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
;
101 memset(state
, 0, sizeof *state
);
103 if (!sampler
|| !view
|| !view
->texture
)
106 texture
= view
->texture
;
109 * We don't copy sampler state over unless it is actually enabled, to avoid
110 * spurious recompiles, as the sampler static state is part of the shader
113 * Ideally the state tracker or cso_cache module would make all state
114 * canonical, but until that happens it's better to be safe than sorry here.
116 * XXX: Actually there's much more than can be done here, especially
117 * regarding 1D/2D/3D/CUBE textures, wrap modes, etc.
120 state
->format
= view
->format
;
121 state
->swizzle_r
= view
->swizzle_r
;
122 state
->swizzle_g
= view
->swizzle_g
;
123 state
->swizzle_b
= view
->swizzle_b
;
124 state
->swizzle_a
= view
->swizzle_a
;
126 state
->target
= texture
->target
;
127 state
->pot_width
= util_is_power_of_two(texture
->width0
);
128 state
->pot_height
= util_is_power_of_two(texture
->height0
);
129 state
->pot_depth
= util_is_power_of_two(texture
->depth0
);
131 state
->wrap_s
= sampler
->wrap_s
;
132 state
->wrap_t
= sampler
->wrap_t
;
133 state
->wrap_r
= sampler
->wrap_r
;
134 state
->min_img_filter
= sampler
->min_img_filter
;
135 state
->mag_img_filter
= sampler
->mag_img_filter
;
137 if (view
->u
.tex
.last_level
&& sampler
->max_lod
> 0.0f
) {
138 state
->min_mip_filter
= sampler
->min_mip_filter
;
140 state
->min_mip_filter
= PIPE_TEX_MIPFILTER_NONE
;
143 if (state
->min_mip_filter
!= PIPE_TEX_MIPFILTER_NONE
) {
144 if (sampler
->lod_bias
!= 0.0f
) {
145 state
->lod_bias_non_zero
= 1;
148 /* If min_lod == max_lod we can greatly simplify mipmap selection.
149 * This is a case that occurs during automatic mipmap generation.
151 if (sampler
->min_lod
== sampler
->max_lod
) {
152 state
->min_max_lod_equal
= 1;
154 if (sampler
->min_lod
> 0.0f
) {
155 state
->apply_min_lod
= 1;
158 if (sampler
->max_lod
< (float)view
->u
.tex
.last_level
) {
159 state
->apply_max_lod
= 1;
164 state
->compare_mode
= sampler
->compare_mode
;
165 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
) {
166 state
->compare_func
= sampler
->compare_func
;
169 state
->normalized_coords
= sampler
->normalized_coords
;
172 * FIXME: Handle the remainder of pipe_sampler_view.
178 * Generate code to compute coordinate gradient (rho).
179 * \param derivs partial derivatives of (s, t, r, q) with respect to X and Y
181 * The resulting rho is scalar per quad.
184 lp_build_rho(struct lp_build_sample_context
*bld
,
186 const struct lp_derivatives
*derivs
)
188 struct gallivm_state
*gallivm
= bld
->gallivm
;
189 struct lp_build_context
*int_size_bld
= &bld
->int_size_in_bld
;
190 struct lp_build_context
*float_size_bld
= &bld
->float_size_in_bld
;
191 struct lp_build_context
*float_bld
= &bld
->float_bld
;
192 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
193 struct lp_build_context
*perquadf_bld
= &bld
->perquadf_bld
;
194 const LLVMValueRef
*ddx_ddy
= derivs
->ddx_ddy
;
195 const unsigned dims
= bld
->dims
;
196 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
197 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
198 LLVMValueRef index0
= LLVMConstInt(i32t
, 0, 0);
199 LLVMValueRef index1
= LLVMConstInt(i32t
, 1, 0);
200 LLVMValueRef index2
= LLVMConstInt(i32t
, 2, 0);
201 LLVMValueRef rho_vec
;
202 LLVMValueRef int_size
, float_size
;
204 LLVMValueRef first_level
, first_level_vec
;
205 LLVMValueRef abs_ddx_ddy
[2];
206 unsigned length
= coord_bld
->type
.length
;
207 unsigned num_quads
= length
/ 4;
209 LLVMValueRef i32undef
= LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
210 LLVMValueRef rho_xvec
, rho_yvec
;
212 abs_ddx_ddy
[0] = lp_build_abs(coord_bld
, ddx_ddy
[0]);
214 abs_ddx_ddy
[1] = lp_build_abs(coord_bld
, ddx_ddy
[1]);
217 abs_ddx_ddy
[1] = NULL
;
221 static const unsigned char swizzle1
[] = {
222 0, LP_BLD_SWIZZLE_DONTCARE
,
223 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
225 static const unsigned char swizzle2
[] = {
226 1, LP_BLD_SWIZZLE_DONTCARE
,
227 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
229 rho_xvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle1
);
230 rho_yvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle2
);
232 else if (dims
== 2) {
233 static const unsigned char swizzle1
[] = {
235 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
237 static const unsigned char swizzle2
[] = {
239 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
241 rho_xvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle1
);
242 rho_yvec
= lp_build_swizzle_aos(coord_bld
, abs_ddx_ddy
[0], swizzle2
);
245 LLVMValueRef shuffles1
[LP_MAX_VECTOR_LENGTH
];
246 LLVMValueRef shuffles2
[LP_MAX_VECTOR_LENGTH
];
248 for (i
= 0; i
< num_quads
; i
++) {
249 shuffles1
[4*i
+ 0] = lp_build_const_int32(gallivm
, 4*i
);
250 shuffles1
[4*i
+ 1] = lp_build_const_int32(gallivm
, 4*i
+ 2);
251 shuffles1
[4*i
+ 2] = lp_build_const_int32(gallivm
, length
+ 4*i
);
252 shuffles1
[4*i
+ 3] = i32undef
;
253 shuffles2
[4*i
+ 0] = lp_build_const_int32(gallivm
, 4*i
+ 1);
254 shuffles2
[4*i
+ 1] = lp_build_const_int32(gallivm
, 4*i
+ 3);
255 shuffles2
[4*i
+ 2] = lp_build_const_int32(gallivm
, length
+ 4*i
+ 1);
256 shuffles2
[4*i
+ 3] = i32undef
;
258 rho_xvec
= LLVMBuildShuffleVector(builder
, abs_ddx_ddy
[0], abs_ddx_ddy
[1],
259 LLVMConstVector(shuffles1
, length
), "");
260 rho_yvec
= LLVMBuildShuffleVector(builder
, abs_ddx_ddy
[0], abs_ddx_ddy
[1],
261 LLVMConstVector(shuffles2
, length
), "");
264 rho_vec
= lp_build_max(coord_bld
, rho_xvec
, rho_yvec
);
266 first_level
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
268 first_level_vec
= lp_build_broadcast_scalar(&bld
->int_size_bld
, first_level
);
269 int_size
= lp_build_minify(int_size_bld
, bld
->int_size
, first_level_vec
);
270 float_size
= lp_build_int_to_float(float_size_bld
, int_size
);
272 if (bld
->coord_type
.length
> 4) {
273 /* expand size to each quad */
275 /* could use some broadcast_vector helper for this? */
276 int num_quads
= bld
->coord_type
.length
/ 4;
277 LLVMValueRef src
[LP_MAX_VECTOR_LENGTH
/4];
278 for (i
= 0; i
< num_quads
; i
++) {
281 float_size
= lp_build_concat(bld
->gallivm
, src
, float_size_bld
->type
, num_quads
);
284 float_size
= lp_build_broadcast_scalar(coord_bld
, float_size
);
286 rho_vec
= lp_build_mul(coord_bld
, rho_vec
, float_size
);
293 static const unsigned char swizzle1
[] = {
294 0, LP_BLD_SWIZZLE_DONTCARE
,
295 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
297 static const unsigned char swizzle2
[] = {
298 1, LP_BLD_SWIZZLE_DONTCARE
,
299 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
301 LLVMValueRef rho_s
, rho_t
, rho_r
;
303 rho_s
= lp_build_swizzle_aos(coord_bld
, rho_vec
, swizzle1
);
304 rho_t
= lp_build_swizzle_aos(coord_bld
, rho_vec
, swizzle2
);
306 rho
= lp_build_max(coord_bld
, rho_s
, rho_t
);
309 static const unsigned char swizzle3
[] = {
310 2, LP_BLD_SWIZZLE_DONTCARE
,
311 LP_BLD_SWIZZLE_DONTCARE
, LP_BLD_SWIZZLE_DONTCARE
313 rho_r
= lp_build_swizzle_aos(coord_bld
, rho_vec
, swizzle3
);
314 rho
= lp_build_max(coord_bld
, rho
, rho_r
);
318 rho
= lp_build_pack_aos_scalars(bld
->gallivm
, coord_bld
->type
,
319 perquadf_bld
->type
, rho
, 0);
323 rho_vec
= LLVMBuildExtractElement(builder
, rho_vec
, index0
, "");
325 rho_vec
= lp_build_mul(float_size_bld
, rho_vec
, float_size
);
332 LLVMValueRef rho_s
, rho_t
, rho_r
;
334 rho_s
= LLVMBuildExtractElement(builder
, rho_vec
, index0
, "");
335 rho_t
= LLVMBuildExtractElement(builder
, rho_vec
, index1
, "");
337 rho
= lp_build_max(float_bld
, rho_s
, rho_t
);
340 rho_r
= LLVMBuildExtractElement(builder
, rho_vec
, index2
, "");
341 rho
= lp_build_max(float_bld
, rho
, rho_r
);
352 * Bri-linear lod computation
354 * Use a piece-wise linear approximation of log2 such that:
355 * - round to nearest, for values in the neighborhood of -1, 0, 1, 2, etc.
356 * - linear approximation for values in the neighborhood of 0.5, 1.5., etc,
357 * with the steepness specified in 'factor'
358 * - exact result for 0.5, 1.5, etc.
374 * This is a technique also commonly used in hardware:
375 * - http://ixbtlabs.com/articles2/gffx/nv40-rx800-3.html
377 * TODO: For correctness, this should only be applied when texture is known to
378 * have regular mipmaps, i.e., mipmaps derived from the base level.
380 * TODO: This could be done in fixed point, where applicable.
383 lp_build_brilinear_lod(struct lp_build_context
*bld
,
386 LLVMValueRef
*out_lod_ipart
,
387 LLVMValueRef
*out_lod_fpart
)
389 LLVMValueRef lod_fpart
;
390 double pre_offset
= (factor
- 0.5)/factor
- 0.5;
391 double post_offset
= 1 - factor
;
394 lp_build_printf(bld
->gallivm
, "lod = %f\n", lod
);
397 lod
= lp_build_add(bld
, lod
,
398 lp_build_const_vec(bld
->gallivm
, bld
->type
, pre_offset
));
400 lp_build_ifloor_fract(bld
, lod
, out_lod_ipart
, &lod_fpart
);
402 lod_fpart
= lp_build_mul(bld
, lod_fpart
,
403 lp_build_const_vec(bld
->gallivm
, bld
->type
, factor
));
405 lod_fpart
= lp_build_add(bld
, lod_fpart
,
406 lp_build_const_vec(bld
->gallivm
, bld
->type
, post_offset
));
409 * It's not necessary to clamp lod_fpart since:
410 * - the above expression will never produce numbers greater than one.
411 * - the mip filtering branch is only taken if lod_fpart is positive
414 *out_lod_fpart
= lod_fpart
;
417 lp_build_printf(bld
->gallivm
, "lod_ipart = %i\n", *out_lod_ipart
);
418 lp_build_printf(bld
->gallivm
, "lod_fpart = %f\n\n", *out_lod_fpart
);
424 * Combined log2 and brilinear lod computation.
426 * It's in all identical to calling lp_build_fast_log2() and
427 * lp_build_brilinear_lod() above, but by combining we can compute the integer
428 * and fractional part independently.
431 lp_build_brilinear_rho(struct lp_build_context
*bld
,
434 LLVMValueRef
*out_lod_ipart
,
435 LLVMValueRef
*out_lod_fpart
)
437 LLVMValueRef lod_ipart
;
438 LLVMValueRef lod_fpart
;
440 const double pre_factor
= (2*factor
- 0.5)/(M_SQRT2
*factor
);
441 const double post_offset
= 1 - 2*factor
;
443 assert(bld
->type
.floating
);
445 assert(lp_check_value(bld
->type
, rho
));
448 * The pre factor will make the intersections with the exact powers of two
449 * happen precisely where we want then to be, which means that the integer
450 * part will not need any post adjustments.
452 rho
= lp_build_mul(bld
, rho
,
453 lp_build_const_vec(bld
->gallivm
, bld
->type
, pre_factor
));
455 /* ipart = ifloor(log2(rho)) */
456 lod_ipart
= lp_build_extract_exponent(bld
, rho
, 0);
458 /* fpart = rho / 2**ipart */
459 lod_fpart
= lp_build_extract_mantissa(bld
, rho
);
461 lod_fpart
= lp_build_mul(bld
, lod_fpart
,
462 lp_build_const_vec(bld
->gallivm
, bld
->type
, factor
));
464 lod_fpart
= lp_build_add(bld
, lod_fpart
,
465 lp_build_const_vec(bld
->gallivm
, bld
->type
, post_offset
));
468 * Like lp_build_brilinear_lod, it's not necessary to clamp lod_fpart since:
469 * - the above expression will never produce numbers greater than one.
470 * - the mip filtering branch is only taken if lod_fpart is positive
473 *out_lod_ipart
= lod_ipart
;
474 *out_lod_fpart
= lod_fpart
;
479 * Generate code to compute texture level of detail (lambda).
480 * \param derivs partial derivatives of (s, t, r, q) with respect to X and Y
481 * \param lod_bias optional float vector with the shader lod bias
482 * \param explicit_lod optional float vector with the explicit lod
483 * \param width scalar int texture width
484 * \param height scalar int texture height
485 * \param depth scalar int texture depth
487 * The resulting lod is scalar per quad, so only the first value per quad
488 * passed in from lod_bias, explicit_lod is used.
491 lp_build_lod_selector(struct lp_build_sample_context
*bld
,
493 const struct lp_derivatives
*derivs
,
494 LLVMValueRef lod_bias
, /* optional */
495 LLVMValueRef explicit_lod
, /* optional */
497 LLVMValueRef
*out_lod_ipart
,
498 LLVMValueRef
*out_lod_fpart
)
501 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
502 struct lp_build_context
*perquadf_bld
= &bld
->perquadf_bld
;
505 *out_lod_ipart
= bld
->perquadi_bld
.zero
;
506 *out_lod_fpart
= perquadf_bld
->zero
;
508 if (bld
->static_state
->min_max_lod_equal
) {
509 /* User is forcing sampling from a particular mipmap level.
510 * This is hit during mipmap generation.
512 LLVMValueRef min_lod
=
513 bld
->dynamic_state
->min_lod(bld
->dynamic_state
, bld
->gallivm
, unit
);
515 lod
= lp_build_broadcast_scalar(perquadf_bld
, min_lod
);
519 lod
= lp_build_pack_aos_scalars(bld
->gallivm
, bld
->coord_bld
.type
,
520 perquadf_bld
->type
, explicit_lod
, 0);
525 rho
= lp_build_rho(bld
, unit
, derivs
);
528 * Compute lod = log2(rho)
532 !bld
->static_state
->lod_bias_non_zero
&&
533 !bld
->static_state
->apply_max_lod
&&
534 !bld
->static_state
->apply_min_lod
) {
536 * Special case when there are no post-log2 adjustments, which
537 * saves instructions but keeping the integer and fractional lod
538 * computations separate from the start.
541 if (mip_filter
== PIPE_TEX_MIPFILTER_NONE
||
542 mip_filter
== PIPE_TEX_MIPFILTER_NEAREST
) {
543 *out_lod_ipart
= lp_build_ilog2(perquadf_bld
, rho
);
544 *out_lod_fpart
= perquadf_bld
->zero
;
547 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
&&
548 !(gallivm_debug
& GALLIVM_DEBUG_NO_BRILINEAR
)) {
549 lp_build_brilinear_rho(perquadf_bld
, rho
, BRILINEAR_FACTOR
,
550 out_lod_ipart
, out_lod_fpart
);
556 lod
= lp_build_log2(perquadf_bld
, rho
);
559 lod
= lp_build_fast_log2(perquadf_bld
, rho
);
562 /* add shader lod bias */
564 lod_bias
= lp_build_pack_aos_scalars(bld
->gallivm
, bld
->coord_bld
.type
,
565 perquadf_bld
->type
, lod_bias
, 0);
566 lod
= LLVMBuildFAdd(builder
, lod
, lod_bias
, "shader_lod_bias");
570 /* add sampler lod bias */
571 if (bld
->static_state
->lod_bias_non_zero
) {
572 LLVMValueRef sampler_lod_bias
=
573 bld
->dynamic_state
->lod_bias(bld
->dynamic_state
, bld
->gallivm
, unit
);
574 sampler_lod_bias
= lp_build_broadcast_scalar(perquadf_bld
,
576 lod
= LLVMBuildFAdd(builder
, lod
, sampler_lod_bias
, "sampler_lod_bias");
580 if (bld
->static_state
->apply_max_lod
) {
581 LLVMValueRef max_lod
=
582 bld
->dynamic_state
->max_lod(bld
->dynamic_state
, bld
->gallivm
, unit
);
583 max_lod
= lp_build_broadcast_scalar(perquadf_bld
, max_lod
);
585 lod
= lp_build_min(perquadf_bld
, lod
, max_lod
);
587 if (bld
->static_state
->apply_min_lod
) {
588 LLVMValueRef min_lod
=
589 bld
->dynamic_state
->min_lod(bld
->dynamic_state
, bld
->gallivm
, unit
);
590 min_lod
= lp_build_broadcast_scalar(perquadf_bld
, min_lod
);
592 lod
= lp_build_max(perquadf_bld
, lod
, min_lod
);
596 if (mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) {
597 if (!(gallivm_debug
& GALLIVM_DEBUG_NO_BRILINEAR
)) {
598 lp_build_brilinear_lod(perquadf_bld
, lod
, BRILINEAR_FACTOR
,
599 out_lod_ipart
, out_lod_fpart
);
602 lp_build_ifloor_fract(perquadf_bld
, lod
, out_lod_ipart
, out_lod_fpart
);
605 lp_build_name(*out_lod_fpart
, "lod_fpart");
608 *out_lod_ipart
= lp_build_iround(perquadf_bld
, lod
);
611 lp_build_name(*out_lod_ipart
, "lod_ipart");
618 * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
619 * mipmap level index.
620 * Note: this is all scalar per quad code.
621 * \param lod_ipart int texture level of detail
622 * \param level_out returns integer
625 lp_build_nearest_mip_level(struct lp_build_sample_context
*bld
,
627 LLVMValueRef lod_ipart
,
628 LLVMValueRef
*level_out
)
630 struct lp_build_context
*perquadi_bld
= &bld
->perquadi_bld
;
631 LLVMValueRef first_level
, last_level
, level
;
633 first_level
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
635 last_level
= bld
->dynamic_state
->last_level(bld
->dynamic_state
,
637 first_level
= lp_build_broadcast_scalar(perquadi_bld
, first_level
);
638 last_level
= lp_build_broadcast_scalar(perquadi_bld
, last_level
);
640 level
= lp_build_add(perquadi_bld
, lod_ipart
, first_level
);
642 /* clamp level to legal range of levels */
643 *level_out
= lp_build_clamp(perquadi_bld
, level
, first_level
, last_level
);
648 * For PIPE_TEX_MIPFILTER_LINEAR, convert per-quad int LOD(s) to two (per-quad)
649 * (adjacent) mipmap level indexes, and fix up float lod part accordingly.
650 * Later, we'll sample from those two mipmap levels and interpolate between them.
653 lp_build_linear_mip_levels(struct lp_build_sample_context
*bld
,
655 LLVMValueRef lod_ipart
,
656 LLVMValueRef
*lod_fpart_inout
,
657 LLVMValueRef
*level0_out
,
658 LLVMValueRef
*level1_out
)
660 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
661 struct lp_build_context
*perquadi_bld
= &bld
->perquadi_bld
;
662 struct lp_build_context
*perquadf_bld
= &bld
->perquadf_bld
;
663 LLVMValueRef first_level
, last_level
;
664 LLVMValueRef clamp_min
;
665 LLVMValueRef clamp_max
;
667 first_level
= bld
->dynamic_state
->first_level(bld
->dynamic_state
,
669 last_level
= bld
->dynamic_state
->last_level(bld
->dynamic_state
,
671 first_level
= lp_build_broadcast_scalar(perquadi_bld
, first_level
);
672 last_level
= lp_build_broadcast_scalar(perquadi_bld
, last_level
);
674 *level0_out
= lp_build_add(perquadi_bld
, lod_ipart
, first_level
);
675 *level1_out
= lp_build_add(perquadi_bld
, *level0_out
, perquadi_bld
->one
);
678 * Clamp both *level0_out and *level1_out to [first_level, last_level], with
679 * the minimum number of comparisons, and zeroing lod_fpart in the extreme
680 * ends in the process.
684 * This code (vector select in particular) only works with llvm 3.1
685 * (if there's more than one quad, with x86 backend). Might consider
686 * converting to our lp_bld_logic helpers.
688 #if HAVE_LLVM < 0x0301
689 assert(perquadi_bld
->type
.length
== 1);
692 /* *level0_out < first_level */
693 clamp_min
= LLVMBuildICmp(builder
, LLVMIntSLT
,
694 *level0_out
, first_level
,
695 "clamp_lod_to_first");
697 *level0_out
= LLVMBuildSelect(builder
, clamp_min
,
698 first_level
, *level0_out
, "");
700 *level1_out
= LLVMBuildSelect(builder
, clamp_min
,
701 first_level
, *level1_out
, "");
703 *lod_fpart_inout
= LLVMBuildSelect(builder
, clamp_min
,
704 perquadf_bld
->zero
, *lod_fpart_inout
, "");
706 /* *level0_out >= last_level */
707 clamp_max
= LLVMBuildICmp(builder
, LLVMIntSGE
,
708 *level0_out
, last_level
,
709 "clamp_lod_to_last");
711 *level0_out
= LLVMBuildSelect(builder
, clamp_max
,
712 last_level
, *level0_out
, "");
714 *level1_out
= LLVMBuildSelect(builder
, clamp_max
,
715 last_level
, *level1_out
, "");
717 *lod_fpart_inout
= LLVMBuildSelect(builder
, clamp_max
,
718 perquadf_bld
->zero
, *lod_fpart_inout
, "");
720 lp_build_name(*level0_out
, "sampler%u_miplevel0", unit
);
721 lp_build_name(*level1_out
, "sampler%u_miplevel1", unit
);
722 lp_build_name(*lod_fpart_inout
, "sampler%u_mipweight", unit
);
727 * Return pointer to a single mipmap level.
728 * \param level integer mipmap level
731 lp_build_get_mipmap_level(struct lp_build_sample_context
*bld
,
734 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
735 LLVMValueRef indexes
[2], data_ptr
, mip_offset
;
737 indexes
[0] = lp_build_const_int32(bld
->gallivm
, 0);
739 mip_offset
= LLVMBuildGEP(builder
, bld
->mip_offsets
, indexes
, 2, "");
740 mip_offset
= LLVMBuildLoad(builder
, mip_offset
, "");
741 data_ptr
= LLVMBuildGEP(builder
, bld
->base_ptr
, &mip_offset
, 1, "");
746 * Return (per-pixel) offsets to mip levels.
747 * \param level integer mipmap level
750 lp_build_get_mip_offsets(struct lp_build_sample_context
*bld
,
753 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
754 LLVMValueRef indexes
[2], offsets
, offset1
;
756 indexes
[0] = lp_build_const_int32(bld
->gallivm
, 0);
757 if (bld
->num_lods
== 1) {
759 offset1
= LLVMBuildGEP(builder
, bld
->mip_offsets
, indexes
, 2, "");
760 offset1
= LLVMBuildLoad(builder
, offset1
, "");
761 offsets
= lp_build_broadcast_scalar(&bld
->int_coord_bld
, offset1
);
763 else if (bld
->num_lods
== bld
->coord_bld
.type
.length
/ 4) {
766 offsets
= bld
->int_coord_bld
.undef
;
767 for (i
= 0; i
< bld
->num_lods
; i
++) {
768 LLVMValueRef indexi
= lp_build_const_int32(bld
->gallivm
, i
);
769 LLVMValueRef indexo
= lp_build_const_int32(bld
->gallivm
, 4 * i
);
770 indexes
[1] = LLVMBuildExtractElement(builder
, level
, indexi
, "");
771 offset1
= LLVMBuildGEP(builder
, bld
->mip_offsets
, indexes
, 2, "");
772 offset1
= LLVMBuildLoad(builder
, offset1
, "");
773 offsets
= LLVMBuildInsertElement(builder
, offsets
, offset1
, indexo
, "");
775 offsets
= lp_build_swizzle_scalar_aos(&bld
->int_coord_bld
, offsets
, 0);
780 assert (bld
->num_lods
== bld
->coord_bld
.type
.length
);
782 offsets
= bld
->int_coord_bld
.undef
;
783 for (i
= 0; i
< bld
->num_lods
; i
++) {
784 LLVMValueRef indexi
= lp_build_const_int32(bld
->gallivm
, i
);
785 indexes
[1] = LLVMBuildExtractElement(builder
, level
, indexi
, "");
786 offset1
= LLVMBuildGEP(builder
, bld
->mip_offsets
, indexes
, 2, "");
787 offset1
= LLVMBuildLoad(builder
, offset1
, "");
788 offsets
= LLVMBuildInsertElement(builder
, offsets
, offset1
, indexi
, "");
796 * Codegen equivalent for u_minify().
797 * Return max(1, base_size >> level);
800 lp_build_minify(struct lp_build_context
*bld
,
801 LLVMValueRef base_size
,
804 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
805 assert(lp_check_value(bld
->type
, base_size
));
806 assert(lp_check_value(bld
->type
, level
));
808 if (level
== bld
->zero
) {
809 /* if we're using mipmap level zero, no minification is needed */
814 LLVMBuildLShr(builder
, base_size
, level
, "minify");
815 assert(bld
->type
.sign
);
816 size
= lp_build_max(bld
, size
, bld
->one
);
823 * Dereference stride_array[mipmap_level] array to get a stride.
824 * Return stride as a vector.
827 lp_build_get_level_stride_vec(struct lp_build_sample_context
*bld
,
828 LLVMValueRef stride_array
, LLVMValueRef level
)
830 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
831 LLVMValueRef indexes
[2], stride
, stride1
;
832 indexes
[0] = lp_build_const_int32(bld
->gallivm
, 0);
833 if (bld
->num_lods
== 1) {
835 stride1
= LLVMBuildGEP(builder
, stride_array
, indexes
, 2, "");
836 stride1
= LLVMBuildLoad(builder
, stride1
, "");
837 stride
= lp_build_broadcast_scalar(&bld
->int_coord_bld
, stride1
);
839 else if (bld
->num_lods
== bld
->coord_bld
.type
.length
/ 4) {
840 LLVMValueRef stride1
;
843 stride
= bld
->int_coord_bld
.undef
;
844 for (i
= 0; i
< bld
->num_lods
; i
++) {
845 LLVMValueRef indexi
= lp_build_const_int32(bld
->gallivm
, i
);
846 LLVMValueRef indexo
= lp_build_const_int32(bld
->gallivm
, i
);
847 indexes
[1] = LLVMBuildExtractElement(builder
, level
, indexi
, "");
848 stride1
= LLVMBuildGEP(builder
, stride_array
, indexes
, 2, "");
849 stride1
= LLVMBuildLoad(builder
, stride1
, "");
850 stride
= LLVMBuildInsertElement(builder
, stride
, stride1
, indexo
, "");
852 stride
= lp_build_swizzle_scalar_aos(&bld
->int_coord_bld
, stride
, 0);
855 LLVMValueRef stride1
;
858 assert (bld
->num_lods
== bld
->coord_bld
.type
.length
);
860 stride
= bld
->int_coord_bld
.undef
;
861 for (i
= 0; i
< bld
->coord_bld
.type
.length
; i
++) {
862 LLVMValueRef indexi
= lp_build_const_int32(bld
->gallivm
, i
);
863 indexes
[1] = LLVMBuildExtractElement(builder
, level
, indexi
, "");
864 stride1
= LLVMBuildGEP(builder
, stride_array
, indexes
, 2, "");
865 stride1
= LLVMBuildLoad(builder
, stride1
, "");
866 stride
= LLVMBuildInsertElement(builder
, stride
, stride1
, indexi
, "");
874 * When sampling a mipmap, we need to compute the width, height, depth
875 * of the source levels from the level indexes. This helper function
879 lp_build_mipmap_level_sizes(struct lp_build_sample_context
*bld
,
881 LLVMValueRef
*out_size
,
882 LLVMValueRef
*row_stride_vec
,
883 LLVMValueRef
*img_stride_vec
)
885 const unsigned dims
= bld
->dims
;
886 LLVMValueRef ilevel_vec
;
889 * Compute width, height, depth at mipmap level 'ilevel'
891 if (bld
->num_lods
== 1) {
892 ilevel_vec
= lp_build_broadcast_scalar(&bld
->int_size_bld
, ilevel
);
893 *out_size
= lp_build_minify(&bld
->int_size_bld
, bld
->int_size
, ilevel_vec
);
896 LLVMValueRef int_size_vec
;
897 LLVMValueRef tmp
[LP_MAX_VECTOR_LENGTH
];
898 unsigned num_quads
= bld
->coord_bld
.type
.length
/ 4;
901 if (bld
->num_lods
== num_quads
) {
903 * XXX: this should be #ifndef SANE_INSTRUCTION_SET.
904 * intel "forgot" the variable shift count instruction until avx2.
905 * A harmless 8x32 shift gets translated into 32 instructions
906 * (16 extracts, 8 scalar shifts, 8 inserts), llvm is apparently
907 * unable to recognize if there are really just 2 different shift
908 * count values. So do the shift 4-wide before expansion.
910 struct lp_build_context bld4
;
911 struct lp_type type4
;
913 type4
= bld
->int_coord_bld
.type
;
916 lp_build_context_init(&bld4
, bld
->gallivm
, type4
);
918 if (bld
->dims
== 1) {
919 assert(bld
->int_size_in_bld
.type
.length
== 1);
920 int_size_vec
= lp_build_broadcast_scalar(&bld4
,
924 assert(bld
->int_size_in_bld
.type
.length
== 4);
925 int_size_vec
= bld
->int_size
;
928 for (i
= 0; i
< num_quads
; i
++) {
929 LLVMValueRef ileveli
;
930 LLVMValueRef indexi
= lp_build_const_int32(bld
->gallivm
, i
);
932 ileveli
= lp_build_extract_broadcast(bld
->gallivm
,
933 bld
->perquadi_bld
.type
,
937 tmp
[i
] = lp_build_minify(&bld4
, int_size_vec
, ileveli
);
940 * out_size is [w0, h0, d0, _, w1, h1, d1, _, ...] vector for dims > 1,
941 * [w0, w0, w0, w0, w1, w1, w1, w1, ...] otherwise.
943 *out_size
= lp_build_concat(bld
->gallivm
,
949 /* FIXME: this is terrible and results in _huge_ vector
950 * (for the dims > 1 case).
951 * Should refactor this (together with extract_image_sizes) and do
952 * something more useful. Could for instance if we have width,height
953 * with 4-wide vector pack all elements into a 8xi16 vector
954 * (on which we can still do useful math) instead of using a 16xi32
956 * FIXME: some callers can't handle this yet.
957 * For dims == 1 this will create [w0, w1, w2, w3, ...] vector.
958 * For dims > 1 this will create [w0, h0, d0, _, w1, h1, d1, _, ...] vector.
960 assert(bld
->num_lods
== bld
->coord_bld
.type
.length
);
961 if (bld
->dims
== 1) {
962 assert(bld
->int_size_bld
.type
.length
== 1);
963 int_size_vec
= lp_build_broadcast_scalar(&bld
->int_coord_bld
,
965 /* vector shift with variable shift count alert... */
966 *out_size
= lp_build_minify(&bld
->int_coord_bld
, int_size_vec
, ilevel
);
969 LLVMValueRef ilevel1
;
970 for (i
= 0; i
< bld
->num_lods
; i
++) {
971 LLVMValueRef indexi
= lp_build_const_int32(bld
->gallivm
, i
);
972 ilevel1
= lp_build_extract_broadcast(bld
->gallivm
, bld
->int_coord_type
,
973 bld
->int_size_in_bld
.type
, ilevel
, indexi
);
974 tmp
[i
] = bld
->int_size
;
975 tmp
[i
] = lp_build_minify(&bld
->int_size_in_bld
, tmp
[i
], ilevel1
);
977 int_size_vec
= lp_build_concat(bld
->gallivm
,
979 bld
->int_size_in_bld
.type
,
986 *row_stride_vec
= lp_build_get_level_stride_vec(bld
,
987 bld
->row_stride_array
,
989 if (dims
== 3 || bld
->static_state
->target
== PIPE_TEXTURE_CUBE
) {
990 *img_stride_vec
= lp_build_get_level_stride_vec(bld
,
991 bld
->img_stride_array
,
999 * Extract and broadcast texture size.
1001 * @param size_type type of the texture size vector (either
1002 * bld->int_size_type or bld->float_size_type)
1003 * @param coord_type type of the texture size vector (either
1004 * bld->int_coord_type or bld->coord_type)
1005 * @param size vector with the texture size (width, height, depth)
1008 lp_build_extract_image_sizes(struct lp_build_sample_context
*bld
,
1009 struct lp_build_context
*size_bld
,
1010 struct lp_type coord_type
,
1012 LLVMValueRef
*out_width
,
1013 LLVMValueRef
*out_height
,
1014 LLVMValueRef
*out_depth
)
1016 const unsigned dims
= bld
->dims
;
1017 LLVMTypeRef i32t
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
1018 struct lp_type size_type
= size_bld
->type
;
1020 if (bld
->num_lods
== 1) {
1021 *out_width
= lp_build_extract_broadcast(bld
->gallivm
,
1025 LLVMConstInt(i32t
, 0, 0));
1027 *out_height
= lp_build_extract_broadcast(bld
->gallivm
,
1031 LLVMConstInt(i32t
, 1, 0));
1033 *out_depth
= lp_build_extract_broadcast(bld
->gallivm
,
1037 LLVMConstInt(i32t
, 2, 0));
1042 unsigned num_quads
= bld
->coord_bld
.type
.length
/ 4;
1047 else if (bld
->num_lods
== num_quads
) {
1048 *out_width
= lp_build_swizzle_scalar_aos(size_bld
, size
, 0);
1050 *out_height
= lp_build_swizzle_scalar_aos(size_bld
, size
, 1);
1052 *out_depth
= lp_build_swizzle_scalar_aos(size_bld
, size
, 2);
1057 assert(bld
->num_lods
== bld
->coord_type
.length
);
1058 *out_width
= lp_build_pack_aos_scalars(bld
->gallivm
, size_type
,
1059 coord_type
, size
, 0);
1061 *out_width
= lp_build_pack_aos_scalars(bld
->gallivm
, size_type
,
1062 coord_type
, size
, 1);
1064 *out_width
= lp_build_pack_aos_scalars(bld
->gallivm
, size_type
,
1065 coord_type
, size
, 2);
1074 * Unnormalize coords.
1076 * @param flt_size vector with the integer texture size (width, height, depth)
1079 lp_build_unnormalized_coords(struct lp_build_sample_context
*bld
,
1080 LLVMValueRef flt_size
,
1085 const unsigned dims
= bld
->dims
;
1087 LLVMValueRef height
;
1090 lp_build_extract_image_sizes(bld
,
1091 &bld
->float_size_bld
,
1098 /* s = s * width, t = t * height */
1099 *s
= lp_build_mul(&bld
->coord_bld
, *s
, width
);
1101 *t
= lp_build_mul(&bld
->coord_bld
, *t
, height
);
1103 *r
= lp_build_mul(&bld
->coord_bld
, *r
, depth
);
1109 /** Helper used by lp_build_cube_lookup() */
1111 lp_build_cube_imapos(struct lp_build_context
*coord_bld
, LLVMValueRef coord
)
1113 /* ima = +0.5 / abs(coord); */
1114 LLVMValueRef posHalf
= lp_build_const_vec(coord_bld
->gallivm
, coord_bld
->type
, 0.5);
1115 LLVMValueRef absCoord
= lp_build_abs(coord_bld
, coord
);
1116 LLVMValueRef ima
= lp_build_div(coord_bld
, posHalf
, absCoord
);
1120 /** Helper used by lp_build_cube_lookup() */
1122 lp_build_cube_imaneg(struct lp_build_context
*coord_bld
, LLVMValueRef coord
)
1124 /* ima = -0.5 / abs(coord); */
1125 LLVMValueRef negHalf
= lp_build_const_vec(coord_bld
->gallivm
, coord_bld
->type
, -0.5);
1126 LLVMValueRef absCoord
= lp_build_abs(coord_bld
, coord
);
1127 LLVMValueRef ima
= lp_build_div(coord_bld
, negHalf
, absCoord
);
1132 * Helper used by lp_build_cube_lookup()
1133 * FIXME: the sign here can also be 0.
1134 * Arithmetically this could definitely make a difference. Either
1135 * fix the comment or use other (simpler) sign function, not sure
1136 * which one it should be.
1137 * \param sign scalar +1 or -1
1138 * \param coord float vector
1139 * \param ima float vector
1142 lp_build_cube_coord(struct lp_build_context
*coord_bld
,
1143 LLVMValueRef sign
, int negate_coord
,
1144 LLVMValueRef coord
, LLVMValueRef ima
)
1146 /* return negate(coord) * ima * sign + 0.5; */
1147 LLVMValueRef half
= lp_build_const_vec(coord_bld
->gallivm
, coord_bld
->type
, 0.5);
1150 assert(negate_coord
== +1 || negate_coord
== -1);
1152 if (negate_coord
== -1) {
1153 coord
= lp_build_negate(coord_bld
, coord
);
1156 res
= lp_build_mul(coord_bld
, coord
, ima
);
1158 sign
= lp_build_broadcast_scalar(coord_bld
, sign
);
1159 res
= lp_build_mul(coord_bld
, res
, sign
);
1161 res
= lp_build_add(coord_bld
, res
, half
);
1167 /** Helper used by lp_build_cube_lookup()
1168 * Return (major_coord >= 0) ? pos_face : neg_face;
1171 lp_build_cube_face(struct lp_build_sample_context
*bld
,
1172 LLVMValueRef major_coord
,
1173 unsigned pos_face
, unsigned neg_face
)
1175 struct gallivm_state
*gallivm
= bld
->gallivm
;
1176 LLVMBuilderRef builder
= gallivm
->builder
;
1177 LLVMValueRef cmp
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
1179 bld
->float_bld
.zero
, "");
1180 LLVMValueRef pos
= lp_build_const_int32(gallivm
, pos_face
);
1181 LLVMValueRef neg
= lp_build_const_int32(gallivm
, neg_face
);
1182 LLVMValueRef res
= LLVMBuildSelect(builder
, cmp
, pos
, neg
, "");
1189 * Generate code to do cube face selection and compute per-face texcoords.
1192 lp_build_cube_lookup(struct lp_build_sample_context
*bld
,
1197 LLVMValueRef
*face_s
,
1198 LLVMValueRef
*face_t
)
1200 struct lp_build_context
*coord_bld
= &bld
->coord_bld
;
1201 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1202 struct gallivm_state
*gallivm
= bld
->gallivm
;
1203 LLVMValueRef rx
, ry
, rz
;
1204 LLVMValueRef tmp
[4], rxyz
, arxyz
;
1207 * Use the average of the four pixel's texcoords to choose the face.
1208 * Slight simplification just calculate the sum, skip scaling.
1213 rxyz
= lp_build_hadd_partial4(&bld
->coord_bld
, tmp
, 3);
1214 arxyz
= lp_build_abs(&bld
->coord_bld
, rxyz
);
1216 if (coord_bld
->type
.length
> 4) {
1217 struct lp_build_context
*cint_bld
= &bld
->int_coord_bld
;
1218 struct lp_type intctype
= cint_bld
->type
;
1219 LLVMValueRef signrxs
, signrys
, signrzs
, signrxyz
, sign
;
1220 LLVMValueRef arxs
, arys
, arzs
;
1221 LLVMValueRef arx_ge_ary
, maxarxsarys
, arz_ge_arx_ary
;
1222 LLVMValueRef snewx
, tnewx
, snewy
, tnewy
, snewz
, tnewz
;
1223 LLVMValueRef ryneg
, rzneg
;
1224 LLVMValueRef ma
, ima
;
1225 LLVMValueRef posHalf
= lp_build_const_vec(gallivm
, coord_bld
->type
, 0.5);
1226 LLVMValueRef signmask
= lp_build_const_int_vec(gallivm
, intctype
,
1227 1 << (intctype
.width
- 1));
1228 LLVMValueRef signshift
= lp_build_const_int_vec(gallivm
, intctype
,
1230 LLVMValueRef facex
= lp_build_const_int_vec(gallivm
, intctype
, PIPE_TEX_FACE_POS_X
);
1231 LLVMValueRef facey
= lp_build_const_int_vec(gallivm
, intctype
, PIPE_TEX_FACE_POS_Y
);
1232 LLVMValueRef facez
= lp_build_const_int_vec(gallivm
, intctype
, PIPE_TEX_FACE_POS_Z
);
1234 assert(PIPE_TEX_FACE_NEG_X
== PIPE_TEX_FACE_POS_X
+ 1);
1235 assert(PIPE_TEX_FACE_NEG_Y
== PIPE_TEX_FACE_POS_Y
+ 1);
1236 assert(PIPE_TEX_FACE_NEG_Z
== PIPE_TEX_FACE_POS_Z
+ 1);
1238 rx
= LLVMBuildBitCast(builder
, s
, lp_build_vec_type(gallivm
, intctype
), "");
1239 ry
= LLVMBuildBitCast(builder
, t
, lp_build_vec_type(gallivm
, intctype
), "");
1240 rz
= LLVMBuildBitCast(builder
, r
, lp_build_vec_type(gallivm
, intctype
), "");
1241 ryneg
= LLVMBuildXor(builder
, ry
, signmask
, "");
1242 rzneg
= LLVMBuildXor(builder
, rz
, signmask
, "");
1244 /* the sign bit comes from the averaged vector (per quad),
1245 * as does the decision which face to use */
1246 signrxyz
= LLVMBuildBitCast(builder
, rxyz
, lp_build_vec_type(gallivm
, intctype
), "");
1247 signrxyz
= LLVMBuildAnd(builder
, signrxyz
, signmask
, "");
1249 arxs
= lp_build_swizzle_scalar_aos(coord_bld
, arxyz
, 0);
1250 arys
= lp_build_swizzle_scalar_aos(coord_bld
, arxyz
, 1);
1251 arzs
= lp_build_swizzle_scalar_aos(coord_bld
, arxyz
, 2);
1254 * select x if x >= y else select y
1255 * select previous result if y >= max(x,y) else select z
1257 arx_ge_ary
= lp_build_cmp(coord_bld
, PIPE_FUNC_GEQUAL
, arxs
, arys
);
1258 maxarxsarys
= lp_build_max(coord_bld
, arxs
, arys
);
1259 arz_ge_arx_ary
= lp_build_cmp(coord_bld
, PIPE_FUNC_GEQUAL
, maxarxsarys
, arzs
);
1262 * compute all possible new s/t coords
1263 * snewx = signrx * -rz;
1266 * tnewy = signry * rz;
1267 * snewz = signrz * rx;
1270 signrxs
= lp_build_swizzle_scalar_aos(cint_bld
, signrxyz
, 0);
1271 snewx
= LLVMBuildXor(builder
, signrxs
, rzneg
, "");
1274 signrys
= lp_build_swizzle_scalar_aos(cint_bld
, signrxyz
, 1);
1276 tnewy
= LLVMBuildXor(builder
, signrys
, rz
, "");
1278 signrzs
= lp_build_swizzle_scalar_aos(cint_bld
, signrxyz
, 2);
1279 snewz
= LLVMBuildXor(builder
, signrzs
, rx
, "");
1282 /* XXX on x86 unclear if we should cast the values back to float
1283 * or not - on some cpus (nehalem) pblendvb has twice the throughput
1284 * of blendvps though on others there just might be domain
1285 * transition penalties when using it (this depends on what llvm
1286 * will chose for the bit ops above so there appears no "right way",
1287 * but given the boatload of selects let's just use the int type).
1289 * Unfortunately we also need the sign bit of the summed coords.
1291 *face_s
= lp_build_select(cint_bld
, arx_ge_ary
, snewx
, snewy
);
1292 *face_t
= lp_build_select(cint_bld
, arx_ge_ary
, tnewx
, tnewy
);
1293 ma
= lp_build_select(coord_bld
, arx_ge_ary
, s
, t
);
1294 *face
= lp_build_select(cint_bld
, arx_ge_ary
, facex
, facey
);
1295 sign
= lp_build_select(cint_bld
, arx_ge_ary
, signrxs
, signrys
);
1297 *face_s
= lp_build_select(cint_bld
, arz_ge_arx_ary
, *face_s
, snewz
);
1298 *face_t
= lp_build_select(cint_bld
, arz_ge_arx_ary
, *face_t
, tnewz
);
1299 ma
= lp_build_select(coord_bld
, arz_ge_arx_ary
, ma
, r
);
1300 *face
= lp_build_select(cint_bld
, arz_ge_arx_ary
, *face
, facez
);
1301 sign
= lp_build_select(cint_bld
, arz_ge_arx_ary
, sign
, signrzs
);
1303 *face_s
= LLVMBuildBitCast(builder
, *face_s
,
1304 lp_build_vec_type(gallivm
, coord_bld
->type
), "");
1305 *face_t
= LLVMBuildBitCast(builder
, *face_t
,
1306 lp_build_vec_type(gallivm
, coord_bld
->type
), "");
1308 /* add +1 for neg face */
1309 /* XXX with AVX probably want to use another select here -
1310 * as long as we ensure vblendvps gets used we can actually
1311 * skip the comparison and just use sign as a "mask" directly.
1313 sign
= LLVMBuildLShr(builder
, sign
, signshift
, "");
1314 *face
= LLVMBuildOr(builder
, *face
, sign
, "face");
1316 ima
= lp_build_cube_imapos(coord_bld
, ma
);
1318 *face_s
= lp_build_mul(coord_bld
, *face_s
, ima
);
1319 *face_s
= lp_build_add(coord_bld
, *face_s
, posHalf
);
1320 *face_t
= lp_build_mul(coord_bld
, *face_t
, ima
);
1321 *face_t
= lp_build_add(coord_bld
, *face_t
, posHalf
);
1325 struct lp_build_if_state if_ctx
;
1326 LLVMValueRef face_s_var
;
1327 LLVMValueRef face_t_var
;
1328 LLVMValueRef face_var
;
1329 LLVMValueRef arx_ge_ary_arz
, ary_ge_arx_arz
;
1330 LLVMValueRef shuffles
[4];
1331 LLVMValueRef arxy_ge_aryx
, arxy_ge_arzz
, arxy_ge_arxy_arzz
;
1332 LLVMValueRef arxyxy
, aryxzz
, arxyxy_ge_aryxzz
;
1333 struct lp_build_context
*float_bld
= &bld
->float_bld
;
1335 assert(bld
->coord_bld
.type
.length
== 4);
1337 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
1338 shuffles
[1] = lp_build_const_int32(gallivm
, 1);
1339 shuffles
[2] = lp_build_const_int32(gallivm
, 0);
1340 shuffles
[3] = lp_build_const_int32(gallivm
, 1);
1341 arxyxy
= LLVMBuildShuffleVector(builder
, arxyz
, arxyz
, LLVMConstVector(shuffles
, 4), "");
1342 shuffles
[0] = lp_build_const_int32(gallivm
, 1);
1343 shuffles
[1] = lp_build_const_int32(gallivm
, 0);
1344 shuffles
[2] = lp_build_const_int32(gallivm
, 2);
1345 shuffles
[3] = lp_build_const_int32(gallivm
, 2);
1346 aryxzz
= LLVMBuildShuffleVector(builder
, arxyz
, arxyz
, LLVMConstVector(shuffles
, 4), "");
1347 arxyxy_ge_aryxzz
= lp_build_cmp(&bld
->coord_bld
, PIPE_FUNC_GEQUAL
, arxyxy
, aryxzz
);
1349 shuffles
[0] = lp_build_const_int32(gallivm
, 0);
1350 shuffles
[1] = lp_build_const_int32(gallivm
, 1);
1351 arxy_ge_aryx
= LLVMBuildShuffleVector(builder
, arxyxy_ge_aryxzz
, arxyxy_ge_aryxzz
,
1352 LLVMConstVector(shuffles
, 2), "");
1353 shuffles
[0] = lp_build_const_int32(gallivm
, 2);
1354 shuffles
[1] = lp_build_const_int32(gallivm
, 3);
1355 arxy_ge_arzz
= LLVMBuildShuffleVector(builder
, arxyxy_ge_aryxzz
, arxyxy_ge_aryxzz
,
1356 LLVMConstVector(shuffles
, 2), "");
1357 arxy_ge_arxy_arzz
= LLVMBuildAnd(builder
, arxy_ge_aryx
, arxy_ge_arzz
, "");
1359 arx_ge_ary_arz
= LLVMBuildExtractElement(builder
, arxy_ge_arxy_arzz
,
1360 lp_build_const_int32(gallivm
, 0), "");
1361 arx_ge_ary_arz
= LLVMBuildICmp(builder
, LLVMIntNE
, arx_ge_ary_arz
,
1362 lp_build_const_int32(gallivm
, 0), "");
1363 ary_ge_arx_arz
= LLVMBuildExtractElement(builder
, arxy_ge_arxy_arzz
,
1364 lp_build_const_int32(gallivm
, 1), "");
1365 ary_ge_arx_arz
= LLVMBuildICmp(builder
, LLVMIntNE
, ary_ge_arx_arz
,
1366 lp_build_const_int32(gallivm
, 0), "");
1367 face_s_var
= lp_build_alloca(gallivm
, bld
->coord_bld
.vec_type
, "face_s_var");
1368 face_t_var
= lp_build_alloca(gallivm
, bld
->coord_bld
.vec_type
, "face_t_var");
1369 face_var
= lp_build_alloca(gallivm
, bld
->int_bld
.vec_type
, "face_var");
1371 lp_build_if(&if_ctx
, gallivm
, arx_ge_ary_arz
);
1374 LLVMValueRef sign
, ima
;
1375 rx
= LLVMBuildExtractElement(builder
, rxyz
,
1376 lp_build_const_int32(gallivm
, 0), "");
1378 sign
= lp_build_sgn(float_bld
, rx
);
1379 ima
= lp_build_cube_imaneg(coord_bld
, s
);
1380 *face_s
= lp_build_cube_coord(coord_bld
, sign
, +1, r
, ima
);
1381 *face_t
= lp_build_cube_coord(coord_bld
, NULL
, +1, t
, ima
);
1382 *face
= lp_build_cube_face(bld
, rx
,
1383 PIPE_TEX_FACE_POS_X
,
1384 PIPE_TEX_FACE_NEG_X
);
1385 LLVMBuildStore(builder
, *face_s
, face_s_var
);
1386 LLVMBuildStore(builder
, *face_t
, face_t_var
);
1387 LLVMBuildStore(builder
, *face
, face_var
);
1389 lp_build_else(&if_ctx
);
1391 struct lp_build_if_state if_ctx2
;
1393 lp_build_if(&if_ctx2
, gallivm
, ary_ge_arx_arz
);
1395 LLVMValueRef sign
, ima
;
1397 ry
= LLVMBuildExtractElement(builder
, rxyz
,
1398 lp_build_const_int32(gallivm
, 1), "");
1399 sign
= lp_build_sgn(float_bld
, ry
);
1400 ima
= lp_build_cube_imaneg(coord_bld
, t
);
1401 *face_s
= lp_build_cube_coord(coord_bld
, NULL
, -1, s
, ima
);
1402 *face_t
= lp_build_cube_coord(coord_bld
, sign
, -1, r
, ima
);
1403 *face
= lp_build_cube_face(bld
, ry
,
1404 PIPE_TEX_FACE_POS_Y
,
1405 PIPE_TEX_FACE_NEG_Y
);
1406 LLVMBuildStore(builder
, *face_s
, face_s_var
);
1407 LLVMBuildStore(builder
, *face_t
, face_t_var
);
1408 LLVMBuildStore(builder
, *face
, face_var
);
1410 lp_build_else(&if_ctx2
);
1413 LLVMValueRef sign
, ima
;
1414 rz
= LLVMBuildExtractElement(builder
, rxyz
,
1415 lp_build_const_int32(gallivm
, 2), "");
1416 sign
= lp_build_sgn(float_bld
, rz
);
1417 ima
= lp_build_cube_imaneg(coord_bld
, r
);
1418 *face_s
= lp_build_cube_coord(coord_bld
, sign
, -1, s
, ima
);
1419 *face_t
= lp_build_cube_coord(coord_bld
, NULL
, +1, t
, ima
);
1420 *face
= lp_build_cube_face(bld
, rz
,
1421 PIPE_TEX_FACE_POS_Z
,
1422 PIPE_TEX_FACE_NEG_Z
);
1423 LLVMBuildStore(builder
, *face_s
, face_s_var
);
1424 LLVMBuildStore(builder
, *face_t
, face_t_var
);
1425 LLVMBuildStore(builder
, *face
, face_var
);
1427 lp_build_endif(&if_ctx2
);
1430 lp_build_endif(&if_ctx
);
1432 *face_s
= LLVMBuildLoad(builder
, face_s_var
, "face_s");
1433 *face_t
= LLVMBuildLoad(builder
, face_t_var
, "face_t");
1434 *face
= LLVMBuildLoad(builder
, face_var
, "face");
1435 *face
= lp_build_broadcast_scalar(&bld
->int_coord_bld
, *face
);
1441 * Compute the partial offset of a pixel block along an arbitrary axis.
1443 * @param coord coordinate in pixels
1444 * @param stride number of bytes between rows of successive pixel blocks
1445 * @param block_length number of pixels in a pixels block along the coordinate
1447 * @param out_offset resulting relative offset of the pixel block in bytes
1448 * @param out_subcoord resulting sub-block pixel coordinate
1451 lp_build_sample_partial_offset(struct lp_build_context
*bld
,
1452 unsigned block_length
,
1454 LLVMValueRef stride
,
1455 LLVMValueRef
*out_offset
,
1456 LLVMValueRef
*out_subcoord
)
1458 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
1459 LLVMValueRef offset
;
1460 LLVMValueRef subcoord
;
1462 if (block_length
== 1) {
1463 subcoord
= bld
->zero
;
1467 * Pixel blocks have power of two dimensions. LLVM should convert the
1468 * rem/div to bit arithmetic.
1469 * TODO: Verify this.
1470 * It does indeed BUT it does transform it to scalar (and back) when doing so
1471 * (using roughly extract, shift/and, mov, unpack) (llvm 2.7).
1472 * The generated code looks seriously unfunny and is quite expensive.
1475 LLVMValueRef block_width
= lp_build_const_int_vec(bld
->type
, block_length
);
1476 subcoord
= LLVMBuildURem(builder
, coord
, block_width
, "");
1477 coord
= LLVMBuildUDiv(builder
, coord
, block_width
, "");
1479 unsigned logbase2
= util_logbase2(block_length
);
1480 LLVMValueRef block_shift
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, logbase2
);
1481 LLVMValueRef block_mask
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
, block_length
- 1);
1482 subcoord
= LLVMBuildAnd(builder
, coord
, block_mask
, "");
1483 coord
= LLVMBuildLShr(builder
, coord
, block_shift
, "");
1487 offset
= lp_build_mul(bld
, coord
, stride
);
1490 assert(out_subcoord
);
1492 *out_offset
= offset
;
1493 *out_subcoord
= subcoord
;
1498 * Compute the offset of a pixel block.
1500 * x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
1502 * Returns the relative offset and i,j sub-block coordinates
1505 lp_build_sample_offset(struct lp_build_context
*bld
,
1506 const struct util_format_description
*format_desc
,
1510 LLVMValueRef y_stride
,
1511 LLVMValueRef z_stride
,
1512 LLVMValueRef
*out_offset
,
1513 LLVMValueRef
*out_i
,
1514 LLVMValueRef
*out_j
)
1516 LLVMValueRef x_stride
;
1517 LLVMValueRef offset
;
1519 x_stride
= lp_build_const_vec(bld
->gallivm
, bld
->type
,
1520 format_desc
->block
.bits
/8);
1522 lp_build_sample_partial_offset(bld
,
1523 format_desc
->block
.width
,
1527 if (y
&& y_stride
) {
1528 LLVMValueRef y_offset
;
1529 lp_build_sample_partial_offset(bld
,
1530 format_desc
->block
.height
,
1533 offset
= lp_build_add(bld
, offset
, y_offset
);
1539 if (z
&& z_stride
) {
1540 LLVMValueRef z_offset
;
1542 lp_build_sample_partial_offset(bld
,
1543 1, /* pixel blocks are always 2D */
1546 offset
= lp_build_add(bld
, offset
, z_offset
);
1549 *out_offset
= offset
;