projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: get rid of depth swizzling.
[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 LLVMValueRef depth_stride,
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 outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
245 struct lp_build_for_loop_state loop_state;
246 struct lp_build_mask_context mask;
247 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
248 shader->info.base.num_inputs < 3 &&
249 shader->info.base.num_instructions < 8);
250 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
251 util_blend_state_is_dual(&key->blend, 0);
252 unsigned attrib;
253 unsigned chan;
254 unsigned cbuf;
255 unsigned depth_mode;
256
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
292 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
293 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
294
295 vec_type = lp_build_vec_type(gallivm, type);
296
297 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
298
299 lp_build_for_loop_begin(&loop_state, gallivm,
300 lp_build_const_int32(gallivm, 0),
301 LLVMIntULT,
302 num_loop,
303 lp_build_const_int32(gallivm, 1));
304
305 mask_ptr = LLVMBuildGEP(builder, mask_store,
306 &loop_state.counter, 1, "mask_ptr");
307 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
308
309 memset(outputs, 0, sizeof outputs);
310
311 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
312 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
313 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
314 lp_build_vec_type(gallivm,
315 type),
316 num_loop, "color");
317 }
318 }
319 if (dual_source_blend) {
320 assert(key->nr_cbufs <= 1);
321 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
322 out_color[1][chan] = lp_build_array_alloca(gallivm,
323 lp_build_vec_type(gallivm,
324 type),
325 num_loop, "color1");
326 }
327 }
328
329
330 /* 'mask' will control execution based on quad's pixel alive/killed state */
331 lp_build_mask_begin(&mask, gallivm, type, mask_val);
332
333 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
334 lp_build_mask_check(&mask);
335
336 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
337 z = interp->pos[2];
338
339 if (depth_mode & EARLY_DEPTH_TEST) {
340 LLVMValueRef zs_dst_val;
341 zs_dst_val = lp_build_depth_stencil_load_swizzled(gallivm, type,
342 zs_format_desc,
343 depth_ptr, depth_stride,
344 loop_state.counter);
345 lp_build_depth_stencil_test(gallivm,
346 &key->depth,
347 key->stencil,
348 type,
349 zs_format_desc,
350 &mask,
351 stencil_refs,
352 z,
353 zs_dst_val,
354 facing,
355 &zs_value,
356 !simple_shader);
357
358 if (depth_mode & EARLY_DEPTH_WRITE) {
359 lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
360 NULL, loop_state.counter,
361 depth_ptr, depth_stride, 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 LLVMValueRef zs_dst_val;
398 int pos0 = find_output_by_semantic(&shader->info.base,
399 TGSI_SEMANTIC_POSITION,
400 0);
401
402 if (pos0 != -1 && outputs[pos0][2]) {
403 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
404 }
405
406 zs_dst_val = lp_build_depth_stencil_load_swizzled(gallivm, type,
407 zs_format_desc,
408 depth_ptr, depth_stride,
409 loop_state.counter);
410
411 lp_build_depth_stencil_test(gallivm,
412 &key->depth,
413 key->stencil,
414 type,
415 zs_format_desc,
416 &mask,
417 stencil_refs,
418 z,
419 zs_dst_val,
420 facing,
421 &zs_value,
422 !simple_shader);
423 /* Late Z write */
424 if (depth_mode & LATE_DEPTH_WRITE) {
425 lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
426 NULL, loop_state.counter,
427 depth_ptr, depth_stride, zs_value);
428 }
429 }
430 else if ((depth_mode & EARLY_DEPTH_TEST) &&
431 (depth_mode & LATE_DEPTH_WRITE))
432 {
433 /* Need to apply a reduced mask to the depth write. Reload the
434 * depth value, update from zs_value with the new mask value and
435 * write that out.
436 */
437 lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
438 &mask, loop_state.counter,
439 depth_ptr, depth_stride, zs_value);
440 }
441
442
443 /* Color write */
444 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
445 {
446 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
447 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
448 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
449 {
450 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
451 if(outputs[attrib][chan]) {
452 /* XXX: just initialize outputs to point at colors[] and
453 * skip this.
454 */
455 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
456 LLVMValueRef color_ptr;
457 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
458 &loop_state.counter, 1, "");
459 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
460 LLVMBuildStore(builder, out, color_ptr);
461 }
462 }
463 }
464 }
465
466 if (key->occlusion_count) {
467 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
468 lp_build_name(counter, "counter");
469 lp_build_occlusion_count(gallivm, type,
470 lp_build_mask_value(&mask), counter);
471 }
472
473 mask_val = lp_build_mask_end(&mask);
474 LLVMBuildStore(builder, mask_val, mask_ptr);
475 lp_build_for_loop_end(&loop_state);
476 }
477
478
479 /**
480 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
481 *
482 * Fragment Shader outputs pixels in small 2x2 blocks
483 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
484 *
485 * However in memory pixels are stored in rows
486 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
487 *
488 * @param type fragment shader type (4x or 8x float)
489 * @param num_fs number of fs_src
490 * @param dst_channels number of output channels
491 * @param fs_src output from fragment shader
492 * @param dst pointer to store result
493 * @param pad_inline is channel padding inline or at end of row
494 * @return the number of dsts
495 */
496 static int
497 generate_fs_twiddle(struct gallivm_state *gallivm,
498 struct lp_type type,
499 unsigned num_fs,
500 unsigned dst_channels,
501 LLVMValueRef fs_src[][4],
502 LLVMValueRef* dst,
503 bool pad_inline)
504 {
505 LLVMValueRef src[16];
506
507 bool swizzle_pad;
508 bool twiddle;
509 bool split;
510
511 unsigned pixels = num_fs == 4 ? 1 : 2;
512 unsigned reorder_group;
513 unsigned src_channels;
514 unsigned src_count;
515 unsigned i;
516
517 src_channels = dst_channels < 3 ? dst_channels : 4;
518 src_count = num_fs * src_channels;
519
520 assert(pixels == 2 || num_fs == 4);
521 assert(num_fs * src_channels <= Elements(src));
522
523 /*
524 * Transpose from SoA -> AoS
525 */
526 for (i = 0; i < num_fs; ++i) {
527 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
528 }
529
530 /*
531 * Pick transformation options
532 */
533 swizzle_pad = false;
534 twiddle = false;
535 split = false;
536 reorder_group = 0;
537
538 if (dst_channels == 1) {
539 twiddle = true;
540
541 if (pixels == 2) {
542 split = true;
543 }
544 } else if (dst_channels == 2) {
545 if (pixels == 1) {
546 reorder_group = 1;
547 }
548 } else if (dst_channels > 2) {
549 if (pixels == 1) {
550 reorder_group = 2;
551 } else {
552 twiddle = true;
553 }
554
555 if (!pad_inline && dst_channels == 3 && pixels > 1) {
556 swizzle_pad = true;
557 }
558 }
559
560 /*
561 * Split the src in half
562 */
563 if (split) {
564 for (i = num_fs; i > 0; --i) {
565 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
566 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
567 }
568
569 src_count *= 2;
570 type.length = 4;
571 }
572
573 /*
574 * Ensure pixels are in memory order
575 */
576 if (reorder_group) {
577 /* Twiddle pixels by reordering the array, e.g.:
578 *
579 * src_count = 8 -> 0 2 1 3 4 6 5 7
580 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
581 */
582 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
583
584 for (i = 0; i < src_count; ++i) {
585 unsigned group = i / reorder_group;
586 unsigned block = (group / 4) * 4 * reorder_group;
587 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
588 dst[i] = src[j];
589 }
590 } else if (twiddle) {
591 /* Twiddle pixels across elements of array */
592 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
593 } else {
594 /* Do nothing */
595 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
596 }
597
598 /*
599 * Moves any padding between pixels to the end
600 * e.g. RGBXRGBX -> RGBRGBXX
601 */
602 if (swizzle_pad) {
603 unsigned char swizzles[16];
604 unsigned elems = pixels * dst_channels;
605
606 for (i = 0; i < type.length; ++i) {
607 if (i < elems)
608 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
609 else
610 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
611 }
612
613 for (i = 0; i < src_count; ++i) {
614 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
615 }
616 }
617
618 return src_count;
619 }
620
621
622 /**
623 * Load an unswizzled block of pixels from memory
624 */
625 static void
626 load_unswizzled_block(struct gallivm_state *gallivm,
627 LLVMValueRef base_ptr,
628 LLVMValueRef stride,
629 unsigned block_width,
630 unsigned block_height,
631 LLVMValueRef* dst,
632 struct lp_type dst_type,
633 unsigned dst_count,
634 unsigned dst_alignment)
635 {
636 LLVMBuilderRef builder = gallivm->builder;
637 unsigned row_size = dst_count / block_height;
638 unsigned i;
639
640 /* Ensure block exactly fits into dst */
641 assert((block_width * block_height) % dst_count == 0);
642
643 for (i = 0; i < dst_count; ++i) {
644 unsigned x = i % row_size;
645 unsigned y = i / row_size;
646
647 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
648 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
649
650 LLVMValueRef gep[2];
651 LLVMValueRef dst_ptr;
652
653 gep[0] = lp_build_const_int32(gallivm, 0);
654 gep[1] = LLVMBuildAdd(builder, bx, by, "");
655
656 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
657 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
658
659 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
660
661 lp_set_load_alignment(dst[i], dst_alignment);
662 }
663 }
664
665
666 /**
667 * Store an unswizzled block of pixels to memory
668 */
669 static void
670 store_unswizzled_block(struct gallivm_state *gallivm,
671 LLVMValueRef base_ptr,
672 LLVMValueRef stride,
673 unsigned block_width,
674 unsigned block_height,
675 LLVMValueRef* src,
676 struct lp_type src_type,
677 unsigned src_count,
678 unsigned src_alignment)
679 {
680 LLVMBuilderRef builder = gallivm->builder;
681 unsigned row_size = src_count / block_height;
682 unsigned i;
683
684 /* Ensure src exactly fits into block */
685 assert((block_width * block_height) % src_count == 0);
686
687 for (i = 0; i < src_count; ++i) {
688 unsigned x = i % row_size;
689 unsigned y = i / row_size;
690
691 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
692 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
693
694 LLVMValueRef gep[2];
695 LLVMValueRef src_ptr;
696
697 gep[0] = lp_build_const_int32(gallivm, 0);
698 gep[1] = LLVMBuildAdd(builder, bx, by, "");
699
700 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
701 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
702
703 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
704
705 lp_set_store_alignment(src_ptr, src_alignment);
706 }
707 }
708
709
710 /**
711 * Checks if a format description is an arithmetic format
712 *
713 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
714 */
715 static INLINE boolean
716 is_arithmetic_format(const struct util_format_description *format_desc)
717 {
718 boolean arith = false;
719 unsigned i;
720
721 for (i = 0; i < format_desc->nr_channels; ++i) {
722 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
723 arith |= (format_desc->channel[i].size % 8) != 0;
724 }
725
726 return arith;
727 }
728
729
730 /**
731 * Retrieves the type representing the memory layout for a format
732 *
733 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
734 */
735 static INLINE void
736 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
737 struct lp_type* type)
738 {
739 unsigned i;
740 unsigned chan;
741
742 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
743 /* just make this a 32bit uint */
744 type->floating = false;
745 type->fixed = false;
746 type->sign = false;
747 type->norm = false;
748 type->width = 32;
749 type->length = 1;
750 return;
751 }
752
753 for (i = 0; i < 4; i++)
754 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
755 break;
756 chan = i;
757
758 memset(type, 0, sizeof(struct lp_type));
759 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
760 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
761 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
762 type->norm = format_desc->channel[chan].normalized;
763
764 if (is_arithmetic_format(format_desc)) {
765 type->width = 0;
766 type->length = 1;
767
768 for (i = 0; i < format_desc->nr_channels; ++i) {
769 type->width += format_desc->channel[i].size;
770 }
771 } else {
772 type->width = format_desc->channel[chan].size;
773 type->length = format_desc->nr_channels;
774 }
775 }
776
777
778 /**
779 * Retrieves the type for a format which is usable in the blending code.
780 *
781 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
782 */
783 static INLINE void
784 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
785 struct lp_type* type)
786 {
787 unsigned i;
788 unsigned chan;
789
790 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
791 /* always use ordinary floats for blending */
792 type->floating = true;
793 type->fixed = false;
794 type->sign = true;
795 type->norm = false;
796 type->width = 32;
797 type->length = 4;
798 return;
799 }
800
801 for (i = 0; i < 4; i++)
802 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
803 break;
804 chan = i;
805
806 memset(type, 0, sizeof(struct lp_type));
807 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
808 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
809 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
810 type->norm = format_desc->channel[chan].normalized;
811 type->width = format_desc->channel[chan].size;
812 type->length = format_desc->nr_channels;
813
814 for (i = 1; i < format_desc->nr_channels; ++i) {
815 if (format_desc->channel[i].size > type->width)
816 type->width = format_desc->channel[i].size;
817 }
818
819 if (type->floating) {
820 type->width = 32;
821 } else {
822 if (type->width <= 8) {
823 type->width = 8;
824 } else if (type->width <= 16) {
825 type->width = 16;
826 } else {
827 type->width = 32;
828 }
829 }
830
831 if (is_arithmetic_format(format_desc) && type->length == 3) {
832 type->length = 4;
833 }
834 }
835
836
837 /**
838 * Scale a normalized value from src_bits to dst_bits
839 */
840 static INLINE LLVMValueRef
841 scale_bits(struct gallivm_state *gallivm,
842 int src_bits,
843 int dst_bits,
844 LLVMValueRef src,
845 struct lp_type src_type)
846 {
847 LLVMBuilderRef builder = gallivm->builder;
848 LLVMValueRef result = src;
849
850 if (dst_bits < src_bits) {
851 /* Scale down by LShr */
852 result = LLVMBuildLShr(builder,
853 src,
854 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
855 "");
856 } else if (dst_bits > src_bits) {
857 /* Scale up bits */
858 int db = dst_bits - src_bits;
859
860 /* Shift left by difference in bits */
861 result = LLVMBuildShl(builder,
862 src,
863 lp_build_const_int_vec(gallivm, src_type, db),
864 "");
865
866 if (db < src_bits) {
867 /* Enough bits in src to fill the remainder */
868 LLVMValueRef lower = LLVMBuildLShr(builder,
869 src,
870 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
871 "");
872
873 result = LLVMBuildOr(builder, result, lower, "");
874 } else if (db > src_bits) {
875 /* Need to repeatedly copy src bits to fill remainder in dst */
876 unsigned n;
877
878 for (n = src_bits; n < dst_bits; n *= 2) {
879 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
880
881 result = LLVMBuildOr(builder,
882 result,
883 LLVMBuildLShr(builder, result, shuv, ""),
884 "");
885 }
886 }
887 }
888
889 return result;
890 }
891
892
893 /**
894 * Convert from memory format to blending format
895 *
896 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
897 */
898 static void
899 convert_to_blend_type(struct gallivm_state *gallivm,
900 const struct util_format_description *src_fmt,
901 struct lp_type src_type,
902 struct lp_type dst_type,
903 LLVMValueRef* src, // and dst
904 unsigned num_srcs)
905 {
906 LLVMValueRef *dst = src;
907 LLVMBuilderRef builder = gallivm->builder;
908 struct lp_type blend_type;
909 struct lp_type mem_type;
910 unsigned i, j, k;
911 unsigned pixels = 16 / num_srcs;
912 bool is_arith;
913
914 /*
915 * full custom path for packed floats - none of the later functions would do
916 * anything useful, and given the lp_type representation they can't be fixed.
917 */
918 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
919 LLVMValueRef tmpsrc[4];
920 /*
921 * This is pretty suboptimal for this case blending in SoA would be much
922 * better, since conversion gets us SoA values so need to convert back.
923 */
924 assert(src_type.width == 32);
925 assert(dst_type.floating);
926 assert(dst_type.width == 32);
927 assert(dst_type.length % 4 == 0);
928 for (i = 0; i < 4; i++) {
929 tmpsrc[i] = src[i];
930 }
931 for (i = 0; i < num_srcs / 4; i++) {
932 LLVMValueRef tmpsoa[4];
933 LLVMValueRef tmps = tmpsrc[i];
934 if (num_srcs == 8) {
935 LLVMValueRef shuffles[8];
936 unsigned j;
937 /* fetch was 4 values but need 8-wide output values */
938 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
939 /*
940 * for 8-wide aos transpose would give us wrong order not matching
941 * incoming converted fs values and mask. ARGH.
942 */
943 for (j = 0; j < 4; j++) {
944 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
945 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
946 }
947 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
948 LLVMConstVector(shuffles, 8), "");
949 }
950 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
951 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
952 }
953 return;
954 }
955
956 lp_mem_type_from_format_desc(src_fmt, &mem_type);
957 lp_blend_type_from_format_desc(src_fmt, &blend_type);
958
959 /* Is the format arithmetic */
960 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
961 is_arith &= !(mem_type.width == 16 && mem_type.floating);
962
963 /* Pad if necessary */
964 if (!is_arith && src_type.length < dst_type.length) {
965 for (i = 0; i < num_srcs; ++i) {
966 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
967 }
968
969 src_type.length = dst_type.length;
970 }
971
972 /* Special case for half-floats */
973 if (mem_type.width == 16 && mem_type.floating) {
974 assert(blend_type.width == 32 && blend_type.floating);
975 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
976 is_arith = false;
977 }
978
979 if (!is_arith) {
980 return;
981 }
982
983 src_type.width = blend_type.width * blend_type.length;
984 blend_type.length *= pixels;
985 src_type.length *= pixels / (src_type.length / mem_type.length);
986
987 for (i = 0; i < num_srcs; ++i) {
988 LLVMValueRef chans[4];
989 LLVMValueRef res = NULL;
990 unsigned sa = 0;
991
992 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
993
994 for (j = 0; j < src_fmt->nr_channels; ++j) {
995 unsigned mask = 0;
996
997 for (k = 0; k < src_fmt->channel[j].size; ++k) {
998 mask |= 1 << k;
999 }
1000
1001 /* Extract bits from source */
1002 chans[j] = LLVMBuildLShr(builder,
1003 dst[i],
1004 lp_build_const_int_vec(gallivm, src_type, sa),
1005 "");
1006
1007 chans[j] = LLVMBuildAnd(builder,
1008 chans[j],
1009 lp_build_const_int_vec(gallivm, src_type, mask),
1010 "");
1011
1012 /* Scale bits */
1013 if (src_type.norm) {
1014 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1015 blend_type.width, chans[j], src_type);
1016 }
1017
1018 /* Insert bits into correct position */
1019 chans[j] = LLVMBuildShl(builder,
1020 chans[j],
1021 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1022 "");
1023
1024 sa += src_fmt->channel[j].size;
1025
1026 if (j == 0) {
1027 res = chans[j];
1028 } else {
1029 res = LLVMBuildOr(builder, res, chans[j], "");
1030 }
1031 }
1032
1033 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1034 }
1035 }
1036
1037
1038 /**
1039 * Convert from blending format to memory format
1040 *
1041 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1042 */
1043 static void
1044 convert_from_blend_type(struct gallivm_state *gallivm,
1045 const struct util_format_description *src_fmt,
1046 struct lp_type src_type,
1047 struct lp_type dst_type,
1048 LLVMValueRef* src, // and dst
1049 unsigned num_srcs)
1050 {
1051 LLVMValueRef* dst = src;
1052 unsigned i, j, k;
1053 struct lp_type mem_type;
1054 struct lp_type blend_type;
1055 LLVMBuilderRef builder = gallivm->builder;
1056 unsigned pixels = 16 / num_srcs;
1057 bool is_arith;
1058
1059 /*
1060 * full custom path for packed floats - none of the later functions would do
1061 * anything useful, and given the lp_type representation they can't be fixed.
1062 */
1063 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1064 /*
1065 * This is pretty suboptimal for this case blending in SoA would be much
1066 * better - we need to transpose the AoS values back to SoA values for
1067 * conversion/packing.
1068 */
1069 assert(src_type.floating);
1070 assert(src_type.width == 32);
1071 assert(src_type.length % 4 == 0);
1072 assert(dst_type.width == 32);
1073 for (i = 0; i < num_srcs / 4; i++) {
1074 LLVMValueRef tmpsoa[4], tmpdst;
1075 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1076 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1077 if (num_srcs == 8) {
1078 LLVMValueRef tmpaos, shuffles[8];
1079 unsigned j;
1080 /*
1081 * for 8-wide aos transpose has given us wrong order not matching
1082 * output order. HMPF. Also need to split the output values manually.
1083 */
1084 for (j = 0; j < 4; j++) {
1085 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1086 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1087 }
1088 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1089 LLVMConstVector(shuffles, 8), "");
1090 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1091 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1092 }
1093 else {
1094 src[i] = tmpdst;
1095 }
1096 }
1097 return;
1098 }
1099
1100 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1101 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1102
1103 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1104
1105 /* Special case for half-floats */
1106 if (mem_type.width == 16 && mem_type.floating) {
1107 int length = dst_type.length;
1108 assert(blend_type.width == 32 && blend_type.floating);
1109
1110 dst_type.length = src_type.length;
1111
1112 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1113
1114 dst_type.length = length;
1115 is_arith = false;
1116 }
1117
1118 /* Remove any padding */
1119 if (!is_arith && (src_type.length % mem_type.length)) {
1120 src_type.length -= (src_type.length % mem_type.length);
1121
1122 for (i = 0; i < num_srcs; ++i) {
1123 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1124 }
1125 }
1126
1127 /* No bit arithmetic to do */
1128 if (!is_arith) {
1129 return;
1130 }
1131
1132 src_type.length = pixels;
1133 src_type.width = blend_type.length * blend_type.width;
1134 dst_type.length = pixels;
1135
1136 for (i = 0; i < num_srcs; ++i) {
1137 LLVMValueRef chans[4];
1138 LLVMValueRef res = NULL;
1139 unsigned sa = 0;
1140
1141 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1142
1143 for (j = 0; j < src_fmt->nr_channels; ++j) {
1144 unsigned mask = 0;
1145
1146 assert(blend_type.width > src_fmt->channel[j].size);
1147
1148 for (k = 0; k < blend_type.width; ++k) {
1149 mask |= 1 << k;
1150 }
1151
1152 /* Extract bits */
1153 chans[j] = LLVMBuildLShr(builder,
1154 dst[i],
1155 lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
1156 "");
1157
1158 chans[j] = LLVMBuildAnd(builder,
1159 chans[j],
1160 lp_build_const_int_vec(gallivm, src_type, mask),
1161 "");
1162
1163 /* Scale down bits */
1164 if (src_type.norm) {
1165 chans[j] = scale_bits(gallivm, blend_type.width,
1166 src_fmt->channel[j].size, chans[j], src_type);
1167 }
1168
1169 /* Insert bits */
1170 chans[j] = LLVMBuildShl(builder,
1171 chans[j],
1172 lp_build_const_int_vec(gallivm, src_type, sa),
1173 "");
1174
1175 sa += src_fmt->channel[j].size;
1176
1177 if (j == 0) {
1178 res = chans[j];
1179 } else {
1180 res = LLVMBuildOr(builder, res, chans[j], "");
1181 }
1182 }
1183
1184 assert (dst_type.width != 24);
1185
1186 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1187 }
1188 }
1189
1190
1191 /**
1192 * Convert alpha to same blend type as src
1193 */
1194 static void
1195 convert_alpha(struct gallivm_state *gallivm,
1196 struct lp_type row_type,
1197 struct lp_type alpha_type,
1198 const unsigned block_size,
1199 const unsigned block_height,
1200 const unsigned src_count,
1201 const unsigned dst_channels,
1202 const bool pad_inline,
1203 LLVMValueRef* src_alpha)
1204 {
1205 LLVMBuilderRef builder = gallivm->builder;
1206 unsigned i, j;
1207 unsigned length = row_type.length;
1208 row_type.length = alpha_type.length;
1209
1210 /* Twiddle the alpha to match pixels */
1211 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
1212
1213 for (i = 0; i < 4; ++i) {
1214 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1215 }
1216
1217 alpha_type = row_type;
1218 row_type.length = length;
1219
1220 /* If only one channel we can only need the single alpha value per pixel */
1221 if (src_count == 1) {
1222 assert(dst_channels == 1);
1223
1224 lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
1225 } else {
1226 /* If there are more srcs than rows then we need to split alpha up */
1227 if (src_count > block_height) {
1228 for (i = src_count; i > 0; --i) {
1229 unsigned pixels = block_size / src_count;
1230 unsigned idx = i - 1;
1231
1232 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1233 }
1234 }
1235
1236 /* If there is a src for each pixel broadcast the alpha across whole row */
1237 if (src_count == block_size) {
1238 for (i = 0; i < src_count; ++i) {
1239 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1240 }
1241 } else {
1242 unsigned pixels = block_size / src_count;
1243 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1244 unsigned alpha_span = 1;
1245 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1246
1247 /* Check if we need 2 src_alphas for our shuffles */
1248 if (pixels > alpha_type.length) {
1249 alpha_span = 2;
1250 }
1251
1252 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1253 for (j = 0; j < row_type.length; ++j) {
1254 if (j < pixels * channels) {
1255 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1256 } else {
1257 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1258 }
1259 }
1260
1261 for (i = 0; i < src_count; ++i) {
1262 unsigned idx1 = i, idx2 = i;
1263
1264 if (alpha_span > 1){
1265 idx1 *= alpha_span;
1266 idx2 = idx1 + 1;
1267 }
1268
1269 src_alpha[i] = LLVMBuildShuffleVector(builder,
1270 src_alpha[idx1],
1271 src_alpha[idx2],
1272 LLVMConstVector(shuffles, row_type.length),
1273 "");
1274 }
1275 }
1276 }
1277 }
1278
1279
1280 /**
1281 * Generates the blend function for unswizzled colour buffers
1282 * Also generates the read & write from colour buffer
1283 */
1284 static void
1285 generate_unswizzled_blend(struct gallivm_state *gallivm,
1286 unsigned rt,
1287 struct lp_fragment_shader_variant *variant,
1288 enum pipe_format out_format,
1289 unsigned int num_fs,
1290 struct lp_type fs_type,
1291 LLVMValueRef* fs_mask,
1292 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1293 LLVMValueRef context_ptr,
1294 LLVMValueRef color_ptr,
1295 LLVMValueRef stride,
1296 unsigned partial_mask,
1297 boolean do_branch)
1298 {
1299 const unsigned alpha_channel = 3;
1300 const unsigned block_width = 4;
1301 const unsigned block_height = 4;
1302 const unsigned block_size = block_width * block_height;
1303 const unsigned lp_integer_vector_width = 128;
1304
1305 LLVMBuilderRef builder = gallivm->builder;
1306 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1307 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1308 LLVMValueRef src_alpha[4 * 4];
1309 LLVMValueRef src1_alpha[4 * 4];
1310 LLVMValueRef src_mask[4 * 4];
1311 LLVMValueRef src[4 * 4];
1312 LLVMValueRef src1[4 * 4];
1313 LLVMValueRef dst[4 * 4];
1314 LLVMValueRef blend_color;
1315 LLVMValueRef blend_alpha;
1316 LLVMValueRef i32_zero;
1317 LLVMValueRef check_mask;
1318
1319 struct lp_build_mask_context mask_ctx;
1320 struct lp_type mask_type;
1321 struct lp_type blend_type;
1322 struct lp_type row_type;
1323 struct lp_type dst_type;
1324
1325 unsigned char swizzle[TGSI_NUM_CHANNELS];
1326 unsigned vector_width;
1327 unsigned src_channels = TGSI_NUM_CHANNELS;
1328 unsigned dst_channels;
1329 unsigned dst_count;
1330 unsigned src_count;
1331 unsigned i, j;
1332
1333 const struct util_format_description* out_format_desc = util_format_description(out_format);
1334
1335 unsigned dst_alignment;
1336
1337 bool pad_inline = is_arithmetic_format(out_format_desc);
1338 bool has_alpha = false;
1339 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1340 util_blend_state_is_dual(&variant->key.blend, 0);
1341
1342 mask_type = lp_int32_vec4_type();
1343 mask_type.length = fs_type.length;
1344
1345 /* Compute the alignment of the destination pointer in bytes */
1346 #if 0
1347 dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1348 #else
1349 /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
1350 dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1351 #endif
1352 /* Force power-of-two alignment by extracting only the least-significant-bit */
1353 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1354 /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
1355 dst_alignment = MIN2(dst_alignment, 16);
1356
1357 /* Do not bother executing code when mask is empty.. */
1358 if (do_branch) {
1359 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1360
1361 for (i = 0; i < num_fs; ++i) {
1362 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1363 }
1364
1365 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1366 lp_build_mask_check(&mask_ctx);
1367 }
1368
1369 partial_mask |= !variant->opaque;
1370 i32_zero = lp_build_const_int32(gallivm, 0);
1371
1372 /* Get type from output format */
1373 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1374 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1375
1376 row_type.length = fs_type.length;
1377 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1378
1379 /* Compute correct swizzle and count channels */
1380 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1381 dst_channels = 0;
1382
1383 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1384 /* Ensure channel is used */
1385 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1386 continue;
1387 }
1388
1389 /* Ensure not already written to (happens in case with GL_ALPHA) */
1390 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1391 continue;
1392 }
1393
1394 /* Ensure we havn't already found all channels */
1395 if (dst_channels >= out_format_desc->nr_channels) {
1396 continue;
1397 }
1398
1399 swizzle[out_format_desc->swizzle[i]] = i;
1400 ++dst_channels;
1401
1402 if (i == alpha_channel) {
1403 has_alpha = true;
1404 }
1405 }
1406
1407 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1408 /* the code above can't work for layout_other */
1409 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1410 has_alpha = true;
1411 swizzle[0] = 0;
1412 swizzle[1] = 1;
1413 swizzle[2] = 2;
1414 swizzle[3] = 3;
1415 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1416 }
1417
1418 /* If 3 channels then pad to include alpha for 4 element transpose */
1419 if (dst_channels == 3 && !has_alpha) {
1420 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1421 if (swizzle[i] > TGSI_NUM_CHANNELS)
1422 swizzle[i] = 3;
1423 }
1424 if (out_format_desc->nr_channels == 4) {
1425 dst_channels = 4;
1426 }
1427 }
1428
1429 /*
1430 * Load shader output
1431 */
1432 for (i = 0; i < num_fs; ++i) {
1433 /* Always load alpha for use in blending */
1434 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1435
1436 /* Load each channel */
1437 for (j = 0; j < dst_channels; ++j) {
1438 assert(swizzle[j] < 4);
1439 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1440 }
1441
1442 /* If 3 channels then pad to include alpha for 4 element transpose */
1443 /*
1444 * XXX If we include that here maybe could actually use it instead of
1445 * separate alpha for blending?
1446 */
1447 if (dst_channels == 3 && !has_alpha) {
1448 fs_src[i][3] = alpha;
1449 }
1450
1451 /* We split the row_mask and row_alpha as we want 128bit interleave */
1452 if (fs_type.length == 8) {
1453 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1454 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1455
1456 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1457 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1458 } else {
1459 src_mask[i] = fs_mask[i];
1460 src_alpha[i] = alpha;
1461 }
1462 }
1463 if (dual_source_blend) {
1464 /* same as above except different src/dst, skip masks and comments... */
1465 for (i = 0; i < num_fs; ++i) {
1466 LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1467
1468 for (j = 0; j < dst_channels; ++j) {
1469 assert(swizzle[j] < 4);
1470 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1471 }
1472 if (dst_channels == 3 && !has_alpha) {
1473 fs_src1[i][3] = alpha;
1474 }
1475 if (fs_type.length == 8) {
1476 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1477 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1478 } else {
1479 src1_alpha[i] = alpha;
1480 }
1481 }
1482 }
1483
1484 if (util_format_is_pure_integer(out_format)) {
1485 /*
1486 * In this case fs_type was really ints or uints disguised as floats,
1487 * fix that up now.
1488 */
1489 fs_type.floating = 0;
1490 fs_type.sign = dst_type.sign;
1491 for (i = 0; i < num_fs; ++i) {
1492 for (j = 0; j < dst_channels; ++j) {
1493 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1494 lp_build_vec_type(gallivm, fs_type), "");
1495 }
1496 if (dst_channels == 3 && !has_alpha) {
1497 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1498 lp_build_vec_type(gallivm, fs_type), "");
1499 }
1500 }
1501 }
1502
1503 /*
1504 * Pixel twiddle from fragment shader order to memory order
1505 */
1506 src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
1507 if (dual_source_blend) {
1508 generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
1509 }
1510
1511 src_channels = dst_channels < 3 ? dst_channels : 4;
1512 if (src_count != num_fs * src_channels) {
1513 unsigned ds = src_count / (num_fs * src_channels);
1514 row_type.length /= ds;
1515 fs_type.length = row_type.length;
1516 }
1517
1518 blend_type = row_type;
1519 mask_type.length = 4;
1520
1521 /* Convert src to row_type */
1522 if (dual_source_blend) {
1523 struct lp_type old_row_type = row_type;
1524 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1525 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1526 }
1527 else {
1528 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1529 }
1530
1531 /* If the rows are not an SSE vector, combine them to become SSE size! */
1532 if ((row_type.width * row_type.length) % 128) {
1533 unsigned bits = row_type.width * row_type.length;
1534 unsigned combined;
1535
1536 dst_count = src_count / (vector_width / bits);
1537 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1538 if (dual_source_blend) {
1539 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1540 }
1541
1542 row_type.length *= combined;
1543 src_count /= combined;
1544
1545 bits = row_type.width * row_type.length;
1546 assert(bits == 128 || bits == 256);
1547 }
1548
1549
1550 /*
1551 * Blend Colour conversion
1552 */
1553 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1554 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1555 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1556
1557 /* Convert */
1558 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1559
1560 /* Extract alpha */
1561 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1562
1563 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1564 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1565 if (pad_inline) {
1566 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1567 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1568 } else {
1569 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1570 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1571 }
1572
1573 /*
1574 * Mask conversion
1575 */
1576 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
1577
1578 if (src_count < block_height) {
1579 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1580 } else if (src_count > block_height) {
1581 for (i = src_count; i > 0; --i) {
1582 unsigned pixels = block_size / src_count;
1583 unsigned idx = i - 1;
1584
1585 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
1586 }
1587 }
1588
1589 assert(mask_type.width == 32);
1590
1591 for (i = 0; i < src_count; ++i) {
1592 unsigned pixels = block_size / src_count;
1593 unsigned pixel_width = row_type.width * dst_channels;
1594
1595 if (pixel_width == 24) {
1596 mask_type.width = 8;
1597 mask_type.length = vector_width / mask_type.width;
1598 } else {
1599 mask_type.length = pixels;
1600 mask_type.width = row_type.width * dst_channels;
1601
1602 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1603
1604 mask_type.length *= dst_channels;
1605 mask_type.width /= dst_channels;
1606 }
1607
1608 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1609 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1610 }
1611
1612 /*
1613 * Alpha conversion
1614 */
1615 if (!has_alpha) {
1616 struct lp_type alpha_type = fs_type;
1617 alpha_type.length = 4;
1618 convert_alpha(gallivm, row_type, alpha_type,
1619 block_size, block_height,
1620 src_count, dst_channels,
1621 pad_inline, src_alpha);
1622 if (dual_source_blend) {
1623 convert_alpha(gallivm, row_type, alpha_type,
1624 block_size, block_height,
1625 src_count, dst_channels,
1626 pad_inline, src1_alpha);
1627 }
1628 }
1629
1630
1631 /*
1632 * Load dst from memory
1633 */
1634 if (src_count < block_height) {
1635 dst_count = block_height;
1636 } else {
1637 dst_count = src_count;
1638 }
1639
1640 dst_type.length *= 16 / dst_count;
1641
1642 if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
1643 /*
1644 * we need multiple values at once for the conversion, so can as well
1645 * load them vectorized here too instead of concatenating later.
1646 * (Still need concatenation later for 8-wide vectors).
1647 */
1648 dst_count = block_height;
1649 dst_type.length = block_width;
1650 }
1651
1652 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1653 dst, dst_type, dst_count, dst_alignment);
1654
1655
1656 /*
1657 * Convert from dst/output format to src/blending format.
1658 *
1659 * This is necessary as we can only read 1 row from memory at a time,
1660 * so the minimum dst_count will ever be at this point is 4.
1661 *
1662 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1663 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1664 * on all 16 pixels in that single vector at once.
1665 */
1666 if (dst_count > src_count) {
1667 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1668 }
1669
1670 /*
1671 * Blending
1672 */
1673 /* XXX this is broken for RGB8 formats -
1674 * they get expanded from 12 to 16 elements (to include alpha)
1675 * by convert_to_blend_type then reduced to 15 instead of 12
1676 * by convert_from_blend_type (a simple fix though breaks A8...).
1677 * R16G16B16 also crashes differently however something going wrong
1678 * inside llvm handling npot vector sizes seemingly.
1679 * It seems some cleanup could be done here (like skipping conversion/blend
1680 * when not needed).
1681 */
1682 convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
1683
1684 for (i = 0; i < src_count; ++i) {
1685 dst[i] = lp_build_blend_aos(gallivm,
1686 &variant->key.blend,
1687 out_format,
1688 row_type,
1689 rt,
1690 src[i],
1691 has_alpha ? NULL : src_alpha[i],
1692 src1[i],
1693 has_alpha ? NULL : src1_alpha[i],
1694 dst[i],
1695 partial_mask ? src_mask[i] : NULL,
1696 blend_color,
1697 has_alpha ? NULL : blend_alpha,
1698 swizzle,
1699 pad_inline ? 4 : dst_channels);
1700 }
1701
1702 convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
1703
1704 /* Split the blend rows back to memory rows */
1705 if (dst_count > src_count) {
1706 row_type.length = dst_type.length * (dst_count / src_count);
1707
1708 if (src_count == 1) {
1709 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1710 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1711
1712 row_type.length /= 2;
1713 src_count *= 2;
1714 }
1715
1716 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1717 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1718 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1719 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1720
1721 row_type.length /= 2;
1722 src_count *= 2;
1723 }
1724
1725
1726 /*
1727 * Store blend result to memory
1728 */
1729 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1730 dst, dst_type, dst_count, dst_alignment);
1731
1732 if (do_branch) {
1733 lp_build_mask_end(&mask_ctx);
1734 }
1735 }
1736
1737
1738 /**
1739 * Generate the runtime callable function for the whole fragment pipeline.
1740 * Note that the function which we generate operates on a block of 16
1741 * pixels at at time. The block contains 2x2 quads. Each quad contains
1742 * 2x2 pixels.
1743 */
1744 static void
1745 generate_fragment(struct llvmpipe_context *lp,
1746 struct lp_fragment_shader *shader,
1747 struct lp_fragment_shader_variant *variant,
1748 unsigned partial_mask)
1749 {
1750 struct gallivm_state *gallivm = variant->gallivm;
1751 const struct lp_fragment_shader_variant_key *key = &variant->key;
1752 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1753 char func_name[256];
1754 struct lp_type fs_type;
1755 struct lp_type blend_type;
1756 LLVMTypeRef fs_elem_type;
1757 LLVMTypeRef blend_vec_type;
1758 LLVMTypeRef arg_types[13];
1759 LLVMTypeRef func_type;
1760 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1761 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1762 LLVMValueRef context_ptr;
1763 LLVMValueRef x;
1764 LLVMValueRef y;
1765 LLVMValueRef a0_ptr;
1766 LLVMValueRef dadx_ptr;
1767 LLVMValueRef dady_ptr;
1768 LLVMValueRef color_ptr_ptr;
1769 LLVMValueRef stride_ptr;
1770 LLVMValueRef depth_ptr;
1771 LLVMValueRef depth_stride;
1772 LLVMValueRef mask_input;
1773 LLVMValueRef thread_data_ptr;
1774 LLVMBasicBlockRef block;
1775 LLVMBuilderRef builder;
1776 struct lp_build_sampler_soa *sampler;
1777 struct lp_build_interp_soa_context interp;
1778 LLVMValueRef fs_mask[16 / 4];
1779 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1780 LLVMValueRef function;
1781 LLVMValueRef facing;
1782 unsigned num_fs;
1783 unsigned i;
1784 unsigned chan;
1785 unsigned cbuf;
1786 boolean cbuf0_write_all;
1787 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
1788 util_blend_state_is_dual(&key->blend, 0);
1789
1790 assert(lp_native_vector_width / 32 >= 4);
1791
1792 /* Adjust color input interpolation according to flatshade state:
1793 */
1794 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1795 for (i = 0; i < shader->info.base.num_inputs; i++) {
1796 if (inputs[i].interp == LP_INTERP_COLOR) {
1797 if (key->flatshade)
1798 inputs[i].interp = LP_INTERP_CONSTANT;
1799 else
1800 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1801 }
1802 }
1803
1804 /* check if writes to cbuf[0] are to be copied to all cbufs */
1805 cbuf0_write_all = FALSE;
1806 for (i = 0;i < shader->info.base.num_properties; i++) {
1807 if (shader->info.base.properties[i].name ==
1808 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1809 cbuf0_write_all = TRUE;
1810 break;
1811 }
1812 }
1813
1814 /* TODO: actually pick these based on the fs and color buffer
1815 * characteristics. */
1816
1817 memset(&fs_type, 0, sizeof fs_type);
1818 fs_type.floating = TRUE; /* floating point values */
1819 fs_type.sign = TRUE; /* values are signed */
1820 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1821 fs_type.width = 32; /* 32-bit float */
1822 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1823 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
1824
1825 memset(&blend_type, 0, sizeof blend_type);
1826 blend_type.floating = FALSE; /* values are integers */
1827 blend_type.sign = FALSE; /* values are unsigned */
1828 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1829 blend_type.width = 8; /* 8-bit ubyte values */
1830 blend_type.length = 16; /* 16 elements per vector */
1831
1832 /*
1833 * Generate the function prototype. Any change here must be reflected in
1834 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1835 */
1836
1837 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1838
1839 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1840
1841 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1842 shader->no, variant->no, partial_mask ? "partial" : "whole");
1843
1844 arg_types[0] = variant->jit_context_ptr_type; /* context */
1845 arg_types[1] = int32_type; /* x */
1846 arg_types[2] = int32_type; /* y */
1847 arg_types[3] = int32_type; /* facing */
1848 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1849 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
1850 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
1851 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
1852 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
1853 arg_types[9] = int32_type; /* mask_input */
1854 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
1855 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
1856 arg_types[12] = int32_type; /* depth_stride */
1857
1858 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
1859 arg_types, Elements(arg_types), 0);
1860
1861 function = LLVMAddFunction(gallivm->module, func_name, func_type);
1862 LLVMSetFunctionCallConv(function, LLVMCCallConv);
1863
1864 variant->function[partial_mask] = function;
1865
1866 /* XXX: need to propagate noalias down into color param now we are
1867 * passing a pointer-to-pointer?
1868 */
1869 for(i = 0; i < Elements(arg_types); ++i)
1870 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1871 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
1872
1873 context_ptr = LLVMGetParam(function, 0);
1874 x = LLVMGetParam(function, 1);
1875 y = LLVMGetParam(function, 2);
1876 facing = LLVMGetParam(function, 3);
1877 a0_ptr = LLVMGetParam(function, 4);
1878 dadx_ptr = LLVMGetParam(function, 5);
1879 dady_ptr = LLVMGetParam(function, 6);
1880 color_ptr_ptr = LLVMGetParam(function, 7);
1881 depth_ptr = LLVMGetParam(function, 8);
1882 mask_input = LLVMGetParam(function, 9);
1883 thread_data_ptr = LLVMGetParam(function, 10);
1884 stride_ptr = LLVMGetParam(function, 11);
1885 depth_stride = LLVMGetParam(function, 12);
1886
1887 lp_build_name(context_ptr, "context");
1888 lp_build_name(x, "x");
1889 lp_build_name(y, "y");
1890 lp_build_name(a0_ptr, "a0");
1891 lp_build_name(dadx_ptr, "dadx");
1892 lp_build_name(dady_ptr, "dady");
1893 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
1894 lp_build_name(depth_ptr, "depth");
1895 lp_build_name(thread_data_ptr, "thread_data");
1896 lp_build_name(mask_input, "mask_input");
1897 lp_build_name(stride_ptr, "stride_ptr");
1898 lp_build_name(depth_stride, "depth_stride");
1899
1900 /*
1901 * Function body
1902 */
1903
1904 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
1905 builder = gallivm->builder;
1906 assert(builder);
1907 LLVMPositionBuilderAtEnd(builder, block);
1908
1909 /* code generated texture sampling */
1910 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
1911
1912 {
1913 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
1914 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
1915 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
1916 num_loop, "mask_store");
1917 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
1918
1919 /*
1920 * The shader input interpolation info is not explicitely baked in the
1921 * shader key, but everything it derives from (TGSI, and flatshade) is
1922 * already included in the shader key.
1923 */
1924 lp_build_interp_soa_init(&interp,
1925 gallivm,
1926 shader->info.base.num_inputs,
1927 inputs,
1928 shader->info.base.pixel_center_integer,
1929 builder, fs_type,
1930 a0_ptr, dadx_ptr, dady_ptr,
1931 x, y);
1932
1933 for (i = 0; i < num_fs; i++) {
1934 LLVMValueRef mask;
1935 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1936 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
1937 &indexi, 1, "mask_ptr");
1938
1939 if (partial_mask) {
1940 mask = generate_quad_mask(gallivm, fs_type,
1941 i*fs_type.length/4, mask_input);
1942 }
1943 else {
1944 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
1945 }
1946 LLVMBuildStore(builder, mask, mask_ptr);
1947 }
1948
1949 generate_fs_loop(gallivm,
1950 shader, key,
1951 builder,
1952 fs_type,
1953 context_ptr,
1954 num_loop,
1955 &interp,
1956 sampler,
1957 mask_store, /* output */
1958 color_store,
1959 depth_ptr,
1960 depth_stride,
1961 facing,
1962 thread_data_ptr);
1963
1964 for (i = 0; i < num_fs; i++) {
1965 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1966 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
1967 &indexi, 1, "");
1968 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
1969 /* This is fucked up need to reorganize things */
1970 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1971 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1972 ptr = LLVMBuildGEP(builder,
1973 color_store[cbuf * !cbuf0_write_all][chan],
1974 &indexi, 1, "");
1975 fs_out_color[cbuf][chan][i] = ptr;
1976 }
1977 }
1978 if (dual_source_blend) {
1979 /* only support one dual source blend target hence always use output 1 */
1980 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1981 ptr = LLVMBuildGEP(builder,
1982 color_store[1][chan],
1983 &indexi, 1, "");
1984 fs_out_color[1][chan][i] = ptr;
1985 }
1986 }
1987 }
1988 }
1989
1990 sampler->destroy(sampler);
1991
1992 /* Loop over color outputs / color buffers to do blending.
1993 */
1994 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
1995 LLVMValueRef color_ptr;
1996 LLVMValueRef stride;
1997 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
1998
1999 boolean do_branch = ((key->depth.enabled
2000 || key->stencil[0].enabled
2001 || key->alpha.enabled)
2002 && !shader->info.base.uses_kill);
2003
2004 color_ptr = LLVMBuildLoad(builder,
2005 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2006 "");
2007
2008 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2009
2010 stride = LLVMBuildLoad(builder,
2011 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2012 "");
2013
2014 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2015 num_fs, fs_type, fs_mask, fs_out_color,
2016 context_ptr, color_ptr, stride, partial_mask, do_branch);
2017 }
2018
2019 LLVMBuildRetVoid(builder);
2020
2021 gallivm_verify_function(gallivm, function);
2022
2023 variant->nr_instrs += lp_build_count_instructions(function);
2024 }
2025
2026
2027 static void
2028 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2029 {
2030 unsigned i;
2031
2032 debug_printf("fs variant %p:\n", (void *) key);
2033
2034 if (key->flatshade) {
2035 debug_printf("flatshade = 1\n");
2036 }
2037 for (i = 0; i < key->nr_cbufs; ++i) {
2038 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2039 }
2040 if (key->depth.enabled) {
2041 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2042 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2043 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2044 }
2045
2046 for (i = 0; i < 2; ++i) {
2047 if (key->stencil[i].enabled) {
2048 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2049 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2050 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2051 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2052 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2053 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2054 }
2055 }
2056
2057 if (key->alpha.enabled) {
2058 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2059 }
2060
2061 if (key->occlusion_count) {
2062 debug_printf("occlusion_count = 1\n");
2063 }
2064
2065 if (key->blend.logicop_enable) {
2066 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2067 }
2068 else if (key->blend.rt[0].blend_enable) {
2069 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2070 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2071 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2072 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2073 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2074 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2075 }
2076 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2077 for (i = 0; i < key->nr_samplers; ++i) {
2078 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2079 debug_printf("sampler[%u] = \n", i);
2080 debug_printf(" .wrap = %s %s %s\n",
2081 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2082 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2083 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2084 debug_printf(" .min_img_filter = %s\n",
2085 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2086 debug_printf(" .min_mip_filter = %s\n",
2087 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2088 debug_printf(" .mag_img_filter = %s\n",
2089 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2090 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2091 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2092 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2093 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2094 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2095 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2096 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2097 }
2098 for (i = 0; i < key->nr_sampler_views; ++i) {
2099 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2100 debug_printf("texture[%u] = \n", i);
2101 debug_printf(" .format = %s\n",
2102 util_format_name(texture->format));
2103 debug_printf(" .target = %s\n",
2104 util_dump_tex_target(texture->target, TRUE));
2105 debug_printf(" .level_zero_only = %u\n",
2106 texture->level_zero_only);
2107 debug_printf(" .pot = %u %u %u\n",
2108 texture->pot_width,
2109 texture->pot_height,
2110 texture->pot_depth);
2111 }
2112 }
2113
2114
2115 void
2116 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2117 {
2118 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2119 variant->shader->no, variant->no);
2120 tgsi_dump(variant->shader->base.tokens, 0);
2121 dump_fs_variant_key(&variant->key);
2122 debug_printf("variant->opaque = %u\n", variant->opaque);
2123 debug_printf("\n");
2124 }
2125
2126
2127 /**
2128 * Generate a new fragment shader variant from the shader code and
2129 * other state indicated by the key.
2130 */
2131 static struct lp_fragment_shader_variant *
2132 generate_variant(struct llvmpipe_context *lp,
2133 struct lp_fragment_shader *shader,
2134 const struct lp_fragment_shader_variant_key *key)
2135 {
2136 struct lp_fragment_shader_variant *variant;
2137 const struct util_format_description *cbuf0_format_desc;
2138 boolean fullcolormask;
2139
2140 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2141 if(!variant)
2142 return NULL;
2143
2144 variant->gallivm = gallivm_create();
2145 if (!variant->gallivm) {
2146 FREE(variant);
2147 return NULL;
2148 }
2149
2150 variant->shader = shader;
2151 variant->list_item_global.base = variant;
2152 variant->list_item_local.base = variant;
2153 variant->no = shader->variants_created++;
2154
2155 memcpy(&variant->key, key, shader->variant_key_size);
2156
2157 /*
2158 * Determine whether we are touching all channels in the color buffer.
2159 */
2160 fullcolormask = FALSE;
2161 if (key->nr_cbufs == 1) {
2162 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2163 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2164 }
2165
2166 variant->opaque =
2167 !key->blend.logicop_enable &&
2168 !key->blend.rt[0].blend_enable &&
2169 fullcolormask &&
2170 !key->stencil[0].enabled &&
2171 !key->alpha.enabled &&
2172 !key->depth.enabled &&
2173 !shader->info.base.uses_kill
2174 ? TRUE : FALSE;
2175
2176 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2177 lp_debug_fs_variant(variant);
2178 }
2179
2180 lp_jit_init_types(variant);
2181
2182 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2183 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2184
2185 if (variant->jit_function[RAST_WHOLE] == NULL) {
2186 if (variant->opaque) {
2187 /* Specialized shader, which doesn't need to read the color buffer. */
2188 generate_fragment(lp, shader, variant, RAST_WHOLE);
2189 }
2190 }
2191
2192 /*
2193 * Compile everything
2194 */
2195
2196 gallivm_compile_module(variant->gallivm);
2197
2198 if (variant->function[RAST_EDGE_TEST]) {
2199 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2200 gallivm_jit_function(variant->gallivm,
2201 variant->function[RAST_EDGE_TEST]);
2202 }
2203
2204 if (variant->function[RAST_WHOLE]) {
2205 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2206 gallivm_jit_function(variant->gallivm,
2207 variant->function[RAST_WHOLE]);
2208 } else if (!variant->jit_function[RAST_WHOLE]) {
2209 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2210 }
2211
2212 return variant;
2213 }
2214
2215
2216 static void *
2217 llvmpipe_create_fs_state(struct pipe_context *pipe,
2218 const struct pipe_shader_state *templ)
2219 {
2220 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2221 struct lp_fragment_shader *shader;
2222 int nr_samplers;
2223 int nr_sampler_views;
2224 int i;
2225
2226 shader = CALLOC_STRUCT(lp_fragment_shader);
2227 if (!shader)
2228 return NULL;
2229
2230 shader->no = fs_no++;
2231 make_empty_list(&shader->variants);
2232
2233 /* get/save the summary info for this shader */
2234 lp_build_tgsi_info(templ->tokens, &shader->info);
2235
2236 /* we need to keep a local copy of the tokens */
2237 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2238
2239 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2240 if (shader->draw_data == NULL) {
2241 FREE((void *) shader->base.tokens);
2242 FREE(shader);
2243 return NULL;
2244 }
2245
2246 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2247 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2248
2249 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2250 state[MAX2(nr_samplers, nr_sampler_views)]);
2251
2252 for (i = 0; i < shader->info.base.num_inputs; i++) {
2253 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2254 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2255
2256 switch (shader->info.base.input_interpolate[i]) {
2257 case TGSI_INTERPOLATE_CONSTANT:
2258 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2259 break;
2260 case TGSI_INTERPOLATE_LINEAR:
2261 shader->inputs[i].interp = LP_INTERP_LINEAR;
2262 break;
2263 case TGSI_INTERPOLATE_PERSPECTIVE:
2264 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2265 break;
2266 case TGSI_INTERPOLATE_COLOR:
2267 shader->inputs[i].interp = LP_INTERP_COLOR;
2268 break;
2269 default:
2270 assert(0);
2271 break;
2272 }
2273
2274 switch (shader->info.base.input_semantic_name[i]) {
2275 case TGSI_SEMANTIC_FACE:
2276 shader->inputs[i].interp = LP_INTERP_FACING;
2277 break;
2278 case TGSI_SEMANTIC_POSITION:
2279 /* Position was already emitted above
2280 */
2281 shader->inputs[i].interp = LP_INTERP_POSITION;
2282 shader->inputs[i].src_index = 0;
2283 continue;
2284 }
2285
2286 shader->inputs[i].src_index = i+1;
2287 }
2288
2289 if (LP_DEBUG & DEBUG_TGSI) {
2290 unsigned attrib;
2291 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2292 shader->no, (void *) shader);
2293 tgsi_dump(templ->tokens, 0);
2294 debug_printf("usage masks:\n");
2295 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2296 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2297 debug_printf(" IN[%u].%s%s%s%s\n",
2298 attrib,
2299 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2300 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2301 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2302 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2303 }
2304 debug_printf("\n");
2305 }
2306
2307 return shader;
2308 }
2309
2310
2311 static void
2312 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2313 {
2314 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2315
2316 if (llvmpipe->fs == fs)
2317 return;
2318
2319 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2320
2321 draw_bind_fragment_shader(llvmpipe->draw,
2322 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2323
2324 llvmpipe->dirty |= LP_NEW_FS;
2325 }
2326
2327
2328 /**
2329 * Remove shader variant from two lists: the shader's variant list
2330 * and the context's variant list.
2331 */
2332 void
2333 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2334 struct lp_fragment_shader_variant *variant)
2335 {
2336 unsigned i;
2337
2338 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2339 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2340 " #%u v total cached #%u\n",
2341 variant->shader->no,
2342 variant->no,
2343 variant->shader->variants_created,
2344 variant->shader->variants_cached,
2345 lp->nr_fs_variants);
2346 }
2347
2348 /* free all the variant's JIT'd functions */
2349 for (i = 0; i < Elements(variant->function); i++) {
2350 if (variant->function[i]) {
2351 gallivm_free_function(variant->gallivm,
2352 variant->function[i],
2353 variant->jit_function[i]);
2354 }
2355 }
2356
2357 gallivm_destroy(variant->gallivm);
2358
2359 /* remove from shader's list */
2360 remove_from_list(&variant->list_item_local);
2361 variant->shader->variants_cached--;
2362
2363 /* remove from context's list */
2364 remove_from_list(&variant->list_item_global);
2365 lp->nr_fs_variants--;
2366 lp->nr_fs_instrs -= variant->nr_instrs;
2367
2368 FREE(variant);
2369 }
2370
2371
2372 static void
2373 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2374 {
2375 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2376 struct lp_fragment_shader *shader = fs;
2377 struct lp_fs_variant_list_item *li;
2378
2379 assert(fs != llvmpipe->fs);
2380
2381 /*
2382 * XXX: we need to flush the context until we have some sort of reference
2383 * counting in fragment shaders as they may still be binned
2384 * Flushing alone might not sufficient we need to wait on it too.
2385 */
2386 llvmpipe_finish(pipe, __FUNCTION__);
2387
2388 /* Delete all the variants */
2389 li = first_elem(&shader->variants);
2390 while(!at_end(&shader->variants, li)) {
2391 struct lp_fs_variant_list_item *next = next_elem(li);
2392 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2393 li = next;
2394 }
2395
2396 /* Delete draw module's data */
2397 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2398
2399 assert(shader->variants_cached == 0);
2400 FREE((void *) shader->base.tokens);
2401 FREE(shader);
2402 }
2403
2404
2405
2406 static void
2407 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2408 uint shader, uint index,
2409 struct pipe_constant_buffer *cb)
2410 {
2411 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2412 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2413
2414 assert(shader < PIPE_SHADER_TYPES);
2415 assert(index < Elements(llvmpipe->constants[shader]));
2416
2417 /* note: reference counting */
2418 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2419
2420 if (shader == PIPE_SHADER_VERTEX ||
2421 shader == PIPE_SHADER_GEOMETRY) {
2422 /* Pass the constants to the 'draw' module */
2423 const unsigned size = cb ? cb->buffer_size : 0;
2424 const ubyte *data;
2425
2426 if (constants) {
2427 data = (ubyte *) llvmpipe_resource_data(constants);
2428 }
2429 else if (cb && cb->user_buffer) {
2430 data = (ubyte *) cb->user_buffer;
2431 }
2432 else {
2433 data = NULL;
2434 }
2435
2436 if (data)
2437 data += cb->buffer_offset;
2438
2439 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2440 index, data, size);
2441 }
2442
2443 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2444
2445 if (cb && cb->user_buffer) {
2446 pipe_resource_reference(&constants, NULL);
2447 }
2448 }
2449
2450
2451 /**
2452 * Return the blend factor equivalent to a destination alpha of one.
2453 */
2454 static INLINE unsigned
2455 force_dst_alpha_one(unsigned factor)
2456 {
2457 switch(factor) {
2458 case PIPE_BLENDFACTOR_DST_ALPHA:
2459 return PIPE_BLENDFACTOR_ONE;
2460 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2461 return PIPE_BLENDFACTOR_ZERO;
2462 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2463 return PIPE_BLENDFACTOR_ZERO;
2464 }
2465
2466 return factor;
2467 }
2468
2469
2470 /**
2471 * We need to generate several variants of the fragment pipeline to match
2472 * all the combinations of the contributing state atoms.
2473 *
2474 * TODO: there is actually no reason to tie this to context state -- the
2475 * generated code could be cached globally in the screen.
2476 */
2477 static void
2478 make_variant_key(struct llvmpipe_context *lp,
2479 struct lp_fragment_shader *shader,
2480 struct lp_fragment_shader_variant_key *key)
2481 {
2482 unsigned i;
2483
2484 memset(key, 0, shader->variant_key_size);
2485
2486 if (lp->framebuffer.zsbuf) {
2487 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
2488 const struct util_format_description *zsbuf_desc =
2489 util_format_description(zsbuf_format);
2490
2491 if (lp->depth_stencil->depth.enabled &&
2492 util_format_has_depth(zsbuf_desc)) {
2493 key->zsbuf_format = zsbuf_format;
2494 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2495 }
2496 if (lp->depth_stencil->stencil[0].enabled &&
2497 util_format_has_stencil(zsbuf_desc)) {
2498 key->zsbuf_format = zsbuf_format;
2499 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2500 }
2501 }
2502
2503 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2504 if (!lp->framebuffer.nr_cbufs ||
2505 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2506 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2507 }
2508 if(key->alpha.enabled)
2509 key->alpha.func = lp->depth_stencil->alpha.func;
2510 /* alpha.ref_value is passed in jit_context */
2511
2512 key->flatshade = lp->rasterizer->flatshade;
2513 if (lp->active_occlusion_query) {
2514 key->occlusion_count = TRUE;
2515 }
2516
2517 if (lp->framebuffer.nr_cbufs) {
2518 memcpy(&key->blend, lp->blend, sizeof key->blend);
2519 }
2520
2521 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2522
2523 if (!key->blend.independent_blend_enable) {
2524 /* we always need independent blend otherwise the fixups below won't work */
2525 for (i = 1; i < key->nr_cbufs; i++) {
2526 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2527 }
2528 key->blend.independent_blend_enable = 1;
2529 }
2530
2531 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2532 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2533 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2534 const struct util_format_description *format_desc;
2535
2536 key->cbuf_format[i] = format;
2537
2538 format_desc = util_format_description(format);
2539 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2540 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2541
2542 /*
2543 * Mask out color channels not present in the color buffer.
2544 */
2545 blend_rt->colormask &= util_format_colormask(format_desc);
2546
2547 /*
2548 * Disable blend for integer formats.
2549 */
2550 if (util_format_is_pure_integer(format)) {
2551 blend_rt->blend_enable = 0;
2552 }
2553
2554 /*
2555 * Our swizzled render tiles always have an alpha channel, but the linear
2556 * render target format often does not, so force here the dst alpha to be
2557 * one.
2558 *
2559 * This is not a mere optimization. Wrong results will be produced if the
2560 * dst alpha is used, the dst format does not have alpha, and the previous
2561 * rendering was not flushed from the swizzled to linear buffer. For
2562 * example, NonPowTwo DCT.
2563 *
2564 * TODO: This should be generalized to all channels for better
2565 * performance, but only alpha causes correctness issues.
2566 *
2567 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2568 */
2569 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2570 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2571 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2572 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2573 blend_rt->alpha_func = blend_rt->rgb_func;
2574 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2575 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2576 }
2577 }
2578
2579 /* This value will be the same for all the variants of a given shader:
2580 */
2581 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2582
2583 for(i = 0; i < key->nr_samplers; ++i) {
2584 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2585 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2586 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2587 }
2588 }
2589
2590 /*
2591 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2592 * are dx10-style? Can't really have mixed opcodes, at least not
2593 * if we want to skip the holes here (without rescanning tgsi).
2594 */
2595 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2596 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2597 for(i = 0; i < key->nr_sampler_views; ++i) {
2598 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2599 lp_sampler_static_texture_state(&key->state[i].texture_state,
2600 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2601 }
2602 }
2603 }
2604 else {
2605 key->nr_sampler_views = key->nr_samplers;
2606 for(i = 0; i < key->nr_sampler_views; ++i) {
2607 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2608 lp_sampler_static_texture_state(&key->state[i].texture_state,
2609 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2610 }
2611 }
2612 }
2613 }
2614
2615
2616
2617 /**
2618 * Update fragment shader state. This is called just prior to drawing
2619 * something when some fragment-related state has changed.
2620 */
2621 void
2622 llvmpipe_update_fs(struct llvmpipe_context *lp)
2623 {
2624 struct lp_fragment_shader *shader = lp->fs;
2625 struct lp_fragment_shader_variant_key key;
2626 struct lp_fragment_shader_variant *variant = NULL;
2627 struct lp_fs_variant_list_item *li;
2628
2629 make_variant_key(lp, shader, &key);
2630
2631 /* Search the variants for one which matches the key */
2632 li = first_elem(&shader->variants);
2633 while(!at_end(&shader->variants, li)) {
2634 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2635 variant = li->base;
2636 break;
2637 }
2638 li = next_elem(li);
2639 }
2640
2641 if (variant) {
2642 /* Move this variant to the head of the list to implement LRU
2643 * deletion of shader's when we have too many.
2644 */
2645 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2646 }
2647 else {
2648 /* variant not found, create it now */
2649 int64_t t0, t1, dt;
2650 unsigned i;
2651 unsigned variants_to_cull;
2652
2653 if (0) {
2654 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2655 lp->nr_fs_variants,
2656 lp->nr_fs_instrs,
2657 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2658 }
2659
2660 /* First, check if we've exceeded the max number of shader variants.
2661 * If so, free 25% of them (the least recently used ones).
2662 */
2663 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2664
2665 if (variants_to_cull ||
2666 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2667 struct pipe_context *pipe = &lp->pipe;
2668
2669 /*
2670 * XXX: we need to flush the context until we have some sort of
2671 * reference counting in fragment shaders as they may still be binned
2672 * Flushing alone might not be sufficient we need to wait on it too.
2673 */
2674 llvmpipe_finish(pipe, __FUNCTION__);
2675
2676 /*
2677 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2678 * number of shader variants (potentially all of them) could be
2679 * pending for destruction on flush.
2680 */
2681
2682 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2683 struct lp_fs_variant_list_item *item;
2684 if (is_empty_list(&lp->fs_variants_list)) {
2685 break;
2686 }
2687 item = last_elem(&lp->fs_variants_list);
2688 assert(item);
2689 assert(item->base);
2690 llvmpipe_remove_shader_variant(lp, item->base);
2691 }
2692 }
2693
2694 /*
2695 * Generate the new variant.
2696 */
2697 t0 = os_time_get();
2698 variant = generate_variant(lp, shader, &key);
2699 t1 = os_time_get();
2700 dt = t1 - t0;
2701 LP_COUNT_ADD(llvm_compile_time, dt);
2702 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2703
2704 llvmpipe_variant_count++;
2705
2706 /* Put the new variant into the list */
2707 if (variant) {
2708 insert_at_head(&shader->variants, &variant->list_item_local);
2709 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2710 lp->nr_fs_variants++;
2711 lp->nr_fs_instrs += variant->nr_instrs;
2712 shader->variants_cached++;
2713 }
2714 }
2715
2716 /* Bind this variant */
2717 lp_setup_set_fs_variant(lp->setup, variant);
2718 }
2719
2720
2721
2722
2723
2724 void
2725 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2726 {
2727 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2728 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2729 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2730
2731 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2732 }