[spirv] Move external arrays into seperate buffers #4121
Conversation
|
This pull request is being automatically deployed with Vercel (learn more). 🔍 Inspect: https://vercel.com/taichi-api-docs-preview/taichi/EcS3q6NV4qW9rAWMVzQSoBb8WKeL [Deployment for a075b18 canceled] |
|
/format |
|
✔️ Deploy Preview for docsite-preview ready! 🔨 Explore the source changes: a075b18 🔍 Inspect the deploy log: https://app.netlify.com/sites/docsite-preview/deploys/61f28860cd02c400080717de 😎 Browse the preview: https://deploy-preview-4121--docsite-preview.netlify.app |
bobcao3
changed the title
|
/rerun |
| @@ -5,6 +5,7 @@ | |||
| #include <functional> | |||
| #include <optional> | |||
| #include <atomic> | |||
| #include <stack> | |||
There was a problem hiding this comment.
It's used in this class
| @@ -427,9 +441,24 @@ void VkRuntime::launch_kernel(KernelHandle handle, RuntimeContext *host_ctx) { | |||
| &ti_kernel->ti_kernel_attribs().ctx_attribs, host_ctx, device_, | |||
| host_result_buffer_, ctx_buffer.get(), ctx_buffer_host.get()); | |||
|
|
|||
| std::unordered_map<int, DeviceAllocation> ext_arrays; | |||
There was a problem hiding this comment.
FYI in opengl we don't redo device allocation if the size doesn't change. maybe it helps with perf a little bit? https://github.com/taichi-dev/taichi/blob/master/taichi/backends/opengl/opengl_api.cpp#L400 (can be a followup ofc)
There was a problem hiding this comment.
Let's leave it for later, the balance between memory usage and allocation performance is hard
| void *host_ptr = host_ctx_->get_arg<void *>(i); | ||
| std::memcpy(host_ptr, device_ptr, ret.stride); | ||
| // void *host_ptr = host_ctx_->get_arg<void *>(i); | ||
| // std::memcpy(host_ptr, device_ptr, ret.stride); |
There was a problem hiding this comment.
We could remove useless comments in the codes
| Context, | ||
| ListGen, | ||
| }; | ||
| enum class BufferType { Root, GlobalTmps, Context, ListGen, ExtArr }; |
There was a problem hiding this comment.
OOC, do we really support ListGen buffer in SPIR-V codegen..?
There was a problem hiding this comment.
We have very low level all dense->bitmasked support
|
/format |
|
/rerun |
Related issue = #3642
First step towards ndarray on SPIRV backends.
Also fixed a few small things to speed up compilation a small bit.
cc: @xwang186