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_coro.h"
41 #include "gallivm/lp_bld_swizzle.h"
42 #include "gallivm/lp_bld_struct.h"
43 #include "gallivm/lp_bld_type.h"
44 #include "gallivm/lp_bld_flow.h"
45 #include "gallivm/lp_bld_debug.h"
46 #include "gallivm/lp_bld_tgsi.h"
47 #include "gallivm/lp_bld_nir.h"
48 #include "gallivm/lp_bld_printf.h"
49 #include "gallivm/lp_bld_intr.h"
50 #include "gallivm/lp_bld_init.h"
51 #include "gallivm/lp_bld_type.h"
52 #include "gallivm/lp_bld_pack.h"
53 #include "gallivm/lp_bld_format.h"
54 #include "gallivm/lp_bld_misc.h"
55 #include "gallivm/lp_bld_jit_sample.h"
56 #include "tgsi/tgsi_exec.h"
57 #include "tgsi/tgsi_dump.h"
58
59 #include "util/u_math.h"
60 #include "util/u_pointer.h"
61 #include "util/u_string.h"
62 #include "nir_serialize.h"
63 #include "util/mesa-sha1.h"
64 #define DEBUG_STORE 0
65
66
67 static void
68 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
69
70
71 struct draw_gs_llvm_iface {
72 struct lp_build_gs_iface base;
73
74 struct draw_gs_llvm_variant *variant;
75 LLVMValueRef input;
76 };
77
78
79 static inline const struct draw_gs_llvm_iface *
draw_gs_llvm_iface(const struct lp_build_gs_iface * iface)80 draw_gs_llvm_iface(const struct lp_build_gs_iface *iface)
81 {
82 return (const struct draw_gs_llvm_iface *)iface;
83 }
84
85
86 struct draw_tcs_llvm_iface {
87 struct lp_build_tcs_iface base;
88
89 struct draw_tcs_llvm_variant *variant;
90 LLVMValueRef input;
91 LLVMValueRef output;
92 };
93
94
95 static inline const struct draw_tcs_llvm_iface *
draw_tcs_llvm_iface(const struct lp_build_tcs_iface * iface)96 draw_tcs_llvm_iface(const struct lp_build_tcs_iface *iface)
97 {
98 return (const struct draw_tcs_llvm_iface *)iface;
99 }
100
101
102 struct draw_tes_llvm_iface {
103 struct lp_build_tes_iface base;
104
105 struct draw_tes_llvm_variant *variant;
106 LLVMValueRef input;
107 };
108
109
110 static inline const struct draw_tes_llvm_iface *
draw_tes_llvm_iface(const struct lp_build_tes_iface * iface)111 draw_tes_llvm_iface(const struct lp_build_tes_iface *iface)
112 {
113 return (const struct draw_tes_llvm_iface *)iface;
114 }
115
116
117 /**
118 * Create LLVM type for draw_vertex_buffer.
119 */
120 static LLVMTypeRef
create_jit_dvbuffer_type(struct gallivm_state * gallivm,const char * struct_name)121 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
122 const char *struct_name)
123 {
124 LLVMTargetDataRef target = gallivm->target;
125 LLVMTypeRef dvbuffer_type;
126 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
127 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
128
129 elem_types[DRAW_JIT_DVBUFFER_MAP] =
130 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
131 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
132
133 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
134 ARRAY_SIZE(elem_types), 0);
135
136 (void) target; /* silence unused var warning for non-debug build */
137 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
138 target, dvbuffer_type,
139 DRAW_JIT_DVBUFFER_MAP);
140 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
141 target, dvbuffer_type,
142 DRAW_JIT_DVBUFFER_SIZE);
143
144 return dvbuffer_type;
145 }
146
147 /**
148 * Create LLVM type for struct draw_jit_context
149 */
150 static LLVMTypeRef
create_vs_jit_context_type(struct gallivm_state * gallivm,const char * struct_name)151 create_vs_jit_context_type(struct gallivm_state *gallivm, const char *struct_name)
152 {
153 LLVMTargetDataRef target = gallivm->target;
154 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
155 LLVMTypeRef elem_types[DRAW_VS_JIT_CTX_NUM_FIELDS];
156
157 elem_types[DRAW_VS_JIT_CTX_PLANES] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4), DRAW_TOTAL_CLIP_PLANES), 0);
158 elem_types[DRAW_VS_JIT_CTX_VIEWPORT] = LLVMPointerType(float_type, 0);
159
160 LLVMTypeRef context_type = LLVMStructTypeInContext(gallivm->context, elem_types, ARRAY_SIZE(elem_types), 0);
161
162 (void) target; /* silence unused var warning for non-debug build */
163 LP_CHECK_MEMBER_OFFSET(struct draw_vs_jit_context, planes,
164 target, context_type, DRAW_VS_JIT_CTX_PLANES);
165 LP_CHECK_MEMBER_OFFSET(struct draw_vs_jit_context, viewports,
166 target, context_type, DRAW_VS_JIT_CTX_VIEWPORT);
167 LP_CHECK_STRUCT_SIZE(struct draw_vs_jit_context,
168 target, context_type);
169
170 return context_type;
171 }
172
173
174 /**
175 * Create LLVM type for struct draw_gs_jit_context
176 */
177 static LLVMTypeRef
create_gs_jit_context_type(struct gallivm_state * gallivm,unsigned vector_length,const char * struct_name)178 create_gs_jit_context_type(struct gallivm_state *gallivm,
179 unsigned vector_length,
180 const char *struct_name)
181 {
182 LLVMTargetDataRef target = gallivm->target;
183 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
184 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
185 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
186 LLVMTypeRef context_type;
187
188 elem_types[DRAW_GS_JIT_CTX_PLANES] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
189 DRAW_TOTAL_CLIP_PLANES), 0);
190 elem_types[DRAW_GS_JIT_CTX_VIEWPORT] = LLVMPointerType(float_type, 0); /* viewports */
191
192
193 elem_types[DRAW_GS_JIT_CTX_PRIM_LENGTHS] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
194 elem_types[DRAW_GS_JIT_CTX_EMITTED_VERTICES] = LLVMPointerType(LLVMVectorType(int_type,
195 vector_length), 0);
196 elem_types[DRAW_GS_JIT_CTX_EMITTED_PRIMS] = LLVMPointerType(LLVMVectorType(int_type,
197 vector_length), 0);
198
199 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
200 ARRAY_SIZE(elem_types), 0);
201
202 (void) target; /* silence unused var warning for non-debug build */
203 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
204 target, context_type, DRAW_GS_JIT_CTX_PLANES);
205 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
206 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
207 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
208 target, context_type,
209 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
210 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
211 target, context_type,
212 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
213 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
214 target, context_type,
215 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
216 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
217 target, context_type);
218 return context_type;
219 }
220
221
222 static LLVMTypeRef
create_gs_jit_input_type_deref(struct gallivm_state * gallivm)223 create_gs_jit_input_type_deref(struct gallivm_state *gallivm)
224 {
225 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
226 LLVMTypeRef input_array;
227
228 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
229 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
230 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
231 return input_array;
232 }
233
234
235 static LLVMTypeRef
create_gs_jit_input_type(struct gallivm_state * gallivm)236 create_gs_jit_input_type(struct gallivm_state *gallivm)
237 {
238 return LLVMPointerType(create_gs_jit_input_type_deref(gallivm), 0); /* num vertices per prim */
239 }
240
241
242 /**
243 * Create LLVM type for struct pipe_vertex_buffer
244 */
245 static LLVMTypeRef
create_jit_vertex_buffer_type(struct gallivm_state * gallivm,const char * struct_name)246 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
247 const char *struct_name)
248 {
249 LLVMTargetDataRef target = gallivm->target;
250 LLVMTypeRef elem_types[3];
251 LLVMTypeRef vb_type;
252
253 elem_types[0] = LLVMInt8TypeInContext(gallivm->context);
254 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
255 elem_types[2] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
256
257 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
258 ARRAY_SIZE(elem_types), 0);
259
260 (void) target; /* silence unused var warning for non-debug build */
261 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
262 target, vb_type, 0);
263 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
264 target, vb_type, 1);
265 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
266 target, vb_type, 2);
267
268 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
269
270 return vb_type;
271 }
272
273
274 static LLVMTypeRef
create_tcs_jit_input_type_deref(struct gallivm_state * gallivm)275 create_tcs_jit_input_type_deref(struct gallivm_state *gallivm)
276 {
277 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
278 LLVMTypeRef input_array;
279
280 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
281 input_array = LLVMArrayType(input_array, NUM_TCS_INPUTS); /* num attrs per vertex */
282 return input_array;
283 }
284
285
286 static LLVMTypeRef
create_tcs_jit_input_type(struct gallivm_state * gallivm)287 create_tcs_jit_input_type(struct gallivm_state *gallivm)
288 {
289 return LLVMPointerType(create_tcs_jit_input_type_deref(gallivm), 0); /* num vertices per prim */
290 }
291
292
293 static LLVMTypeRef
create_tcs_jit_output_type_deref(struct gallivm_state * gallivm)294 create_tcs_jit_output_type_deref(struct gallivm_state *gallivm)
295 {
296 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
297 LLVMTypeRef output_array;
298
299 output_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
300 output_array = LLVMArrayType(output_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
301 return output_array;
302 }
303
304
305 static LLVMTypeRef
create_tcs_jit_output_type(struct gallivm_state * gallivm)306 create_tcs_jit_output_type(struct gallivm_state *gallivm)
307 {
308 return LLVMPointerType(create_tcs_jit_output_type_deref(gallivm), 0); /* num vertices per prim */
309 }
310
311
312 static LLVMTypeRef
create_tes_jit_input_deref_type(struct gallivm_state * gallivm)313 create_tes_jit_input_deref_type(struct gallivm_state *gallivm)
314 {
315 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
316 LLVMTypeRef input_array;
317
318 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
319 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
320
321 return input_array;
322 }
323
324
325 /**
326 * Create LLVM types for various structures.
327 */
328 static void
create_vs_jit_types(struct draw_llvm_variant * variant)329 create_vs_jit_types(struct draw_llvm_variant *variant)
330 {
331 struct gallivm_state *gallivm = variant->gallivm;
332
333 variant->context_type = create_vs_jit_context_type(gallivm, "draw_vs_jit_context");
334 variant->context_ptr_type = LLVMPointerType(variant->context_type, 0);
335
336 variant->resources_type = lp_build_jit_resources_type(gallivm);
337 variant->resources_ptr_type = LLVMPointerType(variant->resources_type, 0);
338
339 variant->buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
340 variant->buffer_ptr_type = LLVMPointerType(variant->buffer_type, 0);
341
342 variant->vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
343 variant->vb_ptr_type = LLVMPointerType(variant->vb_type, 0);
344 }
345
346
347 static LLVMTypeRef
get_context_ptr_type(struct draw_llvm_variant * variant)348 get_context_ptr_type(struct draw_llvm_variant *variant)
349 {
350 if (!variant->context_ptr_type)
351 create_vs_jit_types(variant);
352 return variant->context_ptr_type;
353 }
354
355
356 static LLVMTypeRef
get_buffer_ptr_type(struct draw_llvm_variant * variant)357 get_buffer_ptr_type(struct draw_llvm_variant *variant)
358 {
359 if (!variant->buffer_ptr_type)
360 create_vs_jit_types(variant);
361 return variant->buffer_ptr_type;
362 }
363
364
365 static LLVMTypeRef
get_vb_ptr_type(struct draw_llvm_variant * variant)366 get_vb_ptr_type(struct draw_llvm_variant *variant)
367 {
368 if (!variant->vb_ptr_type)
369 create_vs_jit_types(variant);
370 return variant->vb_ptr_type;
371 }
372
373 static LLVMTypeRef
get_vertex_header_ptr_type(struct draw_llvm_variant * variant)374 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
375 {
376 assert(variant->vertex_header_ptr_type);
377 return variant->vertex_header_ptr_type;
378 }
379
380
381 /**
382 * Create per-context LLVM info.
383 */
384 struct draw_llvm *
draw_llvm_create(struct draw_context * draw,LLVMContextRef context)385 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
386 {
387 struct draw_llvm *llvm;
388
389 if (!lp_build_init())
390 return NULL;
391
392 llvm = CALLOC_STRUCT(draw_llvm);
393 if (!llvm)
394 return NULL;
395
396 llvm->draw = draw;
397
398 llvm->context = context;
399 if (!llvm->context) {
400 llvm->context = LLVMContextCreate();
401
402 #if LLVM_VERSION_MAJOR == 15
403 LLVMContextSetOpaquePointers(llvm->context, false);
404 #endif
405
406 llvm->context_owned = true;
407 }
408 if (!llvm->context)
409 goto fail;
410
411 llvm->nr_variants = 0;
412 list_inithead(&llvm->vs_variants_list.list);
413
414 llvm->nr_gs_variants = 0;
415 list_inithead(&llvm->gs_variants_list.list);
416
417 llvm->nr_tcs_variants = 0;
418 list_inithead(&llvm->tcs_variants_list.list);
419
420 llvm->nr_tes_variants = 0;
421 list_inithead(&llvm->tes_variants_list.list);
422
423 return llvm;
424
425 fail:
426 draw_llvm_destroy(llvm);
427 return NULL;
428 }
429
430
431 /**
432 * Free per-context LLVM info.
433 */
434 void
draw_llvm_destroy(struct draw_llvm * llvm)435 draw_llvm_destroy(struct draw_llvm *llvm)
436 {
437 if (llvm->context_owned)
438 LLVMContextDispose(llvm->context);
439 llvm->context = NULL;
440
441 /* XXX free other draw_llvm data? */
442 FREE(llvm);
443 }
444
445
446 static void
draw_get_ir_cache_key(struct nir_shader * nir,const void * key,size_t key_size,uint32_t val_32bit,unsigned char ir_sha1_cache_key[20])447 draw_get_ir_cache_key(struct nir_shader *nir,
448 const void *key, size_t key_size,
449 uint32_t val_32bit,
450 unsigned char ir_sha1_cache_key[20])
451 {
452 struct blob blob = { 0 };
453 unsigned ir_size;
454 void *ir_binary;
455
456 blob_init(&blob);
457 nir_serialize(&blob, nir, true);
458 ir_binary = blob.data;
459 ir_size = blob.size;
460
461 struct mesa_sha1 ctx;
462 _mesa_sha1_init(&ctx);
463 _mesa_sha1_update(&ctx, key, key_size);
464 _mesa_sha1_update(&ctx, ir_binary, ir_size);
465 _mesa_sha1_update(&ctx, &val_32bit, 4);
466 _mesa_sha1_final(&ctx, ir_sha1_cache_key);
467
468 blob_finish(&blob);
469 }
470
471
472 /**
473 * Create LLVM-generated code for a vertex shader.
474 */
475 struct draw_llvm_variant *
draw_llvm_create_variant(struct draw_llvm * llvm,unsigned num_inputs,const struct draw_llvm_variant_key * key)476 draw_llvm_create_variant(struct draw_llvm *llvm,
477 unsigned num_inputs,
478 const struct draw_llvm_variant_key *key)
479 {
480 struct draw_llvm_variant *variant;
481 struct llvm_vertex_shader *shader =
482 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
483 char module_name[64];
484 unsigned char ir_sha1_cache_key[20];
485 struct lp_cached_code cached = { 0 };
486 bool needs_caching = false;
487 variant = MALLOC(sizeof *variant +
488 shader->variant_key_size -
489 sizeof variant->key);
490 if (!variant)
491 return NULL;
492
493 variant->llvm = llvm;
494 variant->shader = shader;
495 memcpy(&variant->key, key, shader->variant_key_size);
496
497 snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
498 variant->shader->variants_cached);
499
500 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
501 draw_get_ir_cache_key(shader->base.state.ir.nir,
502 key,
503 shader->variant_key_size,
504 num_inputs,
505 ir_sha1_cache_key);
506
507 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
508 &cached,
509 ir_sha1_cache_key);
510 if (!cached.data_size)
511 needs_caching = true;
512 }
513 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
514
515 create_vs_jit_types(variant);
516
517 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
518 if (llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_TGSI)
519 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
520 else
521 nir_print_shader(llvm->draw->vs.vertex_shader->state.ir.nir, stderr);
522 draw_llvm_dump_variant_key(&variant->key);
523 }
524
525 variant->vertex_header_type = lp_build_create_jit_vertex_header_type(variant->gallivm, num_inputs);
526 variant->vertex_header_ptr_type = LLVMPointerType(variant->vertex_header_type, 0);
527
528 draw_llvm_generate(llvm, variant);
529
530 gallivm_compile_module(variant->gallivm);
531
532 variant->jit_func = (draw_jit_vert_func)
533 gallivm_jit_function(variant->gallivm, variant->function);
534
535 if (needs_caching)
536 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
537 &cached,
538 ir_sha1_cache_key);
539 gallivm_free_ir(variant->gallivm);
540
541 variant->list_item_global.base = variant;
542 variant->list_item_local.base = variant;
543 /*variant->no = */shader->variants_created++;
544 variant->list_item_global.base = variant;
545
546 return variant;
547 }
548
549
550 static void
do_clamp_vertex_color(struct gallivm_state * gallivm,struct lp_type type,const struct tgsi_shader_info * info,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS])551 do_clamp_vertex_color(struct gallivm_state *gallivm,
552 struct lp_type type,
553 const struct tgsi_shader_info *info,
554 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS])
555 {
556 LLVMBuilderRef builder = gallivm->builder;
557 LLVMValueRef out;
558 unsigned chan, attrib;
559 struct lp_build_context bld;
560 lp_build_context_init(&bld, gallivm, type);
561
562 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
563 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
564 if (outputs[attrib][chan]) {
565 switch (info->output_semantic_name[attrib]) {
566 case TGSI_SEMANTIC_COLOR:
567 case TGSI_SEMANTIC_BCOLOR:
568 out = LLVMBuildLoad2(builder, LLVMTypeOf(bld.zero), outputs[attrib][chan], "");
569 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
570 LLVMBuildStore(builder, out, outputs[attrib][chan]);
571 break;
572 }
573 }
574 }
575 }
576 }
577
578
579 static void
generate_vs(struct draw_llvm_variant * variant,LLVMBuilderRef builder,struct lp_type vs_type,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],const LLVMValueRef (* inputs)[TGSI_NUM_CHANNELS],const struct lp_bld_tgsi_system_values * system_values,LLVMValueRef context_ptr,LLVMValueRef resources_ptr,const struct lp_build_sampler_soa * draw_sampler,const struct lp_build_image_soa * draw_image,bool clamp_vertex_color,struct lp_build_mask_context * bld_mask)580 generate_vs(struct draw_llvm_variant *variant,
581 LLVMBuilderRef builder,
582 struct lp_type vs_type,
583 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
584 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
585 const struct lp_bld_tgsi_system_values *system_values,
586 LLVMValueRef context_ptr,
587 LLVMValueRef resources_ptr,
588 const struct lp_build_sampler_soa *draw_sampler,
589 const struct lp_build_image_soa *draw_image,
590 bool clamp_vertex_color,
591 struct lp_build_mask_context *bld_mask)
592 {
593 struct draw_llvm *llvm = variant->llvm;
594 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
595 LLVMValueRef consts_ptr =
596 lp_jit_resources_constants(variant->gallivm, variant->resources_type, resources_ptr);
597 LLVMValueRef ssbos_ptr =
598 lp_jit_resources_ssbos(variant->gallivm, variant->resources_type, resources_ptr);
599 struct draw_llvm_variant_key *key = &variant->key;
600
601 struct lp_build_tgsi_params params;
602 memset(¶ms, 0, sizeof(params));
603
604 params.type = vs_type;
605 params.mask = bld_mask;
606 params.consts_ptr = consts_ptr;
607 params.system_values = system_values;
608 params.inputs = inputs;
609 params.num_inputs = key->nr_vertex_elements;
610 params.context_type = variant->context_type;
611 params.context_ptr = context_ptr;
612 params.resources_type = variant->resources_type;
613 params.resources_ptr = resources_ptr;
614 params.sampler = draw_sampler;
615 params.info = &llvm->draw->vs.vertex_shader->info;
616 params.ssbo_ptr = ssbos_ptr;
617 params.image = draw_image;
618 params.aniso_filter_table = lp_jit_resources_aniso_filter_table(variant->gallivm,
619 variant->resources_type,
620 resources_ptr);
621
622 if (llvm->draw->vs.vertex_shader->state.ir.nir &&
623 llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_NIR) {
624 lp_build_nir_soa(variant->gallivm,
625 llvm->draw->vs.vertex_shader->state.ir.nir,
626 ¶ms,
627 outputs);
628 } else {
629 lp_build_tgsi_soa(variant->gallivm,
630 tokens,
631 ¶ms,
632 outputs);
633 }
634
635 if (clamp_vertex_color) {
636 const struct tgsi_shader_info *info = &llvm->draw->vs.vertex_shader->info;
637 do_clamp_vertex_color(variant->gallivm,
638 vs_type, info,
639 outputs);
640 }
641 }
642
643
644 static void
fetch_instanced(struct gallivm_state * gallivm,const struct util_format_description * format_desc,struct lp_type vs_type,LLVMValueRef vb_stride,LLVMValueRef map_ptr,LLVMValueRef buffer_size_adj,LLVMValueRef * inputs,LLVMValueRef index)645 fetch_instanced(struct gallivm_state *gallivm,
646 const struct util_format_description *format_desc,
647 struct lp_type vs_type,
648 LLVMValueRef vb_stride,
649 LLVMValueRef map_ptr,
650 LLVMValueRef buffer_size_adj,
651 LLVMValueRef *inputs,
652 LLVMValueRef index)
653 {
654 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
655 LLVMTypeRef aosf_t, aosi_t;
656 LLVMValueRef zero = LLVMConstNull(i32_t);
657 LLVMBuilderRef builder = gallivm->builder;
658 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
659 unsigned i;
660
661 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
662 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
663
664 /* This mul can overflow. Wraparound is ok. */
665 stride = LLVMBuildMul(builder, vb_stride, index, "");
666
667 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
668 stride, buffer_size_adj,
669 "buffer_overflowed");
670
671 if (0) {
672 lp_build_print_value(gallivm, " instance index = ", index);
673 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
674 }
675
676 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
677 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
678 stride = LLVMBuildAnd(builder, stride, index_valid, "");
679
680 aos = lp_build_fetch_rgba_aos(gallivm,
681 format_desc,
682 lp_float32_vec4_type(),
683 false,
684 map_ptr,
685 stride, zero, zero,
686 NULL);
687
688 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
689 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
690 aos = LLVMBuildAnd(builder, aos, index_valid, "");
691 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
692
693 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
694 LLVMValueRef index = lp_build_const_int32(gallivm, i);
695 inputs[i] = lp_build_extract_broadcast(gallivm,
696 lp_float32_vec4_type(),
697 vs_type, aos, index);
698 }
699 }
700
701
702 static void
fetch_vector(struct gallivm_state * gallivm,const struct util_format_description * format_desc,struct lp_type vs_type,LLVMValueRef vb_stride,LLVMValueRef map_ptr,LLVMValueRef buffer_size_adj,LLVMValueRef * inputs,LLVMValueRef indices)703 fetch_vector(struct gallivm_state *gallivm,
704 const struct util_format_description *format_desc,
705 struct lp_type vs_type,
706 LLVMValueRef vb_stride,
707 LLVMValueRef map_ptr,
708 LLVMValueRef buffer_size_adj,
709 LLVMValueRef *inputs,
710 LLVMValueRef indices)
711 {
712 LLVMBuilderRef builder = gallivm->builder;
713 struct lp_build_context blduivec;
714 struct lp_type fetch_type = vs_type;
715 LLVMValueRef offset, valid_mask;
716 unsigned i;
717
718 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
719
720 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
721 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
722
723 /* This mul can overflow. Wraparound is ok. */
724 offset = lp_build_mul(&blduivec, vb_stride, indices);
725
726 valid_mask = lp_build_compare(gallivm, blduivec.type,
727 PIPE_FUNC_LESS, offset, buffer_size_adj);
728
729 /* not valid elements use offset 0 */
730 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
731
732 if (0) {
733 lp_build_print_value(gallivm, " indices = ", indices);
734 lp_build_print_value(gallivm, " offsets = ", offset);
735 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
736 }
737
738 /*
739 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
740 * This should always produce better code.
741 */
742
743 /* The type handling is annoying here... */
744 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
745 format_desc->channel[0].pure_integer) {
746 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
747 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
748 } else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
749 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
750 }
751 }
752
753 lp_build_fetch_rgba_soa(gallivm, format_desc,
754 fetch_type, false, map_ptr, offset,
755 blduivec.zero, blduivec.zero,
756 NULL, inputs);
757
758 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
759 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
760 lp_build_vec_type(gallivm, vs_type), "");
761 }
762
763 /* out-of-bound fetches return all zeros */
764 for (i = 0; i < format_desc->nr_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 static void
store_aos(struct gallivm_state * gallivm,bool is_per_prim,LLVMTypeRef io_type,LLVMValueRef io_ptr,LLVMValueRef index,LLVMValueRef value)774 store_aos(struct gallivm_state *gallivm,
775 bool is_per_prim,
776 LLVMTypeRef io_type,
777 LLVMValueRef io_ptr,
778 LLVMValueRef index,
779 LLVMValueRef value)
780 {
781 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
782 LLVMBuilderRef builder = gallivm->builder;
783 LLVMValueRef data_ptr;
784 LLVMTypeRef data_type;
785 LLVMValueRef indices[3];
786
787 indices[0] = lp_build_const_int32(gallivm, 0);
788 indices[1] = index;
789 indices[2] = lp_build_const_int32(gallivm, 0);
790
791 if (!is_per_prim) {
792 data_ptr = lp_jit_vertex_header_data(gallivm, io_type, io_ptr);
793 data_type = LLVMStructGetTypeAtIndex(io_type, LP_JIT_VERTEX_HEADER_DATA);
794 } else {
795 data_ptr = io_ptr;
796 data_type = io_type;
797 }
798
799 data_ptr = LLVMBuildGEP2(builder, data_type, data_ptr, indices, 3, "");
800 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
801
802 #if DEBUG_STORE
803 if (is_per_prim)
804 lp_build_printf(gallivm, " ---- %p storing prim attribute %d (io = %p)\n", data_ptr, index, io_ptr);
805 else
806 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
807 #endif
808
809 /* Unaligned store due to the vertex header */
810 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
811 }
812
813
814 /**
815 * Adjust the mask to architecture endianess. The mask will the store in struct:
816 *
817 * struct vertex_header {
818 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
819 * unsigned edgeflag:1;
820 * unsigned pad:1;
821 * unsigned vertex_id:16;
822 * [...]
823 * }
824 *
825 * On little-endian machine nothing needs to done, however on bit-endian machine
826 * the mask's fields need to be adjusted with the algorithm:
827 *
828 * uint32_t reverse (uint32_t x)
829 * {
830 * return (x >> 16) | // vertex_id
831 * ((x & 0x3fff) << 18) | // clipmask
832 * ((x & 0x4000) << 3) | // edgeflag
833 * ((x & 0x8000) << 1); // pad
834 * }
835 */
836 static LLVMValueRef
adjust_mask(struct gallivm_state * gallivm,LLVMValueRef mask)837 adjust_mask(struct gallivm_state *gallivm,
838 LLVMValueRef mask)
839 {
840 #if UTIL_ARCH_BIG_ENDIAN
841 LLVMBuilderRef builder = gallivm->builder;
842 LLVMValueRef vertex_id;
843 LLVMValueRef clipmask;
844 LLVMValueRef pad;
845 LLVMValueRef edgeflag;
846
847 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
848 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
849 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
850 if (0) {
851 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
852 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 1), "");
853 }
854 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
855 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 3), "");
856
857 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
858 if (0) {
859 mask = LLVMBuildOr(builder, mask, pad, "");
860 }
861 mask = LLVMBuildOr(builder, mask, edgeflag, "");
862 #endif
863 return mask;
864 }
865
866
867 void
draw_store_aos_array(struct gallivm_state * gallivm,struct lp_type soa_type,LLVMTypeRef io_type,LLVMValueRef io_ptr,LLVMValueRef * indices,LLVMValueRef * aos,int attrib,LLVMValueRef clipmask,bool need_edgeflag,bool is_per_prim)868 draw_store_aos_array(struct gallivm_state *gallivm,
869 struct lp_type soa_type,
870 LLVMTypeRef io_type,
871 LLVMValueRef io_ptr,
872 LLVMValueRef *indices,
873 LLVMValueRef* aos,
874 int attrib,
875 LLVMValueRef clipmask,
876 bool need_edgeflag, bool is_per_prim)
877 {
878 LLVMBuilderRef builder = gallivm->builder;
879 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
880 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
881 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
882 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
883 int vector_length = soa_type.length;
884
885 assert(TGSI_NUM_CHANNELS == 4);
886
887 for (int i = 0; i < vector_length; i++) {
888 linear_inds[i] = lp_build_const_int32(gallivm, i);
889 if (indices) {
890 inds[i] = indices[i];
891 } else {
892 inds[i] = linear_inds[i];
893 }
894 io_ptrs[i] = LLVMBuildGEP2(builder, io_type, io_ptr, &inds[i], 1, "");
895 }
896
897 if (attrib == 0 && !is_per_prim) {
898 /* store vertex header for each of the n vertices */
899 LLVMValueRef val, cliptmp;
900 int vertex_id_pad_edgeflag;
901
902 /* If this assertion fails, it means we need to update the bit twidding
903 * code here. See struct vertex_header in draw_private.h.
904 */
905 assert(DRAW_TOTAL_CLIP_PLANES==14);
906 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
907 if (!need_edgeflag) {
908 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
909 } else {
910 vertex_id_pad_edgeflag = (0xffff << 16);
911 }
912 if (vector_length == 1)
913 val = lp_build_const_int32(gallivm, vertex_id_pad_edgeflag);
914 else
915 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
916 vertex_id_pad_edgeflag);
917
918 /* OR with the clipmask */
919 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
920 for (unsigned i = 0; i < vector_length; i++) {
921 LLVMValueRef id_ptr = lp_jit_vertex_header_id(gallivm, io_type, io_ptrs[i]);
922 if (vector_length > 1)
923 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
924 else
925 val = cliptmp;
926 val = adjust_mask(gallivm, val);
927 #if DEBUG_STORE
928 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
929 io_ptrs[i], inds[i], val);
930 #endif
931 LLVMBuildStore(builder, val, id_ptr);
932 }
933 }
934
935 /* store for each of the n vertices */
936 for (int i = 0; i < vector_length; i++) {
937 store_aos(gallivm, is_per_prim, io_type, io_ptrs[i], attr_index, aos[i]);
938 }
939 }
940
941
942 static void
convert_to_aos(struct gallivm_state * gallivm,LLVMTypeRef io_type,LLVMValueRef io,LLVMValueRef * indices,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],LLVMValueRef clipmask,int num_outputs,struct lp_type soa_type,int primid_slot,bool need_edgeflag)943 convert_to_aos(struct gallivm_state *gallivm,
944 LLVMTypeRef io_type,
945 LLVMValueRef io,
946 LLVMValueRef *indices,
947 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
948 LLVMValueRef clipmask,
949 int num_outputs,
950 struct lp_type soa_type,
951 int primid_slot,
952 bool need_edgeflag)
953 {
954 LLVMBuilderRef builder = gallivm->builder;
955
956 #if DEBUG_STORE
957 lp_build_printf(gallivm, " # storing begin\n");
958 #endif
959 for (unsigned attrib = 0; attrib < num_outputs; ++attrib) {
960 LLVMValueRef soa[TGSI_NUM_CHANNELS];
961 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
962 for (unsigned chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
963 if (outputs[attrib][chan]) {
964 LLVMTypeRef single_type = (attrib == primid_slot) ? lp_build_int_vec_type(gallivm, soa_type) : lp_build_vec_type(gallivm, soa_type);
965 LLVMValueRef out = LLVMBuildLoad2(builder, single_type, outputs[attrib][chan], "");
966 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
967 #if DEBUG_STORE
968 lp_build_printf(gallivm, "output %d : %d ",
969 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
970 attrib, 0),
971 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
972 chan, 0));
973 lp_build_print_value(gallivm, "val = ", out);
974 {
975 LLVMValueRef iv =
976 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
977
978 lp_build_print_value(gallivm, " ival = ", iv);
979 }
980 #endif
981 soa[chan] = out;
982 } else {
983 soa[chan] = 0;
984 }
985 }
986
987
988 if (soa_type.length == TGSI_NUM_CHANNELS) {
989 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
990 } else {
991 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
992
993 for (unsigned i = 0; i < soa_type.length; ++i) {
994 aos[i] = lp_build_extract_range(gallivm,
995 soa[i % TGSI_NUM_CHANNELS],
996 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
997 TGSI_NUM_CHANNELS);
998 }
999 }
1000
1001 draw_store_aos_array(gallivm,
1002 soa_type,
1003 io_type,
1004 io,
1005 indices,
1006 aos,
1007 attrib,
1008 clipmask,
1009 need_edgeflag, false);
1010 }
1011 #if DEBUG_STORE
1012 lp_build_printf(gallivm, " # storing end\n");
1013 #endif
1014 }
1015
1016
1017 /**
1018 * Stores original vertex positions in clip coordinates
1019 */
1020 static void
store_clip(struct gallivm_state * gallivm,const struct lp_type vs_type,LLVMTypeRef io_type,LLVMValueRef io_ptr,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],int idx)1021 store_clip(struct gallivm_state *gallivm,
1022 const struct lp_type vs_type,
1023 LLVMTypeRef io_type,
1024 LLVMValueRef io_ptr,
1025 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1026 int idx)
1027 {
1028 LLVMBuilderRef builder = gallivm->builder;
1029 LLVMValueRef soa[4];
1030 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1031 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1032 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1033 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1034 LLVMTypeRef clip_ptr_type =
1035 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1036 4), 0);
1037
1038 for (int i = 0; i < vs_type.length; i++) {
1039 inds[i] = lp_build_const_int32(gallivm, i);
1040 io_ptrs[i] = LLVMBuildGEP2(builder, io_type, io_ptr, &inds[i], 1, "");
1041 }
1042
1043 LLVMTypeRef single_type = lp_build_vec_type(gallivm, vs_type);
1044 soa[0] = LLVMBuildLoad2(builder, single_type, outputs[idx][0], ""); /*x0 x1 .. xn*/
1045 soa[1] = LLVMBuildLoad2(builder, single_type, outputs[idx][1], ""); /*y0 y1 .. yn*/
1046 soa[2] = LLVMBuildLoad2(builder, single_type, outputs[idx][2], ""); /*z0 z1 .. zn*/
1047 soa[3] = LLVMBuildLoad2(builder, single_type, outputs[idx][3], ""); /*w0 w1 .. wn*/
1048
1049 for (int i = 0; i < vs_type.length; i++) {
1050 clip_ptrs[i] = lp_jit_vertex_header_clip_pos(gallivm, io_type, io_ptrs[i]);
1051 }
1052
1053 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1054 for (int i = 0; i < vs_type.length; ++i) {
1055 aos[i] = lp_build_extract_range(gallivm,
1056 soa[i % TGSI_NUM_CHANNELS],
1057 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1058 TGSI_NUM_CHANNELS);
1059 }
1060
1061 for (int j = 0; j < vs_type.length; j++) {
1062 LLVMValueRef clip_ptr;
1063
1064 clip_ptr = LLVMBuildPointerCast(builder, clip_ptrs[j], clip_ptr_type, "");
1065
1066 /* Unaligned store */
1067 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1068 }
1069 }
1070
1071
1072 /**
1073 * Transforms the outputs for viewport mapping
1074 */
1075 static void
generate_viewport(struct draw_llvm_variant * variant,LLVMBuilderRef builder,struct lp_type vs_type,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],LLVMValueRef context_ptr)1076 generate_viewport(struct draw_llvm_variant *variant,
1077 LLVMBuilderRef builder,
1078 struct lp_type vs_type,
1079 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1080 LLVMValueRef context_ptr)
1081 {
1082 struct gallivm_state *gallivm = variant->gallivm;
1083 struct lp_type f32_type = vs_type;
1084 const unsigned pos = variant->llvm->draw->vs.position_output;
1085 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1086 LLVMValueRef out3 = LLVMBuildLoad2(builder, vs_type_llvm, outputs[pos][3], ""); /*w0 w1 .. wn*/
1087 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1088 LLVMValueRef vp_ptr = draw_vs_jit_context_viewports(variant, context_ptr);
1089
1090 /* We treat pipe_viewport_state as a float array */
1091 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1092 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1093
1094 /* for 1/w convention*/
1095 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1096 LLVMBuildStore(builder, out3, outputs[pos][3]);
1097
1098 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, vs_type);
1099
1100 /* Viewport Mapping */
1101 for (unsigned i = 0; i < 3; i++) {
1102 LLVMValueRef out = LLVMBuildLoad2(builder, vs_type_llvm, outputs[pos][i], ""); /*x0 x1 .. xn*/
1103 LLVMValueRef scale;
1104 LLVMValueRef trans;
1105 LLVMValueRef scale_i;
1106 LLVMValueRef trans_i;
1107 LLVMValueRef index;
1108
1109 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1110 scale_i = LLVMBuildGEP2(builder, elem_type, vp_ptr, &index, 1, "");
1111
1112 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1113 trans_i = LLVMBuildGEP2(builder, elem_type, vp_ptr, &index, 1, "");
1114
1115 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1116 LLVMBuildLoad2(builder, elem_type, scale_i, "scale"));
1117 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1118 LLVMBuildLoad2(builder, elem_type, trans_i, "trans"));
1119
1120 /* divide by w */
1121 out = LLVMBuildFMul(builder, out, out3, "");
1122 /* mult by scale, add translation */
1123 out = lp_build_fmuladd(builder, out, scale, trans);
1124
1125 /* store transformed outputs */
1126 LLVMBuildStore(builder, out, outputs[pos][i]);
1127 }
1128
1129 }
1130
1131
1132 /**
1133 * Returns clipmask as nxi32 bitmask for the n vertices
1134 */
1135 static LLVMValueRef
generate_clipmask(struct draw_llvm * llvm,struct gallivm_state * gallivm,struct lp_type vs_type,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],struct draw_llvm_variant_key * key,LLVMTypeRef context_type,LLVMValueRef context_ptr,bool * have_clipdist)1136 generate_clipmask(struct draw_llvm *llvm,
1137 struct gallivm_state *gallivm,
1138 struct lp_type vs_type,
1139 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1140 struct draw_llvm_variant_key *key,
1141 LLVMTypeRef context_type,
1142 LLVMValueRef context_ptr,
1143 bool *have_clipdist)
1144 {
1145 LLVMBuilderRef builder = gallivm->builder;
1146 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1147 LLVMValueRef test, temp;
1148 LLVMValueRef zero, shift;
1149 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1150 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1151 LLVMValueRef plane1, planes, plane_ptr;
1152 struct lp_type f32_type = vs_type;
1153 struct lp_type i32_type = lp_int_type(vs_type);
1154 const unsigned pos = llvm->draw->vs.position_output;
1155 const unsigned cv = llvm->draw->vs.clipvertex_output;
1156 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1157 bool have_cd = false;
1158 bool clip_user = key->clip_user;
1159 unsigned ucp_enable = key->ucp_enable;
1160 unsigned cd[2];
1161
1162 cd[0] = llvm->draw->vs.ccdistance_output[0];
1163 cd[1] = llvm->draw->vs.ccdistance_output[1];
1164
1165 if (cd[0] != pos || cd[1] != pos)
1166 have_cd = true;
1167
1168 if (num_written_clipdistance && !clip_user) {
1169 clip_user = true;
1170 ucp_enable = (1 << num_written_clipdistance) - 1;
1171 }
1172
1173 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1174 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1175 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1176 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1177
1178 LLVMTypeRef vec_type = LLVMTypeOf(zero);
1179
1180 /*
1181 * load clipvertex and position from correct locations.
1182 * if they are the same just load them once.
1183 */
1184 pos_x = LLVMBuildLoad2(builder, vec_type, outputs[pos][0], ""); /*x0 x1 .. xn */
1185 pos_y = LLVMBuildLoad2(builder, vec_type, outputs[pos][1], ""); /*y0 y1 .. yn */
1186 pos_z = LLVMBuildLoad2(builder, vec_type, outputs[pos][2], ""); /*z0 z1 .. zn */
1187 pos_w = LLVMBuildLoad2(builder, vec_type, outputs[pos][3], ""); /*w0 w1 .. wn */
1188
1189 if (clip_user && cv != pos) {
1190 cv_x = LLVMBuildLoad2(builder, vec_type, outputs[cv][0], ""); /*x0 x1 .. xn */
1191 cv_y = LLVMBuildLoad2(builder, vec_type, outputs[cv][1], ""); /*y0 y1 .. yn */
1192 cv_z = LLVMBuildLoad2(builder, vec_type, outputs[cv][2], ""); /*z0 z1 .. zn */
1193 cv_w = LLVMBuildLoad2(builder, vec_type, outputs[cv][3], ""); /*w0 w1 .. wn */
1194 } else {
1195 cv_x = pos_x;
1196 cv_y = pos_y;
1197 cv_z = pos_z;
1198 cv_w = pos_w;
1199 }
1200
1201 /*
1202 * Be careful with the comparisons and NaNs (using llvm's unordered
1203 * comparisons here).
1204 */
1205 /* Cliptest, for hardwired planes */
1206 /*
1207 * XXX should take guardband into account (currently not in key).
1208 * Otherwise might run the draw pipeline stages for nothing.
1209 */
1210 if (key->clip_xy) {
1211 /* plane 1 */
1212 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1213 temp = shift;
1214 test = LLVMBuildAnd(builder, test, temp, "");
1215 mask = test;
1216
1217 /* plane 2 */
1218 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1219 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1220 temp = LLVMBuildShl(builder, temp, shift, "");
1221 test = LLVMBuildAnd(builder, test, temp, "");
1222 mask = LLVMBuildOr(builder, mask, test, "");
1223
1224 /* plane 3 */
1225 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1226 temp = LLVMBuildShl(builder, temp, shift, "");
1227 test = LLVMBuildAnd(builder, test, temp, "");
1228 mask = LLVMBuildOr(builder, mask, test, "");
1229
1230 /* plane 4 */
1231 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1232 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1233 temp = LLVMBuildShl(builder, temp, shift, "");
1234 test = LLVMBuildAnd(builder, test, temp, "");
1235 mask = LLVMBuildOr(builder, mask, test, "");
1236 }
1237
1238 if (key->clip_z) {
1239 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1240 if (key->clip_halfz) {
1241 /* plane 5 */
1242 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1243 test = LLVMBuildAnd(builder, test, temp, "");
1244 mask = LLVMBuildOr(builder, mask, test, "");
1245 } else {
1246 /* plane 5 */
1247 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1248 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1249 test = LLVMBuildAnd(builder, test, temp, "");
1250 mask = LLVMBuildOr(builder, mask, test, "");
1251 }
1252 /* plane 6 */
1253 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1254 temp = LLVMBuildShl(builder, temp, shift, "");
1255 test = LLVMBuildAnd(builder, test, temp, "");
1256 mask = LLVMBuildOr(builder, mask, test, "");
1257 }
1258
1259 if (clip_user) {
1260 LLVMValueRef planes_ptr = draw_vs_jit_context_planes(gallivm, context_type, context_ptr);
1261 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
1262 LLVMTypeRef planes_type = LLVMArrayType(LLVMArrayType(float_type, 4), DRAW_TOTAL_CLIP_PLANES);
1263 LLVMValueRef indices[3];
1264 LLVMValueRef is_nan_or_inf;
1265
1266 /* userclip planes */
1267 while (ucp_enable) {
1268 unsigned plane_idx = ffs(ucp_enable)-1;
1269 ucp_enable &= ~(1 << plane_idx);
1270 plane_idx += 6;
1271
1272 if (have_cd && num_written_clipdistance) {
1273 LLVMValueRef clipdist;
1274 int i;
1275 i = plane_idx - 6;
1276
1277 *have_clipdist = true;
1278 if (i < 4) {
1279 clipdist = LLVMBuildLoad2(builder, vec_type, outputs[cd[0]][i], "");
1280 } else {
1281 clipdist = LLVMBuildLoad2(builder, vec_type, outputs[cd[1]][i-4], "");
1282 }
1283 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1284 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1285 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1286 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1287 test = LLVMBuildAnd(builder, test, temp, "");
1288 mask = LLVMBuildOr(builder, mask, test, "");
1289 } else {
1290 LLVMTypeRef vs_elem_type = lp_build_elem_type(gallivm, vs_type);
1291 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1292 LLVMValueRef sum = NULL;
1293 indices[0] = lp_build_const_int32(gallivm, 0);
1294 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1295
1296 for (int i = 0; i < 4; ++i) {
1297 indices[2] = lp_build_const_int32(gallivm, i);
1298 plane_ptr = LLVMBuildGEP2(builder, planes_type, planes_ptr, indices, 3, "");
1299 plane1 = LLVMBuildLoad2(builder, vs_elem_type, plane_ptr,
1300 (const char *[]){"plane_x", "plane_y", "plane_z", "plane_w"}[i]);
1301 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1302 if (i == 0) {
1303 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1304 } else {
1305 sum = lp_build_fmuladd(builder, planes,
1306 (LLVMValueRef[]){cv_x, cv_y, cv_z, cv_w}[i], sum);
1307 }
1308 }
1309
1310 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1311 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1312 test = LLVMBuildAnd(builder, test, temp, "");
1313 mask = LLVMBuildOr(builder, mask, test, "");
1314 }
1315 }
1316 }
1317 if (key->need_edgeflags) {
1318 /*
1319 * This isn't really part of clipmask but stored the same in vertex
1320 * header later, so do it here.
1321 */
1322 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1323 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1324 LLVMValueRef edgeflag = LLVMBuildLoad2(builder, vec_type, outputs[edge_attr][0], "");
1325 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1326 temp = lp_build_const_int_vec(gallivm, i32_type,
1327 1LL << DRAW_TOTAL_CLIP_PLANES);
1328 test = LLVMBuildAnd(builder, test, temp, "");
1329 mask = LLVMBuildOr(builder, mask, test, "");
1330 }
1331 return mask;
1332 }
1333
1334
1335 /**
1336 * Returns boolean if any clipping has occurred
1337 * Used zero/one i8 value to represent boolean
1338 */
1339 static LLVMValueRef
clipmask_booli8(struct gallivm_state * gallivm,const struct lp_type vs_type,LLVMTypeRef clipmask_bool_type,LLVMValueRef clipmask_bool_ptr,bool edgeflag_in_clipmask)1340 clipmask_booli8(struct gallivm_state *gallivm,
1341 const struct lp_type vs_type,
1342 LLVMTypeRef clipmask_bool_type,
1343 LLVMValueRef clipmask_bool_ptr,
1344 bool edgeflag_in_clipmask)
1345 {
1346 LLVMBuilderRef builder = gallivm->builder;
1347 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1348 LLVMValueRef clipmask_bool = LLVMBuildLoad2(builder, clipmask_bool_type, clipmask_bool_ptr, "");
1349 LLVMValueRef ret;
1350 struct lp_build_context bldivec;
1351
1352 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1353
1354 /*
1355 * We need to invert the edgeflag bit from the clipmask here
1356 * (because the result is really if we want to run the pipeline or not
1357 * and we (may) need it if edgeflag was 0).
1358 */
1359 if (edgeflag_in_clipmask) {
1360 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1361 1LL << DRAW_TOTAL_CLIP_PLANES);
1362 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1363 }
1364
1365 /*
1366 * XXX: probably should mask off bits from the mask which come from
1367 * vertices which were beyond the count (i.e. indices_valid for
1368 * linear fetches, for elts ones we don't have the correct mask
1369 * right now). Otherwise might run the pipeline for nothing,
1370 * though everything should still work.
1371 */
1372 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1373 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1374 return ret;
1375 }
1376
1377
1378 static LLVMValueRef
draw_gs_llvm_fetch_input(const struct lp_build_gs_iface * gs_iface,struct lp_build_context * bld,bool is_vindex_indirect,LLVMValueRef vertex_index,bool is_aindex_indirect,LLVMValueRef attrib_index,LLVMValueRef swizzle_index)1379 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface *gs_iface,
1380 struct lp_build_context * bld,
1381 bool is_vindex_indirect,
1382 LLVMValueRef vertex_index,
1383 bool is_aindex_indirect,
1384 LLVMValueRef attrib_index,
1385 LLVMValueRef swizzle_index)
1386 {
1387 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1388 struct gallivm_state *gallivm = bld->gallivm;
1389 LLVMBuilderRef builder = gallivm->builder;
1390 LLVMValueRef indices[3];
1391 LLVMValueRef res;
1392 struct lp_type type = bld->type;
1393
1394 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
1395 LLVMTypeRef channel_vec_type = LLVMVectorType(float_type, TGSI_NUM_CHANNELS);
1396 LLVMTypeRef input_array_type = create_gs_jit_input_type_deref(gallivm);
1397
1398 if (is_vindex_indirect || is_aindex_indirect) {
1399 res = bld->zero;
1400 for (int i = 0; i < type.length; ++i) {
1401 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1402 LLVMValueRef vert_chan_index = vertex_index;
1403 LLVMValueRef attr_chan_index = attrib_index;
1404 LLVMValueRef channel_vec, value;
1405
1406 if (is_vindex_indirect) {
1407 vert_chan_index = LLVMBuildExtractElement(builder,
1408 vertex_index, idx, "");
1409 }
1410 if (is_aindex_indirect) {
1411 attr_chan_index = LLVMBuildExtractElement(builder,
1412 attrib_index, idx, "");
1413 }
1414
1415 indices[0] = vert_chan_index;
1416 indices[1] = attr_chan_index;
1417 indices[2] = swizzle_index;
1418
1419 channel_vec = LLVMBuildGEP2(builder, input_array_type, gs->input, indices, 3, "");
1420 channel_vec = LLVMBuildLoad2(builder, channel_vec_type, channel_vec, "");
1421 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1422
1423 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1424 }
1425 } else {
1426 indices[0] = vertex_index;
1427 indices[1] = attrib_index;
1428 indices[2] = swizzle_index;
1429
1430 res = LLVMBuildGEP2(builder, input_array_type, gs->input, indices, 3, "");
1431 res = LLVMBuildLoad2(builder, channel_vec_type, res, "");
1432 }
1433
1434 return res;
1435 }
1436
1437
1438 static void
draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface * gs_base,struct lp_build_context * bld,LLVMValueRef (* outputs)[4],LLVMValueRef emitted_vertices_vec,LLVMValueRef mask_vec,LLVMValueRef stream_id)1439 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface *gs_base,
1440 struct lp_build_context * bld,
1441 LLVMValueRef (*outputs)[4],
1442 LLVMValueRef emitted_vertices_vec,
1443 LLVMValueRef mask_vec, LLVMValueRef stream_id)
1444 {
1445 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1446 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1447 struct gallivm_state *gallivm = variant->gallivm;
1448 LLVMBuilderRef builder = gallivm->builder;
1449 struct lp_type gs_type = bld->type;
1450 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1451 lp_int_type(gs_type), 0);
1452 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1453 LLVMValueRef next_prim_offset =
1454 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1455 LLVMValueRef io = variant->io_ptr;
1456 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1457
1458 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
1459 for (unsigned i = 0; i < gs_type.length; ++i) {
1460 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1461 LLVMValueRef currently_emitted =
1462 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1463 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1464 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1465 indices[i] = LLVMBuildSelect(builder, LLVMBuildExtractElement(builder, cond, ind, ""), indices[i],
1466 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary - 1), "");
1467 }
1468
1469 LLVMValueRef stream_idx = LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), "");
1470 LLVMValueRef cnd = LLVMBuildICmp(builder, LLVMIntULT, stream_idx, lp_build_const_int32(gallivm, variant->shader->base.num_vertex_streams), "");
1471 struct lp_build_if_state if_ctx;
1472 lp_build_if(&if_ctx, gallivm, cnd);
1473 io = lp_build_pointer_get2(builder, variant->vertex_header_ptr_type,
1474 io, LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), ""));
1475
1476 if (variant->key.clamp_vertex_color) {
1477 do_clamp_vertex_color(gallivm, gs_type,
1478 gs_info, outputs);
1479 }
1480 convert_to_aos(gallivm, variant->vertex_header_type,
1481 io, indices,
1482 outputs, clipmask,
1483 gs_info->num_outputs, gs_type,
1484 -1,
1485 false);
1486 lp_build_endif(&if_ctx);
1487 }
1488
1489
1490 static void
draw_gs_llvm_end_primitive(const struct lp_build_gs_iface * gs_base,struct lp_build_context * bld,LLVMValueRef total_emitted_vertices_vec_ptr,LLVMValueRef verts_per_prim_vec,LLVMValueRef emitted_prims_vec,LLVMValueRef mask_vec,unsigned stream)1491 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface *gs_base,
1492 struct lp_build_context * bld,
1493 LLVMValueRef total_emitted_vertices_vec_ptr,
1494 LLVMValueRef verts_per_prim_vec,
1495 LLVMValueRef emitted_prims_vec,
1496 LLVMValueRef mask_vec, unsigned stream)
1497 {
1498 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1499 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1500 struct gallivm_state *gallivm = variant->gallivm;
1501 LLVMBuilderRef builder = gallivm->builder;
1502 LLVMValueRef prim_lengts_ptr =
1503 draw_gs_jit_prim_lengths(variant, variant->context_ptr);
1504
1505 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
1506 for (unsigned i = 0; i < bld->type.length; ++i) {
1507 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1508 LLVMValueRef prims_emitted =
1509 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1510 LLVMValueRef store_ptr;
1511 LLVMValueRef num_vertices =
1512 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1513
1514 LLVMValueRef this_cond = LLVMBuildExtractElement(gallivm->builder, cond, ind, "");
1515 struct lp_build_if_state ifthen;
1516 lp_build_if(&ifthen, gallivm, this_cond);
1517 prims_emitted = LLVMBuildMul(gallivm->builder, prims_emitted, lp_build_const_int32(gallivm, variant->shader->base.num_vertex_streams), "");
1518 prims_emitted = LLVMBuildAdd(gallivm->builder, prims_emitted, lp_build_const_int32(gallivm, stream), "");
1519 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
1520 LLVMTypeRef prim_lengths_type = LLVMPointerType(int_type, 0);
1521 store_ptr = LLVMBuildGEP2(builder, prim_lengths_type, prim_lengts_ptr, &prims_emitted, 1, "");
1522 store_ptr = LLVMBuildLoad2(builder, prim_lengths_type, store_ptr, "");
1523 store_ptr = LLVMBuildGEP2(builder, int_type, store_ptr, &ind, 1, "");
1524 LLVMBuildStore(builder, num_vertices, store_ptr);
1525 lp_build_endif(&ifthen);
1526 }
1527 }
1528
1529
1530 static void
draw_gs_llvm_epilogue(const struct lp_build_gs_iface * gs_base,LLVMValueRef total_emitted_vertices_vec,LLVMValueRef emitted_prims_vec,unsigned stream)1531 draw_gs_llvm_epilogue(const struct lp_build_gs_iface *gs_base,
1532 LLVMValueRef total_emitted_vertices_vec,
1533 LLVMValueRef emitted_prims_vec, unsigned stream)
1534 {
1535 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1536 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1537 struct gallivm_state *gallivm = variant->gallivm;
1538 LLVMBuilderRef builder = gallivm->builder;
1539 LLVMValueRef emitted_verts_ptr =
1540 draw_gs_jit_emitted_vertices(variant, variant->context_ptr);
1541 LLVMValueRef emitted_prims_ptr =
1542 draw_gs_jit_emitted_prims(variant, variant->context_ptr);
1543 LLVMValueRef stream_val = lp_build_const_int32(gallivm, stream);
1544
1545 emitted_verts_ptr = LLVMBuildGEP2(builder, LLVMTypeOf(total_emitted_vertices_vec), emitted_verts_ptr, &stream_val, 1, "");
1546 emitted_prims_ptr = LLVMBuildGEP2(builder, LLVMTypeOf(emitted_prims_vec), emitted_prims_ptr, &stream_val, 1, "");
1547
1548 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1549 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1550 }
1551
1552
1553 static void
draw_llvm_generate(struct draw_llvm * llvm,struct draw_llvm_variant * variant)1554 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1555 {
1556 struct gallivm_state *gallivm = variant->gallivm;
1557 LLVMContextRef context = gallivm->context;
1558 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1559 LLVMTypeRef arg_types[14];
1560 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1561 LLVMTypeRef func_type;
1562 LLVMValueRef context_ptr;
1563 LLVMValueRef resources_ptr;
1564 LLVMBasicBlockRef block;
1565 LLVMBuilderRef builder;
1566 char func_name[64];
1567 struct lp_type vs_type;
1568 LLVMValueRef count, fetch_elts, start;
1569 LLVMValueRef vertex_id_offset;
1570 LLVMValueRef stride, step, io_itr;
1571 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1572 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1573 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1574 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1575 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1576 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1577 LLVMValueRef fake_buf_ptr, fake_buf;
1578
1579 struct draw_context *draw = llvm->draw;
1580 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1581 unsigned i, j;
1582 struct lp_build_context bld, blduivec;
1583 struct lp_build_loop_state lp_loop;
1584 struct lp_build_if_state if_ctx;
1585 const int vector_length = lp_native_vector_width / 32;
1586 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1587 struct lp_build_sampler_soa *sampler = 0;
1588 struct lp_build_image_soa *image = NULL;
1589 LLVMValueRef ret, clipmask_bool_ptr;
1590 struct draw_llvm_variant_key *key = &variant->key;
1591 /* If geometry shader is present we need to skip both the viewport
1592 * transformation and clipping otherwise the inputs to the geometry
1593 * shader will be incorrect.
1594 * The code can't handle vp transform when vs writes vp index neither
1595 * (though this would be fixable here, but couldn't just broadcast
1596 * the values).
1597 */
1598 const bool bypass_viewport = key->has_gs_or_tes || key->bypass_viewport ||
1599 vs_info->writes_viewport_index;
1600 const bool enable_cliptest = !key->has_gs_or_tes && (key->clip_xy ||
1601 key->clip_z ||
1602 key->clip_user ||
1603 key->need_edgeflags);
1604 LLVMValueRef variant_func;
1605 const unsigned pos = draw->vs.position_output;
1606 const unsigned cv = draw->vs.clipvertex_output;
1607 bool have_clipdist = false;
1608 struct lp_bld_tgsi_system_values system_values;
1609
1610 memset(&system_values, 0, sizeof(system_values));
1611 memset(&outputs, 0, sizeof(outputs));
1612 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant");
1613
1614 i = 0;
1615 arg_types[i++] = get_context_ptr_type(variant); /* context */
1616 arg_types[i++] = variant->resources_ptr_type; /* context */
1617 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1618 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1619 arg_types[i++] = int32_type; /* count */
1620 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1621 arg_types[i++] = int32_type; /* stride */
1622 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1623 arg_types[i++] = int32_type; /* instance_id */
1624 arg_types[i++] = int32_type; /* vertex_id_offset */
1625 arg_types[i++] = int32_type; /* start_instance */
1626 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1627 arg_types[i++] = int32_type; /* draw_id */
1628 arg_types[i++] = int32_type; /* view_id */
1629 assert(i == ARRAY_SIZE(arg_types));
1630
1631 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1632 arg_types, num_arg_types, 0);
1633
1634 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1635 variant->function = variant_func;
1636
1637 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1638 for (i = 0; i < num_arg_types; ++i)
1639 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1640 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1641
1642 if (gallivm->cache && gallivm->cache->data_size)
1643 return;
1644
1645 context_ptr = LLVMGetParam(variant_func, 0);
1646 resources_ptr = LLVMGetParam(variant_func, 1);
1647 io_ptr = LLVMGetParam(variant_func, 2);
1648 vbuffers_ptr = LLVMGetParam(variant_func, 3);
1649 count = LLVMGetParam(variant_func, 4);
1650 start = LLVMGetParam(variant_func, 5);
1651 /*
1652 * XXX: stride is actually unused. The stride we use is strictly calculated
1653 * from the number of outputs (including the draw_extra outputs).
1654 * Should probably fix some day (we need a new vs just because of extra
1655 * outputs which the generated vs won't touch).
1656 */
1657 stride = LLVMGetParam(variant_func, 6);
1658 vb_ptr = LLVMGetParam(variant_func, 7);
1659 system_values.instance_id = LLVMGetParam(variant_func, 8);
1660 vertex_id_offset = LLVMGetParam(variant_func, 9);
1661 system_values.base_instance = LLVMGetParam(variant_func, 10);
1662 fetch_elts = LLVMGetParam(variant_func, 11);
1663 system_values.draw_id = LLVMGetParam(variant_func, 12);
1664 system_values.view_index = LLVMGetParam(variant_func, 13);
1665
1666 lp_build_name(context_ptr, "context");
1667 lp_build_name(resources_ptr, "resources");
1668 lp_build_name(io_ptr, "io");
1669 lp_build_name(vbuffers_ptr, "vbuffers");
1670 lp_build_name(count, "count");
1671 lp_build_name(start, "start");
1672 lp_build_name(stride, "stride");
1673 lp_build_name(vb_ptr, "vb");
1674 lp_build_name(system_values.instance_id, "instance_id");
1675 lp_build_name(vertex_id_offset, "vertex_id_offset");
1676 lp_build_name(system_values.base_instance, "start_instance");
1677 lp_build_name(fetch_elts, "fetch_elts");
1678 lp_build_name(system_values.draw_id, "draw_id");
1679
1680 /*
1681 * Function body
1682 */
1683
1684 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1685 builder = gallivm->builder;
1686 LLVMPositionBuilderAtEnd(builder, block);
1687
1688 memset(&vs_type, 0, sizeof vs_type);
1689 vs_type.floating = true; /* floating point values */
1690 vs_type.sign = true; /* values are signed */
1691 vs_type.norm = false; /* values are not limited to [0,1] or [-1,1] */
1692 vs_type.width = 32; /* 32-bit float */
1693 vs_type.length = vector_length;
1694
1695 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1696 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1697
1698 /* hold temporary "bool" clipmask */
1699 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1700
1701 fake_buf = lp_build_alloca_undef(gallivm,
1702 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1703 fake_buf = LLVMBuildBitCast(builder, fake_buf,
1704 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1705 fake_buf_ptr = LLVMBuildGEP2(builder, LLVMInt8TypeInContext(context), fake_buf, &bld.zero, 1, "");
1706
1707 /* code generated texture sampling */
1708 sampler = lp_bld_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key),
1709 MAX2(key->nr_samplers,
1710 key->nr_sampler_views));
1711 image = lp_bld_llvm_image_soa_create(draw_llvm_variant_key_images(key),
1712 key->nr_images);
1713
1714 step = lp_build_const_int32(gallivm, vector_length);
1715
1716 ind_vec = blduivec.undef;
1717 for (i = 0; i < vs_type.length; i++) {
1718 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1719 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1720 }
1721
1722 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1723 LLVMConstPointerNull(arg_types[11]), fetch_elts, "");
1724
1725 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1726 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1727 /*
1728 * Only needed for non-indexed path.
1729 */
1730 start_vec = lp_build_broadcast_scalar(&blduivec, start);
1731
1732 /*
1733 * Pre-calculate everything which is constant per shader invocation.
1734 */
1735 for (j = 0; j < key->nr_vertex_elements; ++j) {
1736 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1737 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1738 struct pipe_vertex_element *velem = &key->vertex_element[j];
1739 LLVMValueRef vb_index =
1740 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1741 LLVMValueRef bsize = lp_build_const_int32(gallivm,
1742 util_format_get_blocksize(velem->src_format));
1743 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1744 velem->src_offset);
1745 LLVMValueRef src_stride = lp_build_const_int32(gallivm,
1746 velem->src_stride);
1747 struct lp_build_if_state if_ctx;
1748
1749 if (velem->src_format != PIPE_FORMAT_NONE) {
1750 vbuffer_ptr = LLVMBuildGEP2(builder, variant->buffer_type, vbuffers_ptr, &vb_index, 1, "");
1751 vb_info = LLVMBuildGEP2(builder, variant->vb_type, vb_ptr, &vb_index, 1, "");
1752 vb_stride[j] = src_stride;
1753 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, variant->vb_type, vb_info);
1754 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, variant->buffer_type, vbuffer_ptr);
1755 buffer_size = draw_jit_dvbuffer_size(gallivm, variant->buffer_type, vbuffer_ptr);
1756
1757 ofbit = NULL;
1758 /*
1759 * We'll set buffer_size_adj to zero if we have of, so it will
1760 * always overflow later automatically without having to keep ofbit.
1761 * Overflows (with normal wraparound) doing the actual offset
1762 * calculation should be ok, just not for the buffer size calc.
1763 * It would also be possible to detect such overflows and return
1764 * zeros if that happens, but this would be more complex.
1765 */
1766 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1767 tmp = lp_build_sub(&bld, bsize, bld.one);
1768 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1769 &ofbit);
1770 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1771 buf_offset, &ofbit);
1772
1773 /*
1774 * We can't easily set fake vertex buffers outside the generated code.
1775 * Hence, set fake vertex buffers here instead basically, so fetch
1776 * code can always fetch using offset 0, eliminating all control flow
1777 * inside the main loop.
1778 * (Alternatively, could have control flow per vector skipping fetch
1779 * if ofbit is true.)
1780 */
1781 if (velem->instance_divisor) {
1782 /*
1783 * Index is equal to the start instance plus the number of current
1784 * instance divided by the divisor. In this case we compute it as:
1785 * index = start_instance + (instance_id / divisor).
1786 * Note we could actually do the fetch here, outside the loop -
1787 * it's all constant, hopefully llvm recognizes this.
1788 */
1789 LLVMValueRef current_instance;
1790 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1791 lp_build_const_int32(gallivm,
1792 velem->instance_divisor),
1793 "instance_divisor");
1794 instance_index[j] = lp_build_uadd_overflow(gallivm, system_values.base_instance,
1795 current_instance, &ofbit);
1796 }
1797
1798 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
1799 buffer_size_adj[j], "");
1800
1801 LLVMTypeRef byte_type = LLVMInt8TypeInContext(context);
1802 LLVMTypeRef byte_ptr_type = LLVMPointerType(byte_type, 0);
1803 temp_ptr = lp_build_alloca_undef(gallivm, byte_ptr_type, "");
1804
1805 lp_build_if(&if_ctx, gallivm, ofbit);
1806 {
1807 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1808 }
1809 lp_build_else(&if_ctx);
1810 {
1811 map_ptr[j] = LLVMBuildGEP2(builder, byte_type, map_ptr[j], &buf_offset, 1, "");
1812 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1813 }
1814 lp_build_endif(&if_ctx);
1815 map_ptr[j] = LLVMBuildLoad2(builder, byte_ptr_type, temp_ptr, "map_ptr");
1816
1817 if (0) {
1818 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1819 lp_build_const_int32(gallivm, j),
1820 vb_index, vb_stride[j]);
1821 lp_build_printf(gallivm,
1822 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1823 vb_buffer_offset, src_offset, buf_offset);
1824 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
1825 buffer_size, bsize);
1826 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
1827 }
1828 }
1829 }
1830
1831 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1832 {
1833 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1834 LLVMValueRef io;
1835 LLVMValueRef clipmask; /* holds the clipmask value */
1836 LLVMValueRef true_index_array, index_store;
1837 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1838
1839 io_itr = lp_loop.counter;
1840
1841 io = LLVMBuildGEP2(builder, variant->vertex_header_type, io_ptr, &io_itr, 1, "");
1842 #if DEBUG_STORE
1843 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1844 io_itr, io, lp_loop.counter);
1845 #endif
1846
1847 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
1848 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1849
1850 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
1851 /*
1852 * Limit indices to fetch_max, otherwise might try to access indices
1853 * beyond index buffer (or rather vsplit elt buffer) size.
1854 * Could probably safely (?) skip this for non-indexed draws and
1855 * simplify things minimally (by removing it could combine the ind_vec
1856 * and start_vec adds). I think the only effect for non-indexed draws will
1857 * be that for the invalid elements they will be all fetched from the
1858 * same location as the last valid one, but noone should really care.
1859 */
1860 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
1861
1862 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
1863
1864 lp_build_if(&if_ctx, gallivm, have_elts);
1865 {
1866 /*
1867 * Note: you'd expect some comparison/clamp against fetch_elt_max
1868 * here.
1869 * There used to be one here but it was incorrect: overflow was
1870 * detected if index > fetch_elt_max - but the correct condition
1871 * would be index >= fetch_elt_max (since this is just size of elts
1872 * buffer / element size).
1873 * Using the correct condition however will cause failures - due to
1874 * vsplit/vcache code which rebases indices. So, as an example, if
1875 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1876 * replace all invalid indices with 0 - which in case of elt_bias
1877 * not being zero will get a different fetch index than the valid
1878 * index 0. So, just rely on vsplit code preventing out-of-bounds
1879 * fetches. This is also why it's safe to do elts fetch even if there
1880 * was no index buffer bound - the real buffer is never seen here, at
1881 * least not if there are index buffer overflows...
1882 */
1883
1884 /*
1885 * XXX should not have to do this, as scale can be handled
1886 * natively by loads (hits asserts though).
1887 */
1888 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
1889 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1890 LLVMPointerType(LLVMInt8TypeInContext(context),
1891 0), "");
1892 tmp = lp_build_gather(gallivm, vs_type.length,
1893 32, bld.type, true,
1894 fetch_elts, tmp, false);
1895 LLVMBuildStore(builder, tmp, index_store);
1896 }
1897 lp_build_else(&if_ctx);
1898 {
1899 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
1900 LLVMBuildStore(builder, tmp, index_store);
1901 }
1902 lp_build_endif(&if_ctx);
1903
1904 true_index_array = LLVMBuildLoad2(builder, blduivec.vec_type, index_store, "");
1905
1906 for (j = 0; j < key->nr_vertex_elements; ++j) {
1907 struct pipe_vertex_element *velem = &key->vertex_element[j];
1908 const struct util_format_description *format_desc =
1909 util_format_description(velem->src_format);
1910
1911 if (format_desc->format == PIPE_FORMAT_NONE) {
1912 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1913 inputs[j][i] = lp_build_zero(gallivm, vs_type);
1914 }
1915 } else if (velem->instance_divisor) {
1916 fetch_instanced(gallivm, format_desc, vs_type,
1917 vb_stride[j], map_ptr[j],
1918 buffer_size_adj[j],
1919 inputs[j], instance_index[j]);
1920 } else {
1921 fetch_vector(gallivm, format_desc, vs_type,
1922 vb_stride[j], map_ptr[j],
1923 buffer_size_adj[j],
1924 inputs[j], true_index_array);
1925 }
1926 }
1927
1928 struct lp_build_mask_context mask;
1929
1930 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
1931 /* In the paths with elts vertex id has to be unaffected by the
1932 * index bias and because indices inside our elements array have
1933 * already had index bias applied we need to subtract it here to
1934 * get back to the original index.
1935 * In the linear paths vertex id has to be unaffected by the
1936 * original start index and because we abuse the 'start' variable
1937 * to either represent the actual start index or the index at which
1938 * the primitive was split (we split rendering into chunks of at
1939 * most 4095-vertices) we need to back out the original start
1940 * index out of our vertex id here.
1941 * for ARB_shader_draw_parameters, base_vertex should be 0 for
1942 * non-indexed draws.
1943 */
1944 LLVMValueRef base_vertex = lp_build_select(&bld, have_elts, vertex_id_offset, lp_build_const_int32(gallivm, 0));
1945 system_values.basevertex = lp_build_broadcast_scalar(&blduivec, base_vertex);
1946
1947 /* first vertex is for Vulkan base vertex support */
1948 LLVMValueRef first_vertex = vertex_id_offset;
1949 system_values.firstvertex = lp_build_broadcast_scalar(&blduivec, first_vertex);
1950
1951 system_values.vertex_id = true_index_array;
1952 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1953 lp_build_broadcast_scalar(&blduivec, vertex_id_offset), "");
1954
1955 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1956 generate_vs(variant,
1957 builder,
1958 vs_type,
1959 outputs,
1960 ptr_aos,
1961 &system_values,
1962 context_ptr,
1963 resources_ptr,
1964 sampler,
1965 image,
1966 key->clamp_vertex_color,
1967 &mask);
1968
1969 lp_build_mask_end(&mask);
1970 if (pos != -1 && cv != -1) {
1971 /* store original positions in clip before further manipulation */
1972 store_clip(gallivm, vs_type, variant->vertex_header_type, io, outputs, pos);
1973
1974 /* do cliptest */
1975 if (enable_cliptest) {
1976 LLVMValueRef temp = LLVMBuildLoad2(builder, blduivec.vec_type, clipmask_bool_ptr, "");
1977 /* allocate clipmask, assign it integer type */
1978 clipmask = generate_clipmask(llvm,
1979 gallivm,
1980 vs_type,
1981 outputs,
1982 key,
1983 variant->context_type,
1984 context_ptr, &have_clipdist);
1985 temp = LLVMBuildOr(builder, clipmask, temp, "");
1986 /* store temporary clipping boolean value */
1987 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1988 } else {
1989 clipmask = blduivec.zero;
1990 }
1991
1992 /* do viewport mapping */
1993 if (!bypass_viewport) {
1994 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1995 }
1996 } else {
1997 clipmask = blduivec.zero;
1998 }
1999
2000 /* store clipmask in vertex header,
2001 * original positions in clip
2002 * and transformed positions in data
2003 */
2004 convert_to_aos(gallivm, variant->vertex_header_type, io, NULL, outputs, clipmask,
2005 vs_info->num_outputs, vs_type, -1,
2006 enable_cliptest && key->need_edgeflags);
2007 }
2008 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
2009
2010 lp_bld_llvm_sampler_soa_destroy(sampler);
2011 lp_bld_llvm_image_soa_destroy(image);
2012
2013 /* return clipping boolean value for function */
2014 ret = clipmask_booli8(gallivm, vs_type, blduivec.vec_type, clipmask_bool_ptr,
2015 enable_cliptest && key->need_edgeflags);
2016
2017 LLVMBuildRet(builder, ret);
2018
2019 gallivm_verify_function(gallivm, variant_func);
2020 }
2021
2022
2023 struct draw_llvm_variant_key *
draw_llvm_make_variant_key(struct draw_llvm * llvm,char * store)2024 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2025 {
2026 struct draw_llvm_variant_key *key;
2027 struct lp_sampler_static_state *draw_sampler;
2028 struct lp_image_static_state *draw_image;
2029
2030 key = (struct draw_llvm_variant_key *)store;
2031
2032 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2033
2034
2035 /* will have to rig this up properly later */
2036 key->clip_xy = llvm->draw->clip_xy;
2037 key->clip_z = llvm->draw->clip_z;
2038 key->clip_user = llvm->draw->clip_user;
2039 key->bypass_viewport = llvm->draw->bypass_viewport;
2040 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2041 /* XXX assumes edgeflag output not at 0 */
2042 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? true : false);
2043 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2044 key->has_gs_or_tes = llvm->draw->gs.geometry_shader != NULL || llvm->draw->tes.tess_eval_shader != NULL;
2045 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2046
2047 key->clamp_vertex_color = !key->has_gs_or_tes &&
2048 llvm->draw->rasterizer->clamp_vertex_color;
2049
2050 /* All variants of this shader will have the same value for
2051 * nr_samplers. Not yet trying to compact away holes in the
2052 * sampler array.
2053 */
2054 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2055 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2056 key->nr_sampler_views =
2057 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2058 } else {
2059 key->nr_sampler_views = key->nr_samplers;
2060 }
2061
2062 key->nr_images = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2063
2064 /* Presumably all variants of the shader should have the same
2065 * number of vertex elements - ie the number of shader inputs.
2066 * NOTE: we NEED to store the needed number of needed inputs
2067 * here, not the number of provided elements to match keysize
2068 * (and the offset of sampler state in the key).
2069 * If we have excess number of vertex elements, this is valid,
2070 * but the excess ones don't matter.
2071 * If we don't have enough vertex elements (which looks not really
2072 * valid but we'll handle it gracefully) fill out missing ones with
2073 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2074 */
2075 key->nr_vertex_elements =
2076 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2077
2078 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2079 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2080 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2081 memset(key->vertex_element, 0,
2082 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2083 }
2084 memcpy(key->vertex_element,
2085 llvm->draw->pt.vertex_element,
2086 sizeof(struct pipe_vertex_element) *
2087 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2088
2089 draw_sampler = draw_llvm_variant_key_samplers(key);
2090 memset(draw_sampler, 0,
2091 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2092
2093 for (unsigned i = 0 ; i < key->nr_samplers; i++) {
2094 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2095 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2096 }
2097 for (unsigned i = 0 ; i < key->nr_sampler_views; i++) {
2098 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2099 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2100 }
2101
2102 draw_image = draw_llvm_variant_key_images(key);
2103 memset(draw_image, 0,
2104 key->nr_images * sizeof *draw_image);
2105 for (unsigned i = 0; i < key->nr_images; i++) {
2106 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2107 llvm->draw->images[PIPE_SHADER_VERTEX][i]);
2108 }
2109 return key;
2110 }
2111
2112
2113 void
draw_llvm_dump_variant_key(struct draw_llvm_variant_key * key)2114 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2115 {
2116 struct lp_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2117 struct lp_image_static_state *image = draw_llvm_variant_key_images(key);
2118 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2119 debug_printf("clip_xy = %u\n", key->clip_xy);
2120 debug_printf("clip_z = %u\n", key->clip_z);
2121 debug_printf("clip_user = %u\n", key->clip_user);
2122 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2123 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2124 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2125 debug_printf("has_gs_or_tes = %u\n", key->has_gs_or_tes);
2126 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2127
2128 for (unsigned i = 0 ; i < key->nr_vertex_elements; i++) {
2129 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2130 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2131 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2132 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2133 }
2134
2135 for (unsigned i = 0 ; i < key->nr_sampler_views; i++) {
2136 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2137 }
2138
2139 for (unsigned i = 0 ; i < key->nr_images; i++)
2140 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2141 }
2142
2143
2144 void
draw_llvm_set_mapped_texture(struct draw_context * draw,enum pipe_shader_type shader_stage,unsigned sview_idx,uint32_t width,uint32_t height,uint32_t depth,uint32_t first_level,uint32_t last_level,uint32_t num_samples,uint32_t sample_stride,const void * base_ptr,uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])2145 draw_llvm_set_mapped_texture(struct draw_context *draw,
2146 enum pipe_shader_type shader_stage,
2147 unsigned sview_idx,
2148 uint32_t width, uint32_t height, uint32_t depth,
2149 uint32_t first_level, uint32_t last_level,
2150 uint32_t num_samples,
2151 uint32_t sample_stride,
2152 const void *base_ptr,
2153 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2154 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2155 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2156 {
2157 struct lp_jit_texture *jit_tex;
2158
2159 assert(shader_stage < DRAW_MAX_SHADER_STAGE);
2160 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_resources[shader_stage].textures));
2161
2162 jit_tex = &draw->llvm->jit_resources[shader_stage].textures[sview_idx];
2163 jit_tex->width = width;
2164 jit_tex->height = height;
2165 jit_tex->depth = depth;
2166 jit_tex->first_level = first_level;
2167 jit_tex->last_level = last_level;
2168 jit_tex->mip_offsets[0] = 0;
2169 jit_tex->base = base_ptr;
2170 if (num_samples > 1) {
2171 jit_tex->mip_offsets[LP_JIT_TEXTURE_SAMPLE_STRIDE] = sample_stride;
2172 jit_tex->row_stride[0] = row_stride[0];
2173 jit_tex->img_stride[0] = img_stride[0];
2174 jit_tex->last_level = num_samples;
2175 } else {
2176 for (unsigned j = first_level; j <= last_level; j++) {
2177 jit_tex->mip_offsets[j] = mip_offsets[j];
2178 jit_tex->row_stride[j] = row_stride[j];
2179 jit_tex->img_stride[j] = img_stride[j];
2180 }
2181 }
2182 }
2183
2184
2185 void
draw_llvm_set_mapped_image(struct draw_context * draw,enum pipe_shader_type shader_stage,unsigned idx,uint32_t width,uint32_t height,uint32_t depth,const void * base_ptr,uint32_t row_stride,uint32_t img_stride,uint32_t num_samples,uint32_t sample_stride)2186 draw_llvm_set_mapped_image(struct draw_context *draw,
2187 enum pipe_shader_type shader_stage,
2188 unsigned idx,
2189 uint32_t width, uint32_t height, uint32_t depth,
2190 const void *base_ptr,
2191 uint32_t row_stride,
2192 uint32_t img_stride,
2193 uint32_t num_samples,
2194 uint32_t sample_stride)
2195 {
2196 struct lp_jit_image *jit_image;
2197
2198 assert(shader_stage < DRAW_MAX_SHADER_STAGE);
2199 assert(idx < ARRAY_SIZE(draw->llvm->jit_resources[shader_stage].images));
2200
2201 jit_image = &draw->llvm->jit_resources[shader_stage].images[idx];
2202
2203 jit_image->width = width;
2204 jit_image->height = height;
2205 jit_image->depth = depth;
2206 jit_image->base = base_ptr;
2207
2208 jit_image->row_stride = row_stride;
2209 jit_image->img_stride = img_stride;
2210 jit_image->num_samples = num_samples;
2211 jit_image->sample_stride = sample_stride;
2212 }
2213
2214
2215 void
draw_llvm_set_sampler_state(struct draw_context * draw,enum pipe_shader_type shader_type)2216 draw_llvm_set_sampler_state(struct draw_context *draw,
2217 enum pipe_shader_type shader_type)
2218 {
2219 assert(shader_type < DRAW_MAX_SHADER_STAGE);
2220 for (unsigned i = 0; i < draw->num_samplers[shader_type]; i++) {
2221 struct lp_jit_sampler *jit_sam = &draw->llvm->jit_resources[shader_type].samplers[i];
2222
2223 if (draw->samplers[shader_type][i]) {
2224 const struct pipe_sampler_state *s
2225 = draw->samplers[shader_type][i];
2226 jit_sam->min_lod = s->min_lod;
2227 jit_sam->max_lod = s->max_lod;
2228 jit_sam->lod_bias = s->lod_bias;
2229 jit_sam->max_aniso = s->max_anisotropy;
2230 COPY_4V(jit_sam->border_color, s->border_color.f);
2231 }
2232 }
2233 }
2234
2235
2236 void
draw_llvm_destroy_variant(struct draw_llvm_variant * variant)2237 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2238 {
2239 struct draw_llvm *llvm = variant->llvm;
2240
2241 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2242 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2243 variant->shader->variants_cached, llvm->nr_variants);
2244 }
2245
2246 gallivm_destroy(variant->gallivm);
2247
2248 list_del(&variant->list_item_local.list);
2249 variant->shader->variants_cached--;
2250 list_del(&variant->list_item_global.list);
2251 llvm->nr_variants--;
2252 FREE(variant);
2253 }
2254
2255
2256 /**
2257 * Create LLVM types for various structures.
2258 */
2259 static void
create_gs_jit_types(struct draw_gs_llvm_variant * var)2260 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2261 {
2262 struct gallivm_state *gallivm = var->gallivm;
2263
2264 var->context_type = create_gs_jit_context_type(gallivm,
2265 var->shader->base.vector_length,
2266 "draw_gs_jit_context");
2267 var->context_ptr_type = LLVMPointerType(var->context_type, 0);
2268
2269 var->resources_type = lp_build_jit_resources_type(gallivm);
2270 var->resources_ptr_type = LLVMPointerType(var->resources_type, 0);
2271 var->input_array_type = create_gs_jit_input_type(gallivm);
2272 }
2273
2274
2275 static LLVMTypeRef
get_gs_context_ptr_type(struct draw_gs_llvm_variant * variant)2276 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2277 {
2278 if (!variant->context_ptr_type)
2279 create_gs_jit_types(variant);
2280 return variant->context_ptr_type;
2281 }
2282
2283
2284 static LLVMValueRef
generate_mask_value(struct draw_gs_llvm_variant * variant,struct lp_type gs_type)2285 generate_mask_value(struct draw_gs_llvm_variant *variant,
2286 struct lp_type gs_type)
2287 {
2288 struct gallivm_state *gallivm = variant->gallivm;
2289 LLVMBuilderRef builder = gallivm->builder;
2290 struct lp_type mask_type = lp_int_type(gs_type);
2291 LLVMValueRef num_prims;
2292 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2293
2294 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2295 variant->num_prims);
2296 for (unsigned i = 0; i < gs_type.length; i++) {
2297 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2298 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2299 }
2300 mask_val = lp_build_compare(gallivm, mask_type,
2301 PIPE_FUNC_GREATER, num_prims, mask_val);
2302
2303 return mask_val;
2304 }
2305
2306
2307 static void
draw_gs_llvm_generate(struct draw_llvm * llvm,struct draw_gs_llvm_variant * variant)2308 draw_gs_llvm_generate(struct draw_llvm *llvm,
2309 struct draw_gs_llvm_variant *variant)
2310 {
2311 struct gallivm_state *gallivm = variant->gallivm;
2312 LLVMContextRef context = gallivm->context;
2313 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2314 LLVMTypeRef arg_types[9];
2315 LLVMTypeRef func_type;
2316 LLVMValueRef variant_func;
2317 LLVMValueRef context_ptr;
2318 LLVMValueRef resources_ptr;
2319 LLVMValueRef prim_id_ptr;
2320 LLVMBasicBlockRef block;
2321 LLVMBuilderRef builder;
2322 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2323 struct lp_build_sampler_soa *sampler = 0;
2324 struct lp_build_image_soa *image = NULL;
2325 struct lp_build_context bld;
2326 struct lp_bld_tgsi_system_values system_values;
2327 char func_name[64];
2328 struct lp_type gs_type;
2329 struct draw_gs_llvm_iface gs_iface;
2330 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2331 LLVMValueRef consts_ptr;
2332 LLVMValueRef ssbos_ptr;
2333 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2334 struct lp_build_mask_context mask;
2335 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2336 unsigned vector_length = variant->shader->base.vector_length;
2337
2338 memset(&system_values, 0, sizeof(system_values));
2339 memset(&outputs, 0, sizeof(outputs));
2340
2341 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant");
2342
2343 assert(variant->vertex_header_ptr_type);
2344
2345 LLVMTypeRef prim_id_type = LLVMVectorType(int32_type, vector_length);
2346 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2347 arg_types[1] = variant->resources_ptr_type;
2348 arg_types[2] = variant->input_array_type; /* input */
2349 arg_types[3] = LLVMPointerType(variant->vertex_header_ptr_type, 0); /* vertex_header */
2350 arg_types[4] = int32_type; /* num_prims */
2351 arg_types[5] = int32_type; /* instance_id */
2352 arg_types[6] = LLVMPointerType(prim_id_type, 0); /* prim_id_ptr */
2353 arg_types[7] = int32_type;
2354 arg_types[8] = int32_type;
2355
2356 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2357
2358 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2359
2360 variant->function = variant_func;
2361
2362 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2363
2364 for (unsigned i = 0; i < ARRAY_SIZE(arg_types); ++i)
2365 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2366 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2367
2368 if (gallivm->cache && gallivm->cache->data_size)
2369 return;
2370 context_ptr = LLVMGetParam(variant_func, 0);
2371 resources_ptr = LLVMGetParam(variant_func, 1);
2372 input_array = LLVMGetParam(variant_func, 2);
2373 io_ptr = LLVMGetParam(variant_func, 3);
2374 num_prims = LLVMGetParam(variant_func, 4);
2375 system_values.instance_id = LLVMGetParam(variant_func, 5);
2376 prim_id_ptr = LLVMGetParam(variant_func, 6);
2377 system_values.invocation_id = LLVMGetParam(variant_func, 7);
2378 system_values.view_index = LLVMGetParam(variant_func, 8);
2379
2380 lp_build_name(context_ptr, "context");
2381 lp_build_name(resources_ptr, "resources");
2382 lp_build_name(input_array, "input");
2383 lp_build_name(io_ptr, "io");
2384 lp_build_name(num_prims, "num_prims");
2385 lp_build_name(system_values.instance_id, "instance_id");
2386 lp_build_name(prim_id_ptr, "prim_id_ptr");
2387 lp_build_name(system_values.invocation_id, "invocation_id");
2388 lp_build_name(system_values.view_index, "view_index");
2389
2390 variant->context_ptr = context_ptr;
2391 variant->io_ptr = io_ptr;
2392 variant->num_prims = num_prims;
2393
2394 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2395 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2396 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2397 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2398 gs_iface.input = input_array;
2399 gs_iface.variant = variant;
2400
2401 /*
2402 * Function body
2403 */
2404
2405 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2406 builder = gallivm->builder;
2407 LLVMPositionBuilderAtEnd(builder, block);
2408
2409 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2410
2411 memset(&gs_type, 0, sizeof gs_type);
2412 gs_type.floating = true; /* floating point values */
2413 gs_type.sign = true; /* values are signed */
2414 gs_type.norm = false; /* values are not limited to [0,1] or [-1,1] */
2415 gs_type.width = 32; /* 32-bit float */
2416 gs_type.length = vector_length;
2417
2418 consts_ptr = lp_jit_resources_constants(gallivm, variant->resources_type, resources_ptr);
2419
2420 ssbos_ptr = lp_jit_resources_ssbos(gallivm, variant->resources_type, resources_ptr);
2421
2422 /* code generated texture sampling */
2423 sampler = lp_bld_llvm_sampler_soa_create(variant->key.samplers,
2424 MAX2(variant->key.nr_samplers,
2425 variant->key.nr_sampler_views));
2426 image = lp_bld_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant->key),
2427 variant->key.nr_images);
2428 mask_val = generate_mask_value(variant, gs_type);
2429 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2430
2431 if (gs_info->uses_primid) {
2432 system_values.prim_id = LLVMBuildLoad2(builder, prim_id_type, prim_id_ptr, "prim_id");
2433 }
2434
2435 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2436 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2437 tgsi_dump(tokens, 0);
2438 else
2439 nir_print_shader(llvm->draw->gs.geometry_shader->state.ir.nir, stderr);
2440 draw_gs_llvm_dump_variant_key(&variant->key);
2441 }
2442
2443 struct lp_build_tgsi_params params;
2444 memset(¶ms, 0, sizeof(params));
2445
2446 params.type = gs_type;
2447 params.mask = &mask;
2448 params.consts_ptr = consts_ptr;
2449 params.system_values = &system_values;
2450 params.context_type = variant->context_type;
2451 params.context_ptr = context_ptr;
2452 params.resources_type = variant->resources_type;
2453 params.resources_ptr = resources_ptr;
2454 params.sampler = sampler;
2455 params.info = &llvm->draw->gs.geometry_shader->info;
2456 params.gs_iface = (const struct lp_build_gs_iface *)&gs_iface;
2457 params.ssbo_ptr = ssbos_ptr;
2458 params.image = image;
2459 params.gs_vertex_streams = variant->shader->base.num_vertex_streams;
2460 params.aniso_filter_table = lp_jit_resources_aniso_filter_table(gallivm,
2461 variant->resources_type,
2462 resources_ptr);
2463
2464
2465 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2466 lp_build_tgsi_soa(variant->gallivm,
2467 tokens,
2468 ¶ms,
2469 outputs);
2470 else
2471 lp_build_nir_soa(variant->gallivm,
2472 llvm->draw->gs.geometry_shader->state.ir.nir,
2473 ¶ms,
2474 outputs);
2475
2476 lp_bld_llvm_sampler_soa_destroy(sampler);
2477 lp_bld_llvm_image_soa_destroy(image);
2478
2479 lp_build_mask_end(&mask);
2480
2481 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2482
2483 gallivm_verify_function(gallivm, variant_func);
2484 }
2485
2486
2487 struct draw_gs_llvm_variant *
draw_gs_llvm_create_variant(struct draw_llvm * llvm,unsigned num_outputs,const struct draw_gs_llvm_variant_key * key)2488 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2489 unsigned num_outputs,
2490 const struct draw_gs_llvm_variant_key *key)
2491 {
2492 struct draw_gs_llvm_variant *variant;
2493 struct llvm_geometry_shader *shader =
2494 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2495 char module_name[64];
2496 unsigned char ir_sha1_cache_key[20];
2497 struct lp_cached_code cached = { 0 };
2498 bool needs_caching = false;
2499
2500 variant = MALLOC(sizeof *variant +
2501 shader->variant_key_size -
2502 sizeof variant->key);
2503 if (!variant)
2504 return NULL;
2505
2506 variant->llvm = llvm;
2507 variant->shader = shader;
2508
2509 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2510 variant->shader->variants_cached);
2511
2512 memcpy(&variant->key, key, shader->variant_key_size);
2513
2514 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
2515 draw_get_ir_cache_key(shader->base.state.ir.nir,
2516 key,
2517 shader->variant_key_size,
2518 num_outputs,
2519 ir_sha1_cache_key);
2520
2521 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
2522 &cached,
2523 ir_sha1_cache_key);
2524 if (!cached.data_size)
2525 needs_caching = true;
2526 }
2527 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
2528
2529 create_gs_jit_types(variant);
2530
2531 variant->vertex_header_type = lp_build_create_jit_vertex_header_type(variant->gallivm, num_outputs);
2532 variant->vertex_header_ptr_type = LLVMPointerType(variant->vertex_header_type, 0);
2533
2534 draw_gs_llvm_generate(llvm, variant);
2535
2536 gallivm_compile_module(variant->gallivm);
2537
2538 variant->jit_func = (draw_gs_jit_func)
2539 gallivm_jit_function(variant->gallivm, variant->function);
2540
2541 if (needs_caching)
2542 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
2543 &cached,
2544 ir_sha1_cache_key);
2545 gallivm_free_ir(variant->gallivm);
2546
2547 variant->list_item_global.base = variant;
2548 variant->list_item_local.base = variant;
2549 /*variant->no = */shader->variants_created++;
2550 variant->list_item_global.base = variant;
2551
2552 return variant;
2553 }
2554
2555
2556 void
draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant * variant)2557 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2558 {
2559 struct draw_llvm *llvm = variant->llvm;
2560
2561 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2562 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2563 variant->shader->variants_cached, llvm->nr_gs_variants);
2564 }
2565
2566 gallivm_destroy(variant->gallivm);
2567
2568 list_del(&variant->list_item_local.list);
2569 variant->shader->variants_cached--;
2570 list_del(&variant->list_item_global.list);
2571 llvm->nr_gs_variants--;
2572 FREE(variant);
2573 }
2574
2575
2576 struct draw_gs_llvm_variant_key *
draw_gs_llvm_make_variant_key(struct draw_llvm * llvm,char * store)2577 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2578 {
2579 struct draw_gs_llvm_variant_key *key;
2580 struct lp_sampler_static_state *draw_sampler;
2581 struct lp_image_static_state *draw_image;
2582
2583 key = (struct draw_gs_llvm_variant_key *)store;
2584
2585 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2586
2587 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2588
2589 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color;
2590
2591 /* All variants of this shader will have the same value for
2592 * nr_samplers. Not yet trying to compact away holes in the
2593 * sampler array.
2594 */
2595 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2596 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2597 key->nr_sampler_views =
2598 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2599 } else {
2600 key->nr_sampler_views = key->nr_samplers;
2601 }
2602
2603 key->nr_images = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2604
2605 draw_sampler = key->samplers;
2606
2607 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2608
2609 for (unsigned i = 0 ; i < key->nr_samplers; i++) {
2610 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2611 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2612 }
2613 for (unsigned i = 0 ; i < key->nr_sampler_views; i++) {
2614 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2615 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2616 }
2617
2618 draw_image = draw_gs_llvm_variant_key_images(key);
2619 memset(draw_image, 0,
2620 key->nr_images * sizeof *draw_image);
2621 for (unsigned i = 0; i < key->nr_images; i++) {
2622 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2623 llvm->draw->images[PIPE_SHADER_GEOMETRY][i]);
2624 }
2625 return key;
2626 }
2627
2628
2629 void
draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key * key)2630 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2631 {
2632 struct lp_sampler_static_state *sampler = key->samplers;
2633 struct lp_image_static_state *image = draw_gs_llvm_variant_key_images(key);
2634
2635 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2636 for (unsigned i = 0 ; i < key->nr_sampler_views; i++) {
2637 debug_printf("sampler[%i].src_format = %s\n", i,
2638 util_format_name(sampler[i].texture_state.format));
2639 }
2640
2641 for (unsigned i = 0 ; i < key->nr_images; i++)
2642 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2643
2644 }
2645
2646
2647 static void
create_tcs_jit_types(struct draw_tcs_llvm_variant * var)2648 create_tcs_jit_types(struct draw_tcs_llvm_variant *var)
2649 {
2650 struct gallivm_state *gallivm = var->gallivm;
2651
2652 var->resources_type = lp_build_jit_resources_type(gallivm);
2653 var->resources_ptr_type = LLVMPointerType(var->resources_type, 0);
2654 var->input_array_type = create_tcs_jit_input_type(gallivm);
2655 var->output_array_type = create_tcs_jit_output_type(gallivm);
2656 }
2657
2658
2659 static LLVMTypeRef
get_tcs_resources_ptr_type(struct draw_tcs_llvm_variant * variant)2660 get_tcs_resources_ptr_type(struct draw_tcs_llvm_variant *variant)
2661 {
2662 if (!variant->resources_ptr_type)
2663 create_tcs_jit_types(variant);
2664 return variant->resources_ptr_type;
2665 }
2666
2667
2668 static LLVMValueRef
draw_tcs_llvm_emit_fetch_input(const struct lp_build_tcs_iface * tes_iface,struct lp_build_context * bld,bool is_vindex_indirect,LLVMValueRef vertex_index,bool is_aindex_indirect,LLVMValueRef attrib_index,bool is_sindex_indirect,LLVMValueRef swizzle_index)2669 draw_tcs_llvm_emit_fetch_input(const struct lp_build_tcs_iface *tes_iface,
2670 struct lp_build_context *bld,
2671 bool is_vindex_indirect,
2672 LLVMValueRef vertex_index,
2673 bool is_aindex_indirect,
2674 LLVMValueRef attrib_index,
2675 bool is_sindex_indirect,
2676 LLVMValueRef swizzle_index)
2677 {
2678 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
2679 struct gallivm_state *gallivm = bld->gallivm;
2680 LLVMBuilderRef builder = gallivm->builder;
2681 LLVMValueRef indices[3];
2682 LLVMValueRef res;
2683 struct lp_type type = bld->type;
2684 LLVMTypeRef input_type = create_tcs_jit_input_type_deref(gallivm);
2685 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
2686
2687 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
2688 res = bld->zero;
2689 for (int i = 0; i < type.length; ++i) {
2690 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2691 LLVMValueRef vert_chan_index = vertex_index;
2692 LLVMValueRef attr_chan_index = attrib_index;
2693 LLVMValueRef swiz_chan_index = swizzle_index;
2694 LLVMValueRef channel_vec;
2695
2696 if (is_vindex_indirect) {
2697 vert_chan_index = LLVMBuildExtractElement(builder,
2698 vertex_index, idx, "");
2699 }
2700 if (is_aindex_indirect) {
2701 attr_chan_index = LLVMBuildExtractElement(builder,
2702 attrib_index, idx, "");
2703 }
2704 if (is_sindex_indirect) {
2705 swiz_chan_index = LLVMBuildExtractElement(builder,
2706 swizzle_index, idx, "");
2707 }
2708
2709 indices[0] = vert_chan_index;
2710 indices[1] = attr_chan_index;
2711 indices[2] = swiz_chan_index;
2712
2713 channel_vec = LLVMBuildGEP2(builder, input_type, tcs->input, indices, 3, "");
2714 channel_vec = LLVMBuildLoad2(builder, float_type, channel_vec, "");
2715 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
2716 }
2717 } else {
2718 indices[0] = vertex_index;
2719 indices[1] = attrib_index;
2720 indices[2] = swizzle_index;
2721 res = LLVMBuildGEP2(builder, input_type, tcs->input, indices, 3, "");
2722 res = LLVMBuildLoad2(builder, float_type, res, "");
2723 res = lp_build_broadcast_scalar(bld, res);
2724 }
2725 return res;
2726 }
2727
2728
2729 static LLVMValueRef
draw_tcs_llvm_emit_fetch_output(const struct lp_build_tcs_iface * tes_iface,struct lp_build_context * bld,bool is_vindex_indirect,LLVMValueRef vertex_index,bool is_aindex_indirect,LLVMValueRef attrib_index,bool is_sindex_indirect,LLVMValueRef swizzle_index,uint32_t name)2730 draw_tcs_llvm_emit_fetch_output(const struct lp_build_tcs_iface *tes_iface,
2731 struct lp_build_context *bld,
2732 bool is_vindex_indirect,
2733 LLVMValueRef vertex_index,
2734 bool is_aindex_indirect,
2735 LLVMValueRef attrib_index,
2736 bool is_sindex_indirect,
2737 LLVMValueRef swizzle_index,
2738 uint32_t name)
2739 {
2740 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
2741 struct gallivm_state *gallivm = bld->gallivm;
2742 LLVMBuilderRef builder = gallivm->builder;
2743 LLVMValueRef indices[3];
2744 LLVMValueRef res;
2745 struct lp_type type = bld->type;
2746 LLVMTypeRef output_type = create_tcs_jit_output_type_deref(gallivm);
2747 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
2748
2749 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
2750 res = bld->zero;
2751 for (int i = 0; i < type.length; ++i) {
2752 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2753 LLVMValueRef vert_chan_index = vertex_index;
2754 LLVMValueRef attr_chan_index = attrib_index;
2755 LLVMValueRef swiz_chan_index = swizzle_index;
2756 LLVMValueRef channel_vec;
2757
2758 if (is_vindex_indirect) {
2759 vert_chan_index = LLVMBuildExtractElement(builder,
2760 vertex_index, idx, "");
2761 }
2762 if (is_aindex_indirect) {
2763 attr_chan_index = LLVMBuildExtractElement(builder,
2764 attrib_index, idx, "");
2765 }
2766 if (is_sindex_indirect) {
2767 swiz_chan_index = LLVMBuildExtractElement(builder,
2768 swizzle_index, idx, "");
2769 }
2770
2771 indices[0] = vert_chan_index;
2772 indices[1] = attr_chan_index;
2773 indices[2] = swiz_chan_index;
2774
2775 channel_vec = LLVMBuildGEP2(builder, output_type, tcs->output, indices, 3, "");
2776 channel_vec = LLVMBuildLoad2(builder, float_type, channel_vec, "");
2777
2778 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
2779 }
2780 } else {
2781 indices[0] = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
2782 indices[1] = attrib_index;
2783 indices[2] = swizzle_index;
2784
2785 res = LLVMBuildGEP2(builder, output_type, tcs->output, indices, 3, "");
2786 res = LLVMBuildLoad2(builder, float_type, res, "");
2787 res = lp_build_broadcast_scalar(bld, res);
2788 }
2789 return res;
2790 }
2791
2792
2793 static void
draw_tcs_llvm_emit_store_output(const struct lp_build_tcs_iface * tes_iface,struct lp_build_context * bld,unsigned name,bool is_vindex_indirect,LLVMValueRef vertex_index,bool is_aindex_indirect,LLVMValueRef attrib_index,bool is_sindex_indirect,LLVMValueRef swizzle_index,LLVMValueRef value,LLVMValueRef mask_vec)2794 draw_tcs_llvm_emit_store_output(const struct lp_build_tcs_iface *tes_iface,
2795 struct lp_build_context *bld,
2796 unsigned name,
2797 bool is_vindex_indirect,
2798 LLVMValueRef vertex_index,
2799 bool is_aindex_indirect,
2800 LLVMValueRef attrib_index,
2801 bool is_sindex_indirect,
2802 LLVMValueRef swizzle_index,
2803 LLVMValueRef value,
2804 LLVMValueRef mask_vec)
2805 {
2806 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
2807 struct gallivm_state *gallivm = bld->gallivm;
2808 LLVMBuilderRef builder = gallivm->builder;
2809 LLVMValueRef indices[3];
2810 LLVMValueRef res;
2811 struct lp_type type = bld->type;
2812 LLVMTypeRef output_type = create_tcs_jit_output_type_deref(gallivm);
2813
2814 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
2815 for (int i = 0; i < type.length; ++i) {
2816 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2817 LLVMValueRef vert_chan_index = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
2818 LLVMValueRef attr_chan_index = attrib_index;
2819 LLVMValueRef swiz_chan_index = swizzle_index;
2820 LLVMValueRef channel_vec;
2821
2822 if (is_vindex_indirect) {
2823 vert_chan_index = LLVMBuildExtractElement(builder,
2824 vertex_index, idx, "");
2825 }
2826 if (is_aindex_indirect) {
2827 attr_chan_index = LLVMBuildExtractElement(builder,
2828 attrib_index, idx, "");
2829 }
2830
2831 if (is_sindex_indirect) {
2832 swiz_chan_index = LLVMBuildExtractElement(builder,
2833 swizzle_index, idx, "");
2834 }
2835
2836 indices[0] = vert_chan_index;
2837 indices[1] = attr_chan_index;
2838 indices[2] = swiz_chan_index;
2839
2840 channel_vec = LLVMBuildGEP2(builder, output_type, tcs->output, indices, 3, "");
2841
2842 res = LLVMBuildExtractElement(builder, value, idx, "");
2843
2844 struct lp_build_if_state ifthen;
2845 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
2846 cond = LLVMBuildExtractElement(gallivm->builder, cond, idx, "");
2847 lp_build_if(&ifthen, gallivm, cond);
2848 LLVMBuildStore(builder, res, channel_vec);
2849 lp_build_endif(&ifthen);
2850 }
2851 } else {
2852 indices[0] = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
2853 indices[1] = attrib_index;
2854 indices[2] = swizzle_index;
2855
2856 res = LLVMBuildGEP2(builder, output_type, tcs->output, indices, 3, "");
2857 for (unsigned i = 0; i < type.length; ++i) {
2858 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2859 LLVMValueRef val = LLVMBuildExtractElement(builder, value, idx, "");
2860
2861 struct lp_build_if_state ifthen;
2862 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
2863 cond = LLVMBuildExtractElement(gallivm->builder, cond, idx, "");
2864 lp_build_if(&ifthen, gallivm, cond);
2865 LLVMBuildStore(builder, val, res);
2866 lp_build_endif(&ifthen);
2867 }
2868 }
2869 }
2870
2871
2872 static LLVMValueRef
generate_tcs_mask_value(struct draw_tcs_llvm_variant * variant,struct lp_type tcs_type,LLVMValueRef limit,LLVMValueRef loop_counter)2873 generate_tcs_mask_value(struct draw_tcs_llvm_variant *variant,
2874 struct lp_type tcs_type, LLVMValueRef limit, LLVMValueRef loop_counter)
2875 {
2876 struct gallivm_state *gallivm = variant->gallivm;
2877 LLVMBuilderRef builder = gallivm->builder;
2878 struct lp_type mask_type = lp_int_type(tcs_type);
2879 LLVMValueRef num_vecs;
2880 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2881
2882 num_vecs = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type), limit);
2883 for (unsigned i = 0; i < tcs_type.length; i++) {
2884 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2885 mask_val = LLVMBuildInsertElement(builder, mask_val, LLVMBuildAdd(builder, loop_counter, idx, ""), idx, "");
2886 }
2887 mask_val = lp_build_compare(gallivm, mask_type,
2888 PIPE_FUNC_GREATER, num_vecs, mask_val);
2889
2890 return mask_val;
2891 }
2892
2893
2894 static void
draw_tcs_llvm_generate(struct draw_llvm * llvm,struct draw_tcs_llvm_variant * variant)2895 draw_tcs_llvm_generate(struct draw_llvm *llvm,
2896 struct draw_tcs_llvm_variant *variant)
2897 {
2898 struct gallivm_state *gallivm = variant->gallivm;
2899 LLVMContextRef context = gallivm->context;
2900 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2901 LLVMTypeRef arg_types[7];
2902 LLVMTypeRef func_type, coro_func_type;
2903 LLVMValueRef variant_func, variant_coro;
2904 LLVMValueRef resources_ptr;
2905 LLVMValueRef view_index;
2906 LLVMValueRef input_array, output_array, prim_id, patch_vertices_in;
2907 LLVMValueRef mask_val;
2908 LLVMBasicBlockRef block;
2909 LLVMBuilderRef builder;
2910 struct lp_build_context bld, bldvec;
2911 struct lp_build_sampler_soa *sampler = 0;
2912 struct lp_build_image_soa *image = NULL;
2913 struct lp_bld_tgsi_system_values system_values;
2914 char func_name[64], func_name_coro[64];
2915 struct draw_tcs_llvm_iface tcs_iface;
2916 struct lp_build_mask_context mask;
2917 LLVMValueRef consts_ptr;
2918 LLVMValueRef ssbos_ptr;
2919 struct lp_type tcs_type;
2920 unsigned vector_length = variant->shader->base.vector_length;
2921
2922 memset(&system_values, 0, sizeof(system_values));
2923
2924 snprintf(func_name, sizeof(func_name), "draw_llvm_tcs_variant");
2925
2926 snprintf(func_name_coro, sizeof(func_name_coro), "draw_llvm_tcs_coro_variant");
2927
2928 arg_types[0] = get_tcs_resources_ptr_type(variant); /* context */
2929 arg_types[1] = variant->input_array_type; /* input */
2930 arg_types[2] = variant->output_array_type;
2931 arg_types[3] = int32_type;
2932 arg_types[4] = int32_type;
2933 arg_types[5] = int32_type;
2934 arg_types[6] = int32_type; /* coroutine only */
2935
2936 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types) - 1, 0);
2937
2938 coro_func_type = LLVMFunctionType(LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), arg_types, ARRAY_SIZE(arg_types), 0);
2939
2940 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2941
2942 variant_coro = LLVMAddFunction(gallivm->module, func_name_coro, coro_func_type);
2943
2944 variant->function = variant_func;
2945 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2946
2947 LLVMSetFunctionCallConv(variant_coro, LLVMCCallConv);
2948
2949 lp_build_coro_add_presplit(variant_coro);
2950
2951 for (unsigned i = 0; i < ARRAY_SIZE(arg_types); ++i) {
2952 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
2953 lp_add_function_attr(variant_coro, i + 1, LP_FUNC_ATTR_NOALIAS);
2954 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2955 }
2956 }
2957
2958 if (gallivm->cache && gallivm->cache->data_size)
2959 return;
2960 resources_ptr = LLVMGetParam(variant_func, 0);
2961 input_array = LLVMGetParam(variant_func, 1);
2962 output_array = LLVMGetParam(variant_func, 2);
2963 prim_id = LLVMGetParam(variant_func, 3);
2964 patch_vertices_in = LLVMGetParam(variant_func, 4);
2965 view_index = LLVMGetParam(variant_func, 5);
2966
2967 lp_build_name(resources_ptr, "resources");
2968 lp_build_name(input_array, "input");
2969 lp_build_name(output_array, "output");
2970 lp_build_name(prim_id, "prim_id");
2971 lp_build_name(patch_vertices_in, "patch_vertices_in");
2972 lp_build_name(view_index, "view_index");
2973
2974 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2975 builder = gallivm->builder;
2976 LLVMPositionBuilderAtEnd(builder, block);
2977
2978 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2979
2980 memset(&tcs_type, 0, sizeof tcs_type);
2981 tcs_type.floating = true; /* floating point values */
2982 tcs_type.sign = true; /* values are signed */
2983 tcs_type.norm = false; /* values are not limited to [0,1] or [-1,1] */
2984 tcs_type.width = 32; /* 32-bit float */
2985 tcs_type.length = vector_length;
2986
2987 lp_build_context_init(&bldvec, variant->gallivm, lp_int_type(tcs_type));
2988
2989 LLVMValueRef count = lp_build_const_int32(gallivm, variant->shader->base.vertices_out);
2990 LLVMValueRef step = lp_build_const_int32(gallivm, vector_length);
2991
2992 struct lp_build_loop_state loop_state[2];
2993 LLVMValueRef num_inner_loop;
2994 unsigned count_align = util_align_npot(variant->shader->base.vertices_out, tcs_type.length);
2995 num_inner_loop = lp_build_const_int32(gallivm, count_align / tcs_type.length);
2996 LLVMTypeRef hdl_ptr_type = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
2997 LLVMValueRef coro_hdls = LLVMBuildArrayAlloca(gallivm->builder, hdl_ptr_type, num_inner_loop, "coro_hdls");
2998 unsigned end_coroutine = INT_MAX;
2999 lp_build_loop_begin(&loop_state[1], gallivm,
3000 lp_build_const_int32(gallivm, 0)); /* coroutine reentry loop */
3001 lp_build_loop_begin(&loop_state[0], gallivm,
3002 lp_build_const_int32(gallivm, 0)); /* inner loop */
3003 {
3004 LLVMValueRef args[7];
3005 args[0] = resources_ptr;
3006 args[1] = input_array;
3007 args[2] = output_array;
3008 args[3] = prim_id;
3009 args[4] = patch_vertices_in;
3010 args[5] = view_index;
3011 args[6] = loop_state[0].counter;
3012 LLVMValueRef coro_entry = LLVMBuildGEP2(builder, hdl_ptr_type, coro_hdls, &loop_state[0].counter, 1, "");
3013 LLVMValueRef coro_hdl = LLVMBuildLoad2(builder, hdl_ptr_type, coro_entry, "coro_hdl");
3014
3015 struct lp_build_if_state ifstate;
3016 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntEQ, loop_state[1].counter,
3017 lp_build_const_int32(gallivm, 0), "");
3018 /* first time here - call the coroutine function entry point */
3019 lp_build_if(&ifstate, gallivm, cmp);
3020 LLVMValueRef coro_ret = LLVMBuildCall2(builder, coro_func_type, variant_coro, args, 7, "");
3021 LLVMBuildStore(builder, coro_ret, coro_entry);
3022 lp_build_else(&ifstate);
3023 /* subsequent calls for this invocation - check if done. */
3024 LLVMValueRef coro_done = lp_build_coro_done(gallivm, coro_hdl);
3025 struct lp_build_if_state ifstate2;
3026 lp_build_if(&ifstate2, gallivm, coro_done);
3027 /* if done destroy and force loop exit */
3028 lp_build_coro_destroy(gallivm, coro_hdl);
3029 lp_build_loop_force_set_counter(&loop_state[1], lp_build_const_int32(gallivm, end_coroutine - 1));
3030 lp_build_else(&ifstate2);
3031 /* otherwise resume the coroutine */
3032 lp_build_coro_resume(gallivm, coro_hdl);
3033 lp_build_endif(&ifstate2);
3034 lp_build_endif(&ifstate);
3035 lp_build_loop_force_reload_counter(&loop_state[1]);
3036 }
3037 lp_build_loop_end_cond(&loop_state[0],
3038 num_inner_loop,
3039 NULL, LLVMIntUGE);
3040 lp_build_loop_end_cond(&loop_state[1],
3041 lp_build_const_int32(gallivm, end_coroutine),
3042 NULL, LLVMIntEQ);
3043 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
3044
3045 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_coro, "entry");
3046 LLVMPositionBuilderAtEnd(builder, block);
3047
3048 resources_ptr = LLVMGetParam(variant_coro, 0);
3049 input_array = LLVMGetParam(variant_coro, 1);
3050 output_array = LLVMGetParam(variant_coro, 2);
3051 prim_id = LLVMGetParam(variant_coro, 3);
3052 patch_vertices_in = LLVMGetParam(variant_coro, 4);
3053 view_index = LLVMGetParam(variant_coro, 5);
3054
3055 consts_ptr = lp_jit_resources_constants(gallivm, variant->resources_type, resources_ptr);
3056
3057 ssbos_ptr = lp_jit_resources_ssbos(gallivm, variant->resources_type, resources_ptr);
3058 sampler = lp_bld_llvm_sampler_soa_create(variant->key.samplers,
3059 MAX2(variant->key.nr_samplers,
3060 variant->key.nr_sampler_views));
3061 image = lp_bld_llvm_image_soa_create(draw_tcs_llvm_variant_key_images(&variant->key),
3062 variant->key.nr_images);
3063
3064 LLVMValueRef counter = LLVMGetParam(variant_coro, 6);
3065 LLVMValueRef invocvec = LLVMGetUndef(LLVMVectorType(int32_type, vector_length));
3066 for (unsigned i = 0; i < vector_length; i++) {
3067 LLVMValueRef loop_iter = lp_build_const_int32(gallivm, i);
3068 LLVMValueRef idx = LLVMBuildAdd(builder, LLVMBuildMul(builder, counter, step, ""), loop_iter, "");
3069 invocvec = LLVMBuildInsertElement(builder, invocvec, idx, loop_iter, "");
3070 }
3071
3072 system_values.invocation_id = invocvec;
3073 system_values.prim_id = lp_build_broadcast_scalar(&bldvec, prim_id);
3074 system_values.view_index = view_index;
3075 system_values.vertices_in = lp_build_broadcast_scalar(&bldvec, patch_vertices_in);
3076 tcs_iface.input = input_array;
3077 tcs_iface.output = output_array;
3078 tcs_iface.base.emit_fetch_input = draw_tcs_llvm_emit_fetch_input;
3079 tcs_iface.base.emit_fetch_output = draw_tcs_llvm_emit_fetch_output;
3080 tcs_iface.base.emit_store_output = draw_tcs_llvm_emit_store_output;
3081
3082
3083 {
3084 LLVMValueRef coro_id = lp_build_coro_id(gallivm);
3085 LLVMValueRef coro_hdl = lp_build_coro_begin_alloc_mem(gallivm, coro_id);
3086
3087 mask_val = generate_tcs_mask_value(variant, tcs_type, count, LLVMBuildMul(builder, counter, step, ""));
3088 lp_build_mask_begin(&mask, gallivm, tcs_type, mask_val);
3089
3090 struct lp_build_coro_suspend_info coro_info;
3091
3092 LLVMBasicBlockRef sus_block = LLVMAppendBasicBlockInContext(gallivm->context, variant_coro, "suspend");
3093 LLVMBasicBlockRef clean_block = LLVMAppendBasicBlockInContext(gallivm->context, variant_coro, "cleanup");
3094
3095 coro_info.suspend = sus_block;
3096 coro_info.cleanup = clean_block;
3097
3098 struct lp_build_tgsi_params params;
3099 memset(¶ms, 0, sizeof(params));
3100
3101 params.type = tcs_type;
3102 params.mask = &mask;
3103 params.consts_ptr = consts_ptr;
3104 params.system_values = &system_values;
3105 params.resources_type = variant->resources_type;
3106 params.resources_ptr = resources_ptr;
3107 params.sampler = sampler;
3108 params.info = &llvm->draw->tcs.tess_ctrl_shader->info;
3109 params.ssbo_ptr = ssbos_ptr;
3110 params.image = image;
3111 params.coro = &coro_info;
3112 params.tcs_iface = &tcs_iface.base;
3113 params.aniso_filter_table = lp_jit_resources_aniso_filter_table(gallivm,
3114 variant->resources_type,
3115 resources_ptr);
3116
3117 lp_build_nir_soa(variant->gallivm,
3118 llvm->draw->tcs.tess_ctrl_shader->state.ir.nir,
3119 ¶ms, NULL);
3120
3121 lp_build_mask_end(&mask);
3122
3123 lp_build_coro_suspend_switch(gallivm, &coro_info, NULL, true);
3124 LLVMPositionBuilderAtEnd(builder, clean_block);
3125
3126 lp_build_coro_free_mem(gallivm, coro_id, coro_hdl);
3127
3128 LLVMBuildBr(builder, sus_block);
3129 LLVMPositionBuilderAtEnd(builder, sus_block);
3130
3131 lp_build_coro_end(gallivm, coro_hdl);
3132 LLVMBuildRet(builder, coro_hdl);
3133 }
3134
3135 lp_bld_llvm_sampler_soa_destroy(sampler);
3136 lp_bld_llvm_image_soa_destroy(image);
3137 gallivm_verify_function(gallivm, variant_func);
3138 gallivm_verify_function(gallivm, variant_coro);
3139 }
3140
3141
3142 struct draw_tcs_llvm_variant *
draw_tcs_llvm_create_variant(struct draw_llvm * llvm,unsigned num_outputs,const struct draw_tcs_llvm_variant_key * key)3143 draw_tcs_llvm_create_variant(struct draw_llvm *llvm,
3144 unsigned num_outputs,
3145 const struct draw_tcs_llvm_variant_key *key)
3146 {
3147 struct draw_tcs_llvm_variant *variant;
3148 struct llvm_tess_ctrl_shader *shader = llvm_tess_ctrl_shader(llvm->draw->tcs.tess_ctrl_shader);
3149 char module_name[64];
3150 unsigned char ir_sha1_cache_key[20];
3151 struct lp_cached_code cached = { 0 };
3152 bool needs_caching = false;
3153
3154 variant = MALLOC(sizeof *variant +
3155 shader->variant_key_size - sizeof variant->key);
3156 if (!variant)
3157 return NULL;
3158
3159 variant->llvm = llvm;
3160 variant->shader = shader;
3161
3162 snprintf(module_name, sizeof(module_name), "draw_llvm_tcs_variant%u",
3163 variant->shader->variants_cached);
3164
3165 memcpy(&variant->key, key, shader->variant_key_size);
3166
3167 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
3168 draw_get_ir_cache_key(shader->base.state.ir.nir,
3169 key,
3170 shader->variant_key_size,
3171 num_outputs,
3172 ir_sha1_cache_key);
3173
3174 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
3175 &cached,
3176 ir_sha1_cache_key);
3177 if (!cached.data_size)
3178 needs_caching = true;
3179 }
3180
3181 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
3182
3183 create_tcs_jit_types(variant);
3184
3185 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
3186 nir_print_shader(llvm->draw->tcs.tess_ctrl_shader->state.ir.nir, stderr);
3187 draw_tcs_llvm_dump_variant_key(&variant->key);
3188 }
3189
3190 draw_tcs_llvm_generate(llvm, variant);
3191
3192 gallivm_compile_module(variant->gallivm);
3193
3194 variant->jit_func = (draw_tcs_jit_func)
3195 gallivm_jit_function(variant->gallivm, variant->function);
3196
3197 if (needs_caching)
3198 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
3199 &cached,
3200 ir_sha1_cache_key);
3201 gallivm_free_ir(variant->gallivm);
3202
3203 variant->list_item_global.base = variant;
3204 variant->list_item_local.base = variant;
3205 /*variant->no = */shader->variants_created++;
3206 variant->list_item_global.base = variant;
3207
3208 return variant;
3209 }
3210
3211
3212 void
draw_tcs_llvm_destroy_variant(struct draw_tcs_llvm_variant * variant)3213 draw_tcs_llvm_destroy_variant(struct draw_tcs_llvm_variant *variant)
3214 {
3215 struct draw_llvm *llvm = variant->llvm;
3216
3217 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
3218 debug_printf("Deleting TCS variant: %u tcs variants,\t%u total variants\n",
3219 variant->shader->variants_cached, llvm->nr_tcs_variants);
3220 }
3221
3222 gallivm_destroy(variant->gallivm);
3223
3224 list_del(&variant->list_item_local.list);
3225 variant->shader->variants_cached--;
3226 list_del(&variant->list_item_global.list);
3227 llvm->nr_tcs_variants--;
3228 FREE(variant);
3229 }
3230
3231
3232 struct draw_tcs_llvm_variant_key *
draw_tcs_llvm_make_variant_key(struct draw_llvm * llvm,char * store)3233 draw_tcs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
3234 {
3235 unsigned i;
3236 struct draw_tcs_llvm_variant_key *key;
3237 struct lp_sampler_static_state *draw_sampler;
3238 struct lp_image_static_state *draw_image;
3239
3240 key = (struct draw_tcs_llvm_variant_key *)store;
3241
3242 memset(key, 0, offsetof(struct draw_tcs_llvm_variant_key, samplers[0]));
3243
3244 /* All variants of this shader will have the same value for
3245 * nr_samplers. Not yet trying to compact away holes in the
3246 * sampler array.
3247 */
3248 key->nr_samplers = llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
3249 if (llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3250 key->nr_sampler_views =
3251 llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3252 } else {
3253 key->nr_sampler_views = key->nr_samplers;
3254 }
3255
3256 key->nr_images = llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
3257
3258 draw_sampler = key->samplers;
3259
3260 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
3261
3262 for (i = 0 ; i < key->nr_samplers; i++) {
3263 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
3264 llvm->draw->samplers[PIPE_SHADER_TESS_CTRL][i]);
3265 }
3266 for (i = 0 ; i < key->nr_sampler_views; i++) {
3267 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
3268 llvm->draw->sampler_views[PIPE_SHADER_TESS_CTRL][i]);
3269 }
3270
3271 draw_image = draw_tcs_llvm_variant_key_images(key);
3272 memset(draw_image, 0,
3273 key->nr_images * sizeof *draw_image);
3274 for (i = 0; i < key->nr_images; i++) {
3275 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
3276 llvm->draw->images[PIPE_SHADER_TESS_CTRL][i]);
3277 }
3278 return key;
3279 }
3280
3281
3282 void
draw_tcs_llvm_dump_variant_key(struct draw_tcs_llvm_variant_key * key)3283 draw_tcs_llvm_dump_variant_key(struct draw_tcs_llvm_variant_key *key)
3284 {
3285 struct lp_sampler_static_state *sampler = key->samplers;
3286 struct lp_image_static_state *image = draw_tcs_llvm_variant_key_images(key);
3287 for (unsigned i = 0 ; i < key->nr_sampler_views; i++) {
3288 debug_printf("sampler[%i].src_format = %s\n", i,
3289 util_format_name(sampler[i].texture_state.format));
3290 }
3291
3292 for (unsigned i = 0 ; i < key->nr_images; i++)
3293 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
3294 }
3295
3296
3297 static void
create_tes_jit_types(struct draw_tes_llvm_variant * var)3298 create_tes_jit_types(struct draw_tes_llvm_variant *var)
3299 {
3300 struct gallivm_state *gallivm = var->gallivm;
3301
3302 var->resources_type = lp_build_jit_resources_type(gallivm);
3303 var->resources_ptr_type = LLVMPointerType(var->resources_type, 0);
3304 var->input_array_deref_type = create_tes_jit_input_deref_type(gallivm);
3305 var->input_array_type = LLVMPointerType(var->input_array_deref_type, 0); /* num vertices per prim */
3306 }
3307
3308
3309 static LLVMTypeRef
get_tes_resources_ptr_type(struct draw_tes_llvm_variant * variant)3310 get_tes_resources_ptr_type(struct draw_tes_llvm_variant *variant)
3311 {
3312 if (!variant->resources_ptr_type)
3313 create_tes_jit_types(variant);
3314 return variant->resources_ptr_type;
3315 }
3316
3317
3318 static LLVMValueRef
generate_tes_mask_value(struct draw_tes_llvm_variant * variant,struct lp_type tes_type,LLVMValueRef limit,LLVMValueRef loop_counter)3319 generate_tes_mask_value(struct draw_tes_llvm_variant *variant,
3320 struct lp_type tes_type, LLVMValueRef limit, LLVMValueRef loop_counter)
3321 {
3322 struct gallivm_state *gallivm = variant->gallivm;
3323 LLVMBuilderRef builder = gallivm->builder;
3324 struct lp_type mask_type = lp_int_type(tes_type);
3325 LLVMValueRef num_prims;
3326 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
3327 unsigned i;
3328
3329 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type), limit);
3330 for (i = 0; i < tes_type.length; i++) {
3331 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3332 mask_val = LLVMBuildInsertElement(builder, mask_val, LLVMBuildAdd(builder, loop_counter, idx, ""), idx, "");
3333 }
3334 mask_val = lp_build_compare(gallivm, mask_type,
3335 PIPE_FUNC_GREATER, num_prims, mask_val);
3336
3337 return mask_val;
3338 }
3339
3340
3341 static LLVMValueRef
draw_tes_llvm_fetch_vertex_input(const struct lp_build_tes_iface * tes_iface,struct lp_build_context * bld,bool is_vindex_indirect,LLVMValueRef vertex_index,bool is_aindex_indirect,LLVMValueRef attrib_index,bool is_sindex_indirect,LLVMValueRef swizzle_index)3342 draw_tes_llvm_fetch_vertex_input(const struct lp_build_tes_iface *tes_iface,
3343 struct lp_build_context *bld,
3344 bool is_vindex_indirect,
3345 LLVMValueRef vertex_index,
3346 bool is_aindex_indirect,
3347 LLVMValueRef attrib_index,
3348 bool is_sindex_indirect,
3349 LLVMValueRef swizzle_index)
3350 {
3351 const struct draw_tes_llvm_iface *tes = draw_tes_llvm_iface(tes_iface);
3352 struct gallivm_state *gallivm = bld->gallivm;
3353 LLVMBuilderRef builder = gallivm->builder;
3354 LLVMValueRef indices[3];
3355 LLVMValueRef res;
3356 struct lp_type type = bld->type;
3357
3358 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
3359 res = bld->zero;
3360
3361 for (int i = 0; i < type.length; ++i) {
3362 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3363 LLVMValueRef vert_chan_index = vertex_index;
3364 LLVMValueRef attr_chan_index = attrib_index;
3365 LLVMValueRef swiz_chan_index = swizzle_index;
3366 LLVMValueRef channel_vec;
3367
3368 if (is_vindex_indirect) {
3369 vert_chan_index = LLVMBuildExtractElement(builder,
3370 vertex_index, idx, "");
3371 }
3372 if (is_aindex_indirect) {
3373 attr_chan_index = LLVMBuildExtractElement(builder,
3374 attrib_index, idx, "");
3375 }
3376 if (is_sindex_indirect) {
3377 swiz_chan_index = LLVMBuildExtractElement(builder,
3378 swizzle_index, idx, "");
3379 }
3380
3381 indices[0] = vert_chan_index;
3382 indices[1] = attr_chan_index;
3383 indices[2] = swiz_chan_index;
3384
3385 channel_vec = LLVMBuildGEP2(builder, tes->variant->input_array_deref_type, tes->input, indices, 3, "");
3386 channel_vec = LLVMBuildLoad2(builder, LLVMFloatTypeInContext(gallivm->context), channel_vec, "");
3387
3388 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3389 }
3390 } else {
3391 indices[0] = vertex_index;
3392 indices[1] = attrib_index;
3393 indices[2] = swizzle_index;
3394
3395 res = LLVMBuildGEP2(builder, tes->variant->input_array_deref_type, tes->input, indices, 3, "");
3396 res = LLVMBuildLoad2(builder, LLVMFloatTypeInContext(gallivm->context), res, "");
3397 res = lp_build_broadcast_scalar(bld, res);
3398 }
3399 return res;
3400 }
3401
3402
3403 static LLVMValueRef
draw_tes_llvm_fetch_patch_input(const struct lp_build_tes_iface * tes_iface,struct lp_build_context * bld,bool is_aindex_indirect,LLVMValueRef attrib_index,LLVMValueRef swizzle_index)3404 draw_tes_llvm_fetch_patch_input(const struct lp_build_tes_iface *tes_iface,
3405 struct lp_build_context *bld,
3406 bool is_aindex_indirect,
3407 LLVMValueRef attrib_index,
3408 LLVMValueRef swizzle_index)
3409 {
3410 const struct draw_tes_llvm_iface *tes = draw_tes_llvm_iface(tes_iface);
3411 struct gallivm_state *gallivm = bld->gallivm;
3412 LLVMBuilderRef builder = gallivm->builder;
3413 LLVMValueRef indices[3];
3414 LLVMValueRef res;
3415 struct lp_type type = bld->type;
3416
3417 if (is_aindex_indirect) {
3418 res = bld->zero;
3419
3420 for (int i = 0; i < type.length; ++i) {
3421 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3422 LLVMValueRef attr_chan_index = attrib_index;
3423 LLVMValueRef channel_vec;
3424
3425 if (is_aindex_indirect) {
3426 attr_chan_index = LLVMBuildExtractElement(builder,
3427 attrib_index, idx, "");
3428 }
3429
3430 indices[0] = lp_build_const_int32(gallivm, 0);
3431 indices[1] = attr_chan_index;
3432 indices[2] = swizzle_index;
3433
3434 channel_vec = LLVMBuildGEP2(builder, tes->variant->input_array_deref_type, tes->input, indices, 3, "");
3435 channel_vec = LLVMBuildLoad2(builder, LLVMFloatTypeInContext(gallivm->context), channel_vec, "");
3436
3437 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3438 }
3439 } else {
3440 indices[0] = lp_build_const_int32(gallivm, 0);
3441 indices[1] = attrib_index;
3442 indices[2] = swizzle_index;
3443
3444 res = LLVMBuildGEP2(builder, tes->variant->input_array_deref_type, tes->input, indices, 3, "");
3445 res = LLVMBuildLoad2(builder, LLVMFloatTypeInContext(gallivm->context), res, "");
3446 res = lp_build_broadcast_scalar(bld, res);
3447 }
3448 return res;
3449 }
3450
3451
3452 static void
draw_tes_llvm_generate(struct draw_llvm * llvm,struct draw_tes_llvm_variant * variant)3453 draw_tes_llvm_generate(struct draw_llvm *llvm,
3454 struct draw_tes_llvm_variant *variant)
3455 {
3456 struct gallivm_state *gallivm = variant->gallivm;
3457 LLVMContextRef context = gallivm->context;
3458 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
3459 LLVMTypeRef flt_type = LLVMFloatTypeInContext(context);
3460 LLVMTypeRef arg_types[11];
3461 LLVMTypeRef func_type;
3462 LLVMValueRef variant_func;
3463 LLVMValueRef resources_ptr;
3464 LLVMValueRef tess_coord[2], io_ptr, input_array, num_tess_coord;
3465 LLVMValueRef view_index;
3466 LLVMValueRef tess_inner, tess_outer, prim_id, patch_vertices_in;
3467 LLVMBasicBlockRef block;
3468 LLVMBuilderRef builder;
3469 LLVMValueRef mask_val;
3470 struct lp_build_context bld, bldvec;
3471 struct lp_build_sampler_soa *sampler = 0;
3472 struct lp_build_image_soa *image = NULL;
3473 struct lp_bld_tgsi_system_values system_values;
3474 char func_name[64];
3475 unsigned i;
3476 struct draw_tes_llvm_iface tes_iface;
3477 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
3478 struct lp_build_mask_context mask;
3479 LLVMValueRef consts_ptr;
3480 LLVMValueRef ssbos_ptr;
3481 LLVMValueRef step;
3482 struct lp_type tes_type;
3483 unsigned vector_length = variant->shader->base.vector_length;
3484 int primid_slot = -1;
3485
3486 memset(&system_values, 0, sizeof(system_values));
3487 memset(&outputs, 0, sizeof(outputs));
3488
3489 snprintf(func_name, sizeof(func_name), "draw_llvm_tes_variant");
3490
3491 LLVMTypeRef tess_outer_deref_type = LLVMArrayType(flt_type, 4);
3492 LLVMTypeRef tess_inner_deref_type = LLVMArrayType(flt_type, 2);
3493
3494 arg_types[0] = get_tes_resources_ptr_type(variant); /* context */
3495 arg_types[1] = variant->input_array_type; /* input */
3496 arg_types[2] = variant->vertex_header_ptr_type;
3497 arg_types[3] = int32_type;
3498 arg_types[4] = int32_type;
3499 arg_types[5] = LLVMPointerType(flt_type, 0);
3500 arg_types[6] = LLVMPointerType(flt_type, 0);
3501 arg_types[7] = LLVMPointerType(tess_outer_deref_type, 0);
3502 arg_types[8] = LLVMPointerType(tess_inner_deref_type, 0);
3503 arg_types[9] = int32_type;
3504 arg_types[10] = int32_type;
3505
3506 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
3507 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
3508
3509 variant->function = variant_func;
3510 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
3511
3512 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
3513 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
3514 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
3515
3516 if (gallivm->cache && gallivm->cache->data_size)
3517 return;
3518 resources_ptr = LLVMGetParam(variant_func, 0);
3519 input_array = LLVMGetParam(variant_func, 1);
3520 io_ptr = LLVMGetParam(variant_func, 2);
3521 prim_id = LLVMGetParam(variant_func, 3);
3522 num_tess_coord = LLVMGetParam(variant_func, 4);
3523 tess_coord[0] = LLVMGetParam(variant_func, 5);
3524 tess_coord[1] = LLVMGetParam(variant_func, 6);
3525 tess_outer = LLVMGetParam(variant_func, 7);
3526 tess_inner = LLVMGetParam(variant_func, 8);
3527 patch_vertices_in = LLVMGetParam(variant_func, 9);
3528 view_index = LLVMGetParam(variant_func, 10);
3529
3530 lp_build_name(resources_ptr, "resources");
3531 lp_build_name(input_array, "input");
3532 lp_build_name(io_ptr, "io");
3533 lp_build_name(prim_id, "prim_id");
3534 lp_build_name(num_tess_coord, "num_tess_coord");
3535 lp_build_name(tess_coord[0], "tess_coord[0]");
3536 lp_build_name(tess_coord[1], "tess_coord[1]");
3537 lp_build_name(tess_outer, "tess_outer");
3538 lp_build_name(tess_inner, "tess_inner");
3539 lp_build_name(patch_vertices_in, "patch_vertices_in");
3540 lp_build_name(view_index, "view_index");
3541
3542 tes_iface.base.fetch_vertex_input = draw_tes_llvm_fetch_vertex_input;
3543 tes_iface.base.fetch_patch_input = draw_tes_llvm_fetch_patch_input;
3544 tes_iface.input = input_array;
3545 tes_iface.variant = variant;
3546
3547 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
3548 builder = gallivm->builder;
3549 LLVMPositionBuilderAtEnd(builder, block);
3550
3551 lp_build_context_init(&bld, gallivm, lp_type_int(32));
3552
3553 memset(&tes_type, 0, sizeof tes_type);
3554 tes_type.floating = true; /* floating point values */
3555 tes_type.sign = true; /* values are signed */
3556 tes_type.norm = false; /* values are not limited to [0,1] or [-1,1] */
3557 tes_type.width = 32; /* 32-bit float */
3558 tes_type.length = vector_length;
3559
3560 lp_build_context_init(&bldvec, variant->gallivm, lp_int_type(tes_type));
3561 consts_ptr = lp_jit_resources_constants(gallivm, variant->resources_type, resources_ptr);
3562
3563 ssbos_ptr = lp_jit_resources_ssbos(gallivm, variant->resources_type, resources_ptr);
3564
3565 sampler = lp_bld_llvm_sampler_soa_create(variant->key.samplers,
3566 MAX2(variant->key.nr_samplers,
3567 variant->key.nr_sampler_views));
3568 image = lp_bld_llvm_image_soa_create(draw_tes_llvm_variant_key_images(&variant->key),
3569 variant->key.nr_images);
3570 step = lp_build_const_int32(gallivm, vector_length);
3571
3572 system_values.tess_outer = LLVMBuildLoad2(builder, tess_outer_deref_type, tess_outer, "");
3573 system_values.tess_inner = LLVMBuildLoad2(builder, tess_inner_deref_type, tess_inner, "");
3574
3575 system_values.prim_id = lp_build_broadcast_scalar(&bldvec, prim_id);
3576
3577 system_values.view_index = view_index;
3578
3579 system_values.vertices_in = lp_build_broadcast_scalar(&bldvec, patch_vertices_in);
3580
3581 if (variant->key.primid_needed) {
3582 int slot = variant->key.primid_output;
3583 for (unsigned i = 0; i < 4; i++) {
3584 outputs[slot][i] = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, tes_type), "primid");
3585 LLVMBuildStore(builder, system_values.prim_id, outputs[slot][i]);
3586 }
3587 primid_slot = slot;
3588 }
3589 struct lp_build_loop_state lp_loop;
3590 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
3591 {
3592 LLVMValueRef io;
3593
3594 io = LLVMBuildGEP2(builder, variant->vertex_header_type, io_ptr, &lp_loop.counter, 1, "");
3595 mask_val = generate_tes_mask_value(variant, tes_type, num_tess_coord, lp_loop.counter);
3596 lp_build_mask_begin(&mask, gallivm, tes_type, mask_val);
3597
3598 system_values.tess_coord = LLVMGetUndef(LLVMArrayType(LLVMVectorType(flt_type, vector_length), 3));
3599 for (i = 0; i < 3; i++) {
3600 LLVMValueRef tess_coord_chan = LLVMGetUndef(LLVMVectorType(flt_type, vector_length));
3601 for (unsigned j = 0; j < vector_length; j++) {
3602 LLVMValueRef idx = LLVMBuildAdd(builder, lp_loop.counter, lp_build_const_int32(gallivm, j), "");
3603 LLVMValueRef tc_val;
3604 if (i == 2) {
3605 if (variant->shader->base.prim_mode == MESA_PRIM_TRIANGLES) {
3606 tc_val = lp_build_const_float(gallivm, 1.0);
3607 tc_val = LLVMBuildFSub(builder, tc_val, lp_build_pointer_get2(builder, flt_type, tess_coord[0], idx), "");
3608 tc_val = LLVMBuildFSub(builder, tc_val, lp_build_pointer_get2(builder, flt_type, tess_coord[1], idx), "");
3609 } else
3610 tc_val = lp_build_const_float(gallivm, 0.0);
3611 } else
3612 tc_val = lp_build_pointer_get2(builder, flt_type, tess_coord[i], idx);
3613
3614 tess_coord_chan = LLVMBuildInsertElement(builder, tess_coord_chan, tc_val, lp_build_const_int32(gallivm, j), "");
3615 }
3616 system_values.tess_coord = LLVMBuildInsertValue(builder, system_values.tess_coord, tess_coord_chan, i, "");
3617 }
3618
3619 struct lp_build_tgsi_params params;
3620 memset(¶ms, 0, sizeof(params));
3621
3622 params.type = tes_type;
3623 params.mask = &mask;
3624 params.consts_ptr = consts_ptr;
3625 params.system_values = &system_values;
3626 params.resources_type = variant->resources_type;
3627 params.resources_ptr = resources_ptr;
3628 params.sampler = sampler;
3629 params.info = &llvm->draw->tes.tess_eval_shader->info;
3630 params.ssbo_ptr = ssbos_ptr;
3631 params.image = image;
3632 params.tes_iface = &tes_iface.base;
3633 params.aniso_filter_table = lp_jit_resources_aniso_filter_table(gallivm, variant->resources_type, resources_ptr);
3634
3635 lp_build_nir_soa(variant->gallivm,
3636 llvm->draw->tes.tess_eval_shader->state.ir.nir,
3637 ¶ms,
3638 outputs);
3639
3640 lp_build_mask_end(&mask);
3641
3642 if (variant->key.clamp_vertex_color) {
3643 const struct tgsi_shader_info *info = &llvm->draw->tes.tess_eval_shader->info;
3644 do_clamp_vertex_color(variant->gallivm,
3645 tes_type, info,
3646 outputs);
3647 }
3648 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
3649 lp_int_type(tes_type), 0);
3650
3651 convert_to_aos(gallivm, variant->vertex_header_type, io, NULL, outputs, clipmask,
3652 draw_total_tes_outputs(llvm->draw), tes_type, primid_slot, false);
3653 }
3654 lp_build_loop_end_cond(&lp_loop, num_tess_coord, step, LLVMIntUGE);
3655 lp_bld_llvm_sampler_soa_destroy(sampler);
3656 lp_bld_llvm_image_soa_destroy(image);
3657
3658 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
3659 gallivm_verify_function(gallivm, variant_func);
3660 }
3661
3662
3663 struct draw_tes_llvm_variant *
draw_tes_llvm_create_variant(struct draw_llvm * llvm,unsigned num_outputs,const struct draw_tes_llvm_variant_key * key)3664 draw_tes_llvm_create_variant(struct draw_llvm *llvm,
3665 unsigned num_outputs,
3666 const struct draw_tes_llvm_variant_key *key)
3667 {
3668 struct draw_tes_llvm_variant *variant;
3669 struct llvm_tess_eval_shader *shader = llvm_tess_eval_shader(llvm->draw->tes.tess_eval_shader);
3670 char module_name[64];
3671 unsigned char ir_sha1_cache_key[20];
3672 struct lp_cached_code cached = { 0 };
3673 bool needs_caching = false;
3674
3675 variant = MALLOC(sizeof *variant +
3676 shader->variant_key_size - sizeof variant->key);
3677 if (!variant)
3678 return NULL;
3679
3680 variant->llvm = llvm;
3681 variant->shader = shader;
3682
3683 snprintf(module_name, sizeof(module_name), "draw_llvm_tes_variant%u",
3684 variant->shader->variants_cached);
3685
3686 memcpy(&variant->key, key, shader->variant_key_size);
3687 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
3688 draw_get_ir_cache_key(shader->base.state.ir.nir,
3689 key,
3690 shader->variant_key_size,
3691 num_outputs,
3692 ir_sha1_cache_key);
3693
3694 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
3695 &cached,
3696 ir_sha1_cache_key);
3697 if (!cached.data_size)
3698 needs_caching = true;
3699 }
3700 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
3701
3702 create_tes_jit_types(variant);
3703
3704 variant->vertex_header_type = lp_build_create_jit_vertex_header_type(variant->gallivm, num_outputs);
3705 variant->vertex_header_ptr_type = LLVMPointerType(variant->vertex_header_type, 0);
3706
3707 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
3708 nir_print_shader(llvm->draw->tes.tess_eval_shader->state.ir.nir, stderr);
3709 draw_tes_llvm_dump_variant_key(&variant->key);
3710 }
3711
3712 draw_tes_llvm_generate(llvm, variant);
3713
3714 gallivm_compile_module(variant->gallivm);
3715
3716 variant->jit_func = (draw_tes_jit_func)
3717 gallivm_jit_function(variant->gallivm, variant->function);
3718
3719 if (needs_caching)
3720 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
3721 &cached,
3722 ir_sha1_cache_key);
3723 gallivm_free_ir(variant->gallivm);
3724
3725 variant->list_item_global.base = variant;
3726 variant->list_item_local.base = variant;
3727 /*variant->no = */shader->variants_created++;
3728 variant->list_item_global.base = variant;
3729
3730 return variant;
3731 }
3732
3733
3734 void
draw_tes_llvm_destroy_variant(struct draw_tes_llvm_variant * variant)3735 draw_tes_llvm_destroy_variant(struct draw_tes_llvm_variant *variant)
3736 {
3737 struct draw_llvm *llvm = variant->llvm;
3738
3739 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
3740 debug_printf("Deleting TES variant: %u tes variants,\t%u total variants\n",
3741 variant->shader->variants_cached, llvm->nr_tes_variants);
3742 }
3743
3744 gallivm_destroy(variant->gallivm);
3745
3746 list_del(&variant->list_item_local.list);
3747 variant->shader->variants_cached--;
3748 list_del(&variant->list_item_global.list);
3749 llvm->nr_tes_variants--;
3750 FREE(variant);
3751 }
3752
3753
3754 struct draw_tes_llvm_variant_key *
draw_tes_llvm_make_variant_key(struct draw_llvm * llvm,char * store)3755 draw_tes_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
3756 {
3757 struct draw_tes_llvm_variant_key *key;
3758 struct lp_sampler_static_state *draw_sampler;
3759 struct lp_image_static_state *draw_image;
3760
3761 key = (struct draw_tes_llvm_variant_key *)store;
3762
3763 memset(key, 0, offsetof(struct draw_tes_llvm_variant_key, samplers[0]));
3764
3765 int primid_output = draw_find_shader_output(llvm->draw, TGSI_SEMANTIC_PRIMID, 0);
3766 if (primid_output >= 0) {
3767 key->primid_output = primid_output;
3768 key->primid_needed = true;
3769 }
3770
3771 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color &&
3772 llvm->draw->gs.geometry_shader == NULL;
3773
3774 /* All variants of this shader will have the same value for
3775 * nr_samplers. Not yet trying to compact away holes in the
3776 * sampler array.
3777 */
3778 key->nr_samplers = llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
3779 if (llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3780 key->nr_sampler_views =
3781 llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3782 } else {
3783 key->nr_sampler_views = key->nr_samplers;
3784 }
3785
3786 key->nr_images = llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
3787
3788 draw_sampler = key->samplers;
3789
3790 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
3791
3792 for (unsigned i = 0 ; i < key->nr_samplers; i++) {
3793 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
3794 llvm->draw->samplers[PIPE_SHADER_TESS_EVAL][i]);
3795 }
3796 for (unsigned i = 0 ; i < key->nr_sampler_views; i++) {
3797 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
3798 llvm->draw->sampler_views[PIPE_SHADER_TESS_EVAL][i]);
3799 }
3800
3801 draw_image = draw_tes_llvm_variant_key_images(key);
3802 memset(draw_image, 0,
3803 key->nr_images * sizeof *draw_image);
3804 for (unsigned i = 0; i < key->nr_images; i++) {
3805 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
3806 llvm->draw->images[PIPE_SHADER_TESS_EVAL][i]);
3807 }
3808 return key;
3809 }
3810
3811
3812 void
draw_tes_llvm_dump_variant_key(struct draw_tes_llvm_variant_key * key)3813 draw_tes_llvm_dump_variant_key(struct draw_tes_llvm_variant_key *key)
3814 {
3815 struct lp_sampler_static_state *sampler = key->samplers;
3816 struct lp_image_static_state *image = draw_tes_llvm_variant_key_images(key);
3817
3818 if (key->primid_needed)
3819 debug_printf("prim id output %d\n", key->primid_output);
3820 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
3821 for (unsigned i = 0 ; i < key->nr_sampler_views; i++) {
3822 debug_printf("sampler[%i].src_format = %s\n", i,
3823 util_format_name(sampler[i].texture_state.format));
3824 }
3825
3826 for (unsigned i = 0 ; i < key->nr_images; i++)
3827 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
3828
3829 }
3830