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