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 int dims
= texture_dims(static_state
->target
);
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 /* fract = coord - floor(coord) */
206 fract
= lp_build_sub(coord_bld
, coord
, lp_build_floor(coord_bld
, coord
));
208 /* flr = ifloor(coord); */
209 flr
= lp_build_ifloor(coord_bld
, coord
);
211 /* isOdd = flr & 1 */
212 isOdd
= LLVMBuildAnd(bld
->builder
, flr
, int_coord_bld
->one
, "");
214 /* make coord positive or negative depending on isOdd */
215 coord
= lp_build_set_sign(coord_bld
, fract
, isOdd
);
217 /* convert isOdd to float */
218 isOdd
= lp_build_int_to_float(coord_bld
, isOdd
);
220 /* add isOdd to coord */
221 coord
= lp_build_add(coord_bld
, coord
, isOdd
);
228 * Build LLVM code for texture wrap mode for linear filtering.
229 * \param x0_out returns first integer texcoord
230 * \param x1_out returns second integer texcoord
231 * \param weight_out returns linear interpolation weight
234 lp_build_sample_wrap_linear(struct lp_build_sample_context
*bld
,
239 LLVMValueRef
*x0_out
,
240 LLVMValueRef
*x1_out
,
241 LLVMValueRef
*weight_out
)
243 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
244 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
245 struct lp_build_context
*uint_coord_bld
= &bld
->uint_coord_bld
;
246 LLVMValueRef half
= lp_build_const_vec(coord_bld
->type
, 0.5);
247 LLVMValueRef length_f
= lp_build_int_to_float(coord_bld
, length
);
248 LLVMValueRef length_minus_one
= lp_build_sub(uint_coord_bld
, length
, uint_coord_bld
->one
);
249 LLVMValueRef coord0
, coord1
, weight
;
252 case PIPE_TEX_WRAP_REPEAT
:
253 /* mul by size and subtract 0.5 */
254 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
255 coord
= lp_build_sub(coord_bld
, coord
, half
);
256 /* convert to int, compute lerp weight */
257 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
258 coord1
= lp_build_add(uint_coord_bld
, coord0
, uint_coord_bld
->one
);
261 coord0
= LLVMBuildAnd(bld
->builder
, coord0
, length_minus_one
, "");
262 coord1
= LLVMBuildAnd(bld
->builder
, coord1
, length_minus_one
, "");
265 /* Add a bias to the texcoord to handle negative coords */
266 LLVMValueRef bias
= lp_build_mul_imm(uint_coord_bld
, length
, 1024);
267 coord0
= LLVMBuildAdd(bld
->builder
, coord0
, bias
, "");
268 coord1
= LLVMBuildAdd(bld
->builder
, coord1
, bias
, "");
269 coord0
= LLVMBuildURem(bld
->builder
, coord0
, length
, "");
270 coord1
= LLVMBuildURem(bld
->builder
, coord1
, length
, "");
274 case PIPE_TEX_WRAP_CLAMP
:
275 if (bld
->static_state
->normalized_coords
) {
276 /* scale coord to length */
277 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
280 /* clamp to [0, length] */
281 coord
= lp_build_clamp(coord_bld
, coord
, coord_bld
->zero
, length_f
);
283 coord
= lp_build_sub(coord_bld
, coord
, half
);
285 /* convert to int, compute lerp weight */
286 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
287 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
290 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
291 if (bld
->static_state
->normalized_coords
) {
293 coord
= lp_build_clamp(coord_bld
, coord
, coord_bld
->zero
, coord_bld
->one
);
294 /* mul by tex size and subtract 0.5 */
295 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
296 coord
= lp_build_sub(coord_bld
, coord
, half
);
299 LLVMValueRef min
, max
;
300 /* clamp to [0.5, length - 0.5] */
302 max
= lp_build_sub(coord_bld
, length_f
, min
);
303 coord
= lp_build_clamp(coord_bld
, coord
, min
, max
);
305 /* convert to int, compute lerp weight */
306 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
307 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
308 /* coord0 = max(coord0, 0) */
309 coord0
= lp_build_max(int_coord_bld
, coord0
, int_coord_bld
->zero
);
310 /* coord1 = min(coord1, length-1) */
311 coord1
= lp_build_min(int_coord_bld
, coord1
, length_minus_one
);
314 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
316 LLVMValueRef min
, max
;
317 if (bld
->static_state
->normalized_coords
) {
318 /* scale coord to length */
319 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
321 /* clamp to [-0.5, length + 0.5] */
322 min
= lp_build_const_vec(coord_bld
->type
, -0.5F
);
323 max
= lp_build_sub(coord_bld
, length_f
, min
);
324 coord
= lp_build_clamp(coord_bld
, coord
, min
, max
);
325 coord
= lp_build_sub(coord_bld
, coord
, half
);
326 /* convert to int, compute lerp weight */
327 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
328 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
332 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
333 /* compute mirror function */
334 coord
= lp_build_coord_mirror(bld
, coord
);
336 /* scale coord to length */
337 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
338 coord
= lp_build_sub(coord_bld
, coord
, half
);
340 /* convert to int, compute lerp weight */
341 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
342 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
344 /* coord0 = max(coord0, 0) */
345 coord0
= lp_build_max(int_coord_bld
, coord0
, int_coord_bld
->zero
);
346 /* coord1 = min(coord1, length-1) */
347 coord1
= lp_build_min(int_coord_bld
, coord1
, length_minus_one
);
350 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
351 coord
= lp_build_abs(coord_bld
, coord
);
353 if (bld
->static_state
->normalized_coords
) {
354 /* scale coord to length */
355 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
358 /* clamp to [0, length] */
359 coord
= lp_build_min(coord_bld
, coord
, length_f
);
361 coord
= lp_build_sub(coord_bld
, coord
, half
);
363 /* convert to int, compute lerp weight */
364 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
365 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
368 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
370 LLVMValueRef min
, max
;
372 coord
= lp_build_abs(coord_bld
, coord
);
374 if (bld
->static_state
->normalized_coords
) {
375 /* scale coord to length */
376 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
379 /* clamp to [0.5, length - 0.5] */
381 max
= lp_build_sub(coord_bld
, length_f
, min
);
382 coord
= lp_build_clamp(coord_bld
, coord
, min
, max
);
384 coord
= lp_build_sub(coord_bld
, coord
, half
);
386 /* convert to int, compute lerp weight */
387 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
388 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
392 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
394 LLVMValueRef min
, max
;
396 coord
= lp_build_abs(coord_bld
, coord
);
398 if (bld
->static_state
->normalized_coords
) {
399 /* scale coord to length */
400 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
403 /* clamp to [-0.5, length + 0.5] */
404 min
= lp_build_negate(coord_bld
, half
);
405 max
= lp_build_sub(coord_bld
, length_f
, min
);
406 coord
= lp_build_clamp(coord_bld
, coord
, min
, max
);
408 coord
= lp_build_sub(coord_bld
, coord
, half
);
410 /* convert to int, compute lerp weight */
411 lp_build_ifloor_fract(coord_bld
, coord
, &coord0
, &weight
);
412 coord1
= lp_build_add(int_coord_bld
, coord0
, int_coord_bld
->one
);
425 *weight_out
= weight
;
430 * Build LLVM code for texture wrap mode for nearest filtering.
431 * \param coord the incoming texcoord (nominally in [0,1])
432 * \param length the texture size along one dimension, as int vector
433 * \param is_pot if TRUE, length is a power of two
434 * \param wrap_mode one of PIPE_TEX_WRAP_x
437 lp_build_sample_wrap_nearest(struct lp_build_sample_context
*bld
,
443 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
444 struct lp_build_context
*int_coord_bld
= &bld
->int_coord_bld
;
445 struct lp_build_context
*uint_coord_bld
= &bld
->uint_coord_bld
;
446 LLVMValueRef length_f
= lp_build_int_to_float(coord_bld
, length
);
447 LLVMValueRef length_minus_one
= lp_build_sub(uint_coord_bld
, length
, uint_coord_bld
->one
);
451 case PIPE_TEX_WRAP_REPEAT
:
452 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
453 icoord
= lp_build_ifloor(coord_bld
, coord
);
455 icoord
= LLVMBuildAnd(bld
->builder
, icoord
, length_minus_one
, "");
457 /* Add a bias to the texcoord to handle negative coords */
458 LLVMValueRef bias
= lp_build_mul_imm(uint_coord_bld
, length
, 1024);
459 icoord
= LLVMBuildAdd(bld
->builder
, icoord
, bias
, "");
460 icoord
= LLVMBuildURem(bld
->builder
, icoord
, length
, "");
464 case PIPE_TEX_WRAP_CLAMP
:
465 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
466 if (bld
->static_state
->normalized_coords
) {
467 /* scale coord to length */
468 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
472 icoord
= lp_build_ifloor(coord_bld
, coord
);
474 /* clamp to [0, length - 1]. */
475 icoord
= lp_build_clamp(int_coord_bld
, icoord
, int_coord_bld
->zero
,
479 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
480 /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
482 LLVMValueRef min
, max
;
484 if (bld
->static_state
->normalized_coords
) {
485 /* scale coord to length */
486 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
489 icoord
= lp_build_ifloor(coord_bld
, coord
);
491 /* clamp to [-1, length] */
492 min
= lp_build_negate(int_coord_bld
, int_coord_bld
->one
);
494 icoord
= lp_build_clamp(int_coord_bld
, icoord
, min
, max
);
498 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
499 /* compute mirror function */
500 coord
= lp_build_coord_mirror(bld
, coord
);
502 /* scale coord to length */
503 assert(bld
->static_state
->normalized_coords
);
504 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
506 icoord
= lp_build_ifloor(coord_bld
, coord
);
508 /* clamp to [0, length - 1] */
509 icoord
= lp_build_min(int_coord_bld
, icoord
, length_minus_one
);
512 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
513 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
514 coord
= lp_build_abs(coord_bld
, coord
);
516 if (bld
->static_state
->normalized_coords
) {
517 /* scale coord to length */
518 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
521 icoord
= lp_build_ifloor(coord_bld
, coord
);
523 /* clamp to [0, length - 1] */
524 icoord
= lp_build_min(int_coord_bld
, icoord
, length_minus_one
);
527 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
528 coord
= lp_build_abs(coord_bld
, coord
);
530 if (bld
->static_state
->normalized_coords
) {
531 /* scale coord to length */
532 coord
= lp_build_mul(coord_bld
, coord
, length_f
);
535 icoord
= lp_build_ifloor(coord_bld
, coord
);
537 /* clamp to [0, length] */
538 icoord
= lp_build_min(int_coord_bld
, icoord
, length
);
551 * Generate code to sample a mipmap level with nearest filtering.
552 * If sampling a cube texture, r = cube face in [0,5].
555 lp_build_sample_image_nearest(struct lp_build_sample_context
*bld
,
557 LLVMValueRef width_vec
,
558 LLVMValueRef height_vec
,
559 LLVMValueRef depth_vec
,
560 LLVMValueRef row_stride_vec
,
561 LLVMValueRef img_stride_vec
,
562 LLVMValueRef data_ptr
,
566 LLVMValueRef colors_out
[4])
568 const int dims
= texture_dims(bld
->static_state
->target
);
569 LLVMValueRef x
, y
, z
;
572 * Compute integer texcoords.
574 x
= lp_build_sample_wrap_nearest(bld
, s
, width_vec
,
575 bld
->static_state
->pot_width
,
576 bld
->static_state
->wrap_s
);
577 lp_build_name(x
, "tex.x.wrapped");
580 y
= lp_build_sample_wrap_nearest(bld
, t
, height_vec
,
581 bld
->static_state
->pot_height
,
582 bld
->static_state
->wrap_t
);
583 lp_build_name(y
, "tex.y.wrapped");
586 z
= lp_build_sample_wrap_nearest(bld
, r
, depth_vec
,
587 bld
->static_state
->pot_depth
,
588 bld
->static_state
->wrap_r
);
589 lp_build_name(z
, "tex.z.wrapped");
591 else if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
603 * Get texture colors.
605 lp_build_sample_texel_soa(bld
, unit
,
606 width_vec
, height_vec
, depth_vec
,
608 row_stride_vec
, img_stride_vec
,
609 data_ptr
, colors_out
);
614 * Generate code to sample a mipmap level with linear filtering.
615 * If sampling a cube texture, r = cube face in [0,5].
618 lp_build_sample_image_linear(struct lp_build_sample_context
*bld
,
620 LLVMValueRef width_vec
,
621 LLVMValueRef height_vec
,
622 LLVMValueRef depth_vec
,
623 LLVMValueRef row_stride_vec
,
624 LLVMValueRef img_stride_vec
,
625 LLVMValueRef data_ptr
,
629 LLVMValueRef colors_out
[4])
631 const int dims
= texture_dims(bld
->static_state
->target
);
632 LLVMValueRef x0
, y0
, z0
, x1
, y1
, z1
;
633 LLVMValueRef s_fpart
, t_fpart
, r_fpart
;
634 LLVMValueRef neighbors
[2][2][4];
638 * Compute integer texcoords.
640 lp_build_sample_wrap_linear(bld
, s
, width_vec
,
641 bld
->static_state
->pot_width
,
642 bld
->static_state
->wrap_s
,
644 lp_build_name(x0
, "tex.x0.wrapped");
645 lp_build_name(x1
, "tex.x1.wrapped");
648 lp_build_sample_wrap_linear(bld
, t
, height_vec
,
649 bld
->static_state
->pot_height
,
650 bld
->static_state
->wrap_t
,
652 lp_build_name(y0
, "tex.y0.wrapped");
653 lp_build_name(y1
, "tex.y1.wrapped");
656 lp_build_sample_wrap_linear(bld
, r
, depth_vec
,
657 bld
->static_state
->pot_depth
,
658 bld
->static_state
->wrap_r
,
660 lp_build_name(z0
, "tex.z0.wrapped");
661 lp_build_name(z1
, "tex.z1.wrapped");
663 else if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
664 z0
= z1
= r
; /* cube face */
673 y0
= y1
= t_fpart
= NULL
;
674 z0
= z1
= r_fpart
= NULL
;
678 * Get texture colors.
680 /* get x0/x1 texels */
681 lp_build_sample_texel_soa(bld
, unit
,
682 width_vec
, height_vec
, depth_vec
,
684 row_stride_vec
, img_stride_vec
,
685 data_ptr
, neighbors
[0][0]);
686 lp_build_sample_texel_soa(bld
, unit
,
687 width_vec
, height_vec
, depth_vec
,
689 row_stride_vec
, img_stride_vec
,
690 data_ptr
, neighbors
[0][1]);
693 /* Interpolate two samples from 1D image to produce one color */
694 for (chan
= 0; chan
< 4; chan
++) {
695 colors_out
[chan
] = lp_build_lerp(&bld
->texel_bld
, s_fpart
,
696 neighbors
[0][0][chan
],
697 neighbors
[0][1][chan
]);
702 LLVMValueRef colors0
[4];
704 /* get x0/x1 texels at y1 */
705 lp_build_sample_texel_soa(bld
, unit
,
706 width_vec
, height_vec
, depth_vec
,
708 row_stride_vec
, img_stride_vec
,
709 data_ptr
, neighbors
[1][0]);
710 lp_build_sample_texel_soa(bld
, unit
,
711 width_vec
, height_vec
, depth_vec
,
713 row_stride_vec
, img_stride_vec
,
714 data_ptr
, neighbors
[1][1]);
716 /* Bilinear interpolate the four samples from the 2D image / 3D slice */
717 for (chan
= 0; chan
< 4; chan
++) {
718 colors0
[chan
] = lp_build_lerp_2d(&bld
->texel_bld
,
720 neighbors
[0][0][chan
],
721 neighbors
[0][1][chan
],
722 neighbors
[1][0][chan
],
723 neighbors
[1][1][chan
]);
727 LLVMValueRef neighbors1
[2][2][4];
728 LLVMValueRef colors1
[4];
730 /* get x0/x1/y0/y1 texels at z1 */
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
, neighbors1
[0][0]);
736 lp_build_sample_texel_soa(bld
, unit
,
737 width_vec
, height_vec
, depth_vec
,
739 row_stride_vec
, img_stride_vec
,
740 data_ptr
, neighbors1
[0][1]);
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
, neighbors1
[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
, neighbors1
[1][1]);
752 /* Bilinear interpolate the four samples from the second Z slice */
753 for (chan
= 0; chan
< 4; chan
++) {
754 colors1
[chan
] = lp_build_lerp_2d(&bld
->texel_bld
,
756 neighbors1
[0][0][chan
],
757 neighbors1
[0][1][chan
],
758 neighbors1
[1][0][chan
],
759 neighbors1
[1][1][chan
]);
762 /* Linearly interpolate the two samples from the two 3D slices */
763 for (chan
= 0; chan
< 4; chan
++) {
764 colors_out
[chan
] = lp_build_lerp(&bld
->texel_bld
,
766 colors0
[chan
], colors1
[chan
]);
771 for (chan
= 0; chan
< 4; chan
++) {
772 colors_out
[chan
] = colors0
[chan
];
780 * Sample the texture/mipmap using given image filter and mip filter.
781 * data0_ptr and data1_ptr point to the two mipmap levels to sample
782 * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
783 * If we're using nearest miplevel sampling the '1' values will be null/unused.
786 lp_build_sample_mipmap(struct lp_build_sample_context
*bld
,
793 LLVMValueRef lod_fpart
,
794 LLVMValueRef width0_vec
,
795 LLVMValueRef width1_vec
,
796 LLVMValueRef height0_vec
,
797 LLVMValueRef height1_vec
,
798 LLVMValueRef depth0_vec
,
799 LLVMValueRef depth1_vec
,
800 LLVMValueRef row_stride0_vec
,
801 LLVMValueRef row_stride1_vec
,
802 LLVMValueRef img_stride0_vec
,
803 LLVMValueRef img_stride1_vec
,
804 LLVMValueRef data_ptr0
,
805 LLVMValueRef data_ptr1
,
806 LLVMValueRef
*colors_out
)
808 LLVMValueRef colors0
[4], colors1
[4];
811 if (img_filter
== PIPE_TEX_FILTER_NEAREST
) {
812 /* sample the first mipmap level */
813 lp_build_sample_image_nearest(bld
, unit
,
814 width0_vec
, height0_vec
, depth0_vec
,
815 row_stride0_vec
, img_stride0_vec
,
816 data_ptr0
, s
, t
, r
, colors0
);
818 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
819 /* sample the second mipmap level */
820 lp_build_sample_image_nearest(bld
, unit
,
821 width1_vec
, height1_vec
, depth1_vec
,
822 row_stride1_vec
, img_stride1_vec
,
823 data_ptr1
, s
, t
, r
, colors1
);
827 assert(img_filter
== PIPE_TEX_FILTER_LINEAR
);
829 /* sample the first mipmap level */
830 lp_build_sample_image_linear(bld
, unit
,
831 width0_vec
, height0_vec
, depth0_vec
,
832 row_stride0_vec
, img_stride0_vec
,
833 data_ptr0
, s
, t
, r
, colors0
);
835 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
836 /* sample the second mipmap level */
837 lp_build_sample_image_linear(bld
, unit
,
838 width1_vec
, height1_vec
, depth1_vec
,
839 row_stride1_vec
, img_stride1_vec
,
840 data_ptr1
, s
, t
, r
, colors1
);
844 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
845 /* interpolate samples from the two mipmap levels */
846 for (chan
= 0; chan
< 4; chan
++) {
847 colors_out
[chan
] = lp_build_lerp(&bld
->texel_bld
, lod_fpart
,
848 colors0
[chan
], colors1
[chan
]);
852 /* use first/only level's colors */
853 for (chan
= 0; chan
< 4; chan
++) {
854 colors_out
[chan
] = colors0
[chan
];
862 * General texture sampling codegen.
863 * This function handles texture sampling for all texture targets (1D,
864 * 2D, 3D, cube) and all filtering modes.
867 lp_build_sample_general(struct lp_build_sample_context
*bld
,
872 const LLVMValueRef
*ddx
,
873 const LLVMValueRef
*ddy
,
874 LLVMValueRef lod_bias
, /* optional */
875 LLVMValueRef explicit_lod
, /* optional */
879 LLVMValueRef width_vec
,
880 LLVMValueRef height_vec
,
881 LLVMValueRef depth_vec
,
882 LLVMValueRef row_stride_array
,
883 LLVMValueRef img_stride_array
,
884 LLVMValueRef data_array
,
885 LLVMValueRef
*colors_out
)
887 struct lp_build_context
*int_bld
= &bld
->int_bld
;
888 const unsigned mip_filter
= bld
->static_state
->min_mip_filter
;
889 const unsigned min_filter
= bld
->static_state
->min_img_filter
;
890 const unsigned mag_filter
= bld
->static_state
->mag_img_filter
;
891 const int dims
= texture_dims(bld
->static_state
->target
);
892 LLVMValueRef lod_ipart
= NULL
, lod_fpart
= NULL
;
893 LLVMValueRef ilevel0
, ilevel1
= NULL
;
894 LLVMValueRef width0_vec
= NULL
, height0_vec
= NULL
, depth0_vec
= NULL
;
895 LLVMValueRef width1_vec
= NULL
, height1_vec
= NULL
, depth1_vec
= NULL
;
896 LLVMValueRef row_stride0_vec
= NULL
, row_stride1_vec
= NULL
;
897 LLVMValueRef img_stride0_vec
= NULL
, img_stride1_vec
= NULL
;
898 LLVMValueRef data_ptr0
, data_ptr1
= NULL
;
899 LLVMValueRef face_ddx
[4], face_ddy
[4];
902 printf("%s mip %d min %d mag %d\n", __FUNCTION__,
903 mip_filter, min_filter, mag_filter);
907 * Choose cube face, recompute texcoords and derivatives for the chosen face.
909 if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
910 LLVMValueRef face
, face_s
, face_t
;
911 lp_build_cube_lookup(bld
, s
, t
, r
, &face
, &face_s
, &face_t
);
912 s
= face_s
; /* vec */
913 t
= face_t
; /* vec */
914 /* use 'r' to indicate cube face */
915 r
= lp_build_broadcast_scalar(&bld
->int_coord_bld
, face
); /* vec */
917 /* recompute ddx, ddy using the new (s,t) face texcoords */
918 face_ddx
[0] = lp_build_ddx(&bld
->coord_bld
, s
);
919 face_ddx
[1] = lp_build_ddx(&bld
->coord_bld
, t
);
922 face_ddy
[0] = lp_build_ddy(&bld
->coord_bld
, s
);
923 face_ddy
[1] = lp_build_ddy(&bld
->coord_bld
, t
);
931 * Compute the level of detail (float).
933 if (min_filter
!= mag_filter
||
934 mip_filter
!= PIPE_TEX_MIPFILTER_NONE
) {
935 /* Need to compute lod either to choose mipmap levels or to
936 * distinguish between minification/magnification with one mipmap level.
938 lp_build_lod_selector(bld
, unit
, ddx
, ddy
,
939 lod_bias
, explicit_lod
,
940 width
, height
, depth
,
942 &lod_ipart
, &lod_fpart
);
944 lod_ipart
= LLVMConstInt(LLVMInt32Type(), 0, 0);
948 * Compute integer mipmap level(s) to fetch texels from.
950 if (mip_filter
== PIPE_TEX_MIPFILTER_NONE
) {
951 /* always use mip level 0 */
952 if (bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
953 /* XXX this is a work-around for an apparent bug in LLVM 2.7.
954 * We should be able to set ilevel0 = const(0) but that causes
955 * bad x86 code to be emitted.
958 lp_build_nearest_mip_level(bld
, unit
, lod_ipart
, &ilevel0
);
961 ilevel0
= LLVMConstInt(LLVMInt32Type(), 0, 0);
966 if (mip_filter
== PIPE_TEX_MIPFILTER_NEAREST
) {
967 lp_build_nearest_mip_level(bld
, unit
, lod_ipart
, &ilevel0
);
970 assert(mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
);
971 lp_build_linear_mip_levels(bld
, unit
, lod_ipart
, &ilevel0
, &ilevel1
);
972 lod_fpart
= lp_build_broadcast_scalar(&bld
->coord_bld
, lod_fpart
);
976 /* compute image size(s) of source mipmap level(s) */
977 lp_build_mipmap_level_sizes(bld
, dims
, width_vec
, height_vec
, depth_vec
,
979 row_stride_array
, img_stride_array
,
980 &width0_vec
, &width1_vec
,
981 &height0_vec
, &height1_vec
,
982 &depth0_vec
, &depth1_vec
,
983 &row_stride0_vec
, &row_stride1_vec
,
984 &img_stride0_vec
, &img_stride1_vec
);
987 * Get pointer(s) to image data for mipmap level(s).
989 data_ptr0
= lp_build_get_mipmap_level(bld
, data_array
, ilevel0
);
990 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
991 data_ptr1
= lp_build_get_mipmap_level(bld
, data_array
, ilevel1
);
995 * Get/interpolate texture colors.
997 if (min_filter
== mag_filter
) {
998 /* no need to distinquish between minification and magnification */
999 lp_build_sample_mipmap(bld
, unit
,
1000 min_filter
, mip_filter
, s
, t
, r
, lod_fpart
,
1001 width0_vec
, width1_vec
,
1002 height0_vec
, height1_vec
,
1003 depth0_vec
, depth1_vec
,
1004 row_stride0_vec
, row_stride1_vec
,
1005 img_stride0_vec
, img_stride1_vec
,
1006 data_ptr0
, data_ptr1
,
1010 /* Emit conditional to choose min image filter or mag image filter
1011 * depending on the lod being >0 or <= 0, respectively.
1013 struct lp_build_flow_context
*flow_ctx
;
1014 struct lp_build_if_state if_ctx
;
1015 LLVMValueRef minify
;
1017 flow_ctx
= lp_build_flow_create(bld
->builder
);
1018 lp_build_flow_scope_begin(flow_ctx
);
1020 lp_build_flow_scope_declare(flow_ctx
, &colors_out
[0]);
1021 lp_build_flow_scope_declare(flow_ctx
, &colors_out
[1]);
1022 lp_build_flow_scope_declare(flow_ctx
, &colors_out
[2]);
1023 lp_build_flow_scope_declare(flow_ctx
, &colors_out
[3]);
1025 /* minify = lod >= 0.0 */
1026 minify
= LLVMBuildICmp(bld
->builder
, LLVMIntSGE
,
1027 lod_ipart
, int_bld
->zero
, "");
1029 lp_build_if(&if_ctx
, flow_ctx
, bld
->builder
, minify
);
1031 /* Use the minification filter */
1032 lp_build_sample_mipmap(bld
, unit
,
1033 min_filter
, mip_filter
,
1035 width0_vec
, width1_vec
,
1036 height0_vec
, height1_vec
,
1037 depth0_vec
, depth1_vec
,
1038 row_stride0_vec
, row_stride1_vec
,
1039 img_stride0_vec
, img_stride1_vec
,
1040 data_ptr0
, data_ptr1
,
1043 lp_build_else(&if_ctx
);
1045 /* Use the magnification filter */
1046 lp_build_sample_mipmap(bld
, unit
,
1047 mag_filter
, mip_filter
,
1049 width0_vec
, width1_vec
,
1050 height0_vec
, height1_vec
,
1051 depth0_vec
, depth1_vec
,
1052 row_stride0_vec
, row_stride1_vec
,
1053 img_stride0_vec
, img_stride1_vec
,
1054 data_ptr0
, data_ptr1
,
1057 lp_build_endif(&if_ctx
);
1059 lp_build_flow_scope_end(flow_ctx
);
1060 lp_build_flow_destroy(flow_ctx
);
1066 * Do shadow test/comparison.
1067 * \param p the texcoord Z (aka R, aka P) component
1068 * \param texel the texel to compare against (use the X channel)
1071 lp_build_sample_compare(struct lp_build_sample_context
*bld
,
1073 LLVMValueRef texel
[4])
1075 struct lp_build_context
*texel_bld
= &bld
->texel_bld
;
1077 const unsigned chan
= 0;
1079 if (bld
->static_state
->compare_mode
== PIPE_TEX_COMPARE_NONE
)
1084 LLVMValueRef indx
= lp_build_const_int32(0);
1085 LLVMValueRef coord
= LLVMBuildExtractElement(bld
->builder
, p
, indx
, "");
1086 LLVMValueRef tex
= LLVMBuildExtractElement(bld
->builder
,
1087 texel
[chan
], indx
, "");
1088 lp_build_printf(bld
->builder
, "shadow compare coord %f to texture %f\n",
1092 /* result = (p FUNC texel) ? 1 : 0 */
1093 res
= lp_build_cmp(texel_bld
, bld
->static_state
->compare_func
,
1095 res
= lp_build_select(texel_bld
, res
, texel_bld
->one
, texel_bld
->zero
);
1097 /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1101 texel
[3] = texel_bld
->one
;
1106 * Just set texels to white instead of actually sampling the texture.
1110 lp_build_sample_nop(struct lp_type type
,
1111 LLVMValueRef texel_out
[4])
1113 LLVMValueRef one
= lp_build_one(type
);
1116 for (chan
= 0; chan
< 4; chan
++) {
1117 texel_out
[chan
] = one
;
1123 * Build texture sampling code.
1124 * 'texel' will return a vector of four LLVMValueRefs corresponding to
1126 * \param type vector float type to use for coords, etc.
1127 * \param ddx partial derivatives of (s,t,r,q) with respect to x
1128 * \param ddy partial derivatives of (s,t,r,q) with respect to y
1131 lp_build_sample_soa(LLVMBuilderRef builder
,
1132 const struct lp_sampler_static_state
*static_state
,
1133 struct lp_sampler_dynamic_state
*dynamic_state
,
1134 struct lp_type type
,
1136 unsigned num_coords
,
1137 const LLVMValueRef
*coords
,
1138 const LLVMValueRef ddx
[4],
1139 const LLVMValueRef ddy
[4],
1140 LLVMValueRef lod_bias
, /* optional */
1141 LLVMValueRef explicit_lod
, /* optional */
1142 LLVMValueRef texel_out
[4])
1144 unsigned dims
= texture_dims(static_state
->target
);
1145 struct lp_build_sample_context bld
;
1146 LLVMTypeRef i32t
= LLVMInt32Type();
1147 LLVMValueRef width
, width_vec
;
1148 LLVMValueRef height
, height_vec
;
1149 LLVMValueRef depth
, depth_vec
;
1150 LLVMValueRef row_stride_array
, img_stride_array
;
1151 LLVMValueRef data_array
;
1155 struct lp_type float_vec_type
;
1158 enum pipe_format fmt
= static_state
->format
;
1159 debug_printf("Sample from %s\n", util_format_name(fmt
));
1162 assert(type
.floating
);
1164 /* Setup our build context */
1165 memset(&bld
, 0, sizeof bld
);
1166 bld
.builder
= builder
;
1167 bld
.static_state
= static_state
;
1168 bld
.dynamic_state
= dynamic_state
;
1169 bld
.format_desc
= util_format_description(static_state
->format
);
1171 bld
.float_type
= lp_type_float(32);
1172 bld
.int_type
= lp_type_int(32);
1173 bld
.coord_type
= type
;
1174 bld
.uint_coord_type
= lp_uint_type(type
);
1175 bld
.int_coord_type
= lp_int_type(type
);
1176 bld
.float_size_type
= lp_type_float(32);
1177 bld
.float_size_type
.length
= dims
> 1 ? 4 : 1;
1178 bld
.uint_size_type
= lp_uint_type(bld
.float_size_type
);
1179 bld
.texel_type
= type
;
1181 float_vec_type
= lp_type_float_vec(32);
1183 lp_build_context_init(&bld
.float_bld
, builder
, bld
.float_type
);
1184 lp_build_context_init(&bld
.float_vec_bld
, builder
, float_vec_type
);
1185 lp_build_context_init(&bld
.int_bld
, builder
, bld
.int_type
);
1186 lp_build_context_init(&bld
.coord_bld
, builder
, bld
.coord_type
);
1187 lp_build_context_init(&bld
.uint_coord_bld
, builder
, bld
.uint_coord_type
);
1188 lp_build_context_init(&bld
.int_coord_bld
, builder
, bld
.int_coord_type
);
1189 lp_build_context_init(&bld
.uint_size_bld
, builder
, bld
.uint_size_type
);
1190 lp_build_context_init(&bld
.float_size_bld
, builder
, bld
.float_size_type
);
1191 lp_build_context_init(&bld
.texel_bld
, builder
, bld
.texel_type
);
1193 /* Get the dynamic state */
1194 width
= dynamic_state
->width(dynamic_state
, builder
, unit
);
1195 height
= dynamic_state
->height(dynamic_state
, builder
, unit
);
1196 depth
= dynamic_state
->depth(dynamic_state
, builder
, unit
);
1197 row_stride_array
= dynamic_state
->row_stride(dynamic_state
, builder
, unit
);
1198 img_stride_array
= dynamic_state
->img_stride(dynamic_state
, builder
, unit
);
1199 data_array
= dynamic_state
->data_ptr(dynamic_state
, builder
, unit
);
1200 /* Note that data_array is an array[level] of pointers to texture images */
1206 /* width, height, depth as single uint vector */
1208 bld
.uint_size
= width
;
1211 bld
.uint_size
= LLVMBuildInsertElement(builder
, bld
.uint_size_bld
.undef
,
1212 width
, LLVMConstInt(i32t
, 0, 0), "");
1214 bld
.uint_size
= LLVMBuildInsertElement(builder
, bld
.uint_size
,
1215 height
, LLVMConstInt(i32t
, 1, 0), "");
1217 bld
.uint_size
= LLVMBuildInsertElement(builder
, bld
.uint_size
,
1218 depth
, LLVMConstInt(i32t
, 2, 0), "");
1223 /* width, height, depth as uint vectors */
1224 width_vec
= lp_build_broadcast_scalar(&bld
.uint_coord_bld
, width
);
1225 height_vec
= lp_build_broadcast_scalar(&bld
.uint_coord_bld
, height
);
1226 depth_vec
= lp_build_broadcast_scalar(&bld
.uint_coord_bld
, depth
);
1229 /* For debug: no-op texture sampling */
1230 lp_build_sample_nop(bld
.texel_type
, texel_out
);
1232 else if (util_format_fits_8unorm(bld
.format_desc
) &&
1233 lp_is_simple_wrap_mode(static_state
->wrap_s
) &&
1234 lp_is_simple_wrap_mode(static_state
->wrap_t
)) {
1235 /* do sampling/filtering with fixed pt arithmetic */
1236 lp_build_sample_aos(&bld
, unit
, s
, t
, r
, ddx
, ddy
,
1237 lod_bias
, explicit_lod
,
1238 width
, height
, depth
,
1239 width_vec
, height_vec
, depth_vec
,
1240 row_stride_array
, img_stride_array
,
1241 data_array
, texel_out
);
1245 if ((gallivm_debug
& GALLIVM_DEBUG_PERF
) &&
1246 util_format_fits_8unorm(bld
.format_desc
)) {
1247 debug_printf("%s: using floating point linear filtering for %s\n",
1248 __FUNCTION__
, bld
.format_desc
->short_name
);
1249 debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n",
1250 static_state
->min_img_filter
,
1251 static_state
->mag_img_filter
,
1252 static_state
->min_mip_filter
,
1253 static_state
->wrap_s
,
1254 static_state
->wrap_t
);
1257 lp_build_sample_general(&bld
, unit
, s
, t
, r
, ddx
, ddy
,
1258 lod_bias
, explicit_lod
,
1259 width
, height
, depth
,
1260 width_vec
, height_vec
, depth_vec
,
1261 row_stride_array
, img_stride_array
,
1266 lp_build_sample_compare(&bld
, r
, texel_out
);