projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: Make it more clear that *_jit_context points to pipe_viewport_state structures.
[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); /* 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 draw_sampler,
624 &llvm->draw->vs.vertex_shader->info,
625 NULL);
626
627 {
628 LLVMValueRef out;
629 unsigned chan, attrib;
630 struct lp_build_context bld;
631 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
632 lp_build_context_init(&bld, variant->gallivm, vs_type);
633
634 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
635 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
636 if (outputs[attrib][chan]) {
637 switch (info->output_semantic_name[attrib]) {
638 case TGSI_SEMANTIC_COLOR:
639 case TGSI_SEMANTIC_BCOLOR:
640 if (clamp_vertex_color) {
641 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
642 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
643 LLVMBuildStore(builder, out, outputs[attrib][chan]);
644 }
645 break;
646 }
647 }
648 }
649 }
650 }
651 }
652
653 static void
654 generate_fetch(struct gallivm_state *gallivm,
655 struct draw_context *draw,
656 LLVMValueRef vbuffers_ptr,
657 LLVMValueRef *res,
658 struct pipe_vertex_element *velem,
659 LLVMValueRef vbuf,
660 LLVMValueRef index,
661 LLVMValueRef instance_id,
662 LLVMValueRef start_instance)
663 {
664 const struct util_format_description *format_desc =
665 util_format_description(velem->src_format);
666 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
667 LLVMBuilderRef builder = gallivm->builder;
668 LLVMValueRef indices =
669 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
670 velem->vertex_buffer_index, 0);
671 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
672 &indices, 1, "");
673 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
674 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
675 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
676 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
677 LLVMValueRef stride;
678 LLVMValueRef buffer_overflowed;
679 LLVMValueRef needed_buffer_size;
680 LLVMValueRef temp_ptr =
681 lp_build_alloca(gallivm,
682 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
683 LLVMValueRef ofbit = NULL;
684 struct lp_build_if_state if_ctx;
685
686 if (velem->instance_divisor) {
687 /* Index is equal to the start instance plus the number of current
688 * instance divided by the divisor. In this case we compute it as:
689 * index = start_instance + (instance_id / divisor)
690 */
691 LLVMValueRef current_instance;
692 current_instance = LLVMBuildUDiv(builder, instance_id,
693 lp_build_const_int32(gallivm, velem->instance_divisor),
694 "instance_divisor");
695 index = lp_build_uadd_overflow(gallivm, start_instance,
696 current_instance, &ofbit);
697 }
698
699 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
700 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
701 stride = lp_build_uadd_overflow(
702 gallivm, stride,
703 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
704 needed_buffer_size = lp_build_uadd_overflow(
705 gallivm, stride,
706 lp_build_const_int32(gallivm,
707 util_format_get_blocksize(velem->src_format)),
708 &ofbit);
709
710 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
711 needed_buffer_size, buffer_size,
712 "buffer_overflowed");
713 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
714 #if 0
715 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
716 index, vb_stride);
717 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
718 vb_buffer_offset,
719 lp_build_const_int32(gallivm, velem->src_offset));
720 lp_build_print_value(gallivm, " blocksize = ",
721 lp_build_const_int32(
722 gallivm,
723 util_format_get_blocksize(velem->src_format)));
724 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
725 lp_build_printf(gallivm, " stride = %u\n", stride);
726 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
727 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
728 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
729 #endif
730
731 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
732 {
733 LLVMValueRef val =
734 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
735 LLVMBuildStore(builder, val, temp_ptr);
736 }
737 lp_build_else(&if_ctx);
738 {
739 LLVMValueRef val;
740 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
741
742 val = lp_build_fetch_rgba_aos(gallivm,
743 format_desc,
744 lp_float32_vec4_type(),
745 map_ptr,
746 zero, zero, zero);
747 LLVMBuildStore(builder, val, temp_ptr);
748 }
749 lp_build_endif(&if_ctx);
750
751 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
752 }
753
754 static void
755 convert_to_soa(struct gallivm_state *gallivm,
756 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
757 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
758 unsigned num_attribs, const struct lp_type soa_type)
759 {
760 unsigned i, j, k;
761 struct lp_type aos_channel_type = soa_type;
762
763 debug_assert(TGSI_NUM_CHANNELS == 4);
764 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
765
766 aos_channel_type.length >>= 1;
767
768 for (i = 0; i < num_attribs; ++i) {
769 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
770 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
771
772 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
773 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
774
775 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
776
777 for (k = 0; k < pixels_per_channel; ++k) {
778 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
779 }
780
781 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
782 }
783
784 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
785 }
786 }
787
788
789 static void
790 store_aos(struct gallivm_state *gallivm,
791 LLVMValueRef io_ptr,
792 LLVMValueRef index,
793 LLVMValueRef value)
794 {
795 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
796 LLVMBuilderRef builder = gallivm->builder;
797 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
798 LLVMValueRef indices[3];
799
800 indices[0] = lp_build_const_int32(gallivm, 0);
801 indices[1] = index;
802 indices[2] = lp_build_const_int32(gallivm, 0);
803
804 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
805 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
806
807 #if DEBUG_STORE
808 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
809 #endif
810
811 /* Unaligned store due to the vertex header */
812 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
813 }
814
815 /**
816 * Adjust the mask to architecture endianess. The mask will the store in struct:
817 *
818 * struct vertex_header {
819 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
820 * unsigned edgeflag:1;
821 * unsigned have_clipdist:1;
822 * unsigned vertex_id:16;
823 * [...]
824 * }
825 *
826 * On little-endian machine nothing needs to done, however on bit-endian machine
827 * the mask's fields need to be adjusted with the algorithm:
828 *
829 * uint32_t reverse (uint32_t x)
830 * {
831 * return (x >> 16) | // vertex_id
832 * ((x & 0x3fff) << 18) | // clipmask
833 * ((x & 0x4000) << 3) | // have_clipdist
834 * ((x & 0x8000) << 1); // edgeflag
835 * }
836 */
837 static LLVMValueRef
838 adjust_mask(struct gallivm_state *gallivm,
839 LLVMValueRef mask)
840 {
841 #ifdef PIPE_ARCH_BIG_ENDIAN
842 LLVMBuilderRef builder = gallivm->builder;
843 LLVMValueRef vertex_id;
844 LLVMValueRef clipmask;
845 LLVMValueRef have_clipdist;
846 LLVMValueRef edgeflag;
847
848 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
849 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
850 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
851 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
852 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
853 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
854 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
855
856 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
857 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
858 mask = LLVMBuildOr(builder, mask, edgeflag, "");
859 #endif
860 return mask;
861 }
862
863 static void
864 store_aos_array(struct gallivm_state *gallivm,
865 struct lp_type soa_type,
866 LLVMValueRef io_ptr,
867 LLVMValueRef *indices,
868 LLVMValueRef* aos,
869 int attrib,
870 int num_outputs,
871 LLVMValueRef clipmask,
872 boolean have_clipdist)
873 {
874 LLVMBuilderRef builder = gallivm->builder;
875 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
876 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
877 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
878 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
879 int vector_length = soa_type.length;
880 int i;
881
882 debug_assert(TGSI_NUM_CHANNELS == 4);
883
884 for (i = 0; i < vector_length; i++) {
885 linear_inds[i] = lp_build_const_int32(gallivm, i);
886 if (indices) {
887 inds[i] = indices[i];
888 } else {
889 inds[i] = linear_inds[i];
890 }
891 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
892 }
893
894 if (attrib == 0) {
895 /* store vertex header for each of the n vertices */
896 LLVMValueRef val, cliptmp;
897 int vertex_id_pad_edgeflag;
898
899 /* If this assertion fails, it means we need to update the bit twidding
900 * code here. See struct vertex_header in draw_private.h.
901 */
902 assert(DRAW_TOTAL_CLIP_PLANES==14);
903 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
904 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
905 if (have_clipdist)
906 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
907 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
908 /* OR with the clipmask */
909 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
910 for (i = 0; i < vector_length; i++) {
911 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
912 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
913 val = adjust_mask(gallivm, val);
914 #if DEBUG_STORE
915 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
916 io_ptrs[i], inds[i], val);
917 #endif
918 LLVMBuildStore(builder, val, id_ptr);
919 }
920 }
921
922 /* store for each of the n vertices */
923 for (i = 0; i < vector_length; i++) {
924 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
925 }
926 }
927
928
929 static void
930 convert_to_aos(struct gallivm_state *gallivm,
931 LLVMValueRef io,
932 LLVMValueRef *indices,
933 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
934 LLVMValueRef clipmask,
935 int num_outputs,
936 struct lp_type soa_type,
937 boolean have_clipdist)
938 {
939 LLVMBuilderRef builder = gallivm->builder;
940 unsigned chan, attrib, i;
941
942 #if DEBUG_STORE
943 lp_build_printf(gallivm, " # storing begin\n");
944 #endif
945 for (attrib = 0; attrib < num_outputs; ++attrib) {
946 LLVMValueRef soa[TGSI_NUM_CHANNELS];
947 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
948 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
949 if (outputs[attrib][chan]) {
950 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
951 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
952 #if DEBUG_STORE
953 lp_build_printf(gallivm, "output %d : %d ",
954 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
955 attrib, 0),
956 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
957 chan, 0));
958 lp_build_print_value(gallivm, "val = ", out);
959 {
960 LLVMValueRef iv =
961 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
962
963 lp_build_print_value(gallivm, " ival = ", iv);
964 }
965 #endif
966 soa[chan] = out;
967 }
968 else {
969 soa[chan] = 0;
970 }
971 }
972
973
974 if (soa_type.length == TGSI_NUM_CHANNELS) {
975 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
976 } else {
977 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
978
979 for (i = 0; i < soa_type.length; ++i) {
980 aos[i] = lp_build_extract_range(gallivm,
981 soa[i % TGSI_NUM_CHANNELS],
982 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
983 TGSI_NUM_CHANNELS);
984 }
985 }
986
987 store_aos_array(gallivm,
988 soa_type,
989 io, indices,
990 aos,
991 attrib,
992 num_outputs,
993 clipmask, have_clipdist);
994 }
995 #if DEBUG_STORE
996 lp_build_printf(gallivm, " # storing end\n");
997 #endif
998 }
999
1000
1001 /**
1002 * Stores original vertex positions in clip coordinates
1003 */
1004 static void
1005 store_clip(struct gallivm_state *gallivm,
1006 const struct lp_type vs_type,
1007 LLVMValueRef io_ptr,
1008 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1009 boolean pre_clip_pos, int idx)
1010 {
1011 LLVMBuilderRef builder = gallivm->builder;
1012 LLVMValueRef soa[4];
1013 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1014 LLVMValueRef indices[2];
1015 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1016 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1017 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1018 LLVMTypeRef clip_ptr_type =
1019 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1020 4), 0);
1021 int i, j;
1022
1023 indices[0] =
1024 indices[1] = lp_build_const_int32(gallivm, 0);
1025
1026 for (i = 0; i < vs_type.length; i++) {
1027 inds[i] = lp_build_const_int32(gallivm, i);
1028 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1029 }
1030
1031 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1032 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1033 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1034 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1035
1036 if (!pre_clip_pos) {
1037 for (i = 0; i < vs_type.length; i++) {
1038 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1039 }
1040 } else {
1041 for (i = 0; i < vs_type.length; i++) {
1042 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1043 }
1044 }
1045
1046 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1047 for (i = 0; i < vs_type.length; ++i) {
1048 aos[i] = lp_build_extract_range(gallivm,
1049 soa[i % TGSI_NUM_CHANNELS],
1050 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1051 TGSI_NUM_CHANNELS);
1052 }
1053
1054 for (j = 0; j < vs_type.length; j++) {
1055 LLVMValueRef clip_ptr;
1056
1057 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1058 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1059
1060 /* Unaligned store */
1061 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1062 }
1063 }
1064
1065
1066 /**
1067 * Transforms the outputs for viewport mapping
1068 */
1069 static void
1070 generate_viewport(struct draw_llvm_variant *variant,
1071 LLVMBuilderRef builder,
1072 struct lp_type vs_type,
1073 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1074 LLVMValueRef context_ptr)
1075 {
1076 int i;
1077 struct gallivm_state *gallivm = variant->gallivm;
1078 struct lp_type f32_type = vs_type;
1079 const unsigned pos = variant->llvm->draw->vs.position_output;
1080 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1081 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1082 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1083 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1084
1085 /* We treat pipe_viewport_state as a float array */
1086 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1087 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1088
1089 /* for 1/w convention*/
1090 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1091 LLVMBuildStore(builder, out3, outputs[pos][3]);
1092
1093 /* Viewport Mapping */
1094 for (i=0; i<3; i++) {
1095 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1096 LLVMValueRef scale;
1097 LLVMValueRef trans;
1098 LLVMValueRef scale_i;
1099 LLVMValueRef trans_i;
1100 LLVMValueRef index;
1101
1102 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1103 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1104
1105 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1106 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1107
1108 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1109 LLVMBuildLoad(builder, scale_i, "scale"));
1110 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1111 LLVMBuildLoad(builder, trans_i, "trans"));
1112
1113 /* divide by w */
1114 out = LLVMBuildFMul(builder, out, out3, "");
1115 /* mult by scale */
1116 out = LLVMBuildFMul(builder, out, scale, "");
1117 /* add translation */
1118 out = LLVMBuildFAdd(builder, out, trans, "");
1119
1120 /* store transformed outputs */
1121 LLVMBuildStore(builder, out, outputs[pos][i]);
1122 }
1123
1124 }
1125
1126
1127 /**
1128 * Returns clipmask as nxi32 bitmask for the n vertices
1129 */
1130 static LLVMValueRef
1131 generate_clipmask(struct draw_llvm *llvm,
1132 struct gallivm_state *gallivm,
1133 struct lp_type vs_type,
1134 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1135 boolean clip_xy,
1136 boolean clip_z,
1137 boolean clip_user,
1138 boolean clip_halfz,
1139 unsigned ucp_enable,
1140 LLVMValueRef context_ptr,
1141 boolean *have_clipdist)
1142 {
1143 LLVMBuilderRef builder = gallivm->builder;
1144 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1145 LLVMValueRef test, temp;
1146 LLVMValueRef zero, shift;
1147 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1148 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1149 LLVMValueRef plane1, planes, plane_ptr, sum;
1150 struct lp_type f32_type = vs_type;
1151 struct lp_type i32_type = lp_int_type(vs_type);
1152 const unsigned pos = llvm->draw->vs.position_output;
1153 const unsigned cv = llvm->draw->vs.clipvertex_output;
1154 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1155 bool have_cd = false;
1156 unsigned cd[2];
1157
1158 cd[0] = llvm->draw->vs.clipdistance_output[0];
1159 cd[1] = llvm->draw->vs.clipdistance_output[1];
1160
1161 if (cd[0] != pos || cd[1] != pos)
1162 have_cd = true;
1163
1164 if (num_written_clipdistance && !clip_user) {
1165 clip_user = true;
1166 ucp_enable = (1 << num_written_clipdistance) - 1;
1167 }
1168
1169 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1170 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1171 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1172 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1173
1174 /*
1175 * load clipvertex and position from correct locations.
1176 * if they are the same just load them once.
1177 */
1178 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1179 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1180 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1181 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1182
1183 if (clip_user && cv != pos) {
1184 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1185 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1186 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1187 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1188 } else {
1189 cv_x = pos_x;
1190 cv_y = pos_y;
1191 cv_z = pos_z;
1192 cv_w = pos_w;
1193 }
1194
1195 /* Cliptest, for hardwired planes */
1196 if (clip_xy) {
1197 /* plane 1 */
1198 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1199 temp = shift;
1200 test = LLVMBuildAnd(builder, test, temp, "");
1201 mask = test;
1202
1203 /* plane 2 */
1204 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1205 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1206 temp = LLVMBuildShl(builder, temp, shift, "");
1207 test = LLVMBuildAnd(builder, test, temp, "");
1208 mask = LLVMBuildOr(builder, mask, test, "");
1209
1210 /* plane 3 */
1211 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1212 temp = LLVMBuildShl(builder, temp, shift, "");
1213 test = LLVMBuildAnd(builder, test, temp, "");
1214 mask = LLVMBuildOr(builder, mask, test, "");
1215
1216 /* plane 4 */
1217 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1218 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1219 temp = LLVMBuildShl(builder, temp, shift, "");
1220 test = LLVMBuildAnd(builder, test, temp, "");
1221 mask = LLVMBuildOr(builder, mask, test, "");
1222 }
1223
1224 if (clip_z) {
1225 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1226 if (clip_halfz) {
1227 /* plane 5 */
1228 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1229 test = LLVMBuildAnd(builder, test, temp, "");
1230 mask = LLVMBuildOr(builder, mask, test, "");
1231 }
1232 else {
1233 /* plane 5 */
1234 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1235 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1236 test = LLVMBuildAnd(builder, test, temp, "");
1237 mask = LLVMBuildOr(builder, mask, test, "");
1238 }
1239 /* plane 6 */
1240 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1241 temp = LLVMBuildShl(builder, temp, shift, "");
1242 test = LLVMBuildAnd(builder, test, temp, "");
1243 mask = LLVMBuildOr(builder, mask, test, "");
1244 }
1245
1246 if (clip_user) {
1247 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1248 LLVMValueRef indices[3];
1249 LLVMValueRef is_nan_or_inf;
1250
1251 /* userclip planes */
1252 while (ucp_enable) {
1253 unsigned plane_idx = ffs(ucp_enable)-1;
1254 ucp_enable &= ~(1 << plane_idx);
1255 plane_idx += 6;
1256
1257 if (have_cd && num_written_clipdistance) {
1258 LLVMValueRef clipdist;
1259 int i;
1260 i = plane_idx - 6;
1261
1262 *have_clipdist = TRUE;
1263 if (i < 4) {
1264 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1265 } else {
1266 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1267 }
1268 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1269 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1270 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1271 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1272 test = LLVMBuildAnd(builder, test, temp, "");
1273 mask = LLVMBuildOr(builder, mask, test, "");
1274 } else {
1275 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1276 indices[0] = lp_build_const_int32(gallivm, 0);
1277 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1278
1279 indices[2] = lp_build_const_int32(gallivm, 0);
1280 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1281 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1282 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1283 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1284
1285 indices[2] = lp_build_const_int32(gallivm, 1);
1286 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1287 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1288 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1289 test = LLVMBuildFMul(builder, planes, cv_y, "");
1290 sum = LLVMBuildFAdd(builder, sum, test, "");
1291
1292 indices[2] = lp_build_const_int32(gallivm, 2);
1293 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1294 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1295 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1296 test = LLVMBuildFMul(builder, planes, cv_z, "");
1297 sum = LLVMBuildFAdd(builder, sum, test, "");
1298
1299 indices[2] = lp_build_const_int32(gallivm, 3);
1300 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1301 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1302 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1303 test = LLVMBuildFMul(builder, planes, cv_w, "");
1304 sum = LLVMBuildFAdd(builder, sum, test, "");
1305
1306 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1307 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1308 test = LLVMBuildAnd(builder, test, temp, "");
1309 mask = LLVMBuildOr(builder, mask, test, "");
1310 }
1311 }
1312 }
1313 return mask;
1314 }
1315
1316
1317 /**
1318 * Returns boolean if any clipping has occurred
1319 * Used zero/non-zero i32 value to represent boolean
1320 */
1321 static LLVMValueRef
1322 clipmask_booli32(struct gallivm_state *gallivm,
1323 const struct lp_type vs_type,
1324 LLVMValueRef clipmask_bool_ptr)
1325 {
1326 LLVMBuilderRef builder = gallivm->builder;
1327 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1328 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1329 LLVMValueRef ret = LLVMConstNull(int32_type);
1330 LLVMValueRef temp;
1331 int i;
1332
1333 /*
1334 * Can do this with log2(vector length) pack instructions and one extract
1335 * (as we don't actually need a or) with sse2 which would be way better.
1336 */
1337 for (i=0; i < vs_type.length; i++) {
1338 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1339 lp_build_const_int32(gallivm, i) , "");
1340 ret = LLVMBuildOr(builder, ret, temp, "");
1341 }
1342 return ret;
1343 }
1344
1345 static LLVMValueRef
1346 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1347 struct lp_build_tgsi_context * bld_base,
1348 boolean is_vindex_indirect,
1349 LLVMValueRef vertex_index,
1350 boolean is_aindex_indirect,
1351 LLVMValueRef attrib_index,
1352 LLVMValueRef swizzle_index)
1353 {
1354 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1355 struct gallivm_state *gallivm = bld_base->base.gallivm;
1356 LLVMBuilderRef builder = gallivm->builder;
1357 LLVMValueRef indices[3];
1358 LLVMValueRef res;
1359 struct lp_type type = bld_base->base.type;
1360
1361 if (is_vindex_indirect || is_aindex_indirect) {
1362 int i;
1363 res = bld_base->base.zero;
1364 for (i = 0; i < type.length; ++i) {
1365 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1366 LLVMValueRef vert_chan_index = vertex_index;
1367 LLVMValueRef attr_chan_index = attrib_index;
1368 LLVMValueRef channel_vec, value;
1369
1370 if (is_vindex_indirect) {
1371 vert_chan_index = LLVMBuildExtractElement(builder,
1372 vertex_index, idx, "");
1373 }
1374 if (is_aindex_indirect) {
1375 attr_chan_index = LLVMBuildExtractElement(builder,
1376 attrib_index, idx, "");
1377 }
1378
1379 indices[0] = vert_chan_index;
1380 indices[1] = attr_chan_index;
1381 indices[2] = swizzle_index;
1382
1383 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1384 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1385 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1386
1387 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1388 }
1389 } else {
1390 indices[0] = vertex_index;
1391 indices[1] = attrib_index;
1392 indices[2] = swizzle_index;
1393
1394 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1395 res = LLVMBuildLoad(builder, res, "");
1396 }
1397
1398 return res;
1399 }
1400
1401 static void
1402 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1403 struct lp_build_tgsi_context * bld_base,
1404 LLVMValueRef (*outputs)[4],
1405 LLVMValueRef emitted_vertices_vec)
1406 {
1407 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1408 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1409 struct gallivm_state *gallivm = variant->gallivm;
1410 LLVMBuilderRef builder = gallivm->builder;
1411 struct lp_type gs_type = bld_base->base.type;
1412 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1413 lp_int_type(gs_type), 0);
1414 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1415 LLVMValueRef next_prim_offset =
1416 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1417 LLVMValueRef io = variant->io_ptr;
1418 unsigned i;
1419 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1420
1421 for (i = 0; i < gs_type.length; ++i) {
1422 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1423 LLVMValueRef currently_emitted =
1424 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1425 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1426 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1427 }
1428
1429 convert_to_aos(gallivm, io, indices,
1430 outputs, clipmask,
1431 gs_info->num_outputs, gs_type,
1432 FALSE);
1433 }
1434
1435 static void
1436 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1437 struct lp_build_tgsi_context * bld_base,
1438 LLVMValueRef verts_per_prim_vec,
1439 LLVMValueRef emitted_prims_vec)
1440 {
1441 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1442 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1443 struct gallivm_state *gallivm = variant->gallivm;
1444 LLVMBuilderRef builder = gallivm->builder;
1445 LLVMValueRef prim_lengts_ptr =
1446 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1447 unsigned i;
1448
1449 for (i = 0; i < bld_base->base.type.length; ++i) {
1450 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1451 LLVMValueRef prims_emitted =
1452 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1453 LLVMValueRef store_ptr;
1454 LLVMValueRef num_vertices =
1455 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1456
1457 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1458 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1459 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1460 LLVMBuildStore(builder, num_vertices, store_ptr);
1461 }
1462 }
1463
1464 static void
1465 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1466 struct lp_build_tgsi_context * bld_base,
1467 LLVMValueRef total_emitted_vertices_vec,
1468 LLVMValueRef emitted_prims_vec)
1469 {
1470 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1471 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1472 struct gallivm_state *gallivm = variant->gallivm;
1473 LLVMBuilderRef builder = gallivm->builder;
1474 LLVMValueRef emitted_verts_ptr =
1475 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1476 LLVMValueRef emitted_prims_ptr =
1477 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1478 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1479
1480 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1481 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1482
1483 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1484 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1485 }
1486
1487 static void
1488 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1489 boolean elts)
1490 {
1491 struct gallivm_state *gallivm = variant->gallivm;
1492 LLVMContextRef context = gallivm->context;
1493 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1494 LLVMTypeRef arg_types[11];
1495 unsigned num_arg_types =
1496 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1497 LLVMTypeRef func_type;
1498 LLVMValueRef context_ptr;
1499 LLVMBasicBlockRef block;
1500 LLVMBuilderRef builder;
1501 char func_name[64];
1502 struct lp_type vs_type;
1503 LLVMValueRef end, start;
1504 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1505 LLVMValueRef vertex_id_offset, start_instance;
1506 LLVMValueRef stride, step, io_itr;
1507 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1508 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1509 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1510 struct draw_context *draw = llvm->draw;
1511 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1512 unsigned i, j;
1513 struct lp_build_context bld;
1514 struct lp_build_loop_state lp_loop;
1515 const int vector_length = lp_native_vector_width / 32;
1516 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1517 LLVMValueRef fetch_max;
1518 struct lp_build_sampler_soa *sampler = 0;
1519 LLVMValueRef ret, clipmask_bool_ptr;
1520 struct draw_llvm_variant_key *key = &variant->key;
1521 /* If geometry shader is present we need to skip both the viewport
1522 * transformation and clipping otherwise the inputs to the geometry
1523 * shader will be incorrect.
1524 */
1525 const boolean bypass_viewport = key->has_gs || key->bypass_viewport;
1526 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1527 key->clip_z ||
1528 key->clip_user);
1529 LLVMValueRef variant_func;
1530 const unsigned pos = llvm->draw->vs.position_output;
1531 const unsigned cv = llvm->draw->vs.clipvertex_output;
1532 boolean have_clipdist = FALSE;
1533 struct lp_bld_tgsi_system_values system_values;
1534
1535 memset(&system_values, 0, sizeof(system_values));
1536
1537 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1538 variant->shader->variants_cached, elts ? "elts" : "linear");
1539
1540 i = 0;
1541 arg_types[i++] = get_context_ptr_type(variant); /* context */
1542 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1543 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1544 if (elts) {
1545 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1546 arg_types[i++] = int32_type; /* fetch_elt_max */
1547 } else
1548 arg_types[i++] = int32_type; /* start */
1549 arg_types[i++] = int32_type; /* fetch_count / count */
1550 arg_types[i++] = int32_type; /* stride */
1551 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1552 arg_types[i++] = int32_type; /* instance_id */
1553 arg_types[i++] = int32_type; /* vertex_id_offset */
1554 arg_types[i++] = int32_type; /* start_instance */
1555
1556 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1557
1558 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1559
1560 if (elts)
1561 variant->function_elts = variant_func;
1562 else
1563 variant->function = variant_func;
1564
1565 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1566 for (i = 0; i < num_arg_types; ++i)
1567 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1568 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1569 LLVMNoAliasAttribute);
1570
1571 context_ptr = LLVMGetParam(variant_func, 0);
1572 io_ptr = LLVMGetParam(variant_func, 1);
1573 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1574 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1575 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1576 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1577 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1578 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1579
1580 lp_build_name(context_ptr, "context");
1581 lp_build_name(io_ptr, "io");
1582 lp_build_name(vbuffers_ptr, "vbuffers");
1583 lp_build_name(stride, "stride");
1584 lp_build_name(vb_ptr, "vb");
1585 lp_build_name(system_values.instance_id, "instance_id");
1586 lp_build_name(vertex_id_offset, "vertex_id_offset");
1587 lp_build_name(start_instance, "start_instance");
1588
1589 if (elts) {
1590 fetch_elts = LLVMGetParam(variant_func, 3);
1591 fetch_elt_max = LLVMGetParam(variant_func, 4);
1592 fetch_count = LLVMGetParam(variant_func, 5);
1593 lp_build_name(fetch_elts, "fetch_elts");
1594 lp_build_name(fetch_elt_max, "fetch_elt_max");
1595 lp_build_name(fetch_count, "fetch_count");
1596 start = count = NULL;
1597 }
1598 else {
1599 start = LLVMGetParam(variant_func, 3);
1600 count = LLVMGetParam(variant_func, 4);
1601 lp_build_name(start, "start");
1602 lp_build_name(count, "count");
1603 fetch_elts = fetch_count = NULL;
1604 }
1605
1606 /*
1607 * Function body
1608 */
1609
1610 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1611 builder = gallivm->builder;
1612 LLVMPositionBuilderAtEnd(builder, block);
1613
1614 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1615
1616 memset(&vs_type, 0, sizeof vs_type);
1617 vs_type.floating = TRUE; /* floating point values */
1618 vs_type.sign = TRUE; /* values are signed */
1619 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1620 vs_type.width = 32; /* 32-bit float */
1621 vs_type.length = vector_length;
1622
1623 /* hold temporary "bool" clipmask */
1624 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1625 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1626
1627 /* code generated texture sampling */
1628 sampler = draw_llvm_sampler_soa_create(
1629 draw_llvm_variant_key_samplers(key),
1630 context_ptr);
1631
1632 if (elts) {
1633 start = zero;
1634 end = fetch_count;
1635 count = fetch_count;
1636 }
1637 else {
1638 end = lp_build_add(&bld, start, count);
1639 }
1640
1641 step = lp_build_const_int32(gallivm, vector_length);
1642
1643 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1644
1645 lp_build_loop_begin(&lp_loop, gallivm, zero);
1646 {
1647 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1648 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1649 LLVMValueRef io;
1650 LLVMValueRef clipmask; /* holds the clipmask value */
1651 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1652
1653 io_itr = lp_loop.counter;
1654
1655 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1656 #if DEBUG_STORE
1657 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1658 io_itr, io, lp_loop.counter);
1659 #endif
1660 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1661 for (i = 0; i < vector_length; ++i) {
1662 LLVMValueRef vert_index =
1663 LLVMBuildAdd(builder,
1664 lp_loop.counter,
1665 lp_build_const_int32(gallivm, i), "");
1666 LLVMValueRef true_index =
1667 LLVMBuildAdd(builder, start, vert_index, "");
1668 LLVMValueRef vertex_id;
1669
1670 /* make sure we're not out of bounds which can happen
1671 * if fetch_count % 4 != 0, because on the last iteration
1672 * a few of the 4 vertex fetches will be out of bounds */
1673 true_index = lp_build_min(&bld, true_index, fetch_max);
1674
1675 if (elts) {
1676 LLVMValueRef fetch_ptr;
1677 LLVMValueRef index_overflowed;
1678 LLVMValueRef index_ptr =
1679 lp_build_alloca(
1680 gallivm,
1681 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1682 struct lp_build_if_state if_ctx;
1683 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1684 true_index, fetch_elt_max,
1685 "index_overflowed");
1686
1687 lp_build_if(&if_ctx, gallivm, index_overflowed);
1688 {
1689 /* Generate maximum possible index so that
1690 * generate_fetch can treat it just like
1691 * any other overflow and return zeros.
1692 * We don't have to worry about the restart
1693 * primitive index because it has already been
1694 * handled
1695 */
1696 LLVMValueRef val =
1697 lp_build_const_int32(gallivm, 0xffffffff);
1698 LLVMBuildStore(builder, val, index_ptr);
1699 }
1700 lp_build_else(&if_ctx);
1701 {
1702 LLVMValueRef val;
1703 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1704 &true_index, 1, "");
1705 val = LLVMBuildLoad(builder, fetch_ptr, "");
1706 LLVMBuildStore(builder, val, index_ptr);
1707 }
1708 lp_build_endif(&if_ctx);
1709 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1710 }
1711 /* in the paths with elts vertex id has to be unaffected by the
1712 * index bias and because indices inside our elements array have
1713 * already had index bias applied we need to subtract it here to
1714 * get back to the original index.
1715 * in the linear paths vertex id has to be unaffected by the
1716 * original start index and because we abuse the 'start' variable
1717 * to either represent the actual start index or the index at which
1718 * the primitive was split (we split rendering into chunks of at
1719 * most 4095-vertices) we need to back out the original start
1720 * index out of our vertex id here.
1721 */
1722 vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
1723
1724 system_values.vertex_id = LLVMBuildInsertElement(
1725 gallivm->builder,
1726 system_values.vertex_id, vertex_id,
1727 lp_build_const_int32(gallivm, i), "");
1728
1729 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1730 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1731 LLVMValueRef vb_index =
1732 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1733 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1734 generate_fetch(gallivm, draw, vbuffers_ptr,
1735 &aos_attribs[j][i], velem, vb, true_index,
1736 system_values.instance_id, start_instance);
1737 }
1738 }
1739 convert_to_soa(gallivm, aos_attribs, inputs,
1740 draw->pt.nr_vertex_elements, vs_type);
1741
1742 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1743 generate_vs(variant,
1744 builder,
1745 vs_type,
1746 outputs,
1747 ptr_aos,
1748 &system_values,
1749 context_ptr,
1750 sampler,
1751 key->clamp_vertex_color);
1752
1753 if (pos != -1 && cv != -1) {
1754 /* store original positions in clip before further manipulation */
1755 store_clip(gallivm, vs_type, io, outputs, FALSE, key->clip_user ? cv : pos);
1756 store_clip(gallivm, vs_type, io, outputs, TRUE, pos);
1757
1758 /* do cliptest */
1759 if (enable_cliptest) {
1760 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1761 /* allocate clipmask, assign it integer type */
1762 clipmask = generate_clipmask(llvm,
1763 gallivm,
1764 vs_type,
1765 outputs,
1766 key->clip_xy,
1767 key->clip_z,
1768 key->clip_user,
1769 key->clip_halfz,
1770 key->ucp_enable,
1771 context_ptr, &have_clipdist);
1772 temp = LLVMBuildOr(builder, clipmask, temp, "");
1773 /* store temporary clipping boolean value */
1774 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1775 }
1776 else {
1777 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1778 }
1779
1780 /* do viewport mapping */
1781 if (!bypass_viewport) {
1782 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1783 }
1784 }
1785 else {
1786 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1787 }
1788
1789 /* store clipmask in vertex header,
1790 * original positions in clip
1791 * and transformed positions in data
1792 */
1793 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1794 vs_info->num_outputs, vs_type,
1795 have_clipdist);
1796 }
1797 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1798
1799 sampler->destroy(sampler);
1800
1801 /* return clipping boolean value for function */
1802 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1803
1804 LLVMBuildRet(builder, ret);
1805
1806 gallivm_verify_function(gallivm, variant_func);
1807 }
1808
1809
1810 struct draw_llvm_variant_key *
1811 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1812 {
1813 unsigned i;
1814 struct draw_llvm_variant_key *key;
1815 struct draw_sampler_static_state *draw_sampler;
1816
1817 key = (struct draw_llvm_variant_key *)store;
1818
1819 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1820
1821 /* Presumably all variants of the shader should have the same
1822 * number of vertex elements - ie the number of shader inputs.
1823 * NOTE: we NEED to store the needed number of needed inputs
1824 * here, not the number of provided elements to match keysize
1825 * (and the offset of sampler state in the key).
1826 */
1827 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1828 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1829
1830 /* will have to rig this up properly later */
1831 key->clip_xy = llvm->draw->clip_xy;
1832 key->clip_z = llvm->draw->clip_z;
1833 key->clip_user = llvm->draw->clip_user;
1834 key->bypass_viewport = llvm->draw->identity_viewport;
1835 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1836 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1837 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1838 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1839 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1840 key->pad1 = 0;
1841
1842 /* All variants of this shader will have the same value for
1843 * nr_samplers. Not yet trying to compact away holes in the
1844 * sampler array.
1845 */
1846 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1847 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1848 key->nr_sampler_views =
1849 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1850 }
1851 else {
1852 key->nr_sampler_views = key->nr_samplers;
1853 }
1854
1855 draw_sampler = draw_llvm_variant_key_samplers(key);
1856
1857 memcpy(key->vertex_element,
1858 llvm->draw->pt.vertex_element,
1859 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1860
1861 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1862
1863 for (i = 0 ; i < key->nr_samplers; i++) {
1864 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1865 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1866 }
1867 for (i = 0 ; i < key->nr_sampler_views; i++) {
1868 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1869 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1870 }
1871
1872 return key;
1873 }
1874
1875
1876 void
1877 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1878 {
1879 unsigned i;
1880 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1881
1882 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1883 debug_printf("clip_xy = %u\n", key->clip_xy);
1884 debug_printf("clip_z = %u\n", key->clip_z);
1885 debug_printf("clip_user = %u\n", key->clip_user);
1886 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1887 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1888 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1889 debug_printf("has_gs = %u\n", key->has_gs);
1890 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1891
1892 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1893 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1894 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1895 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1896 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1897 }
1898
1899 for (i = 0 ; i < key->nr_sampler_views; i++) {
1900 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1901 }
1902 }
1903
1904
1905 void
1906 draw_llvm_set_mapped_texture(struct draw_context *draw,
1907 unsigned shader_stage,
1908 unsigned sview_idx,
1909 uint32_t width, uint32_t height, uint32_t depth,
1910 uint32_t first_level, uint32_t last_level,
1911 const void *base_ptr,
1912 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1913 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1914 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1915 {
1916 unsigned j;
1917 struct draw_jit_texture *jit_tex;
1918
1919 assert(shader_stage == PIPE_SHADER_VERTEX ||
1920 shader_stage == PIPE_SHADER_GEOMETRY);
1921
1922 if (shader_stage == PIPE_SHADER_VERTEX) {
1923 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1924
1925 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1926 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1927 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1928
1929 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1930 } else {
1931 assert(0);
1932 return;
1933 }
1934
1935 jit_tex->width = width;
1936 jit_tex->height = height;
1937 jit_tex->depth = depth;
1938 jit_tex->first_level = first_level;
1939 jit_tex->last_level = last_level;
1940 jit_tex->base = base_ptr;
1941
1942 for (j = first_level; j <= last_level; j++) {
1943 jit_tex->mip_offsets[j] = mip_offsets[j];
1944 jit_tex->row_stride[j] = row_stride[j];
1945 jit_tex->img_stride[j] = img_stride[j];
1946 }
1947 }
1948
1949
1950 void
1951 draw_llvm_set_sampler_state(struct draw_context *draw,
1952 unsigned shader_type)
1953 {
1954 unsigned i;
1955
1956 if (shader_type == PIPE_SHADER_VERTEX) {
1957 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1958 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1959
1960 if (draw->samplers[i]) {
1961 const struct pipe_sampler_state *s
1962 = draw->samplers[PIPE_SHADER_VERTEX][i];
1963 jit_sam->min_lod = s->min_lod;
1964 jit_sam->max_lod = s->max_lod;
1965 jit_sam->lod_bias = s->lod_bias;
1966 COPY_4V(jit_sam->border_color, s->border_color.f);
1967 }
1968 }
1969 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1970 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1971 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1972
1973 if (draw->samplers[i]) {
1974 const struct pipe_sampler_state *s
1975 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1976 jit_sam->min_lod = s->min_lod;
1977 jit_sam->max_lod = s->max_lod;
1978 jit_sam->lod_bias = s->lod_bias;
1979 COPY_4V(jit_sam->border_color, s->border_color.f);
1980 }
1981 }
1982 }
1983 }
1984
1985
1986 void
1987 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1988 {
1989 struct draw_llvm *llvm = variant->llvm;
1990
1991 gallivm_destroy(variant->gallivm);
1992
1993 remove_from_list(&variant->list_item_local);
1994 variant->shader->variants_cached--;
1995 remove_from_list(&variant->list_item_global);
1996 llvm->nr_variants--;
1997 FREE(variant);
1998 }
1999
2000
2001 /**
2002 * Create LLVM types for various structures.
2003 */
2004 static void
2005 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2006 {
2007 struct gallivm_state *gallivm = var->gallivm;
2008 LLVMTypeRef texture_type, sampler_type, context_type;
2009
2010 texture_type = create_jit_texture_type(gallivm, "texture");
2011 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2012
2013 context_type = create_gs_jit_context_type(gallivm,
2014 var->shader->base.vector_length,
2015 texture_type, sampler_type,
2016 "draw_gs_jit_context");
2017 var->context_ptr_type = LLVMPointerType(context_type, 0);
2018
2019 var->input_array_type = create_gs_jit_input_type(gallivm);
2020 }
2021
2022 static LLVMTypeRef
2023 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2024 {
2025 if (!variant->context_ptr_type)
2026 create_gs_jit_types(variant);
2027 return variant->context_ptr_type;
2028 }
2029
2030 static LLVMValueRef
2031 generate_mask_value(struct draw_gs_llvm_variant *variant,
2032 struct lp_type gs_type)
2033 {
2034 struct gallivm_state *gallivm = variant->gallivm;
2035 LLVMBuilderRef builder = gallivm->builder;
2036 struct lp_type mask_type = lp_int_type(gs_type);
2037 LLVMValueRef num_prims;
2038 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2039 unsigned i;
2040
2041 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2042 variant->num_prims);
2043 for (i = 0; i <= gs_type.length; i++) {
2044 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2045 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2046 }
2047 mask_val = lp_build_compare(gallivm, mask_type,
2048 PIPE_FUNC_GREATER, num_prims, mask_val);
2049
2050 return mask_val;
2051 }
2052
2053 static void
2054 draw_gs_llvm_generate(struct draw_llvm *llvm,
2055 struct draw_gs_llvm_variant *variant)
2056 {
2057 struct gallivm_state *gallivm = variant->gallivm;
2058 LLVMContextRef context = gallivm->context;
2059 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2060 LLVMTypeRef arg_types[6];
2061 LLVMTypeRef func_type;
2062 LLVMValueRef variant_func;
2063 LLVMValueRef context_ptr;
2064 LLVMValueRef prim_id_ptr;
2065 LLVMBasicBlockRef block;
2066 LLVMBuilderRef builder;
2067 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2068 struct lp_build_sampler_soa *sampler = 0;
2069 struct lp_build_context bld;
2070 struct lp_bld_tgsi_system_values system_values;
2071 char func_name[64];
2072 struct lp_type gs_type;
2073 unsigned i;
2074 struct draw_gs_llvm_iface gs_iface;
2075 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2076 LLVMValueRef consts_ptr, num_consts_ptr;
2077 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2078 struct lp_build_mask_context mask;
2079 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2080 unsigned vector_length = variant->shader->base.vector_length;
2081
2082 memset(&system_values, 0, sizeof(system_values));
2083
2084 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2085 variant->shader->variants_cached);
2086
2087 assert(variant->vertex_header_ptr_type);
2088
2089 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2090 arg_types[1] = variant->input_array_type; /* input */
2091 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2092 arg_types[3] = int32_type; /* num_prims */
2093 arg_types[4] = int32_type; /* instance_id */
2094 arg_types[5] = LLVMPointerType(
2095 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2096
2097 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2098
2099 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2100
2101 variant->function = variant_func;
2102
2103 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2104
2105 for (i = 0; i < Elements(arg_types); ++i)
2106 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2107 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2108 LLVMNoAliasAttribute);
2109
2110 context_ptr = LLVMGetParam(variant_func, 0);
2111 input_array = LLVMGetParam(variant_func, 1);
2112 io_ptr = LLVMGetParam(variant_func, 2);
2113 num_prims = LLVMGetParam(variant_func, 3);
2114 system_values.instance_id = LLVMGetParam(variant_func, 4);
2115 prim_id_ptr = LLVMGetParam(variant_func, 5);
2116
2117 lp_build_name(context_ptr, "context");
2118 lp_build_name(input_array, "input");
2119 lp_build_name(io_ptr, "io");
2120 lp_build_name(num_prims, "num_prims");
2121 lp_build_name(system_values.instance_id, "instance_id");
2122 lp_build_name(prim_id_ptr, "prim_id_ptr");
2123
2124 variant->context_ptr = context_ptr;
2125 variant->io_ptr = io_ptr;
2126 variant->num_prims = num_prims;
2127
2128 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2129 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2130 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2131 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2132 gs_iface.input = input_array;
2133 gs_iface.variant = variant;
2134
2135 /*
2136 * Function body
2137 */
2138
2139 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2140 builder = gallivm->builder;
2141 LLVMPositionBuilderAtEnd(builder, block);
2142
2143 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2144
2145 memset(&gs_type, 0, sizeof gs_type);
2146 gs_type.floating = TRUE; /* floating point values */
2147 gs_type.sign = TRUE; /* values are signed */
2148 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2149 gs_type.width = 32; /* 32-bit float */
2150 gs_type.length = vector_length;
2151
2152 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2153 num_consts_ptr =
2154 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2155
2156 /* code generated texture sampling */
2157 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2158 context_ptr);
2159
2160 mask_val = generate_mask_value(variant, gs_type);
2161 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2162
2163 if (gs_info->uses_primid) {
2164 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2165 }
2166
2167 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2168 tgsi_dump(tokens, 0);
2169 draw_gs_llvm_dump_variant_key(&variant->key);
2170 }
2171
2172 lp_build_tgsi_soa(variant->gallivm,
2173 tokens,
2174 gs_type,
2175 &mask,
2176 consts_ptr,
2177 num_consts_ptr,
2178 &system_values,
2179 NULL,
2180 outputs,
2181 sampler,
2182 &llvm->draw->gs.geometry_shader->info,
2183 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2184
2185 sampler->destroy(sampler);
2186
2187 lp_build_mask_end(&mask);
2188
2189 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2190
2191 gallivm_verify_function(gallivm, variant_func);
2192 }
2193
2194
2195 struct draw_gs_llvm_variant *
2196 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2197 unsigned num_outputs,
2198 const struct draw_gs_llvm_variant_key *key)
2199 {
2200 struct draw_gs_llvm_variant *variant;
2201 struct llvm_geometry_shader *shader =
2202 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2203 LLVMTypeRef vertex_header;
2204 char module_name[64];
2205
2206 variant = MALLOC(sizeof *variant +
2207 shader->variant_key_size -
2208 sizeof variant->key);
2209 if (variant == NULL)
2210 return NULL;
2211
2212 variant->llvm = llvm;
2213 variant->shader = shader;
2214
2215 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2216 variant->shader->variants_cached);
2217
2218 variant->gallivm = gallivm_create(module_name, llvm->context);
2219
2220 create_gs_jit_types(variant);
2221
2222 memcpy(&variant->key, key, shader->variant_key_size);
2223
2224 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2225
2226 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2227
2228 draw_gs_llvm_generate(llvm, variant);
2229
2230 gallivm_compile_module(variant->gallivm);
2231
2232 variant->jit_func = (draw_gs_jit_func)
2233 gallivm_jit_function(variant->gallivm, variant->function);
2234
2235 gallivm_free_ir(variant->gallivm);
2236
2237 variant->list_item_global.base = variant;
2238 variant->list_item_local.base = variant;
2239 /*variant->no = */shader->variants_created++;
2240 variant->list_item_global.base = variant;
2241
2242 return variant;
2243 }
2244
2245 void
2246 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2247 {
2248 struct draw_llvm *llvm = variant->llvm;
2249
2250 gallivm_destroy(variant->gallivm);
2251
2252 remove_from_list(&variant->list_item_local);
2253 variant->shader->variants_cached--;
2254 remove_from_list(&variant->list_item_global);
2255 llvm->nr_gs_variants--;
2256 FREE(variant);
2257 }
2258
2259 struct draw_gs_llvm_variant_key *
2260 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2261 {
2262 unsigned i;
2263 struct draw_gs_llvm_variant_key *key;
2264 struct draw_sampler_static_state *draw_sampler;
2265
2266 key = (struct draw_gs_llvm_variant_key *)store;
2267
2268 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2269
2270 /* All variants of this shader will have the same value for
2271 * nr_samplers. Not yet trying to compact away holes in the
2272 * sampler array.
2273 */
2274 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2275 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2276 key->nr_sampler_views =
2277 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2278 }
2279 else {
2280 key->nr_sampler_views = key->nr_samplers;
2281 }
2282
2283 draw_sampler = key->samplers;
2284
2285 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2286
2287 for (i = 0 ; i < key->nr_samplers; i++) {
2288 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2289 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2290 }
2291 for (i = 0 ; i < key->nr_sampler_views; i++) {
2292 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2293 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2294 }
2295
2296 return key;
2297 }
2298
2299 void
2300 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2301 {
2302 unsigned i;
2303 struct draw_sampler_static_state *sampler = key->samplers;
2304
2305 for (i = 0 ; i < key->nr_sampler_views; i++) {
2306 debug_printf("sampler[%i].src_format = %s\n", i,
2307 util_format_name(sampler[i].texture_state.format));
2308 }
2309 }