1/*
2 * Copyright 2013 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25#include "si_build_pm4.h"
26#include "util/u_memory.h"
27#include "util/u_suballoc.h"
28
29static void si_set_streamout_enable(struct si_context *sctx, bool enable);
30
31static inline void si_so_target_reference(struct si_streamout_target **dst,
32                                          struct pipe_stream_output_target *src)
33{
34   pipe_so_target_reference((struct pipe_stream_output_target **)dst, src);
35}
36
37static struct pipe_stream_output_target *si_create_so_target(struct pipe_context *ctx,
38                                                             struct pipe_resource *buffer,
39                                                             unsigned buffer_offset,
40                                                             unsigned buffer_size)
41{
42   struct si_streamout_target *t;
43   struct si_resource *buf = si_resource(buffer);
44
45   t = CALLOC_STRUCT(si_streamout_target);
46   if (!t) {
47      return NULL;
48   }
49
50   t->b.reference.count = 1;
51   t->b.context = ctx;
52   pipe_resource_reference(&t->b.buffer, buffer);
53   t->b.buffer_offset = buffer_offset;
54   t->b.buffer_size = buffer_size;
55
56   util_range_add(&buf->b.b, &buf->valid_buffer_range, buffer_offset, buffer_offset + buffer_size);
57   return &t->b;
58}
59
60static void si_so_target_destroy(struct pipe_context *ctx, struct pipe_stream_output_target *target)
61{
62   struct si_streamout_target *t = (struct si_streamout_target *)target;
63   pipe_resource_reference(&t->b.buffer, NULL);
64   si_resource_reference(&t->buf_filled_size, NULL);
65   FREE(t);
66}
67
68void si_streamout_buffers_dirty(struct si_context *sctx)
69{
70   if (!sctx->streamout.enabled_mask)
71      return;
72
73   si_mark_atom_dirty(sctx, &sctx->atoms.s.streamout_begin);
74   si_set_streamout_enable(sctx, true);
75}
76
77static void si_set_streamout_targets(struct pipe_context *ctx, unsigned num_targets,
78                                     struct pipe_stream_output_target **targets,
79                                     const unsigned *offsets)
80{
81   struct si_context *sctx = (struct si_context *)ctx;
82   unsigned old_num_targets = sctx->streamout.num_targets;
83   unsigned i;
84   bool wait_now = false;
85
86   /* We are going to unbind the buffers. Mark which caches need to be flushed. */
87   if (sctx->streamout.num_targets && sctx->streamout.begin_emitted) {
88      /* Since streamout uses vector writes which go through TC L2
89       * and most other clients can use TC L2 as well, we don't need
90       * to flush it.
91       *
92       * The only cases which requires flushing it is VGT DMA index
93       * fetching (on <= GFX7) and indirect draw data, which are rare
94       * cases. Thus, flag the TC L2 dirtiness in the resource and
95       * handle it at draw call time.
96       */
97      for (i = 0; i < sctx->streamout.num_targets; i++)
98         if (sctx->streamout.targets[i])
99            si_resource(sctx->streamout.targets[i]->b.buffer)->TC_L2_dirty = true;
100
101      /* Invalidate the scalar cache in case a streamout buffer is
102       * going to be used as a constant buffer.
103       *
104       * Invalidate vL1, because streamout bypasses it (done by
105       * setting GLC=1 in the store instruction), but vL1 in other
106       * CUs can contain outdated data of streamout buffers.
107       *
108       * VS_PARTIAL_FLUSH is required if the buffers are going to be
109       * used as an input immediately.
110       */
111      sctx->flags |= SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE;
112
113      /* The BUFFER_FILLED_SIZE is written using a PS_DONE event. */
114      if (sctx->screen->use_ngg_streamout) {
115         sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH | SI_CONTEXT_PFP_SYNC_ME;
116
117         /* Wait now. This is needed to make sure that GDS is not
118          * busy at the end of IBs.
119          *
120          * Also, the next streamout operation will overwrite GDS,
121          * so we need to make sure that it's idle.
122          */
123         wait_now = true;
124      } else {
125         sctx->flags |= SI_CONTEXT_VS_PARTIAL_FLUSH | SI_CONTEXT_PFP_SYNC_ME;
126      }
127   }
128
129   /* All readers of the streamout targets need to be finished before we can
130    * start writing to the targets.
131    */
132   if (num_targets) {
133      sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH | SI_CONTEXT_CS_PARTIAL_FLUSH |
134                     SI_CONTEXT_PFP_SYNC_ME;
135   }
136
137   /* Streamout buffers must be bound in 2 places:
138    * 1) in VGT by setting the VGT_STRMOUT registers
139    * 2) as shader resources
140    */
141
142   /* Stop streamout. */
143   if (sctx->streamout.num_targets && sctx->streamout.begin_emitted)
144      si_emit_streamout_end(sctx);
145
146   /* Set the new targets. */
147   unsigned enabled_mask = 0, append_bitmask = 0;
148   for (i = 0; i < num_targets; i++) {
149      si_so_target_reference(&sctx->streamout.targets[i], targets[i]);
150      if (!targets[i])
151         continue;
152
153      si_context_add_resource_size(sctx, targets[i]->buffer);
154      enabled_mask |= 1 << i;
155
156      if (offsets[i] == ((unsigned)-1))
157         append_bitmask |= 1 << i;
158
159      /* Allocate space for the filled buffer size. */
160      struct si_streamout_target *t = sctx->streamout.targets[i];
161      if (!t->buf_filled_size) {
162         unsigned buf_filled_size_size = sctx->screen->use_ngg_streamout ? 8 : 4;
163         u_suballocator_alloc(&sctx->allocator_zeroed_memory, buf_filled_size_size, 4,
164                              &t->buf_filled_size_offset,
165                              (struct pipe_resource **)&t->buf_filled_size);
166      }
167   }
168
169   for (; i < sctx->streamout.num_targets; i++)
170      si_so_target_reference(&sctx->streamout.targets[i], NULL);
171
172   if (!!sctx->streamout.enabled_mask != !!enabled_mask) {
173      sctx->streamout.enabled_mask = enabled_mask;
174      sctx->do_update_shaders = true; /* to keep/remove streamout shader code as an optimization */
175   }
176
177   sctx->streamout.num_targets = num_targets;
178   sctx->streamout.append_bitmask = append_bitmask;
179
180   /* Update dirty state bits. */
181   if (num_targets) {
182      si_streamout_buffers_dirty(sctx);
183   } else {
184      si_set_atom_dirty(sctx, &sctx->atoms.s.streamout_begin, false);
185      si_set_streamout_enable(sctx, false);
186   }
187
188   /* Set the shader resources.*/
189   for (i = 0; i < num_targets; i++) {
190      if (targets[i]) {
191         struct pipe_shader_buffer sbuf;
192         sbuf.buffer = targets[i]->buffer;
193
194         if (sctx->screen->use_ngg_streamout) {
195            sbuf.buffer_offset = targets[i]->buffer_offset;
196            sbuf.buffer_size = targets[i]->buffer_size;
197         } else {
198            sbuf.buffer_offset = 0;
199            sbuf.buffer_size = targets[i]->buffer_offset + targets[i]->buffer_size;
200         }
201
202         si_set_internal_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, &sbuf);
203         si_resource(targets[i]->buffer)->bind_history |= SI_BIND_STREAMOUT_BUFFER;
204      } else {
205         si_set_internal_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, NULL);
206      }
207   }
208   for (; i < old_num_targets; i++)
209      si_set_internal_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, NULL);
210
211   if (wait_now)
212      sctx->emit_cache_flush(sctx, &sctx->gfx_cs);
213}
214
215static void si_flush_vgt_streamout(struct si_context *sctx)
216{
217   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
218   unsigned reg_strmout_cntl;
219
220   radeon_begin(cs);
221
222   /* The register is at different places on different ASICs. */
223   if (sctx->gfx_level >= GFX9) {
224      reg_strmout_cntl = R_0300FC_CP_STRMOUT_CNTL;
225      radeon_emit(PKT3(PKT3_WRITE_DATA, 3, 0));
226      radeon_emit(S_370_DST_SEL(V_370_MEM_MAPPED_REGISTER) | S_370_ENGINE_SEL(V_370_ME));
227      radeon_emit(R_0300FC_CP_STRMOUT_CNTL >> 2);
228      radeon_emit(0);
229      radeon_emit(0);
230   } else if (sctx->gfx_level >= GFX7) {
231      reg_strmout_cntl = R_0300FC_CP_STRMOUT_CNTL;
232      radeon_set_uconfig_reg(reg_strmout_cntl, 0);
233   } else {
234      reg_strmout_cntl = R_0084FC_CP_STRMOUT_CNTL;
235      radeon_set_config_reg(reg_strmout_cntl, 0);
236   }
237
238   radeon_emit(PKT3(PKT3_EVENT_WRITE, 0, 0));
239   radeon_emit(EVENT_TYPE(EVENT_TYPE_SO_VGTSTREAMOUT_FLUSH) | EVENT_INDEX(0));
240
241   radeon_emit(PKT3(PKT3_WAIT_REG_MEM, 5, 0));
242   radeon_emit(WAIT_REG_MEM_EQUAL); /* wait until the register is equal to the reference value */
243   radeon_emit(reg_strmout_cntl >> 2); /* register */
244   radeon_emit(0);
245   radeon_emit(S_0084FC_OFFSET_UPDATE_DONE(1)); /* reference value */
246   radeon_emit(S_0084FC_OFFSET_UPDATE_DONE(1)); /* mask */
247   radeon_emit(4);                              /* poll interval */
248   radeon_end();
249}
250
251static void si_emit_streamout_begin(struct si_context *sctx)
252{
253   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
254   struct si_streamout_target **t = sctx->streamout.targets;
255
256   if (!sctx->screen->use_ngg_streamout)
257      si_flush_vgt_streamout(sctx);
258
259   for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
260      if (!t[i])
261         continue;
262
263      t[i]->stride_in_dw = sctx->streamout.stride_in_dw[i];
264
265      if (sctx->screen->use_ngg_streamout) {
266         bool append = sctx->streamout.append_bitmask & (1 << i);
267         uint64_t va = 0;
268
269         if (append) {
270            radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, t[i]->buf_filled_size,
271                                      RADEON_USAGE_READ | RADEON_PRIO_SO_FILLED_SIZE);
272
273            va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
274         }
275
276         radeon_begin(cs);
277         radeon_emit(PKT3(PKT3_DMA_DATA, 5, 0));
278         radeon_emit(S_411_SRC_SEL(append ? V_411_SRC_ADDR_TC_L2 : V_411_DATA) |
279                     S_411_DST_SEL(V_411_GDS) | S_411_CP_SYNC(1));
280         radeon_emit(va);
281         radeon_emit(va >> 32);
282         radeon_emit(4 * i); /* destination in GDS */
283         radeon_emit(0);
284         radeon_emit(S_415_BYTE_COUNT_GFX9(4));
285         radeon_end();
286      } else {
287         /* Legacy streamout.
288          *
289          * The hw binds streamout buffers as shader resources. VGT only counts primitives
290          * and tells the shader through SGPRs what to do.
291          */
292         radeon_begin(cs);
293         radeon_set_context_reg_seq(R_028AD0_VGT_STRMOUT_BUFFER_SIZE_0 + 16 * i, 2);
294         radeon_emit((t[i]->b.buffer_offset + t[i]->b.buffer_size) >> 2); /* BUFFER_SIZE (in DW) */
295         radeon_emit(sctx->streamout.stride_in_dw[i]);                                    /* VTX_STRIDE (in DW) */
296
297         if (sctx->streamout.append_bitmask & (1 << i) && t[i]->buf_filled_size_valid) {
298            uint64_t va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
299
300            /* Append. */
301            radeon_emit(PKT3(PKT3_STRMOUT_BUFFER_UPDATE, 4, 0));
302            radeon_emit(STRMOUT_SELECT_BUFFER(i) |
303                        STRMOUT_OFFSET_SOURCE(STRMOUT_OFFSET_FROM_MEM)); /* control */
304            radeon_emit(0);                                              /* unused */
305            radeon_emit(0);                                              /* unused */
306            radeon_emit(va);                                             /* src address lo */
307            radeon_emit(va >> 32);                                       /* src address hi */
308
309            radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, t[i]->buf_filled_size,
310                                      RADEON_USAGE_READ | RADEON_PRIO_SO_FILLED_SIZE);
311         } else {
312            /* Start from the beginning. */
313            radeon_emit(PKT3(PKT3_STRMOUT_BUFFER_UPDATE, 4, 0));
314            radeon_emit(STRMOUT_SELECT_BUFFER(i) |
315                        STRMOUT_OFFSET_SOURCE(STRMOUT_OFFSET_FROM_PACKET)); /* control */
316            radeon_emit(0);                                                 /* unused */
317            radeon_emit(0);                                                 /* unused */
318            radeon_emit(t[i]->b.buffer_offset >> 2); /* buffer offset in DW */
319            radeon_emit(0);                          /* unused */
320         }
321         radeon_end_update_context_roll(sctx);
322      }
323   }
324
325   sctx->streamout.begin_emitted = true;
326}
327
328void si_emit_streamout_end(struct si_context *sctx)
329{
330   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
331   struct si_streamout_target **t = sctx->streamout.targets;
332
333   if (!sctx->screen->use_ngg_streamout)
334      si_flush_vgt_streamout(sctx);
335
336   for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
337      if (!t[i])
338         continue;
339
340      uint64_t va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
341
342      if (sctx->screen->use_ngg_streamout) {
343         /* TODO: PS_DONE doesn't ensure completion of VS if there are no PS waves. */
344         si_cp_release_mem(sctx, &sctx->gfx_cs, V_028A90_PS_DONE, 0, EOP_DST_SEL_TC_L2,
345                           EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM, EOP_DATA_SEL_GDS,
346                           t[i]->buf_filled_size, va, EOP_DATA_GDS(i, 1), 0);
347      } else {
348         radeon_begin(cs);
349         radeon_emit(PKT3(PKT3_STRMOUT_BUFFER_UPDATE, 4, 0));
350         radeon_emit(STRMOUT_SELECT_BUFFER(i) | STRMOUT_OFFSET_SOURCE(STRMOUT_OFFSET_NONE) |
351                     STRMOUT_STORE_BUFFER_FILLED_SIZE); /* control */
352         radeon_emit(va);                                  /* dst address lo */
353         radeon_emit(va >> 32);                            /* dst address hi */
354         radeon_emit(0);                                   /* unused */
355         radeon_emit(0);                                   /* unused */
356
357         /* Zero the buffer size. The counters (primitives generated,
358          * primitives emitted) may be enabled even if there is not
359          * buffer bound. This ensures that the primitives-emitted query
360          * won't increment. */
361         radeon_set_context_reg(R_028AD0_VGT_STRMOUT_BUFFER_SIZE_0 + 16 * i, 0);
362         radeon_end_update_context_roll(sctx);
363
364         radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, t[i]->buf_filled_size,
365                                   RADEON_USAGE_WRITE | RADEON_PRIO_SO_FILLED_SIZE);
366      }
367
368      t[i]->buf_filled_size_valid = true;
369   }
370
371   sctx->streamout.begin_emitted = false;
372}
373
374/* STREAMOUT CONFIG DERIVED STATE
375 *
376 * Streamout must be enabled for the PRIMITIVES_GENERATED query to work.
377 * The buffer mask is an independent state, so no writes occur if there
378 * are no buffers bound.
379 */
380
381static void si_emit_streamout_enable(struct si_context *sctx)
382{
383   assert(!sctx->screen->use_ngg_streamout);
384
385   radeon_begin(&sctx->gfx_cs);
386   radeon_set_context_reg_seq(R_028B94_VGT_STRMOUT_CONFIG, 2);
387   radeon_emit(S_028B94_STREAMOUT_0_EN(si_get_strmout_en(sctx)) |
388               S_028B94_RAST_STREAM(0) |
389               S_028B94_STREAMOUT_1_EN(si_get_strmout_en(sctx)) |
390               S_028B94_STREAMOUT_2_EN(si_get_strmout_en(sctx)) |
391               S_028B94_STREAMOUT_3_EN(si_get_strmout_en(sctx)));
392   radeon_emit(sctx->streamout.hw_enabled_mask & sctx->streamout.enabled_stream_buffers_mask);
393   radeon_end();
394}
395
396static void si_set_streamout_enable(struct si_context *sctx, bool enable)
397{
398   bool old_strmout_en = si_get_strmout_en(sctx);
399   unsigned old_hw_enabled_mask = sctx->streamout.hw_enabled_mask;
400
401   sctx->streamout.streamout_enabled = enable;
402
403   sctx->streamout.hw_enabled_mask =
404      sctx->streamout.enabled_mask | (sctx->streamout.enabled_mask << 4) |
405      (sctx->streamout.enabled_mask << 8) | (sctx->streamout.enabled_mask << 12);
406
407   if (!sctx->screen->use_ngg_streamout &&
408       ((old_strmout_en != si_get_strmout_en(sctx)) ||
409        (old_hw_enabled_mask != sctx->streamout.hw_enabled_mask)))
410      si_mark_atom_dirty(sctx, &sctx->atoms.s.streamout_enable);
411}
412
413void si_update_prims_generated_query_state(struct si_context *sctx, unsigned type, int diff)
414{
415   if (!sctx->screen->use_ngg_streamout && type == PIPE_QUERY_PRIMITIVES_GENERATED) {
416      bool old_strmout_en = si_get_strmout_en(sctx);
417
418      sctx->streamout.num_prims_gen_queries += diff;
419      assert(sctx->streamout.num_prims_gen_queries >= 0);
420
421      sctx->streamout.prims_gen_query_enabled = sctx->streamout.num_prims_gen_queries != 0;
422
423      if (old_strmout_en != si_get_strmout_en(sctx))
424         si_mark_atom_dirty(sctx, &sctx->atoms.s.streamout_enable);
425
426      if (si_update_ngg(sctx)) {
427         si_shader_change_notify(sctx);
428         sctx->do_update_shaders = true;
429      }
430   }
431}
432
433void si_init_streamout_functions(struct si_context *sctx)
434{
435   sctx->b.create_stream_output_target = si_create_so_target;
436   sctx->b.stream_output_target_destroy = si_so_target_destroy;
437   sctx->b.set_stream_output_targets = si_set_streamout_targets;
438   sctx->atoms.s.streamout_begin.emit = si_emit_streamout_begin;
439
440   if (!sctx->screen->use_ngg_streamout)
441      sctx->atoms.s.streamout_enable.emit = si_emit_streamout_enable;
442}
443