projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: support sRGB framebuffers
[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 #include "lp_rast.h"
104
105
106 /** Fragment shader number (for debugging) */
107 static unsigned fs_no = 0;
108
109
110 /**
111 * Expand the relevant bits of mask_input to a n*4-dword mask for the
112 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
113 * quad mask vector to 0 or ~0.
114 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
115 * quad arguments with fs length 8.
116 *
117 * \param first_quad which quad(s) of the quad group to test, in [0,3]
118 * \param mask_input bitwise mask for the whole 4x4 stamp
119 */
120 static LLVMValueRef
121 generate_quad_mask(struct gallivm_state *gallivm,
122 struct lp_type fs_type,
123 unsigned first_quad,
124 LLVMValueRef mask_input) /* int32 */
125 {
126 LLVMBuilderRef builder = gallivm->builder;
127 struct lp_type mask_type;
128 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
129 LLVMValueRef bits[16];
130 LLVMValueRef mask;
131 int shift, i;
132
133 /*
134 * XXX: We'll need a different path for 16 x u8
135 */
136 assert(fs_type.width == 32);
137 assert(fs_type.length <= Elements(bits));
138 mask_type = lp_int_type(fs_type);
139
140 /*
141 * mask_input >>= (quad * 4)
142 */
143 switch (first_quad) {
144 case 0:
145 shift = 0;
146 break;
147 case 1:
148 assert(fs_type.length == 4);
149 shift = 2;
150 break;
151 case 2:
152 shift = 8;
153 break;
154 case 3:
155 assert(fs_type.length == 4);
156 shift = 10;
157 break;
158 default:
159 assert(0);
160 shift = 0;
161 }
162
163 mask_input = LLVMBuildLShr(builder,
164 mask_input,
165 LLVMConstInt(i32t, shift, 0),
166 "");
167
168 /*
169 * mask = { mask_input & (1 << i), for i in [0,3] }
170 */
171 mask = lp_build_broadcast(gallivm,
172 lp_build_vec_type(gallivm, mask_type),
173 mask_input);
174
175 for (i = 0; i < fs_type.length / 4; i++) {
176 unsigned j = 2 * (i % 2) + (i / 2) * 8;
177 bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
178 bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
179 bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
180 bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
181 }
182 mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
183
184 /*
185 * mask = mask != 0 ? ~0 : 0
186 */
187 mask = lp_build_compare(gallivm,
188 mask_type, PIPE_FUNC_NOTEQUAL,
189 mask,
190 lp_build_const_int_vec(gallivm, mask_type, 0));
191
192 return mask;
193 }
194
195
196 #define EARLY_DEPTH_TEST 0x1
197 #define LATE_DEPTH_TEST 0x2
198 #define EARLY_DEPTH_WRITE 0x4
199 #define LATE_DEPTH_WRITE 0x8
200
201 static int
202 find_output_by_semantic( const struct tgsi_shader_info *info,
203 unsigned semantic,
204 unsigned index )
205 {
206 int i;
207
208 for (i = 0; i < info->num_outputs; i++)
209 if (info->output_semantic_name[i] == semantic &&
210 info->output_semantic_index[i] == index)
211 return i;
212
213 return -1;
214 }
215
216
217 /**
218 * Generate the fragment shader, depth/stencil test, and alpha tests.
219 */
220 static void
221 generate_fs_loop(struct gallivm_state *gallivm,
222 struct lp_fragment_shader *shader,
223 const struct lp_fragment_shader_variant_key *key,
224 LLVMBuilderRef builder,
225 struct lp_type type,
226 LLVMValueRef context_ptr,
227 LLVMValueRef num_loop,
228 struct lp_build_interp_soa_context *interp,
229 struct lp_build_sampler_soa *sampler,
230 LLVMValueRef mask_store,
231 LLVMValueRef (*out_color)[4],
232 LLVMValueRef depth_ptr,
233 LLVMValueRef depth_stride,
234 LLVMValueRef facing,
235 LLVMValueRef thread_data_ptr)
236 {
237 const struct util_format_description *zs_format_desc = NULL;
238 const struct tgsi_token *tokens = shader->base.tokens;
239 LLVMTypeRef vec_type;
240 LLVMValueRef mask_ptr, mask_val;
241 LLVMValueRef consts_ptr;
242 LLVMValueRef z;
243 LLVMValueRef z_value, s_value;
244 LLVMValueRef z_fb, s_fb;
245 LLVMValueRef stencil_refs[2];
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 /*
250 * TODO: figure out if simple_shader optimization is really worthwile to
251 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
252 * code since tests tend to take another codepath than real shaders.
253 */
254 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
255 shader->info.base.num_inputs < 3 &&
256 shader->info.base.num_instructions < 8) && 0;
257 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
258 util_blend_state_is_dual(&key->blend, 0);
259 unsigned attrib;
260 unsigned chan;
261 unsigned cbuf;
262 unsigned depth_mode;
263
264 struct lp_bld_tgsi_system_values system_values;
265
266 memset(&system_values, 0, sizeof(system_values));
267
268 if (key->depth.enabled ||
269 key->stencil[0].enabled) {
270
271 zs_format_desc = util_format_description(key->zsbuf_format);
272 assert(zs_format_desc);
273
274 if (!shader->info.base.writes_z) {
275 if (key->alpha.enabled || shader->info.base.uses_kill) {
276 /* With alpha test and kill, can do the depth test early
277 * and hopefully eliminate some quads. But need to do a
278 * special deferred depth write once the final mask value
279 * is known. This only works though if there's either no
280 * stencil test or the stencil value isn't written.
281 */
282 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
283 (key->stencil[1].enabled &&
284 key->stencil[1].writemask)))
285 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
286 else
287 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
288 }
289 else
290 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
291 }
292 else {
293 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
294 }
295
296 if (!(key->depth.enabled && key->depth.writemask) &&
297 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
298 (key->stencil[1].enabled &&
299 key->stencil[1].writemask))))
300 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
301 }
302 else {
303 depth_mode = 0;
304 }
305
306
307 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
308 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
309
310 vec_type = lp_build_vec_type(gallivm, type);
311
312 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
313
314 lp_build_for_loop_begin(&loop_state, gallivm,
315 lp_build_const_int32(gallivm, 0),
316 LLVMIntULT,
317 num_loop,
318 lp_build_const_int32(gallivm, 1));
319
320 mask_ptr = LLVMBuildGEP(builder, mask_store,
321 &loop_state.counter, 1, "mask_ptr");
322 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
323
324 memset(outputs, 0, sizeof outputs);
325
326 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
327 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
328 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
329 lp_build_vec_type(gallivm,
330 type),
331 num_loop, "color");
332 }
333 }
334 if (dual_source_blend) {
335 assert(key->nr_cbufs <= 1);
336 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
337 out_color[1][chan] = lp_build_array_alloca(gallivm,
338 lp_build_vec_type(gallivm,
339 type),
340 num_loop, "color1");
341 }
342 }
343
344
345 /* 'mask' will control execution based on quad's pixel alive/killed state */
346 lp_build_mask_begin(&mask, gallivm, type, mask_val);
347
348 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
349 lp_build_mask_check(&mask);
350
351 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
352 z = interp->pos[2];
353
354 if (depth_mode & EARLY_DEPTH_TEST) {
355 lp_build_depth_stencil_load_swizzled(gallivm, type,
356 zs_format_desc, key->resource_1d,
357 depth_ptr, depth_stride,
358 &z_fb, &s_fb, loop_state.counter);
359 lp_build_depth_stencil_test(gallivm,
360 &key->depth,
361 key->stencil,
362 type,
363 zs_format_desc,
364 &mask,
365 stencil_refs,
366 z, z_fb, s_fb,
367 facing,
368 &z_value, &s_value,
369 !simple_shader);
370
371 if (depth_mode & EARLY_DEPTH_WRITE) {
372 lp_build_depth_stencil_write_swizzled(gallivm, type,
373 zs_format_desc, key->resource_1d,
374 NULL, NULL, NULL, loop_state.counter,
375 depth_ptr, depth_stride,
376 z_value, s_value);
377 }
378 /*
379 * Note mask check if stencil is enabled must be after ds write not after
380 * stencil test otherwise new stencil values may not get written if all
381 * fragments got killed by depth/stencil test.
382 */
383 if (!simple_shader && key->stencil[0].enabled)
384 lp_build_mask_check(&mask);
385 }
386
387 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
388
389 /* Build the actual shader */
390 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
391 consts_ptr, &system_values,
392 interp->inputs,
393 outputs, sampler, &shader->info.base, NULL);
394
395 /* Alpha test */
396 if (key->alpha.enabled) {
397 int color0 = find_output_by_semantic(&shader->info.base,
398 TGSI_SEMANTIC_COLOR,
399 0);
400
401 if (color0 != -1 && outputs[color0][3]) {
402 const struct util_format_description *cbuf_format_desc;
403 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
404 LLVMValueRef alpha_ref_value;
405
406 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
407 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
408
409 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
410
411 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
412 &mask, alpha, alpha_ref_value,
413 (depth_mode & LATE_DEPTH_TEST) != 0);
414 }
415 }
416
417 /* Late Z test */
418 if (depth_mode & LATE_DEPTH_TEST) {
419 int pos0 = find_output_by_semantic(&shader->info.base,
420 TGSI_SEMANTIC_POSITION,
421 0);
422
423 if (pos0 != -1 && outputs[pos0][2]) {
424 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
425 }
426
427 lp_build_depth_stencil_load_swizzled(gallivm, type,
428 zs_format_desc, key->resource_1d,
429 depth_ptr, depth_stride,
430 &z_fb, &s_fb, loop_state.counter);
431
432 lp_build_depth_stencil_test(gallivm,
433 &key->depth,
434 key->stencil,
435 type,
436 zs_format_desc,
437 &mask,
438 stencil_refs,
439 z, z_fb, s_fb,
440 facing,
441 &z_value, &s_value,
442 !simple_shader);
443 /* Late Z write */
444 if (depth_mode & LATE_DEPTH_WRITE) {
445 lp_build_depth_stencil_write_swizzled(gallivm, type,
446 zs_format_desc, key->resource_1d,
447 NULL, NULL, NULL, loop_state.counter,
448 depth_ptr, depth_stride,
449 z_value, s_value);
450 }
451 }
452 else if ((depth_mode & EARLY_DEPTH_TEST) &&
453 (depth_mode & LATE_DEPTH_WRITE))
454 {
455 /* Need to apply a reduced mask to the depth write. Reload the
456 * depth value, update from zs_value with the new mask value and
457 * write that out.
458 */
459 lp_build_depth_stencil_write_swizzled(gallivm, type,
460 zs_format_desc, key->resource_1d,
461 &mask, z_fb, s_fb, loop_state.counter,
462 depth_ptr, depth_stride,
463 z_value, s_value);
464 }
465
466
467 /* Color write */
468 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
469 {
470 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
471 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
472 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
473 {
474 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
475 if(outputs[attrib][chan]) {
476 /* XXX: just initialize outputs to point at colors[] and
477 * skip this.
478 */
479 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
480 LLVMValueRef color_ptr;
481 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
482 &loop_state.counter, 1, "");
483 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
484 LLVMBuildStore(builder, out, color_ptr);
485 }
486 }
487 }
488 }
489
490 if (key->occlusion_count) {
491 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
492 lp_build_name(counter, "counter");
493 lp_build_occlusion_count(gallivm, type,
494 lp_build_mask_value(&mask), counter);
495 }
496
497 mask_val = lp_build_mask_end(&mask);
498 LLVMBuildStore(builder, mask_val, mask_ptr);
499 lp_build_for_loop_end(&loop_state);
500 }
501
502
503 /**
504 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
505 *
506 * Fragment Shader outputs pixels in small 2x2 blocks
507 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
508 *
509 * However in memory pixels are stored in rows
510 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
511 *
512 * @param type fragment shader type (4x or 8x float)
513 * @param num_fs number of fs_src
514 * @param is_1d whether we're outputting to a 1d resource
515 * @param dst_channels number of output channels
516 * @param fs_src output from fragment shader
517 * @param dst pointer to store result
518 * @param pad_inline is channel padding inline or at end of row
519 * @return the number of dsts
520 */
521 static int
522 generate_fs_twiddle(struct gallivm_state *gallivm,
523 struct lp_type type,
524 unsigned num_fs,
525 unsigned dst_channels,
526 LLVMValueRef fs_src[][4],
527 LLVMValueRef* dst,
528 bool pad_inline)
529 {
530 LLVMValueRef src[16];
531
532 bool swizzle_pad;
533 bool twiddle;
534 bool split;
535
536 unsigned pixels = type.length / 4;
537 unsigned reorder_group;
538 unsigned src_channels;
539 unsigned src_count;
540 unsigned i;
541
542 src_channels = dst_channels < 3 ? dst_channels : 4;
543 src_count = num_fs * src_channels;
544
545 assert(pixels == 2 || pixels == 1);
546 assert(num_fs * src_channels <= Elements(src));
547
548 /*
549 * Transpose from SoA -> AoS
550 */
551 for (i = 0; i < num_fs; ++i) {
552 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
553 }
554
555 /*
556 * Pick transformation options
557 */
558 swizzle_pad = false;
559 twiddle = false;
560 split = false;
561 reorder_group = 0;
562
563 if (dst_channels == 1) {
564 twiddle = true;
565
566 if (pixels == 2) {
567 split = true;
568 }
569 } else if (dst_channels == 2) {
570 if (pixels == 1) {
571 reorder_group = 1;
572 }
573 } else if (dst_channels > 2) {
574 if (pixels == 1) {
575 reorder_group = 2;
576 } else {
577 twiddle = true;
578 }
579
580 if (!pad_inline && dst_channels == 3 && pixels > 1) {
581 swizzle_pad = true;
582 }
583 }
584
585 /*
586 * Split the src in half
587 */
588 if (split) {
589 for (i = num_fs; i > 0; --i) {
590 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
591 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
592 }
593
594 src_count *= 2;
595 type.length = 4;
596 }
597
598 /*
599 * Ensure pixels are in memory order
600 */
601 if (reorder_group) {
602 /* Twiddle pixels by reordering the array, e.g.:
603 *
604 * src_count = 8 -> 0 2 1 3 4 6 5 7
605 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
606 */
607 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
608
609 for (i = 0; i < src_count; ++i) {
610 unsigned group = i / reorder_group;
611 unsigned block = (group / 4) * 4 * reorder_group;
612 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
613 dst[i] = src[j];
614 }
615 } else if (twiddle) {
616 /* Twiddle pixels across elements of array */
617 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
618 } else {
619 /* Do nothing */
620 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
621 }
622
623 /*
624 * Moves any padding between pixels to the end
625 * e.g. RGBXRGBX -> RGBRGBXX
626 */
627 if (swizzle_pad) {
628 unsigned char swizzles[16];
629 unsigned elems = pixels * dst_channels;
630
631 for (i = 0; i < type.length; ++i) {
632 if (i < elems)
633 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
634 else
635 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
636 }
637
638 for (i = 0; i < src_count; ++i) {
639 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
640 }
641 }
642
643 return src_count;
644 }
645
646
647 /**
648 * Load an unswizzled block of pixels from memory
649 */
650 static void
651 load_unswizzled_block(struct gallivm_state *gallivm,
652 LLVMValueRef base_ptr,
653 LLVMValueRef stride,
654 unsigned block_width,
655 unsigned block_height,
656 LLVMValueRef* dst,
657 struct lp_type dst_type,
658 unsigned dst_count,
659 unsigned dst_alignment)
660 {
661 LLVMBuilderRef builder = gallivm->builder;
662 unsigned row_size = dst_count / block_height;
663 unsigned i;
664
665 /* Ensure block exactly fits into dst */
666 assert((block_width * block_height) % dst_count == 0);
667
668 for (i = 0; i < dst_count; ++i) {
669 unsigned x = i % row_size;
670 unsigned y = i / row_size;
671
672 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
673 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
674
675 LLVMValueRef gep[2];
676 LLVMValueRef dst_ptr;
677
678 gep[0] = lp_build_const_int32(gallivm, 0);
679 gep[1] = LLVMBuildAdd(builder, bx, by, "");
680
681 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
682 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
683
684 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
685
686 lp_set_load_alignment(dst[i], dst_alignment);
687 }
688 }
689
690
691 /**
692 * Store an unswizzled block of pixels to memory
693 */
694 static void
695 store_unswizzled_block(struct gallivm_state *gallivm,
696 LLVMValueRef base_ptr,
697 LLVMValueRef stride,
698 unsigned block_width,
699 unsigned block_height,
700 LLVMValueRef* src,
701 struct lp_type src_type,
702 unsigned src_count,
703 unsigned src_alignment)
704 {
705 LLVMBuilderRef builder = gallivm->builder;
706 unsigned row_size = src_count / block_height;
707 unsigned i;
708
709 /* Ensure src exactly fits into block */
710 assert((block_width * block_height) % src_count == 0);
711
712 for (i = 0; i < src_count; ++i) {
713 unsigned x = i % row_size;
714 unsigned y = i / row_size;
715
716 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
717 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
718
719 LLVMValueRef gep[2];
720 LLVMValueRef src_ptr;
721
722 gep[0] = lp_build_const_int32(gallivm, 0);
723 gep[1] = LLVMBuildAdd(builder, bx, by, "");
724
725 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
726 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
727
728 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
729
730 lp_set_store_alignment(src_ptr, src_alignment);
731 }
732 }
733
734
735 /**
736 * Checks if a format description is an arithmetic format
737 *
738 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
739 */
740 static INLINE boolean
741 is_arithmetic_format(const struct util_format_description *format_desc)
742 {
743 boolean arith = false;
744 unsigned i;
745
746 for (i = 0; i < format_desc->nr_channels; ++i) {
747 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
748 arith |= (format_desc->channel[i].size % 8) != 0;
749 }
750
751 return arith;
752 }
753
754
755 /**
756 * Checks if this format requires special handling due to required expansion
757 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
758 * SoA conversion.
759 */
760 static INLINE boolean
761 format_expands_to_float_soa(const struct util_format_description *format_desc)
762 {
763 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
764 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
765 return true;
766 }
767 return false;
768 }
769
770
771 /**
772 * Retrieves the type representing the memory layout for a format
773 *
774 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
775 */
776 static INLINE void
777 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
778 struct lp_type* type)
779 {
780 unsigned i;
781 unsigned chan;
782
783 if (format_expands_to_float_soa(format_desc)) {
784 /* just make this a 32bit uint */
785 type->floating = false;
786 type->fixed = false;
787 type->sign = false;
788 type->norm = false;
789 type->width = 32;
790 type->length = 1;
791 return;
792 }
793
794 for (i = 0; i < 4; i++)
795 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
796 break;
797 chan = i;
798
799 memset(type, 0, sizeof(struct lp_type));
800 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
801 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
802 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
803 type->norm = format_desc->channel[chan].normalized;
804
805 if (is_arithmetic_format(format_desc)) {
806 type->width = 0;
807 type->length = 1;
808
809 for (i = 0; i < format_desc->nr_channels; ++i) {
810 type->width += format_desc->channel[i].size;
811 }
812 } else {
813 type->width = format_desc->channel[chan].size;
814 type->length = format_desc->nr_channels;
815 }
816 }
817
818
819 /**
820 * Retrieves the type for a format which is usable in the blending code.
821 *
822 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
823 */
824 static INLINE void
825 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
826 struct lp_type* type)
827 {
828 unsigned i;
829 unsigned chan;
830
831 if (format_expands_to_float_soa(format_desc)) {
832 /* always use ordinary floats for blending */
833 type->floating = true;
834 type->fixed = false;
835 type->sign = true;
836 type->norm = false;
837 type->width = 32;
838 type->length = 4;
839 return;
840 }
841
842 for (i = 0; i < 4; i++)
843 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
844 break;
845 chan = i;
846
847 memset(type, 0, sizeof(struct lp_type));
848 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
849 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
850 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
851 type->norm = format_desc->channel[chan].normalized;
852 type->width = format_desc->channel[chan].size;
853 type->length = format_desc->nr_channels;
854
855 for (i = 1; i < format_desc->nr_channels; ++i) {
856 if (format_desc->channel[i].size > type->width)
857 type->width = format_desc->channel[i].size;
858 }
859
860 if (type->floating) {
861 type->width = 32;
862 } else {
863 if (type->width <= 8) {
864 type->width = 8;
865 } else if (type->width <= 16) {
866 type->width = 16;
867 } else {
868 type->width = 32;
869 }
870 }
871
872 if (is_arithmetic_format(format_desc) && type->length == 3) {
873 type->length = 4;
874 }
875 }
876
877
878 /**
879 * Scale a normalized value from src_bits to dst_bits
880 */
881 static INLINE LLVMValueRef
882 scale_bits(struct gallivm_state *gallivm,
883 int src_bits,
884 int dst_bits,
885 LLVMValueRef src,
886 struct lp_type src_type)
887 {
888 LLVMBuilderRef builder = gallivm->builder;
889 LLVMValueRef result = src;
890
891 if (dst_bits < src_bits) {
892 /* Scale down by LShr */
893 result = LLVMBuildLShr(builder,
894 src,
895 lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
896 "");
897 } else if (dst_bits > src_bits) {
898 /* Scale up bits */
899 int db = dst_bits - src_bits;
900
901 /* Shift left by difference in bits */
902 result = LLVMBuildShl(builder,
903 src,
904 lp_build_const_int_vec(gallivm, src_type, db),
905 "");
906
907 if (db < src_bits) {
908 /* Enough bits in src to fill the remainder */
909 LLVMValueRef lower = LLVMBuildLShr(builder,
910 src,
911 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
912 "");
913
914 result = LLVMBuildOr(builder, result, lower, "");
915 } else if (db > src_bits) {
916 /* Need to repeatedly copy src bits to fill remainder in dst */
917 unsigned n;
918
919 for (n = src_bits; n < dst_bits; n *= 2) {
920 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
921
922 result = LLVMBuildOr(builder,
923 result,
924 LLVMBuildLShr(builder, result, shuv, ""),
925 "");
926 }
927 }
928 }
929
930 return result;
931 }
932
933
934 /**
935 * Convert from memory format to blending format
936 *
937 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
938 */
939 static void
940 convert_to_blend_type(struct gallivm_state *gallivm,
941 unsigned block_size,
942 const struct util_format_description *src_fmt,
943 struct lp_type src_type,
944 struct lp_type dst_type,
945 LLVMValueRef* src, // and dst
946 unsigned num_srcs)
947 {
948 LLVMValueRef *dst = src;
949 LLVMBuilderRef builder = gallivm->builder;
950 struct lp_type blend_type;
951 struct lp_type mem_type;
952 unsigned i, j, k;
953 unsigned pixels = block_size / num_srcs;
954 bool is_arith;
955
956 /*
957 * full custom path for packed floats and srgb formats - none of the later
958 * functions would do anything useful, and given the lp_type representation they
959 * can't be fixed. Should really have some SoA blend path for these kind of
960 * formats rather than hacking them in here.
961 */
962 if (format_expands_to_float_soa(src_fmt)) {
963 LLVMValueRef tmpsrc[4];
964 /*
965 * This is pretty suboptimal for this case blending in SoA would be much
966 * better, since conversion gets us SoA values so need to convert back.
967 */
968 assert(src_type.width == 32);
969 assert(dst_type.floating);
970 assert(dst_type.width == 32);
971 assert(dst_type.length % 4 == 0);
972 assert(num_srcs % 4 == 0);
973
974 for (i = 0; i < 4; i++) {
975 tmpsrc[i] = src[i];
976 }
977 for (i = 0; i < num_srcs / 4; i++) {
978 LLVMValueRef tmpsoa[4];
979 LLVMValueRef tmps = tmpsrc[i];
980 if (dst_type.length == 8) {
981 LLVMValueRef shuffles[8];
982 unsigned j;
983 /* fetch was 4 values but need 8-wide output values */
984 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
985 /*
986 * for 8-wide aos transpose would give us wrong order not matching
987 * incoming converted fs values and mask. ARGH.
988 */
989 for (j = 0; j < 4; j++) {
990 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
991 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
992 }
993 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
994 LLVMConstVector(shuffles, 8), "");
995 }
996 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
997 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
998 }
999 else {
1000 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1001 }
1002 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1003 }
1004 return;
1005 }
1006
1007 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1008 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1009
1010 /* Is the format arithmetic */
1011 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1012 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1013
1014 /* Pad if necessary */
1015 if (!is_arith && src_type.length < dst_type.length) {
1016 for (i = 0; i < num_srcs; ++i) {
1017 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1018 }
1019
1020 src_type.length = dst_type.length;
1021 }
1022
1023 /* Special case for half-floats */
1024 if (mem_type.width == 16 && mem_type.floating) {
1025 assert(blend_type.width == 32 && blend_type.floating);
1026 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1027 is_arith = false;
1028 }
1029
1030 if (!is_arith) {
1031 return;
1032 }
1033
1034 src_type.width = blend_type.width * blend_type.length;
1035 blend_type.length *= pixels;
1036 src_type.length *= pixels / (src_type.length / mem_type.length);
1037
1038 for (i = 0; i < num_srcs; ++i) {
1039 LLVMValueRef chans[4];
1040 LLVMValueRef res = NULL;
1041
1042 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1043
1044 for (j = 0; j < src_fmt->nr_channels; ++j) {
1045 unsigned mask = 0;
1046 unsigned sa = src_fmt->channel[j].shift;
1047 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1048 unsigned from_lsb = j;
1049 #else
1050 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1051 #endif
1052
1053 for (k = 0; k < src_fmt->channel[j].size; ++k) {
1054 mask |= 1 << k;
1055 }
1056
1057 /* Extract bits from source */
1058 chans[j] = LLVMBuildLShr(builder,
1059 dst[i],
1060 lp_build_const_int_vec(gallivm, src_type, sa),
1061 "");
1062
1063 chans[j] = LLVMBuildAnd(builder,
1064 chans[j],
1065 lp_build_const_int_vec(gallivm, src_type, mask),
1066 "");
1067
1068 /* Scale bits */
1069 if (src_type.norm) {
1070 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1071 blend_type.width, chans[j], src_type);
1072 }
1073
1074 /* Insert bits into correct position */
1075 chans[j] = LLVMBuildShl(builder,
1076 chans[j],
1077 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1078 "");
1079
1080 if (j == 0) {
1081 res = chans[j];
1082 } else {
1083 res = LLVMBuildOr(builder, res, chans[j], "");
1084 }
1085 }
1086
1087 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1088 }
1089 }
1090
1091
1092 /**
1093 * Convert from blending format to memory format
1094 *
1095 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1096 */
1097 static void
1098 convert_from_blend_type(struct gallivm_state *gallivm,
1099 unsigned block_size,
1100 const struct util_format_description *src_fmt,
1101 struct lp_type src_type,
1102 struct lp_type dst_type,
1103 LLVMValueRef* src, // and dst
1104 unsigned num_srcs)
1105 {
1106 LLVMValueRef* dst = src;
1107 unsigned i, j, k;
1108 struct lp_type mem_type;
1109 struct lp_type blend_type;
1110 LLVMBuilderRef builder = gallivm->builder;
1111 unsigned pixels = block_size / num_srcs;
1112 bool is_arith;
1113
1114 /*
1115 * full custom path for packed floats and srgb formats - none of the later
1116 * functions would do anything useful, and given the lp_type representation they
1117 * can't be fixed. Should really have some SoA blend path for these kind of
1118 * formats rather than hacking them in here.
1119 */
1120 if (format_expands_to_float_soa(src_fmt)) {
1121 /*
1122 * This is pretty suboptimal for this case blending in SoA would be much
1123 * better - we need to transpose the AoS values back to SoA values for
1124 * conversion/packing.
1125 */
1126 assert(src_type.floating);
1127 assert(src_type.width == 32);
1128 assert(src_type.length % 4 == 0);
1129 assert(dst_type.width == 32);
1130
1131 for (i = 0; i < num_srcs / 4; i++) {
1132 LLVMValueRef tmpsoa[4], tmpdst;
1133 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1134 /* really really need SoA here */
1135
1136 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1137 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1138 }
1139 else {
1140 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1141 src_type, tmpsoa);
1142 }
1143
1144 if (src_type.length == 8) {
1145 LLVMValueRef tmpaos, shuffles[8];
1146 unsigned j;
1147 /*
1148 * for 8-wide aos transpose has given us wrong order not matching
1149 * output order. HMPF. Also need to split the output values manually.
1150 */
1151 for (j = 0; j < 4; j++) {
1152 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1153 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1154 }
1155 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1156 LLVMConstVector(shuffles, 8), "");
1157 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1158 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1159 }
1160 else {
1161 src[i] = tmpdst;
1162 }
1163 }
1164 return;
1165 }
1166
1167 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1168 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1169
1170 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1171
1172 /* Special case for half-floats */
1173 if (mem_type.width == 16 && mem_type.floating) {
1174 int length = dst_type.length;
1175 assert(blend_type.width == 32 && blend_type.floating);
1176
1177 dst_type.length = src_type.length;
1178
1179 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1180
1181 dst_type.length = length;
1182 is_arith = false;
1183 }
1184
1185 /* Remove any padding */
1186 if (!is_arith && (src_type.length % mem_type.length)) {
1187 src_type.length -= (src_type.length % mem_type.length);
1188
1189 for (i = 0; i < num_srcs; ++i) {
1190 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1191 }
1192 }
1193
1194 /* No bit arithmetic to do */
1195 if (!is_arith) {
1196 return;
1197 }
1198
1199 src_type.length = pixels;
1200 src_type.width = blend_type.length * blend_type.width;
1201 dst_type.length = pixels;
1202
1203 for (i = 0; i < num_srcs; ++i) {
1204 LLVMValueRef chans[4];
1205 LLVMValueRef res = NULL;
1206
1207 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1208
1209 for (j = 0; j < src_fmt->nr_channels; ++j) {
1210 unsigned mask = 0;
1211 unsigned sa = src_fmt->channel[j].shift;
1212 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1213 unsigned from_lsb = j;
1214 #else
1215 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1216 #endif
1217
1218 assert(blend_type.width > src_fmt->channel[j].size);
1219
1220 for (k = 0; k < blend_type.width; ++k) {
1221 mask |= 1 << k;
1222 }
1223
1224 /* Extract bits */
1225 chans[j] = LLVMBuildLShr(builder,
1226 dst[i],
1227 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1228 "");
1229
1230 chans[j] = LLVMBuildAnd(builder,
1231 chans[j],
1232 lp_build_const_int_vec(gallivm, src_type, mask),
1233 "");
1234
1235 /* Scale down bits */
1236 if (src_type.norm) {
1237 chans[j] = scale_bits(gallivm, blend_type.width,
1238 src_fmt->channel[j].size, chans[j], src_type);
1239 }
1240
1241 /* Insert bits */
1242 chans[j] = LLVMBuildShl(builder,
1243 chans[j],
1244 lp_build_const_int_vec(gallivm, src_type, sa),
1245 "");
1246
1247 sa += src_fmt->channel[j].size;
1248
1249 if (j == 0) {
1250 res = chans[j];
1251 } else {
1252 res = LLVMBuildOr(builder, res, chans[j], "");
1253 }
1254 }
1255
1256 assert (dst_type.width != 24);
1257
1258 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1259 }
1260 }
1261
1262
1263 /**
1264 * Convert alpha to same blend type as src
1265 */
1266 static void
1267 convert_alpha(struct gallivm_state *gallivm,
1268 struct lp_type row_type,
1269 struct lp_type alpha_type,
1270 const unsigned block_size,
1271 const unsigned block_height,
1272 const unsigned src_count,
1273 const unsigned dst_channels,
1274 const bool pad_inline,
1275 LLVMValueRef* src_alpha)
1276 {
1277 LLVMBuilderRef builder = gallivm->builder;
1278 unsigned i, j;
1279 unsigned length = row_type.length;
1280 row_type.length = alpha_type.length;
1281
1282 /* Twiddle the alpha to match pixels */
1283 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1284
1285 /*
1286 * TODO this should use single lp_build_conv call for
1287 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1288 */
1289 for (i = 0; i < block_height; ++i) {
1290 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1291 }
1292
1293 alpha_type = row_type;
1294 row_type.length = length;
1295
1296 /* If only one channel we can only need the single alpha value per pixel */
1297 if (src_count == 1 && dst_channels == 1) {
1298
1299 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1300 } else {
1301 /* If there are more srcs than rows then we need to split alpha up */
1302 if (src_count > block_height) {
1303 for (i = src_count; i > 0; --i) {
1304 unsigned pixels = block_size / src_count;
1305 unsigned idx = i - 1;
1306
1307 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1308 (idx * pixels) % 4, pixels);
1309 }
1310 }
1311
1312 /* If there is a src for each pixel broadcast the alpha across whole row */
1313 if (src_count == block_size) {
1314 for (i = 0; i < src_count; ++i) {
1315 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1316 }
1317 } else {
1318 unsigned pixels = block_size / src_count;
1319 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1320 unsigned alpha_span = 1;
1321 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1322
1323 /* Check if we need 2 src_alphas for our shuffles */
1324 if (pixels > alpha_type.length) {
1325 alpha_span = 2;
1326 }
1327
1328 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1329 for (j = 0; j < row_type.length; ++j) {
1330 if (j < pixels * channels) {
1331 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1332 } else {
1333 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1334 }
1335 }
1336
1337 for (i = 0; i < src_count; ++i) {
1338 unsigned idx1 = i, idx2 = i;
1339
1340 if (alpha_span > 1){
1341 idx1 *= alpha_span;
1342 idx2 = idx1 + 1;
1343 }
1344
1345 src_alpha[i] = LLVMBuildShuffleVector(builder,
1346 src_alpha[idx1],
1347 src_alpha[idx2],
1348 LLVMConstVector(shuffles, row_type.length),
1349 "");
1350 }
1351 }
1352 }
1353 }
1354
1355
1356 /**
1357 * Generates the blend function for unswizzled colour buffers
1358 * Also generates the read & write from colour buffer
1359 */
1360 static void
1361 generate_unswizzled_blend(struct gallivm_state *gallivm,
1362 unsigned rt,
1363 struct lp_fragment_shader_variant *variant,
1364 enum pipe_format out_format,
1365 unsigned int num_fs,
1366 struct lp_type fs_type,
1367 LLVMValueRef* fs_mask,
1368 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1369 LLVMValueRef context_ptr,
1370 LLVMValueRef color_ptr,
1371 LLVMValueRef stride,
1372 unsigned partial_mask,
1373 boolean do_branch)
1374 {
1375 const unsigned alpha_channel = 3;
1376 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1377 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1378 const unsigned block_size = block_width * block_height;
1379 const unsigned lp_integer_vector_width = 128;
1380
1381 LLVMBuilderRef builder = gallivm->builder;
1382 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1383 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1384 LLVMValueRef src_alpha[4 * 4];
1385 LLVMValueRef src1_alpha[4 * 4];
1386 LLVMValueRef src_mask[4 * 4];
1387 LLVMValueRef src[4 * 4];
1388 LLVMValueRef src1[4 * 4];
1389 LLVMValueRef dst[4 * 4];
1390 LLVMValueRef blend_color;
1391 LLVMValueRef blend_alpha;
1392 LLVMValueRef i32_zero;
1393 LLVMValueRef check_mask;
1394 LLVMValueRef undef_src_val;
1395
1396 struct lp_build_mask_context mask_ctx;
1397 struct lp_type mask_type;
1398 struct lp_type blend_type;
1399 struct lp_type row_type;
1400 struct lp_type dst_type;
1401
1402 unsigned char swizzle[TGSI_NUM_CHANNELS];
1403 unsigned vector_width;
1404 unsigned src_channels = TGSI_NUM_CHANNELS;
1405 unsigned dst_channels;
1406 unsigned dst_count;
1407 unsigned src_count;
1408 unsigned i, j;
1409
1410 const struct util_format_description* out_format_desc = util_format_description(out_format);
1411
1412 unsigned dst_alignment;
1413
1414 bool pad_inline = is_arithmetic_format(out_format_desc);
1415 bool has_alpha = false;
1416 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1417 util_blend_state_is_dual(&variant->key.blend, 0);
1418
1419 const boolean is_1d = variant->key.resource_1d;
1420 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
1421
1422 mask_type = lp_int32_vec4_type();
1423 mask_type.length = fs_type.length;
1424
1425 for (i = num_fs; i < num_fullblock_fs; i++) {
1426 fs_mask[i] = lp_build_zero(gallivm, mask_type);
1427 }
1428
1429 /* Do not bother executing code when mask is empty.. */
1430 if (do_branch) {
1431 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1432
1433 for (i = 0; i < num_fullblock_fs; ++i) {
1434 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1435 }
1436
1437 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1438 lp_build_mask_check(&mask_ctx);
1439 }
1440
1441 partial_mask |= !variant->opaque;
1442 i32_zero = lp_build_const_int32(gallivm, 0);
1443
1444 #if HAVE_LLVM < 0x0302
1445 /*
1446 * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
1447 * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
1448 */
1449 undef_src_val = lp_build_zero(gallivm, fs_type);
1450 #else
1451 undef_src_val = lp_build_undef(gallivm, fs_type);
1452 #endif
1453
1454
1455 /* Get type from output format */
1456 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1457 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1458
1459 row_type.length = fs_type.length;
1460 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1461
1462 /* Compute correct swizzle and count channels */
1463 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1464 dst_channels = 0;
1465
1466 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1467 /* Ensure channel is used */
1468 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1469 continue;
1470 }
1471
1472 /* Ensure not already written to (happens in case with GL_ALPHA) */
1473 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1474 continue;
1475 }
1476
1477 /* Ensure we havn't already found all channels */
1478 if (dst_channels >= out_format_desc->nr_channels) {
1479 continue;
1480 }
1481
1482 swizzle[out_format_desc->swizzle[i]] = i;
1483 ++dst_channels;
1484
1485 if (i == alpha_channel) {
1486 has_alpha = true;
1487 }
1488 }
1489
1490 if (format_expands_to_float_soa(out_format_desc)) {
1491 /*
1492 * the code above can't work for layout_other
1493 * for srgb it would sort of work but we short-circuit swizzles, etc.
1494 * as that is done as part of unpack / pack.
1495 */
1496 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1497 has_alpha = true;
1498 swizzle[0] = 0;
1499 swizzle[1] = 1;
1500 swizzle[2] = 2;
1501 swizzle[3] = 3;
1502 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1503 }
1504
1505 /* If 3 channels then pad to include alpha for 4 element transpose */
1506 if (dst_channels == 3 && !has_alpha) {
1507 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1508 if (swizzle[i] > TGSI_NUM_CHANNELS)
1509 swizzle[i] = 3;
1510 }
1511 if (out_format_desc->nr_channels == 4) {
1512 dst_channels = 4;
1513 }
1514 }
1515
1516 /*
1517 * Load shader output
1518 */
1519 for (i = 0; i < num_fullblock_fs; ++i) {
1520 /* Always load alpha for use in blending */
1521 LLVMValueRef alpha;
1522 if (i < num_fs) {
1523 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1524 }
1525 else {
1526 alpha = undef_src_val;
1527 }
1528
1529 /* Load each channel */
1530 for (j = 0; j < dst_channels; ++j) {
1531 assert(swizzle[j] < 4);
1532 if (i < num_fs) {
1533 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1534 }
1535 else {
1536 fs_src[i][j] = undef_src_val;
1537 }
1538 }
1539
1540 /* If 3 channels then pad to include alpha for 4 element transpose */
1541 /*
1542 * XXX If we include that here maybe could actually use it instead of
1543 * separate alpha for blending?
1544 */
1545 if (dst_channels == 3 && !has_alpha) {
1546 fs_src[i][3] = alpha;
1547 }
1548
1549 /* We split the row_mask and row_alpha as we want 128bit interleave */
1550 if (fs_type.length == 8) {
1551 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1552 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1553
1554 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1555 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1556 } else {
1557 src_mask[i] = fs_mask[i];
1558 src_alpha[i] = alpha;
1559 }
1560 }
1561 if (dual_source_blend) {
1562 /* same as above except different src/dst, skip masks and comments... */
1563 for (i = 0; i < num_fullblock_fs; ++i) {
1564 LLVMValueRef alpha;
1565 if (i < num_fs) {
1566 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1567 }
1568 else {
1569 alpha = undef_src_val;
1570 }
1571
1572 for (j = 0; j < dst_channels; ++j) {
1573 assert(swizzle[j] < 4);
1574 if (i < num_fs) {
1575 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1576 }
1577 else {
1578 fs_src1[i][j] = undef_src_val;
1579 }
1580 }
1581 if (dst_channels == 3 && !has_alpha) {
1582 fs_src1[i][3] = alpha;
1583 }
1584 if (fs_type.length == 8) {
1585 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1586 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1587 } else {
1588 src1_alpha[i] = alpha;
1589 }
1590 }
1591 }
1592
1593 if (util_format_is_pure_integer(out_format)) {
1594 /*
1595 * In this case fs_type was really ints or uints disguised as floats,
1596 * fix that up now.
1597 */
1598 fs_type.floating = 0;
1599 fs_type.sign = dst_type.sign;
1600 for (i = 0; i < num_fullblock_fs; ++i) {
1601 for (j = 0; j < dst_channels; ++j) {
1602 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1603 lp_build_vec_type(gallivm, fs_type), "");
1604 }
1605 if (dst_channels == 3 && !has_alpha) {
1606 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1607 lp_build_vec_type(gallivm, fs_type), "");
1608 }
1609 }
1610 }
1611
1612 /*
1613 * Pixel twiddle from fragment shader order to memory order
1614 */
1615 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
1616 dst_channels, fs_src, src, pad_inline);
1617 if (dual_source_blend) {
1618 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
1619 fs_src1, src1, pad_inline);
1620 }
1621
1622 src_channels = dst_channels < 3 ? dst_channels : 4;
1623 if (src_count != num_fullblock_fs * src_channels) {
1624 unsigned ds = src_count / (num_fullblock_fs * src_channels);
1625 row_type.length /= ds;
1626 fs_type.length = row_type.length;
1627 }
1628
1629 blend_type = row_type;
1630 mask_type.length = 4;
1631
1632 /* Convert src to row_type */
1633 if (dual_source_blend) {
1634 struct lp_type old_row_type = row_type;
1635 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1636 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1637 }
1638 else {
1639 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1640 }
1641
1642 /* If the rows are not an SSE vector, combine them to become SSE size! */
1643 if ((row_type.width * row_type.length) % 128) {
1644 unsigned bits = row_type.width * row_type.length;
1645 unsigned combined;
1646
1647 assert(src_count >= (vector_width / bits));
1648
1649 dst_count = src_count / (vector_width / bits);
1650
1651 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1652 if (dual_source_blend) {
1653 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1654 }
1655
1656 row_type.length *= combined;
1657 src_count /= combined;
1658
1659 bits = row_type.width * row_type.length;
1660 assert(bits == 128 || bits == 256);
1661 }
1662
1663
1664 /*
1665 * Blend Colour conversion
1666 */
1667 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1668 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1669 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1670
1671 /* Convert */
1672 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1673
1674 /* Extract alpha */
1675 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1676
1677 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1678 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1679 if (pad_inline) {
1680 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1681 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1682 } else {
1683 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1684 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1685 }
1686
1687 /*
1688 * Mask conversion
1689 */
1690 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
1691
1692 if (src_count < block_height) {
1693 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1694 } else if (src_count > block_height) {
1695 for (i = src_count; i > 0; --i) {
1696 unsigned pixels = block_size / src_count;
1697 unsigned idx = i - 1;
1698
1699 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
1700 (idx * pixels) % 4, pixels);
1701 }
1702 }
1703
1704 assert(mask_type.width == 32);
1705
1706 for (i = 0; i < src_count; ++i) {
1707 unsigned pixels = block_size / src_count;
1708 unsigned pixel_width = row_type.width * dst_channels;
1709
1710 if (pixel_width == 24) {
1711 mask_type.width = 8;
1712 mask_type.length = vector_width / mask_type.width;
1713 } else {
1714 mask_type.length = pixels;
1715 mask_type.width = row_type.width * dst_channels;
1716
1717 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1718
1719 mask_type.length *= dst_channels;
1720 mask_type.width /= dst_channels;
1721 }
1722
1723 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1724 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1725 }
1726
1727 /*
1728 * Alpha conversion
1729 */
1730 if (!has_alpha) {
1731 struct lp_type alpha_type = fs_type;
1732 alpha_type.length = 4;
1733 convert_alpha(gallivm, row_type, alpha_type,
1734 block_size, block_height,
1735 src_count, dst_channels,
1736 pad_inline, src_alpha);
1737 if (dual_source_blend) {
1738 convert_alpha(gallivm, row_type, alpha_type,
1739 block_size, block_height,
1740 src_count, dst_channels,
1741 pad_inline, src1_alpha);
1742 }
1743 }
1744
1745
1746 /*
1747 * Load dst from memory
1748 */
1749 if (src_count < block_height) {
1750 dst_count = block_height;
1751 } else {
1752 dst_count = src_count;
1753 }
1754
1755 dst_type.length *= block_size / dst_count;
1756
1757 if (format_expands_to_float_soa(out_format_desc)) {
1758 /*
1759 * we need multiple values at once for the conversion, so can as well
1760 * load them vectorized here too instead of concatenating later.
1761 * (Still need concatenation later for 8-wide vectors).
1762 */
1763 dst_count = block_height;
1764 dst_type.length = block_width;
1765 }
1766
1767 /*
1768 * Compute the alignment of the destination pointer in bytes
1769 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
1770 * are always aligned by MIN2(16, fetch_width) except for buffers (not
1771 * 1d tex but can't distinguish here) so need to stick with per-pixel
1772 * alignment in this case.
1773 */
1774 if (is_1d) {
1775 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
1776 }
1777 else {
1778 dst_alignment = dst_type.length * dst_type.width / 8;
1779 }
1780 /* Force power-of-two alignment by extracting only the least-significant-bit */
1781 dst_alignment = 1 << (ffs(dst_alignment) - 1);
1782 /*
1783 * Resource base and stride pointers are aligned to 16 bytes, so that's
1784 * the maximum alignment we can guarantee
1785 */
1786 dst_alignment = MIN2(16, dst_alignment);
1787
1788 if (is_1d) {
1789 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
1790 dst, dst_type, dst_count / 4, dst_alignment);
1791 for (i = dst_count / 4; i < dst_count; i++) {
1792 dst[i] = lp_build_undef(gallivm, dst_type);
1793 }
1794
1795 }
1796 else {
1797 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1798 dst, dst_type, dst_count, dst_alignment);
1799 }
1800
1801
1802 /*
1803 * Convert from dst/output format to src/blending format.
1804 *
1805 * This is necessary as we can only read 1 row from memory at a time,
1806 * so the minimum dst_count will ever be at this point is 4.
1807 *
1808 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1809 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1810 * on all 16 pixels in that single vector at once.
1811 */
1812 if (dst_count > src_count) {
1813 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
1814 }
1815
1816 /*
1817 * Blending
1818 */
1819 /* XXX this is broken for RGB8 formats -
1820 * they get expanded from 12 to 16 elements (to include alpha)
1821 * by convert_to_blend_type then reduced to 15 instead of 12
1822 * by convert_from_blend_type (a simple fix though breaks A8...).
1823 * R16G16B16 also crashes differently however something going wrong
1824 * inside llvm handling npot vector sizes seemingly.
1825 * It seems some cleanup could be done here (like skipping conversion/blend
1826 * when not needed).
1827 */
1828 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
1829
1830 for (i = 0; i < src_count; ++i) {
1831 dst[i] = lp_build_blend_aos(gallivm,
1832 &variant->key.blend,
1833 out_format,
1834 row_type,
1835 rt,
1836 src[i],
1837 has_alpha ? NULL : src_alpha[i],
1838 src1[i],
1839 has_alpha ? NULL : src1_alpha[i],
1840 dst[i],
1841 partial_mask ? src_mask[i] : NULL,
1842 blend_color,
1843 has_alpha ? NULL : blend_alpha,
1844 swizzle,
1845 pad_inline ? 4 : dst_channels);
1846 }
1847
1848 convert_from_blend_type(gallivm, block_size, out_format_desc, row_type, dst_type, dst, src_count);
1849
1850 /* Split the blend rows back to memory rows */
1851 if (dst_count > src_count) {
1852 row_type.length = dst_type.length * (dst_count / src_count);
1853
1854 if (src_count == 1) {
1855 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1856 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1857
1858 row_type.length /= 2;
1859 src_count *= 2;
1860 }
1861
1862 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
1863 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
1864 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
1865 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
1866
1867 row_type.length /= 2;
1868 src_count *= 2;
1869 }
1870
1871 /*
1872 * Store blend result to memory
1873 */
1874 if (is_1d) {
1875 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
1876 dst, dst_type, dst_count / 4, dst_alignment);
1877 }
1878 else {
1879 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
1880 dst, dst_type, dst_count, dst_alignment);
1881 }
1882
1883 if (do_branch) {
1884 lp_build_mask_end(&mask_ctx);
1885 }
1886 }
1887
1888
1889 /**
1890 * Generate the runtime callable function for the whole fragment pipeline.
1891 * Note that the function which we generate operates on a block of 16
1892 * pixels at at time. The block contains 2x2 quads. Each quad contains
1893 * 2x2 pixels.
1894 */
1895 static void
1896 generate_fragment(struct llvmpipe_context *lp,
1897 struct lp_fragment_shader *shader,
1898 struct lp_fragment_shader_variant *variant,
1899 unsigned partial_mask)
1900 {
1901 struct gallivm_state *gallivm = variant->gallivm;
1902 const struct lp_fragment_shader_variant_key *key = &variant->key;
1903 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
1904 char func_name[256];
1905 struct lp_type fs_type;
1906 struct lp_type blend_type;
1907 LLVMTypeRef fs_elem_type;
1908 LLVMTypeRef blend_vec_type;
1909 LLVMTypeRef arg_types[13];
1910 LLVMTypeRef func_type;
1911 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1912 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1913 LLVMValueRef context_ptr;
1914 LLVMValueRef x;
1915 LLVMValueRef y;
1916 LLVMValueRef a0_ptr;
1917 LLVMValueRef dadx_ptr;
1918 LLVMValueRef dady_ptr;
1919 LLVMValueRef color_ptr_ptr;
1920 LLVMValueRef stride_ptr;
1921 LLVMValueRef depth_ptr;
1922 LLVMValueRef depth_stride;
1923 LLVMValueRef mask_input;
1924 LLVMValueRef thread_data_ptr;
1925 LLVMBasicBlockRef block;
1926 LLVMBuilderRef builder;
1927 struct lp_build_sampler_soa *sampler;
1928 struct lp_build_interp_soa_context interp;
1929 LLVMValueRef fs_mask[16 / 4];
1930 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
1931 LLVMValueRef function;
1932 LLVMValueRef facing;
1933 unsigned num_fs;
1934 unsigned i;
1935 unsigned chan;
1936 unsigned cbuf;
1937 boolean cbuf0_write_all;
1938 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
1939 util_blend_state_is_dual(&key->blend, 0);
1940
1941 assert(lp_native_vector_width / 32 >= 4);
1942
1943 /* Adjust color input interpolation according to flatshade state:
1944 */
1945 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
1946 for (i = 0; i < shader->info.base.num_inputs; i++) {
1947 if (inputs[i].interp == LP_INTERP_COLOR) {
1948 if (key->flatshade)
1949 inputs[i].interp = LP_INTERP_CONSTANT;
1950 else
1951 inputs[i].interp = LP_INTERP_PERSPECTIVE;
1952 }
1953 }
1954
1955 /* check if writes to cbuf[0] are to be copied to all cbufs */
1956 cbuf0_write_all = FALSE;
1957 for (i = 0;i < shader->info.base.num_properties; i++) {
1958 if (shader->info.base.properties[i].name ==
1959 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
1960 cbuf0_write_all = TRUE;
1961 break;
1962 }
1963 }
1964
1965 /* TODO: actually pick these based on the fs and color buffer
1966 * characteristics. */
1967
1968 memset(&fs_type, 0, sizeof fs_type);
1969 fs_type.floating = TRUE; /* floating point values */
1970 fs_type.sign = TRUE; /* values are signed */
1971 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1972 fs_type.width = 32; /* 32-bit float */
1973 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
1974
1975 memset(&blend_type, 0, sizeof blend_type);
1976 blend_type.floating = FALSE; /* values are integers */
1977 blend_type.sign = FALSE; /* values are unsigned */
1978 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
1979 blend_type.width = 8; /* 8-bit ubyte values */
1980 blend_type.length = 16; /* 16 elements per vector */
1981
1982 /*
1983 * Generate the function prototype. Any change here must be reflected in
1984 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
1985 */
1986
1987 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
1988
1989 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
1990
1991 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
1992 shader->no, variant->no, partial_mask ? "partial" : "whole");
1993
1994 arg_types[0] = variant->jit_context_ptr_type; /* context */
1995 arg_types[1] = int32_type; /* x */
1996 arg_types[2] = int32_type; /* y */
1997 arg_types[3] = int32_type; /* facing */
1998 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
1999 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2000 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2001 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
2002 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2003 arg_types[9] = int32_type; /* mask_input */
2004 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2005 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2006 arg_types[12] = int32_type; /* depth_stride */
2007
2008 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2009 arg_types, Elements(arg_types), 0);
2010
2011 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2012 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2013
2014 variant->function[partial_mask] = function;
2015
2016 /* XXX: need to propagate noalias down into color param now we are
2017 * passing a pointer-to-pointer?
2018 */
2019 for(i = 0; i < Elements(arg_types); ++i)
2020 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2021 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
2022
2023 context_ptr = LLVMGetParam(function, 0);
2024 x = LLVMGetParam(function, 1);
2025 y = LLVMGetParam(function, 2);
2026 facing = LLVMGetParam(function, 3);
2027 a0_ptr = LLVMGetParam(function, 4);
2028 dadx_ptr = LLVMGetParam(function, 5);
2029 dady_ptr = LLVMGetParam(function, 6);
2030 color_ptr_ptr = LLVMGetParam(function, 7);
2031 depth_ptr = LLVMGetParam(function, 8);
2032 mask_input = LLVMGetParam(function, 9);
2033 thread_data_ptr = LLVMGetParam(function, 10);
2034 stride_ptr = LLVMGetParam(function, 11);
2035 depth_stride = LLVMGetParam(function, 12);
2036
2037 lp_build_name(context_ptr, "context");
2038 lp_build_name(x, "x");
2039 lp_build_name(y, "y");
2040 lp_build_name(a0_ptr, "a0");
2041 lp_build_name(dadx_ptr, "dadx");
2042 lp_build_name(dady_ptr, "dady");
2043 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2044 lp_build_name(depth_ptr, "depth");
2045 lp_build_name(thread_data_ptr, "thread_data");
2046 lp_build_name(mask_input, "mask_input");
2047 lp_build_name(stride_ptr, "stride_ptr");
2048 lp_build_name(depth_stride, "depth_stride");
2049
2050 /*
2051 * Function body
2052 */
2053
2054 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2055 builder = gallivm->builder;
2056 assert(builder);
2057 LLVMPositionBuilderAtEnd(builder, block);
2058
2059 /* code generated texture sampling */
2060 sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
2061
2062 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2063 /* for 1d resources only run "upper half" of stamp */
2064 if (key->resource_1d)
2065 num_fs /= 2;
2066
2067 {
2068 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2069 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2070 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2071 num_loop, "mask_store");
2072 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2073
2074 /*
2075 * The shader input interpolation info is not explicitely baked in the
2076 * shader key, but everything it derives from (TGSI, and flatshade) is
2077 * already included in the shader key.
2078 */
2079 lp_build_interp_soa_init(&interp,
2080 gallivm,
2081 shader->info.base.num_inputs,
2082 inputs,
2083 shader->info.base.pixel_center_integer,
2084 builder, fs_type,
2085 a0_ptr, dadx_ptr, dady_ptr,
2086 x, y);
2087
2088 for (i = 0; i < num_fs; i++) {
2089 LLVMValueRef mask;
2090 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2091 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2092 &indexi, 1, "mask_ptr");
2093
2094 if (partial_mask) {
2095 mask = generate_quad_mask(gallivm, fs_type,
2096 i*fs_type.length/4, mask_input);
2097 }
2098 else {
2099 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2100 }
2101 LLVMBuildStore(builder, mask, mask_ptr);
2102 }
2103
2104 generate_fs_loop(gallivm,
2105 shader, key,
2106 builder,
2107 fs_type,
2108 context_ptr,
2109 num_loop,
2110 &interp,
2111 sampler,
2112 mask_store, /* output */
2113 color_store,
2114 depth_ptr,
2115 depth_stride,
2116 facing,
2117 thread_data_ptr);
2118
2119 for (i = 0; i < num_fs; i++) {
2120 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2121 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2122 &indexi, 1, "");
2123 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2124 /* This is fucked up need to reorganize things */
2125 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2126 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2127 ptr = LLVMBuildGEP(builder,
2128 color_store[cbuf * !cbuf0_write_all][chan],
2129 &indexi, 1, "");
2130 fs_out_color[cbuf][chan][i] = ptr;
2131 }
2132 }
2133 if (dual_source_blend) {
2134 /* only support one dual source blend target hence always use output 1 */
2135 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2136 ptr = LLVMBuildGEP(builder,
2137 color_store[1][chan],
2138 &indexi, 1, "");
2139 fs_out_color[1][chan][i] = ptr;
2140 }
2141 }
2142 }
2143 }
2144
2145 sampler->destroy(sampler);
2146
2147 /* Loop over color outputs / color buffers to do blending.
2148 */
2149 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2150 LLVMValueRef color_ptr;
2151 LLVMValueRef stride;
2152 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2153
2154 boolean do_branch = ((key->depth.enabled
2155 || key->stencil[0].enabled
2156 || key->alpha.enabled)
2157 && !shader->info.base.uses_kill);
2158
2159 color_ptr = LLVMBuildLoad(builder,
2160 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
2161 "");
2162
2163 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2164
2165 stride = LLVMBuildLoad(builder,
2166 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2167 "");
2168
2169 generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
2170 num_fs, fs_type, fs_mask, fs_out_color,
2171 context_ptr, color_ptr, stride, partial_mask, do_branch);
2172 }
2173
2174 LLVMBuildRetVoid(builder);
2175
2176 gallivm_verify_function(gallivm, function);
2177
2178 variant->nr_instrs += lp_build_count_instructions(function);
2179 }
2180
2181
2182 static void
2183 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2184 {
2185 unsigned i;
2186
2187 debug_printf("fs variant %p:\n", (void *) key);
2188
2189 if (key->flatshade) {
2190 debug_printf("flatshade = 1\n");
2191 }
2192 for (i = 0; i < key->nr_cbufs; ++i) {
2193 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2194 }
2195 if (key->depth.enabled) {
2196 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2197 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2198 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2199 }
2200
2201 for (i = 0; i < 2; ++i) {
2202 if (key->stencil[i].enabled) {
2203 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2204 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2205 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2206 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2207 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2208 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2209 }
2210 }
2211
2212 if (key->alpha.enabled) {
2213 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2214 }
2215
2216 if (key->occlusion_count) {
2217 debug_printf("occlusion_count = 1\n");
2218 }
2219
2220 if (key->blend.logicop_enable) {
2221 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2222 }
2223 else if (key->blend.rt[0].blend_enable) {
2224 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2225 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2226 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2227 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2228 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2229 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2230 }
2231 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2232 for (i = 0; i < key->nr_samplers; ++i) {
2233 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2234 debug_printf("sampler[%u] = \n", i);
2235 debug_printf(" .wrap = %s %s %s\n",
2236 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2237 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2238 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2239 debug_printf(" .min_img_filter = %s\n",
2240 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2241 debug_printf(" .min_mip_filter = %s\n",
2242 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2243 debug_printf(" .mag_img_filter = %s\n",
2244 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2245 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2246 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2247 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2248 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2249 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2250 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2251 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2252 }
2253 for (i = 0; i < key->nr_sampler_views; ++i) {
2254 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2255 debug_printf("texture[%u] = \n", i);
2256 debug_printf(" .format = %s\n",
2257 util_format_name(texture->format));
2258 debug_printf(" .target = %s\n",
2259 util_dump_tex_target(texture->target, TRUE));
2260 debug_printf(" .level_zero_only = %u\n",
2261 texture->level_zero_only);
2262 debug_printf(" .pot = %u %u %u\n",
2263 texture->pot_width,
2264 texture->pot_height,
2265 texture->pot_depth);
2266 }
2267 }
2268
2269
2270 void
2271 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2272 {
2273 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2274 variant->shader->no, variant->no);
2275 tgsi_dump(variant->shader->base.tokens, 0);
2276 dump_fs_variant_key(&variant->key);
2277 debug_printf("variant->opaque = %u\n", variant->opaque);
2278 debug_printf("\n");
2279 }
2280
2281
2282 /**
2283 * Generate a new fragment shader variant from the shader code and
2284 * other state indicated by the key.
2285 */
2286 static struct lp_fragment_shader_variant *
2287 generate_variant(struct llvmpipe_context *lp,
2288 struct lp_fragment_shader *shader,
2289 const struct lp_fragment_shader_variant_key *key)
2290 {
2291 struct lp_fragment_shader_variant *variant;
2292 const struct util_format_description *cbuf0_format_desc;
2293 boolean fullcolormask;
2294
2295 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2296 if(!variant)
2297 return NULL;
2298
2299 variant->gallivm = gallivm_create();
2300 if (!variant->gallivm) {
2301 FREE(variant);
2302 return NULL;
2303 }
2304
2305 variant->shader = shader;
2306 variant->list_item_global.base = variant;
2307 variant->list_item_local.base = variant;
2308 variant->no = shader->variants_created++;
2309
2310 memcpy(&variant->key, key, shader->variant_key_size);
2311
2312 /*
2313 * Determine whether we are touching all channels in the color buffer.
2314 */
2315 fullcolormask = FALSE;
2316 if (key->nr_cbufs == 1) {
2317 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2318 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2319 }
2320
2321 variant->opaque =
2322 !key->blend.logicop_enable &&
2323 !key->blend.rt[0].blend_enable &&
2324 fullcolormask &&
2325 !key->stencil[0].enabled &&
2326 !key->alpha.enabled &&
2327 !key->depth.enabled &&
2328 !shader->info.base.uses_kill
2329 ? TRUE : FALSE;
2330
2331 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2332 lp_debug_fs_variant(variant);
2333 }
2334
2335 lp_jit_init_types(variant);
2336
2337 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2338 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2339
2340 if (variant->jit_function[RAST_WHOLE] == NULL) {
2341 if (variant->opaque) {
2342 /* Specialized shader, which doesn't need to read the color buffer. */
2343 generate_fragment(lp, shader, variant, RAST_WHOLE);
2344 }
2345 }
2346
2347 /*
2348 * Compile everything
2349 */
2350
2351 gallivm_compile_module(variant->gallivm);
2352
2353 if (variant->function[RAST_EDGE_TEST]) {
2354 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2355 gallivm_jit_function(variant->gallivm,
2356 variant->function[RAST_EDGE_TEST]);
2357 }
2358
2359 if (variant->function[RAST_WHOLE]) {
2360 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2361 gallivm_jit_function(variant->gallivm,
2362 variant->function[RAST_WHOLE]);
2363 } else if (!variant->jit_function[RAST_WHOLE]) {
2364 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2365 }
2366
2367 return variant;
2368 }
2369
2370
2371 static void *
2372 llvmpipe_create_fs_state(struct pipe_context *pipe,
2373 const struct pipe_shader_state *templ)
2374 {
2375 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2376 struct lp_fragment_shader *shader;
2377 int nr_samplers;
2378 int nr_sampler_views;
2379 int i;
2380
2381 shader = CALLOC_STRUCT(lp_fragment_shader);
2382 if (!shader)
2383 return NULL;
2384
2385 shader->no = fs_no++;
2386 make_empty_list(&shader->variants);
2387
2388 /* get/save the summary info for this shader */
2389 lp_build_tgsi_info(templ->tokens, &shader->info);
2390
2391 /* we need to keep a local copy of the tokens */
2392 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2393
2394 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2395 if (shader->draw_data == NULL) {
2396 FREE((void *) shader->base.tokens);
2397 FREE(shader);
2398 return NULL;
2399 }
2400
2401 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2402 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2403
2404 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2405 state[MAX2(nr_samplers, nr_sampler_views)]);
2406
2407 for (i = 0; i < shader->info.base.num_inputs; i++) {
2408 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2409 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2410
2411 switch (shader->info.base.input_interpolate[i]) {
2412 case TGSI_INTERPOLATE_CONSTANT:
2413 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2414 break;
2415 case TGSI_INTERPOLATE_LINEAR:
2416 shader->inputs[i].interp = LP_INTERP_LINEAR;
2417 break;
2418 case TGSI_INTERPOLATE_PERSPECTIVE:
2419 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2420 break;
2421 case TGSI_INTERPOLATE_COLOR:
2422 shader->inputs[i].interp = LP_INTERP_COLOR;
2423 break;
2424 default:
2425 assert(0);
2426 break;
2427 }
2428
2429 switch (shader->info.base.input_semantic_name[i]) {
2430 case TGSI_SEMANTIC_FACE:
2431 shader->inputs[i].interp = LP_INTERP_FACING;
2432 break;
2433 case TGSI_SEMANTIC_POSITION:
2434 /* Position was already emitted above
2435 */
2436 shader->inputs[i].interp = LP_INTERP_POSITION;
2437 shader->inputs[i].src_index = 0;
2438 continue;
2439 }
2440
2441 shader->inputs[i].src_index = i+1;
2442 }
2443
2444 if (LP_DEBUG & DEBUG_TGSI) {
2445 unsigned attrib;
2446 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2447 shader->no, (void *) shader);
2448 tgsi_dump(templ->tokens, 0);
2449 debug_printf("usage masks:\n");
2450 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2451 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2452 debug_printf(" IN[%u].%s%s%s%s\n",
2453 attrib,
2454 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2455 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2456 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2457 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2458 }
2459 debug_printf("\n");
2460 }
2461
2462 return shader;
2463 }
2464
2465
2466 static void
2467 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2468 {
2469 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2470
2471 if (llvmpipe->fs == fs)
2472 return;
2473
2474 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2475
2476 draw_bind_fragment_shader(llvmpipe->draw,
2477 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2478
2479 llvmpipe->dirty |= LP_NEW_FS;
2480 }
2481
2482
2483 /**
2484 * Remove shader variant from two lists: the shader's variant list
2485 * and the context's variant list.
2486 */
2487 void
2488 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2489 struct lp_fragment_shader_variant *variant)
2490 {
2491 unsigned i;
2492
2493 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2494 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2495 " #%u v total cached #%u\n",
2496 variant->shader->no,
2497 variant->no,
2498 variant->shader->variants_created,
2499 variant->shader->variants_cached,
2500 lp->nr_fs_variants);
2501 }
2502
2503 /* free all the variant's JIT'd functions */
2504 for (i = 0; i < Elements(variant->function); i++) {
2505 if (variant->function[i]) {
2506 gallivm_free_function(variant->gallivm,
2507 variant->function[i],
2508 variant->jit_function[i]);
2509 }
2510 }
2511
2512 gallivm_destroy(variant->gallivm);
2513
2514 /* remove from shader's list */
2515 remove_from_list(&variant->list_item_local);
2516 variant->shader->variants_cached--;
2517
2518 /* remove from context's list */
2519 remove_from_list(&variant->list_item_global);
2520 lp->nr_fs_variants--;
2521 lp->nr_fs_instrs -= variant->nr_instrs;
2522
2523 FREE(variant);
2524 }
2525
2526
2527 static void
2528 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2529 {
2530 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2531 struct lp_fragment_shader *shader = fs;
2532 struct lp_fs_variant_list_item *li;
2533
2534 assert(fs != llvmpipe->fs);
2535
2536 /*
2537 * XXX: we need to flush the context until we have some sort of reference
2538 * counting in fragment shaders as they may still be binned
2539 * Flushing alone might not sufficient we need to wait on it too.
2540 */
2541 llvmpipe_finish(pipe, __FUNCTION__);
2542
2543 /* Delete all the variants */
2544 li = first_elem(&shader->variants);
2545 while(!at_end(&shader->variants, li)) {
2546 struct lp_fs_variant_list_item *next = next_elem(li);
2547 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2548 li = next;
2549 }
2550
2551 /* Delete draw module's data */
2552 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2553
2554 assert(shader->variants_cached == 0);
2555 FREE((void *) shader->base.tokens);
2556 FREE(shader);
2557 }
2558
2559
2560
2561 static void
2562 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2563 uint shader, uint index,
2564 struct pipe_constant_buffer *cb)
2565 {
2566 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2567 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2568
2569 assert(shader < PIPE_SHADER_TYPES);
2570 assert(index < Elements(llvmpipe->constants[shader]));
2571
2572 /* note: reference counting */
2573 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2574
2575 if (shader == PIPE_SHADER_VERTEX ||
2576 shader == PIPE_SHADER_GEOMETRY) {
2577 /* Pass the constants to the 'draw' module */
2578 const unsigned size = cb ? cb->buffer_size : 0;
2579 const ubyte *data;
2580
2581 if (constants) {
2582 data = (ubyte *) llvmpipe_resource_data(constants);
2583 }
2584 else if (cb && cb->user_buffer) {
2585 data = (ubyte *) cb->user_buffer;
2586 }
2587 else {
2588 data = NULL;
2589 }
2590
2591 if (data)
2592 data += cb->buffer_offset;
2593
2594 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2595 index, data, size);
2596 }
2597
2598 llvmpipe->dirty |= LP_NEW_CONSTANTS;
2599
2600 if (cb && cb->user_buffer) {
2601 pipe_resource_reference(&constants, NULL);
2602 }
2603 }
2604
2605
2606 /**
2607 * Return the blend factor equivalent to a destination alpha of one.
2608 */
2609 static INLINE unsigned
2610 force_dst_alpha_one(unsigned factor)
2611 {
2612 switch(factor) {
2613 case PIPE_BLENDFACTOR_DST_ALPHA:
2614 return PIPE_BLENDFACTOR_ONE;
2615 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2616 return PIPE_BLENDFACTOR_ZERO;
2617 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2618 return PIPE_BLENDFACTOR_ZERO;
2619 }
2620
2621 return factor;
2622 }
2623
2624
2625 /**
2626 * We need to generate several variants of the fragment pipeline to match
2627 * all the combinations of the contributing state atoms.
2628 *
2629 * TODO: there is actually no reason to tie this to context state -- the
2630 * generated code could be cached globally in the screen.
2631 */
2632 static void
2633 make_variant_key(struct llvmpipe_context *lp,
2634 struct lp_fragment_shader *shader,
2635 struct lp_fragment_shader_variant_key *key)
2636 {
2637 unsigned i;
2638
2639 memset(key, 0, shader->variant_key_size);
2640
2641 if (lp->framebuffer.zsbuf) {
2642 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
2643 const struct util_format_description *zsbuf_desc =
2644 util_format_description(zsbuf_format);
2645
2646 if (lp->depth_stencil->depth.enabled &&
2647 util_format_has_depth(zsbuf_desc)) {
2648 key->zsbuf_format = zsbuf_format;
2649 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2650 }
2651 if (lp->depth_stencil->stencil[0].enabled &&
2652 util_format_has_stencil(zsbuf_desc)) {
2653 key->zsbuf_format = zsbuf_format;
2654 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2655 }
2656 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
2657 key->resource_1d = TRUE;
2658 }
2659 }
2660
2661 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2662 if (!lp->framebuffer.nr_cbufs ||
2663 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2664 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2665 }
2666 if(key->alpha.enabled)
2667 key->alpha.func = lp->depth_stencil->alpha.func;
2668 /* alpha.ref_value is passed in jit_context */
2669
2670 key->flatshade = lp->rasterizer->flatshade;
2671 if (lp->active_occlusion_queries) {
2672 key->occlusion_count = TRUE;
2673 }
2674
2675 if (lp->framebuffer.nr_cbufs) {
2676 memcpy(&key->blend, lp->blend, sizeof key->blend);
2677 }
2678
2679 key->nr_cbufs = lp->framebuffer.nr_cbufs;
2680
2681 if (!key->blend.independent_blend_enable) {
2682 /* we always need independent blend otherwise the fixups below won't work */
2683 for (i = 1; i < key->nr_cbufs; i++) {
2684 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
2685 }
2686 key->blend.independent_blend_enable = 1;
2687 }
2688
2689 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
2690 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
2691 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
2692 const struct util_format_description *format_desc;
2693
2694 key->cbuf_format[i] = format;
2695
2696 /*
2697 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2698 * mixing of 2d textures with height 1 and 1d textures, so make sure
2699 * we pick 1d if any cbuf or zsbuf is 1d.
2700 */
2701 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[0]->texture)) {
2702 key->resource_1d = TRUE;
2703 }
2704
2705 format_desc = util_format_description(format);
2706 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
2707 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
2708
2709 /*
2710 * Mask out color channels not present in the color buffer.
2711 */
2712 blend_rt->colormask &= util_format_colormask(format_desc);
2713
2714 /*
2715 * Disable blend for integer formats.
2716 */
2717 if (util_format_is_pure_integer(format)) {
2718 blend_rt->blend_enable = 0;
2719 }
2720
2721 /*
2722 * Our swizzled render tiles always have an alpha channel, but the linear
2723 * render target format often does not, so force here the dst alpha to be
2724 * one.
2725 *
2726 * This is not a mere optimization. Wrong results will be produced if the
2727 * dst alpha is used, the dst format does not have alpha, and the previous
2728 * rendering was not flushed from the swizzled to linear buffer. For
2729 * example, NonPowTwo DCT.
2730 *
2731 * TODO: This should be generalized to all channels for better
2732 * performance, but only alpha causes correctness issues.
2733 *
2734 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2735 */
2736 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
2737 format_desc->swizzle[3] == format_desc->swizzle[0]) {
2738 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
2739 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
2740 blend_rt->alpha_func = blend_rt->rgb_func;
2741 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
2742 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
2743 }
2744 }
2745
2746 /* This value will be the same for all the variants of a given shader:
2747 */
2748 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2749
2750 for(i = 0; i < key->nr_samplers; ++i) {
2751 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2752 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
2753 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
2754 }
2755 }
2756
2757 /*
2758 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2759 * are dx10-style? Can't really have mixed opcodes, at least not
2760 * if we want to skip the holes here (without rescanning tgsi).
2761 */
2762 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2763 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2764 for(i = 0; i < key->nr_sampler_views; ++i) {
2765 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
2766 lp_sampler_static_texture_state(&key->state[i].texture_state,
2767 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2768 }
2769 }
2770 }
2771 else {
2772 key->nr_sampler_views = key->nr_samplers;
2773 for(i = 0; i < key->nr_sampler_views; ++i) {
2774 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
2775 lp_sampler_static_texture_state(&key->state[i].texture_state,
2776 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
2777 }
2778 }
2779 }
2780 }
2781
2782
2783
2784 /**
2785 * Update fragment shader state. This is called just prior to drawing
2786 * something when some fragment-related state has changed.
2787 */
2788 void
2789 llvmpipe_update_fs(struct llvmpipe_context *lp)
2790 {
2791 struct lp_fragment_shader *shader = lp->fs;
2792 struct lp_fragment_shader_variant_key key;
2793 struct lp_fragment_shader_variant *variant = NULL;
2794 struct lp_fs_variant_list_item *li;
2795
2796 make_variant_key(lp, shader, &key);
2797
2798 /* Search the variants for one which matches the key */
2799 li = first_elem(&shader->variants);
2800 while(!at_end(&shader->variants, li)) {
2801 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
2802 variant = li->base;
2803 break;
2804 }
2805 li = next_elem(li);
2806 }
2807
2808 if (variant) {
2809 /* Move this variant to the head of the list to implement LRU
2810 * deletion of shader's when we have too many.
2811 */
2812 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
2813 }
2814 else {
2815 /* variant not found, create it now */
2816 int64_t t0, t1, dt;
2817 unsigned i;
2818 unsigned variants_to_cull;
2819
2820 if (0) {
2821 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
2822 lp->nr_fs_variants,
2823 lp->nr_fs_instrs,
2824 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
2825 }
2826
2827 /* First, check if we've exceeded the max number of shader variants.
2828 * If so, free 25% of them (the least recently used ones).
2829 */
2830 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
2831
2832 if (variants_to_cull ||
2833 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
2834 struct pipe_context *pipe = &lp->pipe;
2835
2836 /*
2837 * XXX: we need to flush the context until we have some sort of
2838 * reference counting in fragment shaders as they may still be binned
2839 * Flushing alone might not be sufficient we need to wait on it too.
2840 */
2841 llvmpipe_finish(pipe, __FUNCTION__);
2842
2843 /*
2844 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
2845 * number of shader variants (potentially all of them) could be
2846 * pending for destruction on flush.
2847 */
2848
2849 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
2850 struct lp_fs_variant_list_item *item;
2851 if (is_empty_list(&lp->fs_variants_list)) {
2852 break;
2853 }
2854 item = last_elem(&lp->fs_variants_list);
2855 assert(item);
2856 assert(item->base);
2857 llvmpipe_remove_shader_variant(lp, item->base);
2858 }
2859 }
2860
2861 /*
2862 * Generate the new variant.
2863 */
2864 t0 = os_time_get();
2865 variant = generate_variant(lp, shader, &key);
2866 t1 = os_time_get();
2867 dt = t1 - t0;
2868 LP_COUNT_ADD(llvm_compile_time, dt);
2869 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
2870
2871 llvmpipe_variant_count++;
2872
2873 /* Put the new variant into the list */
2874 if (variant) {
2875 insert_at_head(&shader->variants, &variant->list_item_local);
2876 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
2877 lp->nr_fs_variants++;
2878 lp->nr_fs_instrs += variant->nr_instrs;
2879 shader->variants_cached++;
2880 }
2881 }
2882
2883 /* Bind this variant */
2884 lp_setup_set_fs_variant(lp->setup, variant);
2885 }
2886
2887
2888
2889
2890
2891 void
2892 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
2893 {
2894 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
2895 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
2896 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
2897
2898 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
2899 }