projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw,gallivm,llvmpipe: Avoid implicit casts of 32-bit shifts to 64-bits.
[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 FALSE,
746 map_ptr,
747 zero, zero, zero);
748 LLVMBuildStore(builder, val, temp_ptr);
749 }
750 lp_build_endif(&if_ctx);
751
752 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
753 }
754
755 static void
756 convert_to_soa(struct gallivm_state *gallivm,
757 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
758 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
759 unsigned num_attribs, const struct lp_type soa_type)
760 {
761 unsigned i, j, k;
762 struct lp_type aos_channel_type = soa_type;
763
764 debug_assert(TGSI_NUM_CHANNELS == 4);
765 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
766
767 aos_channel_type.length >>= 1;
768
769 for (i = 0; i < num_attribs; ++i) {
770 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
771 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
772
773 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
774 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
775
776 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
777
778 for (k = 0; k < pixels_per_channel; ++k) {
779 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
780 }
781
782 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
783 }
784
785 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
786 }
787 }
788
789
790 static void
791 store_aos(struct gallivm_state *gallivm,
792 LLVMValueRef io_ptr,
793 LLVMValueRef index,
794 LLVMValueRef value)
795 {
796 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
797 LLVMBuilderRef builder = gallivm->builder;
798 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
799 LLVMValueRef indices[3];
800
801 indices[0] = lp_build_const_int32(gallivm, 0);
802 indices[1] = index;
803 indices[2] = lp_build_const_int32(gallivm, 0);
804
805 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
806 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
807
808 #if DEBUG_STORE
809 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
810 #endif
811
812 /* Unaligned store due to the vertex header */
813 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
814 }
815
816 /**
817 * Adjust the mask to architecture endianess. The mask will the store in struct:
818 *
819 * struct vertex_header {
820 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
821 * unsigned edgeflag:1;
822 * unsigned have_clipdist:1;
823 * unsigned vertex_id:16;
824 * [...]
825 * }
826 *
827 * On little-endian machine nothing needs to done, however on bit-endian machine
828 * the mask's fields need to be adjusted with the algorithm:
829 *
830 * uint32_t reverse (uint32_t x)
831 * {
832 * return (x >> 16) | // vertex_id
833 * ((x & 0x3fff) << 18) | // clipmask
834 * ((x & 0x4000) << 3) | // have_clipdist
835 * ((x & 0x8000) << 1); // edgeflag
836 * }
837 */
838 static LLVMValueRef
839 adjust_mask(struct gallivm_state *gallivm,
840 LLVMValueRef mask)
841 {
842 #ifdef PIPE_ARCH_BIG_ENDIAN
843 LLVMBuilderRef builder = gallivm->builder;
844 LLVMValueRef vertex_id;
845 LLVMValueRef clipmask;
846 LLVMValueRef have_clipdist;
847 LLVMValueRef edgeflag;
848
849 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
850 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
851 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
852 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
853 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
854 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
855 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
856
857 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
858 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
859 mask = LLVMBuildOr(builder, mask, edgeflag, "");
860 #endif
861 return mask;
862 }
863
864 static void
865 store_aos_array(struct gallivm_state *gallivm,
866 struct lp_type soa_type,
867 LLVMValueRef io_ptr,
868 LLVMValueRef *indices,
869 LLVMValueRef* aos,
870 int attrib,
871 int num_outputs,
872 LLVMValueRef clipmask,
873 boolean have_clipdist)
874 {
875 LLVMBuilderRef builder = gallivm->builder;
876 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
877 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
878 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
879 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
880 int vector_length = soa_type.length;
881 int i;
882
883 debug_assert(TGSI_NUM_CHANNELS == 4);
884
885 for (i = 0; i < vector_length; i++) {
886 linear_inds[i] = lp_build_const_int32(gallivm, i);
887 if (indices) {
888 inds[i] = indices[i];
889 } else {
890 inds[i] = linear_inds[i];
891 }
892 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
893 }
894
895 if (attrib == 0) {
896 /* store vertex header for each of the n vertices */
897 LLVMValueRef val, cliptmp;
898 int vertex_id_pad_edgeflag;
899
900 /* If this assertion fails, it means we need to update the bit twidding
901 * code here. See struct vertex_header in draw_private.h.
902 */
903 assert(DRAW_TOTAL_CLIP_PLANES==14);
904 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
905 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
906 if (have_clipdist)
907 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
908 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
909 /* OR with the clipmask */
910 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
911 for (i = 0; i < vector_length; i++) {
912 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
913 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
914 val = adjust_mask(gallivm, val);
915 #if DEBUG_STORE
916 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
917 io_ptrs[i], inds[i], val);
918 #endif
919 LLVMBuildStore(builder, val, id_ptr);
920 }
921 }
922
923 /* store for each of the n vertices */
924 for (i = 0; i < vector_length; i++) {
925 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
926 }
927 }
928
929
930 static void
931 convert_to_aos(struct gallivm_state *gallivm,
932 LLVMValueRef io,
933 LLVMValueRef *indices,
934 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
935 LLVMValueRef clipmask,
936 int num_outputs,
937 struct lp_type soa_type,
938 boolean have_clipdist)
939 {
940 LLVMBuilderRef builder = gallivm->builder;
941 unsigned chan, attrib, i;
942
943 #if DEBUG_STORE
944 lp_build_printf(gallivm, " # storing begin\n");
945 #endif
946 for (attrib = 0; attrib < num_outputs; ++attrib) {
947 LLVMValueRef soa[TGSI_NUM_CHANNELS];
948 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
949 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
950 if (outputs[attrib][chan]) {
951 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
952 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
953 #if DEBUG_STORE
954 lp_build_printf(gallivm, "output %d : %d ",
955 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
956 attrib, 0),
957 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
958 chan, 0));
959 lp_build_print_value(gallivm, "val = ", out);
960 {
961 LLVMValueRef iv =
962 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
963
964 lp_build_print_value(gallivm, " ival = ", iv);
965 }
966 #endif
967 soa[chan] = out;
968 }
969 else {
970 soa[chan] = 0;
971 }
972 }
973
974
975 if (soa_type.length == TGSI_NUM_CHANNELS) {
976 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
977 } else {
978 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
979
980 for (i = 0; i < soa_type.length; ++i) {
981 aos[i] = lp_build_extract_range(gallivm,
982 soa[i % TGSI_NUM_CHANNELS],
983 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
984 TGSI_NUM_CHANNELS);
985 }
986 }
987
988 store_aos_array(gallivm,
989 soa_type,
990 io, indices,
991 aos,
992 attrib,
993 num_outputs,
994 clipmask, have_clipdist);
995 }
996 #if DEBUG_STORE
997 lp_build_printf(gallivm, " # storing end\n");
998 #endif
999 }
1000
1001
1002 /**
1003 * Stores original vertex positions in clip coordinates
1004 */
1005 static void
1006 store_clip(struct gallivm_state *gallivm,
1007 const struct lp_type vs_type,
1008 LLVMValueRef io_ptr,
1009 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1010 boolean pre_clip_pos, int idx)
1011 {
1012 LLVMBuilderRef builder = gallivm->builder;
1013 LLVMValueRef soa[4];
1014 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1015 LLVMValueRef indices[2];
1016 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1017 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1018 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1019 LLVMTypeRef clip_ptr_type =
1020 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1021 4), 0);
1022 int i, j;
1023
1024 indices[0] =
1025 indices[1] = lp_build_const_int32(gallivm, 0);
1026
1027 for (i = 0; i < vs_type.length; i++) {
1028 inds[i] = lp_build_const_int32(gallivm, i);
1029 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1030 }
1031
1032 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1033 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1034 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1035 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1036
1037 if (!pre_clip_pos) {
1038 for (i = 0; i < vs_type.length; i++) {
1039 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1040 }
1041 } else {
1042 for (i = 0; i < vs_type.length; i++) {
1043 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1044 }
1045 }
1046
1047 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1048 for (i = 0; i < vs_type.length; ++i) {
1049 aos[i] = lp_build_extract_range(gallivm,
1050 soa[i % TGSI_NUM_CHANNELS],
1051 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1052 TGSI_NUM_CHANNELS);
1053 }
1054
1055 for (j = 0; j < vs_type.length; j++) {
1056 LLVMValueRef clip_ptr;
1057
1058 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1059 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1060
1061 /* Unaligned store */
1062 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1063 }
1064 }
1065
1066
1067 /**
1068 * Transforms the outputs for viewport mapping
1069 */
1070 static void
1071 generate_viewport(struct draw_llvm_variant *variant,
1072 LLVMBuilderRef builder,
1073 struct lp_type vs_type,
1074 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1075 LLVMValueRef context_ptr)
1076 {
1077 int i;
1078 struct gallivm_state *gallivm = variant->gallivm;
1079 struct lp_type f32_type = vs_type;
1080 const unsigned pos = variant->llvm->draw->vs.position_output;
1081 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1082 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1083 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1084 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1085
1086 /* We treat pipe_viewport_state as a float array */
1087 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1088 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1089
1090 /* for 1/w convention*/
1091 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1092 LLVMBuildStore(builder, out3, outputs[pos][3]);
1093
1094 /* Viewport Mapping */
1095 for (i=0; i<3; i++) {
1096 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1097 LLVMValueRef scale;
1098 LLVMValueRef trans;
1099 LLVMValueRef scale_i;
1100 LLVMValueRef trans_i;
1101 LLVMValueRef index;
1102
1103 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1104 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1105
1106 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1107 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1108
1109 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1110 LLVMBuildLoad(builder, scale_i, "scale"));
1111 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1112 LLVMBuildLoad(builder, trans_i, "trans"));
1113
1114 /* divide by w */
1115 out = LLVMBuildFMul(builder, out, out3, "");
1116 /* mult by scale */
1117 out = LLVMBuildFMul(builder, out, scale, "");
1118 /* add translation */
1119 out = LLVMBuildFAdd(builder, out, trans, "");
1120
1121 /* store transformed outputs */
1122 LLVMBuildStore(builder, out, outputs[pos][i]);
1123 }
1124
1125 }
1126
1127
1128 /**
1129 * Returns clipmask as nxi32 bitmask for the n vertices
1130 */
1131 static LLVMValueRef
1132 generate_clipmask(struct draw_llvm *llvm,
1133 struct gallivm_state *gallivm,
1134 struct lp_type vs_type,
1135 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1136 boolean clip_xy,
1137 boolean clip_z,
1138 boolean clip_user,
1139 boolean clip_halfz,
1140 unsigned ucp_enable,
1141 LLVMValueRef context_ptr,
1142 boolean *have_clipdist)
1143 {
1144 LLVMBuilderRef builder = gallivm->builder;
1145 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1146 LLVMValueRef test, temp;
1147 LLVMValueRef zero, shift;
1148 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1149 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1150 LLVMValueRef plane1, planes, plane_ptr, sum;
1151 struct lp_type f32_type = vs_type;
1152 struct lp_type i32_type = lp_int_type(vs_type);
1153 const unsigned pos = llvm->draw->vs.position_output;
1154 const unsigned cv = llvm->draw->vs.clipvertex_output;
1155 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1156 bool have_cd = false;
1157 unsigned cd[2];
1158
1159 cd[0] = llvm->draw->vs.clipdistance_output[0];
1160 cd[1] = llvm->draw->vs.clipdistance_output[1];
1161
1162 if (cd[0] != pos || cd[1] != pos)
1163 have_cd = true;
1164
1165 if (num_written_clipdistance && !clip_user) {
1166 clip_user = true;
1167 ucp_enable = (1 << num_written_clipdistance) - 1;
1168 }
1169
1170 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1171 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1172 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1173 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1174
1175 /*
1176 * load clipvertex and position from correct locations.
1177 * if they are the same just load them once.
1178 */
1179 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1180 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1181 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1182 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1183
1184 if (clip_user && cv != pos) {
1185 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1186 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1187 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1188 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1189 } else {
1190 cv_x = pos_x;
1191 cv_y = pos_y;
1192 cv_z = pos_z;
1193 cv_w = pos_w;
1194 }
1195
1196 /* Cliptest, for hardwired planes */
1197 if (clip_xy) {
1198 /* plane 1 */
1199 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1200 temp = shift;
1201 test = LLVMBuildAnd(builder, test, temp, "");
1202 mask = test;
1203
1204 /* plane 2 */
1205 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1206 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1207 temp = LLVMBuildShl(builder, temp, shift, "");
1208 test = LLVMBuildAnd(builder, test, temp, "");
1209 mask = LLVMBuildOr(builder, mask, test, "");
1210
1211 /* plane 3 */
1212 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1213 temp = LLVMBuildShl(builder, temp, shift, "");
1214 test = LLVMBuildAnd(builder, test, temp, "");
1215 mask = LLVMBuildOr(builder, mask, test, "");
1216
1217 /* plane 4 */
1218 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1219 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1220 temp = LLVMBuildShl(builder, temp, shift, "");
1221 test = LLVMBuildAnd(builder, test, temp, "");
1222 mask = LLVMBuildOr(builder, mask, test, "");
1223 }
1224
1225 if (clip_z) {
1226 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1227 if (clip_halfz) {
1228 /* plane 5 */
1229 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1230 test = LLVMBuildAnd(builder, test, temp, "");
1231 mask = LLVMBuildOr(builder, mask, test, "");
1232 }
1233 else {
1234 /* plane 5 */
1235 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1236 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1237 test = LLVMBuildAnd(builder, test, temp, "");
1238 mask = LLVMBuildOr(builder, mask, test, "");
1239 }
1240 /* plane 6 */
1241 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1242 temp = LLVMBuildShl(builder, temp, shift, "");
1243 test = LLVMBuildAnd(builder, test, temp, "");
1244 mask = LLVMBuildOr(builder, mask, test, "");
1245 }
1246
1247 if (clip_user) {
1248 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1249 LLVMValueRef indices[3];
1250 LLVMValueRef is_nan_or_inf;
1251
1252 /* userclip planes */
1253 while (ucp_enable) {
1254 unsigned plane_idx = ffs(ucp_enable)-1;
1255 ucp_enable &= ~(1 << plane_idx);
1256 plane_idx += 6;
1257
1258 if (have_cd && num_written_clipdistance) {
1259 LLVMValueRef clipdist;
1260 int i;
1261 i = plane_idx - 6;
1262
1263 *have_clipdist = TRUE;
1264 if (i < 4) {
1265 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1266 } else {
1267 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1268 }
1269 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1270 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1271 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1272 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1273 test = LLVMBuildAnd(builder, test, temp, "");
1274 mask = LLVMBuildOr(builder, mask, test, "");
1275 } else {
1276 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1277 indices[0] = lp_build_const_int32(gallivm, 0);
1278 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1279
1280 indices[2] = lp_build_const_int32(gallivm, 0);
1281 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1282 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1283 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1284 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1285
1286 indices[2] = lp_build_const_int32(gallivm, 1);
1287 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1288 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1289 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1290 test = LLVMBuildFMul(builder, planes, cv_y, "");
1291 sum = LLVMBuildFAdd(builder, sum, test, "");
1292
1293 indices[2] = lp_build_const_int32(gallivm, 2);
1294 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1295 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1296 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1297 test = LLVMBuildFMul(builder, planes, cv_z, "");
1298 sum = LLVMBuildFAdd(builder, sum, test, "");
1299
1300 indices[2] = lp_build_const_int32(gallivm, 3);
1301 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1302 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1303 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1304 test = LLVMBuildFMul(builder, planes, cv_w, "");
1305 sum = LLVMBuildFAdd(builder, sum, test, "");
1306
1307 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1308 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1309 test = LLVMBuildAnd(builder, test, temp, "");
1310 mask = LLVMBuildOr(builder, mask, test, "");
1311 }
1312 }
1313 }
1314 return mask;
1315 }
1316
1317
1318 /**
1319 * Returns boolean if any clipping has occurred
1320 * Used zero/non-zero i32 value to represent boolean
1321 */
1322 static LLVMValueRef
1323 clipmask_booli32(struct gallivm_state *gallivm,
1324 const struct lp_type vs_type,
1325 LLVMValueRef clipmask_bool_ptr)
1326 {
1327 LLVMBuilderRef builder = gallivm->builder;
1328 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1329 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1330 LLVMValueRef ret = LLVMConstNull(int32_type);
1331 LLVMValueRef temp;
1332 int i;
1333
1334 /*
1335 * Can do this with log2(vector length) pack instructions and one extract
1336 * (as we don't actually need a or) with sse2 which would be way better.
1337 */
1338 for (i=0; i < vs_type.length; i++) {
1339 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1340 lp_build_const_int32(gallivm, i) , "");
1341 ret = LLVMBuildOr(builder, ret, temp, "");
1342 }
1343 return ret;
1344 }
1345
1346 static LLVMValueRef
1347 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1348 struct lp_build_tgsi_context * bld_base,
1349 boolean is_vindex_indirect,
1350 LLVMValueRef vertex_index,
1351 boolean is_aindex_indirect,
1352 LLVMValueRef attrib_index,
1353 LLVMValueRef swizzle_index)
1354 {
1355 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1356 struct gallivm_state *gallivm = bld_base->base.gallivm;
1357 LLVMBuilderRef builder = gallivm->builder;
1358 LLVMValueRef indices[3];
1359 LLVMValueRef res;
1360 struct lp_type type = bld_base->base.type;
1361
1362 if (is_vindex_indirect || is_aindex_indirect) {
1363 int i;
1364 res = bld_base->base.zero;
1365 for (i = 0; i < type.length; ++i) {
1366 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1367 LLVMValueRef vert_chan_index = vertex_index;
1368 LLVMValueRef attr_chan_index = attrib_index;
1369 LLVMValueRef channel_vec, value;
1370
1371 if (is_vindex_indirect) {
1372 vert_chan_index = LLVMBuildExtractElement(builder,
1373 vertex_index, idx, "");
1374 }
1375 if (is_aindex_indirect) {
1376 attr_chan_index = LLVMBuildExtractElement(builder,
1377 attrib_index, idx, "");
1378 }
1379
1380 indices[0] = vert_chan_index;
1381 indices[1] = attr_chan_index;
1382 indices[2] = swizzle_index;
1383
1384 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1385 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1386 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1387
1388 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1389 }
1390 } else {
1391 indices[0] = vertex_index;
1392 indices[1] = attrib_index;
1393 indices[2] = swizzle_index;
1394
1395 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1396 res = LLVMBuildLoad(builder, res, "");
1397 }
1398
1399 return res;
1400 }
1401
1402 static void
1403 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1404 struct lp_build_tgsi_context * bld_base,
1405 LLVMValueRef (*outputs)[4],
1406 LLVMValueRef emitted_vertices_vec)
1407 {
1408 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1409 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1410 struct gallivm_state *gallivm = variant->gallivm;
1411 LLVMBuilderRef builder = gallivm->builder;
1412 struct lp_type gs_type = bld_base->base.type;
1413 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1414 lp_int_type(gs_type), 0);
1415 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1416 LLVMValueRef next_prim_offset =
1417 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1418 LLVMValueRef io = variant->io_ptr;
1419 unsigned i;
1420 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1421
1422 for (i = 0; i < gs_type.length; ++i) {
1423 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1424 LLVMValueRef currently_emitted =
1425 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1426 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1427 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1428 }
1429
1430 convert_to_aos(gallivm, io, indices,
1431 outputs, clipmask,
1432 gs_info->num_outputs, gs_type,
1433 FALSE);
1434 }
1435
1436 static void
1437 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1438 struct lp_build_tgsi_context * bld_base,
1439 LLVMValueRef verts_per_prim_vec,
1440 LLVMValueRef emitted_prims_vec)
1441 {
1442 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1443 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1444 struct gallivm_state *gallivm = variant->gallivm;
1445 LLVMBuilderRef builder = gallivm->builder;
1446 LLVMValueRef prim_lengts_ptr =
1447 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1448 unsigned i;
1449
1450 for (i = 0; i < bld_base->base.type.length; ++i) {
1451 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1452 LLVMValueRef prims_emitted =
1453 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1454 LLVMValueRef store_ptr;
1455 LLVMValueRef num_vertices =
1456 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1457
1458 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1459 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1460 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1461 LLVMBuildStore(builder, num_vertices, store_ptr);
1462 }
1463 }
1464
1465 static void
1466 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1467 struct lp_build_tgsi_context * bld_base,
1468 LLVMValueRef total_emitted_vertices_vec,
1469 LLVMValueRef emitted_prims_vec)
1470 {
1471 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1472 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1473 struct gallivm_state *gallivm = variant->gallivm;
1474 LLVMBuilderRef builder = gallivm->builder;
1475 LLVMValueRef emitted_verts_ptr =
1476 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1477 LLVMValueRef emitted_prims_ptr =
1478 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1479 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1480
1481 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1482 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1483
1484 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1485 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1486 }
1487
1488 static void
1489 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1490 boolean elts)
1491 {
1492 struct gallivm_state *gallivm = variant->gallivm;
1493 LLVMContextRef context = gallivm->context;
1494 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1495 LLVMTypeRef arg_types[11];
1496 unsigned num_arg_types =
1497 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1498 LLVMTypeRef func_type;
1499 LLVMValueRef context_ptr;
1500 LLVMBasicBlockRef block;
1501 LLVMBuilderRef builder;
1502 char func_name[64];
1503 struct lp_type vs_type;
1504 LLVMValueRef end, start;
1505 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1506 LLVMValueRef vertex_id_offset, start_instance;
1507 LLVMValueRef stride, step, io_itr;
1508 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1509 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1510 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1511 struct draw_context *draw = llvm->draw;
1512 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1513 unsigned i, j;
1514 struct lp_build_context bld;
1515 struct lp_build_loop_state lp_loop;
1516 const int vector_length = lp_native_vector_width / 32;
1517 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1518 LLVMValueRef fetch_max;
1519 struct lp_build_sampler_soa *sampler = 0;
1520 LLVMValueRef ret, clipmask_bool_ptr;
1521 struct draw_llvm_variant_key *key = &variant->key;
1522 /* If geometry shader is present we need to skip both the viewport
1523 * transformation and clipping otherwise the inputs to the geometry
1524 * shader will be incorrect.
1525 * The code can't handle vp transform when vs writes vp index neither
1526 * (though this would be fixable here, but couldn't just broadcast
1527 * the values).
1528 */
1529 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1530 llvm->draw->vs.vertex_shader->info.writes_viewport_index;
1531 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1532 key->clip_z ||
1533 key->clip_user);
1534 LLVMValueRef variant_func;
1535 const unsigned pos = llvm->draw->vs.position_output;
1536 const unsigned cv = llvm->draw->vs.clipvertex_output;
1537 boolean have_clipdist = FALSE;
1538 struct lp_bld_tgsi_system_values system_values;
1539
1540 memset(&system_values, 0, sizeof(system_values));
1541
1542 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1543 variant->shader->variants_cached, elts ? "elts" : "linear");
1544
1545 i = 0;
1546 arg_types[i++] = get_context_ptr_type(variant); /* context */
1547 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1548 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1549 if (elts) {
1550 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1551 arg_types[i++] = int32_type; /* fetch_elt_max */
1552 } else
1553 arg_types[i++] = int32_type; /* start */
1554 arg_types[i++] = int32_type; /* fetch_count / count */
1555 arg_types[i++] = int32_type; /* stride */
1556 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1557 arg_types[i++] = int32_type; /* instance_id */
1558 arg_types[i++] = int32_type; /* vertex_id_offset */
1559 arg_types[i++] = int32_type; /* start_instance */
1560
1561 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1562
1563 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1564
1565 if (elts)
1566 variant->function_elts = variant_func;
1567 else
1568 variant->function = variant_func;
1569
1570 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1571 for (i = 0; i < num_arg_types; ++i)
1572 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1573 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1574 LLVMNoAliasAttribute);
1575
1576 context_ptr = LLVMGetParam(variant_func, 0);
1577 io_ptr = LLVMGetParam(variant_func, 1);
1578 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1579 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1580 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1581 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1582 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1583 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1584
1585 lp_build_name(context_ptr, "context");
1586 lp_build_name(io_ptr, "io");
1587 lp_build_name(vbuffers_ptr, "vbuffers");
1588 lp_build_name(stride, "stride");
1589 lp_build_name(vb_ptr, "vb");
1590 lp_build_name(system_values.instance_id, "instance_id");
1591 lp_build_name(vertex_id_offset, "vertex_id_offset");
1592 lp_build_name(start_instance, "start_instance");
1593
1594 if (elts) {
1595 fetch_elts = LLVMGetParam(variant_func, 3);
1596 fetch_elt_max = LLVMGetParam(variant_func, 4);
1597 fetch_count = LLVMGetParam(variant_func, 5);
1598 lp_build_name(fetch_elts, "fetch_elts");
1599 lp_build_name(fetch_elt_max, "fetch_elt_max");
1600 lp_build_name(fetch_count, "fetch_count");
1601 start = count = NULL;
1602 }
1603 else {
1604 start = LLVMGetParam(variant_func, 3);
1605 count = LLVMGetParam(variant_func, 4);
1606 lp_build_name(start, "start");
1607 lp_build_name(count, "count");
1608 fetch_elts = fetch_count = NULL;
1609 }
1610
1611 /*
1612 * Function body
1613 */
1614
1615 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1616 builder = gallivm->builder;
1617 LLVMPositionBuilderAtEnd(builder, block);
1618
1619 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1620
1621 memset(&vs_type, 0, sizeof vs_type);
1622 vs_type.floating = TRUE; /* floating point values */
1623 vs_type.sign = TRUE; /* values are signed */
1624 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1625 vs_type.width = 32; /* 32-bit float */
1626 vs_type.length = vector_length;
1627
1628 /* hold temporary "bool" clipmask */
1629 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1630 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1631
1632 /* code generated texture sampling */
1633 sampler = draw_llvm_sampler_soa_create(
1634 draw_llvm_variant_key_samplers(key),
1635 context_ptr);
1636
1637 if (elts) {
1638 start = zero;
1639 end = fetch_count;
1640 count = fetch_count;
1641 }
1642 else {
1643 end = lp_build_add(&bld, start, count);
1644 }
1645
1646 step = lp_build_const_int32(gallivm, vector_length);
1647
1648 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1649
1650 lp_build_loop_begin(&lp_loop, gallivm, zero);
1651 {
1652 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1653 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1654 LLVMValueRef io;
1655 LLVMValueRef clipmask; /* holds the clipmask value */
1656 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1657
1658 io_itr = lp_loop.counter;
1659
1660 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1661 #if DEBUG_STORE
1662 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1663 io_itr, io, lp_loop.counter);
1664 #endif
1665 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1666 for (i = 0; i < vector_length; ++i) {
1667 LLVMValueRef vert_index =
1668 LLVMBuildAdd(builder,
1669 lp_loop.counter,
1670 lp_build_const_int32(gallivm, i), "");
1671 LLVMValueRef true_index =
1672 LLVMBuildAdd(builder, start, vert_index, "");
1673 LLVMValueRef vertex_id;
1674
1675 /* make sure we're not out of bounds which can happen
1676 * if fetch_count % 4 != 0, because on the last iteration
1677 * a few of the 4 vertex fetches will be out of bounds */
1678 true_index = lp_build_min(&bld, true_index, fetch_max);
1679
1680 if (elts) {
1681 LLVMValueRef fetch_ptr;
1682 LLVMValueRef index_overflowed;
1683 LLVMValueRef index_ptr =
1684 lp_build_alloca(
1685 gallivm,
1686 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1687 struct lp_build_if_state if_ctx;
1688 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1689 true_index, fetch_elt_max,
1690 "index_overflowed");
1691
1692 lp_build_if(&if_ctx, gallivm, index_overflowed);
1693 {
1694 /* Generate maximum possible index so that
1695 * generate_fetch can treat it just like
1696 * any other overflow and return zeros.
1697 * We don't have to worry about the restart
1698 * primitive index because it has already been
1699 * handled
1700 */
1701 LLVMValueRef val =
1702 lp_build_const_int32(gallivm, 0xffffffff);
1703 LLVMBuildStore(builder, val, index_ptr);
1704 }
1705 lp_build_else(&if_ctx);
1706 {
1707 LLVMValueRef val;
1708 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1709 &true_index, 1, "");
1710 val = LLVMBuildLoad(builder, fetch_ptr, "");
1711 LLVMBuildStore(builder, val, index_ptr);
1712 }
1713 lp_build_endif(&if_ctx);
1714 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1715 }
1716 /* in the paths with elts vertex id has to be unaffected by the
1717 * index bias and because indices inside our elements array have
1718 * already had index bias applied we need to subtract it here to
1719 * get back to the original index.
1720 * in the linear paths vertex id has to be unaffected by the
1721 * original start index and because we abuse the 'start' variable
1722 * to either represent the actual start index or the index at which
1723 * the primitive was split (we split rendering into chunks of at
1724 * most 4095-vertices) we need to back out the original start
1725 * index out of our vertex id here.
1726 */
1727 vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
1728
1729 system_values.vertex_id = LLVMBuildInsertElement(
1730 gallivm->builder,
1731 system_values.vertex_id, vertex_id,
1732 lp_build_const_int32(gallivm, i), "");
1733
1734 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1735 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1736 LLVMValueRef vb_index =
1737 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1738 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1739 generate_fetch(gallivm, draw, vbuffers_ptr,
1740 &aos_attribs[j][i], velem, vb, true_index,
1741 system_values.instance_id, start_instance);
1742 }
1743 }
1744 convert_to_soa(gallivm, aos_attribs, inputs,
1745 draw->pt.nr_vertex_elements, vs_type);
1746
1747 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1748 generate_vs(variant,
1749 builder,
1750 vs_type,
1751 outputs,
1752 ptr_aos,
1753 &system_values,
1754 context_ptr,
1755 sampler,
1756 key->clamp_vertex_color);
1757
1758 if (pos != -1 && cv != -1) {
1759 /* store original positions in clip before further manipulation */
1760 store_clip(gallivm, vs_type, io, outputs, FALSE, key->clip_user ? cv : pos);
1761 store_clip(gallivm, vs_type, io, outputs, TRUE, pos);
1762
1763 /* do cliptest */
1764 if (enable_cliptest) {
1765 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1766 /* allocate clipmask, assign it integer type */
1767 clipmask = generate_clipmask(llvm,
1768 gallivm,
1769 vs_type,
1770 outputs,
1771 key->clip_xy,
1772 key->clip_z,
1773 key->clip_user,
1774 key->clip_halfz,
1775 key->ucp_enable,
1776 context_ptr, &have_clipdist);
1777 temp = LLVMBuildOr(builder, clipmask, temp, "");
1778 /* store temporary clipping boolean value */
1779 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1780 }
1781 else {
1782 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1783 }
1784
1785 /* do viewport mapping */
1786 if (!bypass_viewport) {
1787 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1788 }
1789 }
1790 else {
1791 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1792 }
1793
1794 /* store clipmask in vertex header,
1795 * original positions in clip
1796 * and transformed positions in data
1797 */
1798 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1799 vs_info->num_outputs, vs_type,
1800 have_clipdist);
1801 }
1802 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1803
1804 sampler->destroy(sampler);
1805
1806 /* return clipping boolean value for function */
1807 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1808
1809 LLVMBuildRet(builder, ret);
1810
1811 gallivm_verify_function(gallivm, variant_func);
1812 }
1813
1814
1815 struct draw_llvm_variant_key *
1816 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1817 {
1818 unsigned i;
1819 struct draw_llvm_variant_key *key;
1820 struct draw_sampler_static_state *draw_sampler;
1821
1822 key = (struct draw_llvm_variant_key *)store;
1823
1824 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1825
1826 /* Presumably all variants of the shader should have the same
1827 * number of vertex elements - ie the number of shader inputs.
1828 * NOTE: we NEED to store the needed number of needed inputs
1829 * here, not the number of provided elements to match keysize
1830 * (and the offset of sampler state in the key).
1831 */
1832 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1833 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1834
1835 /* will have to rig this up properly later */
1836 key->clip_xy = llvm->draw->clip_xy;
1837 key->clip_z = llvm->draw->clip_z;
1838 key->clip_user = llvm->draw->clip_user;
1839 key->bypass_viewport = llvm->draw->identity_viewport;
1840 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1841 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1842 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1843 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1844 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1845 key->pad1 = 0;
1846
1847 /* All variants of this shader will have the same value for
1848 * nr_samplers. Not yet trying to compact away holes in the
1849 * sampler array.
1850 */
1851 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1852 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1853 key->nr_sampler_views =
1854 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1855 }
1856 else {
1857 key->nr_sampler_views = key->nr_samplers;
1858 }
1859
1860 draw_sampler = draw_llvm_variant_key_samplers(key);
1861
1862 memcpy(key->vertex_element,
1863 llvm->draw->pt.vertex_element,
1864 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1865
1866 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1867
1868 for (i = 0 ; i < key->nr_samplers; i++) {
1869 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1870 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1871 }
1872 for (i = 0 ; i < key->nr_sampler_views; i++) {
1873 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1874 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1875 }
1876
1877 return key;
1878 }
1879
1880
1881 void
1882 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1883 {
1884 unsigned i;
1885 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1886
1887 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1888 debug_printf("clip_xy = %u\n", key->clip_xy);
1889 debug_printf("clip_z = %u\n", key->clip_z);
1890 debug_printf("clip_user = %u\n", key->clip_user);
1891 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1892 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1893 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1894 debug_printf("has_gs = %u\n", key->has_gs);
1895 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1896
1897 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1898 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1899 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1900 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1901 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1902 }
1903
1904 for (i = 0 ; i < key->nr_sampler_views; i++) {
1905 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1906 }
1907 }
1908
1909
1910 void
1911 draw_llvm_set_mapped_texture(struct draw_context *draw,
1912 unsigned shader_stage,
1913 unsigned sview_idx,
1914 uint32_t width, uint32_t height, uint32_t depth,
1915 uint32_t first_level, uint32_t last_level,
1916 const void *base_ptr,
1917 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1918 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1919 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1920 {
1921 unsigned j;
1922 struct draw_jit_texture *jit_tex;
1923
1924 assert(shader_stage == PIPE_SHADER_VERTEX ||
1925 shader_stage == PIPE_SHADER_GEOMETRY);
1926
1927 if (shader_stage == PIPE_SHADER_VERTEX) {
1928 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1929
1930 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1931 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1932 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1933
1934 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1935 } else {
1936 assert(0);
1937 return;
1938 }
1939
1940 jit_tex->width = width;
1941 jit_tex->height = height;
1942 jit_tex->depth = depth;
1943 jit_tex->first_level = first_level;
1944 jit_tex->last_level = last_level;
1945 jit_tex->base = base_ptr;
1946
1947 for (j = first_level; j <= last_level; j++) {
1948 jit_tex->mip_offsets[j] = mip_offsets[j];
1949 jit_tex->row_stride[j] = row_stride[j];
1950 jit_tex->img_stride[j] = img_stride[j];
1951 }
1952 }
1953
1954
1955 void
1956 draw_llvm_set_sampler_state(struct draw_context *draw,
1957 unsigned shader_type)
1958 {
1959 unsigned i;
1960
1961 if (shader_type == PIPE_SHADER_VERTEX) {
1962 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1963 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1964
1965 if (draw->samplers[i]) {
1966 const struct pipe_sampler_state *s
1967 = draw->samplers[PIPE_SHADER_VERTEX][i];
1968 jit_sam->min_lod = s->min_lod;
1969 jit_sam->max_lod = s->max_lod;
1970 jit_sam->lod_bias = s->lod_bias;
1971 COPY_4V(jit_sam->border_color, s->border_color.f);
1972 }
1973 }
1974 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1975 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1976 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1977
1978 if (draw->samplers[i]) {
1979 const struct pipe_sampler_state *s
1980 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1981 jit_sam->min_lod = s->min_lod;
1982 jit_sam->max_lod = s->max_lod;
1983 jit_sam->lod_bias = s->lod_bias;
1984 COPY_4V(jit_sam->border_color, s->border_color.f);
1985 }
1986 }
1987 }
1988 }
1989
1990
1991 void
1992 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1993 {
1994 struct draw_llvm *llvm = variant->llvm;
1995
1996 gallivm_destroy(variant->gallivm);
1997
1998 remove_from_list(&variant->list_item_local);
1999 variant->shader->variants_cached--;
2000 remove_from_list(&variant->list_item_global);
2001 llvm->nr_variants--;
2002 FREE(variant);
2003 }
2004
2005
2006 /**
2007 * Create LLVM types for various structures.
2008 */
2009 static void
2010 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2011 {
2012 struct gallivm_state *gallivm = var->gallivm;
2013 LLVMTypeRef texture_type, sampler_type, context_type;
2014
2015 texture_type = create_jit_texture_type(gallivm, "texture");
2016 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2017
2018 context_type = create_gs_jit_context_type(gallivm,
2019 var->shader->base.vector_length,
2020 texture_type, sampler_type,
2021 "draw_gs_jit_context");
2022 var->context_ptr_type = LLVMPointerType(context_type, 0);
2023
2024 var->input_array_type = create_gs_jit_input_type(gallivm);
2025 }
2026
2027 static LLVMTypeRef
2028 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2029 {
2030 if (!variant->context_ptr_type)
2031 create_gs_jit_types(variant);
2032 return variant->context_ptr_type;
2033 }
2034
2035 static LLVMValueRef
2036 generate_mask_value(struct draw_gs_llvm_variant *variant,
2037 struct lp_type gs_type)
2038 {
2039 struct gallivm_state *gallivm = variant->gallivm;
2040 LLVMBuilderRef builder = gallivm->builder;
2041 struct lp_type mask_type = lp_int_type(gs_type);
2042 LLVMValueRef num_prims;
2043 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2044 unsigned i;
2045
2046 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2047 variant->num_prims);
2048 for (i = 0; i <= gs_type.length; i++) {
2049 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2050 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2051 }
2052 mask_val = lp_build_compare(gallivm, mask_type,
2053 PIPE_FUNC_GREATER, num_prims, mask_val);
2054
2055 return mask_val;
2056 }
2057
2058 static void
2059 draw_gs_llvm_generate(struct draw_llvm *llvm,
2060 struct draw_gs_llvm_variant *variant)
2061 {
2062 struct gallivm_state *gallivm = variant->gallivm;
2063 LLVMContextRef context = gallivm->context;
2064 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2065 LLVMTypeRef arg_types[6];
2066 LLVMTypeRef func_type;
2067 LLVMValueRef variant_func;
2068 LLVMValueRef context_ptr;
2069 LLVMValueRef prim_id_ptr;
2070 LLVMBasicBlockRef block;
2071 LLVMBuilderRef builder;
2072 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2073 struct lp_build_sampler_soa *sampler = 0;
2074 struct lp_build_context bld;
2075 struct lp_bld_tgsi_system_values system_values;
2076 char func_name[64];
2077 struct lp_type gs_type;
2078 unsigned i;
2079 struct draw_gs_llvm_iface gs_iface;
2080 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2081 LLVMValueRef consts_ptr, num_consts_ptr;
2082 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2083 struct lp_build_mask_context mask;
2084 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2085 unsigned vector_length = variant->shader->base.vector_length;
2086
2087 memset(&system_values, 0, sizeof(system_values));
2088
2089 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2090 variant->shader->variants_cached);
2091
2092 assert(variant->vertex_header_ptr_type);
2093
2094 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2095 arg_types[1] = variant->input_array_type; /* input */
2096 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2097 arg_types[3] = int32_type; /* num_prims */
2098 arg_types[4] = int32_type; /* instance_id */
2099 arg_types[5] = LLVMPointerType(
2100 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2101
2102 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2103
2104 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2105
2106 variant->function = variant_func;
2107
2108 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2109
2110 for (i = 0; i < Elements(arg_types); ++i)
2111 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2112 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2113 LLVMNoAliasAttribute);
2114
2115 context_ptr = LLVMGetParam(variant_func, 0);
2116 input_array = LLVMGetParam(variant_func, 1);
2117 io_ptr = LLVMGetParam(variant_func, 2);
2118 num_prims = LLVMGetParam(variant_func, 3);
2119 system_values.instance_id = LLVMGetParam(variant_func, 4);
2120 prim_id_ptr = LLVMGetParam(variant_func, 5);
2121
2122 lp_build_name(context_ptr, "context");
2123 lp_build_name(input_array, "input");
2124 lp_build_name(io_ptr, "io");
2125 lp_build_name(num_prims, "num_prims");
2126 lp_build_name(system_values.instance_id, "instance_id");
2127 lp_build_name(prim_id_ptr, "prim_id_ptr");
2128
2129 variant->context_ptr = context_ptr;
2130 variant->io_ptr = io_ptr;
2131 variant->num_prims = num_prims;
2132
2133 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2134 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2135 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2136 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2137 gs_iface.input = input_array;
2138 gs_iface.variant = variant;
2139
2140 /*
2141 * Function body
2142 */
2143
2144 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2145 builder = gallivm->builder;
2146 LLVMPositionBuilderAtEnd(builder, block);
2147
2148 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2149
2150 memset(&gs_type, 0, sizeof gs_type);
2151 gs_type.floating = TRUE; /* floating point values */
2152 gs_type.sign = TRUE; /* values are signed */
2153 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2154 gs_type.width = 32; /* 32-bit float */
2155 gs_type.length = vector_length;
2156
2157 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2158 num_consts_ptr =
2159 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2160
2161 /* code generated texture sampling */
2162 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2163 context_ptr);
2164
2165 mask_val = generate_mask_value(variant, gs_type);
2166 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2167
2168 if (gs_info->uses_primid) {
2169 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2170 }
2171
2172 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2173 tgsi_dump(tokens, 0);
2174 draw_gs_llvm_dump_variant_key(&variant->key);
2175 }
2176
2177 lp_build_tgsi_soa(variant->gallivm,
2178 tokens,
2179 gs_type,
2180 &mask,
2181 consts_ptr,
2182 num_consts_ptr,
2183 &system_values,
2184 NULL,
2185 outputs,
2186 sampler,
2187 &llvm->draw->gs.geometry_shader->info,
2188 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2189
2190 sampler->destroy(sampler);
2191
2192 lp_build_mask_end(&mask);
2193
2194 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2195
2196 gallivm_verify_function(gallivm, variant_func);
2197 }
2198
2199
2200 struct draw_gs_llvm_variant *
2201 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2202 unsigned num_outputs,
2203 const struct draw_gs_llvm_variant_key *key)
2204 {
2205 struct draw_gs_llvm_variant *variant;
2206 struct llvm_geometry_shader *shader =
2207 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2208 LLVMTypeRef vertex_header;
2209 char module_name[64];
2210
2211 variant = MALLOC(sizeof *variant +
2212 shader->variant_key_size -
2213 sizeof variant->key);
2214 if (variant == NULL)
2215 return NULL;
2216
2217 variant->llvm = llvm;
2218 variant->shader = shader;
2219
2220 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2221 variant->shader->variants_cached);
2222
2223 variant->gallivm = gallivm_create(module_name, llvm->context);
2224
2225 create_gs_jit_types(variant);
2226
2227 memcpy(&variant->key, key, shader->variant_key_size);
2228
2229 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2230
2231 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2232
2233 draw_gs_llvm_generate(llvm, variant);
2234
2235 gallivm_compile_module(variant->gallivm);
2236
2237 variant->jit_func = (draw_gs_jit_func)
2238 gallivm_jit_function(variant->gallivm, variant->function);
2239
2240 gallivm_free_ir(variant->gallivm);
2241
2242 variant->list_item_global.base = variant;
2243 variant->list_item_local.base = variant;
2244 /*variant->no = */shader->variants_created++;
2245 variant->list_item_global.base = variant;
2246
2247 return variant;
2248 }
2249
2250 void
2251 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2252 {
2253 struct draw_llvm *llvm = variant->llvm;
2254
2255 gallivm_destroy(variant->gallivm);
2256
2257 remove_from_list(&variant->list_item_local);
2258 variant->shader->variants_cached--;
2259 remove_from_list(&variant->list_item_global);
2260 llvm->nr_gs_variants--;
2261 FREE(variant);
2262 }
2263
2264 struct draw_gs_llvm_variant_key *
2265 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2266 {
2267 unsigned i;
2268 struct draw_gs_llvm_variant_key *key;
2269 struct draw_sampler_static_state *draw_sampler;
2270
2271 key = (struct draw_gs_llvm_variant_key *)store;
2272
2273 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2274
2275 /* All variants of this shader will have the same value for
2276 * nr_samplers. Not yet trying to compact away holes in the
2277 * sampler array.
2278 */
2279 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2280 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2281 key->nr_sampler_views =
2282 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2283 }
2284 else {
2285 key->nr_sampler_views = key->nr_samplers;
2286 }
2287
2288 draw_sampler = key->samplers;
2289
2290 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2291
2292 for (i = 0 ; i < key->nr_samplers; i++) {
2293 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2294 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2295 }
2296 for (i = 0 ; i < key->nr_sampler_views; i++) {
2297 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2298 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2299 }
2300
2301 return key;
2302 }
2303
2304 void
2305 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2306 {
2307 unsigned i;
2308 struct draw_sampler_static_state *sampler = key->samplers;
2309
2310 for (i = 0 ; i < key->nr_sampler_views; i++) {
2311 debug_printf("sampler[%i].src_format = %s\n", i,
2312 util_format_name(sampler[i].texture_state.format));
2313 }
2314 }