This is an automated email from the git hooks/post-receive script.
Git pushed a commit to branch master
in repository ffmpeg.
The following commit(s) were added to refs/heads/master by this push:
new 7d3ace8cb2 vulkan_prores: pick the bitstream prefetch depth per device
7d3ace8cb2 is described below
commit 7d3ace8cb21dfbb7a2b1fdb7246bef73ec629a5d
Author: Lynne <[email protected]>
AuthorDate: Sun Sep 20 17:30:24 2026 +0900
Commit: Lynne <[email protected]>
CommitDate: Sun Sep 20 17:51:47 2026 +0900
vulkan_prores: pick the bitstream prefetch depth per device
Same situation as ProRes RAW: with the packet host-mapped, the VLD is
bound by fetch latency over the bus.
Make GET_BITS_SMEM a specialization constant: 32 lines per invocation
on discrete GPUs reading host-mapped packets, 4 otherwise. The 8x8
workgroup makes that 32 KB of shared memory; 64 does not fit, and 40
or 48 measured slower. An integrated GPU loses a third of its speed at
32, so it keeps 4. The depth is clamped to maxComputeSharedMemorySize
for 32 KB devices such as MoltenVK.
5.7K ProRes HQ on an RX 6900 XT: 85 -> 120 fps, bit-identical output.
---
libavcodec/vulkan/prores_vld.comp.glsl | 4 +++-
libavcodec/vulkan_prores.c | 13 ++++++++++++-
2 files changed, 15 insertions(+), 2 deletions(-)
diff --git a/libavcodec/vulkan/prores_vld.comp.glsl
b/libavcodec/vulkan/prores_vld.comp.glsl
index b0b17264a4..cdf1fe9ef9 100644
--- a/libavcodec/vulkan/prores_vld.comp.glsl
+++ b/libavcodec/vulkan/prores_vld.comp.glsl
@@ -19,7 +19,9 @@
#pragma shader_stage(compute)
#extension GL_GOOGLE_include_directive : require
-#define GET_BITS_SMEM 4
+/* Bitstream prefetch depth in 16-byte lines, set by the host */
+layout (constant_id = 1) const uint smem_lines = 4;
+#define GET_BITS_SMEM smem_lines
#include "common.glsl"
layout (constant_id = 0) const bool interlaced = false;
diff --git a/libavcodec/vulkan_prores.c b/libavcodec/vulkan_prores.c
index d8d7808ef8..1ccbb5288b 100644
--- a/libavcodec/vulkan_prores.c
+++ b/libavcodec/vulkan_prores.c
@@ -374,8 +374,19 @@ static int init_decode_shader(AVCodecContext *avctx,
FFVulkanContext *s,
AVHWFramesContext *dec_frames_ctx;
dec_frames_ctx = (AVHWFramesContext *)avctx->hw_frames_ctx->data;
- SPEC_LIST_CREATE(sl, 1, 1*sizeof(uint32_t))
+ /* Discrete GPUs read host-mapped packets over the bus; prefetch more
+ * to hide the latency. 32 lines (32 KB for the 8x8 workgroup) measured
+ * best; integrated GPUs get slower with every extra line. Clamp to the
+ * shared memory limit, leaving 1 KB for the tables. */
+ int smem_lines = 4;
+ if (s->props.properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
&&
+ (s->extensions & FF_VK_EXT_EXTERNAL_HOST_MEMORY)) {
+ uint32_t max_smem =
s->props.properties.limits.maxComputeSharedMemorySize;
+ smem_lines = FFMIN(32, (max_smem - 1024) / (8*8*16));
+ }
+ SPEC_LIST_CREATE(sl, 2, 2*sizeof(uint32_t))
SPEC_LIST_ADD(sl, 0, 32, interlaced);
+ SPEC_LIST_ADD(sl, 1, 32, smem_lines);
ff_vk_shader_load(shd,
VK_SHADER_STAGE_COMPUTE_BIT, sl,
--
To stop receiving notification emails like this one, please contact
[email protected].
_______________________________________________
ffmpeg-cvslog mailing list -- [email protected]
To unsubscribe send an email to [email protected]
