projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: use per-sample position not sample id for interp
[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 /* Allocate color storage for each fragment sample */
448 LLVMValueRef color_store_size = num_loop;
449 if (key->min_samples > 1)
450 color_store_size = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, key->min_samples), "");
451
452 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
453 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
454 out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
455 lp_build_vec_type(gallivm,
456 type),
457 color_store_size, "color");
458 }
459 }
460 if (dual_source_blend) {
461 assert(key->nr_cbufs <= 1);
462 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
463 out_color[1][chan] = lp_build_array_alloca(gallivm,
464 lp_build_vec_type(gallivm,
465 type),
466 color_store_size, "color1");
467 }
468 }
469
470 lp_build_for_loop_begin(&loop_state, gallivm,
471 lp_build_const_int32(gallivm, 0),
472 LLVMIntULT,
473 num_loop,
474 lp_build_const_int32(gallivm, 1));
475
476 if (key->multisample) {
477 /* create shader execution mask by combining all sample masks. */
478 for (unsigned s = 0; s < key->coverage_samples; s++) {
479 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, s), "");
480 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
481 LLVMValueRef s_mask = lp_build_pointer_get(builder, mask_store, s_mask_idx);
482 if (s == 0)
483 mask_val = s_mask;
484 else
485 mask_val = LLVMBuildOr(builder, s_mask, mask_val, "");
486 }
487 } else {
488 mask_ptr = LLVMBuildGEP(builder, mask_store,
489 &loop_state.counter, 1, "mask_ptr");
490 mask_val = LLVMBuildLoad(builder, mask_ptr, "");
491 }
492
493 /* 'mask' will control execution based on quad's pixel alive/killed state */
494 lp_build_mask_begin(&mask, gallivm, type, mask_val);
495
496 if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
497 lp_build_mask_check(&mask);
498
499 /* Create storage for recombining sample masks after early Z pass. */
500 LLVMValueRef s_mask_or = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, type), "cov_mask_early_depth");
501 LLVMBuildStore(builder, LLVMConstNull(lp_build_int_vec_type(gallivm, type)), s_mask_or);
502
503 LLVMValueRef s_mask = NULL, s_mask_ptr = NULL;
504 LLVMValueRef z_sample_value_store = NULL, s_sample_value_store = NULL;
505 LLVMValueRef z_fb_store = NULL, s_fb_store = NULL;
506 LLVMTypeRef z_type = NULL, z_fb_type = NULL;
507
508 /* Run early depth once per sample */
509 if (key->multisample) {
510
511 if (zs_format_desc) {
512 struct lp_type zs_type = lp_depth_type(zs_format_desc, type.length);
513 struct lp_type z_type = zs_type;
514 struct lp_type s_type = zs_type;
515 if (zs_format_desc->block.bits < type.width)
516 z_type.width = type.width;
517 else if (zs_format_desc->block.bits > 32) {
518 z_type.width = z_type.width / 2;
519 s_type.width = s_type.width / 2;
520 s_type.floating = 0;
521 }
522 z_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
523 zs_samples, "z_sample_store");
524 s_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
525 zs_samples, "s_sample_store");
526 z_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, z_type),
527 zs_samples, "z_fb_store");
528 s_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, s_type),
529 zs_samples, "s_fb_store");
530 }
531 lp_build_for_loop_begin(&sample_loop_state, gallivm,
532 lp_build_const_int32(gallivm, 0),
533 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
534 lp_build_const_int32(gallivm, 1));
535
536 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
537 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
538 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
539
540 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
541 s_mask = LLVMBuildAnd(builder, s_mask, mask_val, "");
542 }
543
544
545 /* for multisample Z needs to be interpolated at sample points for testing. */
546 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);
547 z = interp->pos[2];
548
549 depth_ptr = depth_base_ptr;
550 if (key->multisample) {
551 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
552 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
553 }
554
555 if (depth_mode & EARLY_DEPTH_TEST) {
556 /*
557 * Clamp according to ARB_depth_clamp semantics.
558 */
559 if (key->depth_clamp) {
560 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
561 thread_data_ptr, z);
562 }
563 lp_build_depth_stencil_load_swizzled(gallivm, type,
564 zs_format_desc, key->resource_1d,
565 depth_ptr, depth_stride,
566 &z_fb, &s_fb, loop_state.counter);
567 lp_build_depth_stencil_test(gallivm,
568 &key->depth,
569 key->stencil,
570 type,
571 zs_format_desc,
572 key->multisample ? NULL : &mask,
573 &s_mask,
574 stencil_refs,
575 z, z_fb, s_fb,
576 facing,
577 &z_value, &s_value,
578 !simple_shader && !key->multisample);
579
580 if (depth_mode & EARLY_DEPTH_WRITE) {
581 lp_build_depth_stencil_write_swizzled(gallivm, type,
582 zs_format_desc, key->resource_1d,
583 NULL, NULL, NULL, loop_state.counter,
584 depth_ptr, depth_stride,
585 z_value, s_value);
586 }
587 /*
588 * Note mask check if stencil is enabled must be after ds write not after
589 * stencil test otherwise new stencil values may not get written if all
590 * fragments got killed by depth/stencil test.
591 */
592 if (!simple_shader && key->stencil[0].enabled && !key->multisample)
593 lp_build_mask_check(&mask);
594
595 if (key->multisample) {
596 z_fb_type = LLVMTypeOf(z_fb);
597 z_type = LLVMTypeOf(z_value);
598 lp_build_pointer_set(builder, z_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, z_value, lp_build_int_vec_type(gallivm, type), ""));
599 lp_build_pointer_set(builder, s_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, s_value, lp_build_int_vec_type(gallivm, type), ""));
600 lp_build_pointer_set(builder, z_fb_store, sample_loop_state.counter, z_fb);
601 lp_build_pointer_set(builder, s_fb_store, sample_loop_state.counter, s_fb);
602 }
603 }
604
605 if (key->multisample) {
606 /*
607 * Store the post-early Z coverage mask.
608 * Recombine the resulting coverage masks post early Z into the fragment
609 * shader execution mask.
610 */
611 LLVMValueRef tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
612 tmp_s_mask_or = LLVMBuildOr(builder, tmp_s_mask_or, s_mask, "");
613 LLVMBuildStore(builder, tmp_s_mask_or, s_mask_or);
614
615 LLVMBuildStore(builder, s_mask, s_mask_ptr);
616
617 lp_build_for_loop_end(&sample_loop_state);
618
619 /* recombined all the coverage masks in the shader exec mask. */
620 tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
621 lp_build_mask_update(&mask, tmp_s_mask_or);
622
623 if (key->min_samples == 1) {
624 /* for multisample Z needs to be re interpolated at pixel center */
625 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);
626 lp_build_mask_update(&mask, tmp_s_mask_or);
627 }
628 }
629
630 LLVMValueRef out_sample_mask_storage = NULL;
631 if (shader->info.base.writes_samplemask) {
632 out_sample_mask_storage = lp_build_alloca(gallivm, int_vec_type, "write_mask");
633 if (key->min_samples > 1)
634 LLVMBuildStore(builder, LLVMConstNull(int_vec_type), out_sample_mask_storage);
635 }
636
637 if (key->multisample && key->min_samples > 1) {
638 lp_build_for_loop_begin(&sample_loop_state, gallivm,
639 lp_build_const_int32(gallivm, 0),
640 LLVMIntULT,
641 lp_build_const_int32(gallivm, key->min_samples),
642 lp_build_const_int32(gallivm, 1));
643
644 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
645 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
646 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
647 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
648 lp_build_mask_force(&mask, s_mask);
649 lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, sample_loop_state.counter);
650 system_values.sample_id = sample_loop_state.counter;
651 } else
652 system_values.sample_id = lp_build_const_int32(gallivm, 0);
653
654 system_values.sample_pos = sample_pos_array;
655
656 lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, mask_store, sample_loop_state.counter);
657
658 struct lp_build_tgsi_params params;
659 memset(&params, 0, sizeof(params));
660
661 params.type = type;
662 params.mask = &mask;
663 params.consts_ptr = consts_ptr;
664 params.const_sizes_ptr = num_consts_ptr;
665 params.system_values = &system_values;
666 params.inputs = interp->inputs;
667 params.context_ptr = context_ptr;
668 params.thread_data_ptr = thread_data_ptr;
669 params.sampler = sampler;
670 params.info = &shader->info.base;
671 params.ssbo_ptr = ssbo_ptr;
672 params.ssbo_sizes_ptr = num_ssbo_ptr;
673 params.image = image;
674
675 /* Build the actual shader */
676 if (shader->base.type == PIPE_SHADER_IR_TGSI)
677 lp_build_tgsi_soa(gallivm, tokens, &params,
678 outputs);
679 else
680 lp_build_nir_soa(gallivm, shader->base.ir.nir, &params,
681 outputs);
682
683 /* Alpha test */
684 if (key->alpha.enabled) {
685 int color0 = find_output_by_semantic(&shader->info.base,
686 TGSI_SEMANTIC_COLOR,
687 0);
688
689 if (color0 != -1 && outputs[color0][3]) {
690 const struct util_format_description *cbuf_format_desc;
691 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
692 LLVMValueRef alpha_ref_value;
693
694 alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
695 alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
696
697 cbuf_format_desc = util_format_description(key->cbuf_format[0]);
698
699 lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
700 &mask, alpha, alpha_ref_value,
701 (depth_mode & LATE_DEPTH_TEST) != 0);
702 }
703 }
704
705 /* Emulate Alpha to Coverage with Alpha test */
706 if (key->blend.alpha_to_coverage) {
707 int color0 = find_output_by_semantic(&shader->info.base,
708 TGSI_SEMANTIC_COLOR,
709 0);
710
711 if (color0 != -1 && outputs[color0][3]) {
712 LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
713
714 if (!key->multisample) {
715 lp_build_alpha_to_coverage(gallivm, type,
716 &mask, alpha,
717 (depth_mode & LATE_DEPTH_TEST) != 0);
718 } else {
719 lp_build_sample_alpha_to_coverage(gallivm, type, key->coverage_samples, num_loop,
720 loop_state.counter,
721 mask_store, alpha);
722 }
723 }
724 }
725 if (key->blend.alpha_to_one && key->multisample) {
726 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib) {
727 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
728 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
729 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
730 if (outputs[cbuf][3]) {
731 LLVMBuildStore(builder, lp_build_const_vec(gallivm, type, 1.0), outputs[cbuf][3]);
732 }
733 }
734 }
735 if (shader->info.base.writes_samplemask) {
736 LLVMValueRef output_smask = NULL;
737 int smaski = find_output_by_semantic(&shader->info.base,
738 TGSI_SEMANTIC_SAMPLEMASK,
739 0);
740 struct lp_build_context smask_bld;
741 lp_build_context_init(&smask_bld, gallivm, int_type);
742
743 assert(smaski >= 0);
744 output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
745 output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");
746
747 if (key->min_samples > 1) {
748 /* only the bit corresponding to this sample is to be used. */
749 LLVMValueRef tmp_mask = LLVMBuildLoad(builder, out_sample_mask_storage, "tmp_mask");
750 LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
751 LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, lp_build_broadcast(gallivm, int_vec_type, out_smask_idx), "");
752 output_smask = LLVMBuildOr(builder, tmp_mask, smask_bit, "");
753 }
754
755 LLVMBuildStore(builder, output_smask, out_sample_mask_storage);
756 }
757
758 /* Color write - per fragment sample */
759 for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
760 {
761 unsigned cbuf = shader->info.base.output_semantic_index[attrib];
762 if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
763 ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
764 {
765 for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
766 if(outputs[attrib][chan]) {
767 /* XXX: just initialize outputs to point at colors[] and
768 * skip this.
769 */
770 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
771 LLVMValueRef color_ptr;
772 LLVMValueRef color_idx = loop_state.counter;
773 if (key->min_samples > 1)
774 color_idx = LLVMBuildAdd(builder, color_idx,
775 LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
776 color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
777 &color_idx, 1, "");
778 lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
779 LLVMBuildStore(builder, out, color_ptr);
780 }
781 }
782 }
783 }
784
785 if (key->multisample && key->min_samples > 1) {
786 LLVMBuildStore(builder, lp_build_mask_value(&mask), s_mask_ptr);
787 lp_build_for_loop_end(&sample_loop_state);
788 }
789
790 if (key->multisample) {
791 /* execute depth test for each sample */
792 lp_build_for_loop_begin(&sample_loop_state, gallivm,
793 lp_build_const_int32(gallivm, 0),
794 LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
795 lp_build_const_int32(gallivm, 1));
796
797 /* load the per-sample coverage mask */
798 LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
799 s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
800 s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
801
802 /* combine the execution mask post fragment shader with the coverage mask. */
803 s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
804 if (key->min_samples == 1)
805 s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
806
807 /* if the shader writes sample mask use that */
808 if (shader->info.base.writes_samplemask) {
809 LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
810 out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);
811 LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");
812 LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");
813 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");
814 smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");
815
816 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
817 }
818 }
819
820 depth_ptr = depth_base_ptr;
821 if (key->multisample) {
822 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
823 depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
824 }
825
826 /* Late Z test */
827 if (depth_mode & LATE_DEPTH_TEST) {
828 int pos0 = find_output_by_semantic(&shader->info.base,
829 TGSI_SEMANTIC_POSITION,
830 0);
831 int s_out = find_output_by_semantic(&shader->info.base,
832 TGSI_SEMANTIC_STENCIL,
833 0);
834 if (pos0 != -1 && outputs[pos0][2]) {
835 z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
836 }
837 /*
838 * Clamp according to ARB_depth_clamp semantics.
839 */
840 if (key->depth_clamp) {
841 z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
842 thread_data_ptr, z);
843 }
844
845 if (s_out != -1 && outputs[s_out][1]) {
846 /* there's only one value, and spec says to discard additional bits */
847 LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
848 stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
849 stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
850 stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
851 stencil_refs[1] = stencil_refs[0];
852 }
853
854 lp_build_depth_stencil_load_swizzled(gallivm, type,
855 zs_format_desc, key->resource_1d,
856 depth_ptr, depth_stride,
857 &z_fb, &s_fb, loop_state.counter);
858
859 lp_build_depth_stencil_test(gallivm,
860 &key->depth,
861 key->stencil,
862 type,
863 zs_format_desc,
864 key->multisample ? NULL : &mask,
865 &s_mask,
866 stencil_refs,
867 z, z_fb, s_fb,
868 facing,
869 &z_value, &s_value,
870 !simple_shader);
871 /* Late Z write */
872 if (depth_mode & LATE_DEPTH_WRITE) {
873 lp_build_depth_stencil_write_swizzled(gallivm, type,
874 zs_format_desc, key->resource_1d,
875 NULL, NULL, NULL, loop_state.counter,
876 depth_ptr, depth_stride,
877 z_value, s_value);
878 }
879 }
880 else if ((depth_mode & EARLY_DEPTH_TEST) &&
881 (depth_mode & LATE_DEPTH_WRITE))
882 {
883 /* Need to apply a reduced mask to the depth write. Reload the
884 * depth value, update from zs_value with the new mask value and
885 * write that out.
886 */
887 if (key->multisample) {
888 z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;
889 s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);
890 z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");
891 s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);
892 }
893 lp_build_depth_stencil_write_swizzled(gallivm, type,
894 zs_format_desc, key->resource_1d,
895 key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,
896 depth_ptr, depth_stride,
897 z_value, s_value);
898 }
899
900 if (key->occlusion_count) {
901 LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
902 lp_build_name(counter, "counter");
903
904 lp_build_occlusion_count(gallivm, type,
905 key->multisample ? s_mask : lp_build_mask_value(&mask), counter);
906 }
907
908 if (key->multisample) {
909 /* store the sample mask for this loop */
910 LLVMBuildStore(builder, s_mask, s_mask_ptr);
911 lp_build_for_loop_end(&sample_loop_state);
912 }
913
914 mask_val = lp_build_mask_end(&mask);
915 if (!key->multisample)
916 LLVMBuildStore(builder, mask_val, mask_ptr);
917 lp_build_for_loop_end(&loop_state);
918 }
919
920
921 /**
922 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
923 *
924 * Fragment Shader outputs pixels in small 2x2 blocks
925 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
926 *
927 * However in memory pixels are stored in rows
928 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
929 *
930 * @param type fragment shader type (4x or 8x float)
931 * @param num_fs number of fs_src
932 * @param is_1d whether we're outputting to a 1d resource
933 * @param dst_channels number of output channels
934 * @param fs_src output from fragment shader
935 * @param dst pointer to store result
936 * @param pad_inline is channel padding inline or at end of row
937 * @return the number of dsts
938 */
939 static int
940 generate_fs_twiddle(struct gallivm_state *gallivm,
941 struct lp_type type,
942 unsigned num_fs,
943 unsigned dst_channels,
944 LLVMValueRef fs_src[][4],
945 LLVMValueRef* dst,
946 bool pad_inline)
947 {
948 LLVMValueRef src[16];
949
950 bool swizzle_pad;
951 bool twiddle;
952 bool split;
953
954 unsigned pixels = type.length / 4;
955 unsigned reorder_group;
956 unsigned src_channels;
957 unsigned src_count;
958 unsigned i;
959
960 src_channels = dst_channels < 3 ? dst_channels : 4;
961 src_count = num_fs * src_channels;
962
963 assert(pixels == 2 || pixels == 1);
964 assert(num_fs * src_channels <= ARRAY_SIZE(src));
965
966 /*
967 * Transpose from SoA -> AoS
968 */
969 for (i = 0; i < num_fs; ++i) {
970 lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
971 }
972
973 /*
974 * Pick transformation options
975 */
976 swizzle_pad = false;
977 twiddle = false;
978 split = false;
979 reorder_group = 0;
980
981 if (dst_channels == 1) {
982 twiddle = true;
983
984 if (pixels == 2) {
985 split = true;
986 }
987 } else if (dst_channels == 2) {
988 if (pixels == 1) {
989 reorder_group = 1;
990 }
991 } else if (dst_channels > 2) {
992 if (pixels == 1) {
993 reorder_group = 2;
994 } else {
995 twiddle = true;
996 }
997
998 if (!pad_inline && dst_channels == 3 && pixels > 1) {
999 swizzle_pad = true;
1000 }
1001 }
1002
1003 /*
1004 * Split the src in half
1005 */
1006 if (split) {
1007 for (i = num_fs; i > 0; --i) {
1008 src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
1009 src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
1010 }
1011
1012 src_count *= 2;
1013 type.length = 4;
1014 }
1015
1016 /*
1017 * Ensure pixels are in memory order
1018 */
1019 if (reorder_group) {
1020 /* Twiddle pixels by reordering the array, e.g.:
1021 *
1022 * src_count = 8 -> 0 2 1 3 4 6 5 7
1023 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
1024 */
1025 const unsigned reorder_sw[] = { 0, 2, 1, 3 };
1026
1027 for (i = 0; i < src_count; ++i) {
1028 unsigned group = i / reorder_group;
1029 unsigned block = (group / 4) * 4 * reorder_group;
1030 unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
1031 dst[i] = src[j];
1032 }
1033 } else if (twiddle) {
1034 /* Twiddle pixels across elements of array */
1035 /*
1036 * XXX: we should avoid this in some cases, but would need to tell
1037 * lp_build_conv to reorder (or deal with it ourselves).
1038 */
1039 lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
1040 } else {
1041 /* Do nothing */
1042 memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
1043 }
1044
1045 /*
1046 * Moves any padding between pixels to the end
1047 * e.g. RGBXRGBX -> RGBRGBXX
1048 */
1049 if (swizzle_pad) {
1050 unsigned char swizzles[16];
1051 unsigned elems = pixels * dst_channels;
1052
1053 for (i = 0; i < type.length; ++i) {
1054 if (i < elems)
1055 swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
1056 else
1057 swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
1058 }
1059
1060 for (i = 0; i < src_count; ++i) {
1061 dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
1062 }
1063 }
1064
1065 return src_count;
1066 }
1067
1068
1069 /*
1070 * Untwiddle and transpose, much like the above.
1071 * However, this is after conversion, so we get packed vectors.
1072 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1073 * the vectors will look like:
1074 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1075 * be swizzled here). Extending to 16bit should be trivial.
1076 * Should also be extended to handle twice wide vectors with AVX2...
1077 */
1078 static void
1079 fs_twiddle_transpose(struct gallivm_state *gallivm,
1080 struct lp_type type,
1081 LLVMValueRef *src,
1082 unsigned src_count,
1083 LLVMValueRef *dst)
1084 {
1085 unsigned i, j;
1086 struct lp_type type64, type16, type32;
1087 LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
1088 LLVMBuilderRef builder = gallivm->builder;
1089 LLVMValueRef tmp[4], shuf[8];
1090 for (j = 0; j < 2; j++) {
1091 shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
1092 shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
1093 shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
1094 shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
1095 }
1096
1097 assert(src_count == 4 || src_count == 2 || src_count == 1);
1098 assert(type.width == 8);
1099 assert(type.length == 16);
1100
1101 type8_t = lp_build_vec_type(gallivm, type);
1102
1103 type64 = type;
1104 type64.length /= 8;
1105 type64.width *= 8;
1106 type64_t = lp_build_vec_type(gallivm, type64);
1107
1108 type16 = type;
1109 type16.length /= 2;
1110 type16.width *= 2;
1111 type16_t = lp_build_vec_type(gallivm, type16);
1112
1113 type32 = type;
1114 type32.length /= 4;
1115 type32.width *= 4;
1116 type32_t = lp_build_vec_type(gallivm, type32);
1117
1118 lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
1119
1120 if (src_count == 1) {
1121 /* transpose was no-op, just untwiddle */
1122 LLVMValueRef shuf_vec;
1123 shuf_vec = LLVMConstVector(shuf, 8);
1124 tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
1125 tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
1126 dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
1127 } else if (src_count == 2) {
1128 LLVMValueRef shuf_vec;
1129 shuf_vec = LLVMConstVector(shuf, 4);
1130
1131 for (i = 0; i < 2; i++) {
1132 tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
1133 tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
1134 dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
1135 }
1136 } else {
1137 for (j = 0; j < 2; j++) {
1138 LLVMValueRef lo, hi, lo2, hi2;
1139 /*
1140 * Note that if we only really have 3 valid channels (rgb)
1141 * and we don't need alpha we could substitute a undef here
1142 * for the respective channel (causing llvm to drop conversion
1143 * for alpha).
1144 */
1145 /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1146 lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
1147 hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
1148 lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
1149 hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
1150 dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
1151 dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
1152 }
1153 }
1154 }
1155
1156
1157 /**
1158 * Load an unswizzled block of pixels from memory
1159 */
1160 static void
1161 load_unswizzled_block(struct gallivm_state *gallivm,
1162 LLVMValueRef base_ptr,
1163 LLVMValueRef stride,
1164 unsigned block_width,
1165 unsigned block_height,
1166 LLVMValueRef* dst,
1167 struct lp_type dst_type,
1168 unsigned dst_count,
1169 unsigned dst_alignment)
1170 {
1171 LLVMBuilderRef builder = gallivm->builder;
1172 unsigned row_size = dst_count / block_height;
1173 unsigned i;
1174
1175 /* Ensure block exactly fits into dst */
1176 assert((block_width * block_height) % dst_count == 0);
1177
1178 for (i = 0; i < dst_count; ++i) {
1179 unsigned x = i % row_size;
1180 unsigned y = i / row_size;
1181
1182 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
1183 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1184
1185 LLVMValueRef gep[2];
1186 LLVMValueRef dst_ptr;
1187
1188 gep[0] = lp_build_const_int32(gallivm, 0);
1189 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1190
1191 dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1192 dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
1193 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
1194
1195 dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
1196
1197 LLVMSetAlignment(dst[i], dst_alignment);
1198 }
1199 }
1200
1201
1202 /**
1203 * Store an unswizzled block of pixels to memory
1204 */
1205 static void
1206 store_unswizzled_block(struct gallivm_state *gallivm,
1207 LLVMValueRef base_ptr,
1208 LLVMValueRef stride,
1209 unsigned block_width,
1210 unsigned block_height,
1211 LLVMValueRef* src,
1212 struct lp_type src_type,
1213 unsigned src_count,
1214 unsigned src_alignment)
1215 {
1216 LLVMBuilderRef builder = gallivm->builder;
1217 unsigned row_size = src_count / block_height;
1218 unsigned i;
1219
1220 /* Ensure src exactly fits into block */
1221 assert((block_width * block_height) % src_count == 0);
1222
1223 for (i = 0; i < src_count; ++i) {
1224 unsigned x = i % row_size;
1225 unsigned y = i / row_size;
1226
1227 LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
1228 LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1229
1230 LLVMValueRef gep[2];
1231 LLVMValueRef src_ptr;
1232
1233 gep[0] = lp_build_const_int32(gallivm, 0);
1234 gep[1] = LLVMBuildAdd(builder, bx, by, "");
1235
1236 src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1237 src_ptr = LLVMBuildBitCast(builder, src_ptr,
1238 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
1239
1240 src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
1241
1242 LLVMSetAlignment(src_ptr, src_alignment);
1243 }
1244 }
1245
1246
1247 /**
1248 * Checks if a format description is an arithmetic format
1249 *
1250 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
1251 */
1252 static inline boolean
1253 is_arithmetic_format(const struct util_format_description *format_desc)
1254 {
1255 boolean arith = false;
1256 unsigned i;
1257
1258 for (i = 0; i < format_desc->nr_channels; ++i) {
1259 arith |= format_desc->channel[i].size != format_desc->channel[0].size;
1260 arith |= (format_desc->channel[i].size % 8) != 0;
1261 }
1262
1263 return arith;
1264 }
1265
1266
1267 /**
1268 * Checks if this format requires special handling due to required expansion
1269 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
1270 * SoA conversion.
1271 */
1272 static inline boolean
1273 format_expands_to_float_soa(const struct util_format_description *format_desc)
1274 {
1275 if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
1276 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
1277 return true;
1278 }
1279 return false;
1280 }
1281
1282
1283 /**
1284 * Retrieves the type representing the memory layout for a format
1285 *
1286 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
1287 */
1288 static inline void
1289 lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
1290 struct lp_type* type)
1291 {
1292 unsigned i;
1293 unsigned chan;
1294
1295 if (format_expands_to_float_soa(format_desc)) {
1296 /* just make this a uint with width of block */
1297 type->floating = false;
1298 type->fixed = false;
1299 type->sign = false;
1300 type->norm = false;
1301 type->width = format_desc->block.bits;
1302 type->length = 1;
1303 return;
1304 }
1305
1306 for (i = 0; i < 4; i++)
1307 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1308 break;
1309 chan = i;
1310
1311 memset(type, 0, sizeof(struct lp_type));
1312 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1313 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1314 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1315 type->norm = format_desc->channel[chan].normalized;
1316
1317 if (is_arithmetic_format(format_desc)) {
1318 type->width = 0;
1319 type->length = 1;
1320
1321 for (i = 0; i < format_desc->nr_channels; ++i) {
1322 type->width += format_desc->channel[i].size;
1323 }
1324 } else {
1325 type->width = format_desc->channel[chan].size;
1326 type->length = format_desc->nr_channels;
1327 }
1328 }
1329
1330
1331 /**
1332 * Retrieves the type for a format which is usable in the blending code.
1333 *
1334 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1335 */
1336 static inline void
1337 lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1338 struct lp_type* type)
1339 {
1340 unsigned i;
1341 unsigned chan;
1342
1343 if (format_expands_to_float_soa(format_desc)) {
1344 /* always use ordinary floats for blending */
1345 type->floating = true;
1346 type->fixed = false;
1347 type->sign = true;
1348 type->norm = false;
1349 type->width = 32;
1350 type->length = 4;
1351 return;
1352 }
1353
1354 for (i = 0; i < 4; i++)
1355 if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1356 break;
1357 chan = i;
1358
1359 memset(type, 0, sizeof(struct lp_type));
1360 type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1361 type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1362 type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1363 type->norm = format_desc->channel[chan].normalized;
1364 type->width = format_desc->channel[chan].size;
1365 type->length = format_desc->nr_channels;
1366
1367 for (i = 1; i < format_desc->nr_channels; ++i) {
1368 if (format_desc->channel[i].size > type->width)
1369 type->width = format_desc->channel[i].size;
1370 }
1371
1372 if (type->floating) {
1373 type->width = 32;
1374 } else {
1375 if (type->width <= 8) {
1376 type->width = 8;
1377 } else if (type->width <= 16) {
1378 type->width = 16;
1379 } else {
1380 type->width = 32;
1381 }
1382 }
1383
1384 if (is_arithmetic_format(format_desc) && type->length == 3) {
1385 type->length = 4;
1386 }
1387 }
1388
1389
1390 /**
1391 * Scale a normalized value from src_bits to dst_bits.
1392 *
1393 * The exact calculation is
1394 *
1395 * dst = iround(src * dst_mask / src_mask)
1396 *
1397 * or with integer rounding
1398 *
1399 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1400 *
1401 * where
1402 *
1403 * src_mask = (1 << src_bits) - 1
1404 * dst_mask = (1 << dst_bits) - 1
1405 *
1406 * but we try to avoid division and multiplication through shifts.
1407 */
1408 static inline LLVMValueRef
1409 scale_bits(struct gallivm_state *gallivm,
1410 int src_bits,
1411 int dst_bits,
1412 LLVMValueRef src,
1413 struct lp_type src_type)
1414 {
1415 LLVMBuilderRef builder = gallivm->builder;
1416 LLVMValueRef result = src;
1417
1418 if (dst_bits < src_bits) {
1419 int delta_bits = src_bits - dst_bits;
1420
1421 if (delta_bits <= dst_bits) {
1422 /*
1423 * Approximate the rescaling with a single shift.
1424 *
1425 * This gives the wrong rounding.
1426 */
1427
1428 result = LLVMBuildLShr(builder,
1429 src,
1430 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1431 "");
1432
1433 } else {
1434 /*
1435 * Try more accurate rescaling.
1436 */
1437
1438 /*
1439 * Drop the least significant bits to make space for the multiplication.
1440 *
1441 * XXX: A better approach would be to use a wider integer type as intermediate. But
1442 * this is enough to convert alpha from 16bits -> 2 when rendering to
1443 * PIPE_FORMAT_R10G10B10A2_UNORM.
1444 */
1445 result = LLVMBuildLShr(builder,
1446 src,
1447 lp_build_const_int_vec(gallivm, src_type, dst_bits),
1448 "");
1449
1450
1451 result = LLVMBuildMul(builder,
1452 result,
1453 lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1454 "");
1455
1456 /*
1457 * Add a rounding term before the division.
1458 *
1459 * TODO: Handle signed integers too.
1460 */
1461 if (!src_type.sign) {
1462 result = LLVMBuildAdd(builder,
1463 result,
1464 lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1465 "");
1466 }
1467
1468 /*
1469 * Approximate the division by src_mask with a src_bits shift.
1470 *
1471 * Given the src has already been shifted by dst_bits, all we need
1472 * to do is to shift by the difference.
1473 */
1474
1475 result = LLVMBuildLShr(builder,
1476 result,
1477 lp_build_const_int_vec(gallivm, src_type, delta_bits),
1478 "");
1479 }
1480
1481 } else if (dst_bits > src_bits) {
1482 /* Scale up bits */
1483 int db = dst_bits - src_bits;
1484
1485 /* Shift left by difference in bits */
1486 result = LLVMBuildShl(builder,
1487 src,
1488 lp_build_const_int_vec(gallivm, src_type, db),
1489 "");
1490
1491 if (db <= src_bits) {
1492 /* Enough bits in src to fill the remainder */
1493 LLVMValueRef lower = LLVMBuildLShr(builder,
1494 src,
1495 lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1496 "");
1497
1498 result = LLVMBuildOr(builder, result, lower, "");
1499 } else if (db > src_bits) {
1500 /* Need to repeatedly copy src bits to fill remainder in dst */
1501 unsigned n;
1502
1503 for (n = src_bits; n < dst_bits; n *= 2) {
1504 LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1505
1506 result = LLVMBuildOr(builder,
1507 result,
1508 LLVMBuildLShr(builder, result, shuv, ""),
1509 "");
1510 }
1511 }
1512 }
1513
1514 return result;
1515 }
1516
1517 /**
1518 * If RT is a smallfloat (needing denorms) format
1519 */
1520 static inline int
1521 have_smallfloat_format(struct lp_type dst_type,
1522 enum pipe_format format)
1523 {
1524 return ((dst_type.floating && dst_type.width != 32) ||
1525 /* due to format handling hacks this format doesn't have floating set
1526 * here (and actually has width set to 32 too) so special case this. */
1527 (format == PIPE_FORMAT_R11G11B10_FLOAT));
1528 }
1529
1530
1531 /**
1532 * Convert from memory format to blending format
1533 *
1534 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1535 */
1536 static void
1537 convert_to_blend_type(struct gallivm_state *gallivm,
1538 unsigned block_size,
1539 const struct util_format_description *src_fmt,
1540 struct lp_type src_type,
1541 struct lp_type dst_type,
1542 LLVMValueRef* src, // and dst
1543 unsigned num_srcs)
1544 {
1545 LLVMValueRef *dst = src;
1546 LLVMBuilderRef builder = gallivm->builder;
1547 struct lp_type blend_type;
1548 struct lp_type mem_type;
1549 unsigned i, j;
1550 unsigned pixels = block_size / num_srcs;
1551 bool is_arith;
1552
1553 /*
1554 * full custom path for packed floats and srgb formats - none of the later
1555 * functions would do anything useful, and given the lp_type representation they
1556 * can't be fixed. Should really have some SoA blend path for these kind of
1557 * formats rather than hacking them in here.
1558 */
1559 if (format_expands_to_float_soa(src_fmt)) {
1560 LLVMValueRef tmpsrc[4];
1561 /*
1562 * This is pretty suboptimal for this case blending in SoA would be much
1563 * better, since conversion gets us SoA values so need to convert back.
1564 */
1565 assert(src_type.width == 32 || src_type.width == 16);
1566 assert(dst_type.floating);
1567 assert(dst_type.width == 32);
1568 assert(dst_type.length % 4 == 0);
1569 assert(num_srcs % 4 == 0);
1570
1571 if (src_type.width == 16) {
1572 /* expand 4x16bit values to 4x32bit */
1573 struct lp_type type32x4 = src_type;
1574 LLVMTypeRef ltype32x4;
1575 unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1576 type32x4.width = 32;
1577 ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1578 for (i = 0; i < num_fetch; i++) {
1579 src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1580 }
1581 src_type.width = 32;
1582 }
1583 for (i = 0; i < 4; i++) {
1584 tmpsrc[i] = src[i];
1585 }
1586 for (i = 0; i < num_srcs / 4; i++) {
1587 LLVMValueRef tmpsoa[4];
1588 LLVMValueRef tmps = tmpsrc[i];
1589 if (dst_type.length == 8) {
1590 LLVMValueRef shuffles[8];
1591 unsigned j;
1592 /* fetch was 4 values but need 8-wide output values */
1593 tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1594 /*
1595 * for 8-wide aos transpose would give us wrong order not matching
1596 * incoming converted fs values and mask. ARGH.
1597 */
1598 for (j = 0; j < 4; j++) {
1599 shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1600 shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1601 }
1602 tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1603 LLVMConstVector(shuffles, 8), "");
1604 }
1605 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1606 lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1607 }
1608 else {
1609 lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1610 }
1611 lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1612 }
1613 return;
1614 }
1615
1616 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1617 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1618
1619 /* Is the format arithmetic */
1620 is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1621 is_arith &= !(mem_type.width == 16 && mem_type.floating);
1622
1623 /* Pad if necessary */
1624 if (!is_arith && src_type.length < dst_type.length) {
1625 for (i = 0; i < num_srcs; ++i) {
1626 dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1627 }
1628
1629 src_type.length = dst_type.length;
1630 }
1631
1632 /* Special case for half-floats */
1633 if (mem_type.width == 16 && mem_type.floating) {
1634 assert(blend_type.width == 32 && blend_type.floating);
1635 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1636 is_arith = false;
1637 }
1638
1639 if (!is_arith) {
1640 return;
1641 }
1642
1643 src_type.width = blend_type.width * blend_type.length;
1644 blend_type.length *= pixels;
1645 src_type.length *= pixels / (src_type.length / mem_type.length);
1646
1647 for (i = 0; i < num_srcs; ++i) {
1648 LLVMValueRef chans[4];
1649 LLVMValueRef res = NULL;
1650
1651 dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1652
1653 for (j = 0; j < src_fmt->nr_channels; ++j) {
1654 unsigned mask = 0;
1655 unsigned sa = src_fmt->channel[j].shift;
1656 #if UTIL_ARCH_LITTLE_ENDIAN
1657 unsigned from_lsb = j;
1658 #else
1659 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1660 #endif
1661
1662 mask = (1 << src_fmt->channel[j].size) - 1;
1663
1664 /* Extract bits from source */
1665 chans[j] = LLVMBuildLShr(builder,
1666 dst[i],
1667 lp_build_const_int_vec(gallivm, src_type, sa),
1668 "");
1669
1670 chans[j] = LLVMBuildAnd(builder,
1671 chans[j],
1672 lp_build_const_int_vec(gallivm, src_type, mask),
1673 "");
1674
1675 /* Scale bits */
1676 if (src_type.norm) {
1677 chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1678 blend_type.width, chans[j], src_type);
1679 }
1680
1681 /* Insert bits into correct position */
1682 chans[j] = LLVMBuildShl(builder,
1683 chans[j],
1684 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1685 "");
1686
1687 if (j == 0) {
1688 res = chans[j];
1689 } else {
1690 res = LLVMBuildOr(builder, res, chans[j], "");
1691 }
1692 }
1693
1694 dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1695 }
1696 }
1697
1698
1699 /**
1700 * Convert from blending format to memory format
1701 *
1702 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1703 */
1704 static void
1705 convert_from_blend_type(struct gallivm_state *gallivm,
1706 unsigned block_size,
1707 const struct util_format_description *src_fmt,
1708 struct lp_type src_type,
1709 struct lp_type dst_type,
1710 LLVMValueRef* src, // and dst
1711 unsigned num_srcs)
1712 {
1713 LLVMValueRef* dst = src;
1714 unsigned i, j, k;
1715 struct lp_type mem_type;
1716 struct lp_type blend_type;
1717 LLVMBuilderRef builder = gallivm->builder;
1718 unsigned pixels = block_size / num_srcs;
1719 bool is_arith;
1720
1721 /*
1722 * full custom path for packed floats and srgb formats - none of the later
1723 * functions would do anything useful, and given the lp_type representation they
1724 * can't be fixed. Should really have some SoA blend path for these kind of
1725 * formats rather than hacking them in here.
1726 */
1727 if (format_expands_to_float_soa(src_fmt)) {
1728 /*
1729 * This is pretty suboptimal for this case blending in SoA would be much
1730 * better - we need to transpose the AoS values back to SoA values for
1731 * conversion/packing.
1732 */
1733 assert(src_type.floating);
1734 assert(src_type.width == 32);
1735 assert(src_type.length % 4 == 0);
1736 assert(dst_type.width == 32 || dst_type.width == 16);
1737
1738 for (i = 0; i < num_srcs / 4; i++) {
1739 LLVMValueRef tmpsoa[4], tmpdst;
1740 lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
1741 /* really really need SoA here */
1742
1743 if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1744 tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
1745 }
1746 else {
1747 tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
1748 src_type, tmpsoa);
1749 }
1750
1751 if (src_type.length == 8) {
1752 LLVMValueRef tmpaos, shuffles[8];
1753 unsigned j;
1754 /*
1755 * for 8-wide aos transpose has given us wrong order not matching
1756 * output order. HMPF. Also need to split the output values manually.
1757 */
1758 for (j = 0; j < 4; j++) {
1759 shuffles[j * 2] = lp_build_const_int32(gallivm, j);
1760 shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
1761 }
1762 tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
1763 LLVMConstVector(shuffles, 8), "");
1764 src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
1765 src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
1766 }
1767 else {
1768 src[i] = tmpdst;
1769 }
1770 }
1771 if (dst_type.width == 16) {
1772 struct lp_type type16x8 = dst_type;
1773 struct lp_type type32x4 = dst_type;
1774 LLVMTypeRef ltype16x4, ltypei64, ltypei128;
1775 unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1776 type16x8.length = 8;
1777 type32x4.width = 32;
1778 ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
1779 ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
1780 ltype16x4 = lp_build_vec_type(gallivm, dst_type);
1781 /* We could do vector truncation but it doesn't generate very good code */
1782 for (i = 0; i < num_fetch; i++) {
1783 src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
1784 src[i], lp_build_zero(gallivm, type32x4));
1785 src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
1786 src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
1787 src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
1788 }
1789 }
1790 return;
1791 }
1792
1793 lp_mem_type_from_format_desc(src_fmt, &mem_type);
1794 lp_blend_type_from_format_desc(src_fmt, &blend_type);
1795
1796 is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
1797
1798 /* Special case for half-floats */
1799 if (mem_type.width == 16 && mem_type.floating) {
1800 int length = dst_type.length;
1801 assert(blend_type.width == 32 && blend_type.floating);
1802
1803 dst_type.length = src_type.length;
1804
1805 lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1806
1807 dst_type.length = length;
1808 is_arith = false;
1809 }
1810
1811 /* Remove any padding */
1812 if (!is_arith && (src_type.length % mem_type.length)) {
1813 src_type.length -= (src_type.length % mem_type.length);
1814
1815 for (i = 0; i < num_srcs; ++i) {
1816 dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
1817 }
1818 }
1819
1820 /* No bit arithmetic to do */
1821 if (!is_arith) {
1822 return;
1823 }
1824
1825 src_type.length = pixels;
1826 src_type.width = blend_type.length * blend_type.width;
1827 dst_type.length = pixels;
1828
1829 for (i = 0; i < num_srcs; ++i) {
1830 LLVMValueRef chans[4];
1831 LLVMValueRef res = NULL;
1832
1833 dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1834
1835 for (j = 0; j < src_fmt->nr_channels; ++j) {
1836 unsigned mask = 0;
1837 unsigned sa = src_fmt->channel[j].shift;
1838 unsigned sz_a = src_fmt->channel[j].size;
1839 #if UTIL_ARCH_LITTLE_ENDIAN
1840 unsigned from_lsb = j;
1841 #else
1842 unsigned from_lsb = src_fmt->nr_channels - j - 1;
1843 #endif
1844
1845 assert(blend_type.width > src_fmt->channel[j].size);
1846
1847 for (k = 0; k < blend_type.width; ++k) {
1848 mask |= 1 << k;
1849 }
1850
1851 /* Extract bits */
1852 chans[j] = LLVMBuildLShr(builder,
1853 dst[i],
1854 lp_build_const_int_vec(gallivm, src_type,
1855 from_lsb * blend_type.width),
1856 "");
1857
1858 chans[j] = LLVMBuildAnd(builder,
1859 chans[j],
1860 lp_build_const_int_vec(gallivm, src_type, mask),
1861 "");
1862
1863 /* Scale down bits */
1864 if (src_type.norm) {
1865 chans[j] = scale_bits(gallivm, blend_type.width,
1866 src_fmt->channel[j].size, chans[j], src_type);
1867 } else if (!src_type.floating && sz_a < blend_type.width) {
1868 LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);
1869 LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans[j], mask_val, "");
1870 chans[j] = LLVMBuildSelect(builder, mask, mask_val, chans[j], "");
1871 }
1872
1873 /* Insert bits */
1874 chans[j] = LLVMBuildShl(builder,
1875 chans[j],
1876 lp_build_const_int_vec(gallivm, src_type, sa),
1877 "");
1878
1879 sa += src_fmt->channel[j].size;
1880
1881 if (j == 0) {
1882 res = chans[j];
1883 } else {
1884 res = LLVMBuildOr(builder, res, chans[j], "");
1885 }
1886 }
1887
1888 assert (dst_type.width != 24);
1889
1890 dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
1891 }
1892 }
1893
1894
1895 /**
1896 * Convert alpha to same blend type as src
1897 */
1898 static void
1899 convert_alpha(struct gallivm_state *gallivm,
1900 struct lp_type row_type,
1901 struct lp_type alpha_type,
1902 const unsigned block_size,
1903 const unsigned block_height,
1904 const unsigned src_count,
1905 const unsigned dst_channels,
1906 const bool pad_inline,
1907 LLVMValueRef* src_alpha)
1908 {
1909 LLVMBuilderRef builder = gallivm->builder;
1910 unsigned i, j;
1911 unsigned length = row_type.length;
1912 row_type.length = alpha_type.length;
1913
1914 /* Twiddle the alpha to match pixels */
1915 lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
1916
1917 /*
1918 * TODO this should use single lp_build_conv call for
1919 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1920 */
1921 for (i = 0; i < block_height; ++i) {
1922 lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
1923 }
1924
1925 alpha_type = row_type;
1926 row_type.length = length;
1927
1928 /* If only one channel we can only need the single alpha value per pixel */
1929 if (src_count == 1 && dst_channels == 1) {
1930
1931 lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
1932 } else {
1933 /* If there are more srcs than rows then we need to split alpha up */
1934 if (src_count > block_height) {
1935 for (i = src_count; i > 0; --i) {
1936 unsigned pixels = block_size / src_count;
1937 unsigned idx = i - 1;
1938
1939 src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
1940 (idx * pixels) % 4, pixels);
1941 }
1942 }
1943
1944 /* If there is a src for each pixel broadcast the alpha across whole row */
1945 if (src_count == block_size) {
1946 for (i = 0; i < src_count; ++i) {
1947 src_alpha[i] = lp_build_broadcast(gallivm,
1948 lp_build_vec_type(gallivm, row_type), src_alpha[i]);
1949 }
1950 } else {
1951 unsigned pixels = block_size / src_count;
1952 unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
1953 unsigned alpha_span = 1;
1954 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1955
1956 /* Check if we need 2 src_alphas for our shuffles */
1957 if (pixels > alpha_type.length) {
1958 alpha_span = 2;
1959 }
1960
1961 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1962 for (j = 0; j < row_type.length; ++j) {
1963 if (j < pixels * channels) {
1964 shuffles[j] = lp_build_const_int32(gallivm, j / channels);
1965 } else {
1966 shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1967 }
1968 }
1969
1970 for (i = 0; i < src_count; ++i) {
1971 unsigned idx1 = i, idx2 = i;
1972
1973 if (alpha_span > 1){
1974 idx1 *= alpha_span;
1975 idx2 = idx1 + 1;
1976 }
1977
1978 src_alpha[i] = LLVMBuildShuffleVector(builder,
1979 src_alpha[idx1],
1980 src_alpha[idx2],
1981 LLVMConstVector(shuffles, row_type.length),
1982 "");
1983 }
1984 }
1985 }
1986 }
1987
1988
1989 /**
1990 * Generates the blend function for unswizzled colour buffers
1991 * Also generates the read & write from colour buffer
1992 */
1993 static void
1994 generate_unswizzled_blend(struct gallivm_state *gallivm,
1995 unsigned rt,
1996 struct lp_fragment_shader_variant *variant,
1997 enum pipe_format out_format,
1998 unsigned int num_fs,
1999 struct lp_type fs_type,
2000 LLVMValueRef* fs_mask,
2001 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
2002 LLVMValueRef context_ptr,
2003 LLVMValueRef color_ptr,
2004 LLVMValueRef stride,
2005 unsigned partial_mask,
2006 boolean do_branch)
2007 {
2008 const unsigned alpha_channel = 3;
2009 const unsigned block_width = LP_RASTER_BLOCK_SIZE;
2010 const unsigned block_height = LP_RASTER_BLOCK_SIZE;
2011 const unsigned block_size = block_width * block_height;
2012 const unsigned lp_integer_vector_width = 128;
2013
2014 LLVMBuilderRef builder = gallivm->builder;
2015 LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
2016 LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
2017 LLVMValueRef src_alpha[4 * 4];
2018 LLVMValueRef src1_alpha[4 * 4] = { NULL };
2019 LLVMValueRef src_mask[4 * 4];
2020 LLVMValueRef src[4 * 4];
2021 LLVMValueRef src1[4 * 4];
2022 LLVMValueRef dst[4 * 4];
2023 LLVMValueRef blend_color;
2024 LLVMValueRef blend_alpha;
2025 LLVMValueRef i32_zero;
2026 LLVMValueRef check_mask;
2027 LLVMValueRef undef_src_val;
2028
2029 struct lp_build_mask_context mask_ctx;
2030 struct lp_type mask_type;
2031 struct lp_type blend_type;
2032 struct lp_type row_type;
2033 struct lp_type dst_type;
2034 struct lp_type ls_type;
2035
2036 unsigned char swizzle[TGSI_NUM_CHANNELS];
2037 unsigned vector_width;
2038 unsigned src_channels = TGSI_NUM_CHANNELS;
2039 unsigned dst_channels;
2040 unsigned dst_count;
2041 unsigned src_count;
2042 unsigned i, j;
2043
2044 const struct util_format_description* out_format_desc = util_format_description(out_format);
2045
2046 unsigned dst_alignment;
2047
2048 bool pad_inline = is_arithmetic_format(out_format_desc);
2049 bool has_alpha = false;
2050 const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
2051 util_blend_state_is_dual(&variant->key.blend, 0);
2052
2053 const boolean is_1d = variant->key.resource_1d;
2054 boolean twiddle_after_convert = FALSE;
2055 unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
2056 LLVMValueRef fpstate = 0;
2057
2058 /* Get type from output format */
2059 lp_blend_type_from_format_desc(out_format_desc, &row_type);
2060 lp_mem_type_from_format_desc(out_format_desc, &dst_type);
2061
2062 /*
2063 * Technically this code should go into lp_build_smallfloat_to_float
2064 * and lp_build_float_to_smallfloat but due to the
2065 * http://llvm.org/bugs/show_bug.cgi?id=6393
2066 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2067 * So the ordering is important here and there shouldn't be any
2068 * llvm ir instrunctions in this function before
2069 * this, otherwise half-float format conversions won't work
2070 * (again due to llvm bug #6393).
2071 */
2072 if (have_smallfloat_format(dst_type, out_format)) {
2073 /* We need to make sure that denorms are ok for half float
2074 conversions */
2075 fpstate = lp_build_fpstate_get(gallivm);
2076 lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
2077 }
2078
2079 mask_type = lp_int32_vec4_type();
2080 mask_type.length = fs_type.length;
2081
2082 for (i = num_fs; i < num_fullblock_fs; i++) {
2083 fs_mask[i] = lp_build_zero(gallivm, mask_type);
2084 }
2085
2086 /* Do not bother executing code when mask is empty.. */
2087 if (do_branch) {
2088 check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
2089
2090 for (i = 0; i < num_fullblock_fs; ++i) {
2091 check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
2092 }
2093
2094 lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
2095 lp_build_mask_check(&mask_ctx);
2096 }
2097
2098 partial_mask |= !variant->opaque;
2099 i32_zero = lp_build_const_int32(gallivm, 0);
2100
2101 undef_src_val = lp_build_undef(gallivm, fs_type);
2102
2103 row_type.length = fs_type.length;
2104 vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
2105
2106 /* Compute correct swizzle and count channels */
2107 memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
2108 dst_channels = 0;
2109
2110 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
2111 /* Ensure channel is used */
2112 if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
2113 continue;
2114 }
2115
2116 /* Ensure not already written to (happens in case with GL_ALPHA) */
2117 if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
2118 continue;
2119 }
2120
2121 /* Ensure we havn't already found all channels */
2122 if (dst_channels >= out_format_desc->nr_channels) {
2123 continue;
2124 }
2125
2126 swizzle[out_format_desc->swizzle[i]] = i;
2127 ++dst_channels;
2128
2129 if (i == alpha_channel) {
2130 has_alpha = true;
2131 }
2132 }
2133
2134 if (format_expands_to_float_soa(out_format_desc)) {
2135 /*
2136 * the code above can't work for layout_other
2137 * for srgb it would sort of work but we short-circuit swizzles, etc.
2138 * as that is done as part of unpack / pack.
2139 */
2140 dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2141 has_alpha = true;
2142 swizzle[0] = 0;
2143 swizzle[1] = 1;
2144 swizzle[2] = 2;
2145 swizzle[3] = 3;
2146 pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2147 }
2148
2149 /* If 3 channels then pad to include alpha for 4 element transpose */
2150 if (dst_channels == 3) {
2151 assert (!has_alpha);
2152 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2153 if (swizzle[i] > TGSI_NUM_CHANNELS)
2154 swizzle[i] = 3;
2155 }
2156 if (out_format_desc->nr_channels == 4) {
2157 dst_channels = 4;
2158 /*
2159 * We use alpha from the color conversion, not separate one.
2160 * We had to include it for transpose, hence it will get converted
2161 * too (albeit when doing transpose after conversion, that would
2162 * no longer be the case necessarily).
2163 * (It works only with 4 channel dsts, e.g. rgbx formats, because
2164 * otherwise we really have padding, not alpha, included.)
2165 */
2166 has_alpha = true;
2167 }
2168 }
2169
2170 /*
2171 * Load shader output
2172 */
2173 for (i = 0; i < num_fullblock_fs; ++i) {
2174 /* Always load alpha for use in blending */
2175 LLVMValueRef alpha;
2176 if (i < num_fs) {
2177 alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
2178 }
2179 else {
2180 alpha = undef_src_val;
2181 }
2182
2183 /* Load each channel */
2184 for (j = 0; j < dst_channels; ++j) {
2185 assert(swizzle[j] < 4);
2186 if (i < num_fs) {
2187 fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
2188 }
2189 else {
2190 fs_src[i][j] = undef_src_val;
2191 }
2192 }
2193
2194 /* If 3 channels then pad to include alpha for 4 element transpose */
2195 /*
2196 * XXX If we include that here maybe could actually use it instead of
2197 * separate alpha for blending?
2198 * (Difficult though we actually convert pad channels, not alpha.)
2199 */
2200 if (dst_channels == 3 && !has_alpha) {
2201 fs_src[i][3] = alpha;
2202 }
2203
2204 /* We split the row_mask and row_alpha as we want 128bit interleave */
2205 if (fs_type.length == 8) {
2206 src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
2207 0, src_channels);
2208 src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
2209 src_channels, src_channels);
2210
2211 src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2212 src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2213 src_channels, src_channels);
2214 } else {
2215 src_mask[i] = fs_mask[i];
2216 src_alpha[i] = alpha;
2217 }
2218 }
2219 if (dual_source_blend) {
2220 /* same as above except different src/dst, skip masks and comments... */
2221 for (i = 0; i < num_fullblock_fs; ++i) {
2222 LLVMValueRef alpha;
2223 if (i < num_fs) {
2224 alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
2225 }
2226 else {
2227 alpha = undef_src_val;
2228 }
2229
2230 for (j = 0; j < dst_channels; ++j) {
2231 assert(swizzle[j] < 4);
2232 if (i < num_fs) {
2233 fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
2234 }
2235 else {
2236 fs_src1[i][j] = undef_src_val;
2237 }
2238 }
2239 if (dst_channels == 3 && !has_alpha) {
2240 fs_src1[i][3] = alpha;
2241 }
2242 if (fs_type.length == 8) {
2243 src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2244 src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2245 src_channels, src_channels);
2246 } else {
2247 src1_alpha[i] = alpha;
2248 }
2249 }
2250 }
2251
2252 if (util_format_is_pure_integer(out_format)) {
2253 /*
2254 * In this case fs_type was really ints or uints disguised as floats,
2255 * fix that up now.
2256 */
2257 fs_type.floating = 0;
2258 fs_type.sign = dst_type.sign;
2259 for (i = 0; i < num_fullblock_fs; ++i) {
2260 for (j = 0; j < dst_channels; ++j) {
2261 fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
2262 lp_build_vec_type(gallivm, fs_type), "");
2263 }
2264 if (dst_channels == 3 && !has_alpha) {
2265 fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
2266 lp_build_vec_type(gallivm, fs_type), "");
2267 }
2268 }
2269 }
2270
2271 /*
2272 * We actually should generally do conversion first (for non-1d cases)
2273 * when the blend format is 8 or 16 bits. The reason is obvious,
2274 * there's 2 or 4 times less vectors to deal with for the interleave...
2275 * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2276 * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2277 * unpack only with 128bit vectors).
2278 * Note: for 16bit sizes really need matching pack conversion code
2279 */
2280 if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
2281 twiddle_after_convert = TRUE;
2282 }
2283
2284 /*
2285 * Pixel twiddle from fragment shader order to memory order
2286 */
2287 if (!twiddle_after_convert) {
2288 src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
2289 dst_channels, fs_src, src, pad_inline);
2290 if (dual_source_blend) {
2291 generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
2292 fs_src1, src1, pad_inline);
2293 }
2294 } else {
2295 src_count = num_fullblock_fs * dst_channels;
2296 /*
2297 * We reorder things a bit here, so the cases for 4-wide and 8-wide
2298 * (AVX) turn out the same later when untwiddling/transpose (albeit
2299 * for true AVX2 path untwiddle needs to be different).
2300 * For now just order by colors first (so we can use unpack later).
2301 */
2302 for (j = 0; j < num_fullblock_fs; j++) {
2303 for (i = 0; i < dst_channels; i++) {
2304 src[i*num_fullblock_fs + j] = fs_src[j][i];
2305 if (dual_source_blend) {
2306 src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2307 }
2308 }
2309 }
2310 }
2311
2312 src_channels = dst_channels < 3 ? dst_channels : 4;
2313 if (src_count != num_fullblock_fs * src_channels) {
2314 unsigned ds = src_count / (num_fullblock_fs * src_channels);
2315 row_type.length /= ds;
2316 fs_type.length = row_type.length;
2317 }
2318
2319 blend_type = row_type;
2320 mask_type.length = 4;
2321
2322 /* Convert src to row_type */
2323 if (dual_source_blend) {
2324 struct lp_type old_row_type = row_type;
2325 lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2326 src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2327 }
2328 else {
2329 src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2330 }
2331
2332 /* If the rows are not an SSE vector, combine them to become SSE size! */
2333 if ((row_type.width * row_type.length) % 128) {
2334 unsigned bits = row_type.width * row_type.length;
2335 unsigned combined;
2336
2337 assert(src_count >= (vector_width / bits));
2338
2339 dst_count = src_count / (vector_width / bits);
2340
2341 combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2342 if (dual_source_blend) {
2343 lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2344 }
2345
2346 row_type.length *= combined;
2347 src_count /= combined;
2348
2349 bits = row_type.width * row_type.length;
2350 assert(bits == 128 || bits == 256);
2351 }
2352
2353 if (twiddle_after_convert) {
2354 fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2355 if (dual_source_blend) {
2356 fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2357 }
2358 }
2359
2360 /*
2361 * Blend Colour conversion
2362 */
2363 blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2364 blend_color = LLVMBuildPointerCast(builder, blend_color,
2365 LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2366 blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2367 &i32_zero, 1, ""), "");
2368
2369 /* Convert */
2370 lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2371
2372 if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2373 /*
2374 * since blending is done with floats, there was no conversion.
2375 * However, the rules according to fixed point renderbuffers still
2376 * apply, that is we must clamp inputs to 0.0/1.0.
2377 * (This would apply to separate alpha conversion too but we currently
2378 * force has_alpha to be true.)
2379 * TODO: should skip this with "fake" blend, since post-blend conversion
2380 * will clamp anyway.
2381 * TODO: could also skip this if fragment color clamping is enabled. We
2382 * don't support it natively so it gets baked into the shader however, so
2383 * can't really tell here.
2384 */
2385 struct lp_build_context f32_bld;
2386 assert(row_type.floating);
2387 lp_build_context_init(&f32_bld, gallivm, row_type);
2388 for (i = 0; i < src_count; i++) {
2389 src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2390 }
2391 if (dual_source_blend) {
2392 for (i = 0; i < src_count; i++) {
2393 src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2394 }
2395 }
2396 /* probably can't be different than row_type but better safe than sorry... */
2397 lp_build_context_init(&f32_bld, gallivm, blend_type);
2398 blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2399 }
2400
2401 /* Extract alpha */
2402 blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2403
2404 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2405 pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2406 if (pad_inline) {
2407 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2408 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2409 } else {
2410 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2411 blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2412 }
2413
2414 /*
2415 * Mask conversion
2416 */
2417 lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2418
2419 if (src_count < block_height) {
2420 lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2421 } else if (src_count > block_height) {
2422 for (i = src_count; i > 0; --i) {
2423 unsigned pixels = block_size / src_count;
2424 unsigned idx = i - 1;
2425
2426 src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2427 (idx * pixels) % 4, pixels);
2428 }
2429 }
2430
2431 assert(mask_type.width == 32);
2432
2433 for (i = 0; i < src_count; ++i) {
2434 unsigned pixels = block_size / src_count;
2435 unsigned pixel_width = row_type.width * dst_channels;
2436
2437 if (pixel_width == 24) {
2438 mask_type.width = 8;
2439 mask_type.length = vector_width / mask_type.width;
2440 } else {
2441 mask_type.length = pixels;
2442 mask_type.width = row_type.width * dst_channels;
2443
2444 /*
2445 * If mask_type width is smaller than 32bit, this doesn't quite
2446 * generate the most efficient code (could use some pack).
2447 */
2448 src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2449 lp_build_int_vec_type(gallivm, mask_type), "");
2450
2451 mask_type.length *= dst_channels;
2452 mask_type.width /= dst_channels;
2453 }
2454
2455 src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2456 lp_build_int_vec_type(gallivm, mask_type), "");
2457 src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2458 }
2459
2460 /*
2461 * Alpha conversion
2462 */
2463 if (!has_alpha) {
2464 struct lp_type alpha_type = fs_type;
2465 alpha_type.length = 4;
2466 convert_alpha(gallivm, row_type, alpha_type,
2467 block_size, block_height,
2468 src_count, dst_channels,
2469 pad_inline, src_alpha);
2470 if (dual_source_blend) {
2471 convert_alpha(gallivm, row_type, alpha_type,
2472 block_size, block_height,
2473 src_count, dst_channels,
2474 pad_inline, src1_alpha);
2475 }
2476 }
2477
2478
2479 /*
2480 * Load dst from memory
2481 */
2482 if (src_count < block_height) {
2483 dst_count = block_height;
2484 } else {
2485 dst_count = src_count;
2486 }
2487
2488 dst_type.length *= block_size / dst_count;
2489
2490 if (format_expands_to_float_soa(out_format_desc)) {
2491 /*
2492 * we need multiple values at once for the conversion, so can as well
2493 * load them vectorized here too instead of concatenating later.
2494 * (Still need concatenation later for 8-wide vectors).
2495 */
2496 dst_count = block_height;
2497 dst_type.length = block_width;
2498 }
2499
2500 /*
2501 * Compute the alignment of the destination pointer in bytes
2502 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2503 * are always aligned by MIN2(16, fetch_width) except for buffers (not
2504 * 1d tex but can't distinguish here) so need to stick with per-pixel
2505 * alignment in this case.
2506 */
2507 if (is_1d) {
2508 dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2509 }
2510 else {
2511 dst_alignment = dst_type.length * dst_type.width / 8;
2512 }
2513 /* Force power-of-two alignment by extracting only the least-significant-bit */
2514 dst_alignment = 1 << (ffs(dst_alignment) - 1);
2515 /*
2516 * Resource base and stride pointers are aligned to 16 bytes, so that's
2517 * the maximum alignment we can guarantee
2518 */
2519 dst_alignment = MIN2(16, dst_alignment);
2520
2521 ls_type = dst_type;
2522
2523 if (dst_count > src_count) {
2524 if ((dst_type.width == 8 || dst_type.width == 16) &&
2525 util_is_power_of_two_or_zero(dst_type.length) &&
2526 dst_type.length * dst_type.width < 128) {
2527 /*
2528 * Never try to load values as 4xi8 which we will then
2529 * concatenate to larger vectors. This gives llvm a real
2530 * headache (the problem is the type legalizer (?) will
2531 * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2532 * then the shuffles to concatenate are more or less impossible
2533 * - llvm is easily capable of generating a sequence of 32
2534 * pextrb/pinsrb instructions for that. Albeit it appears to
2535 * be fixed in llvm 4.0. So, load and concatenate with 32bit
2536 * width to avoid the trouble (16bit seems not as bad, llvm
2537 * probably recognizes the load+shuffle as only one shuffle
2538 * is necessary, but we can do just the same anyway).
2539 */
2540 ls_type.length = dst_type.length * dst_type.width / 32;
2541 ls_type.width = 32;
2542 }
2543 }
2544
2545 if (is_1d) {
2546 load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2547 dst, ls_type, dst_count / 4, dst_alignment);
2548 for (i = dst_count / 4; i < dst_count; i++) {
2549 dst[i] = lp_build_undef(gallivm, ls_type);
2550 }
2551
2552 }
2553 else {
2554 load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2555 dst, ls_type, dst_count, dst_alignment);
2556 }
2557
2558
2559 /*
2560 * Convert from dst/output format to src/blending format.
2561 *
2562 * This is necessary as we can only read 1 row from memory at a time,
2563 * so the minimum dst_count will ever be at this point is 4.
2564 *
2565 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2566 * this will take the 4 dsts and combine them into 1 src so we can perform blending
2567 * on all 16 pixels in that single vector at once.
2568 */
2569 if (dst_count > src_count) {
2570 if (ls_type.length != dst_type.length && ls_type.length == 1) {
2571 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2572 LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2573 for (i = 0; i < dst_count; i++) {
2574 dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2575 }
2576 }
2577
2578 lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2579
2580 if (ls_type.length != dst_type.length) {
2581 struct lp_type tmp_type = dst_type;
2582 tmp_type.length = dst_type.length * 4 / src_count;
2583 for (i = 0; i < src_count; i++) {
2584 dst[i] = LLVMBuildBitCast(builder, dst[i],
2585 lp_build_vec_type(gallivm, tmp_type), "");
2586 }
2587 }
2588 }
2589
2590 /*
2591 * Blending
2592 */
2593 /* XXX this is broken for RGB8 formats -
2594 * they get expanded from 12 to 16 elements (to include alpha)
2595 * by convert_to_blend_type then reduced to 15 instead of 12
2596 * by convert_from_blend_type (a simple fix though breaks A8...).
2597 * R16G16B16 also crashes differently however something going wrong
2598 * inside llvm handling npot vector sizes seemingly.
2599 * It seems some cleanup could be done here (like skipping conversion/blend
2600 * when not needed).
2601 */
2602 convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2603 row_type, dst, src_count);
2604
2605 /*
2606 * FIXME: Really should get logic ops / masks out of generic blend / row
2607 * format. Logic ops will definitely not work on the blend float format
2608 * used for SRGB here and I think OpenGL expects this to work as expected
2609 * (that is incoming values converted to srgb then logic op applied).
2610 */
2611 for (i = 0; i < src_count; ++i) {
2612 dst[i] = lp_build_blend_aos(gallivm,
2613 &variant->key.blend,
2614 out_format,
2615 row_type,
2616 rt,
2617 src[i],
2618 has_alpha ? NULL : src_alpha[i],
2619 src1[i],
2620 has_alpha ? NULL : src1_alpha[i],
2621 dst[i],
2622 partial_mask ? src_mask[i] : NULL,
2623 blend_color,
2624 has_alpha ? NULL : blend_alpha,
2625 swizzle,
2626 pad_inline ? 4 : dst_channels);
2627 }
2628
2629 convert_from_blend_type(gallivm, block_size, out_format_desc,
2630 row_type, dst_type, dst, src_count);
2631
2632 /* Split the blend rows back to memory rows */
2633 if (dst_count > src_count) {
2634 row_type.length = dst_type.length * (dst_count / src_count);
2635
2636 if (src_count == 1) {
2637 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2638 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2639
2640 row_type.length /= 2;
2641 src_count *= 2;
2642 }
2643
2644 dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2645 dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2646 dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2647 dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2648
2649 row_type.length /= 2;
2650 src_count *= 2;
2651 }
2652
2653 /*
2654 * Store blend result to memory
2655 */
2656 if (is_1d) {
2657 store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2658 dst, dst_type, dst_count / 4, dst_alignment);
2659 }
2660 else {
2661 store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2662 dst, dst_type, dst_count, dst_alignment);
2663 }
2664
2665 if (have_smallfloat_format(dst_type, out_format)) {
2666 lp_build_fpstate_set(gallivm, fpstate);
2667 }
2668
2669 if (do_branch) {
2670 lp_build_mask_end(&mask_ctx);
2671 }
2672 }
2673
2674
2675 /**
2676 * Generate the runtime callable function for the whole fragment pipeline.
2677 * Note that the function which we generate operates on a block of 16
2678 * pixels at at time. The block contains 2x2 quads. Each quad contains
2679 * 2x2 pixels.
2680 */
2681 static void
2682 generate_fragment(struct llvmpipe_context *lp,
2683 struct lp_fragment_shader *shader,
2684 struct lp_fragment_shader_variant *variant,
2685 unsigned partial_mask)
2686 {
2687 struct gallivm_state *gallivm = variant->gallivm;
2688 struct lp_fragment_shader_variant_key *key = &variant->key;
2689 struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2690 char func_name[64];
2691 struct lp_type fs_type;
2692 struct lp_type blend_type;
2693 LLVMTypeRef fs_elem_type;
2694 LLVMTypeRef blend_vec_type;
2695 LLVMTypeRef arg_types[15];
2696 LLVMTypeRef func_type;
2697 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
2698 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
2699 LLVMValueRef context_ptr;
2700 LLVMValueRef x;
2701 LLVMValueRef y;
2702 LLVMValueRef a0_ptr;
2703 LLVMValueRef dadx_ptr;
2704 LLVMValueRef dady_ptr;
2705 LLVMValueRef color_ptr_ptr;
2706 LLVMValueRef stride_ptr;
2707 LLVMValueRef color_sample_stride_ptr;
2708 LLVMValueRef depth_ptr;
2709 LLVMValueRef depth_stride;
2710 LLVMValueRef depth_sample_stride;
2711 LLVMValueRef mask_input;
2712 LLVMValueRef thread_data_ptr;
2713 LLVMBasicBlockRef block;
2714 LLVMBuilderRef builder;
2715 struct lp_build_sampler_soa *sampler;
2716 struct lp_build_image_soa *image;
2717 struct lp_build_interp_soa_context interp;
2718 LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
2719 LLVMValueRef fs_out_color[LP_MAX_SAMPLES][PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
2720 LLVMValueRef function;
2721 LLVMValueRef facing;
2722 unsigned num_fs;
2723 unsigned i;
2724 unsigned chan;
2725 unsigned cbuf;
2726 boolean cbuf0_write_all;
2727 const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
2728 util_blend_state_is_dual(&key->blend, 0);
2729
2730 assert(lp_native_vector_width / 32 >= 4);
2731
2732 /* Adjust color input interpolation according to flatshade state:
2733 */
2734 memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
2735 for (i = 0; i < shader->info.base.num_inputs; i++) {
2736 if (inputs[i].interp == LP_INTERP_COLOR) {
2737 if (key->flatshade)
2738 inputs[i].interp = LP_INTERP_CONSTANT;
2739 else
2740 inputs[i].interp = LP_INTERP_PERSPECTIVE;
2741 }
2742 }
2743
2744 /* check if writes to cbuf[0] are to be copied to all cbufs */
2745 cbuf0_write_all =
2746 shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
2747
2748 /* TODO: actually pick these based on the fs and color buffer
2749 * characteristics. */
2750
2751 memset(&fs_type, 0, sizeof fs_type);
2752 fs_type.floating = TRUE; /* floating point values */
2753 fs_type.sign = TRUE; /* values are signed */
2754 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2755 fs_type.width = 32; /* 32-bit float */
2756 fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
2757
2758 memset(&blend_type, 0, sizeof blend_type);
2759 blend_type.floating = FALSE; /* values are integers */
2760 blend_type.sign = FALSE; /* values are unsigned */
2761 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
2762 blend_type.width = 8; /* 8-bit ubyte values */
2763 blend_type.length = 16; /* 16 elements per vector */
2764
2765 /*
2766 * Generate the function prototype. Any change here must be reflected in
2767 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2768 */
2769
2770 fs_elem_type = lp_build_elem_type(gallivm, fs_type);
2771
2772 blend_vec_type = lp_build_vec_type(gallivm, blend_type);
2773
2774 snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
2775 shader->no, variant->no, partial_mask ? "partial" : "whole");
2776
2777 arg_types[0] = variant->jit_context_ptr_type; /* context */
2778 arg_types[1] = int32_type; /* x */
2779 arg_types[2] = int32_type; /* y */
2780 arg_types[3] = int32_type; /* facing */
2781 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
2782 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
2783 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
2784 arg_types[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0); /* color */
2785 arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
2786 arg_types[9] = LLVMInt64TypeInContext(gallivm->context); /* mask_input */
2787 arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
2788 arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
2789 arg_types[12] = int32_type; /* depth_stride */
2790 arg_types[13] = LLVMPointerType(int32_type, 0); /* color sample strides */
2791 arg_types[14] = int32_type; /* depth sample stride */
2792
2793 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
2794 arg_types, ARRAY_SIZE(arg_types), 0);
2795
2796 function = LLVMAddFunction(gallivm->module, func_name, func_type);
2797 LLVMSetFunctionCallConv(function, LLVMCCallConv);
2798
2799 variant->function[partial_mask] = function;
2800
2801 /* XXX: need to propagate noalias down into color param now we are
2802 * passing a pointer-to-pointer?
2803 */
2804 for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
2805 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2806 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
2807
2808 context_ptr = LLVMGetParam(function, 0);
2809 x = LLVMGetParam(function, 1);
2810 y = LLVMGetParam(function, 2);
2811 facing = LLVMGetParam(function, 3);
2812 a0_ptr = LLVMGetParam(function, 4);
2813 dadx_ptr = LLVMGetParam(function, 5);
2814 dady_ptr = LLVMGetParam(function, 6);
2815 color_ptr_ptr = LLVMGetParam(function, 7);
2816 depth_ptr = LLVMGetParam(function, 8);
2817 mask_input = LLVMGetParam(function, 9);
2818 thread_data_ptr = LLVMGetParam(function, 10);
2819 stride_ptr = LLVMGetParam(function, 11);
2820 depth_stride = LLVMGetParam(function, 12);
2821 color_sample_stride_ptr = LLVMGetParam(function, 13);
2822 depth_sample_stride = LLVMGetParam(function, 14);
2823
2824 lp_build_name(context_ptr, "context");
2825 lp_build_name(x, "x");
2826 lp_build_name(y, "y");
2827 lp_build_name(a0_ptr, "a0");
2828 lp_build_name(dadx_ptr, "dadx");
2829 lp_build_name(dady_ptr, "dady");
2830 lp_build_name(color_ptr_ptr, "color_ptr_ptr");
2831 lp_build_name(depth_ptr, "depth");
2832 lp_build_name(mask_input, "mask_input");
2833 lp_build_name(thread_data_ptr, "thread_data");
2834 lp_build_name(stride_ptr, "stride_ptr");
2835 lp_build_name(depth_stride, "depth_stride");
2836 lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");
2837 lp_build_name(depth_sample_stride, "depth_sample_stride");
2838
2839 /*
2840 * Function body
2841 */
2842
2843 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
2844 builder = gallivm->builder;
2845 assert(builder);
2846 LLVMPositionBuilderAtEnd(builder, block);
2847
2848 /*
2849 * Must not count ps invocations if there's a null shader.
2850 * (It would be ok to count with null shader if there's d/s tests,
2851 * but only if there's d/s buffers too, which is different
2852 * to implicit rasterization disable which must not depend
2853 * on the d/s buffers.)
2854 * Could use popcount on mask, but pixel accuracy is not required.
2855 * Could disable if there's no stats query, but maybe not worth it.
2856 */
2857 if (shader->info.base.num_instructions > 1) {
2858 LLVMValueRef invocs, val;
2859 invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
2860 val = LLVMBuildLoad(builder, invocs, "");
2861 val = LLVMBuildAdd(builder, val,
2862 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
2863 "invoc_count");
2864 LLVMBuildStore(builder, val, invocs);
2865 }
2866
2867 /* code generated texture sampling */
2868 sampler = lp_llvm_sampler_soa_create(key->samplers);
2869 image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key));
2870
2871 num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
2872 /* for 1d resources only run "upper half" of stamp */
2873 if (key->resource_1d)
2874 num_fs /= 2;
2875
2876 {
2877 LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
2878 LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
2879 LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);
2880 LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
2881 num_loop_samp, "mask_store");
2882
2883 LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
2884 LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, LLVMArrayType(flt_type, key->coverage_samples * 2), "");
2885 LLVMValueRef sample_pos_array;
2886
2887 if (key->multisample && key->coverage_samples == 4) {
2888 LLVMValueRef sample_pos_arr[8];
2889 for (unsigned i = 0; i < 4; i++) {
2890 sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);
2891 sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);
2892 }
2893 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);
2894 } else {
2895 LLVMValueRef sample_pos_arr[2];
2896 sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);
2897 sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);
2898 sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);
2899 }
2900 LLVMSetInitializer(glob_sample_pos, sample_pos_array);
2901
2902 LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
2903 boolean pixel_center_integer =
2904 shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
2905
2906 /*
2907 * The shader input interpolation info is not explicitely baked in the
2908 * shader key, but everything it derives from (TGSI, and flatshade) is
2909 * already included in the shader key.
2910 */
2911 lp_build_interp_soa_init(&interp,
2912 gallivm,
2913 shader->info.base.num_inputs,
2914 inputs,
2915 pixel_center_integer,
2916 key->coverage_samples, glob_sample_pos,
2917 num_loop,
2918 key->depth_clamp,
2919 builder, fs_type,
2920 a0_ptr, dadx_ptr, dady_ptr,
2921 x, y);
2922
2923 for (i = 0; i < num_fs; i++) {
2924 if (key->multisample) {
2925 LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");
2926
2927 /*
2928 * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
2929 * store to the per sample mask storage. Or all of them together to generate
2930 * the fragment shader mask. (sample shading TODO).
2931 * Take the incoming state coverage mask into account.
2932 */
2933 for (unsigned s = 0; s < key->coverage_samples; s++) {
2934 LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));
2935 LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,
2936 &sindexi, 1, "sample_mask_ptr");
2937 LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,
2938 i*fs_type.length/4, s, mask_input);
2939
2940 LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");
2941 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");
2942 smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");
2943 smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);
2944
2945 s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
2946 LLVMBuildStore(builder, s_mask, sample_mask_ptr);
2947 }
2948 } else {
2949 LLVMValueRef mask;
2950 LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
2951 LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
2952 &indexi, 1, "mask_ptr");
2953
2954 if (partial_mask) {
2955 mask = generate_quad_mask(gallivm, fs_type,
2956 i*fs_type.length/4, 0, mask_input);
2957 }
2958 else {
2959 mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
2960 }
2961 LLVMBuildStore(builder, mask, mask_ptr);
2962 }
2963 }
2964
2965 generate_fs_loop(gallivm,
2966 shader, key,
2967 builder,
2968 fs_type,
2969 context_ptr,
2970 glob_sample_pos,
2971 num_loop,
2972 &interp,
2973 sampler,
2974 image,
2975 mask_store, /* output */
2976 color_store,
2977 depth_ptr,
2978 depth_stride,
2979 depth_sample_stride,
2980 facing,
2981 thread_data_ptr);
2982
2983 for (i = 0; i < num_fs; i++) {
2984 LLVMValueRef ptr;
2985 for (unsigned s = 0; s < key->coverage_samples; s++) {
2986 int idx = (i + (s * num_fs));
2987 LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
2988 ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");
2989
2990 fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
2991 }
2992
2993 for (unsigned s = 0; s < key->min_samples; s++) {
2994 /* This is fucked up need to reorganize things */
2995 int idx = s * num_fs + i;
2996 LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
2997 for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
2998 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2999 ptr = LLVMBuildGEP(builder,
3000 color_store[cbuf * !cbuf0_write_all][chan],
3001 &sindexi, 1, "");
3002 fs_out_color[s][cbuf][chan][i] = ptr;
3003 }
3004 }
3005 if (dual_source_blend) {
3006 /* only support one dual source blend target hence always use output 1 */
3007 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3008 ptr = LLVMBuildGEP(builder,
3009 color_store[1][chan],
3010 &sindexi, 1, "");
3011 fs_out_color[s][1][chan][i] = ptr;
3012 }
3013 }
3014 }
3015 }
3016 }
3017
3018 sampler->destroy(sampler);
3019 image->destroy(image);
3020 /* Loop over color outputs / color buffers to do blending.
3021 */
3022 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
3023 if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
3024 LLVMValueRef color_ptr;
3025 LLVMValueRef stride;
3026 LLVMValueRef sample_stride = NULL;
3027 LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
3028
3029 boolean do_branch = ((key->depth.enabled
3030 || key->stencil[0].enabled
3031 || key->alpha.enabled)
3032 && !shader->info.base.uses_kill);
3033
3034 color_ptr = LLVMBuildLoad(builder,
3035 LLVMBuildGEP(builder, color_ptr_ptr,
3036 &index, 1, ""),
3037 "");
3038
3039 stride = LLVMBuildLoad(builder,
3040 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
3041 "");
3042
3043 if (key->multisample)
3044 sample_stride = LLVMBuildLoad(builder,
3045 LLVMBuildGEP(builder, color_sample_stride_ptr,
3046 &index, 1, ""), "");
3047
3048 for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {
3049 unsigned mask_idx = num_fs * (key->multisample ? s : 0);
3050 unsigned out_idx = key->min_samples == 1 ? 0 : s;
3051 LLVMValueRef out_ptr = color_ptr;;
3052
3053 if (key->multisample) {
3054 LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");
3055 out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");
3056 }
3057 out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");
3058
3059 lp_build_name(out_ptr, "color_ptr%d", cbuf);
3060
3061 generate_unswizzled_blend(gallivm, cbuf, variant,
3062 key->cbuf_format[cbuf],
3063 num_fs, fs_type, &fs_mask[mask_idx], fs_out_color[out_idx],
3064 context_ptr, out_ptr, stride,
3065 partial_mask, do_branch);
3066 }
3067 }
3068 }
3069
3070 LLVMBuildRetVoid(builder);
3071
3072 gallivm_verify_function(gallivm, function);
3073 }
3074
3075
3076 static void
3077 dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
3078 {
3079 unsigned i;
3080
3081 debug_printf("fs variant %p:\n", (void *) key);
3082
3083 if (key->flatshade) {
3084 debug_printf("flatshade = 1\n");
3085 }
3086 if (key->multisample) {
3087 debug_printf("multisample = 1\n");
3088 debug_printf("coverage samples = %d\n", key->coverage_samples);
3089 debug_printf("min samples = %d\n", key->min_samples);
3090 }
3091 for (i = 0; i < key->nr_cbufs; ++i) {
3092 debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
3093 debug_printf("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);
3094 }
3095 if (key->depth.enabled || key->stencil[0].enabled) {
3096 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
3097 debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);
3098 }
3099 if (key->depth.enabled) {
3100 debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
3101 debug_printf("depth.writemask = %u\n", key->depth.writemask);
3102 }
3103
3104 for (i = 0; i < 2; ++i) {
3105 if (key->stencil[i].enabled) {
3106 debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
3107 debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
3108 debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
3109 debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
3110 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
3111 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
3112 }
3113 }
3114
3115 if (key->alpha.enabled) {
3116 debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
3117 }
3118
3119 if (key->occlusion_count) {
3120 debug_printf("occlusion_count = 1\n");
3121 }
3122
3123 if (key->blend.logicop_enable) {
3124 debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
3125 }
3126 else if (key->blend.rt[0].blend_enable) {
3127 debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));
3128 debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
3129 debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
3130 debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));
3131 debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
3132 debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
3133 }
3134 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
3135 if (key->blend.alpha_to_coverage) {
3136 debug_printf("blend.alpha_to_coverage is enabled\n");
3137 }
3138 for (i = 0; i < key->nr_samplers; ++i) {
3139 const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
3140 debug_printf("sampler[%u] = \n", i);
3141 debug_printf(" .wrap = %s %s %s\n",
3142 util_str_tex_wrap(sampler->wrap_s, TRUE),
3143 util_str_tex_wrap(sampler->wrap_t, TRUE),
3144 util_str_tex_wrap(sampler->wrap_r, TRUE));
3145 debug_printf(" .min_img_filter = %s\n",
3146 util_str_tex_filter(sampler->min_img_filter, TRUE));
3147 debug_printf(" .min_mip_filter = %s\n",
3148 util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
3149 debug_printf(" .mag_img_filter = %s\n",
3150 util_str_tex_filter(sampler->mag_img_filter, TRUE));
3151 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
3152 debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
3153 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
3154 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
3155 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
3156 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
3157 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
3158 }
3159 for (i = 0; i < key->nr_sampler_views; ++i) {
3160 const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
3161 debug_printf("texture[%u] = \n", i);
3162 debug_printf(" .format = %s\n",
3163 util_format_name(texture->format));
3164 debug_printf(" .target = %s\n",
3165 util_str_tex_target(texture->target, TRUE));
3166 debug_printf(" .level_zero_only = %u\n",
3167 texture->level_zero_only);
3168 debug_printf(" .pot = %u %u %u\n",
3169 texture->pot_width,
3170 texture->pot_height,
3171 texture->pot_depth);
3172 }
3173 struct lp_image_static_state *images = lp_fs_variant_key_images(key);
3174 for (i = 0; i < key->nr_images; ++i) {
3175 const struct lp_static_texture_state *image = &images[i].image_state;
3176 debug_printf("image[%u] = \n", i);
3177 debug_printf(" .format = %s\n",
3178 util_format_name(image->format));
3179 debug_printf(" .target = %s\n",
3180 util_str_tex_target(image->target, TRUE));
3181 debug_printf(" .level_zero_only = %u\n",
3182 image->level_zero_only);
3183 debug_printf(" .pot = %u %u %u\n",
3184 image->pot_width,
3185 image->pot_height,
3186 image->pot_depth);
3187 }
3188 }
3189
3190
3191 void
3192 lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
3193 {
3194 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3195 variant->shader->no, variant->no);
3196 if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
3197 tgsi_dump(variant->shader->base.tokens, 0);
3198 else
3199 nir_print_shader(variant->shader->base.ir.nir, stderr);
3200 dump_fs_variant_key(&variant->key);
3201 debug_printf("variant->opaque = %u\n", variant->opaque);
3202 debug_printf("\n");
3203 }
3204
3205
3206 /**
3207 * Generate a new fragment shader variant from the shader code and
3208 * other state indicated by the key.
3209 */
3210 static struct lp_fragment_shader_variant *
3211 generate_variant(struct llvmpipe_context *lp,
3212 struct lp_fragment_shader *shader,
3213 const struct lp_fragment_shader_variant_key *key)
3214 {
3215 struct lp_fragment_shader_variant *variant;
3216 const struct util_format_description *cbuf0_format_desc = NULL;
3217 boolean fullcolormask;
3218 char module_name[64];
3219
3220 variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
3221 if (!variant)
3222 return NULL;
3223
3224 memset(variant, 0, sizeof(*variant));
3225 snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
3226 shader->no, shader->variants_created);
3227
3228 variant->gallivm = gallivm_create(module_name, lp->context);
3229 if (!variant->gallivm) {
3230 FREE(variant);
3231 return NULL;
3232 }
3233
3234 variant->shader = shader;
3235 variant->list_item_global.base = variant;
3236 variant->list_item_local.base = variant;
3237 variant->no = shader->variants_created++;
3238
3239 memcpy(&variant->key, key, shader->variant_key_size);
3240
3241 /*
3242 * Determine whether we are touching all channels in the color buffer.
3243 */
3244 fullcolormask = FALSE;
3245 if (key->nr_cbufs == 1) {
3246 cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
3247 fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
3248 }
3249
3250 variant->opaque =
3251 !key->blend.logicop_enable &&
3252 !key->blend.rt[0].blend_enable &&
3253 fullcolormask &&
3254 !key->stencil[0].enabled &&
3255 !key->alpha.enabled &&
3256 !key->multisample &&
3257 !key->blend.alpha_to_coverage &&
3258 !key->depth.enabled &&
3259 !shader->info.base.uses_kill &&
3260 !shader->info.base.writes_samplemask
3261 ? TRUE : FALSE;
3262
3263 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3264 lp_debug_fs_variant(variant);
3265 }
3266
3267 lp_jit_init_types(variant);
3268
3269 if (variant->jit_function[RAST_EDGE_TEST] == NULL)
3270 generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
3271
3272 if (variant->jit_function[RAST_WHOLE] == NULL) {
3273 if (variant->opaque) {
3274 /* Specialized shader, which doesn't need to read the color buffer. */
3275 generate_fragment(lp, shader, variant, RAST_WHOLE);
3276 }
3277 }
3278
3279 /*
3280 * Compile everything
3281 */
3282
3283 gallivm_compile_module(variant->gallivm);
3284
3285 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
3286
3287 if (variant->function[RAST_EDGE_TEST]) {
3288 variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
3289 gallivm_jit_function(variant->gallivm,
3290 variant->function[RAST_EDGE_TEST]);
3291 }
3292
3293 if (variant->function[RAST_WHOLE]) {
3294 variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
3295 gallivm_jit_function(variant->gallivm,
3296 variant->function[RAST_WHOLE]);
3297 } else if (!variant->jit_function[RAST_WHOLE]) {
3298 variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
3299 }
3300
3301 gallivm_free_ir(variant->gallivm);
3302
3303 return variant;
3304 }
3305
3306
3307 static void *
3308 llvmpipe_create_fs_state(struct pipe_context *pipe,
3309 const struct pipe_shader_state *templ)
3310 {
3311 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3312 struct lp_fragment_shader *shader;
3313 int nr_samplers;
3314 int nr_sampler_views;
3315 int nr_images;
3316 int i;
3317
3318 shader = CALLOC_STRUCT(lp_fragment_shader);
3319 if (!shader)
3320 return NULL;
3321
3322 shader->no = fs_no++;
3323 make_empty_list(&shader->variants);
3324
3325 shader->base.type = templ->type;
3326 if (templ->type == PIPE_SHADER_IR_TGSI) {
3327 /* get/save the summary info for this shader */
3328 lp_build_tgsi_info(templ->tokens, &shader->info);
3329
3330 /* we need to keep a local copy of the tokens */
3331 shader->base.tokens = tgsi_dup_tokens(templ->tokens);
3332 } else {
3333 shader->base.ir.nir = templ->ir.nir;
3334 nir_tgsi_scan_shader(templ->ir.nir, &shader->info.base, true);
3335 }
3336
3337 shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
3338 if (shader->draw_data == NULL) {
3339 FREE((void *) shader->base.tokens);
3340 FREE(shader);
3341 return NULL;
3342 }
3343
3344 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3345 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3346 nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3347 shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
3348
3349 for (i = 0; i < shader->info.base.num_inputs; i++) {
3350 shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
3351 shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
3352 shader->inputs[i].location = shader->info.base.input_interpolate_loc[i];
3353
3354 switch (shader->info.base.input_interpolate[i]) {
3355 case TGSI_INTERPOLATE_CONSTANT:
3356 shader->inputs[i].interp = LP_INTERP_CONSTANT;
3357 break;
3358 case TGSI_INTERPOLATE_LINEAR:
3359 shader->inputs[i].interp = LP_INTERP_LINEAR;
3360 break;
3361 case TGSI_INTERPOLATE_PERSPECTIVE:
3362 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
3363 break;
3364 case TGSI_INTERPOLATE_COLOR:
3365 shader->inputs[i].interp = LP_INTERP_COLOR;
3366 break;
3367 default:
3368 assert(0);
3369 break;
3370 }
3371
3372 switch (shader->info.base.input_semantic_name[i]) {
3373 case TGSI_SEMANTIC_FACE:
3374 shader->inputs[i].interp = LP_INTERP_FACING;
3375 break;
3376 case TGSI_SEMANTIC_POSITION:
3377 /* Position was already emitted above
3378 */
3379 shader->inputs[i].interp = LP_INTERP_POSITION;
3380 shader->inputs[i].src_index = 0;
3381 continue;
3382 }
3383
3384 /* XXX this is a completely pointless index map... */
3385 shader->inputs[i].src_index = i+1;
3386 }
3387
3388 if (LP_DEBUG & DEBUG_TGSI) {
3389 unsigned attrib;
3390 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3391 shader->no, (void *) shader);
3392 tgsi_dump(templ->tokens, 0);
3393 debug_printf("usage masks:\n");
3394 for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
3395 unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
3396 debug_printf(" IN[%u].%s%s%s%s\n",
3397 attrib,
3398 usage_mask & TGSI_WRITEMASK_X ? "x" : "",
3399 usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
3400 usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
3401 usage_mask & TGSI_WRITEMASK_W ? "w" : "");
3402 }
3403 debug_printf("\n");
3404 }
3405
3406 return shader;
3407 }
3408
3409
3410 static void
3411 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
3412 {
3413 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3414 struct lp_fragment_shader *lp_fs = (struct lp_fragment_shader *)fs;
3415 if (llvmpipe->fs == lp_fs)
3416 return;
3417
3418 draw_bind_fragment_shader(llvmpipe->draw,
3419 (lp_fs ? lp_fs->draw_data : NULL));
3420
3421 llvmpipe->fs = lp_fs;
3422
3423 llvmpipe->dirty |= LP_NEW_FS;
3424 }
3425
3426
3427 /**
3428 * Remove shader variant from two lists: the shader's variant list
3429 * and the context's variant list.
3430 */
3431 static void
3432 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
3433 struct lp_fragment_shader_variant *variant)
3434 {
3435 if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3436 debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3437 "v total cached %u inst %u total inst %u\n",
3438 variant->shader->no, variant->no,
3439 variant->shader->variants_created,
3440 variant->shader->variants_cached,
3441 lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);
3442 }
3443
3444 gallivm_destroy(variant->gallivm);
3445
3446 /* remove from shader's list */
3447 remove_from_list(&variant->list_item_local);
3448 variant->shader->variants_cached--;
3449
3450 /* remove from context's list */
3451 remove_from_list(&variant->list_item_global);
3452 lp->nr_fs_variants--;
3453 lp->nr_fs_instrs -= variant->nr_instrs;
3454
3455 FREE(variant);
3456 }
3457
3458
3459 static void
3460 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
3461 {
3462 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3463 struct lp_fragment_shader *shader = fs;
3464 struct lp_fs_variant_list_item *li;
3465
3466 assert(fs != llvmpipe->fs);
3467
3468 /*
3469 * XXX: we need to flush the context until we have some sort of reference
3470 * counting in fragment shaders as they may still be binned
3471 * Flushing alone might not sufficient we need to wait on it too.
3472 */
3473 llvmpipe_finish(pipe, __FUNCTION__);
3474
3475 /* Delete all the variants */
3476 li = first_elem(&shader->variants);
3477 while(!at_end(&shader->variants, li)) {
3478 struct lp_fs_variant_list_item *next = next_elem(li);
3479 llvmpipe_remove_shader_variant(llvmpipe, li->base);
3480 li = next;
3481 }
3482
3483 /* Delete draw module's data */
3484 draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
3485
3486 if (shader->base.ir.nir)
3487 ralloc_free(shader->base.ir.nir);
3488 assert(shader->variants_cached == 0);
3489 FREE((void *) shader->base.tokens);
3490 FREE(shader);
3491 }
3492
3493
3494
3495 static void
3496 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
3497 enum pipe_shader_type shader, uint index,
3498 const struct pipe_constant_buffer *cb)
3499 {
3500 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3501 struct pipe_resource *constants = cb ? cb->buffer : NULL;
3502
3503 assert(shader < PIPE_SHADER_TYPES);
3504 assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
3505
3506 /* note: reference counting */
3507 util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
3508
3509 if (constants) {
3510 if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
3511 debug_printf("Illegal set constant without bind flag\n");
3512 constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
3513 }
3514 }
3515
3516 if (shader == PIPE_SHADER_VERTEX ||
3517 shader == PIPE_SHADER_GEOMETRY ||
3518 shader == PIPE_SHADER_TESS_CTRL ||
3519 shader == PIPE_SHADER_TESS_EVAL) {
3520 /* Pass the constants to the 'draw' module */
3521 const unsigned size = cb ? cb->buffer_size : 0;
3522 const ubyte *data;
3523
3524 if (constants) {
3525 data = (ubyte *) llvmpipe_resource_data(constants);
3526 }
3527 else if (cb && cb->user_buffer) {
3528 data = (ubyte *) cb->user_buffer;
3529 }
3530 else {
3531 data = NULL;
3532 }
3533
3534 if (data)
3535 data += cb->buffer_offset;
3536
3537 draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
3538 index, data, size);
3539 }
3540 else if (shader == PIPE_SHADER_COMPUTE)
3541 llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;
3542 else
3543 llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
3544
3545 if (cb && cb->user_buffer) {
3546 pipe_resource_reference(&constants, NULL);
3547 }
3548 }
3549
3550 static void
3551 llvmpipe_set_shader_buffers(struct pipe_context *pipe,
3552 enum pipe_shader_type shader, unsigned start_slot,
3553 unsigned count, const struct pipe_shader_buffer *buffers,
3554 unsigned writable_bitmask)
3555 {
3556 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3557 unsigned i, idx;
3558 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3559 const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;
3560
3561 util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);
3562
3563 if (shader == PIPE_SHADER_VERTEX ||
3564 shader == PIPE_SHADER_GEOMETRY ||
3565 shader == PIPE_SHADER_TESS_CTRL ||
3566 shader == PIPE_SHADER_TESS_EVAL) {
3567 const unsigned size = buffer ? buffer->buffer_size : 0;
3568 const ubyte *data = NULL;
3569 if (buffer && buffer->buffer)
3570 data = (ubyte *) llvmpipe_resource_data(buffer->buffer);
3571 if (data)
3572 data += buffer->buffer_offset;
3573 draw_set_mapped_shader_buffer(llvmpipe->draw, shader,
3574 i, data, size);
3575 } else if (shader == PIPE_SHADER_COMPUTE) {
3576 llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;
3577 } else if (shader == PIPE_SHADER_FRAGMENT) {
3578 llvmpipe->dirty |= LP_NEW_FS_SSBOS;
3579 }
3580 }
3581 }
3582
3583 static void
3584 llvmpipe_set_shader_images(struct pipe_context *pipe,
3585 enum pipe_shader_type shader, unsigned start_slot,
3586 unsigned count, const struct pipe_image_view *images)
3587 {
3588 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3589 unsigned i, idx;
3590
3591 draw_flush(llvmpipe->draw);
3592 for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
3593 const struct pipe_image_view *image = images ? &images[idx] : NULL;
3594
3595 util_copy_image_view(&llvmpipe->images[shader][i], image);
3596 }
3597
3598 llvmpipe->num_images[shader] = start_slot + count;
3599 if (shader == PIPE_SHADER_VERTEX ||
3600 shader == PIPE_SHADER_GEOMETRY ||
3601 shader == PIPE_SHADER_TESS_CTRL ||
3602 shader == PIPE_SHADER_TESS_EVAL) {
3603 draw_set_images(llvmpipe->draw,
3604 shader,
3605 llvmpipe->images[shader],
3606 start_slot + count);
3607 } else if (shader == PIPE_SHADER_COMPUTE)
3608 llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;
3609 else
3610 llvmpipe->dirty |= LP_NEW_FS_IMAGES;
3611 }
3612
3613 /**
3614 * Return the blend factor equivalent to a destination alpha of one.
3615 */
3616 static inline unsigned
3617 force_dst_alpha_one(unsigned factor, boolean clamped_zero)
3618 {
3619 switch(factor) {
3620 case PIPE_BLENDFACTOR_DST_ALPHA:
3621 return PIPE_BLENDFACTOR_ONE;
3622 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
3623 return PIPE_BLENDFACTOR_ZERO;
3624 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
3625 if (clamped_zero)
3626 return PIPE_BLENDFACTOR_ZERO;
3627 else
3628 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
3629 }
3630
3631 return factor;
3632 }
3633
3634
3635 /**
3636 * We need to generate several variants of the fragment pipeline to match
3637 * all the combinations of the contributing state atoms.
3638 *
3639 * TODO: there is actually no reason to tie this to context state -- the
3640 * generated code could be cached globally in the screen.
3641 */
3642 static struct lp_fragment_shader_variant_key *
3643 make_variant_key(struct llvmpipe_context *lp,
3644 struct lp_fragment_shader *shader,
3645 char *store)
3646 {
3647 unsigned i;
3648 struct lp_fragment_shader_variant_key *key;
3649
3650 key = (struct lp_fragment_shader_variant_key *)store;
3651
3652 memset(key, 0, offsetof(struct lp_fragment_shader_variant_key, samplers[1]));
3653
3654 if (lp->framebuffer.zsbuf) {
3655 enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
3656 const struct util_format_description *zsbuf_desc =
3657 util_format_description(zsbuf_format);
3658
3659 if (lp->depth_stencil->depth.enabled &&
3660 util_format_has_depth(zsbuf_desc)) {
3661 key->zsbuf_format = zsbuf_format;
3662 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
3663 }
3664 if (lp->depth_stencil->stencil[0].enabled &&
3665 util_format_has_stencil(zsbuf_desc)) {
3666 key->zsbuf_format = zsbuf_format;
3667 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
3668 }
3669 if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
3670 key->resource_1d = TRUE;
3671 }
3672 key->zsbuf_nr_samples = util_res_sample_count(lp->framebuffer.zsbuf->texture);
3673 }
3674
3675 /*
3676 * Propagate the depth clamp setting from the rasterizer state.
3677 * depth_clip == 0 implies depth clamping is enabled.
3678 *
3679 * When clip_halfz is enabled, then always clamp the depth values.
3680 *
3681 * XXX: This is incorrect for GL, but correct for d3d10 (depth
3682 * clamp is always active in d3d10, regardless if depth clip is
3683 * enabled or not).
3684 * (GL has an always-on [0,1] clamp on fs depth output instead
3685 * to ensure the depth values stay in range. Doesn't look like
3686 * we do that, though...)
3687 */
3688 if (lp->rasterizer->clip_halfz) {
3689 key->depth_clamp = 1;
3690 } else {
3691 key->depth_clamp = (lp->rasterizer->depth_clip_near == 0) ? 1 : 0;
3692 }
3693
3694 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
3695 if (!lp->framebuffer.nr_cbufs ||
3696 !lp->framebuffer.cbufs[0] ||
3697 !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
3698 key->alpha.enabled = lp->depth_stencil->alpha.enabled;
3699 }
3700 if(key->alpha.enabled)
3701 key->alpha.func = lp->depth_stencil->alpha.func;
3702 /* alpha.ref_value is passed in jit_context */
3703
3704 key->flatshade = lp->rasterizer->flatshade;
3705 key->multisample = lp->rasterizer->multisample;
3706 if (lp->active_occlusion_queries && !lp->queries_disabled) {
3707 key->occlusion_count = TRUE;
3708 }
3709
3710 if (lp->framebuffer.nr_cbufs) {
3711 memcpy(&key->blend, lp->blend, sizeof key->blend);
3712 }
3713
3714 key->coverage_samples = 1;
3715 key->min_samples = 1;
3716 if (key->multisample) {
3717 key->coverage_samples = util_framebuffer_get_num_samples(&lp->framebuffer);
3718 key->min_samples = lp->min_samples == 1 ? 1 : key->coverage_samples;
3719 }
3720 key->nr_cbufs = lp->framebuffer.nr_cbufs;
3721
3722 if (!key->blend.independent_blend_enable) {
3723 /* we always need independent blend otherwise the fixups below won't work */
3724 for (i = 1; i < key->nr_cbufs; i++) {
3725 memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
3726 }
3727 key->blend.independent_blend_enable = 1;
3728 }
3729
3730 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
3731 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
3732
3733 if (lp->framebuffer.cbufs[i]) {
3734 enum pipe_format format = lp->framebuffer.cbufs[i]->format;
3735 const struct util_format_description *format_desc;
3736
3737 key->cbuf_format[i] = format;
3738 key->cbuf_nr_samples[i] = util_res_sample_count(lp->framebuffer.cbufs[i]->texture);
3739
3740 /*
3741 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
3742 * mixing of 2d textures with height 1 and 1d textures, so make sure
3743 * we pick 1d if any cbuf or zsbuf is 1d.
3744 */
3745 if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
3746 key->resource_1d = TRUE;
3747 }
3748
3749 format_desc = util_format_description(format);
3750 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
3751 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
3752
3753 /*
3754 * Mask out color channels not present in the color buffer.
3755 */
3756 blend_rt->colormask &= util_format_colormask(format_desc);
3757
3758 /*
3759 * Disable blend for integer formats.
3760 */
3761 if (util_format_is_pure_integer(format)) {
3762 blend_rt->blend_enable = 0;
3763 }
3764
3765 /*
3766 * Our swizzled render tiles always have an alpha channel, but the
3767 * linear render target format often does not, so force here the dst
3768 * alpha to be one.
3769 *
3770 * This is not a mere optimization. Wrong results will be produced if
3771 * the dst alpha is used, the dst format does not have alpha, and the
3772 * previous rendering was not flushed from the swizzled to linear
3773 * buffer. For example, NonPowTwo DCT.
3774 *
3775 * TODO: This should be generalized to all channels for better
3776 * performance, but only alpha causes correctness issues.
3777 *
3778 * Also, force rgb/alpha func/factors match, to make AoS blending
3779 * easier.
3780 */
3781 if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
3782 format_desc->swizzle[3] == format_desc->swizzle[0]) {
3783 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
3784 boolean clamped_zero = !util_format_is_float(format) &&
3785 !util_format_is_snorm(format);
3786 blend_rt->rgb_src_factor =
3787 force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
3788 blend_rt->rgb_dst_factor =
3789 force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
3790 blend_rt->alpha_func = blend_rt->rgb_func;
3791 blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
3792 blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
3793 }
3794 }
3795 else {
3796 /* no color buffer for this fragment output */
3797 key->cbuf_format[i] = PIPE_FORMAT_NONE;
3798 key->cbuf_nr_samples[i] = 0;
3799 blend_rt->colormask = 0x0;
3800 blend_rt->blend_enable = 0;
3801 }
3802 }
3803
3804 /* This value will be the same for all the variants of a given shader:
3805 */
3806 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3807
3808 struct lp_sampler_static_state *fs_sampler;
3809
3810 fs_sampler = key->samplers;
3811
3812 memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);
3813
3814 for(i = 0; i < key->nr_samplers; ++i) {
3815 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3816 lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,
3817 lp->samplers[PIPE_SHADER_FRAGMENT][i]);
3818 }
3819 }
3820
3821 /*
3822 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
3823 * are dx10-style? Can't really have mixed opcodes, at least not
3824 * if we want to skip the holes here (without rescanning tgsi).
3825 */
3826 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3827 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3828 for(i = 0; i < key->nr_sampler_views; ++i) {
3829 /*
3830 * Note sview may exceed what's representable by file_mask.
3831 * This will still work, the only downside is that not actually
3832 * used views may be included in the shader key.
3833 */
3834 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
3835 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3836 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3837 }
3838 }
3839 }
3840 else {
3841 key->nr_sampler_views = key->nr_samplers;
3842 for(i = 0; i < key->nr_sampler_views; ++i) {
3843 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
3844 lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
3845 lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
3846 }
3847 }
3848 }
3849
3850 struct lp_image_static_state *lp_image;
3851 lp_image = lp_fs_variant_key_images(key);
3852 key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3853 for (i = 0; i < key->nr_images; ++i) {
3854 if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
3855 lp_sampler_static_texture_state_image(&lp_image[i].image_state,
3856 &lp->images[PIPE_SHADER_FRAGMENT][i]);
3857 }
3858 }
3859 return key;
3860 }
3861
3862
3863
3864 /**
3865 * Update fragment shader state. This is called just prior to drawing
3866 * something when some fragment-related state has changed.
3867 */
3868 void
3869 llvmpipe_update_fs(struct llvmpipe_context *lp)
3870 {
3871 struct lp_fragment_shader *shader = lp->fs;
3872 struct lp_fragment_shader_variant_key *key;
3873 struct lp_fragment_shader_variant *variant = NULL;
3874 struct lp_fs_variant_list_item *li;
3875 char store[LP_FS_MAX_VARIANT_KEY_SIZE];
3876
3877 key = make_variant_key(lp, shader, store);
3878
3879 /* Search the variants for one which matches the key */
3880 li = first_elem(&shader->variants);
3881 while(!at_end(&shader->variants, li)) {
3882 if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
3883 variant = li->base;
3884 break;
3885 }
3886 li = next_elem(li);
3887 }
3888
3889 if (variant) {
3890 /* Move this variant to the head of the list to implement LRU
3891 * deletion of shader's when we have too many.
3892 */
3893 move_to_head(&lp->fs_variants_list, &variant->list_item_global);
3894 }
3895 else {
3896 /* variant not found, create it now */
3897 int64_t t0, t1, dt;
3898 unsigned i;
3899 unsigned variants_to_cull;
3900
3901 if (LP_DEBUG & DEBUG_FS) {
3902 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3903 lp->nr_fs_variants,
3904 lp->nr_fs_instrs,
3905 lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
3906 }
3907
3908 /* First, check if we've exceeded the max number of shader variants.
3909 * If so, free 6.25% of them (the least recently used ones).
3910 */
3911 variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
3912
3913 if (variants_to_cull ||
3914 lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
3915 struct pipe_context *pipe = &lp->pipe;
3916
3917 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
3918 debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
3919 "\t%u instrs,\t%u instrs/variant\n",
3920 shader->variants_cached,
3921 lp->nr_fs_variants, lp->nr_fs_instrs,
3922 lp->nr_fs_instrs / lp->nr_fs_variants);
3923 }
3924
3925 /*
3926 * XXX: we need to flush the context until we have some sort of
3927 * reference counting in fragment shaders as they may still be binned
3928 * Flushing alone might not be sufficient we need to wait on it too.
3929 */
3930 llvmpipe_finish(pipe, __FUNCTION__);
3931
3932 /*
3933 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3934 * number of shader variants (potentially all of them) could be
3935 * pending for destruction on flush.
3936 */
3937
3938 for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
3939 struct lp_fs_variant_list_item *item;
3940 if (is_empty_list(&lp->fs_variants_list)) {
3941 break;
3942 }
3943 item = last_elem(&lp->fs_variants_list);
3944 assert(item);
3945 assert(item->base);
3946 llvmpipe_remove_shader_variant(lp, item->base);
3947 }
3948 }
3949
3950 /*
3951 * Generate the new variant.
3952 */
3953 t0 = os_time_get();
3954 variant = generate_variant(lp, shader, key);
3955 t1 = os_time_get();
3956 dt = t1 - t0;
3957 LP_COUNT_ADD(llvm_compile_time, dt);
3958 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
3959
3960 /* Put the new variant into the list */
3961 if (variant) {
3962 insert_at_head(&shader->variants, &variant->list_item_local);
3963 insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
3964 lp->nr_fs_variants++;
3965 lp->nr_fs_instrs += variant->nr_instrs;
3966 shader->variants_cached++;
3967 }
3968 }
3969
3970 /* Bind this variant */
3971 lp_setup_set_fs_variant(lp->setup, variant);
3972 }
3973
3974
3975
3976
3977
3978 void
3979 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
3980 {
3981 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
3982 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
3983 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
3984
3985 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
3986
3987 llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;
3988 llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;
3989 }
3990
3991