2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics to
4 develop this 3D driver.
6 Permission is hereby granted, free of charge, to any person obtaining
7 a 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, sublicense, 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
16 portions of the Software.
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **********************************************************************/
29 * Keith Whitwell <keithw@vmware.com>
33 #include "main/compiler.h"
34 #include "brw_context.h"
37 #include "brw_state.h"
38 #include "program/prog_print.h"
39 #include "program/prog_parameter.h"
41 #include "util/ralloc.h"
43 static inline void assign_vue_slot(struct brw_vue_map
*vue_map
,
46 /* Make sure this varying hasn't been assigned a slot already */
47 assert (vue_map
->varying_to_slot
[varying
] == -1);
49 vue_map
->varying_to_slot
[varying
] = vue_map
->num_slots
;
50 vue_map
->slot_to_varying
[vue_map
->num_slots
++] = varying
;
54 * Compute the VUE map for vertex shader program.
57 brw_compute_vue_map(struct brw_context
*brw
, struct brw_vue_map
*vue_map
,
58 GLbitfield64 slots_valid
)
60 vue_map
->slots_valid
= slots_valid
;
63 /* gl_Layer and gl_ViewportIndex don't get their own varying slots -- they
64 * are stored in the first VUE slot (VARYING_SLOT_PSIZ).
66 slots_valid
&= ~(VARYING_BIT_LAYER
| VARYING_BIT_VIEWPORT
);
68 /* Make sure that the values we store in vue_map->varying_to_slot and
69 * vue_map->slot_to_varying won't overflow the signed chars that are used
70 * to store them. Note that since vue_map->slot_to_varying sometimes holds
71 * values equal to BRW_VARYING_SLOT_COUNT, we need to ensure that
72 * BRW_VARYING_SLOT_COUNT is <= 127, not 128.
74 STATIC_ASSERT(BRW_VARYING_SLOT_COUNT
<= 127);
76 vue_map
->num_slots
= 0;
77 for (i
= 0; i
< BRW_VARYING_SLOT_COUNT
; ++i
) {
78 vue_map
->varying_to_slot
[i
] = -1;
79 vue_map
->slot_to_varying
[i
] = BRW_VARYING_SLOT_COUNT
;
82 /* VUE header: format depends on chip generation and whether clipping is
86 /* There are 8 dwords in VUE header pre-Ironlake:
87 * dword 0-3 is indices, point width, clip flags.
88 * dword 4-7 is ndc position
89 * dword 8-11 is the first vertex data.
91 * On Ironlake the VUE header is nominally 20 dwords, but the hardware
92 * will accept the same header layout as Gen4 [and should be a bit faster]
94 assign_vue_slot(vue_map
, VARYING_SLOT_PSIZ
);
95 assign_vue_slot(vue_map
, BRW_VARYING_SLOT_NDC
);
96 assign_vue_slot(vue_map
, VARYING_SLOT_POS
);
98 /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
99 * dword 0-3 of the header is indices, point width, clip flags.
100 * dword 4-7 is the 4D space position
101 * dword 8-15 of the vertex header is the user clip distance if
103 * dword 8-11 or 16-19 is the first vertex element data we fill.
105 assign_vue_slot(vue_map
, VARYING_SLOT_PSIZ
);
106 assign_vue_slot(vue_map
, VARYING_SLOT_POS
);
107 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0
))
108 assign_vue_slot(vue_map
, VARYING_SLOT_CLIP_DIST0
);
109 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1
))
110 assign_vue_slot(vue_map
, VARYING_SLOT_CLIP_DIST1
);
112 /* front and back colors need to be consecutive so that we can use
113 * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
116 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_COL0
))
117 assign_vue_slot(vue_map
, VARYING_SLOT_COL0
);
118 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_BFC0
))
119 assign_vue_slot(vue_map
, VARYING_SLOT_BFC0
);
120 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_COL1
))
121 assign_vue_slot(vue_map
, VARYING_SLOT_COL1
);
122 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_BFC1
))
123 assign_vue_slot(vue_map
, VARYING_SLOT_BFC1
);
126 /* The hardware doesn't care about the rest of the vertex outputs, so just
127 * assign them contiguously. Don't reassign outputs that already have a
130 * We generally don't need to assign a slot for VARYING_SLOT_CLIP_VERTEX,
131 * since it's encoded as the clip distances by emit_clip_distances().
132 * However, it may be output by transform feedback, and we'd rather not
133 * recompute state when TF changes, so we just always include it.
135 for (int i
= 0; i
< VARYING_SLOT_MAX
; ++i
) {
136 if ((slots_valid
& BITFIELD64_BIT(i
)) &&
137 vue_map
->varying_to_slot
[i
] == -1) {
138 assign_vue_slot(vue_map
, i
);
145 * Decide which set of clip planes should be used when clipping via
146 * gl_Position or gl_ClipVertex.
148 gl_clip_plane
*brw_select_clip_planes(struct gl_context
*ctx
)
150 if (ctx
->_Shader
->CurrentProgram
[MESA_SHADER_VERTEX
]) {
151 /* There is currently a GLSL vertex shader, so clip according to GLSL
152 * rules, which means compare gl_ClipVertex (or gl_Position, if
153 * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
154 * that were stored in EyeUserPlane at the time the clip planes were
157 return ctx
->Transform
.EyeUserPlane
;
159 /* Either we are using fixed function or an ARB vertex program. In
160 * either case the clip planes are going to be compared against
161 * gl_Position (which is in clip coordinates) so we have to clip using
162 * _ClipUserPlane, which was transformed into clip coordinates by Mesa
165 return ctx
->Transform
._ClipUserPlane
;
171 brw_vs_prog_data_compare(const void *in_a
, const void *in_b
)
173 const struct brw_vs_prog_data
*a
= in_a
;
174 const struct brw_vs_prog_data
*b
= in_b
;
176 /* Compare the base structure. */
177 if (!brw_stage_prog_data_compare(&a
->base
.base
, &b
->base
.base
))
180 /* Compare the rest of the struct. */
181 const unsigned offset
= sizeof(struct brw_stage_prog_data
);
182 if (memcmp(((char *) a
) + offset
, ((char *) b
) + offset
,
183 sizeof(struct brw_vs_prog_data
) - offset
)) {
191 do_vs_prog(struct brw_context
*brw
,
192 struct gl_shader_program
*prog
,
193 struct brw_vertex_program
*vp
,
194 struct brw_vs_prog_key
*key
)
197 const GLuint
*program
;
198 struct brw_vs_compile c
;
199 struct brw_vs_prog_data prog_data
;
200 struct brw_stage_prog_data
*stage_prog_data
= &prog_data
.base
.base
;
203 struct gl_shader
*vs
= NULL
;
206 vs
= prog
->_LinkedShaders
[MESA_SHADER_VERTEX
];
208 memset(&c
, 0, sizeof(c
));
209 memcpy(&c
.key
, key
, sizeof(*key
));
210 memset(&prog_data
, 0, sizeof(prog_data
));
212 /* Use ALT floating point mode for ARB programs so that 0^0 == 1. */
214 stage_prog_data
->use_alt_mode
= true;
216 mem_ctx
= ralloc_context(NULL
);
220 /* Allocate the references to the uniforms that will end up in the
221 * prog_data associated with the compiled program, and which will be freed
222 * by the state cache.
226 /* We add padding around uniform values below vec4 size, with the worst
227 * case being a float value that gets blown up to a vec4, so be
230 param_count
= vs
->num_uniform_components
* 4;
233 param_count
= vp
->program
.Base
.Parameters
->NumParameters
* 4;
235 /* vec4_visitor::setup_uniform_clipplane_values() also uploads user clip
236 * planes as uniforms.
238 param_count
+= c
.key
.base
.nr_userclip_plane_consts
* 4;
240 stage_prog_data
->param
=
241 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
242 stage_prog_data
->pull_param
=
243 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
244 stage_prog_data
->nr_params
= param_count
;
246 GLbitfield64 outputs_written
= vp
->program
.Base
.OutputsWritten
;
247 prog_data
.inputs_read
= vp
->program
.Base
.InputsRead
;
249 if (c
.key
.copy_edgeflag
) {
250 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_EDGE
);
251 prog_data
.inputs_read
|= VERT_BIT_EDGEFLAG
;
255 /* Put dummy slots into the VUE for the SF to put the replaced
256 * point sprite coords in. We shouldn't need these dummy slots,
257 * which take up precious URB space, but it would mean that the SF
258 * doesn't get nice aligned pairs of input coords into output
259 * coords, which would be a pain to handle.
261 for (i
= 0; i
< 8; i
++) {
262 if (c
.key
.point_coord_replace
& (1 << i
))
263 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_TEX0
+ i
);
266 /* if back colors are written, allocate slots for front colors too */
267 if (outputs_written
& BITFIELD64_BIT(VARYING_SLOT_BFC0
))
268 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_COL0
);
269 if (outputs_written
& BITFIELD64_BIT(VARYING_SLOT_BFC1
))
270 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_COL1
);
273 /* In order for legacy clipping to work, we need to populate the clip
274 * distance varying slots whenever clipping is enabled, even if the vertex
275 * shader doesn't write to gl_ClipDistance.
277 if (c
.key
.base
.userclip_active
) {
278 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0
);
279 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1
);
282 brw_compute_vue_map(brw
, &prog_data
.base
.vue_map
, outputs_written
);
285 _mesa_fprint_program_opt(stderr
, &c
.vp
->program
.Base
, PROG_PRINT_DEBUG
,
291 program
= brw_vs_emit(brw
, prog
, &c
, &prog_data
, mem_ctx
, &program_size
);
292 if (program
== NULL
) {
293 ralloc_free(mem_ctx
);
297 /* Scratch space is used for register spilling */
298 if (c
.base
.last_scratch
) {
299 perf_debug("Vertex shader triggered register spilling. "
300 "Try reducing the number of live vec4 values to "
301 "improve performance.\n");
303 prog_data
.base
.base
.total_scratch
304 = brw_get_scratch_size(c
.base
.last_scratch
*REG_SIZE
);
306 brw_get_scratch_bo(brw
, &brw
->vs
.base
.scratch_bo
,
307 prog_data
.base
.base
.total_scratch
*
308 brw
->max_vs_threads
);
311 brw_upload_cache(&brw
->cache
, BRW_CACHE_VS_PROG
,
312 &c
.key
, sizeof(c
.key
),
313 program
, program_size
,
314 &prog_data
, sizeof(prog_data
),
315 &brw
->vs
.base
.prog_offset
, &brw
->vs
.prog_data
);
316 ralloc_free(mem_ctx
);
322 key_debug(struct brw_context
*brw
, const char *name
, int a
, int b
)
325 perf_debug(" %s %d->%d\n", name
, a
, b
);
332 brw_vs_debug_recompile(struct brw_context
*brw
,
333 struct gl_shader_program
*prog
,
334 const struct brw_vs_prog_key
*key
)
336 struct brw_cache_item
*c
= NULL
;
337 const struct brw_vs_prog_key
*old_key
= NULL
;
340 perf_debug("Recompiling vertex shader for program %d\n", prog
->Name
);
342 for (unsigned int i
= 0; i
< brw
->cache
.size
; i
++) {
343 for (c
= brw
->cache
.items
[i
]; c
; c
= c
->next
) {
344 if (c
->cache_id
== BRW_CACHE_VS_PROG
) {
347 if (old_key
->base
.program_string_id
== key
->base
.program_string_id
)
356 perf_debug(" Didn't find previous compile in the shader cache for "
361 for (unsigned int i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
362 found
|= key_debug(brw
, "Vertex attrib w/a flags",
363 old_key
->gl_attrib_wa_flags
[i
],
364 key
->gl_attrib_wa_flags
[i
]);
367 found
|= key_debug(brw
, "user clip flags",
368 old_key
->base
.userclip_active
, key
->base
.userclip_active
);
370 found
|= key_debug(brw
, "user clipping planes as push constants",
371 old_key
->base
.nr_userclip_plane_consts
,
372 key
->base
.nr_userclip_plane_consts
);
374 found
|= key_debug(brw
, "copy edgeflag",
375 old_key
->copy_edgeflag
, key
->copy_edgeflag
);
376 found
|= key_debug(brw
, "PointCoord replace",
377 old_key
->point_coord_replace
, key
->point_coord_replace
);
378 found
|= key_debug(brw
, "vertex color clamping",
379 old_key
->clamp_vertex_color
, key
->clamp_vertex_color
);
381 found
|= brw_debug_recompile_sampler_key(brw
, &old_key
->base
.tex
,
385 perf_debug(" Something else\n");
391 brw_setup_vue_key_clip_info(struct brw_context
*brw
,
392 struct brw_vue_prog_key
*key
,
393 bool program_uses_clip_distance
)
395 struct gl_context
*ctx
= &brw
->ctx
;
397 key
->userclip_active
= (ctx
->Transform
.ClipPlanesEnabled
!= 0);
398 if (key
->userclip_active
&& !program_uses_clip_distance
) {
399 key
->nr_userclip_plane_consts
400 = _mesa_logbase2(ctx
->Transform
.ClipPlanesEnabled
) + 1;
405 brw_upload_vs_prog(struct brw_context
*brw
)
407 struct gl_context
*ctx
= &brw
->ctx
;
408 struct brw_vs_prog_key key
;
409 /* BRW_NEW_VERTEX_PROGRAM */
410 struct brw_vertex_program
*vp
=
411 (struct brw_vertex_program
*)brw
->vertex_program
;
412 struct gl_program
*prog
= (struct gl_program
*) brw
->vertex_program
;
415 if (!brw_state_dirty(brw
,
422 BRW_NEW_VERTEX_PROGRAM
|
423 BRW_NEW_VS_ATTRIB_WORKAROUNDS
))
426 memset(&key
, 0, sizeof(key
));
428 /* Just upload the program verbatim for now. Always send it all
429 * the inputs it asks for, whether they are varying or not.
431 key
.base
.program_string_id
= vp
->id
;
432 brw_setup_vue_key_clip_info(brw
, &key
.base
,
433 vp
->program
.Base
.UsesClipDistanceOut
);
437 key
.copy_edgeflag
= (ctx
->Polygon
.FrontMode
!= GL_FILL
||
438 ctx
->Polygon
.BackMode
!= GL_FILL
);
441 if (prog
->OutputsWritten
& (VARYING_BIT_COL0
| VARYING_BIT_COL1
|
442 VARYING_BIT_BFC0
| VARYING_BIT_BFC1
)) {
443 /* _NEW_LIGHT | _NEW_BUFFERS */
444 key
.clamp_vertex_color
= ctx
->Light
._ClampVertexColor
;
448 if (brw
->gen
< 6 && ctx
->Point
.PointSprite
) {
449 for (i
= 0; i
< 8; i
++) {
450 if (ctx
->Point
.CoordReplace
[i
])
451 key
.point_coord_replace
|= (1 << i
);
456 brw_populate_sampler_prog_key_data(ctx
, prog
, brw
->vs
.base
.sampler_count
,
459 /* BRW_NEW_VS_ATTRIB_WORKAROUNDS */
460 memcpy(key
.gl_attrib_wa_flags
, brw
->vb
.attrib_wa_flags
,
461 sizeof(brw
->vb
.attrib_wa_flags
));
463 if (!brw_search_cache(&brw
->cache
, BRW_CACHE_VS_PROG
,
465 &brw
->vs
.base
.prog_offset
, &brw
->vs
.prog_data
)) {
467 do_vs_prog(brw
, ctx
->_Shader
->CurrentProgram
[MESA_SHADER_VERTEX
], vp
,
472 brw
->vs
.base
.prog_data
= &brw
->vs
.prog_data
->base
.base
;
474 if (memcmp(&brw
->vs
.prog_data
->base
.vue_map
, &brw
->vue_map_geom_out
,
475 sizeof(brw
->vue_map_geom_out
)) != 0) {
476 brw
->vue_map_vs
= brw
->vs
.prog_data
->base
.vue_map
;
477 brw
->state
.dirty
.brw
|= BRW_NEW_VUE_MAP_VS
;
479 /* No geometry shader support, so the VS VUE map is the VUE map for
480 * the output of the "geometry" portion of the pipeline.
482 brw
->vue_map_geom_out
= brw
->vue_map_vs
;
483 brw
->state
.dirty
.brw
|= BRW_NEW_VUE_MAP_GEOM_OUT
;
489 brw_vs_precompile(struct gl_context
*ctx
,
490 struct gl_shader_program
*shader_prog
,
491 struct gl_program
*prog
)
493 struct brw_context
*brw
= brw_context(ctx
);
494 struct brw_vs_prog_key key
;
495 uint32_t old_prog_offset
= brw
->vs
.base
.prog_offset
;
496 struct brw_vs_prog_data
*old_prog_data
= brw
->vs
.prog_data
;
499 struct gl_vertex_program
*vp
= (struct gl_vertex_program
*) prog
;
500 struct brw_vertex_program
*bvp
= brw_vertex_program(vp
);
502 memset(&key
, 0, sizeof(key
));
504 brw_vue_setup_prog_key_for_precompile(ctx
, &key
.base
, bvp
->id
, &vp
->Base
);
505 key
.clamp_vertex_color
=
506 (prog
->OutputsWritten
& (VARYING_BIT_COL0
| VARYING_BIT_COL1
|
507 VARYING_BIT_BFC0
| VARYING_BIT_BFC1
));
509 success
= do_vs_prog(brw
, shader_prog
, bvp
, &key
);
511 brw
->vs
.base
.prog_offset
= old_prog_offset
;
512 brw
->vs
.prog_data
= old_prog_data
;