projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Revert "draw: use vectorized calculations for fetch"
[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 const struct util_format_description *format_desc,
662 LLVMValueRef vb_stride,
663 LLVMValueRef stride_fixed,
664 LLVMValueRef map_ptr,
665 LLVMValueRef buffer_size_adj,
666 LLVMValueRef ofbit,
667 LLVMValueRef *res,
668 LLVMValueRef index)
669 {
670 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
671 LLVMBuilderRef builder = gallivm->builder;
672 LLVMValueRef stride;
673 LLVMValueRef buffer_overflowed;
674 LLVMValueRef temp_ptr =
675 lp_build_alloca(gallivm,
676 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
677 struct lp_build_if_state if_ctx;
678
679 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
680 stride = lp_build_uadd_overflow(gallivm, stride, stride_fixed, &ofbit);
681
682 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
683 stride, buffer_size_adj,
684 "buffer_overflowed");
685 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
686
687 if (0) {
688 lp_build_printf(gallivm, " stride = %u\n", stride);
689 lp_build_printf(gallivm, " buffer size adj = %u\n", buffer_size_adj);
690 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
691 }
692
693 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
694 {
695 LLVMValueRef val =
696 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
697 LLVMBuildStore(builder, val, temp_ptr);
698 }
699 lp_build_else(&if_ctx);
700 {
701 LLVMValueRef val;
702 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
703
704 val = lp_build_fetch_rgba_aos(gallivm,
705 format_desc,
706 lp_float32_vec4_type(),
707 FALSE,
708 map_ptr,
709 zero, zero, zero,
710 NULL);
711 LLVMBuildStore(builder, val, temp_ptr);
712 }
713 lp_build_endif(&if_ctx);
714
715 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
716 }
717
718 static void
719 convert_to_soa(struct gallivm_state *gallivm,
720 LLVMValueRef src_aos[LP_MAX_VECTOR_WIDTH / 32],
721 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
722 unsigned attrib, const struct lp_type soa_type)
723 {
724 unsigned j, k;
725 struct lp_type aos_channel_type = soa_type;
726
727 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
728 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
729
730 debug_assert(TGSI_NUM_CHANNELS == 4);
731 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
732
733 aos_channel_type.length >>= 1;
734
735 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
736 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
737
738 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
739
740 for (k = 0; k < pixels_per_channel; ++k) {
741 channel[k] = src_aos[j + TGSI_NUM_CHANNELS * k];
742 }
743
744 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
745 }
746
747 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[attrib]);
748 }
749
750
751 static void
752 store_aos(struct gallivm_state *gallivm,
753 LLVMValueRef io_ptr,
754 LLVMValueRef index,
755 LLVMValueRef value)
756 {
757 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
758 LLVMBuilderRef builder = gallivm->builder;
759 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
760 LLVMValueRef indices[3];
761
762 indices[0] = lp_build_const_int32(gallivm, 0);
763 indices[1] = index;
764 indices[2] = lp_build_const_int32(gallivm, 0);
765
766 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
767 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
768
769 #if DEBUG_STORE
770 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
771 #endif
772
773 /* Unaligned store due to the vertex header */
774 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
775 }
776
777 /**
778 * Adjust the mask to architecture endianess. The mask will the store in struct:
779 *
780 * struct vertex_header {
781 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
782 * unsigned edgeflag:1;
783 * unsigned pad:1;
784 * unsigned vertex_id:16;
785 * [...]
786 * }
787 *
788 * On little-endian machine nothing needs to done, however on bit-endian machine
789 * the mask's fields need to be adjusted with the algorithm:
790 *
791 * uint32_t reverse (uint32_t x)
792 * {
793 * return (x >> 16) | // vertex_id
794 * ((x & 0x3fff) << 18) | // clipmask
795 * ((x & 0x4000) << 3) | // pad
796 * ((x & 0x8000) << 1); // edgeflag
797 * }
798 */
799 static LLVMValueRef
800 adjust_mask(struct gallivm_state *gallivm,
801 LLVMValueRef mask)
802 {
803 #ifdef PIPE_ARCH_BIG_ENDIAN
804 LLVMBuilderRef builder = gallivm->builder;
805 LLVMValueRef vertex_id;
806 LLVMValueRef clipmask;
807 LLVMValueRef pad;
808 LLVMValueRef edgeflag;
809
810 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
811 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
812 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
813 if (0) {
814 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
815 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
816 }
817 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
818 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
819
820 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
821 if (0) {
822 mask = LLVMBuildOr(builder, mask, pad, "");
823 }
824 mask = LLVMBuildOr(builder, mask, edgeflag, "");
825 #endif
826 return mask;
827 }
828
829 static void
830 store_aos_array(struct gallivm_state *gallivm,
831 struct lp_type soa_type,
832 LLVMValueRef io_ptr,
833 LLVMValueRef *indices,
834 LLVMValueRef* aos,
835 int attrib,
836 int num_outputs,
837 LLVMValueRef clipmask,
838 boolean need_edgeflag)
839 {
840 LLVMBuilderRef builder = gallivm->builder;
841 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
842 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
843 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
844 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
845 int vector_length = soa_type.length;
846 int i;
847
848 debug_assert(TGSI_NUM_CHANNELS == 4);
849
850 for (i = 0; i < vector_length; i++) {
851 linear_inds[i] = lp_build_const_int32(gallivm, i);
852 if (indices) {
853 inds[i] = indices[i];
854 } else {
855 inds[i] = linear_inds[i];
856 }
857 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
858 }
859
860 if (attrib == 0) {
861 /* store vertex header for each of the n vertices */
862 LLVMValueRef val, cliptmp;
863 int vertex_id_pad_edgeflag;
864
865 /* If this assertion fails, it means we need to update the bit twidding
866 * code here. See struct vertex_header in draw_private.h.
867 */
868 assert(DRAW_TOTAL_CLIP_PLANES==14);
869 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
870 if (!need_edgeflag) {
871 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
872 }
873 else {
874 vertex_id_pad_edgeflag = (0xffff << 16);
875 }
876 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
877 vertex_id_pad_edgeflag);
878 /* OR with the clipmask */
879 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
880 for (i = 0; i < vector_length; i++) {
881 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
882 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
883 val = adjust_mask(gallivm, val);
884 #if DEBUG_STORE
885 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
886 io_ptrs[i], inds[i], val);
887 #endif
888 LLVMBuildStore(builder, val, id_ptr);
889 }
890 }
891
892 /* store for each of the n vertices */
893 for (i = 0; i < vector_length; i++) {
894 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
895 }
896 }
897
898
899 static void
900 convert_to_aos(struct gallivm_state *gallivm,
901 LLVMValueRef io,
902 LLVMValueRef *indices,
903 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
904 LLVMValueRef clipmask,
905 int num_outputs,
906 struct lp_type soa_type,
907 boolean need_edgeflag)
908 {
909 LLVMBuilderRef builder = gallivm->builder;
910 unsigned chan, attrib, i;
911
912 #if DEBUG_STORE
913 lp_build_printf(gallivm, " # storing begin\n");
914 #endif
915 for (attrib = 0; attrib < num_outputs; ++attrib) {
916 LLVMValueRef soa[TGSI_NUM_CHANNELS];
917 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
918 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
919 if (outputs[attrib][chan]) {
920 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
921 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
922 #if DEBUG_STORE
923 lp_build_printf(gallivm, "output %d : %d ",
924 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
925 attrib, 0),
926 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
927 chan, 0));
928 lp_build_print_value(gallivm, "val = ", out);
929 {
930 LLVMValueRef iv =
931 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
932
933 lp_build_print_value(gallivm, " ival = ", iv);
934 }
935 #endif
936 soa[chan] = out;
937 }
938 else {
939 soa[chan] = 0;
940 }
941 }
942
943
944 if (soa_type.length == TGSI_NUM_CHANNELS) {
945 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
946 } else {
947 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
948
949 for (i = 0; i < soa_type.length; ++i) {
950 aos[i] = lp_build_extract_range(gallivm,
951 soa[i % TGSI_NUM_CHANNELS],
952 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
953 TGSI_NUM_CHANNELS);
954 }
955 }
956
957 store_aos_array(gallivm,
958 soa_type,
959 io, indices,
960 aos,
961 attrib,
962 num_outputs,
963 clipmask,
964 need_edgeflag);
965 }
966 #if DEBUG_STORE
967 lp_build_printf(gallivm, " # storing end\n");
968 #endif
969 }
970
971
972 /**
973 * Stores original vertex positions in clip coordinates
974 */
975 static void
976 store_clip(struct gallivm_state *gallivm,
977 const struct lp_type vs_type,
978 LLVMValueRef io_ptr,
979 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
980 int idx)
981 {
982 LLVMBuilderRef builder = gallivm->builder;
983 LLVMValueRef soa[4];
984 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
985 LLVMValueRef indices[2];
986 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
987 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
988 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
989 LLVMTypeRef clip_ptr_type =
990 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
991 4), 0);
992 int i, j;
993
994 indices[0] =
995 indices[1] = lp_build_const_int32(gallivm, 0);
996
997 for (i = 0; i < vs_type.length; i++) {
998 inds[i] = lp_build_const_int32(gallivm, i);
999 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1000 }
1001
1002 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1003 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1004 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1005 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1006
1007 for (i = 0; i < vs_type.length; i++) {
1008 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1009 }
1010
1011 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1012 for (i = 0; i < vs_type.length; ++i) {
1013 aos[i] = lp_build_extract_range(gallivm,
1014 soa[i % TGSI_NUM_CHANNELS],
1015 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1016 TGSI_NUM_CHANNELS);
1017 }
1018
1019 for (j = 0; j < vs_type.length; j++) {
1020 LLVMValueRef clip_ptr;
1021
1022 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1023 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1024
1025 /* Unaligned store */
1026 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1027 }
1028 }
1029
1030
1031 /**
1032 * Transforms the outputs for viewport mapping
1033 */
1034 static void
1035 generate_viewport(struct draw_llvm_variant *variant,
1036 LLVMBuilderRef builder,
1037 struct lp_type vs_type,
1038 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1039 LLVMValueRef context_ptr)
1040 {
1041 int i;
1042 struct gallivm_state *gallivm = variant->gallivm;
1043 struct lp_type f32_type = vs_type;
1044 const unsigned pos = variant->llvm->draw->vs.position_output;
1045 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1046 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1047 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1048 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1049
1050 /* We treat pipe_viewport_state as a float array */
1051 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1052 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1053
1054 /* for 1/w convention*/
1055 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1056 LLVMBuildStore(builder, out3, outputs[pos][3]);
1057
1058 /* Viewport Mapping */
1059 for (i=0; i<3; i++) {
1060 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1061 LLVMValueRef scale;
1062 LLVMValueRef trans;
1063 LLVMValueRef scale_i;
1064 LLVMValueRef trans_i;
1065 LLVMValueRef index;
1066
1067 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1068 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1069
1070 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1071 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1072
1073 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1074 LLVMBuildLoad(builder, scale_i, "scale"));
1075 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1076 LLVMBuildLoad(builder, trans_i, "trans"));
1077
1078 /* divide by w */
1079 out = LLVMBuildFMul(builder, out, out3, "");
1080 /* mult by scale, add translation */
1081 out = lp_build_fmuladd(builder, out, scale, trans);
1082
1083 /* store transformed outputs */
1084 LLVMBuildStore(builder, out, outputs[pos][i]);
1085 }
1086
1087 }
1088
1089
1090 /**
1091 * Returns clipmask as nxi32 bitmask for the n vertices
1092 */
1093 static LLVMValueRef
1094 generate_clipmask(struct draw_llvm *llvm,
1095 struct gallivm_state *gallivm,
1096 struct lp_type vs_type,
1097 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1098 struct draw_llvm_variant_key *key,
1099 LLVMValueRef context_ptr,
1100 boolean *have_clipdist)
1101 {
1102 LLVMBuilderRef builder = gallivm->builder;
1103 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1104 LLVMValueRef test, temp;
1105 LLVMValueRef zero, shift;
1106 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1107 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1108 LLVMValueRef plane1, planes, plane_ptr, sum;
1109 struct lp_type f32_type = vs_type;
1110 struct lp_type i32_type = lp_int_type(vs_type);
1111 const unsigned pos = llvm->draw->vs.position_output;
1112 const unsigned cv = llvm->draw->vs.clipvertex_output;
1113 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1114 boolean have_cd = false;
1115 boolean clip_user = key->clip_user;
1116 unsigned ucp_enable = key->ucp_enable;
1117 unsigned cd[2];
1118
1119 cd[0] = llvm->draw->vs.ccdistance_output[0];
1120 cd[1] = llvm->draw->vs.ccdistance_output[1];
1121
1122 if (cd[0] != pos || cd[1] != pos)
1123 have_cd = true;
1124
1125 if (num_written_clipdistance && !clip_user) {
1126 clip_user = true;
1127 ucp_enable = (1 << num_written_clipdistance) - 1;
1128 }
1129
1130 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1131 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1132 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1133 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1134
1135 /*
1136 * load clipvertex and position from correct locations.
1137 * if they are the same just load them once.
1138 */
1139 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1140 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1141 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1142 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1143
1144 if (clip_user && cv != pos) {
1145 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1146 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1147 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1148 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1149 } else {
1150 cv_x = pos_x;
1151 cv_y = pos_y;
1152 cv_z = pos_z;
1153 cv_w = pos_w;
1154 }
1155
1156 /*
1157 * Be careful with the comparisons and NaNs (using llvm's unordered
1158 * comparisons here).
1159 */
1160 /* Cliptest, for hardwired planes */
1161 /*
1162 * XXX should take guardband into account (currently not in key).
1163 * Otherwise might run the draw pipeline stages for nothing.
1164 */
1165 if (key->clip_xy) {
1166 /* plane 1 */
1167 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1168 temp = shift;
1169 test = LLVMBuildAnd(builder, test, temp, "");
1170 mask = test;
1171
1172 /* plane 2 */
1173 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1174 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1175 temp = LLVMBuildShl(builder, temp, shift, "");
1176 test = LLVMBuildAnd(builder, test, temp, "");
1177 mask = LLVMBuildOr(builder, mask, test, "");
1178
1179 /* plane 3 */
1180 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1181 temp = LLVMBuildShl(builder, temp, shift, "");
1182 test = LLVMBuildAnd(builder, test, temp, "");
1183 mask = LLVMBuildOr(builder, mask, test, "");
1184
1185 /* plane 4 */
1186 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1187 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1188 temp = LLVMBuildShl(builder, temp, shift, "");
1189 test = LLVMBuildAnd(builder, test, temp, "");
1190 mask = LLVMBuildOr(builder, mask, test, "");
1191 }
1192
1193 if (key->clip_z) {
1194 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1195 if (key->clip_halfz) {
1196 /* plane 5 */
1197 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1198 test = LLVMBuildAnd(builder, test, temp, "");
1199 mask = LLVMBuildOr(builder, mask, test, "");
1200 }
1201 else {
1202 /* plane 5 */
1203 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1204 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1205 test = LLVMBuildAnd(builder, test, temp, "");
1206 mask = LLVMBuildOr(builder, mask, test, "");
1207 }
1208 /* plane 6 */
1209 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1210 temp = LLVMBuildShl(builder, temp, shift, "");
1211 test = LLVMBuildAnd(builder, test, temp, "");
1212 mask = LLVMBuildOr(builder, mask, test, "");
1213 }
1214
1215 if (clip_user) {
1216 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1217 LLVMValueRef indices[3];
1218 LLVMValueRef is_nan_or_inf;
1219
1220 /* userclip planes */
1221 while (ucp_enable) {
1222 unsigned plane_idx = ffs(ucp_enable)-1;
1223 ucp_enable &= ~(1 << plane_idx);
1224 plane_idx += 6;
1225
1226 if (have_cd && num_written_clipdistance) {
1227 LLVMValueRef clipdist;
1228 int i;
1229 i = plane_idx - 6;
1230
1231 *have_clipdist = TRUE;
1232 if (i < 4) {
1233 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1234 } else {
1235 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1236 }
1237 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1238 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1239 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1240 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1241 test = LLVMBuildAnd(builder, test, temp, "");
1242 mask = LLVMBuildOr(builder, mask, test, "");
1243 } else {
1244 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1245 indices[0] = lp_build_const_int32(gallivm, 0);
1246 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1247
1248 indices[2] = lp_build_const_int32(gallivm, 0);
1249 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1250 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1251 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1252 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1253
1254 indices[2] = lp_build_const_int32(gallivm, 1);
1255 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1256 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1257 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1258 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1259
1260 indices[2] = lp_build_const_int32(gallivm, 2);
1261 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1262 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1263 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1264 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1265
1266 indices[2] = lp_build_const_int32(gallivm, 3);
1267 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1268 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1269 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1270 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1271
1272 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1273 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1274 test = LLVMBuildAnd(builder, test, temp, "");
1275 mask = LLVMBuildOr(builder, mask, test, "");
1276 }
1277 }
1278 }
1279 if (key->need_edgeflags) {
1280 /*
1281 * This isn't really part of clipmask but stored the same in vertex
1282 * header later, so do it here.
1283 */
1284 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1285 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1286 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1287 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1288 temp = lp_build_const_int_vec(gallivm, i32_type,
1289 1LL << DRAW_TOTAL_CLIP_PLANES);
1290 test = LLVMBuildAnd(builder, test, temp, "");
1291 mask = LLVMBuildOr(builder, mask, test, "");
1292 }
1293 return mask;
1294 }
1295
1296
1297 /**
1298 * Returns boolean if any clipping has occurred
1299 * Used zero/non-zero i32 value to represent boolean
1300 */
1301 static LLVMValueRef
1302 clipmask_booli32(struct gallivm_state *gallivm,
1303 const struct lp_type vs_type,
1304 LLVMValueRef clipmask_bool_ptr,
1305 boolean edgeflag_in_clipmask)
1306 {
1307 LLVMBuilderRef builder = gallivm->builder;
1308 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1309 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1310 LLVMValueRef ret = LLVMConstNull(int32_type);
1311 LLVMValueRef temp;
1312 int i;
1313
1314 /*
1315 * We need to invert the edgeflag bit from the clipmask here
1316 * (because the result is really if we want to run the pipeline or not
1317 * and we (may) need it if edgeflag was 0).
1318 */
1319 if (edgeflag_in_clipmask) {
1320 struct lp_type i32_type = lp_int_type(vs_type);
1321 LLVMValueRef edge = lp_build_const_int_vec(gallivm, i32_type,
1322 1LL << DRAW_TOTAL_CLIP_PLANES);
1323 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1324 }
1325 /*
1326 * Could do much better with just cmp/movmskps.
1327 */
1328 for (i=0; i < vs_type.length; i++) {
1329 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1330 lp_build_const_int32(gallivm, i) , "");
1331 ret = LLVMBuildOr(builder, ret, temp, "");
1332 }
1333 return ret;
1334 }
1335
1336 static LLVMValueRef
1337 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1338 struct lp_build_tgsi_context * bld_base,
1339 boolean is_vindex_indirect,
1340 LLVMValueRef vertex_index,
1341 boolean is_aindex_indirect,
1342 LLVMValueRef attrib_index,
1343 LLVMValueRef swizzle_index)
1344 {
1345 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1346 struct gallivm_state *gallivm = bld_base->base.gallivm;
1347 LLVMBuilderRef builder = gallivm->builder;
1348 LLVMValueRef indices[3];
1349 LLVMValueRef res;
1350 struct lp_type type = bld_base->base.type;
1351
1352 if (is_vindex_indirect || is_aindex_indirect) {
1353 int i;
1354 res = bld_base->base.zero;
1355 for (i = 0; i < type.length; ++i) {
1356 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1357 LLVMValueRef vert_chan_index = vertex_index;
1358 LLVMValueRef attr_chan_index = attrib_index;
1359 LLVMValueRef channel_vec, value;
1360
1361 if (is_vindex_indirect) {
1362 vert_chan_index = LLVMBuildExtractElement(builder,
1363 vertex_index, idx, "");
1364 }
1365 if (is_aindex_indirect) {
1366 attr_chan_index = LLVMBuildExtractElement(builder,
1367 attrib_index, idx, "");
1368 }
1369
1370 indices[0] = vert_chan_index;
1371 indices[1] = attr_chan_index;
1372 indices[2] = swizzle_index;
1373
1374 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1375 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1376 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1377
1378 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1379 }
1380 } else {
1381 indices[0] = vertex_index;
1382 indices[1] = attrib_index;
1383 indices[2] = swizzle_index;
1384
1385 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1386 res = LLVMBuildLoad(builder, res, "");
1387 }
1388
1389 return res;
1390 }
1391
1392 static void
1393 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1394 struct lp_build_tgsi_context * bld_base,
1395 LLVMValueRef (*outputs)[4],
1396 LLVMValueRef emitted_vertices_vec)
1397 {
1398 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1399 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1400 struct gallivm_state *gallivm = variant->gallivm;
1401 LLVMBuilderRef builder = gallivm->builder;
1402 struct lp_type gs_type = bld_base->base.type;
1403 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1404 lp_int_type(gs_type), 0);
1405 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1406 LLVMValueRef next_prim_offset =
1407 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1408 LLVMValueRef io = variant->io_ptr;
1409 unsigned i;
1410 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1411
1412 for (i = 0; i < gs_type.length; ++i) {
1413 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1414 LLVMValueRef currently_emitted =
1415 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1416 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1417 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1418 }
1419
1420 convert_to_aos(gallivm, io, indices,
1421 outputs, clipmask,
1422 gs_info->num_outputs, gs_type,
1423 FALSE);
1424 }
1425
1426 static void
1427 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1428 struct lp_build_tgsi_context * bld_base,
1429 LLVMValueRef verts_per_prim_vec,
1430 LLVMValueRef emitted_prims_vec)
1431 {
1432 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1433 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1434 struct gallivm_state *gallivm = variant->gallivm;
1435 LLVMBuilderRef builder = gallivm->builder;
1436 LLVMValueRef prim_lengts_ptr =
1437 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1438 unsigned i;
1439
1440 for (i = 0; i < bld_base->base.type.length; ++i) {
1441 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1442 LLVMValueRef prims_emitted =
1443 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1444 LLVMValueRef store_ptr;
1445 LLVMValueRef num_vertices =
1446 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1447
1448 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1449 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1450 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1451 LLVMBuildStore(builder, num_vertices, store_ptr);
1452 }
1453 }
1454
1455 static void
1456 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1457 struct lp_build_tgsi_context * bld_base,
1458 LLVMValueRef total_emitted_vertices_vec,
1459 LLVMValueRef emitted_prims_vec)
1460 {
1461 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1462 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1463 struct gallivm_state *gallivm = variant->gallivm;
1464 LLVMBuilderRef builder = gallivm->builder;
1465 LLVMValueRef emitted_verts_ptr =
1466 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1467 LLVMValueRef emitted_prims_ptr =
1468 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1469 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1470
1471 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1472 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1473
1474 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1475 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1476 }
1477
1478 static void
1479 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1480 boolean elts)
1481 {
1482 struct gallivm_state *gallivm = variant->gallivm;
1483 LLVMContextRef context = gallivm->context;
1484 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1485 LLVMTypeRef arg_types[11];
1486 unsigned num_arg_types =
1487 elts ? ARRAY_SIZE(arg_types) : ARRAY_SIZE(arg_types) - 1;
1488 LLVMTypeRef func_type;
1489 LLVMValueRef context_ptr;
1490 LLVMBasicBlockRef block;
1491 LLVMBuilderRef builder;
1492 char func_name[64];
1493 struct lp_type vs_type;
1494 LLVMValueRef end, start;
1495 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1496 LLVMValueRef vertex_id_offset, start_instance;
1497 LLVMValueRef stride, step, io_itr;
1498 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1499 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1500 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1501 LLVMValueRef vb_stride[PIPE_MAX_SHADER_INPUTS];
1502 LLVMValueRef map_ptr[PIPE_MAX_SHADER_INPUTS];
1503 LLVMValueRef buffer_size_adj[PIPE_MAX_SHADER_INPUTS];
1504 LLVMValueRef stride_fixed[PIPE_MAX_SHADER_INPUTS];
1505 LLVMValueRef ofbit[PIPE_MAX_SHADER_INPUTS];
1506 LLVMValueRef instance_index[PIPE_MAX_SHADER_INPUTS];
1507
1508 struct draw_context *draw = llvm->draw;
1509 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1510 unsigned i, j;
1511 struct lp_build_context bld;
1512 struct lp_build_loop_state lp_loop;
1513 const int vector_length = lp_native_vector_width / 32;
1514 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1515 LLVMValueRef fetch_max;
1516 struct lp_build_sampler_soa *sampler = 0;
1517 LLVMValueRef ret, clipmask_bool_ptr;
1518 struct draw_llvm_variant_key *key = &variant->key;
1519 /* If geometry shader is present we need to skip both the viewport
1520 * transformation and clipping otherwise the inputs to the geometry
1521 * shader will be incorrect.
1522 * The code can't handle vp transform when vs writes vp index neither
1523 * (though this would be fixable here, but couldn't just broadcast
1524 * the values).
1525 */
1526 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1527 vs_info->writes_viewport_index;
1528 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1529 key->clip_z ||
1530 key->clip_user ||
1531 key->need_edgeflags);
1532 LLVMValueRef variant_func;
1533 const unsigned pos = draw->vs.position_output;
1534 const unsigned cv = draw->vs.clipvertex_output;
1535 boolean have_clipdist = FALSE;
1536 struct lp_bld_tgsi_system_values system_values;
1537
1538 memset(&system_values, 0, sizeof(system_values));
1539
1540 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1541 variant->shader->variants_cached, elts ? "elts" : "linear");
1542
1543 i = 0;
1544 arg_types[i++] = get_context_ptr_type(variant); /* context */
1545 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1546 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1547 if (elts) {
1548 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1549 arg_types[i++] = int32_type; /* fetch_elt_max */
1550 } else
1551 arg_types[i++] = int32_type; /* start */
1552 arg_types[i++] = int32_type; /* fetch_count / count */
1553 arg_types[i++] = int32_type; /* stride */
1554 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1555 arg_types[i++] = int32_type; /* instance_id */
1556 arg_types[i++] = int32_type; /* vertex_id_offset */
1557 arg_types[i++] = int32_type; /* start_instance */
1558
1559 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1560
1561 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1562
1563 if (elts)
1564 variant->function_elts = variant_func;
1565 else
1566 variant->function = variant_func;
1567
1568 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1569 for (i = 0; i < num_arg_types; ++i)
1570 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1571 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1572 LLVMNoAliasAttribute);
1573
1574 context_ptr = LLVMGetParam(variant_func, 0);
1575 io_ptr = LLVMGetParam(variant_func, 1);
1576 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1577 /*
1578 * XXX: stride is actually unused. The stride we use is strictly calculated
1579 * from the number of outputs (including the draw_extra outputs).
1580 * Should probably fix some day (we need a new vs just because of extra
1581 * outputs which the generated vs won't touch).
1582 */
1583 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1584 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1585 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1586 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1587 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1588
1589 lp_build_name(context_ptr, "context");
1590 lp_build_name(io_ptr, "io");
1591 lp_build_name(vbuffers_ptr, "vbuffers");
1592 lp_build_name(stride, "stride");
1593 lp_build_name(vb_ptr, "vb");
1594 lp_build_name(system_values.instance_id, "instance_id");
1595 lp_build_name(vertex_id_offset, "vertex_id_offset");
1596 lp_build_name(start_instance, "start_instance");
1597
1598 if (elts) {
1599 fetch_elts = LLVMGetParam(variant_func, 3);
1600 fetch_elt_max = LLVMGetParam(variant_func, 4);
1601 fetch_count = LLVMGetParam(variant_func, 5);
1602 lp_build_name(fetch_elts, "fetch_elts");
1603 lp_build_name(fetch_elt_max, "fetch_elt_max");
1604 lp_build_name(fetch_count, "fetch_count");
1605 start = count = NULL;
1606 }
1607 else {
1608 start = LLVMGetParam(variant_func, 3);
1609 count = LLVMGetParam(variant_func, 4);
1610 lp_build_name(start, "start");
1611 lp_build_name(count, "count");
1612 fetch_elts = fetch_count = NULL;
1613 }
1614
1615 /*
1616 * Function body
1617 */
1618
1619 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1620 builder = gallivm->builder;
1621 LLVMPositionBuilderAtEnd(builder, block);
1622
1623 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1624
1625 memset(&vs_type, 0, sizeof vs_type);
1626 vs_type.floating = TRUE; /* floating point values */
1627 vs_type.sign = TRUE; /* values are signed */
1628 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1629 vs_type.width = 32; /* 32-bit float */
1630 vs_type.length = vector_length;
1631
1632 /* hold temporary "bool" clipmask */
1633 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1634 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1635
1636 /* code generated texture sampling */
1637 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1638
1639 if (elts) {
1640 start = zero;
1641 end = fetch_count;
1642 count = fetch_count;
1643 }
1644 else {
1645 end = lp_build_add(&bld, start, count);
1646 }
1647
1648 step = lp_build_const_int32(gallivm, vector_length);
1649
1650 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1651
1652 /*
1653 * Pre-calculate everything which is constant per shader invocation.
1654 */
1655 for (j = 0; j < key->nr_vertex_elements; ++j) {
1656 LLVMValueRef vb_buffer_offset, buffer_size;
1657 LLVMValueRef vb_info, vbuffer_ptr;
1658 struct pipe_vertex_element *velem = &key->vertex_element[j];
1659 LLVMValueRef vb_index =
1660 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1661 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1662 util_format_get_blocksize(velem->src_format));
1663 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1664 velem->src_offset);
1665
1666 if (velem->src_format != PIPE_FORMAT_NONE) {
1667 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1668 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1669 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1670 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1671 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1672 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1673
1674 ofbit[j] = NULL;
1675 stride_fixed[j] = lp_build_uadd_overflow(gallivm, vb_buffer_offset,
1676 src_offset, &ofbit[j]);
1677 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, bsize,
1678 &ofbit[j]);
1679
1680 if (velem->instance_divisor) {
1681 /* Index is equal to the start instance plus the number of current
1682 * instance divided by the divisor. In this case we compute it as:
1683 * index = start_instance + (instance_id / divisor)
1684 */
1685 LLVMValueRef current_instance;
1686 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1687 lp_build_const_int32(gallivm,
1688 velem->instance_divisor),
1689 "instance_divisor");
1690 instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
1691 current_instance, &ofbit[j]);
1692 }
1693
1694 if (0) {
1695 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
1696 vb_index, vb_stride[j]);
1697 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
1698 vb_buffer_offset, src_offset);
1699 lp_build_print_value(gallivm, " blocksize = ", bsize);
1700 lp_build_printf(gallivm, " instance_id = %u\n",
1701 system_values.instance_id);
1702 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
1703 }
1704 }
1705 }
1706
1707 lp_build_loop_begin(&lp_loop, gallivm, zero);
1708 {
1709 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1710 LLVMValueRef aos_attribs[LP_MAX_VECTOR_WIDTH / 32] = { 0 };
1711 LLVMValueRef io;
1712 LLVMValueRef clipmask; /* holds the clipmask value */
1713 LLVMValueRef true_index_array = lp_build_zero(gallivm,
1714 lp_type_uint_vec(32, 32*vector_length));
1715 LLVMValueRef true_indices[LP_MAX_VECTOR_WIDTH / 32];
1716 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1717
1718 io_itr = lp_loop.counter;
1719
1720 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1721 #if DEBUG_STORE
1722 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1723 io_itr, io, lp_loop.counter);
1724 #endif
1725
1726 for (i = 0; i < vector_length; ++i) {
1727 LLVMValueRef vert_index =
1728 LLVMBuildAdd(builder,
1729 lp_loop.counter,
1730 lp_build_const_int32(gallivm, i), "");
1731 LLVMValueRef true_index =
1732 LLVMBuildAdd(builder, start, vert_index, "");
1733
1734 /* make sure we're not out of bounds which can happen
1735 * if fetch_count % 4 != 0, because on the last iteration
1736 * a few of the 4 vertex fetches will be out of bounds */
1737 true_index = lp_build_min(&bld, true_index, fetch_max);
1738
1739 if (elts) {
1740 LLVMValueRef fetch_ptr;
1741 LLVMValueRef index_overflowed;
1742 LLVMValueRef index_ptr =
1743 lp_build_alloca(
1744 gallivm,
1745 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1746 struct lp_build_if_state if_ctx;
1747 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1748 true_index, fetch_elt_max,
1749 "index_overflowed");
1750
1751 lp_build_if(&if_ctx, gallivm, index_overflowed);
1752 {
1753 /* Generate maximum possible index so that
1754 * generate_fetch can treat it just like
1755 * any other overflow and return zeros.
1756 * We don't have to worry about the restart
1757 * primitive index because it has already been
1758 * handled
1759 */
1760 LLVMValueRef val =
1761 lp_build_const_int32(gallivm, 0xffffffff);
1762 LLVMBuildStore(builder, val, index_ptr);
1763 }
1764 lp_build_else(&if_ctx);
1765 {
1766 LLVMValueRef val;
1767 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1768 &true_index, 1, "");
1769 val = LLVMBuildLoad(builder, fetch_ptr, "");
1770 LLVMBuildStore(builder, val, index_ptr);
1771 }
1772 lp_build_endif(&if_ctx);
1773 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1774 }
1775 true_indices[i] = true_index;
1776 true_index_array = LLVMBuildInsertElement(
1777 gallivm->builder, true_index_array, true_index,
1778 lp_build_const_int32(gallivm, i), "");
1779 }
1780
1781 for (j = 0; j < key->nr_vertex_elements; ++j) {
1782 struct pipe_vertex_element *velem = &key->vertex_element[j];
1783 const struct util_format_description *format_desc =
1784 util_format_description(velem->src_format);
1785
1786 for (i = 0; i < vector_length; ++i) {
1787 if (format_desc->format == PIPE_FORMAT_NONE) {
1788 aos_attribs[i] = lp_build_const_vec(gallivm,
1789 lp_float32_vec4_type(), 0);
1790 }
1791 else {
1792 generate_fetch(gallivm, format_desc,
1793 vb_stride[j], stride_fixed[j], map_ptr[j],
1794 buffer_size_adj[j], ofbit[j], &aos_attribs[i],
1795 velem->instance_divisor ?
1796 instance_index[j] : true_indices[i]);
1797 }
1798 }
1799 convert_to_soa(gallivm, aos_attribs, inputs, j, vs_type);
1800 }
1801
1802 /* In the paths with elts vertex id has to be unaffected by the
1803 * index bias and because indices inside our elements array have
1804 * already had index bias applied we need to subtract it here to
1805 * get back to the original index.
1806 * in the linear paths vertex id has to be unaffected by the
1807 * original start index and because we abuse the 'start' variable
1808 * to either represent the actual start index or the index at which
1809 * the primitive was split (we split rendering into chunks of at
1810 * most 4095-vertices) we need to back out the original start
1811 * index out of our vertex id here.
1812 */
1813 system_values.basevertex = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm,
1814 lp_type_uint_vec(32, 32*vector_length)),
1815 vertex_id_offset);
1816 system_values.vertex_id = true_index_array;
1817 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1818 system_values.basevertex, "");
1819
1820 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1821 generate_vs(variant,
1822 builder,
1823 vs_type,
1824 outputs,
1825 ptr_aos,
1826 &system_values,
1827 context_ptr,
1828 sampler,
1829 key->clamp_vertex_color);
1830
1831 if (pos != -1 && cv != -1) {
1832 /* store original positions in clip before further manipulation */
1833 store_clip(gallivm, vs_type, io, outputs, pos);
1834
1835 /* do cliptest */
1836 if (enable_cliptest) {
1837 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1838 /* allocate clipmask, assign it integer type */
1839 clipmask = generate_clipmask(llvm,
1840 gallivm,
1841 vs_type,
1842 outputs,
1843 key,
1844 context_ptr, &have_clipdist);
1845 temp = LLVMBuildOr(builder, clipmask, temp, "");
1846 /* store temporary clipping boolean value */
1847 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1848 }
1849 else {
1850 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1851 }
1852
1853 /* do viewport mapping */
1854 if (!bypass_viewport) {
1855 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1856 }
1857 }
1858 else {
1859 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1860 }
1861
1862 /* store clipmask in vertex header,
1863 * original positions in clip
1864 * and transformed positions in data
1865 */
1866 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1867 vs_info->num_outputs, vs_type,
1868 enable_cliptest && key->need_edgeflags);
1869 }
1870 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1871
1872 sampler->destroy(sampler);
1873
1874 /* return clipping boolean value for function */
1875 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr,
1876 enable_cliptest && key->need_edgeflags);
1877
1878 LLVMBuildRet(builder, ret);
1879
1880 gallivm_verify_function(gallivm, variant_func);
1881 }
1882
1883
1884 struct draw_llvm_variant_key *
1885 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1886 {
1887 unsigned i;
1888 struct draw_llvm_variant_key *key;
1889 struct draw_sampler_static_state *draw_sampler;
1890
1891 key = (struct draw_llvm_variant_key *)store;
1892
1893 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
1894
1895 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1896
1897 /* will have to rig this up properly later */
1898 key->clip_xy = llvm->draw->clip_xy;
1899 key->clip_z = llvm->draw->clip_z;
1900 key->clip_user = llvm->draw->clip_user;
1901 key->bypass_viewport = llvm->draw->bypass_viewport;
1902 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1903 /* XXX assumes edgeflag output not at 0 */
1904 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1905 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1906 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1907 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1908
1909 /* All variants of this shader will have the same value for
1910 * nr_samplers. Not yet trying to compact away holes in the
1911 * sampler array.
1912 */
1913 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1914 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1915 key->nr_sampler_views =
1916 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1917 }
1918 else {
1919 key->nr_sampler_views = key->nr_samplers;
1920 }
1921
1922 /* Presumably all variants of the shader should have the same
1923 * number of vertex elements - ie the number of shader inputs.
1924 * NOTE: we NEED to store the needed number of needed inputs
1925 * here, not the number of provided elements to match keysize
1926 * (and the offset of sampler state in the key).
1927 * If we have excess number of vertex elements, this is valid,
1928 * but the excess ones don't matter.
1929 * If we don't have enough vertex elements (which looks not really
1930 * valid but we'll handle it gracefully) fill out missing ones with
1931 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
1932 */
1933 key->nr_vertex_elements =
1934 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1935
1936 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
1937 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
1938 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
1939 memset(key->vertex_element, 0,
1940 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1941 }
1942 memcpy(key->vertex_element,
1943 llvm->draw->pt.vertex_element,
1944 sizeof(struct pipe_vertex_element) *
1945 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
1946
1947 draw_sampler = draw_llvm_variant_key_samplers(key);
1948 memset(draw_sampler, 0,
1949 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1950
1951 for (i = 0 ; i < key->nr_samplers; i++) {
1952 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1953 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1954 }
1955 for (i = 0 ; i < key->nr_sampler_views; i++) {
1956 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1957 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1958 }
1959
1960 return key;
1961 }
1962
1963
1964 void
1965 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1966 {
1967 unsigned i;
1968 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1969
1970 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1971 debug_printf("clip_xy = %u\n", key->clip_xy);
1972 debug_printf("clip_z = %u\n", key->clip_z);
1973 debug_printf("clip_user = %u\n", key->clip_user);
1974 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1975 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1976 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1977 debug_printf("has_gs = %u\n", key->has_gs);
1978 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1979
1980 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1981 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1982 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1983 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1984 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1985 }
1986
1987 for (i = 0 ; i < key->nr_sampler_views; i++) {
1988 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1989 }
1990 }
1991
1992
1993 void
1994 draw_llvm_set_mapped_texture(struct draw_context *draw,
1995 unsigned shader_stage,
1996 unsigned sview_idx,
1997 uint32_t width, uint32_t height, uint32_t depth,
1998 uint32_t first_level, uint32_t last_level,
1999 const void *base_ptr,
2000 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2001 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2002 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2003 {
2004 unsigned j;
2005 struct draw_jit_texture *jit_tex;
2006
2007 assert(shader_stage == PIPE_SHADER_VERTEX ||
2008 shader_stage == PIPE_SHADER_GEOMETRY);
2009
2010 if (shader_stage == PIPE_SHADER_VERTEX) {
2011 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2012
2013 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2014 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2015 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2016
2017 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2018 } else {
2019 assert(0);
2020 return;
2021 }
2022
2023 jit_tex->width = width;
2024 jit_tex->height = height;
2025 jit_tex->depth = depth;
2026 jit_tex->first_level = first_level;
2027 jit_tex->last_level = last_level;
2028 jit_tex->base = base_ptr;
2029
2030 for (j = first_level; j <= last_level; j++) {
2031 jit_tex->mip_offsets[j] = mip_offsets[j];
2032 jit_tex->row_stride[j] = row_stride[j];
2033 jit_tex->img_stride[j] = img_stride[j];
2034 }
2035 }
2036
2037
2038 void
2039 draw_llvm_set_sampler_state(struct draw_context *draw,
2040 unsigned shader_type)
2041 {
2042 unsigned i;
2043
2044 if (shader_type == PIPE_SHADER_VERTEX) {
2045 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2046 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2047
2048 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2049 const struct pipe_sampler_state *s
2050 = draw->samplers[PIPE_SHADER_VERTEX][i];
2051 jit_sam->min_lod = s->min_lod;
2052 jit_sam->max_lod = s->max_lod;
2053 jit_sam->lod_bias = s->lod_bias;
2054 COPY_4V(jit_sam->border_color, s->border_color.f);
2055 }
2056 }
2057 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2058 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2059 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2060
2061 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2062 const struct pipe_sampler_state *s
2063 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2064 jit_sam->min_lod = s->min_lod;
2065 jit_sam->max_lod = s->max_lod;
2066 jit_sam->lod_bias = s->lod_bias;
2067 COPY_4V(jit_sam->border_color, s->border_color.f);
2068 }
2069 }
2070 }
2071 }
2072
2073
2074 void
2075 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2076 {
2077 struct draw_llvm *llvm = variant->llvm;
2078
2079 gallivm_destroy(variant->gallivm);
2080
2081 remove_from_list(&variant->list_item_local);
2082 variant->shader->variants_cached--;
2083 remove_from_list(&variant->list_item_global);
2084 llvm->nr_variants--;
2085 FREE(variant);
2086 }
2087
2088
2089 /**
2090 * Create LLVM types for various structures.
2091 */
2092 static void
2093 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2094 {
2095 struct gallivm_state *gallivm = var->gallivm;
2096 LLVMTypeRef texture_type, sampler_type, context_type;
2097
2098 texture_type = create_jit_texture_type(gallivm, "texture");
2099 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2100
2101 context_type = create_gs_jit_context_type(gallivm,
2102 var->shader->base.vector_length,
2103 texture_type, sampler_type,
2104 "draw_gs_jit_context");
2105 var->context_ptr_type = LLVMPointerType(context_type, 0);
2106
2107 var->input_array_type = create_gs_jit_input_type(gallivm);
2108 }
2109
2110 static LLVMTypeRef
2111 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2112 {
2113 if (!variant->context_ptr_type)
2114 create_gs_jit_types(variant);
2115 return variant->context_ptr_type;
2116 }
2117
2118 static LLVMValueRef
2119 generate_mask_value(struct draw_gs_llvm_variant *variant,
2120 struct lp_type gs_type)
2121 {
2122 struct gallivm_state *gallivm = variant->gallivm;
2123 LLVMBuilderRef builder = gallivm->builder;
2124 struct lp_type mask_type = lp_int_type(gs_type);
2125 LLVMValueRef num_prims;
2126 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2127 unsigned i;
2128
2129 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2130 variant->num_prims);
2131 for (i = 0; i < gs_type.length; i++) {
2132 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2133 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2134 }
2135 mask_val = lp_build_compare(gallivm, mask_type,
2136 PIPE_FUNC_GREATER, num_prims, mask_val);
2137
2138 return mask_val;
2139 }
2140
2141 static void
2142 draw_gs_llvm_generate(struct draw_llvm *llvm,
2143 struct draw_gs_llvm_variant *variant)
2144 {
2145 struct gallivm_state *gallivm = variant->gallivm;
2146 LLVMContextRef context = gallivm->context;
2147 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2148 LLVMTypeRef arg_types[7];
2149 LLVMTypeRef func_type;
2150 LLVMValueRef variant_func;
2151 LLVMValueRef context_ptr;
2152 LLVMValueRef prim_id_ptr;
2153 LLVMBasicBlockRef block;
2154 LLVMBuilderRef builder;
2155 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2156 struct lp_build_sampler_soa *sampler = 0;
2157 struct lp_build_context bld;
2158 struct lp_bld_tgsi_system_values system_values;
2159 char func_name[64];
2160 struct lp_type gs_type;
2161 unsigned i;
2162 struct draw_gs_llvm_iface gs_iface;
2163 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2164 LLVMValueRef consts_ptr, num_consts_ptr;
2165 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2166 struct lp_build_mask_context mask;
2167 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2168 unsigned vector_length = variant->shader->base.vector_length;
2169
2170 memset(&system_values, 0, sizeof(system_values));
2171
2172 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2173 variant->shader->variants_cached);
2174
2175 assert(variant->vertex_header_ptr_type);
2176
2177 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2178 arg_types[1] = variant->input_array_type; /* input */
2179 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2180 arg_types[3] = int32_type; /* num_prims */
2181 arg_types[4] = int32_type; /* instance_id */
2182 arg_types[5] = LLVMPointerType(
2183 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2184 arg_types[6] = int32_type;
2185
2186 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2187
2188 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2189
2190 variant->function = variant_func;
2191
2192 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2193
2194 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2195 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2196 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2197 LLVMNoAliasAttribute);
2198
2199 context_ptr = LLVMGetParam(variant_func, 0);
2200 input_array = LLVMGetParam(variant_func, 1);
2201 io_ptr = LLVMGetParam(variant_func, 2);
2202 num_prims = LLVMGetParam(variant_func, 3);
2203 system_values.instance_id = LLVMGetParam(variant_func, 4);
2204 prim_id_ptr = LLVMGetParam(variant_func, 5);
2205 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2206
2207 lp_build_name(context_ptr, "context");
2208 lp_build_name(input_array, "input");
2209 lp_build_name(io_ptr, "io");
2210 lp_build_name(num_prims, "num_prims");
2211 lp_build_name(system_values.instance_id, "instance_id");
2212 lp_build_name(prim_id_ptr, "prim_id_ptr");
2213 lp_build_name(system_values.invocation_id, "invocation_id");
2214
2215 variant->context_ptr = context_ptr;
2216 variant->io_ptr = io_ptr;
2217 variant->num_prims = num_prims;
2218
2219 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2220 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2221 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2222 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2223 gs_iface.input = input_array;
2224 gs_iface.variant = variant;
2225
2226 /*
2227 * Function body
2228 */
2229
2230 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2231 builder = gallivm->builder;
2232 LLVMPositionBuilderAtEnd(builder, block);
2233
2234 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2235
2236 memset(&gs_type, 0, sizeof gs_type);
2237 gs_type.floating = TRUE; /* floating point values */
2238 gs_type.sign = TRUE; /* values are signed */
2239 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2240 gs_type.width = 32; /* 32-bit float */
2241 gs_type.length = vector_length;
2242
2243 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2244 num_consts_ptr =
2245 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2246
2247 /* code generated texture sampling */
2248 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2249
2250 mask_val = generate_mask_value(variant, gs_type);
2251 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2252
2253 if (gs_info->uses_primid) {
2254 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2255 }
2256
2257 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2258 tgsi_dump(tokens, 0);
2259 draw_gs_llvm_dump_variant_key(&variant->key);
2260 }
2261
2262 lp_build_tgsi_soa(variant->gallivm,
2263 tokens,
2264 gs_type,
2265 &mask,
2266 consts_ptr,
2267 num_consts_ptr,
2268 &system_values,
2269 NULL,
2270 outputs,
2271 context_ptr,
2272 NULL,
2273 sampler,
2274 &llvm->draw->gs.geometry_shader->info,
2275 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2276
2277 sampler->destroy(sampler);
2278
2279 lp_build_mask_end(&mask);
2280
2281 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2282
2283 gallivm_verify_function(gallivm, variant_func);
2284 }
2285
2286
2287 struct draw_gs_llvm_variant *
2288 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2289 unsigned num_outputs,
2290 const struct draw_gs_llvm_variant_key *key)
2291 {
2292 struct draw_gs_llvm_variant *variant;
2293 struct llvm_geometry_shader *shader =
2294 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2295 LLVMTypeRef vertex_header;
2296 char module_name[64];
2297
2298 variant = MALLOC(sizeof *variant +
2299 shader->variant_key_size -
2300 sizeof variant->key);
2301 if (!variant)
2302 return NULL;
2303
2304 variant->llvm = llvm;
2305 variant->shader = shader;
2306
2307 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2308 variant->shader->variants_cached);
2309
2310 variant->gallivm = gallivm_create(module_name, llvm->context);
2311
2312 create_gs_jit_types(variant);
2313
2314 memcpy(&variant->key, key, shader->variant_key_size);
2315
2316 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2317
2318 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2319
2320 draw_gs_llvm_generate(llvm, variant);
2321
2322 gallivm_compile_module(variant->gallivm);
2323
2324 variant->jit_func = (draw_gs_jit_func)
2325 gallivm_jit_function(variant->gallivm, variant->function);
2326
2327 gallivm_free_ir(variant->gallivm);
2328
2329 variant->list_item_global.base = variant;
2330 variant->list_item_local.base = variant;
2331 /*variant->no = */shader->variants_created++;
2332 variant->list_item_global.base = variant;
2333
2334 return variant;
2335 }
2336
2337 void
2338 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2339 {
2340 struct draw_llvm *llvm = variant->llvm;
2341
2342 gallivm_destroy(variant->gallivm);
2343
2344 remove_from_list(&variant->list_item_local);
2345 variant->shader->variants_cached--;
2346 remove_from_list(&variant->list_item_global);
2347 llvm->nr_gs_variants--;
2348 FREE(variant);
2349 }
2350
2351 struct draw_gs_llvm_variant_key *
2352 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2353 {
2354 unsigned i;
2355 struct draw_gs_llvm_variant_key *key;
2356 struct draw_sampler_static_state *draw_sampler;
2357
2358 key = (struct draw_gs_llvm_variant_key *)store;
2359
2360 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2361
2362 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2363
2364 /* All variants of this shader will have the same value for
2365 * nr_samplers. Not yet trying to compact away holes in the
2366 * sampler array.
2367 */
2368 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2369 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2370 key->nr_sampler_views =
2371 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2372 }
2373 else {
2374 key->nr_sampler_views = key->nr_samplers;
2375 }
2376
2377 draw_sampler = key->samplers;
2378
2379 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2380
2381 for (i = 0 ; i < key->nr_samplers; i++) {
2382 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2383 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2384 }
2385 for (i = 0 ; i < key->nr_sampler_views; i++) {
2386 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2387 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2388 }
2389
2390 return key;
2391 }
2392
2393 void
2394 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2395 {
2396 unsigned i;
2397 struct draw_sampler_static_state *sampler = key->samplers;
2398
2399 for (i = 0 ; i < key->nr_sampler_views; i++) {
2400 debug_printf("sampler[%i].src_format = %s\n", i,
2401 util_format_name(sampler[i].texture_state.format));
2402 }
2403 }