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->uint_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->uint_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,
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 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
242 LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
243 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
244 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
245 LLVMValueRef coord0, coord1, weight;
248 case PIPE_TEX_WRAP_REPEAT:
249 /* mul by size and subtract 0.5 */
250 coord = lp_build_mul(coord_bld, coord, length_f);
251 coord = lp_build_sub(coord_bld, coord, half);
252 /* convert to int, compute lerp weight */
253 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
254 coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
257 coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
258 coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
261 /* Add a bias to the texcoord to handle negative coords */
262 LLVMValueRef bias = lp_build_mul_imm(uint_coord_bld, length, 1024);
263 coord0 = LLVMBuildAdd(bld->builder, coord0, bias, "");
264 coord1 = LLVMBuildAdd(bld->builder, coord1, bias, "");
265 coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
266 coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
270 case PIPE_TEX_WRAP_CLAMP:
271 if (bld->static_state->normalized_coords) {
272 /* scale coord to length */
273 coord = lp_build_mul(coord_bld, coord, length_f);
276 /* clamp to [0, length] */
277 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
279 coord = lp_build_sub(coord_bld, coord, half);
281 /* convert to int, compute lerp weight */
282 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
283 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
286 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
287 if (bld->static_state->normalized_coords) {
289 coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
290 /* mul by tex size and subtract 0.5 */
291 coord = lp_build_mul(coord_bld, coord, length_f);
292 coord = lp_build_sub(coord_bld, coord, half);
294 /* XXX this is odd normalized ranges from -0.5 to length-0.5 after denorm
295 but non-normalized ranges from to 0.5 to length-0.5 after clamp */
297 LLVMValueRef min, max;
298 /* clamp to [0.5, length - 0.5] */
300 max = lp_build_sub(coord_bld, length_f, min);
301 coord = lp_build_clamp(coord_bld, coord, min, max);
303 /* convert to int, compute lerp weight */
304 lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
305 coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
306 /* coord0 = max(coord0, 0) */
307 coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
308 /* coord1 = min(coord1, length-1) */
309 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(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,
437 struct lp_build_context *coord_bld = &bld->coord_bld;
438 struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
439 struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
440 LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
441 LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_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(bld->builder, icoord, length_minus_one, "");
451 /* Add a bias to the texcoord to handle negative coords */
452 LLVMValueRef bias = lp_build_mul_imm(uint_coord_bld, length, 1024);
453 icoord = LLVMBuildAdd(bld->builder, icoord, bias, "");
454 icoord = LLVMBuildURem(bld->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,
555 LLVMValueRef width_vec,
556 LLVMValueRef height_vec,
557 LLVMValueRef depth_vec,
558 LLVMValueRef row_stride_vec,
559 LLVMValueRef img_stride_vec,
560 LLVMValueRef data_ptr,
564 LLVMValueRef colors_out[4])
566 const unsigned dims = bld->dims;
567 LLVMValueRef x, y, z;
570 * Compute integer texcoords.
572 x = lp_build_sample_wrap_nearest(bld, s, width_vec,
573 bld->static_state->pot_width,
574 bld->static_state->wrap_s);
575 lp_build_name(x, "tex.x.wrapped");
578 y = lp_build_sample_wrap_nearest(bld, t, height_vec,
579 bld->static_state->pot_height,
580 bld->static_state->wrap_t);
581 lp_build_name(y, "tex.y.wrapped");
584 z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
585 bld->static_state->pot_depth,
586 bld->static_state->wrap_r);
587 lp_build_name(z, "tex.z.wrapped");
589 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
601 * Get texture colors.
603 lp_build_sample_texel_soa(bld, unit,
604 width_vec, height_vec, depth_vec,
606 row_stride_vec, img_stride_vec,
607 data_ptr, colors_out);
612 * Generate code to sample a mipmap level with linear filtering.
613 * If sampling a cube texture, r = cube face in [0,5].
616 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
618 LLVMValueRef width_vec,
619 LLVMValueRef height_vec,
620 LLVMValueRef depth_vec,
621 LLVMValueRef row_stride_vec,
622 LLVMValueRef img_stride_vec,
623 LLVMValueRef data_ptr,
627 LLVMValueRef colors_out[4])
629 const unsigned dims = bld->dims;
630 LLVMValueRef x0, y0, z0, x1, y1, z1;
631 LLVMValueRef s_fpart, t_fpart, r_fpart;
632 LLVMValueRef neighbors[2][2][4];
636 * Compute integer texcoords.
638 lp_build_sample_wrap_linear(bld, s, width_vec,
639 bld->static_state->pot_width,
640 bld->static_state->wrap_s,
642 lp_build_name(x0, "tex.x0.wrapped");
643 lp_build_name(x1, "tex.x1.wrapped");
646 lp_build_sample_wrap_linear(bld, t, height_vec,
647 bld->static_state->pot_height,
648 bld->static_state->wrap_t,
650 lp_build_name(y0, "tex.y0.wrapped");
651 lp_build_name(y1, "tex.y1.wrapped");
654 lp_build_sample_wrap_linear(bld, r, depth_vec,
655 bld->static_state->pot_depth,
656 bld->static_state->wrap_r,
658 lp_build_name(z0, "tex.z0.wrapped");
659 lp_build_name(z1, "tex.z1.wrapped");
661 else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
662 z0 = z1 = r; /* cube face */
671 y0 = y1 = t_fpart = NULL;
672 z0 = z1 = r_fpart = NULL;
676 * Get texture colors.
678 /* get x0/x1 texels */
679 lp_build_sample_texel_soa(bld, unit,
680 width_vec, height_vec, depth_vec,
682 row_stride_vec, img_stride_vec,
683 data_ptr, neighbors[0][0]);
684 lp_build_sample_texel_soa(bld, unit,
685 width_vec, height_vec, depth_vec,
687 row_stride_vec, img_stride_vec,
688 data_ptr, neighbors[0][1]);
691 /* Interpolate two samples from 1D image to produce one color */
692 for (chan = 0; chan < 4; chan++) {
693 colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
694 neighbors[0][0][chan],
695 neighbors[0][1][chan]);
700 LLVMValueRef colors0[4];
702 /* get x0/x1 texels at y1 */
703 lp_build_sample_texel_soa(bld, unit,
704 width_vec, height_vec, depth_vec,
706 row_stride_vec, img_stride_vec,
707 data_ptr, neighbors[1][0]);
708 lp_build_sample_texel_soa(bld, unit,
709 width_vec, height_vec, depth_vec,
711 row_stride_vec, img_stride_vec,
712 data_ptr, neighbors[1][1]);
714 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
715 for (chan = 0; chan < 4; chan++) {
716 colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
718 neighbors[0][0][chan],
719 neighbors[0][1][chan],
720 neighbors[1][0][chan],
721 neighbors[1][1][chan]);
725 LLVMValueRef neighbors1[2][2][4];
726 LLVMValueRef colors1[4];
728 /* get x0/x1/y0/y1 texels at z1 */
729 lp_build_sample_texel_soa(bld, unit,
730 width_vec, height_vec, depth_vec,
732 row_stride_vec, img_stride_vec,
733 data_ptr, neighbors1[0][0]);
734 lp_build_sample_texel_soa(bld, unit,
735 width_vec, height_vec, depth_vec,
737 row_stride_vec, img_stride_vec,
738 data_ptr, neighbors1[0][1]);
739 lp_build_sample_texel_soa(bld, unit,
740 width_vec, height_vec, depth_vec,
742 row_stride_vec, img_stride_vec,
743 data_ptr, neighbors1[1][0]);
744 lp_build_sample_texel_soa(bld, unit,
745 width_vec, height_vec, depth_vec,
747 row_stride_vec, img_stride_vec,
748 data_ptr, neighbors1[1][1]);
750 /* Bilinear interpolate the four samples from the second Z slice */
751 for (chan = 0; chan < 4; chan++) {
752 colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
754 neighbors1[0][0][chan],
755 neighbors1[0][1][chan],
756 neighbors1[1][0][chan],
757 neighbors1[1][1][chan]);
760 /* Linearly interpolate the two samples from the two 3D slices */
761 for (chan = 0; chan < 4; chan++) {
762 colors_out[chan] = lp_build_lerp(&bld->texel_bld,
764 colors0[chan], colors1[chan]);
769 for (chan = 0; chan < 4; chan++) {
770 colors_out[chan] = colors0[chan];
778 * Sample the texture/mipmap using given image filter and mip filter.
779 * data0_ptr and data1_ptr point to the two mipmap levels to sample
780 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
781 * If we're using nearest miplevel sampling the '1' values will be null/unused.
784 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
791 LLVMValueRef ilevel0,
792 LLVMValueRef ilevel1,
793 LLVMValueRef lod_fpart,
794 LLVMValueRef *colors_out)
796 LLVMBuilderRef builder = bld->builder;
797 LLVMValueRef width0_vec;
798 LLVMValueRef width1_vec;
799 LLVMValueRef height0_vec;
800 LLVMValueRef height1_vec;
801 LLVMValueRef depth0_vec;
802 LLVMValueRef depth1_vec;
803 LLVMValueRef row_stride0_vec;
804 LLVMValueRef row_stride1_vec;
805 LLVMValueRef img_stride0_vec;
806 LLVMValueRef img_stride1_vec;
807 LLVMValueRef data_ptr0;
808 LLVMValueRef data_ptr1;
809 LLVMValueRef colors0[4], colors1[4];
812 /* sample the first mipmap level */
813 lp_build_mipmap_level_sizes(bld, ilevel0,
814 &width0_vec, &height0_vec, &depth0_vec,
815 &row_stride0_vec, &img_stride0_vec);
816 data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
817 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
818 lp_build_sample_image_nearest(bld, unit,
819 width0_vec, height0_vec, depth0_vec,
820 row_stride0_vec, img_stride0_vec,
825 assert(img_filter == PIPE_TEX_FILTER_LINEAR);
826 lp_build_sample_image_linear(bld, unit,
827 width0_vec, height0_vec, depth0_vec,
828 row_stride0_vec, img_stride0_vec,
833 /* Store the first level's colors in the output variables */
834 for (chan = 0; chan < 4; chan++) {
835 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
838 if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
839 struct lp_build_flow_context *flow_ctx;
840 struct lp_build_if_state if_ctx;
841 LLVMValueRef need_lerp;
843 flow_ctx = lp_build_flow_create(builder);
845 /* need_lerp = lod_fpart > 0 */
846 need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
851 lp_build_if(&if_ctx, flow_ctx, builder, need_lerp);
853 /* sample the second mipmap level */
854 lp_build_mipmap_level_sizes(bld, ilevel1,
855 &width1_vec, &height1_vec, &depth1_vec,
856 &row_stride1_vec, &img_stride1_vec);
857 data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
858 if (img_filter == PIPE_TEX_FILTER_NEAREST) {
859 lp_build_sample_image_nearest(bld, unit,
860 width1_vec, height1_vec, depth1_vec,
861 row_stride1_vec, img_stride1_vec,
866 lp_build_sample_image_linear(bld, unit,
867 width1_vec, height1_vec, depth1_vec,
868 row_stride1_vec, img_stride1_vec,
873 /* interpolate samples from the two mipmap levels */
875 lod_fpart = lp_build_broadcast_scalar(&bld->texel_bld, lod_fpart);
877 for (chan = 0; chan < 4; chan++) {
878 colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
879 colors0[chan], colors1[chan]);
880 LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
883 lp_build_endif(&if_ctx);
885 lp_build_flow_destroy(flow_ctx);
892 * General texture sampling codegen.
893 * This function handles texture sampling for all texture targets (1D,
894 * 2D, 3D, cube) and all filtering modes.
897 lp_build_sample_general(struct lp_build_sample_context *bld,
902 const LLVMValueRef *ddx,
903 const LLVMValueRef *ddy,
904 LLVMValueRef lod_bias, /* optional */
905 LLVMValueRef explicit_lod, /* optional */
906 LLVMValueRef *colors_out)
908 struct lp_build_context *int_bld = &bld->int_bld;
909 LLVMBuilderRef builder = bld->builder;
910 const unsigned mip_filter = bld->static_state->min_mip_filter;
911 const unsigned min_filter = bld->static_state->min_img_filter;
912 const unsigned mag_filter = bld->static_state->mag_img_filter;
913 LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
914 LLVMValueRef ilevel0, ilevel1 = NULL;
915 LLVMValueRef face_ddx[4], face_ddy[4];
916 LLVMValueRef texels[4];
917 LLVMTypeRef i32t = LLVMInt32Type();
918 LLVMValueRef i32t_zero = LLVMConstInt(i32t, 0, 0);
922 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
923 mip_filter, min_filter, mag_filter);
927 * Choose cube face, recompute texcoords and derivatives for the chosen face.
929 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
930 LLVMValueRef face, face_s, face_t;
931 lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
932 s = face_s; /* vec */
933 t = face_t; /* vec */
934 /* use 'r' to indicate cube face */
935 r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
937 /* recompute ddx, ddy using the new (s,t) face texcoords */
938 face_ddx[0] = lp_build_ddx(&bld->coord_bld, s);
939 face_ddx[1] = lp_build_ddx(&bld->coord_bld, t);
942 face_ddy[0] = lp_build_ddy(&bld->coord_bld, s);
943 face_ddy[1] = lp_build_ddy(&bld->coord_bld, t);
951 * Compute the level of detail (float).
953 if (min_filter != mag_filter ||
954 mip_filter != PIPE_TEX_MIPFILTER_NONE) {
955 /* Need to compute lod either to choose mipmap levels or to
956 * distinguish between minification/magnification with one mipmap level.
958 lp_build_lod_selector(bld, unit, ddx, ddy,
959 lod_bias, explicit_lod,
961 &lod_ipart, &lod_fpart);
963 lod_ipart = i32t_zero;
967 * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
969 switch (mip_filter) {
971 assert(0 && "bad mip_filter value in lp_build_sample_soa()");
973 case PIPE_TEX_MIPFILTER_NONE:
974 /* always use mip level 0 */
975 if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
976 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
977 * We should be able to set ilevel0 = const(0) but that causes
978 * bad x86 code to be emitted.
981 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
987 case PIPE_TEX_MIPFILTER_NEAREST:
989 lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
991 case PIPE_TEX_MIPFILTER_LINEAR:
994 lp_build_linear_mip_levels(bld, unit,
995 lod_ipart, &lod_fpart,
1001 * Get/interpolate texture colors.
1004 for (chan = 0; chan < 4; ++chan) {
1005 texels[chan] = lp_build_alloca(builder, bld->texel_bld.vec_type, "");
1006 lp_build_name(texels[chan], "sampler%u_texel_%c_var", unit, "xyzw"[chan]);
1009 if (min_filter == mag_filter) {
1010 /* no need to distinquish between minification and magnification */
1011 lp_build_sample_mipmap(bld, unit,
1012 min_filter, mip_filter,
1014 ilevel0, ilevel1, lod_fpart,
1018 /* Emit conditional to choose min image filter or mag image filter
1019 * depending on the lod being > 0 or <= 0, respectively.
1021 struct lp_build_flow_context *flow_ctx;
1022 struct lp_build_if_state if_ctx;
1023 LLVMValueRef minify;
1025 flow_ctx = lp_build_flow_create(builder);
1027 /* minify = lod >= 0.0 */
1028 minify = LLVMBuildICmp(builder, LLVMIntSGE,
1029 lod_ipart, int_bld->zero, "");
1031 lp_build_if(&if_ctx, flow_ctx, builder, minify);
1033 /* Use the minification filter */
1034 lp_build_sample_mipmap(bld, unit,
1035 min_filter, mip_filter,
1037 ilevel0, ilevel1, lod_fpart,
1040 lp_build_else(&if_ctx);
1042 /* Use the magnification filter */
1043 lp_build_sample_mipmap(bld, unit,
1044 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1046 i32t_zero, NULL, NULL,
1049 lp_build_endif(&if_ctx);
1051 lp_build_flow_destroy(flow_ctx);
1054 for (chan = 0; chan < 4; ++chan) {
1055 colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1056 lp_build_name(colors_out[chan], "sampler%u_texel_%c", unit, "xyzw"[chan]);
1062 * Do shadow test/comparison.
1063 * \param p the texcoord Z (aka R, aka P) component
1064 * \param texel the texel to compare against (use the X channel)
1067 lp_build_sample_compare(struct lp_build_sample_context *bld,
1069 LLVMValueRef texel[4])
1071 struct lp_build_context *texel_bld = &bld->texel_bld;
1073 const unsigned chan = 0;
1075 if (bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1080 LLVMValueRef indx = lp_build_const_int32(0);
1081 LLVMValueRef coord = LLVMBuildExtractElement(bld->builder, p, indx, "");
1082 LLVMValueRef tex = LLVMBuildExtractElement(bld->builder,
1083 texel[chan], indx, "");
1084 lp_build_printf(bld->builder, "shadow compare coord %f to texture %f\n",
1088 /* result = (p FUNC texel) ? 1 : 0 */
1089 res = lp_build_cmp(texel_bld, bld->static_state->compare_func,
1091 res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1093 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1097 texel[3] = texel_bld->one;
1102 * Just set texels to white instead of actually sampling the texture.
1106 lp_build_sample_nop(struct lp_type type,
1107 LLVMValueRef texel_out[4])
1109 LLVMValueRef one = lp_build_one(type);
1112 for (chan = 0; chan < 4; chan++) {
1113 texel_out[chan] = one;
1119 * Build texture sampling code.
1120 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1122 * \param type vector float type to use for coords, etc.
1123 * \param ddx partial derivatives of (s,t,r,q) with respect to x
1124 * \param ddy partial derivatives of (s,t,r,q) with respect to y
1127 lp_build_sample_soa(LLVMBuilderRef builder,
1128 const struct lp_sampler_static_state *static_state,
1129 struct lp_sampler_dynamic_state *dynamic_state,
1130 struct lp_type type,
1132 unsigned num_coords,
1133 const LLVMValueRef *coords,
1134 const LLVMValueRef ddx[4],
1135 const LLVMValueRef ddy[4],
1136 LLVMValueRef lod_bias, /* optional */
1137 LLVMValueRef explicit_lod, /* optional */
1138 LLVMValueRef texel_out[4])
1140 unsigned dims = texture_dims(static_state->target);
1141 struct lp_build_sample_context bld;
1142 LLVMTypeRef i32t = LLVMInt32Type();
1147 struct lp_type float_vec_type;
1150 enum pipe_format fmt = static_state->format;
1151 debug_printf("Sample from %s\n", util_format_name(fmt));
1154 assert(type.floating);
1156 /* Setup our build context */
1157 memset(&bld, 0, sizeof bld);
1158 bld.builder = builder;
1159 bld.static_state = static_state;
1160 bld.dynamic_state = dynamic_state;
1161 bld.format_desc = util_format_description(static_state->format);
1164 bld.float_type = lp_type_float(32);
1165 bld.int_type = lp_type_int(32);
1166 bld.coord_type = type;
1167 bld.uint_coord_type = lp_uint_type(type);
1168 bld.int_coord_type = lp_int_type(type);
1169 bld.float_size_type = lp_type_float(32);
1170 bld.float_size_type.length = dims > 1 ? 4 : 1;
1171 bld.int_size_type = lp_int_type(bld.float_size_type);
1172 bld.texel_type = type;
1174 float_vec_type = lp_type_float_vec(32);
1176 lp_build_context_init(&bld.float_bld, builder, bld.float_type);
1177 lp_build_context_init(&bld.float_vec_bld, builder, float_vec_type);
1178 lp_build_context_init(&bld.int_bld, builder, bld.int_type);
1179 lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
1180 lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
1181 lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
1182 lp_build_context_init(&bld.int_size_bld, builder, bld.int_size_type);
1183 lp_build_context_init(&bld.float_size_bld, builder, bld.float_size_type);
1184 lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
1186 /* Get the dynamic state */
1187 bld.width = dynamic_state->width(dynamic_state, builder, unit);
1188 bld.height = dynamic_state->height(dynamic_state, builder, unit);
1189 bld.depth = dynamic_state->depth(dynamic_state, builder, unit);
1190 bld.row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
1191 bld.img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit);
1192 bld.data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
1193 /* Note that data_array is an array[level] of pointers to texture images */
1199 /* width, height, depth as single int vector */
1201 bld.int_size = bld.width;
1204 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_bld.undef,
1205 bld.width, LLVMConstInt(i32t, 0, 0), "");
1207 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1208 bld.height, LLVMConstInt(i32t, 1, 0), "");
1210 bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1211 bld.depth, LLVMConstInt(i32t, 2, 0), "");
1217 /* For debug: no-op texture sampling */
1218 lp_build_sample_nop(bld.texel_type, texel_out);
1220 else if (util_format_fits_8unorm(bld.format_desc) &&
1221 lp_is_simple_wrap_mode(static_state->wrap_s) &&
1222 lp_is_simple_wrap_mode(static_state->wrap_t)) {
1223 /* do sampling/filtering with fixed pt arithmetic */
1224 lp_build_sample_aos(&bld, unit, s, t, r, ddx, ddy,
1225 lod_bias, explicit_lod,
1230 if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1231 util_format_fits_8unorm(bld.format_desc)) {
1232 debug_printf("%s: using floating point linear filtering for %s\n",
1233 __FUNCTION__, bld.format_desc->short_name);
1234 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1235 static_state->min_img_filter,
1236 static_state->mag_img_filter,
1237 static_state->min_mip_filter,
1238 static_state->wrap_s,
1239 static_state->wrap_t);
1242 lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
1243 lod_bias, explicit_lod,
1247 lp_build_sample_compare(&bld, r, texel_out);