[Mesa-dev] [PATCH v3 24/25] panfrost: Support batch pipelining
Boris Brezillon
boris.brezillon at collabora.com
Thu Sep 5 19:41:49 UTC 2019
We adjust the code to explicitly request flush of batches accessing
BOs they care about. Thanks to that, we can get rid of the implicit
serialization done in panfrost_batch_submit() and
panfrost_set_framebuffer_state(). Finally, panfrost_flush() is
changed to to flush all pending batches.
Signed-off-by: Boris Brezillon <boris.brezillon at collabora.com>
---
src/gallium/drivers/panfrost/pan_compute.c | 2 +-
src/gallium/drivers/panfrost/pan_context.c | 145 +++++++++++++-------
src/gallium/drivers/panfrost/pan_job.c | 15 +-
src/gallium/drivers/panfrost/pan_resource.c | 26 ++--
4 files changed, 115 insertions(+), 73 deletions(-)
diff --git a/src/gallium/drivers/panfrost/pan_compute.c b/src/gallium/drivers/panfrost/pan_compute.c
index 4639c1b03c38..036dffbb17be 100644
--- a/src/gallium/drivers/panfrost/pan_compute.c
+++ b/src/gallium/drivers/panfrost/pan_compute.c
@@ -133,7 +133,7 @@ panfrost_launch_grid(struct pipe_context *pipe,
/* Queue the job */
panfrost_scoreboard_queue_compute_job(batch, transfer);
- panfrost_flush(pipe, NULL, PIPE_FLUSH_END_OF_FRAME);
+ panfrost_flush_all_batches(ctx, true);
}
void
diff --git a/src/gallium/drivers/panfrost/pan_context.c b/src/gallium/drivers/panfrost/pan_context.c
index 02726e7cd349..993744a1ffd0 100644
--- a/src/gallium/drivers/panfrost/pan_context.c
+++ b/src/gallium/drivers/panfrost/pan_context.c
@@ -150,6 +150,28 @@ panfrost_emit_mfbd(struct panfrost_batch *batch, unsigned vertex_count)
return framebuffer;
}
+static void
+panfrost_flush_fbo_deps(struct panfrost_context *ctx)
+{
+ struct pipe_framebuffer_state *fb = &ctx->pipe_framebuffer;
+ for (unsigned i = 0; i < fb->nr_cbufs; i++) {
+ if (!fb->cbufs[i])
+ continue;
+
+ struct panfrost_resource *rsrc = pan_resource(fb->cbufs[i]->texture);
+
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, true);
+ panfrost_flush_batches_reading_bo(ctx, rsrc->bo, true);
+ }
+
+ if (fb->zsbuf) {
+ struct panfrost_resource *rsrc = pan_resource(fb->zsbuf->texture);
+
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, true);
+ panfrost_flush_batches_reading_bo(ctx, rsrc->bo, true);
+ }
+}
+
static void
panfrost_clear(
struct pipe_context *pipe,
@@ -160,6 +182,7 @@ panfrost_clear(
struct panfrost_context *ctx = pan_context(pipe);
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
+ panfrost_flush_fbo_deps(ctx);
panfrost_batch_add_fbo_bos(batch);
panfrost_batch_clear(batch, buffers, color, depth, stencil);
}
@@ -1324,10 +1347,9 @@ panfrost_flush(
unsigned flags)
{
struct panfrost_context *ctx = pan_context(pipe);
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- /* Submit the frame itself */
- panfrost_batch_submit(batch);
+ /* Submit all pending jobs */
+ panfrost_flush_all_batches(ctx, false);
if (fence) {
struct panfrost_fence *f = panfrost_fence_create(ctx);
@@ -1433,6 +1455,71 @@ panfrost_statistics_record(
ctx->tf_prims_generated += prims;
}
+static void
+panfrost_flush_draw_deps(struct panfrost_context *ctx, const struct pipe_draw_info *info)
+{
+ struct panfrost_resource *rsrc;
+
+ if (ctx->wallpaper_batch)
+ return;
+
+ panfrost_flush_fbo_deps(ctx);
+
+ for (unsigned stage = 0; stage < PIPE_SHADER_TYPES; stage++) {
+ for (unsigned i = 0; i < ctx->sampler_view_count[stage]; i++) {
+ struct panfrost_sampler_view *view = ctx->sampler_views[stage][i];
+
+ if (!view)
+ continue;
+
+ rsrc = pan_resource(view->base.texture);
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, true);
+ }
+
+ for (unsigned i = 0; i < 32; i++) {
+ if (!(ctx->ssbo_mask[stage] & (1 << i)))
+ continue;
+
+ rsrc = pan_resource(ctx->ssbo[stage][i].buffer);
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, true);
+ panfrost_flush_batches_reading_bo(ctx, rsrc->bo, true);
+ }
+ }
+
+ if (info->index_size && !info->has_user_indices) {
+ struct panfrost_resource *rsrc = pan_resource(info->index.resource);
+
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, true);
+ }
+
+ for (unsigned i = 0; ctx->vertex && i < ctx->vertex->num_elements; i++) {
+ struct pipe_vertex_element *velem = &ctx->vertex->pipe[i];
+ unsigned vbi = velem->vertex_buffer_index;
+
+ if (!(ctx->vb_mask & (1 << vbi)))
+ continue;
+
+ struct pipe_vertex_buffer *buf = &ctx->vertex_buffers[vbi];
+
+ if (!buf->buffer.resource)
+ continue;
+
+ rsrc = pan_resource(buf->buffer.resource);
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, true);
+ }
+
+ for (unsigned i = 0; i < ctx->streamout.num_targets; i++) {
+ struct pipe_stream_output_target *target = ctx->streamout.targets[i];
+
+ if (!target)
+ continue;
+
+ rsrc = pan_resource(target->buffer);
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, true);
+ panfrost_flush_batches_reading_bo(ctx, rsrc->bo, true);
+ }
+}
+
static void
panfrost_draw_vbo(
struct pipe_context *pipe,
@@ -1477,6 +1564,7 @@ panfrost_draw_vbo(
}
}
+ panfrost_flush_draw_deps(ctx, info);
ctx->payloads[PIPE_SHADER_VERTEX].offset_start = info->start;
ctx->payloads[PIPE_SHADER_FRAGMENT].offset_start = info->start;
@@ -2252,50 +2340,10 @@ panfrost_set_framebuffer_state(struct pipe_context *pctx,
{
struct panfrost_context *ctx = pan_context(pctx);
- /* Flush when switching framebuffers, but not if the framebuffer
- * state is being restored by u_blitter
- */
-
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- bool is_scanout = panfrost_batch_is_scanout(batch);
- bool has_draws = batch->last_job.gpu;
-
- /* Bail out early when the current and new states are the same. */
- if (util_framebuffer_state_equal(&ctx->pipe_framebuffer, fb))
- return;
-
- /* The wallpaper logic sets a new FB state before doing the blit and
- * restore the old one when it's done. Those FB states are reported to
- * be different because the surface they are pointing to are different,
- * but those surfaces actually point to the same cbufs/zbufs. In that
- * case we definitely don't want new FB descs to be emitted/attached
- * since the job is expected to be flushed just after the blit is done,
- * so let's just copy the new state and return here.
- */
- if (ctx->wallpaper_batch) {
- util_copy_framebuffer_state(&ctx->pipe_framebuffer, fb);
- return;
- }
-
- if (!is_scanout || has_draws)
- panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME);
- else
- assert(!ctx->payloads[PIPE_SHADER_VERTEX].postfix.framebuffer &&
- !ctx->payloads[PIPE_SHADER_FRAGMENT].postfix.framebuffer);
-
- /* Invalidate the FBO job cache since we've just been assigned a new
- * FB state.
- */
- ctx->batch = NULL;
-
+ panfrost_hint_afbc(pan_screen(pctx->screen), fb);
util_copy_framebuffer_state(&ctx->pipe_framebuffer, fb);
-
- /* Given that we're rendering, we'd love to have compression */
- struct panfrost_screen *screen = pan_screen(ctx->base.screen);
-
- panfrost_hint_afbc(screen, &ctx->pipe_framebuffer);
- for (unsigned i = 0; i < PIPE_SHADER_TYPES; ++i)
- ctx->payloads[i].postfix.framebuffer = 0;
+ ctx->batch = NULL;
+ panfrost_invalidate_frame(ctx);
}
static void *
@@ -2513,6 +2561,7 @@ panfrost_get_query_result(struct pipe_context *pipe,
bool wait,
union pipe_query_result *vresult)
{
+ struct panfrost_context *ctx = pan_context(pipe);
struct panfrost_query *query = (struct panfrost_query *) q;
@@ -2521,7 +2570,7 @@ panfrost_get_query_result(struct pipe_context *pipe,
case PIPE_QUERY_OCCLUSION_PREDICATE:
case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
/* Flush first */
- panfrost_flush(pipe, NULL, PIPE_FLUSH_END_OF_FRAME);
+ panfrost_flush_all_batches(ctx, true);
/* Read back the query results */
unsigned *result = (unsigned *) query->transfer.cpu;
@@ -2537,7 +2586,7 @@ panfrost_get_query_result(struct pipe_context *pipe,
case PIPE_QUERY_PRIMITIVES_GENERATED:
case PIPE_QUERY_PRIMITIVES_EMITTED:
- panfrost_flush(pipe, NULL, PIPE_FLUSH_END_OF_FRAME);
+ panfrost_flush_all_batches(ctx, true);
vresult->u64 = query->end - query->start;
break;
diff --git a/src/gallium/drivers/panfrost/pan_job.c b/src/gallium/drivers/panfrost/pan_job.c
index 8eda110542c3..c40f3f12ec13 100644
--- a/src/gallium/drivers/panfrost/pan_job.c
+++ b/src/gallium/drivers/panfrost/pan_job.c
@@ -384,21 +384,14 @@ panfrost_batch_submit(struct panfrost_batch *batch)
fprintf(stderr, "panfrost_batch_submit failed: %d\n", ret);
out:
- if (ctx->batch == batch)
- panfrost_invalidate_frame(ctx);
-
/* The job has been submitted, let's invalidate the current FBO job
* cache.
*/
- assert(!ctx->batch || batch == ctx->batch);
- ctx->batch = NULL;
+ if (ctx->batch == batch) {
+ panfrost_invalidate_frame(ctx);
+ ctx->batch = NULL;
+ }
- /* We always stall the pipeline for correct results since pipelined
- * rendering is quite broken right now (to be fixed by the panfrost_job
- * refactor, just take the perf hit for correctness)
- */
- drmSyncobjWait(pan_screen(ctx->base.screen)->fd, &ctx->out_sync, 1,
- INT64_MAX, 0, NULL);
panfrost_free_batch(batch);
}
diff --git a/src/gallium/drivers/panfrost/pan_resource.c b/src/gallium/drivers/panfrost/pan_resource.c
index a5869768f9de..25b72478dccd 100644
--- a/src/gallium/drivers/panfrost/pan_resource.c
+++ b/src/gallium/drivers/panfrost/pan_resource.c
@@ -578,10 +578,8 @@ panfrost_transfer_map(struct pipe_context *pctx,
is_bound |= fb->cbufs[c]->texture == resource;
}
- if (is_bound && (usage & PIPE_TRANSFER_READ)) {
- assert(level == 0);
- panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME);
- }
+ if (is_bound && (usage & PIPE_TRANSFER_READ))
+ assert(level == 0);
/* TODO: Respect usage flags */
@@ -594,9 +592,10 @@ panfrost_transfer_map(struct pipe_context *pctx,
/* No flush for writes to uninitialized */
} else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
if (usage & PIPE_TRANSFER_WRITE) {
- /* STUB: flush reading */
- //printf("debug: missed reading flush %d\n", resource->target);
+ panfrost_flush_batch_writing_bo(ctx, bo, false);
+ panfrost_flush_batches_reading_bo(ctx, bo, false);
} else if (usage & PIPE_TRANSFER_READ) {
+ panfrost_flush_batch_writing_bo(ctx, bo, false);
/* STUB: flush writing */
//printf("debug: missed writing flush %d (%d-%d)\n", resource->target, box->x, box->x + box->width);
} else {
@@ -604,6 +603,8 @@ panfrost_transfer_map(struct pipe_context *pctx,
}
}
+ panfrost_bo_wait(bo, INT64_MAX);
+
if (rsrc->layout != PAN_LINEAR) {
/* Non-linear resources need to be indirectly mapped */
@@ -748,11 +749,8 @@ panfrost_generate_mipmap(
* reorder-type optimizations in place. But for now prioritize
* correctness. */
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- bool has_draws = batch->last_job.gpu;
-
- if (has_draws)
- panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME);
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, false);
+ panfrost_bo_wait(rsrc->bo, INT64_MAX);
/* We've flushed the original buffer if needed, now trigger a blit */
@@ -765,8 +763,10 @@ panfrost_generate_mipmap(
/* If the blit was successful, flush once more. If it wasn't, well, let
* the state tracker deal with it. */
- if (blit_res)
- panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME);
+ if (blit_res) {
+ panfrost_flush_batch_writing_bo(ctx, rsrc->bo, false);
+ panfrost_bo_wait(rsrc->bo, INT64_MAX);
+ }
return blit_res;
}
--
2.21.0
More information about the mesa-dev
mailing list