bb-ops 0.3.4

Canonical concrete components for the bytesandbrains framework — syscalls, wire transport, backends, protocols, role implementations.
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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210

//! Per-op kernel dispatch for the `CpuBackend`.
//!
//! The 51-op `ai.onnx v1` catalog is grouped into category modules.
//! This module routes each declared op to its category-specific
//! kernel; categories that don't yet have implementations return a
//! clear "not yet implemented" error.
//!
//! **Attribute access limitation.** `BackendRuntime::dispatch_atomic`
//! receives `(op_type, inputs, attrs)` - no NodeProto, no attribute
//! map. Ops whose semantics depend on attributes (Reshape's `dims`,
//! Softmax's `axis`, Conv's `kernel_shape`, etc.) cannot be fully
//! implemented within this dispatch shape. They report
//! `RequiresAttributes` so callers can route them through a
//! follow-up channel.

mod elementwise;
mod linalg;
mod shape;

use bb_ir::proto::onnx::{attribute_proto::AttributeType, AttributeProto};
use bb_runtime::atomic::DispatchResult;
use bb_runtime::bus::OpError;
use bb_runtime::slot_value::SlotValue;

use crate::backends::cpu::CpuBackend;
use crate::backends::cpu::CpuTensor;

/// Find an attribute by name in the supplied `attrs` slice.
fn find_attr<'a>(attrs: &'a [AttributeProto], name: &str) -> Option<&'a AttributeProto> {
    attrs.iter().find(|a| a.name == name)
}

fn need_int_attr(op: &str, attrs: &[AttributeProto], name: &str) -> Result<i64, OpError> {
    let a = find_attr(attrs, name).ok_or_else(|| OpError {
        detail: format!("{op}: missing `{name}` attribute"),
        ..Default::default()
    })?;
    if a.r#type != AttributeType::Int as i32 {
        return Err(OpError {
            detail: format!("{op}: `{name}` attribute must be INT"),
            ..Default::default()
        });
    }
    Ok(a.i)
}

fn need_ints_attr(op: &str, attrs: &[AttributeProto], name: &str) -> Result<Vec<i64>, OpError> {
    let a = find_attr(attrs, name).ok_or_else(|| OpError {
        detail: format!("{op}: missing `{name}` attribute"),
        ..Default::default()
    })?;
    if a.r#type != AttributeType::Ints as i32 {
        return Err(OpError {
            detail: format!("{op}: `{name}` attribute must be INTS"),
            ..Default::default()
        });
    }
    Ok(a.ints.clone())
}

fn opt_float_attr(attrs: &[AttributeProto], name: &str, default: f32) -> f32 {
    find_attr(attrs, name)
        .filter(|a| a.r#type == AttributeType::Float as i32)
        .map(|a| a.f)
        .unwrap_or(default)
}

fn opt_int_attr(attrs: &[AttributeProto], name: &str, default: i64) -> i64 {
    find_attr(attrs, name)
        .filter(|a| a.r#type == AttributeType::Int as i32)
        .map(|a| a.i)
        .unwrap_or(default)
}

/// Downcast a `&dyn SlotValue` to `&CpuTensor`, raising `OpError`
/// with a clear message on type mismatch.
fn as_cpu_tensor<'a>(op: &str, role: &str, h: &'a dyn SlotValue) -> Result<&'a CpuTensor, OpError> {
    h.as_any()
        .downcast_ref::<CpuTensor>()
        .ok_or_else(|| OpError {
            detail: format!("{op}: {role} is not a CpuTensor"),
            ..Default::default()
        })
}

fn need_two_inputs<'a>(
    op: &str,
    inputs: &'a [(&str, &dyn SlotValue)],
) -> Result<(&'a CpuTensor, &'a CpuTensor), OpError> {
    if inputs.len() < 2 {
        return Err(OpError {
            detail: format!("{op}: requires two inputs, got {}", inputs.len()),
            ..Default::default()
        });
    }
    let a = as_cpu_tensor(op, "input 0", inputs[0].1)?;
    let b = as_cpu_tensor(op, "input 1", inputs[1].1)?;
    Ok((a, b))
}

fn need_one_input<'a>(
    op: &str,
    inputs: &'a [(&str, &dyn SlotValue)],
) -> Result<&'a CpuTensor, OpError> {
    if inputs.is_empty() {
        return Err(OpError {
            detail: format!("{op}: requires one input, got 0"),
            ..Default::default()
        });
    }
    as_cpu_tensor(op, "input 0", inputs[0].1)
}

fn out(name: &str, tensor: CpuTensor) -> DispatchResult {
    DispatchResult::Immediate(vec![(name.to_string(), Box::new(tensor))])
}

/// Route `op_type` to the matching kernel. `attrs` is the
/// NodeProto's attribute slice — ops whose semantics depend on
/// attributes (Reshape's `dims`, Softmax's `axis`, Gemm's
/// `alpha`/`beta`, …) read it directly. The framework's per-op
/// dispatch path threads `ctx.current.node_attributes` here; the
/// `execute_graph` walker threads `node.attribute` per node.
pub fn dispatch(
    backend: &CpuBackend,
    op_type: &str,
    inputs: &[(&str, &dyn SlotValue)],
    attrs: &[AttributeProto],
) -> Result<DispatchResult, OpError> {
    match op_type {
        // --- Element-wise binary -----------------------------------
        "Add" => Ok(out("C", elementwise::add(backend, op_type, inputs)?)),
        "Sub" => Ok(out("C", elementwise::sub(backend, op_type, inputs)?)),
        "Mul" => Ok(out("C", elementwise::mul(backend, op_type, inputs)?)),
        "Div" => Ok(out("C", elementwise::div(backend, op_type, inputs)?)),
        "Pow" => Ok(out("C", elementwise::pow(backend, op_type, inputs)?)),

        // --- Element-wise unary ------------------------------------
        "Neg" => Ok(out("Y", elementwise::neg(backend, op_type, inputs)?)),
        "Abs" => Ok(out("Y", elementwise::abs(backend, op_type, inputs)?)),
        "Sqrt" => Ok(out("Y", elementwise::sqrt(backend, op_type, inputs)?)),
        "Exp" => Ok(out("Y", elementwise::exp(backend, op_type, inputs)?)),
        "Log" => Ok(out("Y", elementwise::log(backend, op_type, inputs)?)),

        // --- Activations -------------------------------------------
        "Relu" => Ok(out("Y", elementwise::relu(backend, op_type, inputs)?)),
        "Sigmoid" => Ok(out("Y", elementwise::sigmoid(backend, op_type, inputs)?)),
        "Tanh" => Ok(out("Y", elementwise::tanh(backend, op_type, inputs)?)),
        "Gelu" => Ok(out("Y", elementwise::gelu(backend, op_type, inputs)?)),
        "Identity" => Ok(out("Y", elementwise::identity(backend, op_type, inputs)?)),
        "Softmax" => Ok(out("Y", shape::softmax(backend, op_type, inputs, attrs)?)),
        "LeakyRelu" => Ok(out(
            "Y",
            shape::leaky_relu(backend, op_type, inputs, attrs)?,
        )),

        // --- Element-wise comparison -------------------------------
        "Equal" => Ok(out("C", elementwise::equal(backend, op_type, inputs)?)),
        "Greater" => Ok(out("C", elementwise::greater(backend, op_type, inputs)?)),
        "Less" => Ok(out("C", elementwise::less(backend, op_type, inputs)?)),

        // --- Linear algebra ----------------------------------------
        "MatMul" => Ok(out("Y", linalg::matmul(backend, op_type, inputs)?)),
        "Dot" => Ok(out("Y", linalg::dot(backend, op_type, inputs)?)),
        "Gemm" => Ok(out("Y", shape::gemm(backend, op_type, inputs, attrs)?)),

        // --- Reductions --------------------------------------------
        "ReduceSum" => Ok(out("Y", linalg::reduce_sum(backend, op_type, inputs)?)),
        "ReduceMean" => Ok(out("Y", linalg::reduce_mean(backend, op_type, inputs)?)),
        "ReduceMax" => Ok(out("Y", linalg::reduce_max(backend, op_type, inputs)?)),
        "ReduceMin" => Ok(out("Y", linalg::reduce_min(backend, op_type, inputs)?)),

        // --- Shape / structural ------------------------------------
        "Reshape" => Ok(out("Y", shape::reshape(backend, op_type, inputs, attrs)?)),
        "Transpose" => Ok(out("Y", shape::transpose(backend, op_type, inputs, attrs)?)),
        "Concat" => Ok(out("Y", shape::concat(backend, op_type, inputs, attrs)?)),
        "Squeeze" => Ok(out("Y", shape::squeeze(backend, op_type, inputs, attrs)?)),
        "Unsqueeze" => Ok(out("Y", shape::unsqueeze(backend, op_type, inputs, attrs)?)),
        "Cast" => Ok(out("Y", shape::cast(backend, op_type, inputs, attrs)?)),
        "Slice" => Ok(out("Y", shape::slice(backend, op_type, inputs, attrs)?)),
        "Split" => Ok(shape::split(backend, op_type, inputs, attrs)?),

        // --- Indexing ----------------------------------------------
        "Gather" => Ok(out("Y", shape::gather(backend, op_type, inputs, attrs)?)),

        // --- Pooling without attributes ----------------------------
        "GlobalAveragePool" => Ok(out(
            "Y",
            linalg::global_average_pool(backend, op_type, inputs)?,
        )),

        // --- Creation ----------------------------------------------
        "Zeros" => Ok(out("Y", shape::zeros(backend, op_type, attrs)?)),
        "Ones" => Ok(out("Y", shape::ones(backend, op_type, attrs)?)),
        "Constant" => Ok(out("Y", shape::constant(backend, op_type, attrs)?)),

        // `BackendSubgraph` is routed by the engine's
        // `invoke_backend_subgraph` path, not this dispatch.
        // BatchNorm / LayerNorm / Conv / MaxPool / AveragePool /
        // Scatter / If / Loop are NOT in this backend's declared
        // opset — `Node::ready`'s per-`BackendSubgraph` check
        // catches them at install time. They surface here only as
        // a defensive fallthrough.
        other => Err(OpError {
            detail: format!("CpuBackend: unsupported op_type '{other}'"),
            ..Default::default()
        }),
    }
}