projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: don't pretend have_clipdist is per-vertex
[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)
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 pad: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) | // pad
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 pad;
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 if (0) {
860 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
861 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
862 }
863 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
864 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
865
866 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
867 if (0) {
868 mask = LLVMBuildOr(builder, mask, pad, "");
869 }
870 mask = LLVMBuildOr(builder, mask, edgeflag, "");
871 #endif
872 return mask;
873 }
874
875 static void
876 store_aos_array(struct gallivm_state *gallivm,
877 struct lp_type soa_type,
878 LLVMValueRef io_ptr,
879 LLVMValueRef *indices,
880 LLVMValueRef* aos,
881 int attrib,
882 int num_outputs,
883 LLVMValueRef clipmask)
884 {
885 LLVMBuilderRef builder = gallivm->builder;
886 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
887 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
888 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
889 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
890 int vector_length = soa_type.length;
891 int i;
892
893 debug_assert(TGSI_NUM_CHANNELS == 4);
894
895 for (i = 0; i < vector_length; i++) {
896 linear_inds[i] = lp_build_const_int32(gallivm, i);
897 if (indices) {
898 inds[i] = indices[i];
899 } else {
900 inds[i] = linear_inds[i];
901 }
902 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
903 }
904
905 if (attrib == 0) {
906 /* store vertex header for each of the n vertices */
907 LLVMValueRef val, cliptmp;
908 int vertex_id_pad_edgeflag;
909
910 /* If this assertion fails, it means we need to update the bit twidding
911 * code here. See struct vertex_header in draw_private.h.
912 */
913 assert(DRAW_TOTAL_CLIP_PLANES==14);
914 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
915 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
916 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
917 /* OR with the clipmask */
918 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
919 for (i = 0; i < vector_length; i++) {
920 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
921 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
922 val = adjust_mask(gallivm, val);
923 #if DEBUG_STORE
924 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
925 io_ptrs[i], inds[i], val);
926 #endif
927 LLVMBuildStore(builder, val, id_ptr);
928 }
929 }
930
931 /* store for each of the n vertices */
932 for (i = 0; i < vector_length; i++) {
933 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
934 }
935 }
936
937
938 static void
939 convert_to_aos(struct gallivm_state *gallivm,
940 LLVMValueRef io,
941 LLVMValueRef *indices,
942 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
943 LLVMValueRef clipmask,
944 int num_outputs,
945 struct lp_type soa_type,
946 boolean have_clipdist)
947 {
948 LLVMBuilderRef builder = gallivm->builder;
949 unsigned chan, attrib, i;
950
951 #if DEBUG_STORE
952 lp_build_printf(gallivm, " # storing begin\n");
953 #endif
954 for (attrib = 0; attrib < num_outputs; ++attrib) {
955 LLVMValueRef soa[TGSI_NUM_CHANNELS];
956 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
957 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
958 if (outputs[attrib][chan]) {
959 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
960 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
961 #if DEBUG_STORE
962 lp_build_printf(gallivm, "output %d : %d ",
963 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
964 attrib, 0),
965 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
966 chan, 0));
967 lp_build_print_value(gallivm, "val = ", out);
968 {
969 LLVMValueRef iv =
970 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
971
972 lp_build_print_value(gallivm, " ival = ", iv);
973 }
974 #endif
975 soa[chan] = out;
976 }
977 else {
978 soa[chan] = 0;
979 }
980 }
981
982
983 if (soa_type.length == TGSI_NUM_CHANNELS) {
984 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
985 } else {
986 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
987
988 for (i = 0; i < soa_type.length; ++i) {
989 aos[i] = lp_build_extract_range(gallivm,
990 soa[i % TGSI_NUM_CHANNELS],
991 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
992 TGSI_NUM_CHANNELS);
993 }
994 }
995
996 store_aos_array(gallivm,
997 soa_type,
998 io, indices,
999 aos,
1000 attrib,
1001 num_outputs,
1002 clipmask);
1003 }
1004 #if DEBUG_STORE
1005 lp_build_printf(gallivm, " # storing end\n");
1006 #endif
1007 }
1008
1009
1010 /**
1011 * Stores original vertex positions in clip coordinates
1012 */
1013 static void
1014 store_clip(struct gallivm_state *gallivm,
1015 const struct lp_type vs_type,
1016 LLVMValueRef io_ptr,
1017 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1018 boolean pre_clip_pos, int idx)
1019 {
1020 LLVMBuilderRef builder = gallivm->builder;
1021 LLVMValueRef soa[4];
1022 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1023 LLVMValueRef indices[2];
1024 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1025 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1026 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1027 LLVMTypeRef clip_ptr_type =
1028 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1029 4), 0);
1030 int i, j;
1031
1032 indices[0] =
1033 indices[1] = lp_build_const_int32(gallivm, 0);
1034
1035 for (i = 0; i < vs_type.length; i++) {
1036 inds[i] = lp_build_const_int32(gallivm, i);
1037 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1038 }
1039
1040 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1041 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1042 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1043 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1044
1045 if (!pre_clip_pos) {
1046 for (i = 0; i < vs_type.length; i++) {
1047 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1048 }
1049 } else {
1050 for (i = 0; i < vs_type.length; i++) {
1051 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1052 }
1053 }
1054
1055 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1056 for (i = 0; i < vs_type.length; ++i) {
1057 aos[i] = lp_build_extract_range(gallivm,
1058 soa[i % TGSI_NUM_CHANNELS],
1059 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1060 TGSI_NUM_CHANNELS);
1061 }
1062
1063 for (j = 0; j < vs_type.length; j++) {
1064 LLVMValueRef clip_ptr;
1065
1066 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1067 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1068
1069 /* Unaligned store */
1070 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1071 }
1072 }
1073
1074
1075 /**
1076 * Transforms the outputs for viewport mapping
1077 */
1078 static void
1079 generate_viewport(struct draw_llvm_variant *variant,
1080 LLVMBuilderRef builder,
1081 struct lp_type vs_type,
1082 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1083 LLVMValueRef context_ptr)
1084 {
1085 int i;
1086 struct gallivm_state *gallivm = variant->gallivm;
1087 struct lp_type f32_type = vs_type;
1088 const unsigned pos = variant->llvm->draw->vs.position_output;
1089 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1090 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1091 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1092 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1093
1094 /* We treat pipe_viewport_state as a float array */
1095 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1096 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1097
1098 /* for 1/w convention*/
1099 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1100 LLVMBuildStore(builder, out3, outputs[pos][3]);
1101
1102 /* Viewport Mapping */
1103 for (i=0; i<3; i++) {
1104 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1105 LLVMValueRef scale;
1106 LLVMValueRef trans;
1107 LLVMValueRef scale_i;
1108 LLVMValueRef trans_i;
1109 LLVMValueRef index;
1110
1111 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1112 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1113
1114 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1115 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1116
1117 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1118 LLVMBuildLoad(builder, scale_i, "scale"));
1119 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1120 LLVMBuildLoad(builder, trans_i, "trans"));
1121
1122 /* divide by w */
1123 out = LLVMBuildFMul(builder, out, out3, "");
1124 /* mult by scale */
1125 out = LLVMBuildFMul(builder, out, scale, "");
1126 /* add translation */
1127 out = LLVMBuildFAdd(builder, out, trans, "");
1128
1129 /* store transformed outputs */
1130 LLVMBuildStore(builder, out, outputs[pos][i]);
1131 }
1132
1133 }
1134
1135
1136 /**
1137 * Returns clipmask as nxi32 bitmask for the n vertices
1138 */
1139 static LLVMValueRef
1140 generate_clipmask(struct draw_llvm *llvm,
1141 struct gallivm_state *gallivm,
1142 struct lp_type vs_type,
1143 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1144 boolean clip_xy,
1145 boolean clip_z,
1146 boolean clip_user,
1147 boolean clip_halfz,
1148 unsigned ucp_enable,
1149 LLVMValueRef context_ptr,
1150 boolean *have_clipdist)
1151 {
1152 LLVMBuilderRef builder = gallivm->builder;
1153 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1154 LLVMValueRef test, temp;
1155 LLVMValueRef zero, shift;
1156 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1157 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1158 LLVMValueRef plane1, planes, plane_ptr, sum;
1159 struct lp_type f32_type = vs_type;
1160 struct lp_type i32_type = lp_int_type(vs_type);
1161 const unsigned pos = llvm->draw->vs.position_output;
1162 const unsigned cv = llvm->draw->vs.clipvertex_output;
1163 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1164 bool have_cd = false;
1165 unsigned cd[2];
1166
1167 cd[0] = llvm->draw->vs.clipdistance_output[0];
1168 cd[1] = llvm->draw->vs.clipdistance_output[1];
1169
1170 if (cd[0] != pos || cd[1] != pos)
1171 have_cd = true;
1172
1173 if (num_written_clipdistance && !clip_user) {
1174 clip_user = true;
1175 ucp_enable = (1 << num_written_clipdistance) - 1;
1176 }
1177
1178 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1179 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1180 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1181 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1182
1183 /*
1184 * load clipvertex and position from correct locations.
1185 * if they are the same just load them once.
1186 */
1187 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1188 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1189 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1190 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1191
1192 if (clip_user && cv != pos) {
1193 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1194 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1195 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1196 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1197 } else {
1198 cv_x = pos_x;
1199 cv_y = pos_y;
1200 cv_z = pos_z;
1201 cv_w = pos_w;
1202 }
1203
1204 /* Cliptest, for hardwired planes */
1205 if (clip_xy) {
1206 /* plane 1 */
1207 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1208 temp = shift;
1209 test = LLVMBuildAnd(builder, test, temp, "");
1210 mask = test;
1211
1212 /* plane 2 */
1213 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1214 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1215 temp = LLVMBuildShl(builder, temp, shift, "");
1216 test = LLVMBuildAnd(builder, test, temp, "");
1217 mask = LLVMBuildOr(builder, mask, test, "");
1218
1219 /* plane 3 */
1220 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1221 temp = LLVMBuildShl(builder, temp, shift, "");
1222 test = LLVMBuildAnd(builder, test, temp, "");
1223 mask = LLVMBuildOr(builder, mask, test, "");
1224
1225 /* plane 4 */
1226 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1227 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1228 temp = LLVMBuildShl(builder, temp, shift, "");
1229 test = LLVMBuildAnd(builder, test, temp, "");
1230 mask = LLVMBuildOr(builder, mask, test, "");
1231 }
1232
1233 if (clip_z) {
1234 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1235 if (clip_halfz) {
1236 /* plane 5 */
1237 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1238 test = LLVMBuildAnd(builder, test, temp, "");
1239 mask = LLVMBuildOr(builder, mask, test, "");
1240 }
1241 else {
1242 /* plane 5 */
1243 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1244 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1245 test = LLVMBuildAnd(builder, test, temp, "");
1246 mask = LLVMBuildOr(builder, mask, test, "");
1247 }
1248 /* plane 6 */
1249 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1250 temp = LLVMBuildShl(builder, temp, shift, "");
1251 test = LLVMBuildAnd(builder, test, temp, "");
1252 mask = LLVMBuildOr(builder, mask, test, "");
1253 }
1254
1255 if (clip_user) {
1256 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1257 LLVMValueRef indices[3];
1258 LLVMValueRef is_nan_or_inf;
1259
1260 /* userclip planes */
1261 while (ucp_enable) {
1262 unsigned plane_idx = ffs(ucp_enable)-1;
1263 ucp_enable &= ~(1 << plane_idx);
1264 plane_idx += 6;
1265
1266 if (have_cd && num_written_clipdistance) {
1267 LLVMValueRef clipdist;
1268 int i;
1269 i = plane_idx - 6;
1270
1271 *have_clipdist = TRUE;
1272 if (i < 4) {
1273 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1274 } else {
1275 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1276 }
1277 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1278 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1279 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1280 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1281 test = LLVMBuildAnd(builder, test, temp, "");
1282 mask = LLVMBuildOr(builder, mask, test, "");
1283 } else {
1284 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1285 indices[0] = lp_build_const_int32(gallivm, 0);
1286 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1287
1288 indices[2] = lp_build_const_int32(gallivm, 0);
1289 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1290 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1291 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1292 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1293
1294 indices[2] = lp_build_const_int32(gallivm, 1);
1295 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1296 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1297 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1298 test = LLVMBuildFMul(builder, planes, cv_y, "");
1299 sum = LLVMBuildFAdd(builder, sum, test, "");
1300
1301 indices[2] = lp_build_const_int32(gallivm, 2);
1302 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1303 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1304 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1305 test = LLVMBuildFMul(builder, planes, cv_z, "");
1306 sum = LLVMBuildFAdd(builder, sum, test, "");
1307
1308 indices[2] = lp_build_const_int32(gallivm, 3);
1309 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1310 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1311 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1312 test = LLVMBuildFMul(builder, planes, cv_w, "");
1313 sum = LLVMBuildFAdd(builder, sum, test, "");
1314
1315 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1316 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1317 test = LLVMBuildAnd(builder, test, temp, "");
1318 mask = LLVMBuildOr(builder, mask, test, "");
1319 }
1320 }
1321 }
1322 return mask;
1323 }
1324
1325
1326 /**
1327 * Returns boolean if any clipping has occurred
1328 * Used zero/non-zero i32 value to represent boolean
1329 */
1330 static LLVMValueRef
1331 clipmask_booli32(struct gallivm_state *gallivm,
1332 const struct lp_type vs_type,
1333 LLVMValueRef clipmask_bool_ptr)
1334 {
1335 LLVMBuilderRef builder = gallivm->builder;
1336 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1337 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1338 LLVMValueRef ret = LLVMConstNull(int32_type);
1339 LLVMValueRef temp;
1340 int i;
1341
1342 /*
1343 * Can do this with log2(vector length) pack instructions and one extract
1344 * (as we don't actually need a or) with sse2 which would be way better.
1345 */
1346 for (i=0; i < vs_type.length; i++) {
1347 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1348 lp_build_const_int32(gallivm, i) , "");
1349 ret = LLVMBuildOr(builder, ret, temp, "");
1350 }
1351 return ret;
1352 }
1353
1354 static LLVMValueRef
1355 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1356 struct lp_build_tgsi_context * bld_base,
1357 boolean is_vindex_indirect,
1358 LLVMValueRef vertex_index,
1359 boolean is_aindex_indirect,
1360 LLVMValueRef attrib_index,
1361 LLVMValueRef swizzle_index)
1362 {
1363 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1364 struct gallivm_state *gallivm = bld_base->base.gallivm;
1365 LLVMBuilderRef builder = gallivm->builder;
1366 LLVMValueRef indices[3];
1367 LLVMValueRef res;
1368 struct lp_type type = bld_base->base.type;
1369
1370 if (is_vindex_indirect || is_aindex_indirect) {
1371 int i;
1372 res = bld_base->base.zero;
1373 for (i = 0; i < type.length; ++i) {
1374 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1375 LLVMValueRef vert_chan_index = vertex_index;
1376 LLVMValueRef attr_chan_index = attrib_index;
1377 LLVMValueRef channel_vec, value;
1378
1379 if (is_vindex_indirect) {
1380 vert_chan_index = LLVMBuildExtractElement(builder,
1381 vertex_index, idx, "");
1382 }
1383 if (is_aindex_indirect) {
1384 attr_chan_index = LLVMBuildExtractElement(builder,
1385 attrib_index, idx, "");
1386 }
1387
1388 indices[0] = vert_chan_index;
1389 indices[1] = attr_chan_index;
1390 indices[2] = swizzle_index;
1391
1392 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1393 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1394 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1395
1396 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1397 }
1398 } else {
1399 indices[0] = vertex_index;
1400 indices[1] = attrib_index;
1401 indices[2] = swizzle_index;
1402
1403 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1404 res = LLVMBuildLoad(builder, res, "");
1405 }
1406
1407 return res;
1408 }
1409
1410 static void
1411 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1412 struct lp_build_tgsi_context * bld_base,
1413 LLVMValueRef (*outputs)[4],
1414 LLVMValueRef emitted_vertices_vec)
1415 {
1416 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1417 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1418 struct gallivm_state *gallivm = variant->gallivm;
1419 LLVMBuilderRef builder = gallivm->builder;
1420 struct lp_type gs_type = bld_base->base.type;
1421 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1422 lp_int_type(gs_type), 0);
1423 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1424 LLVMValueRef next_prim_offset =
1425 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1426 LLVMValueRef io = variant->io_ptr;
1427 unsigned i;
1428 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1429
1430 for (i = 0; i < gs_type.length; ++i) {
1431 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1432 LLVMValueRef currently_emitted =
1433 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1434 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1435 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1436 }
1437
1438 convert_to_aos(gallivm, io, indices,
1439 outputs, clipmask,
1440 gs_info->num_outputs, gs_type,
1441 FALSE);
1442 }
1443
1444 static void
1445 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1446 struct lp_build_tgsi_context * bld_base,
1447 LLVMValueRef verts_per_prim_vec,
1448 LLVMValueRef emitted_prims_vec)
1449 {
1450 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1451 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1452 struct gallivm_state *gallivm = variant->gallivm;
1453 LLVMBuilderRef builder = gallivm->builder;
1454 LLVMValueRef prim_lengts_ptr =
1455 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1456 unsigned i;
1457
1458 for (i = 0; i < bld_base->base.type.length; ++i) {
1459 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1460 LLVMValueRef prims_emitted =
1461 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1462 LLVMValueRef store_ptr;
1463 LLVMValueRef num_vertices =
1464 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1465
1466 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1467 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1468 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1469 LLVMBuildStore(builder, num_vertices, store_ptr);
1470 }
1471 }
1472
1473 static void
1474 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1475 struct lp_build_tgsi_context * bld_base,
1476 LLVMValueRef total_emitted_vertices_vec,
1477 LLVMValueRef emitted_prims_vec)
1478 {
1479 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1480 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1481 struct gallivm_state *gallivm = variant->gallivm;
1482 LLVMBuilderRef builder = gallivm->builder;
1483 LLVMValueRef emitted_verts_ptr =
1484 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1485 LLVMValueRef emitted_prims_ptr =
1486 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1487 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1488
1489 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1490 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1491
1492 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1493 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1494 }
1495
1496 static void
1497 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1498 boolean elts)
1499 {
1500 struct gallivm_state *gallivm = variant->gallivm;
1501 LLVMContextRef context = gallivm->context;
1502 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1503 LLVMTypeRef arg_types[11];
1504 unsigned num_arg_types =
1505 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1506 LLVMTypeRef func_type;
1507 LLVMValueRef context_ptr;
1508 LLVMBasicBlockRef block;
1509 LLVMBuilderRef builder;
1510 char func_name[64];
1511 struct lp_type vs_type;
1512 LLVMValueRef end, start;
1513 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1514 LLVMValueRef vertex_id_offset, start_instance;
1515 LLVMValueRef stride, step, io_itr;
1516 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1517 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1518 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1519 struct draw_context *draw = llvm->draw;
1520 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1521 unsigned i, j;
1522 struct lp_build_context bld;
1523 struct lp_build_loop_state lp_loop;
1524 const int vector_length = lp_native_vector_width / 32;
1525 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1526 LLVMValueRef fetch_max;
1527 struct lp_build_sampler_soa *sampler = 0;
1528 LLVMValueRef ret, clipmask_bool_ptr;
1529 struct draw_llvm_variant_key *key = &variant->key;
1530 /* If geometry shader is present we need to skip both the viewport
1531 * transformation and clipping otherwise the inputs to the geometry
1532 * shader will be incorrect.
1533 * The code can't handle vp transform when vs writes vp index neither
1534 * (though this would be fixable here, but couldn't just broadcast
1535 * the values).
1536 */
1537 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1538 llvm->draw->vs.vertex_shader->info.writes_viewport_index;
1539 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1540 key->clip_z ||
1541 key->clip_user);
1542 LLVMValueRef variant_func;
1543 const unsigned pos = llvm->draw->vs.position_output;
1544 const unsigned cv = llvm->draw->vs.clipvertex_output;
1545 boolean have_clipdist = FALSE;
1546 struct lp_bld_tgsi_system_values system_values;
1547
1548 memset(&system_values, 0, sizeof(system_values));
1549
1550 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1551 variant->shader->variants_cached, elts ? "elts" : "linear");
1552
1553 i = 0;
1554 arg_types[i++] = get_context_ptr_type(variant); /* context */
1555 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1556 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1557 if (elts) {
1558 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1559 arg_types[i++] = int32_type; /* fetch_elt_max */
1560 } else
1561 arg_types[i++] = int32_type; /* start */
1562 arg_types[i++] = int32_type; /* fetch_count / count */
1563 arg_types[i++] = int32_type; /* stride */
1564 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1565 arg_types[i++] = int32_type; /* instance_id */
1566 arg_types[i++] = int32_type; /* vertex_id_offset */
1567 arg_types[i++] = int32_type; /* start_instance */
1568
1569 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1570
1571 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1572
1573 if (elts)
1574 variant->function_elts = variant_func;
1575 else
1576 variant->function = variant_func;
1577
1578 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1579 for (i = 0; i < num_arg_types; ++i)
1580 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1581 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1582 LLVMNoAliasAttribute);
1583
1584 context_ptr = LLVMGetParam(variant_func, 0);
1585 io_ptr = LLVMGetParam(variant_func, 1);
1586 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1587 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1588 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1589 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1590 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1591 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1592
1593 lp_build_name(context_ptr, "context");
1594 lp_build_name(io_ptr, "io");
1595 lp_build_name(vbuffers_ptr, "vbuffers");
1596 lp_build_name(stride, "stride");
1597 lp_build_name(vb_ptr, "vb");
1598 lp_build_name(system_values.instance_id, "instance_id");
1599 lp_build_name(vertex_id_offset, "vertex_id_offset");
1600 lp_build_name(start_instance, "start_instance");
1601
1602 if (elts) {
1603 fetch_elts = LLVMGetParam(variant_func, 3);
1604 fetch_elt_max = LLVMGetParam(variant_func, 4);
1605 fetch_count = LLVMGetParam(variant_func, 5);
1606 lp_build_name(fetch_elts, "fetch_elts");
1607 lp_build_name(fetch_elt_max, "fetch_elt_max");
1608 lp_build_name(fetch_count, "fetch_count");
1609 start = count = NULL;
1610 }
1611 else {
1612 start = LLVMGetParam(variant_func, 3);
1613 count = LLVMGetParam(variant_func, 4);
1614 lp_build_name(start, "start");
1615 lp_build_name(count, "count");
1616 fetch_elts = fetch_count = NULL;
1617 }
1618
1619 /*
1620 * Function body
1621 */
1622
1623 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1624 builder = gallivm->builder;
1625 LLVMPositionBuilderAtEnd(builder, block);
1626
1627 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1628
1629 memset(&vs_type, 0, sizeof vs_type);
1630 vs_type.floating = TRUE; /* floating point values */
1631 vs_type.sign = TRUE; /* values are signed */
1632 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1633 vs_type.width = 32; /* 32-bit float */
1634 vs_type.length = vector_length;
1635
1636 /* hold temporary "bool" clipmask */
1637 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1638 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1639
1640 /* code generated texture sampling */
1641 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1642
1643 if (elts) {
1644 start = zero;
1645 end = fetch_count;
1646 count = fetch_count;
1647 }
1648 else {
1649 end = lp_build_add(&bld, start, count);
1650 }
1651
1652 step = lp_build_const_int32(gallivm, vector_length);
1653
1654 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1655
1656 lp_build_loop_begin(&lp_loop, gallivm, zero);
1657 {
1658 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1659 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1660 LLVMValueRef io;
1661 LLVMValueRef clipmask; /* holds the clipmask value */
1662 LLVMValueRef true_index_array = lp_build_zero(gallivm,
1663 lp_type_uint_vec(32, 32*vector_length));
1664 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1665
1666 io_itr = lp_loop.counter;
1667
1668 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1669 #if DEBUG_STORE
1670 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1671 io_itr, io, lp_loop.counter);
1672 #endif
1673 for (i = 0; i < vector_length; ++i) {
1674 LLVMValueRef vert_index =
1675 LLVMBuildAdd(builder,
1676 lp_loop.counter,
1677 lp_build_const_int32(gallivm, i), "");
1678 LLVMValueRef true_index =
1679 LLVMBuildAdd(builder, start, vert_index, "");
1680
1681 /* make sure we're not out of bounds which can happen
1682 * if fetch_count % 4 != 0, because on the last iteration
1683 * a few of the 4 vertex fetches will be out of bounds */
1684 true_index = lp_build_min(&bld, true_index, fetch_max);
1685
1686 if (elts) {
1687 LLVMValueRef fetch_ptr;
1688 LLVMValueRef index_overflowed;
1689 LLVMValueRef index_ptr =
1690 lp_build_alloca(
1691 gallivm,
1692 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1693 struct lp_build_if_state if_ctx;
1694 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1695 true_index, fetch_elt_max,
1696 "index_overflowed");
1697
1698 lp_build_if(&if_ctx, gallivm, index_overflowed);
1699 {
1700 /* Generate maximum possible index so that
1701 * generate_fetch can treat it just like
1702 * any other overflow and return zeros.
1703 * We don't have to worry about the restart
1704 * primitive index because it has already been
1705 * handled
1706 */
1707 LLVMValueRef val =
1708 lp_build_const_int32(gallivm, 0xffffffff);
1709 LLVMBuildStore(builder, val, index_ptr);
1710 }
1711 lp_build_else(&if_ctx);
1712 {
1713 LLVMValueRef val;
1714 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1715 &true_index, 1, "");
1716 val = LLVMBuildLoad(builder, fetch_ptr, "");
1717 LLVMBuildStore(builder, val, index_ptr);
1718 }
1719 lp_build_endif(&if_ctx);
1720 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1721 }
1722 true_index_array = LLVMBuildInsertElement(
1723 gallivm->builder, true_index_array, true_index,
1724 lp_build_const_int32(gallivm, i), "");
1725
1726 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1727 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1728 LLVMValueRef vb_index =
1729 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1730 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1731 generate_fetch(gallivm, draw, vbuffers_ptr,
1732 &aos_attribs[j][i], velem, vb, true_index,
1733 system_values.instance_id, start_instance);
1734 }
1735 }
1736 convert_to_soa(gallivm, aos_attribs, inputs,
1737 draw->pt.nr_vertex_elements, vs_type);
1738
1739 /* In the paths with elts vertex id has to be unaffected by the
1740 * index bias and because indices inside our elements array have
1741 * already had index bias applied we need to subtract it here to
1742 * get back to the original index.
1743 * in the linear paths vertex id has to be unaffected by the
1744 * original start index and because we abuse the 'start' variable
1745 * to either represent the actual start index or the index at which
1746 * the primitive was split (we split rendering into chunks of at
1747 * most 4095-vertices) we need to back out the original start
1748 * index out of our vertex id here.
1749 */
1750 system_values.basevertex = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm,
1751 lp_type_uint_vec(32, 32*vector_length)),
1752 vertex_id_offset);
1753 system_values.vertex_id = true_index_array;
1754 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1755 system_values.basevertex, "");
1756
1757 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1758 generate_vs(variant,
1759 builder,
1760 vs_type,
1761 outputs,
1762 ptr_aos,
1763 &system_values,
1764 context_ptr,
1765 sampler,
1766 key->clamp_vertex_color);
1767
1768 if (pos != -1 && cv != -1) {
1769 /* store original positions in clip before further manipulation */
1770 store_clip(gallivm, vs_type, io, outputs, FALSE, key->clip_user ? cv : pos);
1771 store_clip(gallivm, vs_type, io, outputs, TRUE, pos);
1772
1773 /* do cliptest */
1774 if (enable_cliptest) {
1775 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1776 /* allocate clipmask, assign it integer type */
1777 clipmask = generate_clipmask(llvm,
1778 gallivm,
1779 vs_type,
1780 outputs,
1781 key->clip_xy,
1782 key->clip_z,
1783 key->clip_user,
1784 key->clip_halfz,
1785 key->ucp_enable,
1786 context_ptr, &have_clipdist);
1787 temp = LLVMBuildOr(builder, clipmask, temp, "");
1788 /* store temporary clipping boolean value */
1789 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1790 }
1791 else {
1792 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1793 }
1794
1795 /* do viewport mapping */
1796 if (!bypass_viewport) {
1797 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1798 }
1799 }
1800 else {
1801 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1802 }
1803
1804 /* store clipmask in vertex header,
1805 * original positions in clip
1806 * and transformed positions in data
1807 */
1808 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1809 vs_info->num_outputs, vs_type,
1810 have_clipdist);
1811 }
1812 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1813
1814 sampler->destroy(sampler);
1815
1816 /* return clipping boolean value for function */
1817 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1818
1819 LLVMBuildRet(builder, ret);
1820
1821 gallivm_verify_function(gallivm, variant_func);
1822 }
1823
1824
1825 struct draw_llvm_variant_key *
1826 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1827 {
1828 unsigned i;
1829 struct draw_llvm_variant_key *key;
1830 struct draw_sampler_static_state *draw_sampler;
1831
1832 key = (struct draw_llvm_variant_key *)store;
1833
1834 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1835
1836 /* Presumably all variants of the shader should have the same
1837 * number of vertex elements - ie the number of shader inputs.
1838 * NOTE: we NEED to store the needed number of needed inputs
1839 * here, not the number of provided elements to match keysize
1840 * (and the offset of sampler state in the key).
1841 */
1842 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1843 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1844
1845 /* will have to rig this up properly later */
1846 key->clip_xy = llvm->draw->clip_xy;
1847 key->clip_z = llvm->draw->clip_z;
1848 key->clip_user = llvm->draw->clip_user;
1849 key->bypass_viewport = llvm->draw->bypass_viewport;
1850 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1851 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1852 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1853 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1854 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1855 key->pad1 = 0;
1856
1857 /* All variants of this shader will have the same value for
1858 * nr_samplers. Not yet trying to compact away holes in the
1859 * sampler array.
1860 */
1861 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1862 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1863 key->nr_sampler_views =
1864 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1865 }
1866 else {
1867 key->nr_sampler_views = key->nr_samplers;
1868 }
1869
1870 draw_sampler = draw_llvm_variant_key_samplers(key);
1871
1872 memcpy(key->vertex_element,
1873 llvm->draw->pt.vertex_element,
1874 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1875
1876 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1877
1878 for (i = 0 ; i < key->nr_samplers; i++) {
1879 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1880 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1881 }
1882 for (i = 0 ; i < key->nr_sampler_views; i++) {
1883 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1884 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1885 }
1886
1887 return key;
1888 }
1889
1890
1891 void
1892 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1893 {
1894 unsigned i;
1895 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1896
1897 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1898 debug_printf("clip_xy = %u\n", key->clip_xy);
1899 debug_printf("clip_z = %u\n", key->clip_z);
1900 debug_printf("clip_user = %u\n", key->clip_user);
1901 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1902 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1903 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1904 debug_printf("has_gs = %u\n", key->has_gs);
1905 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1906
1907 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1908 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1909 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1910 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1911 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1912 }
1913
1914 for (i = 0 ; i < key->nr_sampler_views; i++) {
1915 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1916 }
1917 }
1918
1919
1920 void
1921 draw_llvm_set_mapped_texture(struct draw_context *draw,
1922 unsigned shader_stage,
1923 unsigned sview_idx,
1924 uint32_t width, uint32_t height, uint32_t depth,
1925 uint32_t first_level, uint32_t last_level,
1926 const void *base_ptr,
1927 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1928 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1929 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1930 {
1931 unsigned j;
1932 struct draw_jit_texture *jit_tex;
1933
1934 assert(shader_stage == PIPE_SHADER_VERTEX ||
1935 shader_stage == PIPE_SHADER_GEOMETRY);
1936
1937 if (shader_stage == PIPE_SHADER_VERTEX) {
1938 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1939
1940 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1941 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1942 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1943
1944 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1945 } else {
1946 assert(0);
1947 return;
1948 }
1949
1950 jit_tex->width = width;
1951 jit_tex->height = height;
1952 jit_tex->depth = depth;
1953 jit_tex->first_level = first_level;
1954 jit_tex->last_level = last_level;
1955 jit_tex->base = base_ptr;
1956
1957 for (j = first_level; j <= last_level; j++) {
1958 jit_tex->mip_offsets[j] = mip_offsets[j];
1959 jit_tex->row_stride[j] = row_stride[j];
1960 jit_tex->img_stride[j] = img_stride[j];
1961 }
1962 }
1963
1964
1965 void
1966 draw_llvm_set_sampler_state(struct draw_context *draw,
1967 unsigned shader_type)
1968 {
1969 unsigned i;
1970
1971 if (shader_type == PIPE_SHADER_VERTEX) {
1972 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1973 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1974
1975 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
1976 const struct pipe_sampler_state *s
1977 = draw->samplers[PIPE_SHADER_VERTEX][i];
1978 jit_sam->min_lod = s->min_lod;
1979 jit_sam->max_lod = s->max_lod;
1980 jit_sam->lod_bias = s->lod_bias;
1981 COPY_4V(jit_sam->border_color, s->border_color.f);
1982 }
1983 }
1984 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1985 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1986 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1987
1988 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
1989 const struct pipe_sampler_state *s
1990 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1991 jit_sam->min_lod = s->min_lod;
1992 jit_sam->max_lod = s->max_lod;
1993 jit_sam->lod_bias = s->lod_bias;
1994 COPY_4V(jit_sam->border_color, s->border_color.f);
1995 }
1996 }
1997 }
1998 }
1999
2000
2001 void
2002 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2003 {
2004 struct draw_llvm *llvm = variant->llvm;
2005
2006 gallivm_destroy(variant->gallivm);
2007
2008 remove_from_list(&variant->list_item_local);
2009 variant->shader->variants_cached--;
2010 remove_from_list(&variant->list_item_global);
2011 llvm->nr_variants--;
2012 FREE(variant);
2013 }
2014
2015
2016 /**
2017 * Create LLVM types for various structures.
2018 */
2019 static void
2020 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2021 {
2022 struct gallivm_state *gallivm = var->gallivm;
2023 LLVMTypeRef texture_type, sampler_type, context_type;
2024
2025 texture_type = create_jit_texture_type(gallivm, "texture");
2026 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2027
2028 context_type = create_gs_jit_context_type(gallivm,
2029 var->shader->base.vector_length,
2030 texture_type, sampler_type,
2031 "draw_gs_jit_context");
2032 var->context_ptr_type = LLVMPointerType(context_type, 0);
2033
2034 var->input_array_type = create_gs_jit_input_type(gallivm);
2035 }
2036
2037 static LLVMTypeRef
2038 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2039 {
2040 if (!variant->context_ptr_type)
2041 create_gs_jit_types(variant);
2042 return variant->context_ptr_type;
2043 }
2044
2045 static LLVMValueRef
2046 generate_mask_value(struct draw_gs_llvm_variant *variant,
2047 struct lp_type gs_type)
2048 {
2049 struct gallivm_state *gallivm = variant->gallivm;
2050 LLVMBuilderRef builder = gallivm->builder;
2051 struct lp_type mask_type = lp_int_type(gs_type);
2052 LLVMValueRef num_prims;
2053 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2054 unsigned i;
2055
2056 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2057 variant->num_prims);
2058 for (i = 0; i < gs_type.length; i++) {
2059 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2060 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2061 }
2062 mask_val = lp_build_compare(gallivm, mask_type,
2063 PIPE_FUNC_GREATER, num_prims, mask_val);
2064
2065 return mask_val;
2066 }
2067
2068 static void
2069 draw_gs_llvm_generate(struct draw_llvm *llvm,
2070 struct draw_gs_llvm_variant *variant)
2071 {
2072 struct gallivm_state *gallivm = variant->gallivm;
2073 LLVMContextRef context = gallivm->context;
2074 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2075 LLVMTypeRef arg_types[7];
2076 LLVMTypeRef func_type;
2077 LLVMValueRef variant_func;
2078 LLVMValueRef context_ptr;
2079 LLVMValueRef prim_id_ptr;
2080 LLVMBasicBlockRef block;
2081 LLVMBuilderRef builder;
2082 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2083 struct lp_build_sampler_soa *sampler = 0;
2084 struct lp_build_context bld;
2085 struct lp_bld_tgsi_system_values system_values;
2086 char func_name[64];
2087 struct lp_type gs_type;
2088 unsigned i;
2089 struct draw_gs_llvm_iface gs_iface;
2090 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2091 LLVMValueRef consts_ptr, num_consts_ptr;
2092 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2093 struct lp_build_mask_context mask;
2094 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2095 unsigned vector_length = variant->shader->base.vector_length;
2096
2097 memset(&system_values, 0, sizeof(system_values));
2098
2099 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2100 variant->shader->variants_cached);
2101
2102 assert(variant->vertex_header_ptr_type);
2103
2104 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2105 arg_types[1] = variant->input_array_type; /* input */
2106 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2107 arg_types[3] = int32_type; /* num_prims */
2108 arg_types[4] = int32_type; /* instance_id */
2109 arg_types[5] = LLVMPointerType(
2110 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2111 arg_types[6] = int32_type;
2112
2113 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2114
2115 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2116
2117 variant->function = variant_func;
2118
2119 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2120
2121 for (i = 0; i < Elements(arg_types); ++i)
2122 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2123 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2124 LLVMNoAliasAttribute);
2125
2126 context_ptr = LLVMGetParam(variant_func, 0);
2127 input_array = LLVMGetParam(variant_func, 1);
2128 io_ptr = LLVMGetParam(variant_func, 2);
2129 num_prims = LLVMGetParam(variant_func, 3);
2130 system_values.instance_id = LLVMGetParam(variant_func, 4);
2131 prim_id_ptr = LLVMGetParam(variant_func, 5);
2132 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2133
2134 lp_build_name(context_ptr, "context");
2135 lp_build_name(input_array, "input");
2136 lp_build_name(io_ptr, "io");
2137 lp_build_name(num_prims, "num_prims");
2138 lp_build_name(system_values.instance_id, "instance_id");
2139 lp_build_name(prim_id_ptr, "prim_id_ptr");
2140 lp_build_name(system_values.invocation_id, "invocation_id");
2141
2142 variant->context_ptr = context_ptr;
2143 variant->io_ptr = io_ptr;
2144 variant->num_prims = num_prims;
2145
2146 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2147 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2148 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2149 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2150 gs_iface.input = input_array;
2151 gs_iface.variant = variant;
2152
2153 /*
2154 * Function body
2155 */
2156
2157 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2158 builder = gallivm->builder;
2159 LLVMPositionBuilderAtEnd(builder, block);
2160
2161 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2162
2163 memset(&gs_type, 0, sizeof gs_type);
2164 gs_type.floating = TRUE; /* floating point values */
2165 gs_type.sign = TRUE; /* values are signed */
2166 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2167 gs_type.width = 32; /* 32-bit float */
2168 gs_type.length = vector_length;
2169
2170 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2171 num_consts_ptr =
2172 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2173
2174 /* code generated texture sampling */
2175 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2176
2177 mask_val = generate_mask_value(variant, gs_type);
2178 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2179
2180 if (gs_info->uses_primid) {
2181 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2182 }
2183
2184 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2185 tgsi_dump(tokens, 0);
2186 draw_gs_llvm_dump_variant_key(&variant->key);
2187 }
2188
2189 lp_build_tgsi_soa(variant->gallivm,
2190 tokens,
2191 gs_type,
2192 &mask,
2193 consts_ptr,
2194 num_consts_ptr,
2195 &system_values,
2196 NULL,
2197 outputs,
2198 context_ptr,
2199 NULL,
2200 sampler,
2201 &llvm->draw->gs.geometry_shader->info,
2202 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2203
2204 sampler->destroy(sampler);
2205
2206 lp_build_mask_end(&mask);
2207
2208 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2209
2210 gallivm_verify_function(gallivm, variant_func);
2211 }
2212
2213
2214 struct draw_gs_llvm_variant *
2215 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2216 unsigned num_outputs,
2217 const struct draw_gs_llvm_variant_key *key)
2218 {
2219 struct draw_gs_llvm_variant *variant;
2220 struct llvm_geometry_shader *shader =
2221 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2222 LLVMTypeRef vertex_header;
2223 char module_name[64];
2224
2225 variant = MALLOC(sizeof *variant +
2226 shader->variant_key_size -
2227 sizeof variant->key);
2228 if (!variant)
2229 return NULL;
2230
2231 variant->llvm = llvm;
2232 variant->shader = shader;
2233
2234 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2235 variant->shader->variants_cached);
2236
2237 variant->gallivm = gallivm_create(module_name, llvm->context);
2238
2239 create_gs_jit_types(variant);
2240
2241 memcpy(&variant->key, key, shader->variant_key_size);
2242
2243 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2244
2245 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2246
2247 draw_gs_llvm_generate(llvm, variant);
2248
2249 gallivm_compile_module(variant->gallivm);
2250
2251 variant->jit_func = (draw_gs_jit_func)
2252 gallivm_jit_function(variant->gallivm, variant->function);
2253
2254 gallivm_free_ir(variant->gallivm);
2255
2256 variant->list_item_global.base = variant;
2257 variant->list_item_local.base = variant;
2258 /*variant->no = */shader->variants_created++;
2259 variant->list_item_global.base = variant;
2260
2261 return variant;
2262 }
2263
2264 void
2265 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2266 {
2267 struct draw_llvm *llvm = variant->llvm;
2268
2269 gallivm_destroy(variant->gallivm);
2270
2271 remove_from_list(&variant->list_item_local);
2272 variant->shader->variants_cached--;
2273 remove_from_list(&variant->list_item_global);
2274 llvm->nr_gs_variants--;
2275 FREE(variant);
2276 }
2277
2278 struct draw_gs_llvm_variant_key *
2279 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2280 {
2281 unsigned i;
2282 struct draw_gs_llvm_variant_key *key;
2283 struct draw_sampler_static_state *draw_sampler;
2284
2285 key = (struct draw_gs_llvm_variant_key *)store;
2286
2287 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2288
2289 /* All variants of this shader will have the same value for
2290 * nr_samplers. Not yet trying to compact away holes in the
2291 * sampler array.
2292 */
2293 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2294 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2295 key->nr_sampler_views =
2296 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2297 }
2298 else {
2299 key->nr_sampler_views = key->nr_samplers;
2300 }
2301
2302 draw_sampler = key->samplers;
2303
2304 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2305
2306 for (i = 0 ; i < key->nr_samplers; i++) {
2307 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2308 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2309 }
2310 for (i = 0 ; i < key->nr_sampler_views; i++) {
2311 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2312 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2313 }
2314
2315 return key;
2316 }
2317
2318 void
2319 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2320 {
2321 unsigned i;
2322 struct draw_sampler_static_state *sampler = key->samplers;
2323
2324 for (i = 0 ; i < key->nr_sampler_views; i++) {
2325 debug_printf("sampler[%i].src_format = %s\n", i,
2326 util_format_name(sampler[i].texture_state.format));
2327 }
2328 }