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      if (sctx->screen->use_ngg_streamout)
134         si_allocate_gds(sctx);
135
136      sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH | SI_CONTEXT_CS_PARTIAL_FLUSH |
137                     SI_CONTEXT_PFP_SYNC_ME;
138   }
139
140   /* Streamout buffers must be bound in 2 places:
141    * 1) in VGT by setting the VGT_STRMOUT registers
142    * 2) as shader resources
143    */
144
145   /* Stop streamout. */
146   if (sctx->streamout.num_targets && sctx->streamout.begin_emitted)
147      si_emit_streamout_end(sctx);
148
149   /* Set the new targets. */
150   unsigned enabled_mask = 0, append_bitmask = 0;
151   for (i = 0; i < num_targets; i++) {
152      si_so_target_reference(&sctx->streamout.targets[i], targets[i]);
153      if (!targets[i])
154         continue;
155
156      si_context_add_resource_size(sctx, targets[i]->buffer);
157      enabled_mask |= 1 << i;
158
159      if (offsets[i] == ((unsigned)-1))
160         append_bitmask |= 1 << i;
161
162      /* Allocate space for the filled buffer size. */
163      struct si_streamout_target *t = sctx->streamout.targets[i];
164      if (!t->buf_filled_size) {
165         unsigned buf_filled_size_size = sctx->screen->use_ngg_streamout ? 8 : 4;
166         u_suballocator_alloc(&sctx->allocator_zeroed_memory, buf_filled_size_size, 4,
167                              &t->buf_filled_size_offset,
168                              (struct pipe_resource **)&t->buf_filled_size);
169      }
170   }
171
172   for (; i < sctx->streamout.num_targets; i++)
173      si_so_target_reference(&sctx->streamout.targets[i], NULL);
174
175   sctx->streamout.enabled_mask = enabled_mask;
176   sctx->streamout.num_targets = num_targets;
177   sctx->streamout.append_bitmask = append_bitmask;
178
179   /* Update dirty state bits. */
180   if (num_targets) {
181      si_streamout_buffers_dirty(sctx);
182   } else {
183      si_set_atom_dirty(sctx, &sctx->atoms.s.streamout_begin, false);
184      si_set_streamout_enable(sctx, false);
185   }
186
187   /* Set the shader resources.*/
188   for (i = 0; i < num_targets; i++) {
189      if (targets[i]) {
190         struct pipe_shader_buffer sbuf;
191         sbuf.buffer = targets[i]->buffer;
192
193         if (sctx->screen->use_ngg_streamout) {
194            sbuf.buffer_offset = targets[i]->buffer_offset;
195            sbuf.buffer_size = targets[i]->buffer_size;
196         } else {
197            sbuf.buffer_offset = 0;
198            sbuf.buffer_size = targets[i]->buffer_offset + targets[i]->buffer_size;
199         }
200
201         si_set_internal_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, &sbuf);
202         si_resource(targets[i]->buffer)->bind_history |= PIPE_BIND_STREAM_OUTPUT;
203      } else {
204         si_set_internal_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, NULL);
205      }
206   }
207   for (; i < old_num_targets; i++)
208      si_set_internal_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, NULL);
209
210   if (wait_now)
211      sctx->emit_cache_flush(sctx, &sctx->gfx_cs);
212}
213
214static void gfx10_emit_streamout_begin(struct si_context *sctx)
215{
216   struct si_streamout_target **t = sctx->streamout.targets;
217   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
218   unsigned last_target = 0;
219
220   for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
221      if (t[i])
222         last_target = i;
223   }
224
225   radeon_begin(cs);
226
227   for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
228      if (!t[i])
229         continue;
230
231      t[i]->stride_in_dw = sctx->streamout.stride_in_dw[i];
232
233      bool append = sctx->streamout.append_bitmask & (1 << i);
234      uint64_t va = 0;
235
236      if (append) {
237         radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, t[i]->buf_filled_size, RADEON_USAGE_READ,
238                                   RADEON_PRIO_SO_FILLED_SIZE);
239
240         va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
241      }
242
243      radeon_emit(PKT3(PKT3_DMA_DATA, 5, 0));
244      radeon_emit(S_411_SRC_SEL(append ? V_411_SRC_ADDR_TC_L2 : V_411_DATA) |
245                  S_411_DST_SEL(V_411_GDS) | S_411_CP_SYNC(i == last_target));
246      radeon_emit(va);
247      radeon_emit(va >> 32);
248      radeon_emit(4 * i); /* destination in GDS */
249      radeon_emit(0);
250      radeon_emit(S_415_BYTE_COUNT_GFX9(4) | S_415_DISABLE_WR_CONFIRM_GFX9(i != last_target));
251   }
252   radeon_end();
253
254   sctx->streamout.begin_emitted = true;
255}
256
257static void gfx10_emit_streamout_end(struct si_context *sctx)
258{
259   struct si_streamout_target **t = sctx->streamout.targets;
260
261   for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
262      if (!t[i])
263         continue;
264
265      uint64_t va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
266
267      si_cp_release_mem(sctx, &sctx->gfx_cs, V_028A90_PS_DONE, 0, EOP_DST_SEL_TC_L2,
268                        EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM, EOP_DATA_SEL_GDS,
269                        t[i]->buf_filled_size, va, EOP_DATA_GDS(i, 1), 0);
270
271      t[i]->buf_filled_size_valid = true;
272   }
273
274   sctx->streamout.begin_emitted = false;
275}
276
277static void si_flush_vgt_streamout(struct si_context *sctx)
278{
279   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
280   unsigned reg_strmout_cntl;
281
282   radeon_begin(cs);
283
284   /* The register is at different places on different ASICs. */
285   if (sctx->chip_class >= GFX7) {
286      reg_strmout_cntl = R_0300FC_CP_STRMOUT_CNTL;
287      radeon_set_uconfig_reg(reg_strmout_cntl, 0);
288   } else {
289      reg_strmout_cntl = R_0084FC_CP_STRMOUT_CNTL;
290      radeon_set_config_reg(reg_strmout_cntl, 0);
291   }
292
293   radeon_emit(PKT3(PKT3_EVENT_WRITE, 0, 0));
294   radeon_emit(EVENT_TYPE(EVENT_TYPE_SO_VGTSTREAMOUT_FLUSH) | EVENT_INDEX(0));
295
296   radeon_emit(PKT3(PKT3_WAIT_REG_MEM, 5, 0));
297   radeon_emit(WAIT_REG_MEM_EQUAL); /* wait until the register is equal to the reference value */
298   radeon_emit(reg_strmout_cntl >> 2); /* register */
299   radeon_emit(0);
300   radeon_emit(S_0084FC_OFFSET_UPDATE_DONE(1)); /* reference value */
301   radeon_emit(S_0084FC_OFFSET_UPDATE_DONE(1)); /* mask */
302   radeon_emit(4);                              /* poll interval */
303   radeon_end();
304}
305
306static void si_emit_streamout_begin(struct si_context *sctx)
307{
308   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
309   struct si_streamout_target **t = sctx->streamout.targets;
310   uint16_t *stride_in_dw = sctx->streamout.stride_in_dw;
311   unsigned i;
312
313   si_flush_vgt_streamout(sctx);
314
315   radeon_begin(cs);
316
317   for (i = 0; i < sctx->streamout.num_targets; i++) {
318      if (!t[i])
319         continue;
320
321      t[i]->stride_in_dw = stride_in_dw[i];
322
323      /* AMD GCN binds streamout buffers as shader resources.
324       * VGT only counts primitives and tells the shader
325       * through SGPRs what to do. */
326      radeon_set_context_reg_seq(R_028AD0_VGT_STRMOUT_BUFFER_SIZE_0 + 16 * i, 2);
327      radeon_emit((t[i]->b.buffer_offset + t[i]->b.buffer_size) >> 2); /* BUFFER_SIZE (in DW) */
328      radeon_emit(stride_in_dw[i]);                                    /* VTX_STRIDE (in DW) */
329
330      if (sctx->streamout.append_bitmask & (1 << i) && t[i]->buf_filled_size_valid) {
331         uint64_t va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
332
333         /* Append. */
334         radeon_emit(PKT3(PKT3_STRMOUT_BUFFER_UPDATE, 4, 0));
335         radeon_emit(STRMOUT_SELECT_BUFFER(i) |
336                     STRMOUT_OFFSET_SOURCE(STRMOUT_OFFSET_FROM_MEM)); /* control */
337         radeon_emit(0);                                              /* unused */
338         radeon_emit(0);                                              /* unused */
339         radeon_emit(va);                                             /* src address lo */
340         radeon_emit(va >> 32);                                       /* src address hi */
341
342         radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, t[i]->buf_filled_size, RADEON_USAGE_READ,
343                                   RADEON_PRIO_SO_FILLED_SIZE);
344      } else {
345         /* Start from the beginning. */
346         radeon_emit(PKT3(PKT3_STRMOUT_BUFFER_UPDATE, 4, 0));
347         radeon_emit(STRMOUT_SELECT_BUFFER(i) |
348                     STRMOUT_OFFSET_SOURCE(STRMOUT_OFFSET_FROM_PACKET)); /* control */
349         radeon_emit(0);                                                 /* unused */
350         radeon_emit(0);                                                 /* unused */
351         radeon_emit(t[i]->b.buffer_offset >> 2); /* buffer offset in DW */
352         radeon_emit(0);                          /* unused */
353      }
354   }
355   radeon_end();
356
357   sctx->streamout.begin_emitted = true;
358}
359
360void si_emit_streamout_end(struct si_context *sctx)
361{
362   if (sctx->screen->use_ngg_streamout) {
363      gfx10_emit_streamout_end(sctx);
364      return;
365   }
366
367   struct radeon_cmdbuf *cs = &sctx->gfx_cs;
368   struct si_streamout_target **t = sctx->streamout.targets;
369   unsigned i;
370   uint64_t va;
371
372   si_flush_vgt_streamout(sctx);
373
374   radeon_begin(cs);
375
376   for (i = 0; i < sctx->streamout.num_targets; i++) {
377      if (!t[i])
378         continue;
379
380      va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
381      radeon_emit(PKT3(PKT3_STRMOUT_BUFFER_UPDATE, 4, 0));
382      radeon_emit(STRMOUT_SELECT_BUFFER(i) | STRMOUT_OFFSET_SOURCE(STRMOUT_OFFSET_NONE) |
383                  STRMOUT_STORE_BUFFER_FILLED_SIZE); /* control */
384      radeon_emit(va);                                  /* dst address lo */
385      radeon_emit(va >> 32);                            /* dst address hi */
386      radeon_emit(0);                                   /* unused */
387      radeon_emit(0);                                   /* unused */
388
389      radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, t[i]->buf_filled_size, RADEON_USAGE_WRITE,
390                                RADEON_PRIO_SO_FILLED_SIZE);
391
392      /* Zero the buffer size. The counters (primitives generated,
393       * primitives emitted) may be enabled even if there is not
394       * buffer bound. This ensures that the primitives-emitted query
395       * won't increment. */
396      radeon_set_context_reg(R_028AD0_VGT_STRMOUT_BUFFER_SIZE_0 + 16 * i, 0);
397
398      t[i]->buf_filled_size_valid = true;
399   }
400   radeon_end_update_context_roll(sctx);
401
402   sctx->streamout.begin_emitted = false;
403}
404
405/* STREAMOUT CONFIG DERIVED STATE
406 *
407 * Streamout must be enabled for the PRIMITIVES_GENERATED query to work.
408 * The buffer mask is an independent state, so no writes occur if there
409 * are no buffers bound.
410 */
411
412static void si_emit_streamout_enable(struct si_context *sctx)
413{
414   assert(!sctx->screen->use_ngg_streamout);
415
416   radeon_begin(&sctx->gfx_cs);
417   radeon_set_context_reg_seq(R_028B94_VGT_STRMOUT_CONFIG, 2);
418   radeon_emit(S_028B94_STREAMOUT_0_EN(si_get_strmout_en(sctx)) |
419               S_028B94_RAST_STREAM(0) |
420               S_028B94_STREAMOUT_1_EN(si_get_strmout_en(sctx)) |
421               S_028B94_STREAMOUT_2_EN(si_get_strmout_en(sctx)) |
422               S_028B94_STREAMOUT_3_EN(si_get_strmout_en(sctx)));
423   radeon_emit(sctx->streamout.hw_enabled_mask & sctx->streamout.enabled_stream_buffers_mask);
424   radeon_end();
425}
426
427static void si_set_streamout_enable(struct si_context *sctx, bool enable)
428{
429   bool old_strmout_en = si_get_strmout_en(sctx);
430   unsigned old_hw_enabled_mask = sctx->streamout.hw_enabled_mask;
431
432   sctx->streamout.streamout_enabled = enable;
433
434   sctx->streamout.hw_enabled_mask =
435      sctx->streamout.enabled_mask | (sctx->streamout.enabled_mask << 4) |
436      (sctx->streamout.enabled_mask << 8) | (sctx->streamout.enabled_mask << 12);
437
438   if (!sctx->screen->use_ngg_streamout &&
439       ((old_strmout_en != si_get_strmout_en(sctx)) ||
440        (old_hw_enabled_mask != sctx->streamout.hw_enabled_mask)))
441      si_mark_atom_dirty(sctx, &sctx->atoms.s.streamout_enable);
442}
443
444void si_update_prims_generated_query_state(struct si_context *sctx, unsigned type, int diff)
445{
446   if (!sctx->screen->use_ngg_streamout && type == PIPE_QUERY_PRIMITIVES_GENERATED) {
447      bool old_strmout_en = si_get_strmout_en(sctx);
448
449      sctx->streamout.num_prims_gen_queries += diff;
450      assert(sctx->streamout.num_prims_gen_queries >= 0);
451
452      sctx->streamout.prims_gen_query_enabled = sctx->streamout.num_prims_gen_queries != 0;
453
454      if (old_strmout_en != si_get_strmout_en(sctx))
455         si_mark_atom_dirty(sctx, &sctx->atoms.s.streamout_enable);
456
457      if (si_update_ngg(sctx)) {
458         si_shader_change_notify(sctx);
459         sctx->do_update_shaders = true;
460      }
461   }
462}
463
464void si_init_streamout_functions(struct si_context *sctx)
465{
466   sctx->b.create_stream_output_target = si_create_so_target;
467   sctx->b.stream_output_target_destroy = si_so_target_destroy;
468   sctx->b.set_stream_output_targets = si_set_streamout_targets;
469
470   if (sctx->screen->use_ngg_streamout) {
471      sctx->atoms.s.streamout_begin.emit = gfx10_emit_streamout_begin;
472   } else {
473      sctx->atoms.s.streamout_begin.emit = si_emit_streamout_begin;
474      sctx->atoms.s.streamout_enable.emit = si_emit_streamout_enable;
475   }
476}
477