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 *
draw_gs_llvm_iface(const struct lp_build_gs_iface * 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 *
draw_tcs_llvm_iface(const struct lp_build_tcs_iface * 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 *
draw_tes_llvm_iface(const struct lp_build_tes_iface * 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
create_jit_dvbuffer_type(struct gallivm_state * gallivm,const char * struct_name)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
create_jit_texture_type(struct gallivm_state * gallivm,const char * struct_name)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
create_jit_sampler_type(struct gallivm_state * gallivm,const char * struct_name)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] =
223 elem_types[DRAW_JIT_SAMPLER_MAX_ANISO] = LLVMFloatTypeInContext(gallivm->context);
224 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
225 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
226
227 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
228 ARRAY_SIZE(elem_types), 0);
229
230 (void) target; /* silence unused var warning for non-debug build */
231 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
232 target, sampler_type,
233 DRAW_JIT_SAMPLER_MIN_LOD);
234 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
235 target, sampler_type,
236 DRAW_JIT_SAMPLER_MAX_LOD);
237 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
238 target, sampler_type,
239 DRAW_JIT_SAMPLER_LOD_BIAS);
240 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
241 target, sampler_type,
242 DRAW_JIT_SAMPLER_BORDER_COLOR);
243 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_aniso,
244 target, sampler_type,
245 DRAW_JIT_SAMPLER_MAX_ANISO);
246
247 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
248
249 return sampler_type;
250 }
251
252 /**
253 * Create LLVM type for struct draw_jit_texture
254 */
255 static LLVMTypeRef
create_jit_image_type(struct gallivm_state * gallivm,const char * struct_name)256 create_jit_image_type(struct gallivm_state *gallivm, const char *struct_name)
257 {
258 LLVMTargetDataRef target = gallivm->target;
259 LLVMTypeRef image_type;
260 LLVMTypeRef elem_types[DRAW_JIT_IMAGE_NUM_FIELDS];
261 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
262
263 elem_types[DRAW_JIT_IMAGE_WIDTH] =
264 elem_types[DRAW_JIT_IMAGE_HEIGHT] =
265 elem_types[DRAW_JIT_IMAGE_DEPTH] =
266 elem_types[DRAW_JIT_IMAGE_ROW_STRIDE] =
267 elem_types[DRAW_JIT_IMAGE_IMG_STRIDE] =
268 elem_types[DRAW_JIT_IMAGE_NUM_SAMPLES] =
269 elem_types[DRAW_JIT_IMAGE_SAMPLE_STRIDE] = int32_type;
270 elem_types[DRAW_JIT_IMAGE_BASE] =
271 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
272
273 image_type = LLVMStructTypeInContext(gallivm->context, elem_types,
274 ARRAY_SIZE(elem_types), 0);
275
276 (void) target; /* silence unused var warning for non-debug build */
277 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, width,
278 target, image_type,
279 DRAW_JIT_IMAGE_WIDTH);
280 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, height,
281 target, image_type,
282 DRAW_JIT_IMAGE_HEIGHT);
283 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, depth,
284 target, image_type,
285 DRAW_JIT_IMAGE_DEPTH);
286 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, base,
287 target, image_type,
288 DRAW_JIT_IMAGE_BASE);
289 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, row_stride,
290 target, image_type,
291 DRAW_JIT_IMAGE_ROW_STRIDE);
292 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, img_stride,
293 target, image_type,
294 DRAW_JIT_IMAGE_IMG_STRIDE);
295 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, num_samples,
296 target, image_type,
297 DRAW_JIT_IMAGE_NUM_SAMPLES);
298 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image, sample_stride,
299 target, image_type,
300 DRAW_JIT_IMAGE_SAMPLE_STRIDE);
301
302 LP_CHECK_STRUCT_SIZE(struct draw_jit_image, target, image_type);
303
304 return image_type;
305 }
306
307 /**
308 * Create LLVM type for struct draw_jit_context
309 */
310 static LLVMTypeRef
create_jit_context_type(struct gallivm_state * gallivm,LLVMTypeRef texture_type,LLVMTypeRef sampler_type,LLVMTypeRef image_type,const char * struct_name)311 create_jit_context_type(struct gallivm_state *gallivm,
312 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
313 LLVMTypeRef image_type,
314 const char *struct_name)
315 {
316 LLVMTargetDataRef target = gallivm->target;
317 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
318 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
319 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
320 LLVMTypeRef context_type;
321
322 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
323 LP_MAX_TGSI_CONST_BUFFERS);
324 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
325 LP_MAX_TGSI_CONST_BUFFERS);
326 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
327 DRAW_TOTAL_CLIP_PLANES), 0);
328 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
329 elem_types[4] = LLVMArrayType(texture_type,
330 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
331 elem_types[5] = LLVMArrayType(sampler_type,
332 PIPE_MAX_SAMPLERS); /* samplers */
333 elem_types[6] = LLVMArrayType(image_type,
334 PIPE_MAX_SHADER_IMAGES); /* images */
335 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* vs_ssbo */
336 LP_MAX_TGSI_SHADER_BUFFERS);
337 elem_types[8] = LLVMArrayType(int_type, /* num_vs_ssbos */
338 LP_MAX_TGSI_SHADER_BUFFERS);
339 elem_types[9] = LLVMPointerType(float_type, 0); /* aniso table */
340 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
341 ARRAY_SIZE(elem_types), 0);
342
343 (void) target; /* silence unused var warning for non-debug build */
344 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
345 target, context_type, DRAW_JIT_CTX_CONSTANTS);
346 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
347 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
348 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
349 target, context_type, DRAW_JIT_CTX_PLANES);
350 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
351 target, context_type, DRAW_JIT_CTX_VIEWPORT);
352 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
353 target, context_type,
354 DRAW_JIT_CTX_TEXTURES);
355 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
356 target, context_type,
357 DRAW_JIT_CTX_SAMPLERS);
358 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, images,
359 target, context_type, DRAW_JIT_CTX_IMAGES);
360 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_ssbos,
361 target, context_type, DRAW_JIT_CTX_SSBOS);
362 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_ssbos,
363 target, context_type, DRAW_JIT_CTX_NUM_SSBOS);
364 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, aniso_filter_table,
365 target, context_type, DRAW_JIT_CTX_ANISO_FILTER_TABLE);
366 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
367 target, context_type);
368
369 return context_type;
370 }
371
372
373 /**
374 * Create LLVM type for struct draw_gs_jit_context
375 */
376 static LLVMTypeRef
create_gs_jit_context_type(struct gallivm_state * gallivm,unsigned vector_length,LLVMTypeRef texture_type,LLVMTypeRef sampler_type,LLVMTypeRef image_type,const char * struct_name)377 create_gs_jit_context_type(struct gallivm_state *gallivm,
378 unsigned vector_length,
379 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
380 LLVMTypeRef image_type,
381 const char *struct_name)
382 {
383 LLVMTargetDataRef target = gallivm->target;
384 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
385 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
386 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
387 LLVMTypeRef context_type;
388
389 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
390 LP_MAX_TGSI_CONST_BUFFERS);
391 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
392 LP_MAX_TGSI_CONST_BUFFERS);
393 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
394 DRAW_TOTAL_CLIP_PLANES), 0);
395 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
396
397 elem_types[4] = LLVMArrayType(texture_type,
398 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
399 elem_types[5] = LLVMArrayType(sampler_type,
400 PIPE_MAX_SAMPLERS); /* samplers */
401 elem_types[6] = LLVMArrayType(image_type,
402 PIPE_MAX_SHADER_IMAGES); /* images */
403 elem_types[7] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
404 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
405 vector_length), 0);
406 elem_types[9] = LLVMPointerType(LLVMVectorType(int_type,
407 vector_length), 0);
408
409 elem_types[10] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
410 LP_MAX_TGSI_SHADER_BUFFERS);
411 elem_types[11] = LLVMArrayType(int_type, /* num_ssbos */
412 LP_MAX_TGSI_SHADER_BUFFERS);
413 elem_types[12] = LLVMPointerType(float_type, 0); /* aniso table */
414 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
415 ARRAY_SIZE(elem_types), 0);
416
417 (void) target; /* silence unused var warning for non-debug build */
418 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
419 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
420 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
421 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
422 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
423 target, context_type, DRAW_GS_JIT_CTX_PLANES);
424 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
425 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
426 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
427 target, context_type,
428 DRAW_GS_JIT_CTX_TEXTURES);
429 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
430 target, context_type,
431 DRAW_GS_JIT_CTX_SAMPLERS);
432 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
433 target, context_type,
434 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
435 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
436 target, context_type,
437 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
438 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
439 target, context_type,
440 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
441 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, ssbos,
442 target, context_type, DRAW_GS_JIT_CTX_SSBOS);
443 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_ssbos,
444 target, context_type, DRAW_GS_JIT_CTX_NUM_SSBOS);
445 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, images,
446 target, context_type, DRAW_GS_JIT_CTX_IMAGES);
447 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, aniso_filter_table,
448 target, context_type, DRAW_GS_JIT_CTX_ANISO_FILTER_TABLE);
449 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
450 target, context_type);
451
452 return context_type;
453 }
454
455
456 static LLVMTypeRef
create_gs_jit_input_type(struct gallivm_state * gallivm)457 create_gs_jit_input_type(struct gallivm_state *gallivm)
458 {
459 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
460 LLVMTypeRef input_array;
461
462 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
463 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
464 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
465 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
466
467 return input_array;
468 }
469
470 /**
471 * Create LLVM type for struct pipe_vertex_buffer
472 */
473 static LLVMTypeRef
create_jit_vertex_buffer_type(struct gallivm_state * gallivm,const char * struct_name)474 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
475 const char *struct_name)
476 {
477 LLVMTargetDataRef target = gallivm->target;
478 LLVMTypeRef elem_types[4];
479 LLVMTypeRef vb_type;
480
481 elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
482 elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
483 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
484 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
485
486 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
487 ARRAY_SIZE(elem_types), 0);
488
489 (void) target; /* silence unused var warning for non-debug build */
490 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
491 target, vb_type, 0);
492 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
493 target, vb_type, 1);
494 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
495 target, vb_type, 2);
496 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
497 target, vb_type, 3);
498
499 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
500
501 return vb_type;
502 }
503
504
505 /**
506 * Create LLVM type for struct vertex_header;
507 */
508 static LLVMTypeRef
create_jit_vertex_header(struct gallivm_state * gallivm,int data_elems)509 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
510 {
511 LLVMTargetDataRef target = gallivm->target;
512 LLVMTypeRef elem_types[3];
513 LLVMTypeRef vertex_header;
514 char struct_name[24];
515
516 snprintf(struct_name, 23, "vertex_header%d", data_elems);
517
518 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
519 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
520 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
521
522 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
523 ARRAY_SIZE(elem_types), 0);
524
525 /* these are bit-fields and we can't take address of them
526 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
527 target, vertex_header,
528 DRAW_JIT_VERTEX_CLIPMASK);
529 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
530 target, vertex_header,
531 DRAW_JIT_VERTEX_EDGEFLAG);
532 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
533 target, vertex_header,
534 DRAW_JIT_VERTEX_PAD);
535 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
536 target, vertex_header,
537 DRAW_JIT_VERTEX_VERTEX_ID);
538 */
539 (void) target; /* silence unused var warning for non-debug build */
540 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
541 target, vertex_header,
542 DRAW_JIT_VERTEX_CLIP_POS);
543 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
544 target, vertex_header,
545 DRAW_JIT_VERTEX_DATA);
546
547 assert(LLVMABISizeOfType(target, vertex_header) ==
548 offsetof(struct vertex_header, data[data_elems]));
549
550 return vertex_header;
551 }
552
553 /**
554 * Create LLVM type for struct draw_tcs_jit_context
555 */
556 static LLVMTypeRef
create_tcs_jit_context_type(struct gallivm_state * gallivm,unsigned vector_length,LLVMTypeRef texture_type,LLVMTypeRef sampler_type,LLVMTypeRef image_type,const char * struct_name)557 create_tcs_jit_context_type(struct gallivm_state *gallivm,
558 unsigned vector_length,
559 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
560 LLVMTypeRef image_type,
561 const char *struct_name)
562 {
563 LLVMTargetDataRef target = gallivm->target;
564 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
565 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
566 LLVMTypeRef elem_types[DRAW_TCS_JIT_CTX_NUM_FIELDS];
567 LLVMTypeRef context_type;
568
569 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
570 LP_MAX_TGSI_CONST_BUFFERS);
571 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
572 LP_MAX_TGSI_CONST_BUFFERS);
573 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
574 elem_types[3] = LLVMInt32TypeInContext(gallivm->context);
575
576 elem_types[4] = LLVMArrayType(texture_type,
577 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
578 elem_types[5] = LLVMArrayType(sampler_type,
579 PIPE_MAX_SAMPLERS); /* samplers */
580 elem_types[6] = LLVMArrayType(image_type,
581 PIPE_MAX_SHADER_IMAGES); /* images */
582
583 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
584 LP_MAX_TGSI_SHADER_BUFFERS);
585 elem_types[8] = LLVMArrayType(int_type, /* num_ssbos */
586 LP_MAX_TGSI_SHADER_BUFFERS);
587 elem_types[9] = LLVMPointerType(float_type, 0); /* aniso table */
588 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
589 ARRAY_SIZE(elem_types), 0);
590
591 (void) target; /* silence unused var warning for non-debug build */
592 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, constants,
593 target, context_type, DRAW_TCS_JIT_CTX_CONSTANTS);
594 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, num_constants,
595 target, context_type, DRAW_TCS_JIT_CTX_NUM_CONSTANTS);
596 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, textures,
597 target, context_type,
598 DRAW_TCS_JIT_CTX_TEXTURES);
599 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, samplers,
600 target, context_type,
601 DRAW_TCS_JIT_CTX_SAMPLERS);
602 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, ssbos,
603 target, context_type, DRAW_TCS_JIT_CTX_SSBOS);
604 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, num_ssbos,
605 target, context_type, DRAW_TCS_JIT_CTX_NUM_SSBOS);
606 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, images,
607 target, context_type, DRAW_TCS_JIT_CTX_IMAGES);
608 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, aniso_filter_table,
609 target, context_type, DRAW_TCS_JIT_CTX_ANISO_FILTER_TABLE);
610 LP_CHECK_STRUCT_SIZE(struct draw_tcs_jit_context,
611 target, context_type);
612
613 return context_type;
614 }
615
616 static LLVMTypeRef
create_tcs_jit_input_type(struct gallivm_state * gallivm)617 create_tcs_jit_input_type(struct gallivm_state *gallivm)
618 {
619 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
620 LLVMTypeRef input_array;
621
622 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
623 input_array = LLVMArrayType(input_array, NUM_TCS_INPUTS); /* num attrs per vertex */
624 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
625
626 return input_array;
627 }
628
629 static LLVMTypeRef
create_tcs_jit_output_type(struct gallivm_state * gallivm)630 create_tcs_jit_output_type(struct gallivm_state *gallivm)
631 {
632 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
633 LLVMTypeRef output_array;
634
635 output_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
636 output_array = LLVMArrayType(output_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
637 output_array = LLVMPointerType(output_array, 0); /* num vertices per prim */
638
639 return output_array;
640 }
641
642 static LLVMTypeRef
create_tes_jit_input_type(struct gallivm_state * gallivm)643 create_tes_jit_input_type(struct gallivm_state *gallivm)
644 {
645 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
646 LLVMTypeRef input_array;
647
648 input_array = LLVMArrayType(float_type, TGSI_NUM_CHANNELS); /* num channels */
649 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
650 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
651
652 return input_array;
653 }
654
655 /**
656 * Create LLVM type for struct draw_tes_jit_context
657 */
658 static LLVMTypeRef
create_tes_jit_context_type(struct gallivm_state * gallivm,unsigned vector_length,LLVMTypeRef texture_type,LLVMTypeRef sampler_type,LLVMTypeRef image_type,const char * struct_name)659 create_tes_jit_context_type(struct gallivm_state *gallivm,
660 unsigned vector_length,
661 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
662 LLVMTypeRef image_type,
663 const char *struct_name)
664 {
665 LLVMTargetDataRef target = gallivm->target;
666 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
667 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
668 LLVMTypeRef elem_types[DRAW_TCS_JIT_CTX_NUM_FIELDS];
669 LLVMTypeRef context_type;
670
671 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
672 LP_MAX_TGSI_CONST_BUFFERS);
673 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
674 LP_MAX_TGSI_CONST_BUFFERS);
675 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
676 elem_types[3] = LLVMInt32TypeInContext(gallivm->context);
677
678 elem_types[4] = LLVMArrayType(texture_type,
679 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
680 elem_types[5] = LLVMArrayType(sampler_type,
681 PIPE_MAX_SAMPLERS); /* samplers */
682 elem_types[6] = LLVMArrayType(image_type,
683 PIPE_MAX_SHADER_IMAGES); /* images */
684
685 elem_types[7] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
686 LP_MAX_TGSI_SHADER_BUFFERS);
687 elem_types[8] = LLVMArrayType(int_type, /* num_ssbos */
688 LP_MAX_TGSI_SHADER_BUFFERS);
689 elem_types[9] = LLVMPointerType(float_type, 0); /* aniso table */
690 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
691 ARRAY_SIZE(elem_types), 0);
692
693 (void) target; /* silence unused var warning for non-debug build */
694 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, constants,
695 target, context_type, DRAW_TCS_JIT_CTX_CONSTANTS);
696 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, num_constants,
697 target, context_type, DRAW_TCS_JIT_CTX_NUM_CONSTANTS);
698 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, textures,
699 target, context_type,
700 DRAW_TCS_JIT_CTX_TEXTURES);
701 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, samplers,
702 target, context_type,
703 DRAW_TCS_JIT_CTX_SAMPLERS);
704 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, ssbos,
705 target, context_type, DRAW_TCS_JIT_CTX_SSBOS);
706 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, num_ssbos,
707 target, context_type, DRAW_TCS_JIT_CTX_NUM_SSBOS);
708 LP_CHECK_MEMBER_OFFSET(struct draw_tes_jit_context, images,
709 target, context_type, DRAW_TCS_JIT_CTX_IMAGES);
710 LP_CHECK_MEMBER_OFFSET(struct draw_tcs_jit_context, aniso_filter_table,
711 target, context_type, DRAW_TCS_JIT_CTX_ANISO_FILTER_TABLE);
712 LP_CHECK_STRUCT_SIZE(struct draw_tes_jit_context,
713 target, context_type);
714
715 return context_type;
716 }
717
718 /**
719 * Create LLVM types for various structures.
720 */
721 static void
create_jit_types(struct draw_llvm_variant * variant)722 create_jit_types(struct draw_llvm_variant *variant)
723 {
724 struct gallivm_state *gallivm = variant->gallivm;
725 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
726 vb_type, image_type;
727
728 texture_type = create_jit_texture_type(gallivm, "texture");
729 sampler_type = create_jit_sampler_type(gallivm, "sampler");
730 image_type = create_jit_image_type(gallivm, "image");
731
732 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
733 image_type,
734 "draw_jit_context");
735 variant->context_ptr_type = LLVMPointerType(context_type, 0);
736
737 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
738 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
739
740 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
741 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
742 }
743
744
745 static LLVMTypeRef
get_context_ptr_type(struct draw_llvm_variant * variant)746 get_context_ptr_type(struct draw_llvm_variant *variant)
747 {
748 if (!variant->context_ptr_type)
749 create_jit_types(variant);
750 return variant->context_ptr_type;
751 }
752
753
754 static LLVMTypeRef
get_buffer_ptr_type(struct draw_llvm_variant * variant)755 get_buffer_ptr_type(struct draw_llvm_variant *variant)
756 {
757 if (!variant->buffer_ptr_type)
758 create_jit_types(variant);
759 return variant->buffer_ptr_type;
760 }
761
762
763 static LLVMTypeRef
get_vb_ptr_type(struct draw_llvm_variant * variant)764 get_vb_ptr_type(struct draw_llvm_variant *variant)
765 {
766 if (!variant->vb_ptr_type)
767 create_jit_types(variant);
768 return variant->vb_ptr_type;
769 }
770
771 static LLVMTypeRef
get_vertex_header_ptr_type(struct draw_llvm_variant * variant)772 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
773 {
774 if (!variant->vertex_header_ptr_type)
775 create_jit_types(variant);
776 return variant->vertex_header_ptr_type;
777 }
778
779
780 /**
781 * Create per-context LLVM info.
782 */
783 struct draw_llvm *
draw_llvm_create(struct draw_context * draw,LLVMContextRef context)784 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
785 {
786 struct draw_llvm *llvm;
787
788 if (!lp_build_init())
789 return NULL;
790
791 llvm = CALLOC_STRUCT( draw_llvm );
792 if (!llvm)
793 return NULL;
794
795 llvm->draw = draw;
796
797 llvm->context = context;
798 if (!llvm->context) {
799 llvm->context = LLVMContextCreate();
800 llvm->context_owned = true;
801 }
802 if (!llvm->context)
803 goto fail;
804
805 llvm->nr_variants = 0;
806 make_empty_list(&llvm->vs_variants_list);
807
808 llvm->nr_gs_variants = 0;
809 make_empty_list(&llvm->gs_variants_list);
810
811 llvm->nr_tcs_variants = 0;
812 make_empty_list(&llvm->tcs_variants_list);
813
814 llvm->nr_tes_variants = 0;
815 make_empty_list(&llvm->tes_variants_list);
816
817 return llvm;
818
819 fail:
820 draw_llvm_destroy(llvm);
821 return NULL;
822 }
823
824
825 /**
826 * Free per-context LLVM info.
827 */
828 void
draw_llvm_destroy(struct draw_llvm * llvm)829 draw_llvm_destroy(struct draw_llvm *llvm)
830 {
831 if (llvm->context_owned)
832 LLVMContextDispose(llvm->context);
833 llvm->context = NULL;
834
835 /* XXX free other draw_llvm data? */
836 FREE(llvm);
837 }
838
839 static void
draw_get_ir_cache_key(struct nir_shader * nir,const void * key,size_t key_size,uint32_t val_32bit,unsigned char ir_sha1_cache_key[20])840 draw_get_ir_cache_key(struct nir_shader *nir,
841 const void *key, size_t key_size,
842 uint32_t val_32bit,
843 unsigned char ir_sha1_cache_key[20])
844 {
845 struct blob blob = { 0 };
846 unsigned ir_size;
847 void *ir_binary;
848
849 blob_init(&blob);
850 nir_serialize(&blob, nir, true);
851 ir_binary = blob.data;
852 ir_size = blob.size;
853
854 struct mesa_sha1 ctx;
855 _mesa_sha1_init(&ctx);
856 _mesa_sha1_update(&ctx, key, key_size);
857 _mesa_sha1_update(&ctx, ir_binary, ir_size);
858 _mesa_sha1_update(&ctx, &val_32bit, 4);
859 _mesa_sha1_final(&ctx, ir_sha1_cache_key);
860
861 blob_finish(&blob);
862 }
863
864 /**
865 * Create LLVM-generated code for a vertex shader.
866 */
867 struct draw_llvm_variant *
draw_llvm_create_variant(struct draw_llvm * llvm,unsigned num_inputs,const struct draw_llvm_variant_key * key)868 draw_llvm_create_variant(struct draw_llvm *llvm,
869 unsigned num_inputs,
870 const struct draw_llvm_variant_key *key)
871 {
872 struct draw_llvm_variant *variant;
873 struct llvm_vertex_shader *shader =
874 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
875 LLVMTypeRef vertex_header;
876 char module_name[64];
877 unsigned char ir_sha1_cache_key[20];
878 struct lp_cached_code cached = { 0 };
879 bool needs_caching = false;
880 variant = MALLOC(sizeof *variant +
881 shader->variant_key_size -
882 sizeof variant->key);
883 if (!variant)
884 return NULL;
885
886 variant->llvm = llvm;
887 variant->shader = shader;
888 memcpy(&variant->key, key, shader->variant_key_size);
889
890 snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
891 variant->shader->variants_cached);
892
893 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
894 draw_get_ir_cache_key(shader->base.state.ir.nir,
895 key,
896 shader->variant_key_size,
897 num_inputs,
898 ir_sha1_cache_key);
899
900 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
901 &cached,
902 ir_sha1_cache_key);
903 if (!cached.data_size)
904 needs_caching = true;
905 }
906 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
907
908 create_jit_types(variant);
909
910 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
911 if (llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_TGSI)
912 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
913 else
914 nir_print_shader(llvm->draw->vs.vertex_shader->state.ir.nir, stderr);
915 draw_llvm_dump_variant_key(&variant->key);
916 }
917
918 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
919
920 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
921
922 draw_llvm_generate(llvm, variant);
923
924 gallivm_compile_module(variant->gallivm);
925
926 variant->jit_func = (draw_jit_vert_func)
927 gallivm_jit_function(variant->gallivm, variant->function);
928
929 if (needs_caching)
930 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
931 &cached,
932 ir_sha1_cache_key);
933 gallivm_free_ir(variant->gallivm);
934
935 variant->list_item_global.base = variant;
936 variant->list_item_local.base = variant;
937 /*variant->no = */shader->variants_created++;
938 variant->list_item_global.base = variant;
939
940 return variant;
941 }
942
943 static void
do_clamp_vertex_color(struct gallivm_state * gallivm,struct lp_type type,const struct tgsi_shader_info * info,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS])944 do_clamp_vertex_color(struct gallivm_state *gallivm,
945 struct lp_type type,
946 const struct tgsi_shader_info *info,
947 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS])
948 {
949 LLVMBuilderRef builder = gallivm->builder;
950 LLVMValueRef out;
951 unsigned chan, attrib;
952 struct lp_build_context bld;
953 lp_build_context_init(&bld, gallivm, type);
954
955 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
956 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
957 if (outputs[attrib][chan]) {
958 switch (info->output_semantic_name[attrib]) {
959 case TGSI_SEMANTIC_COLOR:
960 case TGSI_SEMANTIC_BCOLOR:
961 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
962 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
963 LLVMBuildStore(builder, out, outputs[attrib][chan]);
964 break;
965 }
966 }
967 }
968 }
969 }
970
971 static void
generate_vs(struct draw_llvm_variant * variant,LLVMBuilderRef builder,struct lp_type vs_type,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],const LLVMValueRef (* inputs)[TGSI_NUM_CHANNELS],const struct lp_bld_tgsi_system_values * system_values,LLVMValueRef context_ptr,const struct lp_build_sampler_soa * draw_sampler,const struct lp_build_image_soa * draw_image,boolean clamp_vertex_color,struct lp_build_mask_context * bld_mask)972 generate_vs(struct draw_llvm_variant *variant,
973 LLVMBuilderRef builder,
974 struct lp_type vs_type,
975 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
976 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
977 const struct lp_bld_tgsi_system_values *system_values,
978 LLVMValueRef context_ptr,
979 const struct lp_build_sampler_soa *draw_sampler,
980 const struct lp_build_image_soa *draw_image,
981 boolean clamp_vertex_color,
982 struct lp_build_mask_context *bld_mask)
983 {
984 struct draw_llvm *llvm = variant->llvm;
985 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
986 LLVMValueRef consts_ptr =
987 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
988 LLVMValueRef num_consts_ptr =
989 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
990 LLVMValueRef ssbos_ptr =
991 draw_jit_context_vs_ssbos(variant->gallivm, context_ptr);
992 LLVMValueRef num_ssbos_ptr =
993 draw_jit_context_num_vs_ssbos(variant->gallivm, context_ptr);
994
995 struct lp_build_tgsi_params params;
996 memset(¶ms, 0, sizeof(params));
997
998 params.type = vs_type;
999 params.mask = bld_mask;
1000 params.consts_ptr = consts_ptr;
1001 params.const_sizes_ptr = num_consts_ptr;
1002 params.system_values = system_values;
1003 params.inputs = inputs;
1004 params.context_ptr = context_ptr;
1005 params.sampler = draw_sampler;
1006 params.info = &llvm->draw->vs.vertex_shader->info;
1007 params.ssbo_ptr = ssbos_ptr;
1008 params.ssbo_sizes_ptr = num_ssbos_ptr;
1009 params.image = draw_image;
1010 params.aniso_filter_table = draw_jit_context_aniso_filter_table(variant->gallivm, context_ptr);
1011
1012 if (llvm->draw->vs.vertex_shader->state.ir.nir &&
1013 llvm->draw->vs.vertex_shader->state.type == PIPE_SHADER_IR_NIR)
1014 lp_build_nir_soa(variant->gallivm,
1015 llvm->draw->vs.vertex_shader->state.ir.nir,
1016 ¶ms,
1017 outputs);
1018 else
1019 lp_build_tgsi_soa(variant->gallivm,
1020 tokens,
1021 ¶ms,
1022 outputs);
1023
1024 if (clamp_vertex_color) {
1025 const struct tgsi_shader_info *info = &llvm->draw->vs.vertex_shader->info;
1026 do_clamp_vertex_color(variant->gallivm,
1027 vs_type, info,
1028 outputs);
1029 }
1030 }
1031
1032
1033 static void
fetch_instanced(struct gallivm_state * gallivm,const struct util_format_description * format_desc,struct lp_type vs_type,LLVMValueRef vb_stride,LLVMValueRef map_ptr,LLVMValueRef buffer_size_adj,LLVMValueRef * inputs,LLVMValueRef index)1034 fetch_instanced(struct gallivm_state *gallivm,
1035 const struct util_format_description *format_desc,
1036 struct lp_type vs_type,
1037 LLVMValueRef vb_stride,
1038 LLVMValueRef map_ptr,
1039 LLVMValueRef buffer_size_adj,
1040 LLVMValueRef *inputs,
1041 LLVMValueRef index)
1042 {
1043 LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
1044 LLVMTypeRef aosf_t, aosi_t;
1045 LLVMValueRef zero = LLVMConstNull(i32_t);
1046 LLVMBuilderRef builder = gallivm->builder;
1047 LLVMValueRef stride, buffer_overflowed, aos, index_valid;
1048 unsigned i;
1049
1050 aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
1051 aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
1052
1053 /* This mul can overflow. Wraparound is ok. */
1054 stride = LLVMBuildMul(builder, vb_stride, index, "");
1055
1056 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
1057 stride, buffer_size_adj,
1058 "buffer_overflowed");
1059
1060 if (0) {
1061 lp_build_print_value(gallivm, " instance index = ", index);
1062 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
1063 }
1064
1065 index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
1066 index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
1067 stride = LLVMBuildAnd(builder, stride, index_valid, "");
1068
1069 aos = lp_build_fetch_rgba_aos(gallivm,
1070 format_desc,
1071 lp_float32_vec4_type(),
1072 FALSE,
1073 map_ptr,
1074 stride, zero, zero,
1075 NULL);
1076
1077 index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
1078 aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
1079 aos = LLVMBuildAnd(builder, aos, index_valid, "");
1080 aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
1081
1082 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1083 LLVMValueRef index = lp_build_const_int32(gallivm, i);
1084 inputs[i] = lp_build_extract_broadcast(gallivm,
1085 lp_float32_vec4_type(),
1086 vs_type, aos, index);
1087 }
1088 }
1089
1090
1091 static void
fetch_vector(struct gallivm_state * gallivm,const struct util_format_description * format_desc,struct lp_type vs_type,LLVMValueRef vb_stride,LLVMValueRef map_ptr,LLVMValueRef buffer_size_adj,LLVMValueRef * inputs,LLVMValueRef indices)1092 fetch_vector(struct gallivm_state *gallivm,
1093 const struct util_format_description *format_desc,
1094 struct lp_type vs_type,
1095 LLVMValueRef vb_stride,
1096 LLVMValueRef map_ptr,
1097 LLVMValueRef buffer_size_adj,
1098 LLVMValueRef *inputs,
1099 LLVMValueRef indices)
1100 {
1101 LLVMBuilderRef builder = gallivm->builder;
1102 struct lp_build_context blduivec;
1103 struct lp_type fetch_type = vs_type;
1104 LLVMValueRef offset, valid_mask;
1105 unsigned i;
1106
1107 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1108
1109 vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
1110 buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
1111
1112 /* This mul can overflow. Wraparound is ok. */
1113 offset = lp_build_mul(&blduivec, vb_stride, indices);
1114
1115 valid_mask = lp_build_compare(gallivm, blduivec.type,
1116 PIPE_FUNC_LESS, offset, buffer_size_adj);
1117
1118 /* not valid elements use offset 0 */
1119 offset = LLVMBuildAnd(builder, offset, valid_mask, "");
1120
1121 if (0) {
1122 lp_build_print_value(gallivm, " indices = ", indices);
1123 lp_build_print_value(gallivm, " offsets = ", offset);
1124 lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
1125 }
1126
1127 /*
1128 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
1129 * This should always produce better code.
1130 */
1131
1132 /* The type handling is annoying here... */
1133 if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
1134 format_desc->channel[0].pure_integer) {
1135 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
1136 fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
1137 }
1138 else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1139 fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
1140 }
1141 }
1142
1143 lp_build_fetch_rgba_soa(gallivm, format_desc,
1144 fetch_type, FALSE, map_ptr, offset,
1145 blduivec.zero, blduivec.zero,
1146 NULL, inputs);
1147
1148 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1149 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
1150 lp_build_vec_type(gallivm, vs_type), "");
1151 }
1152
1153 /* out-of-bound fetches return all zeros */
1154 for (i = 0; i < format_desc->nr_channels; i++) {
1155 inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
1156 inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
1157 inputs[i] = LLVMBuildBitCast(builder, inputs[i],
1158 lp_build_vec_type(gallivm, vs_type), "");
1159 }
1160 }
1161
1162
1163 static void
store_aos(struct gallivm_state * gallivm,LLVMValueRef io_ptr,LLVMValueRef index,LLVMValueRef value)1164 store_aos(struct gallivm_state *gallivm,
1165 LLVMValueRef io_ptr,
1166 LLVMValueRef index,
1167 LLVMValueRef value)
1168 {
1169 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
1170 LLVMBuilderRef builder = gallivm->builder;
1171 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
1172 LLVMValueRef indices[3];
1173
1174 indices[0] = lp_build_const_int32(gallivm, 0);
1175 indices[1] = index;
1176 indices[2] = lp_build_const_int32(gallivm, 0);
1177
1178 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
1179 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
1180
1181 #if DEBUG_STORE
1182 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
1183 #endif
1184
1185 /* Unaligned store due to the vertex header */
1186 LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
1187 }
1188
1189 /**
1190 * Adjust the mask to architecture endianess. The mask will the store in struct:
1191 *
1192 * struct vertex_header {
1193 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
1194 * unsigned edgeflag:1;
1195 * unsigned pad:1;
1196 * unsigned vertex_id:16;
1197 * [...]
1198 * }
1199 *
1200 * On little-endian machine nothing needs to done, however on bit-endian machine
1201 * the mask's fields need to be adjusted with the algorithm:
1202 *
1203 * uint32_t reverse (uint32_t x)
1204 * {
1205 * return (x >> 16) | // vertex_id
1206 * ((x & 0x3fff) << 18) | // clipmask
1207 * ((x & 0x4000) << 3) | // edgeflag
1208 * ((x & 0x8000) << 1); // pad
1209 * }
1210 */
1211 static LLVMValueRef
adjust_mask(struct gallivm_state * gallivm,LLVMValueRef mask)1212 adjust_mask(struct gallivm_state *gallivm,
1213 LLVMValueRef mask)
1214 {
1215 #if UTIL_ARCH_BIG_ENDIAN
1216 LLVMBuilderRef builder = gallivm->builder;
1217 LLVMValueRef vertex_id;
1218 LLVMValueRef clipmask;
1219 LLVMValueRef pad;
1220 LLVMValueRef edgeflag;
1221
1222 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
1223 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
1224 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
1225 if (0) {
1226 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
1227 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 1), "");
1228 }
1229 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
1230 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 3), "");
1231
1232 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
1233 if (0) {
1234 mask = LLVMBuildOr(builder, mask, pad, "");
1235 }
1236 mask = LLVMBuildOr(builder, mask, edgeflag, "");
1237 #endif
1238 return mask;
1239 }
1240
1241 static void
store_aos_array(struct gallivm_state * gallivm,struct lp_type soa_type,LLVMValueRef io_ptr,LLVMValueRef * indices,LLVMValueRef * aos,int attrib,int num_outputs,LLVMValueRef clipmask,boolean need_edgeflag)1242 store_aos_array(struct gallivm_state *gallivm,
1243 struct lp_type soa_type,
1244 LLVMValueRef io_ptr,
1245 LLVMValueRef *indices,
1246 LLVMValueRef* aos,
1247 int attrib,
1248 int num_outputs,
1249 LLVMValueRef clipmask,
1250 boolean need_edgeflag)
1251 {
1252 LLVMBuilderRef builder = gallivm->builder;
1253 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
1254 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1255 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
1256 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1257 int vector_length = soa_type.length;
1258 int i;
1259
1260 debug_assert(TGSI_NUM_CHANNELS == 4);
1261
1262 for (i = 0; i < vector_length; i++) {
1263 linear_inds[i] = lp_build_const_int32(gallivm, i);
1264 if (indices) {
1265 inds[i] = indices[i];
1266 } else {
1267 inds[i] = linear_inds[i];
1268 }
1269 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1270 }
1271
1272 if (attrib == 0) {
1273 /* store vertex header for each of the n vertices */
1274 LLVMValueRef val, cliptmp;
1275 int vertex_id_pad_edgeflag;
1276
1277 /* If this assertion fails, it means we need to update the bit twidding
1278 * code here. See struct vertex_header in draw_private.h.
1279 */
1280 assert(DRAW_TOTAL_CLIP_PLANES==14);
1281 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
1282 if (!need_edgeflag) {
1283 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
1284 }
1285 else {
1286 vertex_id_pad_edgeflag = (0xffff << 16);
1287 }
1288 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
1289 vertex_id_pad_edgeflag);
1290 /* OR with the clipmask */
1291 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
1292 for (i = 0; i < vector_length; i++) {
1293 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
1294 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
1295 val = adjust_mask(gallivm, val);
1296 #if DEBUG_STORE
1297 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
1298 io_ptrs[i], inds[i], val);
1299 #endif
1300 LLVMBuildStore(builder, val, id_ptr);
1301 }
1302 }
1303
1304 /* store for each of the n vertices */
1305 for (i = 0; i < vector_length; i++) {
1306 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
1307 }
1308 }
1309
1310
1311 static void
convert_to_aos(struct gallivm_state * gallivm,LLVMValueRef io,LLVMValueRef * indices,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],LLVMValueRef clipmask,int num_outputs,struct lp_type soa_type,boolean need_edgeflag)1312 convert_to_aos(struct gallivm_state *gallivm,
1313 LLVMValueRef io,
1314 LLVMValueRef *indices,
1315 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1316 LLVMValueRef clipmask,
1317 int num_outputs,
1318 struct lp_type soa_type,
1319 boolean need_edgeflag)
1320 {
1321 LLVMBuilderRef builder = gallivm->builder;
1322 unsigned chan, attrib, i;
1323
1324 #if DEBUG_STORE
1325 lp_build_printf(gallivm, " # storing begin\n");
1326 #endif
1327 for (attrib = 0; attrib < num_outputs; ++attrib) {
1328 LLVMValueRef soa[TGSI_NUM_CHANNELS];
1329 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
1330 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1331 if (outputs[attrib][chan]) {
1332 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
1333 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
1334 #if DEBUG_STORE
1335 lp_build_printf(gallivm, "output %d : %d ",
1336 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1337 attrib, 0),
1338 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1339 chan, 0));
1340 lp_build_print_value(gallivm, "val = ", out);
1341 {
1342 LLVMValueRef iv =
1343 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
1344
1345 lp_build_print_value(gallivm, " ival = ", iv);
1346 }
1347 #endif
1348 soa[chan] = out;
1349 }
1350 else {
1351 soa[chan] = 0;
1352 }
1353 }
1354
1355
1356 if (soa_type.length == TGSI_NUM_CHANNELS) {
1357 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1358 } else {
1359 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1360
1361 for (i = 0; i < soa_type.length; ++i) {
1362 aos[i] = lp_build_extract_range(gallivm,
1363 soa[i % TGSI_NUM_CHANNELS],
1364 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1365 TGSI_NUM_CHANNELS);
1366 }
1367 }
1368
1369 store_aos_array(gallivm,
1370 soa_type,
1371 io, indices,
1372 aos,
1373 attrib,
1374 num_outputs,
1375 clipmask,
1376 need_edgeflag);
1377 }
1378 #if DEBUG_STORE
1379 lp_build_printf(gallivm, " # storing end\n");
1380 #endif
1381 }
1382
1383
1384 /**
1385 * Stores original vertex positions in clip coordinates
1386 */
1387 static void
store_clip(struct gallivm_state * gallivm,const struct lp_type vs_type,LLVMValueRef io_ptr,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],int idx)1388 store_clip(struct gallivm_state *gallivm,
1389 const struct lp_type vs_type,
1390 LLVMValueRef io_ptr,
1391 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1392 int idx)
1393 {
1394 LLVMBuilderRef builder = gallivm->builder;
1395 LLVMValueRef soa[4];
1396 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1397 LLVMValueRef indices[2];
1398 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1399 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1400 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1401 LLVMTypeRef clip_ptr_type =
1402 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1403 4), 0);
1404 int i, j;
1405
1406 indices[0] =
1407 indices[1] = lp_build_const_int32(gallivm, 0);
1408
1409 for (i = 0; i < vs_type.length; i++) {
1410 inds[i] = lp_build_const_int32(gallivm, i);
1411 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1412 }
1413
1414 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1415 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1416 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1417 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1418
1419 for (i = 0; i < vs_type.length; i++) {
1420 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1421 }
1422
1423 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1424 for (i = 0; i < vs_type.length; ++i) {
1425 aos[i] = lp_build_extract_range(gallivm,
1426 soa[i % TGSI_NUM_CHANNELS],
1427 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1428 TGSI_NUM_CHANNELS);
1429 }
1430
1431 for (j = 0; j < vs_type.length; j++) {
1432 LLVMValueRef clip_ptr;
1433
1434 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1435 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1436
1437 /* Unaligned store */
1438 LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1439 }
1440 }
1441
1442
1443 /**
1444 * Transforms the outputs for viewport mapping
1445 */
1446 static void
generate_viewport(struct draw_llvm_variant * variant,LLVMBuilderRef builder,struct lp_type vs_type,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],LLVMValueRef context_ptr)1447 generate_viewport(struct draw_llvm_variant *variant,
1448 LLVMBuilderRef builder,
1449 struct lp_type vs_type,
1450 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1451 LLVMValueRef context_ptr)
1452 {
1453 int i;
1454 struct gallivm_state *gallivm = variant->gallivm;
1455 struct lp_type f32_type = vs_type;
1456 const unsigned pos = variant->llvm->draw->vs.position_output;
1457 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1458 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1459 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1460 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1461
1462 /* We treat pipe_viewport_state as a float array */
1463 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1464 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1465
1466 /* for 1/w convention*/
1467 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1468 LLVMBuildStore(builder, out3, outputs[pos][3]);
1469
1470 /* Viewport Mapping */
1471 for (i=0; i<3; i++) {
1472 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1473 LLVMValueRef scale;
1474 LLVMValueRef trans;
1475 LLVMValueRef scale_i;
1476 LLVMValueRef trans_i;
1477 LLVMValueRef index;
1478
1479 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1480 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1481
1482 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1483 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1484
1485 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1486 LLVMBuildLoad(builder, scale_i, "scale"));
1487 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1488 LLVMBuildLoad(builder, trans_i, "trans"));
1489
1490 /* divide by w */
1491 out = LLVMBuildFMul(builder, out, out3, "");
1492 /* mult by scale, add translation */
1493 out = lp_build_fmuladd(builder, out, scale, trans);
1494
1495 /* store transformed outputs */
1496 LLVMBuildStore(builder, out, outputs[pos][i]);
1497 }
1498
1499 }
1500
1501
1502 /**
1503 * Returns clipmask as nxi32 bitmask for the n vertices
1504 */
1505 static LLVMValueRef
generate_clipmask(struct draw_llvm * llvm,struct gallivm_state * gallivm,struct lp_type vs_type,LLVMValueRef (* outputs)[TGSI_NUM_CHANNELS],struct draw_llvm_variant_key * key,LLVMValueRef context_ptr,boolean * have_clipdist)1506 generate_clipmask(struct draw_llvm *llvm,
1507 struct gallivm_state *gallivm,
1508 struct lp_type vs_type,
1509 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1510 struct draw_llvm_variant_key *key,
1511 LLVMValueRef context_ptr,
1512 boolean *have_clipdist)
1513 {
1514 LLVMBuilderRef builder = gallivm->builder;
1515 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1516 LLVMValueRef test, temp;
1517 LLVMValueRef zero, shift;
1518 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1519 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1520 LLVMValueRef plane1, planes, plane_ptr, sum;
1521 struct lp_type f32_type = vs_type;
1522 struct lp_type i32_type = lp_int_type(vs_type);
1523 const unsigned pos = llvm->draw->vs.position_output;
1524 const unsigned cv = llvm->draw->vs.clipvertex_output;
1525 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1526 boolean have_cd = false;
1527 boolean clip_user = key->clip_user;
1528 unsigned ucp_enable = key->ucp_enable;
1529 unsigned cd[2];
1530
1531 cd[0] = llvm->draw->vs.ccdistance_output[0];
1532 cd[1] = llvm->draw->vs.ccdistance_output[1];
1533
1534 if (cd[0] != pos || cd[1] != pos)
1535 have_cd = true;
1536
1537 if (num_written_clipdistance && !clip_user) {
1538 clip_user = true;
1539 ucp_enable = (1 << num_written_clipdistance) - 1;
1540 }
1541
1542 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1543 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1544 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1545 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1546
1547 /*
1548 * load clipvertex and position from correct locations.
1549 * if they are the same just load them once.
1550 */
1551 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1552 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1553 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1554 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1555
1556 if (clip_user && cv != pos) {
1557 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1558 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1559 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1560 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1561 } else {
1562 cv_x = pos_x;
1563 cv_y = pos_y;
1564 cv_z = pos_z;
1565 cv_w = pos_w;
1566 }
1567
1568 /*
1569 * Be careful with the comparisons and NaNs (using llvm's unordered
1570 * comparisons here).
1571 */
1572 /* Cliptest, for hardwired planes */
1573 /*
1574 * XXX should take guardband into account (currently not in key).
1575 * Otherwise might run the draw pipeline stages for nothing.
1576 */
1577 if (key->clip_xy) {
1578 /* plane 1 */
1579 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1580 temp = shift;
1581 test = LLVMBuildAnd(builder, test, temp, "");
1582 mask = test;
1583
1584 /* plane 2 */
1585 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1586 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1587 temp = LLVMBuildShl(builder, temp, shift, "");
1588 test = LLVMBuildAnd(builder, test, temp, "");
1589 mask = LLVMBuildOr(builder, mask, test, "");
1590
1591 /* plane 3 */
1592 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1593 temp = LLVMBuildShl(builder, temp, shift, "");
1594 test = LLVMBuildAnd(builder, test, temp, "");
1595 mask = LLVMBuildOr(builder, mask, test, "");
1596
1597 /* plane 4 */
1598 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1599 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1600 temp = LLVMBuildShl(builder, temp, shift, "");
1601 test = LLVMBuildAnd(builder, test, temp, "");
1602 mask = LLVMBuildOr(builder, mask, test, "");
1603 }
1604
1605 if (key->clip_z) {
1606 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1607 if (key->clip_halfz) {
1608 /* plane 5 */
1609 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1610 test = LLVMBuildAnd(builder, test, temp, "");
1611 mask = LLVMBuildOr(builder, mask, test, "");
1612 }
1613 else {
1614 /* plane 5 */
1615 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1616 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1617 test = LLVMBuildAnd(builder, test, temp, "");
1618 mask = LLVMBuildOr(builder, mask, test, "");
1619 }
1620 /* plane 6 */
1621 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1622 temp = LLVMBuildShl(builder, temp, shift, "");
1623 test = LLVMBuildAnd(builder, test, temp, "");
1624 mask = LLVMBuildOr(builder, mask, test, "");
1625 }
1626
1627 if (clip_user) {
1628 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1629 LLVMValueRef indices[3];
1630 LLVMValueRef is_nan_or_inf;
1631
1632 /* userclip planes */
1633 while (ucp_enable) {
1634 unsigned plane_idx = ffs(ucp_enable)-1;
1635 ucp_enable &= ~(1 << plane_idx);
1636 plane_idx += 6;
1637
1638 if (have_cd && num_written_clipdistance) {
1639 LLVMValueRef clipdist;
1640 int i;
1641 i = plane_idx - 6;
1642
1643 *have_clipdist = TRUE;
1644 if (i < 4) {
1645 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1646 } else {
1647 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1648 }
1649 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1650 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1651 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1652 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1653 test = LLVMBuildAnd(builder, test, temp, "");
1654 mask = LLVMBuildOr(builder, mask, test, "");
1655 } else {
1656 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1657 indices[0] = lp_build_const_int32(gallivm, 0);
1658 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1659
1660 indices[2] = lp_build_const_int32(gallivm, 0);
1661 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1662 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1663 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1664 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1665
1666 indices[2] = lp_build_const_int32(gallivm, 1);
1667 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1668 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1669 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1670 sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1671
1672 indices[2] = lp_build_const_int32(gallivm, 2);
1673 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1674 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1675 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1676 sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1677
1678 indices[2] = lp_build_const_int32(gallivm, 3);
1679 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1680 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1681 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1682 sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1683
1684 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1685 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1686 test = LLVMBuildAnd(builder, test, temp, "");
1687 mask = LLVMBuildOr(builder, mask, test, "");
1688 }
1689 }
1690 }
1691 if (key->need_edgeflags) {
1692 /*
1693 * This isn't really part of clipmask but stored the same in vertex
1694 * header later, so do it here.
1695 */
1696 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1697 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1698 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1699 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1700 temp = lp_build_const_int_vec(gallivm, i32_type,
1701 1LL << DRAW_TOTAL_CLIP_PLANES);
1702 test = LLVMBuildAnd(builder, test, temp, "");
1703 mask = LLVMBuildOr(builder, mask, test, "");
1704 }
1705 return mask;
1706 }
1707
1708
1709 /**
1710 * Returns boolean if any clipping has occurred
1711 * Used zero/one i8 value to represent boolean
1712 */
1713 static LLVMValueRef
clipmask_booli8(struct gallivm_state * gallivm,const struct lp_type vs_type,LLVMValueRef clipmask_bool_ptr,boolean edgeflag_in_clipmask)1714 clipmask_booli8(struct gallivm_state *gallivm,
1715 const struct lp_type vs_type,
1716 LLVMValueRef clipmask_bool_ptr,
1717 boolean edgeflag_in_clipmask)
1718 {
1719 LLVMBuilderRef builder = gallivm->builder;
1720 LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1721 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1722 LLVMValueRef ret;
1723 struct lp_build_context bldivec;
1724
1725 lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1726
1727 /*
1728 * We need to invert the edgeflag bit from the clipmask here
1729 * (because the result is really if we want to run the pipeline or not
1730 * and we (may) need it if edgeflag was 0).
1731 */
1732 if (edgeflag_in_clipmask) {
1733 LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1734 1LL << DRAW_TOTAL_CLIP_PLANES);
1735 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1736 }
1737
1738 /*
1739 * XXX: probably should mask off bits from the mask which come from
1740 * vertices which were beyond the count (i.e. indices_valid for
1741 * linear fetches, for elts ones we don't have the correct mask
1742 * right now). Otherwise might run the pipeline for nothing,
1743 * though everything should still work.
1744 */
1745 ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1746 ret = LLVMBuildZExt(builder, ret, int8_type, "");
1747 return ret;
1748 }
1749
1750 static LLVMValueRef
draw_gs_llvm_fetch_input(const struct lp_build_gs_iface * gs_iface,struct lp_build_context * bld,boolean is_vindex_indirect,LLVMValueRef vertex_index,boolean is_aindex_indirect,LLVMValueRef attrib_index,LLVMValueRef swizzle_index)1751 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface *gs_iface,
1752 struct lp_build_context * bld,
1753 boolean is_vindex_indirect,
1754 LLVMValueRef vertex_index,
1755 boolean is_aindex_indirect,
1756 LLVMValueRef attrib_index,
1757 LLVMValueRef swizzle_index)
1758 {
1759 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1760 struct gallivm_state *gallivm = bld->gallivm;
1761 LLVMBuilderRef builder = gallivm->builder;
1762 LLVMValueRef indices[3];
1763 LLVMValueRef res;
1764 struct lp_type type = bld->type;
1765
1766 if (is_vindex_indirect || is_aindex_indirect) {
1767 int i;
1768 res = bld->zero;
1769 for (i = 0; i < type.length; ++i) {
1770 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1771 LLVMValueRef vert_chan_index = vertex_index;
1772 LLVMValueRef attr_chan_index = attrib_index;
1773 LLVMValueRef channel_vec, value;
1774
1775 if (is_vindex_indirect) {
1776 vert_chan_index = LLVMBuildExtractElement(builder,
1777 vertex_index, idx, "");
1778 }
1779 if (is_aindex_indirect) {
1780 attr_chan_index = LLVMBuildExtractElement(builder,
1781 attrib_index, idx, "");
1782 }
1783
1784 indices[0] = vert_chan_index;
1785 indices[1] = attr_chan_index;
1786 indices[2] = swizzle_index;
1787
1788 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1789 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1790 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1791
1792 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1793 }
1794 } else {
1795 indices[0] = vertex_index;
1796 indices[1] = attrib_index;
1797 indices[2] = swizzle_index;
1798
1799 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1800 res = LLVMBuildLoad(builder, res, "");
1801 }
1802
1803 return res;
1804 }
1805
1806 static void
draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface * gs_base,struct lp_build_context * bld,LLVMValueRef (* outputs)[4],LLVMValueRef emitted_vertices_vec,LLVMValueRef mask_vec,LLVMValueRef stream_id)1807 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface *gs_base,
1808 struct lp_build_context * bld,
1809 LLVMValueRef (*outputs)[4],
1810 LLVMValueRef emitted_vertices_vec,
1811 LLVMValueRef mask_vec, LLVMValueRef stream_id)
1812 {
1813 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1814 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1815 struct gallivm_state *gallivm = variant->gallivm;
1816 LLVMBuilderRef builder = gallivm->builder;
1817 struct lp_type gs_type = bld->type;
1818 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1819 lp_int_type(gs_type), 0);
1820 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1821 LLVMValueRef next_prim_offset =
1822 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1823 LLVMValueRef io = variant->io_ptr;
1824 unsigned i;
1825 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1826
1827 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
1828 for (i = 0; i < gs_type.length; ++i) {
1829 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1830 LLVMValueRef currently_emitted =
1831 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1832 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1833 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1834 indices[i] = LLVMBuildSelect(builder, LLVMBuildExtractElement(builder, cond, ind, ""), indices[i],
1835 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary - 1), "");
1836 }
1837
1838 LLVMValueRef stream_idx = LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), "");
1839 LLVMValueRef cnd = LLVMBuildICmp(builder, LLVMIntULT, stream_idx, lp_build_const_int32(gallivm, variant->shader->base.num_vertex_streams), "");
1840 struct lp_build_if_state if_ctx;
1841 lp_build_if(&if_ctx, gallivm, cnd);
1842 io = lp_build_pointer_get(builder, io, LLVMBuildExtractElement(builder, stream_id, lp_build_const_int32(gallivm, 0), ""));
1843
1844 if (variant->key.clamp_vertex_color) {
1845 do_clamp_vertex_color(gallivm, gs_type,
1846 gs_info, outputs);
1847 }
1848 convert_to_aos(gallivm, io, indices,
1849 outputs, clipmask,
1850 gs_info->num_outputs, gs_type,
1851 FALSE);
1852 lp_build_endif(&if_ctx);
1853 }
1854
1855 static void
draw_gs_llvm_end_primitive(const struct lp_build_gs_iface * gs_base,struct lp_build_context * bld,LLVMValueRef total_emitted_vertices_vec_ptr,LLVMValueRef verts_per_prim_vec,LLVMValueRef emitted_prims_vec,LLVMValueRef mask_vec,unsigned stream)1856 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface *gs_base,
1857 struct lp_build_context * bld,
1858 LLVMValueRef total_emitted_vertices_vec_ptr,
1859 LLVMValueRef verts_per_prim_vec,
1860 LLVMValueRef emitted_prims_vec,
1861 LLVMValueRef mask_vec, unsigned stream)
1862 {
1863 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1864 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1865 struct gallivm_state *gallivm = variant->gallivm;
1866 LLVMBuilderRef builder = gallivm->builder;
1867 LLVMValueRef prim_lengts_ptr =
1868 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1869 unsigned i;
1870
1871 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
1872 for (i = 0; i < bld->type.length; ++i) {
1873 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1874 LLVMValueRef prims_emitted =
1875 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1876 LLVMValueRef store_ptr;
1877 LLVMValueRef num_vertices =
1878 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1879
1880 LLVMValueRef this_cond = LLVMBuildExtractElement(gallivm->builder, cond, ind, "");
1881 struct lp_build_if_state ifthen;
1882 lp_build_if(&ifthen, gallivm, this_cond);
1883 prims_emitted = LLVMBuildMul(gallivm->builder, prims_emitted, lp_build_const_int32(gallivm, variant->shader->base.num_vertex_streams), "");
1884 prims_emitted = LLVMBuildAdd(gallivm->builder, prims_emitted, lp_build_const_int32(gallivm, stream), "");
1885 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1886 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1887 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1888 LLVMBuildStore(builder, num_vertices, store_ptr);
1889 lp_build_endif(&ifthen);
1890 }
1891 }
1892
1893 static void
draw_gs_llvm_epilogue(const struct lp_build_gs_iface * gs_base,LLVMValueRef total_emitted_vertices_vec,LLVMValueRef emitted_prims_vec,unsigned stream)1894 draw_gs_llvm_epilogue(const struct lp_build_gs_iface *gs_base,
1895 LLVMValueRef total_emitted_vertices_vec,
1896 LLVMValueRef emitted_prims_vec, unsigned stream)
1897 {
1898 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1899 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1900 struct gallivm_state *gallivm = variant->gallivm;
1901 LLVMBuilderRef builder = gallivm->builder;
1902 LLVMValueRef emitted_verts_ptr =
1903 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1904 LLVMValueRef emitted_prims_ptr =
1905 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1906 LLVMValueRef stream_val = lp_build_const_int32(gallivm, stream);
1907
1908 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &stream_val, 1, "");
1909 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &stream_val, 1, "");
1910
1911 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1912 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1913 }
1914
1915 static void
draw_llvm_generate(struct draw_llvm * llvm,struct draw_llvm_variant * variant)1916 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1917 {
1918 struct gallivm_state *gallivm = variant->gallivm;
1919 LLVMContextRef context = gallivm->context;
1920 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1921 LLVMTypeRef arg_types[13];
1922 unsigned num_arg_types = ARRAY_SIZE(arg_types);
1923 LLVMTypeRef func_type;
1924 LLVMValueRef context_ptr;
1925 LLVMBasicBlockRef block;
1926 LLVMBuilderRef builder;
1927 char func_name[64];
1928 struct lp_type vs_type;
1929 LLVMValueRef count, fetch_elts, start_or_maxelt;
1930 LLVMValueRef vertex_id_offset;
1931 LLVMValueRef stride, step, io_itr;
1932 LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1933 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1934 LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1935 LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1936 LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1937 LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1938 LLVMValueRef fake_buf_ptr, fake_buf;
1939
1940 struct draw_context *draw = llvm->draw;
1941 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1942 unsigned i, j;
1943 struct lp_build_context bld, blduivec;
1944 struct lp_build_loop_state lp_loop;
1945 struct lp_build_if_state if_ctx;
1946 const int vector_length = lp_native_vector_width / 32;
1947 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1948 struct lp_build_sampler_soa *sampler = 0;
1949 struct lp_build_image_soa *image = NULL;
1950 LLVMValueRef ret, clipmask_bool_ptr;
1951 struct draw_llvm_variant_key *key = &variant->key;
1952 /* If geometry shader is present we need to skip both the viewport
1953 * transformation and clipping otherwise the inputs to the geometry
1954 * shader will be incorrect.
1955 * The code can't handle vp transform when vs writes vp index neither
1956 * (though this would be fixable here, but couldn't just broadcast
1957 * the values).
1958 */
1959 const boolean bypass_viewport = key->has_gs_or_tes || key->bypass_viewport ||
1960 vs_info->writes_viewport_index;
1961 const boolean enable_cliptest = !key->has_gs_or_tes && (key->clip_xy ||
1962 key->clip_z ||
1963 key->clip_user ||
1964 key->need_edgeflags);
1965 LLVMValueRef variant_func;
1966 const unsigned pos = draw->vs.position_output;
1967 const unsigned cv = draw->vs.clipvertex_output;
1968 boolean have_clipdist = FALSE;
1969 struct lp_bld_tgsi_system_values system_values;
1970
1971 memset(&system_values, 0, sizeof(system_values));
1972 memset(&outputs, 0, sizeof(outputs));
1973 snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant");
1974
1975 i = 0;
1976 arg_types[i++] = get_context_ptr_type(variant); /* context */
1977 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1978 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1979 arg_types[i++] = int32_type; /* count */
1980 arg_types[i++] = int32_type; /* start/fetch_elt_max */
1981 arg_types[i++] = int32_type; /* stride */
1982 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1983 arg_types[i++] = int32_type; /* instance_id */
1984 arg_types[i++] = int32_type; /* vertex_id_offset */
1985 arg_types[i++] = int32_type; /* start_instance */
1986 arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
1987 arg_types[i++] = int32_type; /* draw_id */
1988 arg_types[i++] = int32_type; /* view_id */
1989
1990 func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1991 arg_types, num_arg_types, 0);
1992
1993 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1994 variant->function = variant_func;
1995
1996 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1997 for (i = 0; i < num_arg_types; ++i)
1998 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1999 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2000
2001 if (gallivm->cache && gallivm->cache->data_size)
2002 return;
2003 context_ptr = LLVMGetParam(variant_func, 0);
2004 io_ptr = LLVMGetParam(variant_func, 1);
2005 vbuffers_ptr = LLVMGetParam(variant_func, 2);
2006 count = LLVMGetParam(variant_func, 3);
2007 /*
2008 * XXX: the maxelt part is unused. Not really useful, since we cannot
2009 * get index buffer overflows due to vsplit (which provides its own
2010 * elts buffer, with a different size than what's passed in here).
2011 */
2012 start_or_maxelt = LLVMGetParam(variant_func, 4);
2013 /*
2014 * XXX: stride is actually unused. The stride we use is strictly calculated
2015 * from the number of outputs (including the draw_extra outputs).
2016 * Should probably fix some day (we need a new vs just because of extra
2017 * outputs which the generated vs won't touch).
2018 */
2019 stride = LLVMGetParam(variant_func, 5);
2020 vb_ptr = LLVMGetParam(variant_func, 6);
2021 system_values.instance_id = LLVMGetParam(variant_func, 7);
2022 vertex_id_offset = LLVMGetParam(variant_func, 8);
2023 system_values.base_instance = LLVMGetParam(variant_func, 9);
2024 fetch_elts = LLVMGetParam(variant_func, 10);
2025 system_values.draw_id = LLVMGetParam(variant_func, 11);
2026 system_values.view_index = LLVMGetParam(variant_func, 12);
2027
2028 lp_build_name(context_ptr, "context");
2029 lp_build_name(io_ptr, "io");
2030 lp_build_name(vbuffers_ptr, "vbuffers");
2031 lp_build_name(count, "count");
2032 lp_build_name(start_or_maxelt, "start_or_maxelt");
2033 lp_build_name(stride, "stride");
2034 lp_build_name(vb_ptr, "vb");
2035 lp_build_name(system_values.instance_id, "instance_id");
2036 lp_build_name(vertex_id_offset, "vertex_id_offset");
2037 lp_build_name(system_values.base_instance, "start_instance");
2038 lp_build_name(fetch_elts, "fetch_elts");
2039 lp_build_name(system_values.draw_id, "draw_id");
2040
2041 /*
2042 * Function body
2043 */
2044
2045 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2046 builder = gallivm->builder;
2047 LLVMPositionBuilderAtEnd(builder, block);
2048
2049 memset(&vs_type, 0, sizeof vs_type);
2050 vs_type.floating = TRUE; /* floating point values */
2051 vs_type.sign = TRUE; /* values are signed */
2052 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2053 vs_type.width = 32; /* 32-bit float */
2054 vs_type.length = vector_length;
2055
2056 lp_build_context_init(&bld, gallivm, lp_type_uint(32));
2057 lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
2058
2059 /* hold temporary "bool" clipmask */
2060 clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
2061
2062 fake_buf = lp_build_alloca_undef(gallivm,
2063 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
2064 fake_buf = LLVMBuildBitCast(builder, fake_buf,
2065 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
2066 fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
2067
2068 /* code generated texture sampling */
2069 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key), key->nr_samplers);
2070
2071 image = draw_llvm_image_soa_create(draw_llvm_variant_key_images(key),
2072 key->nr_images);
2073
2074 step = lp_build_const_int32(gallivm, vector_length);
2075
2076 ind_vec = blduivec.undef;
2077 for (i = 0; i < vs_type.length; i++) {
2078 LLVMValueRef index = lp_build_const_int32(gallivm, i);
2079 ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
2080 }
2081
2082 have_elts = LLVMBuildICmp(builder, LLVMIntNE,
2083 LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
2084
2085 fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
2086 fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
2087 /*
2088 * Only needed for non-indexed path.
2089 */
2090 start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
2091
2092 /*
2093 * Pre-calculate everything which is constant per shader invocation.
2094 */
2095 for (j = 0; j < key->nr_vertex_elements; ++j) {
2096 LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
2097 LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
2098 struct pipe_vertex_element *velem = &key->vertex_element[j];
2099 LLVMValueRef vb_index =
2100 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
2101 LLVMValueRef bsize = lp_build_const_int32(gallivm,
2102 util_format_get_blocksize(velem->src_format));
2103 LLVMValueRef src_offset = lp_build_const_int32(gallivm,
2104 velem->src_offset);
2105 struct lp_build_if_state if_ctx;
2106
2107 if (velem->src_format != PIPE_FORMAT_NONE) {
2108 vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
2109 vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
2110 vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
2111 vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
2112 LLVMInt32TypeInContext(context), "");
2113 vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
2114 map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
2115 buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
2116
2117 ofbit = NULL;
2118 /*
2119 * We'll set buffer_size_adj to zero if we have of, so it will
2120 * always overflow later automatically without having to keep ofbit.
2121 * Overflows (with normal wraparound) doing the actual offset
2122 * calculation should be ok, just not for the buffer size calc.
2123 * It would also be possible to detect such overflows and return
2124 * zeros if that happens, but this would be more complex.
2125 */
2126 buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
2127 tmp = lp_build_sub(&bld, bsize, bld.one);
2128 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
2129 &ofbit);
2130 buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
2131 buf_offset, &ofbit);
2132
2133 /*
2134 * We can't easily set fake vertex buffers outside the generated code.
2135 * Hence, set fake vertex buffers here instead basically, so fetch
2136 * code can always fetch using offset 0, eliminating all control flow
2137 * inside the main loop.
2138 * (Alternatively, could have control flow per vector skipping fetch
2139 * if ofbit is true.)
2140 */
2141 if (velem->instance_divisor) {
2142 /*
2143 * Index is equal to the start instance plus the number of current
2144 * instance divided by the divisor. In this case we compute it as:
2145 * index = start_instance + (instance_id / divisor).
2146 * Note we could actually do the fetch here, outside the loop -
2147 * it's all constant, hopefully llvm recognizes this.
2148 */
2149 LLVMValueRef current_instance;
2150 current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
2151 lp_build_const_int32(gallivm,
2152 velem->instance_divisor),
2153 "instance_divisor");
2154 instance_index[j] = lp_build_uadd_overflow(gallivm, system_values.base_instance,
2155 current_instance, &ofbit);
2156 }
2157
2158 buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
2159 buffer_size_adj[j], "");
2160
2161 temp_ptr = lp_build_alloca_undef(gallivm,
2162 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
2163
2164 lp_build_if(&if_ctx, gallivm, ofbit);
2165 {
2166 LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
2167 }
2168 lp_build_else(&if_ctx);
2169 {
2170 map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
2171 LLVMBuildStore(builder, map_ptr[j], temp_ptr);
2172 }
2173 lp_build_endif(&if_ctx);
2174 map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
2175
2176 if (0) {
2177 lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
2178 lp_build_const_int32(gallivm, j),
2179 vb_index, vb_stride[j]);
2180 lp_build_printf(gallivm,
2181 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
2182 vb_buffer_offset, src_offset, buf_offset);
2183 lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
2184 buffer_size, bsize);
2185 lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
2186 }
2187 }
2188 }
2189
2190 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
2191 {
2192 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
2193 LLVMValueRef io;
2194 LLVMValueRef clipmask; /* holds the clipmask value */
2195 LLVMValueRef true_index_array, index_store;
2196 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
2197
2198 io_itr = lp_loop.counter;
2199
2200 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
2201 #if DEBUG_STORE
2202 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
2203 io_itr, io, lp_loop.counter);
2204 #endif
2205
2206 true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
2207 true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
2208
2209 LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
2210 /*
2211 * Limit indices to fetch_max, otherwise might try to access indices
2212 * beyond index buffer (or rather vsplit elt buffer) size.
2213 * Could probably safely (?) skip this for non-indexed draws and
2214 * simplify things minimally (by removing it could combine the ind_vec
2215 * and start_vec adds). I think the only effect for non-indexed draws will
2216 * be that for the invalid elements they will be all fetched from the
2217 * same location as the last valid one, but noone should really care.
2218 */
2219 true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
2220
2221 index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
2222
2223 lp_build_if(&if_ctx, gallivm, have_elts);
2224 {
2225 /*
2226 * Note: you'd expect some comparison/clamp against fetch_elt_max
2227 * here.
2228 * There used to be one here but it was incorrect: overflow was
2229 * detected if index > fetch_elt_max - but the correct condition
2230 * would be index >= fetch_elt_max (since this is just size of elts
2231 * buffer / element size).
2232 * Using the correct condition however will cause failures - due to
2233 * vsplit/vcache code which rebases indices. So, as an example, if
2234 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
2235 * replace all invalid indices with 0 - which in case of elt_bias
2236 * not being zero will get a different fetch index than the valid
2237 * index 0. So, just rely on vsplit code preventing out-of-bounds
2238 * fetches. This is also why it's safe to do elts fetch even if there
2239 * was no index buffer bound - the real buffer is never seen here, at
2240 * least not if there are index buffer overflows...
2241 */
2242
2243 /*
2244 * XXX should not have to do this, as scale can be handled
2245 * natively by loads (hits asserts though).
2246 */
2247 tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
2248 fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
2249 LLVMPointerType(LLVMInt8TypeInContext(context),
2250 0), "");
2251 tmp = lp_build_gather(gallivm, vs_type.length,
2252 32, bld.type, TRUE,
2253 fetch_elts, tmp, FALSE);
2254 LLVMBuildStore(builder, tmp, index_store);
2255 }
2256 lp_build_else(&if_ctx);
2257 {
2258 tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
2259 LLVMBuildStore(builder, tmp, index_store);
2260 }
2261 lp_build_endif(&if_ctx);
2262
2263 true_index_array = LLVMBuildLoad(builder, index_store, "");
2264
2265 for (j = 0; j < key->nr_vertex_elements; ++j) {
2266 struct pipe_vertex_element *velem = &key->vertex_element[j];
2267 const struct util_format_description *format_desc =
2268 util_format_description(velem->src_format);
2269
2270 if (format_desc->format == PIPE_FORMAT_NONE) {
2271 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2272 inputs[j][i] = lp_build_zero(gallivm, vs_type);
2273 }
2274 }
2275 else if (velem->instance_divisor) {
2276 fetch_instanced(gallivm, format_desc, vs_type,
2277 vb_stride[j], map_ptr[j],
2278 buffer_size_adj[j],
2279 inputs[j], instance_index[j]);
2280 }
2281 else {
2282 fetch_vector(gallivm, format_desc, vs_type,
2283 vb_stride[j], map_ptr[j],
2284 buffer_size_adj[j],
2285 inputs[j], true_index_array);
2286 }
2287 }
2288
2289 struct lp_build_mask_context mask;
2290
2291 lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
2292 /* In the paths with elts vertex id has to be unaffected by the
2293 * index bias and because indices inside our elements array have
2294 * already had index bias applied we need to subtract it here to
2295 * get back to the original index.
2296 * in the linear paths vertex id has to be unaffected by the
2297 * original start index and because we abuse the 'start' variable
2298 * to either represent the actual start index or the index at which
2299 * the primitive was split (we split rendering into chunks of at
2300 * most 4095-vertices) we need to back out the original start
2301 * index out of our vertex id here.
2302 * for ARB_shader_draw_parameters, base_vertex should be 0 for non-indexed draws.
2303 */
2304 LLVMValueRef base_vertex = lp_build_select(&bld, have_elts, vertex_id_offset, lp_build_const_int32(gallivm, 0));
2305 system_values.basevertex = lp_build_broadcast_scalar(&blduivec, base_vertex);
2306 /* first vertex is for Vulkan base vertex support */
2307 LLVMValueRef first_vertex = lp_build_select(&bld, have_elts, vertex_id_offset, start_or_maxelt);
2308 system_values.firstvertex = lp_build_broadcast_scalar(&blduivec, first_vertex);
2309 system_values.vertex_id = true_index_array;
2310 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
2311 lp_build_broadcast_scalar(&blduivec, vertex_id_offset), "");
2312
2313 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
2314 generate_vs(variant,
2315 builder,
2316 vs_type,
2317 outputs,
2318 ptr_aos,
2319 &system_values,
2320 context_ptr,
2321 sampler,
2322 image,
2323 key->clamp_vertex_color,
2324 &mask);
2325
2326 lp_build_mask_end(&mask);
2327 if (pos != -1 && cv != -1) {
2328 /* store original positions in clip before further manipulation */
2329 store_clip(gallivm, vs_type, io, outputs, pos);
2330
2331 /* do cliptest */
2332 if (enable_cliptest) {
2333 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
2334 /* allocate clipmask, assign it integer type */
2335 clipmask = generate_clipmask(llvm,
2336 gallivm,
2337 vs_type,
2338 outputs,
2339 key,
2340 context_ptr, &have_clipdist);
2341 temp = LLVMBuildOr(builder, clipmask, temp, "");
2342 /* store temporary clipping boolean value */
2343 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
2344 }
2345 else {
2346 clipmask = blduivec.zero;
2347 }
2348
2349 /* do viewport mapping */
2350 if (!bypass_viewport) {
2351 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
2352 }
2353 }
2354 else {
2355 clipmask = blduivec.zero;
2356 }
2357
2358 /* store clipmask in vertex header,
2359 * original positions in clip
2360 * and transformed positions in data
2361 */
2362 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
2363 vs_info->num_outputs, vs_type,
2364 enable_cliptest && key->need_edgeflags);
2365 }
2366 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
2367
2368 sampler->destroy(sampler);
2369 image->destroy(image);
2370
2371 /* return clipping boolean value for function */
2372 ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
2373 enable_cliptest && key->need_edgeflags);
2374
2375 LLVMBuildRet(builder, ret);
2376
2377 gallivm_verify_function(gallivm, variant_func);
2378 }
2379
2380
2381 struct draw_llvm_variant_key *
draw_llvm_make_variant_key(struct draw_llvm * llvm,char * store)2382 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2383 {
2384 unsigned i;
2385 struct draw_llvm_variant_key *key;
2386 struct draw_sampler_static_state *draw_sampler;
2387 struct draw_image_static_state *draw_image;
2388
2389 key = (struct draw_llvm_variant_key *)store;
2390
2391 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
2392
2393
2394 /* will have to rig this up properly later */
2395 key->clip_xy = llvm->draw->clip_xy;
2396 key->clip_z = llvm->draw->clip_z;
2397 key->clip_user = llvm->draw->clip_user;
2398 key->bypass_viewport = llvm->draw->bypass_viewport;
2399 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
2400 /* XXX assumes edgeflag output not at 0 */
2401 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
2402 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
2403 key->has_gs_or_tes = llvm->draw->gs.geometry_shader != NULL || llvm->draw->tes.tess_eval_shader != NULL;
2404 key->num_outputs = draw_total_vs_outputs(llvm->draw);
2405
2406 key->clamp_vertex_color = !key->has_gs_or_tes &&
2407 llvm->draw->rasterizer->clamp_vertex_color;
2408
2409 /* All variants of this shader will have the same value for
2410 * nr_samplers. Not yet trying to compact away holes in the
2411 * sampler array.
2412 */
2413 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2414 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2415 key->nr_sampler_views =
2416 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2417 }
2418 else {
2419 key->nr_sampler_views = key->nr_samplers;
2420 }
2421
2422 key->nr_images = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
2423
2424 /* Presumably all variants of the shader should have the same
2425 * number of vertex elements - ie the number of shader inputs.
2426 * NOTE: we NEED to store the needed number of needed inputs
2427 * here, not the number of provided elements to match keysize
2428 * (and the offset of sampler state in the key).
2429 * If we have excess number of vertex elements, this is valid,
2430 * but the excess ones don't matter.
2431 * If we don't have enough vertex elements (which looks not really
2432 * valid but we'll handle it gracefully) fill out missing ones with
2433 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2434 */
2435 key->nr_vertex_elements =
2436 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2437
2438 if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2439 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2440 key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2441 memset(key->vertex_element, 0,
2442 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2443 }
2444 memcpy(key->vertex_element,
2445 llvm->draw->pt.vertex_element,
2446 sizeof(struct pipe_vertex_element) *
2447 MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2448
2449 draw_sampler = draw_llvm_variant_key_samplers(key);
2450 memset(draw_sampler, 0,
2451 MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2452
2453 for (i = 0 ; i < key->nr_samplers; i++) {
2454 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2455 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2456 }
2457 for (i = 0 ; i < key->nr_sampler_views; i++) {
2458 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2459 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2460 }
2461
2462 draw_image = draw_llvm_variant_key_images(key);
2463 memset(draw_image, 0,
2464 key->nr_images * sizeof *draw_image);
2465 for (i = 0; i < key->nr_images; i++) {
2466 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
2467 llvm->draw->images[PIPE_SHADER_VERTEX][i]);
2468 }
2469 return key;
2470 }
2471
2472
2473 void
draw_llvm_dump_variant_key(struct draw_llvm_variant_key * key)2474 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2475 {
2476 unsigned i;
2477 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2478 struct draw_image_static_state *image = draw_llvm_variant_key_images(key);
2479 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2480 debug_printf("clip_xy = %u\n", key->clip_xy);
2481 debug_printf("clip_z = %u\n", key->clip_z);
2482 debug_printf("clip_user = %u\n", key->clip_user);
2483 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2484 debug_printf("clip_halfz = %u\n", key->clip_halfz);
2485 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2486 debug_printf("has_gs_or_tes = %u\n", key->has_gs_or_tes);
2487 debug_printf("ucp_enable = %u\n", key->ucp_enable);
2488
2489 for (i = 0 ; i < key->nr_vertex_elements; i++) {
2490 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2491 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2492 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2493 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2494 }
2495
2496 for (i = 0 ; i < key->nr_sampler_views; i++) {
2497 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2498 }
2499
2500 for (i = 0 ; i < key->nr_images; i++)
2501 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
2502 }
2503
2504
2505 void
draw_llvm_set_mapped_texture(struct draw_context * draw,enum pipe_shader_type shader_stage,unsigned sview_idx,uint32_t width,uint32_t height,uint32_t depth,uint32_t first_level,uint32_t last_level,uint32_t num_samples,uint32_t sample_stride,const void * base_ptr,uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])2506 draw_llvm_set_mapped_texture(struct draw_context *draw,
2507 enum pipe_shader_type shader_stage,
2508 unsigned sview_idx,
2509 uint32_t width, uint32_t height, uint32_t depth,
2510 uint32_t first_level, uint32_t last_level,
2511 uint32_t num_samples,
2512 uint32_t sample_stride,
2513 const void *base_ptr,
2514 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2515 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2516 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2517 {
2518 unsigned j;
2519 struct draw_jit_texture *jit_tex;
2520
2521 switch (shader_stage) {
2522 case PIPE_SHADER_VERTEX:
2523 assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2524 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2525 break;
2526 case PIPE_SHADER_GEOMETRY:
2527 assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2528 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2529 break;
2530 case PIPE_SHADER_TESS_CTRL:
2531 assert(sview_idx < ARRAY_SIZE(draw->llvm->tcs_jit_context.textures));
2532 jit_tex = &draw->llvm->tcs_jit_context.textures[sview_idx];
2533 break;
2534 case PIPE_SHADER_TESS_EVAL:
2535 assert(sview_idx < ARRAY_SIZE(draw->llvm->tes_jit_context.textures));
2536 jit_tex = &draw->llvm->tes_jit_context.textures[sview_idx];
2537 break;
2538 default:
2539 assert(0);
2540 return;
2541 }
2542
2543 jit_tex->width = width;
2544 jit_tex->height = height;
2545 jit_tex->depth = depth;
2546 jit_tex->first_level = first_level;
2547 jit_tex->last_level = last_level;
2548 jit_tex->base = base_ptr;
2549 jit_tex->num_samples = num_samples;
2550 jit_tex->sample_stride = sample_stride;
2551
2552 for (j = first_level; j <= last_level; j++) {
2553 jit_tex->mip_offsets[j] = mip_offsets[j];
2554 jit_tex->row_stride[j] = row_stride[j];
2555 jit_tex->img_stride[j] = img_stride[j];
2556 }
2557 }
2558
2559 void
draw_llvm_set_mapped_image(struct draw_context * draw,enum pipe_shader_type shader_stage,unsigned idx,uint32_t width,uint32_t height,uint32_t depth,const void * base_ptr,uint32_t row_stride,uint32_t img_stride,uint32_t num_samples,uint32_t sample_stride)2560 draw_llvm_set_mapped_image(struct draw_context *draw,
2561 enum pipe_shader_type shader_stage,
2562 unsigned idx,
2563 uint32_t width, uint32_t height, uint32_t depth,
2564 const void *base_ptr,
2565 uint32_t row_stride,
2566 uint32_t img_stride,
2567 uint32_t num_samples,
2568 uint32_t sample_stride)
2569 {
2570 struct draw_jit_image *jit_image;
2571
2572 switch (shader_stage) {
2573 case PIPE_SHADER_VERTEX:
2574 assert(idx < ARRAY_SIZE(draw->llvm->jit_context.images));
2575 jit_image = &draw->llvm->jit_context.images[idx];
2576 break;
2577 case PIPE_SHADER_GEOMETRY:
2578 assert(idx < ARRAY_SIZE(draw->llvm->gs_jit_context.images));
2579 jit_image = &draw->llvm->gs_jit_context.images[idx];
2580 break;
2581 case PIPE_SHADER_TESS_CTRL:
2582 assert(idx < ARRAY_SIZE(draw->llvm->tcs_jit_context.images));
2583 jit_image = &draw->llvm->tcs_jit_context.images[idx];
2584 break;
2585 case PIPE_SHADER_TESS_EVAL:
2586 assert(idx < ARRAY_SIZE(draw->llvm->tes_jit_context.images));
2587 jit_image = &draw->llvm->tes_jit_context.images[idx];
2588 break;
2589 default:
2590 assert(0);
2591 return;
2592 }
2593
2594 jit_image->width = width;
2595 jit_image->height = height;
2596 jit_image->depth = depth;
2597 jit_image->base = base_ptr;
2598
2599 jit_image->row_stride = row_stride;
2600 jit_image->img_stride = img_stride;
2601 jit_image->num_samples = num_samples;
2602 jit_image->sample_stride = sample_stride;
2603 }
2604
2605
2606 void
draw_llvm_set_sampler_state(struct draw_context * draw,enum pipe_shader_type shader_type)2607 draw_llvm_set_sampler_state(struct draw_context *draw,
2608 enum pipe_shader_type shader_type)
2609 {
2610 unsigned i;
2611
2612 switch (shader_type) {
2613 case PIPE_SHADER_VERTEX:
2614 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2615 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2616
2617 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2618 const struct pipe_sampler_state *s
2619 = draw->samplers[PIPE_SHADER_VERTEX][i];
2620 jit_sam->min_lod = s->min_lod;
2621 jit_sam->max_lod = s->max_lod;
2622 jit_sam->lod_bias = s->lod_bias;
2623 jit_sam->max_aniso = s->max_anisotropy;
2624 COPY_4V(jit_sam->border_color, s->border_color.f);
2625 }
2626 }
2627 break;
2628 case PIPE_SHADER_GEOMETRY:
2629 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2630 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2631
2632 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2633 const struct pipe_sampler_state *s
2634 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2635 jit_sam->min_lod = s->min_lod;
2636 jit_sam->max_lod = s->max_lod;
2637 jit_sam->lod_bias = s->lod_bias;
2638 jit_sam->max_aniso = s->max_anisotropy;
2639 COPY_4V(jit_sam->border_color, s->border_color.f);
2640 }
2641 }
2642 break;
2643 case PIPE_SHADER_TESS_CTRL:
2644 for (i = 0; i < draw->num_samplers[PIPE_SHADER_TESS_CTRL]; i++) {
2645 struct draw_jit_sampler *jit_sam = &draw->llvm->tcs_jit_context.samplers[i];
2646
2647 if (draw->samplers[PIPE_SHADER_TESS_CTRL][i]) {
2648 const struct pipe_sampler_state *s
2649 = draw->samplers[PIPE_SHADER_TESS_CTRL][i];
2650 jit_sam->min_lod = s->min_lod;
2651 jit_sam->max_lod = s->max_lod;
2652 jit_sam->lod_bias = s->lod_bias;
2653 jit_sam->max_aniso = s->max_anisotropy;
2654 COPY_4V(jit_sam->border_color, s->border_color.f);
2655 }
2656 }
2657 break;
2658 case PIPE_SHADER_TESS_EVAL:
2659 for (i = 0; i < draw->num_samplers[PIPE_SHADER_TESS_EVAL]; i++) {
2660 struct draw_jit_sampler *jit_sam = &draw->llvm->tes_jit_context.samplers[i];
2661
2662 if (draw->samplers[PIPE_SHADER_TESS_EVAL][i]) {
2663 const struct pipe_sampler_state *s
2664 = draw->samplers[PIPE_SHADER_TESS_EVAL][i];
2665 jit_sam->min_lod = s->min_lod;
2666 jit_sam->max_lod = s->max_lod;
2667 jit_sam->lod_bias = s->lod_bias;
2668 jit_sam->max_aniso = s->max_anisotropy;
2669 COPY_4V(jit_sam->border_color, s->border_color.f);
2670 }
2671 }
2672 break;
2673 default:
2674 assert(0);
2675 break;
2676 }
2677 }
2678
2679
2680 void
draw_llvm_destroy_variant(struct draw_llvm_variant * variant)2681 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2682 {
2683 struct draw_llvm *llvm = variant->llvm;
2684
2685 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2686 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2687 variant->shader->variants_cached, llvm->nr_variants);
2688 }
2689
2690 gallivm_destroy(variant->gallivm);
2691
2692 remove_from_list(&variant->list_item_local);
2693 variant->shader->variants_cached--;
2694 remove_from_list(&variant->list_item_global);
2695 llvm->nr_variants--;
2696 FREE(variant);
2697 }
2698
2699
2700 /**
2701 * Create LLVM types for various structures.
2702 */
2703 static void
create_gs_jit_types(struct draw_gs_llvm_variant * var)2704 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2705 {
2706 struct gallivm_state *gallivm = var->gallivm;
2707 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
2708
2709 texture_type = create_jit_texture_type(gallivm, "texture");
2710 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2711 image_type = create_jit_image_type(gallivm, "image");
2712
2713 context_type = create_gs_jit_context_type(gallivm,
2714 var->shader->base.vector_length,
2715 texture_type, sampler_type,
2716 image_type,
2717 "draw_gs_jit_context");
2718 var->context_ptr_type = LLVMPointerType(context_type, 0);
2719
2720 var->input_array_type = create_gs_jit_input_type(gallivm);
2721 }
2722
2723 static LLVMTypeRef
get_gs_context_ptr_type(struct draw_gs_llvm_variant * variant)2724 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2725 {
2726 if (!variant->context_ptr_type)
2727 create_gs_jit_types(variant);
2728 return variant->context_ptr_type;
2729 }
2730
2731 static LLVMValueRef
generate_mask_value(struct draw_gs_llvm_variant * variant,struct lp_type gs_type)2732 generate_mask_value(struct draw_gs_llvm_variant *variant,
2733 struct lp_type gs_type)
2734 {
2735 struct gallivm_state *gallivm = variant->gallivm;
2736 LLVMBuilderRef builder = gallivm->builder;
2737 struct lp_type mask_type = lp_int_type(gs_type);
2738 LLVMValueRef num_prims;
2739 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2740 unsigned i;
2741
2742 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2743 variant->num_prims);
2744 for (i = 0; i < gs_type.length; i++) {
2745 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2746 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2747 }
2748 mask_val = lp_build_compare(gallivm, mask_type,
2749 PIPE_FUNC_GREATER, num_prims, mask_val);
2750
2751 return mask_val;
2752 }
2753
2754 static void
draw_gs_llvm_generate(struct draw_llvm * llvm,struct draw_gs_llvm_variant * variant)2755 draw_gs_llvm_generate(struct draw_llvm *llvm,
2756 struct draw_gs_llvm_variant *variant)
2757 {
2758 struct gallivm_state *gallivm = variant->gallivm;
2759 LLVMContextRef context = gallivm->context;
2760 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2761 LLVMTypeRef arg_types[8];
2762 LLVMTypeRef func_type;
2763 LLVMValueRef variant_func;
2764 LLVMValueRef context_ptr;
2765 LLVMValueRef prim_id_ptr;
2766 LLVMBasicBlockRef block;
2767 LLVMBuilderRef builder;
2768 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2769 struct lp_build_sampler_soa *sampler = 0;
2770 struct lp_build_image_soa *image = NULL;
2771 struct lp_build_context bld;
2772 struct lp_bld_tgsi_system_values system_values;
2773 char func_name[64];
2774 struct lp_type gs_type;
2775 unsigned i;
2776 struct draw_gs_llvm_iface gs_iface;
2777 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2778 LLVMValueRef consts_ptr, num_consts_ptr;
2779 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
2780 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2781 struct lp_build_mask_context mask;
2782 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2783 unsigned vector_length = variant->shader->base.vector_length;
2784
2785 memset(&system_values, 0, sizeof(system_values));
2786 memset(&outputs, 0, sizeof(outputs));
2787
2788 snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant");
2789
2790 assert(variant->vertex_header_ptr_type);
2791
2792 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2793 arg_types[1] = variant->input_array_type; /* input */
2794 arg_types[2] = LLVMPointerType(variant->vertex_header_ptr_type, 0); /* vertex_header */
2795 arg_types[3] = int32_type; /* num_prims */
2796 arg_types[4] = int32_type; /* instance_id */
2797 arg_types[5] = LLVMPointerType(
2798 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2799 arg_types[6] = int32_type;
2800 arg_types[7] = int32_type;
2801
2802 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2803
2804 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2805
2806 variant->function = variant_func;
2807
2808 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2809
2810 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2811 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2812 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2813
2814 if (gallivm->cache && gallivm->cache->data_size)
2815 return;
2816 context_ptr = LLVMGetParam(variant_func, 0);
2817 input_array = LLVMGetParam(variant_func, 1);
2818 io_ptr = LLVMGetParam(variant_func, 2);
2819 num_prims = LLVMGetParam(variant_func, 3);
2820 system_values.instance_id = LLVMGetParam(variant_func, 4);
2821 prim_id_ptr = LLVMGetParam(variant_func, 5);
2822 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2823 system_values.view_index = LLVMGetParam(variant_func, 7);
2824
2825 lp_build_name(context_ptr, "context");
2826 lp_build_name(input_array, "input");
2827 lp_build_name(io_ptr, "io");
2828 lp_build_name(num_prims, "num_prims");
2829 lp_build_name(system_values.instance_id, "instance_id");
2830 lp_build_name(prim_id_ptr, "prim_id_ptr");
2831 lp_build_name(system_values.invocation_id, "invocation_id");
2832 lp_build_name(system_values.view_index, "view_index");
2833
2834 variant->context_ptr = context_ptr;
2835 variant->io_ptr = io_ptr;
2836 variant->num_prims = num_prims;
2837
2838 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2839 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2840 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2841 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2842 gs_iface.input = input_array;
2843 gs_iface.variant = variant;
2844
2845 /*
2846 * Function body
2847 */
2848
2849 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2850 builder = gallivm->builder;
2851 LLVMPositionBuilderAtEnd(builder, block);
2852
2853 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2854
2855 memset(&gs_type, 0, sizeof gs_type);
2856 gs_type.floating = TRUE; /* floating point values */
2857 gs_type.sign = TRUE; /* values are signed */
2858 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2859 gs_type.width = 32; /* 32-bit float */
2860 gs_type.length = vector_length;
2861
2862 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2863 num_consts_ptr =
2864 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2865
2866 ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
2867 num_ssbos_ptr =
2868 draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
2869
2870 /* code generated texture sampling */
2871 sampler = draw_llvm_sampler_soa_create(variant->key.samplers, variant->key.nr_samplers);
2872 image = draw_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant->key),
2873 variant->key.nr_images);
2874 mask_val = generate_mask_value(variant, gs_type);
2875 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2876
2877 if (gs_info->uses_primid) {
2878 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2879 }
2880
2881 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2882 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2883 tgsi_dump(tokens, 0);
2884 else
2885 nir_print_shader(llvm->draw->gs.geometry_shader->state.ir.nir, stderr);
2886 draw_gs_llvm_dump_variant_key(&variant->key);
2887 }
2888
2889 struct lp_build_tgsi_params params;
2890 memset(¶ms, 0, sizeof(params));
2891
2892 params.type = gs_type;
2893 params.mask = &mask;
2894 params.consts_ptr = consts_ptr;
2895 params.const_sizes_ptr = num_consts_ptr;
2896 params.system_values = &system_values;
2897 params.context_ptr = context_ptr;
2898 params.sampler = sampler;
2899 params.info = &llvm->draw->gs.geometry_shader->info;
2900 params.gs_iface = (const struct lp_build_gs_iface *)&gs_iface;
2901 params.ssbo_ptr = ssbos_ptr;
2902 params.ssbo_sizes_ptr = num_ssbos_ptr;
2903 params.image = image;
2904 params.gs_vertex_streams = variant->shader->base.num_vertex_streams;
2905 params.aniso_filter_table = draw_gs_jit_context_aniso_filter_table(gallivm, context_ptr);
2906
2907 if (llvm->draw->gs.geometry_shader->state.type == PIPE_SHADER_IR_TGSI)
2908 lp_build_tgsi_soa(variant->gallivm,
2909 tokens,
2910 ¶ms,
2911 outputs);
2912 else
2913 lp_build_nir_soa(variant->gallivm,
2914 llvm->draw->gs.geometry_shader->state.ir.nir,
2915 ¶ms,
2916 outputs);
2917
2918 sampler->destroy(sampler);
2919 image->destroy(image);
2920
2921 lp_build_mask_end(&mask);
2922
2923 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2924
2925 gallivm_verify_function(gallivm, variant_func);
2926 }
2927
2928 struct draw_gs_llvm_variant *
draw_gs_llvm_create_variant(struct draw_llvm * llvm,unsigned num_outputs,const struct draw_gs_llvm_variant_key * key)2929 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2930 unsigned num_outputs,
2931 const struct draw_gs_llvm_variant_key *key)
2932 {
2933 struct draw_gs_llvm_variant *variant;
2934 struct llvm_geometry_shader *shader =
2935 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2936 LLVMTypeRef vertex_header;
2937 char module_name[64];
2938 unsigned char ir_sha1_cache_key[20];
2939 struct lp_cached_code cached = { 0 };
2940 bool needs_caching = false;
2941
2942 variant = MALLOC(sizeof *variant +
2943 shader->variant_key_size -
2944 sizeof variant->key);
2945 if (!variant)
2946 return NULL;
2947
2948 variant->llvm = llvm;
2949 variant->shader = shader;
2950
2951 snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2952 variant->shader->variants_cached);
2953
2954 memcpy(&variant->key, key, shader->variant_key_size);
2955
2956 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
2957 draw_get_ir_cache_key(shader->base.state.ir.nir,
2958 key,
2959 shader->variant_key_size,
2960 num_outputs,
2961 ir_sha1_cache_key);
2962
2963 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
2964 &cached,
2965 ir_sha1_cache_key);
2966 if (!cached.data_size)
2967 needs_caching = true;
2968 }
2969 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
2970
2971 create_gs_jit_types(variant);
2972
2973 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2974
2975 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2976
2977 draw_gs_llvm_generate(llvm, variant);
2978
2979 gallivm_compile_module(variant->gallivm);
2980
2981 variant->jit_func = (draw_gs_jit_func)
2982 gallivm_jit_function(variant->gallivm, variant->function);
2983
2984 if (needs_caching)
2985 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
2986 &cached,
2987 ir_sha1_cache_key);
2988 gallivm_free_ir(variant->gallivm);
2989
2990 variant->list_item_global.base = variant;
2991 variant->list_item_local.base = variant;
2992 /*variant->no = */shader->variants_created++;
2993 variant->list_item_global.base = variant;
2994
2995 return variant;
2996 }
2997
2998 void
draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant * variant)2999 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
3000 {
3001 struct draw_llvm *llvm = variant->llvm;
3002
3003 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
3004 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
3005 variant->shader->variants_cached, llvm->nr_gs_variants);
3006 }
3007
3008 gallivm_destroy(variant->gallivm);
3009
3010 remove_from_list(&variant->list_item_local);
3011 variant->shader->variants_cached--;
3012 remove_from_list(&variant->list_item_global);
3013 llvm->nr_gs_variants--;
3014 FREE(variant);
3015 }
3016
3017 struct draw_gs_llvm_variant_key *
draw_gs_llvm_make_variant_key(struct draw_llvm * llvm,char * store)3018 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
3019 {
3020 unsigned i;
3021 struct draw_gs_llvm_variant_key *key;
3022 struct draw_sampler_static_state *draw_sampler;
3023 struct draw_image_static_state *draw_image;
3024
3025 key = (struct draw_gs_llvm_variant_key *)store;
3026
3027 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
3028
3029 key->num_outputs = draw_total_gs_outputs(llvm->draw);
3030
3031 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color;
3032
3033 /* All variants of this shader will have the same value for
3034 * nr_samplers. Not yet trying to compact away holes in the
3035 * sampler array.
3036 */
3037 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
3038 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3039 key->nr_sampler_views =
3040 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3041 }
3042 else {
3043 key->nr_sampler_views = key->nr_samplers;
3044 }
3045
3046 key->nr_images = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
3047
3048 draw_sampler = key->samplers;
3049
3050 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
3051
3052 for (i = 0 ; i < key->nr_samplers; i++) {
3053 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
3054 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
3055 }
3056 for (i = 0 ; i < key->nr_sampler_views; i++) {
3057 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
3058 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
3059 }
3060
3061 draw_image = draw_gs_llvm_variant_key_images(key);
3062 memset(draw_image, 0,
3063 key->nr_images * sizeof *draw_image);
3064 for (i = 0; i < key->nr_images; i++) {
3065 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
3066 llvm->draw->images[PIPE_SHADER_GEOMETRY][i]);
3067 }
3068 return key;
3069 }
3070
3071 void
draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key * key)3072 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
3073 {
3074 unsigned i;
3075 struct draw_sampler_static_state *sampler = key->samplers;
3076 struct draw_image_static_state *image = draw_gs_llvm_variant_key_images(key);
3077
3078 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
3079 for (i = 0 ; i < key->nr_sampler_views; i++) {
3080 debug_printf("sampler[%i].src_format = %s\n", i,
3081 util_format_name(sampler[i].texture_state.format));
3082 }
3083
3084 for (i = 0 ; i < key->nr_images; i++)
3085 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
3086
3087 }
3088
3089 static void
create_tcs_jit_types(struct draw_tcs_llvm_variant * var)3090 create_tcs_jit_types(struct draw_tcs_llvm_variant *var)
3091 {
3092 struct gallivm_state *gallivm = var->gallivm;
3093 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
3094
3095 texture_type = create_jit_texture_type(gallivm, "texture");
3096 sampler_type = create_jit_sampler_type(gallivm, "sampler");
3097 image_type = create_jit_image_type(gallivm, "image");
3098
3099 context_type = create_tcs_jit_context_type(gallivm,
3100 0,
3101 texture_type, sampler_type,
3102 image_type,
3103 "draw_tcs_jit_context");
3104 var->input_array_type = create_tcs_jit_input_type(gallivm);
3105 var->output_array_type = create_tcs_jit_output_type(gallivm);
3106 var->context_ptr_type = LLVMPointerType(context_type, 0);
3107 }
3108
3109 static LLVMTypeRef
get_tcs_context_ptr_type(struct draw_tcs_llvm_variant * variant)3110 get_tcs_context_ptr_type(struct draw_tcs_llvm_variant *variant)
3111 {
3112 if (!variant->context_ptr_type)
3113 create_tcs_jit_types(variant);
3114 return variant->context_ptr_type;
3115 }
3116
3117 static LLVMValueRef
draw_tcs_llvm_emit_fetch_input(const struct lp_build_tcs_iface * tes_iface,struct lp_build_context * bld,boolean is_vindex_indirect,LLVMValueRef vertex_index,boolean is_aindex_indirect,LLVMValueRef attrib_index,boolean is_sindex_indirect,LLVMValueRef swizzle_index)3118 draw_tcs_llvm_emit_fetch_input(const struct lp_build_tcs_iface *tes_iface,
3119 struct lp_build_context *bld,
3120 boolean is_vindex_indirect,
3121 LLVMValueRef vertex_index,
3122 boolean is_aindex_indirect,
3123 LLVMValueRef attrib_index,
3124 boolean is_sindex_indirect,
3125 LLVMValueRef swizzle_index)
3126 {
3127 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
3128 struct gallivm_state *gallivm = bld->gallivm;
3129 LLVMBuilderRef builder = gallivm->builder;
3130 LLVMValueRef indices[3];
3131 LLVMValueRef res;
3132 struct lp_type type = bld->type;
3133
3134 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
3135 int i;
3136
3137 res = bld->zero;
3138 for (i = 0; i < type.length; ++i) {
3139 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3140 LLVMValueRef vert_chan_index = vertex_index;
3141 LLVMValueRef attr_chan_index = attrib_index;
3142 LLVMValueRef swiz_chan_index = swizzle_index;
3143 LLVMValueRef channel_vec;
3144
3145 if (is_vindex_indirect) {
3146 vert_chan_index = LLVMBuildExtractElement(builder,
3147 vertex_index, idx, "");
3148 }
3149 if (is_aindex_indirect) {
3150 attr_chan_index = LLVMBuildExtractElement(builder,
3151 attrib_index, idx, "");
3152 }
3153 if (is_sindex_indirect) {
3154 swiz_chan_index = LLVMBuildExtractElement(builder,
3155 swizzle_index, idx, "");
3156 }
3157
3158 indices[0] = vert_chan_index;
3159 indices[1] = attr_chan_index;
3160 indices[2] = swiz_chan_index;
3161
3162 channel_vec = LLVMBuildGEP(builder, tcs->input, indices, 3, "");
3163 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
3164
3165 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3166 }
3167 } else {
3168 indices[0] = vertex_index;
3169 indices[1] = attrib_index;
3170 indices[2] = swizzle_index;
3171
3172 res = LLVMBuildGEP(builder, tcs->input, indices, 3, "");
3173 res = LLVMBuildLoad(builder, res, "");
3174 res = lp_build_broadcast_scalar(bld, res);
3175 }
3176 return res;
3177 }
3178
3179 static LLVMValueRef
draw_tcs_llvm_emit_fetch_output(const struct lp_build_tcs_iface * tes_iface,struct lp_build_context * bld,boolean is_vindex_indirect,LLVMValueRef vertex_index,boolean is_aindex_indirect,LLVMValueRef attrib_index,boolean is_sindex_indirect,LLVMValueRef swizzle_index,uint32_t name)3180 draw_tcs_llvm_emit_fetch_output(const struct lp_build_tcs_iface *tes_iface,
3181 struct lp_build_context *bld,
3182 boolean is_vindex_indirect,
3183 LLVMValueRef vertex_index,
3184 boolean is_aindex_indirect,
3185 LLVMValueRef attrib_index,
3186 boolean is_sindex_indirect,
3187 LLVMValueRef swizzle_index,
3188 uint32_t name)
3189 {
3190 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
3191 struct gallivm_state *gallivm = bld->gallivm;
3192 LLVMBuilderRef builder = gallivm->builder;
3193 LLVMValueRef indices[3];
3194 LLVMValueRef res;
3195 struct lp_type type = bld->type;
3196
3197 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
3198 int i;
3199
3200 res = bld->zero;
3201 for (i = 0; i < type.length; ++i) {
3202 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3203 LLVMValueRef vert_chan_index = vertex_index;
3204 LLVMValueRef attr_chan_index = attrib_index;
3205 LLVMValueRef swiz_chan_index = swizzle_index;
3206 LLVMValueRef channel_vec;
3207
3208 if (is_vindex_indirect) {
3209 vert_chan_index = LLVMBuildExtractElement(builder,
3210 vertex_index, idx, "");
3211 }
3212 if (is_aindex_indirect) {
3213 attr_chan_index = LLVMBuildExtractElement(builder,
3214 attrib_index, idx, "");
3215 }
3216 if (is_sindex_indirect) {
3217 swiz_chan_index = LLVMBuildExtractElement(builder,
3218 swizzle_index, idx, "");
3219 }
3220
3221 indices[0] = vert_chan_index;
3222 indices[1] = attr_chan_index;
3223 indices[2] = swiz_chan_index;
3224
3225 channel_vec = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3226 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
3227
3228 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3229 }
3230 } else {
3231 indices[0] = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
3232 indices[1] = attrib_index;
3233 indices[2] = swizzle_index;
3234
3235 res = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3236 res = LLVMBuildLoad(builder, res, "");
3237 res = lp_build_broadcast_scalar(bld, res);
3238 }
3239 return res;
3240 }
3241
3242 static void
draw_tcs_llvm_emit_store_output(const struct lp_build_tcs_iface * tes_iface,struct lp_build_context * bld,unsigned name,boolean is_vindex_indirect,LLVMValueRef vertex_index,boolean is_aindex_indirect,LLVMValueRef attrib_index,boolean is_sindex_indirect,LLVMValueRef swizzle_index,LLVMValueRef value,LLVMValueRef mask_vec)3243 draw_tcs_llvm_emit_store_output(const struct lp_build_tcs_iface *tes_iface,
3244 struct lp_build_context *bld,
3245 unsigned name,
3246 boolean is_vindex_indirect,
3247 LLVMValueRef vertex_index,
3248 boolean is_aindex_indirect,
3249 LLVMValueRef attrib_index,
3250 boolean is_sindex_indirect,
3251 LLVMValueRef swizzle_index,
3252 LLVMValueRef value,
3253 LLVMValueRef mask_vec)
3254 {
3255 const struct draw_tcs_llvm_iface *tcs = draw_tcs_llvm_iface(tes_iface);
3256 struct gallivm_state *gallivm = bld->gallivm;
3257 LLVMBuilderRef builder = gallivm->builder;
3258 LLVMValueRef indices[3];
3259 LLVMValueRef res;
3260 struct lp_type type = bld->type;
3261
3262 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
3263 int i;
3264
3265 for (i = 0; i < type.length; ++i) {
3266 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3267 LLVMValueRef vert_chan_index = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
3268 LLVMValueRef attr_chan_index = attrib_index;
3269 LLVMValueRef swiz_chan_index = swizzle_index;
3270 LLVMValueRef channel_vec;
3271
3272 if (is_vindex_indirect) {
3273 vert_chan_index = LLVMBuildExtractElement(builder,
3274 vertex_index, idx, "");
3275 }
3276 if (is_aindex_indirect) {
3277 attr_chan_index = LLVMBuildExtractElement(builder,
3278 attrib_index, idx, "");
3279 }
3280
3281 if (is_sindex_indirect) {
3282 swiz_chan_index = LLVMBuildExtractElement(builder,
3283 swizzle_index, idx, "");
3284 }
3285
3286 indices[0] = vert_chan_index;
3287 indices[1] = attr_chan_index;
3288 indices[2] = swiz_chan_index;
3289
3290 channel_vec = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3291
3292 res = LLVMBuildExtractElement(builder, value, idx, "");
3293
3294 struct lp_build_if_state ifthen;
3295 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
3296 cond = LLVMBuildExtractElement(gallivm->builder, cond, idx, "");
3297 lp_build_if(&ifthen, gallivm, cond);
3298 LLVMBuildStore(builder, res, channel_vec);
3299 lp_build_endif(&ifthen);
3300 }
3301 } else {
3302 indices[0] = vertex_index ? vertex_index : lp_build_const_int32(gallivm, 0);
3303 indices[1] = attrib_index;
3304 indices[2] = swizzle_index;
3305
3306 res = LLVMBuildGEP(builder, tcs->output, indices, 3, "");
3307 for (unsigned i = 0; i < type.length; ++i) {
3308 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3309 LLVMValueRef val = LLVMBuildExtractElement(builder, value, idx, "");
3310
3311 struct lp_build_if_state ifthen;
3312 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, mask_vec, lp_build_const_int_vec(gallivm, bld->type, 0), "");
3313 cond = LLVMBuildExtractElement(gallivm->builder, cond, idx, "");
3314 lp_build_if(&ifthen, gallivm, cond);
3315 LLVMBuildStore(builder, val, res);
3316 lp_build_endif(&ifthen);
3317 }
3318 }
3319 }
3320
3321
3322 static LLVMValueRef
generate_tcs_mask_value(struct draw_tcs_llvm_variant * variant,struct lp_type tcs_type,LLVMValueRef limit,LLVMValueRef loop_counter)3323 generate_tcs_mask_value(struct draw_tcs_llvm_variant *variant,
3324 struct lp_type tcs_type, LLVMValueRef limit, LLVMValueRef loop_counter)
3325 {
3326 struct gallivm_state *gallivm = variant->gallivm;
3327 LLVMBuilderRef builder = gallivm->builder;
3328 struct lp_type mask_type = lp_int_type(tcs_type);
3329 LLVMValueRef num_vecs;
3330 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
3331 unsigned i;
3332
3333 num_vecs = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type), limit);
3334 for (i = 0; i < tcs_type.length; i++) {
3335 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3336 mask_val = LLVMBuildInsertElement(builder, mask_val, LLVMBuildAdd(builder, loop_counter, idx, ""), idx, "");
3337 }
3338 mask_val = lp_build_compare(gallivm, mask_type,
3339 PIPE_FUNC_GREATER, num_vecs, mask_val);
3340
3341 return mask_val;
3342 }
3343
3344 static void
draw_tcs_llvm_generate(struct draw_llvm * llvm,struct draw_tcs_llvm_variant * variant)3345 draw_tcs_llvm_generate(struct draw_llvm *llvm,
3346 struct draw_tcs_llvm_variant *variant)
3347 {
3348 struct gallivm_state *gallivm = variant->gallivm;
3349 LLVMContextRef context = gallivm->context;
3350 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
3351 LLVMTypeRef arg_types[7];
3352 LLVMTypeRef func_type, coro_func_type;
3353 LLVMValueRef variant_func, variant_coro;
3354 LLVMValueRef context_ptr;
3355 LLVMValueRef view_index;
3356 LLVMValueRef input_array, output_array, prim_id, patch_vertices_in;
3357 LLVMValueRef mask_val;
3358 LLVMBasicBlockRef block;
3359 LLVMBuilderRef builder;
3360 struct lp_build_context bld, bldvec;
3361 struct lp_build_sampler_soa *sampler = 0;
3362 struct lp_build_image_soa *image = NULL;
3363 struct lp_bld_tgsi_system_values system_values;
3364 char func_name[64], func_name_coro[64];
3365 unsigned i;
3366 struct draw_tcs_llvm_iface tcs_iface;
3367 struct lp_build_mask_context mask;
3368 LLVMValueRef consts_ptr, num_consts_ptr;
3369 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
3370 struct lp_type tcs_type;
3371 unsigned vector_length = variant->shader->base.vector_length;
3372
3373 memset(&system_values, 0, sizeof(system_values));
3374
3375 snprintf(func_name, sizeof(func_name), "draw_llvm_tcs_variant");
3376
3377 snprintf(func_name_coro, sizeof(func_name_coro), "draw_llvm_tcs_coro_variant");
3378
3379 arg_types[0] = get_tcs_context_ptr_type(variant); /* context */
3380 arg_types[1] = variant->input_array_type; /* input */
3381 arg_types[2] = variant->output_array_type;
3382 arg_types[3] = int32_type;
3383 arg_types[4] = int32_type;
3384 arg_types[5] = int32_type;
3385 arg_types[6] = int32_type; /* coroutine only */
3386
3387 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types) - 1, 0);
3388
3389 coro_func_type = LLVMFunctionType(LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), arg_types, ARRAY_SIZE(arg_types), 0);
3390
3391 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
3392
3393 variant_coro = LLVMAddFunction(gallivm->module, func_name_coro, coro_func_type);
3394
3395 variant->function = variant_func;
3396 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
3397
3398 LLVMSetFunctionCallConv(variant_coro, LLVMCCallConv);
3399
3400 for (i = 0; i < ARRAY_SIZE(arg_types); ++i) {
3401 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
3402 lp_add_function_attr(variant_coro, i + 1, LP_FUNC_ATTR_NOALIAS);
3403 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
3404 }
3405 }
3406
3407 if (gallivm->cache && gallivm->cache->data_size)
3408 return;
3409 context_ptr = LLVMGetParam(variant_func, 0);
3410 input_array = LLVMGetParam(variant_func, 1);
3411 output_array = LLVMGetParam(variant_func, 2);
3412 prim_id = LLVMGetParam(variant_func, 3);
3413 patch_vertices_in = LLVMGetParam(variant_func, 4);
3414 view_index = LLVMGetParam(variant_func, 5);
3415
3416 lp_build_name(context_ptr, "context");
3417 lp_build_name(input_array, "input");
3418 lp_build_name(output_array, "output");
3419 lp_build_name(prim_id, "prim_id");
3420 lp_build_name(patch_vertices_in, "patch_vertices_in");
3421 lp_build_name(view_index, "view_index");
3422
3423 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
3424 builder = gallivm->builder;
3425 LLVMPositionBuilderAtEnd(builder, block);
3426
3427 lp_build_context_init(&bld, gallivm, lp_type_int(32));
3428
3429 memset(&tcs_type, 0, sizeof tcs_type);
3430 tcs_type.floating = TRUE; /* floating point values */
3431 tcs_type.sign = TRUE; /* values are signed */
3432 tcs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
3433 tcs_type.width = 32; /* 32-bit float */
3434 tcs_type.length = vector_length;
3435
3436 lp_build_context_init(&bldvec, variant->gallivm, lp_int_type(tcs_type));
3437
3438 LLVMValueRef count = lp_build_const_int32(gallivm, variant->shader->base.vertices_out);
3439 LLVMValueRef step = lp_build_const_int32(gallivm, vector_length);
3440
3441 struct lp_build_loop_state loop_state[2];
3442 LLVMValueRef num_inner_loop;
3443 unsigned count_align = util_align_npot(variant->shader->base.vertices_out, tcs_type.length);
3444 num_inner_loop = lp_build_const_int32(gallivm, count_align / tcs_type.length);
3445 LLVMTypeRef hdl_ptr_type = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
3446 LLVMValueRef coro_hdls = LLVMBuildArrayAlloca(gallivm->builder, hdl_ptr_type, num_inner_loop, "coro_hdls");
3447 unsigned end_coroutine = INT_MAX;
3448 lp_build_loop_begin(&loop_state[1], gallivm,
3449 lp_build_const_int32(gallivm, 0)); /* coroutine reentry loop */
3450 lp_build_loop_begin(&loop_state[0], gallivm,
3451 lp_build_const_int32(gallivm, 0)); /* inner loop */
3452 {
3453 LLVMValueRef args[7];
3454 args[0] = context_ptr;
3455 args[1] = input_array;
3456 args[2] = output_array;
3457 args[3] = prim_id;
3458 args[4] = patch_vertices_in;
3459 args[5] = view_index;
3460 args[6] = loop_state[0].counter;
3461 LLVMValueRef coro_entry = LLVMBuildGEP(builder, coro_hdls, &loop_state[0].counter, 1, "");
3462 LLVMValueRef coro_hdl = LLVMBuildLoad(builder, coro_entry, "coro_hdl");
3463
3464 struct lp_build_if_state ifstate;
3465 LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntEQ, loop_state[1].counter,
3466 lp_build_const_int32(gallivm, 0), "");
3467 /* first time here - call the coroutine function entry point */
3468 lp_build_if(&ifstate, gallivm, cmp);
3469 LLVMValueRef coro_ret = LLVMBuildCall(builder, variant_coro, args, 7, "");
3470 LLVMBuildStore(builder, coro_ret, coro_entry);
3471 lp_build_else(&ifstate);
3472 /* subsequent calls for this invocation - check if done. */
3473 LLVMValueRef coro_done = lp_build_coro_done(gallivm, coro_hdl);
3474 struct lp_build_if_state ifstate2;
3475 lp_build_if(&ifstate2, gallivm, coro_done);
3476 /* if done destroy and force loop exit */
3477 lp_build_coro_destroy(gallivm, coro_hdl);
3478 lp_build_loop_force_set_counter(&loop_state[1], lp_build_const_int32(gallivm, end_coroutine - 1));
3479 lp_build_else(&ifstate2);
3480 /* otherwise resume the coroutine */
3481 lp_build_coro_resume(gallivm, coro_hdl);
3482 lp_build_endif(&ifstate2);
3483 lp_build_endif(&ifstate);
3484 lp_build_loop_force_reload_counter(&loop_state[1]);
3485 }
3486 lp_build_loop_end_cond(&loop_state[0],
3487 num_inner_loop,
3488 NULL, LLVMIntUGE);
3489 lp_build_loop_end_cond(&loop_state[1],
3490 lp_build_const_int32(gallivm, end_coroutine),
3491 NULL, LLVMIntEQ);
3492 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
3493
3494 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_coro, "entry");
3495 LLVMPositionBuilderAtEnd(builder, block);
3496
3497 context_ptr = LLVMGetParam(variant_coro, 0);
3498 input_array = LLVMGetParam(variant_coro, 1);
3499 output_array = LLVMGetParam(variant_coro, 2);
3500 prim_id = LLVMGetParam(variant_coro, 3);
3501 patch_vertices_in = LLVMGetParam(variant_coro, 4);
3502 view_index = LLVMGetParam(variant_coro, 5);
3503
3504 consts_ptr = draw_tcs_jit_context_constants(variant->gallivm, context_ptr);
3505 num_consts_ptr =
3506 draw_tcs_jit_context_num_constants(variant->gallivm, context_ptr);
3507
3508 ssbos_ptr = draw_tcs_jit_context_ssbos(variant->gallivm, context_ptr);
3509 num_ssbos_ptr =
3510 draw_tcs_jit_context_num_ssbos(variant->gallivm, context_ptr);
3511 sampler = draw_llvm_sampler_soa_create(variant->key.samplers, variant->key.nr_samplers);
3512 image = draw_llvm_image_soa_create(draw_tcs_llvm_variant_key_images(&variant->key),
3513 variant->key.nr_images);
3514
3515 LLVMValueRef counter = LLVMGetParam(variant_coro, 6);
3516 LLVMValueRef invocvec = LLVMGetUndef(LLVMVectorType(int32_type, vector_length));
3517 for (i = 0; i < vector_length; i++) {
3518 LLVMValueRef idx = LLVMBuildAdd(builder, LLVMBuildMul(builder, counter, step, ""), lp_build_const_int32(gallivm, i), "");
3519 invocvec = LLVMBuildInsertElement(builder, invocvec, idx, idx, "");
3520 }
3521
3522 system_values.invocation_id = invocvec;
3523 system_values.prim_id = lp_build_broadcast_scalar(&bldvec, prim_id);
3524 system_values.view_index = view_index;
3525 system_values.vertices_in = lp_build_broadcast_scalar(&bldvec, patch_vertices_in);
3526 tcs_iface.input = input_array;
3527 tcs_iface.output = output_array;
3528 tcs_iface.base.emit_fetch_input = draw_tcs_llvm_emit_fetch_input;
3529 tcs_iface.base.emit_fetch_output = draw_tcs_llvm_emit_fetch_output;
3530 tcs_iface.base.emit_store_output = draw_tcs_llvm_emit_store_output;
3531
3532
3533 {
3534 LLVMValueRef coro_id = lp_build_coro_id(gallivm);
3535 LLVMValueRef coro_hdl = lp_build_coro_begin_alloc_mem(gallivm, coro_id);
3536
3537 mask_val = generate_tcs_mask_value(variant, tcs_type, count, LLVMBuildMul(builder, counter, step, ""));
3538 lp_build_mask_begin(&mask, gallivm, tcs_type, mask_val);
3539
3540 struct lp_build_coro_suspend_info coro_info;
3541
3542 LLVMBasicBlockRef sus_block = LLVMAppendBasicBlockInContext(gallivm->context, variant_coro, "suspend");
3543 LLVMBasicBlockRef clean_block = LLVMAppendBasicBlockInContext(gallivm->context, variant_coro, "cleanup");
3544
3545 coro_info.suspend = sus_block;
3546 coro_info.cleanup = clean_block;
3547
3548 struct lp_build_tgsi_params params;
3549 memset(¶ms, 0, sizeof(params));
3550
3551 params.type = tcs_type;
3552 params.mask = &mask;
3553 params.consts_ptr = consts_ptr;
3554 params.const_sizes_ptr = num_consts_ptr;
3555 params.system_values = &system_values;
3556 params.context_ptr = context_ptr;
3557 params.sampler = sampler;
3558 params.info = &llvm->draw->tcs.tess_ctrl_shader->info;
3559 params.ssbo_ptr = ssbos_ptr;
3560 params.ssbo_sizes_ptr = num_ssbos_ptr;
3561 params.image = image;
3562 params.coro = &coro_info;
3563 params.tcs_iface = &tcs_iface.base;
3564 params.aniso_filter_table = draw_tcs_jit_context_aniso_filter_table(gallivm, context_ptr);
3565
3566 lp_build_nir_soa(variant->gallivm,
3567 llvm->draw->tcs.tess_ctrl_shader->state.ir.nir,
3568 ¶ms, NULL);
3569
3570 lp_build_mask_end(&mask);
3571
3572 lp_build_coro_suspend_switch(gallivm, &coro_info, NULL, true);
3573 LLVMPositionBuilderAtEnd(builder, clean_block);
3574
3575 lp_build_coro_free_mem(gallivm, coro_id, coro_hdl);
3576
3577 LLVMBuildBr(builder, sus_block);
3578 LLVMPositionBuilderAtEnd(builder, sus_block);
3579
3580 lp_build_coro_end(gallivm, coro_hdl);
3581 LLVMBuildRet(builder, coro_hdl);
3582 }
3583
3584 sampler->destroy(sampler);
3585 image->destroy(image);
3586 gallivm_verify_function(gallivm, variant_func);
3587 gallivm_verify_function(gallivm, variant_coro);
3588 }
3589
3590 struct draw_tcs_llvm_variant *
draw_tcs_llvm_create_variant(struct draw_llvm * llvm,unsigned num_outputs,const struct draw_tcs_llvm_variant_key * key)3591 draw_tcs_llvm_create_variant(struct draw_llvm *llvm,
3592 unsigned num_outputs,
3593 const struct draw_tcs_llvm_variant_key *key)
3594 {
3595 struct draw_tcs_llvm_variant *variant;
3596 struct llvm_tess_ctrl_shader *shader = llvm_tess_ctrl_shader(llvm->draw->tcs.tess_ctrl_shader);
3597 char module_name[64];
3598 unsigned char ir_sha1_cache_key[20];
3599 struct lp_cached_code cached = { 0 };
3600 bool needs_caching = false;
3601
3602 variant = MALLOC(sizeof *variant +
3603 shader->variant_key_size - sizeof variant->key);
3604 if (!variant)
3605 return NULL;
3606
3607 variant->llvm = llvm;
3608 variant->shader = shader;
3609
3610 snprintf(module_name, sizeof(module_name), "draw_llvm_tcs_variant%u",
3611 variant->shader->variants_cached);
3612
3613 memcpy(&variant->key, key, shader->variant_key_size);
3614
3615 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
3616 draw_get_ir_cache_key(shader->base.state.ir.nir,
3617 key,
3618 shader->variant_key_size,
3619 num_outputs,
3620 ir_sha1_cache_key);
3621
3622 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
3623 &cached,
3624 ir_sha1_cache_key);
3625 if (!cached.data_size)
3626 needs_caching = true;
3627 }
3628
3629 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
3630
3631 create_tcs_jit_types(variant);
3632
3633 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
3634 nir_print_shader(llvm->draw->tcs.tess_ctrl_shader->state.ir.nir, stderr);
3635 draw_tcs_llvm_dump_variant_key(&variant->key);
3636 }
3637
3638 lp_build_coro_declare_malloc_hooks(variant->gallivm);
3639 draw_tcs_llvm_generate(llvm, variant);
3640
3641 gallivm_compile_module(variant->gallivm);
3642
3643 lp_build_coro_add_malloc_hooks(variant->gallivm);
3644 variant->jit_func = (draw_tcs_jit_func)
3645 gallivm_jit_function(variant->gallivm, variant->function);
3646
3647 if (needs_caching)
3648 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
3649 &cached,
3650 ir_sha1_cache_key);
3651 gallivm_free_ir(variant->gallivm);
3652
3653 variant->list_item_global.base = variant;
3654 variant->list_item_local.base = variant;
3655 /*variant->no = */shader->variants_created++;
3656 variant->list_item_global.base = variant;
3657
3658 return variant;
3659 }
3660
3661 void
draw_tcs_llvm_destroy_variant(struct draw_tcs_llvm_variant * variant)3662 draw_tcs_llvm_destroy_variant(struct draw_tcs_llvm_variant *variant)
3663 {
3664 struct draw_llvm *llvm = variant->llvm;
3665
3666 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
3667 debug_printf("Deleting TCS variant: %u tcs variants,\t%u total variants\n",
3668 variant->shader->variants_cached, llvm->nr_tcs_variants);
3669 }
3670
3671 gallivm_destroy(variant->gallivm);
3672
3673 remove_from_list(&variant->list_item_local);
3674 variant->shader->variants_cached--;
3675 remove_from_list(&variant->list_item_global);
3676 llvm->nr_tcs_variants--;
3677 FREE(variant);
3678 }
3679
3680 struct draw_tcs_llvm_variant_key *
draw_tcs_llvm_make_variant_key(struct draw_llvm * llvm,char * store)3681 draw_tcs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
3682 {
3683 unsigned i;
3684 struct draw_tcs_llvm_variant_key *key;
3685 struct draw_sampler_static_state *draw_sampler;
3686 struct draw_image_static_state *draw_image;
3687
3688 key = (struct draw_tcs_llvm_variant_key *)store;
3689
3690 memset(key, 0, offsetof(struct draw_tcs_llvm_variant_key, samplers[0]));
3691
3692 /* All variants of this shader will have the same value for
3693 * nr_samplers. Not yet trying to compact away holes in the
3694 * sampler array.
3695 */
3696 key->nr_samplers = llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
3697 if (llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
3698 key->nr_sampler_views =
3699 llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3700 }
3701 else {
3702 key->nr_sampler_views = key->nr_samplers;
3703 }
3704
3705 key->nr_images = llvm->draw->tcs.tess_ctrl_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
3706
3707 draw_sampler = key->samplers;
3708
3709 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
3710
3711 for (i = 0 ; i < key->nr_samplers; i++) {
3712 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
3713 llvm->draw->samplers[PIPE_SHADER_TESS_CTRL][i]);
3714 }
3715 for (i = 0 ; i < key->nr_sampler_views; i++) {
3716 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
3717 llvm->draw->sampler_views[PIPE_SHADER_TESS_CTRL][i]);
3718 }
3719
3720 draw_image = draw_tcs_llvm_variant_key_images(key);
3721 memset(draw_image, 0,
3722 key->nr_images * sizeof *draw_image);
3723 for (i = 0; i < key->nr_images; i++) {
3724 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
3725 llvm->draw->images[PIPE_SHADER_TESS_CTRL][i]);
3726 }
3727 return key;
3728 }
3729
3730 void
draw_tcs_llvm_dump_variant_key(struct draw_tcs_llvm_variant_key * key)3731 draw_tcs_llvm_dump_variant_key(struct draw_tcs_llvm_variant_key *key)
3732 {
3733 unsigned i;
3734 struct draw_sampler_static_state *sampler = key->samplers;
3735 struct draw_image_static_state *image = draw_tcs_llvm_variant_key_images(key);
3736 for (i = 0 ; i < key->nr_sampler_views; i++) {
3737 debug_printf("sampler[%i].src_format = %s\n", i,
3738 util_format_name(sampler[i].texture_state.format));
3739 }
3740
3741 for (i = 0 ; i < key->nr_images; i++)
3742 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
3743
3744 }
3745
3746 static void
create_tes_jit_types(struct draw_tes_llvm_variant * var)3747 create_tes_jit_types(struct draw_tes_llvm_variant *var)
3748 {
3749 struct gallivm_state *gallivm = var->gallivm;
3750 LLVMTypeRef texture_type, sampler_type, image_type, context_type;
3751
3752 texture_type = create_jit_texture_type(gallivm, "texture");
3753 sampler_type = create_jit_sampler_type(gallivm, "sampler");
3754 image_type = create_jit_image_type(gallivm, "image");
3755
3756 context_type = create_tes_jit_context_type(gallivm,
3757 0,
3758 texture_type, sampler_type,
3759 image_type,
3760 "draw_tes_jit_context");
3761 var->context_ptr_type = LLVMPointerType(context_type, 0);
3762
3763 var->input_array_type = create_tes_jit_input_type(gallivm);
3764 }
3765
3766 static LLVMTypeRef
get_tes_context_ptr_type(struct draw_tes_llvm_variant * variant)3767 get_tes_context_ptr_type(struct draw_tes_llvm_variant *variant)
3768 {
3769 if (!variant->context_ptr_type)
3770 create_tes_jit_types(variant);
3771 return variant->context_ptr_type;
3772 }
3773
3774 static LLVMValueRef
generate_tes_mask_value(struct draw_tes_llvm_variant * variant,struct lp_type tes_type,LLVMValueRef limit,LLVMValueRef loop_counter)3775 generate_tes_mask_value(struct draw_tes_llvm_variant *variant,
3776 struct lp_type tes_type, LLVMValueRef limit, LLVMValueRef loop_counter)
3777 {
3778 struct gallivm_state *gallivm = variant->gallivm;
3779 LLVMBuilderRef builder = gallivm->builder;
3780 struct lp_type mask_type = lp_int_type(tes_type);
3781 LLVMValueRef num_prims;
3782 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
3783 unsigned i;
3784
3785 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type), limit);
3786 for (i = 0; i < tes_type.length; i++) {
3787 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3788 mask_val = LLVMBuildInsertElement(builder, mask_val, LLVMBuildAdd(builder, loop_counter, idx, ""), idx, "");
3789 }
3790 mask_val = lp_build_compare(gallivm, mask_type,
3791 PIPE_FUNC_GREATER, num_prims, mask_val);
3792
3793 return mask_val;
3794 }
3795
3796 static LLVMValueRef
draw_tes_llvm_fetch_vertex_input(const struct lp_build_tes_iface * tes_iface,struct lp_build_context * bld,boolean is_vindex_indirect,LLVMValueRef vertex_index,boolean is_aindex_indirect,LLVMValueRef attrib_index,boolean is_sindex_indirect,LLVMValueRef swizzle_index)3797 draw_tes_llvm_fetch_vertex_input(const struct lp_build_tes_iface *tes_iface,
3798 struct lp_build_context *bld,
3799 boolean is_vindex_indirect,
3800 LLVMValueRef vertex_index,
3801 boolean is_aindex_indirect,
3802 LLVMValueRef attrib_index,
3803 boolean is_sindex_indirect,
3804 LLVMValueRef swizzle_index)
3805 {
3806 const struct draw_tes_llvm_iface *tes = draw_tes_llvm_iface(tes_iface);
3807 struct gallivm_state *gallivm = bld->gallivm;
3808 LLVMBuilderRef builder = gallivm->builder;
3809 LLVMValueRef indices[3];
3810 LLVMValueRef res;
3811 struct lp_type type = bld->type;
3812
3813 if (is_vindex_indirect || is_aindex_indirect || is_sindex_indirect) {
3814 int i;
3815
3816 res = bld->zero;
3817
3818 for (i = 0; i < type.length; ++i) {
3819 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3820 LLVMValueRef vert_chan_index = vertex_index;
3821 LLVMValueRef attr_chan_index = attrib_index;
3822 LLVMValueRef swiz_chan_index = swizzle_index;
3823 LLVMValueRef channel_vec;
3824
3825 if (is_vindex_indirect) {
3826 vert_chan_index = LLVMBuildExtractElement(builder,
3827 vertex_index, idx, "");
3828 }
3829 if (is_aindex_indirect) {
3830 attr_chan_index = LLVMBuildExtractElement(builder,
3831 attrib_index, idx, "");
3832 }
3833 if (is_sindex_indirect) {
3834 swiz_chan_index = LLVMBuildExtractElement(builder,
3835 swizzle_index, idx, "");
3836 }
3837
3838 indices[0] = vert_chan_index;
3839 indices[1] = attr_chan_index;
3840 indices[2] = swiz_chan_index;
3841
3842 channel_vec = LLVMBuildGEP(builder, tes->input, indices, 3, "");
3843 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
3844
3845 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3846 }
3847 } else {
3848 indices[0] = vertex_index;
3849 indices[1] = attrib_index;
3850 indices[2] = swizzle_index;
3851
3852 res = LLVMBuildGEP(builder, tes->input, indices, 3, "");
3853 res = LLVMBuildLoad(builder, res, "");
3854 res = lp_build_broadcast_scalar(bld, res);
3855 }
3856 return res;
3857 }
3858
3859 static LLVMValueRef
draw_tes_llvm_fetch_patch_input(const struct lp_build_tes_iface * tes_iface,struct lp_build_context * bld,boolean is_aindex_indirect,LLVMValueRef attrib_index,LLVMValueRef swizzle_index)3860 draw_tes_llvm_fetch_patch_input(const struct lp_build_tes_iface *tes_iface,
3861 struct lp_build_context *bld,
3862 boolean is_aindex_indirect,
3863 LLVMValueRef attrib_index,
3864 LLVMValueRef swizzle_index)
3865 {
3866 const struct draw_tes_llvm_iface *tes = draw_tes_llvm_iface(tes_iface);
3867 struct gallivm_state *gallivm = bld->gallivm;
3868 LLVMBuilderRef builder = gallivm->builder;
3869 LLVMValueRef indices[3];
3870 LLVMValueRef res;
3871 struct lp_type type = bld->type;
3872
3873 if (is_aindex_indirect) {
3874 int i;
3875
3876 res = bld->zero;
3877
3878 for (i = 0; i < type.length; ++i) {
3879 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
3880 LLVMValueRef attr_chan_index = attrib_index;
3881 LLVMValueRef channel_vec;
3882
3883 if (is_aindex_indirect) {
3884 attr_chan_index = LLVMBuildExtractElement(builder,
3885 attrib_index, idx, "");
3886 }
3887
3888 indices[0] = lp_build_const_int32(gallivm, 0);
3889 indices[1] = attr_chan_index;
3890 indices[2] = swizzle_index;
3891
3892 channel_vec = LLVMBuildGEP(builder, tes->input, indices, 3, "");
3893 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
3894
3895 res = LLVMBuildInsertElement(builder, res, channel_vec, idx, "");
3896 }
3897 } else {
3898 indices[0] = lp_build_const_int32(gallivm, 0);
3899 indices[1] = attrib_index;
3900 indices[2] = swizzle_index;
3901
3902 res = LLVMBuildGEP(builder, tes->input, indices, 3, "");
3903 res = LLVMBuildLoad(builder, res, "");
3904 res = lp_build_broadcast_scalar(bld, res);
3905 }
3906 return res;
3907 }
3908
3909 static void
draw_tes_llvm_generate(struct draw_llvm * llvm,struct draw_tes_llvm_variant * variant)3910 draw_tes_llvm_generate(struct draw_llvm *llvm,
3911 struct draw_tes_llvm_variant *variant)
3912 {
3913 struct gallivm_state *gallivm = variant->gallivm;
3914 LLVMContextRef context = gallivm->context;
3915 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
3916 LLVMTypeRef flt_type = LLVMFloatTypeInContext(context);
3917 LLVMTypeRef arg_types[11];
3918 LLVMTypeRef func_type;
3919 LLVMValueRef variant_func;
3920 LLVMValueRef context_ptr;
3921 LLVMValueRef tess_coord[2], io_ptr, input_array, num_tess_coord;
3922 LLVMValueRef view_index;
3923 LLVMValueRef tess_inner, tess_outer, prim_id, patch_vertices_in;
3924 LLVMBasicBlockRef block;
3925 LLVMBuilderRef builder;
3926 LLVMValueRef mask_val;
3927 struct lp_build_context bld, bldvec;
3928 struct lp_build_sampler_soa *sampler = 0;
3929 struct lp_build_image_soa *image = NULL;
3930 struct lp_bld_tgsi_system_values system_values;
3931 char func_name[64];
3932 unsigned i;
3933 struct draw_tes_llvm_iface tes_iface;
3934 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
3935 struct lp_build_mask_context mask;
3936 LLVMValueRef consts_ptr, num_consts_ptr;
3937 LLVMValueRef ssbos_ptr, num_ssbos_ptr;
3938 LLVMValueRef step;
3939 struct lp_type tes_type;
3940 unsigned vector_length = variant->shader->base.vector_length;
3941
3942 memset(&system_values, 0, sizeof(system_values));
3943 memset(&outputs, 0, sizeof(outputs));
3944
3945 snprintf(func_name, sizeof(func_name), "draw_llvm_tes_variant");
3946
3947 arg_types[0] = get_tes_context_ptr_type(variant); /* context */
3948 arg_types[1] = variant->input_array_type; /* input */
3949 arg_types[2] = variant->vertex_header_ptr_type;
3950 arg_types[3] = int32_type;
3951 arg_types[4] = int32_type;
3952 arg_types[5] = LLVMPointerType(flt_type, 0);
3953 arg_types[6] = LLVMPointerType(flt_type, 0);
3954 arg_types[7] = LLVMPointerType(LLVMArrayType(flt_type, 4), 0);
3955 arg_types[8] = LLVMPointerType(LLVMArrayType(flt_type, 2), 0);
3956 arg_types[9] = int32_type;
3957 arg_types[10] = int32_type;
3958
3959 func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
3960 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
3961
3962 variant->function = variant_func;
3963 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
3964
3965 for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
3966 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
3967 lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
3968
3969 if (gallivm->cache && gallivm->cache->data_size)
3970 return;
3971 context_ptr = LLVMGetParam(variant_func, 0);
3972 input_array = LLVMGetParam(variant_func, 1);
3973 io_ptr = LLVMGetParam(variant_func, 2);
3974 prim_id = LLVMGetParam(variant_func, 3);
3975 num_tess_coord = LLVMGetParam(variant_func, 4);
3976 tess_coord[0] = LLVMGetParam(variant_func, 5);
3977 tess_coord[1] = LLVMGetParam(variant_func, 6);
3978 tess_outer = LLVMGetParam(variant_func, 7);
3979 tess_inner = LLVMGetParam(variant_func, 8);
3980 patch_vertices_in = LLVMGetParam(variant_func, 9);
3981 view_index = LLVMGetParam(variant_func, 10);
3982
3983 lp_build_name(context_ptr, "context");
3984 lp_build_name(input_array, "input");
3985 lp_build_name(io_ptr, "io");
3986 lp_build_name(prim_id, "prim_id");
3987 lp_build_name(num_tess_coord, "num_tess_coord");
3988 lp_build_name(tess_coord[0], "tess_coord[0]");
3989 lp_build_name(tess_coord[1], "tess_coord[1]");
3990 lp_build_name(tess_outer, "tess_outer");
3991 lp_build_name(tess_inner, "tess_inner");
3992 lp_build_name(patch_vertices_in, "patch_vertices_in");
3993 lp_build_name(view_index, "view_index");
3994
3995 tes_iface.base.fetch_vertex_input = draw_tes_llvm_fetch_vertex_input;
3996 tes_iface.base.fetch_patch_input = draw_tes_llvm_fetch_patch_input;
3997 tes_iface.input = input_array;
3998 tes_iface.variant = variant;
3999
4000 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
4001 builder = gallivm->builder;
4002 LLVMPositionBuilderAtEnd(builder, block);
4003
4004 lp_build_context_init(&bld, gallivm, lp_type_int(32));
4005
4006 memset(&tes_type, 0, sizeof tes_type);
4007 tes_type.floating = TRUE; /* floating point values */
4008 tes_type.sign = TRUE; /* values are signed */
4009 tes_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
4010 tes_type.width = 32; /* 32-bit float */
4011 tes_type.length = vector_length;
4012
4013 lp_build_context_init(&bldvec, variant->gallivm, lp_int_type(tes_type));
4014 consts_ptr = draw_tes_jit_context_constants(variant->gallivm, context_ptr);
4015 num_consts_ptr =
4016 draw_tes_jit_context_num_constants(variant->gallivm, context_ptr);
4017
4018 ssbos_ptr = draw_tes_jit_context_ssbos(variant->gallivm, context_ptr);
4019 num_ssbos_ptr =
4020 draw_tes_jit_context_num_ssbos(variant->gallivm, context_ptr);
4021 sampler = draw_llvm_sampler_soa_create(variant->key.samplers, variant->key.nr_samplers);
4022 image = draw_llvm_image_soa_create(draw_tes_llvm_variant_key_images(&variant->key),
4023 variant->key.nr_images);
4024 step = lp_build_const_int32(gallivm, vector_length);
4025
4026 system_values.tess_outer = LLVMBuildLoad(builder, tess_outer, "");
4027 system_values.tess_inner = LLVMBuildLoad(builder, tess_inner, "");
4028
4029 system_values.prim_id = lp_build_broadcast_scalar(&bldvec, prim_id);
4030
4031 system_values.view_index = view_index;
4032
4033 system_values.vertices_in = lp_build_broadcast_scalar(&bldvec, patch_vertices_in);
4034
4035 if (variant->key.primid_needed) {
4036 int slot = variant->key.primid_output;
4037 for (unsigned i = 0; i < 4; i++) {
4038 outputs[slot][i] = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, tes_type), "primid");
4039 LLVMBuildStore(builder, system_values.prim_id, outputs[slot][i]);
4040 }
4041 }
4042 struct lp_build_loop_state lp_loop;
4043 lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
4044 {
4045 LLVMValueRef io;
4046
4047 io = LLVMBuildGEP(builder, io_ptr, &lp_loop.counter, 1, "");
4048 mask_val = generate_tes_mask_value(variant, tes_type, num_tess_coord, lp_loop.counter);
4049 lp_build_mask_begin(&mask, gallivm, tes_type, mask_val);
4050
4051 system_values.tess_coord = LLVMGetUndef(LLVMArrayType(LLVMVectorType(flt_type, vector_length), 3));
4052 for (i = 0; i < 3; i++) {
4053 LLVMValueRef tess_coord_chan = LLVMGetUndef(LLVMVectorType(flt_type, vector_length));
4054 for (unsigned j = 0; j < vector_length; j++) {
4055 LLVMValueRef idx = LLVMBuildAdd(builder, lp_loop.counter, lp_build_const_int32(gallivm, j), "");
4056 LLVMValueRef tc_val;
4057 if (i == 2) {
4058 if (variant->shader->base.prim_mode == PIPE_PRIM_TRIANGLES) {
4059 tc_val = lp_build_const_float(gallivm, 1.0);
4060 tc_val = LLVMBuildFSub(builder, tc_val, lp_build_pointer_get(builder, tess_coord[0], idx), "");
4061 tc_val = LLVMBuildFSub(builder, tc_val, lp_build_pointer_get(builder, tess_coord[1], idx), "");
4062 } else
4063 tc_val = lp_build_const_float(gallivm, 0.0);
4064 } else
4065 tc_val = lp_build_pointer_get(builder, tess_coord[i], idx);
4066
4067 tess_coord_chan = LLVMBuildInsertElement(builder, tess_coord_chan, tc_val, lp_build_const_int32(gallivm, j), "");
4068 }
4069 system_values.tess_coord = LLVMBuildInsertValue(builder, system_values.tess_coord, tess_coord_chan, i, "");
4070 }
4071
4072 struct lp_build_tgsi_params params;
4073 memset(¶ms, 0, sizeof(params));
4074
4075 params.type = tes_type;
4076 params.mask = &mask;
4077 params.consts_ptr = consts_ptr;
4078 params.const_sizes_ptr = num_consts_ptr;
4079 params.system_values = &system_values;
4080 params.context_ptr = context_ptr;
4081 params.sampler = sampler;
4082 params.info = &llvm->draw->tes.tess_eval_shader->info;
4083 params.ssbo_ptr = ssbos_ptr;
4084 params.ssbo_sizes_ptr = num_ssbos_ptr;
4085 params.image = image;
4086 params.tes_iface = &tes_iface.base;
4087 params.aniso_filter_table = draw_tes_jit_context_aniso_filter_table(variant->gallivm, context_ptr);
4088
4089 lp_build_nir_soa(variant->gallivm,
4090 llvm->draw->tes.tess_eval_shader->state.ir.nir,
4091 ¶ms,
4092 outputs);
4093
4094 lp_build_mask_end(&mask);
4095
4096 if (variant->key.clamp_vertex_color) {
4097 const struct tgsi_shader_info *info = &llvm->draw->tes.tess_eval_shader->info;
4098 do_clamp_vertex_color(variant->gallivm,
4099 tes_type, info,
4100 outputs);
4101 }
4102 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
4103 lp_int_type(tes_type), 0);
4104
4105 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
4106 draw_total_tes_outputs(llvm->draw), tes_type, FALSE);
4107 }
4108 lp_build_loop_end_cond(&lp_loop, num_tess_coord, step, LLVMIntUGE);
4109 sampler->destroy(sampler);
4110 image->destroy(image);
4111
4112 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
4113 gallivm_verify_function(gallivm, variant_func);
4114 }
4115
4116 struct draw_tes_llvm_variant *
draw_tes_llvm_create_variant(struct draw_llvm * llvm,unsigned num_outputs,const struct draw_tes_llvm_variant_key * key)4117 draw_tes_llvm_create_variant(struct draw_llvm *llvm,
4118 unsigned num_outputs,
4119 const struct draw_tes_llvm_variant_key *key)
4120 {
4121 struct draw_tes_llvm_variant *variant;
4122 struct llvm_tess_eval_shader *shader = llvm_tess_eval_shader(llvm->draw->tes.tess_eval_shader);
4123 LLVMTypeRef vertex_header;
4124 char module_name[64];
4125 unsigned char ir_sha1_cache_key[20];
4126 struct lp_cached_code cached = { 0 };
4127 bool needs_caching = false;
4128
4129 variant = MALLOC(sizeof *variant +
4130 shader->variant_key_size - sizeof variant->key);
4131 if (!variant)
4132 return NULL;
4133
4134 variant->llvm = llvm;
4135 variant->shader = shader;
4136
4137 snprintf(module_name, sizeof(module_name), "draw_llvm_tes_variant%u",
4138 variant->shader->variants_cached);
4139
4140 memcpy(&variant->key, key, shader->variant_key_size);
4141 if (shader->base.state.ir.nir && llvm->draw->disk_cache_cookie) {
4142 draw_get_ir_cache_key(shader->base.state.ir.nir,
4143 key,
4144 shader->variant_key_size,
4145 num_outputs,
4146 ir_sha1_cache_key);
4147
4148 llvm->draw->disk_cache_find_shader(llvm->draw->disk_cache_cookie,
4149 &cached,
4150 ir_sha1_cache_key);
4151 if (!cached.data_size)
4152 needs_caching = true;
4153 }
4154 variant->gallivm = gallivm_create(module_name, llvm->context, &cached);
4155
4156 create_tes_jit_types(variant);
4157
4158 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
4159
4160 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
4161
4162 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
4163 nir_print_shader(llvm->draw->tes.tess_eval_shader->state.ir.nir, stderr);
4164 draw_tes_llvm_dump_variant_key(&variant->key);
4165 }
4166
4167 draw_tes_llvm_generate(llvm, variant);
4168
4169 gallivm_compile_module(variant->gallivm);
4170
4171 variant->jit_func = (draw_tes_jit_func)
4172 gallivm_jit_function(variant->gallivm, variant->function);
4173
4174 if (needs_caching)
4175 llvm->draw->disk_cache_insert_shader(llvm->draw->disk_cache_cookie,
4176 &cached,
4177 ir_sha1_cache_key);
4178 gallivm_free_ir(variant->gallivm);
4179
4180 variant->list_item_global.base = variant;
4181 variant->list_item_local.base = variant;
4182 /*variant->no = */shader->variants_created++;
4183 variant->list_item_global.base = variant;
4184
4185 return variant;
4186 }
4187
4188 void
draw_tes_llvm_destroy_variant(struct draw_tes_llvm_variant * variant)4189 draw_tes_llvm_destroy_variant(struct draw_tes_llvm_variant *variant)
4190 {
4191 struct draw_llvm *llvm = variant->llvm;
4192
4193 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
4194 debug_printf("Deleting TES variant: %u tes variants,\t%u total variants\n",
4195 variant->shader->variants_cached, llvm->nr_tes_variants);
4196 }
4197
4198 gallivm_destroy(variant->gallivm);
4199
4200 remove_from_list(&variant->list_item_local);
4201 variant->shader->variants_cached--;
4202 remove_from_list(&variant->list_item_global);
4203 llvm->nr_tes_variants--;
4204 FREE(variant);
4205 }
4206
4207 struct draw_tes_llvm_variant_key *
draw_tes_llvm_make_variant_key(struct draw_llvm * llvm,char * store)4208 draw_tes_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
4209 {
4210 unsigned i;
4211 struct draw_tes_llvm_variant_key *key;
4212 struct draw_sampler_static_state *draw_sampler;
4213 struct draw_image_static_state *draw_image;
4214
4215 key = (struct draw_tes_llvm_variant_key *)store;
4216
4217 memset(key, 0, offsetof(struct draw_tes_llvm_variant_key, samplers[0]));
4218
4219 int primid_output = draw_find_shader_output(llvm->draw, TGSI_SEMANTIC_PRIMID, 0);
4220 if (primid_output >= 0) {
4221 key->primid_output = primid_output;
4222 key->primid_needed = true;
4223 }
4224
4225 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color &&
4226 llvm->draw->gs.geometry_shader == NULL;
4227
4228 /* All variants of this shader will have the same value for
4229 * nr_samplers. Not yet trying to compact away holes in the
4230 * sampler array.
4231 */
4232 key->nr_samplers = llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
4233 if (llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
4234 key->nr_sampler_views =
4235 llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
4236 }
4237 else {
4238 key->nr_sampler_views = key->nr_samplers;
4239 }
4240
4241 key->nr_images = llvm->draw->tes.tess_eval_shader->info.file_max[TGSI_FILE_IMAGE] + 1;
4242
4243 draw_sampler = key->samplers;
4244
4245 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
4246
4247 for (i = 0 ; i < key->nr_samplers; i++) {
4248 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
4249 llvm->draw->samplers[PIPE_SHADER_TESS_EVAL][i]);
4250 }
4251 for (i = 0 ; i < key->nr_sampler_views; i++) {
4252 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
4253 llvm->draw->sampler_views[PIPE_SHADER_TESS_EVAL][i]);
4254 }
4255
4256 draw_image = draw_tes_llvm_variant_key_images(key);
4257 memset(draw_image, 0,
4258 key->nr_images * sizeof *draw_image);
4259 for (i = 0; i < key->nr_images; i++) {
4260 lp_sampler_static_texture_state_image(&draw_image[i].image_state,
4261 llvm->draw->images[PIPE_SHADER_TESS_EVAL][i]);
4262 }
4263 return key;
4264 }
4265
4266 void
draw_tes_llvm_dump_variant_key(struct draw_tes_llvm_variant_key * key)4267 draw_tes_llvm_dump_variant_key(struct draw_tes_llvm_variant_key *key)
4268 {
4269 unsigned i;
4270 struct draw_sampler_static_state *sampler = key->samplers;
4271 struct draw_image_static_state *image = draw_tes_llvm_variant_key_images(key);
4272
4273 if (key->primid_needed)
4274 debug_printf("prim id output %d\n", key->primid_output);
4275 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
4276 for (i = 0 ; i < key->nr_sampler_views; i++) {
4277 debug_printf("sampler[%i].src_format = %s\n", i,
4278 util_format_name(sampler[i].texture_state.format));
4279 }
4280
4281 for (i = 0 ; i < key->nr_images; i++)
4282 debug_printf("images[%i].format = %s\n", i, util_format_name(image[i].image_state.format));
4283
4284 }
4285