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 TUNGSTEN GRAPHICS 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 "util/u_math.h"
29 #include "util/u_memory.h"
30 #include "util/u_prim.h"
31 #include "draw/draw_context.h"
32 #include "draw/draw_gs.h"
33 #include "draw/draw_vbuf.h"
34 #include "draw/draw_vertex.h"
35 #include "draw/draw_pt.h"
36 #include "draw/draw_prim_assembler.h"
37 #include "draw/draw_vs.h"
38 #include "draw/draw_llvm.h"
39 #include "gallivm/lp_bld_init.h"
42 struct llvm_middle_end
{
43 struct draw_pt_middle_end base
;
44 struct draw_context
*draw
;
47 struct pt_so_emit
*so_emit
;
48 struct pt_fetch
*fetch
;
49 struct pt_post_vs
*post_vs
;
52 unsigned vertex_data_offset
;
57 struct draw_llvm
*llvm
;
58 struct draw_llvm_variant
*current_variant
;
63 llvm_middle_end_prepare_gs(struct llvm_middle_end
*fpme
)
65 struct draw_context
*draw
= fpme
->draw
;
66 struct draw_geometry_shader
*gs
= draw
->gs
.geometry_shader
;
67 struct draw_gs_llvm_variant_key
*key
;
68 struct draw_gs_llvm_variant
*variant
= NULL
;
69 struct draw_gs_llvm_variant_list_item
*li
;
70 struct llvm_geometry_shader
*shader
= llvm_geometry_shader(gs
);
71 char store
[DRAW_GS_LLVM_MAX_VARIANT_KEY_SIZE
];
74 key
= draw_gs_llvm_make_variant_key(fpme
->llvm
, store
);
76 /* Search shader's list of variants for the key */
77 li
= first_elem(&shader
->variants
);
78 while (!at_end(&shader
->variants
, li
)) {
79 if (memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
87 /* found the variant, move to head of global list (for LRU) */
88 move_to_head(&fpme
->llvm
->gs_variants_list
,
89 &variant
->list_item_global
);
92 /* Need to create new variant */
94 /* First check if we've created too many variants. If so, free
95 * 25% of the LRU to avoid using too much memory.
97 if (fpme
->llvm
->nr_gs_variants
>= DRAW_MAX_SHADER_VARIANTS
) {
99 * XXX: should we flush here ?
101 for (i
= 0; i
< DRAW_MAX_SHADER_VARIANTS
/ 4; i
++) {
102 struct draw_gs_llvm_variant_list_item
*item
;
103 if (is_empty_list(&fpme
->llvm
->gs_variants_list
)) {
106 item
= last_elem(&fpme
->llvm
->gs_variants_list
);
109 draw_gs_llvm_destroy_variant(item
->base
);
113 variant
= draw_gs_llvm_create_variant(fpme
->llvm
, gs
->info
.num_outputs
, key
);
116 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
117 insert_at_head(&fpme
->llvm
->gs_variants_list
,
118 &variant
->list_item_global
);
119 fpme
->llvm
->nr_gs_variants
++;
120 shader
->variants_cached
++;
124 gs
->current_variant
= variant
;
128 * Prepare/validate middle part of the vertex pipeline.
129 * NOTE: if you change this function, also look at the non-LLVM
130 * function fetch_pipeline_prepare() for similar changes.
133 llvm_middle_end_prepare( struct draw_pt_middle_end
*middle
,
136 unsigned *max_vertices
)
138 struct llvm_middle_end
*fpme
= (struct llvm_middle_end
*)middle
;
139 struct draw_context
*draw
= fpme
->draw
;
140 struct draw_vertex_shader
*vs
= draw
->vs
.vertex_shader
;
141 struct draw_geometry_shader
*gs
= draw
->gs
.geometry_shader
;
142 const unsigned out_prim
= gs
? gs
->output_primitive
:
143 u_assembled_prim(in_prim
);
145 /* Add one to num_outputs because the pipeline occasionally tags on
146 * an additional texcoord, eg for AA lines.
148 const unsigned nr
= MAX2( vs
->info
.num_inputs
,
149 vs
->info
.num_outputs
+ 1 );
151 fpme
->input_prim
= in_prim
;
154 /* Always leave room for the vertex header whether we need it or
155 * not. It's hard to get rid of it in particular because of the
156 * viewport code in draw_pt_post_vs.c.
158 fpme
->vertex_size
= sizeof(struct vertex_header
) + nr
* 4 * sizeof(float);
161 draw_pt_post_vs_prepare( fpme
->post_vs
,
166 draw
->identity_viewport
,
167 draw
->rasterizer
->clip_halfz
,
168 (draw
->vs
.edgeflag_output
? TRUE
: FALSE
) );
170 draw_pt_so_emit_prepare( fpme
->so_emit
, gs
== NULL
);
172 if (!(opt
& PT_PIPELINE
)) {
173 draw_pt_emit_prepare( fpme
->emit
,
177 *max_vertices
= MAX2( *max_vertices
, 4096 );
180 /* limit max fetches by limiting max_vertices */
181 *max_vertices
= 4096;
184 /* return even number */
185 *max_vertices
= *max_vertices
& ~1;
187 /* Find/create the vertex shader variant */
189 struct draw_llvm_variant_key
*key
;
190 struct draw_llvm_variant
*variant
= NULL
;
191 struct draw_llvm_variant_list_item
*li
;
192 struct llvm_vertex_shader
*shader
= llvm_vertex_shader(vs
);
193 char store
[DRAW_LLVM_MAX_VARIANT_KEY_SIZE
];
196 key
= draw_llvm_make_variant_key(fpme
->llvm
, store
);
198 /* Search shader's list of variants for the key */
199 li
= first_elem(&shader
->variants
);
200 while (!at_end(&shader
->variants
, li
)) {
201 if (memcmp(&li
->base
->key
, key
, shader
->variant_key_size
) == 0) {
209 /* found the variant, move to head of global list (for LRU) */
210 move_to_head(&fpme
->llvm
->vs_variants_list
,
211 &variant
->list_item_global
);
214 /* Need to create new variant */
216 /* First check if we've created too many variants. If so, free
217 * 25% of the LRU to avoid using too much memory.
219 if (fpme
->llvm
->nr_variants
>= DRAW_MAX_SHADER_VARIANTS
) {
221 * XXX: should we flush here ?
223 for (i
= 0; i
< DRAW_MAX_SHADER_VARIANTS
/ 4; i
++) {
224 struct draw_llvm_variant_list_item
*item
;
225 if (is_empty_list(&fpme
->llvm
->vs_variants_list
)) {
228 item
= last_elem(&fpme
->llvm
->vs_variants_list
);
231 draw_llvm_destroy_variant(item
->base
);
235 variant
= draw_llvm_create_variant(fpme
->llvm
, nr
, key
);
238 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
239 insert_at_head(&fpme
->llvm
->vs_variants_list
,
240 &variant
->list_item_global
);
241 fpme
->llvm
->nr_variants
++;
242 shader
->variants_cached
++;
246 fpme
->current_variant
= variant
;
250 llvm_middle_end_prepare_gs(fpme
);
256 * Bind/update constant buffer pointers, clip planes and viewport dims.
257 * These are "light weight" parameters which aren't baked into the
258 * generated code. Updating these items is much cheaper than revalidating
259 * and rebuilding the generated pipeline code.
262 llvm_middle_end_bind_parameters(struct draw_pt_middle_end
*middle
)
264 struct llvm_middle_end
*fpme
= (struct llvm_middle_end
*)middle
;
265 struct draw_context
*draw
= fpme
->draw
;
268 for (i
= 0; i
< Elements(fpme
->llvm
->jit_context
.vs_constants
); ++i
) {
269 fpme
->llvm
->jit_context
.vs_constants
[i
] = draw
->pt
.user
.vs_constants
[i
];
271 for (i
= 0; i
< Elements(fpme
->llvm
->gs_jit_context
.constants
); ++i
) {
272 fpme
->llvm
->gs_jit_context
.constants
[i
] = draw
->pt
.user
.gs_constants
[i
];
275 fpme
->llvm
->jit_context
.planes
=
276 (float (*)[DRAW_TOTAL_CLIP_PLANES
][4]) draw
->pt
.user
.planes
[0];
277 fpme
->llvm
->gs_jit_context
.planes
=
278 (float (*)[DRAW_TOTAL_CLIP_PLANES
][4]) draw
->pt
.user
.planes
[0];
280 fpme
->llvm
->jit_context
.viewport
= (float *) draw
->viewport
.scale
;
281 fpme
->llvm
->gs_jit_context
.viewport
= (float *) draw
->viewport
.scale
;
285 static void pipeline(struct llvm_middle_end
*llvm
,
286 const struct draw_vertex_info
*vert_info
,
287 const struct draw_prim_info
*prim_info
)
289 if (prim_info
->linear
)
290 draw_pipeline_run_linear( llvm
->draw
,
294 draw_pipeline_run( llvm
->draw
,
299 static void emit(struct pt_emit
*emit
,
300 const struct draw_vertex_info
*vert_info
,
301 const struct draw_prim_info
*prim_info
)
303 if (prim_info
->linear
) {
304 draw_pt_emit_linear(emit
, vert_info
, prim_info
);
307 draw_pt_emit(emit
, vert_info
, prim_info
);
312 llvm_pipeline_generic( struct draw_pt_middle_end
*middle
,
313 const struct draw_fetch_info
*fetch_info
,
314 const struct draw_prim_info
*in_prim_info
)
316 struct llvm_middle_end
*fpme
= (struct llvm_middle_end
*)middle
;
317 struct draw_context
*draw
= fpme
->draw
;
318 struct draw_geometry_shader
*gshader
= draw
->gs
.geometry_shader
;
319 struct draw_prim_info gs_prim_info
;
320 struct draw_vertex_info llvm_vert_info
;
321 struct draw_vertex_info gs_vert_info
;
322 struct draw_vertex_info
*vert_info
;
323 struct draw_prim_info ia_prim_info
;
324 struct draw_vertex_info ia_vert_info
;
325 const struct draw_prim_info
*prim_info
= in_prim_info
;
326 boolean free_prim_info
= FALSE
;
327 unsigned opt
= fpme
->opt
;
328 unsigned clipped
= 0;
330 llvm_vert_info
.count
= fetch_info
->count
;
331 llvm_vert_info
.vertex_size
= fpme
->vertex_size
;
332 llvm_vert_info
.stride
= fpme
->vertex_size
;
333 llvm_vert_info
.verts
=
334 (struct vertex_header
*)MALLOC(fpme
->vertex_size
*
335 align(fetch_info
->count
, lp_native_vector_width
/ 32));
336 if (!llvm_vert_info
.verts
) {
341 if (draw
->collect_statistics
) {
342 draw
->statistics
.ia_vertices
+= fetch_info
->count
;
343 draw
->statistics
.ia_primitives
+=
344 u_decomposed_prims_for_vertices(prim_info
->prim
, prim_info
->count
);
345 draw
->statistics
.vs_invocations
+= fetch_info
->count
;
348 if (fetch_info
->linear
)
349 clipped
= fpme
->current_variant
->jit_func( &fpme
->llvm
->jit_context
,
350 llvm_vert_info
.verts
,
351 (const char **)draw
->pt
.user
.vbuffer
,
355 draw
->pt
.vertex_buffer
,
358 clipped
= fpme
->current_variant
->jit_func_elts( &fpme
->llvm
->jit_context
,
359 llvm_vert_info
.verts
,
360 (const char **)draw
->pt
.user
.vbuffer
,
364 draw
->pt
.vertex_buffer
,
367 /* Finished with fetch and vs:
370 vert_info
= &llvm_vert_info
;
373 if ((opt
& PT_SHADE
) && gshader
) {
374 struct draw_vertex_shader
*vshader
= draw
->vs
.vertex_shader
;
375 draw_geometry_shader_run(gshader
,
376 draw
->pt
.user
.gs_constants
,
377 draw
->pt
.user
.gs_constants_size
,
384 FREE(vert_info
->verts
);
385 vert_info
= &gs_vert_info
;
386 prim_info
= &gs_prim_info
;
388 if (draw_prim_assembler_is_required(draw
, prim_info
, vert_info
)) {
389 draw_prim_assembler_run(draw
, prim_info
, vert_info
,
390 &ia_prim_info
, &ia_vert_info
);
392 if (ia_vert_info
.count
) {
393 FREE(vert_info
->verts
);
394 vert_info
= &ia_vert_info
;
395 prim_info
= &ia_prim_info
;
396 free_prim_info
= TRUE
;
400 if (prim_info
->count
== 0) {
401 debug_printf("GS/IA didn't emit any vertices!\n");
403 FREE(vert_info
->verts
);
404 if (free_prim_info
) {
405 FREE(prim_info
->primitive_lengths
);
410 /* stream output needs to be done before clipping */
411 draw_pt_so_emit( fpme
->so_emit
, vert_info
, prim_info
);
414 * if there's no position, need to stop now, or the latter stages
415 * will try to access non-existent position output.
417 if (draw_current_shader_position_output(draw
) != -1) {
418 if ((opt
& PT_SHADE
) && gshader
) {
419 clipped
= draw_pt_post_vs_run( fpme
->post_vs
, vert_info
);
425 /* Do we need to run the pipeline? Now will come here if clipped
427 if (opt
& PT_PIPELINE
) {
428 pipeline( fpme
, vert_info
, prim_info
);
431 emit( fpme
->emit
, vert_info
, prim_info
);
434 FREE(vert_info
->verts
);
435 if (free_prim_info
) {
436 FREE(prim_info
->primitive_lengths
);
441 static void llvm_middle_end_run( struct draw_pt_middle_end
*middle
,
442 const unsigned *fetch_elts
,
443 unsigned fetch_count
,
444 const ushort
*draw_elts
,
446 unsigned prim_flags
)
448 struct llvm_middle_end
*fpme
= (struct llvm_middle_end
*)middle
;
449 struct draw_fetch_info fetch_info
;
450 struct draw_prim_info prim_info
;
452 fetch_info
.linear
= FALSE
;
453 fetch_info
.start
= 0;
454 fetch_info
.elts
= fetch_elts
;
455 fetch_info
.count
= fetch_count
;
457 prim_info
.linear
= FALSE
;
459 prim_info
.count
= draw_count
;
460 prim_info
.elts
= draw_elts
;
461 prim_info
.prim
= fpme
->input_prim
;
462 prim_info
.flags
= prim_flags
;
463 prim_info
.primitive_count
= 1;
464 prim_info
.primitive_lengths
= &draw_count
;
466 llvm_pipeline_generic( middle
, &fetch_info
, &prim_info
);
470 static void llvm_middle_end_linear_run( struct draw_pt_middle_end
*middle
,
475 struct llvm_middle_end
*fpme
= (struct llvm_middle_end
*)middle
;
476 struct draw_fetch_info fetch_info
;
477 struct draw_prim_info prim_info
;
479 fetch_info
.linear
= TRUE
;
480 fetch_info
.start
= start
;
481 fetch_info
.count
= count
;
482 fetch_info
.elts
= NULL
;
484 prim_info
.linear
= TRUE
;
486 prim_info
.count
= count
;
487 prim_info
.elts
= NULL
;
488 prim_info
.prim
= fpme
->input_prim
;
489 prim_info
.flags
= prim_flags
;
490 prim_info
.primitive_count
= 1;
491 prim_info
.primitive_lengths
= &count
;
493 llvm_pipeline_generic( middle
, &fetch_info
, &prim_info
);
499 llvm_middle_end_linear_run_elts( struct draw_pt_middle_end
*middle
,
502 const ushort
*draw_elts
,
504 unsigned prim_flags
)
506 struct llvm_middle_end
*fpme
= (struct llvm_middle_end
*)middle
;
507 struct draw_fetch_info fetch_info
;
508 struct draw_prim_info prim_info
;
510 fetch_info
.linear
= TRUE
;
511 fetch_info
.start
= start
;
512 fetch_info
.count
= count
;
513 fetch_info
.elts
= NULL
;
515 prim_info
.linear
= FALSE
;
517 prim_info
.count
= draw_count
;
518 prim_info
.elts
= draw_elts
;
519 prim_info
.prim
= fpme
->input_prim
;
520 prim_info
.flags
= prim_flags
;
521 prim_info
.primitive_count
= 1;
522 prim_info
.primitive_lengths
= &draw_count
;
524 llvm_pipeline_generic( middle
, &fetch_info
, &prim_info
);
531 static void llvm_middle_end_finish( struct draw_pt_middle_end
*middle
)
536 static void llvm_middle_end_destroy( struct draw_pt_middle_end
*middle
)
538 struct llvm_middle_end
*fpme
= (struct llvm_middle_end
*)middle
;
541 draw_pt_fetch_destroy( fpme
->fetch
);
544 draw_pt_emit_destroy( fpme
->emit
);
547 draw_pt_so_emit_destroy( fpme
->so_emit
);
550 draw_pt_post_vs_destroy( fpme
->post_vs
);
556 struct draw_pt_middle_end
*
557 draw_pt_fetch_pipeline_or_emit_llvm(struct draw_context
*draw
)
559 struct llvm_middle_end
*fpme
= 0;
564 fpme
= CALLOC_STRUCT( llvm_middle_end
);
568 fpme
->base
.prepare
= llvm_middle_end_prepare
;
569 fpme
->base
.bind_parameters
= llvm_middle_end_bind_parameters
;
570 fpme
->base
.run
= llvm_middle_end_run
;
571 fpme
->base
.run_linear
= llvm_middle_end_linear_run
;
572 fpme
->base
.run_linear_elts
= llvm_middle_end_linear_run_elts
;
573 fpme
->base
.finish
= llvm_middle_end_finish
;
574 fpme
->base
.destroy
= llvm_middle_end_destroy
;
578 fpme
->fetch
= draw_pt_fetch_create( draw
);
582 fpme
->post_vs
= draw_pt_post_vs_create( draw
);
586 fpme
->emit
= draw_pt_emit_create( draw
);
590 fpme
->so_emit
= draw_pt_so_emit_create( draw
);
594 fpme
->llvm
= draw
->llvm
;
598 fpme
->current_variant
= NULL
;
604 llvm_middle_end_destroy( &fpme
->base
);