1 /*
2 * Copyright © 2011 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 /** \file gen6_sol.c
25 *
26 * Code to initialize the binding table entries used by transform feedback.
27 */
28
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"
36
37 static void
gen6_update_sol_surfaces(struct brw_context * brw)38 gen6_update_sol_surfaces(struct brw_context *brw)
39 {
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;
44
45 if (xfb_active) {
46 /* BRW_NEW_TRANSFORM_FEEDBACK */
47 xfb_obj = ctx->TransformFeedback.CurrentObject;
48 linked_xfb_info = xfb_obj->program->sh.LinkedTransformFeedback;
49 }
50
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->programs[MESA_SHADER_GEOMETRY]) {
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);
64 } else {
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);
70 }
71 } else {
72 if (!brw->programs[MESA_SHADER_GEOMETRY])
73 brw->ff_gs.surf_offset[surf_index] = 0;
74 else
75 brw->gs.base.surf_offset[surf_index] = 0;
76 }
77 }
78
79 brw->ctx.NewDriverState |= BRW_NEW_SURFACES;
80 }
81
82 const struct brw_tracked_state gen6_sol_surface = {
83 .dirty = {
84 .mesa = 0,
85 .brw = BRW_NEW_BATCH |
86 BRW_NEW_BLORP |
87 BRW_NEW_TRANSFORM_FEEDBACK,
88 },
89 .emit = gen6_update_sol_surfaces,
90 };
91
92 /**
93 * Constructs the binding table for the WM surface state, which maps unit
94 * numbers to surface state objects.
95 */
96 static void
brw_gs_upload_binding_table(struct brw_context * brw)97 brw_gs_upload_binding_table(struct brw_context *brw)
98 {
99 uint32_t *bind;
100 struct gl_context *ctx = &brw->ctx;
101 const struct gl_program *prog;
102 bool need_binding_table = false;
103
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->programs[MESA_SHADER_GEOMETRY] == NULL).
107 * In this case, we only need surfaces for transform feedback in the
108 * GS stage.
109 * 2) We have a user-provided geometry shader. In this case we may need
110 * surfaces for transform feedback and/or other stuff, like textures,
111 * in the GS stage.
112 */
113
114 if (!brw->programs[MESA_SHADER_GEOMETRY]) {
115 /* BRW_NEW_VERTEX_PROGRAM */
116 prog = ctx->_Shader->CurrentProgram[MESA_SHADER_VERTEX];
117 if (prog) {
118 /* Skip making a binding table if we don't have anything to put in it */
119 const struct gl_transform_feedback_info *linked_xfb_info =
120 prog->sh.LinkedTransformFeedback;
121 need_binding_table = linked_xfb_info->NumOutputs > 0;
122 }
123 if (!need_binding_table) {
124 if (brw->ff_gs.bind_bo_offset != 0) {
125 brw->ctx.NewDriverState |= BRW_NEW_BINDING_TABLE_POINTERS;
126 brw->ff_gs.bind_bo_offset = 0;
127 }
128 return;
129 }
130
131 /* Might want to calculate nr_surfaces first, to avoid taking up so much
132 * space for the binding table. Anyway, in this case we know that we only
133 * use BRW_MAX_SOL_BINDINGS surfaces at most.
134 */
135 bind = brw_state_batch(brw, sizeof(uint32_t) * BRW_MAX_SOL_BINDINGS,
136 32, &brw->ff_gs.bind_bo_offset);
137
138 /* BRW_NEW_SURFACES */
139 memcpy(bind, brw->ff_gs.surf_offset,
140 BRW_MAX_SOL_BINDINGS * sizeof(uint32_t));
141 } else {
142 /* BRW_NEW_GEOMETRY_PROGRAM */
143 prog = ctx->_Shader->CurrentProgram[MESA_SHADER_GEOMETRY];
144 if (prog) {
145 /* Skip making a binding table if we don't have anything to put in it */
146 struct brw_stage_prog_data *prog_data = brw->gs.base.prog_data;
147 const struct gl_transform_feedback_info *linked_xfb_info =
148 prog->sh.LinkedTransformFeedback;
149 need_binding_table = linked_xfb_info->NumOutputs > 0 ||
150 prog_data->binding_table.size_bytes > 0;
151 }
152 if (!need_binding_table) {
153 if (brw->gs.base.bind_bo_offset != 0) {
154 brw->gs.base.bind_bo_offset = 0;
155 brw->ctx.NewDriverState |= BRW_NEW_BINDING_TABLE_POINTERS;
156 }
157 return;
158 }
159
160 /* Might want to calculate nr_surfaces first, to avoid taking up so much
161 * space for the binding table.
162 */
163 bind = brw_state_batch(brw, sizeof(uint32_t) * BRW_MAX_SURFACES,
164 32, &brw->gs.base.bind_bo_offset);
165
166 /* BRW_NEW_SURFACES */
167 memcpy(bind, brw->gs.base.surf_offset,
168 BRW_MAX_SURFACES * sizeof(uint32_t));
169 }
170
171 brw->ctx.NewDriverState |= BRW_NEW_BINDING_TABLE_POINTERS;
172 }
173
174 const struct brw_tracked_state gen6_gs_binding_table = {
175 .dirty = {
176 .mesa = 0,
177 .brw = BRW_NEW_BATCH |
178 BRW_NEW_BLORP |
179 BRW_NEW_GEOMETRY_PROGRAM |
180 BRW_NEW_VERTEX_PROGRAM |
181 BRW_NEW_SURFACES,
182 },
183 .emit = brw_gs_upload_binding_table,
184 };
185
186 struct gl_transform_feedback_object *
brw_new_transform_feedback(struct gl_context * ctx,GLuint name)187 brw_new_transform_feedback(struct gl_context *ctx, GLuint name)
188 {
189 struct brw_context *brw = brw_context(ctx);
190 struct brw_transform_feedback_object *brw_obj =
191 CALLOC_STRUCT(brw_transform_feedback_object);
192 if (!brw_obj)
193 return NULL;
194
195 _mesa_init_transform_feedback_object(&brw_obj->base, name);
196
197 brw_obj->offset_bo =
198 brw_bo_alloc(brw->bufmgr, "transform feedback offsets", 16, 64);
199 brw_obj->prim_count_bo =
200 brw_bo_alloc(brw->bufmgr, "xfb primitive counts", 16384, 64);
201
202 return &brw_obj->base;
203 }
204
205 void
brw_delete_transform_feedback(struct gl_context * ctx,struct gl_transform_feedback_object * obj)206 brw_delete_transform_feedback(struct gl_context *ctx,
207 struct gl_transform_feedback_object *obj)
208 {
209 struct brw_transform_feedback_object *brw_obj =
210 (struct brw_transform_feedback_object *) obj;
211
212 for (unsigned i = 0; i < ARRAY_SIZE(obj->Buffers); i++) {
213 _mesa_reference_buffer_object(ctx, &obj->Buffers[i], NULL);
214 }
215
216 brw_bo_unreference(brw_obj->offset_bo);
217 brw_bo_unreference(brw_obj->prim_count_bo);
218
219 free(brw_obj);
220 }
221
222 /**
223 * Tally the number of primitives generated so far.
224 *
225 * The buffer contains a series of pairs:
226 * (<start0, start1, start2, start3>, <end0, end1, end2, end3>) ;
227 * (<start0, start1, start2, start3>, <end0, end1, end2, end3>) ;
228 *
229 * For each stream, we subtract the pair of values (end - start) to get the
230 * number of primitives generated during one section. We accumulate these
231 * values, adding them up to get the total number of primitives generated.
232 *
233 * Note that we expose one stream pre-Gen7, so the above is just (start, end).
234 */
235 static void
aggregate_transform_feedback_counter(struct brw_context * brw,struct brw_bo * bo,struct brw_transform_feedback_counter * counter)236 aggregate_transform_feedback_counter(
237 struct brw_context *brw,
238 struct brw_bo *bo,
239 struct brw_transform_feedback_counter *counter)
240 {
241 const unsigned streams = brw->ctx.Const.MaxVertexStreams;
242
243 /* If the current batch is still contributing to the number of primitives
244 * generated, flush it now so the results will be present when mapped.
245 */
246 if (brw_batch_references(&brw->batch, bo))
247 intel_batchbuffer_flush(brw);
248
249 if (unlikely(brw->perf_debug && brw_bo_busy(bo)))
250 perf_debug("Stalling for # of transform feedback primitives written.\n");
251
252 uint64_t *prim_counts = brw_bo_map(brw, bo, MAP_READ);
253 prim_counts += counter->bo_start * streams;
254
255 for (unsigned i = counter->bo_start; i + 1 < counter->bo_end; i += 2) {
256 for (unsigned s = 0; s < streams; s++)
257 counter->accum[s] += prim_counts[streams + s] - prim_counts[s];
258
259 prim_counts += 2 * streams;
260 }
261
262 brw_bo_unmap(bo);
263
264 /* We've already gathered up the old data; we can safely overwrite it now. */
265 counter->bo_start = counter->bo_end = 0;
266 }
267
268 /**
269 * Store the SO_NUM_PRIMS_WRITTEN counters for each stream (4 uint64_t values)
270 * to prim_count_bo.
271 *
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.
274 *
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().
278 */
279 void
brw_save_primitives_written_counters(struct brw_context * brw,struct brw_transform_feedback_object * obj)280 brw_save_primitives_written_counters(struct brw_context *brw,
281 struct brw_transform_feedback_object *obj)
282 {
283 const struct gen_device_info *devinfo = &brw->screen->devinfo;
284 const struct gl_context *ctx = &brw->ctx;
285 const int streams = ctx->Const.MaxVertexStreams;
286
287 assert(obj->prim_count_bo != NULL);
288
289 /* Check if there's enough space for a new pair of four values. */
290 if ((obj->counter.bo_end + 2) * streams * sizeof(uint64_t) >=
291 obj->prim_count_bo->size) {
292 aggregate_transform_feedback_counter(brw, obj->prim_count_bo,
293 &obj->previous_counter);
294 aggregate_transform_feedback_counter(brw, obj->prim_count_bo,
295 &obj->counter);
296 }
297
298 /* Flush any drawing so that the counters have the right values. */
299 brw_emit_mi_flush(brw);
300
301 /* Emit MI_STORE_REGISTER_MEM commands to write the values. */
302 if (devinfo->gen >= 7) {
303 for (int i = 0; i < streams; i++) {
304 int offset = (streams * obj->counter.bo_end + i) * sizeof(uint64_t);
305 brw_store_register_mem64(brw, obj->prim_count_bo,
306 GEN7_SO_NUM_PRIMS_WRITTEN(i),
307 offset);
308 }
309 } else {
310 brw_store_register_mem64(brw, obj->prim_count_bo,
311 GEN6_SO_NUM_PRIMS_WRITTEN,
312 obj->counter.bo_end * sizeof(uint64_t));
313 }
314
315 /* Update where to write data to. */
316 obj->counter.bo_end++;
317 }
318
319 static void
compute_vertices_written_so_far(struct brw_context * brw,struct brw_transform_feedback_object * obj,struct brw_transform_feedback_counter * counter,uint64_t * vertices_written)320 compute_vertices_written_so_far(struct brw_context *brw,
321 struct brw_transform_feedback_object *obj,
322 struct brw_transform_feedback_counter *counter,
323 uint64_t *vertices_written)
324 {
325 const struct gl_context *ctx = &brw->ctx;
326 unsigned vertices_per_prim = 0;
327
328 switch (obj->primitive_mode) {
329 case GL_POINTS:
330 vertices_per_prim = 1;
331 break;
332 case GL_LINES:
333 vertices_per_prim = 2;
334 break;
335 case GL_TRIANGLES:
336 vertices_per_prim = 3;
337 break;
338 default:
339 unreachable("Invalid transform feedback primitive mode.");
340 }
341
342 /* Get the number of primitives generated. */
343 aggregate_transform_feedback_counter(brw, obj->prim_count_bo, counter);
344
345 for (int i = 0; i < ctx->Const.MaxVertexStreams; i++) {
346 vertices_written[i] = vertices_per_prim * counter->accum[i];
347 }
348 }
349
350 /**
351 * Compute the number of vertices written by the last transform feedback
352 * begin/end block.
353 */
354 static void
compute_xfb_vertices_written(struct brw_context * brw,struct brw_transform_feedback_object * obj)355 compute_xfb_vertices_written(struct brw_context *brw,
356 struct brw_transform_feedback_object *obj)
357 {
358 if (obj->vertices_written_valid || !obj->base.EndedAnytime)
359 return;
360
361 compute_vertices_written_so_far(brw, obj, &obj->previous_counter,
362 obj->vertices_written);
363 obj->vertices_written_valid = true;
364 }
365
366 /**
367 * GetTransformFeedbackVertexCount() driver hook.
368 *
369 * Returns the number of vertices written to a particular stream by the last
370 * Begin/EndTransformFeedback block. Used to implement DrawTransformFeedback().
371 */
372 GLsizei
brw_get_transform_feedback_vertex_count(struct gl_context * ctx,struct gl_transform_feedback_object * obj,GLuint stream)373 brw_get_transform_feedback_vertex_count(struct gl_context *ctx,
374 struct gl_transform_feedback_object *obj,
375 GLuint stream)
376 {
377 struct brw_context *brw = brw_context(ctx);
378 struct brw_transform_feedback_object *brw_obj =
379 (struct brw_transform_feedback_object *) obj;
380
381 assert(obj->EndedAnytime);
382 assert(stream < ctx->Const.MaxVertexStreams);
383
384 compute_xfb_vertices_written(brw, brw_obj);
385 return brw_obj->vertices_written[stream];
386 }
387
388 void
brw_begin_transform_feedback(struct gl_context * ctx,GLenum mode,struct gl_transform_feedback_object * obj)389 brw_begin_transform_feedback(struct gl_context *ctx, GLenum mode,
390 struct gl_transform_feedback_object *obj)
391 {
392 struct brw_context *brw = brw_context(ctx);
393 const struct gl_program *prog;
394 const struct gl_transform_feedback_info *linked_xfb_info;
395 struct gl_transform_feedback_object *xfb_obj =
396 ctx->TransformFeedback.CurrentObject;
397 struct brw_transform_feedback_object *brw_obj =
398 (struct brw_transform_feedback_object *) xfb_obj;
399
400 assert(brw->screen->devinfo.gen == 6);
401
402 if (ctx->_Shader->CurrentProgram[MESA_SHADER_GEOMETRY]) {
403 /* BRW_NEW_GEOMETRY_PROGRAM */
404 prog = ctx->_Shader->CurrentProgram[MESA_SHADER_GEOMETRY];
405 } else {
406 /* BRW_NEW_VERTEX_PROGRAM */
407 prog = ctx->_Shader->CurrentProgram[MESA_SHADER_VERTEX];
408 }
409 linked_xfb_info = prog->sh.LinkedTransformFeedback;
410
411 /* Compute the maximum number of vertices that we can write without
412 * overflowing any of the buffers currently being used for feedback.
413 */
414 brw_obj->max_index
415 = _mesa_compute_max_transform_feedback_vertices(ctx, xfb_obj,
416 linked_xfb_info);
417
418 /* Initialize the SVBI 0 register to zero and set the maximum index. */
419 BEGIN_BATCH(4);
420 OUT_BATCH(_3DSTATE_GS_SVB_INDEX << 16 | (4 - 2));
421 OUT_BATCH(0); /* SVBI 0 */
422 OUT_BATCH(0); /* starting index */
423 OUT_BATCH(brw_obj->max_index);
424 ADVANCE_BATCH();
425
426 /* Initialize the rest of the unused streams to sane values. Otherwise,
427 * they may indicate that there is no room to write data and prevent
428 * anything from happening at all.
429 */
430 for (int i = 1; i < 4; i++) {
431 BEGIN_BATCH(4);
432 OUT_BATCH(_3DSTATE_GS_SVB_INDEX << 16 | (4 - 2));
433 OUT_BATCH(i << SVB_INDEX_SHIFT);
434 OUT_BATCH(0); /* starting index */
435 OUT_BATCH(0xffffffff);
436 ADVANCE_BATCH();
437 }
438
439 /* Store the starting value of the SO_NUM_PRIMS_WRITTEN counters. */
440 brw_save_primitives_written_counters(brw, brw_obj);
441
442 brw_obj->primitive_mode = mode;
443 }
444
445 void
brw_end_transform_feedback(struct gl_context * ctx,struct gl_transform_feedback_object * obj)446 brw_end_transform_feedback(struct gl_context *ctx,
447 struct gl_transform_feedback_object *obj)
448 {
449 struct brw_context *brw = brw_context(ctx);
450 struct brw_transform_feedback_object *brw_obj =
451 (struct brw_transform_feedback_object *) obj;
452
453 /* Store the ending value of the SO_NUM_PRIMS_WRITTEN counters. */
454 if (!obj->Paused)
455 brw_save_primitives_written_counters(brw, brw_obj);
456
457 /* We've reached the end of a transform feedback begin/end block. This
458 * means that future DrawTransformFeedback() calls will need to pick up the
459 * results of the current counter, and that it's time to roll back the
460 * current primitive counter to zero.
461 */
462 brw_obj->previous_counter = brw_obj->counter;
463 brw_reset_transform_feedback_counter(&brw_obj->counter);
464
465 /* EndTransformFeedback() means that we need to update the number of
466 * vertices written. Since it's only necessary if DrawTransformFeedback()
467 * is called and it means mapping a buffer object, we delay computing it
468 * until it's absolutely necessary to try and avoid stalls.
469 */
470 brw_obj->vertices_written_valid = false;
471 }
472
473 void
brw_pause_transform_feedback(struct gl_context * ctx,struct gl_transform_feedback_object * obj)474 brw_pause_transform_feedback(struct gl_context *ctx,
475 struct gl_transform_feedback_object *obj)
476 {
477 struct brw_context *brw = brw_context(ctx);
478 struct brw_transform_feedback_object *brw_obj =
479 (struct brw_transform_feedback_object *) obj;
480
481 /* Store the temporary ending value of the SO_NUM_PRIMS_WRITTEN counters.
482 * While this operation is paused, other transform feedback actions may
483 * occur, which will contribute to the counters. We need to exclude that
484 * from our counts.
485 */
486 brw_save_primitives_written_counters(brw, brw_obj);
487 }
488
489 void
brw_resume_transform_feedback(struct gl_context * ctx,struct gl_transform_feedback_object * obj)490 brw_resume_transform_feedback(struct gl_context *ctx,
491 struct gl_transform_feedback_object *obj)
492 {
493 struct brw_context *brw = brw_context(ctx);
494 struct brw_transform_feedback_object *brw_obj =
495 (struct brw_transform_feedback_object *) obj;
496
497 /* Reload SVBI 0 with the count of vertices written so far. */
498 uint64_t svbi;
499 compute_vertices_written_so_far(brw, brw_obj, &brw_obj->counter, &svbi);
500
501 BEGIN_BATCH(4);
502 OUT_BATCH(_3DSTATE_GS_SVB_INDEX << 16 | (4 - 2));
503 OUT_BATCH(0); /* SVBI 0 */
504 OUT_BATCH((uint32_t) svbi); /* starting index */
505 OUT_BATCH(brw_obj->max_index);
506 ADVANCE_BATCH();
507
508 /* Initialize the rest of the unused streams to sane values. Otherwise,
509 * they may indicate that there is no room to write data and prevent
510 * anything from happening at all.
511 */
512 for (int i = 1; i < 4; i++) {
513 BEGIN_BATCH(4);
514 OUT_BATCH(_3DSTATE_GS_SVB_INDEX << 16 | (4 - 2));
515 OUT_BATCH(i << SVB_INDEX_SHIFT);
516 OUT_BATCH(0); /* starting index */
517 OUT_BATCH(0xffffffff);
518 ADVANCE_BATCH();
519 }
520
521 /* Store the new starting value of the SO_NUM_PRIMS_WRITTEN counters. */
522 brw_save_primitives_written_counters(brw, brw_obj);
523 }
524