coord = lp_build_mul(coord_bld, coord, length_f);
475 /* use itrunc instead since we clamp to 0 anyway */
476 icoord = lp_build_itrunc(coord_bld, coord);
478 /* clamp to [0, length - 1]. */
479 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
483 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
484 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
486 LLVMValueRef min, max;
488 if (bld->static_state->normalized_coords) {
489 /* scale coord to length */
490 coord = lp_build_mul(coord_bld, coord, length_f);
493 icoord = lp_build_ifloor(coord_bld, coord);
495 /* clamp to [-1, length] */
496 min = lp_build_negate(int_coord_bld, int_coord_bld->one);
498 icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
502 case PIPE_TEX_WRAP_MIRROR_REPEAT:
503 /* compute mirror function */
504 coord = lp_build_coord_mirror(bld, coord);
506 /* scale coord to length */
507 assert(bld->static_state->normalized_coords);
508 coord = lp_build_mul(coord_bld, coord, length_f);
510 /* itrunc == ifloor here */
511 icoord = lp_build_itrunc(coord_bld, coord);
513 /* clamp to [0, length - 1] */
514 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
517 case PIPE_TEX_WRAP_MIRROR_CLAMP:
518 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
519 coord = lp_build_abs(coord_bld, coord);
521 if (bld->static_state->normalized_coords) {
522 /* scale coord to length */
523 coord = lp_build_mul(coord_bld, coord, length_f);
526 /* itrunc == ifloor here */
527 icoord = lp_build_itrunc(coord_bld, coord);
529 /* clamp to [0, length - 1] */
530 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
533 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
534 coord = lp_build_abs(coord_bld, coord);
536 if (bld->static_state->normalized_coords) {
537 /* scale coord to length */
538 coord = lp_build_mul(coord_bld, coord, length_f);
541 /* itrunc == ifloor here */
542 icoord = lp_build_itrunc(coord_bld, coord);
544 /* clamp to [0, length] */
545 icoord = lp_build_min(int_coord_bld, icoord, length);
558 * Generate code to sample a mipmap level with nearest filtering.
559 * If sampling a cube texture, r = cube face in [0,5].
562 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
565 LLVMValueRef row_stride_vec,
566 LLVMValueRef img_stride_vec,
567 LLVMValueRef data_ptr,
571 LLVMValueRef colors_out[4])
573 const unsigned dims = bld->dims;
574 LLVMValueRef width_vec;
575 LLVMValueRef height_vec;
576 LLVMValueRef depth_vec;
577 LLVMValueRef flt_size;
578 LLVMValueRef flt_width_vec;
579 LLVMValueRef flt_height_vec;
580 LLVMValueRef flt_depth_vec;
581 LLVMValueRef x, y, z;
583 lp_build_extract_image_sizes(bld,
587 &width_vec, &height_vec, &depth_vec);
589 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
591 lp_build_extract_image_sizes(bld,
592 bld->float_size_type,
595 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
598 * Compute integer texcoords.
600 x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec,
601 bld->static_state->pot_width,
602 bld->static_state->wrap_s);
603 lp_build_name(x, "tex.x.wrapped");
606 y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec,
607 bld->static_state->pot_height,
608 bld->static_state->wrap_t);
609 lp_build_name(y, "tex.y.wrapped");
612 z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec,
613 bld->static_state->pot_depth,
614 bld->static_state->wrap_r);
615 lp_build_name(z, "tex.z.wrapped");
617 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
629 * Get texture colors.
631 lp_build_sample_texel_soa(bld, unit,
632 width_vec, height_vec, depth_vec,
634 row_stride_vec, img_stride_vec,
635 data_ptr, colors_out);
640 * Generate code to sample a mipmap level with linear filtering.
641 * If sampling a cube texture, r = cube face in [0,5].
644 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
647 LLVMValueRef row_stride_vec,
648 LLVMValueRef img_stride_vec,
649 LLVMValueRef data_ptr,
653 LLVMValueRef colors_out[4])
655 const unsigned dims = bld->dims;
656 LLVMValueRef width_vec;
657 LLVMValueRef height_vec;
658 LLVMValueRef depth_vec;
659 LLVMValueRef flt_size;
660 LLVMValueRef flt_width_vec;
661 LLVMValueRef flt_height_vec;
662 LLVMValueRef flt_depth_vec;
663 LLVMValueRef x0, y0, z0, x1, y1, z1;
664 LLVMValueRef s_fpart, t_fpart, r_fpart;
665 LLVMValueRef neighbors[2][2][4];
668 lp_build_extract_image_sizes(bld,
672 &width_vec, &height_vec, &depth_vec);
674 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
676 lp_build_extract_image_sizes(bld,
677 bld->float_size_type,
680 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
683 * Compute integer texcoords.
685 lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec,
686 bld->static_state->pot_width,
687 bld->static_state->wrap_s,
689 lp_build_name(x0, "tex.x0.wrapped");
690 lp_build_name(x1, "tex.x1.wrapped");
693 lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec,
694 bld->static_state->pot_height,
695 bld->static_state->wrap_t,
697 lp_build_name(y0, "tex.y0.wrapped");
698 lp_build_name(y1, "tex.y1.wrapped");
701 lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec,
702 bld->static_state->pot_depth,
703 bld->static_state->wrap_r,
705 lp_build_name(z0, "tex.z0.wrapped");
706 lp_build_name(z1, "tex.z1.wrapped");
708 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
709 z0 = z1 = r; /* cube face */
718 y0 = y1 = t_fpart = NULL;
719 z0 = z1 = r_fpart = NULL;
723 * Get texture colors.
725 /* get x0/x1 texels */
726 lp_build_sample_texel_soa(bld, unit,
727 width_vec, height_vec, depth_vec,
729 row_stride_vec, img_stride_vec,
730 data_ptr, neighbors[0][0]);
731 lp_build_sample_texel_soa(bld, unit,
732 width_vec, height_vec, depth_vec,
734 row_stride_vec, img_stride_vec,
735 data_ptr, neighbors[0][1]);
738 /* Interpolate two samples from 1D image to produce one color */
739 for (chan = 0; chan < 4; chan++) {
740 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
741 neighbors[0][0][chan],
742 neighbors[0][1][chan]);
747 LLVMValueRef colors0[4];
749 /* get x0/x1 texels at y1 */
750 lp_build_sample_texel_soa(bld, unit,
751 width_vec, height_vec, depth_vec,
753 row_stride_vec, img_stride_vec,
754 data_ptr, neighbors[1][0]);
755 lp_build_sample_texel_soa(bld, unit,
756 width_vec, height_vec, depth_vec,
758 row_stride_vec, img_stride_vec,
759 data_ptr, neighbors[1][1]);
761 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
762 for (chan = 0; chan < 4; chan++) {
763 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
765 neighbors[0][0][chan],
766 neighbors[0][1][chan],
767 neighbors[1][0][chan],
768 neighbors[1][1][chan]);
772 LLVMValueRef neighbors1[2][2][4];
773 LLVMValueRef colors1[4];
775 /* get x0/x1/y0/y1 texels at z1 */
776 lp_build_sample_texel_soa(bld, unit,
777 width_vec, height_vec, depth_vec,
779 row_stride_vec, img_stride_vec,
780 data_ptr, neighbors1[0][0]);
781 lp_build_sample_texel_soa(bld, unit,
782 width_vec, height_vec, depth_vec,
784 row_stride_vec, img_stride_vec,
785 data_ptr, neighbors1[0][1]);
786 lp_build_sample_texel_soa(bld, unit,
787 width_vec, height_vec, depth_vec,
789 row_stride_vec, img_stride_vec,
790 data_ptr, neighbors1[1][0]);
791 lp_build_sample_texel_soa(bld, unit,
792 width_vec, height_vec, depth_vec,
794 row_stride_vec, img_stride_vec,
795 data_ptr, neighbors1[1][1]);
797 /* Bilinear interpolate the four samples from the second Z slice */
798 for (chan = 0; chan < 4; chan++) {
799 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
801 neighbors1[0][0][chan],
802 neighbors1[0][1][chan],
803 neighbors1[1][0][chan],
804 neighbors1[1][1][chan]);
807 /* Linearly interpolate the two samples from the two 3D slices */
808 for (chan = 0; chan < 4; chan++) {
809 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
811 colors0[chan], colors1[chan]);
816 for (chan = 0; chan < 4; chan++) {
817 colors_out[chan] = colors0[chan];
825 * Sample the texture/mipmap using given image filter and mip filter.
826 * data0_ptr and data1_ptr point to the two mipmap levels to sample
827 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
828 * If we're using nearest miplevel sampling the '1' values will be null/unused.
831 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
838 LLVMValueRef ilevel0,
839 LLVMValueRef ilevel1,
840 LLVMValueRef lod_fpart,
841 LLVMValueRef *colors_out)
843 LLVMBuilderRef builder = bld->builder;
846 LLVMValueRef row_stride0_vec;
847 LLVMValueRef row_stride1_vec;
848 LLVMValueRef img_stride0_vec;
849 LLVMValueRef img_stride1_vec;
850 LLVMValueRef data_ptr0;
851 LLVMValueRef data_ptr1;
852 LLVMValueRef colors0[4], colors1[4];
855 /* sample the first mipmap level */
856 lp_build_mipmap_level_sizes(bld, ilevel0,
858 &row_stride0_vec, &img_stride0_vec);
859 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
860 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
861 lp_build_sample_image_nearest(bld, unit,
863 row_stride0_vec, img_stride0_vec,
868 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
869 lp_build_sample_image_linear(bld, unit,
871 row_stride0_vec, img_stride0_vec,
876 /* Store the first level's colors in the output variables */
877 for (chan = 0; chan < 4; chan++) {
878 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
881 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
882 struct lp_build_if_state if_ctx;
883 LLVMValueRef need_lerp;
885 /* need_lerp = lod_fpart > 0 */
886 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
891 lp_build_if(&if_ctx, builder, need_lerp);
893 /* sample the second mipmap level */
894 lp_build_mipmap_level_sizes(bld, ilevel1,
896 &row_stride1_vec, &img_stride1_vec);
897 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
898 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
899 lp_build_sample_image_nearest(bld, unit,
901 row_stride1_vec, img_stride1_vec,
906 lp_build_sample_image_linear(bld, unit,
908 row_stride1_vec, img_stride1_vec,
913 /* interpolate samples from the two mipmap levels */
915 lod_fpart = lp_build_broadcast_scalar(&bld->texel_bld, lod_fpart);
917 for (chan = 0; chan < 4; chan++) {
918 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
919 colors0[chan], colors1[chan]);
920 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
923 lp_build_endif(&if_ctx);
930 * General texture sampling codegen.
931 * This function handles texture sampling for all texture targets (1D,
932 * 2D, 3D, cube) and all filtering modes.
935 lp_build_sample_general(struct lp_build_sample_context *bld,
940 const LLVMValueRef *ddx,
941 const LLVMValueRef *ddy,
942 LLVMValueRef lod_bias, /* optional */
943 LLVMValueRef explicit_lod, /* optional */
944 LLVMValueRef *colors_out)
946 struct lp_build_context *int_bld = &bld->int_bld;
947 LLVMBuilderRef builder = bld->builder;
948 const unsigned mip_filter = bld->static_state->min_mip_filter;
949 const unsigned min_filter = bld->static_state->min_img_filter;
950 const unsigned mag_filter = bld->static_state->mag_img_filter;
951 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
952 LLVMValueRef ilevel0, ilevel1 = NULL;
953 LLVMValueRef face_ddx[4], face_ddy[4];
954 LLVMValueRef texels[4];
955 LLVMTypeRef i32t = LLVMInt32Type();
956 LLVMValueRef i32t_zero = LLVMConstInt(i32t, 0, 0);
960 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
961 mip_filter, min_filter, mag_filter);
965 * Choose cube face, recompute texcoords and derivatives for the chosen face.
967 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
968 LLVMValueRef face, face_s, face_t;
969 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
970 s = face_s; /* vec */
971 t = face_t; /* vec */
972 /* use 'r' to indicate cube face */
973 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
975 /* recompute ddx, ddy using the new (s,t) face texcoords */
976 face_ddx[0] = lp_build_ddx(&bld->coord_bld, s);
977 face_ddx[1] = lp_build_ddx(&bld->coord_bld, t);
980 face_ddy[0] = lp_build_ddy(&bld->coord_bld, s);
981 face_ddy[1] = lp_build_ddy(&bld->coord_bld, t);
989 * Compute the level of detail (float).
991 if (min_filter != mag_filter ||
992 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
993 /* Need to compute lod either to choose mipmap levels or to
994 * distinguish between minification/magnification with one mipmap level.
996 lp_build_lod_selector(bld, unit, ddx, ddy,
997 lod_bias, explicit_lod,
999 &lod_ipart, &lod_fpart);
1001 lod_ipart = i32t_zero;
1005 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1007 switch (mip_filter) {
1009 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1011 case PIPE_TEX_MIPFILTER_NONE:
1012 /* always use mip level 0 */
1013 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1014 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1015 * We should be able to set ilevel0 = const(0) but that causes
1016 * bad x86 code to be emitted.
1019 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
1022 ilevel0 = i32t_zero;
1025 case PIPE_TEX_MIPFILTER_NEAREST:
1027 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
1029 case PIPE_TEX_MIPFILTER_LINEAR:
1032 lp_build_linear_mip_levels(bld, unit,
1033 lod_ipart, &lod_fpart,
1034 &ilevel0, &ilevel1);
1039 * Get/interpolate texture colors.
1042 for (chan = 0; chan < 4; ++chan) {
1043 texels[chan] = lp_build_alloca(builder, bld->texel_bld.vec_type, "");
1044 lp_build_name(texels[chan], "sampler%u_texel_%c_var", unit, "xyzw"[chan]);
1047 if (min_filter == mag_filter) {
1048 /* no need to distinquish between minification and magnification */
1049 lp_build_sample_mipmap(bld, unit,
1050 min_filter, mip_filter,
1052 ilevel0, ilevel1, lod_fpart,
1056 /* Emit conditional to choose min image filter or mag image filter
1057 * depending on the lod being > 0 or <= 0, respectively.
1059 struct lp_build_if_state if_ctx;
1060 LLVMValueRef minify;
1062 /* minify = lod >= 0.0 */
1063 minify = LLVMBuildICmp(builder, LLVMIntSGE,
1064 lod_ipart, int_bld->zero, "");
1066 lp_build_if(&if_ctx, builder, minify);
1068 /* Use the minification filter */
1069 lp_build_sample_mipmap(bld, unit,
1070 min_filter, mip_filter,
1072 ilevel0, ilevel1, lod_fpart,
1075 lp_build_else(&if_ctx);
1077 /* Use the magnification filter */
1078 lp_build_sample_mipmap(bld, unit,
1079 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1081 i32t_zero, NULL, NULL,
1084 lp_build_endif(&if_ctx);
1087 for (chan = 0; chan < 4; ++chan) {
1088 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1089 lp_build_name(colors_out[chan], "sampler%u_texel_%c", unit, "xyzw"[chan]);
1095 * Do shadow test/comparison.
1096 * \param p the texcoord Z (aka R, aka P) component
1097 * \param texel the texel to compare against (use the X channel)
1100 lp_build_sample_compare(struct lp_build_sample_context *bld,
1102 LLVMValueRef texel[4])
1104 struct lp_build_context *texel_bld = &bld->texel_bld;
1106 const unsigned chan = 0;
1108 if (bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1113 LLVMValueRef indx = lp_build_const_int32(0);
1114 LLVMValueRef coord = LLVMBuildExtractElement(bld->builder, p, indx, "");
1115 LLVMValueRef tex = LLVMBuildExtractElement(bld->builder,
1116 texel[chan], indx, "");
1117 lp_build_printf(bld->builder, "shadow compare coord %f to texture %f\n",
1121 /* result = (p FUNC texel) ? 1 : 0 */
1122 res = lp_build_cmp(texel_bld, bld->static_state->compare_func,
1124 res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1126 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1130 texel[3] = texel_bld->one;
1135 * Just set texels to white instead of actually sampling the texture.
1139 lp_build_sample_nop(struct lp_type type,
1140 LLVMValueRef texel_out[4])
1142 LLVMValueRef one = lp_build_one(type);
1145 for (chan = 0; chan < 4; chan++) {
1146 texel_out[chan] = one;
1152 * Build texture sampling code.
1153 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1155 * \param type vector float type to use for coords, etc.
1156 * \param ddx partial derivatives of (s,t,r,q) with respect to x
1157 * \param ddy partial derivatives of (s,t,r,q) with respect to y
1160 lp_build_sample_soa(LLVMBuilderRef builder,
1161 const struct lp_sampler_static_state *static_state,
1162 struct lp_sampler_dynamic_state *dynamic_state,
1163 struct lp_type type,
1165 unsigned num_coords,
1166 const LLVMValueRef *coords,
1167 const LLVMValueRef ddx[4],
1168 const LLVMValueRef ddy[4],
1169 LLVMValueRef lod_bias, /* optional */
1170 LLVMValueRef explicit_lod, /* optional */
1171 LLVMValueRef texel_out[4])
1173 unsigned dims = texture_dims(static_state->target);
1174 struct lp_build_sample_context bld;
1175 LLVMTypeRef i32t = LLVMInt32Type();
1180 struct lp_type float_vec_type;
1183 enum pipe_format fmt = static_state->format;
1184 debug_printf("Sample from %s\n", util_format_name(fmt));
1187 assert(type.floating);
1189 /* Setup our build context */
1190 memset(&bld, 0, sizeof bld);
1191 bld.builder = builder;
1192 bld.static_state = static_state;
1193 bld.dynamic_state = dynamic_state;
1194 bld.format_desc = util_format_description(static_state->format);
1197 bld.float_type = lp_type_float(32);
1198 bld.int_type = lp_type_int(32);
1199 bld.coord_type = type;
1200 bld.int_coord_type = lp_int_type(type);
1201 bld.float_size_type = lp_type_float(32);
1202 bld.float_size_type.length = dims > 1 ? 4 : 1;
1203 bld.int_size_type = lp_int_type(bld.float_size_type);
1204 bld.texel_type = type;
1206 float_vec_type = lp_type_float_vec(32);
1208 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
1209 lp_build_context_init(&bld.float_vec_bld, builder, float_vec_type);
1210 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
1211 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
1212 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
1213 lp_build_context_init(&bld.int_size_bld, builder, bld.int_size_type);
1214 lp_build_context_init(&bld.float_size_bld, builder, bld.float_size_type);
1215 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
1217 /* Get the dynamic state */
1218 bld.width = dynamic_state->width(dynamic_state, builder, unit);
1219 bld.height = dynamic_state->height(dynamic_state, builder, unit);
1220 bld.depth = dynamic_state->depth(dynamic_state, builder, unit);
1221 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
1222 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit);
1223 bld.data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
1224 /* Note that data_array is an array[level] of pointers to texture images */
1230 /* width, height, depth as single int vector */
1232 bld.int_size = bld.width;
1235 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_bld.undef,
1236 bld.width, LLVMConstInt(i32t, 0, 0), "");
1238 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1239 bld.height, LLVMConstInt(i32t, 1, 0), "");
1241 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1242 bld.depth, LLVMConstInt(i32t, 2, 0), "");
1248 /* For debug: no-op texture sampling */
1249 lp_build_sample_nop(bld.texel_type, texel_out);
1251 else if (util_format_fits_8unorm(bld.format_desc) &&
1252 lp_is_simple_wrap_mode(static_state->wrap_s) &&
1253 lp_is_simple_wrap_mode(static_state->wrap_t)) {
1254 /* do sampling/filtering with fixed pt arithmetic */
1255 lp_build_sample_aos(&bld, unit, s, t, r, ddx, ddy,
1256 lod_bias, explicit_lod,
1261 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1262 util_format_fits_8unorm(bld.format_desc)) {
1263 debug_printf("%s: using floating point linear filtering for %s\n",
1264 __FUNCTION__, bld.format_desc->short_name);
1265 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1266 static_state->min_img_filter,
1267 static_state->mag_img_filter,
1268 static_state->min_mip_filter,
1269 static_state->wrap_s,
1270 static_state->wrap_t);
1273 lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
1274 lod_bias, explicit_lod,
1278 lp_build_sample_compare(&bld, r, texel_out);