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