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
211
212
213
214
215
216
217
218
219
220
// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
// SPDX-License-Identifier: MIT OR Apache-2.0
//! OpKinds this backend claims for legalization (`Backend::supported_ops`).
//!
//! Source of truth for the coverage matrix in `docs/op-coverage.md`.
//! Kept in the backend crate so adding an op is a local edit, not a change
//! to `rlx-runtime`'s mega-`backend.rs`.
pub const SUPPORTED_OPS: &[rlx_ir::OpKind] = {
use rlx_ir::OpKind::*;
&[
Input,
Param,
Constant,
Activation,
Cast,
StopGradient,
Binary,
Compare,
Where,
Fma,
ElementwiseRegion,
// Claimed for legalize / FK fusion; `expand::prepare_graph_for_thunks`
// runs `DecomposeFusionRegions` before thunks (CPU has no native FKL
// region executor beyond plain ElementwiseRegion).
TransformRegion,
BatchElementwiseRegion,
MatMul,
DotGeneral,
DenseSolve,
BatchedDenseSolve,
Cholesky,
TriangularSolve,
Det,
LogDet,
Sort,
ArgSort,
Svd,
Qr,
Scan,
ScanBackward,
ScanBackwardXs,
LayerNorm,
LayerNorm2d,
GroupNorm,
BatchNormInference,
RmsNorm,
ResizeNearest2x,
Interpolate3d,
AxialRope2d,
Attention,
Rope,
Reshape,
Transpose,
Narrow,
Concat,
KvAppend,
Expand,
Gather,
Reverse,
Reduce,
Histogram,
Softmax,
Cumsum,
CumProd,
CumMax,
ArgMax,
ArgMin,
TopK,
Sample,
RngNormal,
RngUniform,
Conv,
Im2Col,
ConvTranspose2d,
Conv3d,
ConvTranspose3d,
Pool,
GroupedMatMul,
DequantGroupedMatMul,
DequantGroupedMatMulMlx,
DequantMoEWeights,
ScatterAdd,
ScatterNd,
ScatterElements,
GatherNd,
GatherElements,
LoraMatMul,
DequantMatMul,
SynthMatMul,
SplineActivation,
ScaledMatMul,
ScaledGroupedMatMul,
ScaledQuantize,
ScaledQuantScale,
ScaledDequantize,
Quantize,
Dequantize,
QMatMul,
QConv2d,
SelectiveScan,
GatedDeltaNet,
GatedDeltaNetBackward,
Lstm,
Gru,
Rnn,
Mamba2,
FusedSwiGLU,
FusedMatMulBiasAct,
FusedResidualLN,
FusedResidualRmsNorm,
FusedAttentionBlock,
// Claimed then expanded by `expand::expand_cpu_nop_fused` (CPU thunk
// catch-all would Nop these — same pattern as Metal / OneAPI).
FusedConvBiasAct,
FusedTransformerLayer,
PartitionedConv,
// Claimed; `LowerControlFlow` (fusion + CpuBackend::compile) expands
// to Where / unrolled body replicas before thunks.
If,
While,
AdaLayerNorm,
GatedResidual,
AdaLayerNormBackward,
GatedResidualBackward,
// Backward ops emitted by `rlx_opt::autodiff::grad_with_loss`.
// Their thunks live in rlx-cpu/src/thunk.rs alongside the
// forward kernels; without these entries the legalize step
// below would reject any compiled gradient graph.
ReluBackward,
ActivationBackward,
FakeQuantize,
FakeQuantizeBackward,
// LSQ (learned step size) QAT — native CPU thunks in thunk.rs.
FakeQuantizeLSQ,
FakeQuantizeLSQBackwardX,
FakeQuantizeLSQBackwardScale,
MaxPool2dBackward,
Conv2dBackwardInput,
Conv2dBackwardWeight,
MaxPool3dBackward,
Conv3dBackwardInput,
Conv3dBackwardWeight,
SoftmaxCrossEntropy,
SoftmaxCrossEntropyWithLogits,
SoftmaxCrossEntropyBackward,
AttentionBackward,
AttentionBackwardAll,
LayerNormBackwardInput,
LayerNormBackwardGamma,
BatchNormInferenceBackwardInput,
BatchNormInferenceBackwardGamma,
BatchNormInferenceBackwardBeta,
// GroupNorm backward (native thunks in rlx-cpu/training_bwd):
GroupNormBackwardInput,
GroupNormBackwardGamma,
GroupNormBackwardBeta,
RmsNormBackwardInput,
RmsNormBackwardGamma,
RmsNormBackwardBeta,
RopeBackward,
CumsumBackward,
GatherBackward,
// 3D Gaussian splat CPU reference render/backward (requires `rlx-cpu/splat`).
GaussianSplatRender,
GaussianSplatRenderBackward,
GaussianSplatPrepare,
GaussianSplatRasterize,
// User-registered custom ops dispatched through
// `rlx_cpu::op_registry`. Lowering panics with a clear
// message if the named CPU kernel isn't registered.
Custom,
// User-defined sub-graph with optional override AD rules
// (JAX-shaped custom_vjp / custom_jvp). Body is a regular
// Graph compiled recursively in compile_thunks.
CustomFn,
// FFT primitive (1D last-axis, 2N real-block layout, f64
// power-of-2 sizes). Other backends panic at lowering;
// pin FFT-containing graphs to Device::Cpu for now.
Fft,
FftQ,
FftButterflyStage,
LogMel,
LogMelBackward,
WelchPeaks,
// C64 Wirtinger AD surface. ComplexNormSq is the canonical
// real-valued loss for complex inputs; Conjugate is emitted
// by the new Wirtinger VJP rules for BinaryOp::Mul/Div on
// C64. Both have CPU thunks in rlx-cpu.
ComplexNormSq,
ComplexNormSqBackward,
Conjugate,
// Riemannian / SPD-manifold layers (SPDNet + SPD batch-norm).
// CPU-first (F64); other backends fall through their catch-alls.
BiMap,
ReEig,
LogEig,
SpdBatchNorm,
SpdKarcherMean,
SpdKarcherMeanWeighted,
SpdLogMap,
SpdExpMap,
SpdParallelTransport,
SpdMatrixFnBatch,
ReEigBackward,
LogEigBackward,
SpdBatchNormBackwardX,
SpdBatchNormBackwardG,
SpdLogMapBackward,
SpdExpMapBackward,
SpdParallelTransportBackward,
SpdMatrixFnBatchBackward,
Eigh,
EighBackward,
EighBatch,
EighBatchBackward,
]
};