2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) 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 <keith@tungstengraphics.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 "glsl/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.
56 * Note that consumers of this map using cache keys must include
57 * prog_data->userclip and prog_data->outputs_written in their key
58 * (generated by CACHE_NEW_VS_PROG).
61 brw_compute_vue_map(struct brw_context
*brw
, struct brw_vue_map
*vue_map
,
62 GLbitfield64 slots_valid
, bool userclip_active
)
64 const struct intel_context
*intel
= &brw
->intel
;
66 vue_map
->slots_valid
= slots_valid
;
69 /* Make sure that the values we store in vue_map->varying_to_slot and
70 * vue_map->slot_to_varying won't overflow the signed chars that are used
71 * to store them. Note that since vue_map->slot_to_varying sometimes holds
72 * values equal to BRW_VARYING_SLOT_COUNT, we need to ensure that
73 * BRW_VARYING_SLOT_COUNT is <= 127, not 128.
75 STATIC_ASSERT(BRW_VARYING_SLOT_COUNT
<= 127);
77 vue_map
->num_slots
= 0;
78 for (i
= 0; i
< BRW_VARYING_SLOT_COUNT
; ++i
) {
79 vue_map
->varying_to_slot
[i
] = -1;
80 vue_map
->slot_to_varying
[i
] = BRW_VARYING_SLOT_COUNT
;
83 /* VUE header: format depends on chip generation and whether clipping is
89 /* There are 8 dwords in VUE header pre-Ironlake:
90 * dword 0-3 is indices, point width, clip flags.
91 * dword 4-7 is ndc position
92 * dword 8-11 is the first vertex data.
94 * On Ironlake the VUE header is nominally 20 dwords, but the hardware
95 * will accept the same header layout as Gen4 [and should be a bit faster]
97 assign_vue_slot(vue_map
, VARYING_SLOT_PSIZ
);
98 assign_vue_slot(vue_map
, BRW_VARYING_SLOT_NDC
);
99 assign_vue_slot(vue_map
, VARYING_SLOT_POS
);
103 /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
104 * dword 0-3 of the header is indices, point width, clip flags.
105 * dword 4-7 is the 4D space position
106 * dword 8-15 of the vertex header is the user clip distance if
108 * dword 8-11 or 16-19 is the first vertex element data we fill.
110 assign_vue_slot(vue_map
, VARYING_SLOT_PSIZ
);
111 assign_vue_slot(vue_map
, VARYING_SLOT_POS
);
112 if (userclip_active
) {
113 assign_vue_slot(vue_map
, VARYING_SLOT_CLIP_DIST0
);
114 assign_vue_slot(vue_map
, VARYING_SLOT_CLIP_DIST1
);
116 /* front and back colors need to be consecutive so that we can use
117 * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
120 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_COL0
))
121 assign_vue_slot(vue_map
, VARYING_SLOT_COL0
);
122 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_BFC0
))
123 assign_vue_slot(vue_map
, VARYING_SLOT_BFC0
);
124 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_COL1
))
125 assign_vue_slot(vue_map
, VARYING_SLOT_COL1
);
126 if (slots_valid
& BITFIELD64_BIT(VARYING_SLOT_BFC1
))
127 assign_vue_slot(vue_map
, VARYING_SLOT_BFC1
);
130 assert (!"VUE map not known for this chip generation");
134 /* The hardware doesn't care about the rest of the vertex outputs, so just
135 * assign them contiguously. Don't reassign outputs that already have a
138 * We generally don't need to assign a slot for VARYING_SLOT_CLIP_VERTEX,
139 * since it's encoded as the clip distances by emit_clip_distances().
140 * However, it may be output by transform feedback, and we'd rather not
141 * recompute state when TF changes, so we just always include it.
143 for (int i
= 0; i
< VARYING_SLOT_MAX
; ++i
) {
144 if ((slots_valid
& BITFIELD64_BIT(i
)) &&
145 vue_map
->varying_to_slot
[i
] == -1) {
146 assign_vue_slot(vue_map
, i
);
153 * Decide which set of clip planes should be used when clipping via
154 * gl_Position or gl_ClipVertex.
156 gl_clip_plane
*brw_select_clip_planes(struct gl_context
*ctx
)
158 if (ctx
->Shader
.CurrentVertexProgram
) {
159 /* There is currently a GLSL vertex shader, so clip according to GLSL
160 * rules, which means compare gl_ClipVertex (or gl_Position, if
161 * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
162 * that were stored in EyeUserPlane at the time the clip planes were
165 return ctx
->Transform
.EyeUserPlane
;
167 /* Either we are using fixed function or an ARB vertex program. In
168 * either case the clip planes are going to be compared against
169 * gl_Position (which is in clip coordinates) so we have to clip using
170 * _ClipUserPlane, which was transformed into clip coordinates by Mesa
173 return ctx
->Transform
._ClipUserPlane
;
179 brw_vec4_prog_data_compare(const struct brw_vec4_prog_data
*a
,
180 const struct brw_vec4_prog_data
*b
)
182 /* Compare all the struct up to the pointers. */
183 if (memcmp(a
, b
, offsetof(struct brw_vec4_prog_data
, param
)))
186 if (memcmp(a
->param
, b
->param
, a
->nr_params
* sizeof(void *)))
189 if (memcmp(a
->pull_param
, b
->pull_param
, a
->nr_pull_params
* sizeof(void *)))
197 brw_vs_prog_data_compare(const void *in_a
, const void *in_b
,
198 int aux_size
, const void *in_key
)
200 const struct brw_vs_prog_data
*a
= in_a
;
201 const struct brw_vs_prog_data
*b
= in_b
;
203 /* Compare the base vec4 structure. */
204 if (!brw_vec4_prog_data_compare(&a
->base
, &b
->base
))
207 /* Compare the rest of the struct. */
208 const unsigned offset
= sizeof(struct brw_vec4_prog_data
);
209 if (memcmp(((char *) &a
) + offset
, ((char *) &b
) + offset
,
210 sizeof(struct brw_vs_prog_data
) - offset
)) {
218 do_vs_prog(struct brw_context
*brw
,
219 struct gl_shader_program
*prog
,
220 struct brw_vertex_program
*vp
,
221 struct brw_vs_prog_key
*key
)
223 struct intel_context
*intel
= &brw
->intel
;
225 const GLuint
*program
;
226 struct brw_vs_compile c
;
227 struct brw_vs_prog_data prog_data
;
230 struct gl_shader
*vs
= NULL
;
233 vs
= prog
->_LinkedShaders
[MESA_SHADER_VERTEX
];
235 memset(&c
, 0, sizeof(c
));
236 memcpy(&c
.key
, key
, sizeof(*key
));
237 memset(&prog_data
, 0, sizeof(prog_data
));
239 mem_ctx
= ralloc_context(NULL
);
243 /* Allocate the references to the uniforms that will end up in the
244 * prog_data associated with the compiled program, and which will be freed
245 * by the state cache.
249 /* We add padding around uniform values below vec4 size, with the worst
250 * case being a float value that gets blown up to a vec4, so be
253 param_count
= vs
->num_uniform_components
* 4;
256 param_count
= vp
->program
.Base
.Parameters
->NumParameters
* 4;
258 /* We also upload clip plane data as uniforms */
259 param_count
+= MAX_CLIP_PLANES
* 4;
261 prog_data
.base
.param
= rzalloc_array(NULL
, const float *, param_count
);
262 prog_data
.base
.pull_param
= rzalloc_array(NULL
, const float *, param_count
);
264 GLbitfield64 outputs_written
= vp
->program
.Base
.OutputsWritten
;
265 prog_data
.inputs_read
= vp
->program
.Base
.InputsRead
;
267 if (c
.key
.copy_edgeflag
) {
268 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_EDGE
);
269 prog_data
.inputs_read
|= VERT_BIT_EDGEFLAG
;
272 if (intel
->gen
< 6) {
273 /* Put dummy slots into the VUE for the SF to put the replaced
274 * point sprite coords in. We shouldn't need these dummy slots,
275 * which take up precious URB space, but it would mean that the SF
276 * doesn't get nice aligned pairs of input coords into output
277 * coords, which would be a pain to handle.
279 for (i
= 0; i
< 8; i
++) {
280 if (c
.key
.point_coord_replace
& (1 << i
))
281 outputs_written
|= BITFIELD64_BIT(VARYING_SLOT_TEX0
+ i
);
285 brw_compute_vue_map(brw
, &prog_data
.base
.vue_map
, outputs_written
,
286 c
.key
.base
.userclip_active
);
289 _mesa_fprint_program_opt(stdout
, &c
.vp
->program
.Base
, PROG_PRINT_DEBUG
,
295 program
= brw_vs_emit(brw
, prog
, &c
, &prog_data
, mem_ctx
, &program_size
);
296 if (program
== NULL
) {
297 ralloc_free(mem_ctx
);
301 if (prog_data
.base
.nr_pull_params
)
302 prog_data
.base
.num_surfaces
= 1;
303 if (c
.vp
->program
.Base
.SamplersUsed
)
304 prog_data
.base
.num_surfaces
= SURF_INDEX_VS_TEXTURE(BRW_MAX_TEX_UNIT
);
306 prog
->_LinkedShaders
[MESA_SHADER_VERTEX
]->NumUniformBlocks
) {
307 prog_data
.base
.num_surfaces
=
308 SURF_INDEX_VS_UBO(prog
->_LinkedShaders
[MESA_SHADER_VERTEX
]->NumUniformBlocks
);
311 /* Scratch space is used for register spilling */
312 if (c
.base
.last_scratch
) {
313 perf_debug("Vertex shader triggered register spilling. "
314 "Try reducing the number of live vec4 values to "
315 "improve performance.\n");
317 prog_data
.base
.total_scratch
318 = brw_get_scratch_size(c
.base
.last_scratch
*REG_SIZE
);
320 brw_get_scratch_bo(intel
, &brw
->vs
.scratch_bo
,
321 prog_data
.base
.total_scratch
* brw
->max_vs_threads
);
324 brw_upload_cache(&brw
->cache
, BRW_VS_PROG
,
325 &c
.key
, sizeof(c
.key
),
326 program
, program_size
,
327 &prog_data
, sizeof(prog_data
),
328 &brw
->vs
.prog_offset
, &brw
->vs
.prog_data
);
329 ralloc_free(mem_ctx
);
335 key_debug(struct intel_context
*intel
, const char *name
, int a
, int b
)
338 perf_debug(" %s %d->%d\n", name
, a
, b
);
345 brw_vs_debug_recompile(struct brw_context
*brw
,
346 struct gl_shader_program
*prog
,
347 const struct brw_vs_prog_key
*key
)
349 struct intel_context
*intel
= &brw
->intel
;
350 struct brw_cache_item
*c
= NULL
;
351 const struct brw_vs_prog_key
*old_key
= NULL
;
354 perf_debug("Recompiling vertex shader for program %d\n", prog
->Name
);
356 for (unsigned int i
= 0; i
< brw
->cache
.size
; i
++) {
357 for (c
= brw
->cache
.items
[i
]; c
; c
= c
->next
) {
358 if (c
->cache_id
== BRW_VS_PROG
) {
361 if (old_key
->base
.program_string_id
== key
->base
.program_string_id
)
370 perf_debug(" Didn't find previous compile in the shader cache for "
375 for (unsigned int i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
376 found
|= key_debug(intel
, "Vertex attrib w/a flags",
377 old_key
->gl_attrib_wa_flags
[i
],
378 key
->gl_attrib_wa_flags
[i
]);
381 found
|= key_debug(intel
, "user clip flags",
382 old_key
->base
.userclip_active
, key
->base
.userclip_active
);
384 found
|= key_debug(intel
, "user clipping planes as push constants",
385 old_key
->base
.nr_userclip_plane_consts
,
386 key
->base
.nr_userclip_plane_consts
);
388 found
|= key_debug(intel
, "clip distance enable",
389 old_key
->base
.uses_clip_distance
, key
->base
.uses_clip_distance
);
390 found
|= key_debug(intel
, "clip plane enable bitfield",
391 old_key
->base
.userclip_planes_enabled_gen_4_5
,
392 key
->base
.userclip_planes_enabled_gen_4_5
);
393 found
|= key_debug(intel
, "copy edgeflag",
394 old_key
->copy_edgeflag
, key
->copy_edgeflag
);
395 found
|= key_debug(intel
, "PointCoord replace",
396 old_key
->point_coord_replace
, key
->point_coord_replace
);
397 found
|= key_debug(intel
, "vertex color clamping",
398 old_key
->base
.clamp_vertex_color
, key
->base
.clamp_vertex_color
);
400 found
|= brw_debug_recompile_sampler_key(intel
, &old_key
->base
.tex
,
404 perf_debug(" Something else\n");
408 static void brw_upload_vs_prog(struct brw_context
*brw
)
410 struct intel_context
*intel
= &brw
->intel
;
411 struct gl_context
*ctx
= &intel
->ctx
;
412 struct brw_vs_prog_key key
;
413 /* BRW_NEW_VERTEX_PROGRAM */
414 struct brw_vertex_program
*vp
=
415 (struct brw_vertex_program
*)brw
->vertex_program
;
416 struct gl_program
*prog
= (struct gl_program
*) brw
->vertex_program
;
419 memset(&key
, 0, sizeof(key
));
421 /* Just upload the program verbatim for now. Always send it all
422 * the inputs it asks for, whether they are varying or not.
424 key
.base
.program_string_id
= vp
->id
;
425 key
.base
.userclip_active
= (ctx
->Transform
.ClipPlanesEnabled
!= 0);
426 key
.base
.uses_clip_distance
= vp
->program
.UsesClipDistance
;
427 if (key
.base
.userclip_active
&& !key
.base
.uses_clip_distance
) {
428 if (intel
->gen
< 6) {
429 key
.base
.nr_userclip_plane_consts
430 = _mesa_bitcount_64(ctx
->Transform
.ClipPlanesEnabled
);
431 key
.base
.userclip_planes_enabled_gen_4_5
432 = ctx
->Transform
.ClipPlanesEnabled
;
434 key
.base
.nr_userclip_plane_consts
435 = _mesa_logbase2(ctx
->Transform
.ClipPlanesEnabled
) + 1;
440 if (intel
->gen
< 6) {
441 key
.copy_edgeflag
= (ctx
->Polygon
.FrontMode
!= GL_FILL
||
442 ctx
->Polygon
.BackMode
!= GL_FILL
);
445 /* _NEW_LIGHT | _NEW_BUFFERS */
446 key
.base
.clamp_vertex_color
= ctx
->Light
._ClampVertexColor
;
449 if (intel
->gen
< 6 && ctx
->Point
.PointSprite
) {
450 for (i
= 0; i
< 8; i
++) {
451 if (ctx
->Point
.CoordReplace
[i
])
452 key
.point_coord_replace
|= (1 << i
);
457 brw_populate_sampler_prog_key_data(ctx
, prog
, &key
.base
.tex
);
459 /* BRW_NEW_VERTICES */
460 if (intel
->gen
< 8 && !intel
->is_haswell
) {
461 /* Prior to Haswell, the hardware can't natively support GL_FIXED or
462 * 2_10_10_10_REV vertex formats. Set appropriate workaround flags.
464 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
465 if (!(vp
->program
.Base
.InputsRead
& BITFIELD64_BIT(i
)))
468 uint8_t wa_flags
= 0;
470 switch (brw
->vb
.inputs
[i
].glarray
->Type
) {
473 wa_flags
= brw
->vb
.inputs
[i
].glarray
->Size
;
476 case GL_INT_2_10_10_10_REV
:
477 wa_flags
|= BRW_ATTRIB_WA_SIGN
;
480 case GL_UNSIGNED_INT_2_10_10_10_REV
:
481 if (brw
->vb
.inputs
[i
].glarray
->Format
== GL_BGRA
)
482 wa_flags
|= BRW_ATTRIB_WA_BGRA
;
484 if (brw
->vb
.inputs
[i
].glarray
->Normalized
)
485 wa_flags
|= BRW_ATTRIB_WA_NORMALIZE
;
486 else if (!brw
->vb
.inputs
[i
].glarray
->Integer
)
487 wa_flags
|= BRW_ATTRIB_WA_SCALE
;
492 key
.gl_attrib_wa_flags
[i
] = wa_flags
;
496 if (!brw_search_cache(&brw
->cache
, BRW_VS_PROG
,
498 &brw
->vs
.prog_offset
, &brw
->vs
.prog_data
)) {
499 bool success
= do_vs_prog(brw
, ctx
->Shader
.CurrentVertexProgram
,
504 if (memcmp(&brw
->vs
.prog_data
->base
.vue_map
, &brw
->vue_map_geom_out
,
505 sizeof(brw
->vue_map_geom_out
)) != 0) {
506 brw
->vue_map_geom_out
= brw
->vs
.prog_data
->base
.vue_map
;
507 brw
->state
.dirty
.brw
|= BRW_NEW_VUE_MAP_GEOM_OUT
;
513 const struct brw_tracked_state brw_vs_prog
= {
515 .mesa
= (_NEW_TRANSFORM
| _NEW_POLYGON
| _NEW_POINT
| _NEW_LIGHT
|
518 .brw
= (BRW_NEW_VERTEX_PROGRAM
|
522 .emit
= brw_upload_vs_prog
526 brw_vs_precompile(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
528 struct brw_context
*brw
= brw_context(ctx
);
529 struct brw_vs_prog_key key
;
530 uint32_t old_prog_offset
= brw
->vs
.prog_offset
;
531 struct brw_vs_prog_data
*old_prog_data
= brw
->vs
.prog_data
;
534 if (!prog
->_LinkedShaders
[MESA_SHADER_VERTEX
])
537 struct gl_vertex_program
*vp
= (struct gl_vertex_program
*)
538 prog
->_LinkedShaders
[MESA_SHADER_VERTEX
]->Program
;
539 struct brw_vertex_program
*bvp
= brw_vertex_program(vp
);
541 memset(&key
, 0, sizeof(key
));
543 key
.base
.program_string_id
= bvp
->id
;
544 key
.base
.clamp_vertex_color
= ctx
->API
== API_OPENGL_COMPAT
;
546 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
547 if (vp
->Base
.ShadowSamplers
& (1 << i
)) {
548 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
549 key
.base
.tex
.swizzles
[i
] =
550 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
552 /* Color sampler: assume no swizzling. */
553 key
.base
.tex
.swizzles
[i
] = SWIZZLE_XYZW
;
557 success
= do_vs_prog(brw
, prog
, bvp
, &key
);
559 brw
->vs
.prog_offset
= old_prog_offset
;
560 brw
->vs
.prog_data
= old_prog_data
;
567 brw_vec4_prog_data_free(const struct brw_vec4_prog_data
*prog_data
)
569 ralloc_free((void *)prog_data
->param
);
570 ralloc_free((void *)prog_data
->pull_param
);
575 brw_vs_prog_data_free(const void *in_prog_data
)
577 const struct brw_vs_prog_data
*prog_data
= in_prog_data
;
579 brw_vec4_prog_data_free(&prog_data
->base
);