CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

metal : replace loop of dispatch_async with dispatch_apply (#4934)

Browse Source 
* Replace loop of dispatch_async with dispatch_apply

* Update ggml-metal.m

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Alex Azarov
2024-01-16 14:41:27 +01:00
committed by GitHub
parent 7c8d3abd1a
commit 3a48d558a6
1 changed files with 1439 additions and 1443 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml-metal.m
View File

@@ -737,10 +737,10 @@ static bool ggml_metal_graph_compute(
ctx->command_encoders[i] = [ctx->command_buffers[i] computeCommandEncoderWithDescriptor: edesc];
}
for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
dispatch_apply(n_cb, ctx->d_queue, ^(size_t iter) {
const int cb_idx = iter;
dispatch_async(ctx->d_queue, ^{
size_t offs_src0 = 0;
size_t offs_src1 = 0;
size_t offs_dst = 0;
@@ -2248,10 +2248,6 @@ static bool ggml_metal_graph_compute(
[command_buffer commit];
});
}
// wait for all threads to finish
dispatch_barrier_sync(ctx->d_queue, ^{});
// check status of command buffers
// needed to detect if the device ran out-of-memory for example (#1881)