metal: implement get_timestamp_period #2528
Conversation
|
The numbers do mostly check out, intel's period is 83 1/3, though we might want to clamp it so it doesn't go below 1, AMD's should be 1. |
wgpu-hal/src/metal/mod.rs
Outdated
| } | ||
|
|
||
| let timestamps1 = [std::ptr::null_mut::<Object>(); 2]; | ||
| let () = msg_send![device, sampleTimestamps:×tamps1[0] gpuTimestamp:×tamps1[1]]; |
There was a problem hiding this comment.
shouldn't there be some kind of work between consecutive invocations of the timestamps? or do you think the resolution here is already enough?
There was a problem hiding this comment.
Apple suggested call it at least twice, once when creating a command buffer and a second time inside the command buffer’s completion handler, but that path can't be implemented simply here, and I don't know if it makes sense to create a command buffer for that?
There was a problem hiding this comment.
Well, the whole get_timestamp_period is not in WebGPU. So we don't have to try implementing it, technically, and could come up with an entirely different API. How about no timestamp period at all? User gets some timestamps, and they can use it for comparison, but it's not really tracking absolute time.
There was a problem hiding this comment.
Looks like the timestamps API is still in discussion? gpuweb/gpuweb#2642
Description
Implemented by referring to Apple's Correlating CPU and GPU Timestamps
Metal
sampleTimestamps:gpuTimestamp:API available on iOS | iPadOS 14 and macOS 10.15+, but on old A8 chip, the timestamp value obtained by this API is 0.Testing
On all tested Apple silicon (A-series + M1), obtained cpu_timestamp equals to gpu_timestamp:
On iPad mini4(iPadOS 14.7.1),
sampleTimestamps: gpuTimestamp:alway get 0M1 Mac
get_timestamp_periodalway equal 1.0iPad Pro 2018
get_timestamp_periodalway equal 1.0iPhone 12 Pro
get_timestamp_periodalway equal 1.0Intel Mac with UHD Graphics 630
Intel Mac with AMD Radeon Pro 560X