AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

zink: improve deferred buffer barrier heuristics

Browse Source 
this is only to catch the case of a bound descriptor being written to
by some operation other than its draw/dispatch descriptor bind,
so any non-write binds are ignored

previously those non-write binds were required because of how sync
analysis could drop non-write access, so that is fixed as well

also use the vbo bind count instead of the mask because why not

also also ignore non-write GENERAL image deferred sync because that shouldn't
need anything deferred

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36846>
This commit is contained in:
Mike Blumenkrantz
2025-08-07 16:32:23 -04:00
committed by Marge Bot
parent cf5d41575b
commit 0d7e38f431
1 changed files with 16 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/gallium/drivers/zink/zink_synchronization.cpp
+16 -13
View File
@@ -124,14 +124,18 @@ static void
resource_check_defer_buffer_barrier(struct zink_context *ctx, struct zink_resource *res, VkPipelineStageFlags pipeline)
{
assert(res->obj->is_buffer);
if (res->bind_count[0] - res->so_bind_count > 0) {
if ((res->vbo_bind_mask && !(pipeline & VK_PIPELINE_STAGE_VERTEX_INPUT_BIT)) ||
(util_bitcount(res->vbo_bind_mask) != res->bind_count[0] && !is_shader_pipline_stage(pipeline)))
/* gfx rebind */
/* streamout binds have their own sync: ignore */
unsigned bind_count0 = res->bind_count[0] - res->so_bind_count;
if (bind_count0 > 0) {
/* if has vbo binds and is not a vbo barrier... */
if ((res->vbo_bind_count && !(pipeline & VK_PIPELINE_STAGE_VERTEX_INPUT_BIT)) ||
/* or if has more than just vbo binds and is not a gfx stage */
(res->vbo_bind_count != bind_count0 && !is_shader_pipline_stage(pipeline)))
/* sync on next draw */
_mesa_set_add(ctx->need_barriers[0], res);
}
if (res->bind_count[1] && !(pipeline & VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT))
/* compute rebind */
/* if writing from non-compute and has compute binds, sync on next dispatch */
_mesa_set_add(ctx->need_barriers[1], res);
}
@@ -445,9 +449,9 @@ zink_resource_image_barrier(struct zink_context *ctx, struct zink_resource *res,
if (!GENERAL && !res->obj->needs_zs_evaluate && !zink_resource_image_needs_barrier(res, new_layout, flags, pipeline) &&
(res->queue == zink_screen(ctx->base.screen)->gfx_queue || res->queue == VK_QUEUE_FAMILY_IGNORED))
return;
enum zink_resource_access rw = is_write ? ZINK_RESOURCE_ACCESS_RW : ZINK_RESOURCE_ACCESS_WRITE;
bool completed = zink_resource_usage_check_completion_fast(zink_screen(ctx->base.screen), res, rw);
bool usage_matches = !completed && zink_resource_usage_matches(res, ctx->bs);
bool has_usage = zink_resource_has_usage(res);
bool completed = !has_usage || zink_resource_usage_check_completion_fast(zink_screen(ctx->base.screen), res, ZINK_RESOURCE_ACCESS_RW);
bool usage_matches = zink_resource_usage_matches(res, ctx->bs);
VkCommandBuffer cmdbuf = GENERAL && new_layout == VK_IMAGE_LAYOUT_GENERAL ?
(UNSYNCHRONIZED ? ctx->bs->unsynchronized_cmdbuf : is_write ? zink_get_cmdbuf(ctx, NULL, res) : zink_get_cmdbuf(ctx, res, NULL)) :
update_unordered_access_and_get_cmdbuf<UNSYNCHRONIZED>::apply(ctx, res, usage_matches, is_write);
@@ -614,13 +618,12 @@ void
zink_resource_memory_barrier(struct zink_context *ctx, struct zink_resource *res, VkAccessFlags flags, VkPipelineStageFlags pipeline)
{
bool is_write = zink_resource_access_is_write(flags);
enum zink_resource_access rw = is_write ? ZINK_RESOURCE_ACCESS_RW : ZINK_RESOURCE_ACCESS_WRITE;
bool has_usage = zink_resource_has_usage(res);
bool completed = !has_usage || zink_resource_usage_check_completion_fast(zink_screen(ctx->base.screen), res, rw);
bool usage_matches = !completed && zink_resource_usage_matches(res, ctx->bs);
bool completed = !has_usage || zink_resource_usage_check_completion_fast(zink_screen(ctx->base.screen), res, ZINK_RESOURCE_ACCESS_RW);
bool usage_matches = zink_resource_usage_matches(res, ctx->bs);
if (!usage_matches) {
res->obj->unordered_write = true;
if (is_write || !has_usage || zink_resource_usage_check_completion_fast(zink_screen(ctx->base.screen), res, ZINK_RESOURCE_ACCESS_RW))
if (is_write || completed)
res->obj->unordered_read = true;
}
bool unordered_usage_matches = res->obj->unordered_access && usage_matches;
@@ -682,7 +685,7 @@ zink_resource_memory_barrier(struct zink_context *ctx, struct zink_resource *res
zink_cmd_debug_marker_end(ctx, cmdbuf, marker);
}
if (!UNSYNCHRONIZED) {
if (!UNSYNCHRONIZED && (is_write || res->write_bind_count[0] || res->write_bind_count[1])) {
if (GENERAL_IMAGE)
resource_defer_image_barrier(ctx, res, pipeline);
else