2 * Copyright © 2010 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
24 #include "brw_context.h"
25 #include "intel_batchbuffer.h"
26 #include "intel_fbo.h"
27 #include "intel_reg.h"
30 * According to the latest documentation, any PIPE_CONTROL with the
31 * "Command Streamer Stall" bit set must also have another bit set,
32 * with five different options:
34 * - Render Target Cache Flush
36 * - Stall at Pixel Scoreboard
37 * - Post-Sync Operation
41 * I chose "Stall at Pixel Scoreboard" since we've used it effectively
42 * in the past, but the choice is fairly arbitrary.
45 gen8_add_cs_stall_workaround_bits(uint32_t *flags)
47 uint32_t wa_bits = PIPE_CONTROL_RENDER_TARGET_FLUSH |
48 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
49 PIPE_CONTROL_WRITE_IMMEDIATE |
50 PIPE_CONTROL_WRITE_DEPTH_COUNT |
51 PIPE_CONTROL_WRITE_TIMESTAMP |
52 PIPE_CONTROL_STALL_AT_SCOREBOARD |
53 PIPE_CONTROL_DEPTH_STALL |
54 PIPE_CONTROL_DATA_CACHE_FLUSH;
56 /* If we're doing a CS stall, and don't already have one of the
57 * workaround bits set, add "Stall at Pixel Scoreboard."
59 if ((*flags & PIPE_CONTROL_CS_STALL) != 0 && (*flags & wa_bits) == 0)
60 *flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
63 /* Implement the WaCsStallAtEveryFourthPipecontrol workaround on IVB, BYT:
65 * "Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with
66 * only read-cache-invalidate bit(s) set, must have a CS_STALL bit set."
68 * Note that the kernel does CS stalls between batches, so we only need
69 * to count them within a batch.
72 gen7_cs_stall_every_four_pipe_controls(struct brw_context *brw, uint32_t flags)
74 if (brw->gen == 7 && !brw->is_haswell) {
75 if (flags & PIPE_CONTROL_CS_STALL) {
76 /* If we're doing a CS stall, reset the counter and carry on. */
77 brw->pipe_controls_since_last_cs_stall = 0;
81 /* If this is the fourth pipe control without a CS stall, do one now. */
82 if (++brw->pipe_controls_since_last_cs_stall == 4) {
83 brw->pipe_controls_since_last_cs_stall = 0;
84 return PIPE_CONTROL_CS_STALL;
91 * Emit a PIPE_CONTROL with various flushing flags.
93 * The caller is responsible for deciding what flags are appropriate for the
97 brw_emit_pipe_control_flush(struct brw_context *brw, uint32_t flags)
100 (flags & PIPE_CONTROL_CACHE_FLUSH_BITS) &&
101 (flags & PIPE_CONTROL_CACHE_INVALIDATE_BITS)) {
102 /* A pipe control command with flush and invalidate bits set
103 * simultaneously is an inherently racy operation on Gen6+ if the
104 * contents of the flushed caches were intended to become visible from
105 * any of the invalidated caches. Split it in two PIPE_CONTROLs, the
106 * first one should stall the pipeline to make sure that the flushed R/W
107 * caches are coherent with memory once the specified R/O caches are
108 * invalidated. On pre-Gen6 hardware the (implicit) R/O cache
109 * invalidation seems to happen at the bottom of the pipeline together
110 * with any write cache flush, so this shouldn't be a concern.
112 brw_emit_pipe_control_flush(brw, (flags & PIPE_CONTROL_CACHE_FLUSH_BITS) |
113 PIPE_CONTROL_CS_STALL);
114 flags &= ~(PIPE_CONTROL_CACHE_FLUSH_BITS | PIPE_CONTROL_CS_STALL);
119 gen8_add_cs_stall_workaround_bits(&flags);
122 (flags & PIPE_CONTROL_VF_CACHE_INVALIDATE)) {
123 /* Hardware workaround: SKL
125 * Emit Pipe Control with all bits set to zero before emitting
126 * a Pipe Control with VF Cache Invalidate set.
128 brw_emit_pipe_control_flush(brw, 0);
132 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (6 - 2));
139 } else if (brw->gen >= 6) {
140 flags |= gen7_cs_stall_every_four_pipe_controls(brw, flags);
143 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
151 OUT_BATCH(_3DSTATE_PIPE_CONTROL | flags | (4 - 2));
160 * Emit a PIPE_CONTROL that writes to a buffer object.
162 * \p flags should contain one of the following items:
163 * - PIPE_CONTROL_WRITE_IMMEDIATE
164 * - PIPE_CONTROL_WRITE_TIMESTAMP
165 * - PIPE_CONTROL_WRITE_DEPTH_COUNT
168 brw_emit_pipe_control_write(struct brw_context *brw, uint32_t flags,
169 drm_intel_bo *bo, uint32_t offset,
170 uint32_t imm_lower, uint32_t imm_upper)
174 gen8_add_cs_stall_workaround_bits(&flags);
177 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (6 - 2));
179 OUT_RELOC64(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
181 OUT_BATCH(imm_lower);
182 OUT_BATCH(imm_upper);
184 } else if (brw->gen >= 6) {
185 flags |= gen7_cs_stall_every_four_pipe_controls(brw, flags);
187 /* PPGTT/GGTT is selected by DW2 bit 2 on Sandybridge, but DW1 bit 24
188 * on later platforms. We always use PPGTT on Gen7+.
190 unsigned gen6_gtt = brw->gen == 6 ? PIPE_CONTROL_GLOBAL_GTT_WRITE : 0;
193 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
195 OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
197 OUT_BATCH(imm_lower);
198 OUT_BATCH(imm_upper);
202 OUT_BATCH(_3DSTATE_PIPE_CONTROL | flags | (4 - 2));
203 OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
204 PIPE_CONTROL_GLOBAL_GTT_WRITE | offset);
205 OUT_BATCH(imm_lower);
206 OUT_BATCH(imm_upper);
212 * Restriction [DevSNB, DevIVB]:
214 * Prior to changing Depth/Stencil Buffer state (i.e. any combination of
215 * 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS, 3DSTATE_STENCIL_BUFFER,
216 * 3DSTATE_HIER_DEPTH_BUFFER) SW must first issue a pipelined depth stall
217 * (PIPE_CONTROL with Depth Stall bit set), followed by a pipelined depth
218 * cache flush (PIPE_CONTROL with Depth Flush Bit set), followed by
219 * another pipelined depth stall (PIPE_CONTROL with Depth Stall bit set),
220 * unless SW can otherwise guarantee that the pipeline from WM onwards is
221 * already flushed (e.g., via a preceding MI_FLUSH).
224 brw_emit_depth_stall_flushes(struct brw_context *brw)
226 assert(brw->gen >= 6 && brw->gen <= 9);
228 /* Starting on BDW, these pipe controls are unnecessary.
230 * WM HW will internally manage the draining pipe and flushing of the caches
231 * when this command is issued. The PIPE_CONTROL restrictions are removed.
236 brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_STALL);
237 brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_CACHE_FLUSH);
238 brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_STALL);
242 * From the Ivybridge PRM, Volume 2 Part 1, Section 3.2 (VS Stage Input):
243 * "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth
244 * stall needs to be sent just prior to any 3DSTATE_VS, 3DSTATE_URB_VS,
245 * 3DSTATE_CONSTANT_VS, 3DSTATE_BINDING_TABLE_POINTER_VS,
246 * 3DSTATE_SAMPLER_STATE_POINTER_VS command. Only one PIPE_CONTROL needs
247 * to be sent before any combination of VS associated 3DSTATE."
250 gen7_emit_vs_workaround_flush(struct brw_context *brw)
252 assert(brw->gen == 7);
253 brw_emit_pipe_control_write(brw,
254 PIPE_CONTROL_WRITE_IMMEDIATE
255 | PIPE_CONTROL_DEPTH_STALL,
256 brw->workaround_bo, 0,
262 * Emit a PIPE_CONTROL command for gen7 with the CS Stall bit set.
265 gen7_emit_cs_stall_flush(struct brw_context *brw)
267 brw_emit_pipe_control_write(brw,
268 PIPE_CONTROL_CS_STALL
269 | PIPE_CONTROL_WRITE_IMMEDIATE,
270 brw->workaround_bo, 0,
276 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
277 * implementing two workarounds on gen6. From section 1.4.7.1
278 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
280 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
281 * produced by non-pipelined state commands), software needs to first
282 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
285 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
286 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
288 * And the workaround for these two requires this workaround first:
290 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
291 * BEFORE the pipe-control with a post-sync op and no write-cache
294 * And this last workaround is tricky because of the requirements on
295 * that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
298 * "1 of the following must also be set:
299 * - Render Target Cache Flush Enable ([12] of DW1)
300 * - Depth Cache Flush Enable ([0] of DW1)
301 * - Stall at Pixel Scoreboard ([1] of DW1)
302 * - Depth Stall ([13] of DW1)
303 * - Post-Sync Operation ([13] of DW1)
304 * - Notify Enable ([8] of DW1)"
306 * The cache flushes require the workaround flush that triggered this
307 * one, so we can't use it. Depth stall would trigger the same.
308 * Post-sync nonzero is what triggered this second workaround, so we
309 * can't use that one either. Notify enable is IRQs, which aren't
310 * really our business. That leaves only stall at scoreboard.
313 brw_emit_post_sync_nonzero_flush(struct brw_context *brw)
315 brw_emit_pipe_control_flush(brw,
316 PIPE_CONTROL_CS_STALL |
317 PIPE_CONTROL_STALL_AT_SCOREBOARD);
319 brw_emit_pipe_control_write(brw, PIPE_CONTROL_WRITE_IMMEDIATE,
320 brw->workaround_bo, 0, 0, 0);
323 /* Emit a pipelined flush to either flush render and texture cache for
324 * reading from a FBO-drawn texture, or flush so that frontbuffer
325 * render appears on the screen in DRI1.
327 * This is also used for the always_flush_cache driconf debug option.
330 brw_emit_mi_flush(struct brw_context *brw)
332 if (brw->batch.ring == BLT_RING && brw->gen >= 6) {
334 OUT_BATCH(MI_FLUSH_DW);
340 int flags = PIPE_CONTROL_NO_WRITE | PIPE_CONTROL_RENDER_TARGET_FLUSH;
342 flags |= PIPE_CONTROL_INSTRUCTION_INVALIDATE |
343 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
344 PIPE_CONTROL_VF_CACHE_INVALIDATE |
345 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
346 PIPE_CONTROL_CS_STALL;
349 /* Hardware workaround: SNB B-Spec says:
351 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache
352 * Flush Enable =1, a PIPE_CONTROL with any non-zero
353 * post-sync-op is required.
355 brw_emit_post_sync_nonzero_flush(brw);
358 brw_emit_pipe_control_flush(brw, flags);
363 brw_init_pipe_control(struct brw_context *brw,
364 const struct brw_device_info *devinfo)
366 if (devinfo->gen < 6)
369 /* We can't just use brw_state_batch to get a chunk of space for
370 * the gen6 workaround because it involves actually writing to
371 * the buffer, and the kernel doesn't let us write to the batch.
373 brw->workaround_bo = drm_intel_bo_alloc(brw->bufmgr,
374 "pipe_control workaround",
376 if (brw->workaround_bo == NULL)
379 brw->pipe_controls_since_last_cs_stall = 0;
385 brw_fini_pipe_control(struct brw_context *brw)
387 drm_intel_bo_unreference(brw->workaround_bo);