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