Crate apple_mpsgraph 

mpsgraph-rs

Safe Rust bindings for Apple’s MetalPerformanceShadersGraph framework on macOS.

The GitHub repository is mpsgraph-rs; the published crates.io package is apple-mpsgraph because the short package name is already taken.

Install

cargo add apple-mpsgraph apple-metal

Quick start

use apple_metal::MetalDevice;
use apple_mpsgraph::{data_type, Feed, Graph, TensorData};

let device = MetalDevice::system_default().expect("no Metal device");
let graph = Graph::new().expect("graph");
let input = graph
    .placeholder(Some(&[2, 2]), data_type::FLOAT32, Some("input"))
    .expect("input placeholder");
let bias = graph.constant_scalar(1.0, data_type::FLOAT32).expect("bias");
let added = graph.addition(&input, &bias, Some("add")).expect("add");
let output = graph.relu(&added, Some("relu")).expect("relu");
let data = TensorData::from_f32_slice(&device, &[1.0, -2.0, 3.0, -4.0], &[2, 2])
    .expect("tensor data");
let results = graph
    .run(&[Feed::new(&input, &data)], &[&output])
    .expect("graph run");
let values = results[0].read_f32().expect("read result");
assert_eq!(values, vec![2.0, 0.0, 4.0, 0.0]);

v0.2.1 surface

• Core wrappers for Graph, Tensor, TensorData, Executable, Feed, and FeedDescription
• Metadata and descriptor coverage for GraphDevice, ShapedType, Operation, CompilationDescriptor, ExecutionDescriptor, ExecutableExecutionDescriptor, and ExecutableSerializationDescriptor
• Graph / executable introspection helpers such as placeholder_tensors, feed_tensors, target_tensors, output_types, tensor shape, tensor data_type, and tensor-data graph_device
• Graph construction and execution helpers for:
  • placeholders and constants
  • matrix multiplication
  • unary arithmetic (identity, exponent/log variants, square/sqrt/reciprocal, abs/neg/sign, rounding, trig/hyperbolic, isNaN, isInfinite)
  • binary arithmetic (+, -, *, /, divisionNoNaN, power, min/max, comparisons, logical and/or, atan2, floorModulo, select)
  • activations (reLU, leakyReLU, sigmoid, softMax) and gradient helpers for reLU, sigmoid, and softMax
  • shape ops (reshape, transpose/permute, slice, broadcast, concat, split, stack, pad)
  • reductions (existing sum/max/min/mean plus axis/axes sum/max/min/product)
  • topK
  • call ops via Graph::call plus CompilationDescriptor::set_callable
  • control-flow builders for control dependencies, if/then/else, while, and for
  • gather ops (gather, gatherND, gatherAlongAxis, gatherAlongAxisTensor)
  • descriptor-driven random ops (RandomOpDescriptor, seeded/stateful random tensors, dropout)
  • recurrent layers (singleGateRNN, LSTM, GRU) plus descriptor wrappers
  • 2D convolution, max pooling, and normalization helpers
• Shared constants for MPSDataType, MPSGraphTensorNamedDataLayout, MPSGraphPaddingStyle, graph options, optimization levels, deployment platform values, random distributions, random sampling modes, and RNN activations

This crate still covers a subset of the full SDK. See COVERAGE.md for the audited header-by-header status and deferred areas.

Smoke examples

cargo run --example 01_add_relu
cargo run --example 02_compile_matmul
cargo run --example 03_arithmetic_topk
cargo run --example 04_descriptor_compile
cargo run --example 05_concat_split
cargo run --example 06_control_flow_call
cargo run --example 07_gather_random_rnn

Re-exports

pub use crate::data::TensorData;

pub use crate::error::Error;

pub use crate::error::Result;

pub use crate::control_flow::WhileBeforeResult;

pub use crate::execution::deployment_platform;

pub use crate::execution::graph_options;

pub use crate::execution::optimization;

pub use crate::execution::optimization_profile;

pub use crate::execution::reduced_precision_fast_math;

pub use crate::execution::CompilationDescriptor;

pub use crate::execution::ExecutableExecutionDescriptor;

pub use crate::execution::ExecutableSerializationDescriptor;

pub use crate::execution::ExecutionDescriptor;

pub use crate::graph::data_type;

pub use crate::graph::data_type_size;

pub use crate::graph::padding_style;

pub use crate::graph::tensor_named_data_layout;

pub use crate::graph::Convolution2DDescriptor;

pub use crate::graph::Convolution2DDescriptorInfo;

pub use crate::graph::Executable;

pub use crate::graph::Feed;

pub use crate::graph::FeedDescription;

pub use crate::graph::Graph;

pub use crate::graph::Pooling2DDescriptor;

pub use crate::graph::Pooling2DDescriptorInfo;

pub use crate::graph::Tensor;

pub use crate::ops::BinaryArithmeticOp;

pub use crate::ops::ReductionAxesOp;

pub use crate::ops::ReductionAxisOp;

pub use crate::ops::UnaryArithmeticOp;

pub use crate::random::random_distribution;

pub use crate::random::random_normal_sampling_method;

pub use crate::random::RandomOpDescriptor;

pub use crate::rnn::rnn_activation;

pub use crate::rnn::GRUDescriptor;

pub use crate::rnn::LSTMDescriptor;

pub use crate::rnn::SingleGateRNNDescriptor;

pub use crate::types::graph_device_type;

pub use crate::types::GraphDevice;

pub use crate::types::Operation;

pub use crate::types::ShapedType;

Modules

call

control_flow

data

error

execution

ffi

gather

graph

ops

random

rnn

types