Graph split

ONNX2MPSX.py

@@ -50,6 +50,7 @@ def convert_onnx_to_mpsx(model, halfs):
         onnx.helper.make_graph(
             nodes=model.graph.node,
             name=model.graph.name,
+            value_info=model.graph.value_info,
             inputs=_fp32_to_fp16_info(
                 model.graph.input) if halfs else model.graph.input,
             outputs=_fp32_to_fp16_info(

Sources/MPSX/Foundation/DSL.swift

@@ -412,8 +412,8 @@ public extension MPSGraphTensor {
             self,
             size: [height, width].nsnumbers,
             mode: mode,
-            centerResult: true,
-            alignCorners: false,
+            centerResult: false,
+            alignCorners: true,
             layout: layout,
             name: nil
         )

Sources/MPSX/Foundation/Utilities.swift

@@ -468,3 +468,11 @@ public extension MPSGraphTensor {
         shape?.map(\.intValue) ?? []
     }
 }
+
+public extension MPSCommandBuffer {
+    func commitAndWait() {
+        let _rootCommandBuffer = rootCommandBuffer
+        commitAndContinue()
+        _rootCommandBuffer.waitUntilCompleted()
+    }
+}

Sources/MPSX/ONNX/Experimental/OnnxCustomNode.swift

@@ -0,0 +1,31 @@
+import MetalPerformanceShadersGraph
+
+public protocol OnnxAttributesProvider {
+    func attr(s name: String) -> String
+    func attr(f name: String) -> Float?
+    func attr(floats name: String) -> [Float]?
+    func attr(i name: String) -> Int?
+    func attr(ints name: String) -> [Int]?
+}
+
+public protocol OnnxCustomNode {
+    func preprocess(
+        inputTensor: MPSGraphTensor,
+        inputName: String,
+        graph: MPSGraph
+    ) -> MPSGraphTensor
+
+    func postprocess(
+        outputName: String,
+        outputShape: [Int],
+        requiredDataType: MPSDataType,
+        graph: MPSGraph
+    ) -> (placeholder: MPSGraphTensor, tensor: MPSGraphTensor)
+
+    func eval(
+        inputs: [MPSGraphTensorData],
+        outputShapes: [[Int]],
+        attributesProvider: OnnxAttributesProvider,
+        in commandBuffer: MPSCommandBuffer
+    ) throws -> [MPSGraphTensorData]
+}

Sources/MPSX/ONNX/Experimental/OnnxModel+Split.swift

@@ -0,0 +1,80 @@
+extension OnnxModel {
+    /// Split the graph into subgraphs.
+    /// - Parameter splitOps: The names of the operators in the graph by which you want to split
+    /// - Returns: A map where key is the name of the node (node opType exists in splitOps) and value is an array of input tensors that are "alive" at the time of the split.
+    func split(by splitOps: Set<String>) -> [String: Set<String>] {
+        struct NodeCounter {
+            /// How many times the output of a node will be used as an input
+            var readCount: Int
+            /// Node lifetime: lower bound == first occurrence in the graph, upper bound == last occurrence as input in the graph.
+            var range: ClosedRange<Int>
+        }
+
+        let graph = proto.graph
+
+        var counters: [String: NodeCounter] = [:]
+        counters.reserveCapacity(graph.node.count)
+
+        var nodeIndices: [String: Int] = [:]
+        nodeIndices.reserveCapacity(graph.node.count)
+
+        // STEP 1: calculate counters for every node from start to end
+
+        for (index, node) in graph.node.enumerated() {
+            for outputName in node.output {
+                nodeIndices[outputName] = index
+            }
+
+            for inputName in node.input {
+                // only runtime tensors
+
+                guard initializer[inputName] == nil else {
+                    continue
+                }
+
+                let inputIndex = nodeIndices[inputName, default: 0]
+
+                let counter = counters[inputName, default: .init(readCount: 0, range: inputIndex ... inputIndex)]
+
+                counters[inputName] = .init(
+                    readCount: counter.readCount + 1,
+                    range: counter.range.lowerBound ... index
+                )
+            }
+        }
+
+        // STEP 2: now we have counters for every node and can split graph into subgraphs using this info.
+
+        var subgraphs: [String: Set<String>] = [:]
+
+        for (index, node) in graph.node.enumerated() {
+            for inputName in node.input {
+                guard initializer[inputName] == nil,
+                      var counter = counters[inputName]
+                else {
+                    continue
+                }
+
+                if counter.readCount > 0 {
+                    counter.readCount -= 1
+
+                    counters[inputName] = counter
+                } else {
+                    counters[inputName] = nil
+                }
+            }
+
+            // split
+
+            if splitOps.contains(node.opType) {
+                // TODO: optimize using some tree data structure/sorted collection
+
+                let inputs = counters.filter { $0.value.range.contains(index) }
+
+                subgraphs[node.name] = Set(inputs.map(\.key)).subtracting(Set(node.output))
+            }
+        }
+
+        return subgraphs
+    }
+}