khal-std 0.2.0

GPU standard library for khal compute shaders, with cross-platform primitives compiling to SPIR-V, CUDA PTX, and CPU targets.
Documentation 
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

/// Subgroup (wave/warp) floating-point add reduction.
///
/// Returns the sum of `val` across all invocations in the subgroup.
///
/// On SPIR-V: calls `spirv_std::arch::subgroup_f_add`.
/// On CUDA: uses warp-level `__shfl_xor_sync` butterfly reduction.
/// On CPU: returns `val` unchanged (subgroup size = 1).
#[inline(always)]
pub fn subgroup_f_add(val: f32) -> f32 {
    #[cfg(target_arch = "spirv")]
    {
        spirv_std::arch::subgroup_f_add(val)
    }
    #[cfg(target_arch = "nvptx64")]
    {
        warp_reduce_add(val)
    }
    #[cfg(not(any(target_arch = "spirv", target_arch = "nvptx64")))]
    {
        val
    }
}

/// Subgroup (wave/warp) floating-point max reduction.
///
/// Returns the maximum of `val` across all invocations in the subgroup.
///
/// On SPIR-V: calls `spirv_std::arch::subgroup_f_max`.
/// On CUDA: uses warp-level `__shfl_xor_sync` butterfly reduction.
/// On CPU: returns `val` unchanged (subgroup size = 1).
#[inline(always)]
pub fn subgroup_f_max(val: f32) -> f32 {
    #[cfg(target_arch = "spirv")]
    {
        spirv_std::arch::subgroup_f_max(val)
    }
    #[cfg(target_arch = "nvptx64")]
    {
        warp_reduce_max(val)
    }
    #[cfg(not(any(target_arch = "spirv", target_arch = "nvptx64")))]
    {
        val
    }
}

#[cfg(target_arch = "nvptx64")]
#[inline(always)]
fn shfl_xor_sync(val: f32, lane_mask: u32) -> f32 {
    let result: f32;
    unsafe {
        core::arch::asm!(
            "shfl.sync.bfly.b32 {result}, {val}, {lane_mask}, 0x1f, 0xffffffff;",
            result = out(reg32) result,
            val = in(reg32) val,
            lane_mask = in(reg32) lane_mask,
        );
    }
    result
}

#[cfg(target_arch = "nvptx64")]
#[inline(always)]
fn warp_reduce_add(mut val: f32) -> f32 {
    val += shfl_xor_sync(val, 16);
    val += shfl_xor_sync(val, 8);
    val += shfl_xor_sync(val, 4);
    val += shfl_xor_sync(val, 2);
    val += shfl_xor_sync(val, 1);
    val
}

#[cfg(target_arch = "nvptx64")]
#[inline(always)]
fn warp_reduce_max(mut val: f32) -> f32 {
    let other = shfl_xor_sync(val, 16);
    if other > val {
        val = other;
    }
    let other = shfl_xor_sync(val, 8);
    if other > val {
        val = other;
    }
    let other = shfl_xor_sync(val, 4);
    if other > val {
        val = other;
    }
    let other = shfl_xor_sync(val, 2);
    if other > val {
        val = other;
    }
    let other = shfl_xor_sync(val, 1);
    if other > val {
        val = other;
    }
    val
}