mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-11-13 10:57:15 +00:00
metal : retain src and dst buffers during async ops (#17101)
This commit is contained in:
@@ -289,7 +289,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
|
||||
|
||||
// queue the copy operation into the queue of the Metal context
|
||||
// this will be queued at the end, after any currently ongoing GPU operations
|
||||
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
|
||||
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
|
||||
id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
|
||||
|
||||
[encoder copyFromBuffer:buf_src
|
||||
@@ -300,6 +300,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
|
||||
|
||||
[encoder endEncoding];
|
||||
[cmd_buf commit];
|
||||
[buf_src release];
|
||||
|
||||
// do not wait here for completion
|
||||
//[cmd_buf waitUntilCompleted];
|
||||
@@ -330,7 +331,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
|
||||
|
||||
// queue the copy operation into the queue of the Metal context
|
||||
// this will be queued at the end, after any currently ongoing GPU operations
|
||||
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
|
||||
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
|
||||
id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
|
||||
|
||||
[encoder copyFromBuffer:bid_src.metal
|
||||
@@ -341,6 +342,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
|
||||
|
||||
[encoder endEncoding];
|
||||
[cmd_buf commit];
|
||||
[buf_dst release];
|
||||
|
||||
// do not wait here for completion
|
||||
//[cmd_buf waitUntilCompleted];
|
||||
|
||||
Reference in New Issue
Block a user