Fix cases simple data movement need to trip into MPSGraph.

lib/nnc/cmd/util/mps/ccv_nnc_util_mps.m

@@ -215,6 +215,27 @@ static int _ccv_nnc_format_transform(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint
 		for (i = 0; i < output_size; i++)
 		{
 			const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[i];
+			const ccv_nnc_tensor_view_t* const b = (const ccv_nnc_tensor_view_t*)outputs[i];
+			// If this is just normal data transfer, do this.
+			if (a->info.format == b->info.format && CCV_IS_TENSOR_CONTIGUOUS(a) && CCV_IS_TENSOR_CONTIGUOUS(b) && ccv_nnc_tensor_count(a->info) == ccv_nnc_tensor_count(b->info) && CCV_GET_DATA_TYPE_SIZE(a->info.datatype) == CCV_GET_DATA_TYPE_SIZE(b->info.datatype))
+			{
+				const size_t size = (ssize_t)ccv_nnc_tensor_count(a->info) * CCV_GET_DATA_TYPE_SIZE(a->info.datatype);
+				if (size == 0)
+					continue;
+				const int device_a = CCV_TENSOR_GET_DEVICE_ID(a->info.type);
+				const int device_b = CCV_TENSOR_GET_DEVICE_ID(b->info.type);
+				assert(device_a == device_b);
+				id<MTLBuffer> buffer_a = mpgetbuffer((const ccv_nnc_tensor_t*)a);
+				id<MTLBuffer> buffer_b = mpgetbuffer((const ccv_nnc_tensor_t*)b);
+				const off_t offset_a = mpgetoffset((const ccv_nnc_tensor_t*)a);
+				const off_t offset_b = mpgetoffset((const ccv_nnc_tensor_t*)b);
+				@autoreleasepool {
+					id<MTLBlitCommandEncoder> encoder = [command_buffer blitCommandEncoder];
+					[encoder copyFromBuffer:buffer_a sourceOffset:offset_a toBuffer:buffer_b destinationOffset:offset_b size:size];
+					[encoder endEncoding];
+				}
+				continue;
+			}
 			ccv_nnc_tensor_view_t bt = ccv_nnc_get_tensor_view(outputs[i]);
 			MPSGraph *graph = [MPSGraph new];
 			graph.options = MPSGraphOptionsSynchronizeResults;