ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

metal : use shared buffers between CPU and GPU (#1696)

Browse source 
* Use MTLDevice.newBufferWithBytesNoCopy to share buffers between CPU and GPU

* Page-align buffers used by Metal

* Remove trailing whitespace

* Only import unistd.h for Metal builds

* metal : remove unnecessary copies

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
kiltyj 2023-06-05 13:24:04 -07:00 • committed by GitHub
parent efe0507632
commit 9d0693bce3
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 38 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

17
ggml-metal.m
View file

@ -195,14 +195,25 @@ bool ggml_metal_add_buffer(
}
}
size_t page_size = getpagesize();
size_t aligned_size = size;
if ((aligned_size % page_size) != 0) {
aligned_size += (page_size - (aligned_size % page_size));
}
ctx->buffers[ctx->n_buffers].name = name;
ctx->buffers[ctx->n_buffers].data = data;
ctx->buffers[ctx->n_buffers].size = size;
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytes:data length:size options:MTLResourceStorageModeShared];
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:aligned_size options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
return false;
} else {
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
}
++ctx->n_buffers;
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, size / 1024.0 / 1024.0);
}
return true;