projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: Delete unneeded LLVM stuff earlier.
[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
529 variant = MALLOC(sizeof *variant +
530 shader->variant_key_size -
531 sizeof variant->key);
532 if (variant == NULL)
533 return NULL;
534
535 variant->llvm = llvm;
536
537 variant->gallivm = gallivm_create();
538
539 create_jit_types(variant);
540
541 memcpy(&variant->key, key, shader->variant_key_size);
542
543 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
544
545 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
546
547 draw_llvm_generate(llvm, variant, FALSE); /* linear */
548 draw_llvm_generate(llvm, variant, TRUE); /* elts */
549
550 gallivm_compile_module(variant->gallivm);
551
552 variant->jit_func = (draw_jit_vert_func)
553 gallivm_jit_function(variant->gallivm, variant->function);
554
555 variant->jit_func_elts = (draw_jit_vert_func_elts)
556 gallivm_jit_function(variant->gallivm, variant->function_elts);
557
558 gallivm_free_ir(variant->gallivm);
559
560 variant->shader = shader;
561 variant->list_item_global.base = variant;
562 variant->list_item_local.base = variant;
563 /*variant->no = */shader->variants_created++;
564 variant->list_item_global.base = variant;
565
566 return variant;
567 }
568
569
570 static void
571 generate_vs(struct draw_llvm_variant *variant,
572 LLVMBuilderRef builder,
573 struct lp_type vs_type,
574 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
575 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
576 const struct lp_bld_tgsi_system_values *system_values,
577 LLVMValueRef context_ptr,
578 struct lp_build_sampler_soa *draw_sampler,
579 boolean clamp_vertex_color)
580 {
581 struct draw_llvm *llvm = variant->llvm;
582 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
583 LLVMValueRef consts_ptr =
584 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
585 LLVMValueRef num_consts_ptr =
586 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
587 struct lp_build_sampler_soa *sampler = 0;
588
589 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
590 tgsi_dump(tokens, 0);
591 draw_llvm_dump_variant_key(&variant->key);
592 }
593
594 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
595 sampler = draw_sampler;
596
597 lp_build_tgsi_soa(variant->gallivm,
598 tokens,
599 vs_type,
600 NULL /*struct lp_build_mask_context *mask*/,
601 consts_ptr,
602 num_consts_ptr,
603 system_values,
604 inputs,
605 outputs,
606 sampler,
607 &llvm->draw->vs.vertex_shader->info,
608 NULL);
609
610 {
611 LLVMValueRef out;
612 unsigned chan, attrib;
613 struct lp_build_context bld;
614 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
615 lp_build_context_init(&bld, variant->gallivm, vs_type);
616
617 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
618 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
619 if (outputs[attrib][chan]) {
620 switch (info->output_semantic_name[attrib]) {
621 case TGSI_SEMANTIC_COLOR:
622 case TGSI_SEMANTIC_BCOLOR:
623 if (clamp_vertex_color) {
624 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
625 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
626 LLVMBuildStore(builder, out, outputs[attrib][chan]);
627 }
628 break;
629 }
630 }
631 }
632 }
633 }
634 }
635
636 static void
637 generate_fetch(struct gallivm_state *gallivm,
638 struct draw_context *draw,
639 LLVMValueRef vbuffers_ptr,
640 LLVMValueRef *res,
641 struct pipe_vertex_element *velem,
642 LLVMValueRef vbuf,
643 LLVMValueRef index,
644 LLVMValueRef instance_id)
645 {
646 const struct util_format_description *format_desc =
647 util_format_description(velem->src_format);
648 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
649 LLVMBuilderRef builder = gallivm->builder;
650 LLVMValueRef indices =
651 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
652 velem->vertex_buffer_index, 0);
653 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
654 &indices, 1, "");
655 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
656 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
657 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
658 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
659 LLVMValueRef stride;
660 LLVMValueRef buffer_overflowed;
661 LLVMValueRef needed_buffer_size;
662 LLVMValueRef temp_ptr =
663 lp_build_alloca(gallivm,
664 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
665 LLVMValueRef ofbit = NULL;
666 struct lp_build_if_state if_ctx;
667
668 if (velem->instance_divisor) {
669 /* Index is equal to the start instance plus the number of current
670 * instance divided by the divisor. In this case we compute it as:
671 * index = start_instance + (instance_id / divisor)
672 */
673 LLVMValueRef current_instance;
674 index = lp_build_const_int32(gallivm, draw->start_instance);
675 current_instance = LLVMBuildUDiv(builder, instance_id,
676 lp_build_const_int32(gallivm, velem->instance_divisor),
677 "instance_divisor");
678 index = lp_build_uadd_overflow(gallivm, index, current_instance, &ofbit);
679 }
680
681 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
682 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
683 stride = lp_build_uadd_overflow(
684 gallivm, stride,
685 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
686 needed_buffer_size = lp_build_uadd_overflow(
687 gallivm, stride,
688 lp_build_const_int32(gallivm,
689 util_format_get_blocksize(velem->src_format)),
690 &ofbit);
691
692 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
693 needed_buffer_size, buffer_size,
694 "buffer_overflowed");
695 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
696 #if 0
697 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
698 index, vb_stride);
699 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
700 vb_buffer_offset,
701 lp_build_const_int32(gallivm, velem->src_offset));
702 lp_build_print_value(gallivm, " blocksize = ",
703 lp_build_const_int32(
704 gallivm,
705 util_format_get_blocksize(velem->src_format)));
706 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
707 lp_build_printf(gallivm, " stride = %u\n", stride);
708 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
709 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
710 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
711 #endif
712
713 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
714 {
715 LLVMValueRef val =
716 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
717 LLVMBuildStore(builder, val, temp_ptr);
718 }
719 lp_build_else(&if_ctx);
720 {
721 LLVMValueRef val;
722 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
723
724 val = lp_build_fetch_rgba_aos(gallivm,
725 format_desc,
726 lp_float32_vec4_type(),
727 map_ptr,
728 zero, zero, zero);
729 LLVMBuildStore(builder, val, temp_ptr);
730 }
731 lp_build_endif(&if_ctx);
732
733 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
734 }
735
736 static void
737 convert_to_soa(struct gallivm_state *gallivm,
738 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
739 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
740 unsigned num_attribs, const struct lp_type soa_type)
741 {
742 unsigned i, j, k;
743 struct lp_type aos_channel_type = soa_type;
744
745 debug_assert(TGSI_NUM_CHANNELS == 4);
746 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
747
748 aos_channel_type.length >>= 1;
749
750 for (i = 0; i < num_attribs; ++i) {
751 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
752 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
753
754 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
755 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
756
757 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
758
759 for (k = 0; k < pixels_per_channel; ++k) {
760 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
761 }
762
763 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
764 }
765
766 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
767 }
768 }
769
770
771 static void
772 store_aos(struct gallivm_state *gallivm,
773 LLVMValueRef io_ptr,
774 LLVMValueRef index,
775 LLVMValueRef value)
776 {
777 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
778 LLVMBuilderRef builder = gallivm->builder;
779 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
780 LLVMValueRef indices[3];
781
782 indices[0] = lp_build_const_int32(gallivm, 0);
783 indices[1] = index;
784 indices[2] = lp_build_const_int32(gallivm, 0);
785
786 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
787 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
788
789 #if DEBUG_STORE
790 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
791 #endif
792
793 /* Unaligned store due to the vertex header */
794 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
795 }
796
797 /**
798 * Adjust the mask to architecture endianess. The mask will the store in struct:
799 *
800 * struct vertex_header {
801 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
802 * unsigned edgeflag:1;
803 * unsigned have_clipdist:1;
804 * unsigned vertex_id:16;
805 * [...]
806 * }
807 *
808 * On little-endian machine nothing needs to done, however on bit-endian machine
809 * the mask's fields need to be adjusted with the algorithm:
810 *
811 * uint32_t reverse (uint32_t x)
812 * {
813 * return (x >> 16) | // vertex_id
814 * ((x & 0x3fff) << 18) | // clipmask
815 * ((x & 0x4000) << 3) | // have_clipdist
816 * ((x & 0x8000) << 1); // edgeflag
817 * }
818 */
819 static LLVMValueRef
820 adjust_mask(struct gallivm_state *gallivm,
821 LLVMValueRef mask)
822 {
823 #ifdef PIPE_ARCH_BIG_ENDIAN
824 LLVMBuilderRef builder = gallivm->builder;
825 LLVMValueRef vertex_id;
826 LLVMValueRef clipmask;
827 LLVMValueRef have_clipdist;
828 LLVMValueRef edgeflag;
829
830 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
831 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
832 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
833 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
834 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
835 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
836 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
837
838 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
839 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
840 mask = LLVMBuildOr(builder, mask, edgeflag, "");
841 #endif
842 return mask;
843 }
844
845 static void
846 store_aos_array(struct gallivm_state *gallivm,
847 struct lp_type soa_type,
848 LLVMValueRef io_ptr,
849 LLVMValueRef *indices,
850 LLVMValueRef* aos,
851 int attrib,
852 int num_outputs,
853 LLVMValueRef clipmask,
854 boolean have_clipdist)
855 {
856 LLVMBuilderRef builder = gallivm->builder;
857 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
858 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
859 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
860 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
861 int vector_length = soa_type.length;
862 int i;
863
864 debug_assert(TGSI_NUM_CHANNELS == 4);
865
866 for (i = 0; i < vector_length; i++) {
867 linear_inds[i] = lp_build_const_int32(gallivm, i);
868 if (indices) {
869 inds[i] = indices[i];
870 } else {
871 inds[i] = linear_inds[i];
872 }
873 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
874 }
875
876 if (attrib == 0) {
877 /* store vertex header for each of the n vertices */
878 LLVMValueRef val, cliptmp;
879 int vertex_id_pad_edgeflag;
880
881 /* If this assertion fails, it means we need to update the bit twidding
882 * code here. See struct vertex_header in draw_private.h.
883 */
884 assert(DRAW_TOTAL_CLIP_PLANES==14);
885 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
886 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
887 if (have_clipdist)
888 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
889 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
890 /* OR with the clipmask */
891 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
892 for (i = 0; i < vector_length; i++) {
893 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
894 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
895 val = adjust_mask(gallivm, val);
896 #if DEBUG_STORE
897 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
898 io_ptrs[i], inds[i], val);
899 #endif
900 LLVMBuildStore(builder, val, id_ptr);
901 }
902 }
903
904 /* store for each of the n vertices */
905 for (i = 0; i < vector_length; i++) {
906 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
907 }
908 }
909
910
911 static void
912 convert_to_aos(struct gallivm_state *gallivm,
913 LLVMValueRef io,
914 LLVMValueRef *indices,
915 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
916 LLVMValueRef clipmask,
917 int num_outputs,
918 struct lp_type soa_type,
919 boolean have_clipdist)
920 {
921 LLVMBuilderRef builder = gallivm->builder;
922 unsigned chan, attrib, i;
923
924 #if DEBUG_STORE
925 lp_build_printf(gallivm, " # storing begin\n");
926 #endif
927 for (attrib = 0; attrib < num_outputs; ++attrib) {
928 LLVMValueRef soa[TGSI_NUM_CHANNELS];
929 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
930 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
931 if (outputs[attrib][chan]) {
932 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
933 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
934 #if DEBUG_STORE
935 lp_build_printf(gallivm, "output %d : %d ",
936 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
937 attrib, 0),
938 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
939 chan, 0));
940 lp_build_print_value(gallivm, "val = ", out);
941 {
942 LLVMValueRef iv =
943 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
944
945 lp_build_print_value(gallivm, " ival = ", iv);
946 }
947 #endif
948 soa[chan] = out;
949 }
950 else {
951 soa[chan] = 0;
952 }
953 }
954
955
956 if (soa_type.length == TGSI_NUM_CHANNELS) {
957 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
958 } else {
959 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
960
961 for (i = 0; i < soa_type.length; ++i) {
962 aos[i] = lp_build_extract_range(gallivm,
963 soa[i % TGSI_NUM_CHANNELS],
964 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
965 TGSI_NUM_CHANNELS);
966 }
967 }
968
969 store_aos_array(gallivm,
970 soa_type,
971 io, indices,
972 aos,
973 attrib,
974 num_outputs,
975 clipmask, have_clipdist);
976 }
977 #if DEBUG_STORE
978 lp_build_printf(gallivm, " # storing end\n");
979 #endif
980 }
981
982
983 /**
984 * Stores original vertex positions in clip coordinates
985 */
986 static void
987 store_clip(struct gallivm_state *gallivm,
988 const struct lp_type vs_type,
989 LLVMValueRef io_ptr,
990 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
991 boolean pre_clip_pos, int idx)
992 {
993 LLVMBuilderRef builder = gallivm->builder;
994 LLVMValueRef soa[4];
995 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
996 LLVMValueRef indices[2];
997 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
998 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
999 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1000 LLVMTypeRef clip_ptr_type =
1001 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1002 4), 0);
1003 int i, j;
1004
1005 indices[0] =
1006 indices[1] = lp_build_const_int32(gallivm, 0);
1007
1008 for (i = 0; i < vs_type.length; i++) {
1009 inds[i] = lp_build_const_int32(gallivm, i);
1010 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1011 }
1012
1013 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1014 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1015 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1016 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1017
1018 if (!pre_clip_pos) {
1019 for (i = 0; i < vs_type.length; i++) {
1020 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1021 }
1022 } else {
1023 for (i = 0; i < vs_type.length; i++) {
1024 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1025 }
1026 }
1027
1028 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1029 for (i = 0; i < vs_type.length; ++i) {
1030 aos[i] = lp_build_extract_range(gallivm,
1031 soa[i % TGSI_NUM_CHANNELS],
1032 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1033 TGSI_NUM_CHANNELS);
1034 }
1035
1036 for (j = 0; j < vs_type.length; j++) {
1037 LLVMValueRef clip_ptr;
1038
1039 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1040 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1041
1042 /* Unaligned store */
1043 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1044 }
1045 }
1046
1047
1048 /**
1049 * Transforms the outputs for viewport mapping
1050 */
1051 static void
1052 generate_viewport(struct draw_llvm_variant *variant,
1053 LLVMBuilderRef builder,
1054 struct lp_type vs_type,
1055 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1056 LLVMValueRef context_ptr)
1057 {
1058 int i;
1059 struct gallivm_state *gallivm = variant->gallivm;
1060 struct lp_type f32_type = vs_type;
1061 const unsigned pos = variant->llvm->draw->vs.position_output;
1062 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1063 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1064 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1065 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1066
1067 /* for 1/w convention*/
1068 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1069 LLVMBuildStore(builder, out3, outputs[pos][3]);
1070
1071 /* Viewport Mapping */
1072 for (i=0; i<3; i++) {
1073 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1074 LLVMValueRef scale;
1075 LLVMValueRef trans;
1076 LLVMValueRef scale_i;
1077 LLVMValueRef trans_i;
1078 LLVMValueRef index;
1079
1080 index = lp_build_const_int32(gallivm, i);
1081 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1082
1083 index = lp_build_const_int32(gallivm, i+4);
1084 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1085
1086 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1087 LLVMBuildLoad(builder, scale_i, "scale"));
1088 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1089 LLVMBuildLoad(builder, trans_i, "trans"));
1090
1091 /* divide by w */
1092 out = LLVMBuildFMul(builder, out, out3, "");
1093 /* mult by scale */
1094 out = LLVMBuildFMul(builder, out, scale, "");
1095 /* add translation */
1096 out = LLVMBuildFAdd(builder, out, trans, "");
1097
1098 /* store transformed outputs */
1099 LLVMBuildStore(builder, out, outputs[pos][i]);
1100 }
1101
1102 }
1103
1104
1105 /**
1106 * Returns clipmask as nxi32 bitmask for the n vertices
1107 */
1108 static LLVMValueRef
1109 generate_clipmask(struct draw_llvm *llvm,
1110 struct gallivm_state *gallivm,
1111 struct lp_type vs_type,
1112 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1113 boolean clip_xy,
1114 boolean clip_z,
1115 boolean clip_user,
1116 boolean clip_halfz,
1117 unsigned ucp_enable,
1118 LLVMValueRef context_ptr,
1119 boolean *have_clipdist)
1120 {
1121 LLVMBuilderRef builder = gallivm->builder;
1122 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1123 LLVMValueRef test, temp;
1124 LLVMValueRef zero, shift;
1125 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1126 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1127 LLVMValueRef plane1, planes, plane_ptr, sum;
1128 struct lp_type f32_type = vs_type;
1129 struct lp_type i32_type = lp_int_type(vs_type);
1130 const unsigned pos = llvm->draw->vs.position_output;
1131 const unsigned cv = llvm->draw->vs.clipvertex_output;
1132 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1133 bool have_cd = false;
1134 unsigned cd[2];
1135
1136 cd[0] = llvm->draw->vs.clipdistance_output[0];
1137 cd[1] = llvm->draw->vs.clipdistance_output[1];
1138
1139 if (cd[0] != pos || cd[1] != pos)
1140 have_cd = true;
1141
1142 if (num_written_clipdistance && !clip_user) {
1143 clip_user = true;
1144 ucp_enable = (1 << num_written_clipdistance) - 1;
1145 }
1146
1147 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1148 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1149 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1150 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1151
1152 /*
1153 * load clipvertex and position from correct locations.
1154 * if they are the same just load them once.
1155 */
1156 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1157 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1158 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1159 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1160
1161 if (clip_user && cv != pos) {
1162 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1163 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1164 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1165 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1166 } else {
1167 cv_x = pos_x;
1168 cv_y = pos_y;
1169 cv_z = pos_z;
1170 cv_w = pos_w;
1171 }
1172
1173 /* Cliptest, for hardwired planes */
1174 if (clip_xy) {
1175 /* plane 1 */
1176 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1177 temp = shift;
1178 test = LLVMBuildAnd(builder, test, temp, "");
1179 mask = test;
1180
1181 /* plane 2 */
1182 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1183 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1184 temp = LLVMBuildShl(builder, temp, shift, "");
1185 test = LLVMBuildAnd(builder, test, temp, "");
1186 mask = LLVMBuildOr(builder, mask, test, "");
1187
1188 /* plane 3 */
1189 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1190 temp = LLVMBuildShl(builder, temp, shift, "");
1191 test = LLVMBuildAnd(builder, test, temp, "");
1192 mask = LLVMBuildOr(builder, mask, test, "");
1193
1194 /* plane 4 */
1195 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1196 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1197 temp = LLVMBuildShl(builder, temp, shift, "");
1198 test = LLVMBuildAnd(builder, test, temp, "");
1199 mask = LLVMBuildOr(builder, mask, test, "");
1200 }
1201
1202 if (clip_z) {
1203 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1204 if (clip_halfz) {
1205 /* plane 5 */
1206 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1207 test = LLVMBuildAnd(builder, test, temp, "");
1208 mask = LLVMBuildOr(builder, mask, test, "");
1209 }
1210 else {
1211 /* plane 5 */
1212 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1213 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1214 test = LLVMBuildAnd(builder, test, temp, "");
1215 mask = LLVMBuildOr(builder, mask, test, "");
1216 }
1217 /* plane 6 */
1218 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1219 temp = LLVMBuildShl(builder, temp, shift, "");
1220 test = LLVMBuildAnd(builder, test, temp, "");
1221 mask = LLVMBuildOr(builder, mask, test, "");
1222 }
1223
1224 if (clip_user) {
1225 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1226 LLVMValueRef indices[3];
1227 LLVMValueRef is_nan_or_inf;
1228
1229 /* userclip planes */
1230 while (ucp_enable) {
1231 unsigned plane_idx = ffs(ucp_enable)-1;
1232 ucp_enable &= ~(1 << plane_idx);
1233 plane_idx += 6;
1234
1235 if (have_cd && num_written_clipdistance) {
1236 LLVMValueRef clipdist;
1237 int i;
1238 i = plane_idx - 6;
1239
1240 *have_clipdist = TRUE;
1241 if (i < 4) {
1242 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1243 } else {
1244 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1245 }
1246 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1247 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1248 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1249 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1250 test = LLVMBuildAnd(builder, test, temp, "");
1251 mask = LLVMBuildOr(builder, mask, test, "");
1252 } else {
1253 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1254 indices[0] = lp_build_const_int32(gallivm, 0);
1255 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1256
1257 indices[2] = lp_build_const_int32(gallivm, 0);
1258 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1259 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1260 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1261 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1262
1263 indices[2] = lp_build_const_int32(gallivm, 1);
1264 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1265 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1266 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1267 test = LLVMBuildFMul(builder, planes, cv_y, "");
1268 sum = LLVMBuildFAdd(builder, sum, test, "");
1269
1270 indices[2] = lp_build_const_int32(gallivm, 2);
1271 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1272 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1273 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1274 test = LLVMBuildFMul(builder, planes, cv_z, "");
1275 sum = LLVMBuildFAdd(builder, sum, test, "");
1276
1277 indices[2] = lp_build_const_int32(gallivm, 3);
1278 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1279 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1280 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1281 test = LLVMBuildFMul(builder, planes, cv_w, "");
1282 sum = LLVMBuildFAdd(builder, sum, test, "");
1283
1284 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1285 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1286 test = LLVMBuildAnd(builder, test, temp, "");
1287 mask = LLVMBuildOr(builder, mask, test, "");
1288 }
1289 }
1290 }
1291 return mask;
1292 }
1293
1294
1295 /**
1296 * Returns boolean if any clipping has occurred
1297 * Used zero/non-zero i32 value to represent boolean
1298 */
1299 static LLVMValueRef
1300 clipmask_booli32(struct gallivm_state *gallivm,
1301 const struct lp_type vs_type,
1302 LLVMValueRef clipmask_bool_ptr)
1303 {
1304 LLVMBuilderRef builder = gallivm->builder;
1305 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1306 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1307 LLVMValueRef ret = LLVMConstNull(int32_type);
1308 LLVMValueRef temp;
1309 int i;
1310
1311 /*
1312 * Can do this with log2(vector length) pack instructions and one extract
1313 * (as we don't actually need a or) with sse2 which would be way better.
1314 */
1315 for (i=0; i < vs_type.length; i++) {
1316 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1317 lp_build_const_int32(gallivm, i) , "");
1318 ret = LLVMBuildOr(builder, ret, temp, "");
1319 }
1320 return ret;
1321 }
1322
1323 static LLVMValueRef
1324 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1325 struct lp_build_tgsi_context * bld_base,
1326 boolean is_vindex_indirect,
1327 LLVMValueRef vertex_index,
1328 boolean is_aindex_indirect,
1329 LLVMValueRef attrib_index,
1330 LLVMValueRef swizzle_index)
1331 {
1332 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1333 struct gallivm_state *gallivm = bld_base->base.gallivm;
1334 LLVMBuilderRef builder = gallivm->builder;
1335 LLVMValueRef indices[3];
1336 LLVMValueRef res;
1337 struct lp_type type = bld_base->base.type;
1338
1339 if (is_vindex_indirect || is_aindex_indirect) {
1340 int i;
1341 res = bld_base->base.zero;
1342 for (i = 0; i < type.length; ++i) {
1343 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1344 LLVMValueRef vert_chan_index = vertex_index;
1345 LLVMValueRef attr_chan_index = attrib_index;
1346 LLVMValueRef channel_vec, value;
1347
1348 if (is_vindex_indirect) {
1349 vert_chan_index = LLVMBuildExtractElement(builder,
1350 vertex_index, idx, "");
1351 }
1352 if (is_aindex_indirect) {
1353 attr_chan_index = LLVMBuildExtractElement(builder,
1354 attrib_index, idx, "");
1355 }
1356
1357 indices[0] = vert_chan_index;
1358 indices[1] = attr_chan_index;
1359 indices[2] = swizzle_index;
1360
1361 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1362 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1363 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1364
1365 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1366 }
1367 } else {
1368 indices[0] = vertex_index;
1369 indices[1] = attrib_index;
1370 indices[2] = swizzle_index;
1371
1372 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1373 res = LLVMBuildLoad(builder, res, "");
1374 }
1375
1376 return res;
1377 }
1378
1379 static void
1380 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1381 struct lp_build_tgsi_context * bld_base,
1382 LLVMValueRef (*outputs)[4],
1383 LLVMValueRef emitted_vertices_vec)
1384 {
1385 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1386 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1387 struct gallivm_state *gallivm = variant->gallivm;
1388 LLVMBuilderRef builder = gallivm->builder;
1389 struct lp_type gs_type = bld_base->base.type;
1390 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1391 lp_int_type(gs_type), 0);
1392 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1393 LLVMValueRef next_prim_offset =
1394 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1395 LLVMValueRef io = variant->io_ptr;
1396 unsigned i;
1397 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1398
1399 for (i = 0; i < gs_type.length; ++i) {
1400 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1401 LLVMValueRef currently_emitted =
1402 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1403 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1404 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1405 }
1406
1407 convert_to_aos(gallivm, io, indices,
1408 outputs, clipmask,
1409 gs_info->num_outputs, gs_type,
1410 FALSE);
1411 }
1412
1413 static void
1414 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1415 struct lp_build_tgsi_context * bld_base,
1416 LLVMValueRef verts_per_prim_vec,
1417 LLVMValueRef emitted_prims_vec)
1418 {
1419 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1420 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1421 struct gallivm_state *gallivm = variant->gallivm;
1422 LLVMBuilderRef builder = gallivm->builder;
1423 LLVMValueRef prim_lengts_ptr =
1424 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1425 unsigned i;
1426
1427 for (i = 0; i < bld_base->base.type.length; ++i) {
1428 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1429 LLVMValueRef prims_emitted =
1430 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1431 LLVMValueRef store_ptr;
1432 LLVMValueRef num_vertices =
1433 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1434
1435 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1436 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1437 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1438 LLVMBuildStore(builder, num_vertices, store_ptr);
1439 }
1440 }
1441
1442 static void
1443 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1444 struct lp_build_tgsi_context * bld_base,
1445 LLVMValueRef total_emitted_vertices_vec,
1446 LLVMValueRef emitted_prims_vec)
1447 {
1448 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1449 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1450 struct gallivm_state *gallivm = variant->gallivm;
1451 LLVMBuilderRef builder = gallivm->builder;
1452 LLVMValueRef emitted_verts_ptr =
1453 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1454 LLVMValueRef emitted_prims_ptr =
1455 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1456 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1457
1458 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1459 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1460
1461 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1462 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1463 }
1464
1465 static void
1466 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1467 boolean elts)
1468 {
1469 struct gallivm_state *gallivm = variant->gallivm;
1470 LLVMContextRef context = gallivm->context;
1471 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1472 LLVMTypeRef arg_types[10];
1473 unsigned num_arg_types =
1474 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1475 LLVMTypeRef func_type;
1476 LLVMValueRef context_ptr;
1477 LLVMBasicBlockRef block;
1478 LLVMBuilderRef builder;
1479 struct lp_type vs_type;
1480 LLVMValueRef end, start;
1481 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1482 LLVMValueRef vertex_id_offset;
1483 LLVMValueRef stride, step, io_itr;
1484 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1485 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1486 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1487 struct draw_context *draw = llvm->draw;
1488 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1489 unsigned i, j;
1490 struct lp_build_context bld;
1491 struct lp_build_loop_state lp_loop;
1492 const int vector_length = lp_native_vector_width / 32;
1493 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1494 LLVMValueRef fetch_max;
1495 struct lp_build_sampler_soa *sampler = 0;
1496 LLVMValueRef ret, clipmask_bool_ptr;
1497 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1498 struct draw_llvm_variant_key *key = &variant->key;
1499 /* If geometry shader is present we need to skip both the viewport
1500 * transformation and clipping otherwise the inputs to the geometry
1501 * shader will be incorrect.
1502 */
1503 const boolean bypass_viewport = gs || key->bypass_viewport;
1504 const boolean enable_cliptest = !gs && (key->clip_xy ||
1505 key->clip_z ||
1506 key->clip_user);
1507 LLVMValueRef variant_func;
1508 const unsigned pos = llvm->draw->vs.position_output;
1509 const unsigned cv = llvm->draw->vs.clipvertex_output;
1510 boolean have_clipdist = FALSE;
1511 struct lp_bld_tgsi_system_values system_values;
1512
1513 memset(&system_values, 0, sizeof(system_values));
1514
1515 i = 0;
1516 arg_types[i++] = get_context_ptr_type(variant); /* context */
1517 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1518 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1519 if (elts) {
1520 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1521 arg_types[i++] = int32_type; /* fetch_elt_max */
1522 } else
1523 arg_types[i++] = int32_type; /* start */
1524 arg_types[i++] = int32_type; /* fetch_count / count */
1525 arg_types[i++] = int32_type; /* stride */
1526 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1527 arg_types[i++] = int32_type; /* instance_id */
1528 arg_types[i++] = int32_type; /* vertex_id_offset */
1529
1530 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1531
1532 variant_func = LLVMAddFunction(gallivm->module,
1533 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1534 func_type);
1535
1536 if (elts)
1537 variant->function_elts = variant_func;
1538 else
1539 variant->function = variant_func;
1540
1541 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1542 for (i = 0; i < num_arg_types; ++i)
1543 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1544 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1545 LLVMNoAliasAttribute);
1546
1547 context_ptr = LLVMGetParam(variant_func, 0);
1548 io_ptr = LLVMGetParam(variant_func, 1);
1549 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1550 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1551 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1552 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1553 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1554
1555 lp_build_name(context_ptr, "context");
1556 lp_build_name(io_ptr, "io");
1557 lp_build_name(vbuffers_ptr, "vbuffers");
1558 lp_build_name(stride, "stride");
1559 lp_build_name(vb_ptr, "vb");
1560 lp_build_name(system_values.instance_id, "instance_id");
1561 lp_build_name(vertex_id_offset, "vertex_id_offset");
1562
1563 if (elts) {
1564 fetch_elts = LLVMGetParam(variant_func, 3);
1565 fetch_elt_max = LLVMGetParam(variant_func, 4);
1566 fetch_count = LLVMGetParam(variant_func, 5);
1567 lp_build_name(fetch_elts, "fetch_elts");
1568 lp_build_name(fetch_elt_max, "fetch_elt_max");
1569 lp_build_name(fetch_count, "fetch_count");
1570 start = count = NULL;
1571 }
1572 else {
1573 start = LLVMGetParam(variant_func, 3);
1574 count = LLVMGetParam(variant_func, 4);
1575 lp_build_name(start, "start");
1576 lp_build_name(count, "count");
1577 fetch_elts = fetch_count = NULL;
1578 }
1579
1580 /*
1581 * Function body
1582 */
1583
1584 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1585 builder = gallivm->builder;
1586 LLVMPositionBuilderAtEnd(builder, block);
1587
1588 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1589
1590 memset(&vs_type, 0, sizeof vs_type);
1591 vs_type.floating = TRUE; /* floating point values */
1592 vs_type.sign = TRUE; /* values are signed */
1593 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1594 vs_type.width = 32; /* 32-bit float */
1595 vs_type.length = vector_length;
1596
1597 /* hold temporary "bool" clipmask */
1598 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1599 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1600
1601 /* code generated texture sampling */
1602 sampler = draw_llvm_sampler_soa_create(
1603 draw_llvm_variant_key_samplers(key),
1604 context_ptr);
1605
1606 if (elts) {
1607 start = zero;
1608 end = fetch_count;
1609 count = fetch_count;
1610 }
1611 else {
1612 end = lp_build_add(&bld, start, count);
1613 }
1614
1615 step = lp_build_const_int32(gallivm, vector_length);
1616
1617 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1618
1619 lp_build_loop_begin(&lp_loop, gallivm, zero);
1620 {
1621 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1622 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1623 LLVMValueRef io;
1624 LLVMValueRef clipmask; /* holds the clipmask value */
1625 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1626
1627 io_itr = lp_loop.counter;
1628
1629 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1630 #if DEBUG_STORE
1631 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1632 io_itr, io, lp_loop.counter);
1633 #endif
1634 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1635 for (i = 0; i < vector_length; ++i) {
1636 LLVMValueRef vert_index =
1637 LLVMBuildAdd(builder,
1638 lp_loop.counter,
1639 lp_build_const_int32(gallivm, i), "");
1640 LLVMValueRef true_index =
1641 LLVMBuildAdd(builder, start, vert_index, "");
1642 LLVMValueRef vertex_id;
1643
1644 /* make sure we're not out of bounds which can happen
1645 * if fetch_count % 4 != 0, because on the last iteration
1646 * a few of the 4 vertex fetches will be out of bounds */
1647 true_index = lp_build_min(&bld, true_index, fetch_max);
1648
1649 if (elts) {
1650 LLVMValueRef fetch_ptr;
1651 LLVMValueRef index_overflowed;
1652 LLVMValueRef index_ptr =
1653 lp_build_alloca(
1654 gallivm,
1655 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1656 struct lp_build_if_state if_ctx;
1657 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1658 true_index, fetch_elt_max,
1659 "index_overflowed");
1660
1661 lp_build_if(&if_ctx, gallivm, index_overflowed);
1662 {
1663 /* Generate maximum possible index so that
1664 * generate_fetch can treat it just like
1665 * any other overflow and return zeros.
1666 * We don't have to worry about the restart
1667 * primitive index because it has already been
1668 * handled
1669 */
1670 LLVMValueRef val =
1671 lp_build_const_int32(gallivm, 0xffffffff);
1672 LLVMBuildStore(builder, val, index_ptr);
1673 }
1674 lp_build_else(&if_ctx);
1675 {
1676 LLVMValueRef val;
1677 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1678 &true_index, 1, "");
1679 val = LLVMBuildLoad(builder, fetch_ptr, "");
1680 LLVMBuildStore(builder, val, index_ptr);
1681 }
1682 lp_build_endif(&if_ctx);
1683 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1684 }
1685 /* in the paths with elts vertex id has to be unaffected by the
1686 * index bias and because indices inside our elements array have
1687 * already had index bias applied we need to subtract it here to
1688 * get back to the original index.
1689 * in the linear paths vertex id has to be unaffected by the
1690 * original start index and because we abuse the 'start' variable
1691 * to either represent the actual start index or the index at which
1692 * the primitive was split (we split rendering into chunks of at
1693 * most 4095-vertices) we need to back out the original start
1694 * index out of our vertex id here.
1695 */
1696 vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
1697
1698 system_values.vertex_id = LLVMBuildInsertElement(
1699 gallivm->builder,
1700 system_values.vertex_id, vertex_id,
1701 lp_build_const_int32(gallivm, i), "");
1702
1703 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1704 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1705 LLVMValueRef vb_index =
1706 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1707 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1708 generate_fetch(gallivm, draw, vbuffers_ptr,
1709 &aos_attribs[j][i], velem, vb, true_index,
1710 system_values.instance_id);
1711 }
1712 }
1713 convert_to_soa(gallivm, aos_attribs, inputs,
1714 draw->pt.nr_vertex_elements, vs_type);
1715
1716 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1717 generate_vs(variant,
1718 builder,
1719 vs_type,
1720 outputs,
1721 ptr_aos,
1722 &system_values,
1723 context_ptr,
1724 sampler,
1725 key->clamp_vertex_color);
1726
1727 if (pos != -1 && cv != -1) {
1728 /* store original positions in clip before further manipulation */
1729 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1730 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1731
1732 /* do cliptest */
1733 if (enable_cliptest) {
1734 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1735 /* allocate clipmask, assign it integer type */
1736 clipmask = generate_clipmask(llvm,
1737 gallivm,
1738 vs_type,
1739 outputs,
1740 key->clip_xy,
1741 key->clip_z,
1742 key->clip_user,
1743 key->clip_halfz,
1744 key->ucp_enable,
1745 context_ptr, &have_clipdist);
1746 temp = LLVMBuildOr(builder, clipmask, temp, "");
1747 /* store temporary clipping boolean value */
1748 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1749 }
1750 else {
1751 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1752 }
1753
1754 /* do viewport mapping */
1755 if (!bypass_viewport) {
1756 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1757 }
1758 }
1759 else {
1760 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1761 }
1762
1763 /* store clipmask in vertex header,
1764 * original positions in clip
1765 * and transformed positions in data
1766 */
1767 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1768 vs_info->num_outputs, vs_type,
1769 have_clipdist);
1770 }
1771 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1772
1773 sampler->destroy(sampler);
1774
1775 /* return clipping boolean value for function */
1776 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1777
1778 LLVMBuildRet(builder, ret);
1779
1780 gallivm_verify_function(gallivm, variant_func);
1781 }
1782
1783
1784 struct draw_llvm_variant_key *
1785 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1786 {
1787 unsigned i;
1788 struct draw_llvm_variant_key *key;
1789 struct draw_sampler_static_state *draw_sampler;
1790
1791 key = (struct draw_llvm_variant_key *)store;
1792
1793 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1794
1795 /* Presumably all variants of the shader should have the same
1796 * number of vertex elements - ie the number of shader inputs.
1797 * NOTE: we NEED to store the needed number of needed inputs
1798 * here, not the number of provided elements to match keysize
1799 * (and the offset of sampler state in the key).
1800 */
1801 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1802 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1803
1804 /* will have to rig this up properly later */
1805 key->clip_xy = llvm->draw->clip_xy;
1806 key->clip_z = llvm->draw->clip_z;
1807 key->clip_user = llvm->draw->clip_user;
1808 key->bypass_viewport = llvm->draw->identity_viewport;
1809 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1810 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1811 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1812 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1813 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1814 key->pad1 = 0;
1815
1816 /* All variants of this shader will have the same value for
1817 * nr_samplers. Not yet trying to compact away holes in the
1818 * sampler array.
1819 */
1820 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1821 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1822 key->nr_sampler_views =
1823 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1824 }
1825 else {
1826 key->nr_sampler_views = key->nr_samplers;
1827 }
1828
1829 draw_sampler = draw_llvm_variant_key_samplers(key);
1830
1831 memcpy(key->vertex_element,
1832 llvm->draw->pt.vertex_element,
1833 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1834
1835 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1836
1837 for (i = 0 ; i < key->nr_samplers; i++) {
1838 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1839 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1840 }
1841 for (i = 0 ; i < key->nr_sampler_views; i++) {
1842 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1843 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1844 }
1845
1846 return key;
1847 }
1848
1849
1850 void
1851 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1852 {
1853 unsigned i;
1854 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1855
1856 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1857 debug_printf("clip_xy = %u\n", key->clip_xy);
1858 debug_printf("clip_z = %u\n", key->clip_z);
1859 debug_printf("clip_user = %u\n", key->clip_user);
1860 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1861 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1862 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1863 debug_printf("has_gs = %u\n", key->has_gs);
1864 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1865
1866 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1867 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1868 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1869 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1870 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1871 }
1872
1873 for (i = 0 ; i < key->nr_sampler_views; i++) {
1874 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1875 }
1876 }
1877
1878
1879 void
1880 draw_llvm_set_mapped_texture(struct draw_context *draw,
1881 unsigned shader_stage,
1882 unsigned sview_idx,
1883 uint32_t width, uint32_t height, uint32_t depth,
1884 uint32_t first_level, uint32_t last_level,
1885 const void *base_ptr,
1886 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1887 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1888 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1889 {
1890 unsigned j;
1891 struct draw_jit_texture *jit_tex;
1892
1893 assert(shader_stage == PIPE_SHADER_VERTEX ||
1894 shader_stage == PIPE_SHADER_GEOMETRY);
1895
1896 if (shader_stage == PIPE_SHADER_VERTEX) {
1897 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1898
1899 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1900 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1901 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1902
1903 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1904 } else {
1905 assert(0);
1906 return;
1907 }
1908
1909 jit_tex->width = width;
1910 jit_tex->height = height;
1911 jit_tex->depth = depth;
1912 jit_tex->first_level = first_level;
1913 jit_tex->last_level = last_level;
1914 jit_tex->base = base_ptr;
1915
1916 for (j = first_level; j <= last_level; j++) {
1917 jit_tex->mip_offsets[j] = mip_offsets[j];
1918 jit_tex->row_stride[j] = row_stride[j];
1919 jit_tex->img_stride[j] = img_stride[j];
1920 }
1921 }
1922
1923
1924 void
1925 draw_llvm_set_sampler_state(struct draw_context *draw,
1926 unsigned shader_type)
1927 {
1928 unsigned i;
1929
1930 if (shader_type == PIPE_SHADER_VERTEX) {
1931 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1932 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1933
1934 if (draw->samplers[i]) {
1935 const struct pipe_sampler_state *s
1936 = draw->samplers[PIPE_SHADER_VERTEX][i];
1937 jit_sam->min_lod = s->min_lod;
1938 jit_sam->max_lod = s->max_lod;
1939 jit_sam->lod_bias = s->lod_bias;
1940 COPY_4V(jit_sam->border_color, s->border_color.f);
1941 }
1942 }
1943 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1944 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1945 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1946
1947 if (draw->samplers[i]) {
1948 const struct pipe_sampler_state *s
1949 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1950 jit_sam->min_lod = s->min_lod;
1951 jit_sam->max_lod = s->max_lod;
1952 jit_sam->lod_bias = s->lod_bias;
1953 COPY_4V(jit_sam->border_color, s->border_color.f);
1954 }
1955 }
1956 }
1957 }
1958
1959
1960 void
1961 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1962 {
1963 struct draw_llvm *llvm = variant->llvm;
1964
1965 gallivm_destroy(variant->gallivm);
1966
1967 remove_from_list(&variant->list_item_local);
1968 variant->shader->variants_cached--;
1969 remove_from_list(&variant->list_item_global);
1970 llvm->nr_variants--;
1971 FREE(variant);
1972 }
1973
1974
1975 /**
1976 * Create LLVM types for various structures.
1977 */
1978 static void
1979 create_gs_jit_types(struct draw_gs_llvm_variant *var)
1980 {
1981 struct gallivm_state *gallivm = var->gallivm;
1982 LLVMTypeRef texture_type, sampler_type, context_type;
1983
1984 texture_type = create_jit_texture_type(gallivm, "texture");
1985 sampler_type = create_jit_sampler_type(gallivm, "sampler");
1986
1987 context_type = create_gs_jit_context_type(gallivm,
1988 var->shader->base.vector_length,
1989 texture_type, sampler_type,
1990 "draw_gs_jit_context");
1991 var->context_ptr_type = LLVMPointerType(context_type, 0);
1992
1993 var->input_array_type = create_gs_jit_input_type(gallivm);
1994 }
1995
1996 static LLVMTypeRef
1997 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
1998 {
1999 if (!variant->context_ptr_type)
2000 create_gs_jit_types(variant);
2001 return variant->context_ptr_type;
2002 }
2003
2004 static LLVMValueRef
2005 generate_mask_value(struct draw_gs_llvm_variant *variant,
2006 struct lp_type gs_type)
2007 {
2008 struct gallivm_state *gallivm = variant->gallivm;
2009 LLVMBuilderRef builder = gallivm->builder;
2010 struct lp_type mask_type = lp_int_type(gs_type);
2011 LLVMValueRef num_prims;
2012 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2013 unsigned i;
2014
2015 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2016 variant->num_prims);
2017 for (i = 0; i <= gs_type.length; i++) {
2018 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2019 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2020 }
2021 mask_val = lp_build_compare(gallivm, mask_type,
2022 PIPE_FUNC_GREATER, num_prims, mask_val);
2023
2024 return mask_val;
2025 }
2026
2027 static void
2028 draw_gs_llvm_generate(struct draw_llvm *llvm,
2029 struct draw_gs_llvm_variant *variant)
2030 {
2031 struct gallivm_state *gallivm = variant->gallivm;
2032 LLVMContextRef context = gallivm->context;
2033 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2034 LLVMTypeRef arg_types[6];
2035 LLVMTypeRef func_type;
2036 LLVMValueRef variant_func;
2037 LLVMValueRef context_ptr;
2038 LLVMValueRef prim_id_ptr;
2039 LLVMBasicBlockRef block;
2040 LLVMBuilderRef builder;
2041 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2042 struct lp_build_sampler_soa *sampler = 0;
2043 struct lp_build_context bld;
2044 struct lp_bld_tgsi_system_values system_values;
2045 struct lp_type gs_type;
2046 unsigned i;
2047 struct draw_gs_llvm_iface gs_iface;
2048 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2049 LLVMValueRef consts_ptr, num_consts_ptr;
2050 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2051 struct lp_build_mask_context mask;
2052 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2053 unsigned vector_length = variant->shader->base.vector_length;
2054
2055 memset(&system_values, 0, sizeof(system_values));
2056
2057 assert(variant->vertex_header_ptr_type);
2058
2059 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2060 arg_types[1] = variant->input_array_type; /* input */
2061 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2062 arg_types[3] = int32_type; /* num_prims */
2063 arg_types[4] = int32_type; /* instance_id */
2064 arg_types[5] = LLVMPointerType(
2065 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2066
2067 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2068
2069 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2070 func_type);
2071 variant->function = variant_func;
2072
2073 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2074
2075 for (i = 0; i < Elements(arg_types); ++i)
2076 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2077 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2078 LLVMNoAliasAttribute);
2079
2080 context_ptr = LLVMGetParam(variant_func, 0);
2081 input_array = LLVMGetParam(variant_func, 1);
2082 io_ptr = LLVMGetParam(variant_func, 2);
2083 num_prims = LLVMGetParam(variant_func, 3);
2084 system_values.instance_id = LLVMGetParam(variant_func, 4);
2085 prim_id_ptr = LLVMGetParam(variant_func, 5);
2086
2087 lp_build_name(context_ptr, "context");
2088 lp_build_name(input_array, "input");
2089 lp_build_name(io_ptr, "io");
2090 lp_build_name(num_prims, "num_prims");
2091 lp_build_name(system_values.instance_id, "instance_id");
2092 lp_build_name(prim_id_ptr, "prim_id_ptr");
2093
2094 variant->context_ptr = context_ptr;
2095 variant->io_ptr = io_ptr;
2096 variant->num_prims = num_prims;
2097
2098 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2099 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2100 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2101 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2102 gs_iface.input = input_array;
2103 gs_iface.variant = variant;
2104
2105 /*
2106 * Function body
2107 */
2108
2109 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2110 builder = gallivm->builder;
2111 LLVMPositionBuilderAtEnd(builder, block);
2112
2113 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2114
2115 memset(&gs_type, 0, sizeof gs_type);
2116 gs_type.floating = TRUE; /* floating point values */
2117 gs_type.sign = TRUE; /* values are signed */
2118 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2119 gs_type.width = 32; /* 32-bit float */
2120 gs_type.length = vector_length;
2121
2122 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2123 num_consts_ptr =
2124 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2125
2126 /* code generated texture sampling */
2127 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2128 context_ptr);
2129
2130 mask_val = generate_mask_value(variant, gs_type);
2131 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2132
2133 if (gs_info->uses_primid) {
2134 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2135 }
2136
2137 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2138 tgsi_dump(tokens, 0);
2139 draw_gs_llvm_dump_variant_key(&variant->key);
2140 }
2141
2142 lp_build_tgsi_soa(variant->gallivm,
2143 tokens,
2144 gs_type,
2145 &mask,
2146 consts_ptr,
2147 num_consts_ptr,
2148 &system_values,
2149 NULL,
2150 outputs,
2151 sampler,
2152 &llvm->draw->gs.geometry_shader->info,
2153 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2154
2155 sampler->destroy(sampler);
2156
2157 lp_build_mask_end(&mask);
2158
2159 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2160
2161 gallivm_verify_function(gallivm, variant_func);
2162 }
2163
2164
2165 struct draw_gs_llvm_variant *
2166 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2167 unsigned num_outputs,
2168 const struct draw_gs_llvm_variant_key *key)
2169 {
2170 struct draw_gs_llvm_variant *variant;
2171 struct llvm_geometry_shader *shader =
2172 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2173 LLVMTypeRef vertex_header;
2174
2175 variant = MALLOC(sizeof *variant +
2176 shader->variant_key_size -
2177 sizeof variant->key);
2178 if (variant == NULL)
2179 return NULL;
2180
2181 variant->llvm = llvm;
2182 variant->shader = shader;
2183
2184 variant->gallivm = gallivm_create();
2185
2186 create_gs_jit_types(variant);
2187
2188 memcpy(&variant->key, key, shader->variant_key_size);
2189
2190 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2191
2192 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2193
2194 draw_gs_llvm_generate(llvm, variant);
2195
2196 gallivm_compile_module(variant->gallivm);
2197
2198 variant->jit_func = (draw_gs_jit_func)
2199 gallivm_jit_function(variant->gallivm, variant->function);
2200
2201 gallivm_free_ir(variant->gallivm);
2202
2203 variant->list_item_global.base = variant;
2204 variant->list_item_local.base = variant;
2205 /*variant->no = */shader->variants_created++;
2206 variant->list_item_global.base = variant;
2207
2208 return variant;
2209 }
2210
2211 void
2212 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2213 {
2214 struct draw_llvm *llvm = variant->llvm;
2215
2216 gallivm_destroy(variant->gallivm);
2217
2218 remove_from_list(&variant->list_item_local);
2219 variant->shader->variants_cached--;
2220 remove_from_list(&variant->list_item_global);
2221 llvm->nr_gs_variants--;
2222 FREE(variant);
2223 }
2224
2225 struct draw_gs_llvm_variant_key *
2226 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2227 {
2228 unsigned i;
2229 struct draw_gs_llvm_variant_key *key;
2230 struct draw_sampler_static_state *draw_sampler;
2231
2232 key = (struct draw_gs_llvm_variant_key *)store;
2233
2234 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2235
2236 /* All variants of this shader will have the same value for
2237 * nr_samplers. Not yet trying to compact away holes in the
2238 * sampler array.
2239 */
2240 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2241 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2242 key->nr_sampler_views =
2243 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2244 }
2245 else {
2246 key->nr_sampler_views = key->nr_samplers;
2247 }
2248
2249 draw_sampler = key->samplers;
2250
2251 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2252
2253 for (i = 0 ; i < key->nr_samplers; i++) {
2254 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2255 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2256 }
2257 for (i = 0 ; i < key->nr_sampler_views; i++) {
2258 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2259 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2260 }
2261
2262 return key;
2263 }
2264
2265 void
2266 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2267 {
2268 unsigned i;
2269 struct draw_sampler_static_state *sampler = key->samplers;
2270
2271 for (i = 0 ; i < key->nr_sampler_views; i++) {
2272 debug_printf("sampler[%i].src_format = %s\n", i,
2273 util_format_name(sampler[i].texture_state.format));
2274 }
2275 }