oxibonsai-kernels 0.1.4

1-bit Q1_0_g128 compute kernels (dequant, GEMV, GEMM) for OxiBonsai
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
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

//! Metal MSL kernel sources for FP8 (E4M3FN and E5M2) GEMV operations.
//!
//! Mirrors the CUDA FP8 kernels in `crates/oxibonsai-kernels/src/gpu_backend/cuda_fp8_kernels.rs`.
//!
//! # Block layout (AoS, 34 bytes/block — matches `BlockFP8E4M3` / `BlockFP8E5M2`)
//!
//! ```text
//! Block[i] = [q0, q1, ..., q31, scale_lo, scale_hi]
//!             ^^^^^^^^^^^^^^^^^ 32 FP8 bytes ^^^^   ^^ FP16 LE scale ^^
//! ```
//!
//! This matches the `#[repr(C)]` layout of `BlockFP8E4M3 { qs: [u8; 32], d: f16 }`:
//! `qs` occupies bytes 0-31, `d` (FP16) occupies bytes 32-33.
//!
//! # Dispatch
//!
//! Grid:  `[ceil(n_rows / 8), 1, 1]` — 8 simdgroups per threadgroup, one row per simdgroup
//! Block: `[256, 1, 1]` — 8 simdgroups × 32 lanes
//!
//! Buffer indices (scirs2-core naming convention):
//! - `"x"` → blocks (0, u8*)
//! - `"y"` → input  (1, float*)
//! - `"result"` → output (2, float*)
//! - `"n"` → n_rows (3, scalar uint)
//! - `"k"` → k      (4, scalar uint)

/// Metal MSL kernel: FP8 E4M3FN GEMV.
///
/// Format: s[7] exp[6:3] man[2:0], bias=7
/// - Normal: `(-1)^s * 2^(exp - 7) * (1 + man/8)`
/// - Denorm: `(-1)^s * 2^(-6) * (man/8)`
/// - NaN patterns (0x7F, 0xFF) → 0 for inference
///
/// One simdgroup per row. Each lane strides 32-block chunks.
#[cfg(all(feature = "metal", target_os = "macos"))]
pub const MSL_GEMV_FP8_E4M3_V1: &str = r#"
#include <metal_stdlib>
using namespace metal;

// FP8 E4M3FN decode helper (bias=7, no infinity).
static inline float fp8_e4m3_to_float(uchar b) {
    // NaN patterns: 0x7F and 0xFF → return 0 for inference.
    if (b == 0x7Fu || b == 0xFFu) return 0.0f;
    const uint sign = (uint(b) >> 7u) & 1u;
    const uint exp  = (uint(b) >> 3u) & 15u;  // 4-bit exponent
    const uint mant = uint(b) & 7u;            // 3-bit mantissa
    float val;
    if (exp == 0u) {
        // Denormal: 2^(-6) * (mant / 8)
        val = float(mant) * (1.0f / 8.0f) * (1.0f / 64.0f);
    } else {
        // Normal: 2^(exp - 7) * (1 + mant/8)
        // Assemble as IEEE-754 f32: ((exp - 7 + 127) << 23) | (mant << 20)
        const uint bits = ((exp - 7u + 127u) << 23u) | (mant << 20u);
        val = as_type<float>(bits);
    }
    return sign ? -val : val;
}

kernel void gemv_fp8_e4m3(
    device const uchar* blocks_raw [[buffer(0)]],
    device const float* input      [[buffer(1)]],
    device float* output           [[buffer(2)]],
    constant uint& n_rows          [[buffer(3)]],
    constant uint& k               [[buffer(4)]],
    uint tgid  [[threadgroup_position_in_grid]],
    uint sgid  [[simdgroup_index_in_threadgroup]],
    uint lane  [[thread_index_in_simdgroup]])
{
    const uint row = tgid * 8u + sgid;
    if (row >= n_rows) return;

    const uint blocks_per_row = k >> 5u;  // k / 32
    float local_sum = 0.0f;

    for (uint b = lane; b < blocks_per_row; b += 32u) {
        const uint block_idx = row * blocks_per_row + b;
        const uint base_byte = block_idx * 34u;

        // Scale at bytes 32-33 (after the 32 FP8 weight bytes).
        const ushort scale_bits =
              ushort(blocks_raw[base_byte + 32u])
            | (ushort(blocks_raw[base_byte + 33u]) << 8u);
        const float scale = float(as_type<half>(scale_bits));

        // Dot product: 32 FP8 weights at bytes 0..31 with 32 float inputs.
        const uint inp_base = b * 32u;
        float block_sum = 0.0f;
        for (uint w = 0u; w < 32u; ++w) {
            block_sum += fp8_e4m3_to_float(blocks_raw[base_byte + w]) * input[inp_base + w];
        }
        local_sum += scale * block_sum;
    }

    // Sum across all 32 lanes within the simdgroup.
    float row_sum = simd_sum(local_sum);
    if (lane == 0u) {
        output[row] = row_sum;
    }
}
"#;

/// Metal MSL kernel: FP8 E5M2 GEMV.
///
/// Format: s[7] exp[6:2] man[1:0], bias=15
/// - Normal: `(-1)^s * 2^(exp - 15) * (1 + man/4)`
/// - Denorm: `(-1)^s * 2^(-14) * (man/4)`
/// - Inf/NaN (exp=31) → 0 for inference
///
/// Identical structure to `gemv_fp8_e4m3`, decode path differs only.
#[cfg(all(feature = "metal", target_os = "macos"))]
pub const MSL_GEMV_FP8_E5M2_V1: &str = r#"
#include <metal_stdlib>
using namespace metal;

// FP8 E5M2 decode helper (bias=15, with infinity).
static inline float fp8_e5m2_to_float(uchar b) {
    const uint exp  = (uint(b) >> 2u) & 31u;  // 5-bit exponent
    const uint mant = uint(b) & 3u;            // 2-bit mantissa
    // Inf / NaN: exp = 31 → 0 for inference.
    if (exp == 31u) return 0.0f;
    const uint sign = (uint(b) >> 7u) & 1u;
    float val;
    if (exp == 0u) {
        // Denormal: 2^(-14) * (mant / 4)
        val = float(mant) * (1.0f / 4.0f) * (1.0f / 16384.0f);
    } else {
        // Normal: 2^(exp - 15) * (1 + mant/4)
        // Assemble as IEEE-754 f32: ((exp - 15 + 127) << 23) | (mant << 21)
        const uint bits = ((exp - 15u + 127u) << 23u) | (mant << 21u);
        val = as_type<float>(bits);
    }
    return sign ? -val : val;
}

kernel void gemv_fp8_e5m2(
    device const uchar* blocks_raw [[buffer(0)]],
    device const float* input      [[buffer(1)]],
    device float* output           [[buffer(2)]],
    constant uint& n_rows          [[buffer(3)]],
    constant uint& k               [[buffer(4)]],
    uint tgid  [[threadgroup_position_in_grid]],
    uint sgid  [[simdgroup_index_in_threadgroup]],
    uint lane  [[thread_index_in_simdgroup]])
{
    const uint row = tgid * 8u + sgid;
    if (row >= n_rows) return;

    const uint blocks_per_row = k >> 5u;  // k / 32
    float local_sum = 0.0f;

    for (uint b = lane; b < blocks_per_row; b += 32u) {
        const uint block_idx = row * blocks_per_row + b;
        const uint base_byte = block_idx * 34u;

        const ushort scale_bits =
              ushort(blocks_raw[base_byte + 32u])
            | (ushort(blocks_raw[base_byte + 33u]) << 8u);
        const float scale = float(as_type<half>(scale_bits));

        const uint inp_base = b * 32u;
        float block_sum = 0.0f;
        for (uint w = 0u; w < 32u; ++w) {
            block_sum += fp8_e5m2_to_float(blocks_raw[base_byte + w]) * input[inp_base + w];
        }
        local_sum += scale * block_sum;
    }

    float row_sum = simd_sum(local_sum);
    if (lane == 0u) {
        output[row] = row_sum;
    }
}
"#;

// ═══════════════════════════════════════════════════════════════════════════
// Tests — host-only kernel source string assertions
// ═══════════════════════════════════════════════════════════════════════════

#[cfg(test)]
mod tests {
    #[test]
    #[cfg(all(feature = "metal", target_os = "macos"))]
    fn fp8_kernels_contain_entry_points() {
        use super::*;
        assert!(MSL_GEMV_FP8_E4M3_V1.contains("kernel void gemv_fp8_e4m3"));
        assert!(MSL_GEMV_FP8_E5M2_V1.contains("kernel void gemv_fp8_e5m2"));
        // Decode helpers
        assert!(MSL_GEMV_FP8_E4M3_V1.contains("fp8_e4m3_to_float"));
        assert!(MSL_GEMV_FP8_E5M2_V1.contains("fp8_e5m2_to_float"));
        // Standard buffer-index annotations
        assert!(MSL_GEMV_FP8_E4M3_V1.contains("[[buffer(0)]]"));
        assert!(MSL_GEMV_FP8_E4M3_V1.contains("[[buffer(4)]]"));
        // Block size constant
        assert!(MSL_GEMV_FP8_E4M3_V1.contains("* 34u"));
        assert!(MSL_GEMV_FP8_E5M2_V1.contains("* 34u"));
    }
}