projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: improved handling of undefined inputs
[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 ARRAY_SIZE(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 ARRAY_SIZE(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 ARRAY_SIZE(elem_types), 0);
190
191 (void) target; /* silence unused var warning for non-debug build */
192 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
193 target, sampler_type,
194 DRAW_JIT_SAMPLER_MIN_LOD);
195 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
196 target, sampler_type,
197 DRAW_JIT_SAMPLER_MAX_LOD);
198 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
199 target, sampler_type,
200 DRAW_JIT_SAMPLER_LOD_BIAS);
201 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
202 target, sampler_type,
203 DRAW_JIT_SAMPLER_BORDER_COLOR);
204
205 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
206
207 return sampler_type;
208 }
209
210
211 /**
212 * Create LLVM type for struct draw_jit_context
213 */
214 static LLVMTypeRef
215 create_jit_context_type(struct gallivm_state *gallivm,
216 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
217 const char *struct_name)
218 {
219 LLVMTargetDataRef target = gallivm->target;
220 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
221 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
222 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
223 LLVMTypeRef context_type;
224
225 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
226 LP_MAX_TGSI_CONST_BUFFERS);
227 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
228 LP_MAX_TGSI_CONST_BUFFERS);
229 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
230 DRAW_TOTAL_CLIP_PLANES), 0);
231 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
232 elem_types[4] = LLVMArrayType(texture_type,
233 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
234 elem_types[5] = LLVMArrayType(sampler_type,
235 PIPE_MAX_SAMPLERS); /* samplers */
236 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
237 ARRAY_SIZE(elem_types), 0);
238
239 (void) target; /* silence unused var warning for non-debug build */
240 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
241 target, context_type, DRAW_JIT_CTX_CONSTANTS);
242 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
243 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
244 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
245 target, context_type, DRAW_JIT_CTX_PLANES);
246 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
247 target, context_type, DRAW_JIT_CTX_VIEWPORT);
248 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
249 target, context_type,
250 DRAW_JIT_CTX_TEXTURES);
251 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
252 target, context_type,
253 DRAW_JIT_CTX_SAMPLERS);
254 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
255 target, context_type);
256
257 return context_type;
258 }
259
260
261 /**
262 * Create LLVM type for struct draw_gs_jit_context
263 */
264 static LLVMTypeRef
265 create_gs_jit_context_type(struct gallivm_state *gallivm,
266 unsigned vector_length,
267 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
268 const char *struct_name)
269 {
270 LLVMTargetDataRef target = gallivm->target;
271 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
272 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
273 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
274 LLVMTypeRef context_type;
275
276 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
277 LP_MAX_TGSI_CONST_BUFFERS);
278 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
279 LP_MAX_TGSI_CONST_BUFFERS);
280 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
281 DRAW_TOTAL_CLIP_PLANES), 0);
282 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
283
284 elem_types[4] = LLVMArrayType(texture_type,
285 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
286 elem_types[5] = LLVMArrayType(sampler_type,
287 PIPE_MAX_SAMPLERS); /* samplers */
288
289 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
290 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
291 vector_length), 0);
292 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
293 vector_length), 0);
294
295 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
296 ARRAY_SIZE(elem_types), 0);
297
298 (void) target; /* silence unused var warning for non-debug build */
299 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
300 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
301 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
302 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
303 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
304 target, context_type, DRAW_GS_JIT_CTX_PLANES);
305 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
306 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
307 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
308 target, context_type,
309 DRAW_GS_JIT_CTX_TEXTURES);
310 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
311 target, context_type,
312 DRAW_GS_JIT_CTX_SAMPLERS);
313 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
314 target, context_type,
315 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
316 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
317 target, context_type,
318 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
319 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
320 target, context_type,
321 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
322 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
323 target, context_type);
324
325 return context_type;
326 }
327
328
329 static LLVMTypeRef
330 create_gs_jit_input_type(struct gallivm_state *gallivm)
331 {
332 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
333 LLVMTypeRef input_array;
334
335 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
336 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
337 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
338 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
339
340 return input_array;
341 }
342
343 /**
344 * Create LLVM type for struct pipe_vertex_buffer
345 */
346 static LLVMTypeRef
347 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
348 const char *struct_name)
349 {
350 LLVMTargetDataRef target = gallivm->target;
351 LLVMTypeRef elem_types[4];
352 LLVMTypeRef vb_type;
353
354 elem_types[0] =
355 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
356 elem_types[2] =
357 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
358
359 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
360 ARRAY_SIZE(elem_types), 0);
361
362 (void) target; /* silence unused var warning for non-debug build */
363 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
364 target, vb_type, 0);
365 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
366 target, vb_type, 1);
367
368 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
369
370 return vb_type;
371 }
372
373
374 /**
375 * Create LLVM type for struct vertex_header;
376 */
377 static LLVMTypeRef
378 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
379 {
380 LLVMTargetDataRef target = gallivm->target;
381 LLVMTypeRef elem_types[3];
382 LLVMTypeRef vertex_header;
383 char struct_name[24];
384
385 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
386
387 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
388 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
389 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
390
391 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
392 ARRAY_SIZE(elem_types), 0);
393
394 /* these are bit-fields and we can't take address of them
395 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
396 target, vertex_header,
397 DRAW_JIT_VERTEX_CLIPMASK);
398 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
399 target, vertex_header,
400 DRAW_JIT_VERTEX_EDGEFLAG);
401 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
402 target, vertex_header,
403 DRAW_JIT_VERTEX_PAD);
404 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
405 target, vertex_header,
406 DRAW_JIT_VERTEX_VERTEX_ID);
407 */
408 (void) target; /* silence unused var warning for non-debug build */
409 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
410 target, vertex_header,
411 DRAW_JIT_VERTEX_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)
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->src_format == PIPE_FORMAT_NONE) {
693 *res = lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
694 return;
695 }
696
697 if (velem->instance_divisor) {
698 /* Index is equal to the start instance plus the number of current
699 * instance divided by the divisor. In this case we compute it as:
700 * index = start_instance + (instance_id / divisor)
701 */
702 LLVMValueRef current_instance;
703 current_instance = LLVMBuildUDiv(builder, instance_id,
704 lp_build_const_int32(gallivm, velem->instance_divisor),
705 "instance_divisor");
706 index = lp_build_uadd_overflow(gallivm, start_instance,
707 current_instance, &ofbit);
708 }
709
710 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
711 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
712 stride = lp_build_uadd_overflow(
713 gallivm, stride,
714 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
715 needed_buffer_size = lp_build_uadd_overflow(
716 gallivm, stride,
717 lp_build_const_int32(gallivm,
718 util_format_get_blocksize(velem->src_format)),
719 &ofbit);
720
721 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
722 needed_buffer_size, buffer_size,
723 "buffer_overflowed");
724 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
725 #if 0
726 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
727 index, vb_stride);
728 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
729 vb_buffer_offset,
730 lp_build_const_int32(gallivm, velem->src_offset));
731 lp_build_print_value(gallivm, " blocksize = ",
732 lp_build_const_int32(
733 gallivm,
734 util_format_get_blocksize(velem->src_format)));
735 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
736 lp_build_printf(gallivm, " stride = %u\n", stride);
737 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
738 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
739 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
740 #endif
741
742 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
743 {
744 LLVMValueRef val =
745 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
746 LLVMBuildStore(builder, val, temp_ptr);
747 }
748 lp_build_else(&if_ctx);
749 {
750 LLVMValueRef val;
751 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
752
753 val = lp_build_fetch_rgba_aos(gallivm,
754 format_desc,
755 lp_float32_vec4_type(),
756 FALSE,
757 map_ptr,
758 zero, zero, zero,
759 NULL);
760 LLVMBuildStore(builder, val, temp_ptr);
761 }
762 lp_build_endif(&if_ctx);
763
764 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
765 }
766
767 static void
768 convert_to_soa(struct gallivm_state *gallivm,
769 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
770 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
771 unsigned num_attribs, const struct lp_type soa_type)
772 {
773 unsigned i, j, k;
774 struct lp_type aos_channel_type = soa_type;
775
776 debug_assert(TGSI_NUM_CHANNELS == 4);
777 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
778
779 aos_channel_type.length >>= 1;
780
781 for (i = 0; i < num_attribs; ++i) {
782 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
783 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
784
785 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
786 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
787
788 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
789
790 for (k = 0; k < pixels_per_channel; ++k) {
791 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
792 }
793
794 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
795 }
796
797 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
798 }
799 }
800
801
802 static void
803 store_aos(struct gallivm_state *gallivm,
804 LLVMValueRef io_ptr,
805 LLVMValueRef index,
806 LLVMValueRef value)
807 {
808 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
809 LLVMBuilderRef builder = gallivm->builder;
810 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
811 LLVMValueRef indices[3];
812
813 indices[0] = lp_build_const_int32(gallivm, 0);
814 indices[1] = index;
815 indices[2] = lp_build_const_int32(gallivm, 0);
816
817 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
818 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
819
820 #if DEBUG_STORE
821 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
822 #endif
823
824 /* Unaligned store due to the vertex header */
825 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
826 }
827
828 /**
829 * Adjust the mask to architecture endianess. The mask will the store in struct:
830 *
831 * struct vertex_header {
832 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
833 * unsigned edgeflag:1;
834 * unsigned pad:1;
835 * unsigned vertex_id:16;
836 * [...]
837 * }
838 *
839 * On little-endian machine nothing needs to done, however on bit-endian machine
840 * the mask's fields need to be adjusted with the algorithm:
841 *
842 * uint32_t reverse (uint32_t x)
843 * {
844 * return (x >> 16) | // vertex_id
845 * ((x & 0x3fff) << 18) | // clipmask
846 * ((x & 0x4000) << 3) | // pad
847 * ((x & 0x8000) << 1); // edgeflag
848 * }
849 */
850 static LLVMValueRef
851 adjust_mask(struct gallivm_state *gallivm,
852 LLVMValueRef mask)
853 {
854 #ifdef PIPE_ARCH_BIG_ENDIAN
855 LLVMBuilderRef builder = gallivm->builder;
856 LLVMValueRef vertex_id;
857 LLVMValueRef clipmask;
858 LLVMValueRef pad;
859 LLVMValueRef edgeflag;
860
861 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
862 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
863 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
864 if (0) {
865 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
866 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
867 }
868 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
869 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
870
871 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
872 if (0) {
873 mask = LLVMBuildOr(builder, mask, pad, "");
874 }
875 mask = LLVMBuildOr(builder, mask, edgeflag, "");
876 #endif
877 return mask;
878 }
879
880 static void
881 store_aos_array(struct gallivm_state *gallivm,
882 struct lp_type soa_type,
883 LLVMValueRef io_ptr,
884 LLVMValueRef *indices,
885 LLVMValueRef* aos,
886 int attrib,
887 int num_outputs,
888 LLVMValueRef clipmask,
889 boolean need_edgeflag)
890 {
891 LLVMBuilderRef builder = gallivm->builder;
892 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
893 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
894 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
895 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
896 int vector_length = soa_type.length;
897 int i;
898
899 debug_assert(TGSI_NUM_CHANNELS == 4);
900
901 for (i = 0; i < vector_length; i++) {
902 linear_inds[i] = lp_build_const_int32(gallivm, i);
903 if (indices) {
904 inds[i] = indices[i];
905 } else {
906 inds[i] = linear_inds[i];
907 }
908 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
909 }
910
911 if (attrib == 0) {
912 /* store vertex header for each of the n vertices */
913 LLVMValueRef val, cliptmp;
914 int vertex_id_pad_edgeflag;
915
916 /* If this assertion fails, it means we need to update the bit twidding
917 * code here. See struct vertex_header in draw_private.h.
918 */
919 assert(DRAW_TOTAL_CLIP_PLANES==14);
920 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
921 if (!need_edgeflag) {
922 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
923 }
924 else {
925 vertex_id_pad_edgeflag = (0xffff << 16);
926 }
927 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
928 vertex_id_pad_edgeflag);
929 /* OR with the clipmask */
930 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
931 for (i = 0; i < vector_length; i++) {
932 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
933 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
934 val = adjust_mask(gallivm, val);
935 #if DEBUG_STORE
936 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
937 io_ptrs[i], inds[i], val);
938 #endif
939 LLVMBuildStore(builder, val, id_ptr);
940 }
941 }
942
943 /* store for each of the n vertices */
944 for (i = 0; i < vector_length; i++) {
945 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
946 }
947 }
948
949
950 static void
951 convert_to_aos(struct gallivm_state *gallivm,
952 LLVMValueRef io,
953 LLVMValueRef *indices,
954 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
955 LLVMValueRef clipmask,
956 int num_outputs,
957 struct lp_type soa_type,
958 boolean need_edgeflag)
959 {
960 LLVMBuilderRef builder = gallivm->builder;
961 unsigned chan, attrib, i;
962
963 #if DEBUG_STORE
964 lp_build_printf(gallivm, " # storing begin\n");
965 #endif
966 for (attrib = 0; attrib < num_outputs; ++attrib) {
967 LLVMValueRef soa[TGSI_NUM_CHANNELS];
968 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
969 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
970 if (outputs[attrib][chan]) {
971 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
972 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
973 #if DEBUG_STORE
974 lp_build_printf(gallivm, "output %d : %d ",
975 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
976 attrib, 0),
977 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
978 chan, 0));
979 lp_build_print_value(gallivm, "val = ", out);
980 {
981 LLVMValueRef iv =
982 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
983
984 lp_build_print_value(gallivm, " ival = ", iv);
985 }
986 #endif
987 soa[chan] = out;
988 }
989 else {
990 soa[chan] = 0;
991 }
992 }
993
994
995 if (soa_type.length == TGSI_NUM_CHANNELS) {
996 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
997 } else {
998 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
999
1000 for (i = 0; i < soa_type.length; ++i) {
1001 aos[i] = lp_build_extract_range(gallivm,
1002 soa[i % TGSI_NUM_CHANNELS],
1003 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1004 TGSI_NUM_CHANNELS);
1005 }
1006 }
1007
1008 store_aos_array(gallivm,
1009 soa_type,
1010 io, indices,
1011 aos,
1012 attrib,
1013 num_outputs,
1014 clipmask,
1015 need_edgeflag);
1016 }
1017 #if DEBUG_STORE
1018 lp_build_printf(gallivm, " # storing end\n");
1019 #endif
1020 }
1021
1022
1023 /**
1024 * Stores original vertex positions in clip coordinates
1025 */
1026 static void
1027 store_clip(struct gallivm_state *gallivm,
1028 const struct lp_type vs_type,
1029 LLVMValueRef io_ptr,
1030 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1031 int idx)
1032 {
1033 LLVMBuilderRef builder = gallivm->builder;
1034 LLVMValueRef soa[4];
1035 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1036 LLVMValueRef indices[2];
1037 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1038 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1039 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1040 LLVMTypeRef clip_ptr_type =
1041 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1042 4), 0);
1043 int i, j;
1044
1045 indices[0] =
1046 indices[1] = lp_build_const_int32(gallivm, 0);
1047
1048 for (i = 0; i < vs_type.length; i++) {
1049 inds[i] = lp_build_const_int32(gallivm, i);
1050 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1051 }
1052
1053 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1054 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1055 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1056 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1057
1058 for (i = 0; i < vs_type.length; i++) {
1059 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1060 }
1061
1062 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1063 for (i = 0; i < vs_type.length; ++i) {
1064 aos[i] = lp_build_extract_range(gallivm,
1065 soa[i % TGSI_NUM_CHANNELS],
1066 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1067 TGSI_NUM_CHANNELS);
1068 }
1069
1070 for (j = 0; j < vs_type.length; j++) {
1071 LLVMValueRef clip_ptr;
1072
1073 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1074 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1075
1076 /* Unaligned store */
1077 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1078 }
1079 }
1080
1081
1082 /**
1083 * Transforms the outputs for viewport mapping
1084 */
1085 static void
1086 generate_viewport(struct draw_llvm_variant *variant,
1087 LLVMBuilderRef builder,
1088 struct lp_type vs_type,
1089 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1090 LLVMValueRef context_ptr)
1091 {
1092 int i;
1093 struct gallivm_state *gallivm = variant->gallivm;
1094 struct lp_type f32_type = vs_type;
1095 const unsigned pos = variant->llvm->draw->vs.position_output;
1096 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1097 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1098 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1099 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1100
1101 /* We treat pipe_viewport_state as a float array */
1102 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1103 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1104
1105 /* for 1/w convention*/
1106 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1107 LLVMBuildStore(builder, out3, outputs[pos][3]);
1108
1109 /* Viewport Mapping */
1110 for (i=0; i<3; i++) {
1111 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1112 LLVMValueRef scale;
1113 LLVMValueRef trans;
1114 LLVMValueRef scale_i;
1115 LLVMValueRef trans_i;
1116 LLVMValueRef index;
1117
1118 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1119 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1120
1121 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1122 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1123
1124 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1125 LLVMBuildLoad(builder, scale_i, "scale"));
1126 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1127 LLVMBuildLoad(builder, trans_i, "trans"));
1128
1129 /* divide by w */
1130 out = LLVMBuildFMul(builder, out, out3, "");
1131 /* mult by scale, add translation */
1132 out = lp_build_fmuladd(builder, out, scale, trans);
1133
1134 /* store transformed outputs */
1135 LLVMBuildStore(builder, out, outputs[pos][i]);
1136 }
1137
1138 }
1139
1140
1141 /**
1142 * Returns clipmask as nxi32 bitmask for the n vertices
1143 */
1144 static LLVMValueRef
1145 generate_clipmask(struct draw_llvm *llvm,
1146 struct gallivm_state *gallivm,
1147 struct lp_type vs_type,
1148 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1149 struct draw_llvm_variant_key *key,
1150 LLVMValueRef context_ptr,
1151 boolean *have_clipdist)
1152 {
1153 LLVMBuilderRef builder = gallivm->builder;
1154 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1155 LLVMValueRef test, temp;
1156 LLVMValueRef zero, shift;
1157 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1158 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1159 LLVMValueRef plane1, planes, plane_ptr, sum;
1160 struct lp_type f32_type = vs_type;
1161 struct lp_type i32_type = lp_int_type(vs_type);
1162 const unsigned pos = llvm->draw->vs.position_output;
1163 const unsigned cv = llvm->draw->vs.clipvertex_output;
1164 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1165 boolean have_cd = false;
1166 boolean clip_user = key->clip_user;
1167 unsigned ucp_enable = key->ucp_enable;
1168 unsigned cd[2];
1169
1170 cd[0] = llvm->draw->vs.ccdistance_output[0];
1171 cd[1] = llvm->draw->vs.ccdistance_output[1];
1172
1173 if (cd[0] != pos || cd[1] != pos)
1174 have_cd = true;
1175
1176 if (num_written_clipdistance && !clip_user) {
1177 clip_user = true;
1178 ucp_enable = (1 << num_written_clipdistance) - 1;
1179 }
1180
1181 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1182 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1183 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1184 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1185
1186 /*
1187 * load clipvertex and position from correct locations.
1188 * if they are the same just load them once.
1189 */
1190 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1191 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1192 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1193 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1194
1195 if (clip_user && cv != pos) {
1196 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1197 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1198 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1199 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1200 } else {
1201 cv_x = pos_x;
1202 cv_y = pos_y;
1203 cv_z = pos_z;
1204 cv_w = pos_w;
1205 }
1206
1207 /*
1208 * Be careful with the comparisons and NaNs (using llvm's unordered
1209 * comparisons here).
1210 */
1211 /* Cliptest, for hardwired planes */
1212 /*
1213 * XXX should take guardband into account (currently not in key).
1214 * Otherwise might run the draw pipeline stages for nothing.
1215 */
1216 if (key->clip_xy) {
1217 /* plane 1 */
1218 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1219 temp = shift;
1220 test = LLVMBuildAnd(builder, test, temp, "");
1221 mask = test;
1222
1223 /* plane 2 */
1224 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1225 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1226 temp = LLVMBuildShl(builder, temp, shift, "");
1227 test = LLVMBuildAnd(builder, test, temp, "");
1228 mask = LLVMBuildOr(builder, mask, test, "");
1229
1230 /* plane 3 */
1231 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1232 temp = LLVMBuildShl(builder, temp, shift, "");
1233 test = LLVMBuildAnd(builder, test, temp, "");
1234 mask = LLVMBuildOr(builder, mask, test, "");
1235
1236 /* plane 4 */
1237 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1238 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1239 temp = LLVMBuildShl(builder, temp, shift, "");
1240 test = LLVMBuildAnd(builder, test, temp, "");
1241 mask = LLVMBuildOr(builder, mask, test, "");
1242 }
1243
1244 if (key->clip_z) {
1245 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1246 if (key->clip_halfz) {
1247 /* plane 5 */
1248 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1249 test = LLVMBuildAnd(builder, test, temp, "");
1250 mask = LLVMBuildOr(builder, mask, test, "");
1251 }
1252 else {
1253 /* plane 5 */
1254 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1255 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1256 test = LLVMBuildAnd(builder, test, temp, "");
1257 mask = LLVMBuildOr(builder, mask, test, "");
1258 }
1259 /* plane 6 */
1260 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1261 temp = LLVMBuildShl(builder, temp, shift, "");
1262 test = LLVMBuildAnd(builder, test, temp, "");
1263 mask = LLVMBuildOr(builder, mask, test, "");
1264 }
1265
1266 if (clip_user) {
1267 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1268 LLVMValueRef indices[3];
1269 LLVMValueRef is_nan_or_inf;
1270
1271 /* userclip planes */
1272 while (ucp_enable) {
1273 unsigned plane_idx = ffs(ucp_enable)-1;
1274 ucp_enable &= ~(1 << plane_idx);
1275 plane_idx += 6;
1276
1277 if (have_cd && num_written_clipdistance) {
1278 LLVMValueRef clipdist;
1279 int i;
1280 i = plane_idx - 6;
1281
1282 *have_clipdist = TRUE;
1283 if (i < 4) {
1284 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1285 } else {
1286 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1287 }
1288 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1289 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1290 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1291 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1292 test = LLVMBuildAnd(builder, test, temp, "");
1293 mask = LLVMBuildOr(builder, mask, test, "");
1294 } else {
1295 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1296 indices[0] = lp_build_const_int32(gallivm, 0);
1297 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1298
1299 indices[2] = lp_build_const_int32(gallivm, 0);
1300 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1301 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1302 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1303 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1304
1305 indices[2] = lp_build_const_int32(gallivm, 1);
1306 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1307 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1308 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1309 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1310
1311 indices[2] = lp_build_const_int32(gallivm, 2);
1312 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1313 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1314 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1315 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1316
1317 indices[2] = lp_build_const_int32(gallivm, 3);
1318 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1319 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1320 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1321 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1322
1323 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1324 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1325 test = LLVMBuildAnd(builder, test, temp, "");
1326 mask = LLVMBuildOr(builder, mask, test, "");
1327 }
1328 }
1329 }
1330 if (key->need_edgeflags) {
1331 /*
1332 * This isn't really part of clipmask but stored the same in vertex
1333 * header later, so do it here.
1334 */
1335 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1336 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1337 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1338 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1339 temp = lp_build_const_int_vec(gallivm, i32_type,
1340 1LL << DRAW_TOTAL_CLIP_PLANES);
1341 test = LLVMBuildAnd(builder, test, temp, "");
1342 mask = LLVMBuildOr(builder, mask, test, "");
1343 }
1344 return mask;
1345 }
1346
1347
1348 /**
1349 * Returns boolean if any clipping has occurred
1350 * Used zero/non-zero i32 value to represent boolean
1351 */
1352 static LLVMValueRef
1353 clipmask_booli32(struct gallivm_state *gallivm,
1354 const struct lp_type vs_type,
1355 LLVMValueRef clipmask_bool_ptr,
1356 boolean edgeflag_in_clipmask)
1357 {
1358 LLVMBuilderRef builder = gallivm->builder;
1359 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1360 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1361 LLVMValueRef ret = LLVMConstNull(int32_type);
1362 LLVMValueRef temp;
1363 int i;
1364
1365 /*
1366 * We need to invert the edgeflag bit from the clipmask here
1367 * (because the result is really if we want to run the pipeline or not
1368 * and we (may) need it if edgeflag was 0).
1369 */
1370 if (edgeflag_in_clipmask) {
1371 struct lp_type i32_type = lp_int_type(vs_type);
1372 LLVMValueRef edge = lp_build_const_int_vec(gallivm, i32_type,
1373 1LL << DRAW_TOTAL_CLIP_PLANES);
1374 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1375 }
1376 /*
1377 * Could do much better with just cmp/movmskps.
1378 */
1379 for (i=0; i < vs_type.length; i++) {
1380 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1381 lp_build_const_int32(gallivm, i) , "");
1382 ret = LLVMBuildOr(builder, ret, temp, "");
1383 }
1384 return ret;
1385 }
1386
1387 static LLVMValueRef
1388 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1389 struct lp_build_tgsi_context * bld_base,
1390 boolean is_vindex_indirect,
1391 LLVMValueRef vertex_index,
1392 boolean is_aindex_indirect,
1393 LLVMValueRef attrib_index,
1394 LLVMValueRef swizzle_index)
1395 {
1396 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1397 struct gallivm_state *gallivm = bld_base->base.gallivm;
1398 LLVMBuilderRef builder = gallivm->builder;
1399 LLVMValueRef indices[3];
1400 LLVMValueRef res;
1401 struct lp_type type = bld_base->base.type;
1402
1403 if (is_vindex_indirect || is_aindex_indirect) {
1404 int i;
1405 res = bld_base->base.zero;
1406 for (i = 0; i < type.length; ++i) {
1407 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1408 LLVMValueRef vert_chan_index = vertex_index;
1409 LLVMValueRef attr_chan_index = attrib_index;
1410 LLVMValueRef channel_vec, value;
1411
1412 if (is_vindex_indirect) {
1413 vert_chan_index = LLVMBuildExtractElement(builder,
1414 vertex_index, idx, "");
1415 }
1416 if (is_aindex_indirect) {
1417 attr_chan_index = LLVMBuildExtractElement(builder,
1418 attrib_index, idx, "");
1419 }
1420
1421 indices[0] = vert_chan_index;
1422 indices[1] = attr_chan_index;
1423 indices[2] = swizzle_index;
1424
1425 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1426 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1427 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1428
1429 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1430 }
1431 } else {
1432 indices[0] = vertex_index;
1433 indices[1] = attrib_index;
1434 indices[2] = swizzle_index;
1435
1436 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1437 res = LLVMBuildLoad(builder, res, "");
1438 }
1439
1440 return res;
1441 }
1442
1443 static void
1444 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1445 struct lp_build_tgsi_context * bld_base,
1446 LLVMValueRef (*outputs)[4],
1447 LLVMValueRef emitted_vertices_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 struct lp_type gs_type = bld_base->base.type;
1454 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1455 lp_int_type(gs_type), 0);
1456 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1457 LLVMValueRef next_prim_offset =
1458 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1459 LLVMValueRef io = variant->io_ptr;
1460 unsigned i;
1461 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1462
1463 for (i = 0; i < gs_type.length; ++i) {
1464 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1465 LLVMValueRef currently_emitted =
1466 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1467 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1468 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1469 }
1470
1471 convert_to_aos(gallivm, io, indices,
1472 outputs, clipmask,
1473 gs_info->num_outputs, gs_type,
1474 FALSE);
1475 }
1476
1477 static void
1478 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1479 struct lp_build_tgsi_context * bld_base,
1480 LLVMValueRef verts_per_prim_vec,
1481 LLVMValueRef emitted_prims_vec)
1482 {
1483 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1484 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1485 struct gallivm_state *gallivm = variant->gallivm;
1486 LLVMBuilderRef builder = gallivm->builder;
1487 LLVMValueRef prim_lengts_ptr =
1488 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1489 unsigned i;
1490
1491 for (i = 0; i < bld_base->base.type.length; ++i) {
1492 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1493 LLVMValueRef prims_emitted =
1494 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1495 LLVMValueRef store_ptr;
1496 LLVMValueRef num_vertices =
1497 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1498
1499 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1500 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1501 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1502 LLVMBuildStore(builder, num_vertices, store_ptr);
1503 }
1504 }
1505
1506 static void
1507 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1508 struct lp_build_tgsi_context * bld_base,
1509 LLVMValueRef total_emitted_vertices_vec,
1510 LLVMValueRef emitted_prims_vec)
1511 {
1512 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1513 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1514 struct gallivm_state *gallivm = variant->gallivm;
1515 LLVMBuilderRef builder = gallivm->builder;
1516 LLVMValueRef emitted_verts_ptr =
1517 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1518 LLVMValueRef emitted_prims_ptr =
1519 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1520 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1521
1522 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1523 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1524
1525 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1526 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1527 }
1528
1529 static void
1530 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1531 boolean elts)
1532 {
1533 struct gallivm_state *gallivm = variant->gallivm;
1534 LLVMContextRef context = gallivm->context;
1535 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1536 LLVMTypeRef arg_types[11];
1537 unsigned num_arg_types =
1538 elts ? ARRAY_SIZE(arg_types) : ARRAY_SIZE(arg_types) - 1;
1539 LLVMTypeRef func_type;
1540 LLVMValueRef context_ptr;
1541 LLVMBasicBlockRef block;
1542 LLVMBuilderRef builder;
1543 char func_name[64];
1544 struct lp_type vs_type;
1545 LLVMValueRef end, start;
1546 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1547 LLVMValueRef vertex_id_offset, start_instance;
1548 LLVMValueRef stride, step, io_itr;
1549 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1550 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1551 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1552 struct draw_context *draw = llvm->draw;
1553 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1554 unsigned i, j;
1555 struct lp_build_context bld;
1556 struct lp_build_loop_state lp_loop;
1557 const int vector_length = lp_native_vector_width / 32;
1558 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1559 LLVMValueRef fetch_max;
1560 struct lp_build_sampler_soa *sampler = 0;
1561 LLVMValueRef ret, clipmask_bool_ptr;
1562 struct draw_llvm_variant_key *key = &variant->key;
1563 /* If geometry shader is present we need to skip both the viewport
1564 * transformation and clipping otherwise the inputs to the geometry
1565 * shader will be incorrect.
1566 * The code can't handle vp transform when vs writes vp index neither
1567 * (though this would be fixable here, but couldn't just broadcast
1568 * the values).
1569 */
1570 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1571 llvm->draw->vs.vertex_shader->info.writes_viewport_index;
1572 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1573 key->clip_z ||
1574 key->clip_user ||
1575 key->need_edgeflags);
1576 LLVMValueRef variant_func;
1577 const unsigned pos = llvm->draw->vs.position_output;
1578 const unsigned cv = llvm->draw->vs.clipvertex_output;
1579 boolean have_clipdist = FALSE;
1580 struct lp_bld_tgsi_system_values system_values;
1581
1582 memset(&system_values, 0, sizeof(system_values));
1583
1584 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1585 variant->shader->variants_cached, elts ? "elts" : "linear");
1586
1587 i = 0;
1588 arg_types[i++] = get_context_ptr_type(variant); /* context */
1589 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1590 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1591 if (elts) {
1592 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1593 arg_types[i++] = int32_type; /* fetch_elt_max */
1594 } else
1595 arg_types[i++] = int32_type; /* start */
1596 arg_types[i++] = int32_type; /* fetch_count / count */
1597 arg_types[i++] = int32_type; /* stride */
1598 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1599 arg_types[i++] = int32_type; /* instance_id */
1600 arg_types[i++] = int32_type; /* vertex_id_offset */
1601 arg_types[i++] = int32_type; /* start_instance */
1602
1603 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1604
1605 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1606
1607 if (elts)
1608 variant->function_elts = variant_func;
1609 else
1610 variant->function = variant_func;
1611
1612 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1613 for (i = 0; i < num_arg_types; ++i)
1614 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1615 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1616 LLVMNoAliasAttribute);
1617
1618 context_ptr = LLVMGetParam(variant_func, 0);
1619 io_ptr = LLVMGetParam(variant_func, 1);
1620 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1621 /*
1622 * XXX: stride is actually unused. The stride we use is strictly calculated
1623 * from the number of outputs (including the draw_extra outputs).
1624 * Should probably fix some day (we need a new vs just because of extra
1625 * outputs which the generated vs won't touch).
1626 */
1627 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1628 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1629 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1630 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1631 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1632
1633 lp_build_name(context_ptr, "context");
1634 lp_build_name(io_ptr, "io");
1635 lp_build_name(vbuffers_ptr, "vbuffers");
1636 lp_build_name(stride, "stride");
1637 lp_build_name(vb_ptr, "vb");
1638 lp_build_name(system_values.instance_id, "instance_id");
1639 lp_build_name(vertex_id_offset, "vertex_id_offset");
1640 lp_build_name(start_instance, "start_instance");
1641
1642 if (elts) {
1643 fetch_elts = LLVMGetParam(variant_func, 3);
1644 fetch_elt_max = LLVMGetParam(variant_func, 4);
1645 fetch_count = LLVMGetParam(variant_func, 5);
1646 lp_build_name(fetch_elts, "fetch_elts");
1647 lp_build_name(fetch_elt_max, "fetch_elt_max");
1648 lp_build_name(fetch_count, "fetch_count");
1649 start = count = NULL;
1650 }
1651 else {
1652 start = LLVMGetParam(variant_func, 3);
1653 count = LLVMGetParam(variant_func, 4);
1654 lp_build_name(start, "start");
1655 lp_build_name(count, "count");
1656 fetch_elts = fetch_count = NULL;
1657 }
1658
1659 /*
1660 * Function body
1661 */
1662
1663 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1664 builder = gallivm->builder;
1665 LLVMPositionBuilderAtEnd(builder, block);
1666
1667 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1668
1669 memset(&vs_type, 0, sizeof vs_type);
1670 vs_type.floating = TRUE; /* floating point values */
1671 vs_type.sign = TRUE; /* values are signed */
1672 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1673 vs_type.width = 32; /* 32-bit float */
1674 vs_type.length = vector_length;
1675
1676 /* hold temporary "bool" clipmask */
1677 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1678 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1679
1680 /* code generated texture sampling */
1681 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1682
1683 if (elts) {
1684 start = zero;
1685 end = fetch_count;
1686 count = fetch_count;
1687 }
1688 else {
1689 end = lp_build_add(&bld, start, count);
1690 }
1691
1692 step = lp_build_const_int32(gallivm, vector_length);
1693
1694 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1695
1696 lp_build_loop_begin(&lp_loop, gallivm, zero);
1697 {
1698 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1699 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1700 LLVMValueRef io;
1701 LLVMValueRef clipmask; /* holds the clipmask value */
1702 LLVMValueRef true_index_array = lp_build_zero(gallivm,
1703 lp_type_uint_vec(32, 32*vector_length));
1704 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1705
1706 io_itr = lp_loop.counter;
1707
1708 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1709 #if DEBUG_STORE
1710 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1711 io_itr, io, lp_loop.counter);
1712 #endif
1713
1714 for (i = 0; i < vector_length; ++i) {
1715 LLVMValueRef vert_index =
1716 LLVMBuildAdd(builder,
1717 lp_loop.counter,
1718 lp_build_const_int32(gallivm, i), "");
1719 LLVMValueRef true_index =
1720 LLVMBuildAdd(builder, start, vert_index, "");
1721
1722 /* make sure we're not out of bounds which can happen
1723 * if fetch_count % 4 != 0, because on the last iteration
1724 * a few of the 4 vertex fetches will be out of bounds */
1725 true_index = lp_build_min(&bld, true_index, fetch_max);
1726
1727 if (elts) {
1728 LLVMValueRef fetch_ptr;
1729 LLVMValueRef index_overflowed;
1730 LLVMValueRef index_ptr =
1731 lp_build_alloca(
1732 gallivm,
1733 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1734 struct lp_build_if_state if_ctx;
1735 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1736 true_index, fetch_elt_max,
1737 "index_overflowed");
1738
1739 lp_build_if(&if_ctx, gallivm, index_overflowed);
1740 {
1741 /* Generate maximum possible index so that
1742 * generate_fetch can treat it just like
1743 * any other overflow and return zeros.
1744 * We don't have to worry about the restart
1745 * primitive index because it has already been
1746 * handled
1747 */
1748 LLVMValueRef val =
1749 lp_build_const_int32(gallivm, 0xffffffff);
1750 LLVMBuildStore(builder, val, index_ptr);
1751 }
1752 lp_build_else(&if_ctx);
1753 {
1754 LLVMValueRef val;
1755 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1756 &true_index, 1, "");
1757 val = LLVMBuildLoad(builder, fetch_ptr, "");
1758 LLVMBuildStore(builder, val, index_ptr);
1759 }
1760 lp_build_endif(&if_ctx);
1761 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1762 }
1763 true_index_array = LLVMBuildInsertElement(
1764 gallivm->builder, true_index_array, true_index,
1765 lp_build_const_int32(gallivm, i), "");
1766
1767 for (j = 0; j < key->nr_vertex_elements; ++j) {
1768 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1769 LLVMValueRef vb_index =
1770 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1771 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1772 generate_fetch(gallivm, draw, vbuffers_ptr,
1773 &aos_attribs[j][i], velem, vb, true_index,
1774 system_values.instance_id, start_instance);
1775 }
1776 }
1777 convert_to_soa(gallivm, aos_attribs, inputs,
1778 key->nr_vertex_elements, vs_type);
1779
1780 /* In the paths with elts vertex id has to be unaffected by the
1781 * index bias and because indices inside our elements array have
1782 * already had index bias applied we need to subtract it here to
1783 * get back to the original index.
1784 * in the linear paths vertex id has to be unaffected by the
1785 * original start index and because we abuse the 'start' variable
1786 * to either represent the actual start index or the index at which
1787 * the primitive was split (we split rendering into chunks of at
1788 * most 4095-vertices) we need to back out the original start
1789 * index out of our vertex id here.
1790 */
1791 system_values.basevertex = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm,
1792 lp_type_uint_vec(32, 32*vector_length)),
1793 vertex_id_offset);
1794 system_values.vertex_id = true_index_array;
1795 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1796 system_values.basevertex, "");
1797
1798 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1799 generate_vs(variant,
1800 builder,
1801 vs_type,
1802 outputs,
1803 ptr_aos,
1804 &system_values,
1805 context_ptr,
1806 sampler,
1807 key->clamp_vertex_color);
1808
1809 if (pos != -1 && cv != -1) {
1810 /* store original positions in clip before further manipulation */
1811 store_clip(gallivm, vs_type, io, outputs, pos);
1812
1813 /* do cliptest */
1814 if (enable_cliptest) {
1815 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1816 /* allocate clipmask, assign it integer type */
1817 clipmask = generate_clipmask(llvm,
1818 gallivm,
1819 vs_type,
1820 outputs,
1821 key,
1822 context_ptr, &have_clipdist);
1823 temp = LLVMBuildOr(builder, clipmask, temp, "");
1824 /* store temporary clipping boolean value */
1825 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1826 }
1827 else {
1828 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1829 }
1830
1831 /* do viewport mapping */
1832 if (!bypass_viewport) {
1833 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1834 }
1835 }
1836 else {
1837 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1838 }
1839
1840 /* store clipmask in vertex header,
1841 * original positions in clip
1842 * and transformed positions in data
1843 */
1844 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1845 vs_info->num_outputs, vs_type,
1846 enable_cliptest && key->need_edgeflags);
1847 }
1848 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1849
1850 sampler->destroy(sampler);
1851
1852 /* return clipping boolean value for function */
1853 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr,
1854 enable_cliptest && key->need_edgeflags);
1855
1856 LLVMBuildRet(builder, ret);
1857
1858 gallivm_verify_function(gallivm, variant_func);
1859 }
1860
1861
1862 struct draw_llvm_variant_key *
1863 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1864 {
1865 unsigned i;
1866 struct draw_llvm_variant_key *key;
1867 struct draw_sampler_static_state *draw_sampler;
1868
1869 key = (struct draw_llvm_variant_key *)store;
1870
1871 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
1872
1873 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1874
1875 /* will have to rig this up properly later */
1876 key->clip_xy = llvm->draw->clip_xy;
1877 key->clip_z = llvm->draw->clip_z;
1878 key->clip_user = llvm->draw->clip_user;
1879 key->bypass_viewport = llvm->draw->bypass_viewport;
1880 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1881 /* XXX assumes edgeflag output not at 0 */
1882 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1883 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1884 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1885 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1886
1887 /* All variants of this shader will have the same value for
1888 * nr_samplers. Not yet trying to compact away holes in the
1889 * sampler array.
1890 */
1891 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1892 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1893 key->nr_sampler_views =
1894 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1895 }
1896 else {
1897 key->nr_sampler_views = key->nr_samplers;
1898 }
1899
1900 /* Presumably all variants of the shader should have the same
1901 * number of vertex elements - ie the number of shader inputs.
1902 * NOTE: we NEED to store the needed number of needed inputs
1903 * here, not the number of provided elements to match keysize
1904 * (and the offset of sampler state in the key).
1905 * If we have excess number of vertex elements, this is valid,
1906 * but the excess ones don't matter.
1907 * If we don't have enough vertex elements (which looks not really
1908 * valid but we'll handle it gracefully) fill out missing ones with
1909 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
1910 */
1911 key->nr_vertex_elements =
1912 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1913
1914 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
1915 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
1916 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
1917 memset(key->vertex_element, 0,
1918 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1919 }
1920 memcpy(key->vertex_element,
1921 llvm->draw->pt.vertex_element,
1922 sizeof(struct pipe_vertex_element) *
1923 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
1924
1925 draw_sampler = draw_llvm_variant_key_samplers(key);
1926 memset(draw_sampler, 0,
1927 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1928
1929 for (i = 0 ; i < key->nr_samplers; i++) {
1930 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1931 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1932 }
1933 for (i = 0 ; i < key->nr_sampler_views; i++) {
1934 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1935 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1936 }
1937
1938 return key;
1939 }
1940
1941
1942 void
1943 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1944 {
1945 unsigned i;
1946 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1947
1948 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1949 debug_printf("clip_xy = %u\n", key->clip_xy);
1950 debug_printf("clip_z = %u\n", key->clip_z);
1951 debug_printf("clip_user = %u\n", key->clip_user);
1952 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1953 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1954 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1955 debug_printf("has_gs = %u\n", key->has_gs);
1956 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1957
1958 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1959 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1960 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1961 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1962 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1963 }
1964
1965 for (i = 0 ; i < key->nr_sampler_views; i++) {
1966 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1967 }
1968 }
1969
1970
1971 void
1972 draw_llvm_set_mapped_texture(struct draw_context *draw,
1973 unsigned shader_stage,
1974 unsigned sview_idx,
1975 uint32_t width, uint32_t height, uint32_t depth,
1976 uint32_t first_level, uint32_t last_level,
1977 const void *base_ptr,
1978 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1979 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1980 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1981 {
1982 unsigned j;
1983 struct draw_jit_texture *jit_tex;
1984
1985 assert(shader_stage == PIPE_SHADER_VERTEX ||
1986 shader_stage == PIPE_SHADER_GEOMETRY);
1987
1988 if (shader_stage == PIPE_SHADER_VERTEX) {
1989 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
1990
1991 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1992 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1993 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
1994
1995 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1996 } else {
1997 assert(0);
1998 return;
1999 }
2000
2001 jit_tex->width = width;
2002 jit_tex->height = height;
2003 jit_tex->depth = depth;
2004 jit_tex->first_level = first_level;
2005 jit_tex->last_level = last_level;
2006 jit_tex->base = base_ptr;
2007
2008 for (j = first_level; j <= last_level; j++) {
2009 jit_tex->mip_offsets[j] = mip_offsets[j];
2010 jit_tex->row_stride[j] = row_stride[j];
2011 jit_tex->img_stride[j] = img_stride[j];
2012 }
2013 }
2014
2015
2016 void
2017 draw_llvm_set_sampler_state(struct draw_context *draw,
2018 unsigned shader_type)
2019 {
2020 unsigned i;
2021
2022 if (shader_type == PIPE_SHADER_VERTEX) {
2023 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2024 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2025
2026 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2027 const struct pipe_sampler_state *s
2028 = draw->samplers[PIPE_SHADER_VERTEX][i];
2029 jit_sam->min_lod = s->min_lod;
2030 jit_sam->max_lod = s->max_lod;
2031 jit_sam->lod_bias = s->lod_bias;
2032 COPY_4V(jit_sam->border_color, s->border_color.f);
2033 }
2034 }
2035 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2036 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2037 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2038
2039 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2040 const struct pipe_sampler_state *s
2041 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2042 jit_sam->min_lod = s->min_lod;
2043 jit_sam->max_lod = s->max_lod;
2044 jit_sam->lod_bias = s->lod_bias;
2045 COPY_4V(jit_sam->border_color, s->border_color.f);
2046 }
2047 }
2048 }
2049 }
2050
2051
2052 void
2053 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2054 {
2055 struct draw_llvm *llvm = variant->llvm;
2056
2057 gallivm_destroy(variant->gallivm);
2058
2059 remove_from_list(&variant->list_item_local);
2060 variant->shader->variants_cached--;
2061 remove_from_list(&variant->list_item_global);
2062 llvm->nr_variants--;
2063 FREE(variant);
2064 }
2065
2066
2067 /**
2068 * Create LLVM types for various structures.
2069 */
2070 static void
2071 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2072 {
2073 struct gallivm_state *gallivm = var->gallivm;
2074 LLVMTypeRef texture_type, sampler_type, context_type;
2075
2076 texture_type = create_jit_texture_type(gallivm, "texture");
2077 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2078
2079 context_type = create_gs_jit_context_type(gallivm,
2080 var->shader->base.vector_length,
2081 texture_type, sampler_type,
2082 "draw_gs_jit_context");
2083 var->context_ptr_type = LLVMPointerType(context_type, 0);
2084
2085 var->input_array_type = create_gs_jit_input_type(gallivm);
2086 }
2087
2088 static LLVMTypeRef
2089 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2090 {
2091 if (!variant->context_ptr_type)
2092 create_gs_jit_types(variant);
2093 return variant->context_ptr_type;
2094 }
2095
2096 static LLVMValueRef
2097 generate_mask_value(struct draw_gs_llvm_variant *variant,
2098 struct lp_type gs_type)
2099 {
2100 struct gallivm_state *gallivm = variant->gallivm;
2101 LLVMBuilderRef builder = gallivm->builder;
2102 struct lp_type mask_type = lp_int_type(gs_type);
2103 LLVMValueRef num_prims;
2104 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2105 unsigned i;
2106
2107 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2108 variant->num_prims);
2109 for (i = 0; i < gs_type.length; i++) {
2110 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2111 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2112 }
2113 mask_val = lp_build_compare(gallivm, mask_type,
2114 PIPE_FUNC_GREATER, num_prims, mask_val);
2115
2116 return mask_val;
2117 }
2118
2119 static void
2120 draw_gs_llvm_generate(struct draw_llvm *llvm,
2121 struct draw_gs_llvm_variant *variant)
2122 {
2123 struct gallivm_state *gallivm = variant->gallivm;
2124 LLVMContextRef context = gallivm->context;
2125 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2126 LLVMTypeRef arg_types[7];
2127 LLVMTypeRef func_type;
2128 LLVMValueRef variant_func;
2129 LLVMValueRef context_ptr;
2130 LLVMValueRef prim_id_ptr;
2131 LLVMBasicBlockRef block;
2132 LLVMBuilderRef builder;
2133 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2134 struct lp_build_sampler_soa *sampler = 0;
2135 struct lp_build_context bld;
2136 struct lp_bld_tgsi_system_values system_values;
2137 char func_name[64];
2138 struct lp_type gs_type;
2139 unsigned i;
2140 struct draw_gs_llvm_iface gs_iface;
2141 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2142 LLVMValueRef consts_ptr, num_consts_ptr;
2143 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2144 struct lp_build_mask_context mask;
2145 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2146 unsigned vector_length = variant->shader->base.vector_length;
2147
2148 memset(&system_values, 0, sizeof(system_values));
2149
2150 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2151 variant->shader->variants_cached);
2152
2153 assert(variant->vertex_header_ptr_type);
2154
2155 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2156 arg_types[1] = variant->input_array_type; /* input */
2157 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2158 arg_types[3] = int32_type; /* num_prims */
2159 arg_types[4] = int32_type; /* instance_id */
2160 arg_types[5] = LLVMPointerType(
2161 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2162 arg_types[6] = int32_type;
2163
2164 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2165
2166 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2167
2168 variant->function = variant_func;
2169
2170 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2171
2172 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2173 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2174 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2175 LLVMNoAliasAttribute);
2176
2177 context_ptr = LLVMGetParam(variant_func, 0);
2178 input_array = LLVMGetParam(variant_func, 1);
2179 io_ptr = LLVMGetParam(variant_func, 2);
2180 num_prims = LLVMGetParam(variant_func, 3);
2181 system_values.instance_id = LLVMGetParam(variant_func, 4);
2182 prim_id_ptr = LLVMGetParam(variant_func, 5);
2183 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2184
2185 lp_build_name(context_ptr, "context");
2186 lp_build_name(input_array, "input");
2187 lp_build_name(io_ptr, "io");
2188 lp_build_name(num_prims, "num_prims");
2189 lp_build_name(system_values.instance_id, "instance_id");
2190 lp_build_name(prim_id_ptr, "prim_id_ptr");
2191 lp_build_name(system_values.invocation_id, "invocation_id");
2192
2193 variant->context_ptr = context_ptr;
2194 variant->io_ptr = io_ptr;
2195 variant->num_prims = num_prims;
2196
2197 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2198 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2199 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2200 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2201 gs_iface.input = input_array;
2202 gs_iface.variant = variant;
2203
2204 /*
2205 * Function body
2206 */
2207
2208 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2209 builder = gallivm->builder;
2210 LLVMPositionBuilderAtEnd(builder, block);
2211
2212 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2213
2214 memset(&gs_type, 0, sizeof gs_type);
2215 gs_type.floating = TRUE; /* floating point values */
2216 gs_type.sign = TRUE; /* values are signed */
2217 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2218 gs_type.width = 32; /* 32-bit float */
2219 gs_type.length = vector_length;
2220
2221 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2222 num_consts_ptr =
2223 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2224
2225 /* code generated texture sampling */
2226 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2227
2228 mask_val = generate_mask_value(variant, gs_type);
2229 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2230
2231 if (gs_info->uses_primid) {
2232 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2233 }
2234
2235 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2236 tgsi_dump(tokens, 0);
2237 draw_gs_llvm_dump_variant_key(&variant->key);
2238 }
2239
2240 lp_build_tgsi_soa(variant->gallivm,
2241 tokens,
2242 gs_type,
2243 &mask,
2244 consts_ptr,
2245 num_consts_ptr,
2246 &system_values,
2247 NULL,
2248 outputs,
2249 context_ptr,
2250 NULL,
2251 sampler,
2252 &llvm->draw->gs.geometry_shader->info,
2253 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2254
2255 sampler->destroy(sampler);
2256
2257 lp_build_mask_end(&mask);
2258
2259 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2260
2261 gallivm_verify_function(gallivm, variant_func);
2262 }
2263
2264
2265 struct draw_gs_llvm_variant *
2266 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2267 unsigned num_outputs,
2268 const struct draw_gs_llvm_variant_key *key)
2269 {
2270 struct draw_gs_llvm_variant *variant;
2271 struct llvm_geometry_shader *shader =
2272 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2273 LLVMTypeRef vertex_header;
2274 char module_name[64];
2275
2276 variant = MALLOC(sizeof *variant +
2277 shader->variant_key_size -
2278 sizeof variant->key);
2279 if (!variant)
2280 return NULL;
2281
2282 variant->llvm = llvm;
2283 variant->shader = shader;
2284
2285 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2286 variant->shader->variants_cached);
2287
2288 variant->gallivm = gallivm_create(module_name, llvm->context);
2289
2290 create_gs_jit_types(variant);
2291
2292 memcpy(&variant->key, key, shader->variant_key_size);
2293
2294 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2295
2296 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2297
2298 draw_gs_llvm_generate(llvm, variant);
2299
2300 gallivm_compile_module(variant->gallivm);
2301
2302 variant->jit_func = (draw_gs_jit_func)
2303 gallivm_jit_function(variant->gallivm, variant->function);
2304
2305 gallivm_free_ir(variant->gallivm);
2306
2307 variant->list_item_global.base = variant;
2308 variant->list_item_local.base = variant;
2309 /*variant->no = */shader->variants_created++;
2310 variant->list_item_global.base = variant;
2311
2312 return variant;
2313 }
2314
2315 void
2316 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2317 {
2318 struct draw_llvm *llvm = variant->llvm;
2319
2320 gallivm_destroy(variant->gallivm);
2321
2322 remove_from_list(&variant->list_item_local);
2323 variant->shader->variants_cached--;
2324 remove_from_list(&variant->list_item_global);
2325 llvm->nr_gs_variants--;
2326 FREE(variant);
2327 }
2328
2329 struct draw_gs_llvm_variant_key *
2330 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2331 {
2332 unsigned i;
2333 struct draw_gs_llvm_variant_key *key;
2334 struct draw_sampler_static_state *draw_sampler;
2335
2336 key = (struct draw_gs_llvm_variant_key *)store;
2337
2338 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2339
2340 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2341
2342 /* All variants of this shader will have the same value for
2343 * nr_samplers. Not yet trying to compact away holes in the
2344 * sampler array.
2345 */
2346 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2347 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2348 key->nr_sampler_views =
2349 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2350 }
2351 else {
2352 key->nr_sampler_views = key->nr_samplers;
2353 }
2354
2355 draw_sampler = key->samplers;
2356
2357 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2358
2359 for (i = 0 ; i < key->nr_samplers; i++) {
2360 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2361 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2362 }
2363 for (i = 0 ; i < key->nr_sampler_views; i++) {
2364 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2365 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2366 }
2367
2368 return key;
2369 }
2370
2371 void
2372 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2373 {
2374 unsigned i;
2375 struct draw_sampler_static_state *sampler = key->samplers;
2376
2377 for (i = 0 ; i < key->nr_sampler_views; i++) {
2378 debug_printf("sampler[%i].src_format = %s\n", i,
2379 util_format_name(sampler[i].texture_state.format));
2380 }
2381 }