1 /**************************************************************************
3 * Copyright 2010 VMware, Inc.
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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
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.
26 **************************************************************************/
28 #include "draw_llvm.h"
30 #include "draw_context.h"
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"
55 #include "tgsi/tgsi_exec.h"
56 #include "tgsi/tgsi_dump.h"
58 #include "util/u_math.h"
59 #include "util/u_pointer.h"
60 #include "util/u_string.h"
61 #include "util/simple_list.h"
68 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*var
);
71 struct draw_gs_llvm_iface
{
72 struct lp_build_gs_iface base
;
74 struct draw_gs_llvm_variant
*variant
;
78 static inline const struct draw_gs_llvm_iface
*
79 draw_gs_llvm_iface(const struct lp_build_gs_iface
*iface
)
81 return (const struct draw_gs_llvm_iface
*)iface
;
84 struct draw_tcs_llvm_iface
{
85 struct lp_build_tcs_iface base
;
87 struct draw_tcs_llvm_variant
*variant
;
92 static inline const struct draw_tcs_llvm_iface
*
93 draw_tcs_llvm_iface(const struct lp_build_tcs_iface
*iface
)
95 return (const struct draw_tcs_llvm_iface
*)iface
;
98 struct draw_tes_llvm_iface
{
99 struct lp_build_tes_iface base
;
101 struct draw_tes_llvm_variant
*variant
;
105 static inline const struct draw_tes_llvm_iface
*
106 draw_tes_llvm_iface(const struct lp_build_tes_iface
*iface
)
108 return (const struct draw_tes_llvm_iface
*)iface
;
112 * Create LLVM type for draw_vertex_buffer.
115 create_jit_dvbuffer_type(struct gallivm_state
*gallivm
,
116 const char *struct_name
)
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
);
123 elem_types
[DRAW_JIT_DVBUFFER_MAP
] =
124 LLVMPointerType(LLVMIntTypeInContext(gallivm
->context
, 8), 0);
125 elem_types
[DRAW_JIT_DVBUFFER_SIZE
] = int32_type
;
127 dvbuffer_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
128 ARRAY_SIZE(elem_types
), 0);
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
);
138 return dvbuffer_type
;
142 * Create LLVM type for struct draw_jit_texture
145 create_jit_texture_type(struct gallivm_state
*gallivm
, const char *struct_name
)
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
);
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
);
166 texture_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
167 ARRAY_SIZE(elem_types
), 0);
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
);
204 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture
, target
, texture_type
);
211 * Create LLVM type for struct draw_jit_sampler
214 create_jit_sampler_type(struct gallivm_state
*gallivm
, const char *struct_name
)
216 LLVMTargetDataRef target
= gallivm
->target
;
217 LLVMTypeRef sampler_type
;
218 LLVMTypeRef elem_types
[DRAW_JIT_SAMPLER_NUM_FIELDS
];
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);
226 sampler_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
227 ARRAY_SIZE(elem_types
), 0);
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
);
243 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler
, target
, sampler_type
);
249 * Create LLVM type for struct draw_jit_texture
252 create_jit_image_type(struct gallivm_state
*gallivm
, const char *struct_name
)
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
);
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);
269 image_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
270 ARRAY_SIZE(elem_types
), 0);
272 (void) target
; /* silence unused var warning for non-debug build */
273 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, width
,
275 DRAW_JIT_IMAGE_WIDTH
);
276 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, height
,
278 DRAW_JIT_IMAGE_HEIGHT
);
279 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, depth
,
281 DRAW_JIT_IMAGE_DEPTH
);
282 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, base
,
284 DRAW_JIT_IMAGE_BASE
);
285 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, row_stride
,
287 DRAW_JIT_IMAGE_ROW_STRIDE
);
288 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, img_stride
,
290 DRAW_JIT_IMAGE_IMG_STRIDE
);
291 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, num_samples
,
293 DRAW_JIT_IMAGE_NUM_SAMPLES
);
294 LP_CHECK_MEMBER_OFFSET(struct draw_jit_image
, sample_stride
,
296 DRAW_JIT_IMAGE_SAMPLE_STRIDE
);
298 LP_CHECK_STRUCT_SIZE(struct draw_jit_image
, target
, image_type
);
304 * Create LLVM type for struct draw_jit_context
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
)
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
;
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);
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
);
367 * Create LLVM type for struct draw_gs_jit_context
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
)
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
;
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 */
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
,
399 elem_types
[9] = LLVMPointerType(LLVMVectorType(int_type
,
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
);
407 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
408 ARRAY_SIZE(elem_types
), 0);
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
);
448 create_gs_jit_input_type(struct gallivm_state
*gallivm
)
450 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
451 LLVMTypeRef input_array
;
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 */
462 * Create LLVM type for struct pipe_vertex_buffer
465 create_jit_vertex_buffer_type(struct gallivm_state
*gallivm
,
466 const char *struct_name
)
468 LLVMTargetDataRef target
= gallivm
->target
;
469 LLVMTypeRef elem_types
[4];
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);
477 vb_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
478 ARRAY_SIZE(elem_types
), 0);
480 (void) target
; /* silence unused var warning for non-debug build */
481 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, stride
,
483 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, is_user_buffer
,
485 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer_offset
,
487 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer
.resource
,
490 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer
, target
, vb_type
);
497 * Create LLVM type for struct vertex_header;
500 create_jit_vertex_header(struct gallivm_state
*gallivm
, int data_elems
)
502 LLVMTargetDataRef target
= gallivm
->target
;
503 LLVMTypeRef elem_types
[3];
504 LLVMTypeRef vertex_header
;
505 char struct_name
[24];
507 snprintf(struct_name
, 23, "vertex_header%d", data_elems
);
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
);
513 vertex_header
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
514 ARRAY_SIZE(elem_types
), 0);
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);
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
);
538 assert(LLVMABISizeOfType(target
, vertex_header
) ==
539 offsetof(struct vertex_header
, data
[data_elems
]));
541 return vertex_header
;
545 * Create LLVM type for struct draw_tcs_jit_context
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
)
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
;
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
);
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 */
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
);
579 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
580 ARRAY_SIZE(elem_types
), 0);
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
);
606 create_tcs_jit_input_type(struct gallivm_state
*gallivm
)
608 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
609 LLVMTypeRef input_array
;
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 */
619 create_tcs_jit_output_type(struct gallivm_state
*gallivm
)
621 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
622 LLVMTypeRef output_array
;
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 */
632 create_tes_jit_input_type(struct gallivm_state
*gallivm
)
634 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
635 LLVMTypeRef input_array
;
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 */
645 * Create LLVM type for struct draw_tes_jit_context
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
)
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
;
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
);
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 */
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
);
679 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
680 ARRAY_SIZE(elem_types
), 0);
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
);
706 * Create LLVM types for various structures.
709 create_jit_types(struct draw_llvm_variant
*variant
)
711 struct gallivm_state
*gallivm
= variant
->gallivm
;
712 LLVMTypeRef texture_type
, sampler_type
, context_type
, buffer_type
,
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");
719 context_type
= create_jit_context_type(gallivm
, texture_type
, sampler_type
,
722 variant
->context_ptr_type
= LLVMPointerType(context_type
, 0);
724 buffer_type
= create_jit_dvbuffer_type(gallivm
, "draw_vertex_buffer");
725 variant
->buffer_ptr_type
= LLVMPointerType(buffer_type
, 0);
727 vb_type
= create_jit_vertex_buffer_type(gallivm
, "pipe_vertex_buffer");
728 variant
->vb_ptr_type
= LLVMPointerType(vb_type
, 0);
733 get_context_ptr_type(struct draw_llvm_variant
*variant
)
735 if (!variant
->context_ptr_type
)
736 create_jit_types(variant
);
737 return variant
->context_ptr_type
;
742 get_buffer_ptr_type(struct draw_llvm_variant
*variant
)
744 if (!variant
->buffer_ptr_type
)
745 create_jit_types(variant
);
746 return variant
->buffer_ptr_type
;
751 get_vb_ptr_type(struct draw_llvm_variant
*variant
)
753 if (!variant
->vb_ptr_type
)
754 create_jit_types(variant
);
755 return variant
->vb_ptr_type
;
759 get_vertex_header_ptr_type(struct draw_llvm_variant
*variant
)
761 if (!variant
->vertex_header_ptr_type
)
762 create_jit_types(variant
);
763 return variant
->vertex_header_ptr_type
;
768 * Create per-context LLVM info.
771 draw_llvm_create(struct draw_context
*draw
, LLVMContextRef context
)
773 struct draw_llvm
*llvm
;
775 if (!lp_build_init())
778 llvm
= CALLOC_STRUCT( draw_llvm
);
784 llvm
->context
= context
;
785 if (!llvm
->context
) {
786 llvm
->context
= LLVMContextCreate();
787 llvm
->context_owned
= true;
792 llvm
->nr_variants
= 0;
793 make_empty_list(&llvm
->vs_variants_list
);
795 llvm
->nr_gs_variants
= 0;
796 make_empty_list(&llvm
->gs_variants_list
);
798 llvm
->nr_tcs_variants
= 0;
799 make_empty_list(&llvm
->tcs_variants_list
);
801 llvm
->nr_tes_variants
= 0;
802 make_empty_list(&llvm
->tes_variants_list
);
807 draw_llvm_destroy(llvm
);
813 * Free per-context LLVM info.
816 draw_llvm_destroy(struct draw_llvm
*llvm
)
818 if (llvm
->context_owned
)
819 LLVMContextDispose(llvm
->context
);
820 llvm
->context
= NULL
;
822 /* XXX free other draw_llvm data? */
828 * Create LLVM-generated code for a vertex shader.
830 struct draw_llvm_variant
*
831 draw_llvm_create_variant(struct draw_llvm
*llvm
,
833 const struct draw_llvm_variant_key
*key
)
835 struct draw_llvm_variant
*variant
;
836 struct llvm_vertex_shader
*shader
=
837 llvm_vertex_shader(llvm
->draw
->vs
.vertex_shader
);
838 LLVMTypeRef vertex_header
;
839 char module_name
[64];
841 variant
= MALLOC(sizeof *variant
+
842 shader
->variant_key_size
-
843 sizeof variant
->key
);
847 variant
->llvm
= llvm
;
848 variant
->shader
= shader
;
850 snprintf(module_name
, sizeof(module_name
), "draw_llvm_vs_variant%u",
851 variant
->shader
->variants_cached
);
853 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, NULL
);
855 create_jit_types(variant
);
857 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
859 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
860 if (llvm
->draw
->vs
.vertex_shader
->state
.type
== PIPE_SHADER_IR_TGSI
)
861 tgsi_dump(llvm
->draw
->vs
.vertex_shader
->state
.tokens
, 0);
863 nir_print_shader(llvm
->draw
->vs
.vertex_shader
->state
.ir
.nir
, stderr
);
864 draw_llvm_dump_variant_key(&variant
->key
);
867 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_inputs
);
869 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
871 draw_llvm_generate(llvm
, variant
);
873 gallivm_compile_module(variant
->gallivm
);
875 variant
->jit_func
= (draw_jit_vert_func
)
876 gallivm_jit_function(variant
->gallivm
, variant
->function
);
878 gallivm_free_ir(variant
->gallivm
);
880 variant
->list_item_global
.base
= variant
;
881 variant
->list_item_local
.base
= variant
;
882 /*variant->no = */shader
->variants_created
++;
883 variant
->list_item_global
.base
= variant
;
890 generate_vs(struct draw_llvm_variant
*variant
,
891 LLVMBuilderRef builder
,
892 struct lp_type vs_type
,
893 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
894 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
895 const struct lp_bld_tgsi_system_values
*system_values
,
896 LLVMValueRef context_ptr
,
897 const struct lp_build_sampler_soa
*draw_sampler
,
898 const struct lp_build_image_soa
*draw_image
,
899 boolean clamp_vertex_color
,
900 struct lp_build_mask_context
*bld_mask
)
902 struct draw_llvm
*llvm
= variant
->llvm
;
903 const struct tgsi_token
*tokens
= llvm
->draw
->vs
.vertex_shader
->state
.tokens
;
904 LLVMValueRef consts_ptr
=
905 draw_jit_context_vs_constants(variant
->gallivm
, context_ptr
);
906 LLVMValueRef num_consts_ptr
=
907 draw_jit_context_num_vs_constants(variant
->gallivm
, context_ptr
);
908 LLVMValueRef ssbos_ptr
=
909 draw_jit_context_vs_ssbos(variant
->gallivm
, context_ptr
);
910 LLVMValueRef num_ssbos_ptr
=
911 draw_jit_context_num_vs_ssbos(variant
->gallivm
, context_ptr
);
913 struct lp_build_tgsi_params params
;
914 memset(¶ms
, 0, sizeof(params
));
916 params
.type
= vs_type
;
917 params
.mask
= bld_mask
;
918 params
.consts_ptr
= consts_ptr
;
919 params
.const_sizes_ptr
= num_consts_ptr
;
920 params
.system_values
= system_values
;
921 params
.inputs
= inputs
;
922 params
.context_ptr
= context_ptr
;
923 params
.sampler
= draw_sampler
;
924 params
.info
= &llvm
->draw
->vs
.vertex_shader
->info
;
925 params
.ssbo_ptr
= ssbos_ptr
;
926 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
927 params
.image
= draw_image
;
929 if (llvm
->draw
->vs
.vertex_shader
->state
.ir
.nir
&&
930 llvm
->draw
->vs
.vertex_shader
->state
.type
== PIPE_SHADER_IR_NIR
)
931 lp_build_nir_soa(variant
->gallivm
,
932 llvm
->draw
->vs
.vertex_shader
->state
.ir
.nir
,
936 lp_build_tgsi_soa(variant
->gallivm
,
943 unsigned chan
, attrib
;
944 struct lp_build_context bld
;
945 struct tgsi_shader_info
* info
= &llvm
->draw
->vs
.vertex_shader
->info
;
946 lp_build_context_init(&bld
, variant
->gallivm
, vs_type
);
948 for (attrib
= 0; attrib
< info
->num_outputs
; ++attrib
) {
949 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
950 if (outputs
[attrib
][chan
]) {
951 switch (info
->output_semantic_name
[attrib
]) {
952 case TGSI_SEMANTIC_COLOR
:
953 case TGSI_SEMANTIC_BCOLOR
:
954 if (clamp_vertex_color
) {
955 out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
956 out
= lp_build_clamp(&bld
, out
, bld
.zero
, bld
.one
);
957 LLVMBuildStore(builder
, out
, outputs
[attrib
][chan
]);
969 fetch_instanced(struct gallivm_state
*gallivm
,
970 const struct util_format_description
*format_desc
,
971 struct lp_type vs_type
,
972 LLVMValueRef vb_stride
,
973 LLVMValueRef map_ptr
,
974 LLVMValueRef buffer_size_adj
,
975 LLVMValueRef
*inputs
,
978 LLVMTypeRef i32_t
= LLVMInt32TypeInContext(gallivm
->context
);
979 LLVMTypeRef aosf_t
, aosi_t
;
980 LLVMValueRef zero
= LLVMConstNull(i32_t
);
981 LLVMBuilderRef builder
= gallivm
->builder
;
982 LLVMValueRef stride
, buffer_overflowed
, aos
, index_valid
;
985 aosf_t
= lp_build_vec_type(gallivm
, lp_float32_vec4_type());
986 aosi_t
= lp_build_vec_type(gallivm
, lp_int32_vec4_type());
988 /* This mul can overflow. Wraparound is ok. */
989 stride
= LLVMBuildMul(builder
, vb_stride
, index
, "");
991 buffer_overflowed
= LLVMBuildICmp(builder
, LLVMIntUGE
,
992 stride
, buffer_size_adj
,
993 "buffer_overflowed");
996 lp_build_print_value(gallivm
, " instance index = ", index
);
997 lp_build_print_value(gallivm
, " buffer overflowed = ", buffer_overflowed
);
1000 index_valid
= LLVMBuildNot(builder
, buffer_overflowed
, "");
1001 index_valid
= LLVMBuildSExt(builder
, index_valid
, i32_t
, "");
1002 stride
= LLVMBuildAnd(builder
, stride
, index_valid
, "");
1004 aos
= lp_build_fetch_rgba_aos(gallivm
,
1006 lp_float32_vec4_type(),
1012 index_valid
= lp_build_broadcast(gallivm
, aosi_t
, index_valid
);
1013 aos
= LLVMBuildBitCast(builder
, aos
, aosi_t
, "");
1014 aos
= LLVMBuildAnd(builder
, aos
, index_valid
, "");
1015 aos
= LLVMBuildBitCast(builder
, aos
, aosf_t
, "");
1017 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1018 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
1019 inputs
[i
] = lp_build_extract_broadcast(gallivm
,
1020 lp_float32_vec4_type(),
1021 vs_type
, aos
, index
);
1027 fetch_vector(struct gallivm_state
*gallivm
,
1028 const struct util_format_description
*format_desc
,
1029 struct lp_type vs_type
,
1030 LLVMValueRef vb_stride
,
1031 LLVMValueRef map_ptr
,
1032 LLVMValueRef buffer_size_adj
,
1033 LLVMValueRef
*inputs
,
1034 LLVMValueRef indices
)
1036 LLVMBuilderRef builder
= gallivm
->builder
;
1037 struct lp_build_context blduivec
;
1038 struct lp_type fetch_type
= vs_type
;
1039 LLVMValueRef offset
, valid_mask
;
1042 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
1044 vb_stride
= lp_build_broadcast_scalar(&blduivec
, vb_stride
);
1045 buffer_size_adj
= lp_build_broadcast_scalar(&blduivec
, buffer_size_adj
);
1047 /* This mul can overflow. Wraparound is ok. */
1048 offset
= lp_build_mul(&blduivec
, vb_stride
, indices
);
1050 valid_mask
= lp_build_compare(gallivm
, blduivec
.type
,
1051 PIPE_FUNC_LESS
, offset
, buffer_size_adj
);
1053 /* not valid elements use offset 0 */
1054 offset
= LLVMBuildAnd(builder
, offset
, valid_mask
, "");
1057 lp_build_print_value(gallivm
, " indices = ", indices
);
1058 lp_build_print_value(gallivm
, " offsets = ", offset
);
1059 lp_build_print_value(gallivm
, " valid_mask = ", valid_mask
);
1063 * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
1064 * This should always produce better code.
1067 /* The type handling is annoying here... */
1068 if (format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
&&
1069 format_desc
->channel
[0].pure_integer
) {
1070 if (format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_SIGNED
) {
1071 fetch_type
= lp_type_int_vec(vs_type
.width
, vs_type
.width
* vs_type
.length
);
1073 else if (format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
1074 fetch_type
= lp_type_uint_vec(vs_type
.width
, vs_type
.width
* vs_type
.length
);
1078 lp_build_fetch_rgba_soa(gallivm
, format_desc
,
1079 fetch_type
, FALSE
, map_ptr
, offset
,
1080 blduivec
.zero
, blduivec
.zero
,
1083 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1084 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
],
1085 lp_build_vec_type(gallivm
, vs_type
), "");
1088 /* out-of-bound fetches return all zeros */
1089 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1090 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
], blduivec
.vec_type
, "");
1091 inputs
[i
] = LLVMBuildAnd(builder
, inputs
[i
], valid_mask
, "");
1092 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
],
1093 lp_build_vec_type(gallivm
, vs_type
), "");
1099 store_aos(struct gallivm_state
*gallivm
,
1100 LLVMValueRef io_ptr
,
1104 LLVMTypeRef data_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, lp_float32_vec4_type()), 0);
1105 LLVMBuilderRef builder
= gallivm
->builder
;
1106 LLVMValueRef data_ptr
= draw_jit_header_data(gallivm
, io_ptr
);
1107 LLVMValueRef indices
[3];
1109 indices
[0] = lp_build_const_int32(gallivm
, 0);
1111 indices
[2] = lp_build_const_int32(gallivm
, 0);
1113 data_ptr
= LLVMBuildGEP(builder
, data_ptr
, indices
, 3, "");
1114 data_ptr
= LLVMBuildPointerCast(builder
, data_ptr
, data_ptr_type
, "");
1117 lp_build_printf(gallivm
, " ---- %p storing attribute %d (io = %p)\n", data_ptr
, index
, io_ptr
);
1120 /* Unaligned store due to the vertex header */
1121 LLVMSetAlignment(LLVMBuildStore(builder
, value
, data_ptr
), sizeof(float));
1125 * Adjust the mask to architecture endianess. The mask will the store in struct:
1127 * struct vertex_header {
1128 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
1129 * unsigned edgeflag:1;
1131 * unsigned vertex_id:16;
1135 * On little-endian machine nothing needs to done, however on bit-endian machine
1136 * the mask's fields need to be adjusted with the algorithm:
1138 * uint32_t reverse (uint32_t x)
1140 * return (x >> 16) | // vertex_id
1141 * ((x & 0x3fff) << 18) | // clipmask
1142 * ((x & 0x4000) << 3) | // pad
1143 * ((x & 0x8000) << 1); // edgeflag
1147 adjust_mask(struct gallivm_state
*gallivm
,
1150 #if UTIL_ARCH_BIG_ENDIAN
1151 LLVMBuilderRef builder
= gallivm
->builder
;
1152 LLVMValueRef vertex_id
;
1153 LLVMValueRef clipmask
;
1155 LLVMValueRef edgeflag
;
1157 vertex_id
= LLVMBuildLShr(builder
, mask
, lp_build_const_int32(gallivm
, 16), "");
1158 clipmask
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x3fff), "");
1159 clipmask
= LLVMBuildShl(builder
, clipmask
, lp_build_const_int32(gallivm
, 18), "");
1161 pad
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x4000), "");
1162 pad
= LLVMBuildShl(builder
, pad
, lp_build_const_int32(gallivm
, 3), "");
1164 edgeflag
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x8000), "");
1165 edgeflag
= LLVMBuildShl(builder
, edgeflag
, lp_build_const_int32(gallivm
, 1), "");
1167 mask
= LLVMBuildOr(builder
, vertex_id
, clipmask
, "");
1169 mask
= LLVMBuildOr(builder
, mask
, pad
, "");
1171 mask
= LLVMBuildOr(builder
, mask
, edgeflag
, "");
1177 store_aos_array(struct gallivm_state
*gallivm
,
1178 struct lp_type soa_type
,
1179 LLVMValueRef io_ptr
,
1180 LLVMValueRef
*indices
,
1184 LLVMValueRef clipmask
,
1185 boolean need_edgeflag
)
1187 LLVMBuilderRef builder
= gallivm
->builder
;
1188 LLVMValueRef attr_index
= lp_build_const_int32(gallivm
, attrib
);
1189 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1190 LLVMValueRef linear_inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1191 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1192 int vector_length
= soa_type
.length
;
1195 debug_assert(TGSI_NUM_CHANNELS
== 4);
1197 for (i
= 0; i
< vector_length
; i
++) {
1198 linear_inds
[i
] = lp_build_const_int32(gallivm
, i
);
1200 inds
[i
] = indices
[i
];
1202 inds
[i
] = linear_inds
[i
];
1204 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
1208 /* store vertex header for each of the n vertices */
1209 LLVMValueRef val
, cliptmp
;
1210 int vertex_id_pad_edgeflag
;
1212 /* If this assertion fails, it means we need to update the bit twidding
1213 * code here. See struct vertex_header in draw_private.h.
1215 assert(DRAW_TOTAL_CLIP_PLANES
==14);
1216 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
1217 if (!need_edgeflag
) {
1218 vertex_id_pad_edgeflag
= (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES
);
1221 vertex_id_pad_edgeflag
= (0xffff << 16);
1223 val
= lp_build_const_int_vec(gallivm
, lp_int_type(soa_type
),
1224 vertex_id_pad_edgeflag
);
1225 /* OR with the clipmask */
1226 cliptmp
= LLVMBuildOr(builder
, val
, clipmask
, "");
1227 for (i
= 0; i
< vector_length
; i
++) {
1228 LLVMValueRef id_ptr
= draw_jit_header_id(gallivm
, io_ptrs
[i
]);
1229 val
= LLVMBuildExtractElement(builder
, cliptmp
, linear_inds
[i
], "");
1230 val
= adjust_mask(gallivm
, val
);
1232 lp_build_printf(gallivm
, "io = %p, index %d, clipmask = %x\n",
1233 io_ptrs
[i
], inds
[i
], val
);
1235 LLVMBuildStore(builder
, val
, id_ptr
);
1239 /* store for each of the n vertices */
1240 for (i
= 0; i
< vector_length
; i
++) {
1241 store_aos(gallivm
, io_ptrs
[i
], attr_index
, aos
[i
]);
1247 convert_to_aos(struct gallivm_state
*gallivm
,
1249 LLVMValueRef
*indices
,
1250 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1251 LLVMValueRef clipmask
,
1253 struct lp_type soa_type
,
1254 boolean need_edgeflag
)
1256 LLVMBuilderRef builder
= gallivm
->builder
;
1257 unsigned chan
, attrib
, i
;
1260 lp_build_printf(gallivm
, " # storing begin\n");
1262 for (attrib
= 0; attrib
< num_outputs
; ++attrib
) {
1263 LLVMValueRef soa
[TGSI_NUM_CHANNELS
];
1264 LLVMValueRef aos
[LP_MAX_VECTOR_WIDTH
/ 32];
1265 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1266 if (outputs
[attrib
][chan
]) {
1267 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
1268 lp_build_name(out
, "output%u.%c", attrib
, "xyzw"[chan
]);
1270 lp_build_printf(gallivm
, "output %d : %d ",
1271 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1273 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1275 lp_build_print_value(gallivm
, "val = ", out
);
1278 LLVMBuildBitCast(builder
, out
, lp_build_int_vec_type(gallivm
, soa_type
), "");
1280 lp_build_print_value(gallivm
, " ival = ", iv
);
1291 if (soa_type
.length
== TGSI_NUM_CHANNELS
) {
1292 lp_build_transpose_aos(gallivm
, soa_type
, soa
, aos
);
1294 lp_build_transpose_aos(gallivm
, soa_type
, soa
, soa
);
1296 for (i
= 0; i
< soa_type
.length
; ++i
) {
1297 aos
[i
] = lp_build_extract_range(gallivm
,
1298 soa
[i
% TGSI_NUM_CHANNELS
],
1299 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1304 store_aos_array(gallivm
,
1314 lp_build_printf(gallivm
, " # storing end\n");
1320 * Stores original vertex positions in clip coordinates
1323 store_clip(struct gallivm_state
*gallivm
,
1324 const struct lp_type vs_type
,
1325 LLVMValueRef io_ptr
,
1326 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1329 LLVMBuilderRef builder
= gallivm
->builder
;
1330 LLVMValueRef soa
[4];
1331 LLVMValueRef aos
[LP_MAX_VECTOR_LENGTH
];
1332 LLVMValueRef indices
[2];
1333 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1334 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1335 LLVMValueRef clip_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1336 LLVMTypeRef clip_ptr_type
=
1337 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
),
1342 indices
[1] = lp_build_const_int32(gallivm
, 0);
1344 for (i
= 0; i
< vs_type
.length
; i
++) {
1345 inds
[i
] = lp_build_const_int32(gallivm
, i
);
1346 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
1349 soa
[0] = LLVMBuildLoad(builder
, outputs
[idx
][0], ""); /*x0 x1 .. xn*/
1350 soa
[1] = LLVMBuildLoad(builder
, outputs
[idx
][1], ""); /*y0 y1 .. yn*/
1351 soa
[2] = LLVMBuildLoad(builder
, outputs
[idx
][2], ""); /*z0 z1 .. zn*/
1352 soa
[3] = LLVMBuildLoad(builder
, outputs
[idx
][3], ""); /*w0 w1 .. wn*/
1354 for (i
= 0; i
< vs_type
.length
; i
++) {
1355 clip_ptrs
[i
] = draw_jit_header_clip_pos(gallivm
, io_ptrs
[i
]);
1358 lp_build_transpose_aos(gallivm
, vs_type
, soa
, soa
);
1359 for (i
= 0; i
< vs_type
.length
; ++i
) {
1360 aos
[i
] = lp_build_extract_range(gallivm
,
1361 soa
[i
% TGSI_NUM_CHANNELS
],
1362 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1366 for (j
= 0; j
< vs_type
.length
; j
++) {
1367 LLVMValueRef clip_ptr
;
1369 clip_ptr
= LLVMBuildGEP(builder
, clip_ptrs
[j
], indices
, 2, "clipo");
1370 clip_ptr
= LLVMBuildPointerCast(builder
, clip_ptr
, clip_ptr_type
, "");
1372 /* Unaligned store */
1373 LLVMSetAlignment(LLVMBuildStore(builder
, aos
[j
], clip_ptr
), sizeof(float));
1379 * Transforms the outputs for viewport mapping
1382 generate_viewport(struct draw_llvm_variant
*variant
,
1383 LLVMBuilderRef builder
,
1384 struct lp_type vs_type
,
1385 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1386 LLVMValueRef context_ptr
)
1389 struct gallivm_state
*gallivm
= variant
->gallivm
;
1390 struct lp_type f32_type
= vs_type
;
1391 const unsigned pos
= variant
->llvm
->draw
->vs
.position_output
;
1392 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1393 LLVMValueRef out3
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn*/
1394 LLVMValueRef const1
= lp_build_const_vec(gallivm
, f32_type
, 1.0); /*1.0 1.0 1.0 1.0*/
1395 LLVMValueRef vp_ptr
= draw_jit_context_viewports(gallivm
, context_ptr
);
1397 /* We treat pipe_viewport_state as a float array */
1398 const int scale_index_offset
= offsetof(struct pipe_viewport_state
, scale
) / sizeof(float);
1399 const int trans_index_offset
= offsetof(struct pipe_viewport_state
, translate
) / sizeof(float);
1401 /* for 1/w convention*/
1402 out3
= LLVMBuildFDiv(builder
, const1
, out3
, "");
1403 LLVMBuildStore(builder
, out3
, outputs
[pos
][3]);
1405 /* Viewport Mapping */
1406 for (i
=0; i
<3; i
++) {
1407 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[pos
][i
], ""); /*x0 x1 .. xn*/
1410 LLVMValueRef scale_i
;
1411 LLVMValueRef trans_i
;
1414 index
= lp_build_const_int32(gallivm
, i
+ scale_index_offset
);
1415 scale_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1417 index
= lp_build_const_int32(gallivm
, i
+ trans_index_offset
);
1418 trans_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1420 scale
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1421 LLVMBuildLoad(builder
, scale_i
, "scale"));
1422 trans
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1423 LLVMBuildLoad(builder
, trans_i
, "trans"));
1426 out
= LLVMBuildFMul(builder
, out
, out3
, "");
1427 /* mult by scale, add translation */
1428 out
= lp_build_fmuladd(builder
, out
, scale
, trans
);
1430 /* store transformed outputs */
1431 LLVMBuildStore(builder
, out
, outputs
[pos
][i
]);
1438 * Returns clipmask as nxi32 bitmask for the n vertices
1441 generate_clipmask(struct draw_llvm
*llvm
,
1442 struct gallivm_state
*gallivm
,
1443 struct lp_type vs_type
,
1444 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1445 struct draw_llvm_variant_key
*key
,
1446 LLVMValueRef context_ptr
,
1447 boolean
*have_clipdist
)
1449 LLVMBuilderRef builder
= gallivm
->builder
;
1450 LLVMValueRef mask
; /* stores the <nxi32> clipmasks */
1451 LLVMValueRef test
, temp
;
1452 LLVMValueRef zero
, shift
;
1453 LLVMValueRef pos_x
, pos_y
, pos_z
, pos_w
;
1454 LLVMValueRef cv_x
, cv_y
, cv_z
, cv_w
;
1455 LLVMValueRef plane1
, planes
, plane_ptr
, sum
;
1456 struct lp_type f32_type
= vs_type
;
1457 struct lp_type i32_type
= lp_int_type(vs_type
);
1458 const unsigned pos
= llvm
->draw
->vs
.position_output
;
1459 const unsigned cv
= llvm
->draw
->vs
.clipvertex_output
;
1460 int num_written_clipdistance
= llvm
->draw
->vs
.vertex_shader
->info
.num_written_clipdistance
;
1461 boolean have_cd
= false;
1462 boolean clip_user
= key
->clip_user
;
1463 unsigned ucp_enable
= key
->ucp_enable
;
1466 cd
[0] = llvm
->draw
->vs
.ccdistance_output
[0];
1467 cd
[1] = llvm
->draw
->vs
.ccdistance_output
[1];
1469 if (cd
[0] != pos
|| cd
[1] != pos
)
1472 if (num_written_clipdistance
&& !clip_user
) {
1474 ucp_enable
= (1 << num_written_clipdistance
) - 1;
1477 mask
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1478 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1479 zero
= lp_build_const_vec(gallivm
, f32_type
, 0); /* 0.0f 0.0f 0.0f 0.0f */
1480 shift
= lp_build_const_int_vec(gallivm
, i32_type
, 1); /* 1 1 1 1 */
1483 * load clipvertex and position from correct locations.
1484 * if they are the same just load them once.
1486 pos_x
= LLVMBuildLoad(builder
, outputs
[pos
][0], ""); /*x0 x1 .. xn */
1487 pos_y
= LLVMBuildLoad(builder
, outputs
[pos
][1], ""); /*y0 y1 .. yn */
1488 pos_z
= LLVMBuildLoad(builder
, outputs
[pos
][2], ""); /*z0 z1 .. zn */
1489 pos_w
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn */
1491 if (clip_user
&& cv
!= pos
) {
1492 cv_x
= LLVMBuildLoad(builder
, outputs
[cv
][0], ""); /*x0 x1 .. xn */
1493 cv_y
= LLVMBuildLoad(builder
, outputs
[cv
][1], ""); /*y0 y1 .. yn */
1494 cv_z
= LLVMBuildLoad(builder
, outputs
[cv
][2], ""); /*z0 z1 .. zn */
1495 cv_w
= LLVMBuildLoad(builder
, outputs
[cv
][3], ""); /*w0 w1 .. wn */
1504 * Be careful with the comparisons and NaNs (using llvm's unordered
1505 * comparisons here).
1507 /* Cliptest, for hardwired planes */
1509 * XXX should take guardband into account (currently not in key).
1510 * Otherwise might run the draw pipeline stages for nothing.
1514 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_x
, pos_w
);
1516 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1520 test
= LLVMBuildFAdd(builder
, pos_x
, pos_w
, "");
1521 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1522 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1523 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1524 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1527 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_y
, pos_w
);
1528 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1529 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1530 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1533 test
= LLVMBuildFAdd(builder
, pos_y
, pos_w
, "");
1534 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1535 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1536 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1537 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1541 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 16);
1542 if (key
->clip_halfz
) {
1544 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, pos_z
);
1545 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1546 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1550 test
= LLVMBuildFAdd(builder
, pos_z
, pos_w
, "");
1551 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1552 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1553 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1556 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_z
, pos_w
);
1557 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1558 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1559 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1563 LLVMValueRef planes_ptr
= draw_jit_context_planes(gallivm
, context_ptr
);
1564 LLVMValueRef indices
[3];
1565 LLVMValueRef is_nan_or_inf
;
1567 /* userclip planes */
1568 while (ucp_enable
) {
1569 unsigned plane_idx
= ffs(ucp_enable
)-1;
1570 ucp_enable
&= ~(1 << plane_idx
);
1573 if (have_cd
&& num_written_clipdistance
) {
1574 LLVMValueRef clipdist
;
1578 *have_clipdist
= TRUE
;
1580 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[0]][i
], "");
1582 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[1]][i
-4], "");
1584 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, clipdist
);
1585 is_nan_or_inf
= lp_build_is_inf_or_nan(gallivm
, vs_type
, clipdist
);
1586 test
= LLVMBuildOr(builder
, test
, is_nan_or_inf
, "");
1587 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1588 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1589 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1591 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1592 indices
[0] = lp_build_const_int32(gallivm
, 0);
1593 indices
[1] = lp_build_const_int32(gallivm
, plane_idx
);
1595 indices
[2] = lp_build_const_int32(gallivm
, 0);
1596 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1597 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_x");
1598 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1599 sum
= LLVMBuildFMul(builder
, planes
, cv_x
, "");
1601 indices
[2] = lp_build_const_int32(gallivm
, 1);
1602 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1603 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_y");
1604 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1605 sum
= lp_build_fmuladd(builder
, planes
, cv_y
, sum
);
1607 indices
[2] = lp_build_const_int32(gallivm
, 2);
1608 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1609 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_z");
1610 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1611 sum
= lp_build_fmuladd(builder
, planes
, cv_z
, sum
);
1613 indices
[2] = lp_build_const_int32(gallivm
, 3);
1614 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1615 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_w");
1616 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1617 sum
= lp_build_fmuladd(builder
, planes
, cv_w
, sum
);
1619 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, sum
);
1620 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1621 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1622 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1626 if (key
->need_edgeflags
) {
1628 * This isn't really part of clipmask but stored the same in vertex
1629 * header later, so do it here.
1631 unsigned edge_attr
= llvm
->draw
->vs
.edgeflag_output
;
1632 LLVMValueRef one
= lp_build_const_vec(gallivm
, f32_type
, 1.0);
1633 LLVMValueRef edgeflag
= LLVMBuildLoad(builder
, outputs
[edge_attr
][0], "");
1634 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_EQUAL
, one
, edgeflag
);
1635 temp
= lp_build_const_int_vec(gallivm
, i32_type
,
1636 1LL << DRAW_TOTAL_CLIP_PLANES
);
1637 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1638 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1645 * Returns boolean if any clipping has occurred
1646 * Used zero/one i8 value to represent boolean
1649 clipmask_booli8(struct gallivm_state
*gallivm
,
1650 const struct lp_type vs_type
,
1651 LLVMValueRef clipmask_bool_ptr
,
1652 boolean edgeflag_in_clipmask
)
1654 LLVMBuilderRef builder
= gallivm
->builder
;
1655 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
1656 LLVMValueRef clipmask_bool
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1658 struct lp_build_context bldivec
;
1660 lp_build_context_init(&bldivec
, gallivm
, lp_int_type(vs_type
));
1663 * We need to invert the edgeflag bit from the clipmask here
1664 * (because the result is really if we want to run the pipeline or not
1665 * and we (may) need it if edgeflag was 0).
1667 if (edgeflag_in_clipmask
) {
1668 LLVMValueRef edge
= lp_build_const_int_vec(gallivm
, bldivec
.type
,
1669 1LL << DRAW_TOTAL_CLIP_PLANES
);
1670 clipmask_bool
= LLVMBuildXor(builder
, clipmask_bool
, edge
, "");
1674 * XXX: probably should mask off bits from the mask which come from
1675 * vertices which were beyond the count (i.e. indices_valid for
1676 * linear fetches, for elts ones we don't have the correct mask
1677 * right now). Otherwise might run the pipeline for nothing,
1678 * though everything should still work.
1680 ret
= lp_build_any_true_range(&bldivec
, vs_type
.length
, clipmask_bool
);
1681 ret
= LLVMBuildZExt(builder
, ret
, int8_type
, "");
1686 draw_gs_llvm_fetch_input(const struct lp_build_gs_iface
*gs_iface
,
1687 struct lp_build_context
* bld
,
1688 boolean is_vindex_indirect
,
1689 LLVMValueRef vertex_index
,
1690 boolean is_aindex_indirect
,
1691 LLVMValueRef attrib_index
,
1692 LLVMValueRef swizzle_index
)
1694 const struct draw_gs_llvm_iface
*gs
= draw_gs_llvm_iface(gs_iface
);
1695 struct gallivm_state
*gallivm
= bld
->gallivm
;
1696 LLVMBuilderRef builder
= gallivm
->builder
;
1697 LLVMValueRef indices
[3];
1699 struct lp_type type
= bld
->type
;
1701 if (is_vindex_indirect
|| is_aindex_indirect
) {
1704 for (i
= 0; i
< type
.length
; ++i
) {
1705 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
1706 LLVMValueRef vert_chan_index
= vertex_index
;
1707 LLVMValueRef attr_chan_index
= attrib_index
;
1708 LLVMValueRef channel_vec
, value
;
1710 if (is_vindex_indirect
) {
1711 vert_chan_index
= LLVMBuildExtractElement(builder
,
1712 vertex_index
, idx
, "");
1714 if (is_aindex_indirect
) {
1715 attr_chan_index
= LLVMBuildExtractElement(builder
,
1716 attrib_index
, idx
, "");
1719 indices
[0] = vert_chan_index
;
1720 indices
[1] = attr_chan_index
;
1721 indices
[2] = swizzle_index
;
1723 channel_vec
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1724 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
1725 value
= LLVMBuildExtractElement(builder
, channel_vec
, idx
, "");
1727 res
= LLVMBuildInsertElement(builder
, res
, value
, idx
, "");
1730 indices
[0] = vertex_index
;
1731 indices
[1] = attrib_index
;
1732 indices
[2] = swizzle_index
;
1734 res
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1735 res
= LLVMBuildLoad(builder
, res
, "");
1742 draw_gs_llvm_emit_vertex(const struct lp_build_gs_iface
*gs_base
,
1743 struct lp_build_context
* bld
,
1744 LLVMValueRef (*outputs
)[4],
1745 LLVMValueRef emitted_vertices_vec
,
1746 LLVMValueRef stream_id
)
1748 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1749 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1750 struct gallivm_state
*gallivm
= variant
->gallivm
;
1751 LLVMBuilderRef builder
= gallivm
->builder
;
1752 struct lp_type gs_type
= bld
->type
;
1753 LLVMValueRef clipmask
= lp_build_const_int_vec(gallivm
,
1754 lp_int_type(gs_type
), 0);
1755 LLVMValueRef indices
[LP_MAX_VECTOR_LENGTH
];
1756 LLVMValueRef next_prim_offset
=
1757 lp_build_const_int32(gallivm
, variant
->shader
->base
.primitive_boundary
);
1758 LLVMValueRef io
= variant
->io_ptr
;
1760 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
1762 for (i
= 0; i
< gs_type
.length
; ++i
) {
1763 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1764 LLVMValueRef currently_emitted
=
1765 LLVMBuildExtractElement(builder
, emitted_vertices_vec
, ind
, "");
1766 indices
[i
] = LLVMBuildMul(builder
, ind
, next_prim_offset
, "");
1767 indices
[i
] = LLVMBuildAdd(builder
, indices
[i
], currently_emitted
, "");
1770 LLVMValueRef stream_idx
= LLVMBuildExtractElement(builder
, stream_id
, lp_build_const_int32(gallivm
, 0), "");
1771 LLVMValueRef cnd
= LLVMBuildICmp(builder
, LLVMIntULT
, stream_idx
, lp_build_const_int32(gallivm
, variant
->shader
->base
.num_vertex_streams
), "");
1772 struct lp_build_if_state if_ctx
;
1773 lp_build_if(&if_ctx
, gallivm
, cnd
);
1774 io
= lp_build_pointer_get(builder
, io
, LLVMBuildExtractElement(builder
, stream_id
, lp_build_const_int32(gallivm
, 0), ""));
1776 convert_to_aos(gallivm
, io
, indices
,
1778 gs_info
->num_outputs
, gs_type
,
1780 lp_build_endif(&if_ctx
);
1784 draw_gs_llvm_end_primitive(const struct lp_build_gs_iface
*gs_base
,
1785 struct lp_build_context
* bld
,
1786 LLVMValueRef total_emitted_vertices_vec_ptr
,
1787 LLVMValueRef verts_per_prim_vec
,
1788 LLVMValueRef emitted_prims_vec
,
1789 LLVMValueRef mask_vec
)
1791 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1792 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1793 struct gallivm_state
*gallivm
= variant
->gallivm
;
1794 LLVMBuilderRef builder
= gallivm
->builder
;
1795 LLVMValueRef prim_lengts_ptr
=
1796 draw_gs_jit_prim_lengths(variant
->gallivm
, variant
->context_ptr
);
1799 for (i
= 0; i
< bld
->type
.length
; ++i
) {
1800 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1801 LLVMValueRef prims_emitted
=
1802 LLVMBuildExtractElement(builder
, emitted_prims_vec
, ind
, "");
1803 LLVMValueRef store_ptr
;
1804 LLVMValueRef num_vertices
=
1805 LLVMBuildExtractElement(builder
, verts_per_prim_vec
, ind
, "");
1807 store_ptr
= LLVMBuildGEP(builder
, prim_lengts_ptr
, &prims_emitted
, 1, "");
1808 store_ptr
= LLVMBuildLoad(builder
, store_ptr
, "");
1809 store_ptr
= LLVMBuildGEP(builder
, store_ptr
, &ind
, 1, "");
1810 LLVMBuildStore(builder
, num_vertices
, store_ptr
);
1815 draw_gs_llvm_epilogue(const struct lp_build_gs_iface
*gs_base
,
1816 LLVMValueRef total_emitted_vertices_vec
,
1817 LLVMValueRef emitted_prims_vec
, unsigned stream
)
1819 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1820 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1821 struct gallivm_state
*gallivm
= variant
->gallivm
;
1822 LLVMBuilderRef builder
= gallivm
->builder
;
1823 LLVMValueRef emitted_verts_ptr
=
1824 draw_gs_jit_emitted_vertices(gallivm
, variant
->context_ptr
);
1825 LLVMValueRef emitted_prims_ptr
=
1826 draw_gs_jit_emitted_prims(gallivm
, variant
->context_ptr
);
1827 LLVMValueRef stream_val
= lp_build_const_int32(gallivm
, stream
);
1829 emitted_verts_ptr
= LLVMBuildGEP(builder
, emitted_verts_ptr
, &stream_val
, 1, "");
1830 emitted_prims_ptr
= LLVMBuildGEP(builder
, emitted_prims_ptr
, &stream_val
, 1, "");
1832 LLVMBuildStore(builder
, total_emitted_vertices_vec
, emitted_verts_ptr
);
1833 LLVMBuildStore(builder
, emitted_prims_vec
, emitted_prims_ptr
);
1837 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*variant
)
1839 struct gallivm_state
*gallivm
= variant
->gallivm
;
1840 LLVMContextRef context
= gallivm
->context
;
1841 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
1842 LLVMTypeRef arg_types
[12];
1843 unsigned num_arg_types
= ARRAY_SIZE(arg_types
);
1844 LLVMTypeRef func_type
;
1845 LLVMValueRef context_ptr
;
1846 LLVMBasicBlockRef block
;
1847 LLVMBuilderRef builder
;
1849 struct lp_type vs_type
;
1850 LLVMValueRef count
, fetch_elts
, start_or_maxelt
;
1851 LLVMValueRef vertex_id_offset
;
1852 LLVMValueRef stride
, step
, io_itr
;
1853 LLVMValueRef ind_vec
, start_vec
, have_elts
, fetch_max
, tmp
;
1854 LLVMValueRef io_ptr
, vbuffers_ptr
, vb_ptr
;
1855 LLVMValueRef vb_stride
[PIPE_MAX_ATTRIBS
];
1856 LLVMValueRef map_ptr
[PIPE_MAX_ATTRIBS
];
1857 LLVMValueRef buffer_size_adj
[PIPE_MAX_ATTRIBS
];
1858 LLVMValueRef instance_index
[PIPE_MAX_ATTRIBS
];
1859 LLVMValueRef fake_buf_ptr
, fake_buf
;
1861 struct draw_context
*draw
= llvm
->draw
;
1862 const struct tgsi_shader_info
*vs_info
= &draw
->vs
.vertex_shader
->info
;
1864 struct lp_build_context bld
, blduivec
;
1865 struct lp_build_loop_state lp_loop
;
1866 struct lp_build_if_state if_ctx
;
1867 const int vector_length
= lp_native_vector_width
/ 32;
1868 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
1869 struct lp_build_sampler_soa
*sampler
= 0;
1870 struct lp_build_image_soa
*image
= NULL
;
1871 LLVMValueRef ret
, clipmask_bool_ptr
;
1872 struct draw_llvm_variant_key
*key
= &variant
->key
;
1873 /* If geometry shader is present we need to skip both the viewport
1874 * transformation and clipping otherwise the inputs to the geometry
1875 * shader will be incorrect.
1876 * The code can't handle vp transform when vs writes vp index neither
1877 * (though this would be fixable here, but couldn't just broadcast
1880 const boolean bypass_viewport
= key
->has_gs_or_tes
|| key
->bypass_viewport
||
1881 vs_info
->writes_viewport_index
;
1882 const boolean enable_cliptest
= !key
->has_gs_or_tes
&& (key
->clip_xy
||
1885 key
->need_edgeflags
);
1886 LLVMValueRef variant_func
;
1887 const unsigned pos
= draw
->vs
.position_output
;
1888 const unsigned cv
= draw
->vs
.clipvertex_output
;
1889 boolean have_clipdist
= FALSE
;
1890 struct lp_bld_tgsi_system_values system_values
;
1892 memset(&system_values
, 0, sizeof(system_values
));
1893 memset(&outputs
, 0, sizeof(outputs
));
1894 snprintf(func_name
, sizeof(func_name
), "draw_llvm_vs_variant");
1897 arg_types
[i
++] = get_context_ptr_type(variant
); /* context */
1898 arg_types
[i
++] = get_vertex_header_ptr_type(variant
); /* vertex_header */
1899 arg_types
[i
++] = get_buffer_ptr_type(variant
); /* vbuffers */
1900 arg_types
[i
++] = int32_type
; /* count */
1901 arg_types
[i
++] = int32_type
; /* start/fetch_elt_max */
1902 arg_types
[i
++] = int32_type
; /* stride */
1903 arg_types
[i
++] = get_vb_ptr_type(variant
); /* pipe_vertex_buffer's */
1904 arg_types
[i
++] = int32_type
; /* instance_id */
1905 arg_types
[i
++] = int32_type
; /* vertex_id_offset */
1906 arg_types
[i
++] = int32_type
; /* start_instance */
1907 arg_types
[i
++] = LLVMPointerType(int32_type
, 0); /* fetch_elts */
1908 arg_types
[i
++] = int32_type
; /* draw_id */
1910 func_type
= LLVMFunctionType(LLVMInt8TypeInContext(context
),
1911 arg_types
, num_arg_types
, 0);
1913 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
1914 variant
->function
= variant_func
;
1916 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
1917 for (i
= 0; i
< num_arg_types
; ++i
)
1918 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1919 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
1921 context_ptr
= LLVMGetParam(variant_func
, 0);
1922 io_ptr
= LLVMGetParam(variant_func
, 1);
1923 vbuffers_ptr
= LLVMGetParam(variant_func
, 2);
1924 count
= LLVMGetParam(variant_func
, 3);
1926 * XXX: the maxelt part is unused. Not really useful, since we cannot
1927 * get index buffer overflows due to vsplit (which provides its own
1928 * elts buffer, with a different size than what's passed in here).
1930 start_or_maxelt
= LLVMGetParam(variant_func
, 4);
1932 * XXX: stride is actually unused. The stride we use is strictly calculated
1933 * from the number of outputs (including the draw_extra outputs).
1934 * Should probably fix some day (we need a new vs just because of extra
1935 * outputs which the generated vs won't touch).
1937 stride
= LLVMGetParam(variant_func
, 5);
1938 vb_ptr
= LLVMGetParam(variant_func
, 6);
1939 system_values
.instance_id
= LLVMGetParam(variant_func
, 7);
1940 vertex_id_offset
= LLVMGetParam(variant_func
, 8);
1941 system_values
.base_instance
= LLVMGetParam(variant_func
, 9);
1942 fetch_elts
= LLVMGetParam(variant_func
, 10);
1943 system_values
.draw_id
= LLVMGetParam(variant_func
, 11);
1945 lp_build_name(context_ptr
, "context");
1946 lp_build_name(io_ptr
, "io");
1947 lp_build_name(vbuffers_ptr
, "vbuffers");
1948 lp_build_name(count
, "count");
1949 lp_build_name(start_or_maxelt
, "start_or_maxelt");
1950 lp_build_name(stride
, "stride");
1951 lp_build_name(vb_ptr
, "vb");
1952 lp_build_name(system_values
.instance_id
, "instance_id");
1953 lp_build_name(vertex_id_offset
, "vertex_id_offset");
1954 lp_build_name(system_values
.base_instance
, "start_instance");
1955 lp_build_name(fetch_elts
, "fetch_elts");
1956 lp_build_name(system_values
.draw_id
, "draw_id");
1962 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
1963 builder
= gallivm
->builder
;
1964 LLVMPositionBuilderAtEnd(builder
, block
);
1966 memset(&vs_type
, 0, sizeof vs_type
);
1967 vs_type
.floating
= TRUE
; /* floating point values */
1968 vs_type
.sign
= TRUE
; /* values are signed */
1969 vs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
1970 vs_type
.width
= 32; /* 32-bit float */
1971 vs_type
.length
= vector_length
;
1973 lp_build_context_init(&bld
, gallivm
, lp_type_uint(32));
1974 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
1976 /* hold temporary "bool" clipmask */
1977 clipmask_bool_ptr
= lp_build_alloca(gallivm
, blduivec
.vec_type
, "");
1979 fake_buf
= lp_build_alloca_undef(gallivm
,
1980 LLVMVectorType(LLVMInt64TypeInContext(context
), 4), "");
1981 fake_buf
= LLVMBuildBitCast(builder
, fake_buf
,
1982 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
1983 fake_buf_ptr
= LLVMBuildGEP(builder
, fake_buf
, &bld
.zero
, 1, "");
1985 /* code generated texture sampling */
1986 sampler
= draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key
));
1988 image
= draw_llvm_image_soa_create(draw_llvm_variant_key_images(key
));
1990 step
= lp_build_const_int32(gallivm
, vector_length
);
1992 ind_vec
= blduivec
.undef
;
1993 for (i
= 0; i
< vs_type
.length
; i
++) {
1994 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
1995 ind_vec
= LLVMBuildInsertElement(builder
, ind_vec
, index
, index
, "");
1998 have_elts
= LLVMBuildICmp(builder
, LLVMIntNE
,
1999 LLVMConstPointerNull(arg_types
[10]), fetch_elts
, "");
2001 fetch_max
= LLVMBuildSub(builder
, count
, bld
.one
, "fetch_max");
2002 fetch_max
= lp_build_broadcast_scalar(&blduivec
, fetch_max
);
2004 * Only needed for non-indexed path.
2006 start_vec
= lp_build_broadcast_scalar(&blduivec
, start_or_maxelt
);
2009 * Pre-calculate everything which is constant per shader invocation.
2011 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
2012 LLVMValueRef vb_buffer_offset
, buffer_size
, temp_ptr
;
2013 LLVMValueRef vb_info
, vbuffer_ptr
, buf_offset
, ofbit
;
2014 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
2015 LLVMValueRef vb_index
=
2016 lp_build_const_int32(gallivm
, velem
->vertex_buffer_index
);
2017 LLVMValueRef bsize
= lp_build_const_int32(gallivm
,
2018 util_format_get_blocksize(velem
->src_format
));
2019 LLVMValueRef src_offset
= lp_build_const_int32(gallivm
,
2021 struct lp_build_if_state if_ctx
;
2023 if (velem
->src_format
!= PIPE_FORMAT_NONE
) {
2024 vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffers_ptr
, &vb_index
, 1, "");
2025 vb_info
= LLVMBuildGEP(builder
, vb_ptr
, &vb_index
, 1, "");
2026 vb_stride
[j
] = draw_jit_vbuffer_stride(gallivm
, vb_info
);
2027 vb_stride
[j
] = LLVMBuildZExt(gallivm
->builder
, vb_stride
[j
],
2028 LLVMInt32TypeInContext(context
), "");
2029 vb_buffer_offset
= draw_jit_vbuffer_offset(gallivm
, vb_info
);
2030 map_ptr
[j
] = draw_jit_dvbuffer_map(gallivm
, vbuffer_ptr
);
2031 buffer_size
= draw_jit_dvbuffer_size(gallivm
, vbuffer_ptr
);
2035 * We'll set buffer_size_adj to zero if we have of, so it will
2036 * always overflow later automatically without having to keep ofbit.
2037 * Overflows (with normal wraparound) doing the actual offset
2038 * calculation should be ok, just not for the buffer size calc.
2039 * It would also be possible to detect such overflows and return
2040 * zeros if that happens, but this would be more complex.
2042 buf_offset
= lp_build_add(&bld
, vb_buffer_offset
, src_offset
);
2043 tmp
= lp_build_sub(&bld
, bsize
, bld
.one
);
2044 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size
, tmp
,
2046 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size_adj
[j
],
2047 buf_offset
, &ofbit
);
2050 * We can't easily set fake vertex buffers outside the generated code.
2051 * Hence, set fake vertex buffers here instead basically, so fetch
2052 * code can always fetch using offset 0, eliminating all control flow
2053 * inside the main loop.
2054 * (Alternatively, could have control flow per vector skipping fetch
2055 * if ofbit is true.)
2057 if (velem
->instance_divisor
) {
2059 * Index is equal to the start instance plus the number of current
2060 * instance divided by the divisor. In this case we compute it as:
2061 * index = start_instance + (instance_id / divisor).
2062 * Note we could actually do the fetch here, outside the loop -
2063 * it's all constant, hopefully llvm recognizes this.
2065 LLVMValueRef current_instance
;
2066 current_instance
= LLVMBuildUDiv(builder
, system_values
.instance_id
,
2067 lp_build_const_int32(gallivm
,
2068 velem
->instance_divisor
),
2069 "instance_divisor");
2070 instance_index
[j
] = lp_build_uadd_overflow(gallivm
, system_values
.base_instance
,
2071 current_instance
, &ofbit
);
2074 buffer_size_adj
[j
] = LLVMBuildSelect(builder
, ofbit
, bld
.zero
,
2075 buffer_size_adj
[j
], "");
2077 temp_ptr
= lp_build_alloca_undef(gallivm
,
2078 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
2080 lp_build_if(&if_ctx
, gallivm
, ofbit
);
2082 LLVMBuildStore(builder
, fake_buf_ptr
, temp_ptr
);
2084 lp_build_else(&if_ctx
);
2086 map_ptr
[j
] = LLVMBuildGEP(builder
, map_ptr
[j
], &buf_offset
, 1, "");
2087 LLVMBuildStore(builder
, map_ptr
[j
], temp_ptr
);
2089 lp_build_endif(&if_ctx
);
2090 map_ptr
[j
] = LLVMBuildLoad(builder
, temp_ptr
, "map_ptr");
2093 lp_build_printf(gallivm
, "velem %d, vbuf index = %u, vb_stride = %u\n",
2094 lp_build_const_int32(gallivm
, j
),
2095 vb_index
, vb_stride
[j
]);
2096 lp_build_printf(gallivm
,
2097 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
2098 vb_buffer_offset
, src_offset
, buf_offset
);
2099 lp_build_printf(gallivm
, " buffer size = %u, blocksize = %u\n",
2100 buffer_size
, bsize
);
2101 lp_build_printf(gallivm
, " instance_id = %u\n", system_values
.instance_id
);
2106 lp_build_loop_begin(&lp_loop
, gallivm
, bld
.zero
);
2108 LLVMValueRef inputs
[PIPE_MAX_SHADER_INPUTS
][TGSI_NUM_CHANNELS
];
2110 LLVMValueRef clipmask
; /* holds the clipmask value */
2111 LLVMValueRef true_index_array
, index_store
;
2112 const LLVMValueRef (*ptr_aos
)[TGSI_NUM_CHANNELS
];
2114 io_itr
= lp_loop
.counter
;
2116 io
= LLVMBuildGEP(builder
, io_ptr
, &io_itr
, 1, "");
2118 lp_build_printf(gallivm
, " --- io %d = %p, loop counter %d\n",
2119 io_itr
, io
, lp_loop
.counter
);
2122 true_index_array
= lp_build_broadcast_scalar(&blduivec
, lp_loop
.counter
);
2123 true_index_array
= LLVMBuildAdd(builder
, true_index_array
, ind_vec
, "");
2125 LLVMValueRef exec_mask
= lp_build_cmp(&blduivec
, PIPE_FUNC_LEQUAL
, true_index_array
, fetch_max
);
2127 * Limit indices to fetch_max, otherwise might try to access indices
2128 * beyond index buffer (or rather vsplit elt buffer) size.
2129 * Could probably safely (?) skip this for non-indexed draws and
2130 * simplify things minimally (by removing it could combine the ind_vec
2131 * and start_vec adds). I think the only effect for non-indexed draws will
2132 * be that for the invalid elements they will be all fetched from the
2133 * same location as the last valid one, but noone should really care.
2135 true_index_array
= lp_build_min(&blduivec
, true_index_array
, fetch_max
);
2137 index_store
= lp_build_alloca_undef(gallivm
, blduivec
.vec_type
, "index_store");
2139 lp_build_if(&if_ctx
, gallivm
, have_elts
);
2142 * Note: you'd expect some comparison/clamp against fetch_elt_max
2144 * There used to be one here but it was incorrect: overflow was
2145 * detected if index > fetch_elt_max - but the correct condition
2146 * would be index >= fetch_elt_max (since this is just size of elts
2147 * buffer / element size).
2148 * Using the correct condition however will cause failures - due to
2149 * vsplit/vcache code which rebases indices. So, as an example, if
2150 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
2151 * replace all invalid indices with 0 - which in case of elt_bias
2152 * not being zero will get a different fetch index than the valid
2153 * index 0. So, just rely on vsplit code preventing out-of-bounds
2154 * fetches. This is also why it's safe to do elts fetch even if there
2155 * was no index buffer bound - the real buffer is never seen here, at
2156 * least not if there are index buffer overflows...
2160 * XXX should not have to do this, as scale can be handled
2161 * natively by loads (hits asserts though).
2163 tmp
= lp_build_shl_imm(&blduivec
, true_index_array
, 2);
2164 fetch_elts
= LLVMBuildBitCast(builder
, fetch_elts
,
2165 LLVMPointerType(LLVMInt8TypeInContext(context
),
2167 tmp
= lp_build_gather(gallivm
, vs_type
.length
,
2169 fetch_elts
, tmp
, FALSE
);
2170 LLVMBuildStore(builder
, tmp
, index_store
);
2172 lp_build_else(&if_ctx
);
2174 tmp
= LLVMBuildAdd(builder
, true_index_array
, start_vec
, "");
2175 LLVMBuildStore(builder
, tmp
, index_store
);
2177 lp_build_endif(&if_ctx
);
2179 true_index_array
= LLVMBuildLoad(builder
, index_store
, "");
2181 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
2182 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
2183 const struct util_format_description
*format_desc
=
2184 util_format_description(velem
->src_format
);
2186 if (format_desc
->format
== PIPE_FORMAT_NONE
) {
2187 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
2188 inputs
[j
][i
] = lp_build_zero(gallivm
, vs_type
);
2191 else if (velem
->instance_divisor
) {
2192 fetch_instanced(gallivm
, format_desc
, vs_type
,
2193 vb_stride
[j
], map_ptr
[j
],
2195 inputs
[j
], instance_index
[j
]);
2198 fetch_vector(gallivm
, format_desc
, vs_type
,
2199 vb_stride
[j
], map_ptr
[j
],
2201 inputs
[j
], true_index_array
);
2205 struct lp_build_mask_context mask
;
2207 lp_build_mask_begin(&mask
, gallivm
, vs_type
, exec_mask
);
2208 /* In the paths with elts vertex id has to be unaffected by the
2209 * index bias and because indices inside our elements array have
2210 * already had index bias applied we need to subtract it here to
2211 * get back to the original index.
2212 * in the linear paths vertex id has to be unaffected by the
2213 * original start index and because we abuse the 'start' variable
2214 * to either represent the actual start index or the index at which
2215 * the primitive was split (we split rendering into chunks of at
2216 * most 4095-vertices) we need to back out the original start
2217 * index out of our vertex id here.
2218 * for ARB_shader_draw_parameters, base_vertex should be 0 for non-indexed draws.
2220 LLVMValueRef base_vertex
= lp_build_select(&bld
, have_elts
, vertex_id_offset
, lp_build_const_int32(gallivm
, 0));;
2221 system_values
.basevertex
= lp_build_broadcast_scalar(&blduivec
, base_vertex
);
2222 system_values
.vertex_id
= true_index_array
;
2223 system_values
.vertex_id_nobase
= LLVMBuildSub(builder
, true_index_array
,
2224 lp_build_broadcast_scalar(&blduivec
, vertex_id_offset
), "");
2226 ptr_aos
= (const LLVMValueRef (*)[TGSI_NUM_CHANNELS
]) inputs
;
2227 generate_vs(variant
,
2236 key
->clamp_vertex_color
,
2239 lp_build_mask_end(&mask
);
2240 if (pos
!= -1 && cv
!= -1) {
2241 /* store original positions in clip before further manipulation */
2242 store_clip(gallivm
, vs_type
, io
, outputs
, pos
);
2245 if (enable_cliptest
) {
2246 LLVMValueRef temp
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
2247 /* allocate clipmask, assign it integer type */
2248 clipmask
= generate_clipmask(llvm
,
2253 context_ptr
, &have_clipdist
);
2254 temp
= LLVMBuildOr(builder
, clipmask
, temp
, "");
2255 /* store temporary clipping boolean value */
2256 LLVMBuildStore(builder
, temp
, clipmask_bool_ptr
);
2259 clipmask
= blduivec
.zero
;
2262 /* do viewport mapping */
2263 if (!bypass_viewport
) {
2264 generate_viewport(variant
, builder
, vs_type
, outputs
, context_ptr
);
2268 clipmask
= blduivec
.zero
;
2271 /* store clipmask in vertex header,
2272 * original positions in clip
2273 * and transformed positions in data
2275 convert_to_aos(gallivm
, io
, NULL
, outputs
, clipmask
,
2276 vs_info
->num_outputs
, vs_type
,
2277 enable_cliptest
&& key
->need_edgeflags
);
2279 lp_build_loop_end_cond(&lp_loop
, count
, step
, LLVMIntUGE
);
2281 sampler
->destroy(sampler
);
2282 image
->destroy(image
);
2284 /* return clipping boolean value for function */
2285 ret
= clipmask_booli8(gallivm
, vs_type
, clipmask_bool_ptr
,
2286 enable_cliptest
&& key
->need_edgeflags
);
2288 LLVMBuildRet(builder
, ret
);
2290 gallivm_verify_function(gallivm
, variant_func
);
2294 struct draw_llvm_variant_key
*
2295 draw_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2298 struct draw_llvm_variant_key
*key
;
2299 struct draw_sampler_static_state
*draw_sampler
;
2300 struct draw_image_static_state
*draw_image
;
2302 key
= (struct draw_llvm_variant_key
*)store
;
2304 memset(key
, 0, offsetof(struct draw_llvm_variant_key
, vertex_element
[0]));
2306 key
->clamp_vertex_color
= llvm
->draw
->rasterizer
->clamp_vertex_color
; /**/
2308 /* will have to rig this up properly later */
2309 key
->clip_xy
= llvm
->draw
->clip_xy
;
2310 key
->clip_z
= llvm
->draw
->clip_z
;
2311 key
->clip_user
= llvm
->draw
->clip_user
;
2312 key
->bypass_viewport
= llvm
->draw
->bypass_viewport
;
2313 key
->clip_halfz
= llvm
->draw
->rasterizer
->clip_halfz
;
2314 /* XXX assumes edgeflag output not at 0 */
2315 key
->need_edgeflags
= (llvm
->draw
->vs
.edgeflag_output
? TRUE
: FALSE
);
2316 key
->ucp_enable
= llvm
->draw
->rasterizer
->clip_plane_enable
;
2317 key
->has_gs_or_tes
= llvm
->draw
->gs
.geometry_shader
!= NULL
|| llvm
->draw
->tes
.tess_eval_shader
!= NULL
;
2318 key
->num_outputs
= draw_total_vs_outputs(llvm
->draw
);
2320 /* All variants of this shader will have the same value for
2321 * nr_samplers. Not yet trying to compact away holes in the
2324 key
->nr_samplers
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2325 if (llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2326 key
->nr_sampler_views
=
2327 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2330 key
->nr_sampler_views
= key
->nr_samplers
;
2333 key
->nr_images
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
2335 /* Presumably all variants of the shader should have the same
2336 * number of vertex elements - ie the number of shader inputs.
2337 * NOTE: we NEED to store the needed number of needed inputs
2338 * here, not the number of provided elements to match keysize
2339 * (and the offset of sampler state in the key).
2340 * If we have excess number of vertex elements, this is valid,
2341 * but the excess ones don't matter.
2342 * If we don't have enough vertex elements (which looks not really
2343 * valid but we'll handle it gracefully) fill out missing ones with
2344 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2346 key
->nr_vertex_elements
=
2347 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
2349 if (llvm
->draw
->pt
.nr_vertex_elements
< key
->nr_vertex_elements
) {
2350 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2351 key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
);
2352 memset(key
->vertex_element
, 0,
2353 sizeof(struct pipe_vertex_element
) * key
->nr_vertex_elements
);
2355 memcpy(key
->vertex_element
,
2356 llvm
->draw
->pt
.vertex_element
,
2357 sizeof(struct pipe_vertex_element
) *
2358 MIN2(key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
));
2360 draw_sampler
= draw_llvm_variant_key_samplers(key
);
2361 memset(draw_sampler
, 0,
2362 MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
2364 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
2365 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
2366 llvm
->draw
->samplers
[PIPE_SHADER_VERTEX
][i
]);
2368 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2369 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
2370 llvm
->draw
->sampler_views
[PIPE_SHADER_VERTEX
][i
]);
2373 draw_image
= draw_llvm_variant_key_images(key
);
2374 memset(draw_image
, 0,
2375 key
->nr_images
* sizeof *draw_image
);
2376 for (i
= 0; i
< key
->nr_images
; i
++) {
2377 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
2378 llvm
->draw
->images
[PIPE_SHADER_VERTEX
][i
]);
2385 draw_llvm_dump_variant_key(struct draw_llvm_variant_key
*key
)
2388 struct draw_sampler_static_state
*sampler
= draw_llvm_variant_key_samplers(key
);
2389 struct draw_image_static_state
*image
= draw_llvm_variant_key_images(key
);
2390 debug_printf("clamp_vertex_color = %u\n", key
->clamp_vertex_color
);
2391 debug_printf("clip_xy = %u\n", key
->clip_xy
);
2392 debug_printf("clip_z = %u\n", key
->clip_z
);
2393 debug_printf("clip_user = %u\n", key
->clip_user
);
2394 debug_printf("bypass_viewport = %u\n", key
->bypass_viewport
);
2395 debug_printf("clip_halfz = %u\n", key
->clip_halfz
);
2396 debug_printf("need_edgeflags = %u\n", key
->need_edgeflags
);
2397 debug_printf("has_gs_or_tes = %u\n", key
->has_gs_or_tes
);
2398 debug_printf("ucp_enable = %u\n", key
->ucp_enable
);
2400 for (i
= 0 ; i
< key
->nr_vertex_elements
; i
++) {
2401 debug_printf("vertex_element[%i].src_offset = %u\n", i
, key
->vertex_element
[i
].src_offset
);
2402 debug_printf("vertex_element[%i].instance_divisor = %u\n", i
, key
->vertex_element
[i
].instance_divisor
);
2403 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i
, key
->vertex_element
[i
].vertex_buffer_index
);
2404 debug_printf("vertex_element[%i].src_format = %s\n", i
, util_format_name(key
->vertex_element
[i
].src_format
));
2407 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2408 debug_printf("sampler[%i].src_format = %s\n", i
, util_format_name(sampler
[i
].texture_state
.format
));
2411 for (i
= 0 ; i
< key
->nr_images
; i
++)
2412 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));
2417 draw_llvm_set_mapped_texture(struct draw_context
*draw
,
2418 enum pipe_shader_type shader_stage
,
2420 uint32_t width
, uint32_t height
, uint32_t depth
,
2421 uint32_t first_level
, uint32_t last_level
,
2422 uint32_t num_samples
,
2423 uint32_t sample_stride
,
2424 const void *base_ptr
,
2425 uint32_t row_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2426 uint32_t img_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2427 uint32_t mip_offsets
[PIPE_MAX_TEXTURE_LEVELS
])
2430 struct draw_jit_texture
*jit_tex
;
2432 switch (shader_stage
) {
2433 case PIPE_SHADER_VERTEX
:
2434 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->jit_context
.textures
));
2435 jit_tex
= &draw
->llvm
->jit_context
.textures
[sview_idx
];
2437 case PIPE_SHADER_GEOMETRY
:
2438 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->gs_jit_context
.textures
));
2439 jit_tex
= &draw
->llvm
->gs_jit_context
.textures
[sview_idx
];
2441 case PIPE_SHADER_TESS_CTRL
:
2442 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->tcs_jit_context
.textures
));
2443 jit_tex
= &draw
->llvm
->tcs_jit_context
.textures
[sview_idx
];
2445 case PIPE_SHADER_TESS_EVAL
:
2446 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->tes_jit_context
.textures
));
2447 jit_tex
= &draw
->llvm
->tes_jit_context
.textures
[sview_idx
];
2454 jit_tex
->width
= width
;
2455 jit_tex
->height
= height
;
2456 jit_tex
->depth
= depth
;
2457 jit_tex
->first_level
= first_level
;
2458 jit_tex
->last_level
= last_level
;
2459 jit_tex
->base
= base_ptr
;
2460 jit_tex
->num_samples
= num_samples
;
2461 jit_tex
->sample_stride
= sample_stride
;
2463 for (j
= first_level
; j
<= last_level
; j
++) {
2464 jit_tex
->mip_offsets
[j
] = mip_offsets
[j
];
2465 jit_tex
->row_stride
[j
] = row_stride
[j
];
2466 jit_tex
->img_stride
[j
] = img_stride
[j
];
2471 draw_llvm_set_mapped_image(struct draw_context
*draw
,
2472 enum pipe_shader_type shader_stage
,
2474 uint32_t width
, uint32_t height
, uint32_t depth
,
2475 const void *base_ptr
,
2476 uint32_t row_stride
,
2477 uint32_t img_stride
,
2478 uint32_t num_samples
,
2479 uint32_t sample_stride
)
2481 struct draw_jit_image
*jit_image
;
2483 switch (shader_stage
) {
2484 case PIPE_SHADER_VERTEX
:
2485 assert(idx
< ARRAY_SIZE(draw
->llvm
->jit_context
.images
));
2486 jit_image
= &draw
->llvm
->jit_context
.images
[idx
];
2488 case PIPE_SHADER_GEOMETRY
:
2489 assert(idx
< ARRAY_SIZE(draw
->llvm
->gs_jit_context
.images
));
2490 jit_image
= &draw
->llvm
->gs_jit_context
.images
[idx
];
2492 case PIPE_SHADER_TESS_CTRL
:
2493 assert(idx
< ARRAY_SIZE(draw
->llvm
->tcs_jit_context
.images
));
2494 jit_image
= &draw
->llvm
->tcs_jit_context
.images
[idx
];
2496 case PIPE_SHADER_TESS_EVAL
:
2497 assert(idx
< ARRAY_SIZE(draw
->llvm
->tes_jit_context
.images
));
2498 jit_image
= &draw
->llvm
->tes_jit_context
.images
[idx
];
2505 jit_image
->width
= width
;
2506 jit_image
->height
= height
;
2507 jit_image
->depth
= depth
;
2508 jit_image
->base
= base_ptr
;
2510 jit_image
->row_stride
= row_stride
;
2511 jit_image
->img_stride
= img_stride
;
2512 jit_image
->num_samples
= num_samples
;
2513 jit_image
->sample_stride
= sample_stride
;
2518 draw_llvm_set_sampler_state(struct draw_context
*draw
,
2519 enum pipe_shader_type shader_type
)
2523 switch (shader_type
) {
2524 case PIPE_SHADER_VERTEX
:
2525 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_VERTEX
]; i
++) {
2526 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->jit_context
.samplers
[i
];
2528 if (draw
->samplers
[PIPE_SHADER_VERTEX
][i
]) {
2529 const struct pipe_sampler_state
*s
2530 = draw
->samplers
[PIPE_SHADER_VERTEX
][i
];
2531 jit_sam
->min_lod
= s
->min_lod
;
2532 jit_sam
->max_lod
= s
->max_lod
;
2533 jit_sam
->lod_bias
= s
->lod_bias
;
2534 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2538 case PIPE_SHADER_GEOMETRY
:
2539 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_GEOMETRY
]; i
++) {
2540 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->gs_jit_context
.samplers
[i
];
2542 if (draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]) {
2543 const struct pipe_sampler_state
*s
2544 = draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
];
2545 jit_sam
->min_lod
= s
->min_lod
;
2546 jit_sam
->max_lod
= s
->max_lod
;
2547 jit_sam
->lod_bias
= s
->lod_bias
;
2548 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2552 case PIPE_SHADER_TESS_CTRL
:
2553 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_TESS_CTRL
]; i
++) {
2554 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->tcs_jit_context
.samplers
[i
];
2556 if (draw
->samplers
[PIPE_SHADER_TESS_CTRL
][i
]) {
2557 const struct pipe_sampler_state
*s
2558 = draw
->samplers
[PIPE_SHADER_TESS_CTRL
][i
];
2559 jit_sam
->min_lod
= s
->min_lod
;
2560 jit_sam
->max_lod
= s
->max_lod
;
2561 jit_sam
->lod_bias
= s
->lod_bias
;
2562 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2566 case PIPE_SHADER_TESS_EVAL
:
2567 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_TESS_EVAL
]; i
++) {
2568 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->tes_jit_context
.samplers
[i
];
2570 if (draw
->samplers
[PIPE_SHADER_TESS_EVAL
][i
]) {
2571 const struct pipe_sampler_state
*s
2572 = draw
->samplers
[PIPE_SHADER_TESS_EVAL
][i
];
2573 jit_sam
->min_lod
= s
->min_lod
;
2574 jit_sam
->max_lod
= s
->max_lod
;
2575 jit_sam
->lod_bias
= s
->lod_bias
;
2576 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2588 draw_llvm_destroy_variant(struct draw_llvm_variant
*variant
)
2590 struct draw_llvm
*llvm
= variant
->llvm
;
2592 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2593 debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2594 variant
->shader
->variants_cached
, llvm
->nr_variants
);
2597 gallivm_destroy(variant
->gallivm
);
2599 remove_from_list(&variant
->list_item_local
);
2600 variant
->shader
->variants_cached
--;
2601 remove_from_list(&variant
->list_item_global
);
2602 llvm
->nr_variants
--;
2608 * Create LLVM types for various structures.
2611 create_gs_jit_types(struct draw_gs_llvm_variant
*var
)
2613 struct gallivm_state
*gallivm
= var
->gallivm
;
2614 LLVMTypeRef texture_type
, sampler_type
, image_type
, context_type
;
2616 texture_type
= create_jit_texture_type(gallivm
, "texture");
2617 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
2618 image_type
= create_jit_image_type(gallivm
, "image");
2620 context_type
= create_gs_jit_context_type(gallivm
,
2621 var
->shader
->base
.vector_length
,
2622 texture_type
, sampler_type
,
2624 "draw_gs_jit_context");
2625 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
2627 var
->input_array_type
= create_gs_jit_input_type(gallivm
);
2631 get_gs_context_ptr_type(struct draw_gs_llvm_variant
*variant
)
2633 if (!variant
->context_ptr_type
)
2634 create_gs_jit_types(variant
);
2635 return variant
->context_ptr_type
;
2639 generate_mask_value(struct draw_gs_llvm_variant
*variant
,
2640 struct lp_type gs_type
)
2642 struct gallivm_state
*gallivm
= variant
->gallivm
;
2643 LLVMBuilderRef builder
= gallivm
->builder
;
2644 struct lp_type mask_type
= lp_int_type(gs_type
);
2645 LLVMValueRef num_prims
;
2646 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
2649 num_prims
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
),
2650 variant
->num_prims
);
2651 for (i
= 0; i
< gs_type
.length
; i
++) {
2652 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
2653 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, idx
, idx
, "");
2655 mask_val
= lp_build_compare(gallivm
, mask_type
,
2656 PIPE_FUNC_GREATER
, num_prims
, mask_val
);
2662 draw_gs_llvm_generate(struct draw_llvm
*llvm
,
2663 struct draw_gs_llvm_variant
*variant
)
2665 struct gallivm_state
*gallivm
= variant
->gallivm
;
2666 LLVMContextRef context
= gallivm
->context
;
2667 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
2668 LLVMTypeRef arg_types
[7];
2669 LLVMTypeRef func_type
;
2670 LLVMValueRef variant_func
;
2671 LLVMValueRef context_ptr
;
2672 LLVMValueRef prim_id_ptr
;
2673 LLVMBasicBlockRef block
;
2674 LLVMBuilderRef builder
;
2675 LLVMValueRef io_ptr
, input_array
, num_prims
, mask_val
;
2676 struct lp_build_sampler_soa
*sampler
= 0;
2677 struct lp_build_image_soa
*image
= NULL
;
2678 struct lp_build_context bld
;
2679 struct lp_bld_tgsi_system_values system_values
;
2681 struct lp_type gs_type
;
2683 struct draw_gs_llvm_iface gs_iface
;
2684 const struct tgsi_token
*tokens
= variant
->shader
->base
.state
.tokens
;
2685 LLVMValueRef consts_ptr
, num_consts_ptr
;
2686 LLVMValueRef ssbos_ptr
, num_ssbos_ptr
;
2687 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
2688 struct lp_build_mask_context mask
;
2689 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
2690 unsigned vector_length
= variant
->shader
->base
.vector_length
;
2692 memset(&system_values
, 0, sizeof(system_values
));
2693 memset(&outputs
, 0, sizeof(outputs
));
2695 snprintf(func_name
, sizeof(func_name
), "draw_llvm_gs_variant");
2697 assert(variant
->vertex_header_ptr_type
);
2699 arg_types
[0] = get_gs_context_ptr_type(variant
); /* context */
2700 arg_types
[1] = variant
->input_array_type
; /* input */
2701 arg_types
[2] = LLVMPointerType(variant
->vertex_header_ptr_type
, 0); /* vertex_header */
2702 arg_types
[3] = int32_type
; /* num_prims */
2703 arg_types
[4] = int32_type
; /* instance_id */
2704 arg_types
[5] = LLVMPointerType(
2705 LLVMVectorType(int32_type
, vector_length
), 0); /* prim_id_ptr */
2706 arg_types
[6] = int32_type
;
2708 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
), 0);
2710 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2712 variant
->function
= variant_func
;
2714 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
2716 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2717 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2718 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2720 context_ptr
= LLVMGetParam(variant_func
, 0);
2721 input_array
= LLVMGetParam(variant_func
, 1);
2722 io_ptr
= LLVMGetParam(variant_func
, 2);
2723 num_prims
= LLVMGetParam(variant_func
, 3);
2724 system_values
.instance_id
= LLVMGetParam(variant_func
, 4);
2725 prim_id_ptr
= LLVMGetParam(variant_func
, 5);
2726 system_values
.invocation_id
= LLVMGetParam(variant_func
, 6);
2728 lp_build_name(context_ptr
, "context");
2729 lp_build_name(input_array
, "input");
2730 lp_build_name(io_ptr
, "io");
2731 lp_build_name(num_prims
, "num_prims");
2732 lp_build_name(system_values
.instance_id
, "instance_id");
2733 lp_build_name(prim_id_ptr
, "prim_id_ptr");
2734 lp_build_name(system_values
.invocation_id
, "invocation_id");
2736 variant
->context_ptr
= context_ptr
;
2737 variant
->io_ptr
= io_ptr
;
2738 variant
->num_prims
= num_prims
;
2740 gs_iface
.base
.fetch_input
= draw_gs_llvm_fetch_input
;
2741 gs_iface
.base
.emit_vertex
= draw_gs_llvm_emit_vertex
;
2742 gs_iface
.base
.end_primitive
= draw_gs_llvm_end_primitive
;
2743 gs_iface
.base
.gs_epilogue
= draw_gs_llvm_epilogue
;
2744 gs_iface
.input
= input_array
;
2745 gs_iface
.variant
= variant
;
2751 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
2752 builder
= gallivm
->builder
;
2753 LLVMPositionBuilderAtEnd(builder
, block
);
2755 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
2757 memset(&gs_type
, 0, sizeof gs_type
);
2758 gs_type
.floating
= TRUE
; /* floating point values */
2759 gs_type
.sign
= TRUE
; /* values are signed */
2760 gs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2761 gs_type
.width
= 32; /* 32-bit float */
2762 gs_type
.length
= vector_length
;
2764 consts_ptr
= draw_gs_jit_context_constants(variant
->gallivm
, context_ptr
);
2766 draw_gs_jit_context_num_constants(variant
->gallivm
, context_ptr
);
2768 ssbos_ptr
= draw_gs_jit_context_ssbos(variant
->gallivm
, context_ptr
);
2770 draw_gs_jit_context_num_ssbos(variant
->gallivm
, context_ptr
);
2772 /* code generated texture sampling */
2773 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
);
2774 image
= draw_llvm_image_soa_create(draw_gs_llvm_variant_key_images(&variant
->key
));
2775 mask_val
= generate_mask_value(variant
, gs_type
);
2776 lp_build_mask_begin(&mask
, gallivm
, gs_type
, mask_val
);
2778 if (gs_info
->uses_primid
) {
2779 system_values
.prim_id
= LLVMBuildLoad(builder
, prim_id_ptr
, "prim_id");
2782 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2783 if (llvm
->draw
->gs
.geometry_shader
->state
.type
== PIPE_SHADER_IR_TGSI
)
2784 tgsi_dump(tokens
, 0);
2786 nir_print_shader(llvm
->draw
->gs
.geometry_shader
->state
.ir
.nir
, stderr
);
2787 draw_gs_llvm_dump_variant_key(&variant
->key
);
2790 struct lp_build_tgsi_params params
;
2791 memset(¶ms
, 0, sizeof(params
));
2793 params
.type
= gs_type
;
2794 params
.mask
= &mask
;
2795 params
.consts_ptr
= consts_ptr
;
2796 params
.const_sizes_ptr
= num_consts_ptr
;
2797 params
.system_values
= &system_values
;
2798 params
.context_ptr
= context_ptr
;
2799 params
.sampler
= sampler
;
2800 params
.info
= &llvm
->draw
->gs
.geometry_shader
->info
;
2801 params
.gs_iface
= (const struct lp_build_gs_iface
*)&gs_iface
;
2802 params
.ssbo_ptr
= ssbos_ptr
;
2803 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
2804 params
.image
= image
;
2806 if (llvm
->draw
->gs
.geometry_shader
->state
.type
== PIPE_SHADER_IR_TGSI
)
2807 lp_build_tgsi_soa(variant
->gallivm
,
2812 lp_build_nir_soa(variant
->gallivm
,
2813 llvm
->draw
->gs
.geometry_shader
->state
.ir
.nir
,
2817 sampler
->destroy(sampler
);
2818 image
->destroy(image
);
2820 lp_build_mask_end(&mask
);
2822 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
2824 gallivm_verify_function(gallivm
, variant_func
);
2828 struct draw_gs_llvm_variant
*
2829 draw_gs_llvm_create_variant(struct draw_llvm
*llvm
,
2830 unsigned num_outputs
,
2831 const struct draw_gs_llvm_variant_key
*key
)
2833 struct draw_gs_llvm_variant
*variant
;
2834 struct llvm_geometry_shader
*shader
=
2835 llvm_geometry_shader(llvm
->draw
->gs
.geometry_shader
);
2836 LLVMTypeRef vertex_header
;
2837 char module_name
[64];
2839 variant
= MALLOC(sizeof *variant
+
2840 shader
->variant_key_size
-
2841 sizeof variant
->key
);
2845 variant
->llvm
= llvm
;
2846 variant
->shader
= shader
;
2848 snprintf(module_name
, sizeof(module_name
), "draw_llvm_gs_variant%u",
2849 variant
->shader
->variants_cached
);
2851 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, NULL
);
2853 create_gs_jit_types(variant
);
2855 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2857 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_outputs
);
2859 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
2861 draw_gs_llvm_generate(llvm
, variant
);
2863 gallivm_compile_module(variant
->gallivm
);
2865 variant
->jit_func
= (draw_gs_jit_func
)
2866 gallivm_jit_function(variant
->gallivm
, variant
->function
);
2868 gallivm_free_ir(variant
->gallivm
);
2870 variant
->list_item_global
.base
= variant
;
2871 variant
->list_item_local
.base
= variant
;
2872 /*variant->no = */shader
->variants_created
++;
2873 variant
->list_item_global
.base
= variant
;
2879 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant
*variant
)
2881 struct draw_llvm
*llvm
= variant
->llvm
;
2883 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2884 debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2885 variant
->shader
->variants_cached
, llvm
->nr_gs_variants
);
2888 gallivm_destroy(variant
->gallivm
);
2890 remove_from_list(&variant
->list_item_local
);
2891 variant
->shader
->variants_cached
--;
2892 remove_from_list(&variant
->list_item_global
);
2893 llvm
->nr_gs_variants
--;
2897 struct draw_gs_llvm_variant_key
*
2898 draw_gs_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2901 struct draw_gs_llvm_variant_key
*key
;
2902 struct draw_sampler_static_state
*draw_sampler
;
2903 struct draw_image_static_state
*draw_image
;
2905 key
= (struct draw_gs_llvm_variant_key
*)store
;
2907 memset(key
, 0, offsetof(struct draw_gs_llvm_variant_key
, samplers
[0]));
2909 key
->num_outputs
= draw_total_gs_outputs(llvm
->draw
);
2911 /* All variants of this shader will have the same value for
2912 * nr_samplers. Not yet trying to compact away holes in the
2915 key
->nr_samplers
= llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2916 if (llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2917 key
->nr_sampler_views
=
2918 llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2921 key
->nr_sampler_views
= key
->nr_samplers
;
2924 key
->nr_images
= llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
2926 draw_sampler
= key
->samplers
;
2928 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
2930 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
2931 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
2932 llvm
->draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]);
2934 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2935 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
2936 llvm
->draw
->sampler_views
[PIPE_SHADER_GEOMETRY
][i
]);
2939 draw_image
= draw_gs_llvm_variant_key_images(key
);
2940 memset(draw_image
, 0,
2941 key
->nr_images
* sizeof *draw_image
);
2942 for (i
= 0; i
< key
->nr_images
; i
++) {
2943 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
2944 llvm
->draw
->images
[PIPE_SHADER_GEOMETRY
][i
]);
2950 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key
*key
)
2953 struct draw_sampler_static_state
*sampler
= key
->samplers
;
2954 struct draw_image_static_state
*image
= draw_gs_llvm_variant_key_images(key
);
2955 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2956 debug_printf("sampler[%i].src_format = %s\n", i
,
2957 util_format_name(sampler
[i
].texture_state
.format
));
2960 for (i
= 0 ; i
< key
->nr_images
; i
++)
2961 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));
2966 create_tcs_jit_types(struct draw_tcs_llvm_variant
*var
)
2968 struct gallivm_state
*gallivm
= var
->gallivm
;
2969 LLVMTypeRef texture_type
, sampler_type
, image_type
, context_type
;
2971 texture_type
= create_jit_texture_type(gallivm
, "texture");
2972 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
2973 image_type
= create_jit_image_type(gallivm
, "image");
2975 context_type
= create_tcs_jit_context_type(gallivm
,
2977 texture_type
, sampler_type
,
2979 "draw_tcs_jit_context");
2980 var
->input_array_type
= create_tcs_jit_input_type(gallivm
);
2981 var
->output_array_type
= create_tcs_jit_output_type(gallivm
);
2982 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
2986 get_tcs_context_ptr_type(struct draw_tcs_llvm_variant
*variant
)
2988 if (!variant
->context_ptr_type
)
2989 create_tcs_jit_types(variant
);
2990 return variant
->context_ptr_type
;
2994 draw_tcs_llvm_emit_fetch_input(const struct lp_build_tcs_iface
*tes_iface
,
2995 struct lp_build_context
*bld
,
2996 boolean is_vindex_indirect
,
2997 LLVMValueRef vertex_index
,
2998 boolean is_aindex_indirect
,
2999 LLVMValueRef attrib_index
,
3000 LLVMValueRef swizzle_index
)
3002 const struct draw_tcs_llvm_iface
*tcs
= draw_tcs_llvm_iface(tes_iface
);
3003 struct gallivm_state
*gallivm
= bld
->gallivm
;
3004 LLVMBuilderRef builder
= gallivm
->builder
;
3005 LLVMValueRef indices
[3];
3007 struct lp_type type
= bld
->type
;
3009 if (is_vindex_indirect
|| is_aindex_indirect
) {
3013 for (i
= 0; i
< type
.length
; ++i
) {
3014 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3015 LLVMValueRef vert_chan_index
= vertex_index
;
3016 LLVMValueRef attr_chan_index
= attrib_index
;
3017 LLVMValueRef channel_vec
;
3019 if (is_vindex_indirect
) {
3020 vert_chan_index
= LLVMBuildExtractElement(builder
,
3021 vertex_index
, idx
, "");
3023 if (is_aindex_indirect
) {
3024 attr_chan_index
= LLVMBuildExtractElement(builder
,
3025 attrib_index
, idx
, "");
3028 indices
[0] = vert_chan_index
;
3029 indices
[1] = attr_chan_index
;
3030 indices
[2] = swizzle_index
;
3032 channel_vec
= LLVMBuildGEP(builder
, tcs
->input
, indices
, 3, "");
3033 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3035 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3038 indices
[0] = vertex_index
;
3039 indices
[1] = attrib_index
;
3040 indices
[2] = swizzle_index
;
3042 res
= LLVMBuildGEP(builder
, tcs
->input
, indices
, 3, "");
3043 res
= LLVMBuildLoad(builder
, res
, "");
3044 res
= lp_build_broadcast_scalar(bld
, res
);
3050 draw_tcs_llvm_emit_fetch_output(const struct lp_build_tcs_iface
*tes_iface
,
3051 struct lp_build_context
*bld
,
3052 boolean is_vindex_indirect
,
3053 LLVMValueRef vertex_index
,
3054 boolean is_aindex_indirect
,
3055 LLVMValueRef attrib_index
,
3056 LLVMValueRef swizzle_index
,
3059 const struct draw_tcs_llvm_iface
*tcs
= draw_tcs_llvm_iface(tes_iface
);
3060 struct gallivm_state
*gallivm
= bld
->gallivm
;
3061 LLVMBuilderRef builder
= gallivm
->builder
;
3062 LLVMValueRef indices
[3];
3064 struct lp_type type
= bld
->type
;
3066 if (is_vindex_indirect
|| is_aindex_indirect
) {
3070 for (i
= 0; i
< type
.length
; ++i
) {
3071 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3072 LLVMValueRef vert_chan_index
= vertex_index
;
3073 LLVMValueRef attr_chan_index
= attrib_index
;
3074 LLVMValueRef channel_vec
;
3076 if (is_vindex_indirect
) {
3077 vert_chan_index
= LLVMBuildExtractElement(builder
,
3078 vertex_index
, idx
, "");
3080 if (is_aindex_indirect
) {
3081 attr_chan_index
= LLVMBuildExtractElement(builder
,
3082 attrib_index
, idx
, "");
3085 indices
[0] = vert_chan_index
;
3086 indices
[1] = attr_chan_index
;
3087 indices
[2] = swizzle_index
;
3089 channel_vec
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3090 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3092 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3095 indices
[0] = vertex_index
? vertex_index
: lp_build_const_int32(gallivm
, 0);
3096 indices
[1] = attrib_index
;
3097 indices
[2] = swizzle_index
;
3099 res
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3100 res
= LLVMBuildLoad(builder
, res
, "");
3101 res
= lp_build_broadcast_scalar(bld
, res
);
3107 draw_tcs_llvm_emit_store_output(const struct lp_build_tcs_iface
*tes_iface
,
3108 struct lp_build_context
*bld
,
3110 boolean is_vindex_indirect
,
3111 LLVMValueRef vertex_index
,
3112 boolean is_aindex_indirect
,
3113 LLVMValueRef attrib_index
,
3114 LLVMValueRef swizzle_index
,
3116 LLVMValueRef mask_vec
)
3118 const struct draw_tcs_llvm_iface
*tcs
= draw_tcs_llvm_iface(tes_iface
);
3119 struct gallivm_state
*gallivm
= bld
->gallivm
;
3120 LLVMBuilderRef builder
= gallivm
->builder
;
3121 LLVMValueRef indices
[3];
3123 struct lp_type type
= bld
->type
;
3125 if (is_vindex_indirect
|| is_aindex_indirect
) {
3128 for (i
= 0; i
< type
.length
; ++i
) {
3129 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3130 LLVMValueRef vert_chan_index
= vertex_index
? vertex_index
: lp_build_const_int32(gallivm
, 0);
3131 LLVMValueRef attr_chan_index
= attrib_index
;
3132 LLVMValueRef channel_vec
;
3134 if (is_vindex_indirect
) {
3135 vert_chan_index
= LLVMBuildExtractElement(builder
,
3136 vertex_index
, idx
, "");
3138 if (is_aindex_indirect
) {
3139 attr_chan_index
= LLVMBuildExtractElement(builder
,
3140 attrib_index
, idx
, "");
3143 indices
[0] = vert_chan_index
;
3144 indices
[1] = attr_chan_index
;
3145 indices
[2] = swizzle_index
;
3147 channel_vec
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3149 res
= LLVMBuildExtractElement(builder
, value
, idx
, "");
3151 struct lp_build_if_state ifthen
;
3152 LLVMValueRef cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
, mask_vec
, lp_build_const_int_vec(gallivm
, bld
->type
, 0), "");
3153 cond
= LLVMBuildExtractElement(gallivm
->builder
, cond
, idx
, "");
3154 lp_build_if(&ifthen
, gallivm
, cond
);
3155 LLVMBuildStore(builder
, res
, channel_vec
);
3156 lp_build_endif(&ifthen
);
3159 indices
[0] = vertex_index
? vertex_index
: lp_build_const_int32(gallivm
, 0);
3160 indices
[1] = attrib_index
;
3161 indices
[2] = swizzle_index
;
3163 res
= LLVMBuildGEP(builder
, tcs
->output
, indices
, 3, "");
3164 for (unsigned i
= 0; i
< type
.length
; ++i
) {
3165 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3166 LLVMValueRef val
= LLVMBuildExtractElement(builder
, value
, idx
, "");
3168 struct lp_build_if_state ifthen
;
3169 LLVMValueRef cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
, mask_vec
, lp_build_const_int_vec(gallivm
, bld
->type
, 0), "");
3170 cond
= LLVMBuildExtractElement(gallivm
->builder
, cond
, idx
, "");
3171 lp_build_if(&ifthen
, gallivm
, cond
);
3172 LLVMBuildStore(builder
, val
, res
);
3173 lp_build_endif(&ifthen
);
3180 generate_tcs_mask_value(struct draw_tcs_llvm_variant
*variant
,
3181 struct lp_type tcs_type
, LLVMValueRef limit
, LLVMValueRef loop_counter
)
3183 struct gallivm_state
*gallivm
= variant
->gallivm
;
3184 LLVMBuilderRef builder
= gallivm
->builder
;
3185 struct lp_type mask_type
= lp_int_type(tcs_type
);
3186 LLVMValueRef num_vecs
;
3187 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
3190 num_vecs
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
), limit
);
3191 for (i
= 0; i
< tcs_type
.length
; i
++) {
3192 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3193 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, LLVMBuildAdd(builder
, loop_counter
, idx
, ""), idx
, "");
3195 mask_val
= lp_build_compare(gallivm
, mask_type
,
3196 PIPE_FUNC_GREATER
, num_vecs
, mask_val
);
3202 draw_tcs_llvm_generate(struct draw_llvm
*llvm
,
3203 struct draw_tcs_llvm_variant
*variant
)
3205 struct gallivm_state
*gallivm
= variant
->gallivm
;
3206 LLVMContextRef context
= gallivm
->context
;
3207 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
3208 LLVMTypeRef arg_types
[6];
3209 LLVMTypeRef func_type
, coro_func_type
;
3210 LLVMValueRef variant_func
, variant_coro
;
3211 LLVMValueRef context_ptr
;
3212 LLVMValueRef input_array
, output_array
, prim_id
, patch_vertices_in
;
3213 LLVMValueRef mask_val
;
3214 LLVMBasicBlockRef block
;
3215 LLVMBuilderRef builder
;
3216 struct lp_build_context bld
, bldvec
;
3217 struct lp_build_sampler_soa
*sampler
= 0;
3218 struct lp_build_image_soa
*image
= NULL
;
3219 struct lp_bld_tgsi_system_values system_values
;
3220 char func_name
[64], func_name_coro
[64];
3222 struct draw_tcs_llvm_iface tcs_iface
;
3223 struct lp_build_mask_context mask
;
3224 LLVMValueRef consts_ptr
, num_consts_ptr
;
3225 LLVMValueRef ssbos_ptr
, num_ssbos_ptr
;
3226 struct lp_type tcs_type
;
3227 unsigned vector_length
= variant
->shader
->base
.vector_length
;
3229 memset(&system_values
, 0, sizeof(system_values
));
3231 snprintf(func_name
, sizeof(func_name
), "draw_llvm_tcs_variant");
3233 snprintf(func_name_coro
, sizeof(func_name_coro
), "draw_llvm_tcs_coro_variant");
3235 arg_types
[0] = get_tcs_context_ptr_type(variant
); /* context */
3236 arg_types
[1] = variant
->input_array_type
; /* input */
3237 arg_types
[2] = variant
->output_array_type
;
3238 arg_types
[3] = int32_type
;
3239 arg_types
[4] = int32_type
;
3240 arg_types
[5] = int32_type
; /* coroutine only */
3242 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
) - 1, 0);
3244 coro_func_type
= LLVMFunctionType(LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0), arg_types
, ARRAY_SIZE(arg_types
), 0);
3246 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
3248 variant_coro
= LLVMAddFunction(gallivm
->module
, func_name_coro
, coro_func_type
);
3250 variant
->function
= variant_func
;
3251 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
3253 LLVMSetFunctionCallConv(variant_coro
, LLVMCCallConv
);
3255 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
) {
3256 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
) {
3257 lp_add_function_attr(variant_coro
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
3258 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
3262 context_ptr
= LLVMGetParam(variant_func
, 0);
3263 input_array
= LLVMGetParam(variant_func
, 1);
3264 output_array
= LLVMGetParam(variant_func
, 2);
3265 prim_id
= LLVMGetParam(variant_func
, 3);
3266 patch_vertices_in
= LLVMGetParam(variant_func
, 4);
3268 lp_build_name(context_ptr
, "context");
3269 lp_build_name(input_array
, "input");
3270 lp_build_name(output_array
, "output");
3271 lp_build_name(prim_id
, "prim_id");
3272 lp_build_name(patch_vertices_in
, "patch_vertices_in");
3274 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
3275 builder
= gallivm
->builder
;
3276 LLVMPositionBuilderAtEnd(builder
, block
);
3278 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
3280 memset(&tcs_type
, 0, sizeof tcs_type
);
3281 tcs_type
.floating
= TRUE
; /* floating point values */
3282 tcs_type
.sign
= TRUE
; /* values are signed */
3283 tcs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
3284 tcs_type
.width
= 32; /* 32-bit float */
3285 tcs_type
.length
= vector_length
;
3287 lp_build_context_init(&bldvec
, variant
->gallivm
, lp_int_type(tcs_type
));
3289 LLVMValueRef count
= lp_build_const_int32(gallivm
, variant
->shader
->base
.vertices_out
);
3290 LLVMValueRef step
= lp_build_const_int32(gallivm
, vector_length
);
3292 struct lp_build_loop_state loop_state
[2];
3293 LLVMValueRef num_inner_loop
;
3294 unsigned count_align
= util_align_npot(variant
->shader
->base
.vertices_out
, tcs_type
.length
);
3295 num_inner_loop
= lp_build_const_int32(gallivm
, count_align
/ tcs_type
.length
);
3296 LLVMTypeRef hdl_ptr_type
= LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
3297 LLVMValueRef coro_hdls
= LLVMBuildArrayAlloca(gallivm
->builder
, hdl_ptr_type
, num_inner_loop
, "coro_hdls");
3298 unsigned end_coroutine
= INT_MAX
;
3299 lp_build_loop_begin(&loop_state
[1], gallivm
,
3300 lp_build_const_int32(gallivm
, 0)); /* coroutine reentry loop */
3301 lp_build_loop_begin(&loop_state
[0], gallivm
,
3302 lp_build_const_int32(gallivm
, 0)); /* inner loop */
3304 LLVMValueRef args
[6];
3305 args
[0] = context_ptr
;
3306 args
[1] = input_array
;
3307 args
[2] = output_array
;
3309 args
[4] = patch_vertices_in
;
3310 args
[5] = loop_state
[0].counter
;
3311 LLVMValueRef coro_entry
= LLVMBuildGEP(builder
, coro_hdls
, &loop_state
[0].counter
, 1, "");
3312 LLVMValueRef coro_hdl
= LLVMBuildLoad(builder
, coro_entry
, "coro_hdl");
3314 struct lp_build_if_state ifstate
;
3315 LLVMValueRef cmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, loop_state
[1].counter
,
3316 lp_build_const_int32(gallivm
, 0), "");
3317 /* first time here - call the coroutine function entry point */
3318 lp_build_if(&ifstate
, gallivm
, cmp
);
3319 LLVMValueRef coro_ret
= LLVMBuildCall(builder
, variant_coro
, args
, 6, "");
3320 LLVMBuildStore(builder
, coro_ret
, coro_entry
);
3321 lp_build_else(&ifstate
);
3322 /* subsequent calls for this invocation - check if done. */
3323 LLVMValueRef coro_done
= lp_build_coro_done(gallivm
, coro_hdl
);
3324 struct lp_build_if_state ifstate2
;
3325 lp_build_if(&ifstate2
, gallivm
, coro_done
);
3326 /* if done destroy and force loop exit */
3327 lp_build_coro_destroy(gallivm
, coro_hdl
);
3328 lp_build_loop_force_set_counter(&loop_state
[1], lp_build_const_int32(gallivm
, end_coroutine
- 1));
3329 lp_build_else(&ifstate2
);
3330 /* otherwise resume the coroutine */
3331 lp_build_coro_resume(gallivm
, coro_hdl
);
3332 lp_build_endif(&ifstate2
);
3333 lp_build_endif(&ifstate
);
3334 lp_build_loop_force_reload_counter(&loop_state
[1]);
3336 lp_build_loop_end_cond(&loop_state
[0],
3339 lp_build_loop_end_cond(&loop_state
[1],
3340 lp_build_const_int32(gallivm
, end_coroutine
),
3342 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
3344 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_coro
, "entry");
3345 LLVMPositionBuilderAtEnd(builder
, block
);
3347 context_ptr
= LLVMGetParam(variant_coro
, 0);
3348 input_array
= LLVMGetParam(variant_coro
, 1);
3349 output_array
= LLVMGetParam(variant_coro
, 2);
3350 prim_id
= LLVMGetParam(variant_coro
, 3);
3351 patch_vertices_in
= LLVMGetParam(variant_coro
, 4);
3353 consts_ptr
= draw_tcs_jit_context_constants(variant
->gallivm
, context_ptr
);
3355 draw_tcs_jit_context_num_constants(variant
->gallivm
, context_ptr
);
3357 ssbos_ptr
= draw_tcs_jit_context_ssbos(variant
->gallivm
, context_ptr
);
3359 draw_tcs_jit_context_num_ssbos(variant
->gallivm
, context_ptr
);
3360 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
);
3361 image
= draw_llvm_image_soa_create(draw_tcs_llvm_variant_key_images(&variant
->key
));
3363 LLVMValueRef counter
= LLVMGetParam(variant_coro
, 5);
3364 LLVMValueRef invocvec
= LLVMGetUndef(LLVMVectorType(int32_type
, vector_length
));
3365 for (i
= 0; i
< vector_length
; i
++) {
3366 LLVMValueRef idx
= LLVMBuildAdd(builder
, LLVMBuildMul(builder
, counter
, step
, ""), lp_build_const_int32(gallivm
, i
), "");
3367 invocvec
= LLVMBuildInsertElement(builder
, invocvec
, idx
, idx
, "");
3370 system_values
.invocation_id
= invocvec
;
3371 system_values
.prim_id
= lp_build_broadcast_scalar(&bldvec
, prim_id
);
3372 system_values
.vertices_in
= lp_build_broadcast_scalar(&bldvec
, patch_vertices_in
);
3373 tcs_iface
.input
= input_array
;
3374 tcs_iface
.output
= output_array
;
3375 tcs_iface
.base
.emit_fetch_input
= draw_tcs_llvm_emit_fetch_input
;
3376 tcs_iface
.base
.emit_fetch_output
= draw_tcs_llvm_emit_fetch_output
;
3377 tcs_iface
.base
.emit_store_output
= draw_tcs_llvm_emit_store_output
;
3381 LLVMValueRef coro_id
= lp_build_coro_id(gallivm
);
3382 LLVMValueRef coro_hdl
= lp_build_coro_begin_alloc_mem(gallivm
, coro_id
);
3384 mask_val
= generate_tcs_mask_value(variant
, tcs_type
, count
, LLVMBuildMul(builder
, counter
, step
, ""));
3385 lp_build_mask_begin(&mask
, gallivm
, tcs_type
, mask_val
);
3387 struct lp_build_coro_suspend_info coro_info
;
3389 LLVMBasicBlockRef sus_block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_coro
, "suspend");
3390 LLVMBasicBlockRef clean_block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_coro
, "cleanup");
3392 coro_info
.suspend
= sus_block
;
3393 coro_info
.cleanup
= clean_block
;
3395 struct lp_build_tgsi_params params
;
3396 memset(¶ms
, 0, sizeof(params
));
3398 params
.type
= tcs_type
;
3399 params
.mask
= &mask
;
3400 params
.consts_ptr
= consts_ptr
;
3401 params
.const_sizes_ptr
= num_consts_ptr
;
3402 params
.system_values
= &system_values
;
3403 params
.context_ptr
= context_ptr
;
3404 params
.sampler
= sampler
;
3405 params
.info
= &llvm
->draw
->tcs
.tess_ctrl_shader
->info
;
3406 params
.ssbo_ptr
= ssbos_ptr
;
3407 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
3408 params
.image
= image
;
3409 params
.coro
= &coro_info
;
3410 params
.tcs_iface
= &tcs_iface
.base
;
3412 lp_build_nir_soa(variant
->gallivm
,
3413 llvm
->draw
->tcs
.tess_ctrl_shader
->state
.ir
.nir
,
3416 lp_build_mask_end(&mask
);
3418 lp_build_coro_suspend_switch(gallivm
, &coro_info
, NULL
, true);
3419 LLVMPositionBuilderAtEnd(builder
, clean_block
);
3421 lp_build_coro_free_mem(gallivm
, coro_id
, coro_hdl
);
3423 LLVMBuildBr(builder
, sus_block
);
3424 LLVMPositionBuilderAtEnd(builder
, sus_block
);
3426 lp_build_coro_end(gallivm
, coro_hdl
);
3427 LLVMBuildRet(builder
, coro_hdl
);
3430 sampler
->destroy(sampler
);
3431 image
->destroy(image
);
3432 gallivm_verify_function(gallivm
, variant_func
);
3433 gallivm_verify_function(gallivm
, variant_coro
);
3436 struct draw_tcs_llvm_variant
*
3437 draw_tcs_llvm_create_variant(struct draw_llvm
*llvm
,
3438 unsigned num_outputs
,
3439 const struct draw_tcs_llvm_variant_key
*key
)
3441 struct draw_tcs_llvm_variant
*variant
;
3442 struct llvm_tess_ctrl_shader
*shader
= llvm_tess_ctrl_shader(llvm
->draw
->tcs
.tess_ctrl_shader
);
3443 char module_name
[64];
3445 variant
= MALLOC(sizeof *variant
+
3446 shader
->variant_key_size
- sizeof variant
->key
);
3450 variant
->llvm
= llvm
;
3451 variant
->shader
= shader
;
3453 snprintf(module_name
, sizeof(module_name
), "draw_llvm_tcs_variant%u",
3454 variant
->shader
->variants_cached
);
3456 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, NULL
);
3458 create_tcs_jit_types(variant
);
3460 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3462 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
3463 nir_print_shader(llvm
->draw
->tcs
.tess_ctrl_shader
->state
.ir
.nir
, stderr
);
3464 draw_tcs_llvm_dump_variant_key(&variant
->key
);
3467 lp_build_coro_declare_malloc_hooks(variant
->gallivm
);
3468 draw_tcs_llvm_generate(llvm
, variant
);
3470 gallivm_compile_module(variant
->gallivm
);
3472 lp_build_coro_add_malloc_hooks(variant
->gallivm
);
3473 variant
->jit_func
= (draw_tcs_jit_func
)
3474 gallivm_jit_function(variant
->gallivm
, variant
->function
);
3476 gallivm_free_ir(variant
->gallivm
);
3478 variant
->list_item_global
.base
= variant
;
3479 variant
->list_item_local
.base
= variant
;
3480 /*variant->no = */shader
->variants_created
++;
3481 variant
->list_item_global
.base
= variant
;
3487 draw_tcs_llvm_destroy_variant(struct draw_tcs_llvm_variant
*variant
)
3489 struct draw_llvm
*llvm
= variant
->llvm
;
3491 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
3492 debug_printf("Deleting TCS variant: %u tcs variants,\t%u total variants\n",
3493 variant
->shader
->variants_cached
, llvm
->nr_tcs_variants
);
3496 gallivm_destroy(variant
->gallivm
);
3498 remove_from_list(&variant
->list_item_local
);
3499 variant
->shader
->variants_cached
--;
3500 remove_from_list(&variant
->list_item_global
);
3501 llvm
->nr_tcs_variants
--;
3505 struct draw_tcs_llvm_variant_key
*
3506 draw_tcs_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
3509 struct draw_tcs_llvm_variant_key
*key
;
3510 struct draw_sampler_static_state
*draw_sampler
;
3511 struct draw_image_static_state
*draw_image
;
3513 key
= (struct draw_tcs_llvm_variant_key
*)store
;
3515 memset(key
, 0, offsetof(struct draw_tcs_llvm_variant_key
, samplers
[0]));
3517 /* All variants of this shader will have the same value for
3518 * nr_samplers. Not yet trying to compact away holes in the
3521 key
->nr_samplers
= llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3522 if (llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
3523 key
->nr_sampler_views
=
3524 llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
3527 key
->nr_sampler_views
= key
->nr_samplers
;
3530 key
->nr_images
= llvm
->draw
->tcs
.tess_ctrl_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
3532 draw_sampler
= key
->samplers
;
3534 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
3536 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
3537 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
3538 llvm
->draw
->samplers
[PIPE_SHADER_TESS_CTRL
][i
]);
3540 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
3541 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
3542 llvm
->draw
->sampler_views
[PIPE_SHADER_TESS_CTRL
][i
]);
3545 draw_image
= draw_tcs_llvm_variant_key_images(key
);
3546 memset(draw_image
, 0,
3547 key
->nr_images
* sizeof *draw_image
);
3548 for (i
= 0; i
< key
->nr_images
; i
++) {
3549 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
3550 llvm
->draw
->images
[PIPE_SHADER_TESS_CTRL
][i
]);
3556 draw_tcs_llvm_dump_variant_key(struct draw_tcs_llvm_variant_key
*key
)
3559 struct draw_sampler_static_state
*sampler
= key
->samplers
;
3560 struct draw_image_static_state
*image
= draw_tcs_llvm_variant_key_images(key
);
3561 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
3562 debug_printf("sampler[%i].src_format = %s\n", i
,
3563 util_format_name(sampler
[i
].texture_state
.format
));
3566 for (i
= 0 ; i
< key
->nr_images
; i
++)
3567 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));
3572 create_tes_jit_types(struct draw_tes_llvm_variant
*var
)
3574 struct gallivm_state
*gallivm
= var
->gallivm
;
3575 LLVMTypeRef texture_type
, sampler_type
, image_type
, context_type
;
3577 texture_type
= create_jit_texture_type(gallivm
, "texture");
3578 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
3579 image_type
= create_jit_image_type(gallivm
, "image");
3581 context_type
= create_tes_jit_context_type(gallivm
,
3583 texture_type
, sampler_type
,
3585 "draw_tes_jit_context");
3586 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
3588 var
->input_array_type
= create_tes_jit_input_type(gallivm
);
3592 get_tes_context_ptr_type(struct draw_tes_llvm_variant
*variant
)
3594 if (!variant
->context_ptr_type
)
3595 create_tes_jit_types(variant
);
3596 return variant
->context_ptr_type
;
3600 generate_tes_mask_value(struct draw_tes_llvm_variant
*variant
,
3601 struct lp_type tes_type
, LLVMValueRef limit
, LLVMValueRef loop_counter
)
3603 struct gallivm_state
*gallivm
= variant
->gallivm
;
3604 LLVMBuilderRef builder
= gallivm
->builder
;
3605 struct lp_type mask_type
= lp_int_type(tes_type
);
3606 LLVMValueRef num_prims
;
3607 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
3610 num_prims
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
), limit
);
3611 for (i
= 0; i
< tes_type
.length
; i
++) {
3612 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3613 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, LLVMBuildAdd(builder
, loop_counter
, idx
, ""), idx
, "");
3615 mask_val
= lp_build_compare(gallivm
, mask_type
,
3616 PIPE_FUNC_GREATER
, num_prims
, mask_val
);
3622 draw_tes_llvm_fetch_vertex_input(const struct lp_build_tes_iface
*tes_iface
,
3623 struct lp_build_context
*bld
,
3624 boolean is_vindex_indirect
,
3625 LLVMValueRef vertex_index
,
3626 boolean is_aindex_indirect
,
3627 LLVMValueRef attrib_index
,
3628 LLVMValueRef swizzle_index
)
3630 const struct draw_tes_llvm_iface
*tes
= draw_tes_llvm_iface(tes_iface
);
3631 struct gallivm_state
*gallivm
= bld
->gallivm
;
3632 LLVMBuilderRef builder
= gallivm
->builder
;
3633 LLVMValueRef indices
[3];
3635 struct lp_type type
= bld
->type
;
3637 if (is_vindex_indirect
|| is_aindex_indirect
) {
3642 for (i
= 0; i
< type
.length
; ++i
) {
3643 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3644 LLVMValueRef vert_chan_index
= vertex_index
;
3645 LLVMValueRef attr_chan_index
= attrib_index
;
3646 LLVMValueRef channel_vec
;
3648 if (is_vindex_indirect
) {
3649 vert_chan_index
= LLVMBuildExtractElement(builder
,
3650 vertex_index
, idx
, "");
3652 if (is_aindex_indirect
) {
3653 attr_chan_index
= LLVMBuildExtractElement(builder
,
3654 attrib_index
, idx
, "");
3657 indices
[0] = vert_chan_index
;
3658 indices
[1] = attr_chan_index
;
3659 indices
[2] = swizzle_index
;
3661 channel_vec
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3662 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3664 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3667 indices
[0] = vertex_index
;
3668 indices
[1] = attrib_index
;
3669 indices
[2] = swizzle_index
;
3671 res
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3672 res
= LLVMBuildLoad(builder
, res
, "");
3673 res
= lp_build_broadcast_scalar(bld
, res
);
3679 draw_tes_llvm_fetch_patch_input(const struct lp_build_tes_iface
*tes_iface
,
3680 struct lp_build_context
*bld
,
3681 boolean is_aindex_indirect
,
3682 LLVMValueRef attrib_index
,
3683 LLVMValueRef swizzle_index
)
3685 const struct draw_tes_llvm_iface
*tes
= draw_tes_llvm_iface(tes_iface
);
3686 struct gallivm_state
*gallivm
= bld
->gallivm
;
3687 LLVMBuilderRef builder
= gallivm
->builder
;
3688 LLVMValueRef indices
[3];
3690 struct lp_type type
= bld
->type
;
3692 if (is_aindex_indirect
) {
3697 for (i
= 0; i
< type
.length
; ++i
) {
3698 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
3699 LLVMValueRef attr_chan_index
= attrib_index
;
3700 LLVMValueRef channel_vec
;
3702 if (is_aindex_indirect
) {
3703 attr_chan_index
= LLVMBuildExtractElement(builder
,
3704 attrib_index
, idx
, "");
3707 indices
[0] = lp_build_const_int32(gallivm
, 0);
3708 indices
[1] = attr_chan_index
;
3709 indices
[2] = swizzle_index
;
3711 channel_vec
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3712 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
3714 res
= LLVMBuildInsertElement(builder
, res
, channel_vec
, idx
, "");
3717 indices
[0] = lp_build_const_int32(gallivm
, 0);
3718 indices
[1] = attrib_index
;
3719 indices
[2] = swizzle_index
;
3721 res
= LLVMBuildGEP(builder
, tes
->input
, indices
, 3, "");
3722 res
= LLVMBuildLoad(builder
, res
, "");
3723 res
= lp_build_broadcast_scalar(bld
, res
);
3729 draw_tes_llvm_generate(struct draw_llvm
*llvm
,
3730 struct draw_tes_llvm_variant
*variant
)
3732 struct gallivm_state
*gallivm
= variant
->gallivm
;
3733 LLVMContextRef context
= gallivm
->context
;
3734 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
3735 LLVMTypeRef flt_type
= LLVMFloatTypeInContext(context
);
3736 LLVMTypeRef arg_types
[10];
3737 LLVMTypeRef func_type
;
3738 LLVMValueRef variant_func
;
3739 LLVMValueRef context_ptr
;
3740 LLVMValueRef tess_coord
[2], io_ptr
, input_array
, num_tess_coord
;
3741 LLVMValueRef tess_inner
, tess_outer
, prim_id
, patch_vertices_in
;
3742 LLVMBasicBlockRef block
;
3743 LLVMBuilderRef builder
;
3744 LLVMValueRef mask_val
;
3745 struct lp_build_context bld
, bldvec
;
3746 struct lp_build_sampler_soa
*sampler
= 0;
3747 struct lp_build_image_soa
*image
= NULL
;
3748 struct lp_bld_tgsi_system_values system_values
;
3751 struct draw_tes_llvm_iface tes_iface
;
3752 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
3753 struct lp_build_mask_context mask
;
3754 LLVMValueRef consts_ptr
, num_consts_ptr
;
3755 LLVMValueRef ssbos_ptr
, num_ssbos_ptr
;
3757 struct lp_type tes_type
;
3758 unsigned vector_length
= variant
->shader
->base
.vector_length
;
3760 memset(&system_values
, 0, sizeof(system_values
));
3761 memset(&outputs
, 0, sizeof(outputs
));
3763 snprintf(func_name
, sizeof(func_name
), "draw_llvm_tes_variant");
3765 arg_types
[0] = get_tes_context_ptr_type(variant
); /* context */
3766 arg_types
[1] = variant
->input_array_type
; /* input */
3767 arg_types
[2] = variant
->vertex_header_ptr_type
;
3768 arg_types
[3] = int32_type
;
3769 arg_types
[4] = int32_type
;
3770 arg_types
[5] = LLVMPointerType(flt_type
, 0);
3771 arg_types
[6] = LLVMPointerType(flt_type
, 0);
3772 arg_types
[7] = LLVMPointerType(LLVMArrayType(flt_type
, 4), 0);
3773 arg_types
[8] = LLVMPointerType(LLVMArrayType(flt_type
, 2), 0);
3774 arg_types
[9] = int32_type
;
3776 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
), 0);
3777 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
3779 variant
->function
= variant_func
;
3780 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
3782 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
3783 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
3784 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
3786 context_ptr
= LLVMGetParam(variant_func
, 0);
3787 input_array
= LLVMGetParam(variant_func
, 1);
3788 io_ptr
= LLVMGetParam(variant_func
, 2);
3789 prim_id
= LLVMGetParam(variant_func
, 3);
3790 num_tess_coord
= LLVMGetParam(variant_func
, 4);
3791 tess_coord
[0] = LLVMGetParam(variant_func
, 5);
3792 tess_coord
[1] = LLVMGetParam(variant_func
, 6);
3793 tess_outer
= LLVMGetParam(variant_func
, 7);
3794 tess_inner
= LLVMGetParam(variant_func
, 8);
3795 patch_vertices_in
= LLVMGetParam(variant_func
, 9);
3797 lp_build_name(context_ptr
, "context");
3798 lp_build_name(input_array
, "input");
3799 lp_build_name(io_ptr
, "io");
3800 lp_build_name(prim_id
, "prim_id");
3801 lp_build_name(num_tess_coord
, "num_tess_coord");
3802 lp_build_name(tess_coord
[0], "tess_coord[0]");
3803 lp_build_name(tess_coord
[1], "tess_coord[1]");
3804 lp_build_name(tess_outer
, "tess_outer");
3805 lp_build_name(tess_inner
, "tess_inner");
3806 lp_build_name(patch_vertices_in
, "patch_vertices_in");
3808 tes_iface
.base
.fetch_vertex_input
= draw_tes_llvm_fetch_vertex_input
;
3809 tes_iface
.base
.fetch_patch_input
= draw_tes_llvm_fetch_patch_input
;
3810 tes_iface
.input
= input_array
;
3811 tes_iface
.variant
= variant
;
3813 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
3814 builder
= gallivm
->builder
;
3815 LLVMPositionBuilderAtEnd(builder
, block
);
3817 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
3819 memset(&tes_type
, 0, sizeof tes_type
);
3820 tes_type
.floating
= TRUE
; /* floating point values */
3821 tes_type
.sign
= TRUE
; /* values are signed */
3822 tes_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
3823 tes_type
.width
= 32; /* 32-bit float */
3824 tes_type
.length
= vector_length
;
3826 lp_build_context_init(&bldvec
, variant
->gallivm
, lp_int_type(tes_type
));
3827 consts_ptr
= draw_tes_jit_context_constants(variant
->gallivm
, context_ptr
);
3829 draw_tes_jit_context_num_constants(variant
->gallivm
, context_ptr
);
3831 ssbos_ptr
= draw_tes_jit_context_ssbos(variant
->gallivm
, context_ptr
);
3833 draw_tes_jit_context_num_ssbos(variant
->gallivm
, context_ptr
);
3834 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
);
3835 image
= draw_llvm_image_soa_create(draw_tes_llvm_variant_key_images(&variant
->key
));
3836 step
= lp_build_const_int32(gallivm
, vector_length
);
3838 system_values
.tess_outer
= LLVMBuildLoad(builder
, tess_outer
, "");
3839 system_values
.tess_inner
= LLVMBuildLoad(builder
, tess_inner
, "");
3841 system_values
.prim_id
= lp_build_broadcast_scalar(&bldvec
, prim_id
);
3843 system_values
.vertices_in
= lp_build_broadcast_scalar(&bldvec
, patch_vertices_in
);
3844 struct lp_build_loop_state lp_loop
;
3845 lp_build_loop_begin(&lp_loop
, gallivm
, bld
.zero
);
3849 io
= LLVMBuildGEP(builder
, io_ptr
, &lp_loop
.counter
, 1, "");
3850 mask_val
= generate_tes_mask_value(variant
, tes_type
, num_tess_coord
, lp_loop
.counter
);
3851 lp_build_mask_begin(&mask
, gallivm
, tes_type
, mask_val
);
3853 system_values
.tess_coord
= LLVMGetUndef(LLVMArrayType(LLVMVectorType(flt_type
, vector_length
), 3));
3854 for (i
= 0; i
< 3; i
++) {
3855 LLVMValueRef tess_coord_chan
= LLVMGetUndef(LLVMVectorType(flt_type
, vector_length
));
3856 for (unsigned j
= 0; j
< vector_length
; j
++) {
3857 LLVMValueRef idx
= LLVMBuildAdd(builder
, lp_loop
.counter
, lp_build_const_int32(gallivm
, j
), "");
3858 LLVMValueRef tc_val
;
3860 if (variant
->shader
->base
.prim_mode
== PIPE_PRIM_TRIANGLES
) {
3861 tc_val
= lp_build_const_float(gallivm
, 1.0);
3862 tc_val
= LLVMBuildFSub(builder
, tc_val
, lp_build_pointer_get(builder
, tess_coord
[0], idx
), "");
3863 tc_val
= LLVMBuildFSub(builder
, tc_val
, lp_build_pointer_get(builder
, tess_coord
[1], idx
), "");
3865 tc_val
= lp_build_const_float(gallivm
, 0.0);
3867 tc_val
= lp_build_pointer_get(builder
, tess_coord
[i
], idx
);
3869 tess_coord_chan
= LLVMBuildInsertElement(builder
, tess_coord_chan
, tc_val
, lp_build_const_int32(gallivm
, j
), "");
3871 system_values
.tess_coord
= LLVMBuildInsertValue(builder
, system_values
.tess_coord
, tess_coord_chan
, i
, "");
3874 struct lp_build_tgsi_params params
;
3875 memset(¶ms
, 0, sizeof(params
));
3877 params
.type
= tes_type
;
3878 params
.mask
= &mask
;
3879 params
.consts_ptr
= consts_ptr
;
3880 params
.const_sizes_ptr
= num_consts_ptr
;
3881 params
.system_values
= &system_values
;
3882 params
.context_ptr
= context_ptr
;
3883 params
.sampler
= sampler
;
3884 params
.info
= &llvm
->draw
->tes
.tess_eval_shader
->info
;
3885 params
.ssbo_ptr
= ssbos_ptr
;
3886 params
.ssbo_sizes_ptr
= num_ssbos_ptr
;
3887 params
.image
= image
;
3888 params
.tes_iface
= &tes_iface
.base
;
3890 lp_build_nir_soa(variant
->gallivm
,
3891 llvm
->draw
->tes
.tess_eval_shader
->state
.ir
.nir
,
3895 lp_build_mask_end(&mask
);
3896 LLVMValueRef clipmask
= lp_build_const_int_vec(gallivm
,
3897 lp_int_type(tes_type
), 0);
3899 convert_to_aos(gallivm
, io
, NULL
, outputs
, clipmask
,
3900 params
.info
->num_outputs
, tes_type
, FALSE
);
3902 lp_build_loop_end_cond(&lp_loop
, num_tess_coord
, step
, LLVMIntUGE
);
3903 sampler
->destroy(sampler
);
3904 image
->destroy(image
);
3906 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
3907 gallivm_verify_function(gallivm
, variant_func
);
3910 struct draw_tes_llvm_variant
*
3911 draw_tes_llvm_create_variant(struct draw_llvm
*llvm
,
3912 unsigned num_outputs
,
3913 const struct draw_tes_llvm_variant_key
*key
)
3915 struct draw_tes_llvm_variant
*variant
;
3916 struct llvm_tess_eval_shader
*shader
= llvm_tess_eval_shader(llvm
->draw
->tes
.tess_eval_shader
);
3917 LLVMTypeRef vertex_header
;
3918 char module_name
[64];
3920 variant
= MALLOC(sizeof *variant
+
3921 shader
->variant_key_size
- sizeof variant
->key
);
3925 variant
->llvm
= llvm
;
3926 variant
->shader
= shader
;
3928 snprintf(module_name
, sizeof(module_name
), "draw_llvm_tes_variant%u",
3929 variant
->shader
->variants_cached
);
3931 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
, NULL
);
3933 create_tes_jit_types(variant
);
3935 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
3937 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_outputs
);
3939 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
3941 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
3942 nir_print_shader(llvm
->draw
->tes
.tess_eval_shader
->state
.ir
.nir
, stderr
);
3943 draw_tes_llvm_dump_variant_key(&variant
->key
);
3946 draw_tes_llvm_generate(llvm
, variant
);
3948 gallivm_compile_module(variant
->gallivm
);
3950 variant
->jit_func
= (draw_tes_jit_func
)
3951 gallivm_jit_function(variant
->gallivm
, variant
->function
);
3953 gallivm_free_ir(variant
->gallivm
);
3955 variant
->list_item_global
.base
= variant
;
3956 variant
->list_item_local
.base
= variant
;
3957 /*variant->no = */shader
->variants_created
++;
3958 variant
->list_item_global
.base
= variant
;
3964 draw_tes_llvm_destroy_variant(struct draw_tes_llvm_variant
*variant
)
3966 struct draw_llvm
*llvm
= variant
->llvm
;
3968 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
3969 debug_printf("Deleting TES variant: %u tes variants,\t%u total variants\n",
3970 variant
->shader
->variants_cached
, llvm
->nr_tes_variants
);
3973 gallivm_destroy(variant
->gallivm
);
3975 remove_from_list(&variant
->list_item_local
);
3976 variant
->shader
->variants_cached
--;
3977 remove_from_list(&variant
->list_item_global
);
3978 llvm
->nr_tes_variants
--;
3982 struct draw_tes_llvm_variant_key
*
3983 draw_tes_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
3986 struct draw_tes_llvm_variant_key
*key
;
3987 struct draw_sampler_static_state
*draw_sampler
;
3988 struct draw_image_static_state
*draw_image
;
3990 key
= (struct draw_tes_llvm_variant_key
*)store
;
3992 memset(key
, 0, offsetof(struct draw_tes_llvm_variant_key
, samplers
[0]));
3994 /* All variants of this shader will have the same value for
3995 * nr_samplers. Not yet trying to compact away holes in the
3998 key
->nr_samplers
= llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
3999 if (llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
4000 key
->nr_sampler_views
=
4001 llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
4004 key
->nr_sampler_views
= key
->nr_samplers
;
4007 key
->nr_images
= llvm
->draw
->tes
.tess_eval_shader
->info
.file_max
[TGSI_FILE_IMAGE
] + 1;
4009 draw_sampler
= key
->samplers
;
4011 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
4013 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
4014 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
4015 llvm
->draw
->samplers
[PIPE_SHADER_TESS_EVAL
][i
]);
4017 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
4018 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
4019 llvm
->draw
->sampler_views
[PIPE_SHADER_TESS_EVAL
][i
]);
4022 draw_image
= draw_tes_llvm_variant_key_images(key
);
4023 memset(draw_image
, 0,
4024 key
->nr_images
* sizeof *draw_image
);
4025 for (i
= 0; i
< key
->nr_images
; i
++) {
4026 lp_sampler_static_texture_state_image(&draw_image
[i
].image_state
,
4027 llvm
->draw
->images
[PIPE_SHADER_TESS_EVAL
][i
]);
4033 draw_tes_llvm_dump_variant_key(struct draw_tes_llvm_variant_key
*key
)
4036 struct draw_sampler_static_state
*sampler
= key
->samplers
;
4037 struct draw_image_static_state
*image
= draw_tes_llvm_variant_key_images(key
);
4038 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
4039 debug_printf("sampler[%i].src_format = %s\n", i
,
4040 util_format_name(sampler
[i
].texture_state
.format
));
4043 for (i
= 0 ; i
< key
->nr_images
; i
++)
4044 debug_printf("images[%i].format = %s\n", i
, util_format_name(image
[i
].image_state
.format
));