* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
*
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
- * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
- * IN THE SOFTWARE.
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+/**
+ * @file iris_pipe_control.c
+ *
+ * PIPE_CONTROL is the main flushing and synchronization primitive on Intel
+ * GPUs. It can invalidate caches, stall until rendering reaches various
+ * stages of completion, write to memory, and other things. In a way, it's
+ * a swiss army knife command - it has all kinds of capabilities, but some
+ * significant limitations as well.
+ *
+ * Unfortunately, it's notoriously complicated and difficult to use. Many
+ * sub-commands can't be used together. Some are meant to be used at the
+ * top of the pipeline (invalidating caches before drawing), while some are
+ * meant to be used at the end (stalling or flushing after drawing).
+ *
+ * Also, there's a list of restrictions a mile long, which vary by generation.
+ * Do this before doing that, or suffer the consequences (usually a GPU hang).
+ *
+ * This file contains helpers for emitting them safely. You can simply call
+ * iris_emit_pipe_control_flush() with the desired operations (as logical
+ * PIPE_CONTROL_* bits), and it will take care of splitting it into multiple
+ * PIPE_CONTROL commands as necessary. The per-generation workarounds are
+ * applied in iris_emit_raw_pipe_control() in iris_state.c.
*/
#include "iris_context.h"
+#include "util/hash_table.h"
+#include "util/set.h"
/**
* Emit a PIPE_CONTROL with various flushing flags.
* given generation.
*/
void
-iris_emit_pipe_control_flush(struct iris_context *ice,
- struct iris_batch *batch,
- uint32_t flags)
+iris_emit_pipe_control_flush(struct iris_batch *batch, uint32_t flags)
{
if ((flags & PIPE_CONTROL_CACHE_FLUSH_BITS) &&
(flags & PIPE_CONTROL_CACHE_INVALIDATE_BITS)) {
* with any write cache flush, so this shouldn't be a concern. In order
* to ensure a full stall, we do an end-of-pipe sync.
*/
- iris_emit_end_of_pipe_sync(ice, batch,
- flags & PIPE_CONTROL_CACHE_FLUSH_BITS);
+ iris_emit_end_of_pipe_sync(batch, flags & PIPE_CONTROL_CACHE_FLUSH_BITS);
flags &= ~(PIPE_CONTROL_CACHE_FLUSH_BITS | PIPE_CONTROL_CS_STALL);
}
- ice->state.emit_raw_pipe_control(batch, flags, NULL, 0, 0);
+ batch->vtbl->emit_raw_pipe_control(batch, flags, NULL, 0, 0);
}
/**
* - PIPE_CONTROL_WRITE_DEPTH_COUNT
*/
void
-iris_emit_pipe_control_write(struct iris_context *ice,
- struct iris_batch *batch, uint32_t flags,
+iris_emit_pipe_control_write(struct iris_batch *batch, uint32_t flags,
struct iris_bo *bo, uint32_t offset,
uint64_t imm)
{
- ice->state.emit_raw_pipe_control(batch, flags, bo, offset, imm);
+ batch->vtbl->emit_raw_pipe_control(batch, flags, bo, offset, imm);
}
/*
* Data" in the PIPE_CONTROL command.
*/
void
-iris_emit_end_of_pipe_sync(struct iris_context *ice,
- struct iris_batch *batch,
- uint32_t flags)
+iris_emit_end_of_pipe_sync(struct iris_batch *batch, uint32_t flags)
{
/* From Sandybridge PRM, volume 2, "1.7.3.1 Writing a Value to Memory":
*
* Data, Required Write Cache Flush bits set)
* - Workload-2 (Can use the data produce or output by Workload-1)
*/
- iris_emit_pipe_control_write(ice, batch, flags | PIPE_CONTROL_CS_STALL |
+ iris_emit_pipe_control_write(batch, flags | PIPE_CONTROL_CS_STALL |
PIPE_CONTROL_WRITE_IMMEDIATE,
batch->screen->workaround_bo, 0, 0);
}
+
+static void
+iris_texture_barrier(struct pipe_context *ctx, unsigned flags)
+{
+ struct iris_context *ice = (void *) ctx;
+ struct iris_batch *render_batch = &ice->batches[IRIS_BATCH_RENDER];
+ struct iris_batch *compute_batch = &ice->batches[IRIS_BATCH_COMPUTE];
+
+ if (render_batch->contains_draw ||
+ render_batch->cache.render->entries ||
+ render_batch->cache.depth->entries) {
+ iris_emit_pipe_control_flush(render_batch,
+ PIPE_CONTROL_DEPTH_CACHE_FLUSH |
+ PIPE_CONTROL_RENDER_TARGET_FLUSH |
+ PIPE_CONTROL_CS_STALL);
+ iris_emit_pipe_control_flush(render_batch,
+ PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
+ }
+
+ if (compute_batch->contains_draw) {
+ iris_emit_pipe_control_flush(compute_batch,
+ PIPE_CONTROL_CS_STALL);
+ iris_emit_pipe_control_flush(compute_batch,
+ PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
+ }
+}
+
+static void
+iris_memory_barrier(struct pipe_context *ctx, unsigned flags)
+{
+ struct iris_context *ice = (void *) ctx;
+ unsigned bits = PIPE_CONTROL_DATA_CACHE_FLUSH | PIPE_CONTROL_CS_STALL;
+
+ if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
+ PIPE_BARRIER_INDEX_BUFFER |
+ PIPE_BARRIER_INDIRECT_BUFFER)) {
+ bits |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
+ }
+
+ if (flags & PIPE_BARRIER_CONSTANT_BUFFER) {
+ bits |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
+ PIPE_CONTROL_CONST_CACHE_INVALIDATE;
+ }
+
+ if (flags & (PIPE_BARRIER_TEXTURE | PIPE_BARRIER_FRAMEBUFFER)) {
+ bits |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
+ PIPE_CONTROL_RENDER_TARGET_FLUSH;
+ }
+
+ for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
+ if (ice->batches[i].contains_draw ||
+ ice->batches[i].cache.render->entries)
+ iris_emit_pipe_control_flush(&ice->batches[i], bits);
+ }
+}
+
+void
+iris_init_flush_functions(struct pipe_context *ctx)
+{
+ ctx->memory_barrier = iris_memory_barrier;
+ ctx->texture_barrier = iris_texture_barrier;
+}