projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: Remove the static interpolation.
[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 */
219 static void
220 generate_fs_loop(struct gallivm_state *gallivm,
221 struct lp_fragment_shader *shader,
222 const struct lp_fragment_shader_variant_key *key,
223 LLVMBuilderRef builder,
224 struct lp_type type,
225 LLVMValueRef context_ptr,
226 LLVMValueRef num_loop,
227 struct lp_build_interp_soa_context *interp,
228 struct lp_build_sampler_soa *sampler,
229 LLVMValueRef mask_store,
230 LLVMValueRef (*out_color)[4],
231 LLVMValueRef depth_ptr,
232 unsigned depth_bits,
233 LLVMValueRef facing,
234 LLVMValueRef thread_data_ptr)
235 {
236 const struct util_format_description *zs_format_desc = NULL;
237 const struct tgsi_token *tokens = shader->base.tokens;
238 LLVMTypeRef vec_type;
239 LLVMValueRef mask_ptr, mask_val;
240 LLVMValueRef consts_ptr;
241 LLVMValueRef z;
242 LLVMValueRef zs_value = NULL;
243 LLVMValueRef stencil_refs[2];
244 LLVMValueRef depth_ptr_i;
245 LLVMValueRef depth_offset;
246 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
247 struct lp_build_for_loop_state loop_state;
248 struct lp_build_mask_context mask;
249 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
250 shader->info.base.num_inputs < 3 &&
251 shader->info.base.num_instructions < 8);
252 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
253 util_blend_state_is_dual(&key->blend, 0);
254 unsigned attrib;
255 unsigned chan;
256 unsigned cbuf;
257 unsigned depth_mode;
258
259 struct lp_bld_tgsi_system_values system_values;
260
261 memset(&system_values, 0, sizeof(system_values));
262
263 if (key->depth.enabled ||
264 key->stencil[0].enabled ||
265 key->stencil[1].enabled) {
266
267 zs_format_desc = util_format_description(key->zsbuf_format);
268 assert(zs_format_desc);
269
270 if (!shader->info.base.writes_z) {
271 if (key->alpha.enabled || shader->info.base.uses_kill)
272 /* With alpha test and kill, can do the depth test early
273 * and hopefully eliminate some quads. But need to do a
274 * special deferred depth write once the final mask value
275 * is known.
276 */
277 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
278 else
279 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
280 }
281 else {
282 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
283 }
284
285 if (!(key->depth.enabled && key->depth.writemask) &&
286 !(key->stencil[0].enabled && key->stencil[0].writemask))
287 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
288 }
289 else {
290 depth_mode = 0;
291 }
292
293
294 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
295 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
296
297 vec_type = lp_build_vec_type(gallivm, type);
298
299 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
300
301 lp_build_for_loop_begin(&loop_state, gallivm,
302 lp_build_const_int32(gallivm, 0),
303 LLVMIntULT,
304 num_loop,
305 lp_build_const_int32(gallivm, 1));
306
307 mask_ptr = LLVMBuildGEP(builder, mask_store,
308 &loop_state.counter, 1, "mask_ptr");
309 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
310
311 depth_offset = LLVMBuildMul(builder, loop_state.counter,
312 lp_build_const_int32(gallivm, depth_bits * type.length),
313 "");
314
315 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
316
317 memset(outputs, 0, sizeof outputs);
318
319 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
320 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
321 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
322 lp_build_vec_type(gallivm,
323 type),
324 num_loop, "color");
325 }
326 }
327 if (dual_source_blend) {
328 assert(key->nr_cbufs <= 1);
329 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
330 out_color[1][chan] = lp_build_array_alloca(gallivm,
331 lp_build_vec_type(gallivm,
332 type),
333 num_loop, "color1");
334 }
335 }
336
337
338 /* 'mask' will control execution based on quad's pixel alive/killed state */
339 lp_build_mask_begin(&mask, gallivm, type, mask_val);
340
341 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
342 lp_build_mask_check(&mask);
343
344 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
345 z = interp->pos[2];
346
347 if (depth_mode & EARLY_DEPTH_TEST) {
348 lp_build_depth_stencil_test(gallivm,
349 &key->depth,
350 key->stencil,
351 type,
352 zs_format_desc,
353 &mask,
354 stencil_refs,
355 z,
356 depth_ptr_i, facing,
357 &zs_value,
358 !simple_shader);
359
360 if (depth_mode & EARLY_DEPTH_WRITE) {
361 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
362 }
363 }
364
365 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
366
367 /* Build the actual shader */
368 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
369 consts_ptr, &system_values,
370 interp->inputs,
371 outputs, sampler, &shader->info.base, NULL);
372
373 /* Alpha test */
374 if (key->alpha.enabled) {
375 int color0 = find_output_by_semantic(&shader->info.base,
376 TGSI_SEMANTIC_COLOR,
377 0);
378
379 if (color0 != -1 && outputs[color0][3]) {
380 const struct util_format_description *cbuf_format_desc;
381 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
382 LLVMValueRef alpha_ref_value;
383
384 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
385 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
386
387 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
388
389 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
390 &mask, alpha, alpha_ref_value,
391 (depth_mode & LATE_DEPTH_TEST) != 0);
392 }
393 }
394
395 /* Late Z test */
396 if (depth_mode & LATE_DEPTH_TEST) {
397 int pos0 = find_output_by_semantic(&shader->info.base,
398 TGSI_SEMANTIC_POSITION,
399 0);
400
401 if (pos0 != -1 && outputs[pos0][2]) {
402 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
403 }
404
405 lp_build_depth_stencil_test(gallivm,
406 &key->depth,
407 key->stencil,
408 type,
409 zs_format_desc,
410 &mask,
411 stencil_refs,
412 z,
413 depth_ptr_i, facing,
414 &zs_value,
415 !simple_shader);
416 /* Late Z write */
417 if (depth_mode & LATE_DEPTH_WRITE) {
418 lp_build_depth_write(gallivm, type, zs_format_desc, depth_ptr_i, zs_value);
419 }
420 }
421 else if ((depth_mode & EARLY_DEPTH_TEST) &&
422 (depth_mode & LATE_DEPTH_WRITE))
423 {
424 /* Need to apply a reduced mask to the depth write. Reload the
425 * depth value, update from zs_value with the new mask value and
426 * write that out.
427 */
428 lp_build_deferred_depth_write(gallivm,
429 type,
430 zs_format_desc,
431 &mask,
432 depth_ptr_i,
433 zs_value);
434 }
435
436
437 /* Color write */
438 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
439 {
440 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
441 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
442 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
443 {
444 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
445 if(outputs[attrib][chan]) {
446 /* XXX: just initialize outputs to point at colors[] and
447 * skip this.
448 */
449 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
450 LLVMValueRef color_ptr;
451 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
452 &loop_state.counter, 1, "");
453 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
454 LLVMBuildStore(builder, out, color_ptr);
455 }
456 }
457 }
458 }
459
460 if (key->occlusion_count) {
461 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
462 lp_build_name(counter, "counter");
463 lp_build_occlusion_count(gallivm, type,
464 lp_build_mask_value(&mask), counter);
465 }
466
467 mask_val = lp_build_mask_end(&mask);
468 LLVMBuildStore(builder, mask_val, mask_ptr);
469 lp_build_for_loop_end(&loop_state);
470 }
471
472
473 /**
474 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
475 *
476 * Fragment Shader outputs pixels in small 2x2 blocks
477 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
478 *
479 * However in memory pixels are stored in rows
480 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
481 *
482 * @param type fragment shader type (4x or 8x float)
483 * @param num_fs number of fs_src
484 * @param dst_channels number of output channels
485 * @param fs_src output from fragment shader
486 * @param dst pointer to store result
487 * @param pad_inline is channel padding inline or at end of row
488 * @return the number of dsts
489 */
490 static int
491 generate_fs_twiddle(struct gallivm_state *gallivm,
492 struct lp_type type,
493 unsigned num_fs,
494 unsigned dst_channels,
495 LLVMValueRef fs_src[][4],
496 LLVMValueRef* dst,
497 bool pad_inline)
498 {
499 LLVMValueRef src[16];
500
501 bool swizzle_pad;
502 bool twiddle;
503 bool split;
504
505 unsigned pixels = num_fs == 4 ? 1 : 2;
506 unsigned reorder_group;
507 unsigned src_channels;
508 unsigned src_count;
509 unsigned i;
510
511 src_channels = dst_channels < 3 ? dst_channels : 4;
512 src_count = num_fs * src_channels;
513
514 assert(pixels == 2 || num_fs == 4);
515 assert(num_fs * src_channels <= Elements(src));
516
517 /*
518 * Transpose from SoA -> AoS
519 */
520 for (i = 0; i < num_fs; ++i) {
521 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
522 }
523
524 /*
525 * Pick transformation options
526 */
527 swizzle_pad = false;
528 twiddle = false;
529 split = false;
530 reorder_group = 0;
531
532 if (dst_channels == 1) {
533 twiddle = true;
534
535 if (pixels == 2) {
536 split = true;
537 }
538 } else if (dst_channels == 2) {
539 if (pixels == 1) {
540 reorder_group = 1;
541 }
542 } else if (dst_channels > 2) {
543 if (pixels == 1) {
544 reorder_group = 2;
545 } else {
546 twiddle = true;
547 }
548
549 if (!pad_inline && dst_channels == 3 && pixels > 1) {
550 swizzle_pad = true;
551 }
552 }
553
554 /*
555 * Split the src in half
556 */
557 if (split) {
558 for (i = num_fs; i > 0; --i) {
559 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
560 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
561 }
562
563 src_count *= 2;
564 type.length = 4;
565 }
566
567 /*
568 * Ensure pixels are in memory order
569 */
570 if (reorder_group) {
571 /* Twiddle pixels by reordering the array, e.g.:
572 *
573 * src_count = 8 -> 0 2 1 3 4 6 5 7
574 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
575 */
576 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
577
578 for (i = 0; i < src_count; ++i) {
579 unsigned group = i / reorder_group;
580 unsigned block = (group / 4) * 4 * reorder_group;
581 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
582 dst[i] = src[j];
583 }
584 } else if (twiddle) {
585 /* Twiddle pixels across elements of array */
586 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
587 } else {
588 /* Do nothing */
589 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
590 }
591
592 /*
593 * Moves any padding between pixels to the end
594 * e.g. RGBXRGBX -> RGBRGBXX
595 */
596 if (swizzle_pad) {
597 unsigned char swizzles[16];
598 unsigned elems = pixels * dst_channels;
599
600 for (i = 0; i < type.length; ++i) {
601 if (i < elems)
602 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
603 else
604 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
605 }
606
607 for (i = 0; i < src_count; ++i) {
608 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
609 }
610 }
611
612 return src_count;
613 }
614
615
616 /**
617 * Load an unswizzled block of pixels from memory
618 */
619 static void
620 load_unswizzled_block(struct gallivm_state *gallivm,
621 LLVMValueRef base_ptr,
622 LLVMValueRef stride,
623 unsigned block_width,
624 unsigned block_height,
625 LLVMValueRef* dst,
626 struct lp_type dst_type,
627 unsigned dst_count,
628 unsigned dst_alignment)
629 {
630 LLVMBuilderRef builder = gallivm->builder;
631 unsigned row_size = dst_count / block_height;
632 unsigned i;
633
634 /* Ensure block exactly fits into dst */
635 assert((block_width * block_height) % dst_count == 0);
636
637 for (i = 0; i < dst_count; ++i) {
638 unsigned x = i % row_size;
639 unsigned y = i / row_size;
640
641 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
642 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
643
644 LLVMValueRef gep[2];
645 LLVMValueRef dst_ptr;
646
647 gep[0] = lp_build_const_int32(gallivm, 0);
648 gep[1] = LLVMBuildAdd(builder, bx, by, "");
649
650 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
651 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
652
653 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
654
655 lp_set_load_alignment(dst[i], dst_alignment);
656 }
657 }
658
659
660 /**
661 * Store an unswizzled block of pixels to memory
662 */
663 static void
664 store_unswizzled_block(struct gallivm_state *gallivm,
665 LLVMValueRef base_ptr,
666 LLVMValueRef stride,
667 unsigned block_width,
668 unsigned block_height,
669 LLVMValueRef* src,
670 struct lp_type src_type,
671 unsigned src_count,
672 unsigned src_alignment)
673 {
674 LLVMBuilderRef builder = gallivm->builder;
675 unsigned row_size = src_count / block_height;
676 unsigned i;
677
678 /* Ensure src exactly fits into block */
679 assert((block_width * block_height) % src_count == 0);
680
681 for (i = 0; i < src_count; ++i) {
682 unsigned x = i % row_size;
683 unsigned y = i / row_size;
684
685 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
686 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
687
688 LLVMValueRef gep[2];
689 LLVMValueRef src_ptr;
690
691 gep[0] = lp_build_const_int32(gallivm, 0);
692 gep[1] = LLVMBuildAdd(builder, bx, by, "");
693
694 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
695 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
696
697 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
698
699 lp_set_store_alignment(src_ptr, src_alignment);
700 }
701 }
702
703
704 /**
705 * Checks if a format description is an arithmetic format
706 *
707 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
708 */
709 static INLINE boolean
710 is_arithmetic_format(const struct util_format_description *format_desc)
711 {
712 boolean arith = false;
713 unsigned i;
714
715 for (i = 0; i < format_desc->nr_channels; ++i) {
716 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
717 arith |= (format_desc->channel[i].size % 8) != 0;
718 }
719
720 return arith;
721 }
722
723
724 /**
725 * Retrieves the type representing the memory layout for a format
726 *
727 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
728 */
729 static INLINE void
730 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
731 struct lp_type* type)
732 {
733 unsigned i;
734 unsigned chan;
735
736 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
737 /* just make this a 32bit uint */
738 type->floating = false;
739 type->fixed = false;
740 type->sign = false;
741 type->norm = false;
742 type->width = 32;
743 type->length = 1;
744 return;
745 }
746
747 for (i = 0; i < 4; i++)
748 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
749 break;
750 chan = i;
751
752 memset(type, 0, sizeof(struct lp_type));
753 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
754 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
755 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
756 type->norm = format_desc->channel[chan].normalized;
757
758 if (is_arithmetic_format(format_desc)) {
759 type->width = 0;
760 type->length = 1;
761
762 for (i = 0; i < format_desc->nr_channels; ++i) {
763 type->width += format_desc->channel[i].size;
764 }
765 } else {
766 type->width = format_desc->channel[chan].size;
767 type->length = format_desc->nr_channels;
768 }
769 }
770
771
772 /**
773 * Retrieves the type for a format which is usable in the blending code.
774 *
775 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
776 */
777 static INLINE void
778 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
779 struct lp_type* type)
780 {
781 unsigned i;
782 unsigned chan;
783
784 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
785 /* always use ordinary floats for blending */
786 type->floating = true;
787 type->fixed = false;
788 type->sign = true;
789 type->norm = false;
790 type->width = 32;
791 type->length = 4;
792 return;
793 }
794
795 for (i = 0; i < 4; i++)
796 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
797 break;
798 chan = i;
799
800 memset(type, 0, sizeof(struct lp_type));
801 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
802 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
803 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
804 type->norm = format_desc->channel[chan].normalized;
805 type->width = format_desc->channel[chan].size;
806 type->length = format_desc->nr_channels;
807
808 for (i = 1; i < format_desc->nr_channels; ++i) {
809 if (format_desc->channel[i].size > type->width)
810 type->width = format_desc->channel[i].size;
811 }
812
813 if (type->floating) {
814 type->width = 32;
815 } else {
816 if (type->width <= 8) {
817 type->width = 8;
818 } else if (type->width <= 16) {
819 type->width = 16;
820 } else {
821 type->width = 32;
822 }
823 }
824
825 if (is_arithmetic_format(format_desc) && type->length == 3) {
826 type->length = 4;
827 }
828 }
829
830
831 /**
832 * Scale a normalized value from src_bits to dst_bits
833 */
834 static INLINE LLVMValueRef
835 scale_bits(struct gallivm_state *gallivm,
836 int src_bits,
837 int dst_bits,
838 LLVMValueRef src,
839 struct lp_type src_type)
840 {
841 LLVMBuilderRef builder = gallivm->builder;
842 LLVMValueRef result = src;
843
844 if (dst_bits < src_bits) {
845 /* Scale down by LShr */
846 result = LLVMBuildLShr(builder,
847 src,
848 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
849 "");
850 } else if (dst_bits > src_bits) {
851 /* Scale up bits */
852 int db = dst_bits - src_bits;
853
854 /* Shift left by difference in bits */
855 result = LLVMBuildShl(builder,
856 src,
857 lp_build_const_int_vec(gallivm, src_type, db),
858 "");
859
860 if (db < src_bits) {
861 /* Enough bits in src to fill the remainder */
862 LLVMValueRef lower = LLVMBuildLShr(builder,
863 src,
864 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
865 "");
866
867 result = LLVMBuildOr(builder, result, lower, "");
868 } else if (db > src_bits) {
869 /* Need to repeatedly copy src bits to fill remainder in dst */
870 unsigned n;
871
872 for (n = src_bits; n < dst_bits; n *= 2) {
873 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
874
875 result = LLVMBuildOr(builder,
876 result,
877 LLVMBuildLShr(builder, result, shuv, ""),
878 "");
879 }
880 }
881 }
882
883 return result;
884 }
885
886
887 /**
888 * Convert from memory format to blending format
889 *
890 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
891 */
892 static void
893 convert_to_blend_type(struct gallivm_state *gallivm,
894 const struct util_format_description *src_fmt,
895 struct lp_type src_type,
896 struct lp_type dst_type,
897 LLVMValueRef* src, // and dst
898 unsigned num_srcs)
899 {
900 LLVMValueRef *dst = src;
901 LLVMBuilderRef builder = gallivm->builder;
902 struct lp_type blend_type;
903 struct lp_type mem_type;
904 unsigned i, j, k;
905 unsigned pixels = 16 / num_srcs;
906 bool is_arith;
907
908 /*
909 * full custom path for packed floats - none of the later functions would do
910 * anything useful, and given the lp_type representation they can't be fixed.
911 */
912 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
913 LLVMValueRef tmpsrc[4];
914 /*
915 * This is pretty suboptimal for this case blending in SoA would be much
916 * better, since conversion gets us SoA values so need to convert back.
917 */
918 assert(src_type.width == 32);
919 assert(dst_type.floating);
920 assert(dst_type.width == 32);
921 assert(dst_type.length % 4 == 0);
922 for (i = 0; i < 4; i++) {
923 tmpsrc[i] = src[i];
924 }
925 for (i = 0; i < num_srcs / 4; i++) {
926 LLVMValueRef tmpsoa[4];
927 LLVMValueRef tmps = tmpsrc[i];
928 if (num_srcs == 8) {
929 LLVMValueRef shuffles[8];
930 unsigned j;
931 /* fetch was 4 values but need 8-wide output values */
932 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
933 /*
934 * for 8-wide aos transpose would give us wrong order not matching
935 * incoming converted fs values and mask. ARGH.
936 */
937 for (j = 0; j < 4; j++) {
938 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
939 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
940 }
941 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
942 LLVMConstVector(shuffles, 8), "");
943 }
944 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
945 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
946 }
947 return;
948 }
949
950 lp_mem_type_from_format_desc(src_fmt, &mem_type);
951 lp_blend_type_from_format_desc(src_fmt, &blend_type);
952
953 /* Is the format arithmetic */
954 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
955 is_arith &= !(mem_type.width == 16 && mem_type.floating);
956
957 /* Pad if necessary */
958 if (!is_arith && src_type.length < dst_type.length) {
959 for (i = 0; i < num_srcs; ++i) {
960 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
961 }
962
963 src_type.length = dst_type.length;
964 }
965
966 /* Special case for half-floats */
967 if (mem_type.width == 16 && mem_type.floating) {
968 assert(blend_type.width == 32 && blend_type.floating);
969 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
970 is_arith = false;
971 }
972
973 if (!is_arith) {
974 return;
975 }
976
977 src_type.width = blend_type.width * blend_type.length;
978 blend_type.length *= pixels;
979 src_type.length *= pixels / (src_type.length / mem_type.length);
980
981 for (i = 0; i < num_srcs; ++i) {
982 LLVMValueRef chans[4];
983 LLVMValueRef res = NULL;
984 unsigned sa = 0;
985
986 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
987
988 for (j = 0; j < src_fmt->nr_channels; ++j) {
989 unsigned mask = 0;
990
991 for (k = 0; k < src_fmt->channel[j].size; ++k) {
992 mask |= 1 << k;
993 }
994
995 /* Extract bits from source */
996 chans[j] = LLVMBuildLShr(builder,
997 dst[i],
998 lp_build_const_int_vec(gallivm, src_type, sa),
999 "");
1000
1001 chans[j] = LLVMBuildAnd(builder,
1002 chans[j],
1003 lp_build_const_int_vec(gallivm, src_type, mask),
1004 "");
1005
1006 /* Scale bits */
1007 if (src_type.norm) {
1008 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1009 blend_type.width, chans[j], src_type);
1010 }
1011
1012 /* Insert bits into correct position */
1013 chans[j] = LLVMBuildShl(builder,
1014 chans[j],
1015 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1016 "");
1017
1018 sa += src_fmt->channel[j].size;
1019
1020 if (j == 0) {
1021 res = chans[j];
1022 } else {
1023 res = LLVMBuildOr(builder, res, chans[j], "");
1024 }
1025 }
1026
1027 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1028 }
1029 }
1030
1031
1032 /**
1033 * Convert from blending format to memory format
1034 *
1035 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1036 */
1037 static void
1038 convert_from_blend_type(struct gallivm_state *gallivm,
1039 const struct util_format_description *src_fmt,
1040 struct lp_type src_type,
1041 struct lp_type dst_type,
1042 LLVMValueRef* src, // and dst
1043 unsigned num_srcs)
1044 {
1045 LLVMValueRef* dst = src;
1046 unsigned i, j, k;
1047 struct lp_type mem_type;
1048 struct lp_type blend_type;
1049 LLVMBuilderRef builder = gallivm->builder;
1050 unsigned pixels = 16 / num_srcs;
1051 bool is_arith;
1052
1053 /*
1054 * full custom path for packed floats - none of the later functions would do
1055 * anything useful, and given the lp_type representation they can't be fixed.
1056 */
1057 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1058 /*
1059 * This is pretty suboptimal for this case blending in SoA would be much
1060 * better - we need to transpose the AoS values back to SoA values for
1061 * conversion/packing.
1062 */
1063 assert(src_type.floating);
1064 assert(src_type.width == 32);
1065 assert(src_type.length % 4 == 0);
1066 assert(dst_type.width == 32);
1067 for (i = 0; i < num_srcs / 4; i++) {
1068 LLVMValueRef tmpsoa[4], tmpdst;
1069 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1070 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1071 if (num_srcs == 8) {
1072 LLVMValueRef tmpaos, shuffles[8];
1073 unsigned j;
1074 /*
1075 * for 8-wide aos transpose has given us wrong order not matching
1076 * output order. HMPF. Also need to split the output values manually.
1077 */
1078 for (j = 0; j < 4; j++) {
1079 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1080 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1081 }
1082 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1083 LLVMConstVector(shuffles, 8), "");
1084 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1085 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1086 }
1087 else {
1088 src[i] = tmpdst;
1089 }
1090 }
1091 return;
1092 }
1093
1094 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1095 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1096
1097 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1098
1099 /* Special case for half-floats */
1100 if (mem_type.width == 16 && mem_type.floating) {
1101 int length = dst_type.length;
1102 assert(blend_type.width == 32 && blend_type.floating);
1103
1104 dst_type.length = src_type.length;
1105
1106 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1107
1108 dst_type.length = length;
1109 is_arith = false;
1110 }
1111
1112 /* Remove any padding */
1113 if (!is_arith && (src_type.length % mem_type.length)) {
1114 src_type.length -= (src_type.length % mem_type.length);
1115
1116 for (i = 0; i < num_srcs; ++i) {
1117 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1118 }
1119 }
1120
1121 /* No bit arithmetic to do */
1122 if (!is_arith) {
1123 return;
1124 }
1125
1126 src_type.length = pixels;
1127 src_type.width = blend_type.length * blend_type.width;
1128 dst_type.length = pixels;
1129
1130 for (i = 0; i < num_srcs; ++i) {
1131 LLVMValueRef chans[4];
1132 LLVMValueRef res = NULL;
1133 unsigned sa = 0;
1134
1135 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1136
1137 for (j = 0; j < src_fmt->nr_channels; ++j) {
1138 unsigned mask = 0;
1139
1140 assert(blend_type.width > src_fmt->channel[j].size);
1141
1142 for (k = 0; k < blend_type.width; ++k) {
1143 mask |= 1 << k;
1144 }
1145
1146 /* Extract bits */
1147 chans[j] = LLVMBuildLShr(builder,
1148 dst[i],
1149 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1150 "");
1151
1152 chans[j] = LLVMBuildAnd(builder,
1153 chans[j],
1154 lp_build_const_int_vec(gallivm, src_type, mask),
1155 "");
1156
1157 /* Scale down bits */
1158 if (src_type.norm) {
1159 chans[j] = scale_bits(gallivm, blend_type.width,
1160 src_fmt->channel[j].size, chans[j], src_type);
1161 }
1162
1163 /* Insert bits */
1164 chans[j] = LLVMBuildShl(builder,
1165 chans[j],
1166 lp_build_const_int_vec(gallivm, src_type, sa),
1167 "");
1168
1169 sa += src_fmt->channel[j].size;
1170
1171 if (j == 0) {
1172 res = chans[j];
1173 } else {
1174 res = LLVMBuildOr(builder, res, chans[j], "");
1175 }
1176 }
1177
1178 assert (dst_type.width != 24);
1179
1180 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1181 }
1182 }
1183
1184
1185 /**
1186 * Convert alpha to same blend type as src
1187 */
1188 static void
1189 convert_alpha(struct gallivm_state *gallivm,
1190 struct lp_type row_type,
1191 struct lp_type alpha_type,
1192 const unsigned block_size,
1193 const unsigned block_height,
1194 const unsigned src_count,
1195 const unsigned dst_channels,
1196 const bool pad_inline,
1197 LLVMValueRef* src_alpha)
1198 {
1199 LLVMBuilderRef builder = gallivm->builder;
1200 unsigned i, j;
1201 unsigned length = row_type.length;
1202 row_type.length = alpha_type.length;
1203
1204 /* Twiddle the alpha to match pixels */
1205 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1206
1207 for (i = 0; i < 4; ++i) {
1208 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1209 }
1210
1211 alpha_type = row_type;
1212 row_type.length = length;
1213
1214 /* If only one channel we can only need the single alpha value per pixel */
1215 if (src_count == 1) {
1216 assert(dst_channels == 1);
1217
1218 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1219 } else {
1220 /* If there are more srcs than rows then we need to split alpha up */
1221 if (src_count > block_height) {
1222 for (i = src_count; i > 0; --i) {
1223 unsigned pixels = block_size / src_count;
1224 unsigned idx = i - 1;
1225
1226 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1227 }
1228 }
1229
1230 /* If there is a src for each pixel broadcast the alpha across whole row */
1231 if (src_count == block_size) {
1232 for (i = 0; i < src_count; ++i) {
1233 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1234 }
1235 } else {
1236 unsigned pixels = block_size / src_count;
1237 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1238 unsigned alpha_span = 1;
1239 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1240
1241 /* Check if we need 2 src_alphas for our shuffles */
1242 if (pixels > alpha_type.length) {
1243 alpha_span = 2;
1244 }
1245
1246 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1247 for (j = 0; j < row_type.length; ++j) {
1248 if (j < pixels * channels) {
1249 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1250 } else {
1251 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1252 }
1253 }
1254
1255 for (i = 0; i < src_count; ++i) {
1256 unsigned idx1 = i, idx2 = i;
1257
1258 if (alpha_span > 1){
1259 idx1 *= alpha_span;
1260 idx2 = idx1 + 1;
1261 }
1262
1263 src_alpha[i] = LLVMBuildShuffleVector(builder,
1264 src_alpha[idx1],
1265 src_alpha[idx2],
1266 LLVMConstVector(shuffles, row_type.length),
1267 "");
1268 }
1269 }
1270 }
1271 }
1272
1273
1274 /**
1275 * Generates the blend function for unswizzled colour buffers
1276 * Also generates the read & write from colour buffer
1277 */
1278 static void
1279 generate_unswizzled_blend(struct gallivm_state *gallivm,
1280 unsigned rt,
1281 struct lp_fragment_shader_variant *variant,
1282 enum pipe_format out_format,
1283 unsigned int num_fs,
1284 struct lp_type fs_type,
1285 LLVMValueRef* fs_mask,
1286 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1287 LLVMValueRef context_ptr,
1288 LLVMValueRef color_ptr,
1289 LLVMValueRef stride,
1290 unsigned partial_mask,
1291 boolean do_branch)
1292 {
1293 const unsigned alpha_channel = 3;
1294 const unsigned block_width = 4;
1295 const unsigned block_height = 4;
1296 const unsigned block_size = block_width * block_height;
1297 const unsigned lp_integer_vector_width = 128;
1298
1299 LLVMBuilderRef builder = gallivm->builder;
1300 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1301 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1302 LLVMValueRef src_alpha[4 * 4];
1303 LLVMValueRef src1_alpha[4 * 4];
1304 LLVMValueRef src_mask[4 * 4];
1305 LLVMValueRef src[4 * 4];
1306 LLVMValueRef src1[4 * 4];
1307 LLVMValueRef dst[4 * 4];
1308 LLVMValueRef blend_color;
1309 LLVMValueRef blend_alpha;
1310 LLVMValueRef i32_zero;
1311 LLVMValueRef check_mask;
1312
1313 struct lp_build_mask_context mask_ctx;
1314 struct lp_type mask_type;
1315 struct lp_type blend_type;
1316 struct lp_type row_type;
1317 struct lp_type dst_type;
1318
1319 unsigned char swizzle[TGSI_NUM_CHANNELS];
1320 unsigned vector_width;
1321 unsigned src_channels = TGSI_NUM_CHANNELS;
1322 unsigned dst_channels;
1323 unsigned dst_count;
1324 unsigned src_count;
1325 unsigned i, j;
1326
1327 const struct util_format_description* out_format_desc = util_format_description(out_format);
1328
1329 unsigned dst_alignment;
1330
1331 bool pad_inline = is_arithmetic_format(out_format_desc);
1332 bool has_alpha = false;
1333 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1334 util_blend_state_is_dual(&variant->key.blend, 0);
1335
1336 mask_type = lp_int32_vec4_type();
1337 mask_type.length = fs_type.length;
1338
1339 /* Compute the alignment of the destination pointer in bytes */
1340 #if 0
1341 dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1342 #else
1343 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1344 dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1345 #endif
1346 /* Force power-of-two alignment by extracting only the least-significant-bit */
1347 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1348 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1349 dst_alignment = MIN2(dst_alignment, 16);
1350
1351 /* Do not bother executing code when mask is empty.. */
1352 if (do_branch) {
1353 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1354
1355 for (i = 0; i < num_fs; ++i) {
1356 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1357 }
1358
1359 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1360 lp_build_mask_check(&mask_ctx);
1361 }
1362
1363 partial_mask |= !variant->opaque;
1364 i32_zero = lp_build_const_int32(gallivm, 0);
1365
1366 /* Get type from output format */
1367 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1368 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1369
1370 row_type.length = fs_type.length;
1371 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1372
1373 /* Compute correct swizzle and count channels */
1374 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1375 dst_channels = 0;
1376
1377 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1378 /* Ensure channel is used */
1379 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1380 continue;
1381 }
1382
1383 /* Ensure not already written to (happens in case with GL_ALPHA) */
1384 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1385 continue;
1386 }
1387
1388 /* Ensure we havn't already found all channels */
1389 if (dst_channels >= out_format_desc->nr_channels) {
1390 continue;
1391 }
1392
1393 swizzle[out_format_desc->swizzle[i]] = i;
1394 ++dst_channels;
1395
1396 if (i == alpha_channel) {
1397 has_alpha = true;
1398 }
1399 }
1400
1401 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1402 /* the code above can't work for layout_other */
1403 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1404 has_alpha = true;
1405 swizzle[0] = 0;
1406 swizzle[1] = 1;
1407 swizzle[2] = 2;
1408 swizzle[3] = 3;
1409 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1410 }
1411
1412 /* If 3 channels then pad to include alpha for 4 element transpose */
1413 if (dst_channels == 3 && !has_alpha) {
1414 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1415 if (swizzle[i] > TGSI_NUM_CHANNELS)
1416 swizzle[i] = 3;
1417 }
1418 if (out_format_desc->nr_channels == 4) {
1419 dst_channels = 4;
1420 }
1421 }
1422
1423 /*
1424 * Load shader output
1425 */
1426 for (i = 0; i < num_fs; ++i) {
1427 /* Always load alpha for use in blending */
1428 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1429
1430 /* Load each channel */
1431 for (j = 0; j < dst_channels; ++j) {
1432 assert(swizzle[j] < 4);
1433 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1434 }
1435
1436 /* If 3 channels then pad to include alpha for 4 element transpose */
1437 /*
1438 * XXX If we include that here maybe could actually use it instead of
1439 * separate alpha for blending?
1440 */
1441 if (dst_channels == 3 && !has_alpha) {
1442 fs_src[i][3] = alpha;
1443 }
1444
1445 /* We split the row_mask and row_alpha as we want 128bit interleave */
1446 if (fs_type.length == 8) {
1447 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1448 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1449
1450 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1451 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1452 } else {
1453 src_mask[i] = fs_mask[i];
1454 src_alpha[i] = alpha;
1455 }
1456 }
1457 if (dual_source_blend) {
1458 /* same as above except different src/dst, skip masks and comments... */
1459 for (i = 0; i < num_fs; ++i) {
1460 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1461
1462 for (j = 0; j < dst_channels; ++j) {
1463 assert(swizzle[j] < 4);
1464 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1465 }
1466 if (dst_channels == 3 && !has_alpha) {
1467 fs_src1[i][3] = alpha;
1468 }
1469 if (fs_type.length == 8) {
1470 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1471 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1472 } else {
1473 src1_alpha[i] = alpha;
1474 }
1475 }
1476 }
1477
1478 if (util_format_is_pure_integer(out_format)) {
1479 /*
1480 * In this case fs_type was really ints or uints disguised as floats,
1481 * fix that up now.
1482 */
1483 fs_type.floating = 0;
1484 fs_type.sign = dst_type.sign;
1485 for (i = 0; i < num_fs; ++i) {
1486 for (j = 0; j < dst_channels; ++j) {
1487 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1488 lp_build_vec_type(gallivm, fs_type), "");
1489 }
1490 if (dst_channels == 3 && !has_alpha) {
1491 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1492 lp_build_vec_type(gallivm, fs_type), "");
1493 }
1494 }
1495 }
1496
1497 /*
1498 * Pixel twiddle from fragment shader order to memory order
1499 */
1500 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1501 if (dual_source_blend) {
1502 generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
1503 }
1504
1505 src_channels = dst_channels < 3 ? dst_channels : 4;
1506 if (src_count != num_fs * src_channels) {
1507 unsigned ds = src_count / (num_fs * src_channels);
1508 row_type.length /= ds;
1509 fs_type.length = row_type.length;
1510 }
1511
1512 blend_type = row_type;
1513 mask_type.length = 4;
1514
1515 /* Convert src to row_type */
1516 if (dual_source_blend) {
1517 struct lp_type old_row_type = row_type;
1518 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1519 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1520 }
1521 else {
1522 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1523 }
1524
1525 /* If the rows are not an SSE vector, combine them to become SSE size! */
1526 if ((row_type.width * row_type.length) % 128) {
1527 unsigned bits = row_type.width * row_type.length;
1528 unsigned combined;
1529
1530 dst_count = src_count / (vector_width / bits);
1531 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1532 if (dual_source_blend) {
1533 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1534 }
1535
1536 row_type.length *= combined;
1537 src_count /= combined;
1538
1539 bits = row_type.width * row_type.length;
1540 assert(bits == 128 || bits == 256);
1541 }
1542
1543
1544 /*
1545 * Blend Colour conversion
1546 */
1547 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1548 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1549 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1550
1551 /* Convert */
1552 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1553
1554 /* Extract alpha */
1555 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1556
1557 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1558 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1559 if (pad_inline) {
1560 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1561 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1562 } else {
1563 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1564 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1565 }
1566
1567 /*
1568 * Mask conversion
1569 */
1570 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1571
1572 if (src_count < block_height) {
1573 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1574 } else if (src_count > block_height) {
1575 for (i = src_count; i > 0; --i) {
1576 unsigned pixels = block_size / src_count;
1577 unsigned idx = i - 1;
1578
1579 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1580 }
1581 }
1582
1583 assert(mask_type.width == 32);
1584
1585 for (i = 0; i < src_count; ++i) {
1586 unsigned pixels = block_size / src_count;
1587 unsigned pixel_width = row_type.width * dst_channels;
1588
1589 if (pixel_width == 24) {
1590 mask_type.width = 8;
1591 mask_type.length = vector_width / mask_type.width;
1592 } else {
1593 mask_type.length = pixels;
1594 mask_type.width = row_type.width * dst_channels;
1595
1596 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1597
1598 mask_type.length *= dst_channels;
1599 mask_type.width /= dst_channels;
1600 }
1601
1602 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1603 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1604 }
1605
1606 /*
1607 * Alpha conversion
1608 */
1609 if (!has_alpha) {
1610 struct lp_type alpha_type = fs_type;
1611 alpha_type.length = 4;
1612 convert_alpha(gallivm, row_type, alpha_type,
1613 block_size, block_height,
1614 src_count, dst_channels,
1615 pad_inline, src_alpha);
1616 if (dual_source_blend) {
1617 convert_alpha(gallivm, row_type, alpha_type,
1618 block_size, block_height,
1619 src_count, dst_channels,
1620 pad_inline, src1_alpha);
1621 }
1622 }
1623
1624
1625 /*
1626 * Load dst from memory
1627 */
1628 if (src_count < block_height) {
1629 dst_count = block_height;
1630 } else {
1631 dst_count = src_count;
1632 }
1633
1634 dst_type.length *= 16 / dst_count;
1635
1636 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1637 /*
1638 * we need multiple values at once for the conversion, so can as well
1639 * load them vectorized here too instead of concatenating later.
1640 * (Still need concatenation later for 8-wide vectors).
1641 */
1642 dst_count = block_height;
1643 dst_type.length = block_width;
1644 }
1645
1646 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1647 dst, dst_type, dst_count, dst_alignment);
1648
1649
1650 /*
1651 * Convert from dst/output format to src/blending format.
1652 *
1653 * This is necessary as we can only read 1 row from memory at a time,
1654 * so the minimum dst_count will ever be at this point is 4.
1655 *
1656 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1657 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1658 * on all 16 pixels in that single vector at once.
1659 */
1660 if (dst_count > src_count) {
1661 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1662 }
1663
1664 /*
1665 * Blending
1666 */
1667 /* XXX this is broken for RGB8 formats -
1668 * they get expanded from 12 to 16 elements (to include alpha)
1669 * by convert_to_blend_type then reduced to 15 instead of 12
1670 * by convert_from_blend_type (a simple fix though breaks A8...).
1671 * R16G16B16 also crashes differently however something going wrong
1672 * inside llvm handling npot vector sizes seemingly.
1673 * It seems some cleanup could be done here (like skipping conversion/blend
1674 * when not needed).
1675 */
1676 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
1677
1678 for (i = 0; i < src_count; ++i) {
1679 dst[i] = lp_build_blend_aos(gallivm,
1680 &variant->key.blend,
1681 out_format,
1682 row_type,
1683 rt,
1684 src[i],
1685 has_alpha ? NULL : src_alpha[i],
1686 src1[i],
1687 has_alpha ? NULL : src1_alpha[i],
1688 dst[i],
1689 partial_mask ? src_mask[i] : NULL,
1690 blend_color,
1691 has_alpha ? NULL : blend_alpha,
1692 swizzle,
1693 pad_inline ? 4 : dst_channels);
1694 }
1695
1696 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
1697
1698 /* Split the blend rows back to memory rows */
1699 if (dst_count > src_count) {
1700 row_type.length = dst_type.length * (dst_count / src_count);
1701
1702 if (src_count == 1) {
1703 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1704 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1705
1706 row_type.length /= 2;
1707 src_count *= 2;
1708 }
1709
1710 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1711 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1712 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1713 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1714
1715 row_type.length /= 2;
1716 src_count *= 2;
1717 }
1718
1719
1720 /*
1721 * Store blend result to memory
1722 */
1723 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1724 dst, dst_type, dst_count, dst_alignment);
1725
1726 if (do_branch) {
1727 lp_build_mask_end(&mask_ctx);
1728 }
1729 }
1730
1731
1732 /**
1733 * Generate the runtime callable function for the whole fragment pipeline.
1734 * Note that the function which we generate operates on a block of 16
1735 * pixels at at time. The block contains 2x2 quads. Each quad contains
1736 * 2x2 pixels.
1737 */
1738 static void
1739 generate_fragment(struct llvmpipe_context *lp,
1740 struct lp_fragment_shader *shader,
1741 struct lp_fragment_shader_variant *variant,
1742 unsigned partial_mask)
1743 {
1744 struct gallivm_state *gallivm = variant->gallivm;
1745 const struct lp_fragment_shader_variant_key *key = &variant->key;
1746 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1747 char func_name[256];
1748 struct lp_type fs_type;
1749 struct lp_type blend_type;
1750 LLVMTypeRef fs_elem_type;
1751 LLVMTypeRef blend_vec_type;
1752 LLVMTypeRef arg_types[12];
1753 LLVMTypeRef func_type;
1754 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1755 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1756 LLVMValueRef context_ptr;
1757 LLVMValueRef x;
1758 LLVMValueRef y;
1759 LLVMValueRef a0_ptr;
1760 LLVMValueRef dadx_ptr;
1761 LLVMValueRef dady_ptr;
1762 LLVMValueRef color_ptr_ptr;
1763 LLVMValueRef stride_ptr;
1764 LLVMValueRef depth_ptr;
1765 LLVMValueRef mask_input;
1766 LLVMValueRef thread_data_ptr;
1767 LLVMBasicBlockRef block;
1768 LLVMBuilderRef builder;
1769 struct lp_build_sampler_soa *sampler;
1770 struct lp_build_interp_soa_context interp;
1771 LLVMValueRef fs_mask[16 / 4];
1772 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1773 LLVMValueRef function;
1774 LLVMValueRef facing;
1775 const struct util_format_description *zs_format_desc;
1776 unsigned num_fs;
1777 unsigned i;
1778 unsigned chan;
1779 unsigned cbuf;
1780 boolean cbuf0_write_all;
1781 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
1782 util_blend_state_is_dual(&key->blend, 0);
1783
1784 assert(lp_native_vector_width / 32 >= 4);
1785
1786 /* Adjust color input interpolation according to flatshade state:
1787 */
1788 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1789 for (i = 0; i < shader->info.base.num_inputs; i++) {
1790 if (inputs[i].interp == LP_INTERP_COLOR) {
1791 if (key->flatshade)
1792 inputs[i].interp = LP_INTERP_CONSTANT;
1793 else
1794 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1795 }
1796 }
1797
1798 /* check if writes to cbuf[0] are to be copied to all cbufs */
1799 cbuf0_write_all = FALSE;
1800 for (i = 0;i < shader->info.base.num_properties; i++) {
1801 if (shader->info.base.properties[i].name ==
1802 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1803 cbuf0_write_all = TRUE;
1804 break;
1805 }
1806 }
1807
1808 /* TODO: actually pick these based on the fs and color buffer
1809 * characteristics. */
1810
1811 memset(&fs_type, 0, sizeof fs_type);
1812 fs_type.floating = TRUE; /* floating point values */
1813 fs_type.sign = TRUE; /* values are signed */
1814 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1815 fs_type.width = 32; /* 32-bit float */
1816 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1817 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1818
1819 memset(&blend_type, 0, sizeof blend_type);
1820 blend_type.floating = FALSE; /* values are integers */
1821 blend_type.sign = FALSE; /* values are unsigned */
1822 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1823 blend_type.width = 8; /* 8-bit ubyte values */
1824 blend_type.length = 16; /* 16 elements per vector */
1825
1826 /*
1827 * Generate the function prototype. Any change here must be reflected in
1828 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1829 */
1830
1831 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1832
1833 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1834
1835 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1836 shader->no, variant->no, partial_mask ? "partial" : "whole");
1837
1838 arg_types[0] = variant->jit_context_ptr_type; /* context */
1839 arg_types[1] = int32_type; /* x */
1840 arg_types[2] = int32_type; /* y */
1841 arg_types[3] = int32_type; /* facing */
1842 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1843 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1844 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1845 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1846 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1847 arg_types[9] = int32_type; /* mask_input */
1848 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
1849 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1850
1851 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1852 arg_types, Elements(arg_types), 0);
1853
1854 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1855 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1856
1857 variant->function[partial_mask] = function;
1858
1859 /* XXX: need to propagate noalias down into color param now we are
1860 * passing a pointer-to-pointer?
1861 */
1862 for(i = 0; i < Elements(arg_types); ++i)
1863 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1864 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1865
1866 context_ptr = LLVMGetParam(function, 0);
1867 x = LLVMGetParam(function, 1);
1868 y = LLVMGetParam(function, 2);
1869 facing = LLVMGetParam(function, 3);
1870 a0_ptr = LLVMGetParam(function, 4);
1871 dadx_ptr = LLVMGetParam(function, 5);
1872 dady_ptr = LLVMGetParam(function, 6);
1873 color_ptr_ptr = LLVMGetParam(function, 7);
1874 depth_ptr = LLVMGetParam(function, 8);
1875 mask_input = LLVMGetParam(function, 9);
1876 thread_data_ptr = LLVMGetParam(function, 10);
1877 stride_ptr = LLVMGetParam(function, 11);
1878
1879 lp_build_name(context_ptr, "context");
1880 lp_build_name(x, "x");
1881 lp_build_name(y, "y");
1882 lp_build_name(a0_ptr, "a0");
1883 lp_build_name(dadx_ptr, "dadx");
1884 lp_build_name(dady_ptr, "dady");
1885 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
1886 lp_build_name(depth_ptr, "depth");
1887 lp_build_name(thread_data_ptr, "thread_data");
1888 lp_build_name(mask_input, "mask_input");
1889 lp_build_name(stride_ptr, "stride_ptr");
1890
1891 /*
1892 * Function body
1893 */
1894
1895 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1896 builder = gallivm->builder;
1897 assert(builder);
1898 LLVMPositionBuilderAtEnd(builder, block);
1899
1900 /* code generated texture sampling */
1901 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
1902
1903 zs_format_desc = util_format_description(key->zsbuf_format);
1904
1905 {
1906 unsigned depth_bits = zs_format_desc->block.bits/8;
1907 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
1908 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
1909 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
1910 num_loop, "mask_store");
1911 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1912
1913 /*
1914 * The shader input interpolation info is not explicitely baked in the
1915 * shader key, but everything it derives from (TGSI, and flatshade) is
1916 * already included in the shader key.
1917 */
1918 lp_build_interp_soa_init(&interp,
1919 gallivm,
1920 shader->info.base.num_inputs,
1921 inputs,
1922 builder, fs_type,
1923 a0_ptr, dadx_ptr, dady_ptr,
1924 x, y);
1925
1926 for (i = 0; i < num_fs; i++) {
1927 LLVMValueRef mask;
1928 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1929 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
1930 &indexi, 1, "mask_ptr");
1931
1932 if (partial_mask) {
1933 mask = generate_quad_mask(gallivm, fs_type,
1934 i*fs_type.length/4, mask_input);
1935 }
1936 else {
1937 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
1938 }
1939 LLVMBuildStore(builder, mask, mask_ptr);
1940 }
1941
1942 generate_fs_loop(gallivm,
1943 shader, key,
1944 builder,
1945 fs_type,
1946 context_ptr,
1947 num_loop,
1948 &interp,
1949 sampler,
1950 mask_store, /* output */
1951 color_store,
1952 depth_ptr,
1953 depth_bits,
1954 facing,
1955 thread_data_ptr);
1956
1957 for (i = 0; i < num_fs; i++) {
1958 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1959 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
1960 &indexi, 1, "");
1961 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
1962 /* This is fucked up need to reorganize things */
1963 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1964 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1965 ptr = LLVMBuildGEP(builder,
1966 color_store[cbuf * !cbuf0_write_all][chan],
1967 &indexi, 1, "");
1968 fs_out_color[cbuf][chan][i] = ptr;
1969 }
1970 }
1971 if (dual_source_blend) {
1972 /* only support one dual source blend target hence always use output 1 */
1973 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1974 ptr = LLVMBuildGEP(builder,
1975 color_store[1][chan],
1976 &indexi, 1, "");
1977 fs_out_color[1][chan][i] = ptr;
1978 }
1979 }
1980 }
1981 }
1982
1983 sampler->destroy(sampler);
1984
1985 /* Loop over color outputs / color buffers to do blending.
1986 */
1987 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1988 LLVMValueRef color_ptr;
1989 LLVMValueRef stride;
1990 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
1991
1992 boolean do_branch = ((key->depth.enabled
1993 || key->stencil[0].enabled
1994 || key->alpha.enabled)
1995 && !shader->info.base.uses_kill);
1996
1997 color_ptr = LLVMBuildLoad(builder,
1998 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
1999 "");
2000
2001 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2002
2003 stride = LLVMBuildLoad(builder,
2004 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2005 "");
2006
2007 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2008 num_fs, fs_type, fs_mask, fs_out_color,
2009 context_ptr, color_ptr, stride, partial_mask, do_branch);
2010 }
2011
2012 LLVMBuildRetVoid(builder);
2013
2014 gallivm_verify_function(gallivm, function);
2015
2016 variant->nr_instrs += lp_build_count_instructions(function);
2017 }
2018
2019
2020 static void
2021 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2022 {
2023 unsigned i;
2024
2025 debug_printf("fs variant %p:\n", (void *) key);
2026
2027 if (key->flatshade) {
2028 debug_printf("flatshade = 1\n");
2029 }
2030 for (i = 0; i < key->nr_cbufs; ++i) {
2031 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2032 }
2033 if (key->depth.enabled) {
2034 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2035 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2036 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2037 }
2038
2039 for (i = 0; i < 2; ++i) {
2040 if (key->stencil[i].enabled) {
2041 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2042 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2043 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2044 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2045 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2046 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2047 }
2048 }
2049
2050 if (key->alpha.enabled) {
2051 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2052 }
2053
2054 if (key->occlusion_count) {
2055 debug_printf("occlusion_count = 1\n");
2056 }
2057
2058 if (key->blend.logicop_enable) {
2059 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2060 }
2061 else if (key->blend.rt[0].blend_enable) {
2062 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2063 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2064 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2065 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2066 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2067 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2068 }
2069 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2070 for (i = 0; i < key->nr_samplers; ++i) {
2071 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2072 debug_printf("sampler[%u] = \n", i);
2073 debug_printf(" .wrap = %s %s %s\n",
2074 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2075 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2076 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2077 debug_printf(" .min_img_filter = %s\n",
2078 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2079 debug_printf(" .min_mip_filter = %s\n",
2080 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2081 debug_printf(" .mag_img_filter = %s\n",
2082 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2083 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2084 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2085 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2086 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2087 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2088 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2089 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2090 }
2091 for (i = 0; i < key->nr_sampler_views; ++i) {
2092 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2093 debug_printf("texture[%u] = \n", i);
2094 debug_printf(" .format = %s\n",
2095 util_format_name(texture->format));
2096 debug_printf(" .target = %s\n",
2097 util_dump_tex_target(texture->target, TRUE));
2098 debug_printf(" .level_zero_only = %u\n",
2099 texture->level_zero_only);
2100 debug_printf(" .pot = %u %u %u\n",
2101 texture->pot_width,
2102 texture->pot_height,
2103 texture->pot_depth);
2104 }
2105 }
2106
2107
2108 void
2109 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2110 {
2111 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2112 variant->shader->no, variant->no);
2113 tgsi_dump(variant->shader->base.tokens, 0);
2114 dump_fs_variant_key(&variant->key);
2115 debug_printf("variant->opaque = %u\n", variant->opaque);
2116 debug_printf("\n");
2117 }
2118
2119
2120 /**
2121 * Generate a new fragment shader variant from the shader code and
2122 * other state indicated by the key.
2123 */
2124 static struct lp_fragment_shader_variant *
2125 generate_variant(struct llvmpipe_context *lp,
2126 struct lp_fragment_shader *shader,
2127 const struct lp_fragment_shader_variant_key *key)
2128 {
2129 struct lp_fragment_shader_variant *variant;
2130 const struct util_format_description *cbuf0_format_desc;
2131 boolean fullcolormask;
2132
2133 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2134 if(!variant)
2135 return NULL;
2136
2137 variant->gallivm = gallivm_create();
2138 if (!variant->gallivm) {
2139 FREE(variant);
2140 return NULL;
2141 }
2142
2143 variant->shader = shader;
2144 variant->list_item_global.base = variant;
2145 variant->list_item_local.base = variant;
2146 variant->no = shader->variants_created++;
2147
2148 memcpy(&variant->key, key, shader->variant_key_size);
2149
2150 /*
2151 * Determine whether we are touching all channels in the color buffer.
2152 */
2153 fullcolormask = FALSE;
2154 if (key->nr_cbufs == 1) {
2155 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2156 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2157 }
2158
2159 variant->opaque =
2160 !key->blend.logicop_enable &&
2161 !key->blend.rt[0].blend_enable &&
2162 fullcolormask &&
2163 !key->stencil[0].enabled &&
2164 !key->alpha.enabled &&
2165 !key->depth.enabled &&
2166 !shader->info.base.uses_kill
2167 ? TRUE : FALSE;
2168
2169 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2170 lp_debug_fs_variant(variant);
2171 }
2172
2173 lp_jit_init_types(variant);
2174
2175 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2176 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2177
2178 if (variant->jit_function[RAST_WHOLE] == NULL) {
2179 if (variant->opaque) {
2180 /* Specialized shader, which doesn't need to read the color buffer. */
2181 generate_fragment(lp, shader, variant, RAST_WHOLE);
2182 }
2183 }
2184
2185 /*
2186 * Compile everything
2187 */
2188
2189 gallivm_compile_module(variant->gallivm);
2190
2191 if (variant->function[RAST_EDGE_TEST]) {
2192 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2193 gallivm_jit_function(variant->gallivm,
2194 variant->function[RAST_EDGE_TEST]);
2195 }
2196
2197 if (variant->function[RAST_WHOLE]) {
2198 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2199 gallivm_jit_function(variant->gallivm,
2200 variant->function[RAST_WHOLE]);
2201 } else if (!variant->jit_function[RAST_WHOLE]) {
2202 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2203 }
2204
2205 return variant;
2206 }
2207
2208
2209 static void *
2210 llvmpipe_create_fs_state(struct pipe_context *pipe,
2211 const struct pipe_shader_state *templ)
2212 {
2213 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2214 struct lp_fragment_shader *shader;
2215 int nr_samplers;
2216 int nr_sampler_views;
2217 int i;
2218
2219 shader = CALLOC_STRUCT(lp_fragment_shader);
2220 if (!shader)
2221 return NULL;
2222
2223 shader->no = fs_no++;
2224 make_empty_list(&shader->variants);
2225
2226 /* get/save the summary info for this shader */
2227 lp_build_tgsi_info(templ->tokens, &shader->info);
2228
2229 /* we need to keep a local copy of the tokens */
2230 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2231
2232 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2233 if (shader->draw_data == NULL) {
2234 FREE((void *) shader->base.tokens);
2235 FREE(shader);
2236 return NULL;
2237 }
2238
2239 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2240 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2241
2242 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2243 state[MAX2(nr_samplers, nr_sampler_views)]);
2244
2245 for (i = 0; i < shader->info.base.num_inputs; i++) {
2246 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2247 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2248
2249 switch (shader->info.base.input_interpolate[i]) {
2250 case TGSI_INTERPOLATE_CONSTANT:
2251 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2252 break;
2253 case TGSI_INTERPOLATE_LINEAR:
2254 shader->inputs[i].interp = LP_INTERP_LINEAR;
2255 break;
2256 case TGSI_INTERPOLATE_PERSPECTIVE:
2257 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2258 break;
2259 case TGSI_INTERPOLATE_COLOR:
2260 shader->inputs[i].interp = LP_INTERP_COLOR;
2261 break;
2262 default:
2263 assert(0);
2264 break;
2265 }
2266
2267 switch (shader->info.base.input_semantic_name[i]) {
2268 case TGSI_SEMANTIC_FACE:
2269 shader->inputs[i].interp = LP_INTERP_FACING;
2270 break;
2271 case TGSI_SEMANTIC_POSITION:
2272 /* Position was already emitted above
2273 */
2274 shader->inputs[i].interp = LP_INTERP_POSITION;
2275 shader->inputs[i].src_index = 0;
2276 continue;
2277 }
2278
2279 shader->inputs[i].src_index = i+1;
2280 }
2281
2282 if (LP_DEBUG & DEBUG_TGSI) {
2283 unsigned attrib;
2284 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2285 shader->no, (void *) shader);
2286 tgsi_dump(templ->tokens, 0);
2287 debug_printf("usage masks:\n");
2288 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2289 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2290 debug_printf(" IN[%u].%s%s%s%s\n",
2291 attrib,
2292 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2293 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2294 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2295 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2296 }
2297 debug_printf("\n");
2298 }
2299
2300 return shader;
2301 }
2302
2303
2304 static void
2305 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2306 {
2307 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2308
2309 if (llvmpipe->fs == fs)
2310 return;
2311
2312 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2313
2314 draw_bind_fragment_shader(llvmpipe->draw,
2315 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2316
2317 llvmpipe->dirty |= LP_NEW_FS;
2318 }
2319
2320
2321 /**
2322 * Remove shader variant from two lists: the shader's variant list
2323 * and the context's variant list.
2324 */
2325 void
2326 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2327 struct lp_fragment_shader_variant *variant)
2328 {
2329 unsigned i;
2330
2331 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2332 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2333 " #%u v total cached #%u\n",
2334 variant->shader->no,
2335 variant->no,
2336 variant->shader->variants_created,
2337 variant->shader->variants_cached,
2338 lp->nr_fs_variants);
2339 }
2340
2341 /* free all the variant's JIT'd functions */
2342 for (i = 0; i < Elements(variant->function); i++) {
2343 if (variant->function[i]) {
2344 gallivm_free_function(variant->gallivm,
2345 variant->function[i],
2346 variant->jit_function[i]);
2347 }
2348 }
2349
2350 gallivm_destroy(variant->gallivm);
2351
2352 /* remove from shader's list */
2353 remove_from_list(&variant->list_item_local);
2354 variant->shader->variants_cached--;
2355
2356 /* remove from context's list */
2357 remove_from_list(&variant->list_item_global);
2358 lp->nr_fs_variants--;
2359 lp->nr_fs_instrs -= variant->nr_instrs;
2360
2361 FREE(variant);
2362 }
2363
2364
2365 static void
2366 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2367 {
2368 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2369 struct lp_fragment_shader *shader = fs;
2370 struct lp_fs_variant_list_item *li;
2371
2372 assert(fs != llvmpipe->fs);
2373
2374 /*
2375 * XXX: we need to flush the context until we have some sort of reference
2376 * counting in fragment shaders as they may still be binned
2377 * Flushing alone might not sufficient we need to wait on it too.
2378 */
2379 llvmpipe_finish(pipe, __FUNCTION__);
2380
2381 /* Delete all the variants */
2382 li = first_elem(&shader->variants);
2383 while(!at_end(&shader->variants, li)) {
2384 struct lp_fs_variant_list_item *next = next_elem(li);
2385 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2386 li = next;
2387 }
2388
2389 /* Delete draw module's data */
2390 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2391
2392 assert(shader->variants_cached == 0);
2393 FREE((void *) shader->base.tokens);
2394 FREE(shader);
2395 }
2396
2397
2398
2399 static void
2400 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2401 uint shader, uint index,
2402 struct pipe_constant_buffer *cb)
2403 {
2404 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2405 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2406
2407 assert(shader < PIPE_SHADER_TYPES);
2408 assert(index < Elements(llvmpipe->constants[shader]));
2409
2410 /* note: reference counting */
2411 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2412
2413 if (shader == PIPE_SHADER_VERTEX ||
2414 shader == PIPE_SHADER_GEOMETRY) {
2415 /* Pass the constants to the 'draw' module */
2416 const unsigned size = cb ? cb->buffer_size : 0;
2417 const ubyte *data;
2418
2419 if (constants) {
2420 data = (ubyte *) llvmpipe_resource_data(constants);
2421 }
2422 else if (cb && cb->user_buffer) {
2423 data = (ubyte *) cb->user_buffer;
2424 }
2425 else {
2426 data = NULL;
2427 }
2428
2429 if (data)
2430 data += cb->buffer_offset;
2431
2432 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2433 index, data, size);
2434 }
2435
2436 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2437
2438 if (cb && cb->user_buffer) {
2439 pipe_resource_reference(&constants, NULL);
2440 }
2441 }
2442
2443
2444 /**
2445 * Return the blend factor equivalent to a destination alpha of one.
2446 */
2447 static INLINE unsigned
2448 force_dst_alpha_one(unsigned factor)
2449 {
2450 switch(factor) {
2451 case PIPE_BLENDFACTOR_DST_ALPHA:
2452 return PIPE_BLENDFACTOR_ONE;
2453 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2454 return PIPE_BLENDFACTOR_ZERO;
2455 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2456 return PIPE_BLENDFACTOR_ZERO;
2457 }
2458
2459 return factor;
2460 }
2461
2462
2463 /**
2464 * We need to generate several variants of the fragment pipeline to match
2465 * all the combinations of the contributing state atoms.
2466 *
2467 * TODO: there is actually no reason to tie this to context state -- the
2468 * generated code could be cached globally in the screen.
2469 */
2470 static void
2471 make_variant_key(struct llvmpipe_context *lp,
2472 struct lp_fragment_shader *shader,
2473 struct lp_fragment_shader_variant_key *key)
2474 {
2475 unsigned i;
2476
2477 memset(key, 0, shader->variant_key_size);
2478
2479 if (lp->framebuffer.zsbuf) {
2480 if (lp->depth_stencil->depth.enabled) {
2481 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2482 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2483 }
2484 if (lp->depth_stencil->stencil[0].enabled) {
2485 key->zsbuf_format = lp->framebuffer.zsbuf->format;
2486 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2487 }
2488 }
2489
2490 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2491 if (!lp->framebuffer.nr_cbufs ||
2492 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2493 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2494 }
2495 if(key->alpha.enabled)
2496 key->alpha.func = lp->depth_stencil->alpha.func;
2497 /* alpha.ref_value is passed in jit_context */
2498
2499 key->flatshade = lp->rasterizer->flatshade;
2500 if (lp->active_occlusion_query) {
2501 key->occlusion_count = TRUE;
2502 }
2503
2504 if (lp->framebuffer.nr_cbufs) {
2505 memcpy(&key->blend, lp->blend, sizeof key->blend);
2506 }
2507
2508 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2509
2510 if (!key->blend.independent_blend_enable) {
2511 /* we always need independent blend otherwise the fixups below won't work */
2512 for (i = 1; i < key->nr_cbufs; i++) {
2513 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2514 }
2515 key->blend.independent_blend_enable = 1;
2516 }
2517
2518 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2519 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2520 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2521 const struct util_format_description *format_desc;
2522
2523 key->cbuf_format[i] = format;
2524
2525 format_desc = util_format_description(format);
2526 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2527 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2528
2529 /*
2530 * Mask out color channels not present in the color buffer.
2531 */
2532 blend_rt->colormask &= util_format_colormask(format_desc);
2533
2534 /*
2535 * Disable blend for integer formats.
2536 */
2537 if (util_format_is_pure_integer(format)) {
2538 blend_rt->blend_enable = 0;
2539 }
2540
2541 /*
2542 * Our swizzled render tiles always have an alpha channel, but the linear
2543 * render target format often does not, so force here the dst alpha to be
2544 * one.
2545 *
2546 * This is not a mere optimization. Wrong results will be produced if the
2547 * dst alpha is used, the dst format does not have alpha, and the previous
2548 * rendering was not flushed from the swizzled to linear buffer. For
2549 * example, NonPowTwo DCT.
2550 *
2551 * TODO: This should be generalized to all channels for better
2552 * performance, but only alpha causes correctness issues.
2553 *
2554 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2555 */
2556 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2557 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2558 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2559 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2560 blend_rt->alpha_func = blend_rt->rgb_func;
2561 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2562 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2563 }
2564 }
2565
2566 /* This value will be the same for all the variants of a given shader:
2567 */
2568 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2569
2570 for(i = 0; i < key->nr_samplers; ++i) {
2571 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2572 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2573 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2574 }
2575 }
2576
2577 /*
2578 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2579 * are dx10-style? Can't really have mixed opcodes, at least not
2580 * if we want to skip the holes here (without rescanning tgsi).
2581 */
2582 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2583 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2584 for(i = 0; i < key->nr_sampler_views; ++i) {
2585 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2586 lp_sampler_static_texture_state(&key->state[i].texture_state,
2587 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2588 }
2589 }
2590 }
2591 else {
2592 key->nr_sampler_views = key->nr_samplers;
2593 for(i = 0; i < key->nr_sampler_views; ++i) {
2594 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2595 lp_sampler_static_texture_state(&key->state[i].texture_state,
2596 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2597 }
2598 }
2599 }
2600 }
2601
2602
2603
2604 /**
2605 * Update fragment shader state. This is called just prior to drawing
2606 * something when some fragment-related state has changed.
2607 */
2608 void
2609 llvmpipe_update_fs(struct llvmpipe_context *lp)
2610 {
2611 struct lp_fragment_shader *shader = lp->fs;
2612 struct lp_fragment_shader_variant_key key;
2613 struct lp_fragment_shader_variant *variant = NULL;
2614 struct lp_fs_variant_list_item *li;
2615
2616 make_variant_key(lp, shader, &key);
2617
2618 /* Search the variants for one which matches the key */
2619 li = first_elem(&shader->variants);
2620 while(!at_end(&shader->variants, li)) {
2621 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2622 variant = li->base;
2623 break;
2624 }
2625 li = next_elem(li);
2626 }
2627
2628 if (variant) {
2629 /* Move this variant to the head of the list to implement LRU
2630 * deletion of shader's when we have too many.
2631 */
2632 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2633 }
2634 else {
2635 /* variant not found, create it now */
2636 int64_t t0, t1, dt;
2637 unsigned i;
2638 unsigned variants_to_cull;
2639
2640 if (0) {
2641 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2642 lp->nr_fs_variants,
2643 lp->nr_fs_instrs,
2644 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2645 }
2646
2647 /* First, check if we've exceeded the max number of shader variants.
2648 * If so, free 25% of them (the least recently used ones).
2649 */
2650 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2651
2652 if (variants_to_cull ||
2653 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2654 struct pipe_context *pipe = &lp->pipe;
2655
2656 /*
2657 * XXX: we need to flush the context until we have some sort of
2658 * reference counting in fragment shaders as they may still be binned
2659 * Flushing alone might not be sufficient we need to wait on it too.
2660 */
2661 llvmpipe_finish(pipe, __FUNCTION__);
2662
2663 /*
2664 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2665 * number of shader variants (potentially all of them) could be
2666 * pending for destruction on flush.
2667 */
2668
2669 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2670 struct lp_fs_variant_list_item *item;
2671 if (is_empty_list(&lp->fs_variants_list)) {
2672 break;
2673 }
2674 item = last_elem(&lp->fs_variants_list);
2675 assert(item);
2676 assert(item->base);
2677 llvmpipe_remove_shader_variant(lp, item->base);
2678 }
2679 }
2680
2681 /*
2682 * Generate the new variant.
2683 */
2684 t0 = os_time_get();
2685 variant = generate_variant(lp, shader, &key);
2686 t1 = os_time_get();
2687 dt = t1 - t0;
2688 LP_COUNT_ADD(llvm_compile_time, dt);
2689 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2690
2691 llvmpipe_variant_count++;
2692
2693 /* Put the new variant into the list */
2694 if (variant) {
2695 insert_at_head(&shader->variants, &variant->list_item_local);
2696 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2697 lp->nr_fs_variants++;
2698 lp->nr_fs_instrs += variant->nr_instrs;
2699 shader->variants_cached++;
2700 }
2701 }
2702
2703 /* Bind this variant */
2704 lp_setup_set_fs_variant(lp->setup, variant);
2705 }
2706
2707
2708
2709
2710
2711 void
2712 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2713 {
2714 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2715 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2716 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2717
2718 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2719 }