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 -- common code.
32 * @author Jose Fonseca <jfonseca@vmware.com>
35 #include "pipe/p_defines.h"
36 #include "pipe/p_state.h"
37 #include "util/u_format.h"
38 #include "util/u_math.h"
39 #include "lp_bld_arit.h"
40 #include "lp_bld_const.h"
41 #include "lp_bld_debug.h"
42 #include "lp_bld_printf.h"
43 #include "lp_bld_flow.h"
44 #include "lp_bld_sample.h"
45 #include "lp_bld_swizzle.h"
46 #include "lp_bld_type.h"
47 #include "lp_bld_logic.h"
48 #include "lp_bld_pack.h"
52 * Bri-linear factor. Should be greater than one.
54 #define BRILINEAR_FACTOR 2
57 * Does the given texture wrap mode allow sampling the texture border color?
58 * XXX maybe move this into gallium util code.
61 lp_sampler_wrap_mode_uses_border_color(unsigned mode
,
62 unsigned min_img_filter
,
63 unsigned mag_img_filter
)
66 case PIPE_TEX_WRAP_REPEAT
:
67 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
68 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
69 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
71 case PIPE_TEX_WRAP_CLAMP
:
72 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
73 if (min_img_filter
== PIPE_TEX_FILTER_NEAREST
&&
74 mag_img_filter
== PIPE_TEX_FILTER_NEAREST
) {
79 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
80 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
83 assert(0 && "unexpected wrap mode");
90 * Initialize lp_sampler_static_state object with the gallium sampler
92 * The former is considered to be static and the later dynamic.
95 lp_sampler_static_state(struct lp_sampler_static_state
*state
,
96 const struct pipe_sampler_view
*view
,
97 const struct pipe_sampler_state
*sampler
)
99 const struct pipe_resource
*texture
= view
->texture
;
101 memset(state
, 0, sizeof *state
);
110 * We don't copy sampler state over unless it is actually enabled, to avoid
111 * spurious recompiles, as the sampler static state is part of the shader
114 * Ideally the state tracker or cso_cache module would make all state
115 * canonical, but until that happens it's better to be safe than sorry here.
117 * XXX: Actually there's much more than can be done here, especially
118 * regarding 1D/2D/3D/CUBE textures, wrap modes, etc.
121 state
->format
= view
->format
;
122 state
->swizzle_r
= view
->swizzle_r
;
123 state
->swizzle_g
= view
->swizzle_g
;
124 state
->swizzle_b
= view
->swizzle_b
;
125 state
->swizzle_a
= view
->swizzle_a
;
127 state
->target
= texture
->target
;
128 state
->pot_width
= util_is_power_of_two(texture
->width0
);
129 state
->pot_height
= util_is_power_of_two(texture
->height0
);
130 state
->pot_depth
= util_is_power_of_two(texture
->depth0
);
132 state
->wrap_s
= sampler
->wrap_s
;
133 state
->wrap_t
= sampler
->wrap_t
;
134 state
->wrap_r
= sampler
->wrap_r
;
135 state
->min_img_filter
= sampler
->min_img_filter
;
136 state
->mag_img_filter
= sampler
->mag_img_filter
;
138 if (view
->u
.tex
.last_level
&& sampler
->max_lod
> 0.0f
) {
139 state
->min_mip_filter
= sampler
->min_mip_filter
;
141 state
->min_mip_filter
= PIPE_TEX_MIPFILTER_NONE
;
144 if (state
->min_mip_filter
!= PIPE_TEX_MIPFILTER_NONE
) {
145 if (sampler
->lod_bias
!= 0.0f
) {
146 state
->lod_bias_non_zero
= 1;
149 /* If min_lod == max_lod we can greatly simplify mipmap selection.
150 * This is a case that occurs during automatic mipmap generation.
152 if (sampler
->min_lod
== sampler
->max_lod
) {
153 state
->min_max_lod_equal
= 1;
155 if (sampler
->min_lod
> 0.0f
) {
156 state
->apply_min_lod
= 1;
159 if (sampler
->max_lod
< (float)view
->u
.tex
.last_level
) {
160 state
->apply_max_lod
= 1;
165 state
->compare_mode
= sampler
->compare_mode
;
166 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
) {
167 state
->compare_func
= sampler
->compare_func
;
170 state
->normalized_coords
= sampler
->normalized_coords
;
173 * FIXME: Handle the remainder of pipe_sampler_view.
179 * Generate code to compute coordinate gradient (rho).
180 * \param derivs partial derivatives of (s, t, r, q) with respect to X and Y
182 * The resulting rho is scalar per quad.
185 lp_build_rho(struct lp_build_sample_context
*bld
,
187 const struct lp_derivatives
*derivs
)
189 struct gallivm_state
*gallivm
= bld
->gallivm
;
190 struct lp_build_context
*int_size_bld
= &bld
->int_size_bld
;
191 struct lp_build_context
*float_size_bld
= &bld
->float_size_bld
;
192 struct lp_build_context
*float_bld
= &bld
->float_bld
;
193 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
194 struct lp_build_context
*perquadf_bld
= &bld
->perquadf_bld
;
195 const LLVMValueRef
*ddx_ddy
= derivs
->ddx_ddy
;
196 const unsigned dims
= bld
->dims
;
197 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
198 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
199 LLVMValueRef index0
= LLVMConstInt(i32t
, 0, 0);
200 LLVMValueRef index1
= LLVMConstInt(i32t
, 1, 0);
201 LLVMValueRef index2
= LLVMConstInt(i32t
, 2, 0);
202 LLVMValueRef rho_vec
;
203 LLVMValueRef int_size
, float_size
;
205 LLVMValueRef first_level
, first_level_vec
;
206 LLVMValueRef abs_ddx_ddy
[2];
207 unsigned length
= coord_bld
->type
.length
;
208 unsigned num_quads
= length
/ 4;
210 LLVMValueRef i32undef
= LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
211 LLVMValueRef rho_xvec
, rho_yvec
;
213 abs_ddx_ddy
[0] = lp_build_abs(coord_bld
, ddx_ddy
[0]);
215 abs_ddx_ddy
[1] = lp_build_abs(coord_bld
, ddx_ddy
[1]);
219 static const unsigned char swizzle1
[] = {
220 0, LP_BLD_SWIZZLE_DONTCARE
,
221 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
223 static const unsigned char swizzle2
[] = {
224 1, LP_BLD_SWIZZLE_DONTCARE
,
225 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
227 rho_xvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle1
);
228 rho_yvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle2
);
230 else if (dims
== 2) {
231 static const unsigned char swizzle1
[] = {
233 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
235 static const unsigned char swizzle2
[] = {
237 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
239 rho_xvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle1
);
240 rho_yvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle2
);
243 LLVMValueRef shuffles1
[LP_MAX_VECTOR_LENGTH
];
244 LLVMValueRef shuffles2
[LP_MAX_VECTOR_LENGTH
];
246 for (i
= 0; i
< num_quads
; i
++) {
247 shuffles1
[4*i
+ 0] = lp_build_const_int32(gallivm
, 4*i
);
248 shuffles1
[4*i
+ 1] = lp_build_const_int32(gallivm
, 4*i
+ 2);
249 shuffles1
[4*i
+ 2] = lp_build_const_int32(gallivm
, length
+ 4*i
);
250 shuffles1
[4*i
+ 3] = i32undef
;
251 shuffles2
[4*i
+ 0] = lp_build_const_int32(gallivm
, 4*i
+ 1);
252 shuffles2
[4*i
+ 1] = lp_build_const_int32(gallivm
, 4*i
+ 3);
253 shuffles2
[4*i
+ 2] = lp_build_const_int32(gallivm
, length
+ 4*i
+ 1);
254 shuffles2
[4*i
+ 3] = i32undef
;
256 rho_xvec
= LLVMBuildShuffleVector(builder
, abs_ddx_ddy
[0], abs_ddx_ddy
[1],
257 LLVMConstVector(shuffles1
, length
), "");
258 rho_yvec
= LLVMBuildShuffleVector(builder
, abs_ddx_ddy
[0], abs_ddx_ddy
[1],
259 LLVMConstVector(shuffles2
, length
), "");
262 rho_vec
= lp_build_max(coord_bld
, rho_xvec
, rho_yvec
);
264 first_level
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
266 first_level_vec
= lp_build_broadcast_scalar(&bld
->int_size_bld
, first_level
);
267 int_size
= lp_build_minify(int_size_bld
, bld
->int_size
, first_level_vec
);
268 float_size
= lp_build_int_to_float(float_size_bld
, int_size
);
270 if (bld
->coord_type
.length
> 4) {
271 /* expand size to each quad */
273 /* could use some broadcast_vector helper for this? */
274 int num_quads
= bld
->coord_type
.length
/ 4;
275 LLVMValueRef src
[LP_MAX_VECTOR_LENGTH
/4];
276 for (i
= 0; i
< num_quads
; i
++) {
279 float_size
= lp_build_concat(bld
->gallivm
, src
, float_size_bld
->type
, num_quads
);
282 float_size
= lp_build_broadcast_scalar(coord_bld
, float_size
);
284 rho_vec
= lp_build_mul(coord_bld
, rho_vec
, float_size
);
291 static const unsigned char swizzle1
[] = {
292 0, LP_BLD_SWIZZLE_DONTCARE
,
293 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
295 static const unsigned char swizzle2
[] = {
296 1, LP_BLD_SWIZZLE_DONTCARE
,
297 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
299 LLVMValueRef rho_s
, rho_t
, rho_r
;
301 rho_s
= lp_build_swizzle_aos(coord_bld
, rho_vec
, swizzle1
);
302 rho_t
= lp_build_swizzle_aos(coord_bld
, rho_vec
, swizzle2
);
304 rho
= lp_build_max(coord_bld
, rho_s
, rho_t
);
307 static const unsigned char swizzle3
[] = {
308 2, LP_BLD_SWIZZLE_DONTCARE
,
309 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
311 rho_r
= lp_build_swizzle_aos(coord_bld
, rho_vec
, swizzle3
);
312 rho
= lp_build_max(coord_bld
, rho
, rho_r
);
316 rho
= lp_build_pack_aos_scalars(bld
->gallivm
, coord_bld
->type
,
317 perquadf_bld
->type
, rho
);
321 rho_vec
= LLVMBuildExtractElement(builder
, rho_vec
, index0
, "");
323 rho_vec
= lp_build_mul(float_size_bld
, rho_vec
, float_size
);
330 LLVMValueRef rho_s
, rho_t
, rho_r
;
332 rho_s
= LLVMBuildExtractElement(builder
, rho_vec
, index0
, "");
333 rho_t
= LLVMBuildExtractElement(builder
, rho_vec
, index1
, "");
335 rho
= lp_build_max(float_bld
, rho_s
, rho_t
);
338 rho_r
= LLVMBuildExtractElement(builder
, rho_vec
, index2
, "");
339 rho
= lp_build_max(float_bld
, rho
, rho_r
);
350 * Bri-linear lod computation
352 * Use a piece-wise linear approximation of log2 such that:
353 * - round to nearest, for values in the neighborhood of -1, 0, 1, 2, etc.
354 * - linear approximation for values in the neighborhood of 0.5, 1.5., etc,
355 * with the steepness specified in 'factor'
356 * - exact result for 0.5, 1.5, etc.
372 * This is a technique also commonly used in hardware:
373 * - http://ixbtlabs.com/articles2/gffx/nv40-rx800-3.html
375 * TODO: For correctness, this should only be applied when texture is known to
376 * have regular mipmaps, i.e., mipmaps derived from the base level.
378 * TODO: This could be done in fixed point, where applicable.
381 lp_build_brilinear_lod(struct lp_build_context
*bld
,
384 LLVMValueRef
*out_lod_ipart
,
385 LLVMValueRef
*out_lod_fpart
)
387 LLVMValueRef lod_fpart
;
388 double pre_offset
= (factor
- 0.5)/factor
- 0.5;
389 double post_offset
= 1 - factor
;
392 lp_build_printf(bld
->gallivm
, "lod = %f\n", lod
);
395 lod
= lp_build_add(bld
, lod
,
396 lp_build_const_vec(bld
->gallivm
, bld
->type
, pre_offset
));
398 lp_build_ifloor_fract(bld
, lod
, out_lod_ipart
, &lod_fpart
);
400 lod_fpart
= lp_build_mul(bld
, lod_fpart
,
401 lp_build_const_vec(bld
->gallivm
, bld
->type
, factor
));
403 lod_fpart
= lp_build_add(bld
, lod_fpart
,
404 lp_build_const_vec(bld
->gallivm
, bld
->type
, post_offset
));
407 * It's not necessary to clamp lod_fpart since:
408 * - the above expression will never produce numbers greater than one.
409 * - the mip filtering branch is only taken if lod_fpart is positive
412 *out_lod_fpart
= lod_fpart
;
415 lp_build_printf(bld
->gallivm
, "lod_ipart = %i\n", *out_lod_ipart
);
416 lp_build_printf(bld
->gallivm
, "lod_fpart = %f\n\n", *out_lod_fpart
);
422 * Combined log2 and brilinear lod computation.
424 * It's in all identical to calling lp_build_fast_log2() and
425 * lp_build_brilinear_lod() above, but by combining we can compute the integer
426 * and fractional part independently.
429 lp_build_brilinear_rho(struct lp_build_context
*bld
,
432 LLVMValueRef
*out_lod_ipart
,
433 LLVMValueRef
*out_lod_fpart
)
435 LLVMValueRef lod_ipart
;
436 LLVMValueRef lod_fpart
;
438 const double pre_factor
= (2*factor
- 0.5)/(M_SQRT2
*factor
);
439 const double post_offset
= 1 - 2*factor
;
441 assert(bld
->type
.floating
);
443 assert(lp_check_value(bld
->type
, rho
));
446 * The pre factor will make the intersections with the exact powers of two
447 * happen precisely where we want then to be, which means that the integer
448 * part will not need any post adjustments.
450 rho
= lp_build_mul(bld
, rho
,
451 lp_build_const_vec(bld
->gallivm
, bld
->type
, pre_factor
));
453 /* ipart = ifloor(log2(rho)) */
454 lod_ipart
= lp_build_extract_exponent(bld
, rho
, 0);
456 /* fpart = rho / 2**ipart */
457 lod_fpart
= lp_build_extract_mantissa(bld
, rho
);
459 lod_fpart
= lp_build_mul(bld
, lod_fpart
,
460 lp_build_const_vec(bld
->gallivm
, bld
->type
, factor
));
462 lod_fpart
= lp_build_add(bld
, lod_fpart
,
463 lp_build_const_vec(bld
->gallivm
, bld
->type
, post_offset
));
466 * Like lp_build_brilinear_lod, it's not necessary to clamp lod_fpart since:
467 * - the above expression will never produce numbers greater than one.
468 * - the mip filtering branch is only taken if lod_fpart is positive
471 *out_lod_ipart
= lod_ipart
;
472 *out_lod_fpart
= lod_fpart
;
477 * Generate code to compute texture level of detail (lambda).
478 * \param derivs partial derivatives of (s, t, r, q) with respect to X and Y
479 * \param lod_bias optional float vector with the shader lod bias
480 * \param explicit_lod optional float vector with the explicit lod
481 * \param width scalar int texture width
482 * \param height scalar int texture height
483 * \param depth scalar int texture depth
485 * The resulting lod is scalar per quad, so only the first value per quad
486 * passed in from lod_bias, explicit_lod is used.
489 lp_build_lod_selector(struct lp_build_sample_context
*bld
,
491 const struct lp_derivatives
*derivs
,
492 LLVMValueRef lod_bias
, /* optional */
493 LLVMValueRef explicit_lod
, /* optional */
495 LLVMValueRef
*out_lod_ipart
,
496 LLVMValueRef
*out_lod_fpart
)
499 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
500 struct lp_build_context
*perquadf_bld
= &bld
->perquadf_bld
;
503 *out_lod_ipart
= bld
->perquadi_bld
.zero
;
504 *out_lod_fpart
= perquadf_bld
->zero
;
506 if (bld
->static_state
->min_max_lod_equal
) {
507 /* User is forcing sampling from a particular mipmap level.
508 * This is hit during mipmap generation.
510 LLVMValueRef min_lod
=
511 bld
->dynamic_state
->min_lod(bld
->dynamic_state
, bld
->gallivm
, unit
);
513 lod
= lp_build_broadcast_scalar(perquadf_bld
, min_lod
);
517 lod
= lp_build_pack_aos_scalars(bld
->gallivm
, bld
->coord_bld
.type
,
518 perquadf_bld
->type
, explicit_lod
);
523 rho
= lp_build_rho(bld
, unit
, derivs
);
526 * Compute lod = log2(rho)
530 !bld
->static_state
->lod_bias_non_zero
&&
531 !bld
->static_state
->apply_max_lod
&&
532 !bld
->static_state
->apply_min_lod
) {
534 * Special case when there are no post-log2 adjustments, which
535 * saves instructions but keeping the integer and fractional lod
536 * computations separate from the start.
539 if (mip_filter
== PIPE_TEX_MIPFILTER_NONE
||
540 mip_filter
== PIPE_TEX_MIPFILTER_NEAREST
) {
541 *out_lod_ipart
= lp_build_ilog2(perquadf_bld
, rho
);
542 *out_lod_fpart
= perquadf_bld
->zero
;
545 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
&&
546 !(gallivm_debug
& GALLIVM_DEBUG_NO_BRILINEAR
)) {
547 lp_build_brilinear_rho(perquadf_bld
, rho
, BRILINEAR_FACTOR
,
548 out_lod_ipart
, out_lod_fpart
);
554 lod
= lp_build_log2(perquadf_bld
, rho
);
557 lod
= lp_build_fast_log2(perquadf_bld
, rho
);
560 /* add shader lod bias */
562 lod_bias
= lp_build_pack_aos_scalars(bld
->gallivm
, bld
->coord_bld
.type
,
563 perquadf_bld
->type
, lod_bias
);
564 lod
= LLVMBuildFAdd(builder
, lod
, lod_bias
, "shader_lod_bias");
568 /* add sampler lod bias */
569 if (bld
->static_state
->lod_bias_non_zero
) {
570 LLVMValueRef sampler_lod_bias
=
571 bld
->dynamic_state
->lod_bias(bld
->dynamic_state
, bld
->gallivm
, unit
);
572 sampler_lod_bias
= lp_build_broadcast_scalar(perquadf_bld
,
574 lod
= LLVMBuildFAdd(builder
, lod
, sampler_lod_bias
, "sampler_lod_bias");
578 if (bld
->static_state
->apply_max_lod
) {
579 LLVMValueRef max_lod
=
580 bld
->dynamic_state
->max_lod(bld
->dynamic_state
, bld
->gallivm
, unit
);
581 max_lod
= lp_build_broadcast_scalar(perquadf_bld
, max_lod
);
583 lod
= lp_build_min(perquadf_bld
, lod
, max_lod
);
585 if (bld
->static_state
->apply_min_lod
) {
586 LLVMValueRef min_lod
=
587 bld
->dynamic_state
->min_lod(bld
->dynamic_state
, bld
->gallivm
, unit
);
588 min_lod
= lp_build_broadcast_scalar(perquadf_bld
, min_lod
);
590 lod
= lp_build_max(perquadf_bld
, lod
, min_lod
);
594 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
595 if (!(gallivm_debug
& GALLIVM_DEBUG_NO_BRILINEAR
)) {
596 lp_build_brilinear_lod(perquadf_bld
, lod
, BRILINEAR_FACTOR
,
597 out_lod_ipart
, out_lod_fpart
);
600 lp_build_ifloor_fract(perquadf_bld
, lod
, out_lod_ipart
, out_lod_fpart
);
603 lp_build_name(*out_lod_fpart
, "lod_fpart");
606 *out_lod_ipart
= lp_build_iround(perquadf_bld
, lod
);
609 lp_build_name(*out_lod_ipart
, "lod_ipart");
616 * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
617 * mipmap level index.
618 * Note: this is all scalar per quad code.
619 * \param lod_ipart int texture level of detail
620 * \param level_out returns integer
623 lp_build_nearest_mip_level(struct lp_build_sample_context
*bld
,
625 LLVMValueRef lod_ipart
,
626 LLVMValueRef
*level_out
)
628 struct lp_build_context
*perquadi_bld
= &bld
->perquadi_bld
;
629 LLVMValueRef first_level
, last_level
, level
;
631 first_level
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
633 last_level
= bld
->dynamic_state
->last_level(bld
->dynamic_state
,
635 first_level
= lp_build_broadcast_scalar(perquadi_bld
, first_level
);
636 last_level
= lp_build_broadcast_scalar(perquadi_bld
, last_level
);
638 level
= lp_build_add(perquadi_bld
, lod_ipart
, first_level
);
640 /* clamp level to legal range of levels */
641 *level_out
= lp_build_clamp(perquadi_bld
, level
, first_level
, last_level
);
646 * For PIPE_TEX_MIPFILTER_LINEAR, convert per-quad int LOD(s) to two (per-quad)
647 * (adjacent) mipmap level indexes, and fix up float lod part accordingly.
648 * Later, we'll sample from those two mipmap levels and interpolate between them.
651 lp_build_linear_mip_levels(struct lp_build_sample_context
*bld
,
653 LLVMValueRef lod_ipart
,
654 LLVMValueRef
*lod_fpart_inout
,
655 LLVMValueRef
*level0_out
,
656 LLVMValueRef
*level1_out
)
658 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
659 struct lp_build_context
*perquadi_bld
= &bld
->perquadi_bld
;
660 struct lp_build_context
*perquadf_bld
= &bld
->perquadf_bld
;
661 LLVMValueRef first_level
, last_level
;
662 LLVMValueRef clamp_min
;
663 LLVMValueRef clamp_max
;
665 first_level
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
667 last_level
= bld
->dynamic_state
->last_level(bld
->dynamic_state
,
669 first_level
= lp_build_broadcast_scalar(perquadi_bld
, first_level
);
670 last_level
= lp_build_broadcast_scalar(perquadi_bld
, last_level
);
672 *level0_out
= lp_build_add(perquadi_bld
, lod_ipart
, first_level
);
673 *level1_out
= lp_build_add(perquadi_bld
, *level0_out
, perquadi_bld
->one
);
676 * Clamp both *level0_out and *level1_out to [first_level, last_level], with
677 * the minimum number of comparisons, and zeroing lod_fpart in the extreme
678 * ends in the process.
682 * This code (vector select in particular) only works with llvm 3.1
683 * (if there's more than one quad, with x86 backend). Might consider
684 * converting to our lp_bld_logic helpers.
686 #if HAVE_LLVM < 0x0301
687 assert(perquadi_bld
->type
.length
== 1);
690 /* *level0_out < first_level */
691 clamp_min
= LLVMBuildICmp(builder
, LLVMIntSLT
,
692 *level0_out
, first_level
,
693 "clamp_lod_to_first");
695 *level0_out
= LLVMBuildSelect(builder
, clamp_min
,
696 first_level
, *level0_out
, "");
698 *level1_out
= LLVMBuildSelect(builder
, clamp_min
,
699 first_level
, *level1_out
, "");
701 *lod_fpart_inout
= LLVMBuildSelect(builder
, clamp_min
,
702 perquadf_bld
->zero
, *lod_fpart_inout
, "");
704 /* *level0_out >= last_level */
705 clamp_max
= LLVMBuildICmp(builder
, LLVMIntSGE
,
706 *level0_out
, last_level
,
707 "clamp_lod_to_last");
709 *level0_out
= LLVMBuildSelect(builder
, clamp_max
,
710 last_level
, *level0_out
, "");
712 *level1_out
= LLVMBuildSelect(builder
, clamp_max
,
713 last_level
, *level1_out
, "");
715 *lod_fpart_inout
= LLVMBuildSelect(builder
, clamp_max
,
716 perquadf_bld
->zero
, *lod_fpart_inout
, "");
718 lp_build_name(*level0_out
, "sampler%u_miplevel0", unit
);
719 lp_build_name(*level1_out
, "sampler%u_miplevel1", unit
);
720 lp_build_name(*lod_fpart_inout
, "sampler%u_mipweight", unit
);
725 * Return pointer to a single mipmap level.
726 * \param data_array array of pointers to mipmap levels
727 * \param level integer mipmap level
730 lp_build_get_mipmap_level(struct lp_build_sample_context
*bld
,
733 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
734 LLVMValueRef indexes
[2], data_ptr
;
736 indexes
[0] = lp_build_const_int32(bld
->gallivm
, 0);
738 data_ptr
= LLVMBuildGEP(builder
, bld
->data_array
, indexes
, 2, "");
739 data_ptr
= LLVMBuildLoad(builder
, data_ptr
, "");
745 * Codegen equivalent for u_minify().
746 * Return max(1, base_size >> level);
749 lp_build_minify(struct lp_build_context
*bld
,
750 LLVMValueRef base_size
,
753 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
754 assert(lp_check_value(bld
->type
, base_size
));
755 assert(lp_check_value(bld
->type
, level
));
757 if (level
== bld
->zero
) {
758 /* if we're using mipmap level zero, no minification is needed */
763 LLVMBuildLShr(builder
, base_size
, level
, "minify");
764 assert(bld
->type
.sign
);
765 size
= lp_build_max(bld
, size
, bld
->one
);
772 * Dereference stride_array[mipmap_level] array to get a stride.
773 * Return stride as a vector.
776 lp_build_get_level_stride_vec(struct lp_build_sample_context
*bld
,
777 LLVMValueRef stride_array
, LLVMValueRef level
)
779 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
780 LLVMValueRef indexes
[2], stride
;
781 indexes
[0] = lp_build_const_int32(bld
->gallivm
, 0);
783 stride
= LLVMBuildGEP(builder
, stride_array
, indexes
, 2, "");
784 stride
= LLVMBuildLoad(builder
, stride
, "");
785 stride
= lp_build_broadcast_scalar(&bld
->int_coord_bld
, stride
);
791 * When sampling a mipmap, we need to compute the width, height, depth
792 * of the source levels from the level indexes. This helper function
796 lp_build_mipmap_level_sizes(struct lp_build_sample_context
*bld
,
798 LLVMValueRef
*out_size
,
799 LLVMValueRef
*row_stride_vec
,
800 LLVMValueRef
*img_stride_vec
)
802 const unsigned dims
= bld
->dims
;
803 LLVMValueRef ilevel_vec
;
805 ilevel_vec
= lp_build_broadcast_scalar(&bld
->int_size_bld
, ilevel
);
808 * Compute width, height, depth at mipmap level 'ilevel'
810 *out_size
= lp_build_minify(&bld
->int_size_bld
, bld
->int_size
, ilevel_vec
);
813 *row_stride_vec
= lp_build_get_level_stride_vec(bld
,
814 bld
->row_stride_array
,
816 if (dims
== 3 || bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
817 *img_stride_vec
= lp_build_get_level_stride_vec(bld
,
818 bld
->img_stride_array
,
826 * Extract and broadcast texture size.
828 * @param size_type type of the texture size vector (either
829 * bld->int_size_type or bld->float_size_type)
830 * @param coord_type type of the texture size vector (either
831 * bld->int_coord_type or bld->coord_type)
832 * @param size vector with the texture size (width, height, depth)
835 lp_build_extract_image_sizes(struct lp_build_sample_context
*bld
,
836 struct lp_type size_type
,
837 struct lp_type coord_type
,
839 LLVMValueRef
*out_width
,
840 LLVMValueRef
*out_height
,
841 LLVMValueRef
*out_depth
)
843 const unsigned dims
= bld
->dims
;
844 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
846 *out_width
= lp_build_extract_broadcast(bld
->gallivm
,
850 LLVMConstInt(i32t
, 0, 0));
852 *out_height
= lp_build_extract_broadcast(bld
->gallivm
,
856 LLVMConstInt(i32t
, 1, 0));
858 *out_depth
= lp_build_extract_broadcast(bld
->gallivm
,
862 LLVMConstInt(i32t
, 2, 0));
869 * Unnormalize coords.
871 * @param flt_size vector with the integer texture size (width, height, depth)
874 lp_build_unnormalized_coords(struct lp_build_sample_context
*bld
,
875 LLVMValueRef flt_size
,
880 const unsigned dims
= bld
->dims
;
885 lp_build_extract_image_sizes(bld
,
886 bld
->float_size_type
,
893 /* s = s * width, t = t * height */
894 *s
= lp_build_mul(&bld
->coord_bld
, *s
, width
);
896 *t
= lp_build_mul(&bld
->coord_bld
, *t
, height
);
898 *r
= lp_build_mul(&bld
->coord_bld
, *r
, depth
);
904 /** Helper used by lp_build_cube_lookup() */
906 lp_build_cube_imapos(struct lp_build_context
*coord_bld
, LLVMValueRef coord
)
908 /* ima = +0.5 / abs(coord); */
909 LLVMValueRef posHalf
= lp_build_const_vec(coord_bld
->gallivm
, coord_bld
->type
, 0.5);
910 LLVMValueRef absCoord
= lp_build_abs(coord_bld
, coord
);
911 LLVMValueRef ima
= lp_build_div(coord_bld
, posHalf
, absCoord
);
915 /** Helper used by lp_build_cube_lookup() */
917 lp_build_cube_imaneg(struct lp_build_context
*coord_bld
, LLVMValueRef coord
)
919 /* ima = -0.5 / abs(coord); */
920 LLVMValueRef negHalf
= lp_build_const_vec(coord_bld
->gallivm
, coord_bld
->type
, -0.5);
921 LLVMValueRef absCoord
= lp_build_abs(coord_bld
, coord
);
922 LLVMValueRef ima
= lp_build_div(coord_bld
, negHalf
, absCoord
);
927 * Helper used by lp_build_cube_lookup()
928 * FIXME: the sign here can also be 0.
929 * Arithmetically this could definitely make a difference. Either
930 * fix the comment or use other (simpler) sign function, not sure
931 * which one it should be.
932 * \param sign scalar +1 or -1
933 * \param coord float vector
934 * \param ima float vector
937 lp_build_cube_coord(struct lp_build_context
*coord_bld
,
938 LLVMValueRef sign
, int negate_coord
,
939 LLVMValueRef coord
, LLVMValueRef ima
)
941 /* return negate(coord) * ima * sign + 0.5; */
942 LLVMValueRef half
= lp_build_const_vec(coord_bld
->gallivm
, coord_bld
->type
, 0.5);
945 assert(negate_coord
== +1 || negate_coord
== -1);
947 if (negate_coord
== -1) {
948 coord
= lp_build_negate(coord_bld
, coord
);
951 res
= lp_build_mul(coord_bld
, coord
, ima
);
953 sign
= lp_build_broadcast_scalar(coord_bld
, sign
);
954 res
= lp_build_mul(coord_bld
, res
, sign
);
956 res
= lp_build_add(coord_bld
, res
, half
);
962 /** Helper used by lp_build_cube_lookup()
963 * Return (major_coord >= 0) ? pos_face : neg_face;
966 lp_build_cube_face(struct lp_build_sample_context
*bld
,
967 LLVMValueRef major_coord
,
968 unsigned pos_face
, unsigned neg_face
)
970 struct gallivm_state
*gallivm
= bld
->gallivm
;
971 LLVMBuilderRef builder
= gallivm
->builder
;
972 LLVMValueRef cmp
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
974 bld
->float_bld
.zero
, "");
975 LLVMValueRef pos
= lp_build_const_int32(gallivm
, pos_face
);
976 LLVMValueRef neg
= lp_build_const_int32(gallivm
, neg_face
);
977 LLVMValueRef res
= LLVMBuildSelect(builder
, cmp
, pos
, neg
, "");
984 * Generate code to do cube face selection and compute per-face texcoords.
987 lp_build_cube_lookup(struct lp_build_sample_context
*bld
,
992 LLVMValueRef
*face_s
,
993 LLVMValueRef
*face_t
)
995 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
996 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
997 struct gallivm_state
*gallivm
= bld
->gallivm
;
998 LLVMValueRef rx
, ry
, rz
;
999 LLVMValueRef tmp
[4], rxyz
, arxyz
;
1002 * Use the average of the four pixel's texcoords to choose the face.
1003 * Slight simplification just calculate the sum, skip scaling.
1008 rxyz
= lp_build_hadd_partial4(&bld
->coord_bld
, tmp
, 3);
1009 arxyz
= lp_build_abs(&bld
->coord_bld
, rxyz
);
1011 if (coord_bld
->type
.length
> 4) {
1012 struct lp_build_context
*cint_bld
= &bld
->int_coord_bld
;
1013 struct lp_type intctype
= cint_bld
->type
;
1014 LLVMValueRef signrxs
, signrys
, signrzs
, signrxyz
, sign
;
1015 LLVMValueRef arxs
, arys
, arzs
;
1016 LLVMValueRef arx_ge_ary
, maxarxsarys
, arz_ge_arx_ary
;
1017 LLVMValueRef snewx
, tnewx
, snewy
, tnewy
, snewz
, tnewz
;
1018 LLVMValueRef ryneg
, rzneg
;
1019 LLVMValueRef ma
, ima
;
1020 LLVMValueRef posHalf
= lp_build_const_vec(gallivm
, coord_bld
->type
, 0.5);
1021 LLVMValueRef signmask
= lp_build_const_int_vec(gallivm
, intctype
,
1022 1 << (intctype
.width
- 1));
1023 LLVMValueRef signshift
= lp_build_const_int_vec(gallivm
, intctype
,
1025 LLVMValueRef facex
= lp_build_const_int_vec(gallivm
, intctype
, PIPE_TEX_FACE_POS_X
);
1026 LLVMValueRef facey
= lp_build_const_int_vec(gallivm
, intctype
, PIPE_TEX_FACE_POS_Y
);
1027 LLVMValueRef facez
= lp_build_const_int_vec(gallivm
, intctype
, PIPE_TEX_FACE_POS_Z
);
1029 assert(PIPE_TEX_FACE_NEG_X
== PIPE_TEX_FACE_POS_X
+ 1);
1030 assert(PIPE_TEX_FACE_NEG_Y
== PIPE_TEX_FACE_POS_Y
+ 1);
1031 assert(PIPE_TEX_FACE_NEG_Z
== PIPE_TEX_FACE_POS_Z
+ 1);
1033 rx
= LLVMBuildBitCast(builder
, s
, lp_build_vec_type(gallivm
, intctype
), "");
1034 ry
= LLVMBuildBitCast(builder
, t
, lp_build_vec_type(gallivm
, intctype
), "");
1035 rz
= LLVMBuildBitCast(builder
, r
, lp_build_vec_type(gallivm
, intctype
), "");
1036 ryneg
= LLVMBuildXor(builder
, ry
, signmask
, "");
1037 rzneg
= LLVMBuildXor(builder
, rz
, signmask
, "");
1039 /* the sign bit comes from the averaged vector (per quad),
1040 * as does the decision which face to use */
1041 signrxyz
= LLVMBuildBitCast(builder
, rxyz
, lp_build_vec_type(gallivm
, intctype
), "");
1042 signrxyz
= LLVMBuildAnd(builder
, signrxyz
, signmask
, "");
1044 arxs
= lp_build_swizzle_scalar_aos(coord_bld
, arxyz
, 0);
1045 arys
= lp_build_swizzle_scalar_aos(coord_bld
, arxyz
, 1);
1046 arzs
= lp_build_swizzle_scalar_aos(coord_bld
, arxyz
, 2);
1049 * select x if x >= y else select y
1050 * select previous result if y >= max(x,y) else select z
1052 arx_ge_ary
= lp_build_cmp(coord_bld
, PIPE_FUNC_GEQUAL
, arxs
, arys
);
1053 maxarxsarys
= lp_build_max(coord_bld
, arxs
, arys
);
1054 arz_ge_arx_ary
= lp_build_cmp(coord_bld
, PIPE_FUNC_GEQUAL
, maxarxsarys
, arzs
);
1057 * compute all possible new s/t coords
1058 * snewx = signrx * -rz;
1061 * tnewy = signry * rz;
1062 * snewz = signrz * rx;
1065 signrxs
= lp_build_swizzle_scalar_aos(cint_bld
, signrxyz
, 0);
1066 snewx
= LLVMBuildXor(builder
, signrxs
, rzneg
, "");
1069 signrys
= lp_build_swizzle_scalar_aos(cint_bld
, signrxyz
, 1);
1071 tnewy
= LLVMBuildXor(builder
, signrys
, rz
, "");
1073 signrzs
= lp_build_swizzle_scalar_aos(cint_bld
, signrxyz
, 2);
1074 snewz
= LLVMBuildXor(builder
, signrzs
, rx
, "");
1077 /* XXX on x86 unclear if we should cast the values back to float
1078 * or not - on some cpus (nehalem) pblendvb has twice the throughput
1079 * of blendvps though on others there just might be domain
1080 * transition penalties when using it (this depends on what llvm
1081 * will chose for the bit ops above so there appears no "right way",
1082 * but given the boatload of selects let's just use the int type).
1084 * Unfortunately we also need the sign bit of the summed coords.
1086 *face_s
= lp_build_select(cint_bld
, arx_ge_ary
, snewx
, snewy
);
1087 *face_t
= lp_build_select(cint_bld
, arx_ge_ary
, tnewx
, tnewy
);
1088 ma
= lp_build_select(coord_bld
, arx_ge_ary
, s
, t
);
1089 *face
= lp_build_select(cint_bld
, arx_ge_ary
, facex
, facey
);
1090 sign
= lp_build_select(cint_bld
, arx_ge_ary
, signrxs
, signrys
);
1092 *face_s
= lp_build_select(cint_bld
, arz_ge_arx_ary
, *face_s
, snewz
);
1093 *face_t
= lp_build_select(cint_bld
, arz_ge_arx_ary
, *face_t
, tnewz
);
1094 ma
= lp_build_select(coord_bld
, arz_ge_arx_ary
, ma
, r
);
1095 *face
= lp_build_select(cint_bld
, arz_ge_arx_ary
, *face
, facez
);
1096 sign
= lp_build_select(cint_bld
, arz_ge_arx_ary
, sign
, signrzs
);
1098 *face_s
= LLVMBuildBitCast(builder
, *face_s
,
1099 lp_build_vec_type(gallivm
, coord_bld
->type
), "");
1100 *face_t
= LLVMBuildBitCast(builder
, *face_t
,
1101 lp_build_vec_type(gallivm
, coord_bld
->type
), "");
1103 /* add +1 for neg face */
1104 /* XXX with AVX probably want to use another select here -
1105 * as long as we ensure vblendvps gets used we can actually
1106 * skip the comparison and just use sign as a "mask" directly.
1108 sign
= LLVMBuildLShr(builder
, sign
, signshift
, "");
1109 *face
= LLVMBuildOr(builder
, *face
, sign
, "face");
1111 ima
= lp_build_cube_imapos(coord_bld
, ma
);
1113 *face_s
= lp_build_mul(coord_bld
, *face_s
, ima
);
1114 *face_s
= lp_build_add(coord_bld
, *face_s
, posHalf
);
1115 *face_t
= lp_build_mul(coord_bld
, *face_t
, ima
);
1116 *face_t
= lp_build_add(coord_bld
, *face_t
, posHalf
);
1120 struct lp_build_if_state if_ctx
;
1121 LLVMValueRef face_s_var
;
1122 LLVMValueRef face_t_var
;
1123 LLVMValueRef face_var
;
1124 LLVMValueRef arx_ge_ary_arz
, ary_ge_arx_arz
;
1125 LLVMValueRef shuffles
[4];
1126 LLVMValueRef arxy_ge_aryx
, arxy_ge_arzz
, arxy_ge_arxy_arzz
;
1127 LLVMValueRef arxyxy
, aryxzz
, arxyxy_ge_aryxzz
;
1128 struct lp_build_context
*float_bld
= &bld
->float_bld
;
1130 assert(bld
->coord_bld
.type
.length
== 4);
1132 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
1133 shuffles
[1] = lp_build_const_int32(gallivm
, 1);
1134 shuffles
[2] = lp_build_const_int32(gallivm
, 0);
1135 shuffles
[3] = lp_build_const_int32(gallivm
, 1);
1136 arxyxy
= LLVMBuildShuffleVector(builder
, arxyz
, arxyz
, LLVMConstVector(shuffles
, 4), "");
1137 shuffles
[0] = lp_build_const_int32(gallivm
, 1);
1138 shuffles
[1] = lp_build_const_int32(gallivm
, 0);
1139 shuffles
[2] = lp_build_const_int32(gallivm
, 2);
1140 shuffles
[3] = lp_build_const_int32(gallivm
, 2);
1141 aryxzz
= LLVMBuildShuffleVector(builder
, arxyz
, arxyz
, LLVMConstVector(shuffles
, 4), "");
1142 arxyxy_ge_aryxzz
= lp_build_cmp(&bld
->coord_bld
, PIPE_FUNC_GEQUAL
, arxyxy
, aryxzz
);
1144 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
1145 shuffles
[1] = lp_build_const_int32(gallivm
, 1);
1146 arxy_ge_aryx
= LLVMBuildShuffleVector(builder
, arxyxy_ge_aryxzz
, arxyxy_ge_aryxzz
,
1147 LLVMConstVector(shuffles
, 2), "");
1148 shuffles
[0] = lp_build_const_int32(gallivm
, 2);
1149 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
1150 arxy_ge_arzz
= LLVMBuildShuffleVector(builder
, arxyxy_ge_aryxzz
, arxyxy_ge_aryxzz
,
1151 LLVMConstVector(shuffles
, 2), "");
1152 arxy_ge_arxy_arzz
= LLVMBuildAnd(builder
, arxy_ge_aryx
, arxy_ge_arzz
, "");
1154 arx_ge_ary_arz
= LLVMBuildExtractElement(builder
, arxy_ge_arxy_arzz
,
1155 lp_build_const_int32(gallivm
, 0), "");
1156 arx_ge_ary_arz
= LLVMBuildICmp(builder
, LLVMIntNE
, arx_ge_ary_arz
,
1157 lp_build_const_int32(gallivm
, 0), "");
1158 ary_ge_arx_arz
= LLVMBuildExtractElement(builder
, arxy_ge_arxy_arzz
,
1159 lp_build_const_int32(gallivm
, 1), "");
1160 ary_ge_arx_arz
= LLVMBuildICmp(builder
, LLVMIntNE
, ary_ge_arx_arz
,
1161 lp_build_const_int32(gallivm
, 0), "");
1162 face_s_var
= lp_build_alloca(gallivm
, bld
->coord_bld
.vec_type
, "face_s_var");
1163 face_t_var
= lp_build_alloca(gallivm
, bld
->coord_bld
.vec_type
, "face_t_var");
1164 face_var
= lp_build_alloca(gallivm
, bld
->int_bld
.vec_type
, "face_var");
1166 lp_build_if(&if_ctx
, gallivm
, arx_ge_ary_arz
);
1169 LLVMValueRef sign
, ima
;
1170 rx
= LLVMBuildExtractElement(builder
, rxyz
,
1171 lp_build_const_int32(gallivm
, 0), "");
1173 sign
= lp_build_sgn(float_bld
, rx
);
1174 ima
= lp_build_cube_imaneg(coord_bld
, s
);
1175 *face_s
= lp_build_cube_coord(coord_bld
, sign
, +1, r
, ima
);
1176 *face_t
= lp_build_cube_coord(coord_bld
, NULL
, +1, t
, ima
);
1177 *face
= lp_build_cube_face(bld
, rx
,
1178 PIPE_TEX_FACE_POS_X
,
1179 PIPE_TEX_FACE_NEG_X
);
1180 LLVMBuildStore(builder
, *face_s
, face_s_var
);
1181 LLVMBuildStore(builder
, *face_t
, face_t_var
);
1182 LLVMBuildStore(builder
, *face
, face_var
);
1184 lp_build_else(&if_ctx
);
1186 struct lp_build_if_state if_ctx2
;
1188 lp_build_if(&if_ctx2
, gallivm
, ary_ge_arx_arz
);
1190 LLVMValueRef sign
, ima
;
1192 ry
= LLVMBuildExtractElement(builder
, rxyz
,
1193 lp_build_const_int32(gallivm
, 1), "");
1194 sign
= lp_build_sgn(float_bld
, ry
);
1195 ima
= lp_build_cube_imaneg(coord_bld
, t
);
1196 *face_s
= lp_build_cube_coord(coord_bld
, NULL
, -1, s
, ima
);
1197 *face_t
= lp_build_cube_coord(coord_bld
, sign
, -1, r
, ima
);
1198 *face
= lp_build_cube_face(bld
, ry
,
1199 PIPE_TEX_FACE_POS_Y
,
1200 PIPE_TEX_FACE_NEG_Y
);
1201 LLVMBuildStore(builder
, *face_s
, face_s_var
);
1202 LLVMBuildStore(builder
, *face_t
, face_t_var
);
1203 LLVMBuildStore(builder
, *face
, face_var
);
1205 lp_build_else(&if_ctx2
);
1208 LLVMValueRef sign
, ima
;
1209 rz
= LLVMBuildExtractElement(builder
, rxyz
,
1210 lp_build_const_int32(gallivm
, 2), "");
1211 sign
= lp_build_sgn(float_bld
, rz
);
1212 ima
= lp_build_cube_imaneg(coord_bld
, r
);
1213 *face_s
= lp_build_cube_coord(coord_bld
, sign
, -1, s
, ima
);
1214 *face_t
= lp_build_cube_coord(coord_bld
, NULL
, +1, t
, ima
);
1215 *face
= lp_build_cube_face(bld
, rz
,
1216 PIPE_TEX_FACE_POS_Z
,
1217 PIPE_TEX_FACE_NEG_Z
);
1218 LLVMBuildStore(builder
, *face_s
, face_s_var
);
1219 LLVMBuildStore(builder
, *face_t
, face_t_var
);
1220 LLVMBuildStore(builder
, *face
, face_var
);
1222 lp_build_endif(&if_ctx2
);
1225 lp_build_endif(&if_ctx
);
1227 *face_s
= LLVMBuildLoad(builder
, face_s_var
, "face_s");
1228 *face_t
= LLVMBuildLoad(builder
, face_t_var
, "face_t");
1229 *face
= LLVMBuildLoad(builder
, face_var
, "face");
1230 *face
= lp_build_broadcast_scalar(&bld
->int_coord_bld
, *face
);
1236 * Compute the partial offset of a pixel block along an arbitrary axis.
1238 * @param coord coordinate in pixels
1239 * @param stride number of bytes between rows of successive pixel blocks
1240 * @param block_length number of pixels in a pixels block along the coordinate
1242 * @param out_offset resulting relative offset of the pixel block in bytes
1243 * @param out_subcoord resulting sub-block pixel coordinate
1246 lp_build_sample_partial_offset(struct lp_build_context
*bld
,
1247 unsigned block_length
,
1249 LLVMValueRef stride
,
1250 LLVMValueRef
*out_offset
,
1251 LLVMValueRef
*out_subcoord
)
1253 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1254 LLVMValueRef offset
;
1255 LLVMValueRef subcoord
;
1257 if (block_length
== 1) {
1258 subcoord
= bld
->zero
;
1262 * Pixel blocks have power of two dimensions. LLVM should convert the
1263 * rem/div to bit arithmetic.
1264 * TODO: Verify this.
1265 * It does indeed BUT it does transform it to scalar (and back) when doing so
1266 * (using roughly extract, shift/and, mov, unpack) (llvm 2.7).
1267 * The generated code looks seriously unfunny and is quite expensive.
1270 LLVMValueRef block_width
= lp_build_const_int_vec(bld
->type
, block_length
);
1271 subcoord
= LLVMBuildURem(builder
, coord
, block_width
, "");
1272 coord
= LLVMBuildUDiv(builder
, coord
, block_width
, "");
1274 unsigned logbase2
= util_logbase2(block_length
);
1275 LLVMValueRef block_shift
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, logbase2
);
1276 LLVMValueRef block_mask
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, block_length
- 1);
1277 subcoord
= LLVMBuildAnd(builder
, coord
, block_mask
, "");
1278 coord
= LLVMBuildLShr(builder
, coord
, block_shift
, "");
1282 offset
= lp_build_mul(bld
, coord
, stride
);
1285 assert(out_subcoord
);
1287 *out_offset
= offset
;
1288 *out_subcoord
= subcoord
;
1293 * Compute the offset of a pixel block.
1295 * x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
1297 * Returns the relative offset and i,j sub-block coordinates
1300 lp_build_sample_offset(struct lp_build_context
*bld
,
1301 const struct util_format_description
*format_desc
,
1305 LLVMValueRef y_stride
,
1306 LLVMValueRef z_stride
,
1307 LLVMValueRef
*out_offset
,
1308 LLVMValueRef
*out_i
,
1309 LLVMValueRef
*out_j
)
1311 LLVMValueRef x_stride
;
1312 LLVMValueRef offset
;
1314 x_stride
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
1315 format_desc
->block
.bits
/8);
1317 lp_build_sample_partial_offset(bld
,
1318 format_desc
->block
.width
,
1322 if (y
&& y_stride
) {
1323 LLVMValueRef y_offset
;
1324 lp_build_sample_partial_offset(bld
,
1325 format_desc
->block
.height
,
1328 offset
= lp_build_add(bld
, offset
, y_offset
);
1334 if (z
&& z_stride
) {
1335 LLVMValueRef z_offset
;
1337 lp_build_sample_partial_offset(bld
,
1338 1, /* pixel blocks are always 2D */
1341 offset
= lp_build_add(bld
, offset
, z_offset
);
1344 *out_offset
= offset
;