Guard against the most recent write operation instead. Accumulate read
bindings to a) not issue unnecessary barriers when a resource has
multiple read-only usages, and b) synchronize with all previous read
bindings when writing.
This is motivated by an issue where a program using a single buffer to
store both indices and vertex attributes causes superfluous barriers on
each draw call.
Signed-off-by: Jan Sikorski <jsikorski(a)codeweavers.com>
---
dlls/wined3d/buffer.c | 25 ++++++++++++++++++++-----
dlls/wined3d/texture.c | 27 +++++++++++++++++++++------
dlls/wined3d/wined3d_private.h | 3 +++
3 files changed, 44 insertions(+), 11 deletions(-)
diff --git a/dlls/wined3d/buffer.c b/dlls/wined3d/buffer.c
index 4420885b15f..da61f36af45 100644
--- a/dlls/wined3d/buffer.c
+++ b/dlls/wined3d/buffer.c
@@ -1582,22 +1582,38 @@ HRESULT wined3d_buffer_vk_init(struct wined3d_buffer_vk *buffer_vk, struct wined
void wined3d_buffer_vk_barrier(struct wined3d_buffer_vk *buffer_vk,
struct wined3d_context_vk *context_vk, uint32_t bind_mask)
{
+ uint32_t src_bind_mask = 0;
+
TRACE("buffer_vk %p, context_vk %p, bind_mask %s.\n",
buffer_vk, context_vk, wined3d_debug_bind_flags(bind_mask));
- if (buffer_vk->bind_mask && buffer_vk->bind_mask != bind_mask)
+ if (bind_mask & ~WINED3D_READ_ONLY_BIND_MASK)
+ {
+ src_bind_mask = buffer_vk->bind_mask & WINED3D_READ_ONLY_BIND_MASK;
+ if (!src_bind_mask)
+ src_bind_mask = buffer_vk->bind_mask;
+
+ buffer_vk->bind_mask = bind_mask;
+ }
+ else if ((buffer_vk->bind_mask & bind_mask) != bind_mask)
+ {
+ src_bind_mask = buffer_vk->bind_mask & ~WINED3D_READ_ONLY_BIND_MASK;
+ buffer_vk->bind_mask |= bind_mask;
+ }
+
+ if (src_bind_mask)
{
const struct wined3d_vk_info *vk_info = context_vk->vk_info;
VkBufferMemoryBarrier vk_barrier;
TRACE(" %s -> %s.\n",
- wined3d_debug_bind_flags(buffer_vk->bind_mask), wined3d_debug_bind_flags(bind_mask));
+ wined3d_debug_bind_flags(src_bind_mask), wined3d_debug_bind_flags(bind_mask));
wined3d_context_vk_end_current_render_pass(context_vk);
vk_barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
vk_barrier.pNext = NULL;
- vk_barrier.srcAccessMask = vk_access_mask_from_bind_flags(buffer_vk->bind_mask);
+ vk_barrier.srcAccessMask = vk_access_mask_from_bind_flags(src_bind_mask);
vk_barrier.dstAccessMask = vk_access_mask_from_bind_flags(bind_mask);
vk_barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
vk_barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
@@ -1605,11 +1621,10 @@ void wined3d_buffer_vk_barrier(struct wined3d_buffer_vk *buffer_vk,
vk_barrier.offset = buffer_vk->bo.buffer_offset;
vk_barrier.size = buffer_vk->b.resource.size;
VK_CALL(vkCmdPipelineBarrier(wined3d_context_vk_get_command_buffer(context_vk),
- vk_pipeline_stage_mask_from_bind_flags(buffer_vk->bind_mask),
+ vk_pipeline_stage_mask_from_bind_flags(src_bind_mask),
vk_pipeline_stage_mask_from_bind_flags(bind_mask),
0, 0, NULL, 1, &vk_barrier, 0, NULL));
}
- buffer_vk->bind_mask = bind_mask;
}
HRESULT CDECL wined3d_buffer_create(struct wined3d_device *device, const struct wined3d_buffer_desc *desc,
diff --git a/dlls/wined3d/texture.c b/dlls/wined3d/texture.c
index c1247fbc56b..f5f7fab49a6 100644
--- a/dlls/wined3d/texture.c
+++ b/dlls/wined3d/texture.c
@@ -5244,15 +5244,31 @@ HRESULT wined3d_texture_vk_init(struct wined3d_texture_vk *texture_vk, struct wi
void wined3d_texture_vk_barrier(struct wined3d_texture_vk *texture_vk,
struct wined3d_context_vk *context_vk, uint32_t bind_mask)
{
- VkImageSubresourceRange vk_range;
+ uint32_t src_bind_mask = 0;
TRACE("texture_vk %p, context_vk %p, bind_mask %s.\n",
texture_vk, context_vk, wined3d_debug_bind_flags(bind_mask));
- if (texture_vk->bind_mask && texture_vk->bind_mask != bind_mask)
+ if (bind_mask & ~WINED3D_READ_ONLY_BIND_MASK)
{
+ src_bind_mask = texture_vk->bind_mask & WINED3D_READ_ONLY_BIND_MASK;
+ if (!src_bind_mask)
+ src_bind_mask = texture_vk->bind_mask;
+
+ texture_vk->bind_mask = bind_mask;
+ }
+ else if ((texture_vk->bind_mask & bind_mask) != bind_mask)
+ {
+ src_bind_mask = texture_vk->bind_mask & ~WINED3D_READ_ONLY_BIND_MASK;
+ texture_vk->bind_mask |= bind_mask;
+ }
+
+ if (src_bind_mask)
+ {
+ VkImageSubresourceRange vk_range;
+
TRACE(" %s -> %s.\n",
- wined3d_debug_bind_flags(texture_vk->bind_mask), wined3d_debug_bind_flags(bind_mask));
+ wined3d_debug_bind_flags(src_bind_mask), wined3d_debug_bind_flags(bind_mask));
vk_range.aspectMask = vk_aspect_mask_from_format(texture_vk->t.resource.format);
vk_range.baseMipLevel = 0;
@@ -5261,12 +5277,11 @@ void wined3d_texture_vk_barrier(struct wined3d_texture_vk *texture_vk,
vk_range.layerCount = VK_REMAINING_ARRAY_LAYERS;
wined3d_context_vk_image_barrier(context_vk, wined3d_context_vk_get_command_buffer(context_vk),
- vk_pipeline_stage_mask_from_bind_flags(texture_vk->bind_mask),
+ vk_pipeline_stage_mask_from_bind_flags(src_bind_mask),
vk_pipeline_stage_mask_from_bind_flags(bind_mask),
- vk_access_mask_from_bind_flags(texture_vk->bind_mask), vk_access_mask_from_bind_flags(bind_mask),
+ vk_access_mask_from_bind_flags(src_bind_mask), vk_access_mask_from_bind_flags(bind_mask),
texture_vk->layout, texture_vk->layout, texture_vk->image.vk_image, &vk_range);
}
- texture_vk->bind_mask = bind_mask;
}
static void ffp_blitter_destroy(struct wined3d_blitter *blitter, struct wined3d_context *context)
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index 86eae149306..6e9f3884724 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -310,6 +310,9 @@ struct min_lookup
extern const struct min_lookup minMipLookup[WINED3D_TEXF_LINEAR + 1] DECLSPEC_HIDDEN;
extern const GLenum magLookup[WINED3D_TEXF_LINEAR + 1] DECLSPEC_HIDDEN;
+static const uint32_t WINED3D_READ_ONLY_BIND_MASK = WINED3D_BIND_VERTEX_BUFFER | WINED3D_BIND_INDEX_BUFFER
+ | WINED3D_BIND_CONSTANT_BUFFER | WINED3D_BIND_SHADER_RESOURCE | WINED3D_BIND_INDIRECT_BUFFER;
+
GLenum wined3d_gl_compare_func(enum wined3d_cmp_func f) DECLSPEC_HIDDEN;
VkAccessFlags vk_access_mask_from_bind_flags(uint32_t bind_flags) DECLSPEC_HIDDEN;
VkCompareOp vk_compare_op_from_wined3d(enum wined3d_cmp_func op) DECLSPEC_HIDDEN;
--
2.30.2