483b631fe8eed5de153db4b5d002add2c8a06342
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "draw_llvm.h"
29
30 #include "draw_context.h"
31 #include "draw_vs.h"
32 #include "draw_gs.h"
33
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_bitarit.h"
37 #include "gallivm/lp_bld_gather.h"
38 #include "gallivm/lp_bld_logic.h"
39 #include "gallivm/lp_bld_const.h"
40 #include "gallivm/lp_bld_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 "tgsi/tgsi_exec.h"
56 #include "tgsi/tgsi_dump.h"
57
58 #include "util/u_math.h"
59 #include "util/u_pointer.h"
60 #include "util/u_string.h"
61 #include "util/simple_list.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 static inline const struct draw_gs_llvm_iface *
79 draw_gs_llvm_iface(const struct lp_build_gs_iface *iface)
80 {
81 return (const struct draw_gs_llvm_iface *)iface;
82 }
83
84 struct draw_tcs_llvm_iface {
85 struct lp_build_tcs_iface base;
86
87 struct draw_tcs_llvm_variant *variant;
88 LLVMValueRef input;
89 LLVMValueRef output;
90 };
91
92 static inline const struct draw_tcs_llvm_iface *
93 draw_tcs_llvm_iface(const struct lp_build_tcs_iface *iface)
94 {
95 return (const struct draw_tcs_llvm_iface *)iface;
96 }
97
98 struct draw_tes_llvm_iface {
99 struct lp_build_tes_iface base;
100
101 struct draw_tes_llvm_variant *variant;
102 LLVMValueRef input;
103 };
104
105 static inline const struct draw_tes_llvm_iface *
106 draw_tes_llvm_iface(const struct lp_build_tes_iface *iface)
107 {
108 return (const struct draw_tes_llvm_iface *)iface;
109 }
110
111 /**
112 * Create LLVM type for draw_vertex_buffer.
113 */
114 static LLVMTypeRef
115 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
116 const char *struct_name)
117 {
118 LLVMTargetDataRef target = gallivm->target;
119 LLVMTypeRef dvbuffer_type;
120 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
121 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
122
123 elem_types[DRAW_JIT_DVBUFFER_MAP] =
124 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
125 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
126
127 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
128 ARRAY_SIZE(elem_types), 0);
129
130 (void) target; /* silence unused var warning for non-debug build */
131 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
132 target, dvbuffer_type,
133 DRAW_JIT_DVBUFFER_MAP);
134 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
135 target, dvbuffer_type,
136 DRAW_JIT_DVBUFFER_SIZE);
137
138 return dvbuffer_type;
139 }
140
141 /**
142 * Create LLVM type for struct draw_jit_texture
143 */
144 static LLVMTypeRef
145 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
146 {
147 LLVMTargetDataRef target = gallivm->target;
148 LLVMTypeRef texture_type;
149 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
150 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
151
152 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
153 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
154 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
155 elem_types[DRAW_JIT_TEXTURE_NUM_SAMPLES] =
156 elem_types[DRAW_JIT_TEXTURE_SAMPLE_STRIDE] =
157 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
158 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
159 elem_types[DRAW_JIT_TEXTURE_BASE] =
160 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
161 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
162 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
163 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
164 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
165
166 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
167 ARRAY_SIZE(elem_types), 0);
168
169 (void) target; /* silence unused var warning for non-debug build */
170 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
171 target, texture_type,
172 DRAW_JIT_TEXTURE_WIDTH);
173 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
174 target, texture_type,
175 DRAW_JIT_TEXTURE_HEIGHT);
176 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
177 target, texture_type,
178 DRAW_JIT_TEXTURE_DEPTH);
179 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
180 target, texture_type,
181 DRAW_JIT_TEXTURE_BASE);
182 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
183 target, texture_type,
184 DRAW_JIT_TEXTURE_ROW_STRIDE);
185 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
186 target, texture_type,
187 DRAW_JIT_TEXTURE_IMG_STRIDE);
188 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
189 target, texture_type,
190 DRAW_JIT_TEXTURE_FIRST_LEVEL);
191 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
192 target, texture_type,
193 DRAW_JIT_TEXTURE_LAST_LEVEL);
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
195 target, texture_type,
196 DRAW_JIT_TEXTURE_MIP_OFFSETS);
197 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, num_samples,
198 target, texture_type,
199 DRAW_JIT_TEXTURE_NUM_SAMPLES);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, sample_stride,
201 target, texture_type,
202 DRAW_JIT_TEXTURE_SAMPLE_STRIDE);
203
204 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
205
206 return texture_type;
207 }
208
209
210 /**
211 * Create LLVM type for struct draw_jit_sampler
212 */
213 static LLVMTypeRef
214 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
215 {
216 LLVMTargetDataRef target = gallivm->target;
217 LLVMTypeRef sampler_type;
218 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
219
220 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
221 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
222 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
223 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
224 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
225
226 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
227 ARRAY_SIZE(elem_types), 0);
228
229 (void) target; /* silence unused var warning for non-debug build */
230 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
231 target, sampler_type,
232 DRAW_JIT_SAMPLER_MIN_LOD);
233 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
234 target, sampler_type,
235 DRAW_JIT_SAMPLER_MAX_LOD);
236 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
237 target, sampler_type,
238 DRAW_JIT_SAMPLER_LOD_BIAS);
239 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
240 target, sampler_type,
241 DRAW_JIT_SAMPLER_BORDER_COLOR);
242
243 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
244
245 return sampler_type;
246 }
247
248 /**
249 * Create LLVM type for struct draw_jit_texture
250 */
251 static LLVMTypeRef
252 create_jit_image_type(struct gallivm_state *gallivm, const char *struct_name)
253 {
254 LLVMTargetDataRef target = gallivm->target;
255 LLVMTypeRef image_type;
256 LLVMTypeRef elem_types[DRAW_JIT_IMAGE_NUM_FIELDS];
257 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
258
259 elem_types[DRAW_JIT_IMAGE_WIDTH] =
260 elem_types[DRAW_JIT_IMAGE_HEIGHT] =
261 elem_types[DRAW_JIT_IMAGE_DEPTH] =
262 elem_types[DRAW_JIT_IMAGE_ROW_STRIDE] =
263 elem_types[DRAW_JIT_IMAGE_IMG_STRIDE] =
264 elem_types[DRAW_JIT_IMAGE_NUM_SAMPLES] =
265 elem_types[DRAW_JIT_IMAGE_SAMPLE_STRIDE] = int32_type;
266 elem_types[DRAW_JIT_IMAGE_BASE] =
267 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
268
269 image_type = LLVMStructTypeInContext(gallivm->context, elem_types,
270 ARRAY_SIZE(elem_types), 0);
271
272 (void) target; /* silence unused var warning for non-debug build */
273 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, width,
274 target, image_type,
275 DRAW_JIT_IMAGE_WIDTH);
276 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, height,
277 target, image_type,
278 DRAW_JIT_IMAGE_HEIGHT);
279 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, depth,
280 target, image_type,
281 DRAW_JIT_IMAGE_DEPTH);
282 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, base,
283 target, image_type,
284 DRAW_JIT_IMAGE_BASE);
285 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, row_stride,
286 target, image_type,
287 DRAW_JIT_IMAGE_ROW_STRIDE);
288 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, img_stride,
289 target, image_type,
290 DRAW_JIT_IMAGE_IMG_STRIDE);
291 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, num_samples,
292 target, image_type,
293 DRAW_JIT_IMAGE_NUM_SAMPLES);
294 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, sample_stride,
295 target, image_type,
296 DRAW_JIT_IMAGE_SAMPLE_STRIDE);
297
298 LP_CHECK_STRUCT_SIZE(struct draw_jit_image, target, image_type);
299
300 return image_type;
301 }
302
303 /**
304 * Create LLVM type for struct draw_jit_context
305 */
306 static LLVMTypeRef
307 create_jit_context_type(struct gallivm_state *gallivm,
308 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
309 LLVMTypeRef image_type,
310 const char *struct_name)
311 {
312 LLVMTargetDataRef target = gallivm->target;
313 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
314 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
315 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
316 LLVMTypeRef context_type;
317
318 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
319 LP_MAX_TGSI_CONST_BUFFERS);
320 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
321 LP_MAX_TGSI_CONST_BUFFERS);
322 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
323 DRAW_TOTAL_CLIP_PLANES), 0);
324 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
325 elem_types[4] = LLVMArrayType(texture_type,
326 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
327 elem_types[5] = LLVMArrayType(sampler_type,
328 PIPE_MAX_SAMPLERS); /* samplers */
329 elem_types[6] = LLVMArrayType(image_type,
330 PIPE_MAX_SHADER_IMAGES); /* images */
331 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* vs_ssbo */
332 LP_MAX_TGSI_SHADER_BUFFERS);
333 elem_types[8] = LLVMArrayType(int_type, /* num_vs_ssbos */
334 LP_MAX_TGSI_SHADER_BUFFERS);
335 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
336 ARRAY_SIZE(elem_types), 0);
337
338 (void) target; /* silence unused var warning for non-debug build */
339 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
340 target, context_type, DRAW_JIT_CTX_CONSTANTS);
341 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
342 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
343 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
344 target, context_type, DRAW_JIT_CTX_PLANES);
345 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
346 target, context_type, DRAW_JIT_CTX_VIEWPORT);
347 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
348 target, context_type,
349 DRAW_JIT_CTX_TEXTURES);
350 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
351 target, context_type,
352 DRAW_JIT_CTX_SAMPLERS);
353 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, images,
354 target, context_type, DRAW_JIT_CTX_IMAGES);
355 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_ssbos,
356 target, context_type, DRAW_JIT_CTX_SSBOS);
357 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_ssbos,
358 target, context_type, DRAW_JIT_CTX_NUM_SSBOS);
359 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
360 target, context_type);
361
362 return context_type;
363 }
364
365
366 /**
367 * Create LLVM type for struct draw_gs_jit_context
368 */
369 static LLVMTypeRef
370 create_gs_jit_context_type(struct gallivm_state *gallivm,
371 unsigned vector_length,
372 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
373 LLVMTypeRef image_type,
374 const char *struct_name)
375 {
376 LLVMTargetDataRef target = gallivm->target;
377 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
378 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
379 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
380 LLVMTypeRef context_type;
381
382 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
383 LP_MAX_TGSI_CONST_BUFFERS);
384 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
385 LP_MAX_TGSI_CONST_BUFFERS);
386 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
387 DRAW_TOTAL_CLIP_PLANES), 0);
388 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
389
390 elem_types[4] = LLVMArrayType(texture_type,
391 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
392 elem_types[5] = LLVMArrayType(sampler_type,
393 PIPE_MAX_SAMPLERS); /* samplers */
394 elem_types[6] = LLVMArrayType(image_type,
395 PIPE_MAX_SHADER_IMAGES); /* images */
396 elem_types[7] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
397 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
398 vector_length), 0);
399 elem_types[9] = LLVMPointerType(LLVMVectorType(int_type,
400 vector_length), 0);
401
402 elem_types[10] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
403 LP_MAX_TGSI_SHADER_BUFFERS);
404 elem_types[11] = LLVMArrayType(int_type, /* num_ssbos */
405 LP_MAX_TGSI_SHADER_BUFFERS);
406
407 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
408 ARRAY_SIZE(elem_types), 0);
409
410 (void) target; /* silence unused var warning for non-debug build */
411 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
412 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
413 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
414 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
415 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
416 target, context_type, DRAW_GS_JIT_CTX_PLANES);
417 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
418 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
419 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
420 target, context_type,
421 DRAW_GS_JIT_CTX_TEXTURES);
422 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
423 target, context_type,
424 DRAW_GS_JIT_CTX_SAMPLERS);
425 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
426 target, context_type,
427 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
428 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
429 target, context_type,
430 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
431 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
432 target, context_type,
433 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
434 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, ssbos,
435 target, context_type, DRAW_GS_JIT_CTX_SSBOS);
436 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_ssbos,
437 target, context_type, DRAW_GS_JIT_CTX_NUM_SSBOS);
438 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, images,
439 target, context_type, DRAW_GS_JIT_CTX_IMAGES);
440 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
441 target, context_type);
442
443 return context_type;
444 }
445
446
447 static LLVMTypeRef
448 create_gs_jit_input_type(struct gallivm_state *gallivm)
449 {
450 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
451 LLVMTypeRef input_array;
452
453 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
454 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
455 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
456 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
457
458 return input_array;
459 }
460
461 /**
462 * Create LLVM type for struct pipe_vertex_buffer
463 */
464 static LLVMTypeRef
465 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
466 const char *struct_name)
467 {
468 LLVMTargetDataRef target = gallivm->target;
469 LLVMTypeRef elem_types[4];
470 LLVMTypeRef vb_type;
471
472 elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
473 elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
474 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
475 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
476
477 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
478 ARRAY_SIZE(elem_types), 0);
479
480 (void) target; /* silence unused var warning for non-debug build */
481 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
482 target, vb_type, 0);
483 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
484 target, vb_type, 1);
485 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
486 target, vb_type, 2);
487 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
488 target, vb_type, 3);
489
490 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
491
492 return vb_type;
493 }
494
495
496 /**
497 * Create LLVM type for struct vertex_header;
498 */
499 static LLVMTypeRef
500 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
501 {
502 LLVMTargetDataRef target = gallivm->target;
503 LLVMTypeRef elem_types[3];
504 LLVMTypeRef vertex_header;
505 char struct_name[24];
506
507 snprintf(struct_name, 23, "vertex_header%d", data_elems);
508
509 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
510 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
511 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
512
513 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
514 ARRAY_SIZE(elem_types), 0);
515
516 /* these are bit-fields and we can't take address of them
517 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
518 target, vertex_header,
519 DRAW_JIT_VERTEX_CLIPMASK);
520 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
521 target, vertex_header,
522 DRAW_JIT_VERTEX_EDGEFLAG);
523 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
524 target, vertex_header,
525 DRAW_JIT_VERTEX_PAD);
526 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
527 target, vertex_header,
528 DRAW_JIT_VERTEX_VERTEX_ID);
529 */
530 (void) target; /* silence unused var warning for non-debug build */
531 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
532 target, vertex_header,
533 DRAW_JIT_VERTEX_CLIP_POS);
534 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
535 target, vertex_header,
536 DRAW_JIT_VERTEX_DATA);
537
538 assert(LLVMABISizeOfType(target, vertex_header) ==
539 offsetof(struct vertex_header, data[data_elems]));
540
541 return vertex_header;
542 }
543
544 /**
545 * Create LLVM type for struct draw_tcs_jit_context
546 */
547 static LLVMTypeRef
548 create_tcs_jit_context_type(struct gallivm_state *gallivm,
549 unsigned vector_length,
550 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
551 LLVMTypeRef image_type,
552 const char *struct_name)
553 {
554 LLVMTargetDataRef target = gallivm->target;
555 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
556 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
557 LLVMTypeRef elem_types[DRAW_TCS_JIT_CTX_NUM_FIELDS];
558 LLVMTypeRef context_type;
559
560 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
561 LP_MAX_TGSI_CONST_BUFFERS);
562 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
563 LP_MAX_TGSI_CONST_BUFFERS);
564 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
565 elem_types[3] = LLVMInt32TypeInContext(gallivm->context);
566
567 elem_types[4] = LLVMArrayType(texture_type,
568 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
569 elem_types[5] = LLVMArrayType(sampler_type,
570 PIPE_MAX_SAMPLERS); /* samplers */
571 elem_types[6] = LLVMArrayType(image_type,
572 PIPE_MAX_SHADER_IMAGES); /* images */
573
574 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
575 LP_MAX_TGSI_SHADER_BUFFERS);
576 elem_types[8] = LLVMArrayType(int_type, /* num_ssbos */
577 LP_MAX_TGSI_SHADER_BUFFERS);
578
579 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
580 ARRAY_SIZE(elem_types), 0);
581
582 (void) target; /* silence unused var warning for non-debug build */
583 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, constants,
584 target, context_type, DRAW_TCS_JIT_CTX_CONSTANTS);
585 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, num_constants,
586 target, context_type, DRAW_TCS_JIT_CTX_NUM_CONSTANTS);
587 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, textures,
588 target, context_type,
589 DRAW_TCS_JIT_CTX_TEXTURES);
590 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, samplers,
591 target, context_type,
592 DRAW_TCS_JIT_CTX_SAMPLERS);
593 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, ssbos,
594 target, context_type, DRAW_TCS_JIT_CTX_SSBOS);
595 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, num_ssbos,
596 target, context_type, DRAW_TCS_JIT_CTX_NUM_SSBOS);
597 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, images,
598 target, context_type, DRAW_TCS_JIT_CTX_IMAGES);
599 LP_CHECK_STRUCT_SIZE(struct draw_tcs_jit_context,
600 target, context_type);
601
602 return context_type;
603 }
604
605 static LLVMTypeRef
606 create_tcs_jit_input_type(struct gallivm_state *gallivm)
607 {
608 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
609 LLVMTypeRef input_array;
610
611 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
612 input_array = LLVMArrayType(input_array, NUM_TCS_INPUTS); /* num attrs per vertex */
613 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
614
615 return input_array;
616 }
617
618 static LLVMTypeRef
619 create_tcs_jit_output_type(struct gallivm_state *gallivm)
620 {
621 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
622 LLVMTypeRef output_array;
623
624 output_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
625 output_array = LLVMArrayType(output_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
626 output_array = LLVMPointerType(output_array, 0); /* num vertices per prim */
627
628 return output_array;
629 }
630
631 static LLVMTypeRef
632 create_tes_jit_input_type(struct gallivm_state *gallivm)
633 {
634 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
635 LLVMTypeRef input_array;
636
637 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
638 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
639 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
640
641 return input_array;
642 }
643
644 /**
645 * Create LLVM type for struct draw_tes_jit_context
646 */
647 static LLVMTypeRef
648 create_tes_jit_context_type(struct gallivm_state *gallivm,
649 unsigned vector_length,
650 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
651 LLVMTypeRef image_type,
652 const char *struct_name)
653 {
654 LLVMTargetDataRef target = gallivm->target;
655 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
656 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
657 LLVMTypeRef elem_types[DRAW_TCS_JIT_CTX_NUM_FIELDS];
658 LLVMTypeRef context_type;
659
660 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
661 LP_MAX_TGSI_CONST_BUFFERS);
662 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
663 LP_MAX_TGSI_CONST_BUFFERS);
664 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
665 elem_types[3] = LLVMInt32TypeInContext(gallivm->context);
666
667 elem_types[4] = LLVMArrayType(texture_type,
668 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
669 elem_types[5] = LLVMArrayType(sampler_type,
670 PIPE_MAX_SAMPLERS); /* samplers */
671 elem_types[6] = LLVMArrayType(image_type,
672 PIPE_MAX_SHADER_IMAGES); /* images */
673
674 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
675 LP_MAX_TGSI_SHADER_BUFFERS);
676 elem_types[8] = LLVMArrayType(int_type, /* num_ssbos */
677 LP_MAX_TGSI_SHADER_BUFFERS);
678
679 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
680 ARRAY_SIZE(elem_types), 0);
681
682 (void) target; /* silence unused var warning for non-debug build */
683 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, constants,
684 target, context_type, DRAW_TCS_JIT_CTX_CONSTANTS);
685 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, num_constants,
686 target, context_type, DRAW_TCS_JIT_CTX_NUM_CONSTANTS);
687 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, textures,
688 target, context_type,
689 DRAW_TCS_JIT_CTX_TEXTURES);
690 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, samplers,
691 target, context_type,
692 DRAW_TCS_JIT_CTX_SAMPLERS);
693 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, ssbos,
694 target, context_type, DRAW_TCS_JIT_CTX_SSBOS);
695 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, num_ssbos,
696 target, context_type, DRAW_TCS_JIT_CTX_NUM_SSBOS);
697 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, images,
698 target, context_type, DRAW_TCS_JIT_CTX_IMAGES);
699 LP_CHECK_STRUCT_SIZE(struct draw_tes_jit_context,
700 target, context_type);
701
702 return context_type;
703 }
704
705 /**
706 * Create LLVM types for various structures.
707 */
708 static void
709 create_jit_types(struct draw_llvm_variant *variant)
710 {
711 struct gallivm_state *gallivm = variant->gallivm;
712 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
713 vb_type, image_type;
714
715 texture_type = create_jit_texture_type(gallivm, "texture");
716 sampler_type = create_jit_sampler_type(gallivm, "sampler");
717 image_type = create_jit_image_type(gallivm, "image");
718
719 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
720 image_type,
721 "draw_jit_context");
722 variant->context_ptr_type = LLVMPointerType(context_type, 0);
723
724 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
725 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
726
727 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
728 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
729 }
730
731
732 static LLVMTypeRef
733 get_context_ptr_type(struct draw_llvm_variant *variant)
734 {
735 if (!variant->context_ptr_type)
736 create_jit_types(variant);
737 return variant->context_ptr_type;
738 }
739
740
741 static LLVMTypeRef
742 get_buffer_ptr_type(struct draw_llvm_variant *variant)
743 {
744 if (!variant->buffer_ptr_type)
745 create_jit_types(variant);
746 return variant->buffer_ptr_type;
747 }
748
749
750 static LLVMTypeRef
751 get_vb_ptr_type(struct draw_llvm_variant *variant)
752 {
753 if (!variant->vb_ptr_type)
754 create_jit_types(variant);
755 return variant->vb_ptr_type;
756 }
757
758 static LLVMTypeRef
759 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
760 {
761 if (!variant->vertex_header_ptr_type)
762 create_jit_types(variant);
763 return variant->vertex_header_ptr_type;
764 }
765
766
767 /**
768 * Create per-context LLVM info.
769 */
770 struct draw_llvm *
771 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
772 {
773 struct draw_llvm *llvm;
774
775 if (!lp_build_init())
776 return NULL;
777
778 llvm = CALLOC_STRUCT( draw_llvm );
779 if (!llvm)
780 return NULL;
781
782 llvm->draw = draw;
783
784 llvm->context = context;
785 if (!llvm->context) {
786 llvm->context = LLVMContextCreate();
787 llvm->context_owned = true;
788 }
789 if (!llvm->context)
790 goto fail;
791
792 llvm->nr_variants = 0;
793 make_empty_list(&llvm->vs_variants_list);
794
795 llvm->nr_gs_variants = 0;
796 make_empty_list(&llvm->gs_variants_list);
797
798 llvm->nr_tcs_variants = 0;
799 make_empty_list(&llvm->tcs_variants_list);
800
801 llvm->nr_tes_variants = 0;
802 make_empty_list(&llvm->tes_variants_list);
803
804 return llvm;
805
806 fail:
807 draw_llvm_destroy(llvm);
808 return NULL;
809 }
810
811
812 /**
813 * Free per-context LLVM info.
814 */
815 void
816 draw_llvm_destroy(struct draw_llvm *llvm)
817 {
818 if (llvm->context_owned)
819 LLVMContextDispose(llvm->context);
820 llvm->context = NULL;
821
822 /* XXX free other draw_llvm data? */
823 FREE(llvm);
824 }
825
826 static void
827 draw_get_ir_cache_key(struct nir_shader *nir,
828 const void *key, size_t key_size,
829 uint32_t val_32bit,
830 unsigned char ir_sha1_cache_key[20])
831 {
832 struct blob blob = { 0 };
833 unsigned ir_size;
834 void *ir_binary;
835
836 blob_init(&blob);
837 nir_serialize(&blob, nir, true);
838 ir_binary = blob.data;
839 ir_size = blob.size;
840
841 struct mesa_sha1 ctx;
842 _mesa_sha1_init(&ctx);
843 _mesa_sha1_update(&ctx, key, key_size);
844 _mesa_sha1_update(&ctx, ir_binary, ir_size);
845 _mesa_sha1_update(&ctx, &val_32bit, 4);
846 _mesa_sha1_final(&ctx, ir_sha1_cache_key);
847
848 blob_finish(&blob);
849 }
850
851 /**
852 * Create LLVM-generated code for a vertex shader.
853 */
854 struct draw_llvm_variant *
855 draw_llvm_create_variant(struct draw_llvm *llvm,
856 unsigned num_inputs,
857 const struct draw_llvm_variant_key *key)
858 {
859 struct draw_llvm_variant *variant;
860 struct llvm_vertex_shader *shader =
861 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
862 LLVMTypeRef vertex_header;
863 char module_name[64];
864 unsigned char ir_sha1_cache_key[20];
865 struct lp_cached_code cached = { 0 };
866 bool needs_caching = false;
867 variant = MALLOC(sizeof *variant +
868 shader->variant_key_size -
869 sizeof variant->key);
870 if (!variant)
871 return NULL;
872
873 variant->llvm = llvm;
874 variant->shader = shader;
875 memcpy(&variant->key, key, shader->variant_key_size);
876
877 snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
878 variant->shader->variants_cached);
879
880 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
881 draw_get_ir_cache_key(shader->base.state.ir.nir,
882 key,
883 shader->variant_key_size,
884 num_inputs,
885 ir_sha1_cache_key);
886
887 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
888 &cached,
889 ir_sha1_cache_key);
890 if (!cached.data_size)
891 needs_caching = true;
892 }
893 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
894
895 create_jit_types(variant);
896
897 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
898 if (llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_TGSI)
899 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
900 else
901 nir_print_shader(llvm->draw->vs.vertex_shader->state.ir.nir, stderr);
902 draw_llvm_dump_variant_key(&variant->key);
903 }
904
905 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
906
907 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
908
909 draw_llvm_generate(llvm, variant);
910
911 gallivm_compile_module(variant->gallivm);
912
913 variant->jit_func = (draw_jit_vert_func)
914 gallivm_jit_function(variant->gallivm, variant->function);
915
916 if (needs_caching)
917 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
918 &cached,
919 ir_sha1_cache_key);
920 gallivm_free_ir(variant->gallivm);
921
922 variant->list_item_global.base = variant;
923 variant->list_item_local.base = variant;
924 /*variant->no = */shader->variants_created++;
925 variant->list_item_global.base = variant;
926
927 return variant;
928 }
929
930
931 static void
932 generate_vs(struct draw_llvm_variant *variant,
933 LLVMBuilderRef builder,
934 struct lp_type vs_type,
935 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
936 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
937 const struct lp_bld_tgsi_system_values *system_values,
938 LLVMValueRef context_ptr,
939 const struct lp_build_sampler_soa *draw_sampler,
940 const struct lp_build_image_soa *draw_image,
941 boolean clamp_vertex_color,
942 struct lp_build_mask_context *bld_mask)
943 {
944 struct draw_llvm *llvm = variant->llvm;
945 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
946 LLVMValueRef consts_ptr =
947 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
948 LLVMValueRef num_consts_ptr =
949 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
950 LLVMValueRef ssbos_ptr =
951 draw_jit_context_vs_ssbos(variant->gallivm, context_ptr);
952 LLVMValueRef num_ssbos_ptr =
953 draw_jit_context_num_vs_ssbos(variant->gallivm, context_ptr);
954
955 struct lp_build_tgsi_params params;
956 memset(&params, 0, sizeof(params));
957
958 params.type = vs_type;
959 params.mask = bld_mask;
960 params.consts_ptr = consts_ptr;
961 params.const_sizes_ptr = num_consts_ptr;
962 params.system_values = system_values;
963 params.inputs = inputs;
964 params.context_ptr = context_ptr;
965 params.sampler = draw_sampler;
966 params.info = &llvm->draw->vs.vertex_shader->info;
967 params.ssbo_ptr = ssbos_ptr;
968 params.ssbo_sizes_ptr = num_ssbos_ptr;
969 params.image = draw_image;
970
971 if (llvm->draw->vs.vertex_shader->state.ir.nir &&
972 llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_NIR)
973 lp_build_nir_soa(variant->gallivm,
974 llvm->draw->vs.vertex_shader->state.ir.nir,
975 &params,
976 outputs);
977 else
978 lp_build_tgsi_soa(variant->gallivm,
979 tokens,
980 &params,
981 outputs);
982
983 {
984 LLVMValueRef out;
985 unsigned chan, attrib;
986 struct lp_build_context bld;
987 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
988 lp_build_context_init(&bld, variant->gallivm, vs_type);
989
990 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
991 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
992 if (outputs[attrib][chan]) {
993 switch (info->output_semantic_name[attrib]) {
994 case TGSI_SEMANTIC_COLOR:
995 case TGSI_SEMANTIC_BCOLOR:
996 if (clamp_vertex_color) {
997 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
998 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
999 LLVMBuildStore(builder, out, outputs[attrib][chan]);
1000 }
1001 break;
1002 }
1003 }
1004 }
1005 }
1006 }
1007 }
1008
1009
1010 static void
1011 fetch_instanced(struct gallivm_state *gallivm,
1012 const struct util_format_description *format_desc,
1013 struct lp_type vs_type,
1014 LLVMValueRef vb_stride,
1015 LLVMValueRef map_ptr,
1016 LLVMValueRef buffer_size_adj,
1017 LLVMValueRef *inputs,
1018 LLVMValueRef index)
1019 {
1020 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
1021 LLVMTypeRef aosf_t, aosi_t;
1022 LLVMValueRef zero = LLVMConstNull(i32_t);
1023 LLVMBuilderRef builder = gallivm->builder;
1024 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
1025 unsigned i;
1026
1027 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
1028 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
1029
1030 /* This mul can overflow. Wraparound is ok. */
1031 stride = LLVMBuildMul(builder, vb_stride, index, "");
1032
1033 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
1034 stride, buffer_size_adj,
1035 "buffer_overflowed");
1036
1037 if (0) {
1038 lp_build_print_value(gallivm, " instance index = ", index);
1039 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
1040 }
1041
1042 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
1043 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
1044 stride = LLVMBuildAnd(builder, stride, index_valid, "");
1045
1046 aos = lp_build_fetch_rgba_aos(gallivm,
1047 format_desc,
1048 lp_float32_vec4_type(),
1049 FALSE,
1050 map_ptr,
1051 stride, zero, zero,
1052 NULL);
1053
1054 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
1055 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
1056 aos = LLVMBuildAnd(builder, aos, index_valid, "");
1057 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
1058
1059 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1060 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1061 inputs[i] = lp_build_extract_broadcast(gallivm,
1062 lp_float32_vec4_type(),
1063 vs_type, aos, index);
1064 }
1065 }
1066
1067
1068 static void
1069 fetch_vector(struct gallivm_state *gallivm,
1070 const struct util_format_description *format_desc,
1071 struct lp_type vs_type,
1072 LLVMValueRef vb_stride,
1073 LLVMValueRef map_ptr,
1074 LLVMValueRef buffer_size_adj,
1075 LLVMValueRef *inputs,
1076 LLVMValueRef indices)
1077 {
1078 LLVMBuilderRef builder = gallivm->builder;
1079 struct lp_build_context blduivec;
1080 struct lp_type fetch_type = vs_type;
1081 LLVMValueRef offset, valid_mask;
1082 unsigned i;
1083
1084 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1085
1086 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
1087 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
1088
1089 /* This mul can overflow. Wraparound is ok. */
1090 offset = lp_build_mul(&blduivec, vb_stride, indices);
1091
1092 valid_mask = lp_build_compare(gallivm, blduivec.type,
1093 PIPE_FUNC_LESS, offset, buffer_size_adj);
1094
1095 /* not valid elements use offset 0 */
1096 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
1097
1098 if (0) {
1099 lp_build_print_value(gallivm, " indices = ", indices);
1100 lp_build_print_value(gallivm, " offsets = ", offset);
1101 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
1102 }
1103
1104 /*
1105 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
1106 * This should always produce better code.
1107 */
1108
1109 /* The type handling is annoying here... */
1110 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
1111 format_desc->channel[0].pure_integer) {
1112 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
1113 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
1114 }
1115 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1116 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
1117 }
1118 }
1119
1120 lp_build_fetch_rgba_soa(gallivm, format_desc,
1121 fetch_type, FALSE, map_ptr, offset,
1122 blduivec.zero, blduivec.zero,
1123 NULL, inputs);
1124
1125 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1126 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
1127 lp_build_vec_type(gallivm, vs_type), "");
1128 }
1129
1130 /* out-of-bound fetches return all zeros */
1131 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1132 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
1133 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
1134 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
1135 lp_build_vec_type(gallivm, vs_type), "");
1136 }
1137 }
1138
1139
1140 static void
1141 store_aos(struct gallivm_state *gallivm,
1142 LLVMValueRef io_ptr,
1143 LLVMValueRef index,
1144 LLVMValueRef value)
1145 {
1146 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
1147 LLVMBuilderRef builder = gallivm->builder;
1148 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
1149 LLVMValueRef indices[3];
1150
1151 indices[0] = lp_build_const_int32(gallivm, 0);
1152 indices[1] = index;
1153 indices[2] = lp_build_const_int32(gallivm, 0);
1154
1155 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
1156 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
1157
1158 #if DEBUG_STORE
1159 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
1160 #endif
1161
1162 /* Unaligned store due to the vertex header */
1163 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
1164 }
1165
1166 /**
1167 * Adjust the mask to architecture endianess. The mask will the store in struct:
1168 *
1169 * struct vertex_header {
1170 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
1171 * unsigned edgeflag:1;
1172 * unsigned pad:1;
1173 * unsigned vertex_id:16;
1174 * [...]
1175 * }
1176 *
1177 * On little-endian machine nothing needs to done, however on bit-endian machine
1178 * the mask's fields need to be adjusted with the algorithm:
1179 *
1180 * uint32_t reverse (uint32_t x)
1181 * {
1182 * return (x >> 16) | // vertex_id
1183 * ((x & 0x3fff) << 18) | // clipmask
1184 * ((x & 0x4000) << 3) | // edgeflag
1185 * ((x & 0x8000) << 1); // pad
1186 * }
1187 */
1188 static LLVMValueRef
1189 adjust_mask(struct gallivm_state *gallivm,
1190 LLVMValueRef mask)
1191 {
1192 #if UTIL_ARCH_BIG_ENDIAN
1193 LLVMBuilderRef builder = gallivm->builder;
1194 LLVMValueRef vertex_id;
1195 LLVMValueRef clipmask;
1196 LLVMValueRef pad;
1197 LLVMValueRef edgeflag;
1198
1199 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
1200 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
1201 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
1202 if (0) {
1203 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
1204 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 1), "");
1205 }
1206 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
1207 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 3), "");
1208
1209 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
1210 if (0) {
1211 mask = LLVMBuildOr(builder, mask, pad, "");
1212 }
1213 mask = LLVMBuildOr(builder, mask, edgeflag, "");
1214 #endif
1215 return mask;
1216 }
1217
1218 static void
1219 store_aos_array(struct gallivm_state *gallivm,
1220 struct lp_type soa_type,
1221 LLVMValueRef io_ptr,
1222 LLVMValueRef *indices,
1223 LLVMValueRef* aos,
1224 int attrib,
1225 int num_outputs,
1226 LLVMValueRef clipmask,
1227 boolean need_edgeflag)
1228 {
1229 LLVMBuilderRef builder = gallivm->builder;
1230 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
1231 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1232 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
1233 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1234 int vector_length = soa_type.length;
1235 int i;
1236
1237 debug_assert(TGSI_NUM_CHANNELS == 4);
1238
1239 for (i = 0; i < vector_length; i++) {
1240 linear_inds[i] = lp_build_const_int32(gallivm, i);
1241 if (indices) {
1242 inds[i] = indices[i];
1243 } else {
1244 inds[i] = linear_inds[i];
1245 }
1246 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1247 }
1248
1249 if (attrib == 0) {
1250 /* store vertex header for each of the n vertices */
1251 LLVMValueRef val, cliptmp;
1252 int vertex_id_pad_edgeflag;
1253
1254 /* If this assertion fails, it means we need to update the bit twidding
1255 * code here. See struct vertex_header in draw_private.h.
1256 */
1257 assert(DRAW_TOTAL_CLIP_PLANES==14);
1258 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
1259 if (!need_edgeflag) {
1260 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
1261 }
1262 else {
1263 vertex_id_pad_edgeflag = (0xffff << 16);
1264 }
1265 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
1266 vertex_id_pad_edgeflag);
1267 /* OR with the clipmask */
1268 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
1269 for (i = 0; i < vector_length; i++) {
1270 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
1271 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
1272 val = adjust_mask(gallivm, val);
1273 #if DEBUG_STORE
1274 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
1275 io_ptrs[i], inds[i], val);
1276 #endif
1277 LLVMBuildStore(builder, val, id_ptr);
1278 }
1279 }
1280
1281 /* store for each of the n vertices */
1282 for (i = 0; i < vector_length; i++) {
1283 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
1284 }
1285 }
1286
1287
1288 static void
1289 convert_to_aos(struct gallivm_state *gallivm,
1290 LLVMValueRef io,
1291 LLVMValueRef *indices,
1292 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1293 LLVMValueRef clipmask,
1294 int num_outputs,
1295 struct lp_type soa_type,
1296 boolean need_edgeflag)
1297 {
1298 LLVMBuilderRef builder = gallivm->builder;
1299 unsigned chan, attrib, i;
1300
1301 #if DEBUG_STORE
1302 lp_build_printf(gallivm, " # storing begin\n");
1303 #endif
1304 for (attrib = 0; attrib < num_outputs; ++attrib) {
1305 LLVMValueRef soa[TGSI_NUM_CHANNELS];
1306 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
1307 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1308 if (outputs[attrib][chan]) {
1309 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
1310 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
1311 #if DEBUG_STORE
1312 lp_build_printf(gallivm, "output %d : %d ",
1313 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1314 attrib, 0),
1315 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1316 chan, 0));
1317 lp_build_print_value(gallivm, "val = ", out);
1318 {
1319 LLVMValueRef iv =
1320 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
1321
1322 lp_build_print_value(gallivm, " ival = ", iv);
1323 }
1324 #endif
1325 soa[chan] = out;
1326 }
1327 else {
1328 soa[chan] = 0;
1329 }
1330 }
1331
1332
1333 if (soa_type.length == TGSI_NUM_CHANNELS) {
1334 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1335 } else {
1336 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1337
1338 for (i = 0; i < soa_type.length; ++i) {
1339 aos[i] = lp_build_extract_range(gallivm,
1340 soa[i % TGSI_NUM_CHANNELS],
1341 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1342 TGSI_NUM_CHANNELS);
1343 }
1344 }
1345
1346 store_aos_array(gallivm,
1347 soa_type,
1348 io, indices,
1349 aos,
1350 attrib,
1351 num_outputs,
1352 clipmask,
1353 need_edgeflag);
1354 }
1355 #if DEBUG_STORE
1356 lp_build_printf(gallivm, " # storing end\n");
1357 #endif
1358 }
1359
1360
1361 /**
1362 * Stores original vertex positions in clip coordinates
1363 */
1364 static void
1365 store_clip(struct gallivm_state *gallivm,
1366 const struct lp_type vs_type,
1367 LLVMValueRef io_ptr,
1368 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1369 int idx)
1370 {
1371 LLVMBuilderRef builder = gallivm->builder;
1372 LLVMValueRef soa[4];
1373 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1374 LLVMValueRef indices[2];
1375 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1376 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1377 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1378 LLVMTypeRef clip_ptr_type =
1379 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1380 4), 0);
1381 int i, j;
1382
1383 indices[0] =
1384 indices[1] = lp_build_const_int32(gallivm, 0);
1385
1386 for (i = 0; i < vs_type.length; i++) {
1387 inds[i] = lp_build_const_int32(gallivm, i);
1388 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1389 }
1390
1391 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1392 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1393 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1394 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1395
1396 for (i = 0; i < vs_type.length; i++) {
1397 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1398 }
1399
1400 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1401 for (i = 0; i < vs_type.length; ++i) {
1402 aos[i] = lp_build_extract_range(gallivm,
1403 soa[i % TGSI_NUM_CHANNELS],
1404 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1405 TGSI_NUM_CHANNELS);
1406 }
1407
1408 for (j = 0; j < vs_type.length; j++) {
1409 LLVMValueRef clip_ptr;
1410
1411 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1412 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1413
1414 /* Unaligned store */
1415 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1416 }
1417 }
1418
1419
1420 /**
1421 * Transforms the outputs for viewport mapping
1422 */
1423 static void
1424 generate_viewport(struct draw_llvm_variant *variant,
1425 LLVMBuilderRef builder,
1426 struct lp_type vs_type,
1427 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1428 LLVMValueRef context_ptr)
1429 {
1430 int i;
1431 struct gallivm_state *gallivm = variant->gallivm;
1432 struct lp_type f32_type = vs_type;
1433 const unsigned pos = variant->llvm->draw->vs.position_output;
1434 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1435 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1436 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1437 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1438
1439 /* We treat pipe_viewport_state as a float array */
1440 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1441 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1442
1443 /* for 1/w convention*/
1444 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1445 LLVMBuildStore(builder, out3, outputs[pos][3]);
1446
1447 /* Viewport Mapping */
1448 for (i=0; i<3; i++) {
1449 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1450 LLVMValueRef scale;
1451 LLVMValueRef trans;
1452 LLVMValueRef scale_i;
1453 LLVMValueRef trans_i;
1454 LLVMValueRef index;
1455
1456 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1457 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1458
1459 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1460 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1461
1462 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1463 LLVMBuildLoad(builder, scale_i, "scale"));
1464 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1465 LLVMBuildLoad(builder, trans_i, "trans"));
1466
1467 /* divide by w */
1468 out = LLVMBuildFMul(builder, out, out3, "");
1469 /* mult by scale, add translation */
1470 out = lp_build_fmuladd(builder, out, scale, trans);
1471
1472 /* store transformed outputs */
1473 LLVMBuildStore(builder, out, outputs[pos][i]);
1474 }
1475
1476 }
1477
1478
1479 /**
1480 * Returns clipmask as nxi32 bitmask for the n vertices
1481 */
1482 static LLVMValueRef
1483 generate_clipmask(struct draw_llvm *llvm,
1484 struct gallivm_state *gallivm,
1485 struct lp_type vs_type,
1486 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1487 struct draw_llvm_variant_key *key,
1488 LLVMValueRef context_ptr,
1489 boolean *have_clipdist)
1490 {
1491 LLVMBuilderRef builder = gallivm->builder;
1492 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1493 LLVMValueRef test, temp;
1494 LLVMValueRef zero, shift;
1495 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1496 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1497 LLVMValueRef plane1, planes, plane_ptr, sum;
1498 struct lp_type f32_type = vs_type;
1499 struct lp_type i32_type = lp_int_type(vs_type);
1500 const unsigned pos = llvm->draw->vs.position_output;
1501 const unsigned cv = llvm->draw->vs.clipvertex_output;
1502 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1503 boolean have_cd = false;
1504 boolean clip_user = key->clip_user;
1505 unsigned ucp_enable = key->ucp_enable;
1506 unsigned cd[2];
1507
1508 cd[0] = llvm->draw->vs.ccdistance_output[0];
1509 cd[1] = llvm->draw->vs.ccdistance_output[1];
1510
1511 if (cd[0] != pos || cd[1] != pos)
1512 have_cd = true;
1513
1514 if (num_written_clipdistance && !clip_user) {
1515 clip_user = true;
1516 ucp_enable = (1 << num_written_clipdistance) - 1;
1517 }
1518
1519 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1520 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1521 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1522 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1523
1524 /*
1525 * load clipvertex and position from correct locations.
1526 * if they are the same just load them once.
1527 */
1528 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1529 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1530 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1531 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1532
1533 if (clip_user && cv != pos) {
1534 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1535 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1536 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1537 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1538 } else {
1539 cv_x = pos_x;
1540 cv_y = pos_y;
1541 cv_z = pos_z;
1542 cv_w = pos_w;
1543 }
1544
1545 /*
1546 * Be careful with the comparisons and NaNs (using llvm's unordered
1547 * comparisons here).
1548 */
1549 /* Cliptest, for hardwired planes */
1550 /*
1551 * XXX should take guardband into account (currently not in key).
1552 * Otherwise might run the draw pipeline stages for nothing.
1553 */
1554 if (key->clip_xy) {
1555 /* plane 1 */
1556 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1557 temp = shift;
1558 test = LLVMBuildAnd(builder, test, temp, "");
1559 mask = test;
1560
1561 /* plane 2 */
1562 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1563 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1564 temp = LLVMBuildShl(builder, temp, shift, "");
1565 test = LLVMBuildAnd(builder, test, temp, "");
1566 mask = LLVMBuildOr(builder, mask, test, "");
1567
1568 /* plane 3 */
1569 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1570 temp = LLVMBuildShl(builder, temp, shift, "");
1571 test = LLVMBuildAnd(builder, test, temp, "");
1572 mask = LLVMBuildOr(builder, mask, test, "");
1573
1574 /* plane 4 */
1575 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1576 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1577 temp = LLVMBuildShl(builder, temp, shift, "");
1578 test = LLVMBuildAnd(builder, test, temp, "");
1579 mask = LLVMBuildOr(builder, mask, test, "");
1580 }
1581
1582 if (key->clip_z) {
1583 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1584 if (key->clip_halfz) {
1585 /* plane 5 */
1586 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1587 test = LLVMBuildAnd(builder, test, temp, "");
1588 mask = LLVMBuildOr(builder, mask, test, "");
1589 }
1590 else {
1591 /* plane 5 */
1592 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1593 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1594 test = LLVMBuildAnd(builder, test, temp, "");
1595 mask = LLVMBuildOr(builder, mask, test, "");
1596 }
1597 /* plane 6 */
1598 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1599 temp = LLVMBuildShl(builder, temp, shift, "");
1600 test = LLVMBuildAnd(builder, test, temp, "");
1601 mask = LLVMBuildOr(builder, mask, test, "");
1602 }
1603
1604 if (clip_user) {
1605 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1606 LLVMValueRef indices[3];
1607 LLVMValueRef is_nan_or_inf;
1608
1609 /* userclip planes */
1610 while (ucp_enable) {
1611 unsigned plane_idx = ffs(ucp_enable)-1;
1612 ucp_enable &= ~(1 << plane_idx);
1613 plane_idx += 6;
1614
1615 if (have_cd && num_written_clipdistance) {
1616 LLVMValueRef clipdist;
1617 int i;
1618 i = plane_idx - 6;
1619
1620 *have_clipdist = TRUE;
1621 if (i < 4) {
1622 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1623 } else {
1624 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1625 }
1626 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1627 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1628 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1629 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1630 test = LLVMBuildAnd(builder, test, temp, "");
1631 mask = LLVMBuildOr(builder, mask, test, "");
1632 } else {
1633 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1634 indices[0] = lp_build_const_int32(gallivm, 0);
1635 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1636
1637 indices[2] = lp_build_const_int32(gallivm, 0);
1638 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1639 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1640 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1641 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1642
1643 indices[2] = lp_build_const_int32(gallivm, 1);
1644 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1645 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1646 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1647 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1648
1649 indices[2] = lp_build_const_int32(gallivm, 2);
1650 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1651 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1652 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1653 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1654
1655 indices[2] = lp_build_const_int32(gallivm, 3);
1656 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1657 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1658 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1659 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1660
1661 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1662 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1663 test = LLVMBuildAnd(builder, test, temp, "");
1664 mask = LLVMBuildOr(builder, mask, test, "");
1665 }
1666 }
1667 }
1668 if (key->need_edgeflags) {
1669 /*
1670 * This isn't really part of clipmask but stored the same in vertex
1671 * header later, so do it here.
1672 */
1673 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1674 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1675 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1676 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1677 temp = lp_build_const_int_vec(gallivm, i32_type,
1678 1LL << DRAW_TOTAL_CLIP_PLANES);
1679 test = LLVMBuildAnd(builder, test, temp, "");
1680 mask = LLVMBuildOr(builder, mask, test, "");
1681 }
1682 return mask;
1683 }
1684
1685
1686 /**
1687 * Returns boolean if any clipping has occurred
1688 * Used zero/one i8 value to represent boolean
1689 */
1690 static LLVMValueRef
1691 clipmask_booli8(struct gallivm_state *gallivm,
1692 const struct lp_type vs_type,
1693 LLVMValueRef clipmask_bool_ptr,
1694 boolean edgeflag_in_clipmask)
1695 {
1696 LLVMBuilderRef builder = gallivm->builder;
1697 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1698 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1699 LLVMValueRef ret;
1700 struct lp_build_context bldivec;
1701
1702 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1703
1704 /*
1705 * We need to invert the edgeflag bit from the clipmask here
1706 * (because the result is really if we want to run the pipeline or not
1707 * and we (may) need it if edgeflag was 0).
1708 */
1709 if (edgeflag_in_clipmask) {
1710 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1711 1LL << DRAW_TOTAL_CLIP_PLANES);
1712 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1713 }
1714
1715 /*
1716 * XXX: probably should mask off bits from the mask which come from
1717 * vertices which were beyond the count (i.e. indices_valid for
1718 * linear fetches, for elts ones we don't have the correct mask
1719 * right now). Otherwise might run the pipeline for nothing,
1720 * though everything should still work.
1721 */
1722 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1723 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1724 return ret;
1725 }
1726
1727 static LLVMValueRef
1728 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface *gs_iface,
1729 struct lp_build_context * bld,
1730 boolean is_vindex_indirect,
1731 LLVMValueRef vertex_index,
1732 boolean is_aindex_indirect,
1733 LLVMValueRef attrib_index,
1734 LLVMValueRef swizzle_index)
1735 {
1736 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1737 struct gallivm_state *gallivm = bld->gallivm;
1738 LLVMBuilderRef builder = gallivm->builder;
1739 LLVMValueRef indices[3];
1740 LLVMValueRef res;
1741 struct lp_type type = bld->type;
1742
1743 if (is_vindex_indirect || is_aindex_indirect) {
1744 int i;
1745 res = bld->zero;
1746 for (i = 0; i < type.length; ++i) {
1747 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1748 LLVMValueRef vert_chan_index = vertex_index;
1749 LLVMValueRef attr_chan_index = attrib_index;
1750 LLVMValueRef channel_vec, value;
1751
1752 if (is_vindex_indirect) {
1753 vert_chan_index = LLVMBuildExtractElement(builder,
1754 vertex_index, idx, "");
1755 }
1756 if (is_aindex_indirect) {
1757 attr_chan_index = LLVMBuildExtractElement(builder,
1758 attrib_index, idx, "");
1759 }
1760
1761 indices[0] = vert_chan_index;
1762 indices[1] = attr_chan_index;
1763 indices[2] = swizzle_index;
1764
1765 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1766 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1767 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1768
1769 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1770 }
1771 } else {
1772 indices[0] = vertex_index;
1773 indices[1] = attrib_index;
1774 indices[2] = swizzle_index;
1775
1776 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1777 res = LLVMBuildLoad(builder, res, "");
1778 }
1779
1780 return res;
1781 }
1782
1783 static void
1784 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface *gs_base,
1785 struct lp_build_context * bld,
1786 LLVMValueRef (*outputs)[4],
1787 LLVMValueRef emitted_vertices_vec,
1788 LLVMValueRef stream_id)
1789 {
1790 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1791 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1792 struct gallivm_state *gallivm = variant->gallivm;
1793 LLVMBuilderRef builder = gallivm->builder;
1794 struct lp_type gs_type = bld->type;
1795 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1796 lp_int_type(gs_type), 0);
1797 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1798 LLVMValueRef next_prim_offset =
1799 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1800 LLVMValueRef io = variant->io_ptr;
1801 unsigned i;
1802 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1803
1804 for (i = 0; i < gs_type.length; ++i) {
1805 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1806 LLVMValueRef currently_emitted =
1807 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1808 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1809 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1810 }
1811
1812 LLVMValueRef stream_idx = LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), "");
1813 LLVMValueRef cnd = LLVMBuildICmp(builder, LLVMIntULT, stream_idx, lp_build_const_int32(gallivm, variant->shader->base.num_vertex_streams), "");
1814 struct lp_build_if_state if_ctx;
1815 lp_build_if(&if_ctx, gallivm, cnd);
1816 io = lp_build_pointer_get(builder, io, LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), ""));
1817
1818 convert_to_aos(gallivm, io, indices,
1819 outputs, clipmask,
1820 gs_info->num_outputs, gs_type,
1821 FALSE);
1822 lp_build_endif(&if_ctx);
1823 }
1824
1825 static void
1826 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface *gs_base,
1827 struct lp_build_context * bld,
1828 LLVMValueRef total_emitted_vertices_vec_ptr,
1829 LLVMValueRef verts_per_prim_vec,
1830 LLVMValueRef emitted_prims_vec,
1831 LLVMValueRef mask_vec)
1832 {
1833 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1834 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1835 struct gallivm_state *gallivm = variant->gallivm;
1836 LLVMBuilderRef builder = gallivm->builder;
1837 LLVMValueRef prim_lengts_ptr =
1838 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1839 unsigned i;
1840
1841 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
1842 for (i = 0; i < bld->type.length; ++i) {
1843 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1844 LLVMValueRef prims_emitted =
1845 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1846 LLVMValueRef store_ptr;
1847 LLVMValueRef num_vertices =
1848 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1849
1850 LLVMValueRef this_cond = LLVMBuildExtractElement(gallivm->builder, cond, ind, "");
1851 struct lp_build_if_state ifthen;
1852 lp_build_if(&ifthen, gallivm, this_cond);
1853 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1854 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1855 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1856 LLVMBuildStore(builder, num_vertices, store_ptr);
1857 lp_build_endif(&ifthen);
1858 }
1859 }
1860
1861 static void
1862 draw_gs_llvm_epilogue(const struct lp_build_gs_iface *gs_base,
1863 LLVMValueRef total_emitted_vertices_vec,
1864 LLVMValueRef emitted_prims_vec, unsigned stream)
1865 {
1866 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1867 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1868 struct gallivm_state *gallivm = variant->gallivm;
1869 LLVMBuilderRef builder = gallivm->builder;
1870 LLVMValueRef emitted_verts_ptr =
1871 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1872 LLVMValueRef emitted_prims_ptr =
1873 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1874 LLVMValueRef stream_val = lp_build_const_int32(gallivm, stream);
1875
1876 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &stream_val, 1, "");
1877 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &stream_val, 1, "");
1878
1879 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1880 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1881 }
1882
1883 static void
1884 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1885 {
1886 struct gallivm_state *gallivm = variant->gallivm;
1887 LLVMContextRef context = gallivm->context;
1888 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1889 LLVMTypeRef arg_types[12];
1890 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1891 LLVMTypeRef func_type;
1892 LLVMValueRef context_ptr;
1893 LLVMBasicBlockRef block;
1894 LLVMBuilderRef builder;
1895 char func_name[64];
1896 struct lp_type vs_type;
1897 LLVMValueRef count, fetch_elts, start_or_maxelt;
1898 LLVMValueRef vertex_id_offset;
1899 LLVMValueRef stride, step, io_itr;
1900 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1901 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1902 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1903 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1904 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1905 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1906 LLVMValueRef fake_buf_ptr, fake_buf;
1907
1908 struct draw_context *draw = llvm->draw;
1909 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1910 unsigned i, j;
1911 struct lp_build_context bld, blduivec;
1912 struct lp_build_loop_state lp_loop;
1913 struct lp_build_if_state if_ctx;
1914 const int vector_length = lp_native_vector_width / 32;
1915 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1916 struct lp_build_sampler_soa *sampler = 0;
1917 struct lp_build_image_soa *image = NULL;
1918 LLVMValueRef ret, clipmask_bool_ptr;
1919 struct draw_llvm_variant_key *key = &variant->key;
1920 /* If geometry shader is present we need to skip both the viewport
1921 * transformation and clipping otherwise the inputs to the geometry
1922 * shader will be incorrect.
1923 * The code can't handle vp transform when vs writes vp index neither
1924 * (though this would be fixable here, but couldn't just broadcast
1925 * the values).
1926 */
1927 const boolean bypass_viewport = key->has_gs_or_tes || key->bypass_viewport ||
1928 vs_info->writes_viewport_index;
1929 const boolean enable_cliptest = !key->has_gs_or_tes && (key->clip_xy ||
1930 key->clip_z ||
1931 key->clip_user ||
1932 key->need_edgeflags);
1933 LLVMValueRef variant_func;
1934 const unsigned pos = draw->vs.position_output;
1935 const unsigned cv = draw->vs.clipvertex_output;
1936 boolean have_clipdist = FALSE;
1937 struct lp_bld_tgsi_system_values system_values;
1938
1939 memset(&system_values, 0, sizeof(system_values));
1940 memset(&outputs, 0, sizeof(outputs));
1941 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant");
1942
1943 i = 0;
1944 arg_types[i++] = get_context_ptr_type(variant); /* context */
1945 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1946 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1947 arg_types[i++] = int32_type; /* count */
1948 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1949 arg_types[i++] = int32_type; /* stride */
1950 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1951 arg_types[i++] = int32_type; /* instance_id */
1952 arg_types[i++] = int32_type; /* vertex_id_offset */
1953 arg_types[i++] = int32_type; /* start_instance */
1954 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1955 arg_types[i++] = int32_type; /* draw_id */
1956
1957 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1958 arg_types, num_arg_types, 0);
1959
1960 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1961 variant->function = variant_func;
1962
1963 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1964 for (i = 0; i < num_arg_types; ++i)
1965 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1966 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1967
1968 if (gallivm->cache && gallivm->cache->data_size)
1969 return;
1970 context_ptr = LLVMGetParam(variant_func, 0);
1971 io_ptr = LLVMGetParam(variant_func, 1);
1972 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1973 count = LLVMGetParam(variant_func, 3);
1974 /*
1975 * XXX: the maxelt part is unused. Not really useful, since we cannot
1976 * get index buffer overflows due to vsplit (which provides its own
1977 * elts buffer, with a different size than what's passed in here).
1978 */
1979 start_or_maxelt = LLVMGetParam(variant_func, 4);
1980 /*
1981 * XXX: stride is actually unused. The stride we use is strictly calculated
1982 * from the number of outputs (including the draw_extra outputs).
1983 * Should probably fix some day (we need a new vs just because of extra
1984 * outputs which the generated vs won't touch).
1985 */
1986 stride = LLVMGetParam(variant_func, 5);
1987 vb_ptr = LLVMGetParam(variant_func, 6);
1988 system_values.instance_id = LLVMGetParam(variant_func, 7);
1989 vertex_id_offset = LLVMGetParam(variant_func, 8);
1990 system_values.base_instance = LLVMGetParam(variant_func, 9);
1991 fetch_elts = LLVMGetParam(variant_func, 10);
1992 system_values.draw_id = LLVMGetParam(variant_func, 11);
1993
1994 lp_build_name(context_ptr, "context");
1995 lp_build_name(io_ptr, "io");
1996 lp_build_name(vbuffers_ptr, "vbuffers");
1997 lp_build_name(count, "count");
1998 lp_build_name(start_or_maxelt, "start_or_maxelt");
1999 lp_build_name(stride, "stride");
2000 lp_build_name(vb_ptr, "vb");
2001 lp_build_name(system_values.instance_id, "instance_id");
2002 lp_build_name(vertex_id_offset, "vertex_id_offset");
2003 lp_build_name(system_values.base_instance, "start_instance");
2004 lp_build_name(fetch_elts, "fetch_elts");
2005 lp_build_name(system_values.draw_id, "draw_id");
2006
2007 /*
2008 * Function body
2009 */
2010
2011 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2012 builder = gallivm->builder;
2013 LLVMPositionBuilderAtEnd(builder, block);
2014
2015 memset(&vs_type, 0, sizeof vs_type);
2016 vs_type.floating = TRUE; /* floating point values */
2017 vs_type.sign = TRUE; /* values are signed */
2018 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2019 vs_type.width = 32; /* 32-bit float */
2020 vs_type.length = vector_length;
2021
2022 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
2023 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
2024
2025 /* hold temporary "bool" clipmask */
2026 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
2027
2028 fake_buf = lp_build_alloca_undef(gallivm,
2029 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
2030 fake_buf = LLVMBuildBitCast(builder, fake_buf,
2031 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
2032 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
2033
2034 /* code generated texture sampling */
2035 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key), key->nr_samplers);
2036
2037 image = draw_llvm_image_soa_create(draw_llvm_variant_key_images(key));
2038
2039 step = lp_build_const_int32(gallivm, vector_length);
2040
2041 ind_vec = blduivec.undef;
2042 for (i = 0; i < vs_type.length; i++) {
2043 LLVMValueRef index = lp_build_const_int32(gallivm, i);
2044 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
2045 }
2046
2047 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
2048 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
2049
2050 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
2051 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
2052 /*
2053 * Only needed for non-indexed path.
2054 */
2055 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
2056
2057 /*
2058 * Pre-calculate everything which is constant per shader invocation.
2059 */
2060 for (j = 0; j < key->nr_vertex_elements; ++j) {
2061 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
2062 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
2063 struct pipe_vertex_element *velem = &key->vertex_element[j];
2064 LLVMValueRef vb_index =
2065 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
2066 LLVMValueRef bsize = lp_build_const_int32(gallivm,
2067 util_format_get_blocksize(velem->src_format));
2068 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
2069 velem->src_offset);
2070 struct lp_build_if_state if_ctx;
2071
2072 if (velem->src_format != PIPE_FORMAT_NONE) {
2073 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
2074 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
2075 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
2076 vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
2077 LLVMInt32TypeInContext(context), "");
2078 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
2079 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
2080 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
2081
2082 ofbit = NULL;
2083 /*
2084 * We'll set buffer_size_adj to zero if we have of, so it will
2085 * always overflow later automatically without having to keep ofbit.
2086 * Overflows (with normal wraparound) doing the actual offset
2087 * calculation should be ok, just not for the buffer size calc.
2088 * It would also be possible to detect such overflows and return
2089 * zeros if that happens, but this would be more complex.
2090 */
2091 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
2092 tmp = lp_build_sub(&bld, bsize, bld.one);
2093 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
2094 &ofbit);
2095 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
2096 buf_offset, &ofbit);
2097
2098 /*
2099 * We can't easily set fake vertex buffers outside the generated code.
2100 * Hence, set fake vertex buffers here instead basically, so fetch
2101 * code can always fetch using offset 0, eliminating all control flow
2102 * inside the main loop.
2103 * (Alternatively, could have control flow per vector skipping fetch
2104 * if ofbit is true.)
2105 */
2106 if (velem->instance_divisor) {
2107 /*
2108 * Index is equal to the start instance plus the number of current
2109 * instance divided by the divisor. In this case we compute it as:
2110 * index = start_instance + (instance_id / divisor).
2111 * Note we could actually do the fetch here, outside the loop -
2112 * it's all constant, hopefully llvm recognizes this.
2113 */
2114 LLVMValueRef current_instance;
2115 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
2116 lp_build_const_int32(gallivm,
2117 velem->instance_divisor),
2118 "instance_divisor");
2119 instance_index[j] = lp_build_uadd_overflow(gallivm, system_values.base_instance,
2120 current_instance, &ofbit);
2121 }
2122
2123 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
2124 buffer_size_adj[j], "");
2125
2126 temp_ptr = lp_build_alloca_undef(gallivm,
2127 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
2128
2129 lp_build_if(&if_ctx, gallivm, ofbit);
2130 {
2131 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
2132 }
2133 lp_build_else(&if_ctx);
2134 {
2135 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
2136 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
2137 }
2138 lp_build_endif(&if_ctx);
2139 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
2140
2141 if (0) {
2142 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
2143 lp_build_const_int32(gallivm, j),
2144 vb_index, vb_stride[j]);
2145 lp_build_printf(gallivm,
2146 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
2147 vb_buffer_offset, src_offset, buf_offset);
2148 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
2149 buffer_size, bsize);
2150 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
2151 }
2152 }
2153 }
2154
2155 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
2156 {
2157 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
2158 LLVMValueRef io;
2159 LLVMValueRef clipmask; /* holds the clipmask value */
2160 LLVMValueRef true_index_array, index_store;
2161 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
2162
2163 io_itr = lp_loop.counter;
2164
2165 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
2166 #if DEBUG_STORE
2167 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
2168 io_itr, io, lp_loop.counter);
2169 #endif
2170
2171 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
2172 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
2173
2174 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
2175 /*
2176 * Limit indices to fetch_max, otherwise might try to access indices
2177 * beyond index buffer (or rather vsplit elt buffer) size.
2178 * Could probably safely (?) skip this for non-indexed draws and
2179 * simplify things minimally (by removing it could combine the ind_vec
2180 * and start_vec adds). I think the only effect for non-indexed draws will
2181 * be that for the invalid elements they will be all fetched from the
2182 * same location as the last valid one, but noone should really care.
2183 */
2184 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
2185
2186 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
2187
2188 lp_build_if(&if_ctx, gallivm, have_elts);
2189 {
2190 /*
2191 * Note: you'd expect some comparison/clamp against fetch_elt_max
2192 * here.
2193 * There used to be one here but it was incorrect: overflow was
2194 * detected if index > fetch_elt_max - but the correct condition
2195 * would be index >= fetch_elt_max (since this is just size of elts
2196 * buffer / element size).
2197 * Using the correct condition however will cause failures - due to
2198 * vsplit/vcache code which rebases indices. So, as an example, if
2199 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
2200 * replace all invalid indices with 0 - which in case of elt_bias
2201 * not being zero will get a different fetch index than the valid
2202 * index 0. So, just rely on vsplit code preventing out-of-bounds
2203 * fetches. This is also why it's safe to do elts fetch even if there
2204 * was no index buffer bound - the real buffer is never seen here, at
2205 * least not if there are index buffer overflows...
2206 */
2207
2208 /*
2209 * XXX should not have to do this, as scale can be handled
2210 * natively by loads (hits asserts though).
2211 */
2212 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
2213 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
2214 LLVMPointerType(LLVMInt8TypeInContext(context),
2215 0), "");
2216 tmp = lp_build_gather(gallivm, vs_type.length,
2217 32, bld.type, TRUE,
2218 fetch_elts, tmp, FALSE);
2219 LLVMBuildStore(builder, tmp, index_store);
2220 }
2221 lp_build_else(&if_ctx);
2222 {
2223 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
2224 LLVMBuildStore(builder, tmp, index_store);
2225 }
2226 lp_build_endif(&if_ctx);
2227
2228 true_index_array = LLVMBuildLoad(builder, index_store, "");
2229
2230 for (j = 0; j < key->nr_vertex_elements; ++j) {
2231 struct pipe_vertex_element *velem = &key->vertex_element[j];
2232 const struct util_format_description *format_desc =
2233 util_format_description(velem->src_format);
2234
2235 if (format_desc->format == PIPE_FORMAT_NONE) {
2236 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2237 inputs[j][i] = lp_build_zero(gallivm, vs_type);
2238 }
2239 }
2240 else if (velem->instance_divisor) {
2241 fetch_instanced(gallivm, format_desc, vs_type,
2242 vb_stride[j], map_ptr[j],
2243 buffer_size_adj[j],
2244 inputs[j], instance_index[j]);
2245 }
2246 else {
2247 fetch_vector(gallivm, format_desc, vs_type,
2248 vb_stride[j], map_ptr[j],
2249 buffer_size_adj[j],
2250 inputs[j], true_index_array);
2251 }
2252 }
2253
2254 struct lp_build_mask_context mask;
2255
2256 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
2257 /* In the paths with elts vertex id has to be unaffected by the
2258 * index bias and because indices inside our elements array have
2259 * already had index bias applied we need to subtract it here to
2260 * get back to the original index.
2261 * in the linear paths vertex id has to be unaffected by the
2262 * original start index and because we abuse the 'start' variable
2263 * to either represent the actual start index or the index at which
2264 * the primitive was split (we split rendering into chunks of at
2265 * most 4095-vertices) we need to back out the original start
2266 * index out of our vertex id here.
2267 * for ARB_shader_draw_parameters, base_vertex should be 0 for non-indexed draws.
2268 */
2269 LLVMValueRef base_vertex = lp_build_select(&bld, have_elts, vertex_id_offset, lp_build_const_int32(gallivm, 0));;
2270 system_values.basevertex = lp_build_broadcast_scalar(&blduivec, base_vertex);
2271 system_values.vertex_id = true_index_array;
2272 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
2273 lp_build_broadcast_scalar(&blduivec, vertex_id_offset), "");
2274
2275 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
2276 generate_vs(variant,
2277 builder,
2278 vs_type,
2279 outputs,
2280 ptr_aos,
2281 &system_values,
2282 context_ptr,
2283 sampler,
2284 image,
2285 key->clamp_vertex_color,
2286 &mask);
2287
2288 lp_build_mask_end(&mask);
2289 if (pos != -1 && cv != -1) {
2290 /* store original positions in clip before further manipulation */
2291 store_clip(gallivm, vs_type, io, outputs, pos);
2292
2293 /* do cliptest */
2294 if (enable_cliptest) {
2295 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
2296 /* allocate clipmask, assign it integer type */
2297 clipmask = generate_clipmask(llvm,
2298 gallivm,
2299 vs_type,
2300 outputs,
2301 key,
2302 context_ptr, &have_clipdist);
2303 temp = LLVMBuildOr(builder, clipmask, temp, "");
2304 /* store temporary clipping boolean value */
2305 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
2306 }
2307 else {
2308 clipmask = blduivec.zero;
2309 }
2310
2311 /* do viewport mapping */
2312 if (!bypass_viewport) {
2313 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
2314 }
2315 }
2316 else {
2317 clipmask = blduivec.zero;
2318 }
2319
2320 /* store clipmask in vertex header,
2321 * original positions in clip
2322 * and transformed positions in data
2323 */
2324 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
2325 vs_info->num_outputs, vs_type,
2326 enable_cliptest && key->need_edgeflags);
2327 }
2328 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
2329
2330 sampler->destroy(sampler);
2331 image->destroy(image);
2332
2333 /* return clipping boolean value for function */
2334 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
2335 enable_cliptest && key->need_edgeflags);
2336
2337 LLVMBuildRet(builder, ret);
2338
2339 gallivm_verify_function(gallivm, variant_func);
2340 }
2341
2342
2343 struct draw_llvm_variant_key *
2344 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2345 {
2346 unsigned i;
2347 struct draw_llvm_variant_key *key;
2348 struct draw_sampler_static_state *draw_sampler;
2349 struct draw_image_static_state *draw_image;
2350
2351 key = (struct draw_llvm_variant_key *)store;
2352
2353 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2354
2355 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
2356
2357 /* will have to rig this up properly later */
2358 key->clip_xy = llvm->draw->clip_xy;
2359 key->clip_z = llvm->draw->clip_z;
2360 key->clip_user = llvm->draw->clip_user;
2361 key->bypass_viewport = llvm->draw->bypass_viewport;
2362 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2363 /* XXX assumes edgeflag output not at 0 */
2364 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2365 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2366 key->has_gs_or_tes = llvm->draw->gs.geometry_shader != NULL || llvm->draw->tes.tess_eval_shader != NULL;
2367 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2368
2369 /* All variants of this shader will have the same value for
2370 * nr_samplers. Not yet trying to compact away holes in the
2371 * sampler array.
2372 */
2373 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2374 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2375 key->nr_sampler_views =
2376 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2377 }
2378 else {
2379 key->nr_sampler_views = key->nr_samplers;
2380 }
2381
2382 key->nr_images = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2383
2384 /* Presumably all variants of the shader should have the same
2385 * number of vertex elements - ie the number of shader inputs.
2386 * NOTE: we NEED to store the needed number of needed inputs
2387 * here, not the number of provided elements to match keysize
2388 * (and the offset of sampler state in the key).
2389 * If we have excess number of vertex elements, this is valid,
2390 * but the excess ones don't matter.
2391 * If we don't have enough vertex elements (which looks not really
2392 * valid but we'll handle it gracefully) fill out missing ones with
2393 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2394 */
2395 key->nr_vertex_elements =
2396 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2397
2398 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2399 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2400 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2401 memset(key->vertex_element, 0,
2402 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2403 }
2404 memcpy(key->vertex_element,
2405 llvm->draw->pt.vertex_element,
2406 sizeof(struct pipe_vertex_element) *
2407 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2408
2409 draw_sampler = draw_llvm_variant_key_samplers(key);
2410 memset(draw_sampler, 0,
2411 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2412
2413 for (i = 0 ; i < key->nr_samplers; i++) {
2414 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2415 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2416 }
2417 for (i = 0 ; i < key->nr_sampler_views; i++) {
2418 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2419 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2420 }
2421
2422 draw_image = draw_llvm_variant_key_images(key);
2423 memset(draw_image, 0,
2424 key->nr_images * sizeof *draw_image);
2425 for (i = 0; i < key->nr_images; i++) {
2426 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2427 llvm->draw->images[PIPE_SHADER_VERTEX][i]);
2428 }
2429 return key;
2430 }
2431
2432
2433 void
2434 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2435 {
2436 unsigned i;
2437 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2438 struct draw_image_static_state *image = draw_llvm_variant_key_images(key);
2439 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2440 debug_printf("clip_xy = %u\n", key->clip_xy);
2441 debug_printf("clip_z = %u\n", key->clip_z);
2442 debug_printf("clip_user = %u\n", key->clip_user);
2443 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2444 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2445 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2446 debug_printf("has_gs_or_tes = %u\n", key->has_gs_or_tes);
2447 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2448
2449 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2450 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2451 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2452 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2453 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2454 }
2455
2456 for (i = 0 ; i < key->nr_sampler_views; i++) {
2457 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2458 }
2459
2460 for (i = 0 ; i < key->nr_images; i++)
2461 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2462 }
2463
2464
2465 void
2466 draw_llvm_set_mapped_texture(struct draw_context *draw,
2467 enum pipe_shader_type shader_stage,
2468 unsigned sview_idx,
2469 uint32_t width, uint32_t height, uint32_t depth,
2470 uint32_t first_level, uint32_t last_level,
2471 uint32_t num_samples,
2472 uint32_t sample_stride,
2473 const void *base_ptr,
2474 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2475 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2476 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2477 {
2478 unsigned j;
2479 struct draw_jit_texture *jit_tex;
2480
2481 switch (shader_stage) {
2482 case PIPE_SHADER_VERTEX:
2483 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2484 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2485 break;
2486 case PIPE_SHADER_GEOMETRY:
2487 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2488 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2489 break;
2490 case PIPE_SHADER_TESS_CTRL:
2491 assert(sview_idx < ARRAY_SIZE(draw->llvm->tcs_jit_context.textures));
2492 jit_tex = &draw->llvm->tcs_jit_context.textures[sview_idx];
2493 break;
2494 case PIPE_SHADER_TESS_EVAL:
2495 assert(sview_idx < ARRAY_SIZE(draw->llvm->tes_jit_context.textures));
2496 jit_tex = &draw->llvm->tes_jit_context.textures[sview_idx];
2497 break;
2498 default:
2499 assert(0);
2500 return;
2501 }
2502
2503 jit_tex->width = width;
2504 jit_tex->height = height;
2505 jit_tex->depth = depth;
2506 jit_tex->first_level = first_level;
2507 jit_tex->last_level = last_level;
2508 jit_tex->base = base_ptr;
2509 jit_tex->num_samples = num_samples;
2510 jit_tex->sample_stride = sample_stride;
2511
2512 for (j = first_level; j <= last_level; j++) {
2513 jit_tex->mip_offsets[j] = mip_offsets[j];
2514 jit_tex->row_stride[j] = row_stride[j];
2515 jit_tex->img_stride[j] = img_stride[j];
2516 }
2517 }
2518
2519 void
2520 draw_llvm_set_mapped_image(struct draw_context *draw,
2521 enum pipe_shader_type shader_stage,
2522 unsigned idx,
2523 uint32_t width, uint32_t height, uint32_t depth,
2524 const void *base_ptr,
2525 uint32_t row_stride,
2526 uint32_t img_stride,
2527 uint32_t num_samples,
2528 uint32_t sample_stride)
2529 {
2530 struct draw_jit_image *jit_image;
2531
2532 switch (shader_stage) {
2533 case PIPE_SHADER_VERTEX:
2534 assert(idx < ARRAY_SIZE(draw->llvm->jit_context.images));
2535 jit_image = &draw->llvm->jit_context.images[idx];
2536 break;
2537 case PIPE_SHADER_GEOMETRY:
2538 assert(idx < ARRAY_SIZE(draw->llvm->gs_jit_context.images));
2539 jit_image = &draw->llvm->gs_jit_context.images[idx];
2540 break;
2541 case PIPE_SHADER_TESS_CTRL:
2542 assert(idx < ARRAY_SIZE(draw->llvm->tcs_jit_context.images));
2543 jit_image = &draw->llvm->tcs_jit_context.images[idx];
2544 break;
2545 case PIPE_SHADER_TESS_EVAL:
2546 assert(idx < ARRAY_SIZE(draw->llvm->tes_jit_context.images));
2547 jit_image = &draw->llvm->tes_jit_context.images[idx];
2548 break;
2549 default:
2550 assert(0);
2551 return;
2552 }
2553
2554 jit_image->width = width;
2555 jit_image->height = height;
2556 jit_image->depth = depth;
2557 jit_image->base = base_ptr;
2558
2559 jit_image->row_stride = row_stride;
2560 jit_image->img_stride = img_stride;
2561 jit_image->num_samples = num_samples;
2562 jit_image->sample_stride = sample_stride;
2563 }
2564
2565
2566 void
2567 draw_llvm_set_sampler_state(struct draw_context *draw,
2568 enum pipe_shader_type shader_type)
2569 {
2570 unsigned i;
2571
2572 switch (shader_type) {
2573 case PIPE_SHADER_VERTEX:
2574 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2575 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2576
2577 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2578 const struct pipe_sampler_state *s
2579 = draw->samplers[PIPE_SHADER_VERTEX][i];
2580 jit_sam->min_lod = s->min_lod;
2581 jit_sam->max_lod = s->max_lod;
2582 jit_sam->lod_bias = s->lod_bias;
2583 COPY_4V(jit_sam->border_color, s->border_color.f);
2584 }
2585 }
2586 break;
2587 case PIPE_SHADER_GEOMETRY:
2588 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2589 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2590
2591 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2592 const struct pipe_sampler_state *s
2593 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2594 jit_sam->min_lod = s->min_lod;
2595 jit_sam->max_lod = s->max_lod;
2596 jit_sam->lod_bias = s->lod_bias;
2597 COPY_4V(jit_sam->border_color, s->border_color.f);
2598 }
2599 }
2600 break;
2601 case PIPE_SHADER_TESS_CTRL:
2602 for (i = 0; i < draw->num_samplers[PIPE_SHADER_TESS_CTRL]; i++) {
2603 struct draw_jit_sampler *jit_sam = &draw->llvm->tcs_jit_context.samplers[i];
2604
2605 if (draw->samplers[PIPE_SHADER_TESS_CTRL][i]) {
2606 const struct pipe_sampler_state *s
2607 = draw->samplers[PIPE_SHADER_TESS_CTRL][i];
2608 jit_sam->min_lod = s->min_lod;
2609 jit_sam->max_lod = s->max_lod;
2610 jit_sam->lod_bias = s->lod_bias;
2611 COPY_4V(jit_sam->border_color, s->border_color.f);
2612 }
2613 }
2614 break;
2615 case PIPE_SHADER_TESS_EVAL:
2616 for (i = 0; i < draw->num_samplers[PIPE_SHADER_TESS_EVAL]; i++) {
2617 struct draw_jit_sampler *jit_sam = &draw->llvm->tes_jit_context.samplers[i];
2618
2619 if (draw->samplers[PIPE_SHADER_TESS_EVAL][i]) {
2620 const struct pipe_sampler_state *s
2621 = draw->samplers[PIPE_SHADER_TESS_EVAL][i];
2622 jit_sam->min_lod = s->min_lod;
2623 jit_sam->max_lod = s->max_lod;
2624 jit_sam->lod_bias = s->lod_bias;
2625 COPY_4V(jit_sam->border_color, s->border_color.f);
2626 }
2627 }
2628 break;
2629 default:
2630 assert(0);
2631 break;
2632 }
2633 }
2634
2635
2636 void
2637 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2638 {
2639 struct draw_llvm *llvm = variant->llvm;
2640
2641 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2642 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2643 variant->shader->variants_cached, llvm->nr_variants);
2644 }
2645
2646 gallivm_destroy(variant->gallivm);
2647
2648 remove_from_list(&variant->list_item_local);
2649 variant->shader->variants_cached--;
2650 remove_from_list(&variant->list_item_global);
2651 llvm->nr_variants--;
2652 FREE(variant);
2653 }
2654
2655
2656 /**
2657 * Create LLVM types for various structures.
2658 */
2659 static void
2660 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2661 {
2662 struct gallivm_state *gallivm = var->gallivm;
2663 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
2664
2665 texture_type = create_jit_texture_type(gallivm, "texture");
2666 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2667 image_type = create_jit_image_type(gallivm, "image");
2668
2669 context_type = create_gs_jit_context_type(gallivm,
2670 var->shader->base.vector_length,
2671 texture_type, sampler_type,
2672 image_type,
2673 "draw_gs_jit_context");
2674 var->context_ptr_type = LLVMPointerType(context_type, 0);
2675
2676 var->input_array_type = create_gs_jit_input_type(gallivm);
2677 }
2678
2679 static LLVMTypeRef
2680 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2681 {
2682 if (!variant->context_ptr_type)
2683 create_gs_jit_types(variant);
2684 return variant->context_ptr_type;
2685 }
2686
2687 static LLVMValueRef
2688 generate_mask_value(struct draw_gs_llvm_variant *variant,
2689 struct lp_type gs_type)
2690 {
2691 struct gallivm_state *gallivm = variant->gallivm;
2692 LLVMBuilderRef builder = gallivm->builder;
2693 struct lp_type mask_type = lp_int_type(gs_type);
2694 LLVMValueRef num_prims;
2695 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2696 unsigned i;
2697
2698 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2699 variant->num_prims);
2700 for (i = 0; i < gs_type.length; i++) {
2701 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2702 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2703 }
2704 mask_val = lp_build_compare(gallivm, mask_type,
2705 PIPE_FUNC_GREATER, num_prims, mask_val);
2706
2707 return mask_val;
2708 }
2709
2710 static void
2711 draw_gs_llvm_generate(struct draw_llvm *llvm,
2712 struct draw_gs_llvm_variant *variant)
2713 {
2714 struct gallivm_state *gallivm = variant->gallivm;
2715 LLVMContextRef context = gallivm->context;
2716 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2717 LLVMTypeRef arg_types[7];
2718 LLVMTypeRef func_type;
2719 LLVMValueRef variant_func;
2720 LLVMValueRef context_ptr;
2721 LLVMValueRef prim_id_ptr;
2722 LLVMBasicBlockRef block;
2723 LLVMBuilderRef builder;
2724 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2725 struct lp_build_sampler_soa *sampler = 0;
2726 struct lp_build_image_soa *image = NULL;
2727 struct lp_build_context bld;
2728 struct lp_bld_tgsi_system_values system_values;
2729 char func_name[64];
2730 struct lp_type gs_type;
2731 unsigned i;
2732 struct draw_gs_llvm_iface gs_iface;
2733 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2734 LLVMValueRef consts_ptr, num_consts_ptr;
2735 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
2736 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2737 struct lp_build_mask_context mask;
2738 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2739 unsigned vector_length = variant->shader->base.vector_length;
2740
2741 memset(&system_values, 0, sizeof(system_values));
2742 memset(&outputs, 0, sizeof(outputs));
2743
2744 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant");
2745
2746 assert(variant->vertex_header_ptr_type);
2747
2748 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2749 arg_types[1] = variant->input_array_type; /* input */
2750 arg_types[2] = LLVMPointerType(variant->vertex_header_ptr_type, 0); /* vertex_header */
2751 arg_types[3] = int32_type; /* num_prims */
2752 arg_types[4] = int32_type; /* instance_id */
2753 arg_types[5] = LLVMPointerType(
2754 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2755 arg_types[6] = int32_type;
2756
2757 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2758
2759 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2760
2761 variant->function = variant_func;
2762
2763 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2764
2765 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2766 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2767 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2768
2769 if (gallivm->cache && gallivm->cache->data_size)
2770 return;
2771 context_ptr = LLVMGetParam(variant_func, 0);
2772 input_array = LLVMGetParam(variant_func, 1);
2773 io_ptr = LLVMGetParam(variant_func, 2);
2774 num_prims = LLVMGetParam(variant_func, 3);
2775 system_values.instance_id = LLVMGetParam(variant_func, 4);
2776 prim_id_ptr = LLVMGetParam(variant_func, 5);
2777 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2778
2779 lp_build_name(context_ptr, "context");
2780 lp_build_name(input_array, "input");
2781 lp_build_name(io_ptr, "io");
2782 lp_build_name(num_prims, "num_prims");
2783 lp_build_name(system_values.instance_id, "instance_id");
2784 lp_build_name(prim_id_ptr, "prim_id_ptr");
2785 lp_build_name(system_values.invocation_id, "invocation_id");
2786
2787 variant->context_ptr = context_ptr;
2788 variant->io_ptr = io_ptr;
2789 variant->num_prims = num_prims;
2790
2791 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2792 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2793 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2794 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2795 gs_iface.input = input_array;
2796 gs_iface.variant = variant;
2797
2798 /*
2799 * Function body
2800 */
2801
2802 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2803 builder = gallivm->builder;
2804 LLVMPositionBuilderAtEnd(builder, block);
2805
2806 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2807
2808 memset(&gs_type, 0, sizeof gs_type);
2809 gs_type.floating = TRUE; /* floating point values */
2810 gs_type.sign = TRUE; /* values are signed */
2811 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2812 gs_type.width = 32; /* 32-bit float */
2813 gs_type.length = vector_length;
2814
2815 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2816 num_consts_ptr =
2817 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2818
2819 ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
2820 num_ssbos_ptr =
2821 draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
2822
2823 /* code generated texture sampling */
2824 sampler = draw_llvm_sampler_soa_create(variant->key.samplers, variant->key.nr_samplers);
2825 image = draw_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant->key));
2826 mask_val = generate_mask_value(variant, gs_type);
2827 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2828
2829 if (gs_info->uses_primid) {
2830 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2831 }
2832
2833 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2834 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2835 tgsi_dump(tokens, 0);
2836 else
2837 nir_print_shader(llvm->draw->gs.geometry_shader->state.ir.nir, stderr);
2838 draw_gs_llvm_dump_variant_key(&variant->key);
2839 }
2840
2841 struct lp_build_tgsi_params params;
2842 memset(&params, 0, sizeof(params));
2843
2844 params.type = gs_type;
2845 params.mask = &mask;
2846 params.consts_ptr = consts_ptr;
2847 params.const_sizes_ptr = num_consts_ptr;
2848 params.system_values = &system_values;
2849 params.context_ptr = context_ptr;
2850 params.sampler = sampler;
2851 params.info = &llvm->draw->gs.geometry_shader->info;
2852 params.gs_iface = (const struct lp_build_gs_iface *)&gs_iface;
2853 params.ssbo_ptr = ssbos_ptr;
2854 params.ssbo_sizes_ptr = num_ssbos_ptr;
2855 params.image = image;
2856
2857 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2858 lp_build_tgsi_soa(variant->gallivm,
2859 tokens,
2860 &params,
2861 outputs);
2862 else
2863 lp_build_nir_soa(variant->gallivm,
2864 llvm->draw->gs.geometry_shader->state.ir.nir,
2865 &params,
2866 outputs);
2867
2868 sampler->destroy(sampler);
2869 image->destroy(image);
2870
2871 lp_build_mask_end(&mask);
2872
2873 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2874
2875 gallivm_verify_function(gallivm, variant_func);
2876 }
2877
2878 struct draw_gs_llvm_variant *
2879 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2880 unsigned num_outputs,
2881 const struct draw_gs_llvm_variant_key *key)
2882 {
2883 struct draw_gs_llvm_variant *variant;
2884 struct llvm_geometry_shader *shader =
2885 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2886 LLVMTypeRef vertex_header;
2887 char module_name[64];
2888 unsigned char ir_sha1_cache_key[20];
2889 struct lp_cached_code cached = { 0 };
2890 bool needs_caching = false;
2891
2892 variant = MALLOC(sizeof *variant +
2893 shader->variant_key_size -
2894 sizeof variant->key);
2895 if (!variant)
2896 return NULL;
2897
2898 variant->llvm = llvm;
2899 variant->shader = shader;
2900
2901 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2902 variant->shader->variants_cached);
2903
2904 memcpy(&variant->key, key, shader->variant_key_size);
2905
2906 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
2907 draw_get_ir_cache_key(shader->base.state.ir.nir,
2908 key,
2909 shader->variant_key_size,
2910 num_outputs,
2911 ir_sha1_cache_key);
2912
2913 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
2914 &cached,
2915 ir_sha1_cache_key);
2916 if (!cached.data_size)
2917 needs_caching = true;
2918 }
2919 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
2920
2921 create_gs_jit_types(variant);
2922
2923 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2924
2925 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2926
2927 draw_gs_llvm_generate(llvm, variant);
2928
2929 gallivm_compile_module(variant->gallivm);
2930
2931 variant->jit_func = (draw_gs_jit_func)
2932 gallivm_jit_function(variant->gallivm, variant->function);
2933
2934 if (needs_caching)
2935 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
2936 &cached,
2937 ir_sha1_cache_key);
2938 gallivm_free_ir(variant->gallivm);
2939
2940 variant->list_item_global.base = variant;
2941 variant->list_item_local.base = variant;
2942 /*variant->no = */shader->variants_created++;
2943 variant->list_item_global.base = variant;
2944
2945 return variant;
2946 }
2947
2948 void
2949 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2950 {
2951 struct draw_llvm *llvm = variant->llvm;
2952
2953 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2954 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2955 variant->shader->variants_cached, llvm->nr_gs_variants);
2956 }
2957
2958 gallivm_destroy(variant->gallivm);
2959
2960 remove_from_list(&variant->list_item_local);
2961 variant->shader->variants_cached--;
2962 remove_from_list(&variant->list_item_global);
2963 llvm->nr_gs_variants--;
2964 FREE(variant);
2965 }
2966
2967 struct draw_gs_llvm_variant_key *
2968 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2969 {
2970 unsigned i;
2971 struct draw_gs_llvm_variant_key *key;
2972 struct draw_sampler_static_state *draw_sampler;
2973 struct draw_image_static_state *draw_image;
2974
2975 key = (struct draw_gs_llvm_variant_key *)store;
2976
2977 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2978
2979 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2980
2981 /* All variants of this shader will have the same value for
2982 * nr_samplers. Not yet trying to compact away holes in the
2983 * sampler array.
2984 */
2985 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2986 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2987 key->nr_sampler_views =
2988 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2989 }
2990 else {
2991 key->nr_sampler_views = key->nr_samplers;
2992 }
2993
2994 key->nr_images = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2995
2996 draw_sampler = key->samplers;
2997
2998 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2999
3000 for (i = 0 ; i < key->nr_samplers; i++) {
3001 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
3002 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
3003 }
3004 for (i = 0 ; i < key->nr_sampler_views; i++) {
3005 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
3006 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
3007 }
3008
3009 draw_image = draw_gs_llvm_variant_key_images(key);
3010 memset(draw_image, 0,
3011 key->nr_images * sizeof *draw_image);
3012 for (i = 0; i < key->nr_images; i++) {
3013 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
3014 llvm->draw->images[PIPE_SHADER_GEOMETRY][i]);
3015 }
3016 return key;
3017 }
3018
3019 void
3020 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
3021 {
3022 unsigned i;
3023 struct draw_sampler_static_state *sampler = key->samplers;
3024 struct draw_image_static_state *image = draw_gs_llvm_variant_key_images(key);
3025 for (i = 0 ; i < key->nr_sampler_views; i++) {
3026 debug_printf("sampler[%i].src_format = %s\n", i,
3027 util_format_name(sampler[i].texture_state.format));
3028 }
3029
3030 for (i = 0 ; i < key->nr_images; i++)
3031 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
3032
3033 }
3034
3035 static void
3036 create_tcs_jit_types(struct draw_tcs_llvm_variant *var)
3037 {
3038 struct gallivm_state *gallivm = var->gallivm;
3039 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
3040
3041 texture_type = create_jit_texture_type(gallivm, "texture");
3042 sampler_type = create_jit_sampler_type(gallivm, "sampler");
3043 image_type = create_jit_image_type(gallivm, "image");
3044
3045 context_type = create_tcs_jit_context_type(gallivm,
3046 0,
3047 texture_type, sampler_type,
3048 image_type,
3049 "draw_tcs_jit_context");
3050 var->input_array_type = create_tcs_jit_input_type(gallivm);
3051 var->output_array_type = create_tcs_jit_output_type(gallivm);
3052 var->context_ptr_type = LLVMPointerType(context_type, 0);
3053 }
3054
3055 static LLVMTypeRef
3056 get_tcs_context_ptr_type(struct draw_tcs_llvm_variant *variant)
3057 {
3058 if (!variant->context_ptr_type)
3059 create_tcs_jit_types(variant);
3060 return variant->context_ptr_type;
3061 }
3062
3063 static LLVMValueRef
3064 draw_tcs_llvm_emit_fetch_input(const struct lp_build_tcs_iface *tes_iface,
3065 struct lp_build_context *bld,
3066 boolean is_vindex_indirect,
3067 LLVMValueRef vertex_index,
3068 boolean is_aindex_indirect,
3069 LLVMValueRef attrib_index,
3070 LLVMValueRef swizzle_index)
3071 {
3072 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
3073 struct gallivm_state *gallivm = bld->gallivm;
3074 LLVMBuilderRef builder = gallivm->builder;
3075 LLVMValueRef indices[3];
3076 LLVMValueRef res;
3077 struct lp_type type = bld->type;
3078
3079 if (is_vindex_indirect || is_aindex_indirect) {
3080 int i;
3081
3082 res = bld->zero;
3083 for (i = 0; i < type.length; ++i) {
3084 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3085 LLVMValueRef vert_chan_index = vertex_index;
3086 LLVMValueRef attr_chan_index = attrib_index;
3087 LLVMValueRef channel_vec;
3088
3089 if (is_vindex_indirect) {
3090 vert_chan_index = LLVMBuildExtractElement(builder,
3091 vertex_index, idx, "");
3092 }
3093 if (is_aindex_indirect) {
3094 attr_chan_index = LLVMBuildExtractElement(builder,
3095 attrib_index, idx, "");
3096 }
3097
3098 indices[0] = vert_chan_index;
3099 indices[1] = attr_chan_index;
3100 indices[2] = swizzle_index;
3101
3102 channel_vec = LLVMBuildGEP(builder, tcs->input, indices, 3, "");
3103 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
3104
3105 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3106 }
3107 } else {
3108 indices[0] = vertex_index;
3109 indices[1] = attrib_index;
3110 indices[2] = swizzle_index;
3111
3112 res = LLVMBuildGEP(builder, tcs->input, indices, 3, "");
3113 res = LLVMBuildLoad(builder, res, "");
3114 res = lp_build_broadcast_scalar(bld, res);
3115 }
3116 return res;
3117 }
3118
3119 static LLVMValueRef
3120 draw_tcs_llvm_emit_fetch_output(const struct lp_build_tcs_iface *tes_iface,
3121 struct lp_build_context *bld,
3122 boolean is_vindex_indirect,
3123 LLVMValueRef vertex_index,
3124 boolean is_aindex_indirect,
3125 LLVMValueRef attrib_index,
3126 LLVMValueRef swizzle_index,
3127 uint32_t name)
3128 {
3129 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
3130 struct gallivm_state *gallivm = bld->gallivm;
3131 LLVMBuilderRef builder = gallivm->builder;
3132 LLVMValueRef indices[3];
3133 LLVMValueRef res;
3134 struct lp_type type = bld->type;
3135
3136 if (is_vindex_indirect || is_aindex_indirect) {
3137 int i;
3138
3139 res = bld->zero;
3140 for (i = 0; i < type.length; ++i) {
3141 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3142 LLVMValueRef vert_chan_index = vertex_index;
3143 LLVMValueRef attr_chan_index = attrib_index;
3144 LLVMValueRef channel_vec;
3145
3146 if (is_vindex_indirect) {
3147 vert_chan_index = LLVMBuildExtractElement(builder,
3148 vertex_index, idx, "");
3149 }
3150 if (is_aindex_indirect) {
3151 attr_chan_index = LLVMBuildExtractElement(builder,
3152 attrib_index, idx, "");
3153 }
3154
3155 indices[0] = vert_chan_index;
3156 indices[1] = attr_chan_index;
3157 indices[2] = swizzle_index;
3158
3159 channel_vec = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3160 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
3161
3162 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3163 }
3164 } else {
3165 indices[0] = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
3166 indices[1] = attrib_index;
3167 indices[2] = swizzle_index;
3168
3169 res = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3170 res = LLVMBuildLoad(builder, res, "");
3171 res = lp_build_broadcast_scalar(bld, res);
3172 }
3173 return res;
3174 }
3175
3176 static void
3177 draw_tcs_llvm_emit_store_output(const struct lp_build_tcs_iface *tes_iface,
3178 struct lp_build_context *bld,
3179 unsigned name,
3180 boolean is_vindex_indirect,
3181 LLVMValueRef vertex_index,
3182 boolean is_aindex_indirect,
3183 LLVMValueRef attrib_index,
3184 LLVMValueRef swizzle_index,
3185 LLVMValueRef value,
3186 LLVMValueRef mask_vec)
3187 {
3188 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
3189 struct gallivm_state *gallivm = bld->gallivm;
3190 LLVMBuilderRef builder = gallivm->builder;
3191 LLVMValueRef indices[3];
3192 LLVMValueRef res;
3193 struct lp_type type = bld->type;
3194
3195 if (is_vindex_indirect || is_aindex_indirect) {
3196 int i;
3197
3198 for (i = 0; i < type.length; ++i) {
3199 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3200 LLVMValueRef vert_chan_index = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
3201 LLVMValueRef attr_chan_index = attrib_index;
3202 LLVMValueRef channel_vec;
3203
3204 if (is_vindex_indirect) {
3205 vert_chan_index = LLVMBuildExtractElement(builder,
3206 vertex_index, idx, "");
3207 }
3208 if (is_aindex_indirect) {
3209 attr_chan_index = LLVMBuildExtractElement(builder,
3210 attrib_index, idx, "");
3211 }
3212
3213 indices[0] = vert_chan_index;
3214 indices[1] = attr_chan_index;
3215 indices[2] = swizzle_index;
3216
3217 channel_vec = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3218
3219 res = LLVMBuildExtractElement(builder, value, idx, "");
3220
3221 struct lp_build_if_state ifthen;
3222 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
3223 cond = LLVMBuildExtractElement(gallivm->builder, cond, idx, "");
3224 lp_build_if(&ifthen, gallivm, cond);
3225 LLVMBuildStore(builder, res, channel_vec);
3226 lp_build_endif(&ifthen);
3227 }
3228 } else {
3229 indices[0] = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
3230 indices[1] = attrib_index;
3231 indices[2] = swizzle_index;
3232
3233 res = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3234 for (unsigned i = 0; i < type.length; ++i) {
3235 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3236 LLVMValueRef val = LLVMBuildExtractElement(builder, value, idx, "");
3237
3238 struct lp_build_if_state ifthen;
3239 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
3240 cond = LLVMBuildExtractElement(gallivm->builder, cond, idx, "");
3241 lp_build_if(&ifthen, gallivm, cond);
3242 LLVMBuildStore(builder, val, res);
3243 lp_build_endif(&ifthen);
3244 }
3245 }
3246 }
3247
3248
3249 static LLVMValueRef
3250 generate_tcs_mask_value(struct draw_tcs_llvm_variant *variant,
3251 struct lp_type tcs_type, LLVMValueRef limit, LLVMValueRef loop_counter)
3252 {
3253 struct gallivm_state *gallivm = variant->gallivm;
3254 LLVMBuilderRef builder = gallivm->builder;
3255 struct lp_type mask_type = lp_int_type(tcs_type);
3256 LLVMValueRef num_vecs;
3257 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
3258 unsigned i;
3259
3260 num_vecs = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type), limit);
3261 for (i = 0; i < tcs_type.length; i++) {
3262 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3263 mask_val = LLVMBuildInsertElement(builder, mask_val, LLVMBuildAdd(builder, loop_counter, idx, ""), idx, "");
3264 }
3265 mask_val = lp_build_compare(gallivm, mask_type,
3266 PIPE_FUNC_GREATER, num_vecs, mask_val);
3267
3268 return mask_val;
3269 }
3270
3271 static void
3272 draw_tcs_llvm_generate(struct draw_llvm *llvm,
3273 struct draw_tcs_llvm_variant *variant)
3274 {
3275 struct gallivm_state *gallivm = variant->gallivm;
3276 LLVMContextRef context = gallivm->context;
3277 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
3278 LLVMTypeRef arg_types[6];
3279 LLVMTypeRef func_type, coro_func_type;
3280 LLVMValueRef variant_func, variant_coro;
3281 LLVMValueRef context_ptr;
3282 LLVMValueRef input_array, output_array, prim_id, patch_vertices_in;
3283 LLVMValueRef mask_val;
3284 LLVMBasicBlockRef block;
3285 LLVMBuilderRef builder;
3286 struct lp_build_context bld, bldvec;
3287 struct lp_build_sampler_soa *sampler = 0;
3288 struct lp_build_image_soa *image = NULL;
3289 struct lp_bld_tgsi_system_values system_values;
3290 char func_name[64], func_name_coro[64];
3291 unsigned i;
3292 struct draw_tcs_llvm_iface tcs_iface;
3293 struct lp_build_mask_context mask;
3294 LLVMValueRef consts_ptr, num_consts_ptr;
3295 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
3296 struct lp_type tcs_type;
3297 unsigned vector_length = variant->shader->base.vector_length;
3298
3299 memset(&system_values, 0, sizeof(system_values));
3300
3301 snprintf(func_name, sizeof(func_name), "draw_llvm_tcs_variant");
3302
3303 snprintf(func_name_coro, sizeof(func_name_coro), "draw_llvm_tcs_coro_variant");
3304
3305 arg_types[0] = get_tcs_context_ptr_type(variant); /* context */
3306 arg_types[1] = variant->input_array_type; /* input */
3307 arg_types[2] = variant->output_array_type;
3308 arg_types[3] = int32_type;
3309 arg_types[4] = int32_type;
3310 arg_types[5] = int32_type; /* coroutine only */
3311
3312 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types) - 1, 0);
3313
3314 coro_func_type = LLVMFunctionType(LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), arg_types, ARRAY_SIZE(arg_types), 0);
3315
3316 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
3317
3318 variant_coro = LLVMAddFunction(gallivm->module, func_name_coro, coro_func_type);
3319
3320 variant->function = variant_func;
3321 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
3322
3323 LLVMSetFunctionCallConv(variant_coro, LLVMCCallConv);
3324
3325 for (i = 0; i < ARRAY_SIZE(arg_types); ++i) {
3326 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
3327 lp_add_function_attr(variant_coro, i + 1, LP_FUNC_ATTR_NOALIAS);
3328 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
3329 }
3330 }
3331
3332 if (gallivm->cache && gallivm->cache->data_size)
3333 return;
3334 context_ptr = LLVMGetParam(variant_func, 0);
3335 input_array = LLVMGetParam(variant_func, 1);
3336 output_array = LLVMGetParam(variant_func, 2);
3337 prim_id = LLVMGetParam(variant_func, 3);
3338 patch_vertices_in = LLVMGetParam(variant_func, 4);
3339
3340 lp_build_name(context_ptr, "context");
3341 lp_build_name(input_array, "input");
3342 lp_build_name(output_array, "output");
3343 lp_build_name(prim_id, "prim_id");
3344 lp_build_name(patch_vertices_in, "patch_vertices_in");
3345
3346 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
3347 builder = gallivm->builder;
3348 LLVMPositionBuilderAtEnd(builder, block);
3349
3350 lp_build_context_init(&bld, gallivm, lp_type_int(32));
3351
3352 memset(&tcs_type, 0, sizeof tcs_type);
3353 tcs_type.floating = TRUE; /* floating point values */
3354 tcs_type.sign = TRUE; /* values are signed */
3355 tcs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
3356 tcs_type.width = 32; /* 32-bit float */
3357 tcs_type.length = vector_length;
3358
3359 lp_build_context_init(&bldvec, variant->gallivm, lp_int_type(tcs_type));
3360
3361 LLVMValueRef count = lp_build_const_int32(gallivm, variant->shader->base.vertices_out);
3362 LLVMValueRef step = lp_build_const_int32(gallivm, vector_length);
3363
3364 struct lp_build_loop_state loop_state[2];
3365 LLVMValueRef num_inner_loop;
3366 unsigned count_align = util_align_npot(variant->shader->base.vertices_out, tcs_type.length);
3367 num_inner_loop = lp_build_const_int32(gallivm, count_align / tcs_type.length);
3368 LLVMTypeRef hdl_ptr_type = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
3369 LLVMValueRef coro_hdls = LLVMBuildArrayAlloca(gallivm->builder, hdl_ptr_type, num_inner_loop, "coro_hdls");
3370 unsigned end_coroutine = INT_MAX;
3371 lp_build_loop_begin(&loop_state[1], gallivm,
3372 lp_build_const_int32(gallivm, 0)); /* coroutine reentry loop */
3373 lp_build_loop_begin(&loop_state[0], gallivm,
3374 lp_build_const_int32(gallivm, 0)); /* inner loop */
3375 {
3376 LLVMValueRef args[6