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
generate_quad_mask(struct gallivm_state * gallivm,struct lp_type fs_type,unsigned first_quad,LLVMValueRef mask_input)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
find_output_by_semantic(const struct tgsi_shader_info * info,unsigned semantic,unsigned index)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
lp_llvm_viewport(LLVMValueRef context_ptr,struct gallivm_state * gallivm,LLVMValueRef viewport_index)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
lp_build_depth_clamp(struct gallivm_state * gallivm,LLVMBuilderRef builder,struct lp_type type,LLVMValueRef context_ptr,LLVMValueRef thread_data_ptr,LLVMValueRef z)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
generate_fs_loop(struct gallivm_state * gallivm,struct lp_fragment_shader * shader,const struct lp_fragment_shader_variant_key * key,LLVMBuilderRef builder,struct lp_type type,LLVMValueRef context_ptr,LLVMValueRef num_loop,struct lp_build_interp_soa_context * interp,struct lp_build_sampler_soa * sampler,LLVMValueRef mask_store,LLVMValueRef (* out_color)[4],LLVMValueRef depth_ptr,LLVMValueRef depth_stride,LLVMValueRef facing,LLVMValueRef thread_data_ptr)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
generate_fs_twiddle(struct gallivm_state * gallivm,struct lp_type type,unsigned num_fs,unsigned dst_channels,LLVMValueRef fs_src[][4],LLVMValueRef * dst,bool pad_inline)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 /*
738 * XXX: we should avoid this in some cases, but would need to tell
739 * lp_build_conv to reorder (or deal with it ourselves).
740 */
741 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
742 } else {
743 /* Do nothing */
744 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
745 }
746
747 /*
748 * Moves any padding between pixels to the end
749 * e.g. RGBXRGBX -> RGBRGBXX
750 */
751 if (swizzle_pad) {
752 unsigned char swizzles[16];
753 unsigned elems = pixels * dst_channels;
754
755 for (i = 0; i < type.length; ++i) {
756 if (i < elems)
757 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
758 else
759 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
760 }
761
762 for (i = 0; i < src_count; ++i) {
763 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
764 }
765 }
766
767 return src_count;
768 }
769
770
771 /*
772 * Untwiddle and transpose, much like the above.
773 * However, this is after conversion, so we get packed vectors.
774 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
775 * the vectors will look like:
776 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
777 * be swizzled here). Extending to 16bit should be trivial.
778 * Should also be extended to handle twice wide vectors with AVX2...
779 */
780 static void
fs_twiddle_transpose(struct gallivm_state * gallivm,struct lp_type type,LLVMValueRef * src,unsigned src_count,LLVMValueRef * dst)781 fs_twiddle_transpose(struct gallivm_state *gallivm,
782 struct lp_type type,
783 LLVMValueRef *src,
784 unsigned src_count,
785 LLVMValueRef *dst)
786 {
787 unsigned i, j;
788 struct lp_type type64, type16, type32;
789 LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
790 LLVMBuilderRef builder = gallivm->builder;
791 LLVMValueRef tmp[4], shuf[8];
792 for (j = 0; j < 2; j++) {
793 shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
794 shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
795 shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
796 shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
797 }
798
799 assert(src_count == 4 || src_count == 2 || src_count == 1);
800 assert(type.width == 8);
801 assert(type.length == 16);
802
803 type8_t = lp_build_vec_type(gallivm, type);
804
805 type64 = type;
806 type64.length /= 8;
807 type64.width *= 8;
808 type64_t = lp_build_vec_type(gallivm, type64);
809
810 type16 = type;
811 type16.length /= 2;
812 type16.width *= 2;
813 type16_t = lp_build_vec_type(gallivm, type16);
814
815 type32 = type;
816 type32.length /= 4;
817 type32.width *= 4;
818 type32_t = lp_build_vec_type(gallivm, type32);
819
820 lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
821
822 if (src_count == 1) {
823 /* transpose was no-op, just untwiddle */
824 LLVMValueRef shuf_vec;
825 shuf_vec = LLVMConstVector(shuf, 8);
826 tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
827 tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
828 dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
829 } else if (src_count == 2) {
830 LLVMValueRef shuf_vec;
831 shuf_vec = LLVMConstVector(shuf, 4);
832
833 for (i = 0; i < 2; i++) {
834 tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
835 tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
836 dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
837 }
838 } else {
839 for (j = 0; j < 2; j++) {
840 LLVMValueRef lo, hi, lo2, hi2;
841 /*
842 * Note that if we only really have 3 valid channels (rgb)
843 * and we don't need alpha we could substitute a undef here
844 * for the respective channel (causing llvm to drop conversion
845 * for alpha).
846 */
847 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
848 lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
849 hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
850 lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
851 hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
852 dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
853 dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
854 }
855 }
856 }
857
858
859 /**
860 * Load an unswizzled block of pixels from memory
861 */
862 static void
load_unswizzled_block(struct gallivm_state * gallivm,LLVMValueRef base_ptr,LLVMValueRef stride,unsigned block_width,unsigned block_height,LLVMValueRef * dst,struct lp_type dst_type,unsigned dst_count,unsigned dst_alignment)863 load_unswizzled_block(struct gallivm_state *gallivm,
864 LLVMValueRef base_ptr,
865 LLVMValueRef stride,
866 unsigned block_width,
867 unsigned block_height,
868 LLVMValueRef* dst,
869 struct lp_type dst_type,
870 unsigned dst_count,
871 unsigned dst_alignment)
872 {
873 LLVMBuilderRef builder = gallivm->builder;
874 unsigned row_size = dst_count / block_height;
875 unsigned i;
876
877 /* Ensure block exactly fits into dst */
878 assert((block_width * block_height) % dst_count == 0);
879
880 for (i = 0; i < dst_count; ++i) {
881 unsigned x = i % row_size;
882 unsigned y = i / row_size;
883
884 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
885 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
886
887 LLVMValueRef gep[2];
888 LLVMValueRef dst_ptr;
889
890 gep[0] = lp_build_const_int32(gallivm, 0);
891 gep[1] = LLVMBuildAdd(builder, bx, by, "");
892
893 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
894 dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
895 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
896
897 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
898
899 LLVMSetAlignment(dst[i], dst_alignment);
900 }
901 }
902
903
904 /**
905 * Store an unswizzled block of pixels to memory
906 */
907 static void
store_unswizzled_block(struct gallivm_state * gallivm,LLVMValueRef base_ptr,LLVMValueRef stride,unsigned block_width,unsigned block_height,LLVMValueRef * src,struct lp_type src_type,unsigned src_count,unsigned src_alignment)908 store_unswizzled_block(struct gallivm_state *gallivm,
909 LLVMValueRef base_ptr,
910 LLVMValueRef stride,
911 unsigned block_width,
912 unsigned block_height,
913 LLVMValueRef* src,
914 struct lp_type src_type,
915 unsigned src_count,
916 unsigned src_alignment)
917 {
918 LLVMBuilderRef builder = gallivm->builder;
919 unsigned row_size = src_count / block_height;
920 unsigned i;
921
922 /* Ensure src exactly fits into block */
923 assert((block_width * block_height) % src_count == 0);
924
925 for (i = 0; i < src_count; ++i) {
926 unsigned x = i % row_size;
927 unsigned y = i / row_size;
928
929 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
930 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
931
932 LLVMValueRef gep[2];
933 LLVMValueRef src_ptr;
934
935 gep[0] = lp_build_const_int32(gallivm, 0);
936 gep[1] = LLVMBuildAdd(builder, bx, by, "");
937
938 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
939 src_ptr = LLVMBuildBitCast(builder, src_ptr,
940 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
941
942 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
943
944 LLVMSetAlignment(src_ptr, src_alignment);
945 }
946 }
947
948
949 /**
950 * Checks if a format description is an arithmetic format
951 *
952 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
953 */
954 static inline boolean
is_arithmetic_format(const struct util_format_description * format_desc)955 is_arithmetic_format(const struct util_format_description *format_desc)
956 {
957 boolean arith = false;
958 unsigned i;
959
960 for (i = 0; i < format_desc->nr_channels; ++i) {
961 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
962 arith |= (format_desc->channel[i].size % 8) != 0;
963 }
964
965 return arith;
966 }
967
968
969 /**
970 * Checks if this format requires special handling due to required expansion
971 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
972 * SoA conversion.
973 */
974 static inline boolean
format_expands_to_float_soa(const struct util_format_description * format_desc)975 format_expands_to_float_soa(const struct util_format_description *format_desc)
976 {
977 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
978 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
979 return true;
980 }
981 return false;
982 }
983
984
985 /**
986 * Retrieves the type representing the memory layout for a format
987 *
988 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
989 */
990 static inline void
lp_mem_type_from_format_desc(const struct util_format_description * format_desc,struct lp_type * type)991 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
992 struct lp_type* type)
993 {
994 unsigned i;
995 unsigned chan;
996
997 if (format_expands_to_float_soa(format_desc)) {
998 /* just make this a uint with width of block */
999 type->floating = false;
1000 type->fixed = false;
1001 type->sign = false;
1002 type->norm = false;
1003 type->width = format_desc->block.bits;
1004 type->length = 1;
1005 return;
1006 }
1007
1008 for (i = 0; i < 4; i++)
1009 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1010 break;
1011 chan = i;
1012
1013 memset(type, 0, sizeof(struct lp_type));
1014 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1015 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1016 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1017 type->norm = format_desc->channel[chan].normalized;
1018
1019 if (is_arithmetic_format(format_desc)) {
1020 type->width = 0;
1021 type->length = 1;
1022
1023 for (i = 0; i < format_desc->nr_channels; ++i) {
1024 type->width += format_desc->channel[i].size;
1025 }
1026 } else {
1027 type->width = format_desc->channel[chan].size;
1028 type->length = format_desc->nr_channels;
1029 }
1030 }
1031
1032
1033 /**
1034 * Retrieves the type for a format which is usable in the blending code.
1035 *
1036 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1037 */
1038 static inline void
lp_blend_type_from_format_desc(const struct util_format_description * format_desc,struct lp_type * type)1039 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1040 struct lp_type* type)
1041 {
1042 unsigned i;
1043 unsigned chan;
1044
1045 if (format_expands_to_float_soa(format_desc)) {
1046 /* always use ordinary floats for blending */
1047 type->floating = true;
1048 type->fixed = false;
1049 type->sign = true;
1050 type->norm = false;
1051 type->width = 32;
1052 type->length = 4;
1053 return;
1054 }
1055
1056 for (i = 0; i < 4; i++)
1057 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1058 break;
1059 chan = i;
1060
1061 memset(type, 0, sizeof(struct lp_type));
1062 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1063 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1064 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1065 type->norm = format_desc->channel[chan].normalized;
1066 type->width = format_desc->channel[chan].size;
1067 type->length = format_desc->nr_channels;
1068
1069 for (i = 1; i < format_desc->nr_channels; ++i) {
1070 if (format_desc->channel[i].size > type->width)
1071 type->width = format_desc->channel[i].size;
1072 }
1073
1074 if (type->floating) {
1075 type->width = 32;
1076 } else {
1077 if (type->width <= 8) {
1078 type->width = 8;
1079 } else if (type->width <= 16) {
1080 type->width = 16;
1081 } else {
1082 type->width = 32;
1083 }
1084 }
1085
1086 if (is_arithmetic_format(format_desc) && type->length == 3) {
1087 type->length = 4;
1088 }
1089 }
1090
1091
1092 /**
1093 * Scale a normalized value from src_bits to dst_bits.
1094 *
1095 * The exact calculation is
1096 *
1097 * dst = iround(src * dst_mask / src_mask)
1098 *
1099 * or with integer rounding
1100 *
1101 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1102 *
1103 * where
1104 *
1105 * src_mask = (1 << src_bits) - 1
1106 * dst_mask = (1 << dst_bits) - 1
1107 *
1108 * but we try to avoid division and multiplication through shifts.
1109 */
1110 static inline LLVMValueRef
scale_bits(struct gallivm_state * gallivm,int src_bits,int dst_bits,LLVMValueRef src,struct lp_type src_type)1111 scale_bits(struct gallivm_state *gallivm,
1112 int src_bits,
1113 int dst_bits,
1114 LLVMValueRef src,
1115 struct lp_type src_type)
1116 {
1117 LLVMBuilderRef builder = gallivm->builder;
1118 LLVMValueRef result = src;
1119
1120 if (dst_bits < src_bits) {
1121 int delta_bits = src_bits - dst_bits;
1122
1123 if (delta_bits <= dst_bits) {
1124 /*
1125 * Approximate the rescaling with a single shift.
1126 *
1127 * This gives the wrong rounding.
1128 */
1129
1130 result = LLVMBuildLShr(builder,
1131 src,
1132 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1133 "");
1134
1135 } else {
1136 /*
1137 * Try more accurate rescaling.
1138 */
1139
1140 /*
1141 * Drop the least significant bits to make space for the multiplication.
1142 *
1143 * XXX: A better approach would be to use a wider integer type as intermediate. But
1144 * this is enough to convert alpha from 16bits -> 2 when rendering to
1145 * PIPE_FORMAT_R10G10B10A2_UNORM.
1146 */
1147 result = LLVMBuildLShr(builder,
1148 src,
1149 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1150 "");
1151
1152
1153 result = LLVMBuildMul(builder,
1154 result,
1155 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1156 "");
1157
1158 /*
1159 * Add a rounding term before the division.
1160 *
1161 * TODO: Handle signed integers too.
1162 */
1163 if (!src_type.sign) {
1164 result = LLVMBuildAdd(builder,
1165 result,
1166 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1167 "");
1168 }
1169
1170 /*
1171 * Approximate the division by src_mask with a src_bits shift.
1172 *
1173 * Given the src has already been shifted by dst_bits, all we need
1174 * to do is to shift by the difference.
1175 */
1176
1177 result = LLVMBuildLShr(builder,
1178 result,
1179 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1180 "");
1181 }
1182
1183 } else if (dst_bits > src_bits) {
1184 /* Scale up bits */
1185 int db = dst_bits - src_bits;
1186
1187 /* Shift left by difference in bits */
1188 result = LLVMBuildShl(builder,
1189 src,
1190 lp_build_const_int_vec(gallivm, src_type, db),
1191 "");
1192
1193 if (db <= src_bits) {
1194 /* Enough bits in src to fill the remainder */
1195 LLVMValueRef lower = LLVMBuildLShr(builder,
1196 src,
1197 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1198 "");
1199
1200 result = LLVMBuildOr(builder, result, lower, "");
1201 } else if (db > src_bits) {
1202 /* Need to repeatedly copy src bits to fill remainder in dst */
1203 unsigned n;
1204
1205 for (n = src_bits; n < dst_bits; n *= 2) {
1206 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1207
1208 result = LLVMBuildOr(builder,
1209 result,
1210 LLVMBuildLShr(builder, result, shuv, ""),
1211 "");
1212 }
1213 }
1214 }
1215
1216 return result;
1217 }
1218
1219 /**
1220 * If RT is a smallfloat (needing denorms) format
1221 */
1222 static inline int
have_smallfloat_format(struct lp_type dst_type,enum pipe_format format)1223 have_smallfloat_format(struct lp_type dst_type,
1224 enum pipe_format format)
1225 {
1226 return ((dst_type.floating && dst_type.width != 32) ||
1227 /* due to format handling hacks this format doesn't have floating set
1228 * here (and actually has width set to 32 too) so special case this. */
1229 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1230 }
1231
1232
1233 /**
1234 * Convert from memory format to blending format
1235 *
1236 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1237 */
1238 static void
convert_to_blend_type(struct gallivm_state * gallivm,unsigned block_size,const struct util_format_description * src_fmt,struct lp_type src_type,struct lp_type dst_type,LLVMValueRef * src,unsigned num_srcs)1239 convert_to_blend_type(struct gallivm_state *gallivm,
1240 unsigned block_size,
1241 const struct util_format_description *src_fmt,
1242 struct lp_type src_type,
1243 struct lp_type dst_type,
1244 LLVMValueRef* src, // and dst
1245 unsigned num_srcs)
1246 {
1247 LLVMValueRef *dst = src;
1248 LLVMBuilderRef builder = gallivm->builder;
1249 struct lp_type blend_type;
1250 struct lp_type mem_type;
1251 unsigned i, j;
1252 unsigned pixels = block_size / num_srcs;
1253 bool is_arith;
1254
1255 /*
1256 * full custom path for packed floats and srgb formats - none of the later
1257 * functions would do anything useful, and given the lp_type representation they
1258 * can't be fixed. Should really have some SoA blend path for these kind of
1259 * formats rather than hacking them in here.
1260 */
1261 if (format_expands_to_float_soa(src_fmt)) {
1262 LLVMValueRef tmpsrc[4];
1263 /*
1264 * This is pretty suboptimal for this case blending in SoA would be much
1265 * better, since conversion gets us SoA values so need to convert back.
1266 */
1267 assert(src_type.width == 32 || src_type.width == 16);
1268 assert(dst_type.floating);
1269 assert(dst_type.width == 32);
1270 assert(dst_type.length % 4 == 0);
1271 assert(num_srcs % 4 == 0);
1272
1273 if (src_type.width == 16) {
1274 /* expand 4x16bit values to 4x32bit */
1275 struct lp_type type32x4 = src_type;
1276 LLVMTypeRef ltype32x4;
1277 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1278 type32x4.width = 32;
1279 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1280 for (i = 0; i < num_fetch; i++) {
1281 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1282 }
1283 src_type.width = 32;
1284 }
1285 for (i = 0; i < 4; i++) {
1286 tmpsrc[i] = src[i];
1287 }
1288 for (i = 0; i < num_srcs / 4; i++) {
1289 LLVMValueRef tmpsoa[4];
1290 LLVMValueRef tmps = tmpsrc[i];
1291 if (dst_type.length == 8) {
1292 LLVMValueRef shuffles[8];
1293 unsigned j;
1294 /* fetch was 4 values but need 8-wide output values */
1295 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1296 /*
1297 * for 8-wide aos transpose would give us wrong order not matching
1298 * incoming converted fs values and mask. ARGH.
1299 */
1300 for (j = 0; j < 4; j++) {
1301 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1302 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1303 }
1304 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1305 LLVMConstVector(shuffles, 8), "");
1306 }
1307 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1308 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1309 }
1310 else {
1311 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1312 }
1313 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1314 }
1315 return;
1316 }
1317
1318 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1319 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1320
1321 /* Is the format arithmetic */
1322 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1323 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1324
1325 /* Pad if necessary */
1326 if (!is_arith && src_type.length < dst_type.length) {
1327 for (i = 0; i < num_srcs; ++i) {
1328 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1329 }
1330
1331 src_type.length = dst_type.length;
1332 }
1333
1334 /* Special case for half-floats */
1335 if (mem_type.width == 16 && mem_type.floating) {
1336 assert(blend_type.width == 32 && blend_type.floating);
1337 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1338 is_arith = false;
1339 }
1340
1341 if (!is_arith) {
1342 return;
1343 }
1344
1345 src_type.width = blend_type.width * blend_type.length;
1346 blend_type.length *= pixels;
1347 src_type.length *= pixels / (src_type.length / mem_type.length);
1348
1349 for (i = 0; i < num_srcs; ++i) {
1350 LLVMValueRef chans[4];
1351 LLVMValueRef res = NULL;
1352
1353 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1354
1355 for (j = 0; j < src_fmt->nr_channels; ++j) {
1356 unsigned mask = 0;
1357 unsigned sa = src_fmt->channel[j].shift;
1358 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1359 unsigned from_lsb = j;
1360 #else
1361 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1362 #endif
1363
1364 mask = (1 << src_fmt->channel[j].size) - 1;
1365
1366 /* Extract bits from source */
1367 chans[j] = LLVMBuildLShr(builder,
1368 dst[i],
1369 lp_build_const_int_vec(gallivm, src_type, sa),
1370 "");
1371
1372 chans[j] = LLVMBuildAnd(builder,
1373 chans[j],
1374 lp_build_const_int_vec(gallivm, src_type, mask),
1375 "");
1376
1377 /* Scale bits */
1378 if (src_type.norm) {
1379 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1380 blend_type.width, chans[j], src_type);
1381 }
1382
1383 /* Insert bits into correct position */
1384 chans[j] = LLVMBuildShl(builder,
1385 chans[j],
1386 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1387 "");
1388
1389 if (j == 0) {
1390 res = chans[j];
1391 } else {
1392 res = LLVMBuildOr(builder, res, chans[j], "");
1393 }
1394 }
1395
1396 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1397 }
1398 }
1399
1400
1401 /**
1402 * Convert from blending format to memory format
1403 *
1404 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1405 */
1406 static void
convert_from_blend_type(struct gallivm_state * gallivm,unsigned block_size,const struct util_format_description * src_fmt,struct lp_type src_type,struct lp_type dst_type,LLVMValueRef * src,unsigned num_srcs)1407 convert_from_blend_type(struct gallivm_state *gallivm,
1408 unsigned block_size,
1409 const struct util_format_description *src_fmt,
1410 struct lp_type src_type,
1411 struct lp_type dst_type,
1412 LLVMValueRef* src, // and dst
1413 unsigned num_srcs)
1414 {
1415 LLVMValueRef* dst = src;
1416 unsigned i, j, k;
1417 struct lp_type mem_type;
1418 struct lp_type blend_type;
1419 LLVMBuilderRef builder = gallivm->builder;
1420 unsigned pixels = block_size / num_srcs;
1421 bool is_arith;
1422
1423 /*
1424 * full custom path for packed floats and srgb formats - none of the later
1425 * functions would do anything useful, and given the lp_type representation they
1426 * can't be fixed. Should really have some SoA blend path for these kind of
1427 * formats rather than hacking them in here.
1428 */
1429 if (format_expands_to_float_soa(src_fmt)) {
1430 /*
1431 * This is pretty suboptimal for this case blending in SoA would be much
1432 * better - we need to transpose the AoS values back to SoA values for
1433 * conversion/packing.
1434 */
1435 assert(src_type.floating);
1436 assert(src_type.width == 32);
1437 assert(src_type.length % 4 == 0);
1438 assert(dst_type.width == 32 || dst_type.width == 16);
1439
1440 for (i = 0; i < num_srcs / 4; i++) {
1441 LLVMValueRef tmpsoa[4], tmpdst;
1442 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1443 /* really really need SoA here */
1444
1445 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1446 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1447 }
1448 else {
1449 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1450 src_type, tmpsoa);
1451 }
1452
1453 if (src_type.length == 8) {
1454 LLVMValueRef tmpaos, shuffles[8];
1455 unsigned j;
1456 /*
1457 * for 8-wide aos transpose has given us wrong order not matching
1458 * output order. HMPF. Also need to split the output values manually.
1459 */
1460 for (j = 0; j < 4; j++) {
1461 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1462 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1463 }
1464 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1465 LLVMConstVector(shuffles, 8), "");
1466 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1467 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1468 }
1469 else {
1470 src[i] = tmpdst;
1471 }
1472 }
1473 if (dst_type.width == 16) {
1474 struct lp_type type16x8 = dst_type;
1475 struct lp_type type32x4 = dst_type;
1476 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1477 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1478 type16x8.length = 8;
1479 type32x4.width = 32;
1480 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1481 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1482 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1483 /* We could do vector truncation but it doesn't generate very good code */
1484 for (i = 0; i < num_fetch; i++) {
1485 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1486 src[i], lp_build_zero(gallivm, type32x4));
1487 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1488 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1489 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1490 }
1491 }
1492 return;
1493 }
1494
1495 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1496 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1497
1498 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1499
1500 /* Special case for half-floats */
1501 if (mem_type.width == 16 && mem_type.floating) {
1502 int length = dst_type.length;
1503 assert(blend_type.width == 32 && blend_type.floating);
1504
1505 dst_type.length = src_type.length;
1506
1507 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1508
1509 dst_type.length = length;
1510 is_arith = false;
1511 }
1512
1513 /* Remove any padding */
1514 if (!is_arith && (src_type.length % mem_type.length)) {
1515 src_type.length -= (src_type.length % mem_type.length);
1516
1517 for (i = 0; i < num_srcs; ++i) {
1518 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1519 }
1520 }
1521
1522 /* No bit arithmetic to do */
1523 if (!is_arith) {
1524 return;
1525 }
1526
1527 src_type.length = pixels;
1528 src_type.width = blend_type.length * blend_type.width;
1529 dst_type.length = pixels;
1530
1531 for (i = 0; i < num_srcs; ++i) {
1532 LLVMValueRef chans[4];
1533 LLVMValueRef res = NULL;
1534
1535 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1536
1537 for (j = 0; j < src_fmt->nr_channels; ++j) {
1538 unsigned mask = 0;
1539 unsigned sa = src_fmt->channel[j].shift;
1540 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1541 unsigned from_lsb = j;
1542 #else
1543 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1544 #endif
1545
1546 assert(blend_type.width > src_fmt->channel[j].size);
1547
1548 for (k = 0; k < blend_type.width; ++k) {
1549 mask |= 1 << k;
1550 }
1551
1552 /* Extract bits */
1553 chans[j] = LLVMBuildLShr(builder,
1554 dst[i],
1555 lp_build_const_int_vec(gallivm, src_type,
1556 from_lsb * blend_type.width),
1557 "");
1558
1559 chans[j] = LLVMBuildAnd(builder,
1560 chans[j],
1561 lp_build_const_int_vec(gallivm, src_type, mask),
1562 "");
1563
1564 /* Scale down bits */
1565 if (src_type.norm) {
1566 chans[j] = scale_bits(gallivm, blend_type.width,
1567 src_fmt->channel[j].size, chans[j], src_type);
1568 }
1569
1570 /* Insert bits */
1571 chans[j] = LLVMBuildShl(builder,
1572 chans[j],
1573 lp_build_const_int_vec(gallivm, src_type, sa),
1574 "");
1575
1576 sa += src_fmt->channel[j].size;
1577
1578 if (j == 0) {
1579 res = chans[j];
1580 } else {
1581 res = LLVMBuildOr(builder, res, chans[j], "");
1582 }
1583 }
1584
1585 assert (dst_type.width != 24);
1586
1587 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1588 }
1589 }
1590
1591
1592 /**
1593 * Convert alpha to same blend type as src
1594 */
1595 static void
convert_alpha(struct gallivm_state * gallivm,struct lp_type row_type,struct lp_type alpha_type,const unsigned block_size,const unsigned block_height,const unsigned src_count,const unsigned dst_channels,const bool pad_inline,LLVMValueRef * src_alpha)1596 convert_alpha(struct gallivm_state *gallivm,
1597 struct lp_type row_type,
1598 struct lp_type alpha_type,
1599 const unsigned block_size,
1600 const unsigned block_height,
1601 const unsigned src_count,
1602 const unsigned dst_channels,
1603 const bool pad_inline,
1604 LLVMValueRef* src_alpha)
1605 {
1606 LLVMBuilderRef builder = gallivm->builder;
1607 unsigned i, j;
1608 unsigned length = row_type.length;
1609 row_type.length = alpha_type.length;
1610
1611 /* Twiddle the alpha to match pixels */
1612 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1613
1614 /*
1615 * TODO this should use single lp_build_conv call for
1616 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1617 */
1618 for (i = 0; i < block_height; ++i) {
1619 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1620 }
1621
1622 alpha_type = row_type;
1623 row_type.length = length;
1624
1625 /* If only one channel we can only need the single alpha value per pixel */
1626 if (src_count == 1 && dst_channels == 1) {
1627
1628 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1629 } else {
1630 /* If there are more srcs than rows then we need to split alpha up */
1631 if (src_count > block_height) {
1632 for (i = src_count; i > 0; --i) {
1633 unsigned pixels = block_size / src_count;
1634 unsigned idx = i - 1;
1635
1636 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1637 (idx * pixels) % 4, pixels);
1638 }
1639 }
1640
1641 /* If there is a src for each pixel broadcast the alpha across whole row */
1642 if (src_count == block_size) {
1643 for (i = 0; i < src_count; ++i) {
1644 src_alpha[i] = lp_build_broadcast(gallivm,
1645 lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1646 }
1647 } else {
1648 unsigned pixels = block_size / src_count;
1649 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1650 unsigned alpha_span = 1;
1651 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1652
1653 /* Check if we need 2 src_alphas for our shuffles */
1654 if (pixels > alpha_type.length) {
1655 alpha_span = 2;
1656 }
1657
1658 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1659 for (j = 0; j < row_type.length; ++j) {
1660 if (j < pixels * channels) {
1661 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1662 } else {
1663 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1664 }
1665 }
1666
1667 for (i = 0; i < src_count; ++i) {
1668 unsigned idx1 = i, idx2 = i;
1669
1670 if (alpha_span > 1){
1671 idx1 *= alpha_span;
1672 idx2 = idx1 + 1;
1673 }
1674
1675 src_alpha[i] = LLVMBuildShuffleVector(builder,
1676 src_alpha[idx1],
1677 src_alpha[idx2],
1678 LLVMConstVector(shuffles, row_type.length),
1679 "");
1680 }
1681 }
1682 }
1683 }
1684
1685
1686 /**
1687 * Generates the blend function for unswizzled colour buffers
1688 * Also generates the read & write from colour buffer
1689 */
1690 static void
generate_unswizzled_blend(struct gallivm_state * gallivm,unsigned rt,struct lp_fragment_shader_variant * variant,enum pipe_format out_format,unsigned int num_fs,struct lp_type fs_type,LLVMValueRef * fs_mask,LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],LLVMValueRef context_ptr,LLVMValueRef color_ptr,LLVMValueRef stride,unsigned partial_mask,boolean do_branch)1691 generate_unswizzled_blend(struct gallivm_state *gallivm,
1692 unsigned rt,
1693 struct lp_fragment_shader_variant *variant,
1694 enum pipe_format out_format,
1695 unsigned int num_fs,
1696 struct lp_type fs_type,
1697 LLVMValueRef* fs_mask,
1698 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1699 LLVMValueRef context_ptr,
1700 LLVMValueRef color_ptr,
1701 LLVMValueRef stride,
1702 unsigned partial_mask,
1703 boolean do_branch)
1704 {
1705 const unsigned alpha_channel = 3;
1706 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1707 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1708 const unsigned block_size = block_width * block_height;
1709 const unsigned lp_integer_vector_width = 128;
1710
1711 LLVMBuilderRef builder = gallivm->builder;
1712 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1713 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1714 LLVMValueRef src_alpha[4 * 4];
1715 LLVMValueRef src1_alpha[4 * 4] = { NULL };
1716 LLVMValueRef src_mask[4 * 4];
1717 LLVMValueRef src[4 * 4];
1718 LLVMValueRef src1[4 * 4];
1719 LLVMValueRef dst[4 * 4];
1720 LLVMValueRef blend_color;
1721 LLVMValueRef blend_alpha;
1722 LLVMValueRef i32_zero;
1723 LLVMValueRef check_mask;
1724 LLVMValueRef undef_src_val;
1725
1726 struct lp_build_mask_context mask_ctx;
1727 struct lp_type mask_type;
1728 struct lp_type blend_type;
1729 struct lp_type row_type;
1730 struct lp_type dst_type;
1731 struct lp_type ls_type;
1732
1733 unsigned char swizzle[TGSI_NUM_CHANNELS];
1734 unsigned vector_width;
1735 unsigned src_channels = TGSI_NUM_CHANNELS;
1736 unsigned dst_channels;
1737 unsigned dst_count;
1738 unsigned src_count;
1739 unsigned i, j;
1740
1741 const struct util_format_description* out_format_desc = util_format_description(out_format);
1742
1743 unsigned dst_alignment;
1744
1745 bool pad_inline = is_arithmetic_format(out_format_desc);
1746 bool has_alpha = false;
1747 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
1748 util_blend_state_is_dual(&variant->key.blend, 0);
1749
1750 const boolean is_1d = variant->key.resource_1d;
1751 boolean twiddle_after_convert = FALSE;
1752 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
1753 LLVMValueRef fpstate = 0;
1754
1755 /* Get type from output format */
1756 lp_blend_type_from_format_desc(out_format_desc, &row_type);
1757 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
1758
1759 /*
1760 * Technically this code should go into lp_build_smallfloat_to_float
1761 * and lp_build_float_to_smallfloat but due to the
1762 * http://llvm.org/bugs/show_bug.cgi?id=6393
1763 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1764 * So the ordering is important here and there shouldn't be any
1765 * llvm ir instrunctions in this function before
1766 * this, otherwise half-float format conversions won't work
1767 * (again due to llvm bug #6393).
1768 */
1769 if (have_smallfloat_format(dst_type, out_format)) {
1770 /* We need to make sure that denorms are ok for half float
1771 conversions */
1772 fpstate = lp_build_fpstate_get(gallivm);
1773 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
1774 }
1775
1776 mask_type = lp_int32_vec4_type();
1777 mask_type.length = fs_type.length;
1778
1779 for (i = num_fs; i < num_fullblock_fs; i++) {
1780 fs_mask[i] = lp_build_zero(gallivm, mask_type);
1781 }
1782
1783 /* Do not bother executing code when mask is empty.. */
1784 if (do_branch) {
1785 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
1786
1787 for (i = 0; i < num_fullblock_fs; ++i) {
1788 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
1789 }
1790
1791 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
1792 lp_build_mask_check(&mask_ctx);
1793 }
1794
1795 partial_mask |= !variant->opaque;
1796 i32_zero = lp_build_const_int32(gallivm, 0);
1797
1798 undef_src_val = lp_build_undef(gallivm, fs_type);
1799
1800 row_type.length = fs_type.length;
1801 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
1802
1803 /* Compute correct swizzle and count channels */
1804 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
1805 dst_channels = 0;
1806
1807 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
1808 /* Ensure channel is used */
1809 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
1810 continue;
1811 }
1812
1813 /* Ensure not already written to (happens in case with GL_ALPHA) */
1814 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
1815 continue;
1816 }
1817
1818 /* Ensure we havn't already found all channels */
1819 if (dst_channels >= out_format_desc->nr_channels) {
1820 continue;
1821 }
1822
1823 swizzle[out_format_desc->swizzle[i]] = i;
1824 ++dst_channels;
1825
1826 if (i == alpha_channel) {
1827 has_alpha = true;
1828 }
1829 }
1830
1831 if (format_expands_to_float_soa(out_format_desc)) {
1832 /*
1833 * the code above can't work for layout_other
1834 * for srgb it would sort of work but we short-circuit swizzles, etc.
1835 * as that is done as part of unpack / pack.
1836 */
1837 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1838 has_alpha = true;
1839 swizzle[0] = 0;
1840 swizzle[1] = 1;
1841 swizzle[2] = 2;
1842 swizzle[3] = 3;
1843 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1844 }
1845
1846 /* If 3 channels then pad to include alpha for 4 element transpose */
1847 if (dst_channels == 3) {
1848 assert (!has_alpha);
1849 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1850 if (swizzle[i] > TGSI_NUM_CHANNELS)
1851 swizzle[i] = 3;
1852 }
1853 if (out_format_desc->nr_channels == 4) {
1854 dst_channels = 4;
1855 /*
1856 * We use alpha from the color conversion, not separate one.
1857 * We had to include it for transpose, hence it will get converted
1858 * too (albeit when doing transpose after conversion, that would
1859 * no longer be the case necessarily).
1860 * (It works only with 4 channel dsts, e.g. rgbx formats, because
1861 * otherwise we really have padding, not alpha, included.)
1862 */
1863 has_alpha = true;
1864 }
1865 }
1866
1867 /*
1868 * Load shader output
1869 */
1870 for (i = 0; i < num_fullblock_fs; ++i) {
1871 /* Always load alpha for use in blending */
1872 LLVMValueRef alpha;
1873 if (i < num_fs) {
1874 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
1875 }
1876 else {
1877 alpha = undef_src_val;
1878 }
1879
1880 /* Load each channel */
1881 for (j = 0; j < dst_channels; ++j) {
1882 assert(swizzle[j] < 4);
1883 if (i < num_fs) {
1884 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
1885 }
1886 else {
1887 fs_src[i][j] = undef_src_val;
1888 }
1889 }
1890
1891 /* If 3 channels then pad to include alpha for 4 element transpose */
1892 /*
1893 * XXX If we include that here maybe could actually use it instead of
1894 * separate alpha for blending?
1895 * (Difficult though we actually convert pad channels, not alpha.)
1896 */
1897 if (dst_channels == 3 && !has_alpha) {
1898 fs_src[i][3] = alpha;
1899 }
1900
1901 /* We split the row_mask and row_alpha as we want 128bit interleave */
1902 if (fs_type.length == 8) {
1903 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
1904 0, src_channels);
1905 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
1906 src_channels, src_channels);
1907
1908 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1909 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
1910 src_channels, src_channels);
1911 } else {
1912 src_mask[i] = fs_mask[i];
1913 src_alpha[i] = alpha;
1914 }
1915 }
1916 if (dual_source_blend) {
1917 /* same as above except different src/dst, skip masks and comments... */
1918 for (i = 0; i < num_fullblock_fs; ++i) {
1919 LLVMValueRef alpha;
1920 if (i < num_fs) {
1921 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
1922 }
1923 else {
1924 alpha = undef_src_val;
1925 }
1926
1927 for (j = 0; j < dst_channels; ++j) {
1928 assert(swizzle[j] < 4);
1929 if (i < num_fs) {
1930 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
1931 }
1932 else {
1933 fs_src1[i][j] = undef_src_val;
1934 }
1935 }
1936 if (dst_channels == 3 && !has_alpha) {
1937 fs_src1[i][3] = alpha;
1938 }
1939 if (fs_type.length == 8) {
1940 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
1941 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
1942 src_channels, src_channels);
1943 } else {
1944 src1_alpha[i] = alpha;
1945 }
1946 }
1947 }
1948
1949 if (util_format_is_pure_integer(out_format)) {
1950 /*
1951 * In this case fs_type was really ints or uints disguised as floats,
1952 * fix that up now.
1953 */
1954 fs_type.floating = 0;
1955 fs_type.sign = dst_type.sign;
1956 for (i = 0; i < num_fullblock_fs; ++i) {
1957 for (j = 0; j < dst_channels; ++j) {
1958 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
1959 lp_build_vec_type(gallivm, fs_type), "");
1960 }
1961 if (dst_channels == 3 && !has_alpha) {
1962 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
1963 lp_build_vec_type(gallivm, fs_type), "");
1964 }
1965 }
1966 }
1967
1968 /*
1969 * We actually should generally do conversion first (for non-1d cases)
1970 * when the blend format is 8 or 16 bits. The reason is obvious,
1971 * there's 2 or 4 times less vectors to deal with for the interleave...
1972 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
1973 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
1974 * unpack only with 128bit vectors).
1975 * Note: for 16bit sizes really need matching pack conversion code
1976 */
1977 if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
1978 twiddle_after_convert = TRUE;
1979 }
1980
1981 /*
1982 * Pixel twiddle from fragment shader order to memory order
1983 */
1984 if (!twiddle_after_convert) {
1985 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
1986 dst_channels, fs_src, src, pad_inline);
1987 if (dual_source_blend) {
1988 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
1989 fs_src1, src1, pad_inline);
1990 }
1991 } else {
1992 src_count = num_fullblock_fs * dst_channels;
1993 /*
1994 * We reorder things a bit here, so the cases for 4-wide and 8-wide
1995 * (AVX) turn out the same later when untwiddling/transpose (albeit
1996 * for true AVX2 path untwiddle needs to be different).
1997 * For now just order by colors first (so we can use unpack later).
1998 */
1999 for (j = 0; j < num_fullblock_fs; j++) {
2000 for (i = 0; i < dst_channels; i++) {
2001 src[i*num_fullblock_fs + j] = fs_src[j][i];
2002 if (dual_source_blend) {
2003 src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2004 }
2005 }
2006 }
2007 }
2008
2009 src_channels = dst_channels < 3 ? dst_channels : 4;
2010 if (src_count != num_fullblock_fs * src_channels) {
2011 unsigned ds = src_count / (num_fullblock_fs * src_channels);
2012 row_type.length /= ds;
2013 fs_type.length = row_type.length;
2014 }
2015
2016 blend_type = row_type;
2017 mask_type.length = 4;
2018
2019 /* Convert src to row_type */
2020 if (dual_source_blend) {
2021 struct lp_type old_row_type = row_type;
2022 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2023 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2024 }
2025 else {
2026 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2027 }
2028
2029 /* If the rows are not an SSE vector, combine them to become SSE size! */
2030 if ((row_type.width * row_type.length) % 128) {
2031 unsigned bits = row_type.width * row_type.length;
2032 unsigned combined;
2033
2034 assert(src_count >= (vector_width / bits));
2035
2036 dst_count = src_count / (vector_width / bits);
2037
2038 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2039 if (dual_source_blend) {
2040 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2041 }
2042
2043 row_type.length *= combined;
2044 src_count /= combined;
2045
2046 bits = row_type.width * row_type.length;
2047 assert(bits == 128 || bits == 256);
2048 }
2049
2050 if (twiddle_after_convert) {
2051 fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2052 if (dual_source_blend) {
2053 fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2054 }
2055 }
2056
2057 /*
2058 * Blend Colour conversion
2059 */
2060 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2061 blend_color = LLVMBuildPointerCast(builder, blend_color,
2062 LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2063 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2064 &i32_zero, 1, ""), "");
2065
2066 /* Convert */
2067 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2068
2069 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2070 /*
2071 * since blending is done with floats, there was no conversion.
2072 * However, the rules according to fixed point renderbuffers still
2073 * apply, that is we must clamp inputs to 0.0/1.0.
2074 * (This would apply to separate alpha conversion too but we currently
2075 * force has_alpha to be true.)
2076 * TODO: should skip this with "fake" blend, since post-blend conversion
2077 * will clamp anyway.
2078 * TODO: could also skip this if fragment color clamping is enabled. We
2079 * don't support it natively so it gets baked into the shader however, so
2080 * can't really tell here.
2081 */
2082 struct lp_build_context f32_bld;
2083 assert(row_type.floating);
2084 lp_build_context_init(&f32_bld, gallivm, row_type);
2085 for (i = 0; i < src_count; i++) {
2086 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2087 }
2088 if (dual_source_blend) {
2089 for (i = 0; i < src_count; i++) {
2090 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2091 }
2092 }
2093 /* probably can't be different than row_type but better safe than sorry... */
2094 lp_build_context_init(&f32_bld, gallivm, blend_type);
2095 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2096 }
2097
2098 /* Extract alpha */
2099 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2100
2101 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2102 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2103 if (pad_inline) {
2104 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2105 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2106 } else {
2107 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2108 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2109 }
2110
2111 /*
2112 * Mask conversion
2113 */
2114 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2115
2116 if (src_count < block_height) {
2117 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2118 } else if (src_count > block_height) {
2119 for (i = src_count; i > 0; --i) {
2120 unsigned pixels = block_size / src_count;
2121 unsigned idx = i - 1;
2122
2123 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2124 (idx * pixels) % 4, pixels);
2125 }
2126 }
2127
2128 assert(mask_type.width == 32);
2129
2130 for (i = 0; i < src_count; ++i) {
2131 unsigned pixels = block_size / src_count;
2132 unsigned pixel_width = row_type.width * dst_channels;
2133
2134 if (pixel_width == 24) {
2135 mask_type.width = 8;
2136 mask_type.length = vector_width / mask_type.width;
2137 } else {
2138 mask_type.length = pixels;
2139 mask_type.width = row_type.width * dst_channels;
2140
2141 /*
2142 * If mask_type width is smaller than 32bit, this doesn't quite
2143 * generate the most efficient code (could use some pack).
2144 */
2145 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2146 lp_build_int_vec_type(gallivm, mask_type), "");
2147
2148 mask_type.length *= dst_channels;
2149 mask_type.width /= dst_channels;
2150 }
2151
2152 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2153 lp_build_int_vec_type(gallivm, mask_type), "");
2154 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2155 }
2156
2157 /*
2158 * Alpha conversion
2159 */
2160 if (!has_alpha) {
2161 struct lp_type alpha_type = fs_type;
2162 alpha_type.length = 4;
2163 convert_alpha(gallivm, row_type, alpha_type,
2164 block_size, block_height,
2165 src_count, dst_channels,
2166 pad_inline, src_alpha);
2167 if (dual_source_blend) {
2168 convert_alpha(gallivm, row_type, alpha_type,
2169 block_size, block_height,
2170 src_count, dst_channels,
2171 pad_inline, src1_alpha);
2172 }
2173 }
2174
2175
2176 /*
2177 * Load dst from memory
2178 */
2179 if (src_count < block_height) {
2180 dst_count = block_height;
2181 } else {
2182 dst_count = src_count;
2183 }
2184
2185 dst_type.length *= block_size / dst_count;
2186
2187 if (format_expands_to_float_soa(out_format_desc)) {
2188 /*
2189 * we need multiple values at once for the conversion, so can as well
2190 * load them vectorized here too instead of concatenating later.
2191 * (Still need concatenation later for 8-wide vectors).
2192 */
2193 dst_count = block_height;
2194 dst_type.length = block_width;
2195 }
2196
2197 /*
2198 * Compute the alignment of the destination pointer in bytes
2199 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2200 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2201 * 1d tex but can't distinguish here) so need to stick with per-pixel
2202 * alignment in this case.
2203 */
2204 if (is_1d) {
2205 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2206 }
2207 else {
2208 dst_alignment = dst_type.length * dst_type.width / 8;
2209 }
2210 /* Force power-of-two alignment by extracting only the least-significant-bit */
2211 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2212 /*
2213 * Resource base and stride pointers are aligned to 16 bytes, so that's
2214 * the maximum alignment we can guarantee
2215 */
2216 dst_alignment = MIN2(16, dst_alignment);
2217
2218 ls_type = dst_type;
2219
2220 if (dst_count > src_count) {
2221 if ((dst_type.width == 8 || dst_type.width == 16) &&
2222 util_is_power_of_two(dst_type.length) &&
2223 dst_type.length * dst_type.width < 128) {
2224 /*
2225 * Never try to load values as 4xi8 which we will then
2226 * concatenate to larger vectors. This gives llvm a real
2227 * headache (the problem is the type legalizer (?) will
2228 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2229 * then the shuffles to concatenate are more or less impossible
2230 * - llvm is easily capable of generating a sequence of 32
2231 * pextrb/pinsrb instructions for that. Albeit it appears to
2232 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2233 * width to avoid the trouble (16bit seems not as bad, llvm
2234 * probably recognizes the load+shuffle as only one shuffle
2235 * is necessary, but we can do just the same anyway).
2236 */
2237 ls_type.length = dst_type.length * dst_type.width / 32;
2238 ls_type.width = 32;
2239 }
2240 }
2241
2242 if (is_1d) {
2243 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2244 dst, ls_type, dst_count / 4, dst_alignment);
2245 for (i = dst_count / 4; i < dst_count; i++) {
2246 dst[i] = lp_build_undef(gallivm, ls_type);
2247 }
2248
2249 }
2250 else {
2251 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2252 dst, ls_type, dst_count, dst_alignment);
2253 }
2254
2255
2256 /*
2257 * Convert from dst/output format to src/blending format.
2258 *
2259 * This is necessary as we can only read 1 row from memory at a time,
2260 * so the minimum dst_count will ever be at this point is 4.
2261 *
2262 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2263 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2264 * on all 16 pixels in that single vector at once.
2265 */
2266 if (dst_count > src_count) {
2267 if (ls_type.length != dst_type.length && ls_type.length == 1) {
2268 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2269 LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2270 for (i = 0; i < dst_count; i++) {
2271 dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2272 }
2273 }
2274
2275 lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2276
2277 if (ls_type.length != dst_type.length) {
2278 struct lp_type tmp_type = dst_type;
2279 tmp_type.length = dst_type.length * 4 / src_count;
2280 for (i = 0; i < src_count; i++) {
2281 dst[i] = LLVMBuildBitCast(builder, dst[i],
2282 lp_build_vec_type(gallivm, tmp_type), "");
2283 }
2284 }
2285 }
2286
2287 /*
2288 * Blending
2289 */
2290 /* XXX this is broken for RGB8 formats -
2291 * they get expanded from 12 to 16 elements (to include alpha)
2292 * by convert_to_blend_type then reduced to 15 instead of 12
2293 * by convert_from_blend_type (a simple fix though breaks A8...).
2294 * R16G16B16 also crashes differently however something going wrong
2295 * inside llvm handling npot vector sizes seemingly.
2296 * It seems some cleanup could be done here (like skipping conversion/blend
2297 * when not needed).
2298 */
2299 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2300 row_type, dst, src_count);
2301
2302 /*
2303 * FIXME: Really should get logic ops / masks out of generic blend / row
2304 * format. Logic ops will definitely not work on the blend float format
2305 * used for SRGB here and I think OpenGL expects this to work as expected
2306 * (that is incoming values converted to srgb then logic op applied).
2307 */
2308 for (i = 0; i < src_count; ++i) {
2309 dst[i] = lp_build_blend_aos(gallivm,
2310 &variant->key.blend,
2311 out_format,
2312 row_type,
2313 rt,
2314 src[i],
2315 has_alpha ? NULL : src_alpha[i],
2316 src1[i],
2317 has_alpha ? NULL : src1_alpha[i],
2318 dst[i],
2319 partial_mask ? src_mask[i] : NULL,
2320 blend_color,
2321 has_alpha ? NULL : blend_alpha,
2322 swizzle,
2323 pad_inline ? 4 : dst_channels);
2324 }
2325
2326 convert_from_blend_type(gallivm, block_size, out_format_desc,
2327 row_type, dst_type, dst, src_count);
2328
2329 /* Split the blend rows back to memory rows */
2330 if (dst_count > src_count) {
2331 row_type.length = dst_type.length * (dst_count / src_count);
2332
2333 if (src_count == 1) {
2334 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2335 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2336
2337 row_type.length /= 2;
2338 src_count *= 2;
2339 }
2340
2341 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2342 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2343 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2344 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2345
2346 row_type.length /= 2;
2347 src_count *= 2;
2348 }
2349
2350 /*
2351 * Store blend result to memory
2352 */
2353 if (is_1d) {
2354 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2355 dst, dst_type, dst_count / 4, dst_alignment);
2356 }
2357 else {
2358 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2359 dst, dst_type, dst_count, dst_alignment);
2360 }
2361
2362 if (have_smallfloat_format(dst_type, out_format)) {
2363 lp_build_fpstate_set(gallivm, fpstate);
2364 }
2365
2366 if (do_branch) {
2367 lp_build_mask_end(&mask_ctx);
2368 }
2369 }
2370
2371
2372 /**
2373 * Generate the runtime callable function for the whole fragment pipeline.
2374 * Note that the function which we generate operates on a block of 16
2375 * pixels at at time. The block contains 2x2 quads. Each quad contains
2376 * 2x2 pixels.
2377 */
2378 static void
generate_fragment(struct llvmpipe_context * lp,struct lp_fragment_shader * shader,struct lp_fragment_shader_variant * variant,unsigned partial_mask)2379 generate_fragment(struct llvmpipe_context *lp,
2380 struct lp_fragment_shader *shader,
2381 struct lp_fragment_shader_variant *variant,
2382 unsigned partial_mask)
2383 {
2384 struct gallivm_state *gallivm = variant->gallivm;
2385 const struct lp_fragment_shader_variant_key *key = &variant->key;
2386 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2387 char func_name[64];
2388 struct lp_type fs_type;
2389 struct lp_type blend_type;
2390 LLVMTypeRef fs_elem_type;
2391 LLVMTypeRef blend_vec_type;
2392 LLVMTypeRef arg_types[13];
2393 LLVMTypeRef func_type;
2394 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2395 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2396 LLVMValueRef context_ptr;
2397 LLVMValueRef x;
2398 LLVMValueRef y;
2399 LLVMValueRef a0_ptr;
2400 LLVMValueRef dadx_ptr;
2401 LLVMValueRef dady_ptr;
2402 LLVMValueRef color_ptr_ptr;
2403 LLVMValueRef stride_ptr;
2404 LLVMValueRef depth_ptr;
2405 LLVMValueRef depth_stride;
2406 LLVMValueRef mask_input;
2407 LLVMValueRef thread_data_ptr;
2408 LLVMBasicBlockRef block;
2409 LLVMBuilderRef builder;
2410 struct lp_build_sampler_soa *sampler;
2411 struct lp_build_interp_soa_context interp;
2412 LLVMValueRef fs_mask[16 / 4];
2413 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2414 LLVMValueRef function;
2415 LLVMValueRef facing;
2416 unsigned num_fs;
2417 unsigned i;
2418 unsigned chan;
2419 unsigned cbuf;
2420 boolean cbuf0_write_all;
2421 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2422 util_blend_state_is_dual(&key->blend, 0);
2423
2424 assert(lp_native_vector_width / 32 >= 4);
2425
2426 /* Adjust color input interpolation according to flatshade state:
2427 */
2428 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2429 for (i = 0; i < shader->info.base.num_inputs; i++) {
2430 if (inputs[i].interp == LP_INTERP_COLOR) {
2431 if (key->flatshade)
2432 inputs[i].interp = LP_INTERP_CONSTANT;
2433 else
2434 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2435 }
2436 }
2437
2438 /* check if writes to cbuf[0] are to be copied to all cbufs */
2439 cbuf0_write_all =
2440 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2441
2442 /* TODO: actually pick these based on the fs and color buffer
2443 * characteristics. */
2444
2445 memset(&fs_type, 0, sizeof fs_type);
2446 fs_type.floating = TRUE; /* floating point values */
2447 fs_type.sign = TRUE; /* values are signed */
2448 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2449 fs_type.width = 32; /* 32-bit float */
2450 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2451
2452 memset(&blend_type, 0, sizeof blend_type);
2453 blend_type.floating = FALSE; /* values are integers */
2454 blend_type.sign = FALSE; /* values are unsigned */
2455 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2456 blend_type.width = 8; /* 8-bit ubyte values */
2457 blend_type.length = 16; /* 16 elements per vector */
2458
2459 /*
2460 * Generate the function prototype. Any change here must be reflected in
2461 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2462 */
2463
2464 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2465
2466 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2467
2468 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2469 shader->no, variant->no, partial_mask ? "partial" : "whole");
2470
2471 arg_types[0] = variant->jit_context_ptr_type; /* context */
2472 arg_types[1] = int32_type; /* x */
2473 arg_types[2] = int32_type; /* y */
2474 arg_types[3] = int32_type; /* facing */
2475 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2476 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2477 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2478 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
2479 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2480 arg_types[9] = int32_type; /* mask_input */
2481 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2482 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2483 arg_types[12] = int32_type; /* depth_stride */
2484
2485 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2486 arg_types, ARRAY_SIZE(arg_types), 0);
2487
2488 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2489 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2490
2491 variant->function[partial_mask] = function;
2492
2493 /* XXX: need to propagate noalias down into color param now we are
2494 * passing a pointer-to-pointer?
2495 */
2496 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2497 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2498 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2499
2500 context_ptr = LLVMGetParam(function, 0);
2501 x = LLVMGetParam(function, 1);
2502 y = LLVMGetParam(function, 2);
2503 facing = LLVMGetParam(function, 3);
2504 a0_ptr = LLVMGetParam(function, 4);
2505 dadx_ptr = LLVMGetParam(function, 5);
2506 dady_ptr = LLVMGetParam(function, 6);
2507 color_ptr_ptr = LLVMGetParam(function, 7);
2508 depth_ptr = LLVMGetParam(function, 8);
2509 mask_input = LLVMGetParam(function, 9);
2510 thread_data_ptr = LLVMGetParam(function, 10);
2511 stride_ptr = LLVMGetParam(function, 11);
2512 depth_stride = LLVMGetParam(function, 12);
2513
2514 lp_build_name(context_ptr, "context");
2515 lp_build_name(x, "x");
2516 lp_build_name(y, "y");
2517 lp_build_name(a0_ptr, "a0");
2518 lp_build_name(dadx_ptr, "dadx");
2519 lp_build_name(dady_ptr, "dady");
2520 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2521 lp_build_name(depth_ptr, "depth");
2522 lp_build_name(mask_input, "mask_input");
2523 lp_build_name(thread_data_ptr, "thread_data");
2524 lp_build_name(stride_ptr, "stride_ptr");
2525 lp_build_name(depth_stride, "depth_stride");
2526
2527 /*
2528 * Function body
2529 */
2530
2531 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2532 builder = gallivm->builder;
2533 assert(builder);
2534 LLVMPositionBuilderAtEnd(builder, block);
2535
2536 /* code generated texture sampling */
2537 sampler = lp_llvm_sampler_soa_create(key->state);
2538
2539 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2540 /* for 1d resources only run "upper half" of stamp */
2541 if (key->resource_1d)
2542 num_fs /= 2;
2543
2544 {
2545 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2546 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2547 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2548 num_loop, "mask_store");
2549 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2550 boolean pixel_center_integer =
2551 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2552
2553 /*
2554 * The shader input interpolation info is not explicitely baked in the
2555 * shader key, but everything it derives from (TGSI, and flatshade) is
2556 * already included in the shader key.
2557 */
2558 lp_build_interp_soa_init(&interp,
2559 gallivm,
2560 shader->info.base.num_inputs,
2561 inputs,
2562 pixel_center_integer,
2563 key->depth_clamp,
2564 builder, fs_type,
2565 a0_ptr, dadx_ptr, dady_ptr,
2566 x, y);
2567
2568 for (i = 0; i < num_fs; i++) {
2569 LLVMValueRef mask;
2570 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2571 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2572 &indexi, 1, "mask_ptr");
2573
2574 if (partial_mask) {
2575 mask = generate_quad_mask(gallivm, fs_type,
2576 i*fs_type.length/4, mask_input);
2577 }
2578 else {
2579 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2580 }
2581 LLVMBuildStore(builder, mask, mask_ptr);
2582 }
2583
2584 generate_fs_loop(gallivm,
2585 shader, key,
2586 builder,
2587 fs_type,
2588 context_ptr,
2589 num_loop,
2590 &interp,
2591 sampler,
2592 mask_store, /* output */
2593 color_store,
2594 depth_ptr,
2595 depth_stride,
2596 facing,
2597 thread_data_ptr);
2598
2599 for (i = 0; i < num_fs; i++) {
2600 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2601 LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
2602 &indexi, 1, "");
2603 fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
2604 /* This is fucked up need to reorganize things */
2605 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2606 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2607 ptr = LLVMBuildGEP(builder,
2608 color_store[cbuf * !cbuf0_write_all][chan],
2609 &indexi, 1, "");
2610 fs_out_color[cbuf][chan][i] = ptr;
2611 }
2612 }
2613 if (dual_source_blend) {
2614 /* only support one dual source blend target hence always use output 1 */
2615 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2616 ptr = LLVMBuildGEP(builder,
2617 color_store[1][chan],
2618 &indexi, 1, "");
2619 fs_out_color[1][chan][i] = ptr;
2620 }
2621 }
2622 }
2623 }
2624
2625 sampler->destroy(sampler);
2626
2627 /* Loop over color outputs / color buffers to do blending.
2628 */
2629 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2630 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
2631 LLVMValueRef color_ptr;
2632 LLVMValueRef stride;
2633 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2634
2635 boolean do_branch = ((key->depth.enabled
2636 || key->stencil[0].enabled
2637 || key->alpha.enabled)
2638 && !shader->info.base.uses_kill);
2639
2640 color_ptr = LLVMBuildLoad(builder,
2641 LLVMBuildGEP(builder, color_ptr_ptr,
2642 &index, 1, ""),
2643 "");
2644
2645 lp_build_name(color_ptr, "color_ptr%d", cbuf);
2646
2647 stride = LLVMBuildLoad(builder,
2648 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
2649 "");
2650
2651 generate_unswizzled_blend(gallivm, cbuf, variant,
2652 key->cbuf_format[cbuf],
2653 num_fs, fs_type, fs_mask, fs_out_color,
2654 context_ptr, color_ptr, stride,
2655 partial_mask, do_branch);
2656 }
2657 }
2658
2659 LLVMBuildRetVoid(builder);
2660
2661 gallivm_verify_function(gallivm, function);
2662 }
2663
2664
2665 static void
dump_fs_variant_key(const struct lp_fragment_shader_variant_key * key)2666 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
2667 {
2668 unsigned i;
2669
2670 debug_printf("fs variant %p:\n", (void *) key);
2671
2672 if (key->flatshade) {
2673 debug_printf("flatshade = 1\n");
2674 }
2675 for (i = 0; i < key->nr_cbufs; ++i) {
2676 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
2677 }
2678 if (key->depth.enabled || key->stencil[0].enabled) {
2679 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
2680 }
2681 if (key->depth.enabled) {
2682 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
2683 debug_printf("depth.writemask = %u\n", key->depth.writemask);
2684 }
2685
2686 for (i = 0; i < 2; ++i) {
2687 if (key->stencil[i].enabled) {
2688 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
2689 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
2690 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
2691 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
2692 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
2693 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
2694 }
2695 }
2696
2697 if (key->alpha.enabled) {
2698 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
2699 }
2700
2701 if (key->occlusion_count) {
2702 debug_printf("occlusion_count = 1\n");
2703 }
2704
2705 if (key->blend.logicop_enable) {
2706 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
2707 }
2708 else if (key->blend.rt[0].blend_enable) {
2709 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
2710 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
2711 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
2712 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
2713 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
2714 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
2715 }
2716 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
2717 if (key->blend.alpha_to_coverage) {
2718 debug_printf("blend.alpha_to_coverage is enabled\n");
2719 }
2720 for (i = 0; i < key->nr_samplers; ++i) {
2721 const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
2722 debug_printf("sampler[%u] = \n", i);
2723 debug_printf(" .wrap = %s %s %s\n",
2724 util_dump_tex_wrap(sampler->wrap_s, TRUE),
2725 util_dump_tex_wrap(sampler->wrap_t, TRUE),
2726 util_dump_tex_wrap(sampler->wrap_r, TRUE));
2727 debug_printf(" .min_img_filter = %s\n",
2728 util_dump_tex_filter(sampler->min_img_filter, TRUE));
2729 debug_printf(" .min_mip_filter = %s\n",
2730 util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
2731 debug_printf(" .mag_img_filter = %s\n",
2732 util_dump_tex_filter(sampler->mag_img_filter, TRUE));
2733 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
2734 debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
2735 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
2736 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
2737 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
2738 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
2739 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
2740 }
2741 for (i = 0; i < key->nr_sampler_views; ++i) {
2742 const struct lp_static_texture_state *texture = &key->state[i].texture_state;
2743 debug_printf("texture[%u] = \n", i);
2744 debug_printf(" .format = %s\n",
2745 util_format_name(texture->format));
2746 debug_printf(" .target = %s\n",
2747 util_dump_tex_target(texture->target, TRUE));
2748 debug_printf(" .level_zero_only = %u\n",
2749 texture->level_zero_only);
2750 debug_printf(" .pot = %u %u %u\n",
2751 texture->pot_width,
2752 texture->pot_height,
2753 texture->pot_depth);
2754 }
2755 }
2756
2757
2758 void
lp_debug_fs_variant(const struct lp_fragment_shader_variant * variant)2759 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
2760 {
2761 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2762 variant->shader->no, variant->no);
2763 tgsi_dump(variant->shader->base.tokens, 0);
2764 dump_fs_variant_key(&variant->key);
2765 debug_printf("variant->opaque = %u\n", variant->opaque);
2766 debug_printf("\n");
2767 }
2768
2769
2770 /**
2771 * Generate a new fragment shader variant from the shader code and
2772 * other state indicated by the key.
2773 */
2774 static struct lp_fragment_shader_variant *
generate_variant(struct llvmpipe_context * lp,struct lp_fragment_shader * shader,const struct lp_fragment_shader_variant_key * key)2775 generate_variant(struct llvmpipe_context *lp,
2776 struct lp_fragment_shader *shader,
2777 const struct lp_fragment_shader_variant_key *key)
2778 {
2779 struct lp_fragment_shader_variant *variant;
2780 const struct util_format_description *cbuf0_format_desc;
2781 boolean fullcolormask;
2782 char module_name[64];
2783
2784 variant = CALLOC_STRUCT(lp_fragment_shader_variant);
2785 if (!variant)
2786 return NULL;
2787
2788 util_snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
2789 shader->no, shader->variants_created);
2790
2791 variant->gallivm = gallivm_create(module_name, lp->context);
2792 if (!variant->gallivm) {
2793 FREE(variant);
2794 return NULL;
2795 }
2796
2797 variant->shader = shader;
2798 variant->list_item_global.base = variant;
2799 variant->list_item_local.base = variant;
2800 variant->no = shader->variants_created++;
2801
2802 memcpy(&variant->key, key, shader->variant_key_size);
2803
2804 /*
2805 * Determine whether we are touching all channels in the color buffer.
2806 */
2807 fullcolormask = FALSE;
2808 if (key->nr_cbufs == 1) {
2809 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
2810 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
2811 }
2812
2813 variant->opaque =
2814 !key->blend.logicop_enable &&
2815 !key->blend.rt[0].blend_enable &&
2816 fullcolormask &&
2817 !key->stencil[0].enabled &&
2818 !key->alpha.enabled &&
2819 !key->blend.alpha_to_coverage &&
2820 !key->depth.enabled &&
2821 !shader->info.base.uses_kill
2822 ? TRUE : FALSE;
2823
2824 if ((shader->info.base.num_tokens <= 1) &&
2825 !key->depth.enabled && !key->stencil[0].enabled) {
2826 variant->ps_inv_multiplier = 0;
2827 } else {
2828 variant->ps_inv_multiplier = 1;
2829 }
2830
2831 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
2832 lp_debug_fs_variant(variant);
2833 }
2834
2835 lp_jit_init_types(variant);
2836
2837 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
2838 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
2839
2840 if (variant->jit_function[RAST_WHOLE] == NULL) {
2841 if (variant->opaque) {
2842 /* Specialized shader, which doesn't need to read the color buffer. */
2843 generate_fragment(lp, shader, variant, RAST_WHOLE);
2844 }
2845 }
2846
2847 /*
2848 * Compile everything
2849 */
2850
2851 gallivm_compile_module(variant->gallivm);
2852
2853 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
2854
2855 if (variant->function[RAST_EDGE_TEST]) {
2856 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
2857 gallivm_jit_function(variant->gallivm,
2858 variant->function[RAST_EDGE_TEST]);
2859 }
2860
2861 if (variant->function[RAST_WHOLE]) {
2862 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
2863 gallivm_jit_function(variant->gallivm,
2864 variant->function[RAST_WHOLE]);
2865 } else if (!variant->jit_function[RAST_WHOLE]) {
2866 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
2867 }
2868
2869 gallivm_free_ir(variant->gallivm);
2870
2871 return variant;
2872 }
2873
2874
2875 static void *
llvmpipe_create_fs_state(struct pipe_context * pipe,const struct pipe_shader_state * templ)2876 llvmpipe_create_fs_state(struct pipe_context *pipe,
2877 const struct pipe_shader_state *templ)
2878 {
2879 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2880 struct lp_fragment_shader *shader;
2881 int nr_samplers;
2882 int nr_sampler_views;
2883 int i;
2884
2885 shader = CALLOC_STRUCT(lp_fragment_shader);
2886 if (!shader)
2887 return NULL;
2888
2889 shader->no = fs_no++;
2890 make_empty_list(&shader->variants);
2891
2892 /* get/save the summary info for this shader */
2893 lp_build_tgsi_info(templ->tokens, &shader->info);
2894
2895 /* we need to keep a local copy of the tokens */
2896 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
2897
2898 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
2899 if (shader->draw_data == NULL) {
2900 FREE((void *) shader->base.tokens);
2901 FREE(shader);
2902 return NULL;
2903 }
2904
2905 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
2906 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2907
2908 shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
2909 state[MAX2(nr_samplers, nr_sampler_views)]);
2910
2911 for (i = 0; i < shader->info.base.num_inputs; i++) {
2912 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
2913 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
2914
2915 switch (shader->info.base.input_interpolate[i]) {
2916 case TGSI_INTERPOLATE_CONSTANT:
2917 shader->inputs[i].interp = LP_INTERP_CONSTANT;
2918 break;
2919 case TGSI_INTERPOLATE_LINEAR:
2920 shader->inputs[i].interp = LP_INTERP_LINEAR;
2921 break;
2922 case TGSI_INTERPOLATE_PERSPECTIVE:
2923 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
2924 break;
2925 case TGSI_INTERPOLATE_COLOR:
2926 shader->inputs[i].interp = LP_INTERP_COLOR;
2927 break;
2928 default:
2929 assert(0);
2930 break;
2931 }
2932
2933 switch (shader->info.base.input_semantic_name[i]) {
2934 case TGSI_SEMANTIC_FACE:
2935 shader->inputs[i].interp = LP_INTERP_FACING;
2936 break;
2937 case TGSI_SEMANTIC_POSITION:
2938 /* Position was already emitted above
2939 */
2940 shader->inputs[i].interp = LP_INTERP_POSITION;
2941 shader->inputs[i].src_index = 0;
2942 continue;
2943 }
2944
2945 /* XXX this is a completely pointless index map... */
2946 shader->inputs[i].src_index = i+1;
2947 }
2948
2949 if (LP_DEBUG & DEBUG_TGSI) {
2950 unsigned attrib;
2951 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2952 shader->no, (void *) shader);
2953 tgsi_dump(templ->tokens, 0);
2954 debug_printf("usage masks:\n");
2955 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
2956 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
2957 debug_printf(" IN[%u].%s%s%s%s\n",
2958 attrib,
2959 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
2960 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
2961 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
2962 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
2963 }
2964 debug_printf("\n");
2965 }
2966
2967 return shader;
2968 }
2969
2970
2971 static void
llvmpipe_bind_fs_state(struct pipe_context * pipe,void * fs)2972 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
2973 {
2974 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
2975
2976 if (llvmpipe->fs == fs)
2977 return;
2978
2979 llvmpipe->fs = (struct lp_fragment_shader *) fs;
2980
2981 draw_bind_fragment_shader(llvmpipe->draw,
2982 (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
2983
2984 llvmpipe->dirty |= LP_NEW_FS;
2985 }
2986
2987
2988 /**
2989 * Remove shader variant from two lists: the shader's variant list
2990 * and the context's variant list.
2991 */
2992 void
llvmpipe_remove_shader_variant(struct llvmpipe_context * lp,struct lp_fragment_shader_variant * variant)2993 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
2994 struct lp_fragment_shader_variant *variant)
2995 {
2996 if (gallivm_debug & GALLIVM_DEBUG_IR) {
2997 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2998 " #%u v total cached #%u\n",
2999 variant->shader->no,
3000 variant->no,
3001 variant->shader->variants_created,
3002 variant->shader->variants_cached,
3003 lp->nr_fs_variants);
3004 }
3005
3006 gallivm_destroy(variant->gallivm);
3007
3008 /* remove from shader's list */
3009 remove_from_list(&variant->list_item_local);
3010 variant->shader->variants_cached--;
3011
3012 /* remove from context's list */
3013 remove_from_list(&variant->list_item_global);
3014 lp->nr_fs_variants--;
3015 lp->nr_fs_instrs -= variant->nr_instrs;
3016
3017 FREE(variant);
3018 }
3019
3020
3021 static void
llvmpipe_delete_fs_state(struct pipe_context * pipe,void * fs)3022 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
3023 {
3024 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3025 struct lp_fragment_shader *shader = fs;
3026 struct lp_fs_variant_list_item *li;
3027
3028 assert(fs != llvmpipe->fs);
3029
3030 /*
3031 * XXX: we need to flush the context until we have some sort of reference
3032 * counting in fragment shaders as they may still be binned
3033 * Flushing alone might not sufficient we need to wait on it too.
3034 */
3035 llvmpipe_finish(pipe, __FUNCTION__);
3036
3037 /* Delete all the variants */
3038 li = first_elem(&shader->variants);
3039 while(!at_end(&shader->variants, li)) {
3040 struct lp_fs_variant_list_item *next = next_elem(li);
3041 llvmpipe_remove_shader_variant(llvmpipe, li->base);
3042 li = next;
3043 }
3044
3045 /* Delete draw module's data */
3046 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
3047
3048 assert(shader->variants_cached == 0);
3049 FREE((void *) shader->base.tokens);
3050 FREE(shader);
3051 }
3052
3053
3054
3055 static void
llvmpipe_set_constant_buffer(struct pipe_context * pipe,uint shader,uint index,const struct pipe_constant_buffer * cb)3056 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
3057 uint shader, uint index,
3058 const struct pipe_constant_buffer *cb)
3059 {
3060 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3061 struct pipe_resource *constants = cb ? cb->buffer : NULL;
3062
3063 assert(shader < PIPE_SHADER_TYPES);
3064 assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
3065
3066 /* note: reference counting */
3067 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
3068
3069 if (constants) {
3070 if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
3071 debug_printf("Illegal set constant without bind flag\n");
3072 constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
3073 }
3074 }
3075
3076 if (shader == PIPE_SHADER_VERTEX ||
3077 shader == PIPE_SHADER_GEOMETRY) {
3078 /* Pass the constants to the 'draw' module */
3079 const unsigned size = cb ? cb->buffer_size : 0;
3080 const ubyte *data;
3081
3082 if (constants) {
3083 data = (ubyte *) llvmpipe_resource_data(constants);
3084 }
3085 else if (cb && cb->user_buffer) {
3086 data = (ubyte *) cb->user_buffer;
3087 }
3088 else {
3089 data = NULL;
3090 }
3091
3092 if (data)
3093 data += cb->buffer_offset;
3094
3095 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
3096 index, data, size);
3097 }
3098 else {
3099 llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
3100 }
3101
3102 if (cb && cb->user_buffer) {
3103 pipe_resource_reference(&constants, NULL);
3104 }
3105 }
3106
3107
3108 /**
3109 * Return the blend factor equivalent to a destination alpha of one.
3110 */
3111 static inline unsigned
force_dst_alpha_one(unsigned factor,boolean clamped_zero)3112 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
3113 {
3114 switch(factor) {
3115 case PIPE_BLENDFACTOR_DST_ALPHA:
3116 return PIPE_BLENDFACTOR_ONE;
3117 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
3118 return PIPE_BLENDFACTOR_ZERO;
3119 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
3120 if (clamped_zero)
3121 return PIPE_BLENDFACTOR_ZERO;
3122 else
3123 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
3124 }
3125
3126 return factor;
3127 }
3128
3129
3130 /**
3131 * We need to generate several variants of the fragment pipeline to match
3132 * all the combinations of the contributing state atoms.
3133 *
3134 * TODO: there is actually no reason to tie this to context state -- the
3135 * generated code could be cached globally in the screen.
3136 */
3137 static void
make_variant_key(struct llvmpipe_context * lp,struct lp_fragment_shader * shader,struct lp_fragment_shader_variant_key * key)3138 make_variant_key(struct llvmpipe_context *lp,
3139 struct lp_fragment_shader *shader,
3140 struct lp_fragment_shader_variant_key *key)
3141 {
3142 unsigned i;
3143
3144 memset(key, 0, shader->variant_key_size);
3145
3146 if (lp->framebuffer.zsbuf) {
3147 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
3148 const struct util_format_description *zsbuf_desc =
3149 util_format_description(zsbuf_format);
3150
3151 if (lp->depth_stencil->depth.enabled &&
3152 util_format_has_depth(zsbuf_desc)) {
3153 key->zsbuf_format = zsbuf_format;
3154 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
3155 }
3156 if (lp->depth_stencil->stencil[0].enabled &&
3157 util_format_has_stencil(zsbuf_desc)) {
3158 key->zsbuf_format = zsbuf_format;
3159 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
3160 }
3161 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
3162 key->resource_1d = TRUE;
3163 }
3164 }
3165
3166 /*
3167 * Propagate the depth clamp setting from the rasterizer state.
3168 * depth_clip == 0 implies depth clamping is enabled.
3169 *
3170 * When clip_halfz is enabled, then always clamp the depth values.
3171 *
3172 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3173 * clamp is always active in d3d10, regardless if depth clip is
3174 * enabled or not).
3175 * (GL has an always-on [0,1] clamp on fs depth output instead
3176 * to ensure the depth values stay in range. Doesn't look like
3177 * we do that, though...)
3178 */
3179 if (lp->rasterizer->clip_halfz) {
3180 key->depth_clamp = 1;
3181 } else {
3182 key->depth_clamp = (lp->rasterizer->depth_clip == 0) ? 1 : 0;
3183 }
3184
3185 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3186 if (!lp->framebuffer.nr_cbufs ||
3187 !lp->framebuffer.cbufs[0] ||
3188 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
3189 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
3190 }
3191 if(key->alpha.enabled)
3192 key->alpha.func = lp->depth_stencil->alpha.func;
3193 /* alpha.ref_value is passed in jit_context */
3194
3195 key->flatshade = lp->rasterizer->flatshade;
3196 if (lp->active_occlusion_queries) {
3197 key->occlusion_count = TRUE;
3198 }
3199
3200 if (lp->framebuffer.nr_cbufs) {
3201 memcpy(&key->blend, lp->blend, sizeof key->blend);
3202 }
3203
3204 key->nr_cbufs = lp->framebuffer.nr_cbufs;
3205
3206 if (!key->blend.independent_blend_enable) {
3207 /* we always need independent blend otherwise the fixups below won't work */
3208 for (i = 1; i < key->nr_cbufs; i++) {
3209 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
3210 }
3211 key->blend.independent_blend_enable = 1;
3212 }
3213
3214 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
3215 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
3216
3217 if (lp->framebuffer.cbufs[i]) {
3218 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
3219 const struct util_format_description *format_desc;
3220
3221 key->cbuf_format[i] = format;
3222
3223 /*
3224 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3225 * mixing of 2d textures with height 1 and 1d textures, so make sure
3226 * we pick 1d if any cbuf or zsbuf is 1d.
3227 */
3228 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
3229 key->resource_1d = TRUE;
3230 }
3231
3232 format_desc = util_format_description(format);
3233 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
3234 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
3235
3236 /*
3237 * Mask out color channels not present in the color buffer.
3238 */
3239 blend_rt->colormask &= util_format_colormask(format_desc);
3240
3241 /*
3242 * Disable blend for integer formats.
3243 */
3244 if (util_format_is_pure_integer(format)) {
3245 blend_rt->blend_enable = 0;
3246 }
3247
3248 /*
3249 * Our swizzled render tiles always have an alpha channel, but the
3250 * linear render target format often does not, so force here the dst
3251 * alpha to be one.
3252 *
3253 * This is not a mere optimization. Wrong results will be produced if
3254 * the dst alpha is used, the dst format does not have alpha, and the
3255 * previous rendering was not flushed from the swizzled to linear
3256 * buffer. For example, NonPowTwo DCT.
3257 *
3258 * TODO: This should be generalized to all channels for better
3259 * performance, but only alpha causes correctness issues.
3260 *
3261 * Also, force rgb/alpha func/factors match, to make AoS blending
3262 * easier.
3263 */
3264 if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
3265 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3266 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3267 boolean clamped_zero = !util_format_is_float(format) &&
3268 !util_format_is_snorm(format);
3269 blend_rt->rgb_src_factor =
3270 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3271 blend_rt->rgb_dst_factor =
3272 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3273 blend_rt->alpha_func = blend_rt->rgb_func;
3274 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3275 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3276 }
3277 }
3278 else {
3279 /* no color buffer for this fragment output */
3280 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3281 blend_rt->colormask = 0x0;
3282 blend_rt->blend_enable = 0;
3283 }
3284 }
3285
3286 /* This value will be the same for all the variants of a given shader:
3287 */
3288 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3289
3290 for(i = 0; i < key->nr_samplers; ++i) {
3291 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3292 lp_sampler_static_sampler_state(&key->state[i].sampler_state,
3293 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3294 }
3295 }
3296
3297 /*
3298 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3299 * are dx10-style? Can't really have mixed opcodes, at least not
3300 * if we want to skip the holes here (without rescanning tgsi).
3301 */
3302 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3303 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3304 for(i = 0; i < key->nr_sampler_views; ++i) {
3305 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
3306 lp_sampler_static_texture_state(&key->state[i].texture_state,
3307 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3308 }
3309 }
3310 }
3311 else {
3312 key->nr_sampler_views = key->nr_samplers;
3313 for(i = 0; i < key->nr_sampler_views; ++i) {
3314 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3315 lp_sampler_static_texture_state(&key->state[i].texture_state,
3316 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3317 }
3318 }
3319 }
3320 }
3321
3322
3323
3324 /**
3325 * Update fragment shader state. This is called just prior to drawing
3326 * something when some fragment-related state has changed.
3327 */
3328 void
llvmpipe_update_fs(struct llvmpipe_context * lp)3329 llvmpipe_update_fs(struct llvmpipe_context *lp)
3330 {
3331 struct lp_fragment_shader *shader = lp->fs;
3332 struct lp_fragment_shader_variant_key key;
3333 struct lp_fragment_shader_variant *variant = NULL;
3334 struct lp_fs_variant_list_item *li;
3335
3336 make_variant_key(lp, shader, &key);
3337
3338 /* Search the variants for one which matches the key */
3339 li = first_elem(&shader->variants);
3340 while(!at_end(&shader->variants, li)) {
3341 if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
3342 variant = li->base;
3343 break;
3344 }
3345 li = next_elem(li);
3346 }
3347
3348 if (variant) {
3349 /* Move this variant to the head of the list to implement LRU
3350 * deletion of shader's when we have too many.
3351 */
3352 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3353 }
3354 else {
3355 /* variant not found, create it now */
3356 int64_t t0, t1, dt;
3357 unsigned i;
3358 unsigned variants_to_cull;
3359
3360 if (0) {
3361 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3362 lp->nr_fs_variants,
3363 lp->nr_fs_instrs,
3364 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3365 }
3366
3367 /* First, check if we've exceeded the max number of shader variants.
3368 * If so, free 25% of them (the least recently used ones).
3369 */
3370 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
3371
3372 if (variants_to_cull ||
3373 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3374 struct pipe_context *pipe = &lp->pipe;
3375
3376 /*
3377 * XXX: we need to flush the context until we have some sort of
3378 * reference counting in fragment shaders as they may still be binned
3379 * Flushing alone might not be sufficient we need to wait on it too.
3380 */
3381 llvmpipe_finish(pipe, __FUNCTION__);
3382
3383 /*
3384 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3385 * number of shader variants (potentially all of them) could be
3386 * pending for destruction on flush.
3387 */
3388
3389 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3390 struct lp_fs_variant_list_item *item;
3391 if (is_empty_list(&lp->fs_variants_list)) {
3392 break;
3393 }
3394 item = last_elem(&lp->fs_variants_list);
3395 assert(item);
3396 assert(item->base);
3397 llvmpipe_remove_shader_variant(lp, item->base);
3398 }
3399 }
3400
3401 /*
3402 * Generate the new variant.
3403 */
3404 t0 = os_time_get();
3405 variant = generate_variant(lp, shader, &key);
3406 t1 = os_time_get();
3407 dt = t1 - t0;
3408 LP_COUNT_ADD(llvm_compile_time, dt);
3409 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3410
3411 /* Put the new variant into the list */
3412 if (variant) {
3413 insert_at_head(&shader->variants, &variant->list_item_local);
3414 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3415 lp->nr_fs_variants++;
3416 lp->nr_fs_instrs += variant->nr_instrs;
3417 shader->variants_cached++;
3418 }
3419 }
3420
3421 /* Bind this variant */
3422 lp_setup_set_fs_variant(lp->setup, variant);
3423 }
3424
3425
3426
3427
3428
3429 void
llvmpipe_init_fs_funcs(struct llvmpipe_context * llvmpipe)3430 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3431 {
3432 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3433 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3434 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3435
3436 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3437 }
3438
3439 /*
3440 * Rasterization is disabled if there is no pixel shader and
3441 * both depth and stencil testing are disabled:
3442 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3443 */
3444 boolean
llvmpipe_rasterization_disabled(struct llvmpipe_context * lp)3445 llvmpipe_rasterization_disabled(struct llvmpipe_context *lp)
3446 {
3447 boolean null_fs = !lp->fs || lp->fs->info.base.num_tokens <= 1;
3448
3449 return (null_fs &&
3450 !lp->depth_stencil->depth.enabled &&
3451 !lp->depth_stencil->stencil[0].enabled);
3452 }
3453