projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: use vectorized calculations for fetch
[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_bitarit.h"
37 #include "gallivm/lp_bld_gather.h"
38 #include "gallivm/lp_bld_logic.h"
39 #include "gallivm/lp_bld_const.h"
40 #include "gallivm/lp_bld_swizzle.h"
41 #include "gallivm/lp_bld_struct.h"
42 #include "gallivm/lp_bld_type.h"
43 #include "gallivm/lp_bld_flow.h"
44 #include "gallivm/lp_bld_debug.h"
45 #include "gallivm/lp_bld_tgsi.h"
46 #include "gallivm/lp_bld_printf.h"
47 #include "gallivm/lp_bld_intr.h"
48 #include "gallivm/lp_bld_init.h"
49 #include "gallivm/lp_bld_type.h"
50 #include "gallivm/lp_bld_pack.h"
51 #include "gallivm/lp_bld_format.h"
52
53 #include "tgsi/tgsi_exec.h"
54 #include "tgsi/tgsi_dump.h"
55
56 #include "util/u_math.h"
57 #include "util/u_pointer.h"
58 #include "util/u_string.h"
59 #include "util/simple_list.h"
60
61
62 #define DEBUG_STORE 0
63
64
65 static void
66 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
67 boolean elts);
68
69
70 struct draw_gs_llvm_iface {
71 struct lp_build_tgsi_gs_iface base;
72
73 struct draw_gs_llvm_variant *variant;
74 LLVMValueRef input;
75 };
76
77 static inline const struct draw_gs_llvm_iface *
78 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
79 {
80 return (const struct draw_gs_llvm_iface *)iface;
81 }
82
83 /**
84 * Create LLVM type for draw_vertex_buffer.
85 */
86 static LLVMTypeRef
87 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
88 const char *struct_name)
89 {
90 LLVMTargetDataRef target = gallivm->target;
91 LLVMTypeRef dvbuffer_type;
92 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
93 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
94
95 elem_types[DRAW_JIT_DVBUFFER_MAP] =
96 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
97 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
98
99 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
100 ARRAY_SIZE(elem_types), 0);
101
102 (void) target; /* silence unused var warning for non-debug build */
103 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
104 target, dvbuffer_type,
105 DRAW_JIT_DVBUFFER_MAP);
106 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
107 target, dvbuffer_type,
108 DRAW_JIT_DVBUFFER_SIZE);
109
110 return dvbuffer_type;
111 }
112
113 /**
114 * Create LLVM type for struct draw_jit_texture
115 */
116 static LLVMTypeRef
117 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
118 {
119 LLVMTargetDataRef target = gallivm->target;
120 LLVMTypeRef texture_type;
121 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
122 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
123
124 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
125 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
126 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
127 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
128 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
129 elem_types[DRAW_JIT_TEXTURE_BASE] =
130 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
131 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
132 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
133 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
134 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
135
136 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
137 ARRAY_SIZE(elem_types), 0);
138
139 (void) target; /* silence unused var warning for non-debug build */
140 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
141 target, texture_type,
142 DRAW_JIT_TEXTURE_WIDTH);
143 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
144 target, texture_type,
145 DRAW_JIT_TEXTURE_HEIGHT);
146 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
147 target, texture_type,
148 DRAW_JIT_TEXTURE_DEPTH);
149 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
150 target, texture_type,
151 DRAW_JIT_TEXTURE_FIRST_LEVEL);
152 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
153 target, texture_type,
154 DRAW_JIT_TEXTURE_LAST_LEVEL);
155 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
156 target, texture_type,
157 DRAW_JIT_TEXTURE_BASE);
158 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
159 target, texture_type,
160 DRAW_JIT_TEXTURE_ROW_STRIDE);
161 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
162 target, texture_type,
163 DRAW_JIT_TEXTURE_IMG_STRIDE);
164 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
165 target, texture_type,
166 DRAW_JIT_TEXTURE_MIP_OFFSETS);
167
168 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
169
170 return texture_type;
171 }
172
173
174 /**
175 * Create LLVM type for struct draw_jit_sampler
176 */
177 static LLVMTypeRef
178 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
179 {
180 LLVMTargetDataRef target = gallivm->target;
181 LLVMTypeRef sampler_type;
182 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
183
184 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
185 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
186 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
187 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
188 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
189
190 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
191 ARRAY_SIZE(elem_types), 0);
192
193 (void) target; /* silence unused var warning for non-debug build */
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
195 target, sampler_type,
196 DRAW_JIT_SAMPLER_MIN_LOD);
197 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
198 target, sampler_type,
199 DRAW_JIT_SAMPLER_MAX_LOD);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
201 target, sampler_type,
202 DRAW_JIT_SAMPLER_LOD_BIAS);
203 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
204 target, sampler_type,
205 DRAW_JIT_SAMPLER_BORDER_COLOR);
206
207 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
208
209 return sampler_type;
210 }
211
212
213 /**
214 * Create LLVM type for struct draw_jit_context
215 */
216 static LLVMTypeRef
217 create_jit_context_type(struct gallivm_state *gallivm,
218 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
219 const char *struct_name)
220 {
221 LLVMTargetDataRef target = gallivm->target;
222 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
223 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
224 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
225 LLVMTypeRef context_type;
226
227 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
228 LP_MAX_TGSI_CONST_BUFFERS);
229 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
230 LP_MAX_TGSI_CONST_BUFFERS);
231 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
232 DRAW_TOTAL_CLIP_PLANES), 0);
233 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
234 elem_types[4] = LLVMArrayType(texture_type,
235 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
236 elem_types[5] = LLVMArrayType(sampler_type,
237 PIPE_MAX_SAMPLERS); /* samplers */
238 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
239 ARRAY_SIZE(elem_types), 0);
240
241 (void) target; /* silence unused var warning for non-debug build */
242 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
243 target, context_type, DRAW_JIT_CTX_CONSTANTS);
244 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
245 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
246 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
247 target, context_type, DRAW_JIT_CTX_PLANES);
248 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
249 target, context_type, DRAW_JIT_CTX_VIEWPORT);
250 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
251 target, context_type,
252 DRAW_JIT_CTX_TEXTURES);
253 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
254 target, context_type,
255 DRAW_JIT_CTX_SAMPLERS);
256 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
257 target, context_type);
258
259 return context_type;
260 }
261
262
263 /**
264 * Create LLVM type for struct draw_gs_jit_context
265 */
266 static LLVMTypeRef
267 create_gs_jit_context_type(struct gallivm_state *gallivm,
268 unsigned vector_length,
269 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
270 const char *struct_name)
271 {
272 LLVMTargetDataRef target = gallivm->target;
273 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
274 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
275 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
276 LLVMTypeRef context_type;
277
278 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
279 LP_MAX_TGSI_CONST_BUFFERS);
280 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
281 LP_MAX_TGSI_CONST_BUFFERS);
282 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
283 DRAW_TOTAL_CLIP_PLANES), 0);
284 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
285
286 elem_types[4] = LLVMArrayType(texture_type,
287 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
288 elem_types[5] = LLVMArrayType(sampler_type,
289 PIPE_MAX_SAMPLERS); /* samplers */
290
291 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
292 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
293 vector_length), 0);
294 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
295 vector_length), 0);
296
297 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
298 ARRAY_SIZE(elem_types), 0);
299
300 (void) target; /* silence unused var warning for non-debug build */
301 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
302 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
303 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
304 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
305 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
306 target, context_type, DRAW_GS_JIT_CTX_PLANES);
307 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
308 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
309 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
310 target, context_type,
311 DRAW_GS_JIT_CTX_TEXTURES);
312 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
313 target, context_type,
314 DRAW_GS_JIT_CTX_SAMPLERS);
315 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
316 target, context_type,
317 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
318 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
319 target, context_type,
320 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
321 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
322 target, context_type,
323 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
324 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
325 target, context_type);
326
327 return context_type;
328 }
329
330
331 static LLVMTypeRef
332 create_gs_jit_input_type(struct gallivm_state *gallivm)
333 {
334 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
335 LLVMTypeRef input_array;
336
337 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
338 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
339 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
340 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
341
342 return input_array;
343 }
344
345 /**
346 * Create LLVM type for struct pipe_vertex_buffer
347 */
348 static LLVMTypeRef
349 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
350 const char *struct_name)
351 {
352 LLVMTargetDataRef target = gallivm->target;
353 LLVMTypeRef elem_types[4];
354 LLVMTypeRef vb_type;
355
356 elem_types[0] =
357 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
358 elem_types[2] =
359 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
360
361 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
362 ARRAY_SIZE(elem_types), 0);
363
364 (void) target; /* silence unused var warning for non-debug build */
365 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
366 target, vb_type, 0);
367 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
368 target, vb_type, 1);
369
370 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
371
372 return vb_type;
373 }
374
375
376 /**
377 * Create LLVM type for struct vertex_header;
378 */
379 static LLVMTypeRef
380 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
381 {
382 LLVMTargetDataRef target = gallivm->target;
383 LLVMTypeRef elem_types[3];
384 LLVMTypeRef vertex_header;
385 char struct_name[24];
386
387 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
388
389 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
390 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
391 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
392
393 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
394 ARRAY_SIZE(elem_types), 0);
395
396 /* these are bit-fields and we can't take address of them
397 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
398 target, vertex_header,
399 DRAW_JIT_VERTEX_CLIPMASK);
400 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
401 target, vertex_header,
402 DRAW_JIT_VERTEX_EDGEFLAG);
403 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
404 target, vertex_header,
405 DRAW_JIT_VERTEX_PAD);
406 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
407 target, vertex_header,
408 DRAW_JIT_VERTEX_VERTEX_ID);
409 */
410 (void) target; /* silence unused var warning for non-debug build */
411 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
412 target, vertex_header,
413 DRAW_JIT_VERTEX_CLIP_POS);
414 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
415 target, vertex_header,
416 DRAW_JIT_VERTEX_DATA);
417
418 assert(LLVMABISizeOfType(target, vertex_header) ==
419 offsetof(struct vertex_header, data[data_elems]));
420
421 return vertex_header;
422 }
423
424
425 /**
426 * Create LLVM types for various structures.
427 */
428 static void
429 create_jit_types(struct draw_llvm_variant *variant)
430 {
431 struct gallivm_state *gallivm = variant->gallivm;
432 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
433 vb_type;
434
435 texture_type = create_jit_texture_type(gallivm, "texture");
436 sampler_type = create_jit_sampler_type(gallivm, "sampler");
437
438 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
439 "draw_jit_context");
440 variant->context_ptr_type = LLVMPointerType(context_type, 0);
441
442 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
443 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
444
445 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
446 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
447 }
448
449
450 static LLVMTypeRef
451 get_context_ptr_type(struct draw_llvm_variant *variant)
452 {
453 if (!variant->context_ptr_type)
454 create_jit_types(variant);
455 return variant->context_ptr_type;
456 }
457
458
459 static LLVMTypeRef
460 get_buffer_ptr_type(struct draw_llvm_variant *variant)
461 {
462 if (!variant->buffer_ptr_type)
463 create_jit_types(variant);
464 return variant->buffer_ptr_type;
465 }
466
467
468 static LLVMTypeRef
469 get_vb_ptr_type(struct draw_llvm_variant *variant)
470 {
471 if (!variant->vb_ptr_type)
472 create_jit_types(variant);
473 return variant->vb_ptr_type;
474 }
475
476 static LLVMTypeRef
477 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
478 {
479 if (!variant->vertex_header_ptr_type)
480 create_jit_types(variant);
481 return variant->vertex_header_ptr_type;
482 }
483
484
485 /**
486 * Create per-context LLVM info.
487 */
488 struct draw_llvm *
489 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
490 {
491 struct draw_llvm *llvm;
492
493 if (!lp_build_init())
494 return NULL;
495
496 llvm = CALLOC_STRUCT( draw_llvm );
497 if (!llvm)
498 return NULL;
499
500 llvm->draw = draw;
501
502 llvm->context = context;
503 if (!llvm->context) {
504 llvm->context = LLVMContextCreate();
505 llvm->context_owned = true;
506 }
507 if (!llvm->context)
508 goto fail;
509
510 llvm->nr_variants = 0;
511 make_empty_list(&llvm->vs_variants_list);
512
513 llvm->nr_gs_variants = 0;
514 make_empty_list(&llvm->gs_variants_list);
515
516 return llvm;
517
518 fail:
519 draw_llvm_destroy(llvm);
520 return NULL;
521 }
522
523
524 /**
525 * Free per-context LLVM info.
526 */
527 void
528 draw_llvm_destroy(struct draw_llvm *llvm)
529 {
530 if (llvm->context_owned)
531 LLVMContextDispose(llvm->context);
532 llvm->context = NULL;
533
534 /* XXX free other draw_llvm data? */
535 FREE(llvm);
536 }
537
538
539 /**
540 * Create LLVM-generated code for a vertex shader.
541 */
542 struct draw_llvm_variant *
543 draw_llvm_create_variant(struct draw_llvm *llvm,
544 unsigned num_inputs,
545 const struct draw_llvm_variant_key *key)
546 {
547 struct draw_llvm_variant *variant;
548 struct llvm_vertex_shader *shader =
549 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
550 LLVMTypeRef vertex_header;
551 char module_name[64];
552
553 variant = MALLOC(sizeof *variant +
554 shader->variant_key_size -
555 sizeof variant->key);
556 if (!variant)
557 return NULL;
558
559 variant->llvm = llvm;
560 variant->shader = shader;
561
562 util_snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
563 variant->shader->variants_cached);
564
565 variant->gallivm = gallivm_create(module_name, llvm->context);
566
567 create_jit_types(variant);
568
569 memcpy(&variant->key, key, shader->variant_key_size);
570
571 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
572 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
573 draw_llvm_dump_variant_key(&variant->key);
574 }
575
576 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
577
578 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
579
580 draw_llvm_generate(llvm, variant, FALSE); /* linear */
581 draw_llvm_generate(llvm, variant, TRUE); /* elts */
582
583 gallivm_compile_module(variant->gallivm);
584
585 variant->jit_func = (draw_jit_vert_func)
586 gallivm_jit_function(variant->gallivm, variant->function);
587
588 variant->jit_func_elts = (draw_jit_vert_func_elts)
589 gallivm_jit_function(variant->gallivm, variant->function_elts);
590
591 gallivm_free_ir(variant->gallivm);
592
593 variant->list_item_global.base = variant;
594 variant->list_item_local.base = variant;
595 /*variant->no = */shader->variants_created++;
596 variant->list_item_global.base = variant;
597
598 return variant;
599 }
600
601
602 static void
603 generate_vs(struct draw_llvm_variant *variant,
604 LLVMBuilderRef builder,
605 struct lp_type vs_type,
606 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
607 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
608 const struct lp_bld_tgsi_system_values *system_values,
609 LLVMValueRef context_ptr,
610 struct lp_build_sampler_soa *draw_sampler,
611 boolean clamp_vertex_color)
612 {
613 struct draw_llvm *llvm = variant->llvm;
614 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
615 LLVMValueRef consts_ptr =
616 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
617 LLVMValueRef num_consts_ptr =
618 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
619
620 lp_build_tgsi_soa(variant->gallivm,
621 tokens,
622 vs_type,
623 NULL /*struct lp_build_mask_context *mask*/,
624 consts_ptr,
625 num_consts_ptr,
626 system_values,
627 inputs,
628 outputs,
629 context_ptr,
630 NULL,
631 draw_sampler,
632 &llvm->draw->vs.vertex_shader->info,
633 NULL);
634
635 {
636 LLVMValueRef out;
637 unsigned chan, attrib;
638 struct lp_build_context bld;
639 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
640 lp_build_context_init(&bld, variant->gallivm, vs_type);
641
642 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
643 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
644 if (outputs[attrib][chan]) {
645 switch (info->output_semantic_name[attrib]) {
646 case TGSI_SEMANTIC_COLOR:
647 case TGSI_SEMANTIC_BCOLOR:
648 if (clamp_vertex_color) {
649 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
650 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
651 LLVMBuildStore(builder, out, outputs[attrib][chan]);
652 }
653 break;
654 }
655 }
656 }
657 }
658 }
659 }
660
661
662 static void
663 convert_to_soa(struct gallivm_state *gallivm,
664 LLVMValueRef src_aos[LP_MAX_VECTOR_WIDTH / 32],
665 LLVMValueRef dst_soa[TGSI_NUM_CHANNELS],
666 const struct lp_type soa_type)
667 {
668 unsigned j, k;
669 struct lp_type aos_channel_type = soa_type;
670
671 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
672 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
673
674 debug_assert(TGSI_NUM_CHANNELS == 4);
675 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
676
677 aos_channel_type.length >>= 1;
678
679 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
680 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
681
682 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
683
684 for (k = 0; k < pixels_per_channel; ++k) {
685 channel[k] = src_aos[j + TGSI_NUM_CHANNELS * k];
686 }
687
688 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
689 }
690
691 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa);
692 }
693
694
695 static void
696 fetch_vector(struct gallivm_state *gallivm,
697 const struct util_format_description *format_desc,
698 struct lp_type vs_type,
699 LLVMValueRef vb_stride,
700 LLVMValueRef map_ptr,
701 LLVMValueRef buffer_size_adj,
702 LLVMValueRef *inputs,
703 LLVMValueRef indices,
704 LLVMValueRef valid_mask)
705 {
706 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
707 LLVMBuilderRef builder = gallivm->builder;
708 struct lp_build_context blduivec;
709 LLVMValueRef offset, tmp;
710 LLVMValueRef aos_fetch[LP_MAX_VECTOR_WIDTH / 32];
711 unsigned i;
712
713 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
714
715 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
716 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
717
718 /*
719 * Sort of interestingly, with interleaved attribs, llvm 3.7+ will
720 * recognize these calculations to be constant with different attribs
721 * (the different offset has been added to map_ptr).
722 * llvm 3.3, however, will not (I can't get llvm 3.4-3.6 to link...)
723 *
724 * XXX: could actually avoid this altogether (replacing by simple
725 * non-widening mul) by precalculating the max index instead outside
726 * the loop (at the cost of one scalar udiv per vertex element).
727 */
728 offset = lp_build_mul_32_lohi(&blduivec, vb_stride, indices, &tmp);
729
730 tmp = lp_build_compare(gallivm, blduivec.type,
731 PIPE_FUNC_EQUAL, tmp, blduivec.zero);
732 valid_mask = LLVMBuildAnd(builder, tmp, valid_mask, "");
733
734 tmp = lp_build_compare(gallivm, blduivec.type,
735 PIPE_FUNC_LEQUAL, offset, buffer_size_adj);
736 valid_mask = LLVMBuildAnd(builder, tmp, valid_mask, "");
737
738 /* not valid elements use offset 0 */
739 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
740
741 if (0) {
742 lp_build_print_value(gallivm, " indices = ", indices);
743 lp_build_print_value(gallivm, " offsets = ", offset);
744 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
745 }
746
747 /*
748 * Note: we probably really want to use SoA fetch, not AoS one (albeit
749 * for most formats it will amount to the same as this isn't very
750 * optimized). But looks dangerous since it assumes alignment.
751 */
752 for (i = 0; i < vs_type.length; i++) {
753 LLVMValueRef offset1, elem;
754 elem = lp_build_const_int32(gallivm, i);
755 offset1 = LLVMBuildExtractElement(builder, offset, elem, "");
756
757 aos_fetch[i] = lp_build_fetch_rgba_aos(gallivm, format_desc,
758 lp_float32_vec4_type(),
759 FALSE, map_ptr, offset1,
760 zero, zero, NULL);
761 }
762 convert_to_soa(gallivm, aos_fetch, inputs, vs_type);
763
764 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
765 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
766 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
767 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
768 lp_build_vec_type(gallivm, vs_type), "");
769
770 }
771 }
772
773
774 static void
775 fetch_instanced(struct gallivm_state *gallivm,
776 const struct util_format_description *format_desc,
777 struct lp_type vs_type,
778 LLVMValueRef vb_stride,
779 LLVMValueRef map_ptr,
780 LLVMValueRef buffer_size_adj,
781 LLVMValueRef ofbit,
782 LLVMValueRef *inputs,
783 LLVMValueRef index)
784 {
785 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
786 LLVMBuilderRef builder = gallivm->builder;
787 LLVMValueRef stride, buffer_overflowed, aos;
788 LLVMValueRef temp_ptr =
789 lp_build_alloca(gallivm,
790 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
791 struct lp_build_if_state if_ctx;
792 unsigned i;
793
794 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
795
796 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
797 stride, buffer_size_adj,
798 "buffer_overflowed");
799 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
800
801 if (0) {
802 lp_build_print_value(gallivm, " instance index = ", index);
803 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
804 }
805
806 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
807 {
808 LLVMValueRef val =
809 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
810 LLVMBuildStore(builder, val, temp_ptr);
811 }
812 lp_build_else(&if_ctx);
813 {
814 LLVMValueRef val;
815
816 val = lp_build_fetch_rgba_aos(gallivm,
817 format_desc,
818 lp_float32_vec4_type(),
819 FALSE,
820 map_ptr,
821 stride, zero, zero,
822 NULL);
823 LLVMBuildStore(builder, val, temp_ptr);
824 }
825 lp_build_endif(&if_ctx);
826
827 aos = LLVMBuildLoad(builder, temp_ptr, "aos");
828
829 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
830 LLVMValueRef index = lp_build_const_int32(gallivm, i);
831 inputs[i] = lp_build_extract_broadcast(gallivm,
832 lp_float32_vec4_type(),
833 vs_type, aos, index);
834 }
835 }
836
837
838 static void
839 store_aos(struct gallivm_state *gallivm,
840 LLVMValueRef io_ptr,
841 LLVMValueRef index,
842 LLVMValueRef value)
843 {
844 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
845 LLVMBuilderRef builder = gallivm->builder;
846 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
847 LLVMValueRef indices[3];
848
849 indices[0] = lp_build_const_int32(gallivm, 0);
850 indices[1] = index;
851 indices[2] = lp_build_const_int32(gallivm, 0);
852
853 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
854 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
855
856 #if DEBUG_STORE
857 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
858 #endif
859
860 /* Unaligned store due to the vertex header */
861 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
862 }
863
864
865 /**
866 * Adjust the mask to architecture endianess. The mask will the store in struct:
867 *
868 * struct vertex_header {
869 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
870 * unsigned edgeflag:1;
871 * unsigned pad:1;
872 * unsigned vertex_id:16;
873 * [...]
874 * }
875 *
876 * On little-endian machine nothing needs to done, however on bit-endian machine
877 * the mask's fields need to be adjusted with the algorithm:
878 *
879 * uint32_t reverse (uint32_t x)
880 * {
881 * return (x >> 16) | // vertex_id
882 * ((x & 0x3fff) << 18) | // clipmask
883 * ((x & 0x4000) << 3) | // pad
884 * ((x & 0x8000) << 1); // edgeflag
885 * }
886 */
887 static LLVMValueRef
888 adjust_mask(struct gallivm_state *gallivm,
889 LLVMValueRef mask)
890 {
891 #ifdef PIPE_ARCH_BIG_ENDIAN
892 LLVMBuilderRef builder = gallivm->builder;
893 LLVMValueRef vertex_id;
894 LLVMValueRef clipmask;
895 LLVMValueRef pad;
896 LLVMValueRef edgeflag;
897
898 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
899 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
900 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
901 if (0) {
902 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
903 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
904 }
905 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
906 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
907
908 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
909 if (0) {
910 mask = LLVMBuildOr(builder, mask, pad, "");
911 }
912 mask = LLVMBuildOr(builder, mask, edgeflag, "");
913 #endif
914 return mask;
915 }
916
917 static void
918 store_aos_array(struct gallivm_state *gallivm,
919 struct lp_type soa_type,
920 LLVMValueRef io_ptr,
921 LLVMValueRef *indices,
922 LLVMValueRef* aos,
923 int attrib,
924 int num_outputs,
925 LLVMValueRef clipmask,
926 boolean need_edgeflag)
927 {
928 LLVMBuilderRef builder = gallivm->builder;
929 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
930 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
931 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
932 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
933 int vector_length = soa_type.length;
934 int i;
935
936 debug_assert(TGSI_NUM_CHANNELS == 4);
937
938 for (i = 0; i < vector_length; i++) {
939 linear_inds[i] = lp_build_const_int32(gallivm, i);
940 if (indices) {
941 inds[i] = indices[i];
942 } else {
943 inds[i] = linear_inds[i];
944 }
945 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
946 }
947
948 if (attrib == 0) {
949 /* store vertex header for each of the n vertices */
950 LLVMValueRef val, cliptmp;
951 int vertex_id_pad_edgeflag;
952
953 /* If this assertion fails, it means we need to update the bit twidding
954 * code here. See struct vertex_header in draw_private.h.
955 */
956 assert(DRAW_TOTAL_CLIP_PLANES==14);
957 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
958 if (!need_edgeflag) {
959 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
960 }
961 else {
962 vertex_id_pad_edgeflag = (0xffff << 16);
963 }
964 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
965 vertex_id_pad_edgeflag);
966 /* OR with the clipmask */
967 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
968 for (i = 0; i < vector_length; i++) {
969 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
970 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
971 val = adjust_mask(gallivm, val);
972 #if DEBUG_STORE
973 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
974 io_ptrs[i], inds[i], val);
975 #endif
976 LLVMBuildStore(builder, val, id_ptr);
977 }
978 }
979
980 /* store for each of the n vertices */
981 for (i = 0; i < vector_length; i++) {
982 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
983 }
984 }
985
986
987 static void
988 convert_to_aos(struct gallivm_state *gallivm,
989 LLVMValueRef io,
990 LLVMValueRef *indices,
991 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
992 LLVMValueRef clipmask,
993 int num_outputs,
994 struct lp_type soa_type,
995 boolean need_edgeflag)
996 {
997 LLVMBuilderRef builder = gallivm->builder;
998 unsigned chan, attrib, i;
999
1000 #if DEBUG_STORE
1001 lp_build_printf(gallivm, " # storing begin\n");
1002 #endif
1003 for (attrib = 0; attrib < num_outputs; ++attrib) {
1004 LLVMValueRef soa[TGSI_NUM_CHANNELS];
1005 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
1006 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1007 if (outputs[attrib][chan]) {
1008 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
1009 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
1010 #if DEBUG_STORE
1011 lp_build_printf(gallivm, "output %d : %d ",
1012 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1013 attrib, 0),
1014 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1015 chan, 0));
1016 lp_build_print_value(gallivm, "val = ", out);
1017 {
1018 LLVMValueRef iv =
1019 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
1020
1021 lp_build_print_value(gallivm, " ival = ", iv);
1022 }
1023 #endif
1024 soa[chan] = out;
1025 }
1026 else {
1027 soa[chan] = 0;
1028 }
1029 }
1030
1031
1032 if (soa_type.length == TGSI_NUM_CHANNELS) {
1033 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1034 } else {
1035 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1036
1037 for (i = 0; i < soa_type.length; ++i) {
1038 aos[i] = lp_build_extract_range(gallivm,
1039 soa[i % TGSI_NUM_CHANNELS],
1040 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1041 TGSI_NUM_CHANNELS);
1042 }
1043 }
1044
1045 store_aos_array(gallivm,
1046 soa_type,
1047 io, indices,
1048 aos,
1049 attrib,
1050 num_outputs,
1051 clipmask,
1052 need_edgeflag);
1053 }
1054 #if DEBUG_STORE
1055 lp_build_printf(gallivm, " # storing end\n");
1056 #endif
1057 }
1058
1059
1060 /**
1061 * Stores original vertex positions in clip coordinates
1062 */
1063 static void
1064 store_clip(struct gallivm_state *gallivm,
1065 const struct lp_type vs_type,
1066 LLVMValueRef io_ptr,
1067 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1068 int idx)
1069 {
1070 LLVMBuilderRef builder = gallivm->builder;
1071 LLVMValueRef soa[4];
1072 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1073 LLVMValueRef indices[2];
1074 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1075 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1076 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1077 LLVMTypeRef clip_ptr_type =
1078 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1079 4), 0);
1080 int i, j;
1081
1082 indices[0] =
1083 indices[1] = lp_build_const_int32(gallivm, 0);
1084
1085 for (i = 0; i < vs_type.length; i++) {
1086 inds[i] = lp_build_const_int32(gallivm, i);
1087 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1088 }
1089
1090 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1091 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1092 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1093 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1094
1095 for (i = 0; i < vs_type.length; i++) {
1096 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1097 }
1098
1099 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1100 for (i = 0; i < vs_type.length; ++i) {
1101 aos[i] = lp_build_extract_range(gallivm,
1102 soa[i % TGSI_NUM_CHANNELS],
1103 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1104 TGSI_NUM_CHANNELS);
1105 }
1106
1107 for (j = 0; j < vs_type.length; j++) {
1108 LLVMValueRef clip_ptr;
1109
1110 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1111 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1112
1113 /* Unaligned store */
1114 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1115 }
1116 }
1117
1118
1119 /**
1120 * Transforms the outputs for viewport mapping
1121 */
1122 static void
1123 generate_viewport(struct draw_llvm_variant *variant,
1124 LLVMBuilderRef builder,
1125 struct lp_type vs_type,
1126 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1127 LLVMValueRef context_ptr)
1128 {
1129 int i;
1130 struct gallivm_state *gallivm = variant->gallivm;
1131 struct lp_type f32_type = vs_type;
1132 const unsigned pos = variant->llvm->draw->vs.position_output;
1133 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1134 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1135 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1136 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1137
1138 /* We treat pipe_viewport_state as a float array */
1139 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1140 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1141
1142 /* for 1/w convention*/
1143 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1144 LLVMBuildStore(builder, out3, outputs[pos][3]);
1145
1146 /* Viewport Mapping */
1147 for (i=0; i<3; i++) {
1148 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1149 LLVMValueRef scale;
1150 LLVMValueRef trans;
1151 LLVMValueRef scale_i;
1152 LLVMValueRef trans_i;
1153 LLVMValueRef index;
1154
1155 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1156 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1157
1158 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1159 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1160
1161 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1162 LLVMBuildLoad(builder, scale_i, "scale"));
1163 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1164 LLVMBuildLoad(builder, trans_i, "trans"));
1165
1166 /* divide by w */
1167 out = LLVMBuildFMul(builder, out, out3, "");
1168 /* mult by scale, add translation */
1169 out = lp_build_fmuladd(builder, out, scale, trans);
1170
1171 /* store transformed outputs */
1172 LLVMBuildStore(builder, out, outputs[pos][i]);
1173 }
1174
1175 }
1176
1177
1178 /**
1179 * Returns clipmask as nxi32 bitmask for the n vertices
1180 */
1181 static LLVMValueRef
1182 generate_clipmask(struct draw_llvm *llvm,
1183 struct gallivm_state *gallivm,
1184 struct lp_type vs_type,
1185 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1186 struct draw_llvm_variant_key *key,
1187 LLVMValueRef context_ptr,
1188 boolean *have_clipdist)
1189 {
1190 LLVMBuilderRef builder = gallivm->builder;
1191 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1192 LLVMValueRef test, temp;
1193 LLVMValueRef zero, shift;
1194 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1195 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1196 LLVMValueRef plane1, planes, plane_ptr, sum;
1197 struct lp_type f32_type = vs_type;
1198 struct lp_type i32_type = lp_int_type(vs_type);
1199 const unsigned pos = llvm->draw->vs.position_output;
1200 const unsigned cv = llvm->draw->vs.clipvertex_output;
1201 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1202 boolean have_cd = false;
1203 boolean clip_user = key->clip_user;
1204 unsigned ucp_enable = key->ucp_enable;
1205 unsigned cd[2];
1206
1207 cd[0] = llvm->draw->vs.ccdistance_output[0];
1208 cd[1] = llvm->draw->vs.ccdistance_output[1];
1209
1210 if (cd[0] != pos || cd[1] != pos)
1211 have_cd = true;
1212
1213 if (num_written_clipdistance && !clip_user) {
1214 clip_user = true;
1215 ucp_enable = (1 << num_written_clipdistance) - 1;
1216 }
1217
1218 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1219 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1220 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1221 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1222
1223 /*
1224 * load clipvertex and position from correct locations.
1225 * if they are the same just load them once.
1226 */
1227 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1228 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1229 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1230 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1231
1232 if (clip_user && cv != pos) {
1233 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1234 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1235 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1236 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1237 } else {
1238 cv_x = pos_x;
1239 cv_y = pos_y;
1240 cv_z = pos_z;
1241 cv_w = pos_w;
1242 }
1243
1244 /*
1245 * Be careful with the comparisons and NaNs (using llvm's unordered
1246 * comparisons here).
1247 */
1248 /* Cliptest, for hardwired planes */
1249 /*
1250 * XXX should take guardband into account (currently not in key).
1251 * Otherwise might run the draw pipeline stages for nothing.
1252 */
1253 if (key->clip_xy) {
1254 /* plane 1 */
1255 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1256 temp = shift;
1257 test = LLVMBuildAnd(builder, test, temp, "");
1258 mask = test;
1259
1260 /* plane 2 */
1261 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1262 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1263 temp = LLVMBuildShl(builder, temp, shift, "");
1264 test = LLVMBuildAnd(builder, test, temp, "");
1265 mask = LLVMBuildOr(builder, mask, test, "");
1266
1267 /* plane 3 */
1268 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1269 temp = LLVMBuildShl(builder, temp, shift, "");
1270 test = LLVMBuildAnd(builder, test, temp, "");
1271 mask = LLVMBuildOr(builder, mask, test, "");
1272
1273 /* plane 4 */
1274 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1275 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1276 temp = LLVMBuildShl(builder, temp, shift, "");
1277 test = LLVMBuildAnd(builder, test, temp, "");
1278 mask = LLVMBuildOr(builder, mask, test, "");
1279 }
1280
1281 if (key->clip_z) {
1282 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1283 if (key->clip_halfz) {
1284 /* plane 5 */
1285 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1286 test = LLVMBuildAnd(builder, test, temp, "");
1287 mask = LLVMBuildOr(builder, mask, test, "");
1288 }
1289 else {
1290 /* plane 5 */
1291 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1292 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1293 test = LLVMBuildAnd(builder, test, temp, "");
1294 mask = LLVMBuildOr(builder, mask, test, "");
1295 }
1296 /* plane 6 */
1297 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1298 temp = LLVMBuildShl(builder, temp, shift, "");
1299 test = LLVMBuildAnd(builder, test, temp, "");
1300 mask = LLVMBuildOr(builder, mask, test, "");
1301 }
1302
1303 if (clip_user) {
1304 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1305 LLVMValueRef indices[3];
1306 LLVMValueRef is_nan_or_inf;
1307
1308 /* userclip planes */
1309 while (ucp_enable) {
1310 unsigned plane_idx = ffs(ucp_enable)-1;
1311 ucp_enable &= ~(1 << plane_idx);
1312 plane_idx += 6;
1313
1314 if (have_cd && num_written_clipdistance) {
1315 LLVMValueRef clipdist;
1316 int i;
1317 i = plane_idx - 6;
1318
1319 *have_clipdist = TRUE;
1320 if (i < 4) {
1321 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1322 } else {
1323 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1324 }
1325 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1326 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1327 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1328 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1329 test = LLVMBuildAnd(builder, test, temp, "");
1330 mask = LLVMBuildOr(builder, mask, test, "");
1331 } else {
1332 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1333 indices[0] = lp_build_const_int32(gallivm, 0);
1334 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1335
1336 indices[2] = lp_build_const_int32(gallivm, 0);
1337 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1338 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1339 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1340 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1341
1342 indices[2] = lp_build_const_int32(gallivm, 1);
1343 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1344 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1345 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1346 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1347
1348 indices[2] = lp_build_const_int32(gallivm, 2);
1349 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1350 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1351 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1352 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1353
1354 indices[2] = lp_build_const_int32(gallivm, 3);
1355 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1356 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1357 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1358 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1359
1360 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1361 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1362 test = LLVMBuildAnd(builder, test, temp, "");
1363 mask = LLVMBuildOr(builder, mask, test, "");
1364 }
1365 }
1366 }
1367 if (key->need_edgeflags) {
1368 /*
1369 * This isn't really part of clipmask but stored the same in vertex
1370 * header later, so do it here.
1371 */
1372 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1373 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1374 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1375 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1376 temp = lp_build_const_int_vec(gallivm, i32_type,
1377 1LL << DRAW_TOTAL_CLIP_PLANES);
1378 test = LLVMBuildAnd(builder, test, temp, "");
1379 mask = LLVMBuildOr(builder, mask, test, "");
1380 }
1381 return mask;
1382 }
1383
1384
1385 /**
1386 * Returns boolean if any clipping has occurred
1387 * Used zero/non-zero i32 value to represent boolean
1388 */
1389 static LLVMValueRef
1390 clipmask_booli32(struct gallivm_state *gallivm,
1391 const struct lp_type vs_type,
1392 LLVMValueRef clipmask_bool_ptr,
1393 boolean edgeflag_in_clipmask)
1394 {
1395 LLVMBuilderRef builder = gallivm->builder;
1396 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1397 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1398 LLVMValueRef ret = LLVMConstNull(int32_type);
1399 LLVMValueRef temp;
1400 int i;
1401
1402 /*
1403 * We need to invert the edgeflag bit from the clipmask here
1404 * (because the result is really if we want to run the pipeline or not
1405 * and we (may) need it if edgeflag was 0).
1406 */
1407 if (edgeflag_in_clipmask) {
1408 struct lp_type i32_type = lp_int_type(vs_type);
1409 LLVMValueRef edge = lp_build_const_int_vec(gallivm, i32_type,
1410 1LL << DRAW_TOTAL_CLIP_PLANES);
1411 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1412 }
1413 /*
1414 * Could do much better with just cmp/movmskps.
1415 */
1416 for (i=0; i < vs_type.length; i++) {
1417 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1418 lp_build_const_int32(gallivm, i) , "");
1419 ret = LLVMBuildOr(builder, ret, temp, "");
1420 }
1421 return ret;
1422 }
1423
1424 static LLVMValueRef
1425 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1426 struct lp_build_tgsi_context * bld_base,
1427 boolean is_vindex_indirect,
1428 LLVMValueRef vertex_index,
1429 boolean is_aindex_indirect,
1430 LLVMValueRef attrib_index,
1431 LLVMValueRef swizzle_index)
1432 {
1433 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1434 struct gallivm_state *gallivm = bld_base->base.gallivm;
1435 LLVMBuilderRef builder = gallivm->builder;
1436 LLVMValueRef indices[3];
1437 LLVMValueRef res;
1438 struct lp_type type = bld_base->base.type;
1439
1440 if (is_vindex_indirect || is_aindex_indirect) {
1441 int i;
1442 res = bld_base->base.zero;
1443 for (i = 0; i < type.length; ++i) {
1444 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1445 LLVMValueRef vert_chan_index = vertex_index;
1446 LLVMValueRef attr_chan_index = attrib_index;
1447 LLVMValueRef channel_vec, value;
1448
1449 if (is_vindex_indirect) {
1450 vert_chan_index = LLVMBuildExtractElement(builder,
1451 vertex_index, idx, "");
1452 }
1453 if (is_aindex_indirect) {
1454 attr_chan_index = LLVMBuildExtractElement(builder,
1455 attrib_index, idx, "");
1456 }
1457
1458 indices[0] = vert_chan_index;
1459 indices[1] = attr_chan_index;
1460 indices[2] = swizzle_index;
1461
1462 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1463 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1464 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1465
1466 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1467 }
1468 } else {
1469 indices[0] = vertex_index;
1470 indices[1] = attrib_index;
1471 indices[2] = swizzle_index;
1472
1473 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1474 res = LLVMBuildLoad(builder, res, "");
1475 }
1476
1477 return res;
1478 }
1479
1480 static void
1481 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1482 struct lp_build_tgsi_context * bld_base,
1483 LLVMValueRef (*outputs)[4],
1484 LLVMValueRef emitted_vertices_vec)
1485 {
1486 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1487 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1488 struct gallivm_state *gallivm = variant->gallivm;
1489 LLVMBuilderRef builder = gallivm->builder;
1490 struct lp_type gs_type = bld_base->base.type;
1491 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1492 lp_int_type(gs_type), 0);
1493 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1494 LLVMValueRef next_prim_offset =
1495 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1496 LLVMValueRef io = variant->io_ptr;
1497 unsigned i;
1498 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1499
1500 for (i = 0; i < gs_type.length; ++i) {
1501 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1502 LLVMValueRef currently_emitted =
1503 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1504 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1505 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1506 }
1507
1508 convert_to_aos(gallivm, io, indices,
1509 outputs, clipmask,
1510 gs_info->num_outputs, gs_type,
1511 FALSE);
1512 }
1513
1514 static void
1515 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1516 struct lp_build_tgsi_context * bld_base,
1517 LLVMValueRef verts_per_prim_vec,
1518 LLVMValueRef emitted_prims_vec)
1519 {
1520 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1521 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1522 struct gallivm_state *gallivm = variant->gallivm;
1523 LLVMBuilderRef builder = gallivm->builder;
1524 LLVMValueRef prim_lengts_ptr =
1525 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1526 unsigned i;
1527
1528 for (i = 0; i < bld_base->base.type.length; ++i) {
1529 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1530 LLVMValueRef prims_emitted =
1531 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1532 LLVMValueRef store_ptr;
1533 LLVMValueRef num_vertices =
1534 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1535
1536 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1537 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1538 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1539 LLVMBuildStore(builder, num_vertices, store_ptr);
1540 }
1541 }
1542
1543 static void
1544 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1545 struct lp_build_tgsi_context * bld_base,
1546 LLVMValueRef total_emitted_vertices_vec,
1547 LLVMValueRef emitted_prims_vec)
1548 {
1549 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1550 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1551 struct gallivm_state *gallivm = variant->gallivm;
1552 LLVMBuilderRef builder = gallivm->builder;
1553 LLVMValueRef emitted_verts_ptr =
1554 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1555 LLVMValueRef emitted_prims_ptr =
1556 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1557 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1558
1559 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1560 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1561
1562 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1563 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1564 }
1565
1566 static void
1567 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1568 boolean elts)
1569 {
1570 struct gallivm_state *gallivm = variant->gallivm;
1571 LLVMContextRef context = gallivm->context;
1572 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1573 LLVMTypeRef arg_types[11];
1574 unsigned num_arg_types =
1575 elts ? ARRAY_SIZE(arg_types) : ARRAY_SIZE(arg_types) - 1;
1576 LLVMTypeRef func_type;
1577 LLVMValueRef context_ptr;
1578 LLVMBasicBlockRef block;
1579 LLVMBuilderRef builder;
1580 char func_name[64];
1581 struct lp_type vs_type;
1582 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1583 LLVMValueRef vertex_id_offset, start_instance, start;
1584 LLVMValueRef stride, step, io_itr;
1585 LLVMValueRef ind_vec;
1586 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1587 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1588 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1589 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1590 LLVMValueRef ofmask[PIPE_MAX_ATTRIBS];
1591 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1592 LLVMValueRef fake_buf_ptr, fake_buf;
1593
1594 struct draw_context *draw = llvm->draw;
1595 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1596 unsigned i, j;
1597 struct lp_build_context bld, bldivec;
1598 struct lp_build_loop_state lp_loop;
1599 const int vector_length = lp_native_vector_width / 32;
1600 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1601 LLVMValueRef fetch_max;
1602 struct lp_build_sampler_soa *sampler = 0;
1603 LLVMValueRef ret, clipmask_bool_ptr;
1604 struct draw_llvm_variant_key *key = &variant->key;
1605 /* If geometry shader is present we need to skip both the viewport
1606 * transformation and clipping otherwise the inputs to the geometry
1607 * shader will be incorrect.
1608 * The code can't handle vp transform when vs writes vp index neither
1609 * (though this would be fixable here, but couldn't just broadcast
1610 * the values).
1611 */
1612 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1613 vs_info->writes_viewport_index;
1614 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1615 key->clip_z ||
1616 key->clip_user ||
1617 key->need_edgeflags);
1618 LLVMValueRef variant_func;
1619 const unsigned pos = draw->vs.position_output;
1620 const unsigned cv = draw->vs.clipvertex_output;
1621 boolean have_clipdist = FALSE;
1622 struct lp_bld_tgsi_system_values system_values;
1623
1624 memset(&system_values, 0, sizeof(system_values));
1625
1626 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1627 variant->shader->variants_cached, elts ? "elts" : "linear");
1628
1629 i = 0;
1630 arg_types[i++] = get_context_ptr_type(variant); /* context */
1631 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1632 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1633 if (elts) {
1634 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1635 arg_types[i++] = int32_type; /* fetch_elt_max */
1636 } else
1637 arg_types[i++] = int32_type; /* start */
1638 arg_types[i++] = int32_type; /* fetch_count / count */
1639 arg_types[i++] = int32_type; /* stride */
1640 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1641 arg_types[i++] = int32_type; /* instance_id */
1642 arg_types[i++] = int32_type; /* vertex_id_offset */
1643 arg_types[i++] = int32_type; /* start_instance */
1644
1645 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1646
1647 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1648
1649 if (elts)
1650 variant->function_elts = variant_func;
1651 else
1652 variant->function = variant_func;
1653
1654 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1655 for (i = 0; i < num_arg_types; ++i)
1656 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1657 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1658 LLVMNoAliasAttribute);
1659
1660 context_ptr = LLVMGetParam(variant_func, 0);
1661 io_ptr = LLVMGetParam(variant_func, 1);
1662 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1663 /*
1664 * XXX: stride is actually unused. The stride we use is strictly calculated
1665 * from the number of outputs (including the draw_extra outputs).
1666 * Should probably fix some day (we need a new vs just because of extra
1667 * outputs which the generated vs won't touch).
1668 */
1669 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1670 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1671 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1672 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1673 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1674
1675 lp_build_name(context_ptr, "context");
1676 lp_build_name(io_ptr, "io");
1677 lp_build_name(vbuffers_ptr, "vbuffers");
1678 lp_build_name(stride, "stride");
1679 lp_build_name(vb_ptr, "vb");
1680 lp_build_name(system_values.instance_id, "instance_id");
1681 lp_build_name(vertex_id_offset, "vertex_id_offset");
1682 lp_build_name(start_instance, "start_instance");
1683
1684 if (elts) {
1685 fetch_elts = LLVMGetParam(variant_func, 3);
1686 fetch_elt_max = LLVMGetParam(variant_func, 4);
1687 fetch_count = LLVMGetParam(variant_func, 5);
1688 lp_build_name(fetch_elts, "fetch_elts");
1689 lp_build_name(fetch_elt_max, "fetch_elt_max");
1690 lp_build_name(fetch_count, "fetch_count");
1691 start = count = NULL;
1692 }
1693 else {
1694 start = LLVMGetParam(variant_func, 3);
1695 count = LLVMGetParam(variant_func, 4);
1696 lp_build_name(start, "start");
1697 lp_build_name(count, "count");
1698 fetch_elts = fetch_count = NULL;
1699 }
1700
1701 /*
1702 * Function body
1703 */
1704
1705 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1706 builder = gallivm->builder;
1707 LLVMPositionBuilderAtEnd(builder, block);
1708
1709 memset(&vs_type, 0, sizeof vs_type);
1710 vs_type.floating = TRUE; /* floating point values */
1711 vs_type.sign = TRUE; /* values are signed */
1712 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1713 vs_type.width = 32; /* 32-bit float */
1714 vs_type.length = vector_length;
1715
1716 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1717 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1718
1719 /* hold temporary "bool" clipmask */
1720 clipmask_bool_ptr = lp_build_alloca(gallivm, bldivec.vec_type, "");
1721
1722 fake_buf = lp_build_alloca_undef(gallivm,
1723 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1724 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1725 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1726 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1727
1728 /* code generated texture sampling */
1729 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1730
1731 step = lp_build_const_int32(gallivm, vector_length);
1732
1733 ind_vec = bldivec.undef;
1734 for (i = 0; i < vs_type.length; i++) {
1735 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1736 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1737 }
1738
1739
1740 if (elts) {
1741 fetch_max = lp_build_min(&bld, fetch_count, fetch_elt_max);
1742 count = fetch_count;
1743 start = bldivec.zero;
1744 }
1745 else {
1746 fetch_max = lp_build_add(&bld, start, count);
1747 start = lp_build_broadcast_scalar(&bldivec, start);
1748 ind_vec = lp_build_add(&bldivec, start, ind_vec);
1749 }
1750
1751 fetch_max = lp_build_broadcast_scalar(&bldivec, fetch_max);
1752
1753 /*
1754 * Pre-calculate everything which is constant per shader invocation.
1755 */
1756 for (j = 0; j < key->nr_vertex_elements; ++j) {
1757 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1758 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1759 struct pipe_vertex_element *velem = &key->vertex_element[j];
1760 LLVMValueRef vb_index =
1761 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1762 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1763 util_format_get_blocksize(velem->src_format));
1764 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1765 velem->src_offset);
1766 struct lp_build_if_state if_ctx;
1767
1768 if (velem->src_format != PIPE_FORMAT_NONE) {
1769 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1770 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1771 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1772 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1773 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1774 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1775
1776 ofbit = NULL;
1777 buf_offset = lp_build_uadd_overflow(gallivm, vb_buffer_offset,
1778 src_offset, &ofbit);
1779 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, bsize,
1780 &ofbit);
1781 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1782 buf_offset, &ofbit);
1783
1784 /*
1785 * Unlike elts, we cannot easily set fake vertex buffers outside
1786 * the generated code. Hence, set fake vertex buffers here instead
1787 * basically, so fetch code can always fetch using offset 0, eliminating
1788 * all control flow (for instanced elements, we keep the control flow
1789 * for now as it's a scalar fetch, so don't need this).
1790 */
1791 if (velem->instance_divisor) {
1792 /* Index is equal to the start instance plus the number of current
1793 * instance divided by the divisor. In this case we compute it as:
1794 * index = start_instance + (instance_id / divisor)
1795 */
1796 LLVMValueRef current_instance;
1797 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1798 lp_build_const_int32(gallivm,
1799 velem->instance_divisor),
1800 "instance_divisor");
1801 instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
1802 current_instance, &ofbit);
1803 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1804
1805 /* This is a scalar fetch, just keep the of bit */
1806 ofmask[j] = ofbit;
1807 }
1808 else {
1809 temp_ptr = lp_build_alloca_undef(gallivm,
1810 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1811
1812 lp_build_if(&if_ctx, gallivm, ofbit);
1813 {
1814 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1815 }
1816 lp_build_else(&if_ctx);
1817 {
1818 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1819 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1820 }
1821 lp_build_endif(&if_ctx);
1822 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
1823
1824 /* expand to vector mask */
1825 ofmask[j] = LLVMBuildZExt(builder, ofbit, int32_type, "");
1826 ofmask[j] = lp_build_broadcast_scalar(&bldivec, ofmask[j]);
1827 }
1828
1829 if (0) {
1830 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1831 lp_build_const_int32(gallivm, j),
1832 vb_index, vb_stride[j]);
1833 lp_build_printf(gallivm,
1834 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1835 vb_buffer_offset, src_offset, buf_offset);
1836 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
1837 buffer_size, bsize);
1838 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
1839 }
1840 }
1841 }
1842
1843 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1844 {
1845 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1846 LLVMValueRef io;
1847 LLVMValueRef clipmask; /* holds the clipmask value */
1848 LLVMValueRef true_index_array;
1849 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1850 LLVMValueRef indices_valid;
1851
1852 io_itr = lp_loop.counter;
1853
1854 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1855 #if DEBUG_STORE
1856 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1857 io_itr, io, lp_loop.counter);
1858 #endif
1859
1860 true_index_array = lp_build_broadcast_scalar(&bldivec, lp_loop.counter);
1861 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1862
1863 indices_valid = lp_build_compare(gallivm, bldivec.type,
1864 PIPE_FUNC_LESS,
1865 true_index_array, fetch_max);
1866
1867 if (elts) {
1868 #if 0
1869 /*
1870 * AVX2 gather would not need zeroing offsets.
1871 * And it would not need dummy index buffer neither.
1872 * But for now don't bother.
1873 */
1874 true_index_array = lp_build_gather_masked(gallivm, vs_type.length,
1875 32, 32, TRUE,
1876 fetch_elts, true_index_array,
1877 FALSE, indices_valid);
1878 #else
1879 true_index_array = LLVMBuildAnd(builder, true_index_array,
1880 indices_valid, "");
1881 /*
1882 * XXX should not have to do this, as scale can be handled
1883 * natively by loads (hits asserts though).
1884 */
1885 true_index_array = lp_build_shl_imm(&bldivec, true_index_array, 2);
1886 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1887 LLVMPointerType(LLVMInt8TypeInContext(context),
1888 0), "");
1889 true_index_array = lp_build_gather(gallivm, vs_type.length,
1890 32, 32, TRUE,
1891 fetch_elts, true_index_array,
1892 FALSE);
1893 #endif
1894 }
1895
1896 for (j = 0; j < key->nr_vertex_elements; ++j) {
1897 struct pipe_vertex_element *velem = &key->vertex_element[j];
1898 const struct util_format_description *format_desc =
1899 util_format_description(velem->src_format);
1900
1901 if (format_desc->format == PIPE_FORMAT_NONE) {
1902 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1903 inputs[j][i] = lp_build_zero(gallivm, vs_type);
1904 }
1905 }
1906 else if (velem->instance_divisor) {
1907 fetch_instanced(gallivm, format_desc, vs_type,
1908 vb_stride[j], map_ptr[j],
1909 buffer_size_adj[j], ofmask[j],
1910 inputs[j], instance_index[j]);
1911 }
1912 else {
1913 indices_valid = lp_build_andnot(&bldivec, indices_valid, ofmask[j]);
1914 fetch_vector(gallivm, format_desc, vs_type,
1915 vb_stride[j], map_ptr[j],
1916 buffer_size_adj[j], inputs[j],
1917 true_index_array, indices_valid);
1918 }
1919 }
1920
1921 /* In the paths with elts vertex id has to be unaffected by the
1922 * index bias and because indices inside our elements array have
1923 * already had index bias applied we need to subtract it here to
1924 * get back to the original index.
1925 * in the linear paths vertex id has to be unaffected by the
1926 * original start index and because we abuse the 'start' variable
1927 * to either represent the actual start index or the index at which
1928 * the primitive was split (we split rendering into chunks of at
1929 * most 4095-vertices) we need to back out the original start
1930 * index out of our vertex id here.
1931 */
1932 system_values.basevertex = lp_build_broadcast_scalar(&bldivec,
1933 vertex_id_offset);
1934 system_values.vertex_id = true_index_array;
1935 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1936 system_values.basevertex, "");
1937
1938 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1939 generate_vs(variant,
1940 builder,
1941 vs_type,
1942 outputs,
1943 ptr_aos,
1944 &system_values,
1945 context_ptr,
1946 sampler,
1947 key->clamp_vertex_color);
1948
1949 if (pos != -1 && cv != -1) {
1950 /* store original positions in clip before further manipulation */
1951 store_clip(gallivm, vs_type, io, outputs, pos);
1952
1953 /* do cliptest */
1954 if (enable_cliptest) {
1955 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1956 /* allocate clipmask, assign it integer type */
1957 clipmask = generate_clipmask(llvm,
1958 gallivm,
1959 vs_type,
1960 outputs,
1961 key,
1962 context_ptr, &have_clipdist);
1963 temp = LLVMBuildOr(builder, clipmask, temp, "");
1964 /* store temporary clipping boolean value */
1965 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1966 }
1967 else {
1968 clipmask = bldivec.zero;
1969 }
1970
1971 /* do viewport mapping */
1972 if (!bypass_viewport) {
1973 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1974 }
1975 }
1976 else {
1977 clipmask = bldivec.zero;
1978 }
1979
1980 /* store clipmask in vertex header,
1981 * original positions in clip
1982 * and transformed positions in data
1983 */
1984 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1985 vs_info->num_outputs, vs_type,
1986 enable_cliptest && key->need_edgeflags);
1987 }
1988 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1989
1990 sampler->destroy(sampler);
1991
1992 /* return clipping boolean value for function */
1993 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr,
1994 enable_cliptest && key->need_edgeflags);
1995
1996 LLVMBuildRet(builder, ret);
1997
1998 gallivm_verify_function(gallivm, variant_func);
1999 }
2000
2001
2002 struct draw_llvm_variant_key *
2003 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2004 {
2005 unsigned i;
2006 struct draw_llvm_variant_key *key;
2007 struct draw_sampler_static_state *draw_sampler;
2008
2009 key = (struct draw_llvm_variant_key *)store;
2010
2011 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2012
2013 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
2014
2015 /* will have to rig this up properly later */
2016 key->clip_xy = llvm->draw->clip_xy;
2017 key->clip_z = llvm->draw->clip_z;
2018 key->clip_user = llvm->draw->clip_user;
2019 key->bypass_viewport = llvm->draw->bypass_viewport;
2020 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2021 /* XXX assumes edgeflag output not at 0 */
2022 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2023 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2024 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
2025 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2026
2027 /* All variants of this shader will have the same value for
2028 * nr_samplers. Not yet trying to compact away holes in the
2029 * sampler array.
2030 */
2031 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2032 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2033 key->nr_sampler_views =
2034 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2035 }
2036 else {
2037 key->nr_sampler_views = key->nr_samplers;
2038 }
2039
2040 /* Presumably all variants of the shader should have the same
2041 * number of vertex elements - ie the number of shader inputs.
2042 * NOTE: we NEED to store the needed number of needed inputs
2043 * here, not the number of provided elements to match keysize
2044 * (and the offset of sampler state in the key).
2045 * If we have excess number of vertex elements, this is valid,
2046 * but the excess ones don't matter.
2047 * If we don't have enough vertex elements (which looks not really
2048 * valid but we'll handle it gracefully) fill out missing ones with
2049 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2050 */
2051 key->nr_vertex_elements =
2052 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2053
2054 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2055 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2056 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2057 memset(key->vertex_element, 0,
2058 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2059 }
2060 memcpy(key->vertex_element,
2061 llvm->draw->pt.vertex_element,
2062 sizeof(struct pipe_vertex_element) *
2063 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2064
2065 draw_sampler = draw_llvm_variant_key_samplers(key);
2066 memset(draw_sampler, 0,
2067 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2068
2069 for (i = 0 ; i < key->nr_samplers; i++) {
2070 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2071 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2072 }
2073 for (i = 0 ; i < key->nr_sampler_views; i++) {
2074 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2075 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2076 }
2077
2078 return key;
2079 }
2080
2081
2082 void
2083 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2084 {
2085 unsigned i;
2086 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2087
2088 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2089 debug_printf("clip_xy = %u\n", key->clip_xy);
2090 debug_printf("clip_z = %u\n", key->clip_z);
2091 debug_printf("clip_user = %u\n", key->clip_user);
2092 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2093 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2094 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2095 debug_printf("has_gs = %u\n", key->has_gs);
2096 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2097
2098 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2099 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2100 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2101 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2102 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2103 }
2104
2105 for (i = 0 ; i < key->nr_sampler_views; i++) {
2106 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2107 }
2108 }
2109
2110
2111 void
2112 draw_llvm_set_mapped_texture(struct draw_context *draw,
2113 unsigned shader_stage,
2114 unsigned sview_idx,
2115 uint32_t width, uint32_t height, uint32_t depth,
2116 uint32_t first_level, uint32_t last_level,
2117 const void *base_ptr,
2118 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2119 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2120 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2121 {
2122 unsigned j;
2123 struct draw_jit_texture *jit_tex;
2124
2125 assert(shader_stage == PIPE_SHADER_VERTEX ||
2126 shader_stage == PIPE_SHADER_GEOMETRY);
2127
2128 if (shader_stage == PIPE_SHADER_VERTEX) {
2129 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2130
2131 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2132 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2133 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2134
2135 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2136 } else {
2137 assert(0);
2138 return;
2139 }
2140
2141 jit_tex->width = width;
2142 jit_tex->height = height;
2143 jit_tex->depth = depth;
2144 jit_tex->first_level = first_level;
2145 jit_tex->last_level = last_level;
2146 jit_tex->base = base_ptr;
2147
2148 for (j = first_level; j <= last_level; j++) {
2149 jit_tex->mip_offsets[j] = mip_offsets[j];
2150 jit_tex->row_stride[j] = row_stride[j];
2151 jit_tex->img_stride[j] = img_stride[j];
2152 }
2153 }
2154
2155
2156 void
2157 draw_llvm_set_sampler_state(struct draw_context *draw,
2158 unsigned shader_type)
2159 {
2160 unsigned i;
2161
2162 if (shader_type == PIPE_SHADER_VERTEX) {
2163 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2164 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2165
2166 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2167 const struct pipe_sampler_state *s
2168 = draw->samplers[PIPE_SHADER_VERTEX][i];
2169 jit_sam->min_lod = s->min_lod;
2170 jit_sam->max_lod = s->max_lod;
2171 jit_sam->lod_bias = s->lod_bias;
2172 COPY_4V(jit_sam->border_color, s->border_color.f);
2173 }
2174 }
2175 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2176 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2177 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2178
2179 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2180 const struct pipe_sampler_state *s
2181 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2182 jit_sam->min_lod = s->min_lod;
2183 jit_sam->max_lod = s->max_lod;
2184 jit_sam->lod_bias = s->lod_bias;
2185 COPY_4V(jit_sam->border_color, s->border_color.f);
2186 }
2187 }
2188 }
2189 }
2190
2191
2192 void
2193 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2194 {
2195 struct draw_llvm *llvm = variant->llvm;
2196
2197 gallivm_destroy(variant->gallivm);
2198
2199 remove_from_list(&variant->list_item_local);
2200 variant->shader->variants_cached--;
2201 remove_from_list(&variant->list_item_global);
2202 llvm->nr_variants--;
2203 FREE(variant);
2204 }
2205
2206
2207 /**
2208 * Create LLVM types for various structures.
2209 */
2210 static void
2211 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2212 {
2213 struct gallivm_state *gallivm = var->gallivm;
2214 LLVMTypeRef texture_type, sampler_type, context_type;
2215
2216 texture_type = create_jit_texture_type(gallivm, "texture");
2217 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2218
2219 context_type = create_gs_jit_context_type(gallivm,
2220 var->shader->base.vector_length,
2221 texture_type, sampler_type,
2222 "draw_gs_jit_context");
2223 var->context_ptr_type = LLVMPointerType(context_type, 0);
2224
2225 var->input_array_type = create_gs_jit_input_type(gallivm);
2226 }
2227
2228 static LLVMTypeRef
2229 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2230 {
2231 if (!variant->context_ptr_type)
2232 create_gs_jit_types(variant);
2233 return variant->context_ptr_type;
2234 }
2235
2236 static LLVMValueRef
2237 generate_mask_value(struct draw_gs_llvm_variant *variant,
2238 struct lp_type gs_type)
2239 {
2240 struct gallivm_state *gallivm = variant->gallivm;
2241 LLVMBuilderRef builder = gallivm->builder;
2242 struct lp_type mask_type = lp_int_type(gs_type);
2243 LLVMValueRef num_prims;
2244 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2245 unsigned i;
2246
2247 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2248 variant->num_prims);
2249 for (i = 0; i < gs_type.length; i++) {
2250 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2251 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2252 }
2253 mask_val = lp_build_compare(gallivm, mask_type,
2254 PIPE_FUNC_GREATER, num_prims, mask_val);
2255
2256 return mask_val;
2257 }
2258
2259 static void
2260 draw_gs_llvm_generate(struct draw_llvm *llvm,
2261 struct draw_gs_llvm_variant *variant)
2262 {
2263 struct gallivm_state *gallivm = variant->gallivm;
2264 LLVMContextRef context = gallivm->context;
2265 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2266 LLVMTypeRef arg_types[7];
2267 LLVMTypeRef func_type;
2268 LLVMValueRef variant_func;
2269 LLVMValueRef context_ptr;
2270 LLVMValueRef prim_id_ptr;
2271 LLVMBasicBlockRef block;
2272 LLVMBuilderRef builder;
2273 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2274 struct lp_build_sampler_soa *sampler = 0;
2275 struct lp_build_context bld;
2276 struct lp_bld_tgsi_system_values system_values;
2277 char func_name[64];
2278 struct lp_type gs_type;
2279 unsigned i;
2280 struct draw_gs_llvm_iface gs_iface;
2281 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2282 LLVMValueRef consts_ptr, num_consts_ptr;
2283 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2284 struct lp_build_mask_context mask;
2285 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2286 unsigned vector_length = variant->shader->base.vector_length;
2287
2288 memset(&system_values, 0, sizeof(system_values));
2289
2290 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2291 variant->shader->variants_cached);
2292
2293 assert(variant->vertex_header_ptr_type);
2294
2295 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2296 arg_types[1] = variant->input_array_type; /* input */
2297 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2298 arg_types[3] = int32_type; /* num_prims */
2299 arg_types[4] = int32_type; /* instance_id */
2300 arg_types[5] = LLVMPointerType(
2301 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2302 arg_types[6] = int32_type;
2303
2304 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2305
2306 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2307
2308 variant->function = variant_func;
2309
2310 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2311
2312 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2313 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2314 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2315 LLVMNoAliasAttribute);
2316
2317 context_ptr = LLVMGetParam(variant_func, 0);
2318 input_array = LLVMGetParam(variant_func, 1);
2319 io_ptr = LLVMGetParam(variant_func, 2);
2320 num_prims = LLVMGetParam(variant_func, 3);
2321 system_values.instance_id = LLVMGetParam(variant_func, 4);
2322 prim_id_ptr = LLVMGetParam(variant_func, 5);
2323 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2324
2325 lp_build_name(context_ptr, "context");
2326 lp_build_name(input_array, "input");
2327 lp_build_name(io_ptr, "io");
2328 lp_build_name(num_prims, "num_prims");
2329 lp_build_name(system_values.instance_id, "instance_id");
2330 lp_build_name(prim_id_ptr, "prim_id_ptr");
2331 lp_build_name(system_values.invocation_id, "invocation_id");
2332
2333 variant->context_ptr = context_ptr;
2334 variant->io_ptr = io_ptr;
2335 variant->num_prims = num_prims;
2336
2337 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2338 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2339 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2340 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2341 gs_iface.input = input_array;
2342 gs_iface.variant = variant;
2343
2344 /*
2345 * Function body
2346 */
2347
2348 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2349 builder = gallivm->builder;
2350 LLVMPositionBuilderAtEnd(builder, block);
2351
2352 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2353
2354 memset(&gs_type, 0, sizeof gs_type);
2355 gs_type.floating = TRUE; /* floating point values */
2356 gs_type.sign = TRUE; /* values are signed */
2357 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2358 gs_type.width = 32; /* 32-bit float */
2359 gs_type.length = vector_length;
2360
2361 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2362 num_consts_ptr =
2363 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2364
2365 /* code generated texture sampling */
2366 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2367
2368 mask_val = generate_mask_value(variant, gs_type);
2369 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2370
2371 if (gs_info->uses_primid) {
2372 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2373 }
2374
2375 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2376 tgsi_dump(tokens, 0);
2377 draw_gs_llvm_dump_variant_key(&variant->key);
2378 }
2379
2380 lp_build_tgsi_soa(variant->gallivm,
2381 tokens,
2382 gs_type,
2383 &mask,
2384 consts_ptr,
2385 num_consts_ptr,
2386 &system_values,
2387 NULL,
2388 outputs,
2389 context_ptr,
2390 NULL,
2391 sampler,
2392 &llvm->draw->gs.geometry_shader->info,
2393 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2394
2395 sampler->destroy(sampler);
2396
2397 lp_build_mask_end(&mask);
2398
2399 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2400
2401 gallivm_verify_function(gallivm, variant_func);
2402 }
2403
2404
2405 struct draw_gs_llvm_variant *
2406 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2407 unsigned num_outputs,
2408 const struct draw_gs_llvm_variant_key *key)
2409 {
2410 struct draw_gs_llvm_variant *variant;
2411 struct llvm_geometry_shader *shader =
2412 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2413 LLVMTypeRef vertex_header;
2414 char module_name[64];
2415
2416 variant = MALLOC(sizeof *variant +
2417 shader->variant_key_size -
2418 sizeof variant->key);
2419 if (!variant)
2420 return NULL;
2421
2422 variant->llvm = llvm;
2423 variant->shader = shader;
2424
2425 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2426 variant->shader->variants_cached);
2427
2428 variant->gallivm = gallivm_create(module_name, llvm->context);
2429
2430 create_gs_jit_types(variant);
2431
2432 memcpy(&variant->key, key, shader->variant_key_size);
2433
2434 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2435
2436 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2437
2438 draw_gs_llvm_generate(llvm, variant);
2439
2440 gallivm_compile_module(variant->gallivm);
2441
2442 variant->jit_func = (draw_gs_jit_func)
2443 gallivm_jit_function(variant->gallivm, variant->function);
2444
2445 gallivm_free_ir(variant->gallivm);
2446
2447 variant->list_item_global.base = variant;
2448 variant->list_item_local.base = variant;
2449 /*variant->no = */shader->variants_created++;
2450 variant->list_item_global.base = variant;
2451
2452 return variant;
2453 }
2454
2455 void
2456 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2457 {
2458 struct draw_llvm *llvm = variant->llvm;
2459
2460 gallivm_destroy(variant->gallivm);
2461
2462 remove_from_list(&variant->list_item_local);
2463 variant->shader->variants_cached--;
2464 remove_from_list(&variant->list_item_global);
2465 llvm->nr_gs_variants--;
2466 FREE(variant);
2467 }
2468
2469 struct draw_gs_llvm_variant_key *
2470 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2471 {
2472 unsigned i;
2473 struct draw_gs_llvm_variant_key *key;
2474 struct draw_sampler_static_state *draw_sampler;
2475
2476 key = (struct draw_gs_llvm_variant_key *)store;
2477
2478 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2479
2480 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2481
2482 /* All variants of this shader will have the same value for
2483 * nr_samplers. Not yet trying to compact away holes in the
2484 * sampler array.
2485 */
2486 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2487 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2488 key->nr_sampler_views =
2489 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2490 }
2491 else {
2492 key->nr_sampler_views = key->nr_samplers;
2493 }
2494
2495 draw_sampler = key->samplers;
2496
2497 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2498
2499 for (i = 0 ; i < key->nr_samplers; i++) {
2500 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2501 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2502 }
2503 for (i = 0 ; i < key->nr_sampler_views; i++) {
2504 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2505 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2506 }
2507
2508 return key;
2509 }
2510
2511 void
2512 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2513 {
2514 unsigned i;
2515 struct draw_sampler_static_state *sampler = key->samplers;
2516
2517 for (i = 0 ; i < key->nr_sampler_views; i++) {
2518 debug_printf("sampler[%i].src_format = %s\n", i,
2519 util_format_name(sampler[i].texture_state.format));
2520 }
2521 }