projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: Fix rendering into PIPE_FORMAT_X8*_UNORM.
[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 "util/u_dual_blend.h"
70 #include "os/os_time.h"
71 #include "pipe/p_shader_tokens.h"
72 #include "draw/draw_context.h"
73 #include "tgsi/tgsi_dump.h"
74 #include "tgsi/tgsi_scan.h"
75 #include "tgsi/tgsi_parse.h"
76 #include "gallivm/lp_bld_type.h"
77 #include "gallivm/lp_bld_const.h"
78 #include "gallivm/lp_bld_conv.h"
79 #include "gallivm/lp_bld_init.h"
80 #include "gallivm/lp_bld_intr.h"
81 #include "gallivm/lp_bld_logic.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_swizzle.h"
84 #include "gallivm/lp_bld_flow.h"
85 #include "gallivm/lp_bld_debug.h"
86 #include "gallivm/lp_bld_arit.h"
87 #include "gallivm/lp_bld_pack.h"
88 #include "gallivm/lp_bld_format.h"
89 #include "gallivm/lp_bld_quad.h"
90
91 #include "lp_bld_alpha.h"
92 #include "lp_bld_blend.h"
93 #include "lp_bld_depth.h"
94 #include "lp_bld_interp.h"
95 #include "lp_context.h"
96 #include "lp_debug.h"
97 #include "lp_perf.h"
98 #include "lp_setup.h"
99 #include "lp_state.h"
100 #include "lp_tex_sample.h"
101 #include "lp_flush.h"
102 #include "lp_state_fs.h"
103
104
105 /** Fragment shader number (for debugging) */
106 static unsigned fs_no = 0;
107
108
109 /**
110 * Expand the relevant bits of mask_input to a n*4-dword mask for the
111 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
112 * quad mask vector to 0 or ~0.
113 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
114 * quad arguments with fs length 8.
115 *
116 * \param first_quad which quad(s) of the quad group to test, in [0,3]
117 * \param mask_input bitwise mask for the whole 4x4 stamp
118 */
119 static LLVMValueRef
120 generate_quad_mask(struct gallivm_state *gallivm,
121 struct lp_type fs_type,
122 unsigned first_quad,
123 LLVMValueRef mask_input) /* int32 */
124 {
125 LLVMBuilderRef builder = gallivm->builder;
126 struct lp_type mask_type;
127 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
128 LLVMValueRef bits[16];
129 LLVMValueRef mask;
130 int shift, i;
131
132 /*
133 * XXX: We'll need a different path for 16 x u8
134 */
135 assert(fs_type.width == 32);
136 assert(fs_type.length <= Elements(bits));
137 mask_type = lp_int_type(fs_type);
138
139 /*
140 * mask_input >>= (quad * 4)
141 */
142 switch (first_quad) {
143 case 0:
144 shift = 0;
145 break;
146 case 1:
147 assert(fs_type.length == 4);
148 shift = 2;
149 break;
150 case 2:
151 shift = 8;
152 break;
153 case 3:
154 assert(fs_type.length == 4);
155 shift = 10;
156 break;
157 default:
158 assert(0);
159 shift = 0;
160 }
161
162 mask_input = LLVMBuildLShr(builder,
163 mask_input,
164 LLVMConstInt(i32t, shift, 0),
165 "");
166
167 /*
168 * mask = { mask_input & (1 << i), for i in [0,3] }
169 */
170 mask = lp_build_broadcast(gallivm,
171 lp_build_vec_type(gallivm, mask_type),
172 mask_input);
173
174 for (i = 0; i < fs_type.length / 4; i++) {
175 unsigned j = 2 * (i % 2) + (i / 2) * 8;
176 bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
177 bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
178 bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
179 bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
180 }
181 mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
182
183 /*
184 * mask = mask != 0 ? ~0 : 0
185 */
186 mask = lp_build_compare(gallivm,
187 mask_type, PIPE_FUNC_NOTEQUAL,
188 mask,
189 lp_build_const_int_vec(gallivm, mask_type, 0));
190
191 return mask;
192 }
193
194
195 #define EARLY_DEPTH_TEST 0x1
196 #define LATE_DEPTH_TEST 0x2
197 #define EARLY_DEPTH_WRITE 0x4
198 #define LATE_DEPTH_WRITE 0x8
199
200 static int
201 find_output_by_semantic( const struct tgsi_shader_info *info,
202 unsigned semantic,
203 unsigned index )
204 {
205 int i;
206
207 for (i = 0; i < info->num_outputs; i++)
208 if (info->output_semantic_name[i] == semantic &&
209 info->output_semantic_index[i] == index)
210 return i;
211
212 return -1;
213 }
214
215
216 /**
217 * Generate the fragment shader, depth/stencil test, and alpha tests.
218 * \param i which quad in the tile, in range [0,3]
219 * \param partial_mask if 1, do mask_input testing
220 */
221 static void
222 generate_fs(struct gallivm_state *gallivm,
223 struct lp_fragment_shader *shader,
224 const struct lp_fragment_shader_variant_key *key,
225 LLVMBuilderRef builder,
226 struct lp_type type,
227 LLVMValueRef context_ptr,
228 unsigned i,
229 struct lp_build_interp_soa_context *interp,
230 struct lp_build_sampler_soa *sampler,
231 LLVMValueRef *pmask,
232 LLVMValueRef (*color)[4],
233 LLVMValueRef depth_ptr,
234 LLVMValueRef facing,
235 unsigned partial_mask,
236 LLVMValueRef mask_input,
237 LLVMValueRef thread_data_ptr)
238 {
239 const struct util_format_description *zs_format_desc = NULL;
240 const struct tgsi_token *tokens = shader->base.tokens;
241 LLVMTypeRef vec_type;
242 LLVMValueRef consts_ptr;
243 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
244 LLVMValueRef z;
245 LLVMValueRef zs_value = NULL;
246 LLVMValueRef stencil_refs[2];
247 struct lp_build_mask_context mask;
248 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
249 shader->info.base.num_inputs < 3 &&
250 shader->info.base.num_instructions < 8);
251 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
252 util_blend_state_is_dual(&key->blend, 0);
253 unsigned attrib;
254 unsigned chan;
255 unsigned cbuf;
256 unsigned depth_mode;
257 struct lp_bld_tgsi_system_values system_values;
258
259 memset(&system_values, 0, sizeof(system_values));
260
261 if (key->depth.enabled ||
262 key->stencil[0].enabled ||
263 key->stencil[1].enabled) {
264
265 zs_format_desc = util_format_description(key->zsbuf_format);
266 assert(zs_format_desc);
267
268 if (!shader->info.base.writes_z) {
269 if (key->alpha.enabled || shader->info.base.uses_kill)
270 /* With alpha test and kill, can do the depth test early
271 * and hopefully eliminate some quads. But need to do a
272 * special deferred depth write once the final mask value
273 * is known.
274 */
275 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
276 else
277 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
278 }
279 else {
280 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
281 }
282
283 if (!(key->depth.enabled && key->depth.writemask) &&
284 !(key->stencil[0].enabled && key->stencil[0].writemask))
285 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
286 }
287 else {
288 depth_mode = 0;
289 }
290
291 assert(i < 4);
292
293 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
294 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
295
296 vec_type = lp_build_vec_type(gallivm, type);
297
298 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
299
300 memset(outputs, 0, sizeof outputs);
301
302 /* Declare the color and z variables */
303 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
304 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
305 color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
306 }
307 }
308 if (dual_source_blend) {
309 assert(key->nr_cbufs <= 1);
310 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
311 color[1][chan] = lp_build_alloca(gallivm, vec_type, "color1");
312 }
313 }
314
315 /* do triangle edge testing */
316 if (partial_mask) {
317 *pmask = generate_quad_mask(gallivm, type,
318 i*type.length/4, mask_input);
319 }
320 else {
321 *pmask = lp_build_const_int_vec(gallivm, type, ~0);
322 }
323
324 /* 'mask' will control execution based on quad's pixel alive/killed state */
325 lp_build_mask_begin(&mask, gallivm, type, *pmask);
326
327 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
328 lp_build_mask_check(&mask);
329
330 lp_build_interp_soa_update_pos(interp, gallivm, i*type.length/4);
331 z = interp->pos[2];
332
333 if (depth_mode & EARLY_DEPTH_TEST) {
334 lp_build_depth_stencil_test(gallivm,
335 &key->depth,
336 key->stencil,
337 type,
338 zs_format_desc,
339 &mask,
340 stencil_refs,
341 z,
342 depth_ptr, facing,
343 &zs_value,
344 !simple_shader);
345
346 if (depth_mode & EARLY_DEPTH_WRITE) {
347 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr, zs_value);
348 }
349 }
350
351 lp_build_interp_soa_update_inputs(interp, gallivm, i*type.length/4);
352
353 /* Build the actual shader */
354 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
355 consts_ptr, &system_values,
356 interp->pos, interp->inputs,
357 outputs, sampler, &shader->info.base);
358
359 /* Alpha test */
360 if (key->alpha.enabled) {
361 int color0 = find_output_by_semantic(&shader->info.base,
362 TGSI_SEMANTIC_COLOR,
363 0);
364
365 if (color0 != -1 && outputs[color0][3]) {
366 const struct util_format_description *cbuf_format_desc;
367 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
368 LLVMValueRef alpha_ref_value;
369
370 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
371 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
372
373 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
374
375 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
376 &mask, alpha, alpha_ref_value,
377 (depth_mode & LATE_DEPTH_TEST) != 0);
378 }
379 }
380
381 /* Late Z test */
382 if (depth_mode & LATE_DEPTH_TEST) {
383 int pos0 = find_output_by_semantic(&shader->info.base,
384 TGSI_SEMANTIC_POSITION,
385 0);
386
387 if (pos0 != -1 && outputs[pos0][2]) {
388 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
389 }
390
391 lp_build_depth_stencil_test(gallivm,
392 &key->depth,
393 key->stencil,
394 type,
395 zs_format_desc,
396 &mask,
397 stencil_refs,
398 z,
399 depth_ptr, facing,
400 &zs_value,
401 !simple_shader);
402 /* Late Z write */
403 if (depth_mode & LATE_DEPTH_WRITE) {
404 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr, zs_value);
405 }
406 }
407 else if ((depth_mode & EARLY_DEPTH_TEST) &&
408 (depth_mode & LATE_DEPTH_WRITE))
409 {
410 /* Need to apply a reduced mask to the depth write. Reload the
411 * depth value, update from zs_value with the new mask value and
412 * write that out.
413 */
414 lp_build_deferred_depth_write(gallivm,
415 type,
416 zs_format_desc,
417 &mask,
418 depth_ptr,
419 zs_value);
420 }
421
422
423 /* Color write */
424 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
425 {
426 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
427 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
428 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
429 {
430 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
431 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
432 if(outputs[attrib][chan]) {
433 /* XXX: just initialize outputs to point at colors[] and
434 * skip this.
435 */
436 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
437 lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
438 LLVMBuildStore(builder, out, color[cbuf][chan]);
439 }
440 }
441 }
442 }
443
444 if (key->occlusion_count) {
445 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
446 lp_build_name(counter, "counter");
447 lp_build_occlusion_count(gallivm, type,
448 lp_build_mask_value(&mask), counter);
449 }
450
451 *pmask = lp_build_mask_end(&mask);
452 }
453
454
455 /**
456 * Generate the fragment shader, depth/stencil test, and alpha tests.
457 */
458 static void
459 generate_fs_loop(struct gallivm_state *gallivm,
460 struct lp_fragment_shader *shader,
461 const struct lp_fragment_shader_variant_key *key,
462 LLVMBuilderRef builder,
463 struct lp_type type,
464 LLVMValueRef context_ptr,
465 LLVMValueRef num_loop,
466 struct lp_build_interp_soa_context *interp,
467 struct lp_build_sampler_soa *sampler,
468 LLVMValueRef mask_store,
469 LLVMValueRef (*out_color)[4],
470 LLVMValueRef depth_ptr,
471 unsigned depth_bits,
472 LLVMValueRef facing,
473 LLVMValueRef thread_data_ptr)
474 {
475 const struct util_format_description *zs_format_desc = NULL;
476 const struct tgsi_token *tokens = shader->base.tokens;
477 LLVMTypeRef vec_type;
478 LLVMValueRef mask_ptr, mask_val;
479 LLVMValueRef consts_ptr;
480 LLVMValueRef z;
481 LLVMValueRef zs_value = NULL;
482 LLVMValueRef stencil_refs[2];
483 LLVMValueRef depth_ptr_i;
484 LLVMValueRef depth_offset;
485 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
486 struct lp_build_for_loop_state loop_state;
487 struct lp_build_mask_context mask;
488 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
489 shader->info.base.num_inputs < 3 &&
490 shader->info.base.num_instructions < 8);
491 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
492 util_blend_state_is_dual(&key->blend, 0);
493 unsigned attrib;
494 unsigned chan;
495 unsigned cbuf;
496 unsigned depth_mode;
497
498 struct lp_bld_tgsi_system_values system_values;
499
500 memset(&system_values, 0, sizeof(system_values));
501
502 if (key->depth.enabled ||
503 key->stencil[0].enabled ||
504 key->stencil[1].enabled) {
505
506 zs_format_desc = util_format_description(key->zsbuf_format);
507 assert(zs_format_desc);
508
509 if (!shader->info.base.writes_z) {
510 if (key->alpha.enabled || shader->info.base.uses_kill)
511 /* With alpha test and kill, can do the depth test early
512 * and hopefully eliminate some quads. But need to do a
513 * special deferred depth write once the final mask value
514 * is known.
515 */
516 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
517 else
518 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
519 }
520 else {
521 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
522 }
523
524 if (!(key->depth.enabled && key->depth.writemask) &&
525 !(key->stencil[0].enabled && key->stencil[0].writemask))
526 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
527 }
528 else {
529 depth_mode = 0;
530 }
531
532
533 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
534 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
535
536 vec_type = lp_build_vec_type(gallivm, type);
537
538 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
539
540 lp_build_for_loop_begin(&loop_state, gallivm,
541 lp_build_const_int32(gallivm, 0),
542 LLVMIntULT,
543 num_loop,
544 lp_build_const_int32(gallivm, 1));
545
546 mask_ptr = LLVMBuildGEP(builder, mask_store,
547 &loop_state.counter, 1, "mask_ptr");
548 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
549
550 depth_offset = LLVMBuildMul(builder, loop_state.counter,
551 lp_build_const_int32(gallivm, depth_bits * type.length),
552 "");
553
554 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
555
556 memset(outputs, 0, sizeof outputs);
557
558 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
559 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
560 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
561 lp_build_vec_type(gallivm,
562 type),
563 num_loop, "color");
564 }
565 }
566 if (dual_source_blend) {
567 assert(key->nr_cbufs <= 1);
568 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
569 out_color[1][chan] = lp_build_array_alloca(gallivm,
570 lp_build_vec_type(gallivm,
571 type),
572 num_loop, "color1");
573 }
574 }
575
576
577 /* 'mask' will control execution based on quad's pixel alive/killed state */
578 lp_build_mask_begin(&mask, gallivm, type, mask_val);
579
580 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
581 lp_build_mask_check(&mask);
582
583 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
584 z = interp->pos[2];
585
586 if (depth_mode & EARLY_DEPTH_TEST) {
587 lp_build_depth_stencil_test(gallivm,
588 &key->depth,
589 key->stencil,
590 type,
591 zs_format_desc,
592 &mask,
593 stencil_refs,
594 z,
595 depth_ptr_i, facing,
596 &zs_value,
597 !simple_shader);
598
599 if (depth_mode & EARLY_DEPTH_WRITE) {
600 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
601 }
602 }
603
604 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
605
606 /* Build the actual shader */
607 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
608 consts_ptr, &system_values,
609 interp->pos, interp->inputs,
610 outputs, sampler, &shader->info.base);
611
612 /* Alpha test */
613 if (key->alpha.enabled) {
614 int color0 = find_output_by_semantic(&shader->info.base,
615 TGSI_SEMANTIC_COLOR,
616 0);
617
618 if (color0 != -1 && outputs[color0][3]) {
619 const struct util_format_description *cbuf_format_desc;
620 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
621 LLVMValueRef alpha_ref_value;
622
623 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
624 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
625
626 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
627
628 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
629 &mask, alpha, alpha_ref_value,
630 (depth_mode & LATE_DEPTH_TEST) != 0);
631 }
632 }
633
634 /* Late Z test */
635 if (depth_mode & LATE_DEPTH_TEST) {
636 int pos0 = find_output_by_semantic(&shader->info.base,
637 TGSI_SEMANTIC_POSITION,
638 0);
639
640 if (pos0 != -1 && outputs[pos0][2]) {
641 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
642 }
643
644 lp_build_depth_stencil_test(gallivm,
645 &key->depth,
646 key->stencil,
647 type,
648 zs_format_desc,
649 &mask,
650 stencil_refs,
651 z,
652 depth_ptr_i, facing,
653 &zs_value,
654 !simple_shader);
655 /* Late Z write */
656 if (depth_mode & LATE_DEPTH_WRITE) {
657 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
658 }
659 }
660 else if ((depth_mode & EARLY_DEPTH_TEST) &&
661 (depth_mode & LATE_DEPTH_WRITE))
662 {
663 /* Need to apply a reduced mask to the depth write. Reload the
664 * depth value, update from zs_value with the new mask value and
665 * write that out.
666 */
667 lp_build_deferred_depth_write(gallivm,
668 type,
669 zs_format_desc,
670 &mask,
671 depth_ptr_i,
672 zs_value);
673 }
674
675
676 /* Color write */
677 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
678 {
679 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
680 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
681 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
682 {
683 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
684 if(outputs[attrib][chan]) {
685 /* XXX: just initialize outputs to point at colors[] and
686 * skip this.
687 */
688 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
689 LLVMValueRef color_ptr;
690 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
691 &loop_state.counter, 1, "");
692 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
693 LLVMBuildStore(builder, out, color_ptr);
694 }
695 }
696 }
697 }
698
699 if (key->occlusion_count) {
700 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
701 lp_build_name(counter, "counter");
702 lp_build_occlusion_count(gallivm, type,
703 lp_build_mask_value(&mask), counter);
704 }
705
706 mask_val = lp_build_mask_end(&mask);
707 LLVMBuildStore(builder, mask_val, mask_ptr);
708 lp_build_for_loop_end(&loop_state);
709 }
710
711
712 /**
713 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
714 *
715 * Fragment Shader outputs pixels in small 2x2 blocks
716 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
717 *
718 * However in memory pixels are stored in rows
719 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
720 *
721 * @param type fragment shader type (4x or 8x float)
722 * @param num_fs number of fs_src
723 * @param dst_channels number of output channels
724 * @param fs_src output from fragment shader
725 * @param dst pointer to store result
726 * @param pad_inline is channel padding inline or at end of row
727 * @return the number of dsts
728 */
729 static int
730 generate_fs_twiddle(struct gallivm_state *gallivm,
731 struct lp_type type,
732 unsigned num_fs,
733 unsigned dst_channels,
734 LLVMValueRef fs_src[][4],
735 LLVMValueRef* dst,
736 bool pad_inline)
737 {
738 LLVMValueRef src[16];
739
740 bool swizzle_pad;
741 bool twiddle;
742 bool split;
743
744 unsigned pixels = num_fs == 4 ? 1 : 2;
745 unsigned reorder_group;
746 unsigned src_channels;
747 unsigned src_count;
748 unsigned i;
749
750 src_channels = dst_channels < 3 ? dst_channels : 4;
751 src_count = num_fs * src_channels;
752
753 assert(pixels == 2 || num_fs == 4);
754 assert(num_fs * src_channels <= Elements(src));
755
756 /*
757 * Transpose from SoA -> AoS
758 */
759 for (i = 0; i < num_fs; ++i) {
760 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
761 }
762
763 /*
764 * Pick transformation options
765 */
766 swizzle_pad = false;
767 twiddle = false;
768 split = false;
769 reorder_group = 0;
770
771 if (dst_channels == 1) {
772 twiddle = true;
773
774 if (pixels == 2) {
775 split = true;
776 }
777 } else if (dst_channels == 2) {
778 if (pixels == 1) {
779 reorder_group = 1;
780 }
781 } else if (dst_channels > 2) {
782 if (pixels == 1) {
783 reorder_group = 2;
784 } else {
785 twiddle = true;
786 }
787
788 if (!pad_inline && dst_channels == 3 && pixels > 1) {
789 swizzle_pad = true;
790 }
791 }
792
793 /*
794 * Split the src in half
795 */
796 if (split) {
797 for (i = num_fs; i > 0; --i) {
798 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
799 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
800 }
801
802 src_count *= 2;
803 type.length = 4;
804 }
805
806 /*
807 * Ensure pixels are in memory order
808 */
809 if (reorder_group) {
810 /* Twiddle pixels by reordering the array, e.g.:
811 *
812 * src_count = 8 -> 0 2 1 3 4 6 5 7
813 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
814 */
815 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
816
817 for (i = 0; i < src_count; ++i) {
818 unsigned group = i / reorder_group;
819 unsigned block = (group / 4) * 4 * reorder_group;
820 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
821 dst[i] = src[j];
822 }
823 } else if (twiddle) {
824 /* Twiddle pixels across elements of array */
825 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
826 } else {
827 /* Do nothing */
828 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
829 }
830
831 /*
832 * Moves any padding between pixels to the end
833 * e.g. RGBXRGBX -> RGBRGBXX
834 */
835 if (swizzle_pad) {
836 unsigned char swizzles[16];
837 unsigned elems = pixels * dst_channels;
838
839 for (i = 0; i < type.length; ++i) {
840 if (i < elems)
841 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
842 else
843 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
844 }
845
846 for (i = 0; i < src_count; ++i) {
847 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
848 }
849 }
850
851 return src_count;
852 }
853
854
855 /**
856 * Load an unswizzled block of pixels from memory
857 */
858 static void
859 load_unswizzled_block(struct gallivm_state *gallivm,
860 LLVMValueRef base_ptr,
861 LLVMValueRef stride,
862 unsigned block_width,
863 unsigned block_height,
864 LLVMValueRef* dst,
865 struct lp_type dst_type,
866 unsigned dst_count,
867 unsigned dst_alignment)
868 {
869 LLVMBuilderRef builder = gallivm->builder;
870 unsigned row_size = dst_count / block_height;
871 unsigned i;
872
873 /* Ensure block exactly fits into dst */
874 assert((block_width * block_height) % dst_count == 0);
875
876 for (i = 0; i < dst_count; ++i) {
877 unsigned x = i % row_size;
878 unsigned y = i / row_size;
879
880 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
881 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
882
883 LLVMValueRef gep[2];
884 LLVMValueRef dst_ptr;
885
886 gep[0] = lp_build_const_int32(gallivm, 0);
887 gep[1] = LLVMBuildAdd(builder, bx, by, "");
888
889 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
890 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
891
892 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
893
894 lp_set_load_alignment(dst[i], dst_alignment);
895 }
896 }
897
898
899 /**
900 * Store an unswizzled block of pixels to memory
901 */
902 static void
903 store_unswizzled_block(struct gallivm_state *gallivm,
904 LLVMValueRef base_ptr,
905 LLVMValueRef stride,
906 unsigned block_width,
907 unsigned block_height,
908 LLVMValueRef* src,
909 struct lp_type src_type,
910 unsigned src_count,
911 unsigned src_alignment)
912 {
913 LLVMBuilderRef builder = gallivm->builder;
914 unsigned row_size = src_count / block_height;
915 unsigned i;
916
917 /* Ensure src exactly fits into block */
918 assert((block_width * block_height) % src_count == 0);
919
920 for (i = 0; i < src_count; ++i) {
921 unsigned x = i % row_size;
922 unsigned y = i / row_size;
923
924 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
925 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
926
927 LLVMValueRef gep[2];
928 LLVMValueRef src_ptr;
929
930 gep[0] = lp_build_const_int32(gallivm, 0);
931 gep[1] = LLVMBuildAdd(builder, bx, by, "");
932
933 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
934 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
935
936 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
937
938 lp_set_store_alignment(src_ptr, src_alignment);
939 }
940 }
941
942
943 /**
944 * Checks if a format description is an arithmetic format
945 *
946 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
947 */
948 static INLINE boolean
949 is_arithmetic_format(const struct util_format_description *format_desc)
950 {
951 boolean arith = false;
952 unsigned i;
953
954 for (i = 0; i < format_desc->nr_channels; ++i) {
955 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
956 arith |= (format_desc->channel[i].size % 8) != 0;
957 }
958
959 return arith;
960 }
961
962
963 /**
964 * Retrieves the type representing the memory layout for a format
965 *
966 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
967 */
968 static INLINE void
969 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
970 struct lp_type* type)
971 {
972 unsigned i;
973 unsigned chan;
974
975 for (i = 0; i < 4; i++)
976 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
977 break;
978 chan = i;
979
980 memset(type, 0, sizeof(struct lp_type));
981 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
982 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
983 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
984 type->norm = format_desc->channel[chan].normalized;
985
986 if (is_arithmetic_format(format_desc)) {
987 type->width = 0;
988 type->length = 1;
989
990 for (i = 0; i < format_desc->nr_channels; ++i) {
991 type->width += format_desc->channel[i].size;
992 }
993 } else {
994 type->width = format_desc->channel[chan].size;
995 type->length = format_desc->nr_channels;
996 }
997 }
998
999
1000 /**
1001 * Retrieves the type for a format which is usable in the blending code.
1002 *
1003 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1004 */
1005 static INLINE void
1006 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1007 struct lp_type* type)
1008 {
1009 unsigned i;
1010 unsigned chan;
1011
1012 for (i = 0; i < 4; i++)
1013 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1014 break;
1015 chan = i;
1016
1017 memset(type, 0, sizeof(struct lp_type));
1018 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1019 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1020 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1021 type->norm = format_desc->channel[chan].normalized;
1022 type->width = format_desc->channel[chan].size;
1023 type->length = format_desc->nr_channels;
1024
1025 for (i = 1; i < format_desc->nr_channels; ++i) {
1026 if (format_desc->channel[i].size > type->width)
1027 type->width = format_desc->channel[i].size;
1028 }
1029
1030 if (type->floating) {
1031 type->width = 32;
1032 } else {
1033 if (type->width <= 8) {
1034 type->width = 8;
1035 } else if (type->width <= 16) {
1036 type->width = 16;
1037 } else {
1038 type->width = 32;
1039 }
1040 }
1041
1042 if (is_arithmetic_format(format_desc) && type->length == 3) {
1043 type->length = 4;
1044 }
1045 }
1046
1047
1048 /**
1049 * Scale a normalized value from src_bits to dst_bits
1050 */
1051 static INLINE LLVMValueRef
1052 scale_bits(struct gallivm_state *gallivm,
1053 int src_bits,
1054 int dst_bits,
1055 LLVMValueRef src,
1056 struct lp_type src_type)
1057 {
1058 LLVMBuilderRef builder = gallivm->builder;
1059 LLVMValueRef result = src;
1060
1061 if (dst_bits < src_bits) {
1062 /* Scale down by LShr */
1063 result = LLVMBuildLShr(builder,
1064 src,
1065 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
1066 "");
1067 } else if (dst_bits > src_bits) {
1068 /* Scale up bits */
1069 int db = dst_bits - src_bits;
1070
1071 /* Shift left by difference in bits */
1072 result = LLVMBuildShl(builder,
1073 src,
1074 lp_build_const_int_vec(gallivm, src_type, db),
1075 "");
1076
1077 if (db < src_bits) {
1078 /* Enough bits in src to fill the remainder */
1079 LLVMValueRef lower = LLVMBuildLShr(builder,
1080 src,
1081 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1082 "");
1083
1084 result = LLVMBuildOr(builder, result, lower, "");
1085 } else if (db > src_bits) {
1086 /* Need to repeatedly copy src bits to fill remainder in dst */
1087 unsigned n;
1088
1089 for (n = src_bits; n < dst_bits; n *= 2) {
1090 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1091
1092 result = LLVMBuildOr(builder,
1093 result,
1094 LLVMBuildLShr(builder, result, shuv, ""),
1095 "");
1096 }
1097 }
1098 }
1099
1100 return result;
1101 }
1102
1103
1104 /**
1105 * Convert from memory format to blending format
1106 *
1107 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1108 */
1109 static void
1110 convert_to_blend_type(struct gallivm_state *gallivm,
1111 const struct util_format_description *src_fmt,
1112 struct lp_type src_type,
1113 struct lp_type dst_type,
1114 LLVMValueRef* src, // and dst
1115 unsigned num_srcs)
1116 {
1117 LLVMValueRef *dst = src;
1118 LLVMBuilderRef builder = gallivm->builder;
1119 struct lp_type blend_type;
1120 struct lp_type mem_type;
1121 unsigned i, j, k;
1122 unsigned pixels = 16 / num_srcs;
1123 bool is_arith;
1124
1125 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1126 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1127
1128 /* Is the format arithmetic */
1129 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1130 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1131
1132 /* Pad if necessary */
1133 if (!is_arith && src_type.length < dst_type.length) {
1134 for (i = 0; i < num_srcs; ++i) {
1135 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1136 }
1137
1138 src_type.length = dst_type.length;
1139 }
1140
1141 /* Special case for half-floats */
1142 if (mem_type.width == 16 && mem_type.floating) {
1143 assert(blend_type.width == 32 && blend_type.floating);
1144 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1145 is_arith = false;
1146 }
1147
1148 if (!is_arith) {
1149 return;
1150 }
1151
1152 src_type.width = blend_type.width * blend_type.length;
1153 blend_type.length *= pixels;
1154 src_type.length *= pixels / (src_type.length / mem_type.length);
1155
1156 for (i = 0; i < num_srcs; ++i) {
1157 LLVMValueRef chans[4];
1158 LLVMValueRef res;
1159 unsigned sa = 0;
1160
1161 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1162
1163 for (j = 0; j < src_fmt->nr_channels; ++j) {
1164 unsigned mask = 0;
1165
1166 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1167 mask |= 1 << k;
1168 }
1169
1170 /* Extract bits from source */
1171 chans[j] = LLVMBuildLShr(builder,
1172 dst[i],
1173 lp_build_const_int_vec(gallivm, src_type, sa),
1174 "");
1175
1176 chans[j] = LLVMBuildAnd(builder,
1177 chans[j],
1178 lp_build_const_int_vec(gallivm, src_type, mask),
1179 "");
1180
1181 /* Scale bits */
1182 if (src_type.norm) {
1183 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1184 blend_type.width, chans[j], src_type);
1185 }
1186
1187 /* Insert bits into correct position */
1188 chans[j] = LLVMBuildShl(builder,
1189 chans[j],
1190 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1191 "");
1192
1193 sa += src_fmt->channel[j].size;
1194
1195 if (j == 0) {
1196 res = chans[j];
1197 } else {
1198 res = LLVMBuildOr(builder, res, chans[j], "");
1199 }
1200 }
1201
1202 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1203 }
1204 }
1205
1206
1207 /**
1208 * Convert from blending format to memory format
1209 *
1210 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1211 */
1212 static void
1213 convert_from_blend_type(struct gallivm_state *gallivm,
1214 const struct util_format_description *src_fmt,
1215 struct lp_type src_type,
1216 struct lp_type dst_type,
1217 LLVMValueRef* src, // and dst
1218 unsigned num_srcs)
1219 {
1220 LLVMValueRef* dst = src;
1221 unsigned i, j, k;
1222 struct lp_type mem_type;
1223 struct lp_type blend_type;
1224 LLVMBuilderRef builder = gallivm->builder;
1225 unsigned pixels = 16 / num_srcs;
1226 bool is_arith;
1227
1228 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1229 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1230
1231 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1232
1233 /* Special case for half-floats */
1234 if (mem_type.width == 16 && mem_type.floating) {
1235 int length = dst_type.length;
1236 assert(blend_type.width == 32 && blend_type.floating);
1237
1238 dst_type.length = src_type.length;
1239
1240 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1241
1242 dst_type.length = length;
1243 is_arith = false;
1244 }
1245
1246 /* Remove any padding */
1247 if (!is_arith && (src_type.length % mem_type.length)) {
1248 src_type.length -= (src_type.length % mem_type.length);
1249
1250 for (i = 0; i < num_srcs; ++i) {
1251 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1252 }
1253 }
1254
1255 /* No bit arithmetic to do */
1256 if (!is_arith) {
1257 return;
1258 }
1259
1260 src_type.length = pixels;
1261 src_type.width = blend_type.length * blend_type.width;
1262 dst_type.length = pixels;
1263
1264 for (i = 0; i < num_srcs; ++i) {
1265 LLVMValueRef chans[4];
1266 LLVMValueRef res;
1267 unsigned sa = 0;
1268
1269 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1270
1271 for (j = 0; j < src_fmt->nr_channels; ++j) {
1272 unsigned mask = 0;
1273
1274 assert(blend_type.width > src_fmt->channel[j].size);
1275
1276 for (k = 0; k < blend_type.width; ++k) {
1277 mask |= 1 << k;
1278 }
1279
1280 /* Extract bits */
1281 chans[j] = LLVMBuildLShr(builder,
1282 dst[i],
1283 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1284 "");
1285
1286 chans[j] = LLVMBuildAnd(builder,
1287 chans[j],
1288 lp_build_const_int_vec(gallivm, src_type, mask),
1289 "");
1290
1291 /* Scale down bits */
1292 if (src_type.norm) {
1293 chans[j] = scale_bits(gallivm, blend_type.width,
1294 src_fmt->channel[j].size, chans[j], src_type);
1295 }
1296
1297 /* Insert bits */
1298 chans[j] = LLVMBuildShl(builder,
1299 chans[j],
1300 lp_build_const_int_vec(gallivm, src_type, sa),
1301 "");
1302
1303 sa += src_fmt->channel[j].size;
1304
1305 if (j == 0) {
1306 res = chans[j];
1307 } else {
1308 res = LLVMBuildOr(builder, res, chans[j], "");
1309 }
1310 }
1311
1312 assert (dst_type.width != 24);
1313
1314 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1315 }
1316 }
1317
1318
1319 /**
1320 * Convert alpha to same blend type as src
1321 */
1322 static void
1323 convert_alpha(struct gallivm_state *gallivm,
1324 struct lp_type row_type,
1325 struct lp_type alpha_type,
1326 const unsigned block_size,
1327 const unsigned block_height,
1328 const unsigned src_count,
1329 const unsigned dst_channels,
1330 const bool pad_inline,
1331 LLVMValueRef* src_alpha)
1332 {
1333 LLVMBuilderRef builder = gallivm->builder;
1334 unsigned i, j;
1335 unsigned length = row_type.length;
1336 row_type.length = alpha_type.length;
1337
1338 /* Twiddle the alpha to match pixels */
1339 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1340
1341 for (i = 0; i < 4; ++i) {
1342 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1343 }
1344
1345 alpha_type = row_type;
1346 row_type.length = length;
1347
1348 /* If only one channel we can only need the single alpha value per pixel */
1349 if (src_count == 1) {
1350 assert(dst_channels == 1);
1351
1352 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1353 } else {
1354 /* If there are more srcs than rows then we need to split alpha up */
1355 if (src_count > block_height) {
1356 for (i = src_count; i > 0; --i) {
1357 unsigned pixels = block_size / src_count;
1358 unsigned idx = i - 1;
1359
1360 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1361 }
1362 }
1363
1364 /* If there is a src for each pixel broadcast the alpha across whole row */
1365 if (src_count == block_size) {
1366 for (i = 0; i < src_count; ++i) {
1367 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1368 }
1369 } else {
1370 unsigned pixels = block_size / src_count;
1371 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1372 unsigned alpha_span = 1;
1373 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1374
1375 /* Check if we need 2 src_alphas for our shuffles */
1376 if (pixels > alpha_type.length) {
1377 alpha_span = 2;
1378 }
1379
1380 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1381 for (j = 0; j < row_type.length; ++j) {
1382 if (j < pixels * channels) {
1383 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1384 } else {
1385 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1386 }
1387 }
1388
1389 for (i = 0; i < src_count; ++i) {
1390 unsigned idx1 = i, idx2 = i;
1391
1392 if (alpha_span > 1){
1393 idx1 *= alpha_span;
1394 idx2 = idx1 + 1;
1395 }
1396
1397 src_alpha[i] = LLVMBuildShuffleVector(builder,
1398 src_alpha[idx1],
1399 src_alpha[idx2],
1400 LLVMConstVector(shuffles, row_type.length),
1401 "");
1402 }
1403 }
1404 }
1405 }
1406
1407
1408 /**
1409 * Generates the blend function for unswizzled colour buffers
1410 * Also generates the read & write from colour buffer
1411 */
1412 static void
1413 generate_unswizzled_blend(struct gallivm_state *gallivm,
1414 unsigned rt,
1415 struct lp_fragment_shader_variant *variant,
1416 enum pipe_format out_format,
1417 unsigned int num_fs,
1418 struct lp_type fs_type,
1419 LLVMValueRef* fs_mask,
1420 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1421 LLVMValueRef context_ptr,
1422 LLVMValueRef color_ptr,
1423 LLVMValueRef stride,
1424 unsigned partial_mask,
1425 boolean do_branch)
1426 {
1427 const unsigned alpha_channel = 3;
1428 const unsigned block_width = 4;
1429 const unsigned block_height = 4;
1430 const unsigned block_size = block_width * block_height;
1431 const unsigned lp_integer_vector_width = 128;
1432
1433 LLVMBuilderRef builder = gallivm->builder;
1434 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1435 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1436 LLVMValueRef src_alpha[4 * 4];
1437 LLVMValueRef src1_alpha[4 * 4];
1438 LLVMValueRef src_mask[4 * 4];
1439 LLVMValueRef src[4 * 4];
1440 LLVMValueRef src1[4 * 4];
1441 LLVMValueRef dst[4 * 4];
1442 LLVMValueRef blend_color;
1443 LLVMValueRef blend_alpha;
1444 LLVMValueRef i32_zero;
1445 LLVMValueRef check_mask;
1446
1447 struct lp_build_mask_context mask_ctx;
1448 struct lp_type mask_type;
1449 struct lp_type blend_type;
1450 struct lp_type row_type;
1451 struct lp_type dst_type;
1452
1453 unsigned char swizzle[TGSI_NUM_CHANNELS];
1454 unsigned vector_width;
1455 unsigned src_channels = TGSI_NUM_CHANNELS;
1456 unsigned dst_channels;
1457 unsigned dst_count;
1458 unsigned src_count;
1459 unsigned i, j;
1460
1461 const struct util_format_description* out_format_desc = util_format_description(out_format);
1462
1463 unsigned dst_alignment;
1464
1465 bool pad_inline = is_arithmetic_format(out_format_desc);
1466 bool has_alpha = false;
1467 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1468 util_blend_state_is_dual(&variant->key.blend, 0);
1469
1470 mask_type = lp_int32_vec4_type();
1471 mask_type.length = fs_type.length;
1472
1473 /* Compute the alignment of the destination pointer in bytes */
1474 #if 0
1475 dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1476 #else
1477 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1478 dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1479 #endif
1480 /* Force power-of-two alignment by extracting only the least-significant-bit */
1481 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1482 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1483 dst_alignment = MIN2(dst_alignment, 16);
1484
1485 /* Do not bother executing code when mask is empty.. */
1486 if (do_branch) {
1487 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1488
1489 for (i = 0; i < num_fs; ++i) {
1490 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1491 }
1492
1493 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1494 lp_build_mask_check(&mask_ctx);
1495 }
1496
1497 partial_mask |= !variant->opaque;
1498 i32_zero = lp_build_const_int32(gallivm, 0);
1499
1500 /* Get type from output format */
1501 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1502 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1503
1504 row_type.length = fs_type.length;
1505 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1506
1507 /* Compute correct swizzle and count channels */
1508 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1509 dst_channels = 0;
1510
1511 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1512 /* Ensure channel is used */
1513 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1514 continue;
1515 }
1516
1517 /* Ensure not already written to (happens in case with GL_ALPHA) */
1518 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1519 continue;
1520 }
1521
1522 /* Ensure we havn't already found all channels */
1523 if (dst_channels >= out_format_desc->nr_channels) {
1524 continue;
1525 }
1526
1527 swizzle[out_format_desc->swizzle[i]] = i;
1528 ++dst_channels;
1529
1530 if (i == alpha_channel) {
1531 has_alpha = true;
1532 }
1533 }
1534
1535 /* If 3 channels then pad to include alpha for 4 element transpose */
1536 if (dst_channels == 3 && !has_alpha) {
1537 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1538 if (swizzle[i] > TGSI_NUM_CHANNELS)
1539 swizzle[i] = 3;
1540 }
1541 if (out_format_desc->nr_channels == 4) {
1542 dst_channels = 4;
1543 }
1544 }
1545
1546 /*
1547 * Load shader output
1548 */
1549 for (i = 0; i < num_fs; ++i) {
1550 /* Always load alpha for use in blending */
1551 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1552
1553 /* Load each channel */
1554 for (j = 0; j < dst_channels; ++j) {
1555 assert(swizzle[j] < 4);
1556 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1557 }
1558
1559 /* If 3 channels then pad to include alpha for 4 element transpose */
1560 /*
1561 * XXX If we include that here maybe could actually use it instead of
1562 * separate alpha for blending?
1563 */
1564 if (dst_channels == 3 && !has_alpha) {
1565 fs_src[i][3] = alpha;
1566 }
1567
1568 /* We split the row_mask and row_alpha as we want 128bit interleave */
1569 if (fs_type.length == 8) {
1570 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1571 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1572
1573 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1574 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1575 } else {
1576 src_mask[i] = fs_mask[i];
1577 src_alpha[i] = alpha;
1578 }
1579 }
1580 if (dual_source_blend) {
1581 /* same as above except different src/dst, skip masks and comments... */
1582 for (i = 0; i < num_fs; ++i) {
1583 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1584
1585 for (j = 0; j < dst_channels; ++j) {
1586 assert(swizzle[j] < 4);
1587 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1588 }
1589 if (dst_channels == 3 && !has_alpha) {
1590 fs_src1[i][3] = alpha;
1591 }
1592 if (fs_type.length == 8) {
1593 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1594 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1595 } else {
1596 src1_alpha[i] = alpha;
1597 }
1598 }
1599 }
1600
1601 if (util_format_is_pure_integer(out_format)) {
1602 /*
1603 * In this case fs_type was really ints or uints disguised as floats,
1604 * fix that up now.
1605 */
1606 fs_type.floating = 0;
1607 fs_type.sign = dst_type.sign;
1608 for (i = 0; i < num_fs; ++i) {
1609 for (j = 0; j < dst_channels; ++j) {
1610 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1611 lp_build_vec_type(gallivm, fs_type), "");
1612 }
1613 if (dst_channels == 3 && !has_alpha) {
1614 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1615 lp_build_vec_type(gallivm, fs_type), "");
1616 }
1617 }
1618 }
1619
1620 /*
1621 * Pixel twiddle from fragment shader order to memory order
1622 */
1623 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1624 if (dual_source_blend) {
1625 generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
1626 }
1627
1628 src_channels = dst_channels < 3 ? dst_channels : 4;
1629 if (src_count != num_fs * src_channels) {
1630 unsigned ds = src_count / (num_fs * src_channels);
1631 row_type.length /= ds;
1632 fs_type.length = row_type.length;
1633 }
1634
1635 blend_type = row_type;
1636 mask_type.length = 4;
1637
1638 /* Convert src to row_type */
1639 if (dual_source_blend) {
1640 struct lp_type old_row_type = row_type;
1641 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1642 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1643 }
1644 else {
1645 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1646 }
1647
1648 /* If the rows are not an SSE vector, combine them to become SSE size! */
1649 if ((row_type.width * row_type.length) % 128) {
1650 unsigned bits = row_type.width * row_type.length;
1651 unsigned combined;
1652
1653 dst_count = src_count / (vector_width / bits);
1654 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1655 if (dual_source_blend) {
1656 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1657 }
1658
1659 row_type.length *= combined;
1660 src_count /= combined;
1661
1662 bits = row_type.width * row_type.length;
1663 assert(bits == 128 || bits == 256);
1664 }
1665
1666
1667 /*
1668 * Blend Colour conversion
1669 */
1670 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1671 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1672 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1673
1674 /* Convert */
1675 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1676
1677 /* Extract alpha */
1678 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1679
1680 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1681 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1682 if (pad_inline) {
1683 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1684 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1685 } else {
1686 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1687 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1688 }
1689
1690 /*
1691 * Mask conversion
1692 */
1693 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1694
1695 if (src_count < block_height) {
1696 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1697 } else if (src_count > block_height) {
1698 for (i = src_count; i > 0; --i) {
1699 unsigned pixels = block_size / src_count;
1700 unsigned idx = i - 1;
1701
1702 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1703 }
1704 }
1705
1706 assert(mask_type.width == 32);
1707
1708 for (i = 0; i < src_count; ++i) {
1709 unsigned pixels = block_size / src_count;
1710 unsigned pixel_width = row_type.width * dst_channels;
1711
1712 if (pixel_width == 24) {
1713 mask_type.width = 8;
1714 mask_type.length = vector_width / mask_type.width;
1715 } else {
1716 mask_type.length = pixels;
1717 mask_type.width = row_type.width * dst_channels;
1718
1719 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1720
1721 mask_type.length *= dst_channels;
1722 mask_type.width /= dst_channels;
1723 }
1724
1725 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1726 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1727 }
1728
1729 /*
1730 * Alpha conversion
1731 */
1732 if (!has_alpha) {
1733 struct lp_type alpha_type = fs_type;
1734 alpha_type.length = 4;
1735 convert_alpha(gallivm, row_type, alpha_type,
1736 block_size, block_height,
1737 src_count, dst_channels,
1738 pad_inline, src_alpha);
1739 if (dual_source_blend) {
1740 convert_alpha(gallivm, row_type, alpha_type,
1741 block_size, block_height,
1742 src_count, dst_channels,
1743 pad_inline, src1_alpha);
1744 }
1745 }
1746
1747
1748 /*
1749 * Load dst from memory
1750 */
1751 if (src_count < block_height) {
1752 dst_count = block_height;
1753 } else {
1754 dst_count = src_count;
1755 }
1756
1757 dst_type.length *= 16 / dst_count;
1758
1759 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1760 dst, dst_type, dst_count, dst_alignment);
1761
1762
1763 /*
1764 * Convert from dst/output format to src/blending format.
1765 *
1766 * This is necessary as we can only read 1 row from memory at a time,
1767 * so the minimum dst_count will ever be at this point is 4.
1768 *
1769 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1770 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1771 * on all 16 pixels in that single vector at once.
1772 */
1773 if (dst_count > src_count) {
1774 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1775 }
1776
1777 /*
1778 * Blending
1779 */
1780 /* XXX this is broken for RGB8 formats -
1781 * they get expanded from 12 to 16 elements (to include alpha)
1782 * by convert_to_blend_type then reduced to 15 instead of 12
1783 * by convert_from_blend_type (a simple fix though breaks A8...).
1784 * R16G16B16 also crashes differently however something going wrong
1785 * inside llvm handling npot vector sizes seemingly.
1786 * It seems some cleanup could be done here (like skipping conversion/blend
1787 * when not needed).
1788 */
1789 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
1790
1791 for (i = 0; i < src_count; ++i) {
1792 dst[i] = lp_build_blend_aos(gallivm,
1793 &variant->key.blend,
1794 out_format,
1795 row_type,
1796 rt,
1797 src[i],
1798 has_alpha ? NULL : src_alpha[i],
1799 src1[i],
1800 has_alpha ? NULL : src1_alpha[i],
1801 dst[i],
1802 partial_mask ? src_mask[i] : NULL,
1803 blend_color,
1804 has_alpha ? NULL : blend_alpha,
1805 swizzle,
1806 pad_inline ? 4 : dst_channels);
1807 }
1808
1809 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
1810
1811 /* Split the blend rows back to memory rows */
1812 if (dst_count > src_count) {
1813 row_type.length = dst_type.length * (dst_count / src_count);
1814
1815 if (src_count == 1) {
1816 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1817 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1818
1819 row_type.length /= 2;
1820 src_count *= 2;
1821 }
1822
1823 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1824 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1825 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1826 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1827
1828 row_type.length /= 2;
1829 src_count *= 2;
1830 }
1831
1832
1833 /*
1834 * Store blend result to memory
1835 */
1836 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1837 dst, dst_type, dst_count, dst_alignment);
1838
1839 if (do_branch) {
1840 lp_build_mask_end(&mask_ctx);
1841 }
1842 }
1843
1844
1845 /**
1846 * Generate the runtime callable function for the whole fragment pipeline.
1847 * Note that the function which we generate operates on a block of 16
1848 * pixels at at time. The block contains 2x2 quads. Each quad contains
1849 * 2x2 pixels.
1850 */
1851 static void
1852 generate_fragment(struct llvmpipe_context *lp,
1853 struct lp_fragment_shader *shader,
1854 struct lp_fragment_shader_variant *variant,
1855 unsigned partial_mask)
1856 {
1857 struct gallivm_state *gallivm = variant->gallivm;
1858 const struct lp_fragment_shader_variant_key *key = &variant->key;
1859 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1860 char func_name[256];
1861 struct lp_type fs_type;
1862 struct lp_type blend_type;
1863 LLVMTypeRef fs_elem_type;
1864 LLVMTypeRef blend_vec_type;
1865 LLVMTypeRef arg_types[12];
1866 LLVMTypeRef func_type;
1867 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1868 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1869 LLVMValueRef context_ptr;
1870 LLVMValueRef x;
1871 LLVMValueRef y;
1872 LLVMValueRef a0_ptr;
1873 LLVMValueRef dadx_ptr;
1874 LLVMValueRef dady_ptr;
1875 LLVMValueRef color_ptr_ptr;
1876 LLVMValueRef stride_ptr;
1877 LLVMValueRef depth_ptr;
1878 LLVMValueRef mask_input;
1879 LLVMValueRef thread_data_ptr;
1880 LLVMBasicBlockRef block;
1881 LLVMBuilderRef builder;
1882 struct lp_build_sampler_soa *sampler;
1883 struct lp_build_interp_soa_context interp;
1884 LLVMValueRef fs_mask[16 / 4];
1885 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1886 LLVMValueRef function;
1887 LLVMValueRef facing;
1888 const struct util_format_description *zs_format_desc;
1889 unsigned num_fs;
1890 unsigned i;
1891 unsigned chan;
1892 unsigned cbuf;
1893 boolean cbuf0_write_all;
1894 boolean try_loop = TRUE;
1895 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
1896 util_blend_state_is_dual(&key->blend, 0);
1897
1898 assert(lp_native_vector_width / 32 >= 4);
1899
1900 /* Adjust color input interpolation according to flatshade state:
1901 */
1902 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1903 for (i = 0; i < shader->info.base.num_inputs; i++) {
1904 if (inputs[i].interp == LP_INTERP_COLOR) {
1905 if (key->flatshade)
1906 inputs[i].interp = LP_INTERP_CONSTANT;
1907 else
1908 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1909 }
1910 }
1911
1912 /* check if writes to cbuf[0] are to be copied to all cbufs */
1913 cbuf0_write_all = FALSE;
1914 for (i = 0;i < shader->info.base.num_properties; i++) {
1915 if (shader->info.base.properties[i].name ==
1916 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1917 cbuf0_write_all = TRUE;
1918 break;
1919 }
1920 }
1921
1922 /* TODO: actually pick these based on the fs and color buffer
1923 * characteristics. */
1924
1925 memset(&fs_type, 0, sizeof fs_type);
1926 fs_type.floating = TRUE; /* floating point values */
1927 fs_type.sign = TRUE; /* values are signed */
1928 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1929 fs_type.width = 32; /* 32-bit float */
1930 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1931 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1932
1933 memset(&blend_type, 0, sizeof blend_type);
1934 blend_type.floating = FALSE; /* values are integers */
1935 blend_type.sign = FALSE; /* values are unsigned */
1936 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1937 blend_type.width = 8; /* 8-bit ubyte values */
1938 blend_type.length = 16; /* 16 elements per vector */
1939
1940 /*
1941 * Generate the function prototype. Any change here must be reflected in
1942 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1943 */
1944
1945 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1946
1947 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1948
1949 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1950 shader->no, variant->no, partial_mask ? "partial" : "whole");
1951
1952 arg_types[0] = variant->jit_context_ptr_type; /* context */
1953 arg_types[1] = int32_type; /* x */
1954 arg_types[2] = int32_type; /* y */
1955 arg_types[3] = int32_type; /* facing */
1956 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1957 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1958 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1959 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1960 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1961 arg_types[9] = int32_type; /* mask_input */
1962 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
1963 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1964
1965 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1966 arg_types, Elements(arg_types), 0);
1967
1968 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1969 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1970
1971 variant->function[partial_mask] = function;
1972
1973 /* XXX: need to propagate noalias down into color param now we are
1974 * passing a pointer-to-pointer?
1975 */
1976 for(i = 0; i < Elements(arg_types); ++i)
1977 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1978 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1979
1980 context_ptr = LLVMGetParam(function, 0);
1981 x = LLVMGetParam(function, 1);
1982 y = LLVMGetParam(function, 2);
1983 facing = LLVMGetParam(function, 3);
1984 a0_ptr = LLVMGetParam(function, 4);
1985 dadx_ptr = LLVMGetParam(function, 5);
1986 dady_ptr = LLVMGetParam(function, 6);
1987 color_ptr_ptr = LLVMGetParam(function, 7);
1988 depth_ptr = LLVMGetParam(function, 8);
1989 mask_input = LLVMGetParam(function, 9);
1990 thread_data_ptr = LLVMGetParam(function, 10);
1991 stride_ptr = LLVMGetParam(function, 11);
1992
1993 lp_build_name(context_ptr, "context");
1994 lp_build_name(x, "x");
1995 lp_build_name(y, "y");
1996 lp_build_name(a0_ptr, "a0");
1997 lp_build_name(dadx_ptr, "dadx");
1998 lp_build_name(dady_ptr, "dady");
1999 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2000 lp_build_name(depth_ptr, "depth");
2001 lp_build_name(thread_data_ptr, "thread_data");
2002 lp_build_name(mask_input, "mask_input");
2003 lp_build_name(stride_ptr, "stride_ptr");
2004
2005 /*
2006 * Function body
2007 */
2008
2009 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2010 builder = gallivm->builder;
2011 assert(builder);
2012 LLVMPositionBuilderAtEnd(builder, block);
2013
2014 /* code generated texture sampling */
2015 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
2016
2017 zs_format_desc = util_format_description(key->zsbuf_format);
2018
2019 if (!try_loop) {
2020 /*
2021 * The shader input interpolation info is not explicitely baked in the
2022 * shader key, but everything it derives from (TGSI, and flatshade) is
2023 * already included in the shader key.
2024 */
2025 lp_build_interp_soa_init(&interp,
2026 gallivm,
2027 shader->info.base.num_inputs,
2028 inputs,
2029 builder, fs_type,
2030 FALSE,
2031 a0_ptr, dadx_ptr, dady_ptr,
2032 x, y);
2033
2034 /* loop over quads in the block */
2035 for(i = 0; i < num_fs; ++i) {
2036 LLVMValueRef depth_offset = LLVMConstInt(int32_type,
2037 i*fs_type.length*zs_format_desc->block.bits/8,
2038 0);
2039 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2040 LLVMValueRef depth_ptr_i;
2041
2042 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
2043
2044 generate_fs(gallivm,
2045 shader, key,
2046 builder,
2047 fs_type,
2048 context_ptr,
2049 i,
2050 &interp,
2051 sampler,
2052 &fs_mask[i], /* output */
2053 out_color,
2054 depth_ptr_i,
2055 facing,
2056 partial_mask,
2057 mask_input,
2058 thread_data_ptr);
2059
2060 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2061 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
2062 fs_out_color[cbuf][chan][i] =
2063 out_color[cbuf * !cbuf0_write_all][chan];
2064 }
2065 if (dual_source_blend) {
2066 /* only support one dual source blend target hence always use output 1 */
2067 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
2068 fs_out_color[1][chan][i] =
2069 out_color[1][chan];
2070 }
2071 }
2072 }
2073 else {
2074 unsigned depth_bits = zs_format_desc->block.bits/8;
2075 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2076 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2077 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2078 num_loop, "mask_store");
2079 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2080
2081 /*
2082 * The shader input interpolation info is not explicitely baked in the
2083 * shader key, but everything it derives from (TGSI, and flatshade) is
2084 * already included in the shader key.
2085 */
2086 lp_build_interp_soa_init(&interp,
2087 gallivm,
2088 shader->info.base.num_inputs,
2089 inputs,
2090 builder, fs_type,
2091 TRUE,
2092 a0_ptr, dadx_ptr, dady_ptr,
2093 x, y);
2094
2095 for (i = 0; i < num_fs; i++) {
2096 LLVMValueRef mask;
2097 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2098 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2099 &indexi, 1, "mask_ptr");
2100
2101 if (partial_mask) {
2102 mask = generate_quad_mask(gallivm, fs_type,
2103 i*fs_type.length/4, mask_input);
2104 }
2105 else {
2106 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2107 }
2108 LLVMBuildStore(builder, mask, mask_ptr);
2109 }
2110
2111 generate_fs_loop(gallivm,
2112 shader, key,
2113 builder,
2114 fs_type,
2115 context_ptr,
2116 num_loop,
2117 &interp,
2118 sampler,
2119 mask_store, /* output */
2120 color_store,
2121 depth_ptr,
2122 depth_bits,
2123 facing,
2124 thread_data_ptr);
2125
2126 for (i = 0; i < num_fs; i++) {
2127 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2128 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2129 &indexi, 1, "");
2130 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2131 /* This is fucked up need to reorganize things */
2132 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2133 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2134 ptr = LLVMBuildGEP(builder,
2135 color_store[cbuf * !cbuf0_write_all][chan],
2136 &indexi, 1, "");
2137 fs_out_color[cbuf][chan][i] = ptr;
2138 }
2139 }
2140 if (dual_source_blend) {
2141 /* only support one dual source blend target hence always use output 1 */
2142 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2143 ptr = LLVMBuildGEP(builder,
2144 color_store[1][chan],
2145 &indexi, 1, "");
2146 fs_out_color[1][chan][i] = ptr;
2147 }
2148 }
2149 }
2150 }
2151
2152 sampler->destroy(sampler);
2153
2154 /* Loop over color outputs / color buffers to do blending.
2155 */
2156 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2157 LLVMValueRef color_ptr;
2158 LLVMValueRef stride;
2159 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2160
2161 boolean do_branch = ((key->depth.enabled
2162 || key->stencil[0].enabled
2163 || key->alpha.enabled)
2164 && !shader->info.base.uses_kill);
2165
2166 color_ptr = LLVMBuildLoad(builder,
2167 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2168 "");
2169
2170 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2171
2172 stride = LLVMBuildLoad(builder,
2173 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2174 "");
2175
2176 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2177 num_fs, fs_type, fs_mask, fs_out_color,
2178 context_ptr, color_ptr, stride, partial_mask, do_branch);
2179 }
2180
2181 LLVMBuildRetVoid(builder);
2182
2183 gallivm_verify_function(gallivm, function);
2184
2185 variant->nr_instrs += lp_build_count_instructions(function);
2186 }
2187
2188
2189 static void
2190 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2191 {
2192 unsigned i;
2193
2194 debug_printf("fs variant %p:\n", (void *) key);
2195
2196 if (key->flatshade) {
2197 debug_printf("flatshade = 1\n");
2198 }
2199 for (i = 0; i < key->nr_cbufs; ++i) {
2200 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2201 }
2202 if (key->depth.enabled) {
2203 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2204 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2205 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2206 }
2207
2208 for (i = 0; i < 2; ++i) {
2209 if (key->stencil[i].enabled) {
2210 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2211 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2212 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2213 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2214 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2215 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2216 }
2217 }
2218
2219 if (key->alpha.enabled) {
2220 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2221 }
2222
2223 if (key->occlusion_count) {
2224 debug_printf("occlusion_count = 1\n");
2225 }
2226
2227 if (key->blend.logicop_enable) {
2228 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2229 }
2230 else if (key->blend.rt[0].blend_enable) {
2231 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2232 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2233 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2234 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2235 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2236 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2237 }
2238 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2239 for (i = 0; i < key->nr_samplers; ++i) {
2240 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2241 debug_printf("sampler[%u] = \n", i);
2242 debug_printf(" .wrap = %s %s %s\n",
2243 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2244 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2245 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2246 debug_printf(" .min_img_filter = %s\n",
2247 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2248 debug_printf(" .min_mip_filter = %s\n",
2249 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2250 debug_printf(" .mag_img_filter = %s\n",
2251 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2252 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2253 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2254 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2255 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2256 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2257 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2258 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2259 }
2260 for (i = 0; i < key->nr_sampler_views; ++i) {
2261 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2262 debug_printf("texture[%u] = \n", i);
2263 debug_printf(" .format = %s\n",
2264 util_format_name(texture->format));
2265 debug_printf(" .target = %s\n",
2266 util_dump_tex_target(texture->target, TRUE));
2267 debug_printf(" .level_zero_only = %u\n",
2268 texture->level_zero_only);
2269 debug_printf(" .pot = %u %u %u\n",
2270 texture->pot_width,
2271 texture->pot_height,
2272 texture->pot_depth);
2273 }
2274 }
2275
2276
2277 void
2278 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2279 {
2280 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2281 variant->shader->no, variant->no);
2282 tgsi_dump(variant->shader->base.tokens, 0);
2283 dump_fs_variant_key(&variant->key);
2284 debug_printf("variant->opaque = %u\n", variant->opaque);
2285 debug_printf("\n");
2286 }
2287
2288
2289 /**
2290 * Generate a new fragment shader variant from the shader code and
2291 * other state indicated by the key.
2292 */
2293 static struct lp_fragment_shader_variant *
2294 generate_variant(struct llvmpipe_context *lp,
2295 struct lp_fragment_shader *shader,
2296 const struct lp_fragment_shader_variant_key *key)
2297 {
2298 struct lp_fragment_shader_variant *variant;
2299 const struct util_format_description *cbuf0_format_desc;
2300 boolean fullcolormask;
2301
2302 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2303 if(!variant)
2304 return NULL;
2305
2306 variant->gallivm = gallivm_create();
2307 if (!variant->gallivm) {
2308 FREE(variant);
2309 return NULL;
2310 }
2311
2312 variant->shader = shader;
2313 variant->list_item_global.base = variant;
2314 variant->list_item_local.base = variant;
2315 variant->no = shader->variants_created++;
2316
2317 memcpy(&variant->key, key, shader->variant_key_size);
2318
2319 /*
2320 * Determine whether we are touching all channels in the color buffer.
2321 */
2322 fullcolormask = FALSE;
2323 if (key->nr_cbufs == 1) {
2324 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2325 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2326 }
2327
2328 variant->opaque =
2329 !key->blend.logicop_enable &&
2330 !key->blend.rt[0].blend_enable &&
2331 fullcolormask &&
2332 !key->stencil[0].enabled &&
2333 !key->alpha.enabled &&
2334 !key->depth.enabled &&
2335 !shader->info.base.uses_kill
2336 ? TRUE : FALSE;
2337
2338 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2339 lp_debug_fs_variant(variant);
2340 }
2341
2342 lp_jit_init_types(variant);
2343
2344 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2345 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2346
2347 if (variant->jit_function[RAST_WHOLE] == NULL) {
2348 if (variant->opaque) {
2349 /* Specialized shader, which doesn't need to read the color buffer. */
2350 generate_fragment(lp, shader, variant, RAST_WHOLE);
2351 }
2352 }
2353
2354 /*
2355 * Compile everything
2356 */
2357
2358 gallivm_compile_module(variant->gallivm);
2359
2360 if (variant->function[RAST_EDGE_TEST]) {
2361 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2362 gallivm_jit_function(variant->gallivm,
2363 variant->function[RAST_EDGE_TEST]);
2364 }
2365
2366 if (variant->function[RAST_WHOLE]) {
2367 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2368 gallivm_jit_function(variant->gallivm,
2369 variant->function[RAST_WHOLE]);
2370 } else if (!variant->jit_function[RAST_WHOLE]) {
2371 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2372 }
2373
2374 return variant;
2375 }
2376
2377
2378 static void *
2379 llvmpipe_create_fs_state(struct pipe_context *pipe,
2380 const struct pipe_shader_state *templ)
2381 {
2382 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2383 struct lp_fragment_shader *shader;
2384 int nr_samplers;
2385 int nr_sampler_views;
2386 int i;
2387
2388 shader = CALLOC_STRUCT(lp_fragment_shader);
2389 if (!shader)
2390 return NULL;
2391
2392 shader->no = fs_no++;
2393 make_empty_list(&shader->variants);
2394
2395 /* get/save the summary info for this shader */
2396 lp_build_tgsi_info(templ->tokens, &shader->info);
2397
2398 /* we need to keep a local copy of the tokens */
2399 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2400
2401 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2402 if (shader->draw_data == NULL) {
2403 FREE((void *) shader->base.tokens);
2404 FREE(shader);
2405 return NULL;
2406 }
2407
2408 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2409 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2410
2411 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2412 state[MAX2(nr_samplers, nr_sampler_views)]);
2413
2414 for (i = 0; i < shader->info.base.num_inputs; i++) {
2415 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2416 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2417
2418 switch (shader->info.base.input_interpolate[i]) {
2419 case TGSI_INTERPOLATE_CONSTANT:
2420 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2421 break;
2422 case TGSI_INTERPOLATE_LINEAR:
2423 shader->inputs[i].interp = LP_INTERP_LINEAR;
2424 break;
2425 case TGSI_INTERPOLATE_PERSPECTIVE:
2426 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2427 break;
2428 case TGSI_INTERPOLATE_COLOR:
2429 shader->inputs[i].interp = LP_INTERP_COLOR;
2430 break;
2431 default:
2432 assert(0);
2433 break;
2434 }
2435
2436 switch (shader->info.base.input_semantic_name[i]) {
2437 case TGSI_SEMANTIC_FACE:
2438 shader->inputs[i].interp = LP_INTERP_FACING;
2439 break;
2440 case TGSI_SEMANTIC_POSITION:
2441 /* Position was already emitted above
2442 */
2443 shader->inputs[i].interp = LP_INTERP_POSITION;
2444 shader->inputs[i].src_index = 0;
2445 continue;
2446 }
2447
2448 shader->inputs[i].src_index = i+1;
2449 }
2450
2451 if (LP_DEBUG & DEBUG_TGSI) {
2452 unsigned attrib;
2453 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2454 shader->no, (void *) shader);
2455 tgsi_dump(templ->tokens, 0);
2456 debug_printf("usage masks:\n");
2457 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2458 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2459 debug_printf(" IN[%u].%s%s%s%s\n",
2460 attrib,
2461 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2462 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2463 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2464 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2465 }
2466 debug_printf("\n");
2467 }
2468
2469 return shader;
2470 }
2471
2472
2473 static void
2474 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2475 {
2476 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2477
2478 if (llvmpipe->fs == fs)
2479 return;
2480
2481 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2482
2483 draw_bind_fragment_shader(llvmpipe->draw,
2484 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2485
2486 llvmpipe->dirty |= LP_NEW_FS;
2487 }
2488
2489
2490 /**
2491 * Remove shader variant from two lists: the shader's variant list
2492 * and the context's variant list.
2493 */
2494 void
2495 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2496 struct lp_fragment_shader_variant *variant)
2497 {
2498 unsigned i;
2499
2500 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2501 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2502 " #%u v total cached #%u\n",
2503 variant->shader->no,
2504 variant->no,
2505 variant->shader->variants_created,
2506 variant->shader->variants_cached,
2507 lp->nr_fs_variants);
2508 }
2509
2510 /* free all the variant's JIT'd functions */
2511 for (i = 0; i < Elements(variant->function); i++) {
2512 if (variant->function[i]) {
2513 gallivm_free_function(variant->gallivm,
2514 variant->function[i],
2515 variant->jit_function[i]);
2516 }
2517 }
2518
2519 gallivm_destroy(variant->gallivm);
2520
2521 /* remove from shader's list */
2522 remove_from_list(&variant->list_item_local);
2523 variant->shader->variants_cached--;
2524
2525 /* remove from context's list */
2526 remove_from_list(&variant->list_item_global);
2527 lp->nr_fs_variants--;
2528 lp->nr_fs_instrs -= variant->nr_instrs;
2529
2530 FREE(variant);
2531 }
2532
2533
2534 static void
2535 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2536 {
2537 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2538 struct lp_fragment_shader *shader = fs;
2539 struct lp_fs_variant_list_item *li;
2540
2541 assert(fs != llvmpipe->fs);
2542
2543 /*
2544 * XXX: we need to flush the context until we have some sort of reference
2545 * counting in fragment shaders as they may still be binned
2546 * Flushing alone might not sufficient we need to wait on it too.
2547 */
2548 llvmpipe_finish(pipe, __FUNCTION__);
2549
2550 /* Delete all the variants */
2551 li = first_elem(&shader->variants);
2552 while(!at_end(&shader->variants, li)) {
2553 struct lp_fs_variant_list_item *next = next_elem(li);
2554 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2555 li = next;
2556 }
2557
2558 /* Delete draw module's data */
2559 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2560
2561 assert(shader->variants_cached == 0);
2562 FREE((void *) shader->base.tokens);
2563 FREE(shader);
2564 }
2565
2566
2567
2568 static void
2569 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2570 uint shader, uint index,
2571 struct pipe_constant_buffer *cb)
2572 {
2573 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2574 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2575
2576 assert(shader < PIPE_SHADER_TYPES);
2577 assert(index < Elements(llvmpipe->constants[shader]));
2578
2579 /* note: reference counting */
2580 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2581
2582 if (shader == PIPE_SHADER_VERTEX ||
2583 shader == PIPE_SHADER_GEOMETRY) {
2584 /* Pass the constants to the 'draw' module */
2585 const unsigned size = cb ? cb->buffer_size : 0;
2586 const ubyte *data;
2587
2588 if (constants) {
2589 data = (ubyte *) llvmpipe_resource_data(constants);
2590 }
2591 else if (cb && cb->user_buffer) {
2592 data = (ubyte *) cb->user_buffer;
2593 }
2594 else {
2595 data = NULL;
2596 }
2597
2598 if (data)
2599 data += cb->buffer_offset;
2600
2601 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2602 index, data, size);
2603 }
2604
2605 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2606
2607 if (cb && cb->user_buffer) {
2608 pipe_resource_reference(&constants, NULL);
2609 }
2610 }
2611
2612
2613 /**
2614 * Return the blend factor equivalent to a destination alpha of one.
2615 */
2616 static INLINE unsigned
2617 force_dst_alpha_one(unsigned factor)
2618 {
2619 switch(factor) {
2620 case PIPE_BLENDFACTOR_DST_ALPHA:
2621 return PIPE_BLENDFACTOR_ONE;
2622 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2623 return PIPE_BLENDFACTOR_ZERO;
2624 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2625 return PIPE_BLENDFACTOR_ZERO;
2626 }
2627
2628 return factor;
2629 }
2630
2631
2632 /**
2633 * We need to generate several variants of the fragment pipeline to match
2634 * all the combinations of the contributing state atoms.
2635 *
2636 * TODO: there is actually no reason to tie this to context state -- the
2637 * generated code could be cached globally in the screen.
2638 */
2639 static void
2640 make_variant_key(struct llvmpipe_context *lp,
2641 struct lp_fragment_shader *shader,
2642 struct lp_fragment_shader_variant_key *key)
2643 {
2644 unsigned i;
2645
2646 memset(key, 0, shader->variant_key_size);
2647
2648 if (lp->framebuffer.zsbuf) {
2649 if (lp->depth_stencil->depth.enabled) {
2650 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2651 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2652 }
2653 if (lp->depth_stencil->stencil[0].enabled) {
2654 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2655 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2656 }
2657 }
2658
2659 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2660 if (!lp->framebuffer.nr_cbufs ||
2661 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2662 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2663 }
2664 if(key->alpha.enabled)
2665 key->alpha.func = lp->depth_stencil->alpha.func;
2666 /* alpha.ref_value is passed in jit_context */
2667
2668 key->flatshade = lp->rasterizer->flatshade;
2669 if (lp->active_occlusion_query) {
2670 key->occlusion_count = TRUE;
2671 }
2672
2673 if (lp->framebuffer.nr_cbufs) {
2674 memcpy(&key->blend, lp->blend, sizeof key->blend);
2675 }
2676
2677 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2678
2679 if (!key->blend.independent_blend_enable) {
2680 /* we always need independent blend otherwise the fixups below won't work */
2681 for (i = 1; i < key->nr_cbufs; i++) {
2682 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2683 }
2684 key->blend.independent_blend_enable = 1;
2685 }
2686
2687 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2688 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2689 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2690 const struct util_format_description *format_desc;
2691
2692 key->cbuf_format[i] = format;
2693
2694 format_desc = util_format_description(format);
2695 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2696 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2697
2698 /*
2699 * Mask out color channels not present in the color buffer.
2700 */
2701 blend_rt->colormask &= util_format_colormask(format_desc);
2702
2703 /*
2704 * Disable blend for integer formats.
2705 */
2706 if (util_format_is_pure_integer(format)) {
2707 blend_rt->blend_enable = 0;
2708 }
2709
2710 /*
2711 * Our swizzled render tiles always have an alpha channel, but the linear
2712 * render target format often does not, so force here the dst alpha to be
2713 * one.
2714 *
2715 * This is not a mere optimization. Wrong results will be produced if the
2716 * dst alpha is used, the dst format does not have alpha, and the previous
2717 * rendering was not flushed from the swizzled to linear buffer. For
2718 * example, NonPowTwo DCT.
2719 *
2720 * TODO: This should be generalized to all channels for better
2721 * performance, but only alpha causes correctness issues.
2722 *
2723 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2724 */
2725 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2726 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2727 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2728 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2729 blend_rt->alpha_func = blend_rt->rgb_func;
2730 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2731 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2732 }
2733 }
2734
2735 /* This value will be the same for all the variants of a given shader:
2736 */
2737 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2738
2739 for(i = 0; i < key->nr_samplers; ++i) {
2740 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2741 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2742 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2743 }
2744 }
2745
2746 /*
2747 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2748 * are dx10-style? Can't really have mixed opcodes, at least not
2749 * if we want to skip the holes here (without rescanning tgsi).
2750 */
2751 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2752 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2753 for(i = 0; i < key->nr_sampler_views; ++i) {
2754 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2755 lp_sampler_static_texture_state(&key->state[i].texture_state,
2756 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2757 }
2758 }
2759 }
2760 else {
2761 key->nr_sampler_views = key->nr_samplers;
2762 for(i = 0; i < key->nr_sampler_views; ++i) {
2763 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2764 lp_sampler_static_texture_state(&key->state[i].texture_state,
2765 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2766 }
2767 }
2768 }
2769 }
2770
2771
2772
2773 /**
2774 * Update fragment shader state. This is called just prior to drawing
2775 * something when some fragment-related state has changed.
2776 */
2777 void
2778 llvmpipe_update_fs(struct llvmpipe_context *lp)
2779 {
2780 struct lp_fragment_shader *shader = lp->fs;
2781 struct lp_fragment_shader_variant_key key;
2782 struct lp_fragment_shader_variant *variant = NULL;
2783 struct lp_fs_variant_list_item *li;
2784
2785 make_variant_key(lp, shader, &key);
2786
2787 /* Search the variants for one which matches the key */
2788 li = first_elem(&shader->variants);
2789 while(!at_end(&shader->variants, li)) {
2790 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2791 variant = li->base;
2792 break;
2793 }
2794 li = next_elem(li);
2795 }
2796
2797 if (variant) {
2798 /* Move this variant to the head of the list to implement LRU
2799 * deletion of shader's when we have too many.
2800 */
2801 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2802 }
2803 else {
2804 /* variant not found, create it now */
2805 int64_t t0, t1, dt;
2806 unsigned i;
2807 unsigned variants_to_cull;
2808
2809 if (0) {
2810 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2811 lp->nr_fs_variants,
2812 lp->nr_fs_instrs,
2813 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2814 }
2815
2816 /* First, check if we've exceeded the max number of shader variants.
2817 * If so, free 25% of them (the least recently used ones).
2818 */
2819 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2820
2821 if (variants_to_cull ||
2822 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2823 struct pipe_context *pipe = &lp->pipe;
2824
2825 /*
2826 * XXX: we need to flush the context until we have some sort of
2827 * reference counting in fragment shaders as they may still be binned
2828 * Flushing alone might not be sufficient we need to wait on it too.
2829 */
2830 llvmpipe_finish(pipe, __FUNCTION__);
2831
2832 /*
2833 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2834 * number of shader variants (potentially all of them) could be
2835 * pending for destruction on flush.
2836 */
2837
2838 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2839 struct lp_fs_variant_list_item *item;
2840 if (is_empty_list(&lp->fs_variants_list)) {
2841 break;
2842 }
2843 item = last_elem(&lp->fs_variants_list);
2844 assert(item);
2845 assert(item->base);
2846 llvmpipe_remove_shader_variant(lp, item->base);
2847 }
2848 }
2849
2850 /*
2851 * Generate the new variant.
2852 */
2853 t0 = os_time_get();
2854 variant = generate_variant(lp, shader, &key);
2855 t1 = os_time_get();
2856 dt = t1 - t0;
2857 LP_COUNT_ADD(llvm_compile_time, dt);
2858 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2859
2860 llvmpipe_variant_count++;
2861
2862 /* Put the new variant into the list */
2863 if (variant) {
2864 insert_at_head(&shader->variants, &variant->list_item_local);
2865 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2866 lp->nr_fs_variants++;
2867 lp->nr_fs_instrs += variant->nr_instrs;
2868 shader->variants_cached++;
2869 }
2870 }
2871
2872 /* Bind this variant */
2873 lp_setup_set_fs_variant(lp->setup, variant);
2874 }
2875
2876
2877
2878
2879
2880 void
2881 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2882 {
2883 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2884 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2885 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2886
2887 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2888 }