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
// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.
//! 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,
TransformRegion,
BatchElementwiseRegion,
MatMul,
ScaledMatMul,
ScaledQuantize,
ScaledQuantScale,
ScaledDequantize,
DotGeneral,
LayerNorm,
LayerNorm2d,
GroupNorm,
RmsNorm,
ResizeNearest2x,
AxialRope2d,
Attention,
AttentionBackward,
RmsNormBackwardInput,
RmsNormBackwardGamma,
RmsNormBackwardBeta,
RopeBackward,
Cumsum,
CumsumBackward,
GatherBackward,
Conv2dBackwardInput,
Conv2dBackwardWeight,
MaxPool2dBackward,
Rope,
Reshape,
Transpose,
Narrow,
Concat,
Expand,
Gather,
Reverse,
Reduce,
Softmax,
SoftmaxCrossEntropy,
SoftmaxCrossEntropyWithLogits,
SoftmaxCrossEntropyBackward,
ArgMax,
ArgMin,
TopK,
Sample,
RngNormal,
RngUniform,
Conv,
Im2Col,
ConvTranspose2d,
Pool,
GroupedMatMul,
DequantGroupedMatMul,
DequantMoEWeights,
ScatterAdd,
ScatterNd,
ScatterElements,
GatherNd,
GatherElements,
DequantMatMul,
GatedDeltaNet,
SelectiveScan,
Lstm,
Gru,
Rnn,
Mamba2,
FusedSwiGLU,
FusedMatMulBiasAct,
FusedResidualLN,
FusedResidualRmsNorm,
// Claimed so the Metal fusion pipeline may emit it;
// `MetalExecutable::compile_inner` decomposes it back to the
// primitive chain (no monolithic fused-attention MSL kernel
// yet — the per-run cost is dominated by wait_until_completed,
// not encode, so a dispatch-wrapper fusion buys nothing).
FusedAttentionBlock,
// DiT adaLN-Zero / gated residual — native MSL kernels avoid
// Expand of `[B,1,D]` modulation over the sequence axis.
AdaLayerNorm,
GatedResidual,
AdaLayerNormBackward,
GatedResidualBackward,
// User-registered custom ops dispatched through
// `rlx_metal::op_registry`. Lowering panics with a clear
// message if the named MetalKernel isn't registered;
// executor inserts a sync point + runs the host kernel
// against the unified-memory arena.
Custom,
// Op::Fft is supported via the same host-fallback pattern
// as Custom: sync the GPU, run rlx-cpu's FFT against the
// unified-memory arena, restart cmd_buf. A native Metal
// compute kernel will replace this when a workload makes
// the sync the bottleneck.
Fft,
// Op::Scan (arbitrary-body recurrence) via the same host
// fallback: compile the body once, loop it on the CPU against
// the unified-memory arena. Enables IIR (`biquad`/`sosfilt`).
Scan,
ScanBackward,
ScanBackwardXs,
LogMel,
LogMelBackward,
WelchPeaks,
// Host-fallback splat (unified-memory arena + rlx-cpu/splat).
GaussianSplatRender,
GaussianSplatRenderBackward,
GaussianSplatPrepare,
GaussianSplatRasterize,
// Core Riemannian / SPD-manifold ops. No MSL eigen kernel; they
// host-fallback to `rlx_cpu::spd` (F64) against the unified-memory
// arena via the same sync pattern as Fft/Custom (see
// `rlx_metal::spd` + `Thunk::SpdHost`). The SPD subgraph's F64
// tensors are widened to f32 for arena planning; the host step does
// the f32↔f64 conversion.
BiMap,
ReEig,
LogEig,
SpdBatchNorm,
SpdKarcherMean,
SpdKarcherMeanWeighted,
SpdLogMap,
SpdExpMap,
SpdParallelTransport,
SpdMatrixFnBatch,
ReEigBackward,
LogEigBackward,
SpdBatchNormBackwardX,
SpdBatchNormBackwardG,
SpdLogMapBackward,
SpdExpMapBackward,
SpdParallelTransportBackward,
SpdMatrixFnBatchBackward,
Eigh,
EighBackward,
EighBatch,
EighBatchBackward,
]
};