1/**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29/**
30 * @file
31 * Code generate the whole fragment pipeline.
32 *
33 * The fragment pipeline consists of the following stages:
34 * - early depth test
35 * - fragment shader
36 * - alpha test
37 * - depth/stencil test
38 * - blending
39 *
40 * This file has only the glue to assemble the fragment pipeline.  The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
45 *
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
51 *
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
56 *
57 * @author Jose Fonseca <jfonseca@vmware.com>
58 */
59
60#include <limits.h>
61#include "pipe/p_defines.h"
62#include "util/u_inlines.h"
63#include "util/u_memory.h"
64#include "util/u_pointer.h"
65#include "util/u_format.h"
66#include "util/u_dump.h"
67#include "util/u_string.h"
68#include "util/u_simple_list.h"
69#include "os/os_time.h"
70#include "pipe/p_shader_tokens.h"
71#include "draw/draw_context.h"
72#include "tgsi/tgsi_dump.h"
73#include "tgsi/tgsi_scan.h"
74#include "tgsi/tgsi_parse.h"
75#include "gallivm/lp_bld_type.h"
76#include "gallivm/lp_bld_const.h"
77#include "gallivm/lp_bld_conv.h"
78#include "gallivm/lp_bld_init.h"
79#include "gallivm/lp_bld_intr.h"
80#include "gallivm/lp_bld_logic.h"
81#include "gallivm/lp_bld_tgsi.h"
82#include "gallivm/lp_bld_swizzle.h"
83#include "gallivm/lp_bld_flow.h"
84#include "gallivm/lp_bld_debug.h"
85
86#include "lp_bld_alpha.h"
87#include "lp_bld_blend.h"
88#include "lp_bld_depth.h"
89#include "lp_bld_interp.h"
90#include "lp_context.h"
91#include "lp_debug.h"
92#include "lp_perf.h"
93#include "lp_setup.h"
94#include "lp_state.h"
95#include "lp_tex_sample.h"
96#include "lp_flush.h"
97#include "lp_state_fs.h"
98
99
100/** Fragment shader number (for debugging) */
101static unsigned fs_no = 0;
102
103
104/**
105 * Expand the relevant bits of mask_input to a n*4-dword mask for the
106 * n*four pixels in n 2x2 quads.  This will set the n*four elements of the
107 * quad mask vector to 0 or ~0.
108 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
109 * quad arguments with fs length 8.
110 *
111 * \param first_quad  which quad(s) of the quad group to test, in [0,3]
112 * \param mask_input  bitwise mask for the whole 4x4 stamp
113 */
114static LLVMValueRef
115generate_quad_mask(struct gallivm_state *gallivm,
116                   struct lp_type fs_type,
117                   unsigned first_quad,
118                   LLVMValueRef mask_input) /* int32 */
119{
120   LLVMBuilderRef builder = gallivm->builder;
121   struct lp_type mask_type;
122   LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
123   LLVMValueRef bits[16];
124   LLVMValueRef mask;
125   int shift, i;
126
127   /*
128    * XXX: We'll need a different path for 16 x u8
129    */
130   assert(fs_type.width == 32);
131   assert(fs_type.length <= Elements(bits));
132   mask_type = lp_int_type(fs_type);
133
134   /*
135    * mask_input >>= (quad * 4)
136    */
137   switch (first_quad) {
138   case 0:
139      shift = 0;
140      break;
141   case 1:
142      assert(fs_type.length == 4);
143      shift = 2;
144      break;
145   case 2:
146      shift = 8;
147      break;
148   case 3:
149      assert(fs_type.length == 4);
150      shift = 10;
151      break;
152   default:
153      assert(0);
154      shift = 0;
155   }
156
157   mask_input = LLVMBuildLShr(builder,
158                              mask_input,
159                              LLVMConstInt(i32t, shift, 0),
160                              "");
161
162   /*
163    * mask = { mask_input & (1 << i), for i in [0,3] }
164    */
165   mask = lp_build_broadcast(gallivm,
166                             lp_build_vec_type(gallivm, mask_type),
167                             mask_input);
168
169   for (i = 0; i < fs_type.length / 4; i++) {
170      unsigned j = 2 * (i % 2) + (i / 2) * 8;
171      bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
172      bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
173      bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
174      bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
175   }
176   mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
177
178   /*
179    * mask = mask != 0 ? ~0 : 0
180    */
181   mask = lp_build_compare(gallivm,
182                           mask_type, PIPE_FUNC_NOTEQUAL,
183                           mask,
184                           lp_build_const_int_vec(gallivm, mask_type, 0));
185
186   return mask;
187}
188
189
190#define EARLY_DEPTH_TEST  0x1
191#define LATE_DEPTH_TEST   0x2
192#define EARLY_DEPTH_WRITE 0x4
193#define LATE_DEPTH_WRITE  0x8
194
195static int
196find_output_by_semantic( const struct tgsi_shader_info *info,
197			 unsigned semantic,
198			 unsigned index )
199{
200   int i;
201
202   for (i = 0; i < info->num_outputs; i++)
203      if (info->output_semantic_name[i] == semantic &&
204	  info->output_semantic_index[i] == index)
205	 return i;
206
207   return -1;
208}
209
210
211/**
212 * Generate the fragment shader, depth/stencil test, and alpha tests.
213 * \param i  which quad in the tile, in range [0,3]
214 * \param partial_mask  if 1, do mask_input testing
215 */
216static void
217generate_fs(struct gallivm_state *gallivm,
218            struct lp_fragment_shader *shader,
219            const struct lp_fragment_shader_variant_key *key,
220            LLVMBuilderRef builder,
221            struct lp_type type,
222            LLVMValueRef context_ptr,
223            unsigned i,
224            struct lp_build_interp_soa_context *interp,
225            struct lp_build_sampler_soa *sampler,
226            LLVMValueRef *pmask,
227            LLVMValueRef (*color)[4],
228            LLVMValueRef depth_ptr,
229            LLVMValueRef facing,
230            unsigned partial_mask,
231            LLVMValueRef mask_input,
232            LLVMValueRef counter)
233{
234   const struct util_format_description *zs_format_desc = NULL;
235   const struct tgsi_token *tokens = shader->base.tokens;
236   LLVMTypeRef vec_type;
237   LLVMValueRef consts_ptr;
238   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
239   LLVMValueRef z;
240   LLVMValueRef zs_value = NULL;
241   LLVMValueRef stencil_refs[2];
242   struct lp_build_mask_context mask;
243   boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
244                            shader->info.base.num_inputs < 3 &&
245                            shader->info.base.num_instructions < 8);
246   unsigned attrib;
247   unsigned chan;
248   unsigned cbuf;
249   unsigned depth_mode;
250   struct lp_bld_tgsi_system_values system_values;
251
252   memset(&system_values, 0, sizeof(system_values));
253
254   if (key->depth.enabled ||
255       key->stencil[0].enabled ||
256       key->stencil[1].enabled) {
257
258      zs_format_desc = util_format_description(key->zsbuf_format);
259      assert(zs_format_desc);
260
261      if (!shader->info.base.writes_z) {
262         if (key->alpha.enabled || shader->info.base.uses_kill)
263            /* With alpha test and kill, can do the depth test early
264             * and hopefully eliminate some quads.  But need to do a
265             * special deferred depth write once the final mask value
266             * is known.
267             */
268            depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
269         else
270            depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
271      }
272      else {
273         depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
274      }
275
276      if (!(key->depth.enabled && key->depth.writemask) &&
277          !(key->stencil[0].enabled && key->stencil[0].writemask))
278         depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
279   }
280   else {
281      depth_mode = 0;
282   }
283
284   assert(i < 4);
285
286   stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
287   stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
288
289   vec_type = lp_build_vec_type(gallivm, type);
290
291   consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
292
293   memset(outputs, 0, sizeof outputs);
294
295   /* Declare the color and z variables */
296   for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
297      for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
298         color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
299      }
300   }
301
302   /* do triangle edge testing */
303   if (partial_mask) {
304      *pmask = generate_quad_mask(gallivm, type,
305                                  i*type.length/4, mask_input);
306   }
307   else {
308      *pmask = lp_build_const_int_vec(gallivm, type, ~0);
309   }
310
311   /* 'mask' will control execution based on quad's pixel alive/killed state */
312   lp_build_mask_begin(&mask, gallivm, type, *pmask);
313
314   if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
315      lp_build_mask_check(&mask);
316
317   lp_build_interp_soa_update_pos(interp, gallivm, i*type.length/4);
318   z = interp->pos[2];
319
320   if (depth_mode & EARLY_DEPTH_TEST) {
321      lp_build_depth_stencil_test(gallivm,
322                                  &key->depth,
323                                  key->stencil,
324                                  type,
325                                  zs_format_desc,
326                                  &mask,
327                                  stencil_refs,
328                                  z,
329                                  depth_ptr, facing,
330                                  &zs_value,
331                                  !simple_shader);
332
333      if (depth_mode & EARLY_DEPTH_WRITE) {
334         lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
335      }
336   }
337
338   lp_build_interp_soa_update_inputs(interp, gallivm, i*type.length/4);
339
340   /* Build the actual shader */
341   lp_build_tgsi_soa(gallivm, tokens, type, &mask,
342                     consts_ptr, &system_values,
343                     interp->pos, interp->inputs,
344                     outputs, sampler, &shader->info.base);
345
346   /* Alpha test */
347   if (key->alpha.enabled) {
348      int color0 = find_output_by_semantic(&shader->info.base,
349                                           TGSI_SEMANTIC_COLOR,
350                                           0);
351
352      if (color0 != -1 && outputs[color0][3]) {
353         const struct util_format_description *cbuf_format_desc;
354         LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
355         LLVMValueRef alpha_ref_value;
356
357         alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
358         alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
359
360         cbuf_format_desc = util_format_description(key->cbuf_format[0]);
361
362         lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
363                             &mask, alpha, alpha_ref_value,
364                             (depth_mode & LATE_DEPTH_TEST) != 0);
365      }
366   }
367
368   /* Late Z test */
369   if (depth_mode & LATE_DEPTH_TEST) {
370      int pos0 = find_output_by_semantic(&shader->info.base,
371                                         TGSI_SEMANTIC_POSITION,
372                                         0);
373
374      if (pos0 != -1 && outputs[pos0][2]) {
375         z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
376      }
377
378      lp_build_depth_stencil_test(gallivm,
379                                  &key->depth,
380                                  key->stencil,
381                                  type,
382                                  zs_format_desc,
383                                  &mask,
384                                  stencil_refs,
385                                  z,
386                                  depth_ptr, facing,
387                                  &zs_value,
388                                  !simple_shader);
389      /* Late Z write */
390      if (depth_mode & LATE_DEPTH_WRITE) {
391         lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
392      }
393   }
394   else if ((depth_mode & EARLY_DEPTH_TEST) &&
395            (depth_mode & LATE_DEPTH_WRITE))
396   {
397      /* Need to apply a reduced mask to the depth write.  Reload the
398       * depth value, update from zs_value with the new mask value and
399       * write that out.
400       */
401      lp_build_deferred_depth_write(gallivm,
402                                    type,
403                                    zs_format_desc,
404                                    &mask,
405                                    depth_ptr,
406                                    zs_value);
407   }
408
409
410   /* Color write  */
411   for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
412   {
413      if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
414          shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
415      {
416         unsigned cbuf = shader->info.base.output_semantic_index[attrib];
417         for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
418            if(outputs[attrib][chan]) {
419               /* XXX: just initialize outputs to point at colors[] and
420                * skip this.
421                */
422               LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
423               lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
424               LLVMBuildStore(builder, out, color[cbuf][chan]);
425            }
426         }
427      }
428   }
429
430   if (counter)
431      lp_build_occlusion_count(gallivm, type,
432                               lp_build_mask_value(&mask), counter);
433
434   *pmask = lp_build_mask_end(&mask);
435}
436
437
438/**
439 * Generate the fragment shader, depth/stencil test, and alpha tests.
440 */
441static void
442generate_fs_loop(struct gallivm_state *gallivm,
443                 struct lp_fragment_shader *shader,
444                 const struct lp_fragment_shader_variant_key *key,
445                 LLVMBuilderRef builder,
446                 struct lp_type type,
447                 LLVMValueRef context_ptr,
448                 LLVMValueRef num_loop,
449                 struct lp_build_interp_soa_context *interp,
450                 struct lp_build_sampler_soa *sampler,
451                 LLVMValueRef mask_store,
452                 LLVMValueRef (*out_color)[4],
453                 LLVMValueRef depth_ptr,
454                 unsigned depth_bits,
455                 LLVMValueRef facing,
456                 LLVMValueRef counter)
457{
458   const struct util_format_description *zs_format_desc = NULL;
459   const struct tgsi_token *tokens = shader->base.tokens;
460   LLVMTypeRef vec_type;
461   LLVMValueRef mask_ptr, mask_val;
462   LLVMValueRef consts_ptr;
463   LLVMValueRef z;
464   LLVMValueRef zs_value = NULL;
465   LLVMValueRef stencil_refs[2];
466   LLVMValueRef depth_ptr_i;
467   LLVMValueRef depth_offset;
468   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
469   struct lp_build_for_loop_state loop_state;
470   struct lp_build_mask_context mask;
471   boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
472                            shader->info.base.num_inputs < 3 &&
473                            shader->info.base.num_instructions < 8);
474   unsigned attrib;
475   unsigned chan;
476   unsigned cbuf;
477   unsigned depth_mode;
478
479   struct lp_bld_tgsi_system_values system_values;
480
481   memset(&system_values, 0, sizeof(system_values));
482
483   if (key->depth.enabled ||
484       key->stencil[0].enabled ||
485       key->stencil[1].enabled) {
486
487      zs_format_desc = util_format_description(key->zsbuf_format);
488      assert(zs_format_desc);
489
490      if (!shader->info.base.writes_z) {
491         if (key->alpha.enabled || shader->info.base.uses_kill)
492            /* With alpha test and kill, can do the depth test early
493             * and hopefully eliminate some quads.  But need to do a
494             * special deferred depth write once the final mask value
495             * is known.
496             */
497            depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
498         else
499            depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
500      }
501      else {
502         depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
503      }
504
505      if (!(key->depth.enabled && key->depth.writemask) &&
506          !(key->stencil[0].enabled && key->stencil[0].writemask))
507         depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
508   }
509   else {
510      depth_mode = 0;
511   }
512
513
514   stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
515   stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
516
517   vec_type = lp_build_vec_type(gallivm, type);
518
519   consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
520
521   lp_build_for_loop_begin(&loop_state, gallivm,
522                           lp_build_const_int32(gallivm, 0),
523                           LLVMIntULT,
524                           num_loop,
525                           lp_build_const_int32(gallivm, 1));
526
527   mask_ptr = LLVMBuildGEP(builder, mask_store,
528                           &loop_state.counter, 1, "mask_ptr");
529   mask_val = LLVMBuildLoad(builder, mask_ptr, "");
530
531   depth_offset = LLVMBuildMul(builder, loop_state.counter,
532                               lp_build_const_int32(gallivm, depth_bits * type.length),
533                               "");
534
535   depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
536
537   memset(outputs, 0, sizeof outputs);
538
539   for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
540      for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
541         out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
542                                                       lp_build_vec_type(gallivm,
543                                                                         type),
544                                                       num_loop, "color");
545      }
546   }
547
548
549
550   /* 'mask' will control execution based on quad's pixel alive/killed state */
551   lp_build_mask_begin(&mask, gallivm, type, mask_val);
552
553   if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
554      lp_build_mask_check(&mask);
555
556   lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
557   z = interp->pos[2];
558
559   if (depth_mode & EARLY_DEPTH_TEST) {
560      lp_build_depth_stencil_test(gallivm,
561                                  &key->depth,
562                                  key->stencil,
563                                  type,
564                                  zs_format_desc,
565                                  &mask,
566                                  stencil_refs,
567                                  z,
568                                  depth_ptr_i, facing,
569                                  &zs_value,
570                                  !simple_shader);
571
572      if (depth_mode & EARLY_DEPTH_WRITE) {
573         lp_build_depth_write(builder, zs_format_desc, depth_ptr_i, zs_value);
574      }
575   }
576
577   lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
578
579   /* Build the actual shader */
580   lp_build_tgsi_soa(gallivm, tokens, type, &mask,
581                     consts_ptr, &system_values,
582                     interp->pos, interp->inputs,
583                     outputs, sampler, &shader->info.base);
584
585   /* Alpha test */
586   if (key->alpha.enabled) {
587      int color0 = find_output_by_semantic(&shader->info.base,
588                                           TGSI_SEMANTIC_COLOR,
589                                           0);
590
591      if (color0 != -1 && outputs[color0][3]) {
592         const struct util_format_description *cbuf_format_desc;
593         LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
594         LLVMValueRef alpha_ref_value;
595
596         alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
597         alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
598
599         cbuf_format_desc = util_format_description(key->cbuf_format[0]);
600
601         lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
602                             &mask, alpha, alpha_ref_value,
603                             (depth_mode & LATE_DEPTH_TEST) != 0);
604      }
605   }
606
607   /* Late Z test */
608   if (depth_mode & LATE_DEPTH_TEST) {
609      int pos0 = find_output_by_semantic(&shader->info.base,
610                                         TGSI_SEMANTIC_POSITION,
611                                         0);
612
613      if (pos0 != -1 && outputs[pos0][2]) {
614         z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
615      }
616
617      lp_build_depth_stencil_test(gallivm,
618                                  &key->depth,
619                                  key->stencil,
620                                  type,
621                                  zs_format_desc,
622                                  &mask,
623                                  stencil_refs,
624                                  z,
625                                  depth_ptr_i, facing,
626                                  &zs_value,
627                                  !simple_shader);
628      /* Late Z write */
629      if (depth_mode & LATE_DEPTH_WRITE) {
630         lp_build_depth_write(builder, zs_format_desc, depth_ptr_i, zs_value);
631      }
632   }
633   else if ((depth_mode & EARLY_DEPTH_TEST) &&
634            (depth_mode & LATE_DEPTH_WRITE))
635   {
636      /* Need to apply a reduced mask to the depth write.  Reload the
637       * depth value, update from zs_value with the new mask value and
638       * write that out.
639       */
640      lp_build_deferred_depth_write(gallivm,
641                                    type,
642                                    zs_format_desc,
643                                    &mask,
644                                    depth_ptr_i,
645                                    zs_value);
646   }
647
648
649   /* Color write  */
650   for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
651   {
652      if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
653          shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
654      {
655         unsigned cbuf = shader->info.base.output_semantic_index[attrib];
656         for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
657            if(outputs[attrib][chan]) {
658               /* XXX: just initialize outputs to point at colors[] and
659                * skip this.
660                */
661               LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
662               LLVMValueRef color_ptr;
663               color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
664                                        &loop_state.counter, 1, "");
665               lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
666               LLVMBuildStore(builder, out, color_ptr);
667            }
668         }
669      }
670   }
671
672   if (key->occlusion_count) {
673      lp_build_name(counter, "counter");
674      lp_build_occlusion_count(gallivm, type,
675                               lp_build_mask_value(&mask), counter);
676   }
677
678   mask_val = lp_build_mask_end(&mask);
679   LLVMBuildStore(builder, mask_val, mask_ptr);
680   lp_build_for_loop_end(&loop_state);
681}
682
683
684/**
685 * Generate color blending and color output.
686 * \param rt  the render target index (to index blend, colormask state)
687 * \param type  the pixel color type
688 * \param context_ptr  pointer to the runtime JIT context
689 * \param mask  execution mask (active fragment/pixel mask)
690 * \param src  colors from the fragment shader
691 * \param dst_ptr  the destination color buffer pointer
692 */
693static void
694generate_blend(struct gallivm_state *gallivm,
695               const struct pipe_blend_state *blend,
696               unsigned rt,
697               LLVMBuilderRef builder,
698               struct lp_type type,
699               LLVMValueRef context_ptr,
700               LLVMValueRef mask,
701               LLVMValueRef *src,
702               LLVMValueRef dst_ptr,
703               boolean do_branch)
704{
705   struct lp_build_context bld;
706   struct lp_build_mask_context mask_ctx;
707   LLVMTypeRef vec_type;
708   LLVMValueRef const_ptr;
709   LLVMValueRef con[4];
710   LLVMValueRef dst[4];
711   LLVMValueRef res[4];
712   unsigned chan;
713
714   lp_build_context_init(&bld, gallivm, type);
715
716   lp_build_mask_begin(&mask_ctx, gallivm, type, mask);
717   if (do_branch)
718      lp_build_mask_check(&mask_ctx);
719
720   vec_type = lp_build_vec_type(gallivm, type);
721
722   const_ptr = lp_jit_context_blend_color(gallivm, context_ptr);
723   const_ptr = LLVMBuildBitCast(builder, const_ptr,
724                                LLVMPointerType(vec_type, 0), "");
725
726   /* load constant blend color and colors from the dest color buffer */
727   for(chan = 0; chan < 4; ++chan) {
728      LLVMValueRef index = lp_build_const_int32(gallivm, chan);
729      con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
730
731      dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
732
733      lp_build_name(con[chan], "con.%c", "rgba"[chan]);
734      lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
735   }
736
737   /* do blend */
738   lp_build_blend_soa(gallivm, blend, type, rt, src, dst, con, res);
739
740   /* store results to color buffer */
741   for(chan = 0; chan < 4; ++chan) {
742      if(blend->rt[rt].colormask & (1 << chan)) {
743         LLVMValueRef index = lp_build_const_int32(gallivm, chan);
744         lp_build_name(res[chan], "res.%c", "rgba"[chan]);
745         res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
746         LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
747      }
748   }
749
750   lp_build_mask_end(&mask_ctx);
751}
752
753
754/**
755 * Generate the runtime callable function for the whole fragment pipeline.
756 * Note that the function which we generate operates on a block of 16
757 * pixels at at time.  The block contains 2x2 quads.  Each quad contains
758 * 2x2 pixels.
759 */
760static void
761generate_fragment(struct llvmpipe_context *lp,
762                  struct lp_fragment_shader *shader,
763                  struct lp_fragment_shader_variant *variant,
764                  unsigned partial_mask)
765{
766   struct gallivm_state *gallivm = variant->gallivm;
767   const struct lp_fragment_shader_variant_key *key = &variant->key;
768   struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
769   char func_name[256];
770   struct lp_type fs_type;
771   struct lp_type blend_type;
772   LLVMTypeRef fs_elem_type;
773   LLVMTypeRef blend_vec_type;
774   LLVMTypeRef arg_types[11];
775   LLVMTypeRef func_type;
776   LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
777   LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
778   LLVMValueRef context_ptr;
779   LLVMValueRef x;
780   LLVMValueRef y;
781   LLVMValueRef a0_ptr;
782   LLVMValueRef dadx_ptr;
783   LLVMValueRef dady_ptr;
784   LLVMValueRef color_ptr_ptr;
785   LLVMValueRef depth_ptr;
786   LLVMValueRef mask_input;
787   LLVMValueRef counter = NULL;
788   LLVMBasicBlockRef block;
789   LLVMBuilderRef builder;
790   struct lp_build_sampler_soa *sampler;
791   struct lp_build_interp_soa_context interp;
792   LLVMValueRef fs_mask[16 / 4];
793   LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
794   LLVMValueRef blend_mask;
795   LLVMValueRef function;
796   LLVMValueRef facing;
797   const struct util_format_description *zs_format_desc;
798   unsigned num_fs;
799   unsigned i;
800   unsigned chan;
801   unsigned cbuf;
802   boolean cbuf0_write_all;
803   boolean try_loop = TRUE;
804
805   assert(lp_native_vector_width / 32 >= 4);
806
807   /* Adjust color input interpolation according to flatshade state:
808    */
809   memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
810   for (i = 0; i < shader->info.base.num_inputs; i++) {
811      if (inputs[i].interp == LP_INTERP_COLOR) {
812	 if (key->flatshade)
813	    inputs[i].interp = LP_INTERP_CONSTANT;
814	 else
815	    inputs[i].interp = LP_INTERP_PERSPECTIVE;
816      }
817   }
818
819   /* check if writes to cbuf[0] are to be copied to all cbufs */
820   cbuf0_write_all = FALSE;
821   for (i = 0;i < shader->info.base.num_properties; i++) {
822      if (shader->info.base.properties[i].name ==
823          TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
824         cbuf0_write_all = TRUE;
825         break;
826      }
827   }
828
829   /* TODO: actually pick these based on the fs and color buffer
830    * characteristics. */
831
832   memset(&fs_type, 0, sizeof fs_type);
833   fs_type.floating = TRUE;      /* floating point values */
834   fs_type.sign = TRUE;          /* values are signed */
835   fs_type.norm = FALSE;         /* values are not limited to [0,1] or [-1,1] */
836   fs_type.width = 32;           /* 32-bit float */
837   fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
838   num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
839
840   memset(&blend_type, 0, sizeof blend_type);
841   blend_type.floating = FALSE; /* values are integers */
842   blend_type.sign = FALSE;     /* values are unsigned */
843   blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
844   blend_type.width = 8;        /* 8-bit ubyte values */
845   blend_type.length = 16;      /* 16 elements per vector */
846
847   /*
848    * Generate the function prototype. Any change here must be reflected in
849    * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
850    */
851
852   fs_elem_type = lp_build_elem_type(gallivm, fs_type);
853
854   blend_vec_type = lp_build_vec_type(gallivm, blend_type);
855
856   util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
857		 shader->no, variant->no, partial_mask ? "partial" : "whole");
858
859   arg_types[0] = variant->jit_context_ptr_type;       /* context */
860   arg_types[1] = int32_type;                          /* x */
861   arg_types[2] = int32_type;                          /* y */
862   arg_types[3] = int32_type;                          /* facing */
863   arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
864   arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
865   arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
866   arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
867   arg_types[8] = LLVMPointerType(int8_type, 0);       /* depth */
868   arg_types[9] = int32_type;                          /* mask_input */
869   arg_types[10] = LLVMPointerType(int32_type, 0);     /* counter */
870
871   func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
872                                arg_types, Elements(arg_types), 0);
873
874   function = LLVMAddFunction(gallivm->module, func_name, func_type);
875   LLVMSetFunctionCallConv(function, LLVMCCallConv);
876
877   variant->function[partial_mask] = function;
878
879   /* XXX: need to propagate noalias down into color param now we are
880    * passing a pointer-to-pointer?
881    */
882   for(i = 0; i < Elements(arg_types); ++i)
883      if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
884         LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
885
886   context_ptr  = LLVMGetParam(function, 0);
887   x            = LLVMGetParam(function, 1);
888   y            = LLVMGetParam(function, 2);
889   facing       = LLVMGetParam(function, 3);
890   a0_ptr       = LLVMGetParam(function, 4);
891   dadx_ptr     = LLVMGetParam(function, 5);
892   dady_ptr     = LLVMGetParam(function, 6);
893   color_ptr_ptr = LLVMGetParam(function, 7);
894   depth_ptr    = LLVMGetParam(function, 8);
895   mask_input   = LLVMGetParam(function, 9);
896
897   lp_build_name(context_ptr, "context");
898   lp_build_name(x, "x");
899   lp_build_name(y, "y");
900   lp_build_name(a0_ptr, "a0");
901   lp_build_name(dadx_ptr, "dadx");
902   lp_build_name(dady_ptr, "dady");
903   lp_build_name(color_ptr_ptr, "color_ptr_ptr");
904   lp_build_name(depth_ptr, "depth");
905   lp_build_name(mask_input, "mask_input");
906
907   if (key->occlusion_count) {
908      counter = LLVMGetParam(function, 10);
909      lp_build_name(counter, "counter");
910   }
911
912   /*
913    * Function body
914    */
915
916   block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
917   builder = gallivm->builder;
918   assert(builder);
919   LLVMPositionBuilderAtEnd(builder, block);
920
921   /* code generated texture sampling */
922   sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
923
924   zs_format_desc = util_format_description(key->zsbuf_format);
925
926   if (!try_loop) {
927      /*
928       * The shader input interpolation info is not explicitely baked in the
929       * shader key, but everything it derives from (TGSI, and flatshade) is
930       * already included in the shader key.
931       */
932      lp_build_interp_soa_init(&interp,
933                               gallivm,
934                               shader->info.base.num_inputs,
935                               inputs,
936                               builder, fs_type,
937                               FALSE,
938                               a0_ptr, dadx_ptr, dady_ptr,
939                               x, y);
940
941      /* loop over quads in the block */
942      for(i = 0; i < num_fs; ++i) {
943         LLVMValueRef depth_offset = LLVMConstInt(int32_type,
944                                                  i*fs_type.length*zs_format_desc->block.bits/8,
945                                                  0);
946         LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
947         LLVMValueRef depth_ptr_i;
948
949         depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
950
951         generate_fs(gallivm,
952                     shader, key,
953                     builder,
954                     fs_type,
955                     context_ptr,
956                     i,
957                     &interp,
958                     sampler,
959                     &fs_mask[i], /* output */
960                     out_color,
961                     depth_ptr_i,
962                     facing,
963                     partial_mask,
964                     mask_input,
965                     counter);
966
967         for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
968            for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
969               fs_out_color[cbuf][chan][i] =
970                  out_color[cbuf * !cbuf0_write_all][chan];
971      }
972   }
973   else {
974      unsigned depth_bits = zs_format_desc->block.bits/8;
975      LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
976      LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
977      LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
978                                                      num_loop, "mask_store");
979      LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
980
981      /*
982       * The shader input interpolation info is not explicitely baked in the
983       * shader key, but everything it derives from (TGSI, and flatshade) is
984       * already included in the shader key.
985       */
986      lp_build_interp_soa_init(&interp,
987                               gallivm,
988                               shader->info.base.num_inputs,
989                               inputs,
990                               builder, fs_type,
991                               TRUE,
992                               a0_ptr, dadx_ptr, dady_ptr,
993                               x, y);
994
995      for (i = 0; i < num_fs; i++) {
996         LLVMValueRef mask;
997         LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
998         LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
999                                              &indexi, 1, "mask_ptr");
1000
1001         if (partial_mask) {
1002            mask = generate_quad_mask(gallivm, fs_type,
1003                                      i*fs_type.length/4, mask_input);
1004         }
1005         else {
1006            mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
1007         }
1008         LLVMBuildStore(builder, mask, mask_ptr);
1009      }
1010
1011      generate_fs_loop(gallivm,
1012                       shader, key,
1013                       builder,
1014                       fs_type,
1015                       context_ptr,
1016                       num_loop,
1017                       &interp,
1018                       sampler,
1019                       mask_store, /* output */
1020                       color_store,
1021                       depth_ptr,
1022                       depth_bits,
1023                       facing,
1024                       counter);
1025
1026      for (i = 0; i < num_fs; i++) {
1027         LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1028         LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
1029                                         &indexi, 1, "");
1030         fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
1031         /* This is fucked up need to reorganize things */
1032         for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1033            for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1034               ptr = LLVMBuildGEP(builder,
1035                                  color_store[cbuf * !cbuf0_write_all][chan],
1036                                  &indexi, 1, "");
1037               fs_out_color[cbuf][chan][i] = ptr;
1038            }
1039         }
1040      }
1041   }
1042
1043   sampler->destroy(sampler);
1044
1045   /* Loop over color outputs / color buffers to do blending.
1046    */
1047   for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1048      LLVMValueRef color_ptr;
1049      LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
1050      LLVMValueRef blend_in_color[TGSI_NUM_CHANNELS];
1051      unsigned rt;
1052
1053      /*
1054       * Convert the fs's output color and mask to fit to the blending type.
1055       */
1056      for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1057         LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
1058
1059         for (i = 0; i < num_fs; i++) {
1060            fs_color_vals[i] =
1061               LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
1062         }
1063
1064         lp_build_conv(gallivm, fs_type, blend_type,
1065                       fs_color_vals,
1066                       num_fs,
1067                       &blend_in_color[chan], 1);
1068
1069         lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
1070      }
1071
1072      if (partial_mask || !variant->opaque) {
1073         lp_build_conv_mask(variant->gallivm, fs_type, blend_type,
1074                            fs_mask, num_fs,
1075                            &blend_mask, 1);
1076      } else {
1077         blend_mask = lp_build_const_int_vec(variant->gallivm, blend_type, ~0);
1078      }
1079
1080      color_ptr = LLVMBuildLoad(builder,
1081                                LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
1082                                "");
1083      lp_build_name(color_ptr, "color_ptr%d", cbuf);
1084
1085      /* which blend/colormask state to use */
1086      rt = key->blend.independent_blend_enable ? cbuf : 0;
1087
1088      /*
1089       * Blending.
1090       */
1091      {
1092         /* Could the 4x4 have been killed?
1093          */
1094         boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
1095                              !key->alpha.enabled &&
1096                              !shader->info.base.uses_kill);
1097
1098         generate_blend(variant->gallivm,
1099                        &key->blend,
1100                        rt,
1101                        builder,
1102                        blend_type,
1103                        context_ptr,
1104                        blend_mask,
1105                        blend_in_color,
1106                        color_ptr,
1107                        do_branch);
1108      }
1109   }
1110
1111   LLVMBuildRetVoid(builder);
1112
1113   gallivm_verify_function(gallivm, function);
1114
1115   variant->nr_instrs += lp_build_count_instructions(function);
1116}
1117
1118
1119static void
1120dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
1121{
1122   unsigned i;
1123
1124   debug_printf("fs variant %p:\n", (void *) key);
1125
1126   if (key->flatshade) {
1127      debug_printf("flatshade = 1\n");
1128   }
1129   for (i = 0; i < key->nr_cbufs; ++i) {
1130      debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
1131   }
1132   if (key->depth.enabled) {
1133      debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
1134      debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
1135      debug_printf("depth.writemask = %u\n", key->depth.writemask);
1136   }
1137
1138   for (i = 0; i < 2; ++i) {
1139      if (key->stencil[i].enabled) {
1140         debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
1141         debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
1142         debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
1143         debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
1144         debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
1145         debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
1146      }
1147   }
1148
1149   if (key->alpha.enabled) {
1150      debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
1151   }
1152
1153   if (key->occlusion_count) {
1154      debug_printf("occlusion_count = 1\n");
1155   }
1156
1157   if (key->blend.logicop_enable) {
1158      debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
1159   }
1160   else if (key->blend.rt[0].blend_enable) {
1161      debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
1162      debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
1163      debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
1164      debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
1165      debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
1166      debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
1167   }
1168   debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
1169   for (i = 0; i < key->nr_samplers; ++i) {
1170      debug_printf("sampler[%u] = \n", i);
1171      debug_printf("  .format = %s\n",
1172                   util_format_name(key->sampler[i].format));
1173      debug_printf("  .target = %s\n",
1174                   util_dump_tex_target(key->sampler[i].target, TRUE));
1175      debug_printf("  .pot = %u %u %u\n",
1176                   key->sampler[i].pot_width,
1177                   key->sampler[i].pot_height,
1178                   key->sampler[i].pot_depth);
1179      debug_printf("  .wrap = %s %s %s\n",
1180                   util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
1181                   util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
1182                   util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
1183      debug_printf("  .min_img_filter = %s\n",
1184                   util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
1185      debug_printf("  .min_mip_filter = %s\n",
1186                   util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
1187      debug_printf("  .mag_img_filter = %s\n",
1188                   util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
1189      if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
1190         debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
1191      debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
1192      debug_printf("  .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
1193      debug_printf("  .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
1194      debug_printf("  .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
1195      debug_printf("  .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
1196   }
1197}
1198
1199
1200void
1201lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
1202{
1203   debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
1204                variant->shader->no, variant->no);
1205   tgsi_dump(variant->shader->base.tokens, 0);
1206   dump_fs_variant_key(&variant->key);
1207   debug_printf("variant->opaque = %u\n", variant->opaque);
1208   debug_printf("\n");
1209}
1210
1211
1212/**
1213 * Generate a new fragment shader variant from the shader code and
1214 * other state indicated by the key.
1215 */
1216static struct lp_fragment_shader_variant *
1217generate_variant(struct llvmpipe_context *lp,
1218                 struct lp_fragment_shader *shader,
1219                 const struct lp_fragment_shader_variant_key *key)
1220{
1221   struct lp_fragment_shader_variant *variant;
1222   const struct util_format_description *cbuf0_format_desc;
1223   boolean fullcolormask;
1224
1225   variant = CALLOC_STRUCT(lp_fragment_shader_variant);
1226   if(!variant)
1227      return NULL;
1228
1229   variant->gallivm = gallivm_create();
1230   if (!variant->gallivm) {
1231      FREE(variant);
1232      return NULL;
1233   }
1234
1235   variant->shader = shader;
1236   variant->list_item_global.base = variant;
1237   variant->list_item_local.base = variant;
1238   variant->no = shader->variants_created++;
1239
1240   memcpy(&variant->key, key, shader->variant_key_size);
1241
1242   /*
1243    * Determine whether we are touching all channels in the color buffer.
1244    */
1245   fullcolormask = FALSE;
1246   if (key->nr_cbufs == 1) {
1247      cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
1248      fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
1249   }
1250
1251   variant->opaque =
1252         !key->blend.logicop_enable &&
1253         !key->blend.rt[0].blend_enable &&
1254         fullcolormask &&
1255         !key->stencil[0].enabled &&
1256         !key->alpha.enabled &&
1257         !key->depth.enabled &&
1258         !shader->info.base.uses_kill
1259         ? TRUE : FALSE;
1260
1261
1262   if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
1263      lp_debug_fs_variant(variant);
1264   }
1265
1266   lp_jit_init_types(variant);
1267
1268   if (variant->jit_function[RAST_EDGE_TEST] == NULL)
1269      generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
1270
1271   if (variant->jit_function[RAST_WHOLE] == NULL) {
1272      if (variant->opaque) {
1273         /* Specialized shader, which doesn't need to read the color buffer. */
1274         generate_fragment(lp, shader, variant, RAST_WHOLE);
1275      }
1276   }
1277
1278   /*
1279    * Compile everything
1280    */
1281
1282   gallivm_compile_module(variant->gallivm);
1283
1284   if (variant->function[RAST_EDGE_TEST]) {
1285      variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
1286            gallivm_jit_function(variant->gallivm,
1287                                 variant->function[RAST_EDGE_TEST]);
1288   }
1289
1290   if (variant->function[RAST_WHOLE]) {
1291         variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
1292               gallivm_jit_function(variant->gallivm,
1293                                    variant->function[RAST_WHOLE]);
1294   } else if (!variant->jit_function[RAST_WHOLE]) {
1295      variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
1296   }
1297
1298   return variant;
1299}
1300
1301
1302static void *
1303llvmpipe_create_fs_state(struct pipe_context *pipe,
1304                         const struct pipe_shader_state *templ)
1305{
1306   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1307   struct lp_fragment_shader *shader;
1308   int nr_samplers;
1309   int i;
1310
1311   shader = CALLOC_STRUCT(lp_fragment_shader);
1312   if (!shader)
1313      return NULL;
1314
1315   shader->no = fs_no++;
1316   make_empty_list(&shader->variants);
1317
1318   /* get/save the summary info for this shader */
1319   lp_build_tgsi_info(templ->tokens, &shader->info);
1320
1321   /* we need to keep a local copy of the tokens */
1322   shader->base.tokens = tgsi_dup_tokens(templ->tokens);
1323
1324   shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
1325   if (shader->draw_data == NULL) {
1326      FREE((void *) shader->base.tokens);
1327      FREE(shader);
1328      return NULL;
1329   }
1330
1331   nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1332
1333   shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
1334				     sampler[nr_samplers]);
1335
1336   for (i = 0; i < shader->info.base.num_inputs; i++) {
1337      shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
1338      shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
1339
1340      switch (shader->info.base.input_interpolate[i]) {
1341      case TGSI_INTERPOLATE_CONSTANT:
1342	 shader->inputs[i].interp = LP_INTERP_CONSTANT;
1343	 break;
1344      case TGSI_INTERPOLATE_LINEAR:
1345	 shader->inputs[i].interp = LP_INTERP_LINEAR;
1346	 break;
1347      case TGSI_INTERPOLATE_PERSPECTIVE:
1348	 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
1349	 break;
1350      case TGSI_INTERPOLATE_COLOR:
1351	 shader->inputs[i].interp = LP_INTERP_COLOR;
1352	 break;
1353      default:
1354	 assert(0);
1355	 break;
1356      }
1357
1358      switch (shader->info.base.input_semantic_name[i]) {
1359      case TGSI_SEMANTIC_FACE:
1360	 shader->inputs[i].interp = LP_INTERP_FACING;
1361	 break;
1362      case TGSI_SEMANTIC_POSITION:
1363	 /* Position was already emitted above
1364	  */
1365	 shader->inputs[i].interp = LP_INTERP_POSITION;
1366	 shader->inputs[i].src_index = 0;
1367	 continue;
1368      }
1369
1370      shader->inputs[i].src_index = i+1;
1371   }
1372
1373   if (LP_DEBUG & DEBUG_TGSI) {
1374      unsigned attrib;
1375      debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
1376                   shader->no, (void *) shader);
1377      tgsi_dump(templ->tokens, 0);
1378      debug_printf("usage masks:\n");
1379      for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
1380         unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
1381         debug_printf("  IN[%u].%s%s%s%s\n",
1382                      attrib,
1383                      usage_mask & TGSI_WRITEMASK_X ? "x" : "",
1384                      usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
1385                      usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
1386                      usage_mask & TGSI_WRITEMASK_W ? "w" : "");
1387      }
1388      debug_printf("\n");
1389   }
1390
1391   return shader;
1392}
1393
1394
1395static void
1396llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1397{
1398   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1399
1400   if (llvmpipe->fs == fs)
1401      return;
1402
1403   draw_flush(llvmpipe->draw);
1404
1405   llvmpipe->fs = (struct lp_fragment_shader *) fs;
1406
1407   draw_bind_fragment_shader(llvmpipe->draw,
1408                             (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
1409
1410   llvmpipe->dirty |= LP_NEW_FS;
1411}
1412
1413
1414/**
1415 * Remove shader variant from two lists: the shader's variant list
1416 * and the context's variant list.
1417 */
1418void
1419llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
1420                               struct lp_fragment_shader_variant *variant)
1421{
1422   unsigned i;
1423
1424   if (gallivm_debug & GALLIVM_DEBUG_IR) {
1425      debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
1426                   " #%u v total cached #%u\n",
1427                   variant->shader->no,
1428                   variant->no,
1429                   variant->shader->variants_created,
1430                   variant->shader->variants_cached,
1431                   lp->nr_fs_variants);
1432   }
1433
1434   /* free all the variant's JIT'd functions */
1435   for (i = 0; i < Elements(variant->function); i++) {
1436      if (variant->function[i]) {
1437         gallivm_free_function(variant->gallivm,
1438                               variant->function[i],
1439                               variant->jit_function[i]);
1440      }
1441   }
1442
1443   gallivm_destroy(variant->gallivm);
1444
1445   /* remove from shader's list */
1446   remove_from_list(&variant->list_item_local);
1447   variant->shader->variants_cached--;
1448
1449   /* remove from context's list */
1450   remove_from_list(&variant->list_item_global);
1451   lp->nr_fs_variants--;
1452   lp->nr_fs_instrs -= variant->nr_instrs;
1453
1454   FREE(variant);
1455}
1456
1457
1458static void
1459llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1460{
1461   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1462   struct lp_fragment_shader *shader = fs;
1463   struct lp_fs_variant_list_item *li;
1464
1465   assert(fs != llvmpipe->fs);
1466
1467   /*
1468    * XXX: we need to flush the context until we have some sort of reference
1469    * counting in fragment shaders as they may still be binned
1470    * Flushing alone might not sufficient we need to wait on it too.
1471    */
1472   llvmpipe_finish(pipe, __FUNCTION__);
1473
1474   /* Delete all the variants */
1475   li = first_elem(&shader->variants);
1476   while(!at_end(&shader->variants, li)) {
1477      struct lp_fs_variant_list_item *next = next_elem(li);
1478      llvmpipe_remove_shader_variant(llvmpipe, li->base);
1479      li = next;
1480   }
1481
1482   /* Delete draw module's data */
1483   draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
1484
1485   assert(shader->variants_cached == 0);
1486   FREE((void *) shader->base.tokens);
1487   FREE(shader);
1488}
1489
1490
1491
1492static void
1493llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1494                             uint shader, uint index,
1495                             struct pipe_constant_buffer *cb)
1496{
1497   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1498   struct pipe_resource *constants = cb ? cb->buffer : NULL;
1499   unsigned size;
1500   const void *data;
1501
1502   if (cb && cb->user_buffer) {
1503      constants = llvmpipe_user_buffer_create(pipe->screen,
1504                                              (void *) cb->user_buffer,
1505                                              cb->buffer_size,
1506                                              PIPE_BIND_CONSTANT_BUFFER);
1507   }
1508
1509   size = constants ? constants->width0 : 0;
1510   data = constants ? llvmpipe_resource_data(constants) : NULL;
1511
1512   assert(shader < PIPE_SHADER_TYPES);
1513   assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1514
1515   if(llvmpipe->constants[shader][index] == constants)
1516      return;
1517
1518   draw_flush(llvmpipe->draw);
1519
1520   /* note: reference counting */
1521   pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1522
1523   if(shader == PIPE_SHADER_VERTEX ||
1524      shader == PIPE_SHADER_GEOMETRY) {
1525      draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1526                                      index, data, size);
1527   }
1528
1529   llvmpipe->dirty |= LP_NEW_CONSTANTS;
1530
1531   if (cb && cb->user_buffer) {
1532      pipe_resource_reference(&constants, NULL);
1533   }
1534}
1535
1536
1537/**
1538 * Return the blend factor equivalent to a destination alpha of one.
1539 */
1540static INLINE unsigned
1541force_dst_alpha_one(unsigned factor)
1542{
1543   switch(factor) {
1544   case PIPE_BLENDFACTOR_DST_ALPHA:
1545      return PIPE_BLENDFACTOR_ONE;
1546   case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1547      return PIPE_BLENDFACTOR_ZERO;
1548   case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1549      return PIPE_BLENDFACTOR_ZERO;
1550   }
1551
1552   return factor;
1553}
1554
1555
1556/**
1557 * We need to generate several variants of the fragment pipeline to match
1558 * all the combinations of the contributing state atoms.
1559 *
1560 * TODO: there is actually no reason to tie this to context state -- the
1561 * generated code could be cached globally in the screen.
1562 */
1563static void
1564make_variant_key(struct llvmpipe_context *lp,
1565                 struct lp_fragment_shader *shader,
1566                 struct lp_fragment_shader_variant_key *key)
1567{
1568   unsigned i;
1569
1570   memset(key, 0, shader->variant_key_size);
1571
1572   if (lp->framebuffer.zsbuf) {
1573      if (lp->depth_stencil->depth.enabled) {
1574         key->zsbuf_format = lp->framebuffer.zsbuf->format;
1575         memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1576      }
1577      if (lp->depth_stencil->stencil[0].enabled) {
1578         key->zsbuf_format = lp->framebuffer.zsbuf->format;
1579         memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1580      }
1581   }
1582
1583   key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1584   if(key->alpha.enabled)
1585      key->alpha.func = lp->depth_stencil->alpha.func;
1586   /* alpha.ref_value is passed in jit_context */
1587
1588   key->flatshade = lp->rasterizer->flatshade;
1589   if (lp->active_query_count) {
1590      key->occlusion_count = TRUE;
1591   }
1592
1593   if (lp->framebuffer.nr_cbufs) {
1594      memcpy(&key->blend, lp->blend, sizeof key->blend);
1595   }
1596
1597   key->nr_cbufs = lp->framebuffer.nr_cbufs;
1598   for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1599      enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1600      struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1601      const struct util_format_description *format_desc;
1602
1603      key->cbuf_format[i] = format;
1604
1605      format_desc = util_format_description(format);
1606      assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1607             format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1608
1609      blend_rt->colormask = lp->blend->rt[i].colormask;
1610
1611      /*
1612       * Mask out color channels not present in the color buffer.
1613       */
1614      blend_rt->colormask &= util_format_colormask(format_desc);
1615
1616      /*
1617       * Our swizzled render tiles always have an alpha channel, but the linear
1618       * render target format often does not, so force here the dst alpha to be
1619       * one.
1620       *
1621       * This is not a mere optimization. Wrong results will be produced if the
1622       * dst alpha is used, the dst format does not have alpha, and the previous
1623       * rendering was not flushed from the swizzled to linear buffer. For
1624       * example, NonPowTwo DCT.
1625       *
1626       * TODO: This should be generalized to all channels for better
1627       * performance, but only alpha causes correctness issues.
1628       *
1629       * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1630       */
1631      if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
1632	  format_desc->swizzle[3] == format_desc->swizzle[0]) {
1633         blend_rt->rgb_src_factor   = force_dst_alpha_one(blend_rt->rgb_src_factor);
1634         blend_rt->rgb_dst_factor   = force_dst_alpha_one(blend_rt->rgb_dst_factor);
1635         blend_rt->alpha_func       = blend_rt->rgb_func;
1636         blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
1637         blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
1638      }
1639   }
1640
1641   /* This value will be the same for all the variants of a given shader:
1642    */
1643   key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1644
1645   for(i = 0; i < key->nr_samplers; ++i) {
1646      if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1647         lp_sampler_static_state(&key->sampler[i],
1648				 lp->sampler_views[PIPE_SHADER_FRAGMENT][i],
1649				 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
1650      }
1651   }
1652}
1653
1654
1655
1656/**
1657 * Update fragment shader state.  This is called just prior to drawing
1658 * something when some fragment-related state has changed.
1659 */
1660void
1661llvmpipe_update_fs(struct llvmpipe_context *lp)
1662{
1663   struct lp_fragment_shader *shader = lp->fs;
1664   struct lp_fragment_shader_variant_key key;
1665   struct lp_fragment_shader_variant *variant = NULL;
1666   struct lp_fs_variant_list_item *li;
1667
1668   make_variant_key(lp, shader, &key);
1669
1670   /* Search the variants for one which matches the key */
1671   li = first_elem(&shader->variants);
1672   while(!at_end(&shader->variants, li)) {
1673      if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1674         variant = li->base;
1675         break;
1676      }
1677      li = next_elem(li);
1678   }
1679
1680   if (variant) {
1681      /* Move this variant to the head of the list to implement LRU
1682       * deletion of shader's when we have too many.
1683       */
1684      move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1685   }
1686   else {
1687      /* variant not found, create it now */
1688      int64_t t0, t1, dt;
1689      unsigned i;
1690      unsigned variants_to_cull;
1691
1692      if (0) {
1693         debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
1694                      lp->nr_fs_variants,
1695                      lp->nr_fs_instrs,
1696                      lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
1697      }
1698
1699      /* First, check if we've exceeded the max number of shader variants.
1700       * If so, free 25% of them (the least recently used ones).
1701       */
1702      variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
1703
1704      if (variants_to_cull ||
1705          lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
1706         struct pipe_context *pipe = &lp->pipe;
1707
1708         /*
1709          * XXX: we need to flush the context until we have some sort of
1710          * reference counting in fragment shaders as they may still be binned
1711          * Flushing alone might not be sufficient we need to wait on it too.
1712          */
1713         llvmpipe_finish(pipe, __FUNCTION__);
1714
1715         /*
1716          * We need to re-check lp->nr_fs_variants because an arbitrarliy large
1717          * number of shader variants (potentially all of them) could be
1718          * pending for destruction on flush.
1719          */
1720
1721         for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
1722            struct lp_fs_variant_list_item *item;
1723            if (is_empty_list(&lp->fs_variants_list)) {
1724               break;
1725            }
1726            item = last_elem(&lp->fs_variants_list);
1727            assert(item);
1728            assert(item->base);
1729            llvmpipe_remove_shader_variant(lp, item->base);
1730         }
1731      }
1732
1733      /*
1734       * Generate the new variant.
1735       */
1736      t0 = os_time_get();
1737      variant = generate_variant(lp, shader, &key);
1738      t1 = os_time_get();
1739      dt = t1 - t0;
1740      LP_COUNT_ADD(llvm_compile_time, dt);
1741      LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1742
1743      llvmpipe_variant_count++;
1744
1745      /* Put the new variant into the list */
1746      if (variant) {
1747         insert_at_head(&shader->variants, &variant->list_item_local);
1748         insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1749         lp->nr_fs_variants++;
1750         lp->nr_fs_instrs += variant->nr_instrs;
1751         shader->variants_cached++;
1752      }
1753   }
1754
1755   /* Bind this variant */
1756   lp_setup_set_fs_variant(lp->setup, variant);
1757}
1758
1759
1760
1761
1762
1763
1764
1765void
1766llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1767{
1768   llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1769   llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1770   llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1771
1772   llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1773}
1774