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 vert_result hasn't been assigned a slot already */
47 assert (vue_map
->vert_result_to_slot
[vert_result
] == -1);
49 vue_map
->vert_result_to_slot
[vert_result
] = vue_map
->num_slots
;
50 vue_map
->slot_to_vert_result
[vue_map
->num_slots
++] = vert_result
;
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_vs_compile
*c
)
63 struct brw_context
*brw
= c
->func
.brw
;
64 const struct intel_context
*intel
= &brw
->intel
;
65 struct brw_vue_map
*vue_map
= &c
->prog_data
.vue_map
;
66 GLbitfield64 outputs_written
= c
->prog_data
.outputs_written
;
69 vue_map
->num_slots
= 0;
70 for (i
= 0; i
< BRW_VERT_RESULT_MAX
; ++i
) {
71 vue_map
->vert_result_to_slot
[i
] = -1;
72 vue_map
->slot_to_vert_result
[i
] = BRW_VERT_RESULT_MAX
;
75 /* VUE header: format depends on chip generation and whether clipping is
80 /* There are 8 dwords in VUE header pre-Ironlake:
81 * dword 0-3 is indices, point width, clip flags.
82 * dword 4-7 is ndc position
83 * dword 8-11 is the first vertex data.
85 assign_vue_slot(vue_map
, VERT_RESULT_PSIZ
);
86 assign_vue_slot(vue_map
, BRW_VERT_RESULT_NDC
);
87 assign_vue_slot(vue_map
, VERT_RESULT_HPOS
);
90 /* There are 20 DWs (D0-D19) in VUE header on Ironlake:
91 * dword 0-3 of the header is indices, point width, clip flags.
92 * dword 4-7 is the ndc position
93 * dword 8-11 of the vertex header is the 4D space position
94 * dword 12-19 of the vertex header is the user clip distance.
95 * dword 20-23 is a pad so that the vertex element data is aligned
96 * dword 24-27 is the first vertex data we fill.
98 * Note: future pipeline stages expect 4D space position to be
99 * contiguous with the other vert_results, so we make dword 24-27 a
100 * duplicate copy of the 4D space position.
102 assign_vue_slot(vue_map
, VERT_RESULT_PSIZ
);
103 assign_vue_slot(vue_map
, BRW_VERT_RESULT_NDC
);
104 assign_vue_slot(vue_map
, BRW_VERT_RESULT_HPOS_DUPLICATE
);
105 assign_vue_slot(vue_map
, VERT_RESULT_CLIP_DIST0
);
106 assign_vue_slot(vue_map
, VERT_RESULT_CLIP_DIST1
);
107 assign_vue_slot(vue_map
, BRW_VERT_RESULT_PAD
);
108 assign_vue_slot(vue_map
, VERT_RESULT_HPOS
);
112 /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
113 * dword 0-3 of the header is indices, point width, clip flags.
114 * dword 4-7 is the 4D space position
115 * dword 8-15 of the vertex header is the user clip distance if
117 * dword 8-11 or 16-19 is the first vertex element data we fill.
119 assign_vue_slot(vue_map
, VERT_RESULT_PSIZ
);
120 assign_vue_slot(vue_map
, VERT_RESULT_HPOS
);
121 if (c
->key
.userclip_active
) {
122 assign_vue_slot(vue_map
, VERT_RESULT_CLIP_DIST0
);
123 assign_vue_slot(vue_map
, VERT_RESULT_CLIP_DIST1
);
125 /* front and back colors need to be consecutive so that we can use
126 * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
129 if (outputs_written
& BITFIELD64_BIT(VERT_RESULT_COL0
))
130 assign_vue_slot(vue_map
, VERT_RESULT_COL0
);
131 if (outputs_written
& BITFIELD64_BIT(VERT_RESULT_BFC0
))
132 assign_vue_slot(vue_map
, VERT_RESULT_BFC0
);
133 if (outputs_written
& BITFIELD64_BIT(VERT_RESULT_COL1
))
134 assign_vue_slot(vue_map
, VERT_RESULT_COL1
);
135 if (outputs_written
& BITFIELD64_BIT(VERT_RESULT_BFC1
))
136 assign_vue_slot(vue_map
, VERT_RESULT_BFC1
);
139 assert (!"VUE map not known for this chip generation");
143 /* The hardware doesn't care about the rest of the vertex outputs, so just
144 * assign them contiguously. Don't reassign outputs that already have a
147 * Also, prior to Gen6, don't assign a slot for VERT_RESULT_CLIP_VERTEX,
148 * since it is unsupported. In Gen6 and above, VERT_RESULT_CLIP_VERTEX may
149 * be needed for transform feedback; since we don't want to have to
150 * recompute the VUE map (and everything that depends on it) when transform
151 * feedback is enabled or disabled, just go ahead and assign a slot for it.
153 for (int i
= 0; i
< VERT_RESULT_MAX
; ++i
) {
154 if (intel
->gen
< 6 && i
== VERT_RESULT_CLIP_VERTEX
)
156 if ((outputs_written
& BITFIELD64_BIT(i
)) &&
157 vue_map
->vert_result_to_slot
[i
] == -1) {
158 assign_vue_slot(vue_map
, i
);
165 * Decide which set of clip planes should be used when clipping via
166 * gl_Position or gl_ClipVertex.
168 gl_clip_plane
*brw_select_clip_planes(struct gl_context
*ctx
)
170 if (ctx
->Shader
.CurrentVertexProgram
) {
171 /* There is currently a GLSL vertex shader, so clip according to GLSL
172 * rules, which means compare gl_ClipVertex (or gl_Position, if
173 * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
174 * that were stored in EyeUserPlane at the time the clip planes were
177 return ctx
->Transform
.EyeUserPlane
;
179 /* Either we are using fixed function or an ARB vertex program. In
180 * either case the clip planes are going to be compared against
181 * gl_Position (which is in clip coordinates) so we have to clip using
182 * _ClipUserPlane, which was transformed into clip coordinates by Mesa
185 return ctx
->Transform
._ClipUserPlane
;
190 brw_vs_prog_data_compare(const void *in_a
, const void *in_b
,
191 int aux_size
, const void *in_key
)
193 const struct brw_vs_prog_data
*a
= in_a
;
194 const struct brw_vs_prog_data
*b
= in_b
;
196 /* Compare all the struct up to the pointers. */
197 if (memcmp(a
, b
, offsetof(struct brw_vs_prog_data
, param
)))
200 if (memcmp(a
->param
, b
->param
, a
->nr_params
* sizeof(void *)))
203 if (memcmp(a
->pull_param
, b
->pull_param
, a
->nr_pull_params
* sizeof(void *)))
210 do_vs_prog(struct brw_context
*brw
,
211 struct gl_shader_program
*prog
,
212 struct brw_vertex_program
*vp
,
213 struct brw_vs_prog_key
*key
)
215 struct gl_context
*ctx
= &brw
->intel
.ctx
;
216 struct intel_context
*intel
= &brw
->intel
;
218 const GLuint
*program
;
219 struct brw_vs_compile c
;
223 struct gl_shader
*vs
= NULL
;
226 vs
= prog
->_LinkedShaders
[MESA_SHADER_VERTEX
];
228 memset(&c
, 0, sizeof(c
));
229 memcpy(&c
.key
, key
, sizeof(*key
));
231 mem_ctx
= ralloc_context(NULL
);
233 brw_init_compile(brw
, &c
.func
, mem_ctx
);
236 /* Allocate the references to the uniforms that will end up in the
237 * prog_data associated with the compiled program, and which will be freed
238 * by the state cache.
242 /* We add padding around uniform values below vec4 size, with the worst
243 * case being a float value that gets blown up to a vec4, so be
246 param_count
= vs
->num_uniform_components
* 4;
248 /* We also upload clip plane data as uniforms */
249 param_count
+= MAX_CLIP_PLANES
* 4;
251 param_count
= vp
->program
.Base
.Parameters
->NumParameters
* 4;
253 c
.prog_data
.param
= rzalloc_array(NULL
, const float *, param_count
);
254 c
.prog_data
.pull_param
= rzalloc_array(NULL
, const float *, param_count
);
256 c
.prog_data
.outputs_written
= vp
->program
.Base
.OutputsWritten
;
257 c
.prog_data
.inputs_read
= vp
->program
.Base
.InputsRead
;
259 if (c
.key
.copy_edgeflag
) {
260 c
.prog_data
.outputs_written
|= BITFIELD64_BIT(VERT_RESULT_EDGE
);
261 c
.prog_data
.inputs_read
|= VERT_BIT_EDGEFLAG
;
264 /* Put dummy slots into the VUE for the SF to put the replaced
265 * point sprite coords in. We shouldn't need these dummy slots,
266 * which take up precious URB space, but it would mean that the SF
267 * doesn't get nice aligned pairs of input coords into output
268 * coords, which would be a pain to handle.
270 for (i
= 0; i
< 8; i
++) {
271 if (c
.key
.point_coord_replace
& (1 << i
))
272 c
.prog_data
.outputs_written
|= BITFIELD64_BIT(VERT_RESULT_TEX0
+ i
);
275 brw_compute_vue_map(&c
);
278 _mesa_fprint_program_opt(stdout
, &c
.vp
->program
.Base
, PROG_PRINT_DEBUG
,
285 if (!brw_vs_emit(prog
, &c
)) {
286 ralloc_free(mem_ctx
);
293 if (c
.prog_data
.nr_pull_params
)
294 c
.prog_data
.num_surfaces
= 1;
295 if (c
.vp
->program
.Base
.SamplersUsed
)
296 c
.prog_data
.num_surfaces
= SURF_INDEX_VS_TEXTURE(BRW_MAX_TEX_UNIT
);
298 prog
->_LinkedShaders
[MESA_SHADER_VERTEX
]->NumUniformBlocks
) {
299 c
.prog_data
.num_surfaces
=
300 SURF_INDEX_VS_UBO(prog
->_LinkedShaders
[MESA_SHADER_VERTEX
]->NumUniformBlocks
);
303 /* Scratch space is used for register spilling */
304 if (c
.last_scratch
) {
305 perf_debug("Vertex shader triggered register spilling. "
306 "Try reducing the number of live vec4 values to "
307 "improve performance.\n");
309 c
.prog_data
.total_scratch
= brw_get_scratch_size(c
.last_scratch
*REG_SIZE
);
311 brw_get_scratch_bo(intel
, &brw
->vs
.scratch_bo
,
312 c
.prog_data
.total_scratch
* brw
->max_vs_threads
);
317 program
= brw_get_program(&c
.func
, &program_size
);
319 /* We upload from &c.prog_data including the constant_map assuming
320 * they're packed together. It would be nice to have a
321 * compile-time assert macro here.
323 assert(c
.constant_map
== (int8_t *)&c
.prog_data
+
324 sizeof(c
.prog_data
));
325 assert(ctx
->Const
.VertexProgram
.MaxNativeParameters
==
326 ARRAY_SIZE(c
.constant_map
));
329 aux_size
= sizeof(c
.prog_data
);
331 aux_size
+= c
.vp
->program
.Base
.Parameters
->NumParameters
;
333 brw_upload_cache(&brw
->cache
, BRW_VS_PROG
,
334 &c
.key
, sizeof(c
.key
),
335 program
, program_size
,
336 &c
.prog_data
, aux_size
,
337 &brw
->vs
.prog_offset
, &brw
->vs
.prog_data
);
338 ralloc_free(mem_ctx
);
344 key_debug(const char *name
, int a
, int b
)
347 perf_debug(" %s %d->%d\n", name
, a
, b
);
354 brw_vs_debug_recompile(struct brw_context
*brw
,
355 struct gl_shader_program
*prog
,
356 const struct brw_vs_prog_key
*key
)
358 struct brw_cache_item
*c
= NULL
;
359 const struct brw_vs_prog_key
*old_key
= NULL
;
362 perf_debug("Recompiling vertex shader for program %d\n", prog
->Name
);
364 for (unsigned int i
= 0; i
< brw
->cache
.size
; i
++) {
365 for (c
= brw
->cache
.items
[i
]; c
; c
= c
->next
) {
366 if (c
->cache_id
== BRW_VS_PROG
) {
369 if (old_key
->program_string_id
== key
->program_string_id
)
378 perf_debug(" Didn't find previous compile in the shader cache for "
383 for (unsigned int i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
384 found
|= key_debug("GL_FIXED rescaling",
385 old_key
->gl_fixed_input_size
[i
],
386 key
->gl_fixed_input_size
[i
]);
389 found
|= key_debug("user clip flags",
390 old_key
->userclip_active
, key
->userclip_active
);
392 found
|= key_debug("user clipping planes as push constants",
393 old_key
->nr_userclip_plane_consts
,
394 key
->nr_userclip_plane_consts
);
396 found
|= key_debug("clip distance enable",
397 old_key
->uses_clip_distance
, key
->uses_clip_distance
);
398 found
|= key_debug("clip plane enable bitfield",
399 old_key
->userclip_planes_enabled_gen_4_5
,
400 key
->userclip_planes_enabled_gen_4_5
);
401 found
|= key_debug("copy edgeflag",
402 old_key
->copy_edgeflag
, key
->copy_edgeflag
);
403 found
|= key_debug("PointCoord replace",
404 old_key
->point_coord_replace
, key
->point_coord_replace
);
405 found
|= key_debug("vertex color clamping",
406 old_key
->clamp_vertex_color
, key
->clamp_vertex_color
);
408 found
|= brw_debug_recompile_sampler_key(&old_key
->tex
, &key
->tex
);
411 perf_debug(" Something else\n");
415 static void brw_upload_vs_prog(struct brw_context
*brw
)
417 struct intel_context
*intel
= &brw
->intel
;
418 struct gl_context
*ctx
= &intel
->ctx
;
419 struct brw_vs_prog_key key
;
420 /* BRW_NEW_VERTEX_PROGRAM */
421 struct brw_vertex_program
*vp
=
422 (struct brw_vertex_program
*)brw
->vertex_program
;
423 struct gl_program
*prog
= (struct gl_program
*) brw
->vertex_program
;
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
.program_string_id
= vp
->id
;
432 key
.userclip_active
= (ctx
->Transform
.ClipPlanesEnabled
!= 0);
433 key
.uses_clip_distance
= vp
->program
.UsesClipDistance
;
434 if (key
.userclip_active
&& !key
.uses_clip_distance
) {
435 if (intel
->gen
< 6) {
436 key
.nr_userclip_plane_consts
437 = _mesa_bitcount_64(ctx
->Transform
.ClipPlanesEnabled
);
438 key
.userclip_planes_enabled_gen_4_5
439 = ctx
->Transform
.ClipPlanesEnabled
;
441 key
.nr_userclip_plane_consts
442 = _mesa_logbase2(ctx
->Transform
.ClipPlanesEnabled
) + 1;
447 if (intel
->gen
< 6) {
448 key
.copy_edgeflag
= (ctx
->Polygon
.FrontMode
!= GL_FILL
||
449 ctx
->Polygon
.BackMode
!= GL_FILL
);
452 /* _NEW_LIGHT | _NEW_BUFFERS */
453 key
.clamp_vertex_color
= ctx
->Light
._ClampVertexColor
;
456 if (ctx
->Point
.PointSprite
) {
457 for (i
= 0; i
< 8; i
++) {
458 if (ctx
->Point
.CoordReplace
[i
])
459 key
.point_coord_replace
|= (1 << i
);
464 brw_populate_sampler_prog_key_data(ctx
, prog
, &key
.tex
);
466 /* BRW_NEW_VERTICES */
467 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
468 if (vp
->program
.Base
.InputsRead
& BITFIELD64_BIT(i
) &&
469 brw
->vb
.inputs
[i
].glarray
->Type
== GL_FIXED
) {
470 key
.gl_fixed_input_size
[i
] = brw
->vb
.inputs
[i
].glarray
->Size
;
474 if (!brw_search_cache(&brw
->cache
, BRW_VS_PROG
,
476 &brw
->vs
.prog_offset
, &brw
->vs
.prog_data
)) {
477 bool success
= do_vs_prog(brw
, ctx
->Shader
.CurrentVertexProgram
,
482 brw
->vs
.constant_map
= ((int8_t *)brw
->vs
.prog_data
+
483 sizeof(*brw
->vs
.prog_data
));
488 const struct brw_tracked_state brw_vs_prog
= {
490 .mesa
= (_NEW_TRANSFORM
| _NEW_POLYGON
| _NEW_POINT
| _NEW_LIGHT
|
493 .brw
= (BRW_NEW_VERTEX_PROGRAM
|
497 .emit
= brw_upload_vs_prog
501 brw_vs_precompile(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
503 struct brw_context
*brw
= brw_context(ctx
);
504 struct brw_vs_prog_key key
;
505 uint32_t old_prog_offset
= brw
->vs
.prog_offset
;
506 struct brw_vs_prog_data
*old_prog_data
= brw
->vs
.prog_data
;
509 if (!prog
->_LinkedShaders
[MESA_SHADER_VERTEX
])
512 struct gl_vertex_program
*vp
= (struct gl_vertex_program
*)
513 prog
->_LinkedShaders
[MESA_SHADER_VERTEX
]->Program
;
514 struct brw_vertex_program
*bvp
= brw_vertex_program(vp
);
516 memset(&key
, 0, sizeof(key
));
518 key
.program_string_id
= bvp
->id
;
519 key
.clamp_vertex_color
= true;
521 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
522 if (vp
->Base
.ShadowSamplers
& (1 << i
)) {
523 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
524 key
.tex
.swizzles
[i
] =
525 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
527 /* Color sampler: assume no swizzling. */
528 key
.tex
.swizzles
[i
] = SWIZZLE_XYZW
;
532 success
= do_vs_prog(brw
, prog
, bvp
, &key
);
534 brw
->vs
.prog_offset
= old_prog_offset
;
535 brw
->vs
.prog_data
= old_prog_data
;
541 brw_vs_prog_data_free(const void *in_prog_data
)
543 const struct brw_vs_prog_data
*prog_data
= in_prog_data
;
545 ralloc_free((void *)prog_data
->param
);
546 ralloc_free((void *)prog_data
->pull_param
);