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_swizzle.h"
41 #include "gallivm/lp_bld_struct.h"
42 #include "gallivm/lp_bld_type.h"
43 #include "gallivm/lp_bld_flow.h"
44 #include "gallivm/lp_bld_debug.h"
45 #include "gallivm/lp_bld_tgsi.h"
46 #include "gallivm/lp_bld_printf.h"
47 #include "gallivm/lp_bld_intr.h"
48 #include "gallivm/lp_bld_init.h"
49 #include "gallivm/lp_bld_type.h"
50 #include "gallivm/lp_bld_pack.h"
51 #include "gallivm/lp_bld_format.h"
53 #include "tgsi/tgsi_exec.h"
54 #include "tgsi/tgsi_dump.h"
56 #include "util/u_math.h"
57 #include "util/u_pointer.h"
58 #include "util/u_string.h"
59 #include "util/simple_list.h"
66 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*var
,
70 struct draw_gs_llvm_iface
{
71 struct lp_build_tgsi_gs_iface base
;
73 struct draw_gs_llvm_variant
*variant
;
77 static inline const struct draw_gs_llvm_iface
*
78 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface
*iface
)
80 return (const struct draw_gs_llvm_iface
*)iface
;
84 * Create LLVM type for draw_vertex_buffer.
87 create_jit_dvbuffer_type(struct gallivm_state
*gallivm
,
88 const char *struct_name
)
90 LLVMTargetDataRef target
= gallivm
->target
;
91 LLVMTypeRef dvbuffer_type
;
92 LLVMTypeRef elem_types
[DRAW_JIT_DVBUFFER_NUM_FIELDS
];
93 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
95 elem_types
[DRAW_JIT_DVBUFFER_MAP
] =
96 LLVMPointerType(LLVMIntTypeInContext(gallivm
->context
, 8), 0);
97 elem_types
[DRAW_JIT_DVBUFFER_SIZE
] = int32_type
;
99 dvbuffer_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
100 ARRAY_SIZE(elem_types
), 0);
102 (void) target
; /* silence unused var warning for non-debug build */
103 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer
, map
,
104 target
, dvbuffer_type
,
105 DRAW_JIT_DVBUFFER_MAP
);
106 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer
, size
,
107 target
, dvbuffer_type
,
108 DRAW_JIT_DVBUFFER_SIZE
);
110 return dvbuffer_type
;
114 * Create LLVM type for struct draw_jit_texture
117 create_jit_texture_type(struct gallivm_state
*gallivm
, const char *struct_name
)
119 LLVMTargetDataRef target
= gallivm
->target
;
120 LLVMTypeRef texture_type
;
121 LLVMTypeRef elem_types
[DRAW_JIT_TEXTURE_NUM_FIELDS
];
122 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
124 elem_types
[DRAW_JIT_TEXTURE_WIDTH
] =
125 elem_types
[DRAW_JIT_TEXTURE_HEIGHT
] =
126 elem_types
[DRAW_JIT_TEXTURE_DEPTH
] =
127 elem_types
[DRAW_JIT_TEXTURE_FIRST_LEVEL
] =
128 elem_types
[DRAW_JIT_TEXTURE_LAST_LEVEL
] = int32_type
;
129 elem_types
[DRAW_JIT_TEXTURE_BASE
] =
130 LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
131 elem_types
[DRAW_JIT_TEXTURE_ROW_STRIDE
] =
132 elem_types
[DRAW_JIT_TEXTURE_IMG_STRIDE
] =
133 elem_types
[DRAW_JIT_TEXTURE_MIP_OFFSETS
] =
134 LLVMArrayType(int32_type
, PIPE_MAX_TEXTURE_LEVELS
);
136 texture_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
137 ARRAY_SIZE(elem_types
), 0);
139 (void) target
; /* silence unused var warning for non-debug build */
140 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, width
,
141 target
, texture_type
,
142 DRAW_JIT_TEXTURE_WIDTH
);
143 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, height
,
144 target
, texture_type
,
145 DRAW_JIT_TEXTURE_HEIGHT
);
146 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, depth
,
147 target
, texture_type
,
148 DRAW_JIT_TEXTURE_DEPTH
);
149 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, first_level
,
150 target
, texture_type
,
151 DRAW_JIT_TEXTURE_FIRST_LEVEL
);
152 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, last_level
,
153 target
, texture_type
,
154 DRAW_JIT_TEXTURE_LAST_LEVEL
);
155 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, base
,
156 target
, texture_type
,
157 DRAW_JIT_TEXTURE_BASE
);
158 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, row_stride
,
159 target
, texture_type
,
160 DRAW_JIT_TEXTURE_ROW_STRIDE
);
161 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, img_stride
,
162 target
, texture_type
,
163 DRAW_JIT_TEXTURE_IMG_STRIDE
);
164 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture
, mip_offsets
,
165 target
, texture_type
,
166 DRAW_JIT_TEXTURE_MIP_OFFSETS
);
168 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture
, target
, texture_type
);
175 * Create LLVM type for struct draw_jit_sampler
178 create_jit_sampler_type(struct gallivm_state
*gallivm
, const char *struct_name
)
180 LLVMTargetDataRef target
= gallivm
->target
;
181 LLVMTypeRef sampler_type
;
182 LLVMTypeRef elem_types
[DRAW_JIT_SAMPLER_NUM_FIELDS
];
184 elem_types
[DRAW_JIT_SAMPLER_MIN_LOD
] =
185 elem_types
[DRAW_JIT_SAMPLER_MAX_LOD
] =
186 elem_types
[DRAW_JIT_SAMPLER_LOD_BIAS
] = LLVMFloatTypeInContext(gallivm
->context
);
187 elem_types
[DRAW_JIT_SAMPLER_BORDER_COLOR
] =
188 LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
190 sampler_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
191 ARRAY_SIZE(elem_types
), 0);
193 (void) target
; /* silence unused var warning for non-debug build */
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, min_lod
,
195 target
, sampler_type
,
196 DRAW_JIT_SAMPLER_MIN_LOD
);
197 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, max_lod
,
198 target
, sampler_type
,
199 DRAW_JIT_SAMPLER_MAX_LOD
);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, lod_bias
,
201 target
, sampler_type
,
202 DRAW_JIT_SAMPLER_LOD_BIAS
);
203 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler
, border_color
,
204 target
, sampler_type
,
205 DRAW_JIT_SAMPLER_BORDER_COLOR
);
207 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler
, target
, sampler_type
);
214 * Create LLVM type for struct draw_jit_context
217 create_jit_context_type(struct gallivm_state
*gallivm
,
218 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
219 const char *struct_name
)
221 LLVMTargetDataRef target
= gallivm
->target
;
222 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
223 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
224 LLVMTypeRef elem_types
[DRAW_JIT_CTX_NUM_FIELDS
];
225 LLVMTypeRef context_type
;
227 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* vs_constants */
228 LP_MAX_TGSI_CONST_BUFFERS
);
229 elem_types
[1] = LLVMArrayType(int_type
, /* num_vs_constants */
230 LP_MAX_TGSI_CONST_BUFFERS
);
231 elem_types
[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
232 DRAW_TOTAL_CLIP_PLANES
), 0);
233 elem_types
[3] = LLVMPointerType(float_type
, 0); /* viewports */
234 elem_types
[4] = LLVMArrayType(texture_type
,
235 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
236 elem_types
[5] = LLVMArrayType(sampler_type
,
237 PIPE_MAX_SAMPLERS
); /* samplers */
238 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
239 ARRAY_SIZE(elem_types
), 0);
241 (void) target
; /* silence unused var warning for non-debug build */
242 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, vs_constants
,
243 target
, context_type
, DRAW_JIT_CTX_CONSTANTS
);
244 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, num_vs_constants
,
245 target
, context_type
, DRAW_JIT_CTX_NUM_CONSTANTS
);
246 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, planes
,
247 target
, context_type
, DRAW_JIT_CTX_PLANES
);
248 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, viewports
,
249 target
, context_type
, DRAW_JIT_CTX_VIEWPORT
);
250 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, textures
,
251 target
, context_type
,
252 DRAW_JIT_CTX_TEXTURES
);
253 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context
, samplers
,
254 target
, context_type
,
255 DRAW_JIT_CTX_SAMPLERS
);
256 LP_CHECK_STRUCT_SIZE(struct draw_jit_context
,
257 target
, context_type
);
264 * Create LLVM type for struct draw_gs_jit_context
267 create_gs_jit_context_type(struct gallivm_state
*gallivm
,
268 unsigned vector_length
,
269 LLVMTypeRef texture_type
, LLVMTypeRef sampler_type
,
270 const char *struct_name
)
272 LLVMTargetDataRef target
= gallivm
->target
;
273 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
274 LLVMTypeRef int_type
= LLVMInt32TypeInContext(gallivm
->context
);
275 LLVMTypeRef elem_types
[DRAW_GS_JIT_CTX_NUM_FIELDS
];
276 LLVMTypeRef context_type
;
278 elem_types
[0] = LLVMArrayType(LLVMPointerType(float_type
, 0), /* constants */
279 LP_MAX_TGSI_CONST_BUFFERS
);
280 elem_types
[1] = LLVMArrayType(int_type
, /* num_constants */
281 LP_MAX_TGSI_CONST_BUFFERS
);
282 elem_types
[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type
, 4),
283 DRAW_TOTAL_CLIP_PLANES
), 0);
284 elem_types
[3] = LLVMPointerType(float_type
, 0); /* viewports */
286 elem_types
[4] = LLVMArrayType(texture_type
,
287 PIPE_MAX_SHADER_SAMPLER_VIEWS
); /* textures */
288 elem_types
[5] = LLVMArrayType(sampler_type
,
289 PIPE_MAX_SAMPLERS
); /* samplers */
291 elem_types
[6] = LLVMPointerType(LLVMPointerType(int_type
, 0), 0);
292 elem_types
[7] = LLVMPointerType(LLVMVectorType(int_type
,
294 elem_types
[8] = LLVMPointerType(LLVMVectorType(int_type
,
297 context_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
298 ARRAY_SIZE(elem_types
), 0);
300 (void) target
; /* silence unused var warning for non-debug build */
301 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, constants
,
302 target
, context_type
, DRAW_GS_JIT_CTX_CONSTANTS
);
303 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, num_constants
,
304 target
, context_type
, DRAW_GS_JIT_CTX_NUM_CONSTANTS
);
305 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, planes
,
306 target
, context_type
, DRAW_GS_JIT_CTX_PLANES
);
307 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, viewports
,
308 target
, context_type
, DRAW_GS_JIT_CTX_VIEWPORT
);
309 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, textures
,
310 target
, context_type
,
311 DRAW_GS_JIT_CTX_TEXTURES
);
312 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, samplers
,
313 target
, context_type
,
314 DRAW_GS_JIT_CTX_SAMPLERS
);
315 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, prim_lengths
,
316 target
, context_type
,
317 DRAW_GS_JIT_CTX_PRIM_LENGTHS
);
318 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, emitted_vertices
,
319 target
, context_type
,
320 DRAW_GS_JIT_CTX_EMITTED_VERTICES
);
321 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context
, emitted_prims
,
322 target
, context_type
,
323 DRAW_GS_JIT_CTX_EMITTED_PRIMS
);
324 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context
,
325 target
, context_type
);
332 create_gs_jit_input_type(struct gallivm_state
*gallivm
)
334 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
335 LLVMTypeRef input_array
;
337 input_array
= LLVMVectorType(float_type
, TGSI_NUM_CHANNELS
); /* num primitives */
338 input_array
= LLVMArrayType(input_array
, TGSI_NUM_CHANNELS
); /* num channels */
339 input_array
= LLVMArrayType(input_array
, PIPE_MAX_SHADER_INPUTS
); /* num attrs per vertex */
340 input_array
= LLVMPointerType(input_array
, 0); /* num vertices per prim */
346 * Create LLVM type for struct pipe_vertex_buffer
349 create_jit_vertex_buffer_type(struct gallivm_state
*gallivm
,
350 const char *struct_name
)
352 LLVMTargetDataRef target
= gallivm
->target
;
353 LLVMTypeRef elem_types
[4];
357 elem_types
[1] = LLVMInt32TypeInContext(gallivm
->context
);
359 elem_types
[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
361 vb_type
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
362 ARRAY_SIZE(elem_types
), 0);
364 (void) target
; /* silence unused var warning for non-debug build */
365 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, stride
,
367 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer
, buffer_offset
,
370 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer
, target
, vb_type
);
377 * Create LLVM type for struct vertex_header;
380 create_jit_vertex_header(struct gallivm_state
*gallivm
, int data_elems
)
382 LLVMTargetDataRef target
= gallivm
->target
;
383 LLVMTypeRef elem_types
[3];
384 LLVMTypeRef vertex_header
;
385 char struct_name
[24];
387 util_snprintf(struct_name
, 23, "vertex_header%d", data_elems
);
389 elem_types
[DRAW_JIT_VERTEX_VERTEX_ID
] = LLVMIntTypeInContext(gallivm
->context
, 32);
390 elem_types
[DRAW_JIT_VERTEX_CLIP_POS
] = LLVMArrayType(LLVMFloatTypeInContext(gallivm
->context
), 4);
391 elem_types
[DRAW_JIT_VERTEX_DATA
] = LLVMArrayType(elem_types
[1], data_elems
);
393 vertex_header
= LLVMStructTypeInContext(gallivm
->context
, elem_types
,
394 ARRAY_SIZE(elem_types
), 0);
396 /* these are bit-fields and we can't take address of them
397 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
398 target, vertex_header,
399 DRAW_JIT_VERTEX_CLIPMASK);
400 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
401 target, vertex_header,
402 DRAW_JIT_VERTEX_EDGEFLAG);
403 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
404 target, vertex_header,
405 DRAW_JIT_VERTEX_PAD);
406 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
407 target, vertex_header,
408 DRAW_JIT_VERTEX_VERTEX_ID);
410 (void) target
; /* silence unused var warning for non-debug build */
411 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, clip_pos
,
412 target
, vertex_header
,
413 DRAW_JIT_VERTEX_CLIP_POS
);
414 LP_CHECK_MEMBER_OFFSET(struct vertex_header
, data
,
415 target
, vertex_header
,
416 DRAW_JIT_VERTEX_DATA
);
418 assert(LLVMABISizeOfType(target
, vertex_header
) ==
419 offsetof(struct vertex_header
, data
[data_elems
]));
421 return vertex_header
;
426 * Create LLVM types for various structures.
429 create_jit_types(struct draw_llvm_variant
*variant
)
431 struct gallivm_state
*gallivm
= variant
->gallivm
;
432 LLVMTypeRef texture_type
, sampler_type
, context_type
, buffer_type
,
435 texture_type
= create_jit_texture_type(gallivm
, "texture");
436 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
438 context_type
= create_jit_context_type(gallivm
, texture_type
, sampler_type
,
440 variant
->context_ptr_type
= LLVMPointerType(context_type
, 0);
442 buffer_type
= create_jit_dvbuffer_type(gallivm
, "draw_vertex_buffer");
443 variant
->buffer_ptr_type
= LLVMPointerType(buffer_type
, 0);
445 vb_type
= create_jit_vertex_buffer_type(gallivm
, "pipe_vertex_buffer");
446 variant
->vb_ptr_type
= LLVMPointerType(vb_type
, 0);
451 get_context_ptr_type(struct draw_llvm_variant
*variant
)
453 if (!variant
->context_ptr_type
)
454 create_jit_types(variant
);
455 return variant
->context_ptr_type
;
460 get_buffer_ptr_type(struct draw_llvm_variant
*variant
)
462 if (!variant
->buffer_ptr_type
)
463 create_jit_types(variant
);
464 return variant
->buffer_ptr_type
;
469 get_vb_ptr_type(struct draw_llvm_variant
*variant
)
471 if (!variant
->vb_ptr_type
)
472 create_jit_types(variant
);
473 return variant
->vb_ptr_type
;
477 get_vertex_header_ptr_type(struct draw_llvm_variant
*variant
)
479 if (!variant
->vertex_header_ptr_type
)
480 create_jit_types(variant
);
481 return variant
->vertex_header_ptr_type
;
486 * Create per-context LLVM info.
489 draw_llvm_create(struct draw_context
*draw
, LLVMContextRef context
)
491 struct draw_llvm
*llvm
;
493 if (!lp_build_init())
496 llvm
= CALLOC_STRUCT( draw_llvm
);
502 llvm
->context
= context
;
503 if (!llvm
->context
) {
504 llvm
->context
= LLVMContextCreate();
505 llvm
->context_owned
= true;
510 llvm
->nr_variants
= 0;
511 make_empty_list(&llvm
->vs_variants_list
);
513 llvm
->nr_gs_variants
= 0;
514 make_empty_list(&llvm
->gs_variants_list
);
519 draw_llvm_destroy(llvm
);
525 * Free per-context LLVM info.
528 draw_llvm_destroy(struct draw_llvm
*llvm
)
530 if (llvm
->context_owned
)
531 LLVMContextDispose(llvm
->context
);
532 llvm
->context
= NULL
;
534 /* XXX free other draw_llvm data? */
540 * Create LLVM-generated code for a vertex shader.
542 struct draw_llvm_variant
*
543 draw_llvm_create_variant(struct draw_llvm
*llvm
,
545 const struct draw_llvm_variant_key
*key
)
547 struct draw_llvm_variant
*variant
;
548 struct llvm_vertex_shader
*shader
=
549 llvm_vertex_shader(llvm
->draw
->vs
.vertex_shader
);
550 LLVMTypeRef vertex_header
;
551 char module_name
[64];
553 variant
= MALLOC(sizeof *variant
+
554 shader
->variant_key_size
-
555 sizeof variant
->key
);
559 variant
->llvm
= llvm
;
560 variant
->shader
= shader
;
562 util_snprintf(module_name
, sizeof(module_name
), "draw_llvm_vs_variant%u",
563 variant
->shader
->variants_cached
);
565 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
);
567 create_jit_types(variant
);
569 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
571 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
572 tgsi_dump(llvm
->draw
->vs
.vertex_shader
->state
.tokens
, 0);
573 draw_llvm_dump_variant_key(&variant
->key
);
576 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_inputs
);
578 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
580 draw_llvm_generate(llvm
, variant
, FALSE
); /* linear */
581 draw_llvm_generate(llvm
, variant
, TRUE
); /* elts */
583 gallivm_compile_module(variant
->gallivm
);
585 variant
->jit_func
= (draw_jit_vert_func
)
586 gallivm_jit_function(variant
->gallivm
, variant
->function
);
588 variant
->jit_func_elts
= (draw_jit_vert_func_elts
)
589 gallivm_jit_function(variant
->gallivm
, variant
->function_elts
);
591 gallivm_free_ir(variant
->gallivm
);
593 variant
->list_item_global
.base
= variant
;
594 variant
->list_item_local
.base
= variant
;
595 /*variant->no = */shader
->variants_created
++;
596 variant
->list_item_global
.base
= variant
;
603 generate_vs(struct draw_llvm_variant
*variant
,
604 LLVMBuilderRef builder
,
605 struct lp_type vs_type
,
606 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
607 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
608 const struct lp_bld_tgsi_system_values
*system_values
,
609 LLVMValueRef context_ptr
,
610 struct lp_build_sampler_soa
*draw_sampler
,
611 boolean clamp_vertex_color
)
613 struct draw_llvm
*llvm
= variant
->llvm
;
614 const struct tgsi_token
*tokens
= llvm
->draw
->vs
.vertex_shader
->state
.tokens
;
615 LLVMValueRef consts_ptr
=
616 draw_jit_context_vs_constants(variant
->gallivm
, context_ptr
);
617 LLVMValueRef num_consts_ptr
=
618 draw_jit_context_num_vs_constants(variant
->gallivm
, context_ptr
);
620 lp_build_tgsi_soa(variant
->gallivm
,
623 NULL
/*struct lp_build_mask_context *mask*/,
632 &llvm
->draw
->vs
.vertex_shader
->info
,
637 unsigned chan
, attrib
;
638 struct lp_build_context bld
;
639 struct tgsi_shader_info
* info
= &llvm
->draw
->vs
.vertex_shader
->info
;
640 lp_build_context_init(&bld
, variant
->gallivm
, vs_type
);
642 for (attrib
= 0; attrib
< info
->num_outputs
; ++attrib
) {
643 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
644 if (outputs
[attrib
][chan
]) {
645 switch (info
->output_semantic_name
[attrib
]) {
646 case TGSI_SEMANTIC_COLOR
:
647 case TGSI_SEMANTIC_BCOLOR
:
648 if (clamp_vertex_color
) {
649 out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
650 out
= lp_build_clamp(&bld
, out
, bld
.zero
, bld
.one
);
651 LLVMBuildStore(builder
, out
, outputs
[attrib
][chan
]);
663 fetch_instanced(struct gallivm_state
*gallivm
,
664 const struct util_format_description
*format_desc
,
665 struct lp_type vs_type
,
666 LLVMValueRef vb_stride
,
667 LLVMValueRef map_ptr
,
668 LLVMValueRef buffer_size_adj
,
670 LLVMValueRef
*inputs
,
673 LLVMValueRef zero
= LLVMConstNull(LLVMInt32TypeInContext(gallivm
->context
));
674 LLVMBuilderRef builder
= gallivm
->builder
;
675 LLVMValueRef stride
, buffer_overflowed
, aos
;
676 LLVMValueRef temp_ptr
=
677 lp_build_alloca(gallivm
,
678 lp_build_vec_type(gallivm
, lp_float32_vec4_type()), "");
679 struct lp_build_if_state if_ctx
;
682 stride
= lp_build_umul_overflow(gallivm
, vb_stride
, index
, &ofbit
);
684 buffer_overflowed
= LLVMBuildICmp(builder
, LLVMIntUGT
,
685 stride
, buffer_size_adj
,
686 "buffer_overflowed");
687 buffer_overflowed
= LLVMBuildOr(builder
, buffer_overflowed
, ofbit
, "");
690 lp_build_print_value(gallivm
, " instance index = ", index
);
691 lp_build_print_value(gallivm
, " buffer overflowed = ", buffer_overflowed
);
694 lp_build_if(&if_ctx
, gallivm
, buffer_overflowed
);
697 lp_build_const_vec(gallivm
, lp_float32_vec4_type(), 0);
698 LLVMBuildStore(builder
, val
, temp_ptr
);
700 lp_build_else(&if_ctx
);
704 val
= lp_build_fetch_rgba_aos(gallivm
,
706 lp_float32_vec4_type(),
711 LLVMBuildStore(builder
, val
, temp_ptr
);
713 lp_build_endif(&if_ctx
);
715 aos
= LLVMBuildLoad(builder
, temp_ptr
, "aos");
717 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
718 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
719 inputs
[i
] = lp_build_extract_broadcast(gallivm
,
720 lp_float32_vec4_type(),
721 vs_type
, aos
, index
);
727 convert_to_soa(struct gallivm_state
*gallivm
,
728 LLVMValueRef src_aos
[LP_MAX_VECTOR_WIDTH
/ 32],
729 LLVMValueRef dst_soa
[TGSI_NUM_CHANNELS
],
730 const struct lp_type soa_type
)
733 struct lp_type aos_channel_type
= soa_type
;
735 LLVMValueRef aos_channels
[TGSI_NUM_CHANNELS
];
736 unsigned pixels_per_channel
= soa_type
.length
/ TGSI_NUM_CHANNELS
;
738 debug_assert(TGSI_NUM_CHANNELS
== 4);
739 debug_assert((soa_type
.length
% TGSI_NUM_CHANNELS
) == 0);
741 aos_channel_type
.length
>>= 1;
743 for (j
= 0; j
< TGSI_NUM_CHANNELS
; ++j
) {
744 LLVMValueRef channel
[LP_MAX_VECTOR_LENGTH
] = { 0 };
746 assert(pixels_per_channel
<= LP_MAX_VECTOR_LENGTH
);
748 for (k
= 0; k
< pixels_per_channel
; ++k
) {
749 channel
[k
] = src_aos
[j
+ TGSI_NUM_CHANNELS
* k
];
752 aos_channels
[j
] = lp_build_concat(gallivm
, channel
, aos_channel_type
, pixels_per_channel
);
755 lp_build_transpose_aos(gallivm
, soa_type
, aos_channels
, dst_soa
);
760 fetch_vector(struct gallivm_state
*gallivm
,
761 const struct util_format_description
*format_desc
,
762 struct lp_type vs_type
,
763 LLVMValueRef vb_stride
,
764 LLVMValueRef map_ptr
,
765 LLVMValueRef buffer_size_adj
,
767 LLVMValueRef
*inputs
,
768 LLVMValueRef indices
)
770 LLVMValueRef zero
= LLVMConstNull(LLVMInt32TypeInContext(gallivm
->context
));
771 LLVMBuilderRef builder
= gallivm
->builder
;
772 struct lp_build_context blduivec
;
773 LLVMValueRef offset
, tmp
, valid_mask
;
774 LLVMValueRef aos_fetch
[LP_MAX_VECTOR_WIDTH
/ 32];
777 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
779 vb_stride
= lp_build_broadcast_scalar(&blduivec
, vb_stride
);
780 buffer_size_adj
= lp_build_broadcast_scalar(&blduivec
, buffer_size_adj
);
783 * Sort of interestingly, with interleaved attribs, llvm 3.7+ will
784 * recognize these calculations to be constant with different attribs
785 * (the different offset has been added to map_ptr).
786 * llvm 3.3, however, will not (I can't get llvm 3.4-3.6 to link...)
788 * XXX: could actually avoid this altogether (replacing by simple
789 * non-widening mul) by precalculating the max index instead outside
790 * the loop (at the cost of one scalar udiv per vertex element).
792 offset
= lp_build_mul_32_lohi_cpu(&blduivec
, vb_stride
, indices
, &tmp
);
794 tmp
= lp_build_compare(gallivm
, blduivec
.type
,
795 PIPE_FUNC_EQUAL
, tmp
, blduivec
.zero
);
796 valid_mask
= lp_build_andnot(&blduivec
, tmp
, ofmask
);
798 tmp
= lp_build_compare(gallivm
, blduivec
.type
,
799 PIPE_FUNC_LEQUAL
, offset
, buffer_size_adj
);
800 valid_mask
= LLVMBuildAnd(builder
, tmp
, valid_mask
, "");
802 /* not valid elements use offset 0 */
803 offset
= LLVMBuildAnd(builder
, offset
, valid_mask
, "");
806 lp_build_print_value(gallivm
, " indices = ", indices
);
807 lp_build_print_value(gallivm
, " offsets = ", offset
);
808 lp_build_print_value(gallivm
, " valid_mask = ", valid_mask
);
812 * Note: we probably really want to use SoA fetch, not AoS one (albeit
813 * for most formats it will amount to the same as this isn't very
814 * optimized). But looks dangerous since it assumes alignment.
816 for (i
= 0; i
< vs_type
.length
; i
++) {
817 LLVMValueRef offset1
, elem
;
818 elem
= lp_build_const_int32(gallivm
, i
);
819 offset1
= LLVMBuildExtractElement(builder
, offset
, elem
, "");
821 aos_fetch
[i
] = lp_build_fetch_rgba_aos(gallivm
, format_desc
,
822 lp_float32_vec4_type(),
823 FALSE
, map_ptr
, offset1
,
826 convert_to_soa(gallivm
, aos_fetch
, inputs
, vs_type
);
828 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
829 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
], blduivec
.vec_type
, "");
830 inputs
[i
] = LLVMBuildAnd(builder
, inputs
[i
], valid_mask
, "");
831 inputs
[i
] = LLVMBuildBitCast(builder
, inputs
[i
],
832 lp_build_vec_type(gallivm
, vs_type
), "");
839 store_aos(struct gallivm_state
*gallivm
,
844 LLVMTypeRef data_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, lp_float32_vec4_type()), 0);
845 LLVMBuilderRef builder
= gallivm
->builder
;
846 LLVMValueRef data_ptr
= draw_jit_header_data(gallivm
, io_ptr
);
847 LLVMValueRef indices
[3];
849 indices
[0] = lp_build_const_int32(gallivm
, 0);
851 indices
[2] = lp_build_const_int32(gallivm
, 0);
853 data_ptr
= LLVMBuildGEP(builder
, data_ptr
, indices
, 3, "");
854 data_ptr
= LLVMBuildPointerCast(builder
, data_ptr
, data_ptr_type
, "");
857 lp_build_printf(gallivm
, " ---- %p storing attribute %d (io = %p)\n", data_ptr
, index
, io_ptr
);
860 /* Unaligned store due to the vertex header */
861 LLVMSetAlignment(LLVMBuildStore(builder
, value
, data_ptr
), sizeof(float));
865 * Adjust the mask to architecture endianess. The mask will the store in struct:
867 * struct vertex_header {
868 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
869 * unsigned edgeflag:1;
871 * unsigned vertex_id:16;
875 * On little-endian machine nothing needs to done, however on bit-endian machine
876 * the mask's fields need to be adjusted with the algorithm:
878 * uint32_t reverse (uint32_t x)
880 * return (x >> 16) | // vertex_id
881 * ((x & 0x3fff) << 18) | // clipmask
882 * ((x & 0x4000) << 3) | // pad
883 * ((x & 0x8000) << 1); // edgeflag
887 adjust_mask(struct gallivm_state
*gallivm
,
890 #ifdef PIPE_ARCH_BIG_ENDIAN
891 LLVMBuilderRef builder
= gallivm
->builder
;
892 LLVMValueRef vertex_id
;
893 LLVMValueRef clipmask
;
895 LLVMValueRef edgeflag
;
897 vertex_id
= LLVMBuildLShr(builder
, mask
, lp_build_const_int32(gallivm
, 16), "");
898 clipmask
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x3fff), "");
899 clipmask
= LLVMBuildShl(builder
, clipmask
, lp_build_const_int32(gallivm
, 18), "");
901 pad
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x4000), "");
902 pad
= LLVMBuildShl(builder
, pad
, lp_build_const_int32(gallivm
, 3), "");
904 edgeflag
= LLVMBuildAnd(builder
, mask
, lp_build_const_int32(gallivm
, 0x8000), "");
905 edgeflag
= LLVMBuildShl(builder
, edgeflag
, lp_build_const_int32(gallivm
, 1), "");
907 mask
= LLVMBuildOr(builder
, vertex_id
, clipmask
, "");
909 mask
= LLVMBuildOr(builder
, mask
, pad
, "");
911 mask
= LLVMBuildOr(builder
, mask
, edgeflag
, "");
917 store_aos_array(struct gallivm_state
*gallivm
,
918 struct lp_type soa_type
,
920 LLVMValueRef
*indices
,
924 LLVMValueRef clipmask
,
925 boolean need_edgeflag
)
927 LLVMBuilderRef builder
= gallivm
->builder
;
928 LLVMValueRef attr_index
= lp_build_const_int32(gallivm
, attrib
);
929 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
930 LLVMValueRef linear_inds
[LP_MAX_VECTOR_WIDTH
/ 32];
931 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
932 int vector_length
= soa_type
.length
;
935 debug_assert(TGSI_NUM_CHANNELS
== 4);
937 for (i
= 0; i
< vector_length
; i
++) {
938 linear_inds
[i
] = lp_build_const_int32(gallivm
, i
);
940 inds
[i
] = indices
[i
];
942 inds
[i
] = linear_inds
[i
];
944 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
948 /* store vertex header for each of the n vertices */
949 LLVMValueRef val
, cliptmp
;
950 int vertex_id_pad_edgeflag
;
952 /* If this assertion fails, it means we need to update the bit twidding
953 * code here. See struct vertex_header in draw_private.h.
955 assert(DRAW_TOTAL_CLIP_PLANES
==14);
956 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
957 if (!need_edgeflag
) {
958 vertex_id_pad_edgeflag
= (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES
);
961 vertex_id_pad_edgeflag
= (0xffff << 16);
963 val
= lp_build_const_int_vec(gallivm
, lp_int_type(soa_type
),
964 vertex_id_pad_edgeflag
);
965 /* OR with the clipmask */
966 cliptmp
= LLVMBuildOr(builder
, val
, clipmask
, "");
967 for (i
= 0; i
< vector_length
; i
++) {
968 LLVMValueRef id_ptr
= draw_jit_header_id(gallivm
, io_ptrs
[i
]);
969 val
= LLVMBuildExtractElement(builder
, cliptmp
, linear_inds
[i
], "");
970 val
= adjust_mask(gallivm
, val
);
972 lp_build_printf(gallivm
, "io = %p, index %d, clipmask = %x\n",
973 io_ptrs
[i
], inds
[i
], val
);
975 LLVMBuildStore(builder
, val
, id_ptr
);
979 /* store for each of the n vertices */
980 for (i
= 0; i
< vector_length
; i
++) {
981 store_aos(gallivm
, io_ptrs
[i
], attr_index
, aos
[i
]);
987 convert_to_aos(struct gallivm_state
*gallivm
,
989 LLVMValueRef
*indices
,
990 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
991 LLVMValueRef clipmask
,
993 struct lp_type soa_type
,
994 boolean need_edgeflag
)
996 LLVMBuilderRef builder
= gallivm
->builder
;
997 unsigned chan
, attrib
, i
;
1000 lp_build_printf(gallivm
, " # storing begin\n");
1002 for (attrib
= 0; attrib
< num_outputs
; ++attrib
) {
1003 LLVMValueRef soa
[TGSI_NUM_CHANNELS
];
1004 LLVMValueRef aos
[LP_MAX_VECTOR_WIDTH
/ 32];
1005 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1006 if (outputs
[attrib
][chan
]) {
1007 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
1008 lp_build_name(out
, "output%u.%c", attrib
, "xyzw"[chan
]);
1010 lp_build_printf(gallivm
, "output %d : %d ",
1011 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1013 LLVMConstInt(LLVMInt32TypeInContext(gallivm
->context
),
1015 lp_build_print_value(gallivm
, "val = ", out
);
1018 LLVMBuildBitCast(builder
, out
, lp_build_int_vec_type(gallivm
, soa_type
), "");
1020 lp_build_print_value(gallivm
, " ival = ", iv
);
1031 if (soa_type
.length
== TGSI_NUM_CHANNELS
) {
1032 lp_build_transpose_aos(gallivm
, soa_type
, soa
, aos
);
1034 lp_build_transpose_aos(gallivm
, soa_type
, soa
, soa
);
1036 for (i
= 0; i
< soa_type
.length
; ++i
) {
1037 aos
[i
] = lp_build_extract_range(gallivm
,
1038 soa
[i
% TGSI_NUM_CHANNELS
],
1039 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1044 store_aos_array(gallivm
,
1054 lp_build_printf(gallivm
, " # storing end\n");
1060 * Stores original vertex positions in clip coordinates
1063 store_clip(struct gallivm_state
*gallivm
,
1064 const struct lp_type vs_type
,
1065 LLVMValueRef io_ptr
,
1066 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1069 LLVMBuilderRef builder
= gallivm
->builder
;
1070 LLVMValueRef soa
[4];
1071 LLVMValueRef aos
[LP_MAX_VECTOR_LENGTH
];
1072 LLVMValueRef indices
[2];
1073 LLVMValueRef io_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1074 LLVMValueRef inds
[LP_MAX_VECTOR_WIDTH
/ 32];
1075 LLVMValueRef clip_ptrs
[LP_MAX_VECTOR_WIDTH
/ 32];
1076 LLVMTypeRef clip_ptr_type
=
1077 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
),
1082 indices
[1] = lp_build_const_int32(gallivm
, 0);
1084 for (i
= 0; i
< vs_type
.length
; i
++) {
1085 inds
[i
] = lp_build_const_int32(gallivm
, i
);
1086 io_ptrs
[i
] = LLVMBuildGEP(builder
, io_ptr
, &inds
[i
], 1, "");
1089 soa
[0] = LLVMBuildLoad(builder
, outputs
[idx
][0], ""); /*x0 x1 .. xn*/
1090 soa
[1] = LLVMBuildLoad(builder
, outputs
[idx
][1], ""); /*y0 y1 .. yn*/
1091 soa
[2] = LLVMBuildLoad(builder
, outputs
[idx
][2], ""); /*z0 z1 .. zn*/
1092 soa
[3] = LLVMBuildLoad(builder
, outputs
[idx
][3], ""); /*w0 w1 .. wn*/
1094 for (i
= 0; i
< vs_type
.length
; i
++) {
1095 clip_ptrs
[i
] = draw_jit_header_clip_pos(gallivm
, io_ptrs
[i
]);
1098 lp_build_transpose_aos(gallivm
, vs_type
, soa
, soa
);
1099 for (i
= 0; i
< vs_type
.length
; ++i
) {
1100 aos
[i
] = lp_build_extract_range(gallivm
,
1101 soa
[i
% TGSI_NUM_CHANNELS
],
1102 (i
/ TGSI_NUM_CHANNELS
) * TGSI_NUM_CHANNELS
,
1106 for (j
= 0; j
< vs_type
.length
; j
++) {
1107 LLVMValueRef clip_ptr
;
1109 clip_ptr
= LLVMBuildGEP(builder
, clip_ptrs
[j
], indices
, 2, "clipo");
1110 clip_ptr
= LLVMBuildPointerCast(builder
, clip_ptr
, clip_ptr_type
, "");
1112 /* Unaligned store */
1113 LLVMSetAlignment(LLVMBuildStore(builder
, aos
[j
], clip_ptr
), sizeof(float));
1119 * Transforms the outputs for viewport mapping
1122 generate_viewport(struct draw_llvm_variant
*variant
,
1123 LLVMBuilderRef builder
,
1124 struct lp_type vs_type
,
1125 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1126 LLVMValueRef context_ptr
)
1129 struct gallivm_state
*gallivm
= variant
->gallivm
;
1130 struct lp_type f32_type
= vs_type
;
1131 const unsigned pos
= variant
->llvm
->draw
->vs
.position_output
;
1132 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1133 LLVMValueRef out3
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn*/
1134 LLVMValueRef const1
= lp_build_const_vec(gallivm
, f32_type
, 1.0); /*1.0 1.0 1.0 1.0*/
1135 LLVMValueRef vp_ptr
= draw_jit_context_viewports(gallivm
, context_ptr
);
1137 /* We treat pipe_viewport_state as a float array */
1138 const int scale_index_offset
= offsetof(struct pipe_viewport_state
, scale
) / sizeof(float);
1139 const int trans_index_offset
= offsetof(struct pipe_viewport_state
, translate
) / sizeof(float);
1141 /* for 1/w convention*/
1142 out3
= LLVMBuildFDiv(builder
, const1
, out3
, "");
1143 LLVMBuildStore(builder
, out3
, outputs
[pos
][3]);
1145 /* Viewport Mapping */
1146 for (i
=0; i
<3; i
++) {
1147 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[pos
][i
], ""); /*x0 x1 .. xn*/
1150 LLVMValueRef scale_i
;
1151 LLVMValueRef trans_i
;
1154 index
= lp_build_const_int32(gallivm
, i
+ scale_index_offset
);
1155 scale_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1157 index
= lp_build_const_int32(gallivm
, i
+ trans_index_offset
);
1158 trans_i
= LLVMBuildGEP(builder
, vp_ptr
, &index
, 1, "");
1160 scale
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1161 LLVMBuildLoad(builder
, scale_i
, "scale"));
1162 trans
= lp_build_broadcast(gallivm
, vs_type_llvm
,
1163 LLVMBuildLoad(builder
, trans_i
, "trans"));
1166 out
= LLVMBuildFMul(builder
, out
, out3
, "");
1167 /* mult by scale, add translation */
1168 out
= lp_build_fmuladd(builder
, out
, scale
, trans
);
1170 /* store transformed outputs */
1171 LLVMBuildStore(builder
, out
, outputs
[pos
][i
]);
1178 * Returns clipmask as nxi32 bitmask for the n vertices
1181 generate_clipmask(struct draw_llvm
*llvm
,
1182 struct gallivm_state
*gallivm
,
1183 struct lp_type vs_type
,
1184 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1185 struct draw_llvm_variant_key
*key
,
1186 LLVMValueRef context_ptr
,
1187 boolean
*have_clipdist
)
1189 LLVMBuilderRef builder
= gallivm
->builder
;
1190 LLVMValueRef mask
; /* stores the <nxi32> clipmasks */
1191 LLVMValueRef test
, temp
;
1192 LLVMValueRef zero
, shift
;
1193 LLVMValueRef pos_x
, pos_y
, pos_z
, pos_w
;
1194 LLVMValueRef cv_x
, cv_y
, cv_z
, cv_w
;
1195 LLVMValueRef plane1
, planes
, plane_ptr
, sum
;
1196 struct lp_type f32_type
= vs_type
;
1197 struct lp_type i32_type
= lp_int_type(vs_type
);
1198 const unsigned pos
= llvm
->draw
->vs
.position_output
;
1199 const unsigned cv
= llvm
->draw
->vs
.clipvertex_output
;
1200 int num_written_clipdistance
= llvm
->draw
->vs
.vertex_shader
->info
.num_written_clipdistance
;
1201 boolean have_cd
= false;
1202 boolean clip_user
= key
->clip_user
;
1203 unsigned ucp_enable
= key
->ucp_enable
;
1206 cd
[0] = llvm
->draw
->vs
.ccdistance_output
[0];
1207 cd
[1] = llvm
->draw
->vs
.ccdistance_output
[1];
1209 if (cd
[0] != pos
|| cd
[1] != pos
)
1212 if (num_written_clipdistance
&& !clip_user
) {
1214 ucp_enable
= (1 << num_written_clipdistance
) - 1;
1217 mask
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1218 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 0);
1219 zero
= lp_build_const_vec(gallivm
, f32_type
, 0); /* 0.0f 0.0f 0.0f 0.0f */
1220 shift
= lp_build_const_int_vec(gallivm
, i32_type
, 1); /* 1 1 1 1 */
1223 * load clipvertex and position from correct locations.
1224 * if they are the same just load them once.
1226 pos_x
= LLVMBuildLoad(builder
, outputs
[pos
][0], ""); /*x0 x1 .. xn */
1227 pos_y
= LLVMBuildLoad(builder
, outputs
[pos
][1], ""); /*y0 y1 .. yn */
1228 pos_z
= LLVMBuildLoad(builder
, outputs
[pos
][2], ""); /*z0 z1 .. zn */
1229 pos_w
= LLVMBuildLoad(builder
, outputs
[pos
][3], ""); /*w0 w1 .. wn */
1231 if (clip_user
&& cv
!= pos
) {
1232 cv_x
= LLVMBuildLoad(builder
, outputs
[cv
][0], ""); /*x0 x1 .. xn */
1233 cv_y
= LLVMBuildLoad(builder
, outputs
[cv
][1], ""); /*y0 y1 .. yn */
1234 cv_z
= LLVMBuildLoad(builder
, outputs
[cv
][2], ""); /*z0 z1 .. zn */
1235 cv_w
= LLVMBuildLoad(builder
, outputs
[cv
][3], ""); /*w0 w1 .. wn */
1244 * Be careful with the comparisons and NaNs (using llvm's unordered
1245 * comparisons here).
1247 /* Cliptest, for hardwired planes */
1249 * XXX should take guardband into account (currently not in key).
1250 * Otherwise might run the draw pipeline stages for nothing.
1254 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_x
, pos_w
);
1256 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1260 test
= LLVMBuildFAdd(builder
, pos_x
, pos_w
, "");
1261 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1262 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1263 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1264 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1267 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_y
, pos_w
);
1268 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1269 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1270 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1273 test
= LLVMBuildFAdd(builder
, pos_y
, pos_w
, "");
1274 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1275 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1276 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1277 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1281 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 16);
1282 if (key
->clip_halfz
) {
1284 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, pos_z
);
1285 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1286 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1290 test
= LLVMBuildFAdd(builder
, pos_z
, pos_w
, "");
1291 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, test
);
1292 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1293 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1296 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, pos_z
, pos_w
);
1297 temp
= LLVMBuildShl(builder
, temp
, shift
, "");
1298 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1299 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1303 LLVMValueRef planes_ptr
= draw_jit_context_planes(gallivm
, context_ptr
);
1304 LLVMValueRef indices
[3];
1305 LLVMValueRef is_nan_or_inf
;
1307 /* userclip planes */
1308 while (ucp_enable
) {
1309 unsigned plane_idx
= ffs(ucp_enable
)-1;
1310 ucp_enable
&= ~(1 << plane_idx
);
1313 if (have_cd
&& num_written_clipdistance
) {
1314 LLVMValueRef clipdist
;
1318 *have_clipdist
= TRUE
;
1320 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[0]][i
], "");
1322 clipdist
= LLVMBuildLoad(builder
, outputs
[cd
[1]][i
-4], "");
1324 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, clipdist
);
1325 is_nan_or_inf
= lp_build_is_inf_or_nan(gallivm
, vs_type
, clipdist
);
1326 test
= LLVMBuildOr(builder
, test
, is_nan_or_inf
, "");
1327 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1328 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1329 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1331 LLVMTypeRef vs_type_llvm
= lp_build_vec_type(gallivm
, vs_type
);
1332 indices
[0] = lp_build_const_int32(gallivm
, 0);
1333 indices
[1] = lp_build_const_int32(gallivm
, plane_idx
);
1335 indices
[2] = lp_build_const_int32(gallivm
, 0);
1336 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1337 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_x");
1338 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1339 sum
= LLVMBuildFMul(builder
, planes
, cv_x
, "");
1341 indices
[2] = lp_build_const_int32(gallivm
, 1);
1342 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1343 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_y");
1344 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1345 sum
= lp_build_fmuladd(builder
, planes
, cv_y
, sum
);
1347 indices
[2] = lp_build_const_int32(gallivm
, 2);
1348 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1349 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_z");
1350 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1351 sum
= lp_build_fmuladd(builder
, planes
, cv_z
, sum
);
1353 indices
[2] = lp_build_const_int32(gallivm
, 3);
1354 plane_ptr
= LLVMBuildGEP(builder
, planes_ptr
, indices
, 3, "");
1355 plane1
= LLVMBuildLoad(builder
, plane_ptr
, "plane_w");
1356 planes
= lp_build_broadcast(gallivm
, vs_type_llvm
, plane1
);
1357 sum
= lp_build_fmuladd(builder
, planes
, cv_w
, sum
);
1359 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_GREATER
, zero
, sum
);
1360 temp
= lp_build_const_int_vec(gallivm
, i32_type
, 1LL << plane_idx
);
1361 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1362 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1366 if (key
->need_edgeflags
) {
1368 * This isn't really part of clipmask but stored the same in vertex
1369 * header later, so do it here.
1371 unsigned edge_attr
= llvm
->draw
->vs
.edgeflag_output
;
1372 LLVMValueRef one
= lp_build_const_vec(gallivm
, f32_type
, 1.0);
1373 LLVMValueRef edgeflag
= LLVMBuildLoad(builder
, outputs
[edge_attr
][0], "");
1374 test
= lp_build_compare(gallivm
, f32_type
, PIPE_FUNC_EQUAL
, one
, edgeflag
);
1375 temp
= lp_build_const_int_vec(gallivm
, i32_type
,
1376 1LL << DRAW_TOTAL_CLIP_PLANES
);
1377 test
= LLVMBuildAnd(builder
, test
, temp
, "");
1378 mask
= LLVMBuildOr(builder
, mask
, test
, "");
1385 * Returns boolean if any clipping has occurred
1386 * Used zero/non-zero i32 value to represent boolean
1389 clipmask_booli32(struct gallivm_state
*gallivm
,
1390 const struct lp_type vs_type
,
1391 LLVMValueRef clipmask_bool_ptr
,
1392 boolean edgeflag_in_clipmask
)
1394 LLVMBuilderRef builder
= gallivm
->builder
;
1395 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
1396 LLVMValueRef clipmask_bool
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1397 LLVMValueRef ret
= LLVMConstNull(int32_type
);
1402 * We need to invert the edgeflag bit from the clipmask here
1403 * (because the result is really if we want to run the pipeline or not
1404 * and we (may) need it if edgeflag was 0).
1406 if (edgeflag_in_clipmask
) {
1407 struct lp_type i32_type
= lp_int_type(vs_type
);
1408 LLVMValueRef edge
= lp_build_const_int_vec(gallivm
, i32_type
,
1409 1LL << DRAW_TOTAL_CLIP_PLANES
);
1410 clipmask_bool
= LLVMBuildXor(builder
, clipmask_bool
, edge
, "");
1413 * Could do much better with just cmp/movmskps.
1415 for (i
=0; i
< vs_type
.length
; i
++) {
1416 temp
= LLVMBuildExtractElement(builder
, clipmask_bool
,
1417 lp_build_const_int32(gallivm
, i
) , "");
1418 ret
= LLVMBuildOr(builder
, ret
, temp
, "");
1424 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface
*gs_iface
,
1425 struct lp_build_tgsi_context
* bld_base
,
1426 boolean is_vindex_indirect
,
1427 LLVMValueRef vertex_index
,
1428 boolean is_aindex_indirect
,
1429 LLVMValueRef attrib_index
,
1430 LLVMValueRef swizzle_index
)
1432 const struct draw_gs_llvm_iface
*gs
= draw_gs_llvm_iface(gs_iface
);
1433 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1434 LLVMBuilderRef builder
= gallivm
->builder
;
1435 LLVMValueRef indices
[3];
1437 struct lp_type type
= bld_base
->base
.type
;
1439 if (is_vindex_indirect
|| is_aindex_indirect
) {
1441 res
= bld_base
->base
.zero
;
1442 for (i
= 0; i
< type
.length
; ++i
) {
1443 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
1444 LLVMValueRef vert_chan_index
= vertex_index
;
1445 LLVMValueRef attr_chan_index
= attrib_index
;
1446 LLVMValueRef channel_vec
, value
;
1448 if (is_vindex_indirect
) {
1449 vert_chan_index
= LLVMBuildExtractElement(builder
,
1450 vertex_index
, idx
, "");
1452 if (is_aindex_indirect
) {
1453 attr_chan_index
= LLVMBuildExtractElement(builder
,
1454 attrib_index
, idx
, "");
1457 indices
[0] = vert_chan_index
;
1458 indices
[1] = attr_chan_index
;
1459 indices
[2] = swizzle_index
;
1461 channel_vec
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1462 channel_vec
= LLVMBuildLoad(builder
, channel_vec
, "");
1463 value
= LLVMBuildExtractElement(builder
, channel_vec
, idx
, "");
1465 res
= LLVMBuildInsertElement(builder
, res
, value
, idx
, "");
1468 indices
[0] = vertex_index
;
1469 indices
[1] = attrib_index
;
1470 indices
[2] = swizzle_index
;
1472 res
= LLVMBuildGEP(builder
, gs
->input
, indices
, 3, "");
1473 res
= LLVMBuildLoad(builder
, res
, "");
1480 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface
*gs_base
,
1481 struct lp_build_tgsi_context
* bld_base
,
1482 LLVMValueRef (*outputs
)[4],
1483 LLVMValueRef emitted_vertices_vec
)
1485 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1486 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1487 struct gallivm_state
*gallivm
= variant
->gallivm
;
1488 LLVMBuilderRef builder
= gallivm
->builder
;
1489 struct lp_type gs_type
= bld_base
->base
.type
;
1490 LLVMValueRef clipmask
= lp_build_const_int_vec(gallivm
,
1491 lp_int_type(gs_type
), 0);
1492 LLVMValueRef indices
[LP_MAX_VECTOR_LENGTH
];
1493 LLVMValueRef next_prim_offset
=
1494 lp_build_const_int32(gallivm
, variant
->shader
->base
.primitive_boundary
);
1495 LLVMValueRef io
= variant
->io_ptr
;
1497 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
1499 for (i
= 0; i
< gs_type
.length
; ++i
) {
1500 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1501 LLVMValueRef currently_emitted
=
1502 LLVMBuildExtractElement(builder
, emitted_vertices_vec
, ind
, "");
1503 indices
[i
] = LLVMBuildMul(builder
, ind
, next_prim_offset
, "");
1504 indices
[i
] = LLVMBuildAdd(builder
, indices
[i
], currently_emitted
, "");
1507 convert_to_aos(gallivm
, io
, indices
,
1509 gs_info
->num_outputs
, gs_type
,
1514 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface
*gs_base
,
1515 struct lp_build_tgsi_context
* bld_base
,
1516 LLVMValueRef verts_per_prim_vec
,
1517 LLVMValueRef emitted_prims_vec
)
1519 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1520 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1521 struct gallivm_state
*gallivm
= variant
->gallivm
;
1522 LLVMBuilderRef builder
= gallivm
->builder
;
1523 LLVMValueRef prim_lengts_ptr
=
1524 draw_gs_jit_prim_lengths(variant
->gallivm
, variant
->context_ptr
);
1527 for (i
= 0; i
< bld_base
->base
.type
.length
; ++i
) {
1528 LLVMValueRef ind
= lp_build_const_int32(gallivm
, i
);
1529 LLVMValueRef prims_emitted
=
1530 LLVMBuildExtractElement(builder
, emitted_prims_vec
, ind
, "");
1531 LLVMValueRef store_ptr
;
1532 LLVMValueRef num_vertices
=
1533 LLVMBuildExtractElement(builder
, verts_per_prim_vec
, ind
, "");
1535 store_ptr
= LLVMBuildGEP(builder
, prim_lengts_ptr
, &prims_emitted
, 1, "");
1536 store_ptr
= LLVMBuildLoad(builder
, store_ptr
, "");
1537 store_ptr
= LLVMBuildGEP(builder
, store_ptr
, &ind
, 1, "");
1538 LLVMBuildStore(builder
, num_vertices
, store_ptr
);
1543 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface
*gs_base
,
1544 struct lp_build_tgsi_context
* bld_base
,
1545 LLVMValueRef total_emitted_vertices_vec
,
1546 LLVMValueRef emitted_prims_vec
)
1548 const struct draw_gs_llvm_iface
*gs_iface
= draw_gs_llvm_iface(gs_base
);
1549 struct draw_gs_llvm_variant
*variant
= gs_iface
->variant
;
1550 struct gallivm_state
*gallivm
= variant
->gallivm
;
1551 LLVMBuilderRef builder
= gallivm
->builder
;
1552 LLVMValueRef emitted_verts_ptr
=
1553 draw_gs_jit_emitted_vertices(gallivm
, variant
->context_ptr
);
1554 LLVMValueRef emitted_prims_ptr
=
1555 draw_gs_jit_emitted_prims(gallivm
, variant
->context_ptr
);
1556 LLVMValueRef zero
= lp_build_const_int32(gallivm
, 0);
1558 emitted_verts_ptr
= LLVMBuildGEP(builder
, emitted_verts_ptr
, &zero
, 0, "");
1559 emitted_prims_ptr
= LLVMBuildGEP(builder
, emitted_prims_ptr
, &zero
, 0, "");
1561 LLVMBuildStore(builder
, total_emitted_vertices_vec
, emitted_verts_ptr
);
1562 LLVMBuildStore(builder
, emitted_prims_vec
, emitted_prims_ptr
);
1566 draw_llvm_generate(struct draw_llvm
*llvm
, struct draw_llvm_variant
*variant
,
1569 struct gallivm_state
*gallivm
= variant
->gallivm
;
1570 LLVMContextRef context
= gallivm
->context
;
1571 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
1572 LLVMTypeRef arg_types
[11];
1573 unsigned num_arg_types
=
1574 elts
? ARRAY_SIZE(arg_types
) : ARRAY_SIZE(arg_types
) - 1;
1575 LLVMTypeRef func_type
;
1576 LLVMValueRef context_ptr
;
1577 LLVMBasicBlockRef block
;
1578 LLVMBuilderRef builder
;
1580 struct lp_type vs_type
;
1581 LLVMValueRef count
, fetch_elts
, fetch_elt_max
, fetch_count
;
1582 LLVMValueRef vertex_id_offset
, start_instance
, start
;
1583 LLVMValueRef stride
, step
, io_itr
;
1584 LLVMValueRef ind_vec
;
1585 LLVMValueRef io_ptr
, vbuffers_ptr
, vb_ptr
;
1586 LLVMValueRef vb_stride
[PIPE_MAX_ATTRIBS
];
1587 LLVMValueRef map_ptr
[PIPE_MAX_ATTRIBS
];
1588 LLVMValueRef buffer_size_adj
[PIPE_MAX_ATTRIBS
];
1589 LLVMValueRef ofmask
[PIPE_MAX_ATTRIBS
];
1590 LLVMValueRef instance_index
[PIPE_MAX_ATTRIBS
];
1591 LLVMValueRef fake_buf_ptr
, fake_buf
;
1593 struct draw_context
*draw
= llvm
->draw
;
1594 const struct tgsi_shader_info
*vs_info
= &draw
->vs
.vertex_shader
->info
;
1596 struct lp_build_context bld
, bldivec
, blduivec
;
1597 struct lp_build_loop_state lp_loop
;
1598 const int vector_length
= lp_native_vector_width
/ 32;
1599 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
1600 LLVMValueRef fetch_max
;
1601 struct lp_build_sampler_soa
*sampler
= 0;
1602 LLVMValueRef ret
, clipmask_bool_ptr
;
1603 struct draw_llvm_variant_key
*key
= &variant
->key
;
1604 /* If geometry shader is present we need to skip both the viewport
1605 * transformation and clipping otherwise the inputs to the geometry
1606 * shader will be incorrect.
1607 * The code can't handle vp transform when vs writes vp index neither
1608 * (though this would be fixable here, but couldn't just broadcast
1611 const boolean bypass_viewport
= key
->has_gs
|| key
->bypass_viewport
||
1612 vs_info
->writes_viewport_index
;
1613 const boolean enable_cliptest
= !key
->has_gs
&& (key
->clip_xy
||
1616 key
->need_edgeflags
);
1617 LLVMValueRef variant_func
;
1618 const unsigned pos
= draw
->vs
.position_output
;
1619 const unsigned cv
= draw
->vs
.clipvertex_output
;
1620 boolean have_clipdist
= FALSE
;
1621 struct lp_bld_tgsi_system_values system_values
;
1623 memset(&system_values
, 0, sizeof(system_values
));
1625 util_snprintf(func_name
, sizeof(func_name
), "draw_llvm_vs_variant%u_%s",
1626 variant
->shader
->variants_cached
, elts
? "elts" : "linear");
1629 arg_types
[i
++] = get_context_ptr_type(variant
); /* context */
1630 arg_types
[i
++] = get_vertex_header_ptr_type(variant
); /* vertex_header */
1631 arg_types
[i
++] = get_buffer_ptr_type(variant
); /* vbuffers */
1633 arg_types
[i
++] = LLVMPointerType(int32_type
, 0);/* fetch_elts */
1634 arg_types
[i
++] = int32_type
; /* fetch_elt_max */
1636 arg_types
[i
++] = int32_type
; /* start */
1637 arg_types
[i
++] = int32_type
; /* fetch_count / count */
1638 arg_types
[i
++] = int32_type
; /* stride */
1639 arg_types
[i
++] = get_vb_ptr_type(variant
); /* pipe_vertex_buffer's */
1640 arg_types
[i
++] = int32_type
; /* instance_id */
1641 arg_types
[i
++] = int32_type
; /* vertex_id_offset */
1642 arg_types
[i
++] = int32_type
; /* start_instance */
1644 func_type
= LLVMFunctionType(int32_type
, arg_types
, num_arg_types
, 0);
1646 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
1649 variant
->function_elts
= variant_func
;
1651 variant
->function
= variant_func
;
1653 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
1654 for (i
= 0; i
< num_arg_types
; ++i
)
1655 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
1656 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
1658 context_ptr
= LLVMGetParam(variant_func
, 0);
1659 io_ptr
= LLVMGetParam(variant_func
, 1);
1660 vbuffers_ptr
= LLVMGetParam(variant_func
, 2);
1662 * XXX: stride is actually unused. The stride we use is strictly calculated
1663 * from the number of outputs (including the draw_extra outputs).
1664 * Should probably fix some day (we need a new vs just because of extra
1665 * outputs which the generated vs won't touch).
1667 stride
= LLVMGetParam(variant_func
, 5 + (elts
? 1 : 0));
1668 vb_ptr
= LLVMGetParam(variant_func
, 6 + (elts
? 1 : 0));
1669 system_values
.instance_id
= LLVMGetParam(variant_func
, 7 + (elts
? 1 : 0));
1670 vertex_id_offset
= LLVMGetParam(variant_func
, 8 + (elts
? 1 : 0));
1671 start_instance
= LLVMGetParam(variant_func
, 9 + (elts
? 1 : 0));
1673 lp_build_name(context_ptr
, "context");
1674 lp_build_name(io_ptr
, "io");
1675 lp_build_name(vbuffers_ptr
, "vbuffers");
1676 lp_build_name(stride
, "stride");
1677 lp_build_name(vb_ptr
, "vb");
1678 lp_build_name(system_values
.instance_id
, "instance_id");
1679 lp_build_name(vertex_id_offset
, "vertex_id_offset");
1680 lp_build_name(start_instance
, "start_instance");
1683 fetch_elts
= LLVMGetParam(variant_func
, 3);
1684 fetch_elt_max
= LLVMGetParam(variant_func
, 4);
1685 fetch_count
= LLVMGetParam(variant_func
, 5);
1686 lp_build_name(fetch_elts
, "fetch_elts");
1687 lp_build_name(fetch_elt_max
, "fetch_elt_max");
1688 lp_build_name(fetch_count
, "fetch_count");
1689 start
= count
= NULL
;
1692 start
= LLVMGetParam(variant_func
, 3);
1693 count
= LLVMGetParam(variant_func
, 4);
1694 lp_build_name(start
, "start");
1695 lp_build_name(count
, "count");
1696 fetch_elts
= fetch_count
= NULL
;
1703 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
1704 builder
= gallivm
->builder
;
1705 LLVMPositionBuilderAtEnd(builder
, block
);
1707 memset(&vs_type
, 0, sizeof vs_type
);
1708 vs_type
.floating
= TRUE
; /* floating point values */
1709 vs_type
.sign
= TRUE
; /* values are signed */
1710 vs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
1711 vs_type
.width
= 32; /* 32-bit float */
1712 vs_type
.length
= vector_length
;
1714 lp_build_context_init(&bld
, gallivm
, lp_type_uint(32));
1715 lp_build_context_init(&bldivec
, gallivm
, lp_int_type(vs_type
));
1716 lp_build_context_init(&blduivec
, gallivm
, lp_uint_type(vs_type
));
1718 /* hold temporary "bool" clipmask */
1719 clipmask_bool_ptr
= lp_build_alloca(gallivm
, blduivec
.vec_type
, "");
1721 fake_buf
= lp_build_alloca_undef(gallivm
,
1722 LLVMVectorType(LLVMInt64TypeInContext(context
), 4), "");
1723 fake_buf
= LLVMBuildBitCast(builder
, fake_buf
,
1724 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
1725 fake_buf_ptr
= LLVMBuildGEP(builder
, fake_buf
, &bld
.zero
, 1, "");
1727 /* code generated texture sampling */
1728 sampler
= draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key
));
1730 step
= lp_build_const_int32(gallivm
, vector_length
);
1732 ind_vec
= blduivec
.undef
;
1733 for (i
= 0; i
< vs_type
.length
; i
++) {
1734 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
1735 ind_vec
= LLVMBuildInsertElement(builder
, ind_vec
, index
, index
, "");
1740 fetch_max
= fetch_count
;
1741 count
= fetch_count
;
1742 start
= blduivec
.zero
;
1745 fetch_max
= lp_build_add(&bld
, start
, count
);
1746 start
= lp_build_broadcast_scalar(&blduivec
, start
);
1747 ind_vec
= lp_build_add(&blduivec
, start
, ind_vec
);
1750 fetch_max
= LLVMBuildSub(builder
, fetch_max
, bld
.one
, "fetch_max");
1751 fetch_max
= lp_build_broadcast_scalar(&blduivec
, fetch_max
);
1754 * Pre-calculate everything which is constant per shader invocation.
1756 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
1757 LLVMValueRef vb_buffer_offset
, buffer_size
, temp_ptr
;
1758 LLVMValueRef vb_info
, vbuffer_ptr
, buf_offset
, ofbit
;
1759 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
1760 LLVMValueRef vb_index
=
1761 lp_build_const_int32(gallivm
, velem
->vertex_buffer_index
);
1762 LLVMValueRef bsize
= lp_build_const_int32(gallivm
,
1763 util_format_get_blocksize(velem
->src_format
));
1764 LLVMValueRef src_offset
= lp_build_const_int32(gallivm
,
1766 struct lp_build_if_state if_ctx
;
1768 if (velem
->src_format
!= PIPE_FORMAT_NONE
) {
1769 vbuffer_ptr
= LLVMBuildGEP(builder
, vbuffers_ptr
, &vb_index
, 1, "");
1770 vb_info
= LLVMBuildGEP(builder
, vb_ptr
, &vb_index
, 1, "");
1771 vb_stride
[j
] = draw_jit_vbuffer_stride(gallivm
, vb_info
);
1772 vb_buffer_offset
= draw_jit_vbuffer_offset(gallivm
, vb_info
);
1773 map_ptr
[j
] = draw_jit_dvbuffer_map(gallivm
, vbuffer_ptr
);
1774 buffer_size
= draw_jit_dvbuffer_size(gallivm
, vbuffer_ptr
);
1777 buf_offset
= lp_build_uadd_overflow(gallivm
, vb_buffer_offset
,
1778 src_offset
, &ofbit
);
1779 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size
, bsize
,
1781 buffer_size_adj
[j
] = lp_build_usub_overflow(gallivm
, buffer_size_adj
[j
],
1782 buf_offset
, &ofbit
);
1785 * We can't easily set fake vertex buffers outside the generated code.
1786 * Hence, set fake vertex buffers here instead basically, so fetch
1787 * code can always fetch using offset 0, eliminating all control flow
1788 * inside the main loop.
1789 * (Alternatively, could have control flow per vector skipping fetch
1790 * if ofbit is true.)
1791 * For instanced elements, we keep the control flow for now as it's a
1792 * scalar fetch, making things easier.
1794 if (velem
->instance_divisor
) {
1795 /* Index is equal to the start instance plus the number of current
1796 * instance divided by the divisor. In this case we compute it as:
1797 * index = start_instance + (instance_id / divisor)
1799 LLVMValueRef current_instance
;
1800 current_instance
= LLVMBuildUDiv(builder
, system_values
.instance_id
,
1801 lp_build_const_int32(gallivm
,
1802 velem
->instance_divisor
),
1803 "instance_divisor");
1804 instance_index
[j
] = lp_build_uadd_overflow(gallivm
, start_instance
,
1805 current_instance
, &ofbit
);
1806 map_ptr
[j
] = LLVMBuildGEP(builder
, map_ptr
[j
], &buf_offset
, 1, "");
1808 /* This is a scalar fetch, just keep the of bit */
1812 temp_ptr
= lp_build_alloca_undef(gallivm
,
1813 LLVMPointerType(LLVMInt8TypeInContext(context
), 0), "");
1815 lp_build_if(&if_ctx
, gallivm
, ofbit
);
1817 LLVMBuildStore(builder
, fake_buf_ptr
, temp_ptr
);
1819 lp_build_else(&if_ctx
);
1821 map_ptr
[j
] = LLVMBuildGEP(builder
, map_ptr
[j
], &buf_offset
, 1, "");
1822 LLVMBuildStore(builder
, map_ptr
[j
], temp_ptr
);
1824 lp_build_endif(&if_ctx
);
1825 map_ptr
[j
] = LLVMBuildLoad(builder
, temp_ptr
, "map_ptr");
1827 /* Expand to vector mask */
1828 ofmask
[j
] = LLVMBuildSExt(builder
, ofbit
, int32_type
, "");
1829 ofmask
[j
] = lp_build_broadcast_scalar(&blduivec
, ofmask
[j
]);
1833 lp_build_printf(gallivm
, "velem %d, vbuf index = %u, vb_stride = %u\n",
1834 lp_build_const_int32(gallivm
, j
),
1835 vb_index
, vb_stride
[j
]);
1836 lp_build_printf(gallivm
,
1837 " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1838 vb_buffer_offset
, src_offset
, buf_offset
);
1839 lp_build_printf(gallivm
, " buffer size = %u, blocksize = %u\n",
1840 buffer_size
, bsize
);
1841 lp_build_printf(gallivm
, " instance_id = %u\n", system_values
.instance_id
);
1846 lp_build_loop_begin(&lp_loop
, gallivm
, bld
.zero
);
1848 LLVMValueRef inputs
[PIPE_MAX_SHADER_INPUTS
][TGSI_NUM_CHANNELS
];
1850 LLVMValueRef clipmask
; /* holds the clipmask value */
1851 LLVMValueRef true_index_array
;
1852 const LLVMValueRef (*ptr_aos
)[TGSI_NUM_CHANNELS
];
1854 io_itr
= lp_loop
.counter
;
1856 io
= LLVMBuildGEP(builder
, io_ptr
, &io_itr
, 1, "");
1858 lp_build_printf(gallivm
, " --- io %d = %p, loop counter %d\n",
1859 io_itr
, io
, lp_loop
.counter
);
1862 true_index_array
= lp_build_broadcast_scalar(&blduivec
, lp_loop
.counter
);
1863 true_index_array
= LLVMBuildAdd(builder
, true_index_array
, ind_vec
, "");
1866 * XXX: This code is really fishy. We are required to use a int min
1867 * here, not uint. The reason is that for some non-indexed draws, we
1868 * might get something like MAX_UINT - 3 as start value (due to start
1869 * vertex). So, the first 3 elements in the vector are huge, and
1870 * limiting them to fetch_max is incorrect. By using int min, we'll
1871 * pick that huge value - we rely on this creating an overflow (which
1872 * is guaranteed) in the stride mul later (using (signed) cmp and
1873 * incorporating the result into ofmask would also work).
1874 * For the later elements, this just wraps around the indices, which
1875 * is apparently ok...
1877 true_index_array
= lp_build_min(&bldivec
, true_index_array
, fetch_max
);
1882 * Note: you'd expect some comparison/clamp against fetch_elt_max
1884 * There used to be one here but it was incorrect: overflow was
1885 * detected if index > fetch_elt_max - but the correct condition
1886 * would be index >= fetch_elt_max (since this is just size of elts
1887 * buffer / element size).
1888 * Using the correct condition however will cause failures - due to
1889 * vsplit/vcache code which rebases indices. So, as an example, if
1890 * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1891 * replace all invalid indices with 0 - which in case of elt_bias
1892 * not being zero will get a different fetch index than the valid
1893 * index 0. So, just rely on vsplit code preventing out-of-bounds
1894 * fetches. This is also why it's safe to do elts fetch even if there
1895 * was no index buffer bound - the real buffer is never seen here.
1899 * XXX should not have to do this, as scale can be handled
1900 * natively by loads (hits asserts though).
1902 true_index_array
= lp_build_shl_imm(&blduivec
, true_index_array
, 2);
1903 fetch_elts
= LLVMBuildBitCast(builder
, fetch_elts
,
1904 LLVMPointerType(LLVMInt8TypeInContext(context
),
1906 true_index_array
= lp_build_gather(gallivm
, vs_type
.length
,
1908 fetch_elts
, true_index_array
,
1912 for (j
= 0; j
< key
->nr_vertex_elements
; ++j
) {
1913 struct pipe_vertex_element
*velem
= &key
->vertex_element
[j
];
1914 const struct util_format_description
*format_desc
=
1915 util_format_description(velem
->src_format
);
1917 if (format_desc
->format
== PIPE_FORMAT_NONE
) {
1918 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1919 inputs
[j
][i
] = lp_build_zero(gallivm
, vs_type
);
1922 else if (velem
->instance_divisor
) {
1923 fetch_instanced(gallivm
, format_desc
, vs_type
,
1924 vb_stride
[j
], map_ptr
[j
],
1925 buffer_size_adj
[j
], ofmask
[j
],
1926 inputs
[j
], instance_index
[j
]);
1929 fetch_vector(gallivm
, format_desc
, vs_type
,
1930 vb_stride
[j
], map_ptr
[j
],
1931 buffer_size_adj
[j
], ofmask
[j
],
1932 inputs
[j
], true_index_array
);
1936 /* In the paths with elts vertex id has to be unaffected by the
1937 * index bias and because indices inside our elements array have
1938 * already had index bias applied we need to subtract it here to
1939 * get back to the original index.
1940 * in the linear paths vertex id has to be unaffected by the
1941 * original start index and because we abuse the 'start' variable
1942 * to either represent the actual start index or the index at which
1943 * the primitive was split (we split rendering into chunks of at
1944 * most 4095-vertices) we need to back out the original start
1945 * index out of our vertex id here.
1947 system_values
.basevertex
= lp_build_broadcast_scalar(&blduivec
,
1949 system_values
.vertex_id
= true_index_array
;
1950 system_values
.vertex_id_nobase
= LLVMBuildSub(builder
, true_index_array
,
1951 system_values
.basevertex
, "");
1953 ptr_aos
= (const LLVMValueRef (*)[TGSI_NUM_CHANNELS
]) inputs
;
1954 generate_vs(variant
,
1962 key
->clamp_vertex_color
);
1964 if (pos
!= -1 && cv
!= -1) {
1965 /* store original positions in clip before further manipulation */
1966 store_clip(gallivm
, vs_type
, io
, outputs
, pos
);
1969 if (enable_cliptest
) {
1970 LLVMValueRef temp
= LLVMBuildLoad(builder
, clipmask_bool_ptr
, "");
1971 /* allocate clipmask, assign it integer type */
1972 clipmask
= generate_clipmask(llvm
,
1977 context_ptr
, &have_clipdist
);
1978 temp
= LLVMBuildOr(builder
, clipmask
, temp
, "");
1979 /* store temporary clipping boolean value */
1980 LLVMBuildStore(builder
, temp
, clipmask_bool_ptr
);
1983 clipmask
= blduivec
.zero
;
1986 /* do viewport mapping */
1987 if (!bypass_viewport
) {
1988 generate_viewport(variant
, builder
, vs_type
, outputs
, context_ptr
);
1992 clipmask
= blduivec
.zero
;
1995 /* store clipmask in vertex header,
1996 * original positions in clip
1997 * and transformed positions in data
1999 convert_to_aos(gallivm
, io
, NULL
, outputs
, clipmask
,
2000 vs_info
->num_outputs
, vs_type
,
2001 enable_cliptest
&& key
->need_edgeflags
);
2003 lp_build_loop_end_cond(&lp_loop
, count
, step
, LLVMIntUGE
);
2005 sampler
->destroy(sampler
);
2007 /* return clipping boolean value for function */
2008 ret
= clipmask_booli32(gallivm
, vs_type
, clipmask_bool_ptr
,
2009 enable_cliptest
&& key
->need_edgeflags
);
2011 LLVMBuildRet(builder
, ret
);
2013 gallivm_verify_function(gallivm
, variant_func
);
2017 struct draw_llvm_variant_key
*
2018 draw_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2021 struct draw_llvm_variant_key
*key
;
2022 struct draw_sampler_static_state
*draw_sampler
;
2024 key
= (struct draw_llvm_variant_key
*)store
;
2026 memset(key
, 0, offsetof(struct draw_llvm_variant_key
, vertex_element
[0]));
2028 key
->clamp_vertex_color
= llvm
->draw
->rasterizer
->clamp_vertex_color
; /**/
2030 /* will have to rig this up properly later */
2031 key
->clip_xy
= llvm
->draw
->clip_xy
;
2032 key
->clip_z
= llvm
->draw
->clip_z
;
2033 key
->clip_user
= llvm
->draw
->clip_user
;
2034 key
->bypass_viewport
= llvm
->draw
->bypass_viewport
;
2035 key
->clip_halfz
= llvm
->draw
->rasterizer
->clip_halfz
;
2036 /* XXX assumes edgeflag output not at 0 */
2037 key
->need_edgeflags
= (llvm
->draw
->vs
.edgeflag_output
? TRUE
: FALSE
);
2038 key
->ucp_enable
= llvm
->draw
->rasterizer
->clip_plane_enable
;
2039 key
->has_gs
= llvm
->draw
->gs
.geometry_shader
!= NULL
;
2040 key
->num_outputs
= draw_total_vs_outputs(llvm
->draw
);
2042 /* All variants of this shader will have the same value for
2043 * nr_samplers. Not yet trying to compact away holes in the
2046 key
->nr_samplers
= llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2047 if (llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2048 key
->nr_sampler_views
=
2049 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2052 key
->nr_sampler_views
= key
->nr_samplers
;
2055 /* Presumably all variants of the shader should have the same
2056 * number of vertex elements - ie the number of shader inputs.
2057 * NOTE: we NEED to store the needed number of needed inputs
2058 * here, not the number of provided elements to match keysize
2059 * (and the offset of sampler state in the key).
2060 * If we have excess number of vertex elements, this is valid,
2061 * but the excess ones don't matter.
2062 * If we don't have enough vertex elements (which looks not really
2063 * valid but we'll handle it gracefully) fill out missing ones with
2064 * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2066 key
->nr_vertex_elements
=
2067 llvm
->draw
->vs
.vertex_shader
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
2069 if (llvm
->draw
->pt
.nr_vertex_elements
< key
->nr_vertex_elements
) {
2070 debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2071 key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
);
2072 memset(key
->vertex_element
, 0,
2073 sizeof(struct pipe_vertex_element
) * key
->nr_vertex_elements
);
2075 memcpy(key
->vertex_element
,
2076 llvm
->draw
->pt
.vertex_element
,
2077 sizeof(struct pipe_vertex_element
) *
2078 MIN2(key
->nr_vertex_elements
, llvm
->draw
->pt
.nr_vertex_elements
));
2080 draw_sampler
= draw_llvm_variant_key_samplers(key
);
2081 memset(draw_sampler
, 0,
2082 MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
2084 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
2085 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
2086 llvm
->draw
->samplers
[PIPE_SHADER_VERTEX
][i
]);
2088 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2089 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
2090 llvm
->draw
->sampler_views
[PIPE_SHADER_VERTEX
][i
]);
2098 draw_llvm_dump_variant_key(struct draw_llvm_variant_key
*key
)
2101 struct draw_sampler_static_state
*sampler
= draw_llvm_variant_key_samplers(key
);
2103 debug_printf("clamp_vertex_color = %u\n", key
->clamp_vertex_color
);
2104 debug_printf("clip_xy = %u\n", key
->clip_xy
);
2105 debug_printf("clip_z = %u\n", key
->clip_z
);
2106 debug_printf("clip_user = %u\n", key
->clip_user
);
2107 debug_printf("bypass_viewport = %u\n", key
->bypass_viewport
);
2108 debug_printf("clip_halfz = %u\n", key
->clip_halfz
);
2109 debug_printf("need_edgeflags = %u\n", key
->need_edgeflags
);
2110 debug_printf("has_gs = %u\n", key
->has_gs
);
2111 debug_printf("ucp_enable = %u\n", key
->ucp_enable
);
2113 for (i
= 0 ; i
< key
->nr_vertex_elements
; i
++) {
2114 debug_printf("vertex_element[%i].src_offset = %u\n", i
, key
->vertex_element
[i
].src_offset
);
2115 debug_printf("vertex_element[%i].instance_divisor = %u\n", i
, key
->vertex_element
[i
].instance_divisor
);
2116 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i
, key
->vertex_element
[i
].vertex_buffer_index
);
2117 debug_printf("vertex_element[%i].src_format = %s\n", i
, util_format_name(key
->vertex_element
[i
].src_format
));
2120 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2121 debug_printf("sampler[%i].src_format = %s\n", i
, util_format_name(sampler
[i
].texture_state
.format
));
2127 draw_llvm_set_mapped_texture(struct draw_context
*draw
,
2128 unsigned shader_stage
,
2130 uint32_t width
, uint32_t height
, uint32_t depth
,
2131 uint32_t first_level
, uint32_t last_level
,
2132 const void *base_ptr
,
2133 uint32_t row_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2134 uint32_t img_stride
[PIPE_MAX_TEXTURE_LEVELS
],
2135 uint32_t mip_offsets
[PIPE_MAX_TEXTURE_LEVELS
])
2138 struct draw_jit_texture
*jit_tex
;
2140 assert(shader_stage
== PIPE_SHADER_VERTEX
||
2141 shader_stage
== PIPE_SHADER_GEOMETRY
);
2143 if (shader_stage
== PIPE_SHADER_VERTEX
) {
2144 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->jit_context
.textures
));
2146 jit_tex
= &draw
->llvm
->jit_context
.textures
[sview_idx
];
2147 } else if (shader_stage
== PIPE_SHADER_GEOMETRY
) {
2148 assert(sview_idx
< ARRAY_SIZE(draw
->llvm
->gs_jit_context
.textures
));
2150 jit_tex
= &draw
->llvm
->gs_jit_context
.textures
[sview_idx
];
2156 jit_tex
->width
= width
;
2157 jit_tex
->height
= height
;
2158 jit_tex
->depth
= depth
;
2159 jit_tex
->first_level
= first_level
;
2160 jit_tex
->last_level
= last_level
;
2161 jit_tex
->base
= base_ptr
;
2163 for (j
= first_level
; j
<= last_level
; j
++) {
2164 jit_tex
->mip_offsets
[j
] = mip_offsets
[j
];
2165 jit_tex
->row_stride
[j
] = row_stride
[j
];
2166 jit_tex
->img_stride
[j
] = img_stride
[j
];
2172 draw_llvm_set_sampler_state(struct draw_context
*draw
,
2173 unsigned shader_type
)
2177 if (shader_type
== PIPE_SHADER_VERTEX
) {
2178 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_VERTEX
]; i
++) {
2179 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->jit_context
.samplers
[i
];
2181 if (draw
->samplers
[PIPE_SHADER_VERTEX
][i
]) {
2182 const struct pipe_sampler_state
*s
2183 = draw
->samplers
[PIPE_SHADER_VERTEX
][i
];
2184 jit_sam
->min_lod
= s
->min_lod
;
2185 jit_sam
->max_lod
= s
->max_lod
;
2186 jit_sam
->lod_bias
= s
->lod_bias
;
2187 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2190 } else if (shader_type
== PIPE_SHADER_GEOMETRY
) {
2191 for (i
= 0; i
< draw
->num_samplers
[PIPE_SHADER_GEOMETRY
]; i
++) {
2192 struct draw_jit_sampler
*jit_sam
= &draw
->llvm
->gs_jit_context
.samplers
[i
];
2194 if (draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]) {
2195 const struct pipe_sampler_state
*s
2196 = draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
];
2197 jit_sam
->min_lod
= s
->min_lod
;
2198 jit_sam
->max_lod
= s
->max_lod
;
2199 jit_sam
->lod_bias
= s
->lod_bias
;
2200 COPY_4V(jit_sam
->border_color
, s
->border_color
.f
);
2208 draw_llvm_destroy_variant(struct draw_llvm_variant
*variant
)
2210 struct draw_llvm
*llvm
= variant
->llvm
;
2212 gallivm_destroy(variant
->gallivm
);
2214 remove_from_list(&variant
->list_item_local
);
2215 variant
->shader
->variants_cached
--;
2216 remove_from_list(&variant
->list_item_global
);
2217 llvm
->nr_variants
--;
2223 * Create LLVM types for various structures.
2226 create_gs_jit_types(struct draw_gs_llvm_variant
*var
)
2228 struct gallivm_state
*gallivm
= var
->gallivm
;
2229 LLVMTypeRef texture_type
, sampler_type
, context_type
;
2231 texture_type
= create_jit_texture_type(gallivm
, "texture");
2232 sampler_type
= create_jit_sampler_type(gallivm
, "sampler");
2234 context_type
= create_gs_jit_context_type(gallivm
,
2235 var
->shader
->base
.vector_length
,
2236 texture_type
, sampler_type
,
2237 "draw_gs_jit_context");
2238 var
->context_ptr_type
= LLVMPointerType(context_type
, 0);
2240 var
->input_array_type
= create_gs_jit_input_type(gallivm
);
2244 get_gs_context_ptr_type(struct draw_gs_llvm_variant
*variant
)
2246 if (!variant
->context_ptr_type
)
2247 create_gs_jit_types(variant
);
2248 return variant
->context_ptr_type
;
2252 generate_mask_value(struct draw_gs_llvm_variant
*variant
,
2253 struct lp_type gs_type
)
2255 struct gallivm_state
*gallivm
= variant
->gallivm
;
2256 LLVMBuilderRef builder
= gallivm
->builder
;
2257 struct lp_type mask_type
= lp_int_type(gs_type
);
2258 LLVMValueRef num_prims
;
2259 LLVMValueRef mask_val
= lp_build_const_vec(gallivm
, mask_type
, 0);
2262 num_prims
= lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, mask_type
),
2263 variant
->num_prims
);
2264 for (i
= 0; i
< gs_type
.length
; i
++) {
2265 LLVMValueRef idx
= lp_build_const_int32(gallivm
, i
);
2266 mask_val
= LLVMBuildInsertElement(builder
, mask_val
, idx
, idx
, "");
2268 mask_val
= lp_build_compare(gallivm
, mask_type
,
2269 PIPE_FUNC_GREATER
, num_prims
, mask_val
);
2275 draw_gs_llvm_generate(struct draw_llvm
*llvm
,
2276 struct draw_gs_llvm_variant
*variant
)
2278 struct gallivm_state
*gallivm
= variant
->gallivm
;
2279 LLVMContextRef context
= gallivm
->context
;
2280 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(context
);
2281 LLVMTypeRef arg_types
[7];
2282 LLVMTypeRef func_type
;
2283 LLVMValueRef variant_func
;
2284 LLVMValueRef context_ptr
;
2285 LLVMValueRef prim_id_ptr
;
2286 LLVMBasicBlockRef block
;
2287 LLVMBuilderRef builder
;
2288 LLVMValueRef io_ptr
, input_array
, num_prims
, mask_val
;
2289 struct lp_build_sampler_soa
*sampler
= 0;
2290 struct lp_build_context bld
;
2291 struct lp_bld_tgsi_system_values system_values
;
2293 struct lp_type gs_type
;
2295 struct draw_gs_llvm_iface gs_iface
;
2296 const struct tgsi_token
*tokens
= variant
->shader
->base
.state
.tokens
;
2297 LLVMValueRef consts_ptr
, num_consts_ptr
;
2298 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
2299 struct lp_build_mask_context mask
;
2300 const struct tgsi_shader_info
*gs_info
= &variant
->shader
->base
.info
;
2301 unsigned vector_length
= variant
->shader
->base
.vector_length
;
2303 memset(&system_values
, 0, sizeof(system_values
));
2305 util_snprintf(func_name
, sizeof(func_name
), "draw_llvm_gs_variant%u",
2306 variant
->shader
->variants_cached
);
2308 assert(variant
->vertex_header_ptr_type
);
2310 arg_types
[0] = get_gs_context_ptr_type(variant
); /* context */
2311 arg_types
[1] = variant
->input_array_type
; /* input */
2312 arg_types
[2] = variant
->vertex_header_ptr_type
; /* vertex_header */
2313 arg_types
[3] = int32_type
; /* num_prims */
2314 arg_types
[4] = int32_type
; /* instance_id */
2315 arg_types
[5] = LLVMPointerType(
2316 LLVMVectorType(int32_type
, vector_length
), 0); /* prim_id_ptr */
2317 arg_types
[6] = int32_type
;
2319 func_type
= LLVMFunctionType(int32_type
, arg_types
, ARRAY_SIZE(arg_types
), 0);
2321 variant_func
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2323 variant
->function
= variant_func
;
2325 LLVMSetFunctionCallConv(variant_func
, LLVMCCallConv
);
2327 for (i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
)
2328 if (LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2329 lp_add_function_attr(variant_func
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
2331 context_ptr
= LLVMGetParam(variant_func
, 0);
2332 input_array
= LLVMGetParam(variant_func
, 1);
2333 io_ptr
= LLVMGetParam(variant_func
, 2);
2334 num_prims
= LLVMGetParam(variant_func
, 3);
2335 system_values
.instance_id
= LLVMGetParam(variant_func
, 4);
2336 prim_id_ptr
= LLVMGetParam(variant_func
, 5);
2337 system_values
.invocation_id
= LLVMGetParam(variant_func
, 6);
2339 lp_build_name(context_ptr
, "context");
2340 lp_build_name(input_array
, "input");
2341 lp_build_name(io_ptr
, "io");
2342 lp_build_name(num_prims
, "num_prims");
2343 lp_build_name(system_values
.instance_id
, "instance_id");
2344 lp_build_name(prim_id_ptr
, "prim_id_ptr");
2345 lp_build_name(system_values
.invocation_id
, "invocation_id");
2347 variant
->context_ptr
= context_ptr
;
2348 variant
->io_ptr
= io_ptr
;
2349 variant
->num_prims
= num_prims
;
2351 gs_iface
.base
.fetch_input
= draw_gs_llvm_fetch_input
;
2352 gs_iface
.base
.emit_vertex
= draw_gs_llvm_emit_vertex
;
2353 gs_iface
.base
.end_primitive
= draw_gs_llvm_end_primitive
;
2354 gs_iface
.base
.gs_epilogue
= draw_gs_llvm_epilogue
;
2355 gs_iface
.input
= input_array
;
2356 gs_iface
.variant
= variant
;
2362 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, variant_func
, "entry");
2363 builder
= gallivm
->builder
;
2364 LLVMPositionBuilderAtEnd(builder
, block
);
2366 lp_build_context_init(&bld
, gallivm
, lp_type_int(32));
2368 memset(&gs_type
, 0, sizeof gs_type
);
2369 gs_type
.floating
= TRUE
; /* floating point values */
2370 gs_type
.sign
= TRUE
; /* values are signed */
2371 gs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2372 gs_type
.width
= 32; /* 32-bit float */
2373 gs_type
.length
= vector_length
;
2375 consts_ptr
= draw_gs_jit_context_constants(variant
->gallivm
, context_ptr
);
2377 draw_gs_jit_context_num_constants(variant
->gallivm
, context_ptr
);
2379 /* code generated texture sampling */
2380 sampler
= draw_llvm_sampler_soa_create(variant
->key
.samplers
);
2382 mask_val
= generate_mask_value(variant
, gs_type
);
2383 lp_build_mask_begin(&mask
, gallivm
, gs_type
, mask_val
);
2385 if (gs_info
->uses_primid
) {
2386 system_values
.prim_id
= LLVMBuildLoad(builder
, prim_id_ptr
, "prim_id");
2389 if (gallivm_debug
& (GALLIVM_DEBUG_TGSI
| GALLIVM_DEBUG_IR
)) {
2390 tgsi_dump(tokens
, 0);
2391 draw_gs_llvm_dump_variant_key(&variant
->key
);
2394 lp_build_tgsi_soa(variant
->gallivm
,
2406 &llvm
->draw
->gs
.geometry_shader
->info
,
2407 (const struct lp_build_tgsi_gs_iface
*)&gs_iface
);
2409 sampler
->destroy(sampler
);
2411 lp_build_mask_end(&mask
);
2413 LLVMBuildRet(builder
, lp_build_zero(gallivm
, lp_type_uint(32)));
2415 gallivm_verify_function(gallivm
, variant_func
);
2419 struct draw_gs_llvm_variant
*
2420 draw_gs_llvm_create_variant(struct draw_llvm
*llvm
,
2421 unsigned num_outputs
,
2422 const struct draw_gs_llvm_variant_key
*key
)
2424 struct draw_gs_llvm_variant
*variant
;
2425 struct llvm_geometry_shader
*shader
=
2426 llvm_geometry_shader(llvm
->draw
->gs
.geometry_shader
);
2427 LLVMTypeRef vertex_header
;
2428 char module_name
[64];
2430 variant
= MALLOC(sizeof *variant
+
2431 shader
->variant_key_size
-
2432 sizeof variant
->key
);
2436 variant
->llvm
= llvm
;
2437 variant
->shader
= shader
;
2439 util_snprintf(module_name
, sizeof(module_name
), "draw_llvm_gs_variant%u",
2440 variant
->shader
->variants_cached
);
2442 variant
->gallivm
= gallivm_create(module_name
, llvm
->context
);
2444 create_gs_jit_types(variant
);
2446 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2448 vertex_header
= create_jit_vertex_header(variant
->gallivm
, num_outputs
);
2450 variant
->vertex_header_ptr_type
= LLVMPointerType(vertex_header
, 0);
2452 draw_gs_llvm_generate(llvm
, variant
);
2454 gallivm_compile_module(variant
->gallivm
);
2456 variant
->jit_func
= (draw_gs_jit_func
)
2457 gallivm_jit_function(variant
->gallivm
, variant
->function
);
2459 gallivm_free_ir(variant
->gallivm
);
2461 variant
->list_item_global
.base
= variant
;
2462 variant
->list_item_local
.base
= variant
;
2463 /*variant->no = */shader
->variants_created
++;
2464 variant
->list_item_global
.base
= variant
;
2470 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant
*variant
)
2472 struct draw_llvm
*llvm
= variant
->llvm
;
2474 gallivm_destroy(variant
->gallivm
);
2476 remove_from_list(&variant
->list_item_local
);
2477 variant
->shader
->variants_cached
--;
2478 remove_from_list(&variant
->list_item_global
);
2479 llvm
->nr_gs_variants
--;
2483 struct draw_gs_llvm_variant_key
*
2484 draw_gs_llvm_make_variant_key(struct draw_llvm
*llvm
, char *store
)
2487 struct draw_gs_llvm_variant_key
*key
;
2488 struct draw_sampler_static_state
*draw_sampler
;
2490 key
= (struct draw_gs_llvm_variant_key
*)store
;
2492 memset(key
, 0, offsetof(struct draw_gs_llvm_variant_key
, samplers
[0]));
2494 key
->num_outputs
= draw_total_gs_outputs(llvm
->draw
);
2496 /* All variants of this shader will have the same value for
2497 * nr_samplers. Not yet trying to compact away holes in the
2500 key
->nr_samplers
= llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2501 if (llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2502 key
->nr_sampler_views
=
2503 llvm
->draw
->gs
.geometry_shader
->info
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2506 key
->nr_sampler_views
= key
->nr_samplers
;
2509 draw_sampler
= key
->samplers
;
2511 memset(draw_sampler
, 0, MAX2(key
->nr_samplers
, key
->nr_sampler_views
) * sizeof *draw_sampler
);
2513 for (i
= 0 ; i
< key
->nr_samplers
; i
++) {
2514 lp_sampler_static_sampler_state(&draw_sampler
[i
].sampler_state
,
2515 llvm
->draw
->samplers
[PIPE_SHADER_GEOMETRY
][i
]);
2517 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2518 lp_sampler_static_texture_state(&draw_sampler
[i
].texture_state
,
2519 llvm
->draw
->sampler_views
[PIPE_SHADER_GEOMETRY
][i
]);
2526 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key
*key
)
2529 struct draw_sampler_static_state
*sampler
= key
->samplers
;
2531 for (i
= 0 ; i
< key
->nr_sampler_views
; i
++) {
2532 debug_printf("sampler[%i].src_format = %s\n", i
,
2533 util_format_name(sampler
[i
].texture_state
.format
));