1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188

#![allow(unused_variables)]

use crate::{
    types::{
        BufferSize, ExecutionTarget, Graph, GraphBuilderArray, GraphEncoding,
        GraphExecutionContext, NnErrno, Tensor, TensorType,
    },
    wasi_ephemeral_nn::WasiEphemeralNn,
};
use std::{collections::HashMap, io::Cursor, sync::Arc};
use tract_core::{
    framework::Framework,
    prelude::{DatumType, TVec, Tensor as TractTensor},
};
use tract_tensorflow::prelude::{InferenceModel, InferenceSimplePlan};
use wiggle::{GuestPtr, GuestSlice};

#[derive(Debug, Default)]
pub struct World {
    pub(crate) wasi_nn_ctx: WasiNnCtx,
}

#[derive(Debug, Default)]
pub struct WasiNnCtx {
    models: HashMap<GraphExecutionContext, ExecutionContext>,
    graphs: HashMap<Graph, InferenceModel>,
    last_id: u32,
}

impl WasiNnCtx {
    fn next_id<T>(&mut self) -> T
    where
        T: From<u32>,
    {
        self.last_id += 1;
        self.last_id.into()
    }
}

impl WasiEphemeralNn for WasiNnCtx {
    fn load<'a>(
        &mut self,
        builders: &GraphBuilderArray<'a>,
        encoding: GraphEncoding,
        target: ExecutionTarget,
    ) -> Result<Graph, NnErrno> {
        assert_eq!(encoding, GraphEncoding::Openvino);
        assert_eq!(target, ExecutionTarget::Cpu);

        let model: GuestSlice<u8> =
            builders.as_ptr().read().unwrap().as_slice().unwrap();
        let mut reader = Cursor::new(model.as_ref());

        let graph = tract_tensorflow::tensorflow()
            .model_for_read(&mut reader)
            .unwrap();

        log::info!("Loaded the model:\n{}", graph);

        let id = self.next_id();
        self.graphs.insert(id, graph);

        Ok(id)
    }

    fn init_execution_context(
        &mut self,
        graph: Graph,
    ) -> Result<GraphExecutionContext, NnErrno> {
        let runnable_graph =
            self.graphs.remove(&graph).unwrap().into_runnable().unwrap();

        let id = self.next_id();
        log::info!("Converted {:?} to {:?}", graph, id);

        self.models.insert(
            id,
            ExecutionContext {
                runnable_graph,
                inputs: TVec::default(),
                outputs: TVec::default(),
            },
        );

        Ok(id)
    }

    fn set_input<'a>(
        &mut self,
        context: GraphExecutionContext,
        index: u32,
        tensor: &Tensor<'a>,
    ) -> Result<(), NnErrno> {
        let ctx = self.models.get_mut(&context).unwrap();

        let Tensor {
            dimensions,
            type_,
            data,
        } = tensor;

        let shape = dimensions.as_slice().unwrap();
        let shape: Vec<usize> = shape.iter().map(|&s| s as usize).collect();

        let content = data.as_slice().unwrap();

        let dt = match type_ {
            TensorType::F16 => DatumType::F16,
            TensorType::F32 => DatumType::F32,
            TensorType::U8 => DatumType::U8,
            TensorType::I32 => DatumType::I32,
        };

        let index = index as usize;

        if ctx.inputs.len() < index + 1 {
            ctx.inputs.resize_with(index + 1, Default::default);
        }

        // This function doesn't actually specify why it is "unsafe", but I'm
        // assuming it's just about ensuring our bytes are valid for the
        // underlying data type.
        let tensor =
            unsafe { TractTensor::from_raw_dt(dt, &shape, &content).unwrap() };
        log::info!("Setting input {} on {:?} to {:?}", index, context, tensor);
        ctx.inputs[index] = tensor;

        Ok(())
    }

    fn get_output<'a>(
        &mut self,
        context: GraphExecutionContext,
        index: u32,
        out_buffer: &GuestPtr<'a, u8>,
        out_buffer_max_size: BufferSize,
    ) -> Result<BufferSize, NnErrno> {
        let ctx = self.models.get(&context).unwrap();
        let output_tensor = ctx.outputs.get(index as usize).unwrap();

        let output = unsafe { output_tensor.as_bytes() };

        let output_length = output.len() as u32;
        assert!(out_buffer_max_size >= output_length);

        out_buffer
            .as_array(output_length)
            .as_slice_mut()
            .unwrap()
            .copy_from_slice(output);

        log::info!(
            "Read output {} from {:?}: {:?}",
            index,
            context,
            output_tensor
        );

        Ok(output_length)
    }

    fn compute(
        &mut self,
        context: GraphExecutionContext,
    ) -> Result<(), NnErrno> {
        let ExecutionContext {
            runnable_graph,
            inputs,
            outputs,
        } = self.models.get_mut(&context).unwrap();

        let inputs = std::mem::take(inputs);
        log::info!("Running {:?} with {:?} ", context, inputs,);

        *outputs = runnable_graph.run(inputs).unwrap();

        log::info!("The result is {:?}", outputs);

        Ok(())
    }
}

#[derive(Debug)]
struct ExecutionContext {
    runnable_graph: InferenceSimplePlan<InferenceModel>,
    inputs: TVec<TractTensor>,
    outputs: TVec<Arc<TractTensor>>,
}