projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
llvmpipe/draw: Fix texture sampling in geometry shaders
[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_logic.h"
36 #include "gallivm/lp_bld_const.h"
37 #include "gallivm/lp_bld_swizzle.h"
38 #include "gallivm/lp_bld_struct.h"
39 #include "gallivm/lp_bld_type.h"
40 #include "gallivm/lp_bld_flow.h"
41 #include "gallivm/lp_bld_debug.h"
42 #include "gallivm/lp_bld_tgsi.h"
43 #include "gallivm/lp_bld_printf.h"
44 #include "gallivm/lp_bld_intr.h"
45 #include "gallivm/lp_bld_init.h"
46 #include "gallivm/lp_bld_type.h"
47 #include "gallivm/lp_bld_pack.h"
48 #include "gallivm/lp_bld_format.h"
49
50 #include "tgsi/tgsi_exec.h"
51 #include "tgsi/tgsi_dump.h"
52
53 #include "util/u_math.h"
54 #include "util/u_pointer.h"
55 #include "util/u_string.h"
56 #include "util/u_simple_list.h"
57
58
59 #define DEBUG_STORE 0
60
61
62 static void
63 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
64 boolean elts);
65
66
67 /**
68 * Create LLVM type for struct draw_jit_texture
69 */
70 static LLVMTypeRef
71 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
72 {
73 LLVMTargetDataRef target = gallivm->target;
74 LLVMTypeRef texture_type;
75 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
76 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
77
78 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
79 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
80 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
81 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
82 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
83 elem_types[DRAW_JIT_TEXTURE_BASE] =
84 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
85 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
86 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
87 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
88 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
89
90 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
91 Elements(elem_types), 0);
92
93 #if HAVE_LLVM < 0x0300
94 LLVMAddTypeName(gallivm->module, struct_name, texture_type);
95
96 /* Make sure the target's struct layout cache doesn't return
97 * stale/invalid data.
98 */
99 LLVMInvalidateStructLayout(gallivm->target, texture_type);
100 #endif
101
102 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
103 target, texture_type,
104 DRAW_JIT_TEXTURE_WIDTH);
105 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
106 target, texture_type,
107 DRAW_JIT_TEXTURE_HEIGHT);
108 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
109 target, texture_type,
110 DRAW_JIT_TEXTURE_DEPTH);
111 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
112 target, texture_type,
113 DRAW_JIT_TEXTURE_FIRST_LEVEL);
114 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
115 target, texture_type,
116 DRAW_JIT_TEXTURE_LAST_LEVEL);
117 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
118 target, texture_type,
119 DRAW_JIT_TEXTURE_BASE);
120 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
121 target, texture_type,
122 DRAW_JIT_TEXTURE_ROW_STRIDE);
123 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
124 target, texture_type,
125 DRAW_JIT_TEXTURE_IMG_STRIDE);
126 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
127 target, texture_type,
128 DRAW_JIT_TEXTURE_MIP_OFFSETS);
129
130 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
131
132 return texture_type;
133 }
134
135
136 /**
137 * Create LLVM type for struct draw_jit_sampler
138 */
139 static LLVMTypeRef
140 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
141 {
142 LLVMTargetDataRef target = gallivm->target;
143 LLVMTypeRef sampler_type;
144 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
145
146 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
147 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
148 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
149 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
150 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
151
152 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
153 Elements(elem_types), 0);
154
155 #if HAVE_LLVM < 0x0300
156 LLVMAddTypeName(gallivm->module, struct_name, sampler_type);
157
158 /* Make sure the target's struct layout cache doesn't return
159 * stale/invalid data.
160 */
161 LLVMInvalidateStructLayout(gallivm->target, sampler_type);
162 #endif
163
164 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
165 target, sampler_type,
166 DRAW_JIT_SAMPLER_MIN_LOD);
167 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
168 target, sampler_type,
169 DRAW_JIT_SAMPLER_MAX_LOD);
170 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
171 target, sampler_type,
172 DRAW_JIT_SAMPLER_LOD_BIAS);
173 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
174 target, sampler_type,
175 DRAW_JIT_SAMPLER_BORDER_COLOR);
176
177 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
178
179 return sampler_type;
180 }
181
182
183 /**
184 * Create LLVM type for struct draw_jit_context
185 */
186 static LLVMTypeRef
187 create_jit_context_type(struct gallivm_state *gallivm,
188 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
189 const char *struct_name)
190 {
191 LLVMTargetDataRef target = gallivm->target;
192 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
193 LLVMTypeRef elem_types[5];
194 LLVMTypeRef context_type;
195
196 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
197 LP_MAX_TGSI_CONST_BUFFERS);
198 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
199 DRAW_TOTAL_CLIP_PLANES), 0);
200 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
201 elem_types[3] = LLVMArrayType(texture_type,
202 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
203 elem_types[4] = LLVMArrayType(sampler_type,
204 PIPE_MAX_SAMPLERS); /* samplers */
205 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
206 Elements(elem_types), 0);
207 #if HAVE_LLVM < 0x0300
208 LLVMAddTypeName(gallivm->module, struct_name, context_type);
209
210 LLVMInvalidateStructLayout(gallivm->target, context_type);
211 #endif
212
213 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
214 target, context_type, 0);
215 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
216 target, context_type, 1);
217 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
218 target, context_type, 2);
219 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
220 target, context_type,
221 DRAW_JIT_CTX_TEXTURES);
222 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
223 target, context_type,
224 DRAW_JIT_CTX_SAMPLERS);
225 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
226 target, context_type);
227
228 return context_type;
229 }
230
231
232 /**
233 * Create LLVM type for struct draw_gs_jit_context
234 */
235 static LLVMTypeRef
236 create_gs_jit_context_type(struct gallivm_state *gallivm,
237 unsigned vector_length,
238 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
239 const char *struct_name)
240 {
241 LLVMTargetDataRef target = gallivm->target;
242 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
243 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
244 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
245 LLVMTypeRef context_type;
246
247 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
248 LP_MAX_TGSI_CONST_BUFFERS);
249 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
250 DRAW_TOTAL_CLIP_PLANES), 0);
251 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
252
253 elem_types[3] = LLVMArrayType(texture_type,
254 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
255 elem_types[4] = LLVMArrayType(sampler_type,
256 PIPE_MAX_SAMPLERS); /* samplers */
257
258 elem_types[5] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
259 elem_types[6] = LLVMPointerType(LLVMVectorType(int_type,
260 vector_length), 0);
261 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
262 vector_length), 0);
263
264 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
265 Elements(elem_types), 0);
266 #if HAVE_LLVM < 0x0300
267 LLVMAddTypeName(gallivm->module, struct_name, context_type);
268
269 LLVMInvalidateStructLayout(gallivm->target, context_type);
270 #endif
271
272 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
273 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
274 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
275 target, context_type, DRAW_GS_JIT_CTX_PLANES);
276 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
277 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
278 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
279 target, context_type,
280 DRAW_GS_JIT_CTX_TEXTURES);
281 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
282 target, context_type,
283 DRAW_GS_JIT_CTX_SAMPLERS);
284 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
285 target, context_type,
286 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
287 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
288 target, context_type,
289 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
290 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
291 target, context_type,
292 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
293 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
294 target, context_type);
295
296 return context_type;
297 }
298
299
300 static LLVMTypeRef
301 create_gs_jit_input_type(struct gallivm_state *gallivm)
302 {
303 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
304 LLVMTypeRef input_array;
305
306 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
307 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
308 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
309 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
310
311 return input_array;
312 }
313
314 /**
315 * Create LLVM type for struct pipe_vertex_buffer
316 */
317 static LLVMTypeRef
318 create_jit_vertex_buffer_type(struct gallivm_state *gallivm, const char *struct_name)
319 {
320 LLVMTargetDataRef target = gallivm->target;
321 LLVMTypeRef elem_types[4];
322 LLVMTypeRef vb_type;
323
324 elem_types[0] =
325 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
326 elem_types[2] =
327 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0); /* vs_constants */
328
329 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
330 Elements(elem_types), 0);
331 #if HAVE_LLVM < 0x0300
332 LLVMAddTypeName(gallivm->module, struct_name, vb_type);
333
334 LLVMInvalidateStructLayout(gallivm->target, vb_type);
335 #endif
336
337 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
338 target, vb_type, 0);
339 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
340 target, vb_type, 1);
341
342 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
343
344 return vb_type;
345 }
346
347
348 /**
349 * Create LLVM type for struct vertex_header;
350 */
351 static LLVMTypeRef
352 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
353 {
354 LLVMTargetDataRef target = gallivm->target;
355 LLVMTypeRef elem_types[4];
356 LLVMTypeRef vertex_header;
357 char struct_name[24];
358
359 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
360
361 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
362 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
363 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
364 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
365
366 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
367 Elements(elem_types), 0);
368 #if HAVE_LLVM < 0x0300
369 LLVMAddTypeName(gallivm->module, struct_name, vertex_header);
370
371 LLVMInvalidateStructLayout(gallivm->target, vertex_header);
372 #endif
373
374 /* these are bit-fields and we can't take address of them
375 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
376 target, vertex_header,
377 DRAW_JIT_VERTEX_CLIPMASK);
378 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
379 target, vertex_header,
380 DRAW_JIT_VERTEX_EDGEFLAG);
381 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
382 target, vertex_header,
383 DRAW_JIT_VERTEX_PAD);
384 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
385 target, vertex_header,
386 DRAW_JIT_VERTEX_VERTEX_ID);
387 */
388 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
389 target, vertex_header,
390 DRAW_JIT_VERTEX_CLIP);
391 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
392 target, vertex_header,
393 DRAW_JIT_VERTEX_PRE_CLIP_POS);
394 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
395 target, vertex_header,
396 DRAW_JIT_VERTEX_DATA);
397
398 assert(LLVMABISizeOfType(target, vertex_header) ==
399 offsetof(struct vertex_header, data[data_elems]));
400
401 return vertex_header;
402 }
403
404
405 /**
406 * Create LLVM types for various structures.
407 */
408 static void
409 create_jit_types(struct draw_llvm_variant *variant)
410 {
411 struct gallivm_state *gallivm = variant->gallivm;
412 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type, vb_type;
413
414 texture_type = create_jit_texture_type(gallivm, "texture");
415 sampler_type = create_jit_sampler_type(gallivm, "sampler");
416
417 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
418 "draw_jit_context");
419 variant->context_ptr_type = LLVMPointerType(context_type, 0);
420
421 buffer_type = LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
422 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
423
424 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
425 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
426 }
427
428
429 static LLVMTypeRef
430 get_context_ptr_type(struct draw_llvm_variant *variant)
431 {
432 if (!variant->context_ptr_type)
433 create_jit_types(variant);
434 return variant->context_ptr_type;
435 }
436
437
438 static LLVMTypeRef
439 get_buffer_ptr_type(struct draw_llvm_variant *variant)
440 {
441 if (!variant->buffer_ptr_type)
442 create_jit_types(variant);
443 return variant->buffer_ptr_type;
444 }
445
446
447 static LLVMTypeRef
448 get_vb_ptr_type(struct draw_llvm_variant *variant)
449 {
450 if (!variant->vb_ptr_type)
451 create_jit_types(variant);
452 return variant->vb_ptr_type;
453 }
454
455 static LLVMTypeRef
456 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
457 {
458 if (!variant->vertex_header_ptr_type)
459 create_jit_types(variant);
460 return variant->vertex_header_ptr_type;
461 }
462
463
464 /**
465 * Create per-context LLVM info.
466 */
467 struct draw_llvm *
468 draw_llvm_create(struct draw_context *draw)
469 {
470 struct draw_llvm *llvm;
471
472 llvm = CALLOC_STRUCT( draw_llvm );
473 if (!llvm)
474 return NULL;
475
476 lp_build_init();
477
478 llvm->draw = draw;
479
480 llvm->nr_variants = 0;
481 make_empty_list(&llvm->vs_variants_list);
482
483 llvm->nr_gs_variants = 0;
484 make_empty_list(&llvm->gs_variants_list);
485
486 return llvm;
487 }
488
489
490 /**
491 * Free per-context LLVM info.
492 */
493 void
494 draw_llvm_destroy(struct draw_llvm *llvm)
495 {
496 /* XXX free other draw_llvm data? */
497 FREE(llvm);
498 }
499
500
501 /**
502 * Create LLVM-generated code for a vertex shader.
503 */
504 struct draw_llvm_variant *
505 draw_llvm_create_variant(struct draw_llvm *llvm,
506 unsigned num_inputs,
507 const struct draw_llvm_variant_key *key)
508 {
509 struct draw_llvm_variant *variant;
510 struct llvm_vertex_shader *shader =
511 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
512 LLVMTypeRef vertex_header;
513
514 variant = MALLOC(sizeof *variant +
515 shader->variant_key_size -
516 sizeof variant->key);
517 if (variant == NULL)
518 return NULL;
519
520 variant->llvm = llvm;
521
522 variant->gallivm = gallivm_create();
523
524 create_jit_types(variant);
525
526 memcpy(&variant->key, key, shader->variant_key_size);
527
528 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
529
530 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
531
532 draw_llvm_generate(llvm, variant, FALSE); /* linear */
533 draw_llvm_generate(llvm, variant, TRUE); /* elts */
534
535 gallivm_compile_module(variant->gallivm);
536
537 variant->jit_func = (draw_jit_vert_func)
538 gallivm_jit_function(variant->gallivm, variant->function);
539
540 variant->jit_func_elts = (draw_jit_vert_func_elts)
541 gallivm_jit_function(variant->gallivm, variant->function_elts);
542
543 variant->shader = shader;
544 variant->list_item_global.base = variant;
545 variant->list_item_local.base = variant;
546 /*variant->no = */shader->variants_created++;
547 variant->list_item_global.base = variant;
548
549 return variant;
550 }
551
552
553 static void
554 generate_vs(struct draw_llvm_variant *variant,
555 LLVMBuilderRef builder,
556 struct lp_type vs_type,
557 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
558 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
559 const struct lp_bld_tgsi_system_values *system_values,
560 LLVMValueRef context_ptr,
561 struct lp_build_sampler_soa *draw_sampler,
562 boolean clamp_vertex_color)
563 {
564 struct draw_llvm *llvm = variant->llvm;
565 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
566 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(variant->gallivm, context_ptr);
567 struct lp_build_sampler_soa *sampler = 0;
568
569 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
570 tgsi_dump(tokens, 0);
571 draw_llvm_dump_variant_key(&variant->key);
572 }
573
574 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
575 sampler = draw_sampler;
576
577 lp_build_tgsi_soa(variant->gallivm,
578 tokens,
579 vs_type,
580 NULL /*struct lp_build_mask_context *mask*/,
581 consts_ptr,
582 system_values,
583 NULL /*pos*/,
584 inputs,
585 outputs,
586 sampler,
587 &llvm->draw->vs.vertex_shader->info,
588 NULL);
589
590 {
591 LLVMValueRef out;
592 unsigned chan, attrib;
593 struct lp_build_context bld;
594 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
595 lp_build_context_init(&bld, variant->gallivm, vs_type);
596
597 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
598 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
599 if (outputs[attrib][chan]) {
600 switch (info->output_semantic_name[attrib]) {
601 case TGSI_SEMANTIC_COLOR:
602 case TGSI_SEMANTIC_BCOLOR:
603 if (clamp_vertex_color) {
604 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
605 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
606 LLVMBuildStore(builder, out, outputs[attrib][chan]);
607 }
608 break;
609 case TGSI_SEMANTIC_FOG:
610 if (chan == 1 || chan == 2)
611 LLVMBuildStore(builder, bld.zero, outputs[attrib][chan]);
612 else if (chan == 3)
613 LLVMBuildStore(builder, bld.one, outputs[attrib][chan]);
614 break;
615 }
616 }
617 }
618 }
619 }
620 }
621
622
623 static void
624 generate_fetch(struct gallivm_state *gallivm,
625 LLVMValueRef vbuffers_ptr,
626 LLVMValueRef *res,
627 struct pipe_vertex_element *velem,
628 LLVMValueRef vbuf,
629 LLVMValueRef index,
630 LLVMValueRef instance_id)
631 {
632 const struct util_format_description *format_desc = util_format_description(velem->src_format);
633 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
634 LLVMBuilderRef builder = gallivm->builder;
635 LLVMValueRef indices =
636 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
637 velem->vertex_buffer_index, 0);
638 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
639 &indices, 1, "");
640 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
641 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
642 LLVMValueRef stride;
643
644 if (velem->instance_divisor) {
645 /* array index = instance_id / instance_divisor */
646 index = LLVMBuildUDiv(builder, instance_id,
647 lp_build_const_int32(gallivm, velem->instance_divisor),
648 "instance_divisor");
649 }
650
651 stride = LLVMBuildMul(builder, vb_stride, index, "");
652
653 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
654
655 stride = LLVMBuildAdd(builder, stride,
656 vb_buffer_offset,
657 "");
658 stride = LLVMBuildAdd(builder, stride,
659 lp_build_const_int32(gallivm, velem->src_offset),
660 "");
661
662 /* lp_build_printf(gallivm, "vbuf index = %d, stride is %d\n", indices, stride);*/
663 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
664
665 *res = lp_build_fetch_rgba_aos(gallivm,
666 format_desc,
667 lp_float32_vec4_type(),
668 vbuffer_ptr,
669 zero, zero, zero);
670 }
671
672 static void
673 convert_to_soa(struct gallivm_state *gallivm,
674 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
675 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
676 unsigned num_attribs, const struct lp_type soa_type)
677 {
678 unsigned i, j, k;
679 struct lp_type aos_channel_type = soa_type;
680
681 debug_assert(TGSI_NUM_CHANNELS == 4);
682 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
683
684 aos_channel_type.length >>= 1;
685
686 for (i = 0; i < num_attribs; ++i) {
687 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
688 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
689
690 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
691 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
692
693 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
694
695 for (k = 0; k < pixels_per_channel; ++k) {
696 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
697 }
698
699 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
700 }
701
702 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
703 }
704 }
705
706
707 static void
708 store_aos(struct gallivm_state *gallivm,
709 LLVMValueRef io_ptr,
710 LLVMValueRef index,
711 LLVMValueRef value)
712 {
713 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
714 LLVMBuilderRef builder = gallivm->builder;
715 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
716 LLVMValueRef indices[3];
717
718 indices[0] = lp_build_const_int32(gallivm, 0);
719 indices[1] = index;
720 indices[2] = lp_build_const_int32(gallivm, 0);
721
722 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
723 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
724
725 #if DEBUG_STORE
726 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
727 #endif
728
729 /* Unaligned store due to the vertex header */
730 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
731 }
732
733 /**
734 * Adjust the mask to architecture endianess. The mask will the store in struct:
735 *
736 * struct vertex_header {
737 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
738 * unsigned edgeflag:1;
739 * unsigned have_clipdist:1;
740 * unsigned vertex_id:16;
741 * [...]
742 * }
743 *
744 * On little-endian machine nothing needs to done, however on bit-endian machine
745 * the mask's fields need to be adjusted with the algorithm:
746 *
747 * uint32_t reverse (uint32_t x)
748 * {
749 * return (x >> 16) | // vertex_id
750 * ((x & 0x3fff) << 18) | // clipmask
751 * ((x & 0x4000) << 3) | // have_clipdist
752 * ((x & 0x8000) << 1); // edgeflag
753 * }
754 */
755 static LLVMValueRef
756 adjust_mask(struct gallivm_state *gallivm,
757 LLVMValueRef mask)
758 {
759 #ifdef PIPE_ARCH_BIG_ENDIAN
760 LLVMBuilderRef builder = gallivm->builder;
761 LLVMValueRef vertex_id;
762 LLVMValueRef clipmask;
763 LLVMValueRef have_clipdist;
764 LLVMValueRef edgeflag;
765
766 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
767 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
768 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
769 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
770 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
771 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
772 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
773
774 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
775 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
776 mask = LLVMBuildOr(builder, mask, edgeflag, "");
777 #endif
778 return mask;
779 }
780
781 static void
782 store_aos_array(struct gallivm_state *gallivm,
783 struct lp_type soa_type,
784 LLVMValueRef io_ptr,
785 LLVMValueRef *indices,
786 LLVMValueRef* aos,
787 int attrib,
788 int num_outputs,
789 LLVMValueRef clipmask,
790 boolean have_clipdist)
791 {
792 LLVMBuilderRef builder = gallivm->builder;
793 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
794 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
795 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
796 int vector_length = soa_type.length;
797 int i;
798
799 debug_assert(TGSI_NUM_CHANNELS == 4);
800
801 for (i = 0; i < vector_length; i++) {
802 if (indices) {
803 inds[i] = indices[i];
804 } else {
805 inds[i] = lp_build_const_int32(gallivm, i);
806 }
807 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
808 }
809
810 if (attrib == 0) {
811 /* store vertex header for each of the n vertices */
812 LLVMValueRef val, cliptmp;
813 int vertex_id_pad_edgeflag;
814
815 /* If this assertion fails, it means we need to update the bit twidding
816 * code here. See struct vertex_header in draw_private.h.
817 */
818 assert(DRAW_TOTAL_CLIP_PLANES==14);
819 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
820 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
821 if (have_clipdist)
822 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
823 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
824 /* OR with the clipmask */
825 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
826 for (i = 0; i < vector_length; i++) {
827 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
828 val = LLVMBuildExtractElement(builder, cliptmp, inds[i], "");
829 val = adjust_mask(gallivm, val);
830 LLVMBuildStore(builder, val, id_ptr);
831 #if DEBUG_STORE
832 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
833 io_ptrs[i], inds[i], val);
834 #endif
835 }
836 }
837
838 /* store for each of the n vertices */
839 for (i = 0; i < vector_length; i++) {
840 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
841 }
842 }
843
844
845 static void
846 convert_to_aos(struct gallivm_state *gallivm,
847 LLVMValueRef io,
848 LLVMValueRef *indices,
849 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
850 LLVMValueRef clipmask,
851 int num_outputs,
852 struct lp_type soa_type,
853 boolean have_clipdist)
854 {
855 LLVMBuilderRef builder = gallivm->builder;
856 unsigned chan, attrib, i;
857
858 #if DEBUG_STORE
859 lp_build_printf(gallivm, " # storing begin\n");
860 #endif
861 for (attrib = 0; attrib < num_outputs; ++attrib) {
862 LLVMValueRef soa[TGSI_NUM_CHANNELS];
863 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
864 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
865 if (outputs[attrib][chan]) {
866 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
867 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
868 #if DEBUG_STORE
869 lp_build_printf(gallivm, "output %d : %d ",
870 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
871 attrib, 0),
872 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
873 chan, 0));
874 lp_build_print_value(gallivm, "val = ", out);
875 #endif
876 soa[chan] = out;
877 }
878 else {
879 soa[chan] = 0;
880 }
881 }
882
883
884 if (soa_type.length == TGSI_NUM_CHANNELS) {
885 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
886 } else {
887 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
888
889 for (i = 0; i < soa_type.length; ++i) {
890 aos[i] = lp_build_extract_range(gallivm,
891 soa[i % TGSI_NUM_CHANNELS],
892 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
893 TGSI_NUM_CHANNELS);
894 }
895 }
896
897 store_aos_array(gallivm,
898 soa_type,
899 io, indices,
900 aos,
901 attrib,
902 num_outputs,
903 clipmask, have_clipdist);
904 }
905 #if DEBUG_STORE
906 lp_build_printf(gallivm, " # storing end\n");
907 #endif
908 }
909
910
911 /**
912 * Stores original vertex positions in clip coordinates
913 */
914 static void
915 store_clip(struct gallivm_state *gallivm,
916 const struct lp_type vs_type,
917 LLVMValueRef io_ptr,
918 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
919 boolean pre_clip_pos, int idx)
920 {
921 LLVMBuilderRef builder = gallivm->builder;
922 LLVMValueRef soa[4];
923 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
924 LLVMValueRef indices[2];
925 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
926 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
927 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
928 int i, j;
929
930 indices[0] =
931 indices[1] = lp_build_const_int32(gallivm, 0);
932
933 for (i = 0; i < vs_type.length; i++) {
934 inds[i] = lp_build_const_int32(gallivm, i);
935 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
936 }
937
938 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
939 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
940 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
941 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
942
943 if (!pre_clip_pos) {
944 for (i = 0; i < vs_type.length; i++) {
945 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
946 }
947 } else {
948 for (i = 0; i < vs_type.length; i++) {
949 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
950 }
951 }
952
953 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
954 for (i = 0; i < vs_type.length; ++i) {
955 aos[i] = lp_build_extract_range(gallivm,
956 soa[i % TGSI_NUM_CHANNELS],
957 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
958 TGSI_NUM_CHANNELS);
959 }
960
961 for (j = 0; j < vs_type.length; j++) {
962 LLVMTypeRef clip_ptr_type = LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0);
963 LLVMValueRef clip_ptr;
964
965 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
966 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
967
968 /* Unaligned store */
969 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
970 }
971 }
972
973
974 /**
975 * Transforms the outputs for viewport mapping
976 */
977 static void
978 generate_viewport(struct draw_llvm_variant *variant,
979 LLVMBuilderRef builder,
980 struct lp_type vs_type,
981 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
982 LLVMValueRef context_ptr)
983 {
984 int i;
985 struct gallivm_state *gallivm = variant->gallivm;
986 struct lp_type f32_type = vs_type;
987 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
988 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 .. wn*/
989 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
990 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
991
992 /* for 1/w convention*/
993 out3 = LLVMBuildFDiv(builder, const1, out3, "");
994 LLVMBuildStore(builder, out3, outputs[0][3]);
995
996 /* Viewport Mapping */
997 for (i=0; i<3; i++) {
998 LLVMValueRef out = LLVMBuildLoad(builder, outputs[0][i], ""); /*x0 x1 .. xn*/
999 LLVMValueRef scale;
1000 LLVMValueRef trans;
1001 LLVMValueRef scale_i;
1002 LLVMValueRef trans_i;
1003 LLVMValueRef index;
1004
1005 index = lp_build_const_int32(gallivm, i);
1006 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1007
1008 index = lp_build_const_int32(gallivm, i+4);
1009 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1010
1011 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1012 LLVMBuildLoad(builder, scale_i, "scale"));
1013 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1014 LLVMBuildLoad(builder, trans_i, "trans"));
1015
1016 /* divide by w */
1017 out = LLVMBuildFMul(builder, out, out3, "");
1018 /* mult by scale */
1019 out = LLVMBuildFMul(builder, out, scale, "");
1020 /* add translation */
1021 out = LLVMBuildFAdd(builder, out, trans, "");
1022
1023 /* store transformed outputs */
1024 LLVMBuildStore(builder, out, outputs[0][i]);
1025 }
1026
1027 }
1028
1029
1030 /**
1031 * Returns clipmask as nxi32 bitmask for the n vertices
1032 */
1033 static LLVMValueRef
1034 generate_clipmask(struct draw_llvm *llvm,
1035 struct gallivm_state *gallivm,
1036 struct lp_type vs_type,
1037 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1038 boolean clip_xy,
1039 boolean clip_z,
1040 boolean clip_user,
1041 boolean clip_halfz,
1042 unsigned ucp_enable,
1043 LLVMValueRef context_ptr,
1044 boolean *have_clipdist)
1045 {
1046 LLVMBuilderRef builder = gallivm->builder;
1047 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1048 LLVMValueRef test, temp;
1049 LLVMValueRef zero, shift;
1050 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1051 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1052 LLVMValueRef plane1, planes, plane_ptr, sum;
1053 struct lp_type f32_type = vs_type;
1054 struct lp_type i32_type = lp_int_type(vs_type);
1055 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1056 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1057 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1058 bool have_cd = false;
1059 unsigned cd[2];
1060
1061 cd[0] = draw_current_shader_clipdistance_output(llvm->draw, 0);
1062 cd[1] = draw_current_shader_clipdistance_output(llvm->draw, 1);
1063
1064 if (cd[0] != pos || cd[1] != pos)
1065 have_cd = true;
1066
1067 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1068 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1069 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1070 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1071
1072 /*
1073 * load clipvertex and position from correct locations.
1074 * if they are the same just load them once.
1075 */
1076 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1077 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1078 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1079 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1080
1081 if (clip_user && cv != pos) {
1082 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1083 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1084 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1085 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1086 } else {
1087 cv_x = pos_x;
1088 cv_y = pos_y;
1089 cv_z = pos_z;
1090 cv_w = pos_w;
1091 }
1092
1093 /* Cliptest, for hardwired planes */
1094 if (clip_xy) {
1095 /* plane 1 */
1096 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1097 temp = shift;
1098 test = LLVMBuildAnd(builder, test, temp, "");
1099 mask = test;
1100
1101 /* plane 2 */
1102 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1103 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1104 temp = LLVMBuildShl(builder, temp, shift, "");
1105 test = LLVMBuildAnd(builder, test, temp, "");
1106 mask = LLVMBuildOr(builder, mask, test, "");
1107
1108 /* plane 3 */
1109 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1110 temp = LLVMBuildShl(builder, temp, shift, "");
1111 test = LLVMBuildAnd(builder, test, temp, "");
1112 mask = LLVMBuildOr(builder, mask, test, "");
1113
1114 /* plane 4 */
1115 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1116 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1117 temp = LLVMBuildShl(builder, temp, shift, "");
1118 test = LLVMBuildAnd(builder, test, temp, "");
1119 mask = LLVMBuildOr(builder, mask, test, "");
1120 }
1121
1122 if (clip_z) {
1123 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1124 if (clip_halfz) {
1125 /* plane 5 */
1126 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1127 test = LLVMBuildAnd(builder, test, temp, "");
1128 mask = LLVMBuildOr(builder, mask, test, "");
1129 }
1130 else {
1131 /* plane 5 */
1132 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1133 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1134 test = LLVMBuildAnd(builder, test, temp, "");
1135 mask = LLVMBuildOr(builder, mask, test, "");
1136 }
1137 /* plane 6 */
1138 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1139 temp = LLVMBuildShl(builder, temp, shift, "");
1140 test = LLVMBuildAnd(builder, test, temp, "");
1141 mask = LLVMBuildOr(builder, mask, test, "");
1142 }
1143
1144 if (clip_user) {
1145 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1146 LLVMValueRef indices[3];
1147
1148 /* userclip planes */
1149 while (ucp_enable) {
1150 unsigned plane_idx = ffs(ucp_enable)-1;
1151 ucp_enable &= ~(1 << plane_idx);
1152 plane_idx += 6;
1153
1154 if (have_cd && num_written_clipdistance) {
1155 LLVMValueRef clipdist;
1156 int i;
1157 i = plane_idx - 6;
1158
1159 *have_clipdist = TRUE;
1160 if (i < 4) {
1161 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1162 } else {
1163 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1164 }
1165 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1166 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1167 test = LLVMBuildAnd(builder, test, temp, "");
1168 mask = LLVMBuildOr(builder, mask, test, "");
1169 } else {
1170 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1171 indices[0] = lp_build_const_int32(gallivm, 0);
1172 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1173
1174 indices[2] = lp_build_const_int32(gallivm, 0);
1175 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1176 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1177 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1178 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1179
1180 indices[2] = lp_build_const_int32(gallivm, 1);
1181 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1182 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1183 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1184 test = LLVMBuildFMul(builder, planes, cv_y, "");
1185 sum = LLVMBuildFAdd(builder, sum, test, "");
1186
1187 indices[2] = lp_build_const_int32(gallivm, 2);
1188 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1189 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1190 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1191 test = LLVMBuildFMul(builder, planes, cv_z, "");
1192 sum = LLVMBuildFAdd(builder, sum, test, "");
1193
1194 indices[2] = lp_build_const_int32(gallivm, 3);
1195 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1196 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1197 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1198 test = LLVMBuildFMul(builder, planes, cv_w, "");
1199 sum = LLVMBuildFAdd(builder, sum, test, "");
1200
1201 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1202 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1203 test = LLVMBuildAnd(builder, test, temp, "");
1204 mask = LLVMBuildOr(builder, mask, test, "");
1205 }
1206 }
1207 }
1208 return mask;
1209 }
1210
1211
1212 /**
1213 * Returns boolean if any clipping has occurred
1214 * Used zero/non-zero i32 value to represent boolean
1215 */
1216 static LLVMValueRef
1217 clipmask_booli32(struct gallivm_state *gallivm,
1218 const struct lp_type vs_type,
1219 LLVMValueRef clipmask_bool_ptr)
1220 {
1221 LLVMBuilderRef builder = gallivm->builder;
1222 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1223 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1224 LLVMValueRef ret = LLVMConstNull(int32_type);
1225 LLVMValueRef temp;
1226 int i;
1227
1228 /*
1229 * Can do this with log2(vector length) pack instructions and one extract
1230 * (as we don't actually need a or) with sse2 which would be way better.
1231 */
1232 for (i=0; i < vs_type.length; i++) {
1233 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1234 lp_build_const_int32(gallivm, i) , "");
1235 ret = LLVMBuildOr(builder, ret, temp, "");
1236 }
1237 return ret;
1238 }
1239
1240 static void
1241 draw_gs_llvm_emit_vertex(struct lp_build_tgsi_context * bld_base,
1242 LLVMValueRef (*outputs)[4],
1243 LLVMValueRef emitted_vertices_vec,
1244 void *user_data)
1245 {
1246 struct draw_gs_llvm_variant *variant =
1247 (struct draw_gs_llvm_variant *)user_data;
1248 struct gallivm_state *gallivm = variant->gallivm;
1249 LLVMBuilderRef builder = gallivm->builder;
1250 struct lp_type gs_type = bld_base->base.type;
1251 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1252 lp_int_type(gs_type), 0);
1253 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1254 LLVMValueRef max_output_vertices =
1255 lp_build_const_int32(gallivm, variant->shader->base.max_output_vertices);
1256 LLVMValueRef io = variant->io_ptr;
1257 unsigned i;
1258 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1259
1260 for (i = 0; i < gs_type.length; ++i) {
1261 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1262 LLVMValueRef currently_emitted =
1263 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1264 indices[i] = LLVMBuildMul(builder, ind, max_output_vertices, "");
1265 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1266 }
1267
1268 convert_to_aos(gallivm, io, indices,
1269 outputs, clipmask,
1270 gs_info->num_outputs, gs_type,
1271 FALSE);
1272 }
1273
1274 static void
1275 draw_gs_llvm_end_primitive(struct lp_build_tgsi_context * bld_base,
1276 LLVMValueRef verts_per_prim_vec,
1277 LLVMValueRef emitted_prims_vec,
1278 void *user_data)
1279 {
1280 struct draw_gs_llvm_variant *variant =
1281 (struct draw_gs_llvm_variant *)user_data;
1282 struct gallivm_state *gallivm = variant->gallivm;
1283 LLVMBuilderRef builder = gallivm->builder;
1284 LLVMValueRef prim_lengts_ptr =
1285 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1286 unsigned i;
1287
1288 for (i = 0; i < bld_base->base.type.length; ++i) {
1289 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1290 LLVMValueRef prims_emitted =
1291 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1292 LLVMValueRef store_ptr;
1293 LLVMValueRef num_vertices =
1294 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1295
1296 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1297 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1298 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1299 LLVMBuildStore(builder, num_vertices, store_ptr);
1300 }
1301 }
1302
1303 static void
1304 draw_gs_llvm_epilogue(struct lp_build_tgsi_context * bld_base,
1305 LLVMValueRef total_emitted_vertices_vec,
1306 LLVMValueRef emitted_prims_vec,
1307 void *user_data)
1308 {
1309 struct draw_gs_llvm_variant *variant =
1310 (struct draw_gs_llvm_variant *)user_data;
1311 struct gallivm_state *gallivm = variant->gallivm;
1312 LLVMBuilderRef builder = gallivm->builder;
1313 LLVMValueRef emitted_verts_ptr =
1314 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1315 LLVMValueRef emitted_prims_ptr =
1316 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1317 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1318
1319 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1320 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1321
1322 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1323 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1324 }
1325
1326 static void
1327 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1328 boolean elts)
1329 {
1330 struct gallivm_state *gallivm = variant->gallivm;
1331 LLVMContextRef context = gallivm->context;
1332 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1333 LLVMTypeRef arg_types[8];
1334 LLVMTypeRef func_type;
1335 LLVMValueRef context_ptr;
1336 LLVMBasicBlockRef block;
1337 LLVMBuilderRef builder;
1338 struct lp_type vs_type;
1339 LLVMValueRef end, start;
1340 LLVMValueRef count, fetch_elts, fetch_count;
1341 LLVMValueRef stride, step, io_itr;
1342 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1343 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1344 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1345 struct draw_context *draw = llvm->draw;
1346 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1347 unsigned i, j;
1348 struct lp_build_context bld;
1349 struct lp_build_loop_state lp_loop;
1350 const int vector_length = lp_native_vector_width / 32;
1351 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1352 LLVMValueRef fetch_max;
1353 struct lp_build_sampler_soa *sampler = 0;
1354 LLVMValueRef ret, clipmask_bool_ptr;
1355 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1356 /* If geometry shader is present we need to skip both the viewport
1357 * transformation and clipping otherwise the inputs to the geometry
1358 * shader will be incorrect.
1359 */
1360 const boolean bypass_viewport = gs || variant->key.bypass_viewport;
1361 const boolean enable_cliptest = !gs && (variant->key.clip_xy ||
1362 variant->key.clip_z ||
1363 variant->key.clip_user);
1364 LLVMValueRef variant_func;
1365 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1366 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1367 boolean have_clipdist = FALSE;
1368 struct lp_bld_tgsi_system_values system_values;
1369
1370 memset(&system_values, 0, sizeof(system_values));
1371
1372 arg_types[0] = get_context_ptr_type(variant); /* context */
1373 arg_types[1] = get_vertex_header_ptr_type(variant); /* vertex_header */
1374 arg_types[2] = get_buffer_ptr_type(variant); /* vbuffers */
1375 if (elts)
1376 arg_types[3] = LLVMPointerType(int32_type, 0);/* fetch_elts * */
1377 else
1378 arg_types[3] = int32_type; /* start */
1379 arg_types[4] = int32_type; /* fetch_count / count */
1380 arg_types[5] = int32_type; /* stride */
1381 arg_types[6] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1382 arg_types[7] = int32_type; /* instance_id */
1383
1384 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1385
1386 variant_func = LLVMAddFunction(gallivm->module,
1387 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1388 func_type);
1389
1390 if (elts)
1391 variant->function_elts = variant_func;
1392 else
1393 variant->function = variant_func;
1394
1395 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1396 for (i = 0; i < Elements(arg_types); ++i)
1397 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1398 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1399 LLVMNoAliasAttribute);
1400
1401 context_ptr = LLVMGetParam(variant_func, 0);
1402 io_ptr = LLVMGetParam(variant_func, 1);
1403 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1404 stride = LLVMGetParam(variant_func, 5);
1405 vb_ptr = LLVMGetParam(variant_func, 6);
1406 system_values.instance_id = LLVMGetParam(variant_func, 7);
1407
1408 lp_build_name(context_ptr, "context");
1409 lp_build_name(io_ptr, "io");
1410 lp_build_name(vbuffers_ptr, "vbuffers");
1411 lp_build_name(stride, "stride");
1412 lp_build_name(vb_ptr, "vb");
1413 lp_build_name(system_values.instance_id, "instance_id");
1414
1415 if (elts) {
1416 fetch_elts = LLVMGetParam(variant_func, 3);
1417 fetch_count = LLVMGetParam(variant_func, 4);
1418 lp_build_name(fetch_elts, "fetch_elts");
1419 lp_build_name(fetch_count, "fetch_count");
1420 start = count = NULL;
1421 }
1422 else {
1423 start = LLVMGetParam(variant_func, 3);
1424 count = LLVMGetParam(variant_func, 4);
1425 lp_build_name(start, "start");
1426 lp_build_name(count, "count");
1427 fetch_elts = fetch_count = NULL;
1428 }
1429
1430 /*
1431 * Function body
1432 */
1433
1434 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1435 builder = gallivm->builder;
1436 LLVMPositionBuilderAtEnd(builder, block);
1437
1438 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1439
1440 memset(&vs_type, 0, sizeof vs_type);
1441 vs_type.floating = TRUE; /* floating point values */
1442 vs_type.sign = TRUE; /* values are signed */
1443 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1444 vs_type.width = 32; /* 32-bit float */
1445 vs_type.length = vector_length;
1446
1447 /* hold temporary "bool" clipmask */
1448 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1449 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1450
1451 /* code generated texture sampling */
1452 sampler = draw_llvm_sampler_soa_create(
1453 draw_llvm_variant_key_samplers(&variant->key),
1454 context_ptr);
1455
1456 if (elts) {
1457 start = zero;
1458 end = fetch_count;
1459 }
1460 else {
1461 end = lp_build_add(&bld, start, count);
1462 }
1463
1464 step = lp_build_const_int32(gallivm, vector_length);
1465
1466 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1467
1468 lp_build_loop_begin(&lp_loop, gallivm, start);
1469 {
1470 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1471 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1472 LLVMValueRef io;
1473 LLVMValueRef clipmask; /* holds the clipmask value */
1474 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1475
1476 if (elts)
1477 io_itr = lp_loop.counter;
1478 else
1479 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
1480
1481 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1482 #if DEBUG_STORE
1483 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1484 io_itr, io, lp_loop.counter);
1485 #endif
1486 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1487 for (i = 0; i < vector_length; ++i) {
1488 LLVMValueRef true_index =
1489 LLVMBuildAdd(builder,
1490 lp_loop.counter,
1491 lp_build_const_int32(gallivm, i), "");
1492
1493 /* make sure we're not out of bounds which can happen
1494 * if fetch_count % 4 != 0, because on the last iteration
1495 * a few of the 4 vertex fetches will be out of bounds */
1496 true_index = lp_build_min(&bld, true_index, fetch_max);
1497
1498 if (elts) {
1499 LLVMValueRef fetch_ptr;
1500 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1501 &true_index, 1, "");
1502 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
1503 }
1504
1505 system_values.vertex_id = LLVMBuildInsertElement(gallivm->builder,
1506 system_values.vertex_id, true_index,
1507 lp_build_const_int32(gallivm, i), "");
1508 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1509 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1510 LLVMValueRef vb_index =
1511 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1512 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1513 generate_fetch(gallivm, vbuffers_ptr,
1514 &aos_attribs[j][i], velem, vb, true_index,
1515 system_values.instance_id);
1516 }
1517 }
1518 convert_to_soa(gallivm, aos_attribs, inputs,
1519 draw->pt.nr_vertex_elements, vs_type);
1520
1521 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1522 generate_vs(variant,
1523 builder,
1524 vs_type,
1525 outputs,
1526 ptr_aos,
1527 &system_values,
1528 context_ptr,
1529 sampler,
1530 variant->key.clamp_vertex_color);
1531
1532 if (pos != -1 && cv != -1) {
1533 /* store original positions in clip before further manipulation */
1534 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1535 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1536
1537 /* do cliptest */
1538 if (enable_cliptest) {
1539 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1540 /* allocate clipmask, assign it integer type */
1541 clipmask = generate_clipmask(llvm,
1542 gallivm,
1543 vs_type,
1544 outputs,
1545 variant->key.clip_xy,
1546 variant->key.clip_z,
1547 variant->key.clip_user,
1548 variant->key.clip_halfz,
1549 variant->key.ucp_enable,
1550 context_ptr, &have_clipdist);
1551 temp = LLVMBuildOr(builder, clipmask, temp, "");
1552 /* store temporary clipping boolean value */
1553 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1554 }
1555 else {
1556 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1557 }
1558
1559 /* do viewport mapping */
1560 if (!bypass_viewport) {
1561 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1562 }
1563 }
1564 else {
1565 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1566 }
1567
1568 /* store clipmask in vertex header,
1569 * original positions in clip
1570 * and transformed positions in data
1571 */
1572 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1573 vs_info->num_outputs, vs_type,
1574 have_clipdist);
1575 }
1576
1577 lp_build_loop_end_cond(&lp_loop, end, step, LLVMIntUGE);
1578
1579 sampler->destroy(sampler);
1580
1581 /* return clipping boolean value for function */
1582 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1583
1584 LLVMBuildRet(builder, ret);
1585
1586 gallivm_verify_function(gallivm, variant_func);
1587 }
1588
1589
1590 struct draw_llvm_variant_key *
1591 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1592 {
1593 unsigned i;
1594 struct draw_llvm_variant_key *key;
1595 struct draw_sampler_static_state *draw_sampler;
1596
1597 key = (struct draw_llvm_variant_key *)store;
1598
1599 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1600
1601 /* Presumably all variants of the shader should have the same
1602 * number of vertex elements - ie the number of shader inputs.
1603 * NOTE: we NEED to store the needed number of needed inputs
1604 * here, not the number of provided elements to match keysize
1605 * (and the offset of sampler state in the key).
1606 */
1607 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1608 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1609
1610 /* will have to rig this up properly later */
1611 key->clip_xy = llvm->draw->clip_xy;
1612 key->clip_z = llvm->draw->clip_z;
1613 key->clip_user = llvm->draw->clip_user;
1614 key->bypass_viewport = llvm->draw->identity_viewport;
1615 key->clip_halfz = !llvm->draw->rasterizer->gl_rasterization_rules;
1616 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1617 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1618 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1619 key->pad1 = 0;
1620
1621 /* All variants of this shader will have the same value for
1622 * nr_samplers. Not yet trying to compact away holes in the
1623 * sampler array.
1624 */
1625 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1626 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1627 key->nr_sampler_views =
1628 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1629 }
1630 else {
1631 key->nr_sampler_views = key->nr_samplers;
1632 }
1633
1634 draw_sampler = draw_llvm_variant_key_samplers(key);
1635
1636 memcpy(key->vertex_element,
1637 llvm->draw->pt.vertex_element,
1638 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1639
1640 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1641
1642 for (i = 0 ; i < key->nr_samplers; i++) {
1643 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1644 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1645 }
1646 for (i = 0 ; i < key->nr_sampler_views; i++) {
1647 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1648 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1649 }
1650
1651 return key;
1652 }
1653
1654
1655 void
1656 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1657 {
1658 unsigned i;
1659 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1660
1661 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1662 debug_printf("clip_xy = %u\n", key->clip_xy);
1663 debug_printf("clip_z = %u\n", key->clip_z);
1664 debug_printf("clip_user = %u\n", key->clip_user);
1665 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1666 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1667 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1668 debug_printf("has_gs = %u\n", key->has_gs);
1669 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1670
1671 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1672 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1673 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1674 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1675 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1676 }
1677
1678 for (i = 0 ; i < key->nr_sampler_views; i++) {
1679 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1680 }
1681 }
1682
1683
1684 void
1685 draw_llvm_set_mapped_texture(struct draw_context *draw,
1686 unsigned shader_stage,
1687 unsigned sview_idx,
1688 uint32_t width, uint32_t height, uint32_t depth,
1689 uint32_t first_level, uint32_t last_level,
1690 const void *base_ptr,
1691 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1692 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1693 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1694 {
1695 unsigned j;
1696 struct draw_jit_texture *jit_tex;
1697
1698 assert(shader_stage == PIPE_SHADER_VERTEX ||
1699 shader_stage == PIPE_SHADER_GEOMETRY);
1700
1701 if (shader_stage == PIPE_SHADER_VERTEX) {
1702 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1703
1704 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1705 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1706 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1707
1708 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1709 }
1710
1711 jit_tex->width = width;
1712 jit_tex->height = height;
1713 jit_tex->depth = depth;
1714 jit_tex->first_level = first_level;
1715 jit_tex->last_level = last_level;
1716 jit_tex->base = base_ptr;
1717
1718 for (j = first_level; j <= last_level; j++) {
1719 jit_tex->mip_offsets[j] = mip_offsets[j];
1720 jit_tex->row_stride[j] = row_stride[j];
1721 jit_tex->img_stride[j] = img_stride[j];
1722 }
1723 }
1724
1725
1726 void
1727 draw_llvm_set_sampler_state(struct draw_context *draw,
1728 unsigned shader_type)
1729 {
1730 unsigned i;
1731
1732 if (shader_type == PIPE_SHADER_VERTEX) {
1733 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1734 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1735
1736 if (draw->samplers[i]) {
1737 const struct pipe_sampler_state *s
1738 = draw->samplers[PIPE_SHADER_VERTEX][i];
1739 jit_sam->min_lod = s->min_lod;
1740 jit_sam->max_lod = s->max_lod;
1741 jit_sam->lod_bias = s->lod_bias;
1742 COPY_4V(jit_sam->border_color, s->border_color.f);
1743 }
1744 }
1745 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1746 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1747 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1748
1749 if (draw->samplers[i]) {
1750 const struct pipe_sampler_state *s
1751 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1752 jit_sam->min_lod = s->min_lod;
1753 jit_sam->max_lod = s->max_lod;
1754 jit_sam->lod_bias = s->lod_bias;
1755 COPY_4V(jit_sam->border_color, s->border_color.f);
1756 }
1757 }
1758 }
1759 }
1760
1761
1762 void
1763 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1764 {
1765 struct draw_llvm *llvm = variant->llvm;
1766
1767 if (variant->function_elts) {
1768 gallivm_free_function(variant->gallivm,
1769 variant->function_elts, variant->jit_func_elts);
1770 }
1771
1772 if (variant->function) {
1773 gallivm_free_function(variant->gallivm,
1774 variant->function, variant->jit_func);
1775 }
1776
1777 gallivm_destroy(variant->gallivm);
1778
1779 remove_from_list(&variant->list_item_local);
1780 variant->shader->variants_cached--;
1781 remove_from_list(&variant->list_item_global);
1782 llvm->nr_variants--;
1783 FREE(variant);
1784 }
1785
1786
1787 /**
1788 * Create LLVM types for various structures.
1789 */
1790 static void
1791 create_gs_jit_types(struct draw_gs_llvm_variant *var)
1792 {
1793 struct gallivm_state *gallivm = var->gallivm;
1794 LLVMTypeRef texture_type, sampler_type, context_type;
1795
1796 texture_type = create_jit_texture_type(gallivm, "texture");
1797 sampler_type = create_jit_sampler_type(gallivm, "sampler");
1798
1799 context_type = create_gs_jit_context_type(gallivm,
1800 var->shader->base.vector_length,
1801 texture_type, sampler_type,
1802 "draw_gs_jit_context");
1803 var->context_ptr_type = LLVMPointerType(context_type, 0);
1804
1805 var->input_array_type = create_gs_jit_input_type(gallivm);
1806 }
1807
1808 static LLVMTypeRef
1809 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
1810 {
1811 if (!variant->context_ptr_type)
1812 create_gs_jit_types(variant);
1813 return variant->context_ptr_type;
1814 }
1815
1816 static LLVMValueRef
1817 generate_mask_value(struct draw_gs_llvm_variant *variant,
1818 struct lp_type gs_type)
1819 {
1820 struct gallivm_state *gallivm = variant->gallivm;
1821 LLVMBuilderRef builder = gallivm->builder;
1822 LLVMValueRef bits[16];
1823 struct lp_type mask_type = lp_int_type(gs_type);
1824 struct lp_type mask_elem_type = lp_elem_type(mask_type);
1825 LLVMValueRef mask_val = lp_build_const_vec(gallivm,
1826 mask_type,
1827 0);
1828 unsigned i;
1829
1830 assert(gs_type.length <= Elements(bits));
1831
1832 for (i = gs_type.length; i >= 1; --i) {
1833 int idx = i - 1;
1834 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1835 bits[idx] = lp_build_compare(gallivm,
1836 mask_elem_type, PIPE_FUNC_GEQUAL,
1837 variant->num_prims, ind);
1838 }
1839 for (i = 0; i < gs_type.length; ++i) {
1840 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1841 mask_val = LLVMBuildInsertElement(builder, mask_val, bits[i], ind, "");
1842 }
1843 mask_val = lp_build_compare(gallivm,
1844 mask_type, PIPE_FUNC_NOTEQUAL,
1845 mask_val,
1846 lp_build_const_int_vec(gallivm, mask_type, 0));
1847
1848 return mask_val;
1849 }
1850
1851 static void
1852 draw_gs_llvm_generate(struct draw_llvm *llvm,
1853 struct draw_gs_llvm_variant *variant)
1854 {
1855 struct gallivm_state *gallivm = variant->gallivm;
1856 LLVMContextRef context = gallivm->context;
1857 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1858 LLVMTypeRef arg_types[5];
1859 LLVMTypeRef func_type;
1860 LLVMValueRef variant_func;
1861 LLVMValueRef context_ptr;
1862 LLVMBasicBlockRef block;
1863 LLVMBuilderRef builder;
1864 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
1865 struct lp_build_sampler_soa *sampler = 0;
1866 struct lp_build_context bld;
1867 struct lp_bld_tgsi_system_values system_values;
1868 struct lp_type gs_type;
1869 unsigned i;
1870 struct lp_build_tgsi_gs_iface gs_iface;
1871 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
1872 LLVMValueRef consts_ptr;
1873 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1874 struct lp_build_mask_context mask;
1875
1876 memset(&system_values, 0, sizeof(system_values));
1877
1878 assert(variant->vertex_header_ptr_type);
1879
1880 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
1881 arg_types[1] = variant->input_array_type; /* input */
1882 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
1883 arg_types[3] = int32_type; /* num_prims */
1884 arg_types[4] = int32_type; /* instance_id */
1885
1886 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1887
1888 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
1889 func_type);
1890 variant->function = variant_func;
1891
1892 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1893
1894 for (i = 0; i < Elements(arg_types); ++i)
1895 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1896 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1897 LLVMNoAliasAttribute);
1898
1899 context_ptr = LLVMGetParam(variant_func, 0);
1900 input_array = LLVMGetParam(variant_func, 1);
1901 io_ptr = LLVMGetParam(variant_func, 2);
1902 num_prims = LLVMGetParam(variant_func, 3);
1903 system_values.instance_id = LLVMGetParam(variant_func, 4);
1904
1905 lp_build_name(context_ptr, "context");
1906 lp_build_name(input_array, "input");
1907 lp_build_name(io_ptr, "io");
1908 lp_build_name(io_ptr, "num_prims");
1909 lp_build_name(system_values.instance_id, "instance_id");
1910
1911 variant->context_ptr = context_ptr;
1912 variant->io_ptr = io_ptr;
1913 variant->num_prims = num_prims;
1914
1915 gs_iface.input = input_array;
1916 gs_iface.emit_vertex = draw_gs_llvm_emit_vertex;
1917 gs_iface.end_primitive = draw_gs_llvm_end_primitive;
1918 gs_iface.gs_epilogue = draw_gs_llvm_epilogue;
1919 gs_iface.user_data = variant;
1920
1921 /*
1922 * Function body
1923 */
1924
1925 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1926 builder = gallivm->builder;
1927 LLVMPositionBuilderAtEnd(builder, block);
1928
1929 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1930
1931 memset(&gs_type, 0, sizeof gs_type);
1932 gs_type.floating = TRUE; /* floating point values */
1933 gs_type.sign = TRUE; /* values are signed */
1934 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1935 gs_type.width = 32; /* 32-bit float */
1936 gs_type.length = variant->shader->base.vector_length;
1937
1938 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
1939
1940 /* code generated texture sampling */
1941 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
1942 context_ptr);
1943
1944 mask_val = generate_mask_value(variant, gs_type);
1945 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
1946
1947 lp_build_tgsi_soa(variant->gallivm,
1948 tokens,
1949 gs_type,
1950 &mask,
1951 consts_ptr,
1952 &system_values,
1953 NULL /*pos*/,
1954 NULL,
1955 outputs,
1956 sampler,
1957 &llvm->draw->gs.geometry_shader->info,
1958 &gs_iface);
1959
1960 sampler->destroy(sampler);
1961
1962 lp_build_mask_end(&mask);
1963
1964 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
1965
1966 gallivm_verify_function(gallivm, variant_func);
1967 }
1968
1969
1970 struct draw_gs_llvm_variant *
1971 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
1972 unsigned num_outputs,
1973 const struct draw_gs_llvm_variant_key *key)
1974 {
1975 struct draw_gs_llvm_variant *variant;
1976 struct llvm_geometry_shader *shader =
1977 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
1978 LLVMTypeRef vertex_header;
1979
1980 variant = MALLOC(sizeof *variant +
1981 shader->variant_key_size -
1982 sizeof variant->key);
1983 if (variant == NULL)
1984 return NULL;
1985
1986 variant->llvm = llvm;
1987 variant->shader = shader;
1988
1989 variant->gallivm = gallivm_create();
1990
1991 create_gs_jit_types(variant);
1992
1993 memcpy(&variant->key, key, shader->variant_key_size);
1994
1995 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
1996
1997 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
1998
1999 draw_gs_llvm_generate(llvm, variant);
2000
2001 gallivm_compile_module(variant->gallivm);
2002
2003 variant->jit_func = (draw_gs_jit_func)
2004 gallivm_jit_function(variant->gallivm, variant->function);
2005
2006 variant->list_item_global.base = variant;
2007 variant->list_item_local.base = variant;
2008 /*variant->no = */shader->variants_created++;
2009 variant->list_item_global.base = variant;
2010
2011 return variant;
2012 }
2013
2014 void
2015 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2016 {
2017 struct draw_llvm *llvm = variant->llvm;
2018
2019 if (variant->function) {
2020 gallivm_free_function(variant->gallivm,
2021 variant->function, variant->jit_func);
2022 }
2023
2024 gallivm_destroy(variant->gallivm);
2025
2026 remove_from_list(&variant->list_item_local);
2027 variant->shader->variants_cached--;
2028 remove_from_list(&variant->list_item_global);
2029 llvm->nr_gs_variants--;
2030 FREE(variant);
2031 }
2032
2033 struct draw_gs_llvm_variant_key *
2034 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2035 {
2036 unsigned i;
2037 struct draw_gs_llvm_variant_key *key;
2038 struct draw_sampler_static_state *draw_sampler;
2039
2040 key = (struct draw_gs_llvm_variant_key *)store;
2041
2042 /* All variants of this shader will have the same value for
2043 * nr_samplers. Not yet trying to compact away holes in the
2044 * sampler array.
2045 */
2046 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2047 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2048 key->nr_sampler_views =
2049 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2050 }
2051 else {
2052 key->nr_sampler_views = key->nr_samplers;
2053 }
2054
2055 draw_sampler = key->samplers;
2056
2057 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2058
2059 for (i = 0 ; i < key->nr_samplers; i++) {
2060 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2061 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2062 }
2063 for (i = 0 ; i < key->nr_sampler_views; i++) {
2064 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2065 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2066 }
2067
2068 return key;
2069 }
2070
2071 void
2072 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2073 {
2074 unsigned i;
2075 struct draw_sampler_static_state *sampler = key->samplers;
2076
2077 for (i = 0 ; i < key->nr_sampler_views; i++) {
2078 debug_printf("sampler[%i].src_format = %s\n", i,
2079 util_format_name(sampler[i].texture_state.format));
2080 }
2081 }