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;
87 LLVMBuilderRef builder = bld->gallivm->builder;
90 LLVMValueRef use_border = NULL;
92 /* use_border = x < 0 || x >= width || y < 0 || y >= height */
93 if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
94 static_state->min_img_filter,
95 static_state->mag_img_filter)) {
97 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
98 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
99 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
103 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
104 static_state->min_img_filter,
105 static_state->mag_img_filter)) {
107 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
108 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
110 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
111 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
114 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
119 lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
120 static_state->min_img_filter,
121 static_state->mag_img_filter)) {
123 b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
124 b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
126 use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
127 use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
130 use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
134 /* convert x,y,z coords to linear offset from start of texture, in bytes */
135 lp_build_sample_offset(&bld->int_coord_bld,
137 x, y, z, y_stride, z_stride,
141 /* If we can sample the border color, it means that texcoords may
142 * lie outside the bounds of the texture image. We need to do
143 * something to prevent reading out of bounds and causing a segfault.
145 * Simply AND the texture coords with !use_border. This will cause
146 * coords which are out of bounds to become zero. Zero's guaranteed
147 * to be inside the texture image.
149 offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
152 lp_build_fetch_rgba_soa(bld->gallivm,
160 * Note: if we find an app which frequently samples the texture border
161 * we might want to implement a true conditional here to avoid sampling
162 * the texture whenever possible (since that's quite a bit of code).
165 * texel = border_color;
168 * texel = sample_texture(coord);
170 * As it is now, we always sample the texture, then selectively replace
171 * the texel color results with the border color.
175 /* select texel color or border color depending on use_border */
176 LLVMValueRef border_color_ptr =
177 bld->dynamic_state->border_color(bld->dynamic_state,
180 for (chan = 0; chan < 4; chan++) {
181 LLVMValueRef border_chan =
182 lp_build_array_get(bld->gallivm, border_color_ptr,
183 lp_build_const_int32(bld->gallivm, chan));
184 LLVMValueRef border_chan_vec =
185 lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan);
186 texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
187 border_chan_vec, texel_out[chan]);
194 * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
197 lp_build_coord_mirror(struct lp_build_sample_context *bld,
200 struct lp_build_context *coord_bld = &bld->coord_bld;
201 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
202 LLVMValueRef fract, flr, isOdd;
204 lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
206 /* isOdd = flr & 1 */
207 isOdd = LLVMBuildAnd(bld->gallivm->builder, flr, int_coord_bld->one, "");
209 /* make coord positive or negative depending on isOdd */
210 coord = lp_build_set_sign(coord_bld, fract, isOdd);
212 /* convert isOdd to float */
213 isOdd = lp_build_int_to_float(coord_bld, isOdd);
215 /* add isOdd to coord */
216 coord = lp_build_add(coord_bld, coord, isOdd);
223 * Build LLVM code for texture wrap mode for linear filtering.
224 * \param x0_out returns first integer texcoord
225 * \param x1_out returns second integer texcoord
226 * \param weight_out returns linear interpolation weight
229 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
232 LLVMValueRef length_f,
235 LLVMValueRef *x0_out,
236 LLVMValueRef *x1_out,
237 LLVMValueRef *weight_out)
239 struct lp_build_context *coord_bld = &bld->coord_bld;
240 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
241 LLVMBuilderRef builder = bld->gallivm->builder;
242 LLVMValueRef half = lp_build_const_vec(bld->gallivm, 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(builder, coord0, length_minus_one, "");
257 coord1 = LLVMBuildAnd(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(builder, coord0, bias, "");
264 coord0 = LLVMBuildURem(builder, coord0, length, "");
265 mask = lp_build_compare(bld->gallivm, int_coord_bld->type,
266 PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
267 coord1 = LLVMBuildAnd(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);
298 /* clamp to length max */
299 coord = lp_build_min(coord_bld, coord, length_f);
301 coord = lp_build_sub(coord_bld, coord, half);
302 /* clamp to [0, length - 0.5] */
303 coord = lp_build_max(coord_bld, coord, coord_bld->zero);
304 /* convert to int, compute lerp weight */
305 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
306 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
307 /* coord1 = min(coord1, length-1) */
308 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
312 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
315 if (bld->static_state->normalized_coords) {
316 /* scale coord to length */
317 coord = lp_build_mul(coord_bld, coord, length_f);
319 /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
320 coord = lp_build_sub(coord_bld, coord, half);
321 min = lp_build_const_vec(bld->gallivm, coord_bld->type, -1.0F);
322 coord = lp_build_clamp(coord_bld, coord, min, length_f);
323 /* convert to int, compute lerp weight */
324 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
325 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
329 case PIPE_TEX_WRAP_MIRROR_REPEAT:
330 /* compute mirror function */
331 coord = lp_build_coord_mirror(bld, coord);
333 /* scale coord to length */
334 coord = lp_build_mul(coord_bld, coord, length_f);
335 coord = lp_build_sub(coord_bld, coord, half);
337 /* convert to int, compute lerp weight */
338 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
339 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
341 /* coord0 = max(coord0, 0) */
342 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
343 /* coord1 = min(coord1, length-1) */
344 coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
347 case PIPE_TEX_WRAP_MIRROR_CLAMP:
348 coord = lp_build_abs(coord_bld, coord);
350 if (bld->static_state->normalized_coords) {
351 /* scale coord to length */
352 coord = lp_build_mul(coord_bld, coord, length_f);
355 /* clamp to [0, length] */
356 coord = lp_build_min(coord_bld, coord, length_f);
358 coord = lp_build_sub(coord_bld, coord, half);
360 /* convert to int, compute lerp weight */
361 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
362 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
365 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
367 LLVMValueRef min, max;
368 struct lp_build_context abs_coord_bld = bld->coord_bld;
369 abs_coord_bld.type.sign = FALSE;
370 coord = lp_build_abs(coord_bld, coord);
372 if (bld->static_state->normalized_coords) {
373 /* scale coord to length */
374 coord = lp_build_mul(coord_bld, coord, length_f);
377 /* clamp to [0.5, length - 0.5] */
379 max = lp_build_sub(coord_bld, length_f, min);
380 coord = lp_build_clamp(coord_bld, coord, min, max);
382 coord = lp_build_sub(coord_bld, coord, half);
384 /* convert to int, compute lerp weight */
385 lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
386 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
390 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
392 coord = lp_build_abs(coord_bld, coord);
394 if (bld->static_state->normalized_coords) {
395 /* scale coord to length */
396 coord = lp_build_mul(coord_bld, coord, length_f);
399 /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
400 /* skip -0.5 clamp (always positive), do sub first */
401 coord = lp_build_sub(coord_bld, coord, half);
402 coord = lp_build_min(coord_bld, coord, length_f);
404 /* convert to int, compute lerp weight */
405 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
406 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
419 *weight_out = weight;
424 * Build LLVM code for texture wrap mode for nearest filtering.
425 * \param coord the incoming texcoord (nominally in [0,1])
426 * \param length the texture size along one dimension, as int vector
427 * \param is_pot if TRUE, length is a power of two
428 * \param wrap_mode one of PIPE_TEX_WRAP_x
431 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
434 LLVMValueRef length_f,
438 struct lp_build_context *coord_bld = &bld->coord_bld;
439 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
440 LLVMBuilderRef builder = bld->gallivm->builder;
441 LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
445 case PIPE_TEX_WRAP_REPEAT:
446 coord = lp_build_mul(coord_bld, coord, length_f);
447 icoord = lp_build_ifloor(coord_bld, coord);
449 icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
451 /* Add a bias to the texcoord to handle negative coords */
452 LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024);
453 icoord = LLVMBuildAdd(builder, icoord, bias, "");
454 icoord = LLVMBuildURem(builder, icoord, length, "");
458 case PIPE_TEX_WRAP_CLAMP:
459 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
460 if (bld->static_state->normalized_coords) {
461 /* scale coord to length */
462 coord = lp_build_mul(coord_bld, coord, length_f);
466 /* use itrunc instead since we clamp to 0 anyway */
467 icoord = lp_build_itrunc(coord_bld, coord);
469 /* clamp to [0, length - 1]. */
470 icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
474 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
475 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
477 LLVMValueRef min, max;
479 if (bld->static_state->normalized_coords) {
480 /* scale coord to length */
481 coord = lp_build_mul(coord_bld, coord, length_f);
484 icoord = lp_build_ifloor(coord_bld, coord);
486 /* clamp to [-1, length] */
487 min = lp_build_negate(int_coord_bld, int_coord_bld->one);
489 icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
493 case PIPE_TEX_WRAP_MIRROR_REPEAT:
494 /* compute mirror function */
495 coord = lp_build_coord_mirror(bld, coord);
497 /* scale coord to length */
498 assert(bld->static_state->normalized_coords);
499 coord = lp_build_mul(coord_bld, coord, length_f);
501 /* itrunc == ifloor here */
502 icoord = lp_build_itrunc(coord_bld, coord);
504 /* clamp to [0, length - 1] */
505 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
508 case PIPE_TEX_WRAP_MIRROR_CLAMP:
509 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
510 coord = lp_build_abs(coord_bld, coord);
512 if (bld->static_state->normalized_coords) {
513 /* scale coord to length */
514 coord = lp_build_mul(coord_bld, coord, length_f);
517 /* itrunc == ifloor here */
518 icoord = lp_build_itrunc(coord_bld, coord);
520 /* clamp to [0, length - 1] */
521 icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
524 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
525 coord = lp_build_abs(coord_bld, coord);
527 if (bld->static_state->normalized_coords) {
528 /* scale coord to length */
529 coord = lp_build_mul(coord_bld, coord, length_f);
532 /* itrunc == ifloor here */
533 icoord = lp_build_itrunc(coord_bld, coord);
535 /* clamp to [0, length] */
536 icoord = lp_build_min(int_coord_bld, icoord, length);
549 * Generate code to sample a mipmap level with nearest filtering.
550 * If sampling a cube texture, r = cube face in [0,5].
553 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
556 LLVMValueRef row_stride_vec,
557 LLVMValueRef img_stride_vec,
558 LLVMValueRef data_ptr,
562 LLVMValueRef colors_out[4])
564 const unsigned dims = bld->dims;
565 LLVMValueRef width_vec;
566 LLVMValueRef height_vec;
567 LLVMValueRef depth_vec;
568 LLVMValueRef flt_size;
569 LLVMValueRef flt_width_vec;
570 LLVMValueRef flt_height_vec;
571 LLVMValueRef flt_depth_vec;
572 LLVMValueRef x, y, z;
574 lp_build_extract_image_sizes(bld,
578 &width_vec, &height_vec, &depth_vec);
580 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
582 lp_build_extract_image_sizes(bld,
583 bld->float_size_type,
586 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
589 * Compute integer texcoords.
591 x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec,
592 bld->static_state->pot_width,
593 bld->static_state->wrap_s);
594 lp_build_name(x, "tex.x.wrapped");
597 y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec,
598 bld->static_state->pot_height,
599 bld->static_state->wrap_t);
600 lp_build_name(y, "tex.y.wrapped");
603 z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec,
604 bld->static_state->pot_depth,
605 bld->static_state->wrap_r);
606 lp_build_name(z, "tex.z.wrapped");
608 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
620 * Get texture colors.
622 lp_build_sample_texel_soa(bld, unit,
623 width_vec, height_vec, depth_vec,
625 row_stride_vec, img_stride_vec,
626 data_ptr, colors_out);
631 * Generate code to sample a mipmap level with linear filtering.
632 * If sampling a cube texture, r = cube face in [0,5].
635 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
638 LLVMValueRef row_stride_vec,
639 LLVMValueRef img_stride_vec,
640 LLVMValueRef data_ptr,
644 LLVMValueRef colors_out[4])
646 const unsigned dims = bld->dims;
647 LLVMValueRef width_vec;
648 LLVMValueRef height_vec;
649 LLVMValueRef depth_vec;
650 LLVMValueRef flt_size;
651 LLVMValueRef flt_width_vec;
652 LLVMValueRef flt_height_vec;
653 LLVMValueRef flt_depth_vec;
654 LLVMValueRef x0, y0, z0, x1, y1, z1;
655 LLVMValueRef s_fpart, t_fpart, r_fpart;
656 LLVMValueRef neighbors[2][2][4];
659 lp_build_extract_image_sizes(bld,
663 &width_vec, &height_vec, &depth_vec);
665 flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
667 lp_build_extract_image_sizes(bld,
668 bld->float_size_type,
671 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
674 * Compute integer texcoords.
676 lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec,
677 bld->static_state->pot_width,
678 bld->static_state->wrap_s,
680 lp_build_name(x0, "tex.x0.wrapped");
681 lp_build_name(x1, "tex.x1.wrapped");
684 lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec,
685 bld->static_state->pot_height,
686 bld->static_state->wrap_t,
688 lp_build_name(y0, "tex.y0.wrapped");
689 lp_build_name(y1, "tex.y1.wrapped");
692 lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec,
693 bld->static_state->pot_depth,
694 bld->static_state->wrap_r,
696 lp_build_name(z0, "tex.z0.wrapped");
697 lp_build_name(z1, "tex.z1.wrapped");
699 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
700 z0 = z1 = r; /* cube face */
709 y0 = y1 = t_fpart = NULL;
710 z0 = z1 = r_fpart = NULL;
714 * Get texture colors.
716 /* get x0/x1 texels */
717 lp_build_sample_texel_soa(bld, unit,
718 width_vec, height_vec, depth_vec,
720 row_stride_vec, img_stride_vec,
721 data_ptr, neighbors[0][0]);
722 lp_build_sample_texel_soa(bld, unit,
723 width_vec, height_vec, depth_vec,
725 row_stride_vec, img_stride_vec,
726 data_ptr, neighbors[0][1]);
729 /* Interpolate two samples from 1D image to produce one color */
730 for (chan = 0; chan < 4; chan++) {
731 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
732 neighbors[0][0][chan],
733 neighbors[0][1][chan]);
738 LLVMValueRef colors0[4];
740 /* get x0/x1 texels at y1 */
741 lp_build_sample_texel_soa(bld, unit,
742 width_vec, height_vec, depth_vec,
744 row_stride_vec, img_stride_vec,
745 data_ptr, neighbors[1][0]);
746 lp_build_sample_texel_soa(bld, unit,
747 width_vec, height_vec, depth_vec,
749 row_stride_vec, img_stride_vec,
750 data_ptr, neighbors[1][1]);
752 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
753 for (chan = 0; chan < 4; chan++) {
754 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
756 neighbors[0][0][chan],
757 neighbors[0][1][chan],
758 neighbors[1][0][chan],
759 neighbors[1][1][chan]);
763 LLVMValueRef neighbors1[2][2][4];
764 LLVMValueRef colors1[4];
766 /* get x0/x1/y0/y1 texels at z1 */
767 lp_build_sample_texel_soa(bld, unit,
768 width_vec, height_vec, depth_vec,
770 row_stride_vec, img_stride_vec,
771 data_ptr, neighbors1[0][0]);
772 lp_build_sample_texel_soa(bld, unit,
773 width_vec, height_vec, depth_vec,
775 row_stride_vec, img_stride_vec,
776 data_ptr, neighbors1[0][1]);
777 lp_build_sample_texel_soa(bld, unit,
778 width_vec, height_vec, depth_vec,
780 row_stride_vec, img_stride_vec,
781 data_ptr, neighbors1[1][0]);
782 lp_build_sample_texel_soa(bld, unit,
783 width_vec, height_vec, depth_vec,
785 row_stride_vec, img_stride_vec,
786 data_ptr, neighbors1[1][1]);
788 /* Bilinear interpolate the four samples from the second Z slice */
789 for (chan = 0; chan < 4; chan++) {
790 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
792 neighbors1[0][0][chan],
793 neighbors1[0][1][chan],
794 neighbors1[1][0][chan],
795 neighbors1[1][1][chan]);
798 /* Linearly interpolate the two samples from the two 3D slices */
799 for (chan = 0; chan < 4; chan++) {
800 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
802 colors0[chan], colors1[chan]);
807 for (chan = 0; chan < 4; chan++) {
808 colors_out[chan] = colors0[chan];
816 * Sample the texture/mipmap using given image filter and mip filter.
817 * data0_ptr and data1_ptr point to the two mipmap levels to sample
818 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
819 * If we're using nearest miplevel sampling the '1' values will be null/unused.
822 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
829 LLVMValueRef ilevel0,
830 LLVMValueRef ilevel1,
831 LLVMValueRef lod_fpart,
832 LLVMValueRef *colors_out)
834 LLVMBuilderRef builder = bld->gallivm->builder;
835 LLVMValueRef size0 = NULL;
836 LLVMValueRef size1 = NULL;
837 LLVMValueRef row_stride0_vec = NULL;
838 LLVMValueRef row_stride1_vec = NULL;
839 LLVMValueRef img_stride0_vec = NULL;
840 LLVMValueRef img_stride1_vec = NULL;
841 LLVMValueRef data_ptr0 = NULL;
842 LLVMValueRef data_ptr1 = NULL;
843 LLVMValueRef colors0[4], colors1[4];
846 /* sample the first mipmap level */
847 lp_build_mipmap_level_sizes(bld, ilevel0,
849 &row_stride0_vec, &img_stride0_vec);
850 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
851 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
852 lp_build_sample_image_nearest(bld, unit,
854 row_stride0_vec, img_stride0_vec,
859 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
860 lp_build_sample_image_linear(bld, unit,
862 row_stride0_vec, img_stride0_vec,
867 /* Store the first level's colors in the output variables */
868 for (chan = 0; chan < 4; chan++) {
869 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
872 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
873 struct lp_build_if_state if_ctx;
874 LLVMValueRef need_lerp;
876 /* need_lerp = lod_fpart > 0 */
877 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
882 lp_build_if(&if_ctx, bld->gallivm, need_lerp);
884 /* sample the second mipmap level */
885 lp_build_mipmap_level_sizes(bld, ilevel1,
887 &row_stride1_vec, &img_stride1_vec);
888 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
889 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
890 lp_build_sample_image_nearest(bld, unit,
892 row_stride1_vec, img_stride1_vec,
897 lp_build_sample_image_linear(bld, unit,
899 row_stride1_vec, img_stride1_vec,
904 /* interpolate samples from the two mipmap levels */
906 lod_fpart = lp_build_broadcast_scalar(&bld->texel_bld, lod_fpart);
908 for (chan = 0; chan < 4; chan++) {
909 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
910 colors0[chan], colors1[chan]);
911 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
914 lp_build_endif(&if_ctx);
921 * General texture sampling codegen.
922 * This function handles texture sampling for all texture targets (1D,
923 * 2D, 3D, cube) and all filtering modes.
926 lp_build_sample_general(struct lp_build_sample_context *bld,
931 const LLVMValueRef *ddx,
932 const LLVMValueRef *ddy,
933 LLVMValueRef lod_bias, /* optional */
934 LLVMValueRef explicit_lod, /* optional */
935 LLVMValueRef *colors_out)
937 struct lp_build_context *int_bld = &bld->int_bld;
938 LLVMBuilderRef builder = bld->gallivm->builder;
939 const unsigned mip_filter = bld->static_state->min_mip_filter;
940 const unsigned min_filter = bld->static_state->min_img_filter;
941 const unsigned mag_filter = bld->static_state->mag_img_filter;
942 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
943 LLVMValueRef ilevel0, ilevel1 = NULL;
944 LLVMValueRef face_ddx[4], face_ddy[4];
945 LLVMValueRef texels[4];
946 LLVMValueRef first_level;
947 LLVMValueRef i32t_zero = lp_build_const_int32(bld->gallivm, 0);
951 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
952 mip_filter, min_filter, mag_filter);
956 * Choose cube face, recompute texcoords and derivatives for the chosen face.
958 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
959 LLVMValueRef face, face_s, face_t;
960 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
961 s = face_s; /* vec */
962 t = face_t; /* vec */
963 /* use 'r' to indicate cube face */
964 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
966 /* recompute ddx, ddy using the new (s,t) face texcoords */
967 face_ddx[0] = lp_build_scalar_ddx(&bld->coord_bld, s);
968 face_ddx[1] = lp_build_scalar_ddx(&bld->coord_bld, t);
971 face_ddy[0] = lp_build_scalar_ddy(&bld->coord_bld, s);
972 face_ddy[1] = lp_build_scalar_ddy(&bld->coord_bld, t);
980 * Compute the level of detail (float).
982 if (min_filter != mag_filter ||
983 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
984 /* Need to compute lod either to choose mipmap levels or to
985 * distinguish between minification/magnification with one mipmap level.
987 lp_build_lod_selector(bld, unit, ddx, ddy,
988 lod_bias, explicit_lod,
990 &lod_ipart, &lod_fpart);
992 lod_ipart = i32t_zero;
996 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
998 switch (mip_filter) {
1000 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1002 case PIPE_TEX_MIPFILTER_NONE:
1003 /* always use mip level 0 */
1004 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1005 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1006 * We should be able to set ilevel0 = const(0) but that causes
1007 * bad x86 code to be emitted.
1010 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
1013 first_level = bld->dynamic_state->first_level(bld->dynamic_state,
1014 bld->gallivm, unit);
1015 ilevel0 = first_level;
1018 case PIPE_TEX_MIPFILTER_NEAREST:
1020 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
1022 case PIPE_TEX_MIPFILTER_LINEAR:
1025 lp_build_linear_mip_levels(bld, unit,
1026 lod_ipart, &lod_fpart,
1027 &ilevel0, &ilevel1);
1032 * Get/interpolate texture colors.
1035 for (chan = 0; chan < 4; ++chan) {
1036 texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
1037 lp_build_name(texels[chan], "sampler%u_texel_%c_var", unit, "xyzw"[chan]);
1040 if (min_filter == mag_filter) {
1041 /* no need to distinquish between minification and magnification */
1042 lp_build_sample_mipmap(bld, unit,
1043 min_filter, mip_filter,
1045 ilevel0, ilevel1, lod_fpart,
1049 /* Emit conditional to choose min image filter or mag image filter
1050 * depending on the lod being > 0 or <= 0, respectively.
1052 struct lp_build_if_state if_ctx;
1053 LLVMValueRef minify;
1055 /* minify = lod >= 0.0 */
1056 minify = LLVMBuildICmp(builder, LLVMIntSGE,
1057 lod_ipart, int_bld->zero, "");
1059 lp_build_if(&if_ctx, bld->gallivm, minify);
1061 /* Use the minification filter */
1062 lp_build_sample_mipmap(bld, unit,
1063 min_filter, mip_filter,
1065 ilevel0, ilevel1, lod_fpart,
1068 lp_build_else(&if_ctx);
1070 /* Use the magnification filter */
1071 lp_build_sample_mipmap(bld, unit,
1072 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1074 ilevel0, NULL, NULL,
1077 lp_build_endif(&if_ctx);
1080 for (chan = 0; chan < 4; ++chan) {
1081 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1082 lp_build_name(colors_out[chan], "sampler%u_texel_%c", unit, "xyzw"[chan]);
1088 * Do shadow test/comparison.
1089 * \param p the texcoord Z (aka R, aka P) component
1090 * \param texel the texel to compare against (use the X channel)
1093 lp_build_sample_compare(struct lp_build_sample_context *bld,
1095 LLVMValueRef texel[4])
1097 struct lp_build_context *texel_bld = &bld->texel_bld;
1098 LLVMBuilderRef builder = bld->gallivm->builder;
1100 const unsigned chan = 0;
1102 if (bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1107 LLVMValueRef indx = lp_build_const_int32(bld->gallivm, 0);
1108 LLVMValueRef coord = LLVMBuildExtractElement(builder, p, indx, "");
1109 LLVMValueRef tex = LLVMBuildExtractElement(builder, texel[chan], indx, "");
1110 lp_build_printf(bld->gallivm, "shadow compare coord %f to texture %f\n",
1114 /* Clamp p coords to [0,1] */
1115 p = lp_build_clamp(&bld->coord_bld, p,
1116 bld->coord_bld.zero,
1117 bld->coord_bld.one);
1119 /* result = (p FUNC texel) ? 1 : 0 */
1120 res = lp_build_cmp(texel_bld, bld->static_state->compare_func,
1122 res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1124 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1128 texel[3] = texel_bld->one;
1133 * Just set texels to white instead of actually sampling the texture.
1137 lp_build_sample_nop(struct gallivm_state *gallivm, struct lp_type type,
1138 LLVMValueRef texel_out[4])
1140 LLVMValueRef one = lp_build_one(gallivm, type);
1143 for (chan = 0; chan < 4; chan++) {
1144 texel_out[chan] = one;
1150 * Build texture sampling code.
1151 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1153 * \param type vector float type to use for coords, etc.
1154 * \param ddx partial derivatives of (s,t,r,q) with respect to x
1155 * \param ddy partial derivatives of (s,t,r,q) with respect to y
1158 lp_build_sample_soa(struct gallivm_state *gallivm,
1159 const struct lp_sampler_static_state *static_state,
1160 struct lp_sampler_dynamic_state *dynamic_state,
1161 struct lp_type type,
1163 unsigned num_coords,
1164 const LLVMValueRef *coords,
1165 const LLVMValueRef ddx[4],
1166 const LLVMValueRef ddy[4],
1167 LLVMValueRef lod_bias, /* optional */
1168 LLVMValueRef explicit_lod, /* optional */
1169 LLVMValueRef texel_out[4])
1171 unsigned dims = texture_dims(static_state->target);
1172 struct lp_build_sample_context bld;
1173 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
1174 LLVMBuilderRef builder = gallivm->builder;
1178 struct lp_type float_vec_type;
1181 enum pipe_format fmt = static_state->format;
1182 debug_printf("Sample from %s\n", util_format_name(fmt));
1185 assert(type.floating);
1187 /* Setup our build context */
1188 memset(&bld, 0, sizeof bld);
1189 bld.gallivm = gallivm;
1190 bld.static_state = static_state;
1191 bld.dynamic_state = dynamic_state;
1192 bld.format_desc = util_format_description(static_state->format);
1195 bld.float_type = lp_type_float(32);
1196 bld.int_type = lp_type_int(32);
1197 bld.coord_type = type;
1198 bld.int_coord_type = lp_int_type(type);
1199 bld.float_size_type = lp_type_float(32);
1200 bld.float_size_type.length = dims > 1 ? 4 : 1;
1201 bld.int_size_type = lp_int_type(bld.float_size_type);
1202 bld.texel_type = type;
1204 float_vec_type = lp_type_float_vec(32);
1206 lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
1207 lp_build_context_init(&bld.float_vec_bld, gallivm, float_vec_type);
1208 lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
1209 lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
1210 lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
1211 lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
1212 lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
1213 lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
1215 /* Get the dynamic state */
1216 bld.width = dynamic_state->width(dynamic_state, gallivm, unit);
1217 bld.height = dynamic_state->height(dynamic_state, gallivm, unit);
1218 bld.depth = dynamic_state->depth(dynamic_state, gallivm, unit);
1219 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, unit);
1220 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, unit);
1221 bld.data_array = dynamic_state->data_ptr(dynamic_state, gallivm, unit);
1222 /* Note that data_array is an array[level] of pointers to texture images */
1228 /* width, height, depth as single int vector */
1230 bld.int_size = bld.width;
1233 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_bld.undef,
1234 bld.width, LLVMConstInt(i32t, 0, 0), "");
1236 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1237 bld.height, LLVMConstInt(i32t, 1, 0), "");
1239 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1240 bld.depth, LLVMConstInt(i32t, 2, 0), "");
1246 /* For debug: no-op texture sampling */
1247 lp_build_sample_nop(gallivm, bld.texel_type, texel_out);
1249 else if (util_format_fits_8unorm(bld.format_desc) &&
1250 lp_is_simple_wrap_mode(static_state->wrap_s) &&
1251 lp_is_simple_wrap_mode(static_state->wrap_t)) {
1252 /* do sampling/filtering with fixed pt arithmetic */
1253 lp_build_sample_aos(&bld, unit, s, t, r, ddx, ddy,
1254 lod_bias, explicit_lod,
1259 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1260 util_format_fits_8unorm(bld.format_desc)) {
1261 debug_printf("%s: using floating point linear filtering for %s\n",
1262 __FUNCTION__, bld.format_desc->short_name);
1263 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1264 static_state->min_img_filter,
1265 static_state->mag_img_filter,
1266 static_state->min_mip_filter,
1267 static_state->wrap_s,
1268 static_state->wrap_t);
1271 lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
1272 lod_bias, explicit_lod,
1276 lp_build_sample_compare(&bld, r, texel_out);
1278 apply_sampler_swizzle(&bld, texel_out);