gallivm: implement aos unpack (to unorm8) for small unorm formats
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_fs.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 VMware, Inc.
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 VMWARE 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/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, bits_vec;
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 <= ARRAY_SIZE(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, 1ULL << (j + 0), 0);
178 bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
179 bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
180 bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
181 }
182 bits_vec = LLVMConstVector(bits, fs_type.length);
183 mask = LLVMBuildAnd(builder, mask, bits_vec, "");
184
185 /*
186 * mask = mask == bits ? ~0 : 0
187 */
188 mask = lp_build_compare(gallivm,
189 mask_type, PIPE_FUNC_EQUAL,
190 mask, bits_vec);
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 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
219 */
220 static LLVMValueRef
221 lp_llvm_viewport(LLVMValueRef context_ptr,
222 struct gallivm_state *gallivm,
223 LLVMValueRef viewport_index)
224 {
225 LLVMBuilderRef builder = gallivm->builder;
226 LLVMValueRef ptr;
227 LLVMValueRef res;
228 struct lp_type viewport_type =
229 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
230
231 ptr = lp_jit_context_viewports(gallivm, context_ptr);
232 ptr = LLVMBuildPointerCast(builder, ptr,
233 LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
234
235 res = lp_build_pointer_get(builder, ptr, viewport_index);
236
237 return res;
238 }
239
240
241 static LLVMValueRef
242 lp_build_depth_clamp(struct gallivm_state *gallivm,
243 LLVMBuilderRef builder,
244 struct lp_type type,
245 LLVMValueRef context_ptr,
246 LLVMValueRef thread_data_ptr,
247 LLVMValueRef z)
248 {
249 LLVMValueRef viewport, min_depth, max_depth;
250 LLVMValueRef viewport_index;
251 struct lp_build_context f32_bld;
252
253 assert(type.floating);
254 lp_build_context_init(&f32_bld, gallivm, type);
255
256 /*
257 * Assumes clamping of the viewport index will occur in setup/gs. Value
258 * is passed through the rasterization stage via lp_rast_shader_inputs.
259 *
260 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
261 * semantics.
262 */
263 viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
264 thread_data_ptr);
265
266 /*
267 * Load the min and max depth from the lp_jit_context.viewports
268 * array of lp_jit_viewport structures.
269 */
270 viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
271
272 /* viewports[viewport_index].min_depth */
273 min_depth = LLVMBuildExtractElement(builder, viewport,
274 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
275 min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
276
277 /* viewports[viewport_index].max_depth */
278 max_depth = LLVMBuildExtractElement(builder, viewport,
279 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
280 max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
281
282 /*
283 * Clamp to the min and max depth values for the given viewport.
284 */
285 return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
286 }
287
288
289 /**
290 * Generate the fragment shader, depth/stencil test, and alpha tests.
291 */
292 static void
293 generate_fs_loop(struct gallivm_state *gallivm,
294 struct lp_fragment_shader *shader,
295 const struct lp_fragment_shader_variant_key *key,
296 LLVMBuilderRef builder,
297 struct lp_type type,
298 LLVMValueRef context_ptr,
299 LLVMValueRef num_loop,
300 struct lp_build_interp_soa_context *interp,
301 struct lp_build_sampler_soa *sampler,
302 LLVMValueRef mask_store,
303 LLVMValueRef (*out_color)[4],
304 LLVMValueRef depth_ptr,
305 LLVMValueRef depth_stride,
306 LLVMValueRef facing,
307 LLVMValueRef thread_data_ptr)
308 {
309 const struct util_format_description *zs_format_desc = NULL;
310 const struct tgsi_token *tokens = shader->base.tokens;
311 struct lp_type int_type = lp_int_type(type);
312 LLVMTypeRef vec_type, int_vec_type;
313 LLVMValueRef mask_ptr, mask_val;
314 LLVMValueRef consts_ptr, num_consts_ptr;
315 LLVMValueRef z;
316 LLVMValueRef z_value, s_value;
317 LLVMValueRef z_fb, s_fb;
318 LLVMValueRef stencil_refs[2];
319 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
320 struct lp_build_for_loop_state loop_state;
321 struct lp_build_mask_context mask;
322 /*
323 * TODO: figure out if simple_shader optimization is really worthwile to
324 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
325 * code since tests tend to take another codepath than real shaders.
326 */
327 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
328 shader->info.base.num_inputs < 3 &&
329 shader->info.base.num_instructions < 8) && 0;
330 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
331 util_blend_state_is_dual(&key->blend, 0);
332 unsigned attrib;
333 unsigned chan;
334 unsigned cbuf;
335 unsigned depth_mode;
336
337 struct lp_bld_tgsi_system_values system_values;
338
339 memset(&system_values, 0, sizeof(system_values));
340
341 if (key->depth.enabled ||
342 key->stencil[0].enabled) {
343
344 zs_format_desc = util_format_description(key->zsbuf_format);
345 assert(zs_format_desc);
346
347 if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
348 if (key->alpha.enabled ||
349 key->blend.alpha_to_coverage ||
350 shader->info.base.uses_kill) {
351 /* With alpha test and kill, can do the depth test early
352 * and hopefully eliminate some quads. But need to do a
353 * special deferred depth write once the final mask value
354 * is known. This only works though if there's either no
355 * stencil test or the stencil value isn't written.
356 */
357 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
358 (key->stencil[1].enabled &&
359 key->stencil[1].writemask)))
360 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
361 else
362 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
363 }
364 else
365 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
366 }
367 else {
368 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
369 }
370
371 if (!(key->depth.enabled && key->depth.writemask) &&
372 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
373 (key->stencil[1].enabled &&
374 key->stencil[1].writemask))))
375 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
376 }
377 else {
378 depth_mode = 0;
379 }
380
381 vec_type = lp_build_vec_type(gallivm, type);
382 int_vec_type = lp_build_vec_type(gallivm, int_type);
383
384 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
385 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
386 /* convert scalar stencil refs into vectors */
387 stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
388 stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
389
390 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
391 num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
392
393 lp_build_for_loop_begin(&loop_state, gallivm,
394 lp_build_const_int32(gallivm, 0),
395 LLVMIntULT,
396 num_loop,
397 lp_build_const_int32(gallivm, 1));
398
399 mask_ptr = LLVMBuildGEP(builder, mask_store,
400 &loop_state.counter, 1, "mask_ptr");
401 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
402
403 memset(outputs, 0, sizeof outputs);
404
405 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
406 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
407 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
408 lp_build_vec_type(gallivm,
409 type),
410 num_loop, "color");
411 }
412 }
413 if (dual_source_blend) {
414 assert(key->nr_cbufs <= 1);
415 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
416 out_color[1][chan] = lp_build_array_alloca(gallivm,
417 lp_build_vec_type(gallivm,
418 type),
419 num_loop, "color1");
420 }
421 }
422
423
424 /* 'mask' will control execution based on quad's pixel alive/killed state */
425 lp_build_mask_begin(&mask, gallivm, type, mask_val);
426
427 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
428 lp_build_mask_check(&mask);
429
430 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
431 z = interp->pos[2];
432
433 if (depth_mode & EARLY_DEPTH_TEST) {
434 /*
435 * Clamp according to ARB_depth_clamp semantics.
436 */
437 if (key->depth_clamp) {
438 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
439 thread_data_ptr, z);
440 }
441 lp_build_depth_stencil_load_swizzled(gallivm, type,
442 zs_format_desc, key->resource_1d,
443 depth_ptr, depth_stride,
444 &z_fb, &s_fb, loop_state.counter);
445 lp_build_depth_stencil_test(gallivm,
446 &key->depth,
447 key->stencil,
448 type,
449 zs_format_desc,
450 &mask,
451 stencil_refs,
452 z, z_fb, s_fb,
453 facing,
454 &z_value, &s_value,
455 !simple_shader);
456
457 if (depth_mode & EARLY_DEPTH_WRITE) {
458 lp_build_depth_stencil_write_swizzled(gallivm, type,
459 zs_format_desc, key->resource_1d,
460 NULL, NULL, NULL, loop_state.counter,
461 depth_ptr, depth_stride,
462 z_value, s_value);
463 }
464 /*
465 * Note mask check if stencil is enabled must be after ds write not after
466 * stencil test otherwise new stencil values may not get written if all
467 * fragments got killed by depth/stencil test.
468 */
469 if (!simple_shader && key->stencil[0].enabled)
470 lp_build_mask_check(&mask);
471 }
472
473 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
474
475 /* Build the actual shader */
476 lp_build_tgsi_soa(gallivm, tokens, type, &mask,
477 consts_ptr, num_consts_ptr, &system_values,
478 interp->inputs,
479 outputs, context_ptr, thread_data_ptr,
480 sampler, &shader->info.base, NULL);
481
482 /* Alpha test */
483 if (key->alpha.enabled) {
484 int color0 = find_output_by_semantic(&shader->info.base,
485 TGSI_SEMANTIC_COLOR,
486 0);
487
488 if (color0 != -1 && outputs[color0][3]) {
489 const struct util_format_description *cbuf_format_desc;
490 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
491 LLVMValueRef alpha_ref_value;
492
493 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
494 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
495
496 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
497
498 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
499 &mask, alpha, alpha_ref_value,
500 (depth_mode & LATE_DEPTH_TEST) != 0);
501 }
502 }
503
504 /* Emulate Alpha to Coverage with Alpha test */
505 if (key->blend.alpha_to_coverage) {
506 int color0 = find_output_by_semantic(&shader->info.base,
507 TGSI_SEMANTIC_COLOR,
508 0);
509
510 if (color0 != -1 && outputs[color0][3]) {
511 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
512
513 lp_build_alpha_to_coverage(gallivm, type,
514 &mask, alpha,
515 (depth_mode & LATE_DEPTH_TEST) != 0);
516 }
517 }
518
519 /* Late Z test */
520 if (depth_mode & LATE_DEPTH_TEST) {
521 int pos0 = find_output_by_semantic(&shader->info.base,
522 TGSI_SEMANTIC_POSITION,
523 0);
524 int s_out = find_output_by_semantic(&shader->info.base,
525 TGSI_SEMANTIC_STENCIL,
526 0);
527 if (pos0 != -1 && outputs[pos0][2]) {
528 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
529 }
530 /*
531 * Clamp according to ARB_depth_clamp semantics.
532 */
533 if (key->depth_clamp) {
534 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
535 thread_data_ptr, z);
536 }
537
538 if (s_out != -1 && outputs[s_out][1]) {
539 /* there's only one value, and spec says to discard additional bits */
540 LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
541 stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
542 stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
543 stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
544 stencil_refs[1] = stencil_refs[0];
545 }
546
547 lp_build_depth_stencil_load_swizzled(gallivm, type,
548 zs_format_desc, key->resource_1d,
549 depth_ptr, depth_stride,
550 &z_fb, &s_fb, loop_state.counter);
551
552 lp_build_depth_stencil_test(gallivm,
553 &key->depth,
554 key->stencil,
555 type,
556 zs_format_desc,
557 &mask,
558 stencil_refs,
559 z, z_fb, s_fb,
560 facing,
561 &z_value, &s_value,
562 !simple_shader);
563 /* Late Z write */
564 if (depth_mode & LATE_DEPTH_WRITE) {
565 lp_build_depth_stencil_write_swizzled(gallivm, type,
566 zs_format_desc, key->resource_1d,
567 NULL, NULL, NULL, loop_state.counter,
568 depth_ptr, depth_stride,
569 z_value, s_value);
570 }
571 }
572 else if ((depth_mode & EARLY_DEPTH_TEST) &&
573 (depth_mode & LATE_DEPTH_WRITE))
574 {
575 /* Need to apply a reduced mask to the depth write. Reload the
576 * depth value, update from zs_value with the new mask value and
577 * write that out.
578 */
579 lp_build_depth_stencil_write_swizzled(gallivm, type,
580 zs_format_desc, key->resource_1d,
581 &mask, z_fb, s_fb, loop_state.counter,
582 depth_ptr, depth_stride,
583 z_value, s_value);
584 }
585
586
587 /* Color write */
588 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
589 {
590 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
591 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
592 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
593 {
594 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
595 if(outputs[attrib][chan]) {
596 /* XXX: just initialize outputs to point at colors[] and
597 * skip this.
598 */
599 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
600 LLVMValueRef color_ptr;
601 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
602 &loop_state.counter, 1, "");
603 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
604 LLVMBuildStore(builder, out, color_ptr);
605 }
606 }
607 }
608 }
609
610 if (key->occlusion_count) {
611 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
612 lp_build_name(counter, "counter");
613 lp_build_occlusion_count(gallivm, type,
614 lp_build_mask_value(&mask), counter);
615 }
616
617 mask_val = lp_build_mask_end(&mask);
618 LLVMBuildStore(builder, mask_val, mask_ptr);
619 lp_build_for_loop_end(&loop_state);
620 }
621
622
623 /**
624 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
625 *
626 * Fragment Shader outputs pixels in small 2x2 blocks
627 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
628 *
629 * However in memory pixels are stored in rows
630 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
631 *
632 * @param type fragment shader type (4x or 8x float)
633 * @param num_fs number of fs_src
634 * @param is_1d whether we're outputting to a 1d resource
635 * @param dst_channels number of output channels
636 * @param fs_src output from fragment shader
637 * @param dst pointer to store result
638 * @param pad_inline is channel padding inline or at end of row
639 * @return the number of dsts
640 */
641 static int
642 generate_fs_twiddle(struct gallivm_state *gallivm,
643 struct lp_type type,
644 unsigned num_fs,
645 unsigned dst_channels,
646 LLVMValueRef fs_src[][4],
647 LLVMValueRef* dst,
648 bool pad_inline)
649 {
650 LLVMValueRef src[16];
651
652 bool swizzle_pad;
653 bool twiddle;
654 bool split;
655
656 unsigned pixels = type.length / 4;
657 unsigned reorder_group;
658 unsigned src_channels;
659 unsigned src_count;
660 unsigned i;
661
662 src_channels = dst_channels < 3 ? dst_channels : 4;
663 src_count = num_fs * src_channels;
664
665 assert(pixels == 2 || pixels == 1);
666 assert(num_fs * src_channels <= ARRAY_SIZE(src));
667
668 /*
669 * Transpose from SoA -> AoS
670 */
671 for (i = 0; i < num_fs; ++i) {
672 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
673 }
674
675 /*
676 * Pick transformation options
677 */
678 swizzle_pad = false;
679 twiddle = false;
680 split = false;
681 reorder_group = 0;
682
683 if (dst_channels == 1) {
684 twiddle = true;
685
686 if (pixels == 2) {
687 split = true;
688 }
689 } else if (dst_channels == 2) {
690 if (pixels == 1) {
691 reorder_group = 1;
692 }
693 } else if (dst_channels > 2) {
694 if (pixels == 1) {
695 reorder_group = 2;
696 } else {
697 twiddle = true;
698 }
699
700 if (!pad_inline && dst_channels == 3 && pixels > 1) {
701 swizzle_pad = true;
702 }
703 }
704
705 /*
706 * Split the src in half
707 */
708 if (split) {
709 for (i = num_fs; i > 0; --i) {
710 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
711 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
712 }
713
714 src_count *= 2;
715 type.length = 4;
716 }
717
718 /*
719 * Ensure pixels are in memory order
720 */
721 if (reorder_group) {
722 /* Twiddle pixels by reordering the array, e.g.:
723 *
724 * src_count = 8 -> 0 2 1 3 4 6 5 7
725 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
726 */
727 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
728
729 for (i = 0; i < src_count; ++i) {
730 unsigned group = i / reorder_group;
731 unsigned block = (group / 4) * 4 * reorder_group;
732 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
733 dst[i] = src[j];
734 }
735 } else if (twiddle) {
736 /* Twiddle pixels across elements of array */
737 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
738 } else {
739 /* Do nothing */
740 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
741 }
742
743 /*
744 * Moves any padding between pixels to the end
745 * e.g. RGBXRGBX -> RGBRGBXX
746 */
747 if (swizzle_pad) {
748 unsigned char swizzles[16];
749 unsigned elems = pixels * dst_channels;
750
751 for (i = 0; i < type.length; ++i) {
752 if (i < elems)
753 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
754 else
755 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
756 }
757
758 for (i = 0; i < src_count; ++i) {
759 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
760 }
761 }
762
763 return src_count;
764 }
765
766
767 /**
768 * Load an unswizzled block of pixels from memory
769 */
770 static void
771 load_unswizzled_block(struct gallivm_state *gallivm,
772 LLVMValueRef base_ptr,
773 LLVMValueRef stride,
774 unsigned block_width,
775 unsigned block_height,
776 LLVMValueRef* dst,
777 struct lp_type dst_type,
778 unsigned dst_count,
779 unsigned dst_alignment)
780 {
781 LLVMBuilderRef builder = gallivm->builder;
782 unsigned row_size = dst_count / block_height;
783 unsigned i;
784
785 /* Ensure block exactly fits into dst */
786 assert((block_width * block_height) % dst_count == 0);
787
788 for (i = 0; i < dst_count; ++i) {
789 unsigned x = i % row_size;
790 unsigned y = i / row_size;
791
792 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
793 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
794
795 LLVMValueRef gep[2];
796 LLVMValueRef dst_ptr;
797
798 gep[0] = lp_build_const_int32(gallivm, 0);
799 gep[1] = LLVMBuildAdd(builder, bx, by, "");
800
801 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
802 dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
803
804 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
805
806 LLVMSetAlignment(dst[i], dst_alignment);
807 }
808 }
809
810
811 /**
812 * Store an unswizzled block of pixels to memory
813 */
814 static void
815 store_unswizzled_block(struct gallivm_state *gallivm,
816 LLVMValueRef base_ptr,
817 LLVMValueRef stride,
818 unsigned block_width,
819 unsigned block_height,
820 LLVMValueRef* src,
821 struct lp_type src_type,
822 unsigned src_count,
823 unsigned src_alignment)
824 {
825 LLVMBuilderRef builder = gallivm->builder;
826 unsigned row_size = src_count / block_height;
827 unsigned i;
828
829 /* Ensure src exactly fits into block */
830 assert((block_width * block_height) % src_count == 0);
831
832 for (i = 0; i < src_count; ++i) {
833 unsigned x = i % row_size;
834 unsigned y = i / row_size;
835
836 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
837 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
838
839 LLVMValueRef gep[2];
840 LLVMValueRef src_ptr;
841
842 gep[0] = lp_build_const_int32(gallivm, 0);
843 gep[1] = LLVMBuildAdd(builder, bx, by, "");
844
845 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
846 src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
847
848 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
849
850 LLVMSetAlignment(src_ptr, src_alignment);
851 }
852 }
853
854
855 /**
856 * Checks if a format description is an arithmetic format
857 *
858 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
859 */
860 static inline boolean
861 is_arithmetic_format(const struct util_format_description *format_desc)
862 {
863 boolean arith = false;
864 unsigned i;
865
866 for (i = 0; i < format_desc->nr_channels; ++i) {
867 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
868 arith |= (format_desc->channel[i].size % 8) != 0;
869 }
870
871 return arith;
872 }
873
874
875 /**
876 * Checks if this format requires special handling due to required expansion
877 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
878 * SoA conversion.
879 */
880 static inline boolean
881 format_expands_to_float_soa(const struct util_format_description *format_desc)
882 {
883 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
884 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
885 return true;
886 }
887 return false;
888 }
889
890
891 /**
892 * Retrieves the type representing the memory layout for a format
893 *
894 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
895 */
896 static inline void
897 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
898 struct lp_type* type)
899 {
900 unsigned i;
901 unsigned chan;
902
903 if (format_expands_to_float_soa(format_desc)) {
904 /* just make this a uint with width of block */
905 type->floating = false;
906 type->fixed = false;
907 type->sign = false;
908 type->norm = false;
909 type->width = format_desc->block.bits;
910 type->length = 1;
911 return;
912 }
913
914 for (i = 0; i < 4; i++)
915 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
916 break;
917 chan = i;
918
919 memset(type, 0, sizeof(struct lp_type));
920 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
921 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
922 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
923 type->norm = format_desc->channel[chan].normalized;
924
925 if (is_arithmetic_format(format_desc)) {
926 type->width = 0;
927 type->length = 1;
928
929 for (i = 0; i < format_desc->nr_channels; ++i) {
930 type->width += format_desc->channel[i].size;
931 }
932 } else {
933 type->width = format_desc->channel[chan].size;
934 type->length = format_desc->nr_channels;
935 }
936 }
937
938
939 /**
940 * Retrieves the type for a format which is usable in the blending code.
941 *
942 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
943 */
944 static inline void
945 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
946 struct lp_type* type)
947 {
948 unsigned i;
949 unsigned chan;
950
951 if (format_expands_to_float_soa(format_desc)) {
952 /* always use ordinary floats for blending */
953 type->floating = true;
954 type->fixed = false;
955 type->sign = true;
956 type->norm = false;
957 type->width = 32;
958 type->length = 4;
959 return;
960 }
961
962 for (i = 0; i < 4; i++)
963 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
964 break;
965 chan = i;
966
967 memset(type, 0, sizeof(struct lp_type));
968 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
969 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
970 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
971 type->norm = format_desc->channel[chan].normalized;
972 type->width = format_desc->channel[chan].size;
973 type->length = format_desc->nr_channels;
974
975 for (i = 1; i < format_desc->nr_channels; ++i) {
976 if (format_desc->channel[i].size > type->width)
977 type->width = format_desc->channel[i].size;
978 }
979
980 if (type->floating) {
981 type->width = 32;
982 } else {
983 if (type->width <= 8) {
984 type->width = 8;
985 } else if (type->width <= 16) {
986 type->width = 16;
987 } else {
988 type->width = 32;
989 }
990 }
991
992 if (is_arithmetic_format(format_desc) && type->length == 3) {
993 type->length = 4;
994 }
995 }
996
997
998 /**
999 * Scale a normalized value from src_bits to dst_bits.
1000 *
1001 * The exact calculation is
1002 *
1003 * dst = iround(src * dst_mask / src_mask)
1004 *
1005 * or with integer rounding
1006 *
1007 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1008 *
1009 * where
1010 *
1011 * src_mask = (1 << src_bits) - 1
1012 * dst_mask = (1 << dst_bits) - 1
1013 *
1014 * but we try to avoid division and multiplication through shifts.
1015 */
1016 static inline LLVMValueRef
1017 scale_bits(struct gallivm_state *gallivm,
1018 int src_bits,
1019 int dst_bits,
1020 LLVMValueRef src,
1021 struct lp_type src_type)
1022 {
1023 LLVMBuilderRef builder = gallivm->builder;
1024 LLVMValueRef result = src;
1025
1026 if (dst_bits < src_bits) {
1027 int delta_bits = src_bits - dst_bits;
1028
1029 if (delta_bits <= dst_bits) {
1030 /*
1031 * Approximate the rescaling with a single shift.
1032 *
1033 * This gives the wrong rounding.
1034 */
1035
1036 result = LLVMBuildLShr(builder,
1037 src,
1038 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1039 "");
1040
1041 } else {
1042 /*
1043 * Try more accurate rescaling.
1044 */
1045
1046 /*
1047 * Drop the least significant bits to make space for the multiplication.
1048 *
1049 * XXX: A better approach would be to use a wider integer type as intermediate. But
1050 * this is enough to convert alpha from 16bits -> 2 when rendering to
1051 * PIPE_FORMAT_R10G10B10A2_UNORM.
1052 */
1053 result = LLVMBuildLShr(builder,
1054 src,
1055 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1056 "");
1057
1058
1059 result = LLVMBuildMul(builder,
1060 result,
1061 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1062 "");
1063
1064 /*
1065 * Add a rounding term before the division.
1066 *
1067 * TODO: Handle signed integers too.
1068 */
1069 if (!src_type.sign) {
1070 result = LLVMBuildAdd(builder,
1071 result,
1072 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1073 "");
1074 }
1075
1076 /*
1077 * Approximate the division by src_mask with a src_bits shift.
1078 *
1079 * Given the src has already been shifted by dst_bits, all we need
1080 * to do is to shift by the difference.
1081 */
1082
1083 result = LLVMBuildLShr(builder,
1084 result,
1085 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1086 "");
1087 }
1088
1089 } else if (dst_bits > src_bits) {
1090 /* Scale up bits */
1091 int db = dst_bits - src_bits;
1092
1093 /* Shift left by difference in bits */
1094 result = LLVMBuildShl(builder,
1095 src,
1096 lp_build_const_int_vec(gallivm, src_type, db),
1097 "");
1098
1099 if (db <= src_bits) {
1100 /* Enough bits in src to fill the remainder */
1101 LLVMValueRef lower = LLVMBuildLShr(builder,
1102 src,
1103 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1104 "");
1105
1106 result = LLVMBuildOr(builder, result, lower, "");
1107 } else if (db > src_bits) {
1108 /* Need to repeatedly copy src bits to fill remainder in dst */
1109 unsigned n;
1110
1111 for (n = src_bits; n < dst_bits; n *= 2) {
1112 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1113
1114 result = LLVMBuildOr(builder,
1115 result,
1116 LLVMBuildLShr(builder, result, shuv, ""),
1117 "");
1118 }
1119 }
1120 }
1121
1122 return result;
1123 }
1124
1125 /**
1126 * If RT is a smallfloat (needing denorms) format
1127 */
1128 static inline int
1129 have_smallfloat_format(struct lp_type dst_type,
1130 enum pipe_format format)
1131 {
1132 return ((dst_type.floating && dst_type.width != 32) ||
1133 /* due to format handling hacks this format doesn't have floating set
1134 * here (and actually has width set to 32 too) so special case this. */
1135 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1136 }
1137
1138
1139 /**
1140 * Convert from memory format to blending format
1141 *
1142 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1143 */
1144 static void
1145 convert_to_blend_type(struct gallivm_state *gallivm,
1146 unsigned block_size,
1147 const struct util_format_description *src_fmt,
1148 struct lp_type src_type,
1149 struct lp_type dst_type,
1150 LLVMValueRef* src, // and dst
1151 unsigned num_srcs)
1152 {
1153 LLVMValueRef *dst = src;
1154 LLVMBuilderRef builder = gallivm->builder;
1155 struct lp_type blend_type;
1156 struct lp_type mem_type;
1157 unsigned i, j;
1158 unsigned pixels = block_size / num_srcs;
1159 bool is_arith;
1160
1161 /*
1162 * full custom path for packed floats and srgb formats - none of the later
1163 * functions would do anything useful, and given the lp_type representation they
1164 * can't be fixed. Should really have some SoA blend path for these kind of
1165 * formats rather than hacking them in here.
1166 */
1167 if (format_expands_to_float_soa(src_fmt)) {
1168 LLVMValueRef tmpsrc[4];
1169 /*
1170 * This is pretty suboptimal for this case blending in SoA would be much
1171 * better, since conversion gets us SoA values so need to convert back.
1172 */
1173 assert(src_type.width == 32 || src_type.width == 16);
1174 assert(dst_type.floating);
1175 assert(dst_type.width == 32);
1176 assert(dst_type.length % 4 == 0);
1177 assert(num_srcs % 4 == 0);
1178
1179 if (src_type.width == 16) {
1180 /* expand 4x16bit values to 4x32bit */
1181 struct lp_type type32x4 = src_type;
1182 LLVMTypeRef ltype32x4;
1183 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1184 type32x4.width = 32;
1185 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1186 for (i = 0; i < num_fetch; i++) {
1187 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1188 }
1189 src_type.width = 32;
1190 }
1191 for (i = 0; i < 4; i++) {
1192 tmpsrc[i] = src[i];
1193 }
1194 for (i = 0; i < num_srcs / 4; i++) {
1195 LLVMValueRef tmpsoa[4];
1196 LLVMValueRef tmps = tmpsrc[i];
1197 if (dst_type.length == 8) {
1198 LLVMValueRef shuffles[8];
1199 unsigned j;
1200 /* fetch was 4 values but need 8-wide output values */
1201 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1202 /*
1203 * for 8-wide aos transpose would give us wrong order not matching
1204 * incoming converted fs values and mask. ARGH.
1205 */
1206 for (j = 0; j < 4; j++) {
1207 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1208 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1209 }
1210 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1211 LLVMConstVector(shuffles, 8), "");
1212 }
1213 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1214 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1215 }
1216 else {
1217 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1218 }
1219 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1220 }
1221 return;
1222 }
1223
1224 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1225 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1226
1227 /* Is the format arithmetic */
1228 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1229 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1230
1231 /* Pad if necessary */
1232 if (!is_arith && src_type.length < dst_type.length) {
1233 for (i = 0; i < num_srcs; ++i) {
1234 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1235 }
1236
1237 src_type.length = dst_type.length;
1238 }
1239
1240 /* Special case for half-floats */
1241 if (mem_type.width == 16 && mem_type.floating) {
1242 assert(blend_type.width == 32 && blend_type.floating);
1243 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1244 is_arith = false;
1245 }
1246
1247 if (!is_arith) {
1248 return;
1249 }
1250
1251 src_type.width = blend_type.width * blend_type.length;
1252 blend_type.length *= pixels;
1253 src_type.length *= pixels / (src_type.length / mem_type.length);
1254
1255 for (i = 0; i < num_srcs; ++i) {
1256 LLVMValueRef chans[4];
1257 LLVMValueRef res = NULL;
1258
1259 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1260
1261 for (j = 0; j < src_fmt->nr_channels; ++j) {
1262 unsigned mask = 0;
1263 unsigned sa = src_fmt->channel[j].shift;
1264 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1265 unsigned from_lsb = j;
1266 #else
1267 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1268 #endif
1269
1270 mask = (1 << src_fmt->channel[j].size) - 1;
1271
1272 /* Extract bits from source */
1273 chans[j] = LLVMBuildLShr(builder,
1274 dst[i],
1275 lp_build_const_int_vec(gallivm, src_type, sa),
1276 "");
1277
1278 chans[j] = LLVMBuildAnd(builder,
1279 chans[j],
1280 lp_build_const_int_vec(gallivm, src_type, mask),
1281 "");
1282
1283 /* Scale bits */
1284 if (src_type.norm) {
1285 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1286 blend_type.width, chans[j], src_type);
1287 }
1288
1289 /* Insert bits into correct position */
1290 chans[j] = LLVMBuildShl(builder,
1291 chans[j],
1292 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1293 "");
1294
1295 if (j == 0) {
1296 res = chans[j];
1297 } else {
1298 res = LLVMBuildOr(builder, res, chans[j], "");
1299 }
1300 }
1301
1302 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1303 }
1304 }
1305
1306
1307 /**
1308 * Convert from blending format to memory format
1309 *
1310 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1311 */
1312 static void
1313 convert_from_blend_type(struct gallivm_state *gallivm,
1314 unsigned block_size,
1315 const struct util_format_description *src_fmt,
1316 struct lp_type src_type,
1317 struct lp_type dst_type,
1318 LLVMValueRef* src, // and dst
1319 unsigned num_srcs)
1320 {
1321 LLVMValueRef* dst = src;
1322 unsigned i, j, k;
1323 struct lp_type mem_type;
1324 struct lp_type blend_type;
1325 LLVMBuilderRef builder = gallivm->builder;
1326 unsigned pixels = block_size / num_srcs;
1327 bool is_arith;
1328
1329 /*
1330 * full custom path for packed floats and srgb formats - none of the later
1331 * functions would do anything useful, and given the lp_type representation they
1332 * can't be fixed. Should really have some SoA blend path for these kind of
1333 * formats rather than hacking them in here.
1334 */
1335 if (format_expands_to_float_soa(src_fmt)) {
1336 /*
1337 * This is pretty suboptimal for this case blending in SoA would be much
1338 * better - we need to transpose the AoS values back to SoA values for
1339 * conversion/packing.
1340 */
1341 assert(src_type.floating);
1342 assert(src_type.width == 32);
1343 assert(src_type.length % 4 == 0);
1344 assert(dst_type.width == 32 || dst_type.width == 16);
1345
1346 for (i = 0; i < num_srcs / 4; i++) {
1347 LLVMValueRef tmpsoa[4], tmpdst;
1348 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1349 /* really really need SoA here */
1350
1351 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1352 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1353 }
1354 else {
1355 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1356 src_type, tmpsoa);
1357 }
1358
1359 if (src_type.length == 8) {
1360 LLVMValueRef tmpaos, shuffles[8];
1361 unsigned j;
1362 /*
1363 * for 8-wide aos transpose has given us wrong order not matching
1364 * output order. HMPF. Also need to split the output values manually.
1365 */
1366 for (j = 0; j < 4; j++) {
1367 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1368 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1369 }
1370 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1371 LLVMConstVector(shuffles, 8), "");
1372 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1373 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1374 }
1375 else {
1376 src[i] = tmpdst;
1377 }
1378 }
1379 if (dst_type.width == 16) {
1380 struct lp_type type16x8 = dst_type;
1381 struct lp_type type32x4 = dst_type;
1382 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1383 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1384 type16x8.length = 8;
1385 type32x4.width = 32;
1386 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1387 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1388 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1389 /* We could do vector truncation but it doesn't generate very good code */
1390 for (i = 0; i < num_fetch; i++) {
1391 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1392 src[i], lp_build_zero(gallivm, type32x4));
1393 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1394 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1395 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1396 }
1397 }
1398 return;
1399 }
1400
1401 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1402 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1403
1404 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1405
1406 /* Special case for half-floats */
1407 if (mem_type.width == 16 && mem_type.floating) {
1408 int length = dst_type.length;
1409 assert(blend_type.width == 32 && blend_type.floating);
1410
1411 dst_type.length = src_type.length;
1412
1413 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1414
1415 dst_type.length = length;
1416 is_arith = false;
1417 }
1418
1419 /* Remove any padding */
1420 if (!is_arith && (src_type.length % mem_type.length)) {
1421 src_type.length -= (src_type.length % mem_type.length);
1422
1423 for (i = 0; i < num_srcs; ++i) {
1424 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1425 }
1426 }
1427
1428 /* No bit arithmetic to do */
1429 if (!is_arith) {
1430 return;
1431 }
1432
1433 src_type.length = pixels;
1434 src_type.width = blend_type.length * blend_type.width;
1435 dst_type.length = pixels;
1436
1437 for (i = 0; i < num_srcs; ++i) {
1438 LLVMValueRef chans[4];
1439 LLVMValueRef res = NULL;
1440
1441 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1442
1443 for (j = 0; j < src_fmt->nr_channels; ++j) {
1444 unsigned mask = 0;
1445 unsigned sa = src_fmt->channel[j].shift;
1446 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1447 unsigned from_lsb = j;
1448 #else
1449 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1450 #endif
1451
1452 assert(blend_type.width > src_fmt->channel[j].size);
1453
1454 for (k = 0; k < blend_type.width; ++k) {
1455 mask |= 1 << k;
1456 }
1457
1458 /* Extract bits */
1459 chans[j] = LLVMBuildLShr(builder,
1460 dst[i],
1461 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1462 "");
1463
1464 chans[j] = LLVMBuildAnd(builder,
1465 chans[j],
1466 lp_build_const_int_vec(gallivm, src_type, mask),
1467 "");
1468
1469 /* Scale down bits */
1470 if (src_type.norm) {
1471 chans[j] = scale_bits(gallivm, blend_type.width,
1472 src_fmt->channel[j].size, chans[j], src_type);
1473 }
1474
1475 /* Insert bits */
1476 chans[j] = LLVMBuildShl(builder,
1477 chans[j],
1478 lp_build_const_int_vec(gallivm, src_type, sa),
1479 "");
1480
1481 sa += src_fmt->channel[j].size;
1482
1483 if (j == 0) {
1484 res = chans[j];
1485 } else {
1486 res = LLVMBuildOr(builder, res, chans[j], "");
1487 }
1488 }
1489
1490 assert (dst_type.width != 24);
1491
1492 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1493 }
1494 }
1495
1496
1497 /**
1498 * Convert alpha to same blend type as src
1499 */
1500 static void
1501 convert_alpha(struct gallivm_state *gallivm,
1502 struct lp_type row_type,
1503 struct lp_type alpha_type,
1504 const unsigned block_size,
1505 const unsigned block_height,
1506 const unsigned src_count,
1507 const unsigned dst_channels,
1508 const bool pad_inline,
1509 LLVMValueRef* src_alpha)
1510 {
1511 LLVMBuilderRef builder = gallivm->builder;
1512 unsigned i, j;
1513 unsigned length = row_type.length;
1514 row_type.length = alpha_type.length;
1515
1516 /* Twiddle the alpha to match pixels */
1517 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1518
1519 /*
1520 * TODO this should use single lp_build_conv call for
1521 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1522 */
1523 for (i = 0; i < block_height; ++i) {
1524 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1525 }
1526
1527 alpha_type = row_type;
1528 row_type.length = length;
1529
1530 /* If only one channel we can only need the single alpha value per pixel */
1531 if (src_count == 1 && dst_channels == 1) {
1532
1533 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1534 } else {
1535 /* If there are more srcs than rows then we need to split alpha up */
1536 if (src_count > block_height) {
1537 for (i = src_count; i > 0; --i) {
1538 unsigned pixels = block_size / src_count;
1539 unsigned idx = i - 1;
1540
1541 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1542 (idx * pixels) % 4, pixels);
1543 }
1544 }
1545
1546 /* If there is a src for each pixel broadcast the alpha across whole row */
1547 if (src_count == block_size) {
1548 for (i = 0; i < src_count; ++i) {
1549 src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1550 }
1551 } else {
1552 unsigned pixels = block_size / src_count;
1553 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1554 unsigned alpha_span = 1;
1555 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1556
1557 /* Check if we need 2 src_alphas for our shuffles */
1558 if (pixels > alpha_type.length) {
1559 alpha_span = 2;
1560 }
1561
1562 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1563 for (j = 0; j < row_type.length; ++j) {
1564 if (j < pixels * channels) {
1565 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1566 } else {
1567 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1568 }
1569 }
1570
1571 for (i = 0; i < src_count; ++i) {
1572 unsigned idx1 = i, idx2 = i;
1573
1574 if (alpha_span > 1){
1575 idx1 *= alpha_span;
1576 idx2 = idx1 + 1;
1577 }
1578
1579 src_alpha[i] = LLVMBuildShuffleVector(builder,
1580 src_alpha[idx1],
1581 src_alpha[idx2],
1582 LLVMConstVector(shuffles, row_type.length),
1583 "");
1584 }
1585 }
1586 }
1587 }
1588
1589
1590 /**
1591 * Generates the blend function for unswizzled colour buffers
1592 * Also generates the read & write from colour buffer
1593 */
1594 static void
1595 generate_unswizzled_blend(struct gallivm_state *gallivm,
1596 unsigned rt,
1597 struct lp_fragment_shader_variant *variant,
1598 enum pipe_format out_format,
1599 unsigned int num_fs,
1600 struct lp_type fs_type,
1601 LLVMValueRef* fs_mask,
1602 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1603 LLVMValueRef context_ptr,
1604 LLVMValueRef color_ptr,
1605 LLVMValueRef stride,
1606 unsigned partial_mask,
1607 boolean do_branch)
1608 {
1609 const unsigned alpha_channel = 3;
1610 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1611 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1612 const unsigned block_size = block_width * block_height;
1613 const unsigned lp_integer_vector_width = 128;
1614
1615 LLVMBuilderRef builder = gallivm->builder;
1616 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1617 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1618 LLVMValueRef src_alpha[4 * 4];
1619 LLVMValueRef src1_alpha[4 * 4] = { NULL };
1620 LLVMValueRef src_mask[4 * 4];
1621 LLVMValueRef src[4 * 4];
1622 LLVMValueRef src1[4 * 4];
1623 LLVMValueRef dst[4 * 4];
1624 LLVMValueRef blend_color;
1625 LLVMValueRef blend_alpha;
1626 LLVMValueRef i32_zero;
1627 LLVMValueRef check_mask;
1628 LLVMValueRef undef_src_val;
1629
1630 struct lp_build_mask_context mask_ctx;
1631 struct lp_type mask_type;
1632 struct lp_type blend_type;
1633 struct lp_type row_type;
1634 struct lp_type dst_type;
1635
1636 unsigned char swizzle[TGSI_NUM_CHANNELS];
1637 unsigned vector_width;
1638 unsigned src_channels = TGSI_NUM_CHANNELS;
1639 unsigned dst_channels;
1640 unsigned dst_count;
1641 unsigned src_count;
1642 unsigned i, j;
1643
1644 const struct util_format_description* out_format_desc = util_format_description(out_format);
1645
1646 unsigned dst_alignment;
1647
1648 bool pad_inline = is_arithmetic_format(out_format_desc);
1649 bool has_alpha = false;
1650 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1651 util_blend_state_is_dual(&variant->key.blend, 0);
1652
1653 const boolean is_1d = variant->key.resource_1d;
1654 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
1655 LLVMValueRef fpstate = 0;
1656
1657 /* Get type from output format */
1658 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1659 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1660
1661 /*
1662 * Technically this code should go into lp_build_smallfloat_to_float
1663 * and lp_build_float_to_smallfloat but due to the
1664 * http://llvm.org/bugs/show_bug.cgi?id=6393
1665 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1666 * So the ordering is important here and there shouldn't be any
1667 * llvm ir instrunctions in this function before
1668 * this, otherwise half-float format conversions won't work
1669 * (again due to llvm bug #6393).
1670 */
1671 if (have_smallfloat_format(dst_type, out_format)) {
1672 /* We need to make sure that denorms are ok for half float
1673 conversions */
1674 fpstate = lp_build_fpstate_get(gallivm);
1675 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
1676 }
1677
1678 mask_type = lp_int32_vec4_type();
1679 mask_type.length = fs_type.length;
1680
1681 for (i = num_fs; i < num_fullblock_fs; i++) {
1682 fs_mask[i] = lp_build_zero(gallivm, mask_type);
1683 }
1684
1685 /* Do not bother executing code when mask is empty.. */
1686 if (do_branch) {
1687 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1688
1689 for (i = 0; i < num_fullblock_fs; ++i) {
1690 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1691 }
1692
1693 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1694 lp_build_mask_check(&mask_ctx);
1695 }
1696
1697 partial_mask |= !variant->opaque;
1698 i32_zero = lp_build_const_int32(gallivm, 0);
1699
1700 undef_src_val = lp_build_undef(gallivm, fs_type);
1701
1702 row_type.length = fs_type.length;
1703 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1704
1705 /* Compute correct swizzle and count channels */
1706 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1707 dst_channels = 0;
1708
1709 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1710 /* Ensure channel is used */
1711 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1712 continue;
1713 }
1714
1715 /* Ensure not already written to (happens in case with GL_ALPHA) */
1716 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1717 continue;
1718 }
1719
1720 /* Ensure we havn't already found all channels */
1721 if (dst_channels >= out_format_desc->nr_channels) {
1722 continue;
1723 }
1724
1725 swizzle[out_format_desc->swizzle[i]] = i;
1726 ++dst_channels;
1727
1728 if (i == alpha_channel) {
1729 has_alpha = true;
1730 }
1731 }
1732
1733 if (format_expands_to_float_soa(out_format_desc)) {
1734 /*
1735 * the code above can't work for layout_other
1736 * for srgb it would sort of work but we short-circuit swizzles, etc.
1737 * as that is done as part of unpack / pack.
1738 */
1739 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1740 has_alpha = true;
1741 swizzle[0] = 0;
1742 swizzle[1] = 1;
1743 swizzle[2] = 2;
1744 swizzle[3] = 3;
1745 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1746 }
1747
1748 /* If 3 channels then pad to include alpha for 4 element transpose */
1749 if (dst_channels == 3 && !has_alpha) {
1750 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1751 if (swizzle[i] > TGSI_NUM_CHANNELS)
1752 swizzle[i] = 3;
1753 }
1754 if (out_format_desc->nr_channels == 4) {
1755 dst_channels = 4;
1756 }
1757 }
1758
1759 /*
1760 * Load shader output
1761 */
1762 for (i = 0; i < num_fullblock_fs; ++i) {
1763 /* Always load alpha for use in blending */
1764 LLVMValueRef alpha;
1765 if (i < num_fs) {
1766 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1767 }
1768 else {
1769 alpha = undef_src_val;
1770 }
1771
1772 /* Load each channel */
1773 for (j = 0; j < dst_channels; ++j) {
1774 assert(swizzle[j] < 4);
1775 if (i < num_fs) {
1776 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1777 }
1778 else {
1779 fs_src[i][j] = undef_src_val;
1780 }
1781 }
1782
1783 /* If 3 channels then pad to include alpha for 4 element transpose */
1784 /*
1785 * XXX If we include that here maybe could actually use it instead of
1786 * separate alpha for blending?
1787 */
1788 if (dst_channels == 3 && !has_alpha) {
1789 fs_src[i][3] = alpha;
1790 }
1791
1792 /* We split the row_mask and row_alpha as we want 128bit interleave */
1793 if (fs_type.length == 8) {
1794 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
1795 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
1796
1797 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1798 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1799 } else {
1800 src_mask[i] = fs_mask[i];
1801 src_alpha[i] = alpha;
1802 }
1803 }
1804 if (dual_source_blend) {
1805 /* same as above except different src/dst, skip masks and comments... */
1806 for (i = 0; i < num_fullblock_fs; ++i) {
1807 LLVMValueRef alpha;
1808 if (i < num_fs) {
1809 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1810 }
1811 else {
1812 alpha = undef_src_val;
1813 }
1814
1815 for (j = 0; j < dst_channels; ++j) {
1816 assert(swizzle[j] < 4);
1817 if (i < num_fs) {
1818 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1819 }
1820 else {
1821 fs_src1[i][j] = undef_src_val;
1822 }
1823 }
1824 if (dst_channels == 3 && !has_alpha) {
1825 fs_src1[i][3] = alpha;
1826 }
1827 if (fs_type.length == 8) {
1828 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1829 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
1830 } else {
1831 src1_alpha[i] = alpha;
1832 }
1833 }
1834 }
1835
1836 if (util_format_is_pure_integer(out_format)) {
1837 /*
1838 * In this case fs_type was really ints or uints disguised as floats,
1839 * fix that up now.
1840 */
1841 fs_type.floating = 0;
1842 fs_type.sign = dst_type.sign;
1843 for (i = 0; i < num_fullblock_fs; ++i) {
1844 for (j = 0; j < dst_channels; ++j) {
1845 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1846 lp_build_vec_type(gallivm, fs_type), "");
1847 }
1848 if (dst_channels == 3 && !has_alpha) {
1849 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1850 lp_build_vec_type(gallivm, fs_type), "");
1851 }
1852 }
1853 }
1854
1855 /*
1856 * Pixel twiddle from fragment shader order to memory order
1857 */
1858 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
1859 dst_channels, fs_src, src, pad_inline);
1860 if (dual_source_blend) {
1861 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
1862 fs_src1, src1, pad_inline);
1863 }
1864
1865 src_channels = dst_channels < 3 ? dst_channels : 4;
1866 if (src_count != num_fullblock_fs * src_channels) {
1867 unsigned ds = src_count / (num_fullblock_fs * src_channels);
1868 row_type.length /= ds;
1869 fs_type.length = row_type.length;
1870 }
1871
1872 blend_type = row_type;
1873 mask_type.length = 4;
1874
1875 /* Convert src to row_type */
1876 if (dual_source_blend) {
1877 struct lp_type old_row_type = row_type;
1878 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1879 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
1880 }
1881 else {
1882 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
1883 }
1884
1885 /* If the rows are not an SSE vector, combine them to become SSE size! */
1886 if ((row_type.width * row_type.length) % 128) {
1887 unsigned bits = row_type.width * row_type.length;
1888 unsigned combined;
1889
1890 assert(src_count >= (vector_width / bits));
1891
1892 dst_count = src_count / (vector_width / bits);
1893
1894 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
1895 if (dual_source_blend) {
1896 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
1897 }
1898
1899 row_type.length *= combined;
1900 src_count /= combined;
1901
1902 bits = row_type.width * row_type.length;
1903 assert(bits == 128 || bits == 256);
1904 }
1905
1906
1907 /*
1908 * Blend Colour conversion
1909 */
1910 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
1911 blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
1912 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
1913
1914 /* Convert */
1915 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
1916
1917 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1918 /*
1919 * since blending is done with floats, there was no conversion.
1920 * However, the rules according to fixed point renderbuffers still
1921 * apply, that is we must clamp inputs to 0.0/1.0.
1922 * (This would apply to separate alpha conversion too but we currently
1923 * force has_alpha to be true.)
1924 * TODO: should skip this with "fake" blend, since post-blend conversion
1925 * will clamp anyway.
1926 * TODO: could also skip this if fragment color clamping is enabled. We
1927 * don't support it natively so it gets baked into the shader however, so
1928 * can't really tell here.
1929 */
1930 struct lp_build_context f32_bld;
1931 assert(row_type.floating);
1932 lp_build_context_init(&f32_bld, gallivm, row_type);
1933 for (i = 0; i < src_count; i++) {
1934 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
1935 }
1936 if (dual_source_blend) {
1937 for (i = 0; i < src_count; i++) {
1938 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
1939 }
1940 }
1941 /* probably can't be different than row_type but better safe than sorry... */
1942 lp_build_context_init(&f32_bld, gallivm, blend_type);
1943 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
1944 }
1945
1946 /* Extract alpha */
1947 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
1948
1949 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1950 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
1951 if (pad_inline) {
1952 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1953 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
1954 } else {
1955 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1956 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
1957 }
1958
1959 /*
1960 * Mask conversion
1961 */
1962 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
1963
1964 if (src_count < block_height) {
1965 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
1966 } else if (src_count > block_height) {
1967 for (i = src_count; i > 0; --i) {
1968 unsigned pixels = block_size / src_count;
1969 unsigned idx = i - 1;
1970
1971 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
1972 (idx * pixels) % 4, pixels);
1973 }
1974 }
1975
1976 assert(mask_type.width == 32);
1977
1978 for (i = 0; i < src_count; ++i) {
1979 unsigned pixels = block_size / src_count;
1980 unsigned pixel_width = row_type.width * dst_channels;
1981
1982 if (pixel_width == 24) {
1983 mask_type.width = 8;
1984 mask_type.length = vector_width / mask_type.width;
1985 } else {
1986 mask_type.length = pixels;
1987 mask_type.width = row_type.width * dst_channels;
1988
1989 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1990
1991 mask_type.length *= dst_channels;
1992 mask_type.width /= dst_channels;
1993 }
1994
1995 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
1996 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
1997 }
1998
1999 /*
2000 * Alpha conversion
2001 */
2002 if (!has_alpha) {
2003 struct lp_type alpha_type = fs_type;
2004 alpha_type.length = 4;
2005 convert_alpha(gallivm, row_type, alpha_type,
2006 block_size, block_height,
2007 src_count, dst_channels,
2008 pad_inline, src_alpha);
2009 if (dual_source_blend) {
2010 convert_alpha(gallivm, row_type, alpha_type,
2011 block_size, block_height,
2012 src_count, dst_channels,
2013 pad_inline, src1_alpha);
2014 }
2015 }
2016
2017
2018 /*
2019 * Load dst from memory
2020 */
2021 if (src_count < block_height) {
2022 dst_count = block_height;
2023 } else {
2024 dst_count = src_count;
2025 }
2026
2027 dst_type.length *= block_size / dst_count;
2028
2029 if (format_expands_to_float_soa(out_format_desc)) {
2030 /*
2031 * we need multiple values at once for the conversion, so can as well
2032 * load them vectorized here too instead of concatenating later.
2033 * (Still need concatenation later for 8-wide vectors).
2034 */
2035 dst_count = block_height;
2036 dst_type.length = block_width;
2037 }
2038
2039 /*
2040 * Compute the alignment of the destination pointer in bytes
2041 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2042 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2043 * 1d tex but can't distinguish here) so need to stick with per-pixel
2044 * alignment in this case.
2045 */
2046 if (is_1d) {
2047 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2048 }
2049 else {
2050 dst_alignment = dst_type.length * dst_type.width / 8;
2051 }
2052 /* Force power-of-two alignment by extracting only the least-significant-bit */
2053 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2054 /*
2055 * Resource base and stride pointers are aligned to 16 bytes, so that's
2056 * the maximum alignment we can guarantee
2057 */
2058 dst_alignment = MIN2(16, dst_alignment);
2059
2060 if (is_1d) {
2061 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2062 dst, dst_type, dst_count / 4, dst_alignment);
2063 for (i = dst_count / 4; i < dst_count; i++) {
2064 dst[i] = lp_build_undef(gallivm, dst_type);
2065 }
2066
2067 }
2068 else {
2069 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2070 dst, dst_type, dst_count, dst_alignment);
2071 }
2072
2073
2074 /*
2075 * Convert from dst/output format to src/blending format.
2076 *
2077 * This is necessary as we can only read 1 row from memory at a time,
2078 * so the minimum dst_count will ever be at this point is 4.
2079 *
2080 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2081 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2082 * on all 16 pixels in that single vector at once.
2083 */
2084 if (dst_count > src_count) {
2085 lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
2086 }
2087
2088 /*
2089 * Blending
2090 */
2091 /* XXX this is broken for RGB8 formats -
2092 * they get expanded from 12 to 16 elements (to include alpha)
2093 * by convert_to_blend_type then reduced to 15 instead of 12
2094 * by convert_from_blend_type (a simple fix though breaks A8...).
2095 * R16G16B16 also crashes differently however something going wrong
2096 * inside llvm handling npot vector sizes seemingly.
2097 * It seems some cleanup could be done here (like skipping conversion/blend
2098 * when not needed).
2099 */
2100 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
2101
2102 /*
2103 * FIXME: Really should get logic ops / masks out of generic blend / row
2104 * format. Logic ops will definitely not work on the blend float format
2105 * used for SRGB here and I think OpenGL expects this to work as expected
2106 * (that is incoming values converted to srgb then logic op applied).
2107 */
2108 for (i = 0; i < src_count; ++i) {
2109 dst[i] = lp_build_blend_aos(gallivm,
2110 &variant->key.blend,
2111 out_format,
2112 row_type,
2113 rt,
2114 src[i],
2115 has_alpha ? NULL : src_alpha[i],
2116 src1[i],
2117 has_alpha ? NULL : src1_alpha[i],
2118 dst[i],
2119 partial_mask ? src_mask[i] : NULL,
2120 blend_color,
2121 has_alpha ? NULL : blend_alpha,
2122 swizzle,
2123 pad_inline ? 4 : dst_channels);
2124 }
2125
2126 convert_from_blend_type(gallivm, block_size, out_format_desc, row_type, dst_type, dst, src_count);
2127
2128 /* Split the blend rows back to memory rows */
2129 if (dst_count > src_count) {
2130 row_type.length = dst_type.length * (dst_count / src_count);
2131
2132 if (src_count == 1) {
2133 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2134 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2135
2136 row_type.length /= 2;
2137 src_count *= 2;
2138 }
2139
2140 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2141 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2142 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2143 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2144
2145 row_type.length /= 2;
2146 src_count *= 2;
2147 }
2148
2149 /*
2150 * Store blend result to memory
2151 */
2152 if (is_1d) {
2153 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2154 dst, dst_type, dst_count / 4, dst_alignment);
2155 }
2156 else {
2157 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2158 dst, dst_type, dst_count, dst_alignment);
2159 }
2160
2161 if (have_smallfloat_format(dst_type, out_format)) {
2162 lp_build_fpstate_set(gallivm, fpstate);
2163 }
2164
2165 if (do_branch) {
2166 lp_build_mask_end(&mask_ctx);
2167 }
2168 }
2169
2170
2171 /**
2172 * Generate the runtime callable function for the whole fragment pipeline.
2173 * Note that the function which we generate operates on a block of 16
2174 * pixels at at time. The block contains 2x2 quads. Each quad contains
2175 * 2x2 pixels.
2176 */
2177 static void
2178 generate_fragment(struct llvmpipe_context *lp,
2179 struct lp_fragment_shader *shader,
2180 struct lp_fragment_shader_variant *variant,
2181 unsigned partial_mask)
2182 {
2183 struct gallivm_state *gallivm = variant->gallivm;
2184 const struct lp_fragment_shader_variant_key *key = &variant->key;
2185 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2186 char func_name[64];
2187 struct lp_type fs_type;
2188 struct lp_type blend_type;
2189 LLVMTypeRef fs_elem_type;
2190 LLVMTypeRef blend_vec_type;
2191 LLVMTypeRef arg_types[13];
2192 LLVMTypeRef func_type;
2193 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2194 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2195 LLVMValueRef context_ptr;
2196 LLVMValueRef x;
2197 LLVMValueRef y;
2198 LLVMValueRef a0_ptr;
2199 LLVMValueRef dadx_ptr;
2200 LLVMValueRef dady_ptr;
2201 LLVMValueRef color_ptr_ptr;
2202 LLVMValueRef stride_ptr;
2203 LLVMValueRef depth_ptr;
2204 LLVMValueRef depth_stride;
2205 LLVMValueRef mask_input;
2206 LLVMValueRef thread_data_ptr;
2207 LLVMBasicBlockRef block;
2208 LLVMBuilderRef builder;
2209 struct lp_build_sampler_soa *sampler;
2210 struct lp_build_interp_soa_context interp;
2211 LLVMValueRef fs_mask[16 / 4];
2212 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2213 LLVMValueRef function;
2214 LLVMValueRef facing;
2215 unsigned num_fs;
2216 unsigned i;
2217 unsigned chan;
2218 unsigned cbuf;
2219 boolean cbuf0_write_all;
2220 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2221 util_blend_state_is_dual(&key->blend, 0);
2222
2223 assert(lp_native_vector_width / 32 >= 4);
2224
2225 /* Adjust color input interpolation according to flatshade state:
2226 */
2227 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2228 for (i = 0; i < shader->info.base.num_inputs; i++) {
2229 if (inputs[i].interp == LP_INTERP_COLOR) {
2230 if (key->flatshade)
2231 inputs[i].interp = LP_INTERP_CONSTANT;
2232 else
2233 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2234 }
2235 }
2236
2237 /* check if writes to cbuf[0] are to be copied to all cbufs */
2238 cbuf0_write_all =
2239 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2240
2241 /* TODO: actually pick these based on the fs and color buffer
2242 * characteristics. */
2243
2244 memset(&fs_type, 0, sizeof fs_type);
2245 fs_type.floating = TRUE; /* floating point values */
2246 fs_type.sign = TRUE; /* values are signed */
2247 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2248 fs_type.width = 32; /* 32-bit float */
2249 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2250
2251 memset(&blend_type, 0, sizeof blend_type);
2252 blend_type.floating = FALSE; /* values are integers */
2253 blend_type.sign = FALSE; /* values are unsigned */
2254 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2255 blend_type.width = 8; /* 8-bit ubyte values */
2256 blend_type.length = 16; /* 16 elements per vector */
2257
2258 /*
2259 * Generate the function prototype. Any change here must be reflected in
2260 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2261 */
2262
2263 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2264
2265 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2266
2267 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2268 shader->no, variant->no, partial_mask ? "partial" : "whole");
2269
2270 arg_types[0] = variant->jit_context_ptr_type; /* context */
2271 arg_types[1] = int32_type; /* x */
2272 arg_types[2] = int32_type; /* y */
2273 arg_types[3] = int32_type; /* facing */
2274 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2275 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2276 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2277 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
2278 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2279 arg_types[9] = int32_type; /* mask_input */
2280 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2281 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2282 arg_types[12] = int32_type; /* depth_stride */
2283
2284 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2285 arg_types, ARRAY_SIZE(arg_types), 0);
2286
2287 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2288 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2289
2290 variant->function[partial_mask] = function;
2291
2292 /* XXX: need to propagate noalias down into color param now we are
2293 * passing a pointer-to-pointer?
2294 */
2295 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2296 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2297 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2298
2299 context_ptr = LLVMGetParam(function, 0);
2300 x = LLVMGetParam(function, 1);
2301 y = LLVMGetParam(function, 2);
2302 facing = LLVMGetParam(function, 3);
2303 a0_ptr = LLVMGetParam(function, 4);
2304 dadx_ptr = LLVMGetParam(function, 5);
2305 dady_ptr = LLVMGetParam(function, 6);
2306 color_ptr_ptr = LLVMGetParam(function, 7);
2307 depth_ptr = LLVMGetParam(function, 8);
2308 mask_input = LLVMGetParam(function, 9);
2309 thread_data_ptr = LLVMGetParam(function, 10);
2310 stride_ptr = LLVMGetParam(function, 11);
2311 depth_stride = LLVMGetParam(function, 12);
2312
2313 lp_build_name(context_ptr, "context");
2314 lp_build_name(x, "x");
2315 lp_build_name(y, "y");
2316 lp_build_name(a0_ptr, "a0");
2317 lp_build_name(dadx_ptr, "dadx");
2318 lp_build_name(dady_ptr, "dady");
2319 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2320 lp_build_name(depth_ptr, "depth");
2321 lp_build_name(mask_input, "mask_input");
2322 lp_build_name(thread_data_ptr, "thread_data");
2323 lp_build_name(stride_ptr, "stride_ptr");
2324 lp_build_name(depth_stride, "depth_stride");
2325
2326 /*
2327 * Function body
2328 */
2329
2330 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2331 builder = gallivm->builder;
2332 assert(builder);
2333 LLVMPositionBuilderAtEnd(builder, block);
2334
2335 /* code generated texture sampling */
2336 sampler = lp_llvm_sampler_soa_create(key->state);
2337
2338 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2339 /* for 1d resources only run "upper half" of stamp */
2340 if (key->resource_1d)
2341 num_fs /= 2;
2342
2343 {
2344 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2345 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2346 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2347 num_loop, "mask_store");
2348 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2349 boolean pixel_center_integer =
2350 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2351
2352 /*
2353 * The shader input interpolation info is not explicitely baked in the
2354 * shader key, but everything it derives from (TGSI, and flatshade) is
2355 * already included in the shader key.
2356 */
2357 lp_build_interp_soa_init(&interp,
2358 gallivm,
2359 shader->info.base.num_inputs,
2360 inputs,
2361 pixel_center_integer,
2362 key->depth_clamp,
2363 builder, fs_type,
2364 a0_ptr, dadx_ptr, dady_ptr,
2365 x, y);
2366
2367 for (i = 0; i < num_fs; i++) {
2368 LLVMValueRef mask;
2369 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2370 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2371 &indexi, 1, "mask_ptr");
2372
2373 if (partial_mask) {
2374 mask = generate_quad_mask(gallivm, fs_type,
2375 i*fs_type.length/4, mask_input);
2376 }
2377 else {
2378 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2379 }
2380 LLVMBuildStore(builder, mask, mask_ptr);
2381 }
2382
2383 generate_fs_loop(gallivm,
2384 shader, key,
2385 builder,
2386 fs_type,
2387 context_ptr,
2388 num_loop,
2389 &interp,
2390 sampler,
2391 mask_store, /* output */
2392 color_store,
2393 depth_ptr,
2394 depth_stride,
2395 facing,
2396 thread_data_ptr);
2397
2398 for (i = 0; i < num_fs; i++) {
2399 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2400 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2401 &indexi, 1, "");
2402 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2403 /* This is fucked up need to reorganize things */
2404 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2405 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2406 ptr = LLVMBuildGEP(builder,
2407 color_store[cbuf * !cbuf0_write_all][chan],
2408 &indexi, 1, "");
2409 fs_out_color[cbuf][chan][i] = ptr;
2410 }
2411 }
2412 if (dual_source_blend) {
2413 /* only support one dual source blend target hence always use output 1 */
2414 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2415 ptr = LLVMBuildGEP(builder,
2416 color_store[1][chan],
2417 &indexi, 1, "");
2418 fs_out_color[1][chan][i] = ptr;
2419 }
2420 }
2421 }
2422 }
2423
2424 sampler->destroy(sampler);
2425
2426 /* Loop over color outputs / color buffers to do blending.
2427 */
2428 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2429 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
2430 LLVMValueRef color_ptr;
2431 LLVMValueRef stride;
2432 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2433
2434 boolean do_branch = ((key->depth.enabled
2435 || key->stencil[0].enabled
2436 || key->alpha.enabled)
2437 && !shader->info.base.uses_kill);
2438
2439 color_ptr = LLVMBuildLoad(builder,
2440 LLVMBuildGEP(builder, color_ptr_ptr,
2441 &index, 1, ""),
2442 "");
2443
2444 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2445
2446 stride = LLVMBuildLoad(builder,
2447 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2448 "");
2449
2450 generate_unswizzled_blend(gallivm, cbuf, variant,
2451 key->cbuf_format[cbuf],
2452 num_fs, fs_type, fs_mask, fs_out_color,
2453 context_ptr, color_ptr, stride,
2454 partial_mask, do_branch);
2455 }
2456 }
2457
2458 LLVMBuildRetVoid(builder);
2459
2460 gallivm_verify_function(gallivm, function);
2461 }
2462
2463
2464 static void
2465 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2466 {
2467 unsigned i;
2468
2469 debug_printf("fs variant %p:\n", (void *) key);
2470
2471 if (key->flatshade) {
2472 debug_printf("flatshade = 1\n");
2473 }
2474 for (i = 0; i < key->nr_cbufs; ++i) {
2475 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2476 }
2477 if (key->depth.enabled || key->stencil[0].enabled) {
2478 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2479 }
2480 if (key->depth.enabled) {
2481 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2482 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2483 }
2484
2485 for (i = 0; i < 2; ++i) {
2486 if (key->stencil[i].enabled) {
2487 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2488 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2489 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2490 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2491 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2492 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2493 }
2494 }
2495
2496 if (key->alpha.enabled) {
2497 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2498 }
2499
2500 if (key->occlusion_count) {
2501 debug_printf("occlusion_count = 1\n");
2502 }
2503
2504 if (key->blend.logicop_enable) {
2505 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2506 }
2507 else if (key->blend.rt[0].blend_enable) {
2508 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2509 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2510 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2511 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2512 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2513 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2514 }
2515 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2516 if (key->blend.alpha_to_coverage) {
2517 debug_printf("blend.alpha_to_coverage is enabled\n");
2518 }
2519 for (i = 0; i < key->nr_samplers; ++i) {
2520 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2521 debug_printf("sampler[%u] = \n", i);
2522 debug_printf(" .wrap = %s %s %s\n",
2523 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2524 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2525 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2526 debug_printf(" .min_img_filter = %s\n",
2527 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2528 debug_printf(" .min_mip_filter = %s\n",
2529 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2530 debug_printf(" .mag_img_filter = %s\n",
2531 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2532 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2533 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2534 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2535 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2536 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2537 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2538 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2539 }
2540 for (i = 0; i < key->nr_sampler_views; ++i) {
2541 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2542 debug_printf("texture[%u] = \n", i);
2543 debug_printf(" .format = %s\n",
2544 util_format_name(texture->format));
2545 debug_printf(" .target = %s\n",
2546 util_dump_tex_target(texture->target, TRUE));
2547 debug_printf(" .level_zero_only = %u\n",
2548 texture->level_zero_only);
2549 debug_printf(" .pot = %u %u %u\n",
2550 texture->pot_width,
2551 texture->pot_height,
2552 texture->pot_depth);
2553 }
2554 }
2555
2556
2557 void
2558 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2559 {
2560 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2561 variant->shader->no, variant->no);
2562 tgsi_dump(variant->shader->base.tokens, 0);
2563 dump_fs_variant_key(&variant->key);
2564 debug_printf("variant->opaque = %u\n", variant->opaque);
2565 debug_printf("\n");
2566 }
2567
2568
2569 /**
2570 * Generate a new fragment shader variant from the shader code and
2571 * other state indicated by the key.
2572 */
2573 static struct lp_fragment_shader_variant *
2574 generate_variant(struct llvmpipe_context *lp,
2575 struct lp_fragment_shader *shader,
2576 const struct lp_fragment_shader_variant_key *key)
2577 {
2578 struct lp_fragment_shader_variant *variant;
2579 const struct util_format_description *cbuf0_format_desc;
2580 boolean fullcolormask;
2581 char module_name[64];
2582
2583 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2584 if (!variant)
2585 return NULL;
2586
2587 util_snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
2588 shader->no, shader->variants_created);
2589
2590 variant->gallivm = gallivm_create(module_name, lp->context);
2591 if (!variant->gallivm) {
2592 FREE(variant);
2593 return NULL;
2594 }
2595
2596 variant->shader = shader;
2597 variant->list_item_global.base = variant;
2598 variant->list_item_local.base = variant;
2599 variant->no = shader->variants_created++;
2600
2601 memcpy(&variant->key, key, shader->variant_key_size);
2602
2603 /*
2604 * Determine whether we are touching all channels in the color buffer.
2605 */
2606 fullcolormask = FALSE;
2607 if (key->nr_cbufs == 1) {
2608 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2609 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2610 }
2611
2612 variant->opaque =
2613 !key->blend.logicop_enable &&
2614 !key->blend.rt[0].blend_enable &&
2615 fullcolormask &&
2616 !key->stencil[0].enabled &&
2617 !key->alpha.enabled &&
2618 !key->blend.alpha_to_coverage &&
2619 !key->depth.enabled &&
2620 !shader->info.base.uses_kill
2621 ? TRUE : FALSE;
2622
2623 if ((shader->info.base.num_tokens <= 1) &&
2624 !key->depth.enabled && !key->stencil[0].enabled) {
2625 variant->ps_inv_multiplier = 0;
2626 } else {
2627 variant->ps_inv_multiplier = 1;
2628 }
2629
2630 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2631 lp_debug_fs_variant(variant);
2632 }
2633
2634 lp_jit_init_types(variant);
2635
2636 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2637 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2638
2639 if (variant->jit_function[RAST_WHOLE] == NULL) {
2640 if (variant->opaque) {
2641 /* Specialized shader, which doesn't need to read the color buffer. */
2642 generate_fragment(lp, shader, variant, RAST_WHOLE);
2643 }
2644 }
2645
2646 /*
2647 * Compile everything
2648 */
2649
2650 gallivm_compile_module(variant->gallivm);
2651
2652 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
2653
2654 if (variant->function[RAST_EDGE_TEST]) {
2655 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2656 gallivm_jit_function(variant->gallivm,
2657 variant->function[RAST_EDGE_TEST]);
2658 }
2659
2660 if (variant->function[RAST_WHOLE]) {
2661 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2662 gallivm_jit_function(variant->gallivm,
2663 variant->function[RAST_WHOLE]);
2664 } else if (!variant->jit_function[RAST_WHOLE]) {
2665 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2666 }
2667
2668 gallivm_free_ir(variant->gallivm);
2669
2670 return variant;
2671 }
2672
2673
2674 static void *
2675 llvmpipe_create_fs_state(struct pipe_context *pipe,
2676 const struct pipe_shader_state *templ)
2677 {
2678 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2679 struct lp_fragment_shader *shader;
2680 int nr_samplers;
2681 int nr_sampler_views;
2682 int i;
2683
2684 shader = CALLOC_STRUCT(lp_fragment_shader);
2685 if (!shader)
2686 return NULL;
2687
2688 shader->no = fs_no++;
2689 make_empty_list(&shader->variants);
2690
2691 /* get/save the summary info for this shader */
2692 lp_build_tgsi_info(templ->tokens, &shader->info);
2693
2694 /* we need to keep a local copy of the tokens */
2695 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2696
2697 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2698 if (shader->draw_data == NULL) {
2699 FREE((void *) shader->base.tokens);
2700 FREE(shader);
2701 return NULL;
2702 }
2703
2704 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2705 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2706
2707 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2708 state[MAX2(nr_samplers, nr_sampler_views)]);
2709
2710 for (i = 0; i < shader->info.base.num_inputs; i++) {
2711 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2712 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2713
2714 switch (shader->info.base.input_interpolate[i]) {
2715 case TGSI_INTERPOLATE_CONSTANT:
2716 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2717 break;
2718 case TGSI_INTERPOLATE_LINEAR:
2719 shader->inputs[i].interp = LP_INTERP_LINEAR;
2720 break;
2721 case TGSI_INTERPOLATE_PERSPECTIVE:
2722 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2723 break;
2724 case TGSI_INTERPOLATE_COLOR:
2725 shader->inputs[i].interp = LP_INTERP_COLOR;
2726 break;
2727 default:
2728 assert(0);
2729 break;
2730 }
2731
2732 switch (shader->info.base.input_semantic_name[i]) {
2733 case TGSI_SEMANTIC_FACE:
2734 shader->inputs[i].interp = LP_INTERP_FACING;
2735 break;
2736 case TGSI_SEMANTIC_POSITION:
2737 /* Position was already emitted above
2738 */
2739 shader->inputs[i].interp = LP_INTERP_POSITION;
2740 shader->inputs[i].src_index = 0;
2741 continue;
2742 }
2743
2744 /* XXX this is a completely pointless index map... */
2745 shader->inputs[i].src_index = i+1;
2746 }
2747
2748 if (LP_DEBUG & DEBUG_TGSI) {
2749 unsigned attrib;
2750 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2751 shader->no, (void *) shader);
2752 tgsi_dump(templ->tokens, 0);
2753 debug_printf("usage masks:\n");
2754 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2755 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2756 debug_printf(" IN[%u].%s%s%s%s\n",
2757 attrib,
2758 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2759 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2760 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2761 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2762 }
2763 debug_printf("\n");
2764 }
2765
2766 return shader;
2767 }
2768
2769
2770 static void
2771 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2772 {
2773 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2774
2775 if (llvmpipe->fs == fs)
2776 return;
2777
2778 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2779
2780 draw_bind_fragment_shader(llvmpipe->draw,
2781 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2782
2783 llvmpipe->dirty |= LP_NEW_FS;
2784 }
2785
2786
2787 /**
2788 * Remove shader variant from two lists: the shader's variant list
2789 * and the context's variant list.
2790 */
2791 void
2792 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2793 struct lp_fragment_shader_variant *variant)
2794 {
2795 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2796 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2797 " #%u v total cached #%u\n",
2798 variant->shader->no,
2799 variant->no,
2800 variant->shader->variants_created,
2801 variant->shader->variants_cached,
2802 lp->nr_fs_variants);
2803 }
2804
2805 gallivm_destroy(variant->gallivm);
2806
2807 /* remove from shader's list */
2808 remove_from_list(&variant->list_item_local);
2809 variant->shader->variants_cached--;
2810
2811 /* remove from context's list */
2812 remove_from_list(&variant->list_item_global);
2813 lp->nr_fs_variants--;
2814 lp->nr_fs_instrs -= variant->nr_instrs;
2815
2816 FREE(variant);
2817 }
2818
2819
2820 static void
2821 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
2822 {
2823 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2824 struct lp_fragment_shader *shader = fs;
2825 struct lp_fs_variant_list_item *li;
2826
2827 assert(fs != llvmpipe->fs);
2828
2829 /*
2830 * XXX: we need to flush the context until we have some sort of reference
2831 * counting in fragment shaders as they may still be binned
2832 * Flushing alone might not sufficient we need to wait on it too.
2833 */
2834 llvmpipe_finish(pipe, __FUNCTION__);
2835
2836 /* Delete all the variants */
2837 li = first_elem(&shader->variants);
2838 while(!at_end(&shader->variants, li)) {
2839 struct lp_fs_variant_list_item *next = next_elem(li);
2840 llvmpipe_remove_shader_variant(llvmpipe, li->base);
2841 li = next;
2842 }
2843
2844 /* Delete draw module's data */
2845 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
2846
2847 assert(shader->variants_cached == 0);
2848 FREE((void *) shader->base.tokens);
2849 FREE(shader);
2850 }
2851
2852
2853
2854 static void
2855 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
2856 uint shader, uint index,
2857 const struct pipe_constant_buffer *cb)
2858 {
2859 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2860 struct pipe_resource *constants = cb ? cb->buffer : NULL;
2861
2862 assert(shader < PIPE_SHADER_TYPES);
2863 assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
2864
2865 /* note: reference counting */
2866 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
2867
2868 if (constants) {
2869 if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
2870 debug_printf("Illegal set constant without bind flag\n");
2871 constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
2872 }
2873 }
2874
2875 if (shader == PIPE_SHADER_VERTEX ||
2876 shader == PIPE_SHADER_GEOMETRY) {
2877 /* Pass the constants to the 'draw' module */
2878 const unsigned size = cb ? cb->buffer_size : 0;
2879 const ubyte *data;
2880
2881 if (constants) {
2882 data = (ubyte *) llvmpipe_resource_data(constants);
2883 }
2884 else if (cb && cb->user_buffer) {
2885 data = (ubyte *) cb->user_buffer;
2886 }
2887 else {
2888 data = NULL;
2889 }
2890
2891 if (data)
2892 data += cb->buffer_offset;
2893
2894 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
2895 index, data, size);
2896 }
2897 else {
2898 llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
2899 }
2900
2901 if (cb && cb->user_buffer) {
2902 pipe_resource_reference(&constants, NULL);
2903 }
2904 }
2905
2906
2907 /**
2908 * Return the blend factor equivalent to a destination alpha of one.
2909 */
2910 static inline unsigned
2911 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
2912 {
2913 switch(factor) {
2914 case PIPE_BLENDFACTOR_DST_ALPHA:
2915 return PIPE_BLENDFACTOR_ONE;
2916 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
2917 return PIPE_BLENDFACTOR_ZERO;
2918 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
2919 if (clamped_zero)
2920 return PIPE_BLENDFACTOR_ZERO;
2921 else
2922 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
2923 }
2924
2925 return factor;
2926 }
2927
2928
2929 /**
2930 * We need to generate several variants of the fragment pipeline to match
2931 * all the combinations of the contributing state atoms.
2932 *
2933 * TODO: there is actually no reason to tie this to context state -- the
2934 * generated code could be cached globally in the screen.
2935 */
2936 static void
2937 make_variant_key(struct llvmpipe_context *lp,
2938 struct lp_fragment_shader *shader,
2939 struct lp_fragment_shader_variant_key *key)
2940 {
2941 unsigned i;
2942
2943 memset(key, 0, shader->variant_key_size);
2944
2945 if (lp->framebuffer.zsbuf) {
2946 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
2947 const struct util_format_description *zsbuf_desc =
2948 util_format_description(zsbuf_format);
2949
2950 if (lp->depth_stencil->depth.enabled &&
2951 util_format_has_depth(zsbuf_desc)) {
2952 key->zsbuf_format = zsbuf_format;
2953 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
2954 }
2955 if (lp->depth_stencil->stencil[0].enabled &&
2956 util_format_has_stencil(zsbuf_desc)) {
2957 key->zsbuf_format = zsbuf_format;
2958 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
2959 }
2960 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
2961 key->resource_1d = TRUE;
2962 }
2963 }
2964
2965 /*
2966 * Propagate the depth clamp setting from the rasterizer state.
2967 * depth_clip == 0 implies depth clamping is enabled.
2968 *
2969 * When clip_halfz is enabled, then always clamp the depth values.
2970 *
2971 * XXX: This is incorrect for GL, but correct for d3d10 (depth
2972 * clamp is always active in d3d10, regardless if depth clip is
2973 * enabled or not).
2974 * (GL has an always-on [0,1] clamp on fs depth output instead
2975 * to ensure the depth values stay in range. Doesn't look like
2976 * we do that, though...)
2977 */
2978 if (lp->rasterizer->clip_halfz) {
2979 key->depth_clamp = 1;
2980 } else {
2981 key->depth_clamp = (lp->rasterizer->depth_clip == 0) ? 1 : 0;
2982 }
2983
2984 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2985 if (!lp->framebuffer.nr_cbufs ||
2986 !lp->framebuffer.cbufs[0] ||
2987 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
2988 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
2989 }
2990 if(key->alpha.enabled)
2991 key->alpha.func = lp->depth_stencil->alpha.func;
2992 /* alpha.ref_value is passed in jit_context */
2993
2994 key->flatshade = lp->rasterizer->flatshade;
2995 if (lp->active_occlusion_queries) {
2996 key->occlusion_count = TRUE;
2997 }
2998
2999 if (lp->framebuffer.nr_cbufs) {
3000 memcpy(&key->blend, lp->blend, sizeof key->blend);
3001 }
3002
3003 key->nr_cbufs = lp->framebuffer.nr_cbufs;
3004
3005 if (!key->blend.independent_blend_enable) {
3006 /* we always need independent blend otherwise the fixups below won't work */
3007 for (i = 1; i < key->nr_cbufs; i++) {
3008 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
3009 }
3010 key->blend.independent_blend_enable = 1;
3011 }
3012
3013 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
3014 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
3015
3016 if (lp->framebuffer.cbufs[i]) {
3017 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
3018 const struct util_format_description *format_desc;
3019
3020 key->cbuf_format[i] = format;
3021
3022 /*
3023 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3024 * mixing of 2d textures with height 1 and 1d textures, so make sure
3025 * we pick 1d if any cbuf or zsbuf is 1d.
3026 */
3027 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
3028 key->resource_1d = TRUE;
3029 }
3030
3031 format_desc = util_format_description(format);
3032 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
3033 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
3034
3035 /*
3036 * Mask out color channels not present in the color buffer.
3037 */
3038 blend_rt->colormask &= util_format_colormask(format_desc);
3039
3040 /*
3041 * Disable blend for integer formats.
3042 */
3043 if (util_format_is_pure_integer(format)) {
3044 blend_rt->blend_enable = 0;
3045 }
3046
3047 /*
3048 * Our swizzled render tiles always have an alpha channel, but the
3049 * linear render target format often does not, so force here the dst
3050 * alpha to be one.
3051 *
3052 * This is not a mere optimization. Wrong results will be produced if
3053 * the dst alpha is used, the dst format does not have alpha, and the
3054 * previous rendering was not flushed from the swizzled to linear
3055 * buffer. For example, NonPowTwo DCT.
3056 *
3057 * TODO: This should be generalized to all channels for better
3058 * performance, but only alpha causes correctness issues.
3059 *
3060 * Also, force rgb/alpha func/factors match, to make AoS blending
3061 * easier.
3062 */
3063 if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
3064 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3065 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3066 boolean clamped_zero = !util_format_is_float(format) &&
3067 !util_format_is_snorm(format);
3068 blend_rt->rgb_src_factor =
3069 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3070 blend_rt->rgb_dst_factor =
3071 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3072 blend_rt->alpha_func = blend_rt->rgb_func;
3073 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3074 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3075 }
3076 }
3077 else {
3078 /* no color buffer for this fragment output */
3079 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3080 blend_rt->colormask = 0x0;
3081 blend_rt->blend_enable = 0;
3082 }
3083 }
3084
3085 /* This value will be the same for all the variants of a given shader:
3086 */
3087 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3088
3089 for(i = 0; i < key->nr_samplers; ++i) {
3090 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3091 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
3092 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3093 }
3094 }
3095
3096 /*
3097 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3098 * are dx10-style? Can't really have mixed opcodes, at least not
3099 * if we want to skip the holes here (without rescanning tgsi).
3100 */
3101 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3102 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3103 for(i = 0; i < key->nr_sampler_views; ++i) {
3104 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
3105 lp_sampler_static_texture_state(&key->state[i].texture_state,
3106 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3107 }
3108 }
3109 }
3110 else {
3111 key->nr_sampler_views = key->nr_samplers;
3112 for(i = 0; i < key->nr_sampler_views; ++i) {
3113 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3114 lp_sampler_static_texture_state(&key->state[i].texture_state,
3115 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3116 }
3117 }
3118 }
3119 }
3120
3121
3122
3123 /**
3124 * Update fragment shader state. This is called just prior to drawing
3125 * something when some fragment-related state has changed.
3126 */
3127 void
3128 llvmpipe_update_fs(struct llvmpipe_context *lp)
3129 {
3130 struct lp_fragment_shader *shader = lp->fs;
3131 struct lp_fragment_shader_variant_key key;
3132 struct lp_fragment_shader_variant *variant = NULL;
3133 struct lp_fs_variant_list_item *li;
3134
3135 make_variant_key(lp, shader, &key);
3136
3137 /* Search the variants for one which matches the key */
3138 li = first_elem(&shader->variants);
3139 while(!at_end(&shader->variants, li)) {
3140 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
3141 variant = li->base;
3142 break;
3143 }
3144 li = next_elem(li);
3145 }
3146
3147 if (variant) {
3148 /* Move this variant to the head of the list to implement LRU
3149 * deletion of shader's when we have too many.
3150 */
3151 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3152 }
3153 else {
3154 /* variant not found, create it now */
3155 int64_t t0, t1, dt;
3156 unsigned i;
3157 unsigned variants_to_cull;
3158
3159 if (0) {
3160 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3161 lp->nr_fs_variants,
3162 lp->nr_fs_instrs,
3163 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3164 }
3165
3166 /* First, check if we've exceeded the max number of shader variants.
3167 * If so, free 25% of them (the least recently used ones).
3168 */
3169 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
3170
3171 if (variants_to_cull ||
3172 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3173 struct pipe_context *pipe = &lp->pipe;
3174
3175 /*
3176 * XXX: we need to flush the context until we have some sort of
3177 * reference counting in fragment shaders as they may still be binned
3178 * Flushing alone might not be sufficient we need to wait on it too.
3179 */
3180 llvmpipe_finish(pipe, __FUNCTION__);
3181
3182 /*
3183 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3184 * number of shader variants (potentially all of them) could be
3185 * pending for destruction on flush.
3186 */
3187
3188 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3189 struct lp_fs_variant_list_item *item;
3190 if (is_empty_list(&lp->fs_variants_list)) {
3191 break;
3192 }
3193 item = last_elem(&lp->fs_variants_list);
3194 assert(item);
3195 assert(item->base);
3196 llvmpipe_remove_shader_variant(lp, item->base);
3197 }
3198 }
3199
3200 /*
3201 * Generate the new variant.
3202 */
3203 t0 = os_time_get();
3204 variant = generate_variant(lp, shader, &key);
3205 t1 = os_time_get();
3206 dt = t1 - t0;
3207 LP_COUNT_ADD(llvm_compile_time, dt);
3208 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3209
3210 /* Put the new variant into the list */
3211 if (variant) {
3212 insert_at_head(&shader->variants, &variant->list_item_local);
3213 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3214 lp->nr_fs_variants++;
3215 lp->nr_fs_instrs += variant->nr_instrs;
3216 shader->variants_cached++;
3217 }
3218 }
3219
3220 /* Bind this variant */
3221 lp_setup_set_fs_variant(lp->setup, variant);
3222 }
3223
3224
3225
3226
3227
3228 void
3229 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3230 {
3231 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3232 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3233 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3234
3235 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3236 }
3237
3238 /*
3239 * Rasterization is disabled if there is no pixel shader and
3240 * both depth and stencil testing are disabled:
3241 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3242 */
3243 boolean
3244 llvmpipe_rasterization_disabled(struct llvmpipe_context *lp)
3245 {
3246 boolean null_fs = !lp->fs || lp->fs->info.base.num_tokens <= 1;
3247
3248 return (null_fs &&
3249 !lp->depth_stencil->depth.enabled &&
3250 !lp->depth_stencil->stencil[0].enabled);
3251 }