• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**************************************************************************
2  *
3  * Copyright 2009 VMware, Inc.
4  * All Rights Reserved.
5  *
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:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
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.
25  *
26  **************************************************************************/
27 
28 /**
29  * @file
30  * Texture sampling -- common code.
31  *
32  * @author Jose Fonseca <jfonseca@vmware.com>
33  */
34 
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 "util/u_cpu_detect.h"
40 #include "lp_bld_arit.h"
41 #include "lp_bld_const.h"
42 #include "lp_bld_debug.h"
43 #include "lp_bld_printf.h"
44 #include "lp_bld_flow.h"
45 #include "lp_bld_sample.h"
46 #include "lp_bld_swizzle.h"
47 #include "lp_bld_type.h"
48 #include "lp_bld_logic.h"
49 #include "lp_bld_pack.h"
50 #include "lp_bld_quad.h"
51 #include "lp_bld_bitarit.h"
52 
53 
54 /*
55  * Bri-linear factor. Should be greater than one.
56  */
57 #define BRILINEAR_FACTOR 2
58 
59 /**
60  * Does the given texture wrap mode allow sampling the texture border color?
61  * XXX maybe move this into gallium util code.
62  */
63 boolean
lp_sampler_wrap_mode_uses_border_color(unsigned mode,unsigned min_img_filter,unsigned mag_img_filter)64 lp_sampler_wrap_mode_uses_border_color(unsigned mode,
65                                        unsigned min_img_filter,
66                                        unsigned mag_img_filter)
67 {
68    switch (mode) {
69    case PIPE_TEX_WRAP_REPEAT:
70    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
71    case PIPE_TEX_WRAP_MIRROR_REPEAT:
72    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
73       return FALSE;
74    case PIPE_TEX_WRAP_CLAMP:
75    case PIPE_TEX_WRAP_MIRROR_CLAMP:
76       if (min_img_filter == PIPE_TEX_FILTER_NEAREST &&
77           mag_img_filter == PIPE_TEX_FILTER_NEAREST) {
78          return FALSE;
79       } else {
80          return TRUE;
81       }
82    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
83    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
84       return TRUE;
85    default:
86       assert(0 && "unexpected wrap mode");
87       return FALSE;
88    }
89 }
90 
91 
92 /**
93  * Initialize lp_sampler_static_texture_state object with the gallium
94  * texture/sampler_view state (this contains the parts which are
95  * considered static).
96  */
97 void
lp_sampler_static_texture_state(struct lp_static_texture_state * state,const struct pipe_sampler_view * view)98 lp_sampler_static_texture_state(struct lp_static_texture_state *state,
99                                 const struct pipe_sampler_view *view)
100 {
101    const struct pipe_resource *texture;
102 
103    memset(state, 0, sizeof *state);
104 
105    if (!view || !view->texture)
106       return;
107 
108    texture = view->texture;
109 
110    state->format            = view->format;
111    state->swizzle_r         = view->swizzle_r;
112    state->swizzle_g         = view->swizzle_g;
113    state->swizzle_b         = view->swizzle_b;
114    state->swizzle_a         = view->swizzle_a;
115 
116    state->target            = view->target;
117    state->pot_width         = util_is_power_of_two(texture->width0);
118    state->pot_height        = util_is_power_of_two(texture->height0);
119    state->pot_depth         = util_is_power_of_two(texture->depth0);
120    state->level_zero_only   = !view->u.tex.last_level;
121 
122    /*
123     * the layer / element / level parameters are all either dynamic
124     * state or handled transparently wrt execution.
125     */
126 }
127 
128 
129 /**
130  * Initialize lp_sampler_static_sampler_state object with the gallium sampler
131  * state (this contains the parts which are considered static).
132  */
133 void
lp_sampler_static_sampler_state(struct lp_static_sampler_state * state,const struct pipe_sampler_state * sampler)134 lp_sampler_static_sampler_state(struct lp_static_sampler_state *state,
135                                 const struct pipe_sampler_state *sampler)
136 {
137    memset(state, 0, sizeof *state);
138 
139    if (!sampler)
140       return;
141 
142    /*
143     * We don't copy sampler state over unless it is actually enabled, to avoid
144     * spurious recompiles, as the sampler static state is part of the shader
145     * key.
146     *
147     * Ideally the state tracker or cso_cache module would make all state
148     * canonical, but until that happens it's better to be safe than sorry here.
149     *
150     * XXX: Actually there's much more than can be done here, especially
151     * regarding 1D/2D/3D/CUBE textures, wrap modes, etc.
152     */
153 
154    state->wrap_s            = sampler->wrap_s;
155    state->wrap_t            = sampler->wrap_t;
156    state->wrap_r            = sampler->wrap_r;
157    state->min_img_filter    = sampler->min_img_filter;
158    state->mag_img_filter    = sampler->mag_img_filter;
159    state->min_mip_filter    = sampler->min_mip_filter;
160    state->seamless_cube_map = sampler->seamless_cube_map;
161 
162    if (sampler->max_lod > 0.0f) {
163       state->max_lod_pos = 1;
164    }
165 
166    if (sampler->lod_bias != 0.0f) {
167       state->lod_bias_non_zero = 1;
168    }
169 
170    if (state->min_mip_filter != PIPE_TEX_MIPFILTER_NONE ||
171        state->min_img_filter != state->mag_img_filter) {
172 
173       /* If min_lod == max_lod we can greatly simplify mipmap selection.
174        * This is a case that occurs during automatic mipmap generation.
175        */
176       if (sampler->min_lod == sampler->max_lod) {
177          state->min_max_lod_equal = 1;
178       } else {
179          if (sampler->min_lod > 0.0f) {
180             state->apply_min_lod = 1;
181          }
182 
183          /*
184           * XXX this won't do anything with the mesa state tracker which always
185           * sets max_lod to not more than actually present mip maps...
186           */
187          if (sampler->max_lod < (PIPE_MAX_TEXTURE_LEVELS - 1)) {
188             state->apply_max_lod = 1;
189          }
190       }
191    }
192 
193    state->compare_mode      = sampler->compare_mode;
194    if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE) {
195       state->compare_func   = sampler->compare_func;
196    }
197 
198    state->normalized_coords = sampler->normalized_coords;
199 }
200 
201 
202 /**
203  * Generate code to compute coordinate gradient (rho).
204  * \param derivs  partial derivatives of (s, t, r, q) with respect to X and Y
205  *
206  * The resulting rho has bld->levelf format (per quad or per element).
207  */
208 static LLVMValueRef
lp_build_rho(struct lp_build_sample_context * bld,unsigned texture_unit,LLVMValueRef s,LLVMValueRef t,LLVMValueRef r,LLVMValueRef cube_rho,const struct lp_derivatives * derivs)209 lp_build_rho(struct lp_build_sample_context *bld,
210              unsigned texture_unit,
211              LLVMValueRef s,
212              LLVMValueRef t,
213              LLVMValueRef r,
214              LLVMValueRef cube_rho,
215              const struct lp_derivatives *derivs)
216 {
217    struct gallivm_state *gallivm = bld->gallivm;
218    struct lp_build_context *int_size_bld = &bld->int_size_in_bld;
219    struct lp_build_context *float_size_bld = &bld->float_size_in_bld;
220    struct lp_build_context *float_bld = &bld->float_bld;
221    struct lp_build_context *coord_bld = &bld->coord_bld;
222    struct lp_build_context *rho_bld = &bld->lodf_bld;
223    const unsigned dims = bld->dims;
224    LLVMValueRef ddx_ddy[2] = {NULL};
225    LLVMBuilderRef builder = bld->gallivm->builder;
226    LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
227    LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
228    LLVMValueRef index1 = LLVMConstInt(i32t, 1, 0);
229    LLVMValueRef index2 = LLVMConstInt(i32t, 2, 0);
230    LLVMValueRef rho_vec;
231    LLVMValueRef int_size, float_size;
232    LLVMValueRef rho;
233    LLVMValueRef first_level, first_level_vec;
234    unsigned length = coord_bld->type.length;
235    unsigned num_quads = length / 4;
236    boolean rho_per_quad = rho_bld->type.length != length;
237    boolean no_rho_opt = bld->no_rho_approx && (dims > 1);
238    unsigned i;
239    LLVMValueRef i32undef = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
240    LLVMValueRef rho_xvec, rho_yvec;
241 
242    /* Note that all simplified calculations will only work for isotropic filtering */
243 
244    /*
245     * rho calcs are always per quad except for explicit derivs (excluding
246     * the messy cube maps for now) when requested.
247     */
248 
249    first_level = bld->dynamic_state->first_level(bld->dynamic_state, bld->gallivm,
250                                                  bld->context_ptr, texture_unit);
251    first_level_vec = lp_build_broadcast_scalar(int_size_bld, first_level);
252    int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec, TRUE);
253    float_size = lp_build_int_to_float(float_size_bld, int_size);
254 
255    if (cube_rho) {
256       LLVMValueRef cubesize;
257       LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
258 
259       /*
260        * Cube map code did already everything except size mul and per-quad extraction.
261        * Luckily cube maps are always quadratic!
262        */
263       if (rho_per_quad) {
264          rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
265                                          rho_bld->type, cube_rho, 0);
266       }
267       else {
268          rho = lp_build_swizzle_scalar_aos(coord_bld, cube_rho, 0, 4);
269       }
270       /* Could optimize this for single quad just skip the broadcast */
271       cubesize = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
272                                             rho_bld->type, float_size, index0);
273       /* skipping sqrt hence returning rho squared */
274       cubesize = lp_build_mul(rho_bld, cubesize, cubesize);
275       rho = lp_build_mul(rho_bld, cubesize, rho);
276    }
277    else if (derivs) {
278       LLVMValueRef ddmax[3], ddx[3], ddy[3];
279       for (i = 0; i < dims; i++) {
280          LLVMValueRef floatdim;
281          LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
282 
283          floatdim = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
284                                                coord_bld->type, float_size, indexi);
285 
286          /*
287           * note that for rho_per_quad case could reduce math (at some shuffle
288           * cost), but for now use same code to per-pixel lod case.
289           */
290          if (no_rho_opt) {
291             ddx[i] = lp_build_mul(coord_bld, floatdim, derivs->ddx[i]);
292             ddy[i] = lp_build_mul(coord_bld, floatdim, derivs->ddy[i]);
293             ddx[i] = lp_build_mul(coord_bld, ddx[i], ddx[i]);
294             ddy[i] = lp_build_mul(coord_bld, ddy[i], ddy[i]);
295          }
296          else {
297             LLVMValueRef tmpx, tmpy;
298             tmpx = lp_build_abs(coord_bld, derivs->ddx[i]);
299             tmpy = lp_build_abs(coord_bld, derivs->ddy[i]);
300             ddmax[i] = lp_build_max(coord_bld, tmpx, tmpy);
301             ddmax[i] = lp_build_mul(coord_bld, floatdim, ddmax[i]);
302          }
303       }
304       if (no_rho_opt) {
305          rho_xvec = lp_build_add(coord_bld, ddx[0], ddx[1]);
306          rho_yvec = lp_build_add(coord_bld, ddy[0], ddy[1]);
307          if (dims > 2) {
308             rho_xvec = lp_build_add(coord_bld, rho_xvec, ddx[2]);
309             rho_yvec = lp_build_add(coord_bld, rho_yvec, ddy[2]);
310          }
311          rho = lp_build_max(coord_bld, rho_xvec, rho_yvec);
312          /* skipping sqrt hence returning rho squared */
313      }
314       else {
315          rho = ddmax[0];
316          if (dims > 1) {
317             rho = lp_build_max(coord_bld, rho, ddmax[1]);
318             if (dims > 2) {
319                rho = lp_build_max(coord_bld, rho, ddmax[2]);
320             }
321          }
322       }
323       if (rho_per_quad) {
324          /*
325           * rho_vec contains per-pixel rho, convert to scalar per quad.
326           */
327          rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
328                                          rho_bld->type, rho, 0);
329       }
330    }
331    else {
332       /*
333        * This looks all a bit complex, but it's not that bad
334        * (the shuffle code makes it look worse than it is).
335        * Still, might not be ideal for all cases.
336        */
337       static const unsigned char swizzle0[] = { /* no-op swizzle */
338          0, LP_BLD_SWIZZLE_DONTCARE,
339          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
340       };
341       static const unsigned char swizzle1[] = {
342          1, LP_BLD_SWIZZLE_DONTCARE,
343          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
344       };
345       static const unsigned char swizzle2[] = {
346          2, LP_BLD_SWIZZLE_DONTCARE,
347          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
348       };
349 
350       if (dims < 2) {
351          ddx_ddy[0] = lp_build_packed_ddx_ddy_onecoord(coord_bld, s);
352       }
353       else if (dims >= 2) {
354          ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(coord_bld, s, t);
355          if (dims > 2) {
356             ddx_ddy[1] = lp_build_packed_ddx_ddy_onecoord(coord_bld, r);
357          }
358       }
359 
360       if (no_rho_opt) {
361          static const unsigned char swizzle01[] = { /* no-op swizzle */
362             0, 1,
363             LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
364          };
365          static const unsigned char swizzle23[] = {
366             2, 3,
367             LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
368          };
369          LLVMValueRef ddx_ddys, ddx_ddyt, floatdim, shuffles[LP_MAX_VECTOR_LENGTH / 4];
370 
371          for (i = 0; i < num_quads; i++) {
372             shuffles[i*4+0] = shuffles[i*4+1] = index0;
373             shuffles[i*4+2] = shuffles[i*4+3] = index1;
374          }
375          floatdim = LLVMBuildShuffleVector(builder, float_size, float_size,
376                                            LLVMConstVector(shuffles, length), "");
377          ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], floatdim);
378          ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], ddx_ddy[0]);
379          ddx_ddys = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle01);
380          ddx_ddyt = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle23);
381          rho_vec = lp_build_add(coord_bld, ddx_ddys, ddx_ddyt);
382 
383          if (dims > 2) {
384             static const unsigned char swizzle02[] = {
385                0, 2,
386                LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
387             };
388             floatdim = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
389                                                   coord_bld->type, float_size, index2);
390             ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], floatdim);
391             ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], ddx_ddy[1]);
392             ddx_ddy[1] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle02);
393             rho_vec = lp_build_add(coord_bld, rho_vec, ddx_ddy[1]);
394          }
395 
396          rho_xvec = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
397          rho_yvec = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
398          rho = lp_build_max(coord_bld, rho_xvec, rho_yvec);
399 
400          if (rho_per_quad) {
401             rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
402                                             rho_bld->type, rho, 0);
403          }
404          else {
405             rho = lp_build_swizzle_scalar_aos(coord_bld, rho, 0, 4);
406          }
407          /* skipping sqrt hence returning rho squared */
408       }
409       else {
410          ddx_ddy[0] = lp_build_abs(coord_bld, ddx_ddy[0]);
411          if (dims > 2) {
412             ddx_ddy[1] = lp_build_abs(coord_bld, ddx_ddy[1]);
413          }
414          else {
415             ddx_ddy[1] = NULL; /* silence compiler warning */
416          }
417 
418          if (dims < 2) {
419             rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle0);
420             rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle2);
421          }
422          else if (dims == 2) {
423             static const unsigned char swizzle02[] = {
424                0, 2,
425                LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
426             };
427             static const unsigned char swizzle13[] = {
428                1, 3,
429                LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
430             };
431             rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle02);
432             rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle13);
433          }
434          else {
435             LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH];
436             LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH];
437             assert(dims == 3);
438             for (i = 0; i < num_quads; i++) {
439                shuffles1[4*i + 0] = lp_build_const_int32(gallivm, 4*i);
440                shuffles1[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 2);
441                shuffles1[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i);
442                shuffles1[4*i + 3] = i32undef;
443                shuffles2[4*i + 0] = lp_build_const_int32(gallivm, 4*i + 1);
444                shuffles2[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 3);
445                shuffles2[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i + 2);
446                shuffles2[4*i + 3] = i32undef;
447             }
448             rho_xvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
449                                               LLVMConstVector(shuffles1, length), "");
450             rho_yvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
451                                               LLVMConstVector(shuffles2, length), "");
452          }
453 
454          rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
455 
456          if (bld->coord_type.length > 4) {
457             /* expand size to each quad */
458             if (dims > 1) {
459                /* could use some broadcast_vector helper for this? */
460                LLVMValueRef src[LP_MAX_VECTOR_LENGTH/4];
461                for (i = 0; i < num_quads; i++) {
462                   src[i] = float_size;
463                }
464                float_size = lp_build_concat(bld->gallivm, src, float_size_bld->type, num_quads);
465             }
466             else {
467                float_size = lp_build_broadcast_scalar(coord_bld, float_size);
468             }
469             rho_vec = lp_build_mul(coord_bld, rho_vec, float_size);
470 
471             if (dims <= 1) {
472                rho = rho_vec;
473             }
474             else {
475                if (dims >= 2) {
476                   LLVMValueRef rho_s, rho_t, rho_r;
477 
478                   rho_s = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
479                   rho_t = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
480 
481                   rho = lp_build_max(coord_bld, rho_s, rho_t);
482 
483                   if (dims >= 3) {
484                      rho_r = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle2);
485                      rho = lp_build_max(coord_bld, rho, rho_r);
486                   }
487                }
488             }
489             if (rho_per_quad) {
490                rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
491                                                rho_bld->type, rho, 0);
492             }
493             else {
494                rho = lp_build_swizzle_scalar_aos(coord_bld, rho, 0, 4);
495             }
496          }
497          else {
498             if (dims <= 1) {
499                rho_vec = LLVMBuildExtractElement(builder, rho_vec, index0, "");
500             }
501             rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
502 
503             if (dims <= 1) {
504                rho = rho_vec;
505             }
506             else {
507                if (dims >= 2) {
508                   LLVMValueRef rho_s, rho_t, rho_r;
509 
510                   rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
511                   rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
512 
513                   rho = lp_build_max(float_bld, rho_s, rho_t);
514 
515                   if (dims >= 3) {
516                      rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
517                      rho = lp_build_max(float_bld, rho, rho_r);
518                   }
519                }
520             }
521             if (!rho_per_quad) {
522                rho = lp_build_broadcast_scalar(rho_bld, rho);
523             }
524          }
525       }
526    }
527 
528    return rho;
529 }
530 
531 
532 /*
533  * Bri-linear lod computation
534  *
535  * Use a piece-wise linear approximation of log2 such that:
536  * - round to nearest, for values in the neighborhood of -1, 0, 1, 2, etc.
537  * - linear approximation for values in the neighborhood of 0.5, 1.5., etc,
538  *   with the steepness specified in 'factor'
539  * - exact result for 0.5, 1.5, etc.
540  *
541  *
542  *   1.0 -              /----*
543  *                     /
544  *                    /
545  *                   /
546  *   0.5 -          *
547  *                 /
548  *                /
549  *               /
550  *   0.0 - *----/
551  *
552  *         |                 |
553  *        2^0               2^1
554  *
555  * This is a technique also commonly used in hardware:
556  * - http://ixbtlabs.com/articles2/gffx/nv40-rx800-3.html
557  *
558  * TODO: For correctness, this should only be applied when texture is known to
559  * have regular mipmaps, i.e., mipmaps derived from the base level.
560  *
561  * TODO: This could be done in fixed point, where applicable.
562  */
563 static void
lp_build_brilinear_lod(struct lp_build_context * bld,LLVMValueRef lod,double factor,LLVMValueRef * out_lod_ipart,LLVMValueRef * out_lod_fpart)564 lp_build_brilinear_lod(struct lp_build_context *bld,
565                        LLVMValueRef lod,
566                        double factor,
567                        LLVMValueRef *out_lod_ipart,
568                        LLVMValueRef *out_lod_fpart)
569 {
570    LLVMValueRef lod_fpart;
571    double pre_offset = (factor - 0.5)/factor - 0.5;
572    double post_offset = 1 - factor;
573 
574    if (0) {
575       lp_build_printf(bld->gallivm, "lod = %f\n", lod);
576    }
577 
578    lod = lp_build_add(bld, lod,
579                       lp_build_const_vec(bld->gallivm, bld->type, pre_offset));
580 
581    lp_build_ifloor_fract(bld, lod, out_lod_ipart, &lod_fpart);
582 
583    lod_fpart = lp_build_mad(bld, lod_fpart,
584                             lp_build_const_vec(bld->gallivm, bld->type, factor),
585                             lp_build_const_vec(bld->gallivm, bld->type, post_offset));
586 
587    /*
588     * It's not necessary to clamp lod_fpart since:
589     * - the above expression will never produce numbers greater than one.
590     * - the mip filtering branch is only taken if lod_fpart is positive
591     */
592 
593    *out_lod_fpart = lod_fpart;
594 
595    if (0) {
596       lp_build_printf(bld->gallivm, "lod_ipart = %i\n", *out_lod_ipart);
597       lp_build_printf(bld->gallivm, "lod_fpart = %f\n\n", *out_lod_fpart);
598    }
599 }
600 
601 
602 /*
603  * Combined log2 and brilinear lod computation.
604  *
605  * It's in all identical to calling lp_build_fast_log2() and
606  * lp_build_brilinear_lod() above, but by combining we can compute the integer
607  * and fractional part independently.
608  */
609 static void
lp_build_brilinear_rho(struct lp_build_context * bld,LLVMValueRef rho,double factor,LLVMValueRef * out_lod_ipart,LLVMValueRef * out_lod_fpart)610 lp_build_brilinear_rho(struct lp_build_context *bld,
611                        LLVMValueRef rho,
612                        double factor,
613                        LLVMValueRef *out_lod_ipart,
614                        LLVMValueRef *out_lod_fpart)
615 {
616    LLVMValueRef lod_ipart;
617    LLVMValueRef lod_fpart;
618 
619    const double pre_factor = (2*factor - 0.5)/(M_SQRT2*factor);
620    const double post_offset = 1 - 2*factor;
621 
622    assert(bld->type.floating);
623 
624    assert(lp_check_value(bld->type, rho));
625 
626    /*
627     * The pre factor will make the intersections with the exact powers of two
628     * happen precisely where we want them to be, which means that the integer
629     * part will not need any post adjustments.
630     */
631    rho = lp_build_mul(bld, rho,
632                       lp_build_const_vec(bld->gallivm, bld->type, pre_factor));
633 
634    /* ipart = ifloor(log2(rho)) */
635    lod_ipart = lp_build_extract_exponent(bld, rho, 0);
636 
637    /* fpart = rho / 2**ipart */
638    lod_fpart = lp_build_extract_mantissa(bld, rho);
639 
640    lod_fpart = lp_build_mad(bld, lod_fpart,
641                             lp_build_const_vec(bld->gallivm, bld->type, factor),
642                             lp_build_const_vec(bld->gallivm, bld->type, post_offset));
643 
644    /*
645     * Like lp_build_brilinear_lod, it's not necessary to clamp lod_fpart since:
646     * - the above expression will never produce numbers greater than one.
647     * - the mip filtering branch is only taken if lod_fpart is positive
648     */
649 
650    *out_lod_ipart = lod_ipart;
651    *out_lod_fpart = lod_fpart;
652 }
653 
654 
655 /**
656  * Fast implementation of iround(log2(sqrt(x))), based on
657  * log2(x^n) == n*log2(x).
658  *
659  * Gives accurate results all the time.
660  * (Could be trivially extended to handle other power-of-two roots.)
661  */
662 static LLVMValueRef
lp_build_ilog2_sqrt(struct lp_build_context * bld,LLVMValueRef x)663 lp_build_ilog2_sqrt(struct lp_build_context *bld,
664                     LLVMValueRef x)
665 {
666    LLVMBuilderRef builder = bld->gallivm->builder;
667    LLVMValueRef ipart;
668    struct lp_type i_type = lp_int_type(bld->type);
669    LLVMValueRef one = lp_build_const_int_vec(bld->gallivm, i_type, 1);
670 
671    assert(bld->type.floating);
672 
673    assert(lp_check_value(bld->type, x));
674 
675    /* ipart = log2(x) + 0.5 = 0.5*(log2(x^2) + 1.0) */
676    ipart = lp_build_extract_exponent(bld, x, 1);
677    ipart = LLVMBuildAShr(builder, ipart, one, "");
678 
679    return ipart;
680 }
681 
682 
683 /**
684  * Generate code to compute texture level of detail (lambda).
685  * \param derivs  partial derivatives of (s, t, r, q) with respect to X and Y
686  * \param lod_bias  optional float vector with the shader lod bias
687  * \param explicit_lod  optional float vector with the explicit lod
688  * \param cube_rho  rho calculated by cube coord mapping (optional)
689  * \param out_lod_ipart  integer part of lod
690  * \param out_lod_fpart  float part of lod (never larger than 1 but may be negative)
691  * \param out_lod_positive  (mask) if lod is positive (i.e. texture is minified)
692  *
693  * The resulting lod can be scalar per quad or be per element.
694  */
695 void
lp_build_lod_selector(struct lp_build_sample_context * bld,boolean is_lodq,unsigned texture_unit,unsigned sampler_unit,LLVMValueRef s,LLVMValueRef t,LLVMValueRef r,LLVMValueRef cube_rho,const struct lp_derivatives * derivs,LLVMValueRef lod_bias,LLVMValueRef explicit_lod,unsigned mip_filter,LLVMValueRef * out_lod,LLVMValueRef * out_lod_ipart,LLVMValueRef * out_lod_fpart,LLVMValueRef * out_lod_positive)696 lp_build_lod_selector(struct lp_build_sample_context *bld,
697                       boolean is_lodq,
698                       unsigned texture_unit,
699                       unsigned sampler_unit,
700                       LLVMValueRef s,
701                       LLVMValueRef t,
702                       LLVMValueRef r,
703                       LLVMValueRef cube_rho,
704                       const struct lp_derivatives *derivs,
705                       LLVMValueRef lod_bias, /* optional */
706                       LLVMValueRef explicit_lod, /* optional */
707                       unsigned mip_filter,
708                       LLVMValueRef *out_lod,
709                       LLVMValueRef *out_lod_ipart,
710                       LLVMValueRef *out_lod_fpart,
711                       LLVMValueRef *out_lod_positive)
712 
713 {
714    LLVMBuilderRef builder = bld->gallivm->builder;
715    struct lp_sampler_dynamic_state *dynamic_state = bld->dynamic_state;
716    struct lp_build_context *lodf_bld = &bld->lodf_bld;
717    LLVMValueRef lod;
718 
719    *out_lod_ipart = bld->lodi_bld.zero;
720    *out_lod_positive = bld->lodi_bld.zero;
721    *out_lod_fpart = lodf_bld->zero;
722 
723    /*
724     * For determining min/mag, we follow GL 4.1 spec, 3.9.12 Texture Magnification:
725     * "Implementations may either unconditionally assume c = 0 for the minification
726     * vs. magnification switch-over point, or may choose to make c depend on the
727     * combination of minification and magnification modes as follows: if the
728     * magnification filter is given by LINEAR and the minification filter is given
729     * by NEAREST_MIPMAP_NEAREST or NEAREST_MIPMAP_LINEAR, then c = 0.5. This is
730     * done to ensure that a minified texture does not appear "sharper" than a
731     * magnified texture. Otherwise c = 0."
732     * And 3.9.11 Texture Minification:
733     * "If lod is less than or equal to the constant c (see section 3.9.12) the
734     * texture is said to be magnified; if it is greater, the texture is minified."
735     * So, using 0 as switchover point always, and using magnification for lod == 0.
736     * Note that the always c = 0 behavior is new (first appearing in GL 3.1 spec),
737     * old GL versions required 0.5 for the modes listed above.
738     * I have no clue about the (undocumented) wishes of d3d9/d3d10 here!
739     */
740 
741    if (bld->static_sampler_state->min_max_lod_equal && !is_lodq) {
742       /* User is forcing sampling from a particular mipmap level.
743        * This is hit during mipmap generation.
744        */
745       LLVMValueRef min_lod =
746          dynamic_state->min_lod(dynamic_state, bld->gallivm,
747                                 bld->context_ptr, sampler_unit);
748 
749       lod = lp_build_broadcast_scalar(lodf_bld, min_lod);
750    }
751    else {
752       if (explicit_lod) {
753          if (bld->num_lods != bld->coord_type.length)
754             lod = lp_build_pack_aos_scalars(bld->gallivm, bld->coord_bld.type,
755                                             lodf_bld->type, explicit_lod, 0);
756          else
757             lod = explicit_lod;
758       }
759       else {
760          LLVMValueRef rho;
761          boolean rho_squared = (bld->no_rho_approx &&
762                                 (bld->dims > 1)) || cube_rho;
763 
764          rho = lp_build_rho(bld, texture_unit, s, t, r, cube_rho, derivs);
765 
766          /*
767           * Compute lod = log2(rho)
768           */
769 
770          if (!lod_bias && !is_lodq &&
771              !bld->static_sampler_state->lod_bias_non_zero &&
772              !bld->static_sampler_state->apply_max_lod &&
773              !bld->static_sampler_state->apply_min_lod) {
774             /*
775              * Special case when there are no post-log2 adjustments, which
776              * saves instructions but keeping the integer and fractional lod
777              * computations separate from the start.
778              */
779 
780             if (mip_filter == PIPE_TEX_MIPFILTER_NONE ||
781                 mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
782                /*
783                 * Don't actually need both values all the time, lod_ipart is
784                 * needed for nearest mipfilter, lod_positive if min != mag.
785                 */
786                if (rho_squared) {
787                   *out_lod_ipart = lp_build_ilog2_sqrt(lodf_bld, rho);
788                }
789                else {
790                   *out_lod_ipart = lp_build_ilog2(lodf_bld, rho);
791                }
792                *out_lod_positive = lp_build_cmp(lodf_bld, PIPE_FUNC_GREATER,
793                                                 rho, lodf_bld->one);
794                return;
795             }
796             if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR &&
797                 !bld->no_brilinear && !rho_squared) {
798                /*
799                 * This can't work if rho is squared. Not sure if it could be
800                 * fixed while keeping it worthwile, could also do sqrt here
801                 * but brilinear and no_rho_opt seems like a combination not
802                 * making much sense anyway so just use ordinary path below.
803                 */
804                lp_build_brilinear_rho(lodf_bld, rho, BRILINEAR_FACTOR,
805                                       out_lod_ipart, out_lod_fpart);
806                *out_lod_positive = lp_build_cmp(lodf_bld, PIPE_FUNC_GREATER,
807                                                 rho, lodf_bld->one);
808                return;
809             }
810          }
811 
812          if (0) {
813             lod = lp_build_log2(lodf_bld, rho);
814          }
815          else {
816             lod = lp_build_fast_log2(lodf_bld, rho);
817          }
818          if (rho_squared) {
819             /* log2(x^2) == 0.5*log2(x) */
820             lod = lp_build_mul(lodf_bld, lod,
821                                lp_build_const_vec(bld->gallivm, lodf_bld->type, 0.5F));
822          }
823 
824          /* add shader lod bias */
825          if (lod_bias) {
826             if (bld->num_lods != bld->coord_type.length)
827                lod_bias = lp_build_pack_aos_scalars(bld->gallivm, bld->coord_bld.type,
828                                                     lodf_bld->type, lod_bias, 0);
829             lod = LLVMBuildFAdd(builder, lod, lod_bias, "shader_lod_bias");
830          }
831       }
832 
833       /* add sampler lod bias */
834       if (bld->static_sampler_state->lod_bias_non_zero) {
835          LLVMValueRef sampler_lod_bias =
836             dynamic_state->lod_bias(dynamic_state, bld->gallivm,
837                                     bld->context_ptr, sampler_unit);
838          sampler_lod_bias = lp_build_broadcast_scalar(lodf_bld,
839                                                       sampler_lod_bias);
840          lod = LLVMBuildFAdd(builder, lod, sampler_lod_bias, "sampler_lod_bias");
841       }
842 
843       if (is_lodq) {
844          *out_lod = lod;
845       }
846 
847       /* clamp lod */
848       if (bld->static_sampler_state->apply_max_lod) {
849          LLVMValueRef max_lod =
850             dynamic_state->max_lod(dynamic_state, bld->gallivm,
851                                    bld->context_ptr, sampler_unit);
852          max_lod = lp_build_broadcast_scalar(lodf_bld, max_lod);
853 
854          lod = lp_build_min(lodf_bld, lod, max_lod);
855       }
856       if (bld->static_sampler_state->apply_min_lod) {
857          LLVMValueRef min_lod =
858             dynamic_state->min_lod(dynamic_state, bld->gallivm,
859                                    bld->context_ptr, sampler_unit);
860          min_lod = lp_build_broadcast_scalar(lodf_bld, min_lod);
861 
862          lod = lp_build_max(lodf_bld, lod, min_lod);
863       }
864 
865       if (is_lodq) {
866          *out_lod_fpart = lod;
867          return;
868       }
869    }
870 
871    *out_lod_positive = lp_build_cmp(lodf_bld, PIPE_FUNC_GREATER,
872                                     lod, lodf_bld->zero);
873 
874    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
875       if (!bld->no_brilinear) {
876          lp_build_brilinear_lod(lodf_bld, lod, BRILINEAR_FACTOR,
877                                 out_lod_ipart, out_lod_fpart);
878       }
879       else {
880          lp_build_ifloor_fract(lodf_bld, lod, out_lod_ipart, out_lod_fpart);
881       }
882 
883       lp_build_name(*out_lod_fpart, "lod_fpart");
884    }
885    else {
886       *out_lod_ipart = lp_build_iround(lodf_bld, lod);
887    }
888 
889    lp_build_name(*out_lod_ipart, "lod_ipart");
890 
891    return;
892 }
893 
894 
895 /**
896  * For PIPE_TEX_MIPFILTER_NEAREST, convert int part of lod
897  * to actual mip level.
898  * Note: this is all scalar per quad code.
899  * \param lod_ipart  int texture level of detail
900  * \param level_out  returns integer
901  * \param out_of_bounds returns per coord out_of_bounds mask if provided
902  */
903 void
lp_build_nearest_mip_level(struct lp_build_sample_context * bld,unsigned texture_unit,LLVMValueRef lod_ipart,LLVMValueRef * level_out,LLVMValueRef * out_of_bounds)904 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
905                            unsigned texture_unit,
906                            LLVMValueRef lod_ipart,
907                            LLVMValueRef *level_out,
908                            LLVMValueRef *out_of_bounds)
909 {
910    struct lp_build_context *leveli_bld = &bld->leveli_bld;
911    struct lp_sampler_dynamic_state *dynamic_state = bld->dynamic_state;
912    LLVMValueRef first_level, last_level, level;
913 
914    first_level = dynamic_state->first_level(dynamic_state, bld->gallivm,
915                                             bld->context_ptr, texture_unit);
916    last_level = dynamic_state->last_level(dynamic_state, bld->gallivm,
917                                           bld->context_ptr, texture_unit);
918    first_level = lp_build_broadcast_scalar(leveli_bld, first_level);
919    last_level = lp_build_broadcast_scalar(leveli_bld, last_level);
920 
921    level = lp_build_add(leveli_bld, lod_ipart, first_level);
922 
923    if (out_of_bounds) {
924       LLVMValueRef out, out1;
925       out = lp_build_cmp(leveli_bld, PIPE_FUNC_LESS, level, first_level);
926       out1 = lp_build_cmp(leveli_bld, PIPE_FUNC_GREATER, level, last_level);
927       out = lp_build_or(leveli_bld, out, out1);
928       if (bld->num_mips == bld->coord_bld.type.length) {
929          *out_of_bounds = out;
930       }
931       else if (bld->num_mips == 1) {
932          *out_of_bounds = lp_build_broadcast_scalar(&bld->int_coord_bld, out);
933       }
934       else {
935          assert(bld->num_mips == bld->coord_bld.type.length / 4);
936          *out_of_bounds = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
937                                                                 leveli_bld->type,
938                                                                 bld->int_coord_bld.type,
939                                                                 out);
940       }
941       level = lp_build_andnot(&bld->int_coord_bld, level, *out_of_bounds);
942       *level_out = level;
943    }
944    else {
945       /* clamp level to legal range of levels */
946       *level_out = lp_build_clamp(leveli_bld, level, first_level, last_level);
947 
948    }
949 }
950 
951 
952 /**
953  * For PIPE_TEX_MIPFILTER_LINEAR, convert per-quad (or per element) int LOD(s)
954  * to two (per-quad) (adjacent) mipmap level indexes, and fix up float lod
955  * part accordingly.
956  * Later, we'll sample from those two mipmap levels and interpolate between them.
957  */
958 void
lp_build_linear_mip_levels(struct lp_build_sample_context * bld,unsigned texture_unit,LLVMValueRef lod_ipart,LLVMValueRef * lod_fpart_inout,LLVMValueRef * level0_out,LLVMValueRef * level1_out)959 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
960                            unsigned texture_unit,
961                            LLVMValueRef lod_ipart,
962                            LLVMValueRef *lod_fpart_inout,
963                            LLVMValueRef *level0_out,
964                            LLVMValueRef *level1_out)
965 {
966    LLVMBuilderRef builder = bld->gallivm->builder;
967    struct lp_sampler_dynamic_state *dynamic_state = bld->dynamic_state;
968    struct lp_build_context *leveli_bld = &bld->leveli_bld;
969    struct lp_build_context *levelf_bld = &bld->levelf_bld;
970    LLVMValueRef first_level, last_level;
971    LLVMValueRef clamp_min;
972    LLVMValueRef clamp_max;
973 
974    assert(bld->num_lods == bld->num_mips);
975 
976    first_level = dynamic_state->first_level(dynamic_state, bld->gallivm,
977                                             bld->context_ptr, texture_unit);
978    last_level = dynamic_state->last_level(dynamic_state, bld->gallivm,
979                                           bld->context_ptr, texture_unit);
980    first_level = lp_build_broadcast_scalar(leveli_bld, first_level);
981    last_level = lp_build_broadcast_scalar(leveli_bld, last_level);
982 
983    *level0_out = lp_build_add(leveli_bld, lod_ipart, first_level);
984    *level1_out = lp_build_add(leveli_bld, *level0_out, leveli_bld->one);
985 
986    /*
987     * Clamp both *level0_out and *level1_out to [first_level, last_level], with
988     * the minimum number of comparisons, and zeroing lod_fpart in the extreme
989     * ends in the process.
990     */
991 
992    /* *level0_out < first_level */
993    clamp_min = LLVMBuildICmp(builder, LLVMIntSLT,
994                              *level0_out, first_level,
995                              "clamp_lod_to_first");
996 
997    *level0_out = LLVMBuildSelect(builder, clamp_min,
998                                  first_level, *level0_out, "");
999 
1000    *level1_out = LLVMBuildSelect(builder, clamp_min,
1001                                  first_level, *level1_out, "");
1002 
1003    *lod_fpart_inout = LLVMBuildSelect(builder, clamp_min,
1004                                       levelf_bld->zero, *lod_fpart_inout, "");
1005 
1006    /* *level0_out >= last_level */
1007    clamp_max = LLVMBuildICmp(builder, LLVMIntSGE,
1008                              *level0_out, last_level,
1009                              "clamp_lod_to_last");
1010 
1011    *level0_out = LLVMBuildSelect(builder, clamp_max,
1012                                  last_level, *level0_out, "");
1013 
1014    *level1_out = LLVMBuildSelect(builder, clamp_max,
1015                                  last_level, *level1_out, "");
1016 
1017    *lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
1018                                       levelf_bld->zero, *lod_fpart_inout, "");
1019 
1020    lp_build_name(*level0_out, "texture%u_miplevel0", texture_unit);
1021    lp_build_name(*level1_out, "texture%u_miplevel1", texture_unit);
1022    lp_build_name(*lod_fpart_inout, "texture%u_mipweight", texture_unit);
1023 }
1024 
1025 
1026 /**
1027  * Return pointer to a single mipmap level.
1028  * \param level  integer mipmap level
1029  */
1030 LLVMValueRef
lp_build_get_mipmap_level(struct lp_build_sample_context * bld,LLVMValueRef level)1031 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
1032                           LLVMValueRef level)
1033 {
1034    LLVMBuilderRef builder = bld->gallivm->builder;
1035    LLVMValueRef indexes[2], data_ptr, mip_offset;
1036 
1037    indexes[0] = lp_build_const_int32(bld->gallivm, 0);
1038    indexes[1] = level;
1039    mip_offset = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1040    mip_offset = LLVMBuildLoad(builder, mip_offset, "");
1041    data_ptr = LLVMBuildGEP(builder, bld->base_ptr, &mip_offset, 1, "");
1042    return data_ptr;
1043 }
1044 
1045 /**
1046  * Return (per-pixel) offsets to mip levels.
1047  * \param level  integer mipmap level
1048  */
1049 LLVMValueRef
lp_build_get_mip_offsets(struct lp_build_sample_context * bld,LLVMValueRef level)1050 lp_build_get_mip_offsets(struct lp_build_sample_context *bld,
1051                          LLVMValueRef level)
1052 {
1053    LLVMBuilderRef builder = bld->gallivm->builder;
1054    LLVMValueRef indexes[2], offsets, offset1;
1055 
1056    indexes[0] = lp_build_const_int32(bld->gallivm, 0);
1057    if (bld->num_mips == 1) {
1058       indexes[1] = level;
1059       offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1060       offset1 = LLVMBuildLoad(builder, offset1, "");
1061       offsets = lp_build_broadcast_scalar(&bld->int_coord_bld, offset1);
1062    }
1063    else if (bld->num_mips == bld->coord_bld.type.length / 4) {
1064       unsigned i;
1065 
1066       offsets = bld->int_coord_bld.undef;
1067       for (i = 0; i < bld->num_mips; i++) {
1068          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1069          LLVMValueRef indexo = lp_build_const_int32(bld->gallivm, 4 * i);
1070          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1071          offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1072          offset1 = LLVMBuildLoad(builder, offset1, "");
1073          offsets = LLVMBuildInsertElement(builder, offsets, offset1, indexo, "");
1074       }
1075       offsets = lp_build_swizzle_scalar_aos(&bld->int_coord_bld, offsets, 0, 4);
1076    }
1077    else {
1078       unsigned i;
1079 
1080       assert (bld->num_mips == bld->coord_bld.type.length);
1081 
1082       offsets = bld->int_coord_bld.undef;
1083       for (i = 0; i < bld->num_mips; i++) {
1084          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1085          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1086          offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
1087          offset1 = LLVMBuildLoad(builder, offset1, "");
1088          offsets = LLVMBuildInsertElement(builder, offsets, offset1, indexi, "");
1089       }
1090    }
1091    return offsets;
1092 }
1093 
1094 
1095 /**
1096  * Codegen equivalent for u_minify().
1097  * @param lod_scalar  if lod is a (broadcasted) scalar
1098  * Return max(1, base_size >> level);
1099  */
1100 LLVMValueRef
lp_build_minify(struct lp_build_context * bld,LLVMValueRef base_size,LLVMValueRef level,boolean lod_scalar)1101 lp_build_minify(struct lp_build_context *bld,
1102                 LLVMValueRef base_size,
1103                 LLVMValueRef level,
1104                 boolean lod_scalar)
1105 {
1106    LLVMBuilderRef builder = bld->gallivm->builder;
1107    assert(lp_check_value(bld->type, base_size));
1108    assert(lp_check_value(bld->type, level));
1109 
1110    if (level == bld->zero) {
1111       /* if we're using mipmap level zero, no minification is needed */
1112       return base_size;
1113    }
1114    else {
1115       LLVMValueRef size;
1116       assert(bld->type.sign);
1117       if (lod_scalar ||
1118          (util_cpu_caps.has_avx2 || !util_cpu_caps.has_sse)) {
1119          size = LLVMBuildLShr(builder, base_size, level, "minify");
1120          size = lp_build_max(bld, size, bld->one);
1121       }
1122       else {
1123          /*
1124           * emulate shift with float mul, since intel "forgot" shifts with
1125           * per-element shift count until avx2, which results in terrible
1126           * scalar extraction (both count and value), scalar shift,
1127           * vector reinsertion. Should not be an issue on any non-x86 cpu
1128           * with a vector instruction set.
1129           * On cpus with AMD's XOP this should also be unnecessary but I'm
1130           * not sure if llvm would emit this with current flags.
1131           */
1132          LLVMValueRef const127, const23, lf;
1133          struct lp_type ftype;
1134          struct lp_build_context fbld;
1135          ftype = lp_type_float_vec(32, bld->type.length * bld->type.width);
1136          lp_build_context_init(&fbld, bld->gallivm, ftype);
1137          const127 = lp_build_const_int_vec(bld->gallivm, bld->type, 127);
1138          const23 = lp_build_const_int_vec(bld->gallivm, bld->type, 23);
1139 
1140          /* calculate 2^(-level) float */
1141          lf = lp_build_sub(bld, const127, level);
1142          lf = lp_build_shl(bld, lf, const23);
1143          lf = LLVMBuildBitCast(builder, lf, fbld.vec_type, "");
1144 
1145          /* finish shift operation by doing float mul */
1146          base_size = lp_build_int_to_float(&fbld, base_size);
1147          size = lp_build_mul(&fbld, base_size, lf);
1148          /*
1149           * do the max also with floats because
1150           * a) non-emulated int max requires sse41
1151           *    (this is actually a lie as we could cast to 16bit values
1152           *    as 16bit is sufficient and 16bit int max is sse2)
1153           * b) with avx we can do int max 4-wide but float max 8-wide
1154           */
1155          size = lp_build_max(&fbld, size, fbld.one);
1156          size = lp_build_itrunc(&fbld, size);
1157       }
1158       return size;
1159    }
1160 }
1161 
1162 
1163 /**
1164  * Dereference stride_array[mipmap_level] array to get a stride.
1165  * Return stride as a vector.
1166  */
1167 static LLVMValueRef
lp_build_get_level_stride_vec(struct lp_build_sample_context * bld,LLVMValueRef stride_array,LLVMValueRef level)1168 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
1169                               LLVMValueRef stride_array, LLVMValueRef level)
1170 {
1171    LLVMBuilderRef builder = bld->gallivm->builder;
1172    LLVMValueRef indexes[2], stride, stride1;
1173    indexes[0] = lp_build_const_int32(bld->gallivm, 0);
1174    if (bld->num_mips == 1) {
1175       indexes[1] = level;
1176       stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
1177       stride1 = LLVMBuildLoad(builder, stride1, "");
1178       stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride1);
1179    }
1180    else if (bld->num_mips == bld->coord_bld.type.length / 4) {
1181       LLVMValueRef stride1;
1182       unsigned i;
1183 
1184       stride = bld->int_coord_bld.undef;
1185       for (i = 0; i < bld->num_mips; i++) {
1186          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1187          LLVMValueRef indexo = lp_build_const_int32(bld->gallivm, 4 * i);
1188          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1189          stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
1190          stride1 = LLVMBuildLoad(builder, stride1, "");
1191          stride = LLVMBuildInsertElement(builder, stride, stride1, indexo, "");
1192       }
1193       stride = lp_build_swizzle_scalar_aos(&bld->int_coord_bld, stride, 0, 4);
1194    }
1195    else {
1196       LLVMValueRef stride1;
1197       unsigned i;
1198 
1199       assert (bld->num_mips == bld->coord_bld.type.length);
1200 
1201       stride = bld->int_coord_bld.undef;
1202       for (i = 0; i < bld->coord_bld.type.length; i++) {
1203          LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1204          indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
1205          stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
1206          stride1 = LLVMBuildLoad(builder, stride1, "");
1207          stride = LLVMBuildInsertElement(builder, stride, stride1, indexi, "");
1208       }
1209    }
1210    return stride;
1211 }
1212 
1213 
1214 /**
1215  * When sampling a mipmap, we need to compute the width, height, depth
1216  * of the source levels from the level indexes.  This helper function
1217  * does that.
1218  */
1219 void
lp_build_mipmap_level_sizes(struct lp_build_sample_context * bld,LLVMValueRef ilevel,LLVMValueRef * out_size,LLVMValueRef * row_stride_vec,LLVMValueRef * img_stride_vec)1220 lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
1221                             LLVMValueRef ilevel,
1222                             LLVMValueRef *out_size,
1223                             LLVMValueRef *row_stride_vec,
1224                             LLVMValueRef *img_stride_vec)
1225 {
1226    const unsigned dims = bld->dims;
1227    LLVMValueRef ilevel_vec;
1228 
1229    /*
1230     * Compute width, height, depth at mipmap level 'ilevel'
1231     */
1232    if (bld->num_mips == 1) {
1233       ilevel_vec = lp_build_broadcast_scalar(&bld->int_size_bld, ilevel);
1234       *out_size = lp_build_minify(&bld->int_size_bld, bld->int_size, ilevel_vec, TRUE);
1235    }
1236    else {
1237       LLVMValueRef int_size_vec;
1238       LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
1239       unsigned num_quads = bld->coord_bld.type.length / 4;
1240       unsigned i;
1241 
1242       if (bld->num_mips == num_quads) {
1243          /*
1244           * XXX: this should be #ifndef SANE_INSTRUCTION_SET.
1245           * intel "forgot" the variable shift count instruction until avx2.
1246           * A harmless 8x32 shift gets translated into 32 instructions
1247           * (16 extracts, 8 scalar shifts, 8 inserts), llvm is apparently
1248           * unable to recognize if there are really just 2 different shift
1249           * count values. So do the shift 4-wide before expansion.
1250           */
1251          struct lp_build_context bld4;
1252          struct lp_type type4;
1253 
1254          type4 = bld->int_coord_bld.type;
1255          type4.length = 4;
1256 
1257          lp_build_context_init(&bld4, bld->gallivm, type4);
1258 
1259          if (bld->dims == 1) {
1260             assert(bld->int_size_in_bld.type.length == 1);
1261             int_size_vec = lp_build_broadcast_scalar(&bld4,
1262                                                      bld->int_size);
1263          }
1264          else {
1265             assert(bld->int_size_in_bld.type.length == 4);
1266             int_size_vec = bld->int_size;
1267          }
1268 
1269          for (i = 0; i < num_quads; i++) {
1270             LLVMValueRef ileveli;
1271             LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1272 
1273             ileveli = lp_build_extract_broadcast(bld->gallivm,
1274                                                  bld->leveli_bld.type,
1275                                                  bld4.type,
1276                                                  ilevel,
1277                                                  indexi);
1278             tmp[i] = lp_build_minify(&bld4, int_size_vec, ileveli, TRUE);
1279          }
1280          /*
1281           * out_size is [w0, h0, d0, _, w1, h1, d1, _, ...] vector for dims > 1,
1282           * [w0, w0, w0, w0, w1, w1, w1, w1, ...] otherwise.
1283           */
1284          *out_size = lp_build_concat(bld->gallivm,
1285                                      tmp,
1286                                      bld4.type,
1287                                      num_quads);
1288       }
1289       else {
1290         /* FIXME: this is terrible and results in _huge_ vector
1291          * (for the dims > 1 case).
1292          * Should refactor this (together with extract_image_sizes) and do
1293          * something more useful. Could for instance if we have width,height
1294          * with 4-wide vector pack all elements into a 8xi16 vector
1295          * (on which we can still do useful math) instead of using a 16xi32
1296          * vector.
1297          * For dims == 1 this will create [w0, w1, w2, w3, ...] vector.
1298          * For dims > 1 this will create [w0, h0, d0, _, w1, h1, d1, _, ...] vector.
1299          */
1300          assert(bld->num_mips == bld->coord_bld.type.length);
1301          if (bld->dims == 1) {
1302             assert(bld->int_size_in_bld.type.length == 1);
1303             int_size_vec = lp_build_broadcast_scalar(&bld->int_coord_bld,
1304                                                      bld->int_size);
1305             *out_size = lp_build_minify(&bld->int_coord_bld, int_size_vec, ilevel, FALSE);
1306          }
1307          else {
1308             LLVMValueRef ilevel1;
1309             for (i = 0; i < bld->num_mips; i++) {
1310                LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
1311                ilevel1 = lp_build_extract_broadcast(bld->gallivm, bld->int_coord_type,
1312                                                     bld->int_size_in_bld.type, ilevel, indexi);
1313                tmp[i] = bld->int_size;
1314                tmp[i] = lp_build_minify(&bld->int_size_in_bld, tmp[i], ilevel1, TRUE);
1315             }
1316             *out_size = lp_build_concat(bld->gallivm, tmp,
1317                                         bld->int_size_in_bld.type,
1318                                         bld->num_mips);
1319          }
1320       }
1321    }
1322 
1323    if (dims >= 2) {
1324       *row_stride_vec = lp_build_get_level_stride_vec(bld,
1325                                                       bld->row_stride_array,
1326                                                       ilevel);
1327    }
1328    if (dims == 3 || has_layer_coord(bld->static_texture_state->target)) {
1329       *img_stride_vec = lp_build_get_level_stride_vec(bld,
1330                                                       bld->img_stride_array,
1331                                                       ilevel);
1332    }
1333 }
1334 
1335 
1336 /**
1337  * Extract and broadcast texture size.
1338  *
1339  * @param size_type   type of the texture size vector (either
1340  *                    bld->int_size_type or bld->float_size_type)
1341  * @param coord_type  type of the texture size vector (either
1342  *                    bld->int_coord_type or bld->coord_type)
1343  * @param size        vector with the texture size (width, height, depth)
1344  */
1345 void
lp_build_extract_image_sizes(struct lp_build_sample_context * bld,struct lp_build_context * size_bld,struct lp_type coord_type,LLVMValueRef size,LLVMValueRef * out_width,LLVMValueRef * out_height,LLVMValueRef * out_depth)1346 lp_build_extract_image_sizes(struct lp_build_sample_context *bld,
1347                              struct lp_build_context *size_bld,
1348                              struct lp_type coord_type,
1349                              LLVMValueRef size,
1350                              LLVMValueRef *out_width,
1351                              LLVMValueRef *out_height,
1352                              LLVMValueRef *out_depth)
1353 {
1354    const unsigned dims = bld->dims;
1355    LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
1356    struct lp_type size_type = size_bld->type;
1357 
1358    if (bld->num_mips == 1) {
1359       *out_width = lp_build_extract_broadcast(bld->gallivm,
1360                                               size_type,
1361                                               coord_type,
1362                                               size,
1363                                               LLVMConstInt(i32t, 0, 0));
1364       if (dims >= 2) {
1365          *out_height = lp_build_extract_broadcast(bld->gallivm,
1366                                                   size_type,
1367                                                   coord_type,
1368                                                   size,
1369                                                   LLVMConstInt(i32t, 1, 0));
1370          if (dims == 3) {
1371             *out_depth = lp_build_extract_broadcast(bld->gallivm,
1372                                                     size_type,
1373                                                     coord_type,
1374                                                     size,
1375                                                     LLVMConstInt(i32t, 2, 0));
1376          }
1377       }
1378    }
1379    else {
1380       unsigned num_quads = bld->coord_bld.type.length / 4;
1381 
1382       if (dims == 1) {
1383          *out_width = size;
1384       }
1385       else if (bld->num_mips == num_quads) {
1386          *out_width = lp_build_swizzle_scalar_aos(size_bld, size, 0, 4);
1387          if (dims >= 2) {
1388             *out_height = lp_build_swizzle_scalar_aos(size_bld, size, 1, 4);
1389             if (dims == 3) {
1390                *out_depth = lp_build_swizzle_scalar_aos(size_bld, size, 2, 4);
1391             }
1392          }
1393       }
1394       else {
1395          assert(bld->num_mips == bld->coord_type.length);
1396          *out_width = lp_build_pack_aos_scalars(bld->gallivm, size_type,
1397                                                 coord_type, size, 0);
1398          if (dims >= 2) {
1399             *out_height = lp_build_pack_aos_scalars(bld->gallivm, size_type,
1400                                                     coord_type, size, 1);
1401             if (dims == 3) {
1402                *out_depth = lp_build_pack_aos_scalars(bld->gallivm, size_type,
1403                                                       coord_type, size, 2);
1404             }
1405          }
1406       }
1407    }
1408 }
1409 
1410 
1411 /**
1412  * Unnormalize coords.
1413  *
1414  * @param flt_size  vector with the integer texture size (width, height, depth)
1415  */
1416 void
lp_build_unnormalized_coords(struct lp_build_sample_context * bld,LLVMValueRef flt_size,LLVMValueRef * s,LLVMValueRef * t,LLVMValueRef * r)1417 lp_build_unnormalized_coords(struct lp_build_sample_context *bld,
1418                              LLVMValueRef flt_size,
1419                              LLVMValueRef *s,
1420                              LLVMValueRef *t,
1421                              LLVMValueRef *r)
1422 {
1423    const unsigned dims = bld->dims;
1424    LLVMValueRef width;
1425    LLVMValueRef height = NULL;
1426    LLVMValueRef depth = NULL;
1427 
1428    lp_build_extract_image_sizes(bld,
1429                                 &bld->float_size_bld,
1430                                 bld->coord_type,
1431                                 flt_size,
1432                                 &width,
1433                                 &height,
1434                                 &depth);
1435 
1436    /* s = s * width, t = t * height */
1437    *s = lp_build_mul(&bld->coord_bld, *s, width);
1438    if (dims >= 2) {
1439       *t = lp_build_mul(&bld->coord_bld, *t, height);
1440       if (dims >= 3) {
1441          *r = lp_build_mul(&bld->coord_bld, *r, depth);
1442       }
1443    }
1444 }
1445 
1446 /**
1447  * Generate new coords and faces for cubemap texels falling off the face.
1448  *
1449  * @param face   face (center) of the pixel
1450  * @param x0     lower x coord
1451  * @param x1     higher x coord (must be x0 + 1)
1452  * @param y0     lower y coord
1453  * @param y1     higher y coord (must be x0 + 1)
1454  * @param max_coord     texture cube (level) size - 1
1455  * @param next_faces    new face values when falling off
1456  * @param next_xcoords  new x coord values when falling off
1457  * @param next_ycoords  new y coord values when falling off
1458  *
1459  * The arrays hold the new values when under/overflow of
1460  * lower x, higher x, lower y, higher y coord would occur (in this order).
1461  * next_xcoords/next_ycoords have two entries each (for both new lower and
1462  * higher coord).
1463  */
1464 void
lp_build_cube_new_coords(struct lp_build_context * ivec_bld,LLVMValueRef face,LLVMValueRef x0,LLVMValueRef x1,LLVMValueRef y0,LLVMValueRef y1,LLVMValueRef max_coord,LLVMValueRef next_faces[4],LLVMValueRef next_xcoords[4][2],LLVMValueRef next_ycoords[4][2])1465 lp_build_cube_new_coords(struct lp_build_context *ivec_bld,
1466                         LLVMValueRef face,
1467                         LLVMValueRef x0,
1468                         LLVMValueRef x1,
1469                         LLVMValueRef y0,
1470                         LLVMValueRef y1,
1471                         LLVMValueRef max_coord,
1472                         LLVMValueRef next_faces[4],
1473                         LLVMValueRef next_xcoords[4][2],
1474                         LLVMValueRef next_ycoords[4][2])
1475 {
1476    /*
1477     * Lookup tables aren't nice for simd code hence try some logic here.
1478     * (Note that while it would not be necessary to do per-sample (4) lookups
1479     * when using a LUT as it's impossible that texels fall off of positive
1480     * and negative edges simultaneously, it would however be necessary to
1481     * do 2 lookups for corner handling as in this case texels both fall off
1482     * of x and y axes.)
1483     */
1484    /*
1485     * Next faces (for face 012345):
1486     * x < 0.0  : 451110
1487     * x >= 1.0 : 540001
1488     * y < 0.0  : 225422
1489     * y >= 1.0 : 334533
1490     * Hence nfx+ (and nfy+) == nfx- (nfy-) xor 1
1491     * nfx-: face > 1 ? (face == 5 ? 0 : 1) : (4 + face & 1)
1492     * nfy+: face & ~4 > 1 ? face + 2 : 3;
1493     * This could also use pshufb instead, but would need (manually coded)
1494     * ssse3 intrinsic (llvm won't do non-constant shuffles).
1495     */
1496    struct gallivm_state *gallivm = ivec_bld->gallivm;
1497    LLVMValueRef sel, sel_f2345, sel_f23, sel_f2, tmpsel, tmp;
1498    LLVMValueRef faceand1, sel_fand1, maxmx0, maxmx1, maxmy0, maxmy1;
1499    LLVMValueRef c2 = lp_build_const_int_vec(gallivm, ivec_bld->type, 2);
1500    LLVMValueRef c3 = lp_build_const_int_vec(gallivm, ivec_bld->type, 3);
1501    LLVMValueRef c4 = lp_build_const_int_vec(gallivm, ivec_bld->type, 4);
1502    LLVMValueRef c5 = lp_build_const_int_vec(gallivm, ivec_bld->type, 5);
1503 
1504    sel = lp_build_cmp(ivec_bld, PIPE_FUNC_EQUAL, face, c5);
1505    tmpsel = lp_build_select(ivec_bld, sel, ivec_bld->zero, ivec_bld->one);
1506    sel_f2345 = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, face, ivec_bld->one);
1507    faceand1 = lp_build_and(ivec_bld, face, ivec_bld->one);
1508    tmp = lp_build_add(ivec_bld, faceand1, c4);
1509    next_faces[0] = lp_build_select(ivec_bld, sel_f2345, tmpsel, tmp);
1510    next_faces[1] = lp_build_xor(ivec_bld, next_faces[0], ivec_bld->one);
1511 
1512    tmp = lp_build_andnot(ivec_bld, face, c4);
1513    sel_f23 = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, tmp, ivec_bld->one);
1514    tmp = lp_build_add(ivec_bld, face, c2);
1515    next_faces[3] = lp_build_select(ivec_bld, sel_f23, tmp, c3);
1516    next_faces[2] = lp_build_xor(ivec_bld, next_faces[3], ivec_bld->one);
1517 
1518    /*
1519     * new xcoords (for face 012345):
1520     * x < 0.0  : max   max   t     max-t max  max
1521     * x >= 1.0 : 0     0     max-t t     0    0
1522     * y < 0.0  : max   0     max-s s     s    max-s
1523     * y >= 1.0 : max   0     s     max-s s    max-s
1524     *
1525     * ncx[1] = face & ~4 > 1 ? (face == 2 ? max-t : t) : 0
1526     * ncx[0] = max - ncx[1]
1527     * ncx[3] = face > 1 ? (face & 1 ? max-s : s) : (face & 1) ? 0 : max
1528     * ncx[2] = face & ~4 > 1 ? max - ncx[3] : ncx[3]
1529     */
1530    sel_f2 = lp_build_cmp(ivec_bld, PIPE_FUNC_EQUAL, face, c2);
1531    maxmy0 = lp_build_sub(ivec_bld, max_coord, y0);
1532    tmp = lp_build_select(ivec_bld, sel_f2, maxmy0, y0);
1533    next_xcoords[1][0] = lp_build_select(ivec_bld, sel_f23, tmp, ivec_bld->zero);
1534    next_xcoords[0][0] = lp_build_sub(ivec_bld, max_coord, next_xcoords[1][0]);
1535    maxmy1 = lp_build_sub(ivec_bld, max_coord, y1);
1536    tmp = lp_build_select(ivec_bld, sel_f2, maxmy1, y1);
1537    next_xcoords[1][1] = lp_build_select(ivec_bld, sel_f23, tmp, ivec_bld->zero);
1538    next_xcoords[0][1] = lp_build_sub(ivec_bld, max_coord, next_xcoords[1][1]);
1539 
1540    sel_fand1 = lp_build_cmp(ivec_bld, PIPE_FUNC_EQUAL, faceand1, ivec_bld->one);
1541 
1542    tmpsel = lp_build_select(ivec_bld, sel_fand1, ivec_bld->zero, max_coord);
1543    maxmx0 = lp_build_sub(ivec_bld, max_coord, x0);
1544    tmp = lp_build_select(ivec_bld, sel_fand1, maxmx0, x0);
1545    next_xcoords[3][0] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1546    tmp = lp_build_sub(ivec_bld, max_coord, next_xcoords[3][0]);
1547    next_xcoords[2][0] = lp_build_select(ivec_bld, sel_f23, tmp, next_xcoords[3][0]);
1548    maxmx1 = lp_build_sub(ivec_bld, max_coord, x1);
1549    tmp = lp_build_select(ivec_bld, sel_fand1, maxmx1, x1);
1550    next_xcoords[3][1] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1551    tmp = lp_build_sub(ivec_bld, max_coord, next_xcoords[3][1]);
1552    next_xcoords[2][1] = lp_build_select(ivec_bld, sel_f23, tmp, next_xcoords[3][1]);
1553 
1554    /*
1555     * new ycoords (for face 012345):
1556     * x < 0.0  : t     t     0     max   t    t
1557     * x >= 1.0 : t     t     0     max   t    t
1558     * y < 0.0  : max-s s     0     max   max  0
1559     * y >= 1.0 : s     max-s 0     max   0    max
1560     *
1561     * ncy[0] = face & ~4 > 1 ? (face == 2 ? 0 : max) : t
1562     * ncy[1] = ncy[0]
1563     * ncy[3] = face > 1 ? (face & 1 ? max : 0) : (face & 1) ? max-s : max
1564     * ncx[2] = face & ~4 > 1 ? max - ncx[3] : ncx[3]
1565     */
1566    tmp = lp_build_select(ivec_bld, sel_f2, ivec_bld->zero, max_coord);
1567    next_ycoords[0][0] = lp_build_select(ivec_bld, sel_f23, tmp, y0);
1568    next_ycoords[1][0] = next_ycoords[0][0];
1569    next_ycoords[0][1] = lp_build_select(ivec_bld, sel_f23, tmp, y1);
1570    next_ycoords[1][1] = next_ycoords[0][1];
1571 
1572    tmpsel = lp_build_select(ivec_bld, sel_fand1, maxmx0, x0);
1573    tmp = lp_build_select(ivec_bld, sel_fand1, max_coord, ivec_bld->zero);
1574    next_ycoords[3][0] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1575    tmp = lp_build_sub(ivec_bld, max_coord, next_ycoords[3][0]);
1576    next_ycoords[2][0] = lp_build_select(ivec_bld, sel_f23, next_ycoords[3][0], tmp);
1577    tmpsel = lp_build_select(ivec_bld, sel_fand1, maxmx1, x1);
1578    tmp = lp_build_select(ivec_bld, sel_fand1, max_coord, ivec_bld->zero);
1579    next_ycoords[3][1] = lp_build_select(ivec_bld, sel_f2345, tmp, tmpsel);
1580    tmp = lp_build_sub(ivec_bld, max_coord, next_ycoords[3][1]);
1581    next_ycoords[2][1] = lp_build_select(ivec_bld, sel_f23, next_ycoords[3][1], tmp);
1582 }
1583 
1584 
1585 /** Helper used by lp_build_cube_lookup() */
1586 static LLVMValueRef
lp_build_cube_imapos(struct lp_build_context * coord_bld,LLVMValueRef coord)1587 lp_build_cube_imapos(struct lp_build_context *coord_bld, LLVMValueRef coord)
1588 {
1589    /* ima = +0.5 / abs(coord); */
1590    LLVMValueRef posHalf = lp_build_const_vec(coord_bld->gallivm, coord_bld->type, 0.5);
1591    LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1592    LLVMValueRef ima = lp_build_div(coord_bld, posHalf, absCoord);
1593    return ima;
1594 }
1595 
1596 
1597 /** Helper for doing 3-wise selection.
1598  * Returns sel1 ? val2 : (sel0 ? val0 : val1).
1599  */
1600 static LLVMValueRef
lp_build_select3(struct lp_build_context * sel_bld,LLVMValueRef sel0,LLVMValueRef sel1,LLVMValueRef val0,LLVMValueRef val1,LLVMValueRef val2)1601 lp_build_select3(struct lp_build_context *sel_bld,
1602                  LLVMValueRef sel0,
1603                  LLVMValueRef sel1,
1604                  LLVMValueRef val0,
1605                  LLVMValueRef val1,
1606                  LLVMValueRef val2)
1607 {
1608    LLVMValueRef tmp;
1609    tmp = lp_build_select(sel_bld, sel0, val0, val1);
1610    return lp_build_select(sel_bld, sel1, val2, tmp);
1611 }
1612 
1613 
1614 /**
1615  * Generate code to do cube face selection and compute per-face texcoords.
1616  */
1617 void
lp_build_cube_lookup(struct lp_build_sample_context * bld,LLVMValueRef * coords,const struct lp_derivatives * derivs_in,LLVMValueRef * rho,struct lp_derivatives * derivs_out,boolean need_derivs)1618 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1619                      LLVMValueRef *coords,
1620                      const struct lp_derivatives *derivs_in, /* optional */
1621                      LLVMValueRef *rho,
1622                      struct lp_derivatives *derivs_out, /* optional */
1623                      boolean need_derivs)
1624 {
1625    struct lp_build_context *coord_bld = &bld->coord_bld;
1626    LLVMBuilderRef builder = bld->gallivm->builder;
1627    struct gallivm_state *gallivm = bld->gallivm;
1628    LLVMValueRef si, ti, ri;
1629 
1630    /*
1631     * Do per-pixel face selection. We cannot however (as we used to do)
1632     * simply calculate the derivs afterwards (which is very bogus for
1633     * explicit derivs btw) because the values would be "random" when
1634     * not all pixels lie on the same face. So what we do here is just
1635     * calculate the derivatives after scaling the coords by the absolute
1636     * value of the inverse major axis, and essentially do rho calculation
1637     * steps as if it were a 3d texture. This is perfect if all pixels hit
1638     * the same face, but not so great at edges, I believe the max error
1639     * should be sqrt(2) with no_rho_approx or 2 otherwise (essentially measuring
1640     * the 3d distance between 2 points on the cube instead of measuring up/down
1641     * the edge). Still this is possibly a win over just selecting the same face
1642     * for all pixels. Unfortunately, something like that doesn't work for
1643     * explicit derivatives.
1644     */
1645    struct lp_build_context *cint_bld = &bld->int_coord_bld;
1646    struct lp_type intctype = cint_bld->type;
1647    LLVMTypeRef coord_vec_type = coord_bld->vec_type;
1648    LLVMTypeRef cint_vec_type = cint_bld->vec_type;
1649    LLVMValueRef as, at, ar, face, face_s, face_t;
1650    LLVMValueRef as_ge_at, maxasat, ar_ge_as_at;
1651    LLVMValueRef snewx, tnewx, snewy, tnewy, snewz, tnewz;
1652    LLVMValueRef tnegi, rnegi;
1653    LLVMValueRef ma, mai, signma, signmabit, imahalfpos;
1654    LLVMValueRef posHalf = lp_build_const_vec(gallivm, coord_bld->type, 0.5);
1655    LLVMValueRef signmask = lp_build_const_int_vec(gallivm, intctype,
1656                                                   1LL << (intctype.width - 1));
1657    LLVMValueRef signshift = lp_build_const_int_vec(gallivm, intctype,
1658                                                    intctype.width -1);
1659    LLVMValueRef facex = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_X);
1660    LLVMValueRef facey = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_Y);
1661    LLVMValueRef facez = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_Z);
1662    LLVMValueRef s = coords[0];
1663    LLVMValueRef t = coords[1];
1664    LLVMValueRef r = coords[2];
1665 
1666    assert(PIPE_TEX_FACE_NEG_X == PIPE_TEX_FACE_POS_X + 1);
1667    assert(PIPE_TEX_FACE_NEG_Y == PIPE_TEX_FACE_POS_Y + 1);
1668    assert(PIPE_TEX_FACE_NEG_Z == PIPE_TEX_FACE_POS_Z + 1);
1669 
1670    /*
1671     * get absolute value (for x/y/z face selection) and sign bit
1672     * (for mirroring minor coords and pos/neg face selection)
1673     * of the original coords.
1674     */
1675    as = lp_build_abs(&bld->coord_bld, s);
1676    at = lp_build_abs(&bld->coord_bld, t);
1677    ar = lp_build_abs(&bld->coord_bld, r);
1678 
1679    /*
1680     * major face determination: select x if x > y else select y
1681     * select z if z >= max(x,y) else select previous result
1682     * if some axis are the same we chose z over y, y over x - the
1683     * dx10 spec seems to ask for it while OpenGL doesn't care (if we
1684     * wouldn't care could save a select or two if using different
1685     * compares and doing at_g_as_ar last since tnewx and tnewz are the
1686     * same).
1687     */
1688    as_ge_at = lp_build_cmp(coord_bld, PIPE_FUNC_GREATER, as, at);
1689    maxasat = lp_build_max(coord_bld, as, at);
1690    ar_ge_as_at = lp_build_cmp(coord_bld, PIPE_FUNC_GEQUAL, ar, maxasat);
1691 
1692    if (need_derivs && (derivs_in || (bld->no_quad_lod && bld->no_rho_approx))) {
1693       /*
1694        * XXX: This is really really complex.
1695        * It is a bit overkill to use this for implicit derivatives as well,
1696        * no way this is worth the cost in practice, but seems to be the
1697        * only way for getting accurate and per-pixel lod values.
1698        */
1699       LLVMValueRef ima, imahalf, tmp, ddx[3], ddy[3];
1700       LLVMValueRef madx, mady, madxdivma, madydivma;
1701       LLVMValueRef sdxi, tdxi, rdxi, sdyi, tdyi, rdyi;
1702       LLVMValueRef tdxnegi, rdxnegi, tdynegi, rdynegi;
1703       LLVMValueRef sdxnewx, sdxnewy, sdxnewz, tdxnewx, tdxnewy, tdxnewz;
1704       LLVMValueRef sdynewx, sdynewy, sdynewz, tdynewx, tdynewy, tdynewz;
1705       LLVMValueRef face_sdx, face_tdx, face_sdy, face_tdy;
1706       /*
1707        * s = 1/2 * ( sc / ma + 1)
1708        * t = 1/2 * ( tc / ma + 1)
1709        *
1710        * s' = 1/2 * (sc' * ma - sc * ma') / ma^2
1711        * t' = 1/2 * (tc' * ma - tc * ma') / ma^2
1712        *
1713        * dx.s = 0.5 * (dx.sc - sc * dx.ma / ma) / ma
1714        * dx.t = 0.5 * (dx.tc - tc * dx.ma / ma) / ma
1715        * dy.s = 0.5 * (dy.sc - sc * dy.ma / ma) / ma
1716        * dy.t = 0.5 * (dy.tc - tc * dy.ma / ma) / ma
1717        */
1718 
1719       /* select ma, calculate ima */
1720       ma = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, s, t, r);
1721       mai = LLVMBuildBitCast(builder, ma, cint_vec_type, "");
1722       signmabit = LLVMBuildAnd(builder, mai, signmask, "");
1723       ima = lp_build_div(coord_bld, coord_bld->one, ma);
1724       imahalf = lp_build_mul(coord_bld, posHalf, ima);
1725       imahalfpos = lp_build_abs(coord_bld, imahalf);
1726 
1727       if (!derivs_in) {
1728          ddx[0] = lp_build_ddx(coord_bld, s);
1729          ddx[1] = lp_build_ddx(coord_bld, t);
1730          ddx[2] = lp_build_ddx(coord_bld, r);
1731          ddy[0] = lp_build_ddy(coord_bld, s);
1732          ddy[1] = lp_build_ddy(coord_bld, t);
1733          ddy[2] = lp_build_ddy(coord_bld, r);
1734       }
1735       else {
1736          ddx[0] = derivs_in->ddx[0];
1737          ddx[1] = derivs_in->ddx[1];
1738          ddx[2] = derivs_in->ddx[2];
1739          ddy[0] = derivs_in->ddy[0];
1740          ddy[1] = derivs_in->ddy[1];
1741          ddy[2] = derivs_in->ddy[2];
1742       }
1743 
1744       /* select major derivatives */
1745       madx = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, ddx[0], ddx[1], ddx[2]);
1746       mady = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, ddy[0], ddy[1], ddy[2]);
1747 
1748       si = LLVMBuildBitCast(builder, s, cint_vec_type, "");
1749       ti = LLVMBuildBitCast(builder, t, cint_vec_type, "");
1750       ri = LLVMBuildBitCast(builder, r, cint_vec_type, "");
1751 
1752       sdxi = LLVMBuildBitCast(builder, ddx[0], cint_vec_type, "");
1753       tdxi = LLVMBuildBitCast(builder, ddx[1], cint_vec_type, "");
1754       rdxi = LLVMBuildBitCast(builder, ddx[2], cint_vec_type, "");
1755 
1756       sdyi = LLVMBuildBitCast(builder, ddy[0], cint_vec_type, "");
1757       tdyi = LLVMBuildBitCast(builder, ddy[1], cint_vec_type, "");
1758       rdyi = LLVMBuildBitCast(builder, ddy[2], cint_vec_type, "");
1759 
1760       /*
1761        * compute all possible new s/t coords, which does the mirroring,
1762        * and do the same for derivs minor axes.
1763        * snewx = signma * -r;
1764        * tnewx = -t;
1765        * snewy = s;
1766        * tnewy = signma * r;
1767        * snewz = signma * s;
1768        * tnewz = -t;
1769        */
1770       tnegi = LLVMBuildXor(builder, ti, signmask, "");
1771       rnegi = LLVMBuildXor(builder, ri, signmask, "");
1772       tdxnegi = LLVMBuildXor(builder, tdxi, signmask, "");
1773       rdxnegi = LLVMBuildXor(builder, rdxi, signmask, "");
1774       tdynegi = LLVMBuildXor(builder, tdyi, signmask, "");
1775       rdynegi = LLVMBuildXor(builder, rdyi, signmask, "");
1776 
1777       snewx = LLVMBuildXor(builder, signmabit, rnegi, "");
1778       tnewx = tnegi;
1779       sdxnewx = LLVMBuildXor(builder, signmabit, rdxnegi, "");
1780       tdxnewx = tdxnegi;
1781       sdynewx = LLVMBuildXor(builder, signmabit, rdynegi, "");
1782       tdynewx = tdynegi;
1783 
1784       snewy = si;
1785       tnewy = LLVMBuildXor(builder, signmabit, ri, "");
1786       sdxnewy = sdxi;
1787       tdxnewy = LLVMBuildXor(builder, signmabit, rdxi, "");
1788       sdynewy = sdyi;
1789       tdynewy = LLVMBuildXor(builder, signmabit, rdyi, "");
1790 
1791       snewz = LLVMBuildXor(builder, signmabit, si, "");
1792       tnewz = tnegi;
1793       sdxnewz = LLVMBuildXor(builder, signmabit, sdxi, "");
1794       tdxnewz = tdxnegi;
1795       sdynewz = LLVMBuildXor(builder, signmabit, sdyi, "");
1796       tdynewz = tdynegi;
1797 
1798       /* select the mirrored values */
1799       face = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, facex, facey, facez);
1800       face_s = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, snewx, snewy, snewz);
1801       face_t = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tnewx, tnewy, tnewz);
1802       face_sdx = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, sdxnewx, sdxnewy, sdxnewz);
1803       face_tdx = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tdxnewx, tdxnewy, tdxnewz);
1804       face_sdy = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, sdynewx, sdynewy, sdynewz);
1805       face_tdy = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tdynewx, tdynewy, tdynewz);
1806 
1807       face_s = LLVMBuildBitCast(builder, face_s, coord_vec_type, "");
1808       face_t = LLVMBuildBitCast(builder, face_t, coord_vec_type, "");
1809       face_sdx = LLVMBuildBitCast(builder, face_sdx, coord_vec_type, "");
1810       face_tdx = LLVMBuildBitCast(builder, face_tdx, coord_vec_type, "");
1811       face_sdy = LLVMBuildBitCast(builder, face_sdy, coord_vec_type, "");
1812       face_tdy = LLVMBuildBitCast(builder, face_tdy, coord_vec_type, "");
1813 
1814       /* deriv math, dx.s = 0.5 * (dx.sc - sc * dx.ma / ma) / ma */
1815       madxdivma = lp_build_mul(coord_bld, madx, ima);
1816       tmp = lp_build_mul(coord_bld, madxdivma, face_s);
1817       tmp = lp_build_sub(coord_bld, face_sdx, tmp);
1818       derivs_out->ddx[0] = lp_build_mul(coord_bld, tmp, imahalf);
1819 
1820       /* dx.t = 0.5 * (dx.tc - tc * dx.ma / ma) / ma */
1821       tmp = lp_build_mul(coord_bld, madxdivma, face_t);
1822       tmp = lp_build_sub(coord_bld, face_tdx, tmp);
1823       derivs_out->ddx[1] = lp_build_mul(coord_bld, tmp, imahalf);
1824 
1825       /* dy.s = 0.5 * (dy.sc - sc * dy.ma / ma) / ma */
1826       madydivma = lp_build_mul(coord_bld, mady, ima);
1827       tmp = lp_build_mul(coord_bld, madydivma, face_s);
1828       tmp = lp_build_sub(coord_bld, face_sdy, tmp);
1829       derivs_out->ddy[0] = lp_build_mul(coord_bld, tmp, imahalf);
1830 
1831       /* dy.t = 0.5 * (dy.tc - tc * dy.ma / ma) / ma */
1832       tmp = lp_build_mul(coord_bld, madydivma, face_t);
1833       tmp = lp_build_sub(coord_bld, face_tdy, tmp);
1834       derivs_out->ddy[1] = lp_build_mul(coord_bld, tmp, imahalf);
1835 
1836       signma = LLVMBuildLShr(builder, mai, signshift, "");
1837       coords[2] = LLVMBuildOr(builder, face, signma, "face");
1838 
1839       /* project coords */
1840       face_s = lp_build_mul(coord_bld, face_s, imahalfpos);
1841       face_t = lp_build_mul(coord_bld, face_t, imahalfpos);
1842 
1843       coords[0] = lp_build_add(coord_bld, face_s, posHalf);
1844       coords[1] = lp_build_add(coord_bld, face_t, posHalf);
1845 
1846       return;
1847    }
1848 
1849    else if (need_derivs) {
1850       LLVMValueRef ddx_ddy[2], tmp[3], rho_vec;
1851       static const unsigned char swizzle0[] = { /* no-op swizzle */
1852          0, LP_BLD_SWIZZLE_DONTCARE,
1853          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1854       };
1855       static const unsigned char swizzle1[] = {
1856          1, LP_BLD_SWIZZLE_DONTCARE,
1857          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1858       };
1859       static const unsigned char swizzle01[] = { /* no-op swizzle */
1860          0, 1,
1861          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1862       };
1863       static const unsigned char swizzle23[] = {
1864          2, 3,
1865          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1866       };
1867       static const unsigned char swizzle02[] = {
1868          0, 2,
1869          LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
1870       };
1871 
1872       /*
1873        * scale the s/t/r coords pre-select/mirror so we can calculate
1874        * "reasonable" derivs.
1875        */
1876       ma = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, s, t, r);
1877       imahalfpos = lp_build_cube_imapos(coord_bld, ma);
1878       s = lp_build_mul(coord_bld, s, imahalfpos);
1879       t = lp_build_mul(coord_bld, t, imahalfpos);
1880       r = lp_build_mul(coord_bld, r, imahalfpos);
1881 
1882       /*
1883        * This isn't quite the same as the "ordinary" (3d deriv) path since we
1884        * know the texture is square which simplifies things (we can omit the
1885        * size mul which happens very early completely here and do it at the
1886        * very end).
1887        * Also always do calculations according to GALLIVM_DEBUG_NO_RHO_APPROX
1888        * since the error can get quite big otherwise at edges.
1889        * (With no_rho_approx max error is sqrt(2) at edges, same as it is
1890        * without no_rho_approx for 2d textures, otherwise it would be factor 2.)
1891        */
1892       ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(coord_bld, s, t);
1893       ddx_ddy[1] = lp_build_packed_ddx_ddy_onecoord(coord_bld, r);
1894 
1895       ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], ddx_ddy[0]);
1896       ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], ddx_ddy[1]);
1897 
1898       tmp[0] = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle01);
1899       tmp[1] = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle23);
1900       tmp[2] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle02);
1901 
1902       rho_vec = lp_build_add(coord_bld, tmp[0], tmp[1]);
1903       rho_vec = lp_build_add(coord_bld, rho_vec, tmp[2]);
1904 
1905       tmp[0] = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
1906       tmp[1] = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
1907       *rho = lp_build_max(coord_bld, tmp[0], tmp[1]);
1908    }
1909 
1910    if (!need_derivs) {
1911       ma = lp_build_select3(coord_bld, as_ge_at, ar_ge_as_at, s, t, r);
1912    }
1913    mai = LLVMBuildBitCast(builder, ma, cint_vec_type, "");
1914    signmabit = LLVMBuildAnd(builder, mai, signmask, "");
1915 
1916    si = LLVMBuildBitCast(builder, s, cint_vec_type, "");
1917    ti = LLVMBuildBitCast(builder, t, cint_vec_type, "");
1918    ri = LLVMBuildBitCast(builder, r, cint_vec_type, "");
1919 
1920    /*
1921     * compute all possible new s/t coords, which does the mirroring
1922     * snewx = signma * -r;
1923     * tnewx = -t;
1924     * snewy = s;
1925     * tnewy = signma * r;
1926     * snewz = signma * s;
1927     * tnewz = -t;
1928     */
1929    tnegi = LLVMBuildXor(builder, ti, signmask, "");
1930    rnegi = LLVMBuildXor(builder, ri, signmask, "");
1931 
1932    snewx = LLVMBuildXor(builder, signmabit, rnegi, "");
1933    tnewx = tnegi;
1934 
1935    snewy = si;
1936    tnewy = LLVMBuildXor(builder, signmabit, ri, "");
1937 
1938    snewz = LLVMBuildXor(builder, signmabit, si, "");
1939    tnewz = tnegi;
1940 
1941    /* select the mirrored values */
1942    face_s = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, snewx, snewy, snewz);
1943    face_t = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, tnewx, tnewy, tnewz);
1944    face = lp_build_select3(cint_bld, as_ge_at, ar_ge_as_at, facex, facey, facez);
1945 
1946    face_s = LLVMBuildBitCast(builder, face_s, coord_vec_type, "");
1947    face_t = LLVMBuildBitCast(builder, face_t, coord_vec_type, "");
1948 
1949    /* add +1 for neg face */
1950    /* XXX with AVX probably want to use another select here -
1951     * as long as we ensure vblendvps gets used we can actually
1952     * skip the comparison and just use sign as a "mask" directly.
1953     */
1954    signma = LLVMBuildLShr(builder, mai, signshift, "");
1955    coords[2] = LLVMBuildOr(builder, face, signma, "face");
1956 
1957    /* project coords */
1958    if (!need_derivs) {
1959       imahalfpos = lp_build_cube_imapos(coord_bld, ma);
1960       face_s = lp_build_mul(coord_bld, face_s, imahalfpos);
1961       face_t = lp_build_mul(coord_bld, face_t, imahalfpos);
1962    }
1963 
1964    coords[0] = lp_build_add(coord_bld, face_s, posHalf);
1965    coords[1] = lp_build_add(coord_bld, face_t, posHalf);
1966 }
1967 
1968 
1969 /**
1970  * Compute the partial offset of a pixel block along an arbitrary axis.
1971  *
1972  * @param coord   coordinate in pixels
1973  * @param stride  number of bytes between rows of successive pixel blocks
1974  * @param block_length  number of pixels in a pixels block along the coordinate
1975  *                      axis
1976  * @param out_offset    resulting relative offset of the pixel block in bytes
1977  * @param out_subcoord  resulting sub-block pixel coordinate
1978  */
1979 void
lp_build_sample_partial_offset(struct lp_build_context * bld,unsigned block_length,LLVMValueRef coord,LLVMValueRef stride,LLVMValueRef * out_offset,LLVMValueRef * out_subcoord)1980 lp_build_sample_partial_offset(struct lp_build_context *bld,
1981                                unsigned block_length,
1982                                LLVMValueRef coord,
1983                                LLVMValueRef stride,
1984                                LLVMValueRef *out_offset,
1985                                LLVMValueRef *out_subcoord)
1986 {
1987    LLVMBuilderRef builder = bld->gallivm->builder;
1988    LLVMValueRef offset;
1989    LLVMValueRef subcoord;
1990 
1991    if (block_length == 1) {
1992       subcoord = bld->zero;
1993    }
1994    else {
1995       /*
1996        * Pixel blocks have power of two dimensions. LLVM should convert the
1997        * rem/div to bit arithmetic.
1998        * TODO: Verify this.
1999        * It does indeed BUT it does transform it to scalar (and back) when doing so
2000        * (using roughly extract, shift/and, mov, unpack) (llvm 2.7).
2001        * The generated code looks seriously unfunny and is quite expensive.
2002        */
2003 #if 0
2004       LLVMValueRef block_width = lp_build_const_int_vec(bld->type, block_length);
2005       subcoord = LLVMBuildURem(builder, coord, block_width, "");
2006       coord    = LLVMBuildUDiv(builder, coord, block_width, "");
2007 #else
2008       unsigned logbase2 = util_logbase2(block_length);
2009       LLVMValueRef block_shift = lp_build_const_int_vec(bld->gallivm, bld->type, logbase2);
2010       LLVMValueRef block_mask = lp_build_const_int_vec(bld->gallivm, bld->type, block_length - 1);
2011       subcoord = LLVMBuildAnd(builder, coord, block_mask, "");
2012       coord = LLVMBuildLShr(builder, coord, block_shift, "");
2013 #endif
2014    }
2015 
2016    offset = lp_build_mul(bld, coord, stride);
2017 
2018    assert(out_offset);
2019    assert(out_subcoord);
2020 
2021    *out_offset = offset;
2022    *out_subcoord = subcoord;
2023 }
2024 
2025 
2026 /**
2027  * Compute the offset of a pixel block.
2028  *
2029  * x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
2030  *
2031  * Returns the relative offset and i,j sub-block coordinates
2032  */
2033 void
lp_build_sample_offset(struct lp_build_context * bld,const struct util_format_description * format_desc,LLVMValueRef x,LLVMValueRef y,LLVMValueRef z,LLVMValueRef y_stride,LLVMValueRef z_stride,LLVMValueRef * out_offset,LLVMValueRef * out_i,LLVMValueRef * out_j)2034 lp_build_sample_offset(struct lp_build_context *bld,
2035                        const struct util_format_description *format_desc,
2036                        LLVMValueRef x,
2037                        LLVMValueRef y,
2038                        LLVMValueRef z,
2039                        LLVMValueRef y_stride,
2040                        LLVMValueRef z_stride,
2041                        LLVMValueRef *out_offset,
2042                        LLVMValueRef *out_i,
2043                        LLVMValueRef *out_j)
2044 {
2045    LLVMValueRef x_stride;
2046    LLVMValueRef offset;
2047 
2048    x_stride = lp_build_const_vec(bld->gallivm, bld->type,
2049                                  format_desc->block.bits/8);
2050 
2051    lp_build_sample_partial_offset(bld,
2052                                   format_desc->block.width,
2053                                   x, x_stride,
2054                                   &offset, out_i);
2055 
2056    if (y && y_stride) {
2057       LLVMValueRef y_offset;
2058       lp_build_sample_partial_offset(bld,
2059                                      format_desc->block.height,
2060                                      y, y_stride,
2061                                      &y_offset, out_j);
2062       offset = lp_build_add(bld, offset, y_offset);
2063    }
2064    else {
2065       *out_j = bld->zero;
2066    }
2067 
2068    if (z && z_stride) {
2069       LLVMValueRef z_offset;
2070       LLVMValueRef k;
2071       lp_build_sample_partial_offset(bld,
2072                                      1, /* pixel blocks are always 2D */
2073                                      z, z_stride,
2074                                      &z_offset, &k);
2075       offset = lp_build_add(bld, offset, z_offset);
2076    }
2077 
2078    *out_offset = offset;
2079 }
2080