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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! GPU driver identity types — enums for driver, compiler, architecture, and strategy.
//!
//! These types answer **"who is driving the hardware?"** and feed into
//! [`super::capabilities::DeviceCapabilities`] for workgroup sizing,
//! precision routing, and FP64 strategy selection.
//!
//! The former `GpuDriverProfile` struct was removed in v0.3.8 (Sprint 18).
//! All consumers migrated to `DeviceCapabilities` in Sprint 14.
pub use GpuArch;
pub use Workaround;
// ── Driver identity ───────────────────────────────────────────────────────────
/// GPU driver/compiler identity.
/// GPU shader compiler backend.
// ── FP64 rate ─────────────────────────────────────────────────────────────────
/// FP64 hardware rate classification.
// ── Eigensolve strategy ───────────────────────────────────────────────────────
/// Eigensolve dispatch strategy chosen based on driver/arch.
// ── FP64 execution strategy ──────────────────────────────────────────────────
/// Hardware-adaptive FP64 execution strategy (physics domain core-streaming discovery).
///
/// On compute-class GPUs (Titan V, V100, MI250) with 1:2 FP64:FP32 hardware,
/// native `f64` is fastest. On consumer GPUs (1:64 ratio), routing bulk math
/// through double-float f32-pair (`Df64`) on the massive FP32 core array
/// delivers ~10x the effective throughput, reserving native `f64` for
/// precision-critical reductions and convergence tests.
// ── Precision routing (toadStool S128 integration) ──────────────────────────
/// Precision routing advice from toadStool S128.
///
/// Captures the three-axis reality of GPU f64: some hardware has native f64
/// everywhere, some has native f64 but broken shared-memory f64 reductions
/// (naga/SPIR-V emit zeros), some can only do DF64, and some are f32-only.
///
/// Use this to route workloads to the correct precision path at dispatch time,
/// especially for reductions that rely on `var<workgroup>` f64 accumulators.