projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe: add cache for compressed textures
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "draw_llvm.h"
29
30 #include "draw_context.h"
31 #include "draw_vs.h"
32 #include "draw_gs.h"
33
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_logic.h"
37 #include "gallivm/lp_bld_const.h"
38 #include "gallivm/lp_bld_swizzle.h"
39 #include "gallivm/lp_bld_struct.h"
40 #include "gallivm/lp_bld_type.h"
41 #include "gallivm/lp_bld_flow.h"
42 #include "gallivm/lp_bld_debug.h"
43 #include "gallivm/lp_bld_tgsi.h"
44 #include "gallivm/lp_bld_printf.h"
45 #include "gallivm/lp_bld_intr.h"
46 #include "gallivm/lp_bld_init.h"
47 #include "gallivm/lp_bld_type.h"
48 #include "gallivm/lp_bld_pack.h"
49 #include "gallivm/lp_bld_format.h"
50
51 #include "tgsi/tgsi_exec.h"
52 #include "tgsi/tgsi_dump.h"
53
54 #include "util/u_math.h"
55 #include "util/u_pointer.h"
56 #include "util/u_string.h"
57 #include "util/simple_list.h"
58
59
60 #define DEBUG_STORE 0
61
62
63 static void
64 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
65 boolean elts);
66
67
68 struct draw_gs_llvm_iface {
69 struct lp_build_tgsi_gs_iface base;
70
71 struct draw_gs_llvm_variant *variant;
72 LLVMValueRef input;
73 };
74
75 static inline const struct draw_gs_llvm_iface *
76 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
77 {
78 return (const struct draw_gs_llvm_iface *)iface;
79 }
80
81 /**
82 * Create LLVM type for draw_vertex_buffer.
83 */
84 static LLVMTypeRef
85 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
86 const char *struct_name)
87 {
88 LLVMTargetDataRef target = gallivm->target;
89 LLVMTypeRef dvbuffer_type;
90 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
91 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
92
93 elem_types[DRAW_JIT_DVBUFFER_MAP] =
94 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
95 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
96
97 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
98 Elements(elem_types), 0);
99
100 (void) target; /* silence unused var warning for non-debug build */
101 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
102 target, dvbuffer_type,
103 DRAW_JIT_DVBUFFER_MAP);
104 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
105 target, dvbuffer_type,
106 DRAW_JIT_DVBUFFER_SIZE);
107
108 return dvbuffer_type;
109 }
110
111 /**
112 * Create LLVM type for struct draw_jit_texture
113 */
114 static LLVMTypeRef
115 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
116 {
117 LLVMTargetDataRef target = gallivm->target;
118 LLVMTypeRef texture_type;
119 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
120 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
121
122 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
123 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
124 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
125 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
126 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
127 elem_types[DRAW_JIT_TEXTURE_BASE] =
128 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
129 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
130 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
131 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
132 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
133
134 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
135 Elements(elem_types), 0);
136
137 (void) target; /* silence unused var warning for non-debug build */
138 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
139 target, texture_type,
140 DRAW_JIT_TEXTURE_WIDTH);
141 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
142 target, texture_type,
143 DRAW_JIT_TEXTURE_HEIGHT);
144 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
145 target, texture_type,
146 DRAW_JIT_TEXTURE_DEPTH);
147 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
148 target, texture_type,
149 DRAW_JIT_TEXTURE_FIRST_LEVEL);
150 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
151 target, texture_type,
152 DRAW_JIT_TEXTURE_LAST_LEVEL);
153 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
154 target, texture_type,
155 DRAW_JIT_TEXTURE_BASE);
156 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
157 target, texture_type,
158 DRAW_JIT_TEXTURE_ROW_STRIDE);
159 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
160 target, texture_type,
161 DRAW_JIT_TEXTURE_IMG_STRIDE);
162 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
163 target, texture_type,
164 DRAW_JIT_TEXTURE_MIP_OFFSETS);
165
166 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
167
168 return texture_type;
169 }
170
171
172 /**
173 * Create LLVM type for struct draw_jit_sampler
174 */
175 static LLVMTypeRef
176 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
177 {
178 LLVMTargetDataRef target = gallivm->target;
179 LLVMTypeRef sampler_type;
180 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
181
182 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
183 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
184 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
185 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
186 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
187
188 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
189 Elements(elem_types), 0);
190
191 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
192 target, sampler_type,
193 DRAW_JIT_SAMPLER_MIN_LOD);
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
195 target, sampler_type,
196 DRAW_JIT_SAMPLER_MAX_LOD);
197 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
198 target, sampler_type,
199 DRAW_JIT_SAMPLER_LOD_BIAS);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
201 target, sampler_type,
202 DRAW_JIT_SAMPLER_BORDER_COLOR);
203
204 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
205
206 return sampler_type;
207 }
208
209
210 /**
211 * Create LLVM type for struct draw_jit_context
212 */
213 static LLVMTypeRef
214 create_jit_context_type(struct gallivm_state *gallivm,
215 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
216 const char *struct_name)
217 {
218 LLVMTargetDataRef target = gallivm->target;
219 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
220 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
221 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
222 LLVMTypeRef context_type;
223
224 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
225 LP_MAX_TGSI_CONST_BUFFERS);
226 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
227 LP_MAX_TGSI_CONST_BUFFERS);
228 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
229 DRAW_TOTAL_CLIP_PLANES), 0);
230 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
231 elem_types[4] = LLVMArrayType(texture_type,
232 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
233 elem_types[5] = LLVMArrayType(sampler_type,
234 PIPE_MAX_SAMPLERS); /* samplers */
235 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
236 Elements(elem_types), 0);
237 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
238 target, context_type, DRAW_JIT_CTX_CONSTANTS);
239 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
240 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
241 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
242 target, context_type, DRAW_JIT_CTX_PLANES);
243 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
244 target, context_type, DRAW_JIT_CTX_VIEWPORT);
245 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
246 target, context_type,
247 DRAW_JIT_CTX_TEXTURES);
248 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
249 target, context_type,
250 DRAW_JIT_CTX_SAMPLERS);
251 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
252 target, context_type);
253
254 return context_type;
255 }
256
257
258 /**
259 * Create LLVM type for struct draw_gs_jit_context
260 */
261 static LLVMTypeRef
262 create_gs_jit_context_type(struct gallivm_state *gallivm,
263 unsigned vector_length,
264 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
265 const char *struct_name)
266 {
267 LLVMTargetDataRef target = gallivm->target;
268 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
269 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
270 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
271 LLVMTypeRef context_type;
272
273 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
274 LP_MAX_TGSI_CONST_BUFFERS);
275 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
276 LP_MAX_TGSI_CONST_BUFFERS);
277 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
278 DRAW_TOTAL_CLIP_PLANES), 0);
279 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
280
281 elem_types[4] = LLVMArrayType(texture_type,
282 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
283 elem_types[5] = LLVMArrayType(sampler_type,
284 PIPE_MAX_SAMPLERS); /* samplers */
285
286 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
287 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
288 vector_length), 0);
289 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
290 vector_length), 0);
291
292 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
293 Elements(elem_types), 0);
294
295 (void) target; /* silence unused var warning for non-debug build */
296 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
297 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
298 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
299 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
300 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
301 target, context_type, DRAW_GS_JIT_CTX_PLANES);
302 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
303 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
304 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
305 target, context_type,
306 DRAW_GS_JIT_CTX_TEXTURES);
307 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
308 target, context_type,
309 DRAW_GS_JIT_CTX_SAMPLERS);
310 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
311 target, context_type,
312 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
313 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
314 target, context_type,
315 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
316 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
317 target, context_type,
318 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
319 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
320 target, context_type);
321
322 return context_type;
323 }
324
325
326 static LLVMTypeRef
327 create_gs_jit_input_type(struct gallivm_state *gallivm)
328 {
329 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
330 LLVMTypeRef input_array;
331
332 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
333 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
334 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
335 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
336
337 return input_array;
338 }
339
340 /**
341 * Create LLVM type for struct pipe_vertex_buffer
342 */
343 static LLVMTypeRef
344 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
345 const char *struct_name)
346 {
347 LLVMTargetDataRef target = gallivm->target;
348 LLVMTypeRef elem_types[4];
349 LLVMTypeRef vb_type;
350
351 elem_types[0] =
352 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
353 elem_types[2] =
354 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
355
356 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
357 Elements(elem_types), 0);
358
359 (void) target; /* silence unused var warning for non-debug build */
360 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
361 target, vb_type, 0);
362 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
363 target, vb_type, 1);
364
365 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
366
367 return vb_type;
368 }
369
370
371 /**
372 * Create LLVM type for struct vertex_header;
373 */
374 static LLVMTypeRef
375 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
376 {
377 LLVMTargetDataRef target = gallivm->target;
378 LLVMTypeRef elem_types[4];
379 LLVMTypeRef vertex_header;
380 char struct_name[24];
381
382 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
383
384 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
385 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
386 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
387 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
388
389 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
390 Elements(elem_types), 0);
391
392 /* these are bit-fields and we can't take address of them
393 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
394 target, vertex_header,
395 DRAW_JIT_VERTEX_CLIPMASK);
396 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
397 target, vertex_header,
398 DRAW_JIT_VERTEX_EDGEFLAG);
399 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
400 target, vertex_header,
401 DRAW_JIT_VERTEX_PAD);
402 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
403 target, vertex_header,
404 DRAW_JIT_VERTEX_VERTEX_ID);
405 */
406 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
407 target, vertex_header,
408 DRAW_JIT_VERTEX_CLIP);
409 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
410 target, vertex_header,
411 DRAW_JIT_VERTEX_PRE_CLIP_POS);
412 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
413 target, vertex_header,
414 DRAW_JIT_VERTEX_DATA);
415
416 assert(LLVMABISizeOfType(target, vertex_header) ==
417 offsetof(struct vertex_header, data[data_elems]));
418
419 return vertex_header;
420 }
421
422
423 /**
424 * Create LLVM types for various structures.
425 */
426 static void
427 create_jit_types(struct draw_llvm_variant *variant)
428 {
429 struct gallivm_state *gallivm = variant->gallivm;
430 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
431 vb_type;
432
433 texture_type = create_jit_texture_type(gallivm, "texture");
434 sampler_type = create_jit_sampler_type(gallivm, "sampler");
435
436 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
437 "draw_jit_context");
438 variant->context_ptr_type = LLVMPointerType(context_type, 0);
439
440 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
441 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
442
443 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
444 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
445 }
446
447
448 static LLVMTypeRef
449 get_context_ptr_type(struct draw_llvm_variant *variant)
450 {
451 if (!variant->context_ptr_type)
452 create_jit_types(variant);
453 return variant->context_ptr_type;
454 }
455
456
457 static LLVMTypeRef
458 get_buffer_ptr_type(struct draw_llvm_variant *variant)
459 {
460 if (!variant->buffer_ptr_type)
461 create_jit_types(variant);
462 return variant->buffer_ptr_type;
463 }
464
465
466 static LLVMTypeRef
467 get_vb_ptr_type(struct draw_llvm_variant *variant)
468 {
469 if (!variant->vb_ptr_type)
470 create_jit_types(variant);
471 return variant->vb_ptr_type;
472 }
473
474 static LLVMTypeRef
475 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
476 {
477 if (!variant->vertex_header_ptr_type)
478 create_jit_types(variant);
479 return variant->vertex_header_ptr_type;
480 }
481
482
483 /**
484 * Create per-context LLVM info.
485 */
486 struct draw_llvm *
487 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
488 {
489 struct draw_llvm *llvm;
490
491 if (!lp_build_init())
492 return NULL;
493
494 llvm = CALLOC_STRUCT( draw_llvm );
495 if (!llvm)
496 return NULL;
497
498 llvm->draw = draw;
499
500 llvm->context = context;
501 if (!llvm->context) {
502 llvm->context = LLVMContextCreate();
503 llvm->context_owned = true;
504 }
505 if (!llvm->context)
506 goto fail;
507
508 llvm->nr_variants = 0;
509 make_empty_list(&llvm->vs_variants_list);
510
511 llvm->nr_gs_variants = 0;
512 make_empty_list(&llvm->gs_variants_list);
513
514 return llvm;
515
516 fail:
517 draw_llvm_destroy(llvm);
518 return NULL;
519 }
520
521
522 /**
523 * Free per-context LLVM info.
524 */
525 void
526 draw_llvm_destroy(struct draw_llvm *llvm)
527 {
528 if (llvm->context_owned)
529 LLVMContextDispose(llvm->context);
530 llvm->context = NULL;
531
532 /* XXX free other draw_llvm data? */
533 FREE(llvm);
534 }
535
536
537 /**
538 * Create LLVM-generated code for a vertex shader.
539 */
540 struct draw_llvm_variant *
541 draw_llvm_create_variant(struct draw_llvm *llvm,
542 unsigned num_inputs,
543 const struct draw_llvm_variant_key *key)
544 {
545 struct draw_llvm_variant *variant;
546 struct llvm_vertex_shader *shader =
547 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
548 LLVMTypeRef vertex_header;
549 char module_name[64];
550
551 variant = MALLOC(sizeof *variant +
552 shader->variant_key_size -
553 sizeof variant->key);
554 if (variant == NULL)
555 return NULL;
556
557 variant->llvm = llvm;
558 variant->shader = shader;
559
560 util_snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
561 variant->shader->variants_cached);
562
563 variant->gallivm = gallivm_create(module_name, llvm->context);
564
565 create_jit_types(variant);
566
567 memcpy(&variant->key, key, shader->variant_key_size);
568
569 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
570 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
571 draw_llvm_dump_variant_key(&variant->key);
572 }
573
574 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
575
576 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
577
578 draw_llvm_generate(llvm, variant, FALSE); /* linear */
579 draw_llvm_generate(llvm, variant, TRUE); /* elts */
580
581 gallivm_compile_module(variant->gallivm);
582
583 variant->jit_func = (draw_jit_vert_func)
584 gallivm_jit_function(variant->gallivm, variant->function);
585
586 variant->jit_func_elts = (draw_jit_vert_func_elts)
587 gallivm_jit_function(variant->gallivm, variant->function_elts);
588
589 gallivm_free_ir(variant->gallivm);
590
591 variant->list_item_global.base = variant;
592 variant->list_item_local.base = variant;
593 /*variant->no = */shader->variants_created++;
594 variant->list_item_global.base = variant;
595
596 return variant;
597 }
598
599
600 static void
601 generate_vs(struct draw_llvm_variant *variant,
602 LLVMBuilderRef builder,
603 struct lp_type vs_type,
604 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
605 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
606 const struct lp_bld_tgsi_system_values *system_values,
607 LLVMValueRef context_ptr,
608 struct lp_build_sampler_soa *draw_sampler,
609 boolean clamp_vertex_color)
610 {
611 struct draw_llvm *llvm = variant->llvm;
612 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
613 LLVMValueRef consts_ptr =
614 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
615 LLVMValueRef num_consts_ptr =
616 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
617
618 lp_build_tgsi_soa(variant->gallivm,
619 tokens,
620 vs_type,
621 NULL /*struct lp_build_mask_context *mask*/,
622 consts_ptr,
623 num_consts_ptr,
624 system_values,
625 inputs,
626 outputs,
627 context_ptr,
628 NULL,
629 draw_sampler,
630 &llvm->draw->vs.vertex_shader->info,
631 NULL);
632
633 {
634 LLVMValueRef out;
635 unsigned chan, attrib;
636 struct lp_build_context bld;
637 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
638 lp_build_context_init(&bld, variant->gallivm, vs_type);
639
640 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
641 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
642 if (outputs[attrib][chan]) {
643 switch (info->output_semantic_name[attrib]) {
644 case TGSI_SEMANTIC_COLOR:
645 case TGSI_SEMANTIC_BCOLOR:
646 if (clamp_vertex_color) {
647 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
648 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
649 LLVMBuildStore(builder, out, outputs[attrib][chan]);
650 }
651 break;
652 }
653 }
654 }
655 }
656 }
657 }
658
659 static void
660 generate_fetch(struct gallivm_state *gallivm,
661 struct draw_context *draw,
662 LLVMValueRef vbuffers_ptr,
663 LLVMValueRef *res,
664 struct pipe_vertex_element *velem,
665 LLVMValueRef vbuf,
666 LLVMValueRef index,
667 LLVMValueRef instance_id,
668 LLVMValueRef start_instance)
669 {
670 const struct util_format_description *format_desc =
671 util_format_description(velem->src_format);
672 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
673 LLVMBuilderRef builder = gallivm->builder;
674 LLVMValueRef indices =
675 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
676 velem->vertex_buffer_index, 0);
677 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
678 &indices, 1, "");
679 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
680 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
681 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
682 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
683 LLVMValueRef stride;
684 LLVMValueRef buffer_overflowed;
685 LLVMValueRef needed_buffer_size;
686 LLVMValueRef temp_ptr =
687 lp_build_alloca(gallivm,
688 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
689 LLVMValueRef ofbit = NULL;
690 struct lp_build_if_state if_ctx;
691
692 if (velem->instance_divisor) {
693 /* Index is equal to the start instance plus the number of current
694 * instance divided by the divisor. In this case we compute it as:
695 * index = start_instance + (instance_id / divisor)
696 */
697 LLVMValueRef current_instance;
698 current_instance = LLVMBuildUDiv(builder, instance_id,
699 lp_build_const_int32(gallivm, velem->instance_divisor),
700 "instance_divisor");
701 index = lp_build_uadd_overflow(gallivm, start_instance,
702 current_instance, &ofbit);
703 }
704
705 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
706 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
707 stride = lp_build_uadd_overflow(
708 gallivm, stride,
709 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
710 needed_buffer_size = lp_build_uadd_overflow(
711 gallivm, stride,
712 lp_build_const_int32(gallivm,
713 util_format_get_blocksize(velem->src_format)),
714 &ofbit);
715
716 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
717 needed_buffer_size, buffer_size,
718 "buffer_overflowed");
719 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
720 #if 0
721 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
722 index, vb_stride);
723 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
724 vb_buffer_offset,
725 lp_build_const_int32(gallivm, velem->src_offset));
726 lp_build_print_value(gallivm, " blocksize = ",
727 lp_build_const_int32(
728 gallivm,
729 util_format_get_blocksize(velem->src_format)));
730 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
731 lp_build_printf(gallivm, " stride = %u\n", stride);
732 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
733 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
734 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
735 #endif
736
737 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
738 {
739 LLVMValueRef val =
740 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
741 LLVMBuildStore(builder, val, temp_ptr);
742 }
743 lp_build_else(&if_ctx);
744 {
745 LLVMValueRef val;
746 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
747
748 val = lp_build_fetch_rgba_aos(gallivm,
749 format_desc,
750 lp_float32_vec4_type(),
751 FALSE,
752 map_ptr,
753 zero, zero, zero,
754 NULL);
755 LLVMBuildStore(builder, val, temp_ptr);
756 }
757 lp_build_endif(&if_ctx);
758
759 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
760 }
761
762 static void
763 convert_to_soa(struct gallivm_state *gallivm,
764 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
765 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
766 unsigned num_attribs, const struct lp_type soa_type)
767 {
768 unsigned i, j, k;
769 struct lp_type aos_channel_type = soa_type;
770
771 debug_assert(TGSI_NUM_CHANNELS == 4);
772 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
773
774 aos_channel_type.length >>= 1;
775
776 for (i = 0; i < num_attribs; ++i) {
777 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
778 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
779
780 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
781 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
782
783 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
784
785 for (k = 0; k < pixels_per_channel; ++k) {
786 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
787 }
788
789 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
790 }
791
792 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
793 }
794 }
795
796
797 static void
798 store_aos(struct gallivm_state *gallivm,
799 LLVMValueRef io_ptr,
800 LLVMValueRef index,
801 LLVMValueRef value)
802 {
803 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
804 LLVMBuilderRef builder = gallivm->builder;
805 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
806 LLVMValueRef indices[3];
807
808 indices[0] = lp_build_const_int32(gallivm, 0);
809 indices[1] = index;
810 indices[2] = lp_build_const_int32(gallivm, 0);
811
812 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
813 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
814
815 #if DEBUG_STORE
816 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
817 #endif
818
819 /* Unaligned store due to the vertex header */
820 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
821 }
822
823 /**
824 * Adjust the mask to architecture endianess. The mask will the store in struct:
825 *
826 * struct vertex_header {
827 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
828 * unsigned edgeflag:1;
829 * unsigned have_clipdist:1;
830 * unsigned vertex_id:16;
831 * [...]
832 * }
833 *
834 * On little-endian machine nothing needs to done, however on bit-endian machine
835 * the mask's fields need to be adjusted with the algorithm:
836 *
837 * uint32_t reverse (uint32_t x)
838 * {
839 * return (x >> 16) | // vertex_id
840 * ((x & 0x3fff) << 18) | // clipmask
841 * ((x & 0x4000) << 3) | // have_clipdist
842 * ((x & 0x8000) << 1); // edgeflag
843 * }
844 */
845 static LLVMValueRef
846 adjust_mask(struct gallivm_state *gallivm,
847 LLVMValueRef mask)
848 {
849 #ifdef PIPE_ARCH_BIG_ENDIAN
850 LLVMBuilderRef builder = gallivm->builder;
851 LLVMValueRef vertex_id;
852 LLVMValueRef clipmask;
853 LLVMValueRef have_clipdist;
854 LLVMValueRef edgeflag;
855
856 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
857 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
858 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
859 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
860 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
861 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
862 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
863
864 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
865 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
866 mask = LLVMBuildOr(builder, mask, edgeflag, "");
867 #endif
868 return mask;
869 }
870
871 static void
872 store_aos_array(struct gallivm_state *gallivm,
873 struct lp_type soa_type,
874 LLVMValueRef io_ptr,
875 LLVMValueRef *indices,
876 LLVMValueRef* aos,
877 int attrib,
878 int num_outputs,
879 LLVMValueRef clipmask,
880 boolean have_clipdist)
881 {
882 LLVMBuilderRef builder = gallivm->builder;
883 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
884 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
885 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
886 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
887 int vector_length = soa_type.length;
888 int i;
889
890 debug_assert(TGSI_NUM_CHANNELS == 4);
891
892 for (i = 0; i < vector_length; i++) {
893 linear_inds[i] = lp_build_const_int32(gallivm, i);
894 if (indices) {
895 inds[i] = indices[i];
896 } else {
897 inds[i] = linear_inds[i];
898 }
899 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
900 }
901
902 if (attrib == 0) {
903 /* store vertex header for each of the n vertices */
904 LLVMValueRef val, cliptmp;
905 int vertex_id_pad_edgeflag;
906
907 /* If this assertion fails, it means we need to update the bit twidding
908 * code here. See struct vertex_header in draw_private.h.
909 */
910 assert(DRAW_TOTAL_CLIP_PLANES==14);
911 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
912 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
913 if (have_clipdist)
914 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
915 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
916 /* OR with the clipmask */
917 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
918 for (i = 0; i < vector_length; i++) {
919 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
920 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
921 val = adjust_mask(gallivm, val);
922 #if DEBUG_STORE
923 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
924 io_ptrs[i], inds[i], val);
925 #endif
926 LLVMBuildStore(builder, val, id_ptr);
927 }
928 }
929
930 /* store for each of the n vertices */
931 for (i = 0; i < vector_length; i++) {
932 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
933 }
934 }
935
936
937 static void
938 convert_to_aos(struct gallivm_state *gallivm,
939 LLVMValueRef io,
940 LLVMValueRef *indices,
941 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
942 LLVMValueRef clipmask,
943 int num_outputs,
944 struct lp_type soa_type,
945 boolean have_clipdist)
946 {
947 LLVMBuilderRef builder = gallivm->builder;
948 unsigned chan, attrib, i;
949
950 #if DEBUG_STORE
951 lp_build_printf(gallivm, " # storing begin\n");
952 #endif
953 for (attrib = 0; attrib < num_outputs; ++attrib) {
954 LLVMValueRef soa[TGSI_NUM_CHANNELS];
955 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
956 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
957 if (outputs[attrib][chan]) {
958 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
959 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
960 #if DEBUG_STORE
961 lp_build_printf(gallivm, "output %d : %d ",
962 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
963 attrib, 0),
964 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
965 chan, 0));
966 lp_build_print_value(gallivm, "val = ", out);
967 {
968 LLVMValueRef iv =
969 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
970
971 lp_build_print_value(gallivm, " ival = ", iv);
972 }
973 #endif
974 soa[chan] = out;
975 }
976 else {
977 soa[chan] = 0;
978 }
979 }
980
981
982 if (soa_type.length == TGSI_NUM_CHANNELS) {
983 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
984 } else {
985 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
986
987 for (i = 0; i < soa_type.length; ++i) {
988 aos[i] = lp_build_extract_range(gallivm,
989 soa[i % TGSI_NUM_CHANNELS],
990 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
991 TGSI_NUM_CHANNELS);
992 }
993 }
994
995 store_aos_array(gallivm,
996 soa_type,
997 io, indices,
998 aos,
999 attrib,
1000 num_outputs,
1001 clipmask, have_clipdist);
1002 }
1003 #if DEBUG_STORE
1004 lp_build_printf(gallivm, " # storing end\n");
1005 #endif
1006 }
1007
1008
1009 /**
1010 * Stores original vertex positions in clip coordinates
1011 */
1012 static void
1013 store_clip(struct gallivm_state *gallivm,
1014 const struct lp_type vs_type,
1015 LLVMValueRef io_ptr,
1016 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1017 boolean pre_clip_pos, int idx)
1018 {
1019 LLVMBuilderRef builder = gallivm->builder;
1020 LLVMValueRef soa[4];
1021 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1022 LLVMValueRef indices[2];
1023 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1024 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1025 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1026 LLVMTypeRef clip_ptr_type =
1027 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1028 4), 0);
1029 int i, j;
1030
1031 indices[0] =
1032 indices[1] = lp_build_const_int32(gallivm, 0);
1033
1034 for (i = 0; i < vs_type.length; i++) {
1035 inds[i] = lp_build_const_int32(gallivm, i);
1036 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1037 }
1038
1039 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1040 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1041 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1042 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1043
1044 if (!pre_clip_pos) {
1045 for (i = 0; i < vs_type.length; i++) {
1046 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1047 }
1048 } else {
1049 for (i = 0; i < vs_type.length; i++) {
1050 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1051 }
1052 }
1053
1054 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1055 for (i = 0; i < vs_type.length; ++i) {
1056 aos[i] = lp_build_extract_range(gallivm,
1057 soa[i % TGSI_NUM_CHANNELS],
1058 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1059 TGSI_NUM_CHANNELS);
1060 }
1061
1062 for (j = 0; j < vs_type.length; j++) {
1063 LLVMValueRef clip_ptr;
1064
1065 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1066 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1067
1068 /* Unaligned store */
1069 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1070 }
1071 }
1072
1073
1074 /**
1075 * Transforms the outputs for viewport mapping
1076 */
1077 static void
1078 generate_viewport(struct draw_llvm_variant *variant,
1079 LLVMBuilderRef builder,
1080 struct lp_type vs_type,
1081 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1082 LLVMValueRef context_ptr)
1083 {
1084 int i;
1085 struct gallivm_state *gallivm = variant->gallivm;
1086 struct lp_type f32_type = vs_type;
1087 const unsigned pos = variant->llvm->draw->vs.position_output;
1088 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1089 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1090 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1091 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1092
1093 /* We treat pipe_viewport_state as a float array */
1094 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1095 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1096
1097 /* for 1/w convention*/
1098 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1099 LLVMBuildStore(builder, out3, outputs[pos][3]);
1100
1101 /* Viewport Mapping */
1102 for (i=0; i<3; i++) {
1103 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1104 LLVMValueRef scale;
1105 LLVMValueRef trans;
1106 LLVMValueRef scale_i;
1107 LLVMValueRef trans_i;
1108 LLVMValueRef index;
1109
1110 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1111 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1112
1113 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1114 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1115
1116 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1117 LLVMBuildLoad(builder, scale_i, "scale"));
1118 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1119 LLVMBuildLoad(builder, trans_i, "trans"));
1120
1121 /* divide by w */
1122 out = LLVMBuildFMul(builder, out, out3, "");
1123 /* mult by scale */
1124 out = LLVMBuildFMul(builder, out, scale, "");
1125 /* add translation */
1126 out = LLVMBuildFAdd(builder, out, trans, "");
1127
1128 /* store transformed outputs */
1129 LLVMBuildStore(builder, out, outputs[pos][i]);
1130 }
1131
1132 }
1133
1134
1135 /**
1136 * Returns clipmask as nxi32 bitmask for the n vertices
1137 */
1138 static LLVMValueRef
1139 generate_clipmask(struct draw_llvm *llvm,
1140 struct gallivm_state *gallivm,
1141 struct lp_type vs_type,
1142 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1143 boolean clip_xy,
1144 boolean clip_z,
1145 boolean clip_user,
1146 boolean clip_halfz,
1147 unsigned ucp_enable,
1148 LLVMValueRef context_ptr,
1149 boolean *have_clipdist)
1150 {
1151 LLVMBuilderRef builder = gallivm->builder;
1152 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1153 LLVMValueRef test, temp;
1154 LLVMValueRef zero, shift;
1155 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1156 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1157 LLVMValueRef plane1, planes, plane_ptr, sum;
1158 struct lp_type f32_type = vs_type;
1159 struct lp_type i32_type = lp_int_type(vs_type);
1160 const unsigned pos = llvm->draw->vs.position_output;
1161 const unsigned cv = llvm->draw->vs.clipvertex_output;
1162 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1163 bool have_cd = false;
1164 unsigned cd[2];
1165
1166 cd[0] = llvm->draw->vs.clipdistance_output[0];
1167 cd[1] = llvm->draw->vs.clipdistance_output[1];
1168
1169 if (cd[0] != pos || cd[1] != pos)
1170 have_cd = true;
1171
1172 if (num_written_clipdistance && !clip_user) {
1173 clip_user = true;
1174 ucp_enable = (1 << num_written_clipdistance) - 1;
1175 }
1176
1177 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1178 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1179 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1180 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1181
1182 /*
1183 * load clipvertex and position from correct locations.
1184 * if they are the same just load them once.
1185 */
1186 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1187 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1188 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1189 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1190
1191 if (clip_user && cv != pos) {
1192 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1193 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1194 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1195 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1196 } else {
1197 cv_x = pos_x;
1198 cv_y = pos_y;
1199 cv_z = pos_z;
1200 cv_w = pos_w;
1201 }
1202
1203 /* Cliptest, for hardwired planes */
1204 if (clip_xy) {
1205 /* plane 1 */
1206 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1207 temp = shift;
1208 test = LLVMBuildAnd(builder, test, temp, "");
1209 mask = test;
1210
1211 /* plane 2 */
1212 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1213 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1214 temp = LLVMBuildShl(builder, temp, shift, "");
1215 test = LLVMBuildAnd(builder, test, temp, "");
1216 mask = LLVMBuildOr(builder, mask, test, "");
1217
1218 /* plane 3 */
1219 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1220 temp = LLVMBuildShl(builder, temp, shift, "");
1221 test = LLVMBuildAnd(builder, test, temp, "");
1222 mask = LLVMBuildOr(builder, mask, test, "");
1223
1224 /* plane 4 */
1225 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1226 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1227 temp = LLVMBuildShl(builder, temp, shift, "");
1228 test = LLVMBuildAnd(builder, test, temp, "");
1229 mask = LLVMBuildOr(builder, mask, test, "");
1230 }
1231
1232 if (clip_z) {
1233 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1234 if (clip_halfz) {
1235 /* plane 5 */
1236 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1237 test = LLVMBuildAnd(builder, test, temp, "");
1238 mask = LLVMBuildOr(builder, mask, test, "");
1239 }
1240 else {
1241 /* plane 5 */
1242 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1243 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1244 test = LLVMBuildAnd(builder, test, temp, "");
1245 mask = LLVMBuildOr(builder, mask, test, "");
1246 }
1247 /* plane 6 */
1248 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1249 temp = LLVMBuildShl(builder, temp, shift, "");
1250 test = LLVMBuildAnd(builder, test, temp, "");
1251 mask = LLVMBuildOr(builder, mask, test, "");
1252 }
1253
1254 if (clip_user) {
1255 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1256 LLVMValueRef indices[3];
1257 LLVMValueRef is_nan_or_inf;
1258
1259 /* userclip planes */
1260 while (ucp_enable) {
1261 unsigned plane_idx = ffs(ucp_enable)-1;
1262 ucp_enable &= ~(1 << plane_idx);
1263 plane_idx += 6;
1264
1265 if (have_cd && num_written_clipdistance) {
1266 LLVMValueRef clipdist;
1267 int i;
1268 i = plane_idx - 6;
1269
1270 *have_clipdist = TRUE;
1271 if (i < 4) {
1272 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1273 } else {
1274 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1275 }
1276 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1277 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1278 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1279 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1280 test = LLVMBuildAnd(builder, test, temp, "");
1281 mask = LLVMBuildOr(builder, mask, test, "");
1282 } else {
1283 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1284 indices[0] = lp_build_const_int32(gallivm, 0);
1285 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1286
1287 indices[2] = lp_build_const_int32(gallivm, 0);
1288 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1289 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1290 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1291 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1292
1293 indices[2] = lp_build_const_int32(gallivm, 1);
1294 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1295 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1296 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1297 test = LLVMBuildFMul(builder, planes, cv_y, "");
1298 sum = LLVMBuildFAdd(builder, sum, test, "");
1299
1300 indices[2] = lp_build_const_int32(gallivm, 2);
1301 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1302 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1303 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1304 test = LLVMBuildFMul(builder, planes, cv_z, "");
1305 sum = LLVMBuildFAdd(builder, sum, test, "");
1306
1307 indices[2] = lp_build_const_int32(gallivm, 3);
1308 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1309 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1310 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1311 test = LLVMBuildFMul(builder, planes, cv_w, "");
1312 sum = LLVMBuildFAdd(builder, sum, test, "");
1313
1314 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1315 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1316 test = LLVMBuildAnd(builder, test, temp, "");
1317 mask = LLVMBuildOr(builder, mask, test, "");
1318 }
1319 }
1320 }
1321 return mask;
1322 }
1323
1324
1325 /**
1326 * Returns boolean if any clipping has occurred
1327 * Used zero/non-zero i32 value to represent boolean
1328 */
1329 static LLVMValueRef
1330 clipmask_booli32(struct gallivm_state *gallivm,
1331 const struct lp_type vs_type,
1332 LLVMValueRef clipmask_bool_ptr)
1333 {
1334 LLVMBuilderRef builder = gallivm->builder;
1335 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1336 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1337 LLVMValueRef ret = LLVMConstNull(int32_type);
1338 LLVMValueRef temp;
1339 int i;
1340
1341 /*
1342 * Can do this with log2(vector length) pack instructions and one extract
1343 * (as we don't actually need a or) with sse2 which would be way better.
1344 */
1345 for (i=0; i < vs_type.length; i++) {
1346 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1347 lp_build_const_int32(gallivm, i) , "");
1348 ret = LLVMBuildOr(builder, ret, temp, "");
1349 }
1350 return ret;
1351 }
1352
1353 static LLVMValueRef
1354 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1355 struct lp_build_tgsi_context * bld_base,
1356 boolean is_vindex_indirect,
1357 LLVMValueRef vertex_index,
1358 boolean is_aindex_indirect,
1359 LLVMValueRef attrib_index,
1360 LLVMValueRef swizzle_index)
1361 {
1362 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1363 struct gallivm_state *gallivm = bld_base->base.gallivm;
1364 LLVMBuilderRef builder = gallivm->builder;
1365 LLVMValueRef indices[3];
1366 LLVMValueRef res;
1367 struct lp_type type = bld_base->base.type;
1368
1369 if (is_vindex_indirect || is_aindex_indirect) {
1370 int i;
1371 res = bld_base->base.zero;
1372 for (i = 0; i < type.length; ++i) {
1373 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1374 LLVMValueRef vert_chan_index = vertex_index;
1375 LLVMValueRef attr_chan_index = attrib_index;
1376 LLVMValueRef channel_vec, value;
1377
1378 if (is_vindex_indirect) {
1379 vert_chan_index = LLVMBuildExtractElement(builder,
1380 vertex_index, idx, "");
1381 }
1382 if (is_aindex_indirect) {
1383 attr_chan_index = LLVMBuildExtractElement(builder,
1384 attrib_index, idx, "");
1385 }
1386
1387 indices[0] = vert_chan_index;
1388 indices[1] = attr_chan_index;
1389 indices[2] = swizzle_index;
1390
1391 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1392 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1393 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1394
1395 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1396 }
1397 } else {
1398 indices[0] = vertex_index;
1399 indices[1] = attrib_index;
1400 indices[2] = swizzle_index;
1401
1402 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1403 res = LLVMBuildLoad(builder, res, "");
1404 }
1405
1406 return res;
1407 }
1408
1409 static void
1410 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1411 struct lp_build_tgsi_context * bld_base,
1412 LLVMValueRef (*outputs)[4],
1413 LLVMValueRef emitted_vertices_vec)
1414 {
1415 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1416 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1417 struct gallivm_state *gallivm = variant->gallivm;
1418 LLVMBuilderRef builder = gallivm->builder;
1419 struct lp_type gs_type = bld_base->base.type;
1420 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1421 lp_int_type(gs_type), 0);
1422 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1423 LLVMValueRef next_prim_offset =
1424 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1425 LLVMValueRef io = variant->io_ptr;
1426 unsigned i;
1427 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1428
1429 for (i = 0; i < gs_type.length; ++i) {
1430 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1431 LLVMValueRef currently_emitted =
1432 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1433 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1434 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1435 }
1436
1437 convert_to_aos(gallivm, io, indices,
1438 outputs, clipmask,
1439 gs_info->num_outputs, gs_type,
1440 FALSE);
1441 }
1442
1443 static void
1444 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1445 struct lp_build_tgsi_context * bld_base,
1446 LLVMValueRef verts_per_prim_vec,
1447 LLVMValueRef emitted_prims_vec)
1448 {
1449 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1450 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1451 struct gallivm_state *gallivm = variant->gallivm;
1452 LLVMBuilderRef builder = gallivm->builder;
1453 LLVMValueRef prim_lengts_ptr =
1454 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1455 unsigned i;
1456
1457 for (i = 0; i < bld_base->base.type.length; ++i) {
1458 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1459 LLVMValueRef prims_emitted =
1460 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1461 LLVMValueRef store_ptr;
1462 LLVMValueRef num_vertices =
1463 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1464
1465 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1466 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1467 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1468 LLVMBuildStore(builder, num_vertices, store_ptr);
1469 }
1470 }
1471
1472 static void
1473 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1474 struct lp_build_tgsi_context * bld_base,
1475 LLVMValueRef total_emitted_vertices_vec,
1476 LLVMValueRef emitted_prims_vec)
1477 {
1478 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1479 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1480 struct gallivm_state *gallivm = variant->gallivm;
1481 LLVMBuilderRef builder = gallivm->builder;
1482 LLVMValueRef emitted_verts_ptr =
1483 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1484 LLVMValueRef emitted_prims_ptr =
1485 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1486 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1487
1488 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1489 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1490
1491 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1492 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1493 }
1494
1495 static void
1496 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1497 boolean elts)
1498 {
1499 struct gallivm_state *gallivm = variant->gallivm;
1500 LLVMContextRef context = gallivm->context;
1501 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1502 LLVMTypeRef arg_types[11];
1503 unsigned num_arg_types =
1504 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1505 LLVMTypeRef func_type;
1506 LLVMValueRef context_ptr;
1507 LLVMBasicBlockRef block;
1508 LLVMBuilderRef builder;
1509 char func_name[64];
1510 struct lp_type vs_type;
1511 LLVMValueRef end, start;
1512 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1513 LLVMValueRef vertex_id_offset, start_instance;
1514 LLVMValueRef stride, step, io_itr;
1515 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1516 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1517 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1518 struct draw_context *draw = llvm->draw;
1519 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1520 unsigned i, j;
1521 struct lp_build_context bld;
1522 struct lp_build_loop_state lp_loop;
1523 const int vector_length = lp_native_vector_width / 32;
1524 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1525 LLVMValueRef fetch_max;
1526 struct lp_build_sampler_soa *sampler = 0;
1527 LLVMValueRef ret, clipmask_bool_ptr;
1528 struct draw_llvm_variant_key *key = &variant->key;
1529 /* If geometry shader is present we need to skip both the viewport
1530 * transformation and clipping otherwise the inputs to the geometry
1531 * shader will be incorrect.
1532 * The code can't handle vp transform when vs writes vp index neither
1533 * (though this would be fixable here, but couldn't just broadcast
1534 * the values).
1535 */
1536 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1537 llvm->draw->vs.vertex_shader->info.writes_viewport_index;
1538 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1539 key->clip_z ||
1540 key->clip_user);
1541 LLVMValueRef variant_func;
1542 const unsigned pos = llvm->draw->vs.position_output;
1543 const unsigned cv = llvm->draw->vs.clipvertex_output;
1544 boolean have_clipdist = FALSE;
1545 struct lp_bld_tgsi_system_values system_values;
1546
1547 memset(&system_values, 0, sizeof(system_values));
1548
1549 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1550 variant->shader->variants_cached, elts ? "elts" : "linear");
1551
1552 i = 0;
1553 arg_types[i++] = get_context_ptr_type(variant); /* context */
1554 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1555 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1556 if (elts) {
1557 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1558 arg_types[i++] = int32_type; /* fetch_elt_max */
1559 } else
1560 arg_types[i++] = int32_type; /* start */
1561 arg_types[i++] = int32_type; /* fetch_count / count */
1562 arg_types[i++] = int32_type; /* stride */
1563 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1564 arg_types[i++] = int32_type; /* instance_id */
1565 arg_types[i++] = int32_type; /* vertex_id_offset */
1566 arg_types[i++] = int32_type; /* start_instance */
1567
1568 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1569
1570 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1571
1572 if (elts)
1573 variant->function_elts = variant_func;
1574 else
1575 variant->function = variant_func;
1576
1577 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1578 for (i = 0; i < num_arg_types; ++i)
1579 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1580 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1581 LLVMNoAliasAttribute);
1582
1583 context_ptr = LLVMGetParam(variant_func, 0);
1584 io_ptr = LLVMGetParam(variant_func, 1);
1585 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1586 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1587 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1588 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1589 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1590 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1591
1592 lp_build_name(context_ptr, "context");
1593 lp_build_name(io_ptr, "io");
1594 lp_build_name(vbuffers_ptr, "vbuffers");
1595 lp_build_name(stride, "stride");
1596 lp_build_name(vb_ptr, "vb");
1597 lp_build_name(system_values.instance_id, "instance_id");
1598 lp_build_name(vertex_id_offset, "vertex_id_offset");
1599 lp_build_name(start_instance, "start_instance");
1600
1601 if (elts) {
1602 fetch_elts = LLVMGetParam(variant_func, 3);
1603 fetch_elt_max = LLVMGetParam(variant_func, 4);
1604 fetch_count = LLVMGetParam(variant_func, 5);
1605 lp_build_name(fetch_elts, "fetch_elts");
1606 lp_build_name(fetch_elt_max, "fetch_elt_max");
1607 lp_build_name(fetch_count, "fetch_count");
1608 start = count = NULL;
1609 }
1610 else {
1611 start = LLVMGetParam(variant_func, 3);
1612 count = LLVMGetParam(variant_func, 4);
1613 lp_build_name(start, "start");
1614 lp_build_name(count, "count");
1615 fetch_elts = fetch_count = NULL;
1616 }
1617
1618 /*
1619 * Function body
1620 */
1621
1622 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1623 builder = gallivm->builder;
1624 LLVMPositionBuilderAtEnd(builder, block);
1625
1626 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1627
1628 memset(&vs_type, 0, sizeof vs_type);
1629 vs_type.floating = TRUE; /* floating point values */
1630 vs_type.sign = TRUE; /* values are signed */
1631 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1632 vs_type.width = 32; /* 32-bit float */
1633 vs_type.length = vector_length;
1634
1635 /* hold temporary "bool" clipmask */
1636 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1637 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1638
1639 /* code generated texture sampling */
1640 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1641
1642 if (elts) {
1643 start = zero;
1644 end = fetch_count;
1645 count = fetch_count;
1646 }
1647 else {
1648 end = lp_build_add(&bld, start, count);
1649 }
1650
1651 step = lp_build_const_int32(gallivm, vector_length);
1652
1653 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1654
1655 lp_build_loop_begin(&lp_loop, gallivm, zero);
1656 {
1657 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1658 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1659 LLVMValueRef io;
1660 LLVMValueRef clipmask; /* holds the clipmask value */
1661 LLVMValueRef true_index_array = lp_build_zero(gallivm,
1662 lp_type_uint_vec(32, 32*vector_length));
1663 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1664
1665 io_itr = lp_loop.counter;
1666
1667 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1668 #if DEBUG_STORE
1669 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1670 io_itr, io, lp_loop.counter);
1671 #endif
1672 for (i = 0; i < vector_length; ++i) {
1673 LLVMValueRef vert_index =
1674 LLVMBuildAdd(builder,
1675 lp_loop.counter,
1676 lp_build_const_int32(gallivm, i), "");
1677 LLVMValueRef true_index =
1678 LLVMBuildAdd(builder, start, vert_index, "");
1679
1680 /* make sure we're not out of bounds which can happen
1681 * if fetch_count % 4 != 0, because on the last iteration
1682 * a few of the 4 vertex fetches will be out of bounds */
1683 true_index = lp_build_min(&bld, true_index, fetch_max);
1684
1685 if (elts) {
1686 LLVMValueRef fetch_ptr;
1687 LLVMValueRef index_overflowed;
1688 LLVMValueRef index_ptr =
1689 lp_build_alloca(
1690 gallivm,
1691 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1692 struct lp_build_if_state if_ctx;
1693 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1694 true_index, fetch_elt_max,
1695 "index_overflowed");
1696
1697 lp_build_if(&if_ctx, gallivm, index_overflowed);
1698 {
1699 /* Generate maximum possible index so that
1700 * generate_fetch can treat it just like
1701 * any other overflow and return zeros.
1702 * We don't have to worry about the restart
1703 * primitive index because it has already been
1704 * handled
1705 */
1706 LLVMValueRef val =
1707 lp_build_const_int32(gallivm, 0xffffffff);
1708 LLVMBuildStore(builder, val, index_ptr);
1709 }
1710 lp_build_else(&if_ctx);
1711 {
1712 LLVMValueRef val;
1713 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1714 &true_index, 1, "");
1715 val = LLVMBuildLoad(builder, fetch_ptr, "");
1716 LLVMBuildStore(builder, val, index_ptr);
1717 }
1718 lp_build_endif(&if_ctx);
1719 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1720 }
1721 true_index_array = LLVMBuildInsertElement(
1722 gallivm->builder, true_index_array, true_index,
1723 lp_build_const_int32(gallivm, i), "");
1724
1725 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1726 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1727 LLVMValueRef vb_index =
1728 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1729 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1730 generate_fetch(gallivm, draw, vbuffers_ptr,
1731 &aos_attribs[j][i], velem, vb, true_index,
1732 system_values.instance_id, start_instance);
1733 }
1734 }
1735 convert_to_soa(gallivm, aos_attribs, inputs,
1736 draw->pt.nr_vertex_elements, vs_type);
1737
1738 /* In the paths with elts vertex id has to be unaffected by the
1739 * index bias and because indices inside our elements array have
1740 * already had index bias applied we need to subtract it here to
1741 * get back to the original index.
1742 * in the linear paths vertex id has to be unaffected by the
1743 * original start index and because we abuse the 'start' variable
1744 * to either represent the actual start index or the index at which
1745 * the primitive was split (we split rendering into chunks of at
1746 * most 4095-vertices) we need to back out the original start
1747 * index out of our vertex id here.
1748 */
1749 system_values.basevertex = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm,
1750 lp_type_uint_vec(32, 32*vector_length)),
1751 vertex_id_offset);
1752 system_values.vertex_id = true_index_array;
1753 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1754 system_values.basevertex, "");
1755
1756 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1757 generate_vs(variant,
1758 builder,
1759 vs_type,
1760 outputs,
1761 ptr_aos,
1762 &system_values,
1763 context_ptr,
1764 sampler,
1765 key->clamp_vertex_color);
1766
1767 if (pos != -1 && cv != -1) {
1768 /* store original positions in clip before further manipulation */
1769 store_clip(gallivm, vs_type, io, outputs, FALSE, key->clip_user ? cv : pos);
1770 store_clip(gallivm, vs_type, io, outputs, TRUE, pos);
1771
1772 /* do cliptest */
1773 if (enable_cliptest) {
1774 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1775 /* allocate clipmask, assign it integer type */
1776 clipmask = generate_clipmask(llvm,
1777 gallivm,
1778 vs_type,
1779 outputs,
1780 key->clip_xy,
1781 key->clip_z,
1782 key->clip_user,
1783 key->clip_halfz,
1784 key->ucp_enable,
1785 context_ptr, &have_clipdist);
1786 temp = LLVMBuildOr(builder, clipmask, temp, "");
1787 /* store temporary clipping boolean value */
1788 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1789 }
1790 else {
1791 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1792 }
1793
1794 /* do viewport mapping */
1795 if (!bypass_viewport) {
1796 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1797 }
1798 }
1799 else {
1800 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1801 }
1802
1803 /* store clipmask in vertex header,
1804 * original positions in clip
1805 * and transformed positions in data
1806 */
1807 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1808 vs_info->num_outputs, vs_type,
1809 have_clipdist);
1810 }
1811 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1812
1813 sampler->destroy(sampler);
1814
1815 /* return clipping boolean value for function */
1816 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1817
1818 LLVMBuildRet(builder, ret);
1819
1820 gallivm_verify_function(gallivm, variant_func);
1821 }
1822
1823
1824 struct draw_llvm_variant_key *
1825 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1826 {
1827 unsigned i;
1828 struct draw_llvm_variant_key *key;
1829 struct draw_sampler_static_state *draw_sampler;
1830
1831 key = (struct draw_llvm_variant_key *)store;
1832
1833 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1834
1835 /* Presumably all variants of the shader should have the same
1836 * number of vertex elements - ie the number of shader inputs.
1837 * NOTE: we NEED to store the needed number of needed inputs
1838 * here, not the number of provided elements to match keysize
1839 * (and the offset of sampler state in the key).
1840 */
1841 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1842 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1843
1844 /* will have to rig this up properly later */
1845 key->clip_xy = llvm->draw->clip_xy;
1846 key->clip_z = llvm->draw->clip_z;
1847 key->clip_user = llvm->draw->clip_user;
1848 key->bypass_viewport = llvm->draw->bypass_viewport;
1849 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1850 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1851 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1852 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1853 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1854 key->pad1 = 0;
1855
1856 /* All variants of this shader will have the same value for
1857 * nr_samplers. Not yet trying to compact away holes in the
1858 * sampler array.
1859 */
1860 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1861 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1862 key->nr_sampler_views =
1863 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1864 }
1865 else {
1866 key->nr_sampler_views = key->nr_samplers;
1867 }
1868
1869 draw_sampler = draw_llvm_variant_key_samplers(key);
1870
1871 memcpy(key->vertex_element,
1872 llvm->draw->pt.vertex_element,
1873 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1874
1875 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1876
1877 for (i = 0 ; i < key->nr_samplers; i++) {
1878 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1879 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1880 }
1881 for (i = 0 ; i < key->nr_sampler_views; i++) {
1882 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1883 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1884 }
1885
1886 return key;
1887 }
1888
1889
1890 void
1891 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1892 {
1893 unsigned i;
1894 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1895
1896 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1897 debug_printf("clip_xy = %u\n", key->clip_xy);
1898 debug_printf("clip_z = %u\n", key->clip_z);
1899 debug_printf("clip_user = %u\n", key->clip_user);
1900 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1901 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1902 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1903 debug_printf("has_gs = %u\n", key->has_gs);
1904 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1905
1906 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1907 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1908 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1909 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1910 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1911 }
1912
1913 for (i = 0 ; i < key->nr_sampler_views; i++) {
1914 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1915 }
1916 }
1917
1918
1919 void
1920 draw_llvm_set_mapped_texture(struct draw_context *draw,
1921 unsigned shader_stage,
1922 unsigned sview_idx,
1923 uint32_t width, uint32_t height, uint32_t depth,
1924 uint32_t first_level, uint32_t last_level,
1925 const void *base_ptr,
1926 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1927 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1928 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1929 {
1930 unsigned j;
1931 struct draw_jit_texture *jit_tex;
1932
1933 assert(shader_stage == PIPE_SHADER_VERTEX ||
1934 shader_stage == PIPE_SHADER_GEOMETRY);
1935
1936 if (shader_stage == PIPE_SHADER_VERTEX) {
1937 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1938
1939 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1940 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1941 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1942
1943 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1944 } else {
1945 assert(0);
1946 return;
1947 }
1948
1949 jit_tex->width = width;
1950 jit_tex->height = height;
1951 jit_tex->depth = depth;
1952 jit_tex->first_level = first_level;
1953 jit_tex->last_level = last_level;
1954 jit_tex->base = base_ptr;
1955
1956 for (j = first_level; j <= last_level; j++) {
1957 jit_tex->mip_offsets[j] = mip_offsets[j];
1958 jit_tex->row_stride[j] = row_stride[j];
1959 jit_tex->img_stride[j] = img_stride[j];
1960 }
1961 }
1962
1963
1964 void
1965 draw_llvm_set_sampler_state(struct draw_context *draw,
1966 unsigned shader_type)
1967 {
1968 unsigned i;
1969
1970 if (shader_type == PIPE_SHADER_VERTEX) {
1971 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1972 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1973
1974 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
1975 const struct pipe_sampler_state *s
1976 = draw->samplers[PIPE_SHADER_VERTEX][i];
1977 jit_sam->min_lod = s->min_lod;
1978 jit_sam->max_lod = s->max_lod;
1979 jit_sam->lod_bias = s->lod_bias;
1980 COPY_4V(jit_sam->border_color, s->border_color.f);
1981 }
1982 }
1983 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1984 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1985 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1986
1987 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
1988 const struct pipe_sampler_state *s
1989 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1990 jit_sam->min_lod = s->min_lod;
1991 jit_sam->max_lod = s->max_lod;
1992 jit_sam->lod_bias = s->lod_bias;
1993 COPY_4V(jit_sam->border_color, s->border_color.f);
1994 }
1995 }
1996 }
1997 }
1998
1999
2000 void
2001 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2002 {
2003 struct draw_llvm *llvm = variant->llvm;
2004
2005 gallivm_destroy(variant->gallivm);
2006
2007 remove_from_list(&variant->list_item_local);
2008 variant->shader->variants_cached--;
2009 remove_from_list(&variant->list_item_global);
2010 llvm->nr_variants--;
2011 FREE(variant);
2012 }
2013
2014
2015 /**
2016 * Create LLVM types for various structures.
2017 */
2018 static void
2019 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2020 {
2021 struct gallivm_state *gallivm = var->gallivm;
2022 LLVMTypeRef texture_type, sampler_type, context_type;
2023
2024 texture_type = create_jit_texture_type(gallivm, "texture");
2025 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2026
2027 context_type = create_gs_jit_context_type(gallivm,
2028 var->shader->base.vector_length,
2029 texture_type, sampler_type,
2030 "draw_gs_jit_context");
2031 var->context_ptr_type = LLVMPointerType(context_type, 0);
2032
2033 var->input_array_type = create_gs_jit_input_type(gallivm);
2034 }
2035
2036 static LLVMTypeRef
2037 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2038 {
2039 if (!variant->context_ptr_type)
2040 create_gs_jit_types(variant);
2041 return variant->context_ptr_type;
2042 }
2043
2044 static LLVMValueRef
2045 generate_mask_value(struct draw_gs_llvm_variant *variant,
2046 struct lp_type gs_type)
2047 {
2048 struct gallivm_state *gallivm = variant->gallivm;
2049 LLVMBuilderRef builder = gallivm->builder;
2050 struct lp_type mask_type = lp_int_type(gs_type);
2051 LLVMValueRef num_prims;
2052 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2053 unsigned i;
2054
2055 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2056 variant->num_prims);
2057 for (i = 0; i < gs_type.length; i++) {
2058 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2059 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2060 }
2061 mask_val = lp_build_compare(gallivm, mask_type,
2062 PIPE_FUNC_GREATER, num_prims, mask_val);
2063
2064 return mask_val;
2065 }
2066
2067 static void
2068 draw_gs_llvm_generate(struct draw_llvm *llvm,
2069 struct draw_gs_llvm_variant *variant)
2070 {
2071 struct gallivm_state *gallivm = variant->gallivm;
2072 LLVMContextRef context = gallivm->context;
2073 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2074 LLVMTypeRef arg_types[7];
2075 LLVMTypeRef func_type;
2076 LLVMValueRef variant_func;
2077 LLVMValueRef context_ptr;
2078 LLVMValueRef prim_id_ptr;
2079 LLVMBasicBlockRef block;
2080 LLVMBuilderRef builder;
2081 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2082 struct lp_build_sampler_soa *sampler = 0;
2083 struct lp_build_context bld;
2084 struct lp_bld_tgsi_system_values system_values;
2085 char func_name[64];
2086 struct lp_type gs_type;
2087 unsigned i;
2088 struct draw_gs_llvm_iface gs_iface;
2089 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2090 LLVMValueRef consts_ptr, num_consts_ptr;
2091 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2092 struct lp_build_mask_context mask;
2093 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2094 unsigned vector_length = variant->shader->base.vector_length;
2095
2096 memset(&system_values, 0, sizeof(system_values));
2097
2098 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2099 variant->shader->variants_cached);
2100
2101 assert(variant->vertex_header_ptr_type);
2102
2103 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2104 arg_types[1] = variant->input_array_type; /* input */
2105 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2106 arg_types[3] = int32_type; /* num_prims */
2107 arg_types[4] = int32_type; /* instance_id */
2108 arg_types[5] = LLVMPointerType(
2109 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2110 arg_types[6] = int32_type;
2111
2112 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2113
2114 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2115
2116 variant->function = variant_func;
2117
2118 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2119
2120 for (i = 0; i < Elements(arg_types); ++i)
2121 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2122 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2123 LLVMNoAliasAttribute);
2124
2125 context_ptr = LLVMGetParam(variant_func, 0);
2126 input_array = LLVMGetParam(variant_func, 1);
2127 io_ptr = LLVMGetParam(variant_func, 2);
2128 num_prims = LLVMGetParam(variant_func, 3);
2129 system_values.instance_id = LLVMGetParam(variant_func, 4);
2130 prim_id_ptr = LLVMGetParam(variant_func, 5);
2131 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2132
2133 lp_build_name(context_ptr, "context");
2134 lp_build_name(input_array, "input");
2135 lp_build_name(io_ptr, "io");
2136 lp_build_name(num_prims, "num_prims");
2137 lp_build_name(system_values.instance_id, "instance_id");
2138 lp_build_name(prim_id_ptr, "prim_id_ptr");
2139 lp_build_name(system_values.invocation_id, "invocation_id");
2140
2141 variant->context_ptr = context_ptr;
2142 variant->io_ptr = io_ptr;
2143 variant->num_prims = num_prims;
2144
2145 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2146 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2147 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2148 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2149 gs_iface.input = input_array;
2150 gs_iface.variant = variant;
2151
2152 /*
2153 * Function body
2154 */
2155
2156 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2157 builder = gallivm->builder;
2158 LLVMPositionBuilderAtEnd(builder, block);
2159
2160 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2161
2162 memset(&gs_type, 0, sizeof gs_type);
2163 gs_type.floating = TRUE; /* floating point values */
2164 gs_type.sign = TRUE; /* values are signed */
2165 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2166 gs_type.width = 32; /* 32-bit float */
2167 gs_type.length = vector_length;
2168
2169 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2170 num_consts_ptr =
2171 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2172
2173 /* code generated texture sampling */
2174 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2175
2176 mask_val = generate_mask_value(variant, gs_type);
2177 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2178
2179 if (gs_info->uses_primid) {
2180 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2181 }
2182
2183 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2184 tgsi_dump(tokens, 0);
2185 draw_gs_llvm_dump_variant_key(&variant->key);
2186 }
2187
2188 lp_build_tgsi_soa(variant->gallivm,
2189 tokens,
2190 gs_type,
2191 &mask,
2192 consts_ptr,
2193 num_consts_ptr,
2194 &system_values,
2195 NULL,
2196 outputs,
2197 context_ptr,
2198 NULL,
2199 sampler,
2200 &llvm->draw->gs.geometry_shader->info,
2201 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2202
2203 sampler->destroy(sampler);
2204
2205 lp_build_mask_end(&mask);
2206
2207 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2208
2209 gallivm_verify_function(gallivm, variant_func);
2210 }
2211
2212
2213 struct draw_gs_llvm_variant *
2214 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2215 unsigned num_outputs,
2216 const struct draw_gs_llvm_variant_key *key)
2217 {
2218 struct draw_gs_llvm_variant *variant;
2219 struct llvm_geometry_shader *shader =
2220 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2221 LLVMTypeRef vertex_header;
2222 char module_name[64];
2223
2224 variant = MALLOC(sizeof *variant +
2225 shader->variant_key_size -
2226 sizeof variant->key);
2227 if (variant == NULL)
2228 return NULL;
2229
2230 variant->llvm = llvm;
2231 variant->shader = shader;
2232
2233 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2234 variant->shader->variants_cached);
2235
2236 variant->gallivm = gallivm_create(module_name, llvm->context);
2237
2238 create_gs_jit_types(variant);
2239
2240 memcpy(&variant->key, key, shader->variant_key_size);
2241
2242 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2243
2244 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2245
2246 draw_gs_llvm_generate(llvm, variant);
2247
2248 gallivm_compile_module(variant->gallivm);
2249
2250 variant->jit_func = (draw_gs_jit_func)
2251 gallivm_jit_function(variant->gallivm, variant->function);
2252
2253 gallivm_free_ir(variant->gallivm);
2254
2255 variant->list_item_global.base = variant;
2256 variant->list_item_local.base = variant;
2257 /*variant->no = */shader->variants_created++;
2258 variant->list_item_global.base = variant;
2259
2260 return variant;
2261 }
2262
2263 void
2264 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2265 {
2266 struct draw_llvm *llvm = variant->llvm;
2267
2268 gallivm_destroy(variant->gallivm);
2269
2270 remove_from_list(&variant->list_item_local);
2271 variant->shader->variants_cached--;
2272 remove_from_list(&variant->list_item_global);
2273 llvm->nr_gs_variants--;
2274 FREE(variant);
2275 }
2276
2277 struct draw_gs_llvm_variant_key *
2278 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2279 {
2280 unsigned i;
2281 struct draw_gs_llvm_variant_key *key;
2282 struct draw_sampler_static_state *draw_sampler;
2283
2284 key = (struct draw_gs_llvm_variant_key *)store;
2285
2286 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2287
2288 /* All variants of this shader will have the same value for
2289 * nr_samplers. Not yet trying to compact away holes in the
2290 * sampler array.
2291 */
2292 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2293 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2294 key->nr_sampler_views =
2295 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2296 }
2297 else {
2298 key->nr_sampler_views = key->nr_samplers;
2299 }
2300
2301 draw_sampler = key->samplers;
2302
2303 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2304
2305 for (i = 0 ; i < key->nr_samplers; i++) {
2306 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2307 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2308 }
2309 for (i = 0 ; i < key->nr_sampler_views; i++) {
2310 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2311 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2312 }
2313
2314 return key;
2315 }
2316
2317 void
2318 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2319 {
2320 unsigned i;
2321 struct draw_sampler_static_state *sampler = key->samplers;
2322
2323 for (i = 0 ; i < key->nr_sampler_views; i++) {
2324 debug_printf("sampler[%i].src_format = %s\n", i,
2325 util_format_name(sampler[i].texture_state.format));
2326 }
2327 }