1/*
2 * Copyright © 2017 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 shall be included
12 * in all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
21 */
22
23/**
24 * @file crocus_pipe_control.c
25 *
26 * PIPE_CONTROL is the main flushing and synchronization primitive on Intel
27 * GPUs.  It can invalidate caches, stall until rendering reaches various
28 * stages of completion, write to memory, and other things.  In a way, it's
29 * a swiss army knife command - it has all kinds of capabilities, but some
30 * significant limitations as well.
31 *
32 * Unfortunately, it's notoriously complicated and difficult to use.  Many
33 * sub-commands can't be used together.  Some are meant to be used at the
34 * top of the pipeline (invalidating caches before drawing), while some are
35 * meant to be used at the end (stalling or flushing after drawing).
36 *
37 * Also, there's a list of restrictions a mile long, which vary by generation.
38 * Do this before doing that, or suffer the consequences (usually a GPU hang).
39 *
40 * This file contains helpers for emitting them safely.  You can simply call
41 * crocus_emit_pipe_control_flush() with the desired operations (as logical
42 * PIPE_CONTROL_* bits), and it will take care of splitting it into multiple
43 * PIPE_CONTROL commands as necessary.  The per-generation workarounds are
44 * applied in crocus_emit_raw_pipe_control() in crocus_state.c.
45 */
46
47#include "crocus_context.h"
48#include "util/hash_table.h"
49#include "util/set.h"
50
51/**
52 * Emit a PIPE_CONTROL with various flushing flags.
53 *
54 * The caller is responsible for deciding what flags are appropriate for the
55 * given generation.
56 */
57void
58crocus_emit_pipe_control_flush(struct crocus_batch *batch,
59                               const char *reason,
60                               uint32_t flags)
61{
62   const struct intel_device_info *devinfo = &batch->screen->devinfo;
63
64   if (devinfo->ver >= 6 &&
65       (flags & PIPE_CONTROL_CACHE_FLUSH_BITS) &&
66       (flags & PIPE_CONTROL_CACHE_INVALIDATE_BITS)) {
67      /* A pipe control command with flush and invalidate bits set
68       * simultaneously is an inherently racy operation on Gen6+ if the
69       * contents of the flushed caches were intended to become visible from
70       * any of the invalidated caches.  Split it in two PIPE_CONTROLs, the
71       * first one should stall the pipeline to make sure that the flushed R/W
72       * caches are coherent with memory once the specified R/O caches are
73       * invalidated.  On pre-Gen6 hardware the (implicit) R/O cache
74       * invalidation seems to happen at the bottom of the pipeline together
75       * with any write cache flush, so this shouldn't be a concern.  In order
76       * to ensure a full stall, we do an end-of-pipe sync.
77       */
78      crocus_emit_end_of_pipe_sync(batch, reason,
79                                   flags & PIPE_CONTROL_CACHE_FLUSH_BITS);
80      flags &= ~(PIPE_CONTROL_CACHE_FLUSH_BITS | PIPE_CONTROL_CS_STALL);
81   }
82
83   batch->screen->vtbl.emit_raw_pipe_control(batch, reason, flags, NULL, 0, 0);
84}
85
86/**
87 * Emit a PIPE_CONTROL that writes to a buffer object.
88 *
89 * \p flags should contain one of the following items:
90 *  - PIPE_CONTROL_WRITE_IMMEDIATE
91 *  - PIPE_CONTROL_WRITE_TIMESTAMP
92 *  - PIPE_CONTROL_WRITE_DEPTH_COUNT
93 */
94void
95crocus_emit_pipe_control_write(struct crocus_batch *batch,
96                               const char *reason, uint32_t flags,
97                               struct crocus_bo *bo, uint32_t offset,
98                               uint64_t imm)
99{
100   batch->screen->vtbl.emit_raw_pipe_control(batch, reason, flags, bo, offset, imm);
101}
102
103/**
104 * Restriction [DevSNB, DevIVB]:
105 *
106 * Prior to changing Depth/Stencil Buffer state (i.e. any combination of
107 * 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS, 3DSTATE_STENCIL_BUFFER,
108 * 3DSTATE_HIER_DEPTH_BUFFER) SW must first issue a pipelined depth stall
109 * (PIPE_CONTROL with Depth Stall bit set), followed by a pipelined depth
110 * cache flush (PIPE_CONTROL with Depth Flush Bit set), followed by
111 * another pipelined depth stall (PIPE_CONTROL with Depth Stall bit set),
112 * unless SW can otherwise guarantee that the pipeline from WM onwards is
113 * already flushed (e.g., via a preceding MI_FLUSH).
114 */
115void
116crocus_emit_depth_stall_flushes(struct crocus_batch *batch)
117{
118   UNUSED const struct intel_device_info *devinfo = &batch->screen->devinfo;
119
120   assert(devinfo->ver >= 6);
121
122   /* Starting on BDW, these pipe controls are unnecessary.
123    *
124    *   WM HW will internally manage the draining pipe and flushing of the caches
125    *   when this command is issued. The PIPE_CONTROL restrictions are removed.
126    */
127   if (devinfo->ver >= 8)
128      return;
129
130   crocus_emit_pipe_control_flush(batch, "depth stall", PIPE_CONTROL_DEPTH_STALL);
131   crocus_emit_pipe_control_flush(batch, "depth stall", PIPE_CONTROL_DEPTH_CACHE_FLUSH);
132   crocus_emit_pipe_control_flush(batch, "depth stall", PIPE_CONTROL_DEPTH_STALL);
133}
134
135/*
136 * From Sandybridge PRM, volume 2, "1.7.2 End-of-Pipe Synchronization":
137 *
138 *  Write synchronization is a special case of end-of-pipe
139 *  synchronization that requires that the render cache and/or depth
140 *  related caches are flushed to memory, where the data will become
141 *  globally visible. This type of synchronization is required prior to
142 *  SW (CPU) actually reading the result data from memory, or initiating
143 *  an operation that will use as a read surface (such as a texture
144 *  surface) a previous render target and/or depth/stencil buffer
145 *
146 * From Haswell PRM, volume 2, part 1, "End-of-Pipe Synchronization":
147 *
148 *  Exercising the write cache flush bits (Render Target Cache Flush
149 *  Enable, Depth Cache Flush Enable, DC Flush) in PIPE_CONTROL only
150 *  ensures the write caches are flushed and doesn't guarantee the data
151 *  is globally visible.
152 *
153 *  SW can track the completion of the end-of-pipe-synchronization by
154 *  using "Notify Enable" and "PostSync Operation - Write Immediate
155 *  Data" in the PIPE_CONTROL command.
156 */
157void
158crocus_emit_end_of_pipe_sync(struct crocus_batch *batch,
159                             const char *reason, uint32_t flags)
160{
161   const struct intel_device_info *devinfo = &batch->screen->devinfo;
162
163   if (devinfo->ver >= 6) {
164      /* From Sandybridge PRM, volume 2, "1.7.3.1 Writing a Value to Memory":
165       *
166       *    "The most common action to perform upon reaching a synchronization
167       *    point is to write a value out to memory. An immediate value
168       *    (included with the synchronization command) may be written."
169       *
170       * From Broadwell PRM, volume 7, "End-of-Pipe Synchronization":
171       *
172       *    "In case the data flushed out by the render engine is to be read
173       *    back in to the render engine in coherent manner, then the render
174       *    engine has to wait for the fence completion before accessing the
175       *    flushed data. This can be achieved by following means on various
176       *    products: PIPE_CONTROL command with CS Stall and the required
177       *    write caches flushed with Post-Sync-Operation as Write Immediate
178       *    Data.
179       *
180       *    Example:
181       *       - Workload-1 (3D/GPGPU/MEDIA)
182       *       - PIPE_CONTROL (CS Stall, Post-Sync-Operation Write Immediate
183       *         Data, Required Write Cache Flush bits set)
184       *       - Workload-2 (Can use the data produce or output by Workload-1)
185       */
186      crocus_emit_pipe_control_write(batch, reason,
187                                     flags | PIPE_CONTROL_CS_STALL |
188                                     PIPE_CONTROL_WRITE_IMMEDIATE,
189                                     batch->ice->workaround_bo,
190                                     batch->ice->workaround_offset, 0);
191
192      if (batch->screen->devinfo.platform == INTEL_PLATFORM_HSW) {
193#define GEN7_3DPRIM_START_INSTANCE      0x243C
194         batch->screen->vtbl.load_register_mem32(batch, GEN7_3DPRIM_START_INSTANCE,
195                                                 batch->ice->workaround_bo,
196                                                 batch->ice->workaround_offset);
197      }
198   } else {
199      /* On gen4-5, a regular pipe control seems to suffice. */
200      crocus_emit_pipe_control_flush(batch, reason, flags);
201   }
202}
203
204/* Emit a pipelined flush to either flush render and texture cache for
205 * reading from a FBO-drawn texture, or flush so that frontbuffer
206 * render appears on the screen in DRI1.
207 *
208 * This is also used for the always_flush_cache driconf debug option.
209 */
210void
211crocus_emit_mi_flush(struct crocus_batch *batch)
212{
213   const struct intel_device_info *devinfo = &batch->screen->devinfo;
214   int flags = PIPE_CONTROL_RENDER_TARGET_FLUSH;
215   if (devinfo->ver >= 6) {
216      flags |= PIPE_CONTROL_INSTRUCTION_INVALIDATE |
217               PIPE_CONTROL_CONST_CACHE_INVALIDATE |
218               PIPE_CONTROL_DATA_CACHE_FLUSH |
219               PIPE_CONTROL_DEPTH_CACHE_FLUSH |
220               PIPE_CONTROL_VF_CACHE_INVALIDATE |
221               PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
222               PIPE_CONTROL_CS_STALL;
223   }
224   crocus_emit_pipe_control_flush(batch, "mi flush", flags);
225}
226
227/**
228 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
229 * implementing two workarounds on gen6.  From section 1.4.7.1
230 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
231 *
232 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
233 * produced by non-pipelined state commands), software needs to first
234 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
235 * 0.
236 *
237 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
238 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
239 *
240 * And the workaround for these two requires this workaround first:
241 *
242 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
243 * BEFORE the pipe-control with a post-sync op and no write-cache
244 * flushes.
245 *
246 * And this last workaround is tricky because of the requirements on
247 * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
248 * volume 2 part 1:
249 *
250 *     "1 of the following must also be set:
251 *      - Render Target Cache Flush Enable ([12] of DW1)
252 *      - Depth Cache Flush Enable ([0] of DW1)
253 *      - Stall at Pixel Scoreboard ([1] of DW1)
254 *      - Depth Stall ([13] of DW1)
255 *      - Post-Sync Operation ([13] of DW1)
256 *      - Notify Enable ([8] of DW1)"
257 *
258 * The cache flushes require the workaround flush that triggered this
259 * one, so we can't use it.  Depth stall would trigger the same.
260 * Post-sync nonzero is what triggered this second workaround, so we
261 * can't use that one either.  Notify enable is IRQs, which aren't
262 * really our business.  That leaves only stall at scoreboard.
263 */
264void
265crocus_emit_post_sync_nonzero_flush(struct crocus_batch *batch)
266{
267   crocus_emit_pipe_control_flush(batch, "nonzero",
268                                  PIPE_CONTROL_CS_STALL |
269                                  PIPE_CONTROL_STALL_AT_SCOREBOARD);
270
271   crocus_emit_pipe_control_write(batch, "nonzero",
272                                  PIPE_CONTROL_WRITE_IMMEDIATE,
273                                  batch->ice->workaround_bo,
274                                  batch->ice->workaround_offset, 0);
275}
276
277/**
278 * Flush and invalidate all caches (for debugging purposes).
279 */
280void
281crocus_flush_all_caches(struct crocus_batch *batch)
282{
283   crocus_emit_pipe_control_flush(batch, "debug: flush all caches",
284                                  PIPE_CONTROL_CS_STALL |
285                                  PIPE_CONTROL_DATA_CACHE_FLUSH |
286                                  PIPE_CONTROL_DEPTH_CACHE_FLUSH |
287                                  PIPE_CONTROL_RENDER_TARGET_FLUSH |
288                                  PIPE_CONTROL_VF_CACHE_INVALIDATE |
289                                  PIPE_CONTROL_INSTRUCTION_INVALIDATE |
290                                  PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
291                                  PIPE_CONTROL_CONST_CACHE_INVALIDATE |
292                                  PIPE_CONTROL_STATE_CACHE_INVALIDATE);
293}
294
295static void
296crocus_texture_barrier(struct pipe_context *ctx, unsigned flags)
297{
298   struct crocus_context *ice = (void *) ctx;
299   struct crocus_batch *render_batch = &ice->batches[CROCUS_BATCH_RENDER];
300   struct crocus_batch *compute_batch = &ice->batches[CROCUS_BATCH_COMPUTE];
301   const struct intel_device_info *devinfo = &render_batch->screen->devinfo;
302
303   if (devinfo->ver < 6) {
304      crocus_emit_mi_flush(render_batch);
305      return;
306   }
307
308   if (render_batch->contains_draw) {
309      crocus_batch_maybe_flush(render_batch, 48);
310      crocus_emit_pipe_control_flush(render_batch,
311                                     "API: texture barrier (1/2)",
312                                     (flags == 1 ? PIPE_CONTROL_DEPTH_CACHE_FLUSH  : 0) |
313                                     PIPE_CONTROL_RENDER_TARGET_FLUSH |
314                                     PIPE_CONTROL_CS_STALL);
315      crocus_emit_pipe_control_flush(render_batch,
316                                     "API: texture barrier (2/2)",
317                                     PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
318   }
319
320   if (compute_batch->contains_draw) {
321      crocus_batch_maybe_flush(compute_batch, 48);
322      crocus_emit_pipe_control_flush(compute_batch,
323                                     "API: texture barrier (1/2)",
324                                     PIPE_CONTROL_CS_STALL);
325      crocus_emit_pipe_control_flush(compute_batch,
326                                     "API: texture barrier (2/2)",
327                                     PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
328   }
329}
330
331static void
332crocus_memory_barrier(struct pipe_context *ctx, unsigned flags)
333{
334   struct crocus_context *ice = (void *) ctx;
335   unsigned bits = PIPE_CONTROL_DATA_CACHE_FLUSH | PIPE_CONTROL_CS_STALL;
336   const struct intel_device_info *devinfo = &ice->batches[0].screen->devinfo;
337
338   assert(devinfo->ver >= 7);
339
340   if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
341                PIPE_BARRIER_INDEX_BUFFER |
342                PIPE_BARRIER_INDIRECT_BUFFER)) {
343      bits |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
344   }
345
346   if (flags & PIPE_BARRIER_CONSTANT_BUFFER) {
347      bits |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
348              PIPE_CONTROL_CONST_CACHE_INVALIDATE;
349   }
350
351   if (flags & (PIPE_BARRIER_TEXTURE | PIPE_BARRIER_FRAMEBUFFER)) {
352      bits |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
353              PIPE_CONTROL_RENDER_TARGET_FLUSH;
354   }
355
356   /* Typed surface messages are handled by the render cache on IVB, so we
357    * need to flush it too.
358    */
359   if (devinfo->verx10 < 75)
360      bits |= PIPE_CONTROL_RENDER_TARGET_FLUSH;
361
362   for (int i = 0; i < ice->batch_count; i++) {
363      if (ice->batches[i].contains_draw) {
364         crocus_batch_maybe_flush(&ice->batches[i], 24);
365         crocus_emit_pipe_control_flush(&ice->batches[i], "API: memory barrier",
366                                        bits);
367      }
368   }
369}
370
371void
372crocus_init_flush_functions(struct pipe_context *ctx)
373{
374   ctx->memory_barrier = crocus_memory_barrier;
375   ctx->texture_barrier = crocus_texture_barrier;
376}
377