1 /**************************************************************************
3 * Copyright 2006 Tungsten Graphics, Inc., Cedar Park, Texas.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 #include "intel_context.h"
29 #include "intel_batchbuffer.h"
30 #include "intel_buffer_objects.h"
31 #include "intel_decode.h"
32 #include "intel_reg.h"
33 #include "intel_bufmgr.h"
34 #include "intel_buffers.h"
36 struct cached_batch_item {
37 struct cached_batch_item *next;
42 static void clear_cache( struct intel_context *intel )
44 struct cached_batch_item *item = intel->batch.cached_items;
47 struct cached_batch_item *next = item->next;
52 intel->batch.cached_items = NULL;
56 intel_batchbuffer_init(struct intel_context *intel)
58 intel_batchbuffer_reset(intel);
60 if (intel->gen == 6) {
61 /* We can't just use brw_state_batch to get a chunk of space for
62 * the gen6 workaround because it involves actually writing to
63 * the buffer, and the kernel doesn't let us write to the batch.
65 intel->batch.workaround_bo = drm_intel_bo_alloc(intel->bufmgr,
72 intel_batchbuffer_reset(struct intel_context *intel)
74 if (intel->batch.last_bo != NULL) {
75 drm_intel_bo_unreference(intel->batch.last_bo);
76 intel->batch.last_bo = NULL;
78 intel->batch.last_bo = intel->batch.bo;
82 intel->batch.bo = drm_intel_bo_alloc(intel->bufmgr, "batchbuffer",
83 intel->maxBatchSize, 4096);
85 intel->batch.reserved_space = BATCH_RESERVED;
86 intel->batch.state_batch_offset = intel->batch.bo->size;
87 intel->batch.used = 0;
91 intel_batchbuffer_free(struct intel_context *intel)
93 drm_intel_bo_unreference(intel->batch.last_bo);
94 drm_intel_bo_unreference(intel->batch.bo);
95 drm_intel_bo_unreference(intel->batch.workaround_bo);
100 /* TODO: Push this whole function into bufmgr.
103 do_flush_locked(struct intel_context *intel)
105 struct intel_batchbuffer *batch = &intel->batch;
108 if (!intel->intelScreen->no_hw) {
111 if (intel->gen < 6 || !batch->is_blit) {
112 ring = I915_EXEC_RENDER;
114 ring = I915_EXEC_BLT;
117 ret = drm_intel_bo_subdata(batch->bo, 0, 4*batch->used, batch->map);
118 if (ret == 0 && batch->state_batch_offset != batch->bo->size) {
119 ret = drm_intel_bo_subdata(batch->bo,
120 batch->state_batch_offset,
121 batch->bo->size - batch->state_batch_offset,
122 (char *)batch->map + batch->state_batch_offset);
126 ret = drm_intel_bo_mrb_exec(batch->bo, 4*batch->used, NULL, 0, 0, ring);
129 if (unlikely(INTEL_DEBUG & DEBUG_BATCH)) {
130 intel_decode(batch->map, batch->used,
132 intel->intelScreen->deviceID, GL_TRUE);
134 if (intel->vtbl.debug_batch != NULL)
135 intel->vtbl.debug_batch(intel);
141 intel->vtbl.new_batch(intel);
145 _intel_batchbuffer_flush(struct intel_context *intel,
146 const char *file, int line)
148 if (intel->batch.used == 0)
151 if (intel->first_post_swapbuffers_batch == NULL) {
152 intel->first_post_swapbuffers_batch = intel->batch.bo;
153 drm_intel_bo_reference(intel->first_post_swapbuffers_batch);
156 if (unlikely(INTEL_DEBUG & DEBUG_BATCH))
157 fprintf(stderr, "%s:%d: Batchbuffer flush with %db used\n", file, line,
158 4*intel->batch.used);
160 intel->batch.reserved_space = 0;
162 if (intel->always_flush_cache) {
163 intel_batchbuffer_emit_mi_flush(intel);
166 /* Mark the end of the buffer. */
167 intel_batchbuffer_emit_dword(intel, MI_BATCH_BUFFER_END);
168 if (intel->batch.used & 1) {
169 /* Round batchbuffer usage to 2 DWORDs. */
170 intel_batchbuffer_emit_dword(intel, MI_NOOP);
173 if (intel->vtbl.finish_batch)
174 intel->vtbl.finish_batch(intel);
176 intel_upload_finish(intel);
178 /* Check that we didn't just wrap our batchbuffer at a bad time. */
179 assert(!intel->no_batch_wrap);
181 do_flush_locked(intel);
183 if (unlikely(INTEL_DEBUG & DEBUG_SYNC)) {
184 fprintf(stderr, "waiting for idle\n");
185 drm_intel_bo_wait_rendering(intel->batch.bo);
190 intel_batchbuffer_reset(intel);
194 /* This is the only way buffers get added to the validate list.
197 intel_batchbuffer_emit_reloc(struct intel_context *intel,
198 drm_intel_bo *buffer,
199 uint32_t read_domains, uint32_t write_domain,
204 ret = drm_intel_bo_emit_reloc(intel->batch.bo, 4*intel->batch.used,
206 read_domains, write_domain);
211 * Using the old buffer offset, write in what the right data would be, in case
212 * the buffer doesn't move and we can short-circuit the relocation processing
215 intel_batchbuffer_emit_dword(intel, buffer->offset + delta);
221 intel_batchbuffer_emit_reloc_fenced(struct intel_context *intel,
222 drm_intel_bo *buffer,
223 uint32_t read_domains,
224 uint32_t write_domain,
229 ret = drm_intel_bo_emit_reloc_fence(intel->batch.bo, 4*intel->batch.used,
231 read_domains, write_domain);
236 * Using the old buffer offset, write in what the right data would
237 * be, in case the buffer doesn't move and we can short-circuit the
238 * relocation processing in the kernel
240 intel_batchbuffer_emit_dword(intel, buffer->offset + delta);
246 intel_batchbuffer_data(struct intel_context *intel,
247 const void *data, GLuint bytes, bool is_blit)
249 assert((bytes & 3) == 0);
250 intel_batchbuffer_require_space(intel, bytes, is_blit);
251 __memcpy(intel->batch.map + intel->batch.used, data, bytes);
252 intel->batch.used += bytes >> 2;
256 intel_batchbuffer_cached_advance(struct intel_context *intel)
258 struct cached_batch_item **prev = &intel->batch.cached_items, *item;
259 uint32_t sz = (intel->batch.used - intel->batch.emit) * sizeof(uint32_t);
260 uint32_t *start = intel->batch.map + intel->batch.emit;
261 uint16_t op = *start >> 16;
267 old = intel->batch.map + item->header;
268 if (op == *old >> 16) {
269 if (item->size == sz && memcmp(old, start, sz) == 0) {
270 if (prev != &intel->batch.cached_items) {
272 item->next = intel->batch.cached_items;
273 intel->batch.cached_items = item;
275 intel->batch.used = intel->batch.emit;
284 item = malloc(sizeof(struct cached_batch_item));
288 item->next = intel->batch.cached_items;
289 intel->batch.cached_items = item;
293 item->header = intel->batch.emit;
297 * Restriction [DevSNB, DevIVB]:
299 * Prior to changing Depth/Stencil Buffer state (i.e. any combination of
300 * 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS, 3DSTATE_STENCIL_BUFFER,
301 * 3DSTATE_HIER_DEPTH_BUFFER) SW must first issue a pipelined depth stall
302 * (PIPE_CONTROL with Depth Stall bit set), followed by a pipelined depth
303 * cache flush (PIPE_CONTROL with Depth Flush Bit set), followed by
304 * another pipelined depth stall (PIPE_CONTROL with Depth Stall bit set),
305 * unless SW can otherwise guarantee that the pipeline from WM onwards is
306 * already flushed (e.g., via a preceding MI_FLUSH).
309 intel_emit_depth_stall_flushes(struct intel_context *intel)
311 assert(intel->gen >= 6 && intel->gen <= 7);
314 OUT_BATCH(_3DSTATE_PIPE_CONTROL);
315 OUT_BATCH(PIPE_CONTROL_DEPTH_STALL);
316 OUT_BATCH(0); /* address */
317 OUT_BATCH(0); /* write data */
321 OUT_BATCH(_3DSTATE_PIPE_CONTROL);
322 OUT_BATCH(PIPE_CONTROL_DEPTH_CACHE_FLUSH);
323 OUT_BATCH(0); /* address */
324 OUT_BATCH(0); /* write data */
328 OUT_BATCH(_3DSTATE_PIPE_CONTROL);
329 OUT_BATCH(PIPE_CONTROL_DEPTH_STALL);
330 OUT_BATCH(0); /* address */
331 OUT_BATCH(0); /* write data */
336 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
337 * implementing two workarounds on gen6. From section 1.4.7.1
338 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
340 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
341 * produced by non-pipelined state commands), software needs to first
342 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
345 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
346 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
348 * And the workaround for these two requires this workaround first:
350 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
351 * BEFORE the pipe-control with a post-sync op and no write-cache
354 * And this last workaround is tricky because of the requirements on
355 * that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
358 * "1 of the following must also be set:
359 * - Render Target Cache Flush Enable ([12] of DW1)
360 * - Depth Cache Flush Enable ([0] of DW1)
361 * - Stall at Pixel Scoreboard ([1] of DW1)
362 * - Depth Stall ([13] of DW1)
363 * - Post-Sync Operation ([13] of DW1)
364 * - Notify Enable ([8] of DW1)"
366 * The cache flushes require the workaround flush that triggered this
367 * one, so we can't use it. Depth stall would trigger the same.
368 * Post-sync nonzero is what triggered this second workaround, so we
369 * can't use that one either. Notify enable is IRQs, which aren't
370 * really our business. That leaves only stall at scoreboard.
373 intel_emit_post_sync_nonzero_flush(struct intel_context *intel)
375 if (!intel->batch.need_workaround_flush)
379 OUT_BATCH(_3DSTATE_PIPE_CONTROL);
380 OUT_BATCH(PIPE_CONTROL_CS_STALL |
381 PIPE_CONTROL_STALL_AT_SCOREBOARD);
382 OUT_BATCH(0); /* address */
383 OUT_BATCH(0); /* write data */
387 OUT_BATCH(_3DSTATE_PIPE_CONTROL);
388 OUT_BATCH(PIPE_CONTROL_WRITE_IMMEDIATE);
389 OUT_RELOC(intel->batch.workaround_bo,
390 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, 0);
391 OUT_BATCH(0); /* write data */
394 intel->batch.need_workaround_flush = false;
397 /* Emit a pipelined flush to either flush render and texture cache for
398 * reading from a FBO-drawn texture, or flush so that frontbuffer
399 * render appears on the screen in DRI1.
401 * This is also used for the always_flush_cache driconf debug option.
404 intel_batchbuffer_emit_mi_flush(struct intel_context *intel)
406 if (intel->gen >= 6) {
407 if (intel->batch.is_blit) {
409 OUT_BATCH(MI_FLUSH_DW);
415 if (intel->gen == 6) {
416 /* Hardware workaround: SNB B-Spec says:
418 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache
419 * Flush Enable =1, a PIPE_CONTROL with any non-zero
420 * post-sync-op is required.
422 intel_emit_post_sync_nonzero_flush(intel);
426 OUT_BATCH(_3DSTATE_PIPE_CONTROL);
427 OUT_BATCH(PIPE_CONTROL_INSTRUCTION_FLUSH |
428 PIPE_CONTROL_WRITE_FLUSH |
429 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
430 PIPE_CONTROL_TC_FLUSH |
431 PIPE_CONTROL_NO_WRITE);
432 OUT_BATCH(0); /* write address */
433 OUT_BATCH(0); /* write data */
436 } else if (intel->gen >= 4) {
438 OUT_BATCH(_3DSTATE_PIPE_CONTROL |
439 PIPE_CONTROL_WRITE_FLUSH |
440 PIPE_CONTROL_NO_WRITE);
441 OUT_BATCH(0); /* write address */
442 OUT_BATCH(0); /* write data */
443 OUT_BATCH(0); /* write data */