2 * Copyright © 2011 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 * Code to initialize the binding table entries used by transform feedback.
29 #include "main/bufferobj.h"
30 #include "main/macros.h"
31 #include "brw_context.h"
32 #include "intel_batchbuffer.h"
33 #include "brw_defines.h"
34 #include "brw_state.h"
35 #include "main/transformfeedback.h"
38 gen6_update_sol_surfaces(struct brw_context
*brw
)
40 struct gl_context
*ctx
= &brw
->ctx
;
41 bool xfb_active
= _mesa_is_xfb_active_and_unpaused(ctx
);
42 struct gl_transform_feedback_object
*xfb_obj
;
43 const struct gl_transform_feedback_info
*linked_xfb_info
= NULL
;
46 /* BRW_NEW_TRANSFORM_FEEDBACK */
47 xfb_obj
= ctx
->TransformFeedback
.CurrentObject
;
48 linked_xfb_info
= xfb_obj
->program
->sh
.LinkedTransformFeedback
;
51 for (int i
= 0; i
< BRW_MAX_SOL_BINDINGS
; ++i
) {
52 const int surf_index
= BRW_GEN6_SOL_BINDING_START
+ i
;
53 if (xfb_active
&& i
< linked_xfb_info
->NumOutputs
) {
54 unsigned buffer
= linked_xfb_info
->Outputs
[i
].OutputBuffer
;
55 unsigned buffer_offset
=
56 xfb_obj
->Offset
[buffer
] / 4 +
57 linked_xfb_info
->Outputs
[i
].DstOffset
;
58 if (brw
->geometry_program
) {
59 brw_update_sol_surface(
60 brw
, xfb_obj
->Buffers
[buffer
],
61 &brw
->gs
.base
.surf_offset
[surf_index
],
62 linked_xfb_info
->Outputs
[i
].NumComponents
,
63 linked_xfb_info
->Buffers
[buffer
].Stride
, buffer_offset
);
65 brw_update_sol_surface(
66 brw
, xfb_obj
->Buffers
[buffer
],
67 &brw
->ff_gs
.surf_offset
[surf_index
],
68 linked_xfb_info
->Outputs
[i
].NumComponents
,
69 linked_xfb_info
->Buffers
[buffer
].Stride
, buffer_offset
);
72 if (!brw
->geometry_program
)
73 brw
->ff_gs
.surf_offset
[surf_index
] = 0;
75 brw
->gs
.base
.surf_offset
[surf_index
] = 0;
79 brw
->ctx
.NewDriverState
|= BRW_NEW_SURFACES
;
82 const struct brw_tracked_state gen6_sol_surface
= {
85 .brw
= BRW_NEW_BATCH
|
87 BRW_NEW_TRANSFORM_FEEDBACK
,
89 .emit
= gen6_update_sol_surfaces
,
93 * Constructs the binding table for the WM surface state, which maps unit
94 * numbers to surface state objects.
97 brw_gs_upload_binding_table(struct brw_context
*brw
)
100 struct gl_context
*ctx
= &brw
->ctx
;
101 const struct gl_program
*prog
;
102 bool need_binding_table
= false;
104 /* We have two scenarios here:
105 * 1) We are using a geometry shader only to implement transform feedback
106 * for a vertex shader (brw->geometry_program == NULL). In this case, we
107 * only need surfaces for transform feedback in the GS stage.
108 * 2) We have a user-provided geometry shader. In this case we may need
109 * surfaces for transform feedback and/or other stuff, like textures,
113 if (!brw
->geometry_program
) {
114 /* BRW_NEW_VERTEX_PROGRAM */
115 prog
= ctx
->_Shader
->CurrentProgram
[MESA_SHADER_VERTEX
];
117 /* Skip making a binding table if we don't have anything to put in it */
118 const struct gl_transform_feedback_info
*linked_xfb_info
=
119 prog
->sh
.LinkedTransformFeedback
;
120 need_binding_table
= linked_xfb_info
->NumOutputs
> 0;
122 if (!need_binding_table
) {
123 if (brw
->ff_gs
.bind_bo_offset
!= 0) {
124 brw
->ctx
.NewDriverState
|= BRW_NEW_BINDING_TABLE_POINTERS
;
125 brw
->ff_gs
.bind_bo_offset
= 0;
130 /* Might want to calculate nr_surfaces first, to avoid taking up so much
131 * space for the binding table. Anyway, in this case we know that we only
132 * use BRW_MAX_SOL_BINDINGS surfaces at most.
134 bind
= brw_state_batch(brw
, sizeof(uint32_t) * BRW_MAX_SOL_BINDINGS
,
135 32, &brw
->ff_gs
.bind_bo_offset
);
137 /* BRW_NEW_SURFACES */
138 memcpy(bind
, brw
->ff_gs
.surf_offset
,
139 BRW_MAX_SOL_BINDINGS
* sizeof(uint32_t));
141 /* BRW_NEW_GEOMETRY_PROGRAM */
142 prog
= ctx
->_Shader
->CurrentProgram
[MESA_SHADER_GEOMETRY
];
144 /* Skip making a binding table if we don't have anything to put in it */
145 struct brw_stage_prog_data
*prog_data
= brw
->gs
.base
.prog_data
;
146 const struct gl_transform_feedback_info
*linked_xfb_info
=
147 prog
->sh
.LinkedTransformFeedback
;
148 need_binding_table
= linked_xfb_info
->NumOutputs
> 0 ||
149 prog_data
->binding_table
.size_bytes
> 0;
151 if (!need_binding_table
) {
152 if (brw
->gs
.base
.bind_bo_offset
!= 0) {
153 brw
->gs
.base
.bind_bo_offset
= 0;
154 brw
->ctx
.NewDriverState
|= BRW_NEW_BINDING_TABLE_POINTERS
;
159 /* Might want to calculate nr_surfaces first, to avoid taking up so much
160 * space for the binding table.
162 bind
= brw_state_batch(brw
, sizeof(uint32_t) * BRW_MAX_SURFACES
,
163 32, &brw
->gs
.base
.bind_bo_offset
);
165 /* BRW_NEW_SURFACES */
166 memcpy(bind
, brw
->gs
.base
.surf_offset
,
167 BRW_MAX_SURFACES
* sizeof(uint32_t));
170 brw
->ctx
.NewDriverState
|= BRW_NEW_BINDING_TABLE_POINTERS
;
173 const struct brw_tracked_state gen6_gs_binding_table
= {
176 .brw
= BRW_NEW_BATCH
|
178 BRW_NEW_GEOMETRY_PROGRAM
|
179 BRW_NEW_VERTEX_PROGRAM
|
182 .emit
= brw_gs_upload_binding_table
,
185 struct gl_transform_feedback_object
*
186 brw_new_transform_feedback(struct gl_context
*ctx
, GLuint name
)
188 struct brw_context
*brw
= brw_context(ctx
);
189 struct brw_transform_feedback_object
*brw_obj
=
190 CALLOC_STRUCT(brw_transform_feedback_object
);
194 _mesa_init_transform_feedback_object(&brw_obj
->base
, name
);
197 brw_bo_alloc(brw
->bufmgr
, "transform feedback offsets", 16, 64);
198 brw_obj
->prim_count_bo
=
199 brw_bo_alloc(brw
->bufmgr
, "xfb primitive counts", 4096, 64);
201 return &brw_obj
->base
;
205 brw_delete_transform_feedback(struct gl_context
*ctx
,
206 struct gl_transform_feedback_object
*obj
)
208 struct brw_transform_feedback_object
*brw_obj
=
209 (struct brw_transform_feedback_object
*) obj
;
211 for (unsigned i
= 0; i
< ARRAY_SIZE(obj
->Buffers
); i
++) {
212 _mesa_reference_buffer_object(ctx
, &obj
->Buffers
[i
], NULL
);
215 brw_bo_unreference(brw_obj
->offset_bo
);
216 brw_bo_unreference(brw_obj
->prim_count_bo
);
222 * Tally the number of primitives generated so far.
224 * The buffer contains a series of pairs:
225 * (<start0, start1, start2, start3>, <end0, end1, end2, end3>) ;
226 * (<start0, start1, start2, start3>, <end0, end1, end2, end3>) ;
228 * For each stream, we subtract the pair of values (end - start) to get the
229 * number of primitives generated during one section. We accumulate these
230 * values, adding them up to get the total number of primitives generated.
232 * Note that we expose one stream pre-Gen7, so the above is just (start, end).
235 tally_prims_generated(struct brw_context
*brw
,
236 struct brw_transform_feedback_object
*obj
)
238 const struct gl_context
*ctx
= &brw
->ctx
;
239 const int streams
= ctx
->Const
.MaxVertexStreams
;
241 /* If the current batch is still contributing to the number of primitives
242 * generated, flush it now so the results will be present when mapped.
244 if (brw_batch_references(&brw
->batch
, obj
->prim_count_bo
))
245 intel_batchbuffer_flush(brw
);
247 if (unlikely(brw
->perf_debug
&& brw_bo_busy(obj
->prim_count_bo
)))
248 perf_debug("Stalling for # of transform feedback primitives written.\n");
250 uint64_t *prim_counts
= brw_bo_map(brw
, obj
->prim_count_bo
, false);
252 assert(obj
->prim_count_buffer_index
% (2 * streams
) == 0);
253 int pairs
= obj
->prim_count_buffer_index
/ (2 * streams
);
255 for (int i
= 0; i
< pairs
; i
++) {
256 for (int s
= 0; s
< streams
; s
++) {
257 obj
->prims_generated
[s
] += prim_counts
[streams
+ s
] - prim_counts
[s
];
259 prim_counts
+= 2 * streams
; /* move to the next pair */
262 brw_bo_unmap(obj
->prim_count_bo
);
264 /* We've already gathered up the old data; we can safely overwrite it now. */
265 obj
->prim_count_buffer_index
= 0;
269 * Store the SO_NUM_PRIMS_WRITTEN counters for each stream (4 uint64_t values)
272 * If prim_count_bo is out of space, gather up the results so far into
273 * prims_generated[] and allocate a new buffer with enough space.
275 * The number of primitives written is used to compute the number of vertices
276 * written to a transform feedback stream, which is required to implement
277 * DrawTransformFeedback().
280 brw_save_primitives_written_counters(struct brw_context
*brw
,
281 struct brw_transform_feedback_object
*obj
)
283 const struct gl_context
*ctx
= &brw
->ctx
;
284 const int streams
= ctx
->Const
.MaxVertexStreams
;
286 assert(obj
->prim_count_bo
!= NULL
);
288 /* Check if there's enough space for a new pair of four values. */
289 if (obj
->prim_count_buffer_index
+ 2 * streams
>= 4096 / sizeof(uint64_t)) {
290 /* Gather up the results so far and release the BO. */
291 tally_prims_generated(brw
, obj
);
294 /* Flush any drawing so that the counters have the right values. */
295 brw_emit_mi_flush(brw
);
297 /* Emit MI_STORE_REGISTER_MEM commands to write the values. */
299 for (int i
= 0; i
< streams
; i
++) {
300 int offset
= (obj
->prim_count_buffer_index
+ i
) * sizeof(uint64_t);
301 brw_store_register_mem64(brw
, obj
->prim_count_bo
,
302 GEN7_SO_NUM_PRIMS_WRITTEN(i
),
306 brw_store_register_mem64(brw
, obj
->prim_count_bo
,
307 GEN6_SO_NUM_PRIMS_WRITTEN
,
308 obj
->prim_count_buffer_index
* sizeof(uint64_t));
311 /* Update where to write data to. */
312 obj
->prim_count_buffer_index
+= streams
;
316 compute_vertices_written_so_far(struct brw_context
*brw
,
317 struct brw_transform_feedback_object
*obj
,
318 uint64_t *vertices_written
)
320 const struct gl_context
*ctx
= &brw
->ctx
;
321 unsigned vertices_per_prim
= 0;
323 switch (obj
->primitive_mode
) {
325 vertices_per_prim
= 1;
328 vertices_per_prim
= 2;
331 vertices_per_prim
= 3;
334 unreachable("Invalid transform feedback primitive mode.");
337 /* Get the number of primitives generated. */
338 tally_prims_generated(brw
, obj
);
340 for (int i
= 0; i
< ctx
->Const
.MaxVertexStreams
; i
++) {
341 vertices_written
[i
] = vertices_per_prim
* obj
->prims_generated
[i
];
346 * Compute the number of vertices written by this transform feedback operation.
349 brw_compute_xfb_vertices_written(struct brw_context
*brw
,
350 struct brw_transform_feedback_object
*obj
)
352 if (obj
->vertices_written_valid
|| !obj
->base
.EndedAnytime
)
355 compute_vertices_written_so_far(brw
, obj
, obj
->vertices_written
);
357 obj
->vertices_written_valid
= true;
361 * GetTransformFeedbackVertexCount() driver hook.
363 * Returns the number of vertices written to a particular stream by the last
364 * Begin/EndTransformFeedback block. Used to implement DrawTransformFeedback().
367 brw_get_transform_feedback_vertex_count(struct gl_context
*ctx
,
368 struct gl_transform_feedback_object
*obj
,
371 struct brw_context
*brw
= brw_context(ctx
);
372 struct brw_transform_feedback_object
*brw_obj
=
373 (struct brw_transform_feedback_object
*) obj
;
375 assert(obj
->EndedAnytime
);
376 assert(stream
< ctx
->Const
.MaxVertexStreams
);
378 brw_compute_xfb_vertices_written(brw
, brw_obj
);
379 return brw_obj
->vertices_written
[stream
];
383 brw_begin_transform_feedback(struct gl_context
*ctx
, GLenum mode
,
384 struct gl_transform_feedback_object
*obj
)
386 struct brw_context
*brw
= brw_context(ctx
);
387 const struct gl_program
*prog
;
388 const struct gl_transform_feedback_info
*linked_xfb_info
;
389 struct gl_transform_feedback_object
*xfb_obj
=
390 ctx
->TransformFeedback
.CurrentObject
;
391 struct brw_transform_feedback_object
*brw_obj
=
392 (struct brw_transform_feedback_object
*) xfb_obj
;
394 assert(brw
->gen
== 6);
396 if (ctx
->_Shader
->CurrentProgram
[MESA_SHADER_GEOMETRY
]) {
397 /* BRW_NEW_GEOMETRY_PROGRAM */
398 prog
= ctx
->_Shader
->CurrentProgram
[MESA_SHADER_GEOMETRY
];
400 /* BRW_NEW_VERTEX_PROGRAM */
401 prog
= ctx
->_Shader
->CurrentProgram
[MESA_SHADER_VERTEX
];
403 linked_xfb_info
= prog
->sh
.LinkedTransformFeedback
;
405 /* Compute the maximum number of vertices that we can write without
406 * overflowing any of the buffers currently being used for feedback.
409 = _mesa_compute_max_transform_feedback_vertices(ctx
, xfb_obj
,
412 /* Initialize the SVBI 0 register to zero and set the maximum index. */
414 OUT_BATCH(_3DSTATE_GS_SVB_INDEX
<< 16 | (4 - 2));
415 OUT_BATCH(0); /* SVBI 0 */
416 OUT_BATCH(0); /* starting index */
417 OUT_BATCH(brw_obj
->max_index
);
420 /* Initialize the rest of the unused streams to sane values. Otherwise,
421 * they may indicate that there is no room to write data and prevent
422 * anything from happening at all.
424 for (int i
= 1; i
< 4; i
++) {
426 OUT_BATCH(_3DSTATE_GS_SVB_INDEX
<< 16 | (4 - 2));
427 OUT_BATCH(i
<< SVB_INDEX_SHIFT
);
428 OUT_BATCH(0); /* starting index */
429 OUT_BATCH(0xffffffff);
433 /* We're about to lose the information needed to compute the number of
434 * vertices written during the last Begin/EndTransformFeedback section,
435 * so we can't delay it any further.
437 brw_compute_xfb_vertices_written(brw
, brw_obj
);
439 /* No primitives have been generated yet. */
440 for (int i
= 0; i
< BRW_MAX_XFB_STREAMS
; i
++) {
441 brw_obj
->prims_generated
[i
] = 0;
444 /* Store the starting value of the SO_NUM_PRIMS_WRITTEN counters. */
445 brw_save_primitives_written_counters(brw
, brw_obj
);
447 brw_obj
->primitive_mode
= mode
;
451 brw_end_transform_feedback(struct gl_context
*ctx
,
452 struct gl_transform_feedback_object
*obj
)
454 struct brw_context
*brw
= brw_context(ctx
);
455 struct brw_transform_feedback_object
*brw_obj
=
456 (struct brw_transform_feedback_object
*) obj
;
458 /* Store the ending value of the SO_NUM_PRIMS_WRITTEN counters. */
460 brw_save_primitives_written_counters(brw
, brw_obj
);
462 /* EndTransformFeedback() means that we need to update the number of
463 * vertices written. Since it's only necessary if DrawTransformFeedback()
464 * is called and it means mapping a buffer object, we delay computing it
465 * until it's absolutely necessary to try and avoid stalls.
467 brw_obj
->vertices_written_valid
= false;
471 brw_pause_transform_feedback(struct gl_context
*ctx
,
472 struct gl_transform_feedback_object
*obj
)
474 struct brw_context
*brw
= brw_context(ctx
);
475 struct brw_transform_feedback_object
*brw_obj
=
476 (struct brw_transform_feedback_object
*) obj
;
478 /* Store the temporary ending value of the SO_NUM_PRIMS_WRITTEN counters.
479 * While this operation is paused, other transform feedback actions may
480 * occur, which will contribute to the counters. We need to exclude that
483 brw_save_primitives_written_counters(brw
, brw_obj
);
487 brw_resume_transform_feedback(struct gl_context
*ctx
,
488 struct gl_transform_feedback_object
*obj
)
490 struct brw_context
*brw
= brw_context(ctx
);
491 struct brw_transform_feedback_object
*brw_obj
=
492 (struct brw_transform_feedback_object
*) obj
;
494 /* Reload SVBI 0 with the count of vertices written so far. */
496 compute_vertices_written_so_far(brw
, brw_obj
, &svbi
);
499 OUT_BATCH(_3DSTATE_GS_SVB_INDEX
<< 16 | (4 - 2));
500 OUT_BATCH(0); /* SVBI 0 */
501 OUT_BATCH((uint32_t) svbi
); /* starting index */
502 OUT_BATCH(brw_obj
->max_index
);
505 /* Initialize the rest of the unused streams to sane values. Otherwise,
506 * they may indicate that there is no room to write data and prevent
507 * anything from happening at all.
509 for (int i
= 1; i
< 4; i
++) {
511 OUT_BATCH(_3DSTATE_GS_SVB_INDEX
<< 16 | (4 - 2));
512 OUT_BATCH(i
<< SVB_INDEX_SHIFT
);
513 OUT_BATCH(0); /* starting index */
514 OUT_BATCH(0xffffffff);
518 /* Store the new starting value of the SO_NUM_PRIMS_WRITTEN counters. */
519 brw_save_primitives_written_counters(brw
, brw_obj
);