projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: disable opaque variant for multisample
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_fs.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 VMware, Inc.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29 /**
30 * @file
31 * Code generate the whole fragment pipeline.
32 *
33 * The fragment pipeline consists of the following stages:
34 * - early depth test
35 * - fragment shader
36 * - alpha test
37 * - depth/stencil test
38 * - blending
39 *
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
45 *
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
51 *
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
56 *
57 * @author Jose Fonseca <jfonseca@vmware.com>
58 */
59
60 #include <limits.h>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/format/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 "util/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_nir.h"
84 #include "gallivm/lp_bld_swizzle.h"
85 #include "gallivm/lp_bld_flow.h"
86 #include "gallivm/lp_bld_debug.h"
87 #include "gallivm/lp_bld_arit.h"
88 #include "gallivm/lp_bld_bitarit.h"
89 #include "gallivm/lp_bld_pack.h"
90 #include "gallivm/lp_bld_format.h"
91 #include "gallivm/lp_bld_quad.h"
92
93 #include "lp_bld_alpha.h"
94 #include "lp_bld_blend.h"
95 #include "lp_bld_depth.h"
96 #include "lp_bld_interp.h"
97 #include "lp_context.h"
98 #include "lp_debug.h"
99 #include "lp_perf.h"
100 #include "lp_setup.h"
101 #include "lp_state.h"
102 #include "lp_tex_sample.h"
103 #include "lp_flush.h"
104 #include "lp_state_fs.h"
105 #include "lp_rast.h"
106 #include "nir/nir_to_tgsi_info.h"
107
108 /** Fragment shader number (for debugging) */
109 static unsigned fs_no = 0;
110
111
112 /**
113 * Expand the relevant bits of mask_input to a n*4-dword mask for the
114 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
115 * quad mask vector to 0 or ~0.
116 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
117 * quad arguments with fs length 8.
118 *
119 * \param first_quad which quad(s) of the quad group to test, in [0,3]
120 * \param mask_input bitwise mask for the whole 4x4 stamp
121 */
122 static LLVMValueRef
123 generate_quad_mask(struct gallivm_state *gallivm,
124 struct lp_type fs_type,
125 unsigned first_quad,
126 unsigned sample,
127 LLVMValueRef mask_input) /* int64 */
128 {
129 LLVMBuilderRef builder = gallivm->builder;
130 struct lp_type mask_type;
131 LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
132 LLVMValueRef bits[16];
133 LLVMValueRef mask, bits_vec;
134 int shift, i;
135
136 /*
137 * XXX: We'll need a different path for 16 x u8
138 */
139 assert(fs_type.width == 32);
140 assert(fs_type.length <= ARRAY_SIZE(bits));
141 mask_type = lp_int_type(fs_type);
142
143 /*
144 * mask_input >>= (quad * 4)
145 */
146 switch (first_quad) {
147 case 0:
148 shift = 0;
149 break;
150 case 1:
151 assert(fs_type.length == 4);
152 shift = 2;
153 break;
154 case 2:
155 shift = 8;
156 break;
157 case 3:
158 assert(fs_type.length == 4);
159 shift = 10;
160 break;
161 default:
162 assert(0);
163 shift = 0;
164 }
165
166 mask_input = LLVMBuildLShr(builder, mask_input, lp_build_const_int64(gallivm, 16 * sample), "");
167 mask_input = LLVMBuildTrunc(builder, mask_input,
168 i32t, "");
169 mask_input = LLVMBuildAnd(builder, mask_input, lp_build_const_int32(gallivm, 0xffff), "");
170
171 mask_input = LLVMBuildLShr(builder,
172 mask_input,
173 LLVMConstInt(i32t, shift, 0),
174 "");
175
176 /*
177 * mask = { mask_input & (1 << i), for i in [0,3] }
178 */
179 mask = lp_build_broadcast(gallivm,
180 lp_build_vec_type(gallivm, mask_type),
181 mask_input);
182
183 for (i = 0; i < fs_type.length / 4; i++) {
184 unsigned j = 2 * (i % 2) + (i / 2) * 8;
185 bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
186 bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
187 bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
188 bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
189 }
190 bits_vec = LLVMConstVector(bits, fs_type.length);
191 mask = LLVMBuildAnd(builder, mask, bits_vec, "");
192
193 /*
194 * mask = mask == bits ? ~0 : 0
195 */
196 mask = lp_build_compare(gallivm,
197 mask_type, PIPE_FUNC_EQUAL,
198 mask, bits_vec);
199
200 return mask;
201 }
202
203
204 #define EARLY_DEPTH_TEST 0x1
205 #define LATE_DEPTH_TEST 0x2
206 #define EARLY_DEPTH_WRITE 0x4
207 #define LATE_DEPTH_WRITE 0x8
208
209 static int
210 find_output_by_semantic( const struct tgsi_shader_info *info,
211 unsigned semantic,
212 unsigned index )
213 {
214 int i;
215
216 for (i = 0; i < info->num_outputs; i++)
217 if (info->output_semantic_name[i] == semantic &&
218 info->output_semantic_index[i] == index)
219 return i;
220
221 return -1;
222 }
223
224
225 /**
226 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
227 */
228 static LLVMValueRef
229 lp_llvm_viewport(LLVMValueRef context_ptr,
230 struct gallivm_state *gallivm,
231 LLVMValueRef viewport_index)
232 {
233 LLVMBuilderRef builder = gallivm->builder;
234 LLVMValueRef ptr;
235 LLVMValueRef res;
236 struct lp_type viewport_type =
237 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
238
239 ptr = lp_jit_context_viewports(gallivm, context_ptr);
240 ptr = LLVMBuildPointerCast(builder, ptr,
241 LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
242
243 res = lp_build_pointer_get(builder, ptr, viewport_index);
244
245 return res;
246 }
247
248
249 static LLVMValueRef
250 lp_build_depth_clamp(struct gallivm_state *gallivm,
251 LLVMBuilderRef builder,
252 struct lp_type type,
253 LLVMValueRef context_ptr,
254 LLVMValueRef thread_data_ptr,
255 LLVMValueRef z)
256 {
257 LLVMValueRef viewport, min_depth, max_depth;
258 LLVMValueRef viewport_index;
259 struct lp_build_context f32_bld;
260
261 assert(type.floating);
262 lp_build_context_init(&f32_bld, gallivm, type);
263
264 /*
265 * Assumes clamping of the viewport index will occur in setup/gs. Value
266 * is passed through the rasterization stage via lp_rast_shader_inputs.
267 *
268 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
269 * semantics.
270 */
271 viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
272 thread_data_ptr);
273
274 /*
275 * Load the min and max depth from the lp_jit_context.viewports
276 * array of lp_jit_viewport structures.
277 */
278 viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
279
280 /* viewports[viewport_index].min_depth */
281 min_depth = LLVMBuildExtractElement(builder, viewport,
282 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
283 min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
284
285 /* viewports[viewport_index].max_depth */
286 max_depth = LLVMBuildExtractElement(builder, viewport,
287 lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
288 max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
289
290 /*
291 * Clamp to the min and max depth values for the given viewport.
292 */
293 return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
294 }
295
296 static void
297 lp_build_sample_alpha_to_coverage(struct gallivm_state *gallivm,
298 struct lp_type type,
299 unsigned coverage_samples,
300 LLVMValueRef num_loop,
301 LLVMValueRef loop_counter,
302 LLVMValueRef coverage_mask_store,
303 LLVMValueRef alpha)
304 {
305 struct lp_build_context bld;
306 LLVMBuilderRef builder = gallivm->builder;
307 float step = 1.0 / coverage_samples;
308
309 lp_build_context_init(&bld, gallivm, type);
310 for (unsigned s = 0; s < coverage_samples; s++) {
311 LLVMValueRef alpha_ref_value = lp_build_const_vec(gallivm, type, step * s);
312 LLVMValueRef test = lp_build_cmp(&bld, PIPE_FUNC_GREATER, alpha, alpha_ref_value);
313
314 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, lp_build_const_int32(gallivm, s), num_loop, "");
315 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_counter, "");
316 LLVMValueRef s_mask_ptr = LLVMBuildGEP(builder, coverage_mask_store, &s_mask_idx, 1, "");
317 LLVMValueRef s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
318 s_mask = LLVMBuildAnd(builder, s_mask, test, "");
319 LLVMBuildStore(builder, s_mask, s_mask_ptr);
320 }
321 }
322
323 /**
324 * Generate the fragment shader, depth/stencil test, and alpha tests.
325 */
326 static void
327 generate_fs_loop(struct gallivm_state *gallivm,
328 struct lp_fragment_shader *shader,
329 const struct lp_fragment_shader_variant_key *key,
330 LLVMBuilderRef builder,
331 struct lp_type type,
332 LLVMValueRef context_ptr,
333 LLVMValueRef sample_pos_array,
334 LLVMValueRef num_loop,
335 struct lp_build_interp_soa_context *interp,
336 const struct lp_build_sampler_soa *sampler,
337 const struct lp_build_image_soa *image,
338 LLVMValueRef mask_store,
339 LLVMValueRef (*out_color)[4],
340 LLVMValueRef depth_base_ptr,
341 LLVMValueRef depth_stride,
342 LLVMValueRef depth_sample_stride,
343 LLVMValueRef facing,
344 LLVMValueRef thread_data_ptr)
345 {
346 const struct util_format_description *zs_format_desc = NULL;
347 const struct tgsi_token *tokens = shader->base.tokens;
348 struct lp_type int_type = lp_int_type(type);
349 LLVMTypeRef vec_type, int_vec_type;
350 LLVMValueRef mask_ptr = NULL, mask_val = NULL;
351 LLVMValueRef consts_ptr, num_consts_ptr;
352 LLVMValueRef ssbo_ptr, num_ssbo_ptr;
353 LLVMValueRef z;
354 LLVMValueRef z_value, s_value;
355 LLVMValueRef z_fb, s_fb;
356 LLVMValueRef depth_ptr;
357 LLVMValueRef stencil_refs[2];
358 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
359 LLVMValueRef zs_samples = lp_build_const_int32(gallivm, key->zsbuf_nr_samples);
360 struct lp_build_for_loop_state loop_state, sample_loop_state;
361 struct lp_build_mask_context mask;
362 /*
363 * TODO: figure out if simple_shader optimization is really worthwile to
364 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
365 * code since tests tend to take another codepath than real shaders.
366 */
367 boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
368 shader->info.base.num_inputs < 3 &&
369 shader->info.base.num_instructions < 8) && 0;
370 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
371 util_blend_state_is_dual(&key->blend, 0);
372 unsigned attrib;
373 unsigned chan;
374 unsigned cbuf;
375 unsigned depth_mode;
376
377 struct lp_bld_tgsi_system_values system_values;
378
379 memset(&system_values, 0, sizeof(system_values));
380
381 /* truncate then sign extend. */
382 system_values.front_facing = LLVMBuildTrunc(gallivm->builder, facing, LLVMInt1TypeInContext(gallivm->context), "");
383 system_values.front_facing = LLVMBuildSExt(gallivm->builder, system_values.front_facing, LLVMInt32TypeInContext(gallivm->context), "");
384
385 if (key->depth.enabled ||
386 key->stencil[0].enabled) {
387
388 zs_format_desc = util_format_description(key->zsbuf_format);
389 assert(zs_format_desc);
390
391 if (shader->info.base.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL])
392 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
393 else if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
394 if (shader->info.base.writes_memory)
395 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
396 else if (key->alpha.enabled ||
397 key->blend.alpha_to_coverage ||
398 shader->info.base.uses_kill ||
399 shader->info.base.writes_samplemask) {
400 /* With alpha test and kill, can do the depth test early
401 * and hopefully eliminate some quads. But need to do a
402 * special deferred depth write once the final mask value
403 * is known. This only works though if there's either no
404 * stencil test or the stencil value isn't written.
405 */
406 if (key->stencil[0].enabled && (key->stencil[0].writemask ||
407 (key->stencil[1].enabled &&
408 key->stencil[1].writemask)))
409 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
410 else
411 depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
412 }
413 else
414 depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
415 }
416 else {
417 depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
418 }
419
420 if (!(key->depth.enabled && key->depth.writemask) &&
421 !(key->stencil[0].enabled && (key->stencil[0].writemask ||
422 (key->stencil[1].enabled &&
423 key->stencil[1].writemask))))
424 depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
425 }
426 else {
427 depth_mode = 0;
428 }
429
430 vec_type = lp_build_vec_type(gallivm, type);
431 int_vec_type = lp_build_vec_type(gallivm, int_type);
432
433 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
434 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
435 /* convert scalar stencil refs into vectors */
436 stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
437 stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
438
439 consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
440 num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
441
442 ssbo_ptr = lp_jit_context_ssbos(gallivm, context_ptr);
443 num_ssbo_ptr = lp_jit_context_num_ssbos(gallivm, context_ptr);
444
445 memset(outputs, 0, sizeof outputs);
446
447 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
448 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
449 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
450 lp_build_vec_type(gallivm,
451 type),
452 num_loop, "color");
453 }
454 }
455 if (dual_source_blend) {
456 assert(key->nr_cbufs <= 1);
457 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
458 out_color[1][chan] = lp_build_array_alloca(gallivm,
459 lp_build_vec_type(gallivm,
460 type),
461 num_loop, "color1");
462 }
463 }
464
465 lp_build_for_loop_begin(&loop_state, gallivm,
466 lp_build_const_int32(gallivm, 0),
467 LLVMIntULT,
468 num_loop,
469 lp_build_const_int32(gallivm, 1));
470
471 if (key->multisample) {
472 /* create shader execution mask by combining all sample masks. */
473 for (unsigned s = 0; s < key->coverage_samples; s++) {
474 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, s), "");
475 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
476 LLVMValueRef s_mask = lp_build_pointer_get(builder, mask_store, s_mask_idx);
477 if (s == 0)
478 mask_val = s_mask;
479 else
480 mask_val = LLVMBuildOr(builder, s_mask, mask_val, "");
481 }
482 } else {
483 mask_ptr = LLVMBuildGEP(builder, mask_store,
484 &loop_state.counter, 1, "mask_ptr");
485 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
486 }
487
488 /* 'mask' will control execution based on quad's pixel alive/killed state */
489 lp_build_mask_begin(&mask, gallivm, type, mask_val);
490
491 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
492 lp_build_mask_check(&mask);
493
494 /* Create storage for recombining sample masks after early Z pass. */
495 LLVMValueRef s_mask_or = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, type), "cov_mask_early_depth");
496 LLVMBuildStore(builder, LLVMConstNull(lp_build_int_vec_type(gallivm, type)), s_mask_or);
497
498 LLVMValueRef s_mask = NULL, s_mask_ptr = NULL;
499 LLVMValueRef z_sample_value_store = NULL, s_sample_value_store = NULL;
500 LLVMValueRef z_fb_store = NULL, s_fb_store = NULL;
501 LLVMTypeRef z_type = NULL, z_fb_type = NULL;
502
503 /* Run early depth once per sample */
504 if (key->multisample) {
505
506 if (zs_format_desc) {
507 struct lp_type zs_type = lp_depth_type(zs_format_desc, type.length);
508 struct lp_type z_type = zs_type;
509 struct lp_type s_type = zs_type;
510 if (zs_format_desc->block.bits < type.width)
511 z_type.width = type.width;
512 else if (zs_format_desc->block.bits > 32) {
513 z_type.width = z_type.width / 2;
514 s_type.width = s_type.width / 2;
515 s_type.floating = 0;
516 }
517 z_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
518 zs_samples, "z_sample_store");
519 s_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
520 zs_samples, "s_sample_store");
521 z_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, z_type),
522 zs_samples, "z_fb_store");
523 s_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, s_type),
524 zs_samples, "s_fb_store");
525 }
526 lp_build_for_loop_begin(&sample_loop_state, gallivm,
527 lp_build_const_int32(gallivm, 0),
528 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
529 lp_build_const_int32(gallivm, 1));
530
531 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
532 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
533 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
534
535 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
536 s_mask = LLVMBuildAnd(builder, s_mask, mask_val, "");
537 }
538
539
540 /* for multisample Z needs to be interpolated at sample points for testing. */
541 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);
542 z = interp->pos[2];
543
544 depth_ptr = depth_base_ptr;
545 if (key->multisample) {
546 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
547 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
548 }
549
550 if (depth_mode & EARLY_DEPTH_TEST) {
551 /*
552 * Clamp according to ARB_depth_clamp semantics.
553 */
554 if (key->depth_clamp) {
555 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
556 thread_data_ptr, z);
557 }
558 lp_build_depth_stencil_load_swizzled(gallivm, type,
559 zs_format_desc, key->resource_1d,
560 depth_ptr, depth_stride,
561 &z_fb, &s_fb, loop_state.counter);
562 lp_build_depth_stencil_test(gallivm,
563 &key->depth,
564 key->stencil,
565 type,
566 zs_format_desc,
567 key->multisample ? NULL : &mask,
568 &s_mask,
569 stencil_refs,
570 z, z_fb, s_fb,
571 facing,
572 &z_value, &s_value,
573 !simple_shader && !key->multisample);
574
575 if (depth_mode & EARLY_DEPTH_WRITE) {
576 lp_build_depth_stencil_write_swizzled(gallivm, type,
577 zs_format_desc, key->resource_1d,
578 NULL, NULL, NULL, loop_state.counter,
579 depth_ptr, depth_stride,
580 z_value, s_value);
581 }
582 /*
583 * Note mask check if stencil is enabled must be after ds write not after
584 * stencil test otherwise new stencil values may not get written if all
585 * fragments got killed by depth/stencil test.
586 */
587 if (!simple_shader && key->stencil[0].enabled && !key->multisample)
588 lp_build_mask_check(&mask);
589
590 if (key->multisample) {
591 z_fb_type = LLVMTypeOf(z_fb);
592 z_type = LLVMTypeOf(z_value);
593 lp_build_pointer_set(builder, z_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, z_value, lp_build_int_vec_type(gallivm, type), ""));
594 lp_build_pointer_set(builder, s_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, s_value, lp_build_int_vec_type(gallivm, type), ""));
595 lp_build_pointer_set(builder, z_fb_store, sample_loop_state.counter, z_fb);
596 lp_build_pointer_set(builder, s_fb_store, sample_loop_state.counter, s_fb);
597 }
598 }
599
600 if (key->multisample) {
601 /*
602 * Store the post-early Z coverage mask.
603 * Recombine the resulting coverage masks post early Z into the fragment
604 * shader execution mask.
605 */
606 LLVMValueRef tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
607 tmp_s_mask_or = LLVMBuildOr(builder, tmp_s_mask_or, s_mask, "");
608 LLVMBuildStore(builder, tmp_s_mask_or, s_mask_or);
609
610 LLVMBuildStore(builder, s_mask, s_mask_ptr);
611
612 lp_build_for_loop_end(&sample_loop_state);
613
614 /* recombined all the coverage masks in the shader exec mask. */
615 tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
616 lp_build_mask_update(&mask, tmp_s_mask_or);
617
618 /* for multisample Z needs to be re interpolated at pixel center */
619 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);
620 }
621
622 LLVMValueRef out_sample_mask_storage = NULL;
623 if (shader->info.base.writes_samplemask) {
624 out_sample_mask_storage = lp_build_alloca(gallivm, int_vec_type, "write_mask");
625 }
626 system_values.sample_pos = sample_pos_array;
627
628 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, mask_store, NULL);
629
630 struct lp_build_tgsi_params params;
631 memset(&params, 0, sizeof(params));
632
633 params.type = type;
634 params.mask = &mask;
635 params.consts_ptr = consts_ptr;
636 params.const_sizes_ptr = num_consts_ptr;
637 params.system_values = &system_values;
638 params.inputs = interp->inputs;
639 params.context_ptr = context_ptr;
640 params.thread_data_ptr = thread_data_ptr;
641 params.sampler = sampler;
642 params.info = &shader->info.base;
643 params.ssbo_ptr = ssbo_ptr;
644 params.ssbo_sizes_ptr = num_ssbo_ptr;
645 params.image = image;
646
647 /* Build the actual shader */
648 if (shader->base.type == PIPE_SHADER_IR_TGSI)
649 lp_build_tgsi_soa(gallivm, tokens, &params,
650 outputs);
651 else
652 lp_build_nir_soa(gallivm, shader->base.ir.nir, &params,
653 outputs);
654
655 /* Alpha test */
656 if (key->alpha.enabled) {
657 int color0 = find_output_by_semantic(&shader->info.base,
658 TGSI_SEMANTIC_COLOR,
659 0);
660
661 if (color0 != -1 && outputs[color0][3]) {
662 const struct util_format_description *cbuf_format_desc;
663 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
664 LLVMValueRef alpha_ref_value;
665
666 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
667 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
668
669 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
670
671 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
672 &mask, alpha, alpha_ref_value,
673 (depth_mode & LATE_DEPTH_TEST) != 0);
674 }
675 }
676
677 /* Emulate Alpha to Coverage with Alpha test */
678 if (key->blend.alpha_to_coverage) {
679 int color0 = find_output_by_semantic(&shader->info.base,
680 TGSI_SEMANTIC_COLOR,
681 0);
682
683 if (color0 != -1 && outputs[color0][3]) {
684 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
685
686 if (!key->multisample) {
687 lp_build_alpha_to_coverage(gallivm, type,
688 &mask, alpha,
689 (depth_mode & LATE_DEPTH_TEST) != 0);
690 } else {
691 lp_build_sample_alpha_to_coverage(gallivm, type, key->coverage_samples, num_loop,
692 loop_state.counter,
693 mask_store, alpha);
694 }
695 }
696 }
697 if (key->blend.alpha_to_one && key->multisample) {
698 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib) {
699 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
700 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
701 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
702 if (outputs[cbuf][3]) {
703 LLVMBuildStore(builder, lp_build_const_vec(gallivm, type, 1.0), outputs[cbuf][3]);
704 }
705 }
706 }
707 if (shader->info.base.writes_samplemask) {
708 LLVMValueRef output_smask = NULL;
709 int smaski = find_output_by_semantic(&shader->info.base,
710 TGSI_SEMANTIC_SAMPLEMASK,
711 0);
712 struct lp_build_context smask_bld;
713 lp_build_context_init(&smask_bld, gallivm, int_type);
714
715 assert(smaski >= 0);
716 output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
717 /*
718 * Pixel is alive according to the first sample in the mask.
719 */
720 output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");
721 if (!key->multisample) {
722 output_smask = lp_build_and(&smask_bld, output_smask, smask_bld.one);
723 output_smask = lp_build_cmp(&smask_bld, PIPE_FUNC_NOTEQUAL, output_smask, smask_bld.zero);
724 lp_build_mask_update(&mask, output_smask);
725 }
726 LLVMBuildStore(builder, output_smask, out_sample_mask_storage);
727 }
728
729 /* Color write */
730 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
731 {
732 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
733 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
734 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
735 {
736 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
737 if(outputs[attrib][chan]) {
738 /* XXX: just initialize outputs to point at colors[] and
739 * skip this.
740 */
741 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
742 LLVMValueRef color_ptr;
743 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
744 &loop_state.counter, 1, "");
745 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
746 LLVMBuildStore(builder, out, color_ptr);
747 }
748 }
749 }
750 }
751
752
753 if (key->multisample) {
754 /* execute depth test for each sample */
755 lp_build_for_loop_begin(&sample_loop_state, gallivm,
756 lp_build_const_int32(gallivm, 0),
757 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
758 lp_build_const_int32(gallivm, 1));
759
760 /* load the per-sample coverage mask */
761 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
762 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
763 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
764
765 /* combine the execution mask post fragment shader with the coverage mask. */
766 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
767 s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
768
769 /* if the shader writes sample mask use that */
770 if (shader->info.base.writes_samplemask) {
771 LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
772 out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);
773 LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");
774 LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");
775 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");
776 smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");
777
778 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
779 }
780 }
781
782 depth_ptr = depth_base_ptr;
783 if (key->multisample) {
784 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
785 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
786 }
787
788 /* Late Z test */
789 if (depth_mode & LATE_DEPTH_TEST) {
790 int pos0 = find_output_by_semantic(&shader->info.base,
791 TGSI_SEMANTIC_POSITION,
792 0);
793 int s_out = find_output_by_semantic(&shader->info.base,
794 TGSI_SEMANTIC_STENCIL,
795 0);
796 if (pos0 != -1 && outputs[pos0][2]) {
797 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
798 }
799 /*
800 * Clamp according to ARB_depth_clamp semantics.
801 */
802 if (key->depth_clamp) {
803 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
804 thread_data_ptr, z);
805 }
806
807 if (s_out != -1 && outputs[s_out][1]) {
808 /* there's only one value, and spec says to discard additional bits */
809 LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
810 stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
811 stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
812 stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
813 stencil_refs[1] = stencil_refs[0];
814 }
815
816 lp_build_depth_stencil_load_swizzled(gallivm, type,
817 zs_format_desc, key->resource_1d,
818 depth_ptr, depth_stride,
819 &z_fb, &s_fb, loop_state.counter);
820
821 lp_build_depth_stencil_test(gallivm,
822 &key->depth,
823 key->stencil,
824 type,
825 zs_format_desc,
826 key->multisample ? NULL : &mask,
827 &s_mask,
828 stencil_refs,
829 z, z_fb, s_fb,
830 facing,
831 &z_value, &s_value,
832 !simple_shader);
833 /* Late Z write */
834 if (depth_mode & LATE_DEPTH_WRITE) {
835 lp_build_depth_stencil_write_swizzled(gallivm, type,
836 zs_format_desc, key->resource_1d,
837 NULL, NULL, NULL, loop_state.counter,
838 depth_ptr, depth_stride,
839 z_value, s_value);
840 }
841 }
842 else if ((depth_mode & EARLY_DEPTH_TEST) &&
843 (depth_mode & LATE_DEPTH_WRITE))
844 {
845 /* Need to apply a reduced mask to the depth write. Reload the
846 * depth value, update from zs_value with the new mask value and
847 * write that out.
848 */
849 if (key->multisample) {
850 z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;
851 s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);
852 z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");
853 s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);
854 }
855 lp_build_depth_stencil_write_swizzled(gallivm, type,
856 zs_format_desc, key->resource_1d,
857 key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,
858 depth_ptr, depth_stride,
859 z_value, s_value);
860 }
861
862 if (key->occlusion_count) {
863 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
864 lp_build_name(counter, "counter");
865
866 lp_build_occlusion_count(gallivm, type,
867 key->multisample ? s_mask : lp_build_mask_value(&mask), counter);
868 }
869
870 if (key->multisample) {
871 /* store the sample mask for this loop */
872 LLVMBuildStore(builder, s_mask, s_mask_ptr);
873 lp_build_for_loop_end(&sample_loop_state);
874 }
875
876 mask_val = lp_build_mask_end(&mask);
877 if (!key->multisample)
878 LLVMBuildStore(builder, mask_val, mask_ptr);
879 lp_build_for_loop_end(&loop_state);
880 }
881
882
883 /**
884 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
885 *
886 * Fragment Shader outputs pixels in small 2x2 blocks
887 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
888 *
889 * However in memory pixels are stored in rows
890 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
891 *
892 * @param type fragment shader type (4x or 8x float)
893 * @param num_fs number of fs_src
894 * @param is_1d whether we're outputting to a 1d resource
895 * @param dst_channels number of output channels
896 * @param fs_src output from fragment shader
897 * @param dst pointer to store result
898 * @param pad_inline is channel padding inline or at end of row
899 * @return the number of dsts
900 */
901 static int
902 generate_fs_twiddle(struct gallivm_state *gallivm,
903 struct lp_type type,
904 unsigned num_fs,
905 unsigned dst_channels,
906 LLVMValueRef fs_src[][4],
907 LLVMValueRef* dst,
908 bool pad_inline)
909 {
910 LLVMValueRef src[16];
911
912 bool swizzle_pad;
913 bool twiddle;
914 bool split;
915
916 unsigned pixels = type.length / 4;
917 unsigned reorder_group;
918 unsigned src_channels;
919 unsigned src_count;
920 unsigned i;
921
922 src_channels = dst_channels < 3 ? dst_channels : 4;
923 src_count = num_fs * src_channels;
924
925 assert(pixels == 2 || pixels == 1);
926 assert(num_fs * src_channels <= ARRAY_SIZE(src));
927
928 /*
929 * Transpose from SoA -> AoS
930 */
931 for (i = 0; i < num_fs; ++i) {
932 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
933 }
934
935 /*
936 * Pick transformation options
937 */
938 swizzle_pad = false;
939 twiddle = false;
940 split = false;
941 reorder_group = 0;
942
943 if (dst_channels == 1) {
944 twiddle = true;
945
946 if (pixels == 2) {
947 split = true;
948 }
949 } else if (dst_channels == 2) {
950 if (pixels == 1) {
951 reorder_group = 1;
952 }
953 } else if (dst_channels > 2) {
954 if (pixels == 1) {
955 reorder_group = 2;
956 } else {
957 twiddle = true;
958 }
959
960 if (!pad_inline && dst_channels == 3 && pixels > 1) {
961 swizzle_pad = true;
962 }
963 }
964
965 /*
966 * Split the src in half
967 */
968 if (split) {
969 for (i = num_fs; i > 0; --i) {
970 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
971 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
972 }
973
974 src_count *= 2;
975 type.length = 4;
976 }
977
978 /*
979 * Ensure pixels are in memory order
980 */
981 if (reorder_group) {
982 /* Twiddle pixels by reordering the array, e.g.:
983 *
984 * src_count = 8 -> 0 2 1 3 4 6 5 7
985 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
986 */
987 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
988
989 for (i = 0; i < src_count; ++i) {
990 unsigned group = i / reorder_group;
991 unsigned block = (group / 4) * 4 * reorder_group;
992 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
993 dst[i] = src[j];
994 }
995 } else if (twiddle) {
996 /* Twiddle pixels across elements of array */
997 /*
998 * XXX: we should avoid this in some cases, but would need to tell
999 * lp_build_conv to reorder (or deal with it ourselves).
1000 */
1001 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
1002 } else {
1003 /* Do nothing */
1004 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
1005 }
1006
1007 /*
1008 * Moves any padding between pixels to the end
1009 * e.g. RGBXRGBX -> RGBRGBXX
1010 */
1011 if (swizzle_pad) {
1012 unsigned char swizzles[16];
1013 unsigned elems = pixels * dst_channels;
1014
1015 for (i = 0; i < type.length; ++i) {
1016 if (i < elems)
1017 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
1018 else
1019 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
1020 }
1021
1022 for (i = 0; i < src_count; ++i) {
1023 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
1024 }
1025 }
1026
1027 return src_count;
1028 }
1029
1030
1031 /*
1032 * Untwiddle and transpose, much like the above.
1033 * However, this is after conversion, so we get packed vectors.
1034 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1035 * the vectors will look like:
1036 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1037 * be swizzled here). Extending to 16bit should be trivial.
1038 * Should also be extended to handle twice wide vectors with AVX2...
1039 */
1040 static void
1041 fs_twiddle_transpose(struct gallivm_state *gallivm,
1042 struct lp_type type,
1043 LLVMValueRef *src,
1044 unsigned src_count,
1045 LLVMValueRef *dst)
1046 {
1047 unsigned i, j;
1048 struct lp_type type64, type16, type32;
1049 LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
1050 LLVMBuilderRef builder = gallivm->builder;
1051 LLVMValueRef tmp[4], shuf[8];
1052 for (j = 0; j < 2; j++) {
1053 shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
1054 shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
1055 shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
1056 shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
1057 }
1058
1059 assert(src_count == 4 || src_count == 2 || src_count == 1);
1060 assert(type.width == 8);
1061 assert(type.length == 16);
1062
1063 type8_t = lp_build_vec_type(gallivm, type);
1064
1065 type64 = type;
1066 type64.length /= 8;
1067 type64.width *= 8;
1068 type64_t = lp_build_vec_type(gallivm, type64);
1069
1070 type16 = type;
1071 type16.length /= 2;
1072 type16.width *= 2;
1073 type16_t = lp_build_vec_type(gallivm, type16);
1074
1075 type32 = type;
1076 type32.length /= 4;
1077 type32.width *= 4;
1078 type32_t = lp_build_vec_type(gallivm, type32);
1079
1080 lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
1081
1082 if (src_count == 1) {
1083 /* transpose was no-op, just untwiddle */
1084 LLVMValueRef shuf_vec;
1085 shuf_vec = LLVMConstVector(shuf, 8);
1086 tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
1087 tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
1088 dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
1089 } else if (src_count == 2) {
1090 LLVMValueRef shuf_vec;
1091 shuf_vec = LLVMConstVector(shuf, 4);
1092
1093 for (i = 0; i < 2; i++) {
1094 tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
1095 tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
1096 dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
1097 }
1098 } else {
1099 for (j = 0; j < 2; j++) {
1100 LLVMValueRef lo, hi, lo2, hi2;
1101 /*
1102 * Note that if we only really have 3 valid channels (rgb)
1103 * and we don't need alpha we could substitute a undef here
1104 * for the respective channel (causing llvm to drop conversion
1105 * for alpha).
1106 */
1107 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1108 lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
1109 hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
1110 lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
1111 hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
1112 dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
1113 dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
1114 }
1115 }
1116 }
1117
1118
1119 /**
1120 * Load an unswizzled block of pixels from memory
1121 */
1122 static void
1123 load_unswizzled_block(struct gallivm_state *gallivm,
1124 LLVMValueRef base_ptr,
1125 LLVMValueRef stride,
1126 unsigned block_width,
1127 unsigned block_height,
1128 LLVMValueRef* dst,
1129 struct lp_type dst_type,
1130 unsigned dst_count,
1131 unsigned dst_alignment)
1132 {
1133 LLVMBuilderRef builder = gallivm->builder;
1134 unsigned row_size = dst_count / block_height;
1135 unsigned i;
1136
1137 /* Ensure block exactly fits into dst */
1138 assert((block_width * block_height) % dst_count == 0);
1139
1140 for (i = 0; i < dst_count; ++i) {
1141 unsigned x = i % row_size;
1142 unsigned y = i / row_size;
1143
1144 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
1145 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1146
1147 LLVMValueRef gep[2];
1148 LLVMValueRef dst_ptr;
1149
1150 gep[0] = lp_build_const_int32(gallivm, 0);
1151 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1152
1153 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1154 dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
1155 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
1156
1157 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
1158
1159 LLVMSetAlignment(dst[i], dst_alignment);
1160 }
1161 }
1162
1163
1164 /**
1165 * Store an unswizzled block of pixels to memory
1166 */
1167 static void
1168 store_unswizzled_block(struct gallivm_state *gallivm,
1169 LLVMValueRef base_ptr,
1170 LLVMValueRef stride,
1171 unsigned block_width,
1172 unsigned block_height,
1173 LLVMValueRef* src,
1174 struct lp_type src_type,
1175 unsigned src_count,
1176 unsigned src_alignment)
1177 {
1178 LLVMBuilderRef builder = gallivm->builder;
1179 unsigned row_size = src_count / block_height;
1180 unsigned i;
1181
1182 /* Ensure src exactly fits into block */
1183 assert((block_width * block_height) % src_count == 0);
1184
1185 for (i = 0; i < src_count; ++i) {
1186 unsigned x = i % row_size;
1187 unsigned y = i / row_size;
1188
1189 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
1190 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1191
1192 LLVMValueRef gep[2];
1193 LLVMValueRef src_ptr;
1194
1195 gep[0] = lp_build_const_int32(gallivm, 0);
1196 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1197
1198 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1199 src_ptr = LLVMBuildBitCast(builder, src_ptr,
1200 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
1201
1202 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
1203
1204 LLVMSetAlignment(src_ptr, src_alignment);
1205 }
1206 }
1207
1208
1209 /**
1210 * Checks if a format description is an arithmetic format
1211 *
1212 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1213 */
1214 static inline boolean
1215 is_arithmetic_format(const struct util_format_description *format_desc)
1216 {
1217 boolean arith = false;
1218 unsigned i;
1219
1220 for (i = 0; i < format_desc->nr_channels; ++i) {
1221 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
1222 arith |= (format_desc->channel[i].size % 8) != 0;
1223 }
1224
1225 return arith;
1226 }
1227
1228
1229 /**
1230 * Checks if this format requires special handling due to required expansion
1231 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1232 * SoA conversion.
1233 */
1234 static inline boolean
1235 format_expands_to_float_soa(const struct util_format_description *format_desc)
1236 {
1237 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
1238 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1239 return true;
1240 }
1241 return false;
1242 }
1243
1244
1245 /**
1246 * Retrieves the type representing the memory layout for a format
1247 *
1248 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1249 */
1250 static inline void
1251 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
1252 struct lp_type* type)
1253 {
1254 unsigned i;
1255 unsigned chan;
1256
1257 if (format_expands_to_float_soa(format_desc)) {
1258 /* just make this a uint with width of block */
1259 type->floating = false;
1260 type->fixed = false;
1261 type->sign = false;
1262 type->norm = false;
1263 type->width = format_desc->block.bits;
1264 type->length = 1;
1265 return;
1266 }
1267
1268 for (i = 0; i < 4; i++)
1269 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1270 break;
1271 chan = i;
1272
1273 memset(type, 0, sizeof(struct lp_type));
1274 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1275 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1276 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1277 type->norm = format_desc->channel[chan].normalized;
1278
1279 if (is_arithmetic_format(format_desc)) {
1280 type->width = 0;
1281 type->length = 1;
1282
1283 for (i = 0; i < format_desc->nr_channels; ++i) {
1284 type->width += format_desc->channel[i].size;
1285 }
1286 } else {
1287 type->width = format_desc->channel[chan].size;
1288 type->length = format_desc->nr_channels;
1289 }
1290 }
1291
1292
1293 /**
1294 * Retrieves the type for a format which is usable in the blending code.
1295 *
1296 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1297 */
1298 static inline void
1299 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1300 struct lp_type* type)
1301 {
1302 unsigned i;
1303 unsigned chan;
1304
1305 if (format_expands_to_float_soa(format_desc)) {
1306 /* always use ordinary floats for blending */
1307 type->floating = true;
1308 type->fixed = false;
1309 type->sign = true;
1310 type->norm = false;
1311 type->width = 32;
1312 type->length = 4;
1313 return;
1314 }
1315
1316 for (i = 0; i < 4; i++)
1317 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1318 break;
1319 chan = i;
1320
1321 memset(type, 0, sizeof(struct lp_type));
1322 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1323 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1324 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1325 type->norm = format_desc->channel[chan].normalized;
1326 type->width = format_desc->channel[chan].size;
1327 type->length = format_desc->nr_channels;
1328
1329 for (i = 1; i < format_desc->nr_channels; ++i) {
1330 if (format_desc->channel[i].size > type->width)
1331 type->width = format_desc->channel[i].size;
1332 }
1333
1334 if (type->floating) {
1335 type->width = 32;
1336 } else {
1337 if (type->width <= 8) {
1338 type->width = 8;
1339 } else if (type->width <= 16) {
1340 type->width = 16;
1341 } else {
1342 type->width = 32;
1343 }
1344 }
1345
1346 if (is_arithmetic_format(format_desc) && type->length == 3) {
1347 type->length = 4;
1348 }
1349 }
1350
1351
1352 /**
1353 * Scale a normalized value from src_bits to dst_bits.
1354 *
1355 * The exact calculation is
1356 *
1357 * dst = iround(src * dst_mask / src_mask)
1358 *
1359 * or with integer rounding
1360 *
1361 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1362 *
1363 * where
1364 *
1365 * src_mask = (1 << src_bits) - 1
1366 * dst_mask = (1 << dst_bits) - 1
1367 *
1368 * but we try to avoid division and multiplication through shifts.
1369 */
1370 static inline LLVMValueRef
1371 scale_bits(struct gallivm_state *gallivm,
1372 int src_bits,
1373 int dst_bits,
1374 LLVMValueRef src,
1375 struct lp_type src_type)
1376 {
1377 LLVMBuilderRef builder = gallivm->builder;
1378 LLVMValueRef result = src;
1379
1380 if (dst_bits < src_bits) {
1381 int delta_bits = src_bits - dst_bits;
1382
1383 if (delta_bits <= dst_bits) {
1384 /*
1385 * Approximate the rescaling with a single shift.
1386 *
1387 * This gives the wrong rounding.
1388 */
1389
1390 result = LLVMBuildLShr(builder,
1391 src,
1392 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1393 "");
1394
1395 } else {
1396 /*
1397 * Try more accurate rescaling.
1398 */
1399
1400 /*
1401 * Drop the least significant bits to make space for the multiplication.
1402 *
1403 * XXX: A better approach would be to use a wider integer type as intermediate. But
1404 * this is enough to convert alpha from 16bits -> 2 when rendering to
1405 * PIPE_FORMAT_R10G10B10A2_UNORM.
1406 */
1407 result = LLVMBuildLShr(builder,
1408 src,
1409 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1410 "");
1411
1412
1413 result = LLVMBuildMul(builder,
1414 result,
1415 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1416 "");
1417
1418 /*
1419 * Add a rounding term before the division.
1420 *
1421 * TODO: Handle signed integers too.
1422 */
1423 if (!src_type.sign) {
1424 result = LLVMBuildAdd(builder,
1425 result,
1426 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1427 "");
1428 }
1429
1430 /*
1431 * Approximate the division by src_mask with a src_bits shift.
1432 *
1433 * Given the src has already been shifted by dst_bits, all we need
1434 * to do is to shift by the difference.
1435 */
1436
1437 result = LLVMBuildLShr(builder,
1438 result,
1439 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1440 "");
1441 }
1442
1443 } else if (dst_bits > src_bits) {
1444 /* Scale up bits */
1445 int db = dst_bits - src_bits;
1446
1447 /* Shift left by difference in bits */
1448 result = LLVMBuildShl(builder,
1449 src,
1450 lp_build_const_int_vec(gallivm, src_type, db),
1451 "");
1452
1453 if (db <= src_bits) {
1454 /* Enough bits in src to fill the remainder */
1455 LLVMValueRef lower = LLVMBuildLShr(builder,
1456 src,
1457 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1458 "");
1459
1460 result = LLVMBuildOr(builder, result, lower, "");
1461 } else if (db > src_bits) {
1462 /* Need to repeatedly copy src bits to fill remainder in dst */
1463 unsigned n;
1464
1465 for (n = src_bits; n < dst_bits; n *= 2) {
1466 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1467
1468 result = LLVMBuildOr(builder,
1469 result,
1470 LLVMBuildLShr(builder, result, shuv, ""),
1471 "");
1472 }
1473 }
1474 }
1475
1476 return result;
1477 }
1478
1479 /**
1480 * If RT is a smallfloat (needing denorms) format
1481 */
1482 static inline int
1483 have_smallfloat_format(struct lp_type dst_type,
1484 enum pipe_format format)
1485 {
1486 return ((dst_type.floating && dst_type.width != 32) ||
1487 /* due to format handling hacks this format doesn't have floating set
1488 * here (and actually has width set to 32 too) so special case this. */
1489 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1490 }
1491
1492
1493 /**
1494 * Convert from memory format to blending format
1495 *
1496 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1497 */
1498 static void
1499 convert_to_blend_type(struct gallivm_state *gallivm,
1500 unsigned block_size,
1501 const struct util_format_description *src_fmt,
1502 struct lp_type src_type,
1503 struct lp_type dst_type,
1504 LLVMValueRef* src, // and dst
1505 unsigned num_srcs)
1506 {
1507 LLVMValueRef *dst = src;
1508 LLVMBuilderRef builder = gallivm->builder;
1509 struct lp_type blend_type;
1510 struct lp_type mem_type;
1511 unsigned i, j;
1512 unsigned pixels = block_size / num_srcs;
1513 bool is_arith;
1514
1515 /*
1516 * full custom path for packed floats and srgb formats - none of the later
1517 * functions would do anything useful, and given the lp_type representation they
1518 * can't be fixed. Should really have some SoA blend path for these kind of
1519 * formats rather than hacking them in here.
1520 */
1521 if (format_expands_to_float_soa(src_fmt)) {
1522 LLVMValueRef tmpsrc[4];
1523 /*
1524 * This is pretty suboptimal for this case blending in SoA would be much
1525 * better, since conversion gets us SoA values so need to convert back.
1526 */
1527 assert(src_type.width == 32 || src_type.width == 16);
1528 assert(dst_type.floating);
1529 assert(dst_type.width == 32);
1530 assert(dst_type.length % 4 == 0);
1531 assert(num_srcs % 4 == 0);
1532
1533 if (src_type.width == 16) {
1534 /* expand 4x16bit values to 4x32bit */
1535 struct lp_type type32x4 = src_type;
1536 LLVMTypeRef ltype32x4;
1537 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1538 type32x4.width = 32;
1539 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1540 for (i = 0; i < num_fetch; i++) {
1541 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1542 }
1543 src_type.width = 32;
1544 }
1545 for (i = 0; i < 4; i++) {
1546 tmpsrc[i] = src[i];
1547 }
1548 for (i = 0; i < num_srcs / 4; i++) {
1549 LLVMValueRef tmpsoa[4];
1550 LLVMValueRef tmps = tmpsrc[i];
1551 if (dst_type.length == 8) {
1552 LLVMValueRef shuffles[8];
1553 unsigned j;
1554 /* fetch was 4 values but need 8-wide output values */
1555 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1556 /*
1557 * for 8-wide aos transpose would give us wrong order not matching
1558 * incoming converted fs values and mask. ARGH.
1559 */
1560 for (j = 0; j < 4; j++) {
1561 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1562 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1563 }
1564 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1565 LLVMConstVector(shuffles, 8), "");
1566 }
1567 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1568 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1569 }
1570 else {
1571 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1572 }
1573 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1574 }
1575 return;
1576 }
1577
1578 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1579 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1580
1581 /* Is the format arithmetic */
1582 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1583 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1584
1585 /* Pad if necessary */
1586 if (!is_arith && src_type.length < dst_type.length) {
1587 for (i = 0; i < num_srcs; ++i) {
1588 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1589 }
1590
1591 src_type.length = dst_type.length;
1592 }
1593
1594 /* Special case for half-floats */
1595 if (mem_type.width == 16 && mem_type.floating) {
1596 assert(blend_type.width == 32 && blend_type.floating);
1597 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1598 is_arith = false;
1599 }
1600
1601 if (!is_arith) {
1602 return;
1603 }
1604
1605 src_type.width = blend_type.width * blend_type.length;
1606 blend_type.length *= pixels;
1607 src_type.length *= pixels / (src_type.length / mem_type.length);
1608
1609 for (i = 0; i < num_srcs; ++i) {
1610 LLVMValueRef chans[4];
1611 LLVMValueRef res = NULL;
1612
1613 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1614
1615 for (j = 0; j < src_fmt->nr_channels; ++j) {
1616 unsigned mask = 0;
1617 unsigned sa = src_fmt->channel[j].shift;
1618 #if UTIL_ARCH_LITTLE_ENDIAN
1619 unsigned from_lsb = j;
1620 #else
1621 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1622 #endif
1623
1624 mask = (1 << src_fmt->channel[j].size) - 1;
1625
1626 /* Extract bits from source */
1627 chans[j] = LLVMBuildLShr(builder,
1628 dst[i],
1629 lp_build_const_int_vec(gallivm, src_type, sa),
1630 "");
1631
1632 chans[j] = LLVMBuildAnd(builder,
1633 chans[j],
1634 lp_build_const_int_vec(gallivm, src_type, mask),
1635 "");
1636
1637 /* Scale bits */
1638 if (src_type.norm) {
1639 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1640 blend_type.width, chans[j], src_type);
1641 }
1642
1643 /* Insert bits into correct position */
1644 chans[j] = LLVMBuildShl(builder,
1645 chans[j],
1646 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1647 "");
1648
1649 if (j == 0) {
1650 res = chans[j];
1651 } else {
1652 res = LLVMBuildOr(builder, res, chans[j], "");
1653 }
1654 }
1655
1656 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1657 }
1658 }
1659
1660
1661 /**
1662 * Convert from blending format to memory format
1663 *
1664 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1665 */
1666 static void
1667 convert_from_blend_type(struct gallivm_state *gallivm,
1668 unsigned block_size,
1669 const struct util_format_description *src_fmt,
1670 struct lp_type src_type,
1671 struct lp_type dst_type,
1672 LLVMValueRef* src, // and dst
1673 unsigned num_srcs)
1674 {
1675 LLVMValueRef* dst = src;
1676 unsigned i, j, k;
1677 struct lp_type mem_type;
1678 struct lp_type blend_type;
1679 LLVMBuilderRef builder = gallivm->builder;
1680 unsigned pixels = block_size / num_srcs;
1681 bool is_arith;
1682
1683 /*
1684 * full custom path for packed floats and srgb formats - none of the later
1685 * functions would do anything useful, and given the lp_type representation they
1686 * can't be fixed. Should really have some SoA blend path for these kind of
1687 * formats rather than hacking them in here.
1688 */
1689 if (format_expands_to_float_soa(src_fmt)) {
1690 /*
1691 * This is pretty suboptimal for this case blending in SoA would be much
1692 * better - we need to transpose the AoS values back to SoA values for
1693 * conversion/packing.
1694 */
1695 assert(src_type.floating);
1696 assert(src_type.width == 32);
1697 assert(src_type.length % 4 == 0);
1698 assert(dst_type.width == 32 || dst_type.width == 16);
1699
1700 for (i = 0; i < num_srcs / 4; i++) {
1701 LLVMValueRef tmpsoa[4], tmpdst;
1702 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1703 /* really really need SoA here */
1704
1705 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1706 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1707 }
1708 else {
1709 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1710 src_type, tmpsoa);
1711 }
1712
1713 if (src_type.length == 8) {
1714 LLVMValueRef tmpaos, shuffles[8];
1715 unsigned j;
1716 /*
1717 * for 8-wide aos transpose has given us wrong order not matching
1718 * output order. HMPF. Also need to split the output values manually.
1719 */
1720 for (j = 0; j < 4; j++) {
1721 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1722 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1723 }
1724 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1725 LLVMConstVector(shuffles, 8), "");
1726 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1727 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1728 }
1729 else {
1730 src[i] = tmpdst;
1731 }
1732 }
1733 if (dst_type.width == 16) {
1734 struct lp_type type16x8 = dst_type;
1735 struct lp_type type32x4 = dst_type;
1736 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1737 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1738 type16x8.length = 8;
1739 type32x4.width = 32;
1740 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1741 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1742 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1743 /* We could do vector truncation but it doesn't generate very good code */
1744 for (i = 0; i < num_fetch; i++) {
1745 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1746 src[i], lp_build_zero(gallivm, type32x4));
1747 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1748 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1749 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1750 }
1751 }
1752 return;
1753 }
1754
1755 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1756 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1757
1758 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1759
1760 /* Special case for half-floats */
1761 if (mem_type.width == 16 && mem_type.floating) {
1762 int length = dst_type.length;
1763 assert(blend_type.width == 32 && blend_type.floating);
1764
1765 dst_type.length = src_type.length;
1766
1767 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1768
1769 dst_type.length = length;
1770 is_arith = false;
1771 }
1772
1773 /* Remove any padding */
1774 if (!is_arith && (src_type.length % mem_type.length)) {
1775 src_type.length -= (src_type.length % mem_type.length);
1776
1777 for (i = 0; i < num_srcs; ++i) {
1778 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1779 }
1780 }
1781
1782 /* No bit arithmetic to do */
1783 if (!is_arith) {
1784 return;
1785 }
1786
1787 src_type.length = pixels;
1788 src_type.width = blend_type.length * blend_type.width;
1789 dst_type.length = pixels;
1790
1791 for (i = 0; i < num_srcs; ++i) {
1792 LLVMValueRef chans[4];
1793 LLVMValueRef res = NULL;
1794
1795 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1796
1797 for (j = 0; j < src_fmt->nr_channels; ++j) {
1798 unsigned mask = 0;
1799 unsigned sa = src_fmt->channel[j].shift;
1800 unsigned sz_a = src_fmt->channel[j].size;
1801 #if UTIL_ARCH_LITTLE_ENDIAN
1802 unsigned from_lsb = j;
1803 #else
1804 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1805 #endif
1806
1807 assert(blend_type.width > src_fmt->channel[j].size);
1808
1809 for (k = 0; k < blend_type.width; ++k) {
1810 mask |= 1 << k;
1811 }
1812
1813 /* Extract bits */
1814 chans[j] = LLVMBuildLShr(builder,
1815 dst[i],
1816 lp_build_const_int_vec(gallivm, src_type,
1817 from_lsb * blend_type.width),
1818 "");
1819
1820 chans[j] = LLVMBuildAnd(builder,
1821 chans[j],
1822 lp_build_const_int_vec(gallivm, src_type, mask),
1823 "");
1824
1825 /* Scale down bits */
1826 if (src_type.norm) {
1827 chans[j] = scale_bits(gallivm, blend_type.width,
1828 src_fmt->channel[j].size, chans[j], src_type);
1829 } else if (!src_type.floating && sz_a < blend_type.width) {
1830 LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);
1831 LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans[j], mask_val, "");
1832 chans[j] = LLVMBuildSelect(builder, mask, mask_val, chans[j], "");
1833 }
1834
1835 /* Insert bits */
1836 chans[j] = LLVMBuildShl(builder,
1837 chans[j],
1838 lp_build_const_int_vec(gallivm, src_type, sa),
1839 "");
1840
1841 sa += src_fmt->channel[j].size;
1842
1843 if (j == 0) {
1844 res = chans[j];
1845 } else {
1846 res = LLVMBuildOr(builder, res, chans[j], "");
1847 }
1848 }
1849
1850 assert (dst_type.width != 24);
1851
1852 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1853 }
1854 }
1855
1856
1857 /**
1858 * Convert alpha to same blend type as src
1859 */
1860 static void
1861 convert_alpha(struct gallivm_state *gallivm,
1862 struct lp_type row_type,
1863 struct lp_type alpha_type,
1864 const unsigned block_size,
1865 const unsigned block_height,
1866 const unsigned src_count,
1867 const unsigned dst_channels,
1868 const bool pad_inline,
1869 LLVMValueRef* src_alpha)
1870 {
1871 LLVMBuilderRef builder = gallivm->builder;
1872 unsigned i, j;
1873 unsigned length = row_type.length;
1874 row_type.length = alpha_type.length;
1875
1876 /* Twiddle the alpha to match pixels */
1877 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1878
1879 /*
1880 * TODO this should use single lp_build_conv call for
1881 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1882 */
1883 for (i = 0; i < block_height; ++i) {
1884 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1885 }
1886
1887 alpha_type = row_type;
1888 row_type.length = length;
1889
1890 /* If only one channel we can only need the single alpha value per pixel */
1891 if (src_count == 1 && dst_channels == 1) {
1892
1893 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1894 } else {
1895 /* If there are more srcs than rows then we need to split alpha up */
1896 if (src_count > block_height) {
1897 for (i = src_count; i > 0; --i) {
1898 unsigned pixels = block_size / src_count;
1899 unsigned idx = i - 1;
1900
1901 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1902 (idx * pixels) % 4, pixels);
1903 }
1904 }
1905
1906 /* If there is a src for each pixel broadcast the alpha across whole row */
1907 if (src_count == block_size) {
1908 for (i = 0; i < src_count; ++i) {
1909 src_alpha[i] = lp_build_broadcast(gallivm,
1910 lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1911 }
1912 } else {
1913 unsigned pixels = block_size / src_count;
1914 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1915 unsigned alpha_span = 1;
1916 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1917
1918 /* Check if we need 2 src_alphas for our shuffles */
1919 if (pixels > alpha_type.length) {
1920 alpha_span = 2;
1921 }
1922
1923 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1924 for (j = 0; j < row_type.length; ++j) {
1925 if (j < pixels * channels) {
1926 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1927 } else {
1928 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1929 }
1930 }
1931
1932 for (i = 0; i < src_count; ++i) {
1933 unsigned idx1 = i, idx2 = i;
1934
1935 if (alpha_span > 1){
1936 idx1 *= alpha_span;
1937 idx2 = idx1 + 1;
1938 }
1939
1940 src_alpha[i] = LLVMBuildShuffleVector(builder,
1941 src_alpha[idx1],
1942 src_alpha[idx2],
1943 LLVMConstVector(shuffles, row_type.length),
1944 "");
1945 }
1946 }
1947 }
1948 }
1949
1950
1951 /**
1952 * Generates the blend function for unswizzled colour buffers
1953 * Also generates the read & write from colour buffer
1954 */
1955 static void
1956 generate_unswizzled_blend(struct gallivm_state *gallivm,
1957 unsigned rt,
1958 struct lp_fragment_shader_variant *variant,
1959 enum pipe_format out_format,
1960 unsigned int num_fs,
1961 struct lp_type fs_type,
1962 LLVMValueRef* fs_mask,
1963 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
1964 LLVMValueRef context_ptr,
1965 LLVMValueRef color_ptr,
1966 LLVMValueRef stride,
1967 unsigned partial_mask,
1968 boolean do_branch)
1969 {
1970 const unsigned alpha_channel = 3;
1971 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
1972 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
1973 const unsigned block_size = block_width * block_height;
1974 const unsigned lp_integer_vector_width = 128;
1975
1976 LLVMBuilderRef builder = gallivm->builder;
1977 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
1978 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
1979 LLVMValueRef src_alpha[4 * 4];
1980 LLVMValueRef src1_alpha[4 * 4] = { NULL };
1981 LLVMValueRef src_mask[4 * 4];
1982 LLVMValueRef src[4 * 4];
1983 LLVMValueRef src1[4 * 4];
1984 LLVMValueRef dst[4 * 4];
1985 LLVMValueRef blend_color;
1986 LLVMValueRef blend_alpha;
1987 LLVMValueRef i32_zero;
1988 LLVMValueRef check_mask;
1989 LLVMValueRef undef_src_val;
1990
1991 struct lp_build_mask_context mask_ctx;
1992 struct lp_type mask_type;
1993 struct lp_type blend_type;
1994 struct lp_type row_type;
1995 struct lp_type dst_type;
1996 struct lp_type ls_type;
1997
1998 unsigned char swizzle[TGSI_NUM_CHANNELS];
1999 unsigned vector_width;
2000 unsigned src_channels = TGSI_NUM_CHANNELS;
2001 unsigned dst_channels;
2002 unsigned dst_count;
2003 unsigned src_count;
2004 unsigned i, j;
2005
2006 const struct util_format_description* out_format_desc = util_format_description(out_format);
2007
2008 unsigned dst_alignment;
2009
2010 bool pad_inline = is_arithmetic_format(out_format_desc);
2011 bool has_alpha = false;
2012 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
2013 util_blend_state_is_dual(&variant->key.blend, 0);
2014
2015 const boolean is_1d = variant->key.resource_1d;
2016 boolean twiddle_after_convert = FALSE;
2017 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
2018 LLVMValueRef fpstate = 0;
2019
2020 /* Get type from output format */
2021 lp_blend_type_from_format_desc(out_format_desc, &row_type);
2022 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
2023
2024 /*
2025 * Technically this code should go into lp_build_smallfloat_to_float
2026 * and lp_build_float_to_smallfloat but due to the
2027 * http://llvm.org/bugs/show_bug.cgi?id=6393
2028 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2029 * So the ordering is important here and there shouldn't be any
2030 * llvm ir instrunctions in this function before
2031 * this, otherwise half-float format conversions won't work
2032 * (again due to llvm bug #6393).
2033 */
2034 if (have_smallfloat_format(dst_type, out_format)) {
2035 /* We need to make sure that denorms are ok for half float
2036 conversions */
2037 fpstate = lp_build_fpstate_get(gallivm);
2038 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
2039 }
2040
2041 mask_type = lp_int32_vec4_type();
2042 mask_type.length = fs_type.length;
2043
2044 for (i = num_fs; i < num_fullblock_fs; i++) {
2045 fs_mask[i] = lp_build_zero(gallivm, mask_type);
2046 }
2047
2048 /* Do not bother executing code when mask is empty.. */
2049 if (do_branch) {
2050 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
2051
2052 for (i = 0; i < num_fullblock_fs; ++i) {
2053 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
2054 }
2055
2056 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
2057 lp_build_mask_check(&mask_ctx);
2058 }
2059
2060 partial_mask |= !variant->opaque;
2061 i32_zero = lp_build_const_int32(gallivm, 0);
2062
2063 undef_src_val = lp_build_undef(gallivm, fs_type);
2064
2065 row_type.length = fs_type.length;
2066 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
2067
2068 /* Compute correct swizzle and count channels */
2069 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
2070 dst_channels = 0;
2071
2072 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
2073 /* Ensure channel is used */
2074 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
2075 continue;
2076 }
2077
2078 /* Ensure not already written to (happens in case with GL_ALPHA) */
2079 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
2080 continue;
2081 }
2082
2083 /* Ensure we havn't already found all channels */
2084 if (dst_channels >= out_format_desc->nr_channels) {
2085 continue;
2086 }
2087
2088 swizzle[out_format_desc->swizzle[i]] = i;
2089 ++dst_channels;
2090
2091 if (i == alpha_channel) {
2092 has_alpha = true;
2093 }
2094 }
2095
2096 if (format_expands_to_float_soa(out_format_desc)) {
2097 /*
2098 * the code above can't work for layout_other
2099 * for srgb it would sort of work but we short-circuit swizzles, etc.
2100 * as that is done as part of unpack / pack.
2101 */
2102 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2103 has_alpha = true;
2104 swizzle[0] = 0;
2105 swizzle[1] = 1;
2106 swizzle[2] = 2;
2107 swizzle[3] = 3;
2108 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2109 }
2110
2111 /* If 3 channels then pad to include alpha for 4 element transpose */
2112 if (dst_channels == 3) {
2113 assert (!has_alpha);
2114 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2115 if (swizzle[i] > TGSI_NUM_CHANNELS)
2116 swizzle[i] = 3;
2117 }
2118 if (out_format_desc->nr_channels == 4) {
2119 dst_channels = 4;
2120 /*
2121 * We use alpha from the color conversion, not separate one.
2122 * We had to include it for transpose, hence it will get converted
2123 * too (albeit when doing transpose after conversion, that would
2124 * no longer be the case necessarily).
2125 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2126 * otherwise we really have padding, not alpha, included.)
2127 */
2128 has_alpha = true;
2129 }
2130 }
2131
2132 /*
2133 * Load shader output
2134 */
2135 for (i = 0; i < num_fullblock_fs; ++i) {
2136 /* Always load alpha for use in blending */
2137 LLVMValueRef alpha;
2138 if (i < num_fs) {
2139 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
2140 }
2141 else {
2142 alpha = undef_src_val;
2143 }
2144
2145 /* Load each channel */
2146 for (j = 0; j < dst_channels; ++j) {
2147 assert(swizzle[j] < 4);
2148 if (i < num_fs) {
2149 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
2150 }
2151 else {
2152 fs_src[i][j] = undef_src_val;
2153 }
2154 }
2155
2156 /* If 3 channels then pad to include alpha for 4 element transpose */
2157 /*
2158 * XXX If we include that here maybe could actually use it instead of
2159 * separate alpha for blending?
2160 * (Difficult though we actually convert pad channels, not alpha.)
2161 */
2162 if (dst_channels == 3 && !has_alpha) {
2163 fs_src[i][3] = alpha;
2164 }
2165
2166 /* We split the row_mask and row_alpha as we want 128bit interleave */
2167 if (fs_type.length == 8) {
2168 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
2169 0, src_channels);
2170 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
2171 src_channels, src_channels);
2172
2173 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2174 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2175 src_channels, src_channels);
2176 } else {
2177 src_mask[i] = fs_mask[i];
2178 src_alpha[i] = alpha;
2179 }
2180 }
2181 if (dual_source_blend) {
2182 /* same as above except different src/dst, skip masks and comments... */
2183 for (i = 0; i < num_fullblock_fs; ++i) {
2184 LLVMValueRef alpha;
2185 if (i < num_fs) {
2186 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
2187 }
2188 else {
2189 alpha = undef_src_val;
2190 }
2191
2192 for (j = 0; j < dst_channels; ++j) {
2193 assert(swizzle[j] < 4);
2194 if (i < num_fs) {
2195 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
2196 }
2197 else {
2198 fs_src1[i][j] = undef_src_val;
2199 }
2200 }
2201 if (dst_channels == 3 && !has_alpha) {
2202 fs_src1[i][3] = alpha;
2203 }
2204 if (fs_type.length == 8) {
2205 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2206 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2207 src_channels, src_channels);
2208 } else {
2209 src1_alpha[i] = alpha;
2210 }
2211 }
2212 }
2213
2214 if (util_format_is_pure_integer(out_format)) {
2215 /*
2216 * In this case fs_type was really ints or uints disguised as floats,
2217 * fix that up now.
2218 */
2219 fs_type.floating = 0;
2220 fs_type.sign = dst_type.sign;
2221 for (i = 0; i < num_fullblock_fs; ++i) {
2222 for (j = 0; j < dst_channels; ++j) {
2223 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
2224 lp_build_vec_type(gallivm, fs_type), "");
2225 }
2226 if (dst_channels == 3 && !has_alpha) {
2227 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
2228 lp_build_vec_type(gallivm, fs_type), "");
2229 }
2230 }
2231 }
2232
2233 /*
2234 * We actually should generally do conversion first (for non-1d cases)
2235 * when the blend format is 8 or 16 bits. The reason is obvious,
2236 * there's 2 or 4 times less vectors to deal with for the interleave...
2237 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2238 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2239 * unpack only with 128bit vectors).
2240 * Note: for 16bit sizes really need matching pack conversion code
2241 */
2242 if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
2243 twiddle_after_convert = TRUE;
2244 }
2245
2246 /*
2247 * Pixel twiddle from fragment shader order to memory order
2248 */
2249 if (!twiddle_after_convert) {
2250 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
2251 dst_channels, fs_src, src, pad_inline);
2252 if (dual_source_blend) {
2253 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
2254 fs_src1, src1, pad_inline);
2255 }
2256 } else {
2257 src_count = num_fullblock_fs * dst_channels;
2258 /*
2259 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2260 * (AVX) turn out the same later when untwiddling/transpose (albeit
2261 * for true AVX2 path untwiddle needs to be different).
2262 * For now just order by colors first (so we can use unpack later).
2263 */
2264 for (j = 0; j < num_fullblock_fs; j++) {
2265 for (i = 0; i < dst_channels; i++) {
2266 src[i*num_fullblock_fs + j] = fs_src[j][i];
2267 if (dual_source_blend) {
2268 src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2269 }
2270 }
2271 }
2272 }
2273
2274 src_channels = dst_channels < 3 ? dst_channels : 4;
2275 if (src_count != num_fullblock_fs * src_channels) {
2276 unsigned ds = src_count / (num_fullblock_fs * src_channels);
2277 row_type.length /= ds;
2278 fs_type.length = row_type.length;
2279 }
2280
2281 blend_type = row_type;
2282 mask_type.length = 4;
2283
2284 /* Convert src to row_type */
2285 if (dual_source_blend) {
2286 struct lp_type old_row_type = row_type;
2287 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2288 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2289 }
2290 else {
2291 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2292 }
2293
2294 /* If the rows are not an SSE vector, combine them to become SSE size! */
2295 if ((row_type.width * row_type.length) % 128) {
2296 unsigned bits = row_type.width * row_type.length;
2297 unsigned combined;
2298
2299 assert(src_count >= (vector_width / bits));
2300
2301 dst_count = src_count / (vector_width / bits);
2302
2303 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2304 if (dual_source_blend) {
2305 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2306 }
2307
2308 row_type.length *= combined;
2309 src_count /= combined;
2310
2311 bits = row_type.width * row_type.length;
2312 assert(bits == 128 || bits == 256);
2313 }
2314
2315 if (twiddle_after_convert) {
2316 fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2317 if (dual_source_blend) {
2318 fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2319 }
2320 }
2321
2322 /*
2323 * Blend Colour conversion
2324 */
2325 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2326 blend_color = LLVMBuildPointerCast(builder, blend_color,
2327 LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2328 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2329 &i32_zero, 1, ""), "");
2330
2331 /* Convert */
2332 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2333
2334 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2335 /*
2336 * since blending is done with floats, there was no conversion.
2337 * However, the rules according to fixed point renderbuffers still
2338 * apply, that is we must clamp inputs to 0.0/1.0.
2339 * (This would apply to separate alpha conversion too but we currently
2340 * force has_alpha to be true.)
2341 * TODO: should skip this with "fake" blend, since post-blend conversion
2342 * will clamp anyway.
2343 * TODO: could also skip this if fragment color clamping is enabled. We
2344 * don't support it natively so it gets baked into the shader however, so
2345 * can't really tell here.
2346 */
2347 struct lp_build_context f32_bld;
2348 assert(row_type.floating);
2349 lp_build_context_init(&f32_bld, gallivm, row_type);
2350 for (i = 0; i < src_count; i++) {
2351 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2352 }
2353 if (dual_source_blend) {
2354 for (i = 0; i < src_count; i++) {
2355 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2356 }
2357 }
2358 /* probably can't be different than row_type but better safe than sorry... */
2359 lp_build_context_init(&f32_bld, gallivm, blend_type);
2360 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2361 }
2362
2363 /* Extract alpha */
2364 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2365
2366 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2367 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2368 if (pad_inline) {
2369 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2370 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2371 } else {
2372 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2373 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2374 }
2375
2376 /*
2377 * Mask conversion
2378 */
2379 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2380
2381 if (src_count < block_height) {
2382 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2383 } else if (src_count > block_height) {
2384 for (i = src_count; i > 0; --i) {
2385 unsigned pixels = block_size / src_count;
2386 unsigned idx = i - 1;
2387
2388 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2389 (idx * pixels) % 4, pixels);
2390 }
2391 }
2392
2393 assert(mask_type.width == 32);
2394
2395 for (i = 0; i < src_count; ++i) {
2396 unsigned pixels = block_size / src_count;
2397 unsigned pixel_width = row_type.width * dst_channels;
2398
2399 if (pixel_width == 24) {
2400 mask_type.width = 8;
2401 mask_type.length = vector_width / mask_type.width;
2402 } else {
2403 mask_type.length = pixels;
2404 mask_type.width = row_type.width * dst_channels;
2405
2406 /*
2407 * If mask_type width is smaller than 32bit, this doesn't quite
2408 * generate the most efficient code (could use some pack).
2409 */
2410 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2411 lp_build_int_vec_type(gallivm, mask_type), "");
2412
2413 mask_type.length *= dst_channels;
2414 mask_type.width /= dst_channels;
2415 }
2416
2417 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2418 lp_build_int_vec_type(gallivm, mask_type), "");
2419 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2420 }
2421
2422 /*
2423 * Alpha conversion
2424 */
2425 if (!has_alpha) {
2426 struct lp_type alpha_type = fs_type;
2427 alpha_type.length = 4;
2428 convert_alpha(gallivm, row_type, alpha_type,
2429 block_size, block_height,
2430 src_count, dst_channels,
2431 pad_inline, src_alpha);
2432 if (dual_source_blend) {
2433 convert_alpha(gallivm, row_type, alpha_type,
2434 block_size, block_height,
2435 src_count, dst_channels,
2436 pad_inline, src1_alpha);
2437 }
2438 }
2439
2440
2441 /*
2442 * Load dst from memory
2443 */
2444 if (src_count < block_height) {
2445 dst_count = block_height;
2446 } else {
2447 dst_count = src_count;
2448 }
2449
2450 dst_type.length *= block_size / dst_count;
2451
2452 if (format_expands_to_float_soa(out_format_desc)) {
2453 /*
2454 * we need multiple values at once for the conversion, so can as well
2455 * load them vectorized here too instead of concatenating later.
2456 * (Still need concatenation later for 8-wide vectors).
2457 */
2458 dst_count = block_height;
2459 dst_type.length = block_width;
2460 }
2461
2462 /*
2463 * Compute the alignment of the destination pointer in bytes
2464 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2465 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2466 * 1d tex but can't distinguish here) so need to stick with per-pixel
2467 * alignment in this case.
2468 */
2469 if (is_1d) {
2470 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2471 }
2472 else {
2473 dst_alignment = dst_type.length * dst_type.width / 8;
2474 }
2475 /* Force power-of-two alignment by extracting only the least-significant-bit */
2476 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2477 /*
2478 * Resource base and stride pointers are aligned to 16 bytes, so that's
2479 * the maximum alignment we can guarantee
2480 */
2481 dst_alignment = MIN2(16, dst_alignment);
2482
2483 ls_type = dst_type;
2484
2485 if (dst_count > src_count) {
2486 if ((dst_type.width == 8 || dst_type.width == 16) &&
2487 util_is_power_of_two_or_zero(dst_type.length) &&
2488 dst_type.length * dst_type.width < 128) {
2489 /*
2490 * Never try to load values as 4xi8 which we will then
2491 * concatenate to larger vectors. This gives llvm a real
2492 * headache (the problem is the type legalizer (?) will
2493 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2494 * then the shuffles to concatenate are more or less impossible
2495 * - llvm is easily capable of generating a sequence of 32
2496 * pextrb/pinsrb instructions for that. Albeit it appears to
2497 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2498 * width to avoid the trouble (16bit seems not as bad, llvm
2499 * probably recognizes the load+shuffle as only one shuffle
2500 * is necessary, but we can do just the same anyway).
2501 */
2502 ls_type.length = dst_type.length * dst_type.width / 32;
2503 ls_type.width = 32;
2504 }
2505 }
2506
2507 if (is_1d) {
2508 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2509 dst, ls_type, dst_count / 4, dst_alignment);
2510 for (i = dst_count / 4; i < dst_count; i++) {
2511 dst[i] = lp_build_undef(gallivm, ls_type);
2512 }
2513
2514 }
2515 else {
2516 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2517 dst, ls_type, dst_count, dst_alignment);
2518 }
2519
2520
2521 /*
2522 * Convert from dst/output format to src/blending format.
2523 *
2524 * This is necessary as we can only read 1 row from memory at a time,
2525 * so the minimum dst_count will ever be at this point is 4.
2526 *
2527 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2528 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2529 * on all 16 pixels in that single vector at once.
2530 */
2531 if (dst_count > src_count) {
2532 if (ls_type.length != dst_type.length && ls_type.length == 1) {
2533 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2534 LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2535 for (i = 0; i < dst_count; i++) {
2536 dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2537 }
2538 }
2539
2540 lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2541
2542 if (ls_type.length != dst_type.length) {
2543 struct lp_type tmp_type = dst_type;
2544 tmp_type.length = dst_type.length * 4 / src_count;
2545 for (i = 0; i < src_count; i++) {
2546 dst[i] = LLVMBuildBitCast(builder, dst[i],
2547 lp_build_vec_type(gallivm, tmp_type), "");
2548 }
2549 }
2550 }
2551
2552 /*
2553 * Blending
2554 */
2555 /* XXX this is broken for RGB8 formats -
2556 * they get expanded from 12 to 16 elements (to include alpha)
2557 * by convert_to_blend_type then reduced to 15 instead of 12
2558 * by convert_from_blend_type (a simple fix though breaks A8...).
2559 * R16G16B16 also crashes differently however something going wrong
2560 * inside llvm handling npot vector sizes seemingly.
2561 * It seems some cleanup could be done here (like skipping conversion/blend
2562 * when not needed).
2563 */
2564 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2565 row_type, dst, src_count);
2566
2567 /*
2568 * FIXME: Really should get logic ops / masks out of generic blend / row
2569 * format. Logic ops will definitely not work on the blend float format
2570 * used for SRGB here and I think OpenGL expects this to work as expected
2571 * (that is incoming values converted to srgb then logic op applied).
2572 */
2573 for (i = 0; i < src_count; ++i) {
2574 dst[i] = lp_build_blend_aos(gallivm,
2575 &variant->key.blend,
2576 out_format,
2577 row_type,
2578 rt,
2579 src[i],
2580 has_alpha ? NULL : src_alpha[i],
2581 src1[i],
2582 has_alpha ? NULL : src1_alpha[i],
2583 dst[i],
2584 partial_mask ? src_mask[i] : NULL,
2585 blend_color,
2586 has_alpha ? NULL : blend_alpha,
2587 swizzle,
2588 pad_inline ? 4 : dst_channels);
2589 }
2590
2591 convert_from_blend_type(gallivm, block_size, out_format_desc,
2592 row_type, dst_type, dst, src_count);
2593
2594 /* Split the blend rows back to memory rows */
2595 if (dst_count > src_count) {
2596 row_type.length = dst_type.length * (dst_count / src_count);
2597
2598 if (src_count == 1) {
2599 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2600 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2601
2602 row_type.length /= 2;
2603 src_count *= 2;
2604 }
2605
2606 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2607 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2608 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2609 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2610
2611 row_type.length /= 2;
2612 src_count *= 2;
2613 }
2614
2615 /*
2616 * Store blend result to memory
2617 */
2618 if (is_1d) {
2619 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2620 dst, dst_type, dst_count / 4, dst_alignment);
2621 }
2622 else {
2623 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2624 dst, dst_type, dst_count, dst_alignment);
2625 }
2626
2627 if (have_smallfloat_format(dst_type, out_format)) {
2628 lp_build_fpstate_set(gallivm, fpstate);
2629 }
2630
2631 if (do_branch) {
2632 lp_build_mask_end(&mask_ctx);
2633 }
2634 }
2635
2636
2637 /**
2638 * Generate the runtime callable function for the whole fragment pipeline.
2639 * Note that the function which we generate operates on a block of 16
2640 * pixels at at time. The block contains 2x2 quads. Each quad contains
2641 * 2x2 pixels.
2642 */
2643 static void
2644 generate_fragment(struct llvmpipe_context *lp,
2645 struct lp_fragment_shader *shader,
2646 struct lp_fragment_shader_variant *variant,
2647 unsigned partial_mask)
2648 {
2649 struct gallivm_state *gallivm = variant->gallivm;
2650 struct lp_fragment_shader_variant_key *key = &variant->key;
2651 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2652 char func_name[64];
2653 struct lp_type fs_type;
2654 struct lp_type blend_type;
2655 LLVMTypeRef fs_elem_type;
2656 LLVMTypeRef blend_vec_type;
2657 LLVMTypeRef arg_types[15];
2658 LLVMTypeRef func_type;
2659 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2660 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2661 LLVMValueRef context_ptr;
2662 LLVMValueRef x;
2663 LLVMValueRef y;
2664 LLVMValueRef a0_ptr;
2665 LLVMValueRef dadx_ptr;
2666 LLVMValueRef dady_ptr;
2667 LLVMValueRef color_ptr_ptr;
2668 LLVMValueRef stride_ptr;
2669 LLVMValueRef color_sample_stride_ptr;
2670 LLVMValueRef depth_ptr;
2671 LLVMValueRef depth_stride;
2672 LLVMValueRef depth_sample_stride;
2673 LLVMValueRef mask_input;
2674 LLVMValueRef thread_data_ptr;
2675 LLVMBasicBlockRef block;
2676 LLVMBuilderRef builder;
2677 struct lp_build_sampler_soa *sampler;
2678 struct lp_build_image_soa *image;
2679 struct lp_build_interp_soa_context interp;
2680 LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
2681 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2682 LLVMValueRef function;
2683 LLVMValueRef facing;
2684 unsigned num_fs;
2685 unsigned i;
2686 unsigned chan;
2687 unsigned cbuf;
2688 boolean cbuf0_write_all;
2689 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2690 util_blend_state_is_dual(&key->blend, 0);
2691
2692 assert(lp_native_vector_width / 32 >= 4);
2693
2694 /* Adjust color input interpolation according to flatshade state:
2695 */
2696 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2697 for (i = 0; i < shader->info.base.num_inputs; i++) {
2698 if (inputs[i].interp == LP_INTERP_COLOR) {
2699 if (key->flatshade)
2700 inputs[i].interp = LP_INTERP_CONSTANT;
2701 else
2702 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2703 }
2704 }
2705
2706 /* check if writes to cbuf[0] are to be copied to all cbufs */
2707 cbuf0_write_all =
2708 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2709
2710 /* TODO: actually pick these based on the fs and color buffer
2711 * characteristics. */
2712
2713 memset(&fs_type, 0, sizeof fs_type);
2714 fs_type.floating = TRUE; /* floating point values */
2715 fs_type.sign = TRUE; /* values are signed */
2716 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2717 fs_type.width = 32; /* 32-bit float */
2718 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2719
2720 memset(&blend_type, 0, sizeof blend_type);
2721 blend_type.floating = FALSE; /* values are integers */
2722 blend_type.sign = FALSE; /* values are unsigned */
2723 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2724 blend_type.width = 8; /* 8-bit ubyte values */
2725 blend_type.length = 16; /* 16 elements per vector */
2726
2727 /*
2728 * Generate the function prototype. Any change here must be reflected in
2729 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2730 */
2731
2732 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2733
2734 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2735
2736 snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2737 shader->no, variant->no, partial_mask ? "partial" : "whole");
2738
2739 arg_types[0] = variant->jit_context_ptr_type; /* context */
2740 arg_types[1] = int32_type; /* x */
2741 arg_types[2] = int32_type; /* y */
2742 arg_types[3] = int32_type; /* facing */
2743 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2744 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2745 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2746 arg_types[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0); /* color */
2747 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2748 arg_types[9] = LLVMInt64TypeInContext(gallivm->context); /* mask_input */
2749 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2750 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2751 arg_types[12] = int32_type; /* depth_stride */
2752 arg_types[13] = LLVMPointerType(int32_type, 0); /* color sample strides */
2753 arg_types[14] = int32_type; /* depth sample stride */
2754
2755 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2756 arg_types, ARRAY_SIZE(arg_types), 0);
2757
2758 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2759 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2760
2761 variant->function[partial_mask] = function;
2762
2763 /* XXX: need to propagate noalias down into color param now we are
2764 * passing a pointer-to-pointer?
2765 */
2766 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2767 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2768 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2769
2770 context_ptr = LLVMGetParam(function, 0);
2771 x = LLVMGetParam(function, 1);
2772 y = LLVMGetParam(function, 2);
2773 facing = LLVMGetParam(function, 3);
2774 a0_ptr = LLVMGetParam(function, 4);
2775 dadx_ptr = LLVMGetParam(function, 5);
2776 dady_ptr = LLVMGetParam(function, 6);
2777 color_ptr_ptr = LLVMGetParam(function, 7);
2778 depth_ptr = LLVMGetParam(function, 8);
2779 mask_input = LLVMGetParam(function, 9);
2780 thread_data_ptr = LLVMGetParam(function, 10);
2781 stride_ptr = LLVMGetParam(function, 11);
2782 depth_stride = LLVMGetParam(function, 12);
2783 color_sample_stride_ptr = LLVMGetParam(function, 13);
2784 depth_sample_stride = LLVMGetParam(function, 14);
2785
2786 lp_build_name(context_ptr, "context");
2787 lp_build_name(x, "x");
2788 lp_build_name(y, "y");
2789 lp_build_name(a0_ptr, "a0");
2790 lp_build_name(dadx_ptr, "dadx");
2791 lp_build_name(dady_ptr, "dady");
2792 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2793 lp_build_name(depth_ptr, "depth");
2794 lp_build_name(mask_input, "mask_input");
2795 lp_build_name(thread_data_ptr, "thread_data");
2796 lp_build_name(stride_ptr, "stride_ptr");
2797 lp_build_name(depth_stride, "depth_stride");
2798 lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");
2799 lp_build_name(depth_sample_stride, "depth_sample_stride");
2800
2801 /*
2802 * Function body
2803 */
2804
2805 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2806 builder = gallivm->builder;
2807 assert(builder);
2808 LLVMPositionBuilderAtEnd(builder, block);
2809
2810 /*
2811 * Must not count ps invocations if there's a null shader.
2812 * (It would be ok to count with null shader if there's d/s tests,
2813 * but only if there's d/s buffers too, which is different
2814 * to implicit rasterization disable which must not depend
2815 * on the d/s buffers.)
2816 * Could use popcount on mask, but pixel accuracy is not required.
2817 * Could disable if there's no stats query, but maybe not worth it.
2818 */
2819 if (shader->info.base.num_instructions > 1) {
2820 LLVMValueRef invocs, val;
2821 invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
2822 val = LLVMBuildLoad(builder, invocs, "");
2823 val = LLVMBuildAdd(builder, val,
2824 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
2825 "invoc_count");
2826 LLVMBuildStore(builder, val, invocs);
2827 }
2828
2829 /* code generated texture sampling */
2830 sampler = lp_llvm_sampler_soa_create(key->samplers);
2831 image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key));
2832
2833 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2834 /* for 1d resources only run "upper half" of stamp */
2835 if (key->resource_1d)
2836 num_fs /= 2;
2837
2838 {
2839 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2840 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2841 LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);
2842 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2843 num_loop_samp, "mask_store");
2844
2845 LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
2846 LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, flt_type, "");
2847 LLVMValueRef sample_pos_array;
2848
2849 if (key->multisample && key->coverage_samples == 4) {
2850 LLVMValueRef sample_pos_arr[8];
2851 for (unsigned i = 0; i < 4; i++) {
2852 sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);
2853 sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);
2854 }
2855 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);
2856 } else {
2857 LLVMValueRef sample_pos_arr[2];
2858 sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);
2859 sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);
2860 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);
2861 }
2862 LLVMSetInitializer(glob_sample_pos, sample_pos_array);
2863
2864 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2865 boolean pixel_center_integer =
2866 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2867
2868 /*
2869 * The shader input interpolation info is not explicitely baked in the
2870 * shader key, but everything it derives from (TGSI, and flatshade) is
2871 * already included in the shader key.
2872 */
2873 lp_build_interp_soa_init(&interp,
2874 gallivm,
2875 shader->info.base.num_inputs,
2876 inputs,
2877 pixel_center_integer,
2878 key->coverage_samples, glob_sample_pos,
2879 num_loop,
2880 key->depth_clamp,
2881 builder, fs_type,
2882 a0_ptr, dadx_ptr, dady_ptr,
2883 x, y);
2884
2885 for (i = 0; i < num_fs; i++) {
2886 if (key->multisample) {
2887 LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");
2888
2889 /*
2890 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2891 * store to the per sample mask storage. Or all of them together to generate
2892 * the fragment shader mask. (sample shading TODO).
2893 * Take the incoming state coverage mask into account.
2894 */
2895 for (unsigned s = 0; s < key->coverage_samples; s++) {
2896 LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));
2897 LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,
2898 &sindexi, 1, "sample_mask_ptr");
2899 LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,
2900 i*fs_type.length/4, s, mask_input);
2901
2902 LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");
2903 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");
2904 smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");
2905 smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);
2906
2907 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
2908 LLVMBuildStore(builder, s_mask, sample_mask_ptr);
2909 }
2910 } else {
2911 LLVMValueRef mask;
2912 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2913 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2914 &indexi, 1, "mask_ptr");
2915
2916 if (partial_mask) {
2917 mask = generate_quad_mask(gallivm, fs_type,
2918 i*fs_type.length/4, 0, mask_input);
2919 }
2920 else {
2921 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2922 }
2923 LLVMBuildStore(builder, mask, mask_ptr);
2924 }
2925 }
2926
2927 generate_fs_loop(gallivm,
2928 shader, key,
2929 builder,
2930 fs_type,
2931 context_ptr,
2932 glob_sample_pos,
2933 num_loop,
2934 &interp,
2935 sampler,
2936 image,
2937 mask_store, /* output */
2938 color_store,
2939 depth_ptr,
2940 depth_stride,
2941 depth_sample_stride,
2942 facing,
2943 thread_data_ptr);
2944
2945 for (i = 0; i < num_fs; i++) {
2946 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2947 LLVMValueRef ptr;
2948 for (unsigned s = 0; s < key->coverage_samples; s++) {
2949 int idx = (i + (s * num_fs));
2950 LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
2951 ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");
2952
2953 fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
2954 }
2955
2956 /* This is fucked up need to reorganize things */
2957 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2958 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2959 ptr = LLVMBuildGEP(builder,
2960 color_store[cbuf * !cbuf0_write_all][chan],
2961 &indexi, 1, "");
2962 fs_out_color[cbuf][chan][i] = ptr;
2963 }
2964 }
2965 if (dual_source_blend) {
2966 /* only support one dual source blend target hence always use output 1 */
2967 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2968 ptr = LLVMBuildGEP(builder,
2969 color_store[1][chan],
2970 &indexi, 1, "");
2971 fs_out_color[1][chan][i] = ptr;
2972 }
2973 }
2974 }
2975 }
2976
2977 sampler->destroy(sampler);
2978 image->destroy(image);
2979 /* Loop over color outputs / color buffers to do blending.
2980 */
2981 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2982 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
2983 LLVMValueRef color_ptr;
2984 LLVMValueRef stride;
2985 LLVMValueRef sample_stride = NULL;
2986 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
2987
2988 boolean do_branch = ((key->depth.enabled
2989 || key->stencil[0].enabled
2990 || key->alpha.enabled)
2991 && !shader->info.base.uses_kill);
2992
2993 color_ptr = LLVMBuildLoad(builder,
2994 LLVMBuildGEP(builder, color_ptr_ptr,
2995 &index, 1, ""),
2996 "");
2997
2998 stride = LLVMBuildLoad(builder,
2999 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
3000 "");
3001
3002 if (key->multisample)
3003 sample_stride = LLVMBuildLoad(builder,
3004 LLVMBuildGEP(builder, color_sample_stride_ptr,
3005 &index, 1, ""), "");
3006
3007 for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {
3008 unsigned mask_idx = num_fs * (key->multisample ? s : 0);
3009 LLVMValueRef out_ptr = color_ptr;;
3010
3011 if (key->multisample) {
3012 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");
3013 out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");
3014 }
3015 out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");
3016
3017 lp_build_name(out_ptr, "color_ptr%d", cbuf);
3018
3019 generate_unswizzled_blend(gallivm, cbuf, variant,
3020 key->cbuf_format[cbuf],
3021 num_fs, fs_type, &fs_mask[mask_idx], fs_out_color,
3022 context_ptr, out_ptr, stride,
3023 partial_mask, do_branch);
3024 }
3025 }
3026 }
3027
3028 LLVMBuildRetVoid(builder);
3029
3030 gallivm_verify_function(gallivm, function);
3031 }
3032
3033
3034 static void
3035 dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
3036 {
3037 unsigned i;
3038
3039 debug_printf("fs variant %p:\n", (void *) key);
3040
3041 if (key->flatshade) {
3042 debug_printf("flatshade = 1\n");
3043 }
3044 if (key->multisample) {
3045 debug_printf("multisample = 1\n");
3046 debug_printf("coverage samples = %d\n", key->coverage_samples);
3047 }
3048 for (i = 0; i < key->nr_cbufs; ++i) {
3049 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
3050 debug_printf("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);
3051 }
3052 if (key->depth.enabled || key->stencil[0].enabled) {
3053 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
3054 debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);
3055 }
3056 if (key->depth.enabled) {
3057 debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
3058 debug_printf("depth.writemask = %u\n", key->depth.writemask);
3059 }
3060
3061 for (i = 0; i < 2; ++i) {
3062 if (key->stencil[i].enabled) {
3063 debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
3064 debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
3065 debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
3066 debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
3067 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
3068 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
3069 }
3070 }
3071
3072 if (key->alpha.enabled) {
3073 debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
3074 }
3075
3076 if (key->occlusion_count) {
3077 debug_printf("occlusion_count = 1\n");
3078 }
3079
3080 if (key->blend.logicop_enable) {
3081 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
3082 }
3083 else if (key->blend.rt[0].blend_enable) {
3084 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));
3085 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
3086 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
3087 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));
3088 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
3089 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
3090 }
3091 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
3092 if (key->blend.alpha_to_coverage) {
3093 debug_printf("blend.alpha_to_coverage is enabled\n");
3094 }
3095 for (i = 0; i < key->nr_samplers; ++i) {
3096 const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
3097 debug_printf("sampler[%u] = \n", i);
3098 debug_printf(" .wrap = %s %s %s\n",
3099 util_str_tex_wrap(sampler->wrap_s, TRUE),
3100 util_str_tex_wrap(sampler->wrap_t, TRUE),
3101 util_str_tex_wrap(sampler->wrap_r, TRUE));
3102 debug_printf(" .min_img_filter = %s\n",
3103 util_str_tex_filter(sampler->min_img_filter, TRUE));
3104 debug_printf(" .min_mip_filter = %s\n",
3105 util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
3106 debug_printf(" .mag_img_filter = %s\n",
3107 util_str_tex_filter(sampler->mag_img_filter, TRUE));
3108 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
3109 debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
3110 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
3111 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
3112 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
3113 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
3114 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
3115 }
3116 for (i = 0; i < key->nr_sampler_views; ++i) {
3117 const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
3118 debug_printf("texture[%u] = \n", i);
3119 debug_printf(" .format = %s\n",
3120 util_format_name(texture->format));
3121 debug_printf(" .target = %s\n",
3122 util_str_tex_target(texture->target, TRUE));
3123 debug_printf(" .level_zero_only = %u\n",
3124 texture->level_zero_only);
3125 debug_printf(" .pot = %u %u %u\n",
3126 texture->pot_width,
3127 texture->pot_height,
3128 texture->pot_depth);
3129 }
3130 struct lp_image_static_state *images = lp_fs_variant_key_images(key);
3131 for (i = 0; i < key->nr_images; ++i) {
3132 const struct lp_static_texture_state *image = &images[i].image_state;
3133 debug_printf("image[%u] = \n", i);
3134 debug_printf(" .format = %s\n",
3135 util_format_name(image->format));
3136 debug_printf(" .target = %s\n",
3137 util_str_tex_target(image->target, TRUE));
3138 debug_printf(" .level_zero_only = %u\n",
3139 image->level_zero_only);
3140 debug_printf(" .pot = %u %u %u\n",
3141 image->pot_width,
3142 image->pot_height,
3143 image->pot_depth);
3144 }
3145 }
3146
3147
3148 void
3149 lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
3150 {
3151 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3152 variant->shader->no, variant->no);
3153 if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
3154 tgsi_dump(variant->shader->base.tokens, 0);
3155 else
3156 nir_print_shader(variant->shader->base.ir.nir, stderr);
3157 dump_fs_variant_key(&variant->key);
3158 debug_printf("variant->opaque = %u\n", variant->opaque);
3159 debug_printf("\n");
3160 }
3161
3162
3163 /**
3164 * Generate a new fragment shader variant from the shader code and
3165 * other state indicated by the key.
3166 */
3167 static struct lp_fragment_shader_variant *
3168 generate_variant(struct llvmpipe_context *lp,
3169 struct lp_fragment_shader *shader,
3170 const struct lp_fragment_shader_variant_key *key)
3171 {
3172 struct lp_fragment_shader_variant *variant;
3173 const struct util_format_description *cbuf0_format_desc = NULL;
3174 boolean fullcolormask;
3175 char module_name[64];
3176
3177 variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
3178 if (!variant)
3179 return NULL;
3180
3181 memset(variant, 0, sizeof(*variant));
3182 snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
3183 shader->no, shader->variants_created);
3184
3185 variant->gallivm = gallivm_create(module_name, lp->context);
3186 if (!variant->gallivm) {
3187 FREE(variant);
3188 return NULL;
3189 }
3190
3191 variant->shader = shader;
3192 variant->list_item_global.base = variant;
3193 variant->list_item_local.base = variant;
3194 variant->no = shader->variants_created++;
3195
3196 memcpy(&variant->key, key, shader->variant_key_size);
3197
3198 /*
3199 * Determine whether we are touching all channels in the color buffer.
3200 */
3201 fullcolormask = FALSE;
3202 if (key->nr_cbufs == 1) {
3203 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
3204 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
3205 }
3206
3207 variant->opaque =
3208 !key->blend.logicop_enable &&
3209 !key->blend.rt[0].blend_enable &&
3210 fullcolormask &&
3211 !key->stencil[0].enabled &&
3212 !key->alpha.enabled &&
3213 !key->multisample &&
3214 !key->blend.alpha_to_coverage &&
3215 !key->depth.enabled &&
3216 !shader->info.base.uses_kill &&
3217 !shader->info.base.writes_samplemask
3218 ? TRUE : FALSE;
3219
3220 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3221 lp_debug_fs_variant(variant);
3222 }
3223
3224 lp_jit_init_types(variant);
3225
3226 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
3227 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
3228
3229 if (variant->jit_function[RAST_WHOLE] == NULL) {
3230 if (variant->opaque) {
3231 /* Specialized shader, which doesn't need to read the color buffer. */
3232 generate_fragment(lp, shader, variant, RAST_WHOLE);
3233 }
3234 }
3235
3236 /*
3237 * Compile everything
3238 */
3239
3240 gallivm_compile_module(variant->gallivm);
3241
3242 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
3243
3244 if (variant->function[RAST_EDGE_TEST]) {
3245 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
3246 gallivm_jit_function(variant->gallivm,
3247 variant->function[RAST_EDGE_TEST]);
3248 }
3249
3250 if (variant->function[RAST_WHOLE]) {
3251 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
3252 gallivm_jit_function(variant->gallivm,
3253 variant->function[RAST_WHOLE]);
3254 } else if (!variant->jit_function[RAST_WHOLE]) {
3255 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
3256 }
3257
3258 gallivm_free_ir(variant->gallivm);
3259
3260 return variant;
3261 }
3262
3263
3264 static void *
3265 llvmpipe_create_fs_state(struct pipe_context *pipe,
3266 const struct pipe_shader_state *templ)
3267 {
3268 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3269 struct lp_fragment_shader *shader;
3270 int nr_samplers;
3271 int nr_sampler_views;
3272 int nr_images;
3273 int i;
3274
3275 shader = CALLOC_STRUCT(lp_fragment_shader);
3276 if (!shader)
3277 return NULL;
3278
3279 shader->no = fs_no++;
3280 make_empty_list(&shader->variants);
3281
3282 shader->base.type = templ->type;
3283 if (templ->type == PIPE_SHADER_IR_TGSI) {
3284 /* get/save the summary info for this shader */
3285 lp_build_tgsi_info(templ->tokens, &shader->info);
3286
3287 /* we need to keep a local copy of the tokens */
3288 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
3289 } else {
3290 shader->base.ir.nir = templ->ir.nir;
3291 nir_tgsi_scan_shader(templ->ir.nir, &shader->info.base, true);
3292 }
3293
3294 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
3295 if (shader->draw_data == NULL) {
3296 FREE((void *) shader->base.tokens);
3297 FREE(shader);
3298 return NULL;
3299 }
3300
3301 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3302 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3303 nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3304 shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
3305
3306 for (i = 0; i < shader->info.base.num_inputs; i++) {
3307 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
3308 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
3309 shader->inputs[i].location = shader->info.base.input_interpolate_loc[i];
3310
3311 switch (shader->info.base.input_interpolate[i]) {
3312 case TGSI_INTERPOLATE_CONSTANT:
3313 shader->inputs[i].interp = LP_INTERP_CONSTANT;
3314 break;
3315 case TGSI_INTERPOLATE_LINEAR:
3316 shader->inputs[i].interp = LP_INTERP_LINEAR;
3317 break;
3318 case TGSI_INTERPOLATE_PERSPECTIVE:
3319 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
3320 break;
3321 case TGSI_INTERPOLATE_COLOR:
3322 shader->inputs[i].interp = LP_INTERP_COLOR;
3323 break;
3324 default:
3325 assert(0);
3326 break;
3327 }
3328
3329 switch (shader->info.base.input_semantic_name[i]) {
3330 case TGSI_SEMANTIC_FACE:
3331 shader->inputs[i].interp = LP_INTERP_FACING;
3332 break;
3333 case TGSI_SEMANTIC_POSITION:
3334 /* Position was already emitted above
3335 */
3336 shader->inputs[i].interp = LP_INTERP_POSITION;
3337 shader->inputs[i].src_index = 0;
3338 continue;
3339 }
3340
3341 /* XXX this is a completely pointless index map... */
3342 shader->inputs[i].src_index = i+1;
3343 }
3344
3345 if (LP_DEBUG & DEBUG_TGSI) {
3346 unsigned attrib;
3347 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3348 shader->no, (void *) shader);
3349 tgsi_dump(templ->tokens, 0);
3350 debug_printf("usage masks:\n");
3351 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
3352 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
3353 debug_printf(" IN[%u].%s%s%s%s\n",
3354 attrib,
3355 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
3356 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
3357 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
3358 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
3359 }
3360 debug_printf("\n");
3361 }
3362
3363 return shader;
3364 }
3365
3366
3367 static void
3368 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
3369 {
3370 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3371 struct lp_fragment_shader *lp_fs = (struct lp_fragment_shader *)fs;
3372 if (llvmpipe->fs == lp_fs)
3373 return;
3374
3375 draw_bind_fragment_shader(llvmpipe->draw,
3376 (lp_fs ? lp_fs->draw_data : NULL));
3377
3378 llvmpipe->fs = lp_fs;
3379
3380 llvmpipe->dirty |= LP_NEW_FS;
3381 }
3382
3383
3384 /**
3385 * Remove shader variant from two lists: the shader's variant list
3386 * and the context's variant list.
3387 */
3388 static void
3389 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
3390 struct lp_fragment_shader_variant *variant)
3391 {
3392 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3393 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3394 "v total cached %u inst %u total inst %u\n",
3395 variant->shader->no, variant->no,
3396 variant->shader->variants_created,
3397 variant->shader->variants_cached,
3398 lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);
3399 }
3400
3401 gallivm_destroy(variant->gallivm);
3402
3403 /* remove from shader's list */
3404 remove_from_list(&variant->list_item_local);
3405 variant->shader->variants_cached--;
3406
3407 /* remove from context's list */
3408 remove_from_list(&variant->list_item_global);
3409 lp->nr_fs_variants--;
3410 lp->nr_fs_instrs -= variant->nr_instrs;
3411
3412 FREE(variant);
3413 }
3414
3415
3416 static void
3417 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
3418 {
3419 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3420 struct lp_fragment_shader *shader = fs;
3421 struct lp_fs_variant_list_item *li;
3422
3423 assert(fs != llvmpipe->fs);
3424
3425 /*
3426 * XXX: we need to flush the context until we have some sort of reference
3427 * counting in fragment shaders as they may still be binned
3428 * Flushing alone might not sufficient we need to wait on it too.
3429 */
3430 llvmpipe_finish(pipe, __FUNCTION__);
3431
3432 /* Delete all the variants */
3433 li = first_elem(&shader->variants);
3434 while(!at_end(&shader->variants, li)) {
3435 struct lp_fs_variant_list_item *next = next_elem(li);
3436 llvmpipe_remove_shader_variant(llvmpipe, li->base);
3437 li = next;
3438 }
3439
3440 /* Delete draw module's data */
3441 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
3442
3443 if (shader->base.ir.nir)
3444 ralloc_free(shader->base.ir.nir);
3445 assert(shader->variants_cached == 0);
3446 FREE((void *) shader->base.tokens);
3447 FREE(shader);
3448 }
3449
3450
3451
3452 static void
3453 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
3454 enum pipe_shader_type shader, uint index,
3455 const struct pipe_constant_buffer *cb)
3456 {
3457 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3458 struct pipe_resource *constants = cb ? cb->buffer : NULL;
3459
3460 assert(shader < PIPE_SHADER_TYPES);
3461 assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
3462
3463 /* note: reference counting */
3464 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
3465
3466 if (constants) {
3467 if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
3468 debug_printf("Illegal set constant without bind flag\n");
3469 constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
3470 }
3471 }
3472
3473 if (shader == PIPE_SHADER_VERTEX ||
3474 shader == PIPE_SHADER_GEOMETRY ||
3475 shader == PIPE_SHADER_TESS_CTRL ||
3476 shader == PIPE_SHADER_TESS_EVAL) {
3477 /* Pass the constants to the 'draw' module */
3478 const unsigned size = cb ? cb->buffer_size : 0;
3479 const ubyte *data;
3480
3481 if (constants) {
3482 data = (ubyte *) llvmpipe_resource_data(constants);
3483 }
3484 else if (cb && cb->user_buffer) {
3485 data = (ubyte *) cb->user_buffer;
3486 }
3487 else {
3488 data = NULL;
3489 }
3490
3491 if (data)
3492 data += cb->buffer_offset;
3493
3494 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
3495 index, data, size);
3496 }
3497 else if (shader == PIPE_SHADER_COMPUTE)
3498 llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;
3499 else
3500 llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
3501
3502 if (cb && cb->user_buffer) {
3503 pipe_resource_reference(&constants, NULL);
3504 }
3505 }
3506
3507 static void
3508 llvmpipe_set_shader_buffers(struct pipe_context *pipe,
3509 enum pipe_shader_type shader, unsigned start_slot,
3510 unsigned count, const struct pipe_shader_buffer *buffers,
3511 unsigned writable_bitmask)
3512 {
3513 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3514 unsigned i, idx;
3515 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3516 const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;
3517
3518 util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);
3519
3520 if (shader == PIPE_SHADER_VERTEX ||
3521 shader == PIPE_SHADER_GEOMETRY ||
3522 shader == PIPE_SHADER_TESS_CTRL ||
3523 shader == PIPE_SHADER_TESS_EVAL) {
3524 const unsigned size = buffer ? buffer->buffer_size : 0;
3525 const ubyte *data = NULL;
3526 if (buffer && buffer->buffer)
3527 data = (ubyte *) llvmpipe_resource_data(buffer->buffer);
3528 if (data)
3529 data += buffer->buffer_offset;
3530 draw_set_mapped_shader_buffer(llvmpipe->draw, shader,
3531 i, data, size);
3532 } else if (shader == PIPE_SHADER_COMPUTE) {
3533 llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;
3534 } else if (shader == PIPE_SHADER_FRAGMENT) {
3535 llvmpipe->dirty |= LP_NEW_FS_SSBOS;
3536 }
3537 }
3538 }
3539
3540 static void
3541 llvmpipe_set_shader_images(struct pipe_context *pipe,
3542 enum pipe_shader_type shader, unsigned start_slot,
3543 unsigned count, const struct pipe_image_view *images)
3544 {
3545 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3546 unsigned i, idx;
3547
3548 draw_flush(llvmpipe->draw);
3549 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3550 const struct pipe_image_view *image = images ? &images[idx] : NULL;
3551
3552 util_copy_image_view(&llvmpipe->images[shader][i], image);
3553 }
3554
3555 llvmpipe->num_images[shader] = start_slot + count;
3556 if (shader == PIPE_SHADER_VERTEX ||
3557 shader == PIPE_SHADER_GEOMETRY ||
3558 shader == PIPE_SHADER_TESS_CTRL ||
3559 shader == PIPE_SHADER_TESS_EVAL) {
3560 draw_set_images(llvmpipe->draw,
3561 shader,
3562 llvmpipe->images[shader],
3563 start_slot + count);
3564 } else if (shader == PIPE_SHADER_COMPUTE)
3565 llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;
3566 else
3567 llvmpipe->dirty |= LP_NEW_FS_IMAGES;
3568 }
3569
3570 /**
3571 * Return the blend factor equivalent to a destination alpha of one.
3572 */
3573 static inline unsigned
3574 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
3575 {
3576 switch(factor) {
3577 case PIPE_BLENDFACTOR_DST_ALPHA:
3578 return PIPE_BLENDFACTOR_ONE;
3579 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
3580 return PIPE_BLENDFACTOR_ZERO;
3581 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
3582 if (clamped_zero)
3583 return PIPE_BLENDFACTOR_ZERO;
3584 else
3585 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
3586 }
3587
3588 return factor;
3589 }
3590
3591
3592 /**
3593 * We need to generate several variants of the fragment pipeline to match
3594 * all the combinations of the contributing state atoms.
3595 *
3596 * TODO: there is actually no reason to tie this to context state -- the
3597 * generated code could be cached globally in the screen.
3598 */
3599 static struct lp_fragment_shader_variant_key *
3600 make_variant_key(struct llvmpipe_context *lp,
3601 struct lp_fragment_shader *shader,
3602 char *store)
3603 {
3604 unsigned i;
3605 struct lp_fragment_shader_variant_key *key;
3606
3607 key = (struct lp_fragment_shader_variant_key *)store;
3608
3609 memset(key, 0, offsetof(struct lp_fragment_shader_variant_key, samplers[1]));
3610
3611 if (lp->framebuffer.zsbuf) {
3612 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
3613 const struct util_format_description *zsbuf_desc =
3614 util_format_description(zsbuf_format);
3615
3616 if (lp->depth_stencil->depth.enabled &&
3617 util_format_has_depth(zsbuf_desc)) {
3618 key->zsbuf_format = zsbuf_format;
3619 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
3620 }
3621 if (lp->depth_stencil->stencil[0].enabled &&
3622 util_format_has_stencil(zsbuf_desc)) {
3623 key->zsbuf_format = zsbuf_format;
3624 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
3625 }
3626 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
3627 key->resource_1d = TRUE;
3628 }
3629 key->zsbuf_nr_samples = util_res_sample_count(lp->framebuffer.zsbuf->texture);
3630 }
3631
3632 /*
3633 * Propagate the depth clamp setting from the rasterizer state.
3634 * depth_clip == 0 implies depth clamping is enabled.
3635 *
3636 * When clip_halfz is enabled, then always clamp the depth values.
3637 *
3638 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3639 * clamp is always active in d3d10, regardless if depth clip is
3640 * enabled or not).
3641 * (GL has an always-on [0,1] clamp on fs depth output instead
3642 * to ensure the depth values stay in range. Doesn't look like
3643 * we do that, though...)
3644 */
3645 if (lp->rasterizer->clip_halfz) {
3646 key->depth_clamp = 1;
3647 } else {
3648 key->depth_clamp = (lp->rasterizer->depth_clip_near == 0) ? 1 : 0;
3649 }
3650
3651 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3652 if (!lp->framebuffer.nr_cbufs ||
3653 !lp->framebuffer.cbufs[0] ||
3654 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
3655 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
3656 }
3657 if(key->alpha.enabled)
3658 key->alpha.func = lp->depth_stencil->alpha.func;
3659 /* alpha.ref_value is passed in jit_context */
3660
3661 key->flatshade = lp->rasterizer->flatshade;
3662 key->multisample = lp->rasterizer->multisample;
3663 if (lp->active_occlusion_queries && !lp->queries_disabled) {
3664 key->occlusion_count = TRUE;
3665 }
3666
3667 if (lp->framebuffer.nr_cbufs) {
3668 memcpy(&key->blend, lp->blend, sizeof key->blend);
3669 }
3670
3671 key->coverage_samples = 1;
3672 if (key->multisample)
3673 key->coverage_samples = util_framebuffer_get_num_samples(&lp->framebuffer);
3674 key->nr_cbufs = lp->framebuffer.nr_cbufs;
3675
3676 if (!key->blend.independent_blend_enable) {
3677 /* we always need independent blend otherwise the fixups below won't work */
3678 for (i = 1; i < key->nr_cbufs; i++) {
3679 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
3680 }
3681 key->blend.independent_blend_enable = 1;
3682 }
3683
3684 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
3685 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
3686
3687 if (lp->framebuffer.cbufs[i]) {
3688 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
3689 const struct util_format_description *format_desc;
3690
3691 key->cbuf_format[i] = format;
3692 key->cbuf_nr_samples[i] = util_res_sample_count(lp->framebuffer.cbufs[i]->texture);
3693
3694 /*
3695 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3696 * mixing of 2d textures with height 1 and 1d textures, so make sure
3697 * we pick 1d if any cbuf or zsbuf is 1d.
3698 */
3699 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
3700 key->resource_1d = TRUE;
3701 }
3702
3703 format_desc = util_format_description(format);
3704 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
3705 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
3706
3707 /*
3708 * Mask out color channels not present in the color buffer.
3709 */
3710 blend_rt->colormask &= util_format_colormask(format_desc);
3711
3712 /*
3713 * Disable blend for integer formats.
3714 */
3715 if (util_format_is_pure_integer(format)) {
3716 blend_rt->blend_enable = 0;
3717 }
3718
3719 /*
3720 * Our swizzled render tiles always have an alpha channel, but the
3721 * linear render target format often does not, so force here the dst
3722 * alpha to be one.
3723 *
3724 * This is not a mere optimization. Wrong results will be produced if
3725 * the dst alpha is used, the dst format does not have alpha, and the
3726 * previous rendering was not flushed from the swizzled to linear
3727 * buffer. For example, NonPowTwo DCT.
3728 *
3729 * TODO: This should be generalized to all channels for better
3730 * performance, but only alpha causes correctness issues.
3731 *
3732 * Also, force rgb/alpha func/factors match, to make AoS blending
3733 * easier.
3734 */
3735 if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
3736 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3737 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3738 boolean clamped_zero = !util_format_is_float(format) &&
3739 !util_format_is_snorm(format);
3740 blend_rt->rgb_src_factor =
3741 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3742 blend_rt->rgb_dst_factor =
3743 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3744 blend_rt->alpha_func = blend_rt->rgb_func;
3745 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3746 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3747 }
3748 }
3749 else {
3750 /* no color buffer for this fragment output */
3751 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3752 key->cbuf_nr_samples[i] = 0;
3753 blend_rt->colormask = 0x0;
3754 blend_rt->blend_enable = 0;
3755 }
3756 }
3757
3758 /* This value will be the same for all the variants of a given shader:
3759 */
3760 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3761
3762 struct lp_sampler_static_state *fs_sampler;
3763
3764 fs_sampler = key->samplers;
3765
3766 memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);
3767
3768 for(i = 0; i < key->nr_samplers; ++i) {
3769 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3770 lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,
3771 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3772 }
3773 }
3774
3775 /*
3776 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3777 * are dx10-style? Can't really have mixed opcodes, at least not
3778 * if we want to skip the holes here (without rescanning tgsi).
3779 */
3780 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3781 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3782 for(i = 0; i < key->nr_sampler_views; ++i) {
3783 /*
3784 * Note sview may exceed what's representable by file_mask.
3785 * This will still work, the only downside is that not actually
3786 * used views may be included in the shader key.
3787 */
3788 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
3789 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3790 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3791 }
3792 }
3793 }
3794 else {
3795 key->nr_sampler_views = key->nr_samplers;
3796 for(i = 0; i < key->nr_sampler_views; ++i) {
3797 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3798 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3799 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3800 }
3801 }
3802 }
3803
3804 struct lp_image_static_state *lp_image;
3805 lp_image = lp_fs_variant_key_images(key);
3806 key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3807 for (i = 0; i < key->nr_images; ++i) {
3808 if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
3809 lp_sampler_static_texture_state_image(&lp_image[i].image_state,
3810 &lp->images[PIPE_SHADER_FRAGMENT][i]);
3811 }
3812 }
3813 return key;
3814 }
3815
3816
3817
3818 /**
3819 * Update fragment shader state. This is called just prior to drawing
3820 * something when some fragment-related state has changed.
3821 */
3822 void
3823 llvmpipe_update_fs(struct llvmpipe_context *lp)
3824 {
3825 struct lp_fragment_shader *shader = lp->fs;
3826 struct lp_fragment_shader_variant_key *key;
3827 struct lp_fragment_shader_variant *variant = NULL;
3828 struct lp_fs_variant_list_item *li;
3829 char store[LP_FS_MAX_VARIANT_KEY_SIZE];
3830
3831 key = make_variant_key(lp, shader, store);
3832
3833 /* Search the variants for one which matches the key */
3834 li = first_elem(&shader->variants);
3835 while(!at_end(&shader->variants, li)) {
3836 if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
3837 variant = li->base;
3838 break;
3839 }
3840 li = next_elem(li);
3841 }
3842
3843 if (variant) {
3844 /* Move this variant to the head of the list to implement LRU
3845 * deletion of shader's when we have too many.
3846 */
3847 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3848 }
3849 else {
3850 /* variant not found, create it now */
3851 int64_t t0, t1, dt;
3852 unsigned i;
3853 unsigned variants_to_cull;
3854
3855 if (LP_DEBUG & DEBUG_FS) {
3856 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3857 lp->nr_fs_variants,
3858 lp->nr_fs_instrs,
3859 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3860 }
3861
3862 /* First, check if we've exceeded the max number of shader variants.
3863 * If so, free 6.25% of them (the least recently used ones).
3864 */
3865 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
3866
3867 if (variants_to_cull ||
3868 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3869 struct pipe_context *pipe = &lp->pipe;
3870
3871 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
3872 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3873 "\t%u instrs,\t%u instrs/variant\n",
3874 shader->variants_cached,
3875 lp->nr_fs_variants, lp->nr_fs_instrs,
3876 lp->nr_fs_instrs / lp->nr_fs_variants);
3877 }
3878
3879 /*
3880 * XXX: we need to flush the context until we have some sort of
3881 * reference counting in fragment shaders as they may still be binned
3882 * Flushing alone might not be sufficient we need to wait on it too.
3883 */
3884 llvmpipe_finish(pipe, __FUNCTION__);
3885
3886 /*
3887 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3888 * number of shader variants (potentially all of them) could be
3889 * pending for destruction on flush.
3890 */
3891
3892 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3893 struct lp_fs_variant_list_item *item;
3894 if (is_empty_list(&lp->fs_variants_list)) {
3895 break;
3896 }
3897 item = last_elem(&lp->fs_variants_list);
3898 assert(item);
3899 assert(item->base);
3900 llvmpipe_remove_shader_variant(lp, item->base);
3901 }
3902 }
3903
3904 /*
3905 * Generate the new variant.
3906 */
3907 t0 = os_time_get();
3908 variant = generate_variant(lp, shader, key);
3909 t1 = os_time_get();
3910 dt = t1 - t0;
3911 LP_COUNT_ADD(llvm_compile_time, dt);
3912 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3913
3914 /* Put the new variant into the list */
3915 if (variant) {
3916 insert_at_head(&shader->variants, &variant->list_item_local);
3917 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3918 lp->nr_fs_variants++;
3919 lp->nr_fs_instrs += variant->nr_instrs;
3920 shader->variants_cached++;
3921 }
3922 }
3923
3924 /* Bind this variant */
3925 lp_setup_set_fs_variant(lp->setup, variant);
3926 }
3927
3928
3929
3930
3931
3932 void
3933 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3934 {
3935 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3936 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3937 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3938
3939 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3940
3941 llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;
3942 llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;
3943 }
3944
3945