More code

Author: christiangnrd

lib/mtl/command_enc/compute4.jl

@@ -44,14 +44,14 @@ function dispatchThreads!(cce::MTL4ComputeCommandEncoder, threadsSize::MTLSize,
 end
 
 # Copy Operations (Blit functionality integrated into compute encoder in Metal 4)
-# function append_copy!(cce::MTL4ComputeCommandEncoder, dst::MTLBuffer, dstOffset::Integer,
-#                       src::MTLBuffer, srcOffset::Integer, size::Integer)
-#     @objc [cce::id{MTL4ComputeCommandEncoder} copyFromBuffer:src::id{MTLBuffer}
-#                                              sourceOffset:srcOffset::NSUInteger
-#                                              toBuffer:dst::id{MTLBuffer}
-#                                              destinationOffset:dstOffset::NSUInteger
-#                                              size:size::NSUInteger]::Nothing
-# end
+function append_copy!(cce::MTL4ComputeCommandEncoder, dst::MTLBuffer, dstOffset::Integer,
+                      src::MTLBuffer, srcOffset::Integer, size::Integer)
+    @objc [cce::id{MTL4ComputeCommandEncoder} copyFromBuffer:src::id{MTLBuffer}
+                                             sourceOffset:srcOffset::NSUInteger
+                                             toBuffer:dst::id{MTLBuffer}
+                                             destinationOffset:dstOffset::NSUInteger
+                                             size:size::NSUInteger]::Nothing
+end
 
 # function append_copy!(cce::MTL4ComputeCommandEncoder, dst::MTLTexture, dstSlice::Integer, dstLevel::Integer, dstOrigin::MTLOrigin,
 #                       src::MTLBuffer, srcOffset::Integer, srcBytesPerRow::Integer, srcBytesPerImage::Integer,

src/memory.jl

@@ -31,7 +31,7 @@ Base.convert(::Type{Ptr{T}}, ptr::MtlPtr) where {T} =
 
 # CPU -> GPU
 function Base.unsafe_copyto!(dev::MTLDevice, dst::MtlPtr{T}, src::Ptr{T}, N::Integer;
-                             queue::MTLCommandQueue=global_queue(dev), async::Bool=false) where T
+                             queue=use_metal4() ? global_queue4(dev) : global_queue(dev), async::Bool=false) where T
     storage_type = dst.buffer.storageMode
     if storage_type == MTL.MTLStorageModePrivate
         # stage through a shared buffer
@@ -53,7 +53,7 @@ end
 
 # GPU -> CPU
 function Base.unsafe_copyto!(dev::MTLDevice, dst::Ptr{T}, src::MtlPtr{T}, N::Integer;
-                             queue::MTLCommandQueue=global_queue(dev), async::Bool=false) where T
+                             queue=use_metal4() ? global_queue4(dev) : global_queue(dev), async::Bool=false) where T
     storage_type = src.buffer.storageMode
     if storage_type == MTL.MTLStorageModePrivate
         # stage through a shared buffer
@@ -88,15 +88,25 @@ end
 # GPU -> GPU
 @autoreleasepool function Base.unsafe_copyto!(dev::MTLDevice, dst::MtlPtr{T},
                                               src::MtlPtr{T}, N::Integer;
-                                              queue::MTLCommandQueue=global_queue(dev),
+                                              queue=use_metal4() ? global_queue4(dev) : global_queue(dev),
                                               async::Bool=false) where T
     if N > 0
-        cmdbuf = MTLCommandBuffer(queue)
-        MTLBlitCommandEncoder(cmdbuf) do enc
-            append_copy!(enc, dst.buffer, dst.offset, src.buffer, src.offset, N * sizeof(T))
+        if queue isa MTL4CommandQueue
+            @info "MTL4"
+            cmdbuf = MTL4CommandBuffer(dev; queue) do cmdbuf
+                MTL4ComputeCommandEncoder(cmdbuf, !async) do enc
+                    append_copy!(enc, dst.buffer, dst.offset, src.buffer, src.offset, N * sizeof(T))
+                end
+            end
+        else
+            @info "MTL3"
+            cmdbuf = MTLCommandBuffer(queue)
+            MTLBlitCommandEncoder(cmdbuf) do enc
+                append_copy!(enc, dst.buffer, dst.offset, src.buffer, src.offset, N * sizeof(T))
+            end
+            commit!(cmdbuf)
+            async || wait_completed(cmdbuf)
         end
-        commit!(cmdbuf)
-        async || wait_completed(cmdbuf)
     end
     return dst
 end

src/state.jl

@@ -100,7 +100,9 @@ end
     end
 
     cmdbuf = commit!(cmdbuf, queue, allocator) do cmdbuf
-        MTL4ComputeCommandEncoder(identity, cmdbuf, #=sync=#true)
+        encoder = MTL4ComputeCommandEncoder(cmdbuf)
+        MTL.barrierAfterQueueStages!(encoder)
+        close(encoder)
     end
     return
 end