projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: Unswizzled rendering.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_fs.c
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 #include "gallivm/lp_bld_arit.h"
86 #include "gallivm/lp_bld_pack.h"
87 #include "gallivm/lp_bld_format.h"
88 #include "gallivm/lp_bld_quad.h"
89
90 #include "lp_bld_alpha.h"
91 #include "lp_bld_blend.h"
92 #include "lp_bld_depth.h"
93 #include "lp_bld_interp.h"
94 #include "lp_context.h"
95 #include "lp_debug.h"
96 #include "lp_perf.h"
97 #include "lp_setup.h"
98 #include "lp_state.h"
99 #include "lp_tex_sample.h"
100 #include "lp_flush.h"
101 #include "lp_state_fs.h"
102
103
104 /** Fragment shader number (for debugging) */
105 static unsigned fs_no = 0;
106
107
108 /**
109 * Expand the relevant bits of mask_input to a n*4-dword mask for the
110 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
111 * quad mask vector to 0 or ~0.
112 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
113 * quad arguments with fs length 8.
114 *
115 * \param first_quad which quad(s) of the quad group to test, in [0,3]
116 * \param mask_input bitwise mask for the whole 4x4 stamp
117 */
118 static LLVMValueRef
119 generate_quad_mask(struct gallivm_state *gallivm,
120 struct lp_type fs_type,
121 unsigned first_quad,
122 LLVMValueRef mask_input) /* int32 */
123 {
124 LLVMBuilderRef builder = gallivm->builder;
125 struct lp_type mask_type;
126 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
127 LLVMValueRef bits[16];
128 LLVMValueRef mask;
129 int shift, i;
130
131 /*
132 * XXX: We'll need a different path for 16 x u8
133 */
134 assert(fs_type.width == 32);
135 assert(fs_type.length <= Elements(bits));
136 mask_type = lp_int_type(fs_type);
137
138 /*
139 * mask_input >>= (quad * 4)
140 */
141 switch (first_quad) {
142 case 0:
143 shift = 0;
144 break;
145 case 1:
146 assert(fs_type.length == 4);
147 shift = 2;
148 break;
149 case 2:
150 shift = 8;
151 break;
152 case 3:
153 assert(fs_type.length == 4);
154 shift = 10;
155 break;
156 default:
157 assert(0);
158 shift = 0;
159 }
160
161 mask_input = LLVMBuildLShr(builder,
162 mask_input,
163 LLVMConstInt(i32t, shift, 0),
164 "");
165
166 /*
167 * mask = { mask_input & (1 << i), for i in [0,3] }
168 */
169 mask = lp_build_broadcast(gallivm,
170 lp_build_vec_type(gallivm, mask_type),
171 mask_input);
172
173 for (i = 0; i < fs_type.length / 4; i++) {
174 unsigned j = 2 * (i % 2) + (i / 2) * 8;
175 bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
176 bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
177 bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
178 bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
179 }
180 mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
181
182 /*
183 * mask = mask != 0 ? ~0 : 0
184 */
185 mask = lp_build_compare(gallivm,
186 mask_type, PIPE_FUNC_NOTEQUAL,
187 mask,
188 lp_build_const_int_vec(gallivm, mask_type, 0));
189
190 return mask;
191 }
192
193
194 #define EARLY_DEPTH_TEST 0x1
195 #define LATE_DEPTH_TEST 0x2
196 #define EARLY_DEPTH_WRITE 0x4
197 #define LATE_DEPTH_WRITE 0x8
198
199 static int
200 find_output_by_semantic( const struct tgsi_shader_info *info,
201 unsigned semantic,
202 unsigned index )
203 {
204 int i;
205
206 for (i = 0; i < info->num_outputs; i++)
207 if (info->output_semantic_name[i] == semantic &&
208 info->output_semantic_index[i] == index)
209 return i;
210
211 return -1;
212 }
213
214
215 /**
216 * Generate the fragment shader, depth/stencil test, and alpha tests.
217 * \param i which quad in the tile, in range [0,3]
218 * \param partial_mask if 1, do mask_input testing
219 */
220 static void
221 generate_fs(struct gallivm_state *gallivm,
222 struct lp_fragment_shader *shader,
223 const struct lp_fragment_shader_variant_key *key,
224 LLVMBuilderRef builder,
225 struct lp_type type,
226 LLVMValueRef context_ptr,
227 unsigned i,
228 struct lp_build_interp_soa_context *interp,
229 struct lp_build_sampler_soa *sampler,
230 LLVMValueRef *pmask,
231 LLVMValueRef (*color)[4],
232 LLVMValueRef depth_ptr,
233 LLVMValueRef facing,
234 unsigned partial_mask,
235 LLVMValueRef mask_input,
236 LLVMValueRef counter)
237 {
238 const struct util_format_description *zs_format_desc = NULL;
239 const struct tgsi_token *tokens = shader->base.tokens;
240 LLVMTypeRef vec_type;
241 LLVMValueRef consts_ptr;
242 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
243 LLVMValueRef z;
244 LLVMValueRef zs_value = NULL;
245 LLVMValueRef stencil_refs[2];
246 struct lp_build_mask_context mask;
247 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
248 shader->info.base.num_inputs < 3 &&
249 shader->info.base.num_instructions < 8);
250 unsigned attrib;
251 unsigned chan;
252 unsigned cbuf;
253 unsigned depth_mode;
254 struct lp_bld_tgsi_system_values system_values;
255
256 memset(&system_values, 0, sizeof(system_values));
257
258 if (key->depth.enabled ||
259 key->stencil[0].enabled ||
260 key->stencil[1].enabled) {
261
262 zs_format_desc = util_format_description(key->zsbuf_format);
263 assert(zs_format_desc);
264
265 if (!shader->info.base.writes_z) {
266 if (key->alpha.enabled || shader->info.base.uses_kill)
267 /* With alpha test and kill, can do the depth test early
268 * and hopefully eliminate some quads. But need to do a
269 * special deferred depth write once the final mask value
270 * is known.
271 */
272 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
273 else
274 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
275 }
276 else {
277 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
278 }
279
280 if (!(key->depth.enabled && key->depth.writemask) &&
281 !(key->stencil[0].enabled && key->stencil[0].writemask))
282 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
283 }
284 else {
285 depth_mode = 0;
286 }
287
288 assert(i < 4);
289
290 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
291 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
292
293 vec_type = lp_build_vec_type(gallivm, type);
294
295 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
296
297 memset(outputs, 0, sizeof outputs);
298
299 /* Declare the color and z variables */
300 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
301 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
302 color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
303 }
304 }
305
306 /* do triangle edge testing */
307 if (partial_mask) {
308 *pmask = generate_quad_mask(gallivm, type,
309 i*type.length/4, mask_input);
310 }
311 else {
312 *pmask = lp_build_const_int_vec(gallivm, type, ~0);
313 }
314
315 /* 'mask' will control execution based on quad's pixel alive/killed state */
316 lp_build_mask_begin(&mask, gallivm, type, *pmask);
317
318 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
319 lp_build_mask_check(&mask);
320
321 lp_build_interp_soa_update_pos(interp, gallivm, i*type.length/4);
322 z = interp->pos[2];
323
324 if (depth_mode & EARLY_DEPTH_TEST) {
325 lp_build_depth_stencil_test(gallivm,
326 &key->depth,
327 key->stencil,
328 type,
329 zs_format_desc,
330 &mask,
331 stencil_refs,
332 z,
333 depth_ptr, facing,
334 &zs_value,
335 !simple_shader);
336
337 if (depth_mode & EARLY_DEPTH_WRITE) {
338 lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
339 }
340 }
341
342 lp_build_interp_soa_update_inputs(interp, gallivm, i*type.length/4);
343
344 /* Build the actual shader */
345 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
346 consts_ptr, &system_values,
347 interp->pos, interp->inputs,
348 outputs, sampler, &shader->info.base);
349
350 /* Alpha test */
351 if (key->alpha.enabled) {
352 int color0 = find_output_by_semantic(&shader->info.base,
353 TGSI_SEMANTIC_COLOR,
354 0);
355
356 if (color0 != -1 && outputs[color0][3]) {
357 const struct util_format_description *cbuf_format_desc;
358 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
359 LLVMValueRef alpha_ref_value;
360
361 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
362 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
363
364 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
365
366 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
367 &mask, alpha, alpha_ref_value,
368 (depth_mode & LATE_DEPTH_TEST) != 0);
369 }
370 }
371
372 /* Late Z test */
373 if (depth_mode & LATE_DEPTH_TEST) {
374 int pos0 = find_output_by_semantic(&shader->info.base,
375 TGSI_SEMANTIC_POSITION,
376 0);
377
378 if (pos0 != -1 && outputs[pos0][2]) {
379 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
380 }
381
382 lp_build_depth_stencil_test(gallivm,
383 &key->depth,
384 key->stencil,
385 type,
386 zs_format_desc,
387 &mask,
388 stencil_refs,
389 z,
390 depth_ptr, facing,
391 &zs_value,
392 !simple_shader);
393 /* Late Z write */
394 if (depth_mode & LATE_DEPTH_WRITE) {
395 lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
396 }
397 }
398 else if ((depth_mode & EARLY_DEPTH_TEST) &&
399 (depth_mode & LATE_DEPTH_WRITE))
400 {
401 /* Need to apply a reduced mask to the depth write. Reload the
402 * depth value, update from zs_value with the new mask value and
403 * write that out.
404 */
405 lp_build_deferred_depth_write(gallivm,
406 type,
407 zs_format_desc,
408 &mask,
409 depth_ptr,
410 zs_value);
411 }
412
413
414 /* Color write */
415 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
416 {
417 if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
418 shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
419 {
420 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
421 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
422 if(outputs[attrib][chan]) {
423 /* XXX: just initialize outputs to point at colors[] and
424 * skip this.
425 */
426 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
427 lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
428 LLVMBuildStore(builder, out, color[cbuf][chan]);
429 }
430 }
431 }
432 }
433
434 if (counter)
435 lp_build_occlusion_count(gallivm, type,
436 lp_build_mask_value(&mask), counter);
437
438 *pmask = lp_build_mask_end(&mask);
439 }
440
441
442 /**
443 * Generate the fragment shader, depth/stencil test, and alpha tests.
444 */
445 static void
446 generate_fs_loop(struct gallivm_state *gallivm,
447 struct lp_fragment_shader *shader,
448 const struct lp_fragment_shader_variant_key *key,
449 LLVMBuilderRef builder,
450 struct lp_type type,
451 LLVMValueRef context_ptr,
452 LLVMValueRef num_loop,
453 struct lp_build_interp_soa_context *interp,
454 struct lp_build_sampler_soa *sampler,
455 LLVMValueRef mask_store,
456 LLVMValueRef (*out_color)[4],
457 LLVMValueRef depth_ptr,
458 unsigned depth_bits,
459 LLVMValueRef facing,
460 LLVMValueRef counter)
461 {
462 const struct util_format_description *zs_format_desc = NULL;
463 const struct tgsi_token *tokens = shader->base.tokens;
464 LLVMTypeRef vec_type;
465 LLVMValueRef mask_ptr, mask_val;
466 LLVMValueRef consts_ptr;
467 LLVMValueRef z;
468 LLVMValueRef zs_value = NULL;
469 LLVMValueRef stencil_refs[2];
470 LLVMValueRef depth_ptr_i;
471 LLVMValueRef depth_offset;
472 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
473 struct lp_build_for_loop_state loop_state;
474 struct lp_build_mask_context mask;
475 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
476 shader->info.base.num_inputs < 3 &&
477 shader->info.base.num_instructions < 8);
478 unsigned attrib;
479 unsigned chan;
480 unsigned cbuf;
481 unsigned depth_mode;
482
483 struct lp_bld_tgsi_system_values system_values;
484
485 memset(&system_values, 0, sizeof(system_values));
486
487 if (key->depth.enabled ||
488 key->stencil[0].enabled ||
489 key->stencil[1].enabled) {
490
491 zs_format_desc = util_format_description(key->zsbuf_format);
492 assert(zs_format_desc);
493
494 if (!shader->info.base.writes_z) {
495 if (key->alpha.enabled || shader->info.base.uses_kill)
496 /* With alpha test and kill, can do the depth test early
497 * and hopefully eliminate some quads. But need to do a
498 * special deferred depth write once the final mask value
499 * is known.
500 */
501 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
502 else
503 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
504 }
505 else {
506 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
507 }
508
509 if (!(key->depth.enabled && key->depth.writemask) &&
510 !(key->stencil[0].enabled && key->stencil[0].writemask))
511 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
512 }
513 else {
514 depth_mode = 0;
515 }
516
517
518 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
519 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
520
521 vec_type = lp_build_vec_type(gallivm, type);
522
523 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
524
525 lp_build_for_loop_begin(&loop_state, gallivm,
526 lp_build_const_int32(gallivm, 0),
527 LLVMIntULT,
528 num_loop,
529 lp_build_const_int32(gallivm, 1));
530
531 mask_ptr = LLVMBuildGEP(builder, mask_store,
532 &loop_state.counter, 1, "mask_ptr");
533 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
534
535 depth_offset = LLVMBuildMul(builder, loop_state.counter,
536 lp_build_const_int32(gallivm, depth_bits * type.length),
537 "");
538
539 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
540
541 memset(outputs, 0, sizeof outputs);
542
543 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
544 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
545 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
546 lp_build_vec_type(gallivm,
547 type),
548 num_loop, "color");
549 }
550 }
551
552
553
554 /* 'mask' will control execution based on quad's pixel alive/killed state */
555 lp_build_mask_begin(&mask, gallivm, type, mask_val);
556
557 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
558 lp_build_mask_check(&mask);
559
560 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
561 z = interp->pos[2];
562
563 if (depth_mode & EARLY_DEPTH_TEST) {
564 lp_build_depth_stencil_test(gallivm,
565 &key->depth,
566 key->stencil,
567 type,
568 zs_format_desc,
569 &mask,
570 stencil_refs,
571 z,
572 depth_ptr_i, facing,
573 &zs_value,
574 !simple_shader);
575
576 if (depth_mode & EARLY_DEPTH_WRITE) {
577 lp_build_depth_write(builder, zs_format_desc, depth_ptr_i, zs_value);
578 }
579 }
580
581 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
582
583 /* Build the actual shader */
584 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
585 consts_ptr, &system_values,
586 interp->pos, interp->inputs,
587 outputs, sampler, &shader->info.base);
588
589 /* Alpha test */
590 if (key->alpha.enabled) {
591 int color0 = find_output_by_semantic(&shader->info.base,
592 TGSI_SEMANTIC_COLOR,
593 0);
594
595 if (color0 != -1 && outputs[color0][3]) {
596 const struct util_format_description *cbuf_format_desc;
597 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
598 LLVMValueRef alpha_ref_value;
599
600 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
601 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
602
603 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
604
605 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
606 &mask, alpha, alpha_ref_value,
607 (depth_mode & LATE_DEPTH_TEST) != 0);
608 }
609 }
610
611 /* Late Z test */
612 if (depth_mode & LATE_DEPTH_TEST) {
613 int pos0 = find_output_by_semantic(&shader->info.base,
614 TGSI_SEMANTIC_POSITION,
615 0);
616
617 if (pos0 != -1 && outputs[pos0][2]) {
618 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
619 }
620
621 lp_build_depth_stencil_test(gallivm,
622 &key->depth,
623 key->stencil,
624 type,
625 zs_format_desc,
626 &mask,
627 stencil_refs,
628 z,
629 depth_ptr_i, facing,
630 &zs_value,
631 !simple_shader);
632 /* Late Z write */
633 if (depth_mode & LATE_DEPTH_WRITE) {
634 lp_build_depth_write(builder, zs_format_desc, depth_ptr_i, zs_value);
635 }
636 }
637 else if ((depth_mode & EARLY_DEPTH_TEST) &&
638 (depth_mode & LATE_DEPTH_WRITE))
639 {
640 /* Need to apply a reduced mask to the depth write. Reload the
641 * depth value, update from zs_value with the new mask value and
642 * write that out.
643 */
644 lp_build_deferred_depth_write(gallivm,
645 type,
646 zs_format_desc,
647 &mask,
648 depth_ptr_i,
649 zs_value);
650 }
651
652
653 /* Color write */
654 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
655 {
656 if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
657 shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
658 {
659 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
660 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
661 if(outputs[attrib][chan]) {
662 /* XXX: just initialize outputs to point at colors[] and
663 * skip this.
664 */
665 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
666 LLVMValueRef color_ptr;
667 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
668 &loop_state.counter, 1, "");
669 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
670 LLVMBuildStore(builder, out, color_ptr);
671 }
672 }
673 }
674 }
675
676 if (key->occlusion_count) {
677 lp_build_name(counter, "counter");
678 lp_build_occlusion_count(gallivm, type,
679 lp_build_mask_value(&mask), counter);
680 }
681
682 mask_val = lp_build_mask_end(&mask);
683 LLVMBuildStore(builder, mask_val, mask_ptr);
684 lp_build_for_loop_end(&loop_state);
685 }
686
687
688 /**
689 * Generate color blending and color output.
690 * \param rt the render target index (to index blend, colormask state)
691 * \param type the pixel color type
692 * \param context_ptr pointer to the runtime JIT context
693 * \param mask execution mask (active fragment/pixel mask)
694 * \param src colors from the fragment shader
695 * \param dst_ptr the destination color buffer pointer
696 */
697 static void
698 generate_blend(struct gallivm_state *gallivm,
699 const struct pipe_blend_state *blend,
700 unsigned rt,
701 LLVMBuilderRef builder,
702 struct lp_type type,
703 LLVMValueRef context_ptr,
704 LLVMValueRef mask,
705 LLVMValueRef *src,
706 LLVMValueRef dst_ptr,
707 boolean do_branch)
708 {
709 struct lp_build_context bld;
710 struct lp_build_mask_context mask_ctx;
711 LLVMTypeRef vec_type;
712 LLVMValueRef const_ptr;
713 LLVMValueRef con[4];
714 LLVMValueRef dst[4];
715 LLVMValueRef res[4];
716 unsigned chan;
717
718 lp_build_context_init(&bld, gallivm, type);
719
720 lp_build_mask_begin(&mask_ctx, gallivm, type, mask);
721 if (do_branch)
722 lp_build_mask_check(&mask_ctx);
723
724 vec_type = lp_build_vec_type(gallivm, type);
725
726 const_ptr = lp_jit_context_u8_blend_color(gallivm, context_ptr);
727 const_ptr = LLVMBuildBitCast(builder, const_ptr,
728 LLVMPointerType(vec_type, 0), "");
729
730 /* load constant blend color and colors from the dest color buffer */
731 for(chan = 0; chan < 4; ++chan) {
732 LLVMValueRef index = lp_build_const_int32(gallivm, chan);
733 con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
734
735 dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
736
737 lp_build_name(con[chan], "con.%c", "rgba"[chan]);
738 lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
739 }
740
741 /* do blend */
742 lp_build_blend_soa(gallivm, blend, type, rt, src, dst, con, res);
743
744 /* store results to color buffer */
745 for(chan = 0; chan < 4; ++chan) {
746 if(blend->rt[rt].colormask & (1 << chan)) {
747 LLVMValueRef index = lp_build_const_int32(gallivm, chan);
748 lp_build_name(res[chan], "res.%c", "rgba"[chan]);
749 res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
750 LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
751 }
752 }
753
754 lp_build_mask_end(&mask_ctx);
755 }
756
757
758 /**
759 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
760 *
761 * Fragment Shader outputs pixels in small 2x2 blocks
762 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
763 *
764 * However in memory pixels are stored in rows
765 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
766 *
767 * @param type fragment shader type (4x or 8x float)
768 * @param num_fs number of fs_src
769 * @param dst_channels number of output channels
770 * @param fs_src output from fragment shader
771 * @param dst pointer to store result
772 * @param pad_inline is channel padding inline or at end of row
773 * @return the number of dsts
774 */
775 static int
776 generate_fs_twiddle(struct gallivm_state *gallivm,
777 struct lp_type type,
778 unsigned num_fs,
779 unsigned dst_channels,
780 LLVMValueRef fs_src[][4],
781 LLVMValueRef* dst,
782 bool pad_inline)
783 {
784 LLVMValueRef src[16];
785
786 bool swizzle_pad;
787 bool twiddle;
788 bool split;
789
790 unsigned pixels = num_fs == 4 ? 1 : 2;
791 unsigned reorder_group;
792 unsigned src_channels;
793 unsigned src_count;
794 unsigned i;
795
796 src_channels = dst_channels < 3 ? dst_channels : 4;
797 src_count = num_fs * src_channels;
798
799 assert(pixels == 2 || num_fs == 4);
800 assert(num_fs * src_channels <= Elements(src));
801
802 /*
803 * Transpose from SoA -> AoS
804 */
805 for (i = 0; i < num_fs; ++i) {
806 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
807 }
808
809 /*
810 * Pick transformation options
811 */
812 swizzle_pad = false;
813 twiddle = false;
814 split = false;
815 reorder_group = 0;
816
817 if (dst_channels == 1) {
818 twiddle = true;
819
820 if (pixels == 2) {
821 split = true;
822 }
823 } else if (dst_channels == 2) {
824 if (pixels == 1) {
825 reorder_group = 1;
826 }
827 } else if (dst_channels > 2) {
828 if (pixels == 1) {
829 reorder_group = 2;
830 } else {
831 twiddle = true;
832 }
833
834 if (!pad_inline && dst_channels == 3 && pixels > 1) {
835 swizzle_pad = true;
836 }
837 }
838
839 /*
840 * Split the src in half
841 */
842 if (split) {
843 for (i = num_fs; i > 0; --i) {
844 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
845 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
846 }
847
848 src_count *= 2;
849 type.length = 4;
850 }
851
852 /*
853 * Ensure pixels are in memory order
854 */
855 if (reorder_group) {
856 /* Twiddle pixels by reordering the array, e.g.:
857 *
858 * src_count = 8 -> 0 2 1 3 4 6 5 7
859 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
860 */
861 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
862
863 for (i = 0; i < src_count; ++i) {
864 unsigned group = i / reorder_group;
865 unsigned block = (group / 4) * 4 * reorder_group;
866 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
867 dst[i] = src[j];
868 }
869 } else if (twiddle) {
870 /* Twiddle pixels across elements of array */
871 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
872 } else {
873 /* Do nothing */
874 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
875 }
876
877 /*
878 * Moves any padding between pixels to the end
879 * e.g. RGBXRGBX -> RGBRGBXX
880 */
881 if (swizzle_pad) {
882 unsigned char swizzles[16];
883 unsigned elems = pixels * dst_channels;
884
885 for (i = 0; i < type.length; ++i) {
886 if (i < elems)
887 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
888 else
889 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
890 }
891
892 for (i = 0; i < src_count; ++i) {
893 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
894 }
895 }
896
897 return src_count;
898 }
899
900
901 /**
902 * Load an unswizzled block of pixels from memory
903 */
904 static void
905 load_unswizzled_block(struct gallivm_state *gallivm,
906 LLVMValueRef base_ptr,
907 LLVMValueRef stride,
908 unsigned block_width,
909 unsigned block_height,
910 LLVMValueRef* dst,
911 struct lp_type dst_type,
912 unsigned dst_count)
913 {
914 LLVMBuilderRef builder = gallivm->builder;
915 unsigned row_size = dst_count / block_height;
916 unsigned i;
917
918 /* Ensure block exactly fits into dst */
919 assert((block_width * block_height) % dst_count == 0);
920
921 for (i = 0; i < dst_count; ++i) {
922 unsigned x = i % row_size;
923 unsigned y = i / row_size;
924
925 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
926 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
927
928 LLVMValueRef gep[2];
929 LLVMValueRef dst_ptr;
930
931 gep[0] = lp_build_const_int32(gallivm, 0);
932 gep[1] = LLVMBuildAdd(builder, bx, by, "");
933
934 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
935 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
936
937 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
938
939 if ((dst_type.length % 3) == 0) {
940 lp_set_load_alignment(dst[i], dst_type.width / 8);
941 }
942 }
943 }
944
945
946 /**
947 * Store an unswizzled block of pixels to memory
948 */
949 static void
950 store_unswizzled_block(struct gallivm_state *gallivm,
951 LLVMValueRef base_ptr,
952 LLVMValueRef stride,
953 unsigned block_width,
954 unsigned block_height,
955 LLVMValueRef* src,
956 struct lp_type src_type,
957 unsigned src_count)
958 {
959 LLVMBuilderRef builder = gallivm->builder;
960 unsigned row_size = src_count / block_height;
961 unsigned i;
962
963 /* Ensure src exactly fits into block */
964 assert((block_width * block_height) % src_count == 0);
965
966 for (i = 0; i < src_count; ++i) {
967 unsigned x = i % row_size;
968 unsigned y = i / row_size;
969
970 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
971 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
972
973 LLVMValueRef gep[2];
974 LLVMValueRef src_ptr;
975
976 gep[0] = lp_build_const_int32(gallivm, 0);
977 gep[1] = LLVMBuildAdd(builder, bx, by, "");
978
979 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
980 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
981
982 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
983
984 if ((src_type.length % 3) == 0) {
985 lp_set_store_alignment(src_ptr, src_type.width / 8);
986 }
987 }
988 }
989
990
991 /**
992 * Checks if a format description is an arithmetic format
993 *
994 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
995 */
996 static INLINE boolean
997 is_arithmetic_format(const struct util_format_description *format_desc)
998 {
999 boolean arith = false;
1000 unsigned i;
1001
1002 for (i = 0; i < format_desc->nr_channels; ++i) {
1003 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
1004 arith |= (format_desc->channel[i].size % 8) != 0;
1005 }
1006
1007 return arith;
1008 }
1009
1010
1011 /**
1012 * Retrieves the type representing the memory layout for a format
1013 *
1014 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1015 */
1016 static INLINE void
1017 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
1018 struct lp_type* type)
1019 {
1020 int i;
1021
1022 memset(type, 0, sizeof(struct lp_type));
1023 type->floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT;
1024 type->fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED;
1025 type->sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED;
1026 type->norm = format_desc->channel[0].normalized;
1027
1028 if (is_arithmetic_format(format_desc)) {
1029 type->width = 0;
1030 type->length = 1;
1031
1032 for (i = 0; i < format_desc->nr_channels; ++i) {
1033 type->width += format_desc->channel[i].size;
1034 }
1035 } else {
1036 type->width = format_desc->channel[0].size;
1037 type->length = format_desc->nr_channels;
1038 }
1039 }
1040
1041
1042 /**
1043 * Retrieves the type for a format which is usable in the blending code.
1044 *
1045 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1046 */
1047 static INLINE void
1048 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1049 struct lp_type* type)
1050 {
1051 int i;
1052
1053 memset(type, 0, sizeof(struct lp_type));
1054 type->floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT;
1055 type->fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED;
1056 type->sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED;
1057 type->norm = format_desc->channel[0].normalized;
1058 type->width = format_desc->channel[0].size;
1059 type->length = format_desc->nr_channels;
1060
1061 for (i = 1; i < format_desc->nr_channels; ++i) {
1062 if (format_desc->channel[i].size > type->width)
1063 type->width = format_desc->channel[i].size;
1064 }
1065
1066 if (type->floating) {
1067 type->width = 32;
1068 } else {
1069 if (type->width <= 8) {
1070 type->width = 8;
1071 } else if (type->width <= 16) {
1072 type->width = 16;
1073 } else {
1074 type->width = 32;
1075 }
1076 }
1077
1078 if (is_arithmetic_format(format_desc) && type->length == 3) {
1079 type->length = 4;
1080 }
1081 }
1082
1083
1084 /**
1085 * Scale a normalised value from src_bits to dst_bits
1086 */
1087 static INLINE LLVMValueRef
1088 scale_bits(struct gallivm_state *gallivm,
1089 int src_bits,
1090 int dst_bits,
1091 LLVMValueRef src,
1092 struct lp_type src_type)
1093 {
1094 LLVMBuilderRef builder = gallivm->builder;
1095 LLVMValueRef result = src;
1096
1097 if (dst_bits < src_bits) {
1098 /* Scale down by LShr */
1099 result = LLVMBuildLShr(builder,
1100 src,
1101 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
1102 "");
1103 } else if (dst_bits > src_bits) {
1104 /* Scale up bits */
1105 int db = dst_bits - src_bits;
1106
1107 /* Shift left by difference in bits */
1108 result = LLVMBuildShl(builder,
1109 src,
1110 lp_build_const_int_vec(gallivm, src_type, db),
1111 "");
1112
1113 if (db < src_bits) {
1114 /* Enough bits in src to fill the remainder */
1115 LLVMValueRef lower = LLVMBuildLShr(builder,
1116 src,
1117 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1118 "");
1119
1120 result = LLVMBuildOr(builder, result, lower, "");
1121 } else if (db > src_bits) {
1122 /* Need to repeatedely copy src bits to fill remainder in dst */
1123 unsigned n;
1124
1125 for (n = src_bits; n < dst_bits; n *= 2) {
1126 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1127
1128 result = LLVMBuildOr(builder,
1129 result,
1130 LLVMBuildLShr(builder, result, shuv, ""),
1131 "");
1132 }
1133 }
1134 }
1135
1136 return result;
1137 }
1138
1139
1140 /**
1141 * Convert from memory format to blending format
1142 *
1143 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1144 */
1145 static void
1146 convert_to_blend_type(struct gallivm_state *gallivm,
1147 const struct util_format_description *src_fmt,
1148 struct lp_type src_type,
1149 struct lp_type dst_type,
1150 LLVMValueRef* src,
1151 unsigned num_srcs,
1152 LLVMValueRef* dst)
1153 {
1154 LLVMBuilderRef builder = gallivm->builder;
1155 struct lp_type blend_type;
1156 struct lp_type mem_type;
1157 unsigned i, j, k;
1158 unsigned pixels = 16 / num_srcs;
1159 bool is_arith;
1160
1161 memcpy(dst, src, sizeof(LLVMValueRef*) * num_srcs);
1162
1163 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1164 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1165
1166 /* Is the format arithmetic */
1167 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1168 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1169
1170 /* Pad if necessary */
1171 if (!is_arith && src_type.length < dst_type.length) {
1172 for (i = 0; i < num_srcs; ++i) {
1173 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1174 }
1175
1176 src_type.length = dst_type.length;
1177 }
1178
1179 /* Special case for half-floats */
1180 if (mem_type.width == 16 && mem_type.floating) {
1181 assert(blend_type.width == 32 && blend_type.floating);
1182 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1183 is_arith = false;
1184 }
1185
1186 if (!is_arith) {
1187 return;
1188 }
1189
1190 src_type.width = blend_type.width * blend_type.length;
1191 blend_type.length *= pixels;
1192 src_type.length *= pixels / (src_type.length / mem_type.length);
1193
1194 for (i = 0; i < num_srcs; ++i) {
1195 LLVMValueRef chans[4];
1196 LLVMValueRef res;
1197 unsigned sa = 0;
1198
1199 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1200
1201 for (j = 0; j < src_fmt->nr_channels; ++j) {
1202 unsigned mask = 0;
1203
1204 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1205 mask |= 1 << k;
1206 }
1207
1208 /* Extract bits from source */
1209 chans[j] = LLVMBuildLShr(builder,
1210 dst[i],
1211 lp_build_const_int_vec(gallivm, src_type, sa),
1212 "");
1213
1214 chans[j] = LLVMBuildAnd(builder,
1215 chans[j],
1216 lp_build_const_int_vec(gallivm, src_type, mask),
1217 "");
1218
1219 /* Scale bits */
1220 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size, blend_type.width, chans[j], src_type);
1221
1222 /* Insert bits into correct position */
1223 chans[j] = LLVMBuildShl(builder,
1224 chans[j],
1225 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1226 "");
1227
1228 sa += src_fmt->channel[j].size;
1229
1230 if (j == 0) {
1231 res = chans[j];
1232 } else {
1233 res = LLVMBuildOr(builder, res, chans[j], "");
1234 }
1235 }
1236
1237 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1238 }
1239 }
1240
1241
1242 /**
1243 * Convert from blending format to memory format
1244 *
1245 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1246 */
1247 static void
1248 convert_from_blend_type(struct gallivm_state *gallivm,
1249 const struct util_format_description *src_fmt,
1250 struct lp_type src_type,
1251 struct lp_type dst_type,
1252 LLVMValueRef* src,
1253 unsigned num_srcs,
1254 LLVMValueRef* dst)
1255 {
1256 unsigned i, j, k;
1257 struct lp_type mem_type;
1258 struct lp_type blend_type;
1259 LLVMBuilderRef builder = gallivm->builder;
1260 unsigned pixels = 16 / num_srcs;
1261 bool is_arith;
1262
1263 memcpy(dst, src, sizeof(LLVMValueRef*) * num_srcs);
1264
1265 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1266 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1267
1268 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1269
1270 /* Special case for half-floats */
1271 if (mem_type.width == 16 && mem_type.floating) {
1272 int length = dst_type.length;
1273 assert(blend_type.width == 32 && blend_type.floating);
1274
1275 dst_type.length = src_type.length;
1276
1277 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1278
1279 dst_type.length = length;
1280 is_arith = false;
1281 }
1282
1283 /* Remove any padding */
1284 if (!is_arith && (src_type.length % mem_type.length)) {
1285 src_type.length -= (src_type.length % mem_type.length);
1286
1287 for (i = 0; i < num_srcs; ++i) {
1288 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1289 }
1290 }
1291
1292 /* No bit arithmitic to do */
1293 if (!is_arith) {
1294 return;
1295 }
1296
1297 src_type.length = pixels;
1298 src_type.width = blend_type.length * blend_type.width;
1299 dst_type.length = pixels;
1300
1301 for (i = 0; i < num_srcs; ++i) {
1302 LLVMValueRef chans[4];
1303 LLVMValueRef res;
1304 unsigned sa = 0;
1305
1306 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1307
1308 for (j = 0; j < src_fmt->nr_channels; ++j) {
1309 unsigned mask = 0;
1310
1311 assert(blend_type.width > src_fmt->channel[j].size);
1312
1313 for (k = 0; k < blend_type.width; ++k) {
1314 mask |= 1 << k;
1315 }
1316
1317 /* Extract bits */
1318 chans[j] = LLVMBuildLShr(builder,
1319 dst[i],
1320 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1321 "");
1322
1323 chans[j] = LLVMBuildAnd(builder,
1324 chans[j],
1325 lp_build_const_int_vec(gallivm, src_type, mask),
1326 "");
1327
1328 /* Scale down bits */
1329 chans[j] = scale_bits(gallivm, blend_type.width, src_fmt->channel[j].size, chans[j], src_type);
1330
1331 /* Insert bits */
1332 chans[j] = LLVMBuildShl(builder,
1333 chans[j],
1334 lp_build_const_int_vec(gallivm, src_type, sa),
1335 "");
1336
1337 sa += src_fmt->channel[j].size;
1338
1339 if (j == 0) {
1340 res = chans[j];
1341 } else {
1342 res = LLVMBuildOr(builder, res, chans[j], "");
1343 }
1344 }
1345
1346 assert (dst_type.width != 24);
1347
1348 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1349 }
1350 }
1351
1352
1353 /**
1354 * Generates the blend function for unswizzled colour buffers
1355 * Also generates the read & write from colour buffer
1356 */
1357 static void
1358 generate_unswizzled_blend(struct gallivm_state *gallivm,
1359 unsigned rt,
1360 struct lp_fragment_shader_variant *variant,
1361 enum pipe_format out_format,
1362 unsigned int num_fs,
1363 struct lp_type fs_type,
1364 LLVMValueRef* fs_mask,
1365 LLVMValueRef fs_out_color[TGSI_NUM_CHANNELS][4],
1366 LLVMValueRef context_ptr,
1367 LLVMValueRef color_ptr,
1368 LLVMValueRef stride,
1369 unsigned partial_mask,
1370 boolean do_branch)
1371 {
1372 const unsigned alpha_channel = 3;
1373 const unsigned block_width = 4;
1374 const unsigned block_height = 4;
1375 const unsigned block_size = block_width * block_height;
1376 const unsigned lp_integer_vector_width = 128;
1377
1378 LLVMBuilderRef builder = gallivm->builder;
1379 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1380 LLVMValueRef src_alpha[block_size];
1381 LLVMValueRef src_mask[block_size];
1382 LLVMValueRef src[block_size];
1383 LLVMValueRef dst[block_size];
1384 LLVMValueRef blend_color;
1385 LLVMValueRef blend_alpha;
1386 LLVMValueRef i32_zero;
1387 LLVMValueRef check_mask;
1388
1389 struct lp_build_mask_context mask_ctx;
1390 struct lp_type mask_type;
1391 struct lp_type blend_type;
1392 struct lp_type alpha_type;
1393 struct lp_type row_type;
1394 struct lp_type dst_type;
1395
1396 unsigned char swizzle[TGSI_NUM_CHANNELS];
1397 unsigned vector_width;
1398 unsigned dst_channels;
1399 unsigned src_channels;
1400 unsigned dst_count;
1401 unsigned src_count;
1402 unsigned i, j;
1403
1404 const struct util_format_description* out_format_desc = util_format_description(out_format);
1405
1406 bool pad_inline = is_arithmetic_format(out_format_desc);
1407 bool has_alpha = false;
1408
1409 src_channels = TGSI_NUM_CHANNELS;
1410 mask_type = lp_int32_vec4_type();
1411 mask_type.length = fs_type.length;
1412
1413 /* Do not bother executing code when mask is empty.. */
1414 if (do_branch) {
1415 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1416
1417 for (i = 0; i < num_fs; ++i) {
1418 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1419 }
1420
1421 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1422 lp_build_mask_check(&mask_ctx);
1423 }
1424
1425 partial_mask |= !variant->opaque;
1426 i32_zero = lp_build_const_int32(gallivm, 0);
1427
1428 /* Get type from output format */
1429 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1430 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1431
1432 row_type.length = fs_type.length;
1433 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1434
1435 /* Compute correct swizzle and count channels */
1436 memset(swizzle, 0xFF, TGSI_NUM_CHANNELS);
1437 dst_channels = 0;
1438
1439 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1440 /* Ensure channel is used */
1441 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1442 continue;
1443 }
1444
1445 /* Ensure not already written to (happens in case with GL_ALPHA) */
1446 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1447 continue;
1448 }
1449
1450 /* Ensure we havn't already found all channels */
1451 if (dst_channels >= out_format_desc->nr_channels) {
1452 continue;
1453 }
1454
1455 swizzle[out_format_desc->swizzle[i]] = i;
1456 ++dst_channels;
1457
1458 if (i == alpha_channel) {
1459 has_alpha = true;
1460 }
1461 }
1462
1463 /* If 3 channels then pad to include alpha for 4 element transpose */
1464 if (dst_channels == 3 && !has_alpha) {
1465 swizzle[3] = 3;
1466
1467 if (out_format_desc->nr_channels == 4) {
1468 dst_channels = 4;
1469 }
1470 }
1471
1472 /*
1473 * Load shader output
1474 */
1475 for (i = 0; i < num_fs; ++i) {
1476 /* Always load alpha for use in blending */
1477 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[alpha_channel][i], "");
1478
1479 /* Load each channel */
1480 for (j = 0; j < dst_channels; ++j) {
1481 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[swizzle[j]][i], "");
1482 }
1483
1484 /* If 3 channels then pad to include alpha for 4 element transpose */
1485 if (dst_channels == 3 && !has_alpha) {
1486 fs_src[i][3] = alpha;
1487 swizzle[3] = 3;
1488 }
1489
1490 /* We split the row_mask and row_alpha as we want 128bit interleave */
1491 if (fs_type.length == 8) {
1492 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1493 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1494
1495 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1496 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1497 } else {
1498 src_mask[i] = fs_mask[i];
1499 src_alpha[i] = alpha;
1500 }
1501 }
1502
1503
1504 /*
1505 * Pixel twiddle from fragment shader order to memory order
1506 */
1507 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1508 src_channels = dst_channels < 3 ? dst_channels : 4;
1509 if (src_count != num_fs * src_channels) {
1510 unsigned ds = src_count / (num_fs * src_channels);
1511 row_type.length /= ds;
1512 fs_type.length = row_type.length;
1513 }
1514
1515 blend_type = row_type;
1516 alpha_type = fs_type;
1517 alpha_type.length = 4;
1518 mask_type.length = 4;
1519
1520 /* Convert src to row_type */
1521 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1522
1523 /* If the rows are not an SSE vector, combine them to become SSE size! */
1524 if ((row_type.width * row_type.length) % 128) {
1525 unsigned bits = row_type.width * row_type.length;
1526 unsigned combined;
1527
1528 dst_count = src_count / (vector_width / bits);
1529 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1530
1531 row_type.length *= combined;
1532 src_count /= combined;
1533
1534 bits = row_type.width * row_type.length;
1535 assert(bits == 128 || bits == 256);
1536 }
1537
1538
1539 /*
1540 * Blend Colour conversion
1541 */
1542 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1543 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1544 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1545
1546 /* Convert */
1547 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1548
1549 /* Extract alpha */
1550 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1551
1552 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1553 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1554 if (pad_inline) {
1555 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1556 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1557 } else {
1558 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1559 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1560 }
1561
1562 /*
1563 * Mask conversion
1564 */
1565 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1566
1567 if (src_count < block_height) {
1568 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1569 } else if (src_count > block_height) {
1570 for (i = src_count; i > 0; --i) {
1571 unsigned pixels = block_size / src_count;
1572 unsigned idx = i - 1;
1573
1574 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1575 }
1576 }
1577
1578 assert(mask_type.width == 32);
1579
1580 for (i = 0; i < src_count; ++i) {
1581 unsigned pixels = block_size / src_count;
1582 unsigned pixel_width = row_type.width * dst_channels;
1583
1584 if (pixel_width == 24) {
1585 mask_type.width = 8;
1586 mask_type.length = vector_width / mask_type.width;
1587 } else {
1588 mask_type.length = pixels;
1589 mask_type.width = row_type.width * dst_channels;
1590
1591 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1592
1593 mask_type.length *= dst_channels;
1594 mask_type.width /= dst_channels;
1595 }
1596
1597 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1598 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1599 }
1600
1601 /*
1602 * Alpha conversion
1603 */
1604 if (!has_alpha) {
1605 unsigned length = row_type.length;
1606 row_type.length = alpha_type.length;
1607
1608 /* Twiddle the alpha to match pixels */
1609 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1610
1611 for (i = 0; i < 4; ++i) {
1612 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1613 }
1614
1615 alpha_type = row_type;
1616 row_type.length = length;
1617
1618 /* If only one channel we can only need the single alpha value per pixel */
1619 if (src_count == 1) {
1620 assert(dst_channels == 1);
1621
1622 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1623 } else {
1624 /* If there are more srcs than rows then we need to split alpha up */
1625 if (src_count > block_height) {
1626 for (i = src_count; i > 0; --i) {
1627 unsigned pixels = block_size / src_count;
1628 unsigned idx = i - 1;
1629
1630 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1631 }
1632 }
1633
1634 /* If there is a src for each pixel broadcast the alpha across whole row */
1635 if (src_count == block_size) {
1636 for (i = 0; i < src_count; ++i) {
1637 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1638 }
1639 } else {
1640 unsigned pixels = block_size / src_count;
1641 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1642 unsigned alpha_span = 1;
1643
1644 /* Check if we need 2 src_alphas for our shuffles */
1645 if (pixels > alpha_type.length) {
1646 alpha_span = 2;
1647 }
1648
1649 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1650 for (i = 0; i < src_count; ++i) {
1651 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1652 unsigned idx1 = i, idx2 = i;
1653
1654 if (alpha_span > 1){
1655 idx1 *= alpha_span;
1656 idx2 = idx1 + 1;
1657 }
1658
1659 for (j = 0; j < row_type.length; ++j) {
1660 if (j < pixels * channels) {
1661 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1662 } else {
1663 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1664 }
1665 }
1666
1667 src_alpha[i] = LLVMBuildShuffleVector(builder,
1668 src_alpha[idx1],
1669 src_alpha[idx2],
1670 LLVMConstVector(shuffles, row_type.length),
1671 "");
1672 }
1673 }
1674 }
1675 }
1676
1677
1678 /*
1679 * Load dst from memory
1680 */
1681 if (src_count < block_height) {
1682 dst_count = block_height;
1683 } else {
1684 dst_count = src_count;
1685 }
1686
1687 dst_type.length *= 16 / dst_count;
1688
1689 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height, dst, dst_type, dst_count);
1690
1691
1692 /*
1693 * Convert from dst/output format to src/blending format.
1694 *
1695 * This is necessary as we can only read 1 row from memory at a time,
1696 * so the minimum dst_count will ever be at this point is 4.
1697 *
1698 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1699 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1700 * on all 16 pixels in that single vector at once.
1701 */
1702 if (dst_count > src_count) {
1703 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1704 }
1705
1706 /*
1707 * Blending
1708 */
1709 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count, dst);
1710
1711 for (i = 0; i < src_count; ++i) {
1712 dst[i] = lp_build_blend_aos(gallivm,
1713 &variant->key.blend,
1714 variant->key.cbuf_format,
1715 row_type,
1716 rt,
1717 src[i],
1718 has_alpha ? NULL : src_alpha[i],
1719 dst[i],
1720 partial_mask ? src_mask[i] : NULL,
1721 blend_color,
1722 has_alpha ? NULL : blend_alpha,
1723 swizzle,
1724 pad_inline ? 4 : dst_channels);
1725 }
1726
1727 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count, dst);
1728
1729 /* Split the blend rows back to memory rows */
1730 if (dst_count > src_count) {
1731 row_type.length = dst_type.length * (dst_count / src_count);
1732
1733 if (src_count == 1) {
1734 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1735 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1736
1737 row_type.length /= 2;
1738 src_count *= 2;
1739 }
1740
1741 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1742 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1743 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1744 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1745
1746 row_type.length /= 2;
1747 src_count *= 2;
1748 }
1749
1750
1751 /*
1752 * Store blend result to memory
1753 */
1754 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height, dst, dst_type, dst_count);
1755
1756 if (do_branch) {
1757 lp_build_mask_end(&mask_ctx);
1758 }
1759 }
1760
1761
1762 /**
1763 * Generate the runtime callable function for the whole fragment pipeline.
1764 * Note that the function which we generate operates on a block of 16
1765 * pixels at at time. The block contains 2x2 quads. Each quad contains
1766 * 2x2 pixels.
1767 */
1768 static void
1769 generate_fragment(struct llvmpipe_context *lp,
1770 struct lp_fragment_shader *shader,
1771 struct lp_fragment_shader_variant *variant,
1772 unsigned partial_mask)
1773 {
1774 struct gallivm_state *gallivm = variant->gallivm;
1775 const struct lp_fragment_shader_variant_key *key = &variant->key;
1776 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1777 char func_name[256];
1778 struct lp_type fs_type;
1779 struct lp_type blend_type;
1780 LLVMTypeRef fs_elem_type;
1781 LLVMTypeRef blend_vec_type;
1782 LLVMTypeRef arg_types[12];
1783 LLVMTypeRef func_type;
1784 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1785 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1786 LLVMValueRef context_ptr;
1787 LLVMValueRef x;
1788 LLVMValueRef y;
1789 LLVMValueRef a0_ptr;
1790 LLVMValueRef dadx_ptr;
1791 LLVMValueRef dady_ptr;
1792 LLVMValueRef color_ptr_ptr;
1793 LLVMValueRef stride_ptr;
1794 LLVMValueRef depth_ptr;
1795 LLVMValueRef mask_input;
1796 LLVMValueRef counter = NULL;
1797 LLVMBasicBlockRef block;
1798 LLVMBuilderRef builder;
1799 struct lp_build_sampler_soa *sampler;
1800 struct lp_build_interp_soa_context interp;
1801 LLVMValueRef fs_mask[16 / 4];
1802 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1803 LLVMValueRef blend_mask;
1804 LLVMValueRef function;
1805 LLVMValueRef facing;
1806 const struct util_format_description *zs_format_desc;
1807 unsigned num_fs;
1808 unsigned i;
1809 unsigned chan;
1810 unsigned cbuf;
1811 boolean cbuf0_write_all;
1812 boolean try_loop = TRUE;
1813
1814 assert(lp_native_vector_width / 32 >= 4);
1815
1816 /* Adjust color input interpolation according to flatshade state:
1817 */
1818 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1819 for (i = 0; i < shader->info.base.num_inputs; i++) {
1820 if (inputs[i].interp == LP_INTERP_COLOR) {
1821 if (key->flatshade)
1822 inputs[i].interp = LP_INTERP_CONSTANT;
1823 else
1824 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1825 }
1826 }
1827
1828 /* check if writes to cbuf[0] are to be copied to all cbufs */
1829 cbuf0_write_all = FALSE;
1830 for (i = 0;i < shader->info.base.num_properties; i++) {
1831 if (shader->info.base.properties[i].name ==
1832 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1833 cbuf0_write_all = TRUE;
1834 break;
1835 }
1836 }
1837
1838 /* TODO: actually pick these based on the fs and color buffer
1839 * characteristics. */
1840
1841 memset(&fs_type, 0, sizeof fs_type);
1842 fs_type.floating = TRUE; /* floating point values */
1843 fs_type.sign = TRUE; /* values are signed */
1844 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1845 fs_type.width = 32; /* 32-bit float */
1846 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1847 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1848
1849 memset(&blend_type, 0, sizeof blend_type);
1850 blend_type.floating = FALSE; /* values are integers */
1851 blend_type.sign = FALSE; /* values are unsigned */
1852 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1853 blend_type.width = 8; /* 8-bit ubyte values */
1854 blend_type.length = 16; /* 16 elements per vector */
1855
1856 /*
1857 * Generate the function prototype. Any change here must be reflected in
1858 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1859 */
1860
1861 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1862
1863 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1864
1865 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1866 shader->no, variant->no, partial_mask ? "partial" : "whole");
1867
1868 arg_types[0] = variant->jit_context_ptr_type; /* context */
1869 arg_types[1] = int32_type; /* x */
1870 arg_types[2] = int32_type; /* y */
1871 arg_types[3] = int32_type; /* facing */
1872 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1873 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1874 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1875 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1876 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1877 arg_types[9] = int32_type; /* mask_input */
1878 arg_types[10] = LLVMPointerType(int32_type, 0); /* counter */
1879 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1880
1881 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1882 arg_types, Elements(arg_types), 0);
1883
1884 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1885 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1886
1887 variant->function[partial_mask] = function;
1888
1889 /* XXX: need to propagate noalias down into color param now we are
1890 * passing a pointer-to-pointer?
1891 */
1892 for(i = 0; i < Elements(arg_types); ++i)
1893 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1894 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1895
1896 context_ptr = LLVMGetParam(function, 0);
1897 x = LLVMGetParam(function, 1);
1898 y = LLVMGetParam(function, 2);
1899 facing = LLVMGetParam(function, 3);
1900 a0_ptr = LLVMGetParam(function, 4);
1901 dadx_ptr = LLVMGetParam(function, 5);
1902 dady_ptr = LLVMGetParam(function, 6);
1903 color_ptr_ptr = LLVMGetParam(function, 7);
1904 depth_ptr = LLVMGetParam(function, 8);
1905 mask_input = LLVMGetParam(function, 9);
1906 stride_ptr = LLVMGetParam(function, 11);
1907
1908 lp_build_name(context_ptr, "context");
1909 lp_build_name(x, "x");
1910 lp_build_name(y, "y");
1911 lp_build_name(a0_ptr, "a0");
1912 lp_build_name(dadx_ptr, "dadx");
1913 lp_build_name(dady_ptr, "dady");
1914 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
1915 lp_build_name(depth_ptr, "depth");
1916 lp_build_name(mask_input, "mask_input");
1917 lp_build_name(stride_ptr, "stride_ptr");
1918
1919 if (key->occlusion_count) {
1920 counter = LLVMGetParam(function, 10);
1921 lp_build_name(counter, "counter");
1922 }
1923
1924 /*
1925 * Function body
1926 */
1927
1928 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1929 builder = gallivm->builder;
1930 assert(builder);
1931 LLVMPositionBuilderAtEnd(builder, block);
1932
1933 /* code generated texture sampling */
1934 sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
1935
1936 zs_format_desc = util_format_description(key->zsbuf_format);
1937
1938 if (!try_loop) {
1939 /*
1940 * The shader input interpolation info is not explicitely baked in the
1941 * shader key, but everything it derives from (TGSI, and flatshade) is
1942 * already included in the shader key.
1943 */
1944 lp_build_interp_soa_init(&interp,
1945 gallivm,
1946 shader->info.base.num_inputs,
1947 inputs,
1948 builder, fs_type,
1949 FALSE,
1950 a0_ptr, dadx_ptr, dady_ptr,
1951 x, y);
1952
1953 /* loop over quads in the block */
1954 for(i = 0; i < num_fs; ++i) {
1955 LLVMValueRef depth_offset = LLVMConstInt(int32_type,
1956 i*fs_type.length*zs_format_desc->block.bits/8,
1957 0);
1958 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1959 LLVMValueRef depth_ptr_i;
1960
1961 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
1962
1963 generate_fs(gallivm,
1964 shader, key,
1965 builder,
1966 fs_type,
1967 context_ptr,
1968 i,
1969 &interp,
1970 sampler,
1971 &fs_mask[i], /* output */
1972 out_color,
1973 depth_ptr_i,
1974 facing,
1975 partial_mask,
1976 mask_input,
1977 counter);
1978
1979 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
1980 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
1981 fs_out_color[cbuf][chan][i] =
1982 out_color[cbuf * !cbuf0_write_all][chan];
1983 }
1984 }
1985 else {
1986 unsigned depth_bits = zs_format_desc->block.bits/8;
1987 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
1988 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
1989 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
1990 num_loop, "mask_store");
1991 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1992
1993 /*
1994 * The shader input interpolation info is not explicitely baked in the
1995 * shader key, but everything it derives from (TGSI, and flatshade) is
1996 * already included in the shader key.
1997 */
1998 lp_build_interp_soa_init(&interp,
1999 gallivm,
2000 shader->info.base.num_inputs,
2001 inputs,
2002 builder, fs_type,
2003 TRUE,
2004 a0_ptr, dadx_ptr, dady_ptr,
2005 x, y);
2006
2007 for (i = 0; i < num_fs; i++) {
2008 LLVMValueRef mask;
2009 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2010 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2011 &indexi, 1, "mask_ptr");
2012
2013 if (partial_mask) {
2014 mask = generate_quad_mask(gallivm, fs_type,
2015 i*fs_type.length/4, mask_input);
2016 }
2017 else {
2018 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2019 }
2020 LLVMBuildStore(builder, mask, mask_ptr);
2021 }
2022
2023 generate_fs_loop(gallivm,
2024 shader, key,
2025 builder,
2026 fs_type,
2027 context_ptr,
2028 num_loop,
2029 &interp,
2030 sampler,
2031 mask_store, /* output */
2032 color_store,
2033 depth_ptr,
2034 depth_bits,
2035 facing,
2036 counter);
2037
2038 for (i = 0; i < num_fs; i++) {
2039 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2040 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2041 &indexi, 1, "");
2042 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2043 /* This is fucked up need to reorganize things */
2044 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2045 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2046 ptr = LLVMBuildGEP(builder,
2047 color_store[cbuf * !cbuf0_write_all][chan],
2048 &indexi, 1, "");
2049 fs_out_color[cbuf][chan][i] = ptr;
2050 }
2051 }
2052 }
2053 }
2054
2055 sampler->destroy(sampler);
2056
2057 /* Loop over color outputs / color buffers to do blending.
2058 */
2059 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2060 LLVMValueRef color_ptr;
2061 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2062 LLVMValueRef blend_in_color[TGSI_NUM_CHANNELS];
2063 unsigned rt = key->blend.independent_blend_enable ? cbuf : 0;
2064
2065 boolean do_branch = ((key->depth.enabled
2066 || key->stencil[0].enabled
2067 || key->alpha.enabled)
2068 && !shader->info.base.uses_kill);
2069
2070 color_ptr = LLVMBuildLoad(builder,
2071 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2072 "");
2073
2074 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2075
2076 if (variant->unswizzled_cbufs & (1 << cbuf)) {
2077 LLVMValueRef stride = LLVMBuildLoad(builder,
2078 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2079 "");
2080
2081 generate_unswizzled_blend(gallivm, rt, variant, key->cbuf_format[cbuf],
2082 num_fs, fs_type, fs_mask, fs_out_color[cbuf],
2083 context_ptr, color_ptr, stride, partial_mask, do_branch);
2084 } else {
2085 /*
2086 * Convert the fs's output color and mask to fit to the blending type.
2087 */
2088 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2089 LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
2090
2091 for (i = 0; i < num_fs; i++) {
2092 fs_color_vals[i] =
2093 LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
2094 }
2095
2096 lp_build_conv(gallivm, fs_type, blend_type,
2097 fs_color_vals,
2098 num_fs,
2099 &blend_in_color[chan], 1);
2100
2101 lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
2102 }
2103
2104 if (partial_mask || !variant->opaque) {
2105 lp_build_conv_mask(gallivm, fs_type, blend_type,
2106 fs_mask, num_fs,
2107 &blend_mask, 1);
2108 } else {
2109 blend_mask = lp_build_const_int_vec(gallivm, blend_type, ~0);
2110 }
2111
2112 generate_blend(gallivm,
2113 &key->blend,
2114 rt,
2115 builder,
2116 blend_type,
2117 context_ptr,
2118 blend_mask,
2119 blend_in_color,
2120 color_ptr,
2121 do_branch);
2122 }
2123 }
2124
2125 LLVMBuildRetVoid(builder);
2126
2127 gallivm_verify_function(gallivm, function);
2128
2129 variant->nr_instrs += lp_build_count_instructions(function);
2130 }
2131
2132
2133 static void
2134 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2135 {
2136 unsigned i;
2137
2138 debug_printf("fs variant %p:\n", (void *) key);
2139
2140 if (key->flatshade) {
2141 debug_printf("flatshade = 1\n");
2142 }
2143 for (i = 0; i < key->nr_cbufs; ++i) {
2144 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2145 }
2146 if (key->depth.enabled) {
2147 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2148 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2149 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2150 }
2151
2152 for (i = 0; i < 2; ++i) {
2153 if (key->stencil[i].enabled) {
2154 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2155 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2156 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2157 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2158 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2159 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2160 }
2161 }
2162
2163 if (key->alpha.enabled) {
2164 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2165 }
2166
2167 if (key->occlusion_count) {
2168 debug_printf("occlusion_count = 1\n");
2169 }
2170
2171 if (key->blend.logicop_enable) {
2172 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2173 }
2174 else if (key->blend.rt[0].blend_enable) {
2175 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2176 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2177 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2178 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2179 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2180 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2181 }
2182 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2183 for (i = 0; i < key->nr_samplers; ++i) {
2184 debug_printf("sampler[%u] = \n", i);
2185 debug_printf(" .format = %s\n",
2186 util_format_name(key->sampler[i].format));
2187 debug_printf(" .target = %s\n",
2188 util_dump_tex_target(key->sampler[i].target, TRUE));
2189 debug_printf(" .pot = %u %u %u\n",
2190 key->sampler[i].pot_width,
2191 key->sampler[i].pot_height,
2192 key->sampler[i].pot_depth);
2193 debug_printf(" .wrap = %s %s %s\n",
2194 util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
2195 util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
2196 util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
2197 debug_printf(" .min_img_filter = %s\n",
2198 util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
2199 debug_printf(" .min_mip_filter = %s\n",
2200 util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
2201 debug_printf(" .mag_img_filter = %s\n",
2202 util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
2203 if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
2204 debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
2205 debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
2206 debug_printf(" .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
2207 debug_printf(" .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
2208 debug_printf(" .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
2209 debug_printf(" .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
2210 }
2211 }
2212
2213
2214 void
2215 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2216 {
2217 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2218 variant->shader->no, variant->no);
2219 tgsi_dump(variant->shader->base.tokens, 0);
2220 dump_fs_variant_key(&variant->key);
2221 debug_printf("variant->opaque = %u\n", variant->opaque);
2222 debug_printf("\n");
2223 }
2224
2225
2226 /**
2227 * Generate a new fragment shader variant from the shader code and
2228 * other state indicated by the key.
2229 */
2230 static struct lp_fragment_shader_variant *
2231 generate_variant(struct llvmpipe_context *lp,
2232 struct lp_fragment_shader *shader,
2233 const struct lp_fragment_shader_variant_key *key)
2234 {
2235 struct lp_fragment_shader_variant *variant;
2236 const struct util_format_description *cbuf0_format_desc;
2237 boolean fullcolormask;
2238 unsigned i;
2239
2240 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2241 if(!variant)
2242 return NULL;
2243
2244 variant->gallivm = gallivm_create();
2245 if (!variant->gallivm) {
2246 FREE(variant);
2247 return NULL;
2248 }
2249
2250 variant->shader = shader;
2251 variant->list_item_global.base = variant;
2252 variant->list_item_local.base = variant;
2253 variant->no = shader->variants_created++;
2254
2255 memcpy(&variant->key, key, shader->variant_key_size);
2256
2257 /*
2258 * Determine whether we are touching all channels in the color buffer.
2259 */
2260 fullcolormask = FALSE;
2261 if (key->nr_cbufs == 1) {
2262 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2263 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2264 }
2265
2266 variant->opaque =
2267 !key->blend.logicop_enable &&
2268 !key->blend.rt[0].blend_enable &&
2269 fullcolormask &&
2270 !key->stencil[0].enabled &&
2271 !key->alpha.enabled &&
2272 !key->depth.enabled &&
2273 !shader->info.base.uses_kill
2274 ? TRUE : FALSE;
2275
2276 for (i = 0; i < key->nr_cbufs; ++i) {
2277 variant->unswizzled_cbufs |= llvmpipe_is_format_unswizzled(key->cbuf_format[i]) << i;
2278 }
2279
2280 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2281 lp_debug_fs_variant(variant);
2282 }
2283
2284 lp_jit_init_types(variant);
2285
2286 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2287 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2288
2289 if (variant->jit_function[RAST_WHOLE] == NULL) {
2290 if (variant->opaque) {
2291 /* Specialized shader, which doesn't need to read the color buffer. */
2292 generate_fragment(lp, shader, variant, RAST_WHOLE);
2293 }
2294 }
2295
2296 /*
2297 * Compile everything
2298 */
2299
2300 gallivm_compile_module(variant->gallivm);
2301
2302 if (variant->function[RAST_EDGE_TEST]) {
2303 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2304 gallivm_jit_function(variant->gallivm,
2305 variant->function[RAST_EDGE_TEST]);
2306 }
2307
2308 if (variant->function[RAST_WHOLE]) {
2309 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2310 gallivm_jit_function(variant->gallivm,
2311 variant->function[RAST_WHOLE]);
2312 } else if (!variant->jit_function[RAST_WHOLE]) {
2313 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2314 }
2315
2316 return variant;
2317 }
2318
2319
2320 static void *
2321 llvmpipe_create_fs_state(struct pipe_context *pipe,
2322 const struct pipe_shader_state *templ)
2323 {
2324 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2325 struct lp_fragment_shader *shader;
2326 int nr_samplers;
2327 int i;
2328
2329 shader = CALLOC_STRUCT(lp_fragment_shader);
2330 if (!shader)
2331 return NULL;
2332
2333 shader->no = fs_no++;
2334 make_empty_list(&shader->variants);
2335
2336 /* get/save the summary info for this shader */
2337 lp_build_tgsi_info(templ->tokens, &shader->info);
2338
2339 /* we need to keep a local copy of the tokens */
2340 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2341
2342 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2343 if (shader->draw_data == NULL) {
2344 FREE((void *) shader->base.tokens);
2345 FREE(shader);
2346 return NULL;
2347 }
2348
2349 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2350
2351 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2352 sampler[nr_samplers]);
2353
2354 for (i = 0; i < shader->info.base.num_inputs; i++) {
2355 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2356 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2357
2358 switch (shader->info.base.input_interpolate[i]) {
2359 case TGSI_INTERPOLATE_CONSTANT:
2360 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2361 break;
2362 case TGSI_INTERPOLATE_LINEAR:
2363 shader->inputs[i].interp = LP_INTERP_LINEAR;
2364 break;
2365 case TGSI_INTERPOLATE_PERSPECTIVE:
2366 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2367 break;
2368 case TGSI_INTERPOLATE_COLOR:
2369 shader->inputs[i].interp = LP_INTERP_COLOR;
2370 break;
2371 default:
2372 assert(0);
2373 break;
2374 }
2375
2376 switch (shader->info.base.input_semantic_name[i]) {
2377 case TGSI_SEMANTIC_FACE:
2378 shader->inputs[i].interp = LP_INTERP_FACING;
2379 break;
2380 case TGSI_SEMANTIC_POSITION:
2381 /* Position was already emitted above
2382 */
2383 shader->inputs[i].interp = LP_INTERP_POSITION;
2384 shader->inputs[i].src_index = 0;
2385 continue;
2386 }
2387
2388 shader->inputs[i].src_index = i+1;
2389 }
2390
2391 if (LP_DEBUG & DEBUG_TGSI) {
2392 unsigned attrib;
2393 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2394 shader->no, (void *) shader);
2395 tgsi_dump(templ->tokens, 0);
2396 debug_printf("usage masks:\n");
2397 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2398 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2399 debug_printf(" IN[%u].%s%s%s%s\n",
2400 attrib,
2401 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2402 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2403 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2404 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2405 }
2406 debug_printf("\n");
2407 }
2408
2409 return shader;
2410 }
2411
2412
2413 static void
2414 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2415 {
2416 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2417
2418 if (llvmpipe->fs == fs)
2419 return;
2420
2421 draw_flush(llvmpipe->draw);
2422
2423 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2424
2425 draw_bind_fragment_shader(llvmpipe->draw,
2426 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2427
2428 llvmpipe->dirty |= LP_NEW_FS;
2429 }
2430
2431
2432 /**
2433 * Remove shader variant from two lists: the shader's variant list
2434 * and the context's variant list.
2435 */
2436 void
2437 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2438 struct lp_fragment_shader_variant *variant)
2439 {
2440 unsigned i;
2441
2442 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2443 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2444 " #%u v total cached #%u\n",
2445 variant->shader->no,
2446 variant->no,
2447 variant->shader->variants_created,
2448 variant->shader->variants_cached,
2449 lp->nr_fs_variants);
2450 }
2451
2452 /* free all the variant's JIT'd functions */
2453 for (i = 0; i < Elements(variant->function); i++) {
2454 if (variant->function[i]) {
2455 gallivm_free_function(variant->gallivm,
2456 variant->function[i],
2457 variant->jit_function[i]);
2458 }
2459 }
2460
2461 gallivm_destroy(variant->gallivm);
2462
2463 /* remove from shader's list */
2464 remove_from_list(&variant->list_item_local);
2465 variant->shader->variants_cached--;
2466
2467 /* remove from context's list */
2468 remove_from_list(&variant->list_item_global);
2469 lp->nr_fs_variants--;
2470 lp->nr_fs_instrs -= variant->nr_instrs;
2471
2472 FREE(variant);
2473 }
2474
2475
2476 static void
2477 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2478 {
2479 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2480 struct lp_fragment_shader *shader = fs;
2481 struct lp_fs_variant_list_item *li;
2482
2483 assert(fs != llvmpipe->fs);
2484
2485 /*
2486 * XXX: we need to flush the context until we have some sort of reference
2487 * counting in fragment shaders as they may still be binned
2488 * Flushing alone might not sufficient we need to wait on it too.
2489 */
2490 llvmpipe_finish(pipe, __FUNCTION__);
2491
2492 /* Delete all the variants */
2493 li = first_elem(&shader->variants);
2494 while(!at_end(&shader->variants, li)) {
2495 struct lp_fs_variant_list_item *next = next_elem(li);
2496 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2497 li = next;
2498 }
2499
2500 /* Delete draw module's data */
2501 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2502
2503 assert(shader->variants_cached == 0);
2504 FREE((void *) shader->base.tokens);
2505 FREE(shader);
2506 }
2507
2508
2509
2510 static void
2511 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2512 uint shader, uint index,
2513 struct pipe_constant_buffer *cb)
2514 {
2515 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2516 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2517 unsigned size;
2518 const void *data;
2519
2520 if (cb && cb->user_buffer) {
2521 constants = llvmpipe_user_buffer_create(pipe->screen,
2522 (void *) cb->user_buffer,
2523 cb->buffer_size,
2524 PIPE_BIND_CONSTANT_BUFFER);
2525 }
2526
2527 size = constants ? constants->width0 : 0;
2528 data = constants ? llvmpipe_resource_data(constants) : NULL;
2529
2530 assert(shader < PIPE_SHADER_TYPES);
2531 assert(index < PIPE_MAX_CONSTANT_BUFFERS);
2532
2533 if(llvmpipe->constants[shader][index] == constants)
2534 return;
2535
2536 draw_flush(llvmpipe->draw);
2537
2538 /* note: reference counting */
2539 pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
2540
2541 if(shader == PIPE_SHADER_VERTEX ||
2542 shader == PIPE_SHADER_GEOMETRY) {
2543 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2544 index, data, size);
2545 }
2546
2547 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2548
2549 if (cb && cb->user_buffer) {
2550 pipe_resource_reference(&constants, NULL);
2551 }
2552 }
2553
2554
2555 /**
2556 * Return the blend factor equivalent to a destination alpha of one.
2557 */
2558 static INLINE unsigned
2559 force_dst_alpha_one(unsigned factor)
2560 {
2561 switch(factor) {
2562 case PIPE_BLENDFACTOR_DST_ALPHA:
2563 return PIPE_BLENDFACTOR_ONE;
2564 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2565 return PIPE_BLENDFACTOR_ZERO;
2566 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2567 return PIPE_BLENDFACTOR_ZERO;
2568 }
2569
2570 return factor;
2571 }
2572
2573
2574 /**
2575 * We need to generate several variants of the fragment pipeline to match
2576 * all the combinations of the contributing state atoms.
2577 *
2578 * TODO: there is actually no reason to tie this to context state -- the
2579 * generated code could be cached globally in the screen.
2580 */
2581 static void
2582 make_variant_key(struct llvmpipe_context *lp,
2583 struct lp_fragment_shader *shader,
2584 struct lp_fragment_shader_variant_key *key)
2585 {
2586 unsigned i;
2587
2588 memset(key, 0, shader->variant_key_size);
2589
2590 if (lp->framebuffer.zsbuf) {
2591 if (lp->depth_stencil->depth.enabled) {
2592 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2593 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2594 }
2595 if (lp->depth_stencil->stencil[0].enabled) {
2596 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2597 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2598 }
2599 }
2600
2601 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2602 if(key->alpha.enabled)
2603 key->alpha.func = lp->depth_stencil->alpha.func;
2604 /* alpha.ref_value is passed in jit_context */
2605
2606 key->flatshade = lp->rasterizer->flatshade;
2607 if (lp->active_query_count) {
2608 key->occlusion_count = TRUE;
2609 }
2610
2611 if (lp->framebuffer.nr_cbufs) {
2612 memcpy(&key->blend, lp->blend, sizeof key->blend);
2613 }
2614
2615 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2616 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2617 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2618 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2619 const struct util_format_description *format_desc;
2620
2621 key->cbuf_format[i] = format;
2622
2623 format_desc = util_format_description(format);
2624 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2625 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2626
2627 blend_rt->colormask = lp->blend->rt[i].colormask;
2628
2629 /*
2630 * Mask out color channels not present in the color buffer.
2631 */
2632 blend_rt->colormask &= util_format_colormask(format_desc);
2633
2634 /*
2635 * Our swizzled render tiles always have an alpha channel, but the linear
2636 * render target format often does not, so force here the dst alpha to be
2637 * one.
2638 *
2639 * This is not a mere optimization. Wrong results will be produced if the
2640 * dst alpha is used, the dst format does not have alpha, and the previous
2641 * rendering was not flushed from the swizzled to linear buffer. For
2642 * example, NonPowTwo DCT.
2643 *
2644 * TODO: This should be generalized to all channels for better
2645 * performance, but only alpha causes correctness issues.
2646 *
2647 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2648 */
2649 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2650 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2651 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2652 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2653 blend_rt->alpha_func = blend_rt->rgb_func;
2654 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2655 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2656 }
2657 }
2658
2659 /* This value will be the same for all the variants of a given shader:
2660 */
2661 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2662
2663 for(i = 0; i < key->nr_samplers; ++i) {
2664 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2665 lp_sampler_static_state(&key->sampler[i],
2666 lp->sampler_views[PIPE_SHADER_FRAGMENT][i],
2667 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2668 }
2669 }
2670 }
2671
2672
2673
2674 /**
2675 * Update fragment shader state. This is called just prior to drawing
2676 * something when some fragment-related state has changed.
2677 */
2678 void
2679 llvmpipe_update_fs(struct llvmpipe_context *lp)
2680 {
2681 struct lp_fragment_shader *shader = lp->fs;
2682 struct lp_fragment_shader_variant_key key;
2683 struct lp_fragment_shader_variant *variant = NULL;
2684 struct lp_fs_variant_list_item *li;
2685
2686 make_variant_key(lp, shader, &key);
2687
2688 /* Search the variants for one which matches the key */
2689 li = first_elem(&shader->variants);
2690 while(!at_end(&shader->variants, li)) {
2691 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2692 variant = li->base;
2693 break;
2694 }
2695 li = next_elem(li);
2696 }
2697
2698 if (variant) {
2699 /* Move this variant to the head of the list to implement LRU
2700 * deletion of shader's when we have too many.
2701 */
2702 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2703 }
2704 else {
2705 /* variant not found, create it now */
2706 int64_t t0, t1, dt;
2707 unsigned i;
2708 unsigned variants_to_cull;
2709
2710 if (0) {
2711 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2712 lp->nr_fs_variants,
2713 lp->nr_fs_instrs,
2714 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2715 }
2716
2717 /* First, check if we've exceeded the max number of shader variants.
2718 * If so, free 25% of them (the least recently used ones).
2719 */
2720 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2721
2722 if (variants_to_cull ||
2723 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2724 struct pipe_context *pipe = &lp->pipe;
2725
2726 /*
2727 * XXX: we need to flush the context until we have some sort of
2728 * reference counting in fragment shaders as they may still be binned
2729 * Flushing alone might not be sufficient we need to wait on it too.
2730 */
2731 llvmpipe_finish(pipe, __FUNCTION__);
2732
2733 /*
2734 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2735 * number of shader variants (potentially all of them) could be
2736 * pending for destruction on flush.
2737 */
2738
2739 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2740 struct lp_fs_variant_list_item *item;
2741 if (is_empty_list(&lp->fs_variants_list)) {
2742 break;
2743 }
2744 item = last_elem(&lp->fs_variants_list);
2745 assert(item);
2746 assert(item->base);
2747 llvmpipe_remove_shader_variant(lp, item->base);
2748 }
2749 }
2750
2751 /*
2752 * Generate the new variant.
2753 */
2754 t0 = os_time_get();
2755 variant = generate_variant(lp, shader, &key);
2756 t1 = os_time_get();
2757 dt = t1 - t0;
2758 LP_COUNT_ADD(llvm_compile_time, dt);
2759 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2760
2761 llvmpipe_variant_count++;
2762
2763 /* Put the new variant into the list */
2764 if (variant) {
2765 insert_at_head(&shader->variants, &variant->list_item_local);
2766 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2767 lp->nr_fs_variants++;
2768 lp->nr_fs_instrs += variant->nr_instrs;
2769 shader->variants_cached++;
2770 }
2771 }
2772
2773 /* Bind this variant */
2774 lp_setup_set_fs_variant(lp->setup, variant);
2775 }
2776
2777
2778
2779
2780
2781
2782
2783 void
2784 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2785 {
2786 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2787 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2788 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2789
2790 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2791 }