oxicuda-dnn 0.2.0

OxiCUDA DNN - GPU-accelerated deep learning primitives (cuDNN equivalent)
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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228

//! Shared PTX helper functions for depthwise separable convolution kernels.

use oxicuda_ptx::ir::PtxType;

use crate::error::{DnnError, DnnResult};

use super::types::ActivationType;

/// Parses precision string to PTX type and element size.
pub(super) fn parse_precision(precision: &str) -> DnnResult<(PtxType, u32)> {
    match precision {
        "f32" => Ok((PtxType::F32, 4)),
        "f64" => Ok((PtxType::F64, 8)),
        other => Err(DnnError::InvalidArgument(format!(
            "unsupported precision: {other}"
        ))),
    }
}

/// Emits inline activation PTX. Returns the register holding the result.
pub(super) fn emit_activation(
    b: &mut oxicuda_ptx::builder::BodyBuilder<'_>,
    float_type: PtxType,
    val: oxicuda_ptx::ir::Register,
    activation: ActivationType,
) -> oxicuda_ptx::ir::Register {
    match activation {
        ActivationType::None => val,
        ActivationType::Relu => {
            // max(0, val)
            let zero = alloc_zero(b, float_type);
            let result = alloc_float(b, float_type);
            if float_type == PtxType::F32 {
                b.raw_ptx(&format!("max.f32 {result}, {val}, {zero};"));
            } else {
                b.raw_ptx(&format!("max.f64 {result}, {val}, {zero};"));
            }
            result
        }
        ActivationType::Relu6 => {
            // min(6, max(0, val))
            let zero = alloc_zero(b, float_type);
            let six = alloc_float_const(b, float_type, 6.0);
            let clamped_lo = alloc_float(b, float_type);
            let result = alloc_float(b, float_type);
            if float_type == PtxType::F32 {
                b.raw_ptx(&format!("max.f32 {clamped_lo}, {val}, {zero};"));
                b.raw_ptx(&format!("min.f32 {result}, {clamped_lo}, {six};"));
            } else {
                b.raw_ptx(&format!("max.f64 {clamped_lo}, {val}, {zero};"));
                b.raw_ptx(&format!("min.f64 {result}, {clamped_lo}, {six};"));
            }
            result
        }
        ActivationType::Silu => {
            // x * sigmoid(x) = x / (1 + exp(-x))
            let neg_x = alloc_float(b, float_type);
            let exp_neg = alloc_float(b, float_type);
            let one = alloc_float_const(b, float_type, 1.0);
            let denom = alloc_float(b, float_type);
            let result = alloc_float(b, float_type);
            if float_type == PtxType::F32 {
                b.raw_ptx(&format!("neg.f32 {neg_x}, {val};"));
                b.raw_ptx(&format!("ex2.approx.f32 {exp_neg}, {neg_x};"));
                b.raw_ptx(&format!("add.rn.f32 {denom}, {exp_neg}, {one};"));
                b.raw_ptx(&format!("div.rn.f32 {result}, {val}, {denom};"));
            } else {
                // f64: use neg + approximate approach
                b.raw_ptx(&format!("neg.f64 {neg_x}, {val};"));
                // Convert to f32 for exp approx, then back
                let tmp32 = b.alloc_reg(PtxType::F32);
                let exp32 = b.alloc_reg(PtxType::F32);
                b.raw_ptx(&format!("cvt.rn.f32.f64 {tmp32}, {neg_x};"));
                b.raw_ptx(&format!("ex2.approx.f32 {exp32}, {tmp32};"));
                b.raw_ptx(&format!("cvt.rn.f64.f32 {exp_neg}, {exp32};"));
                b.raw_ptx(&format!("add.rn.f64 {denom}, {exp_neg}, {one};"));
                b.raw_ptx(&format!("div.rn.f64 {result}, {val}, {denom};"));
            }
            result
        }
        ActivationType::HardSwish => {
            // x * min(6, max(0, x + 3)) / 6
            let three = alloc_float_const(b, float_type, 3.0);
            let six = alloc_float_const(b, float_type, 6.0);
            let zero = alloc_zero(b, float_type);
            let x_plus_3 = alloc_float(b, float_type);
            let clamped_lo = alloc_float(b, float_type);
            let clamped = alloc_float(b, float_type);
            let numerator = alloc_float(b, float_type);
            let result = alloc_float(b, float_type);
            if float_type == PtxType::F32 {
                b.raw_ptx(&format!("add.rn.f32 {x_plus_3}, {val}, {three};"));
                b.raw_ptx(&format!("max.f32 {clamped_lo}, {x_plus_3}, {zero};"));
                b.raw_ptx(&format!("min.f32 {clamped}, {clamped_lo}, {six};"));
                b.raw_ptx(&format!("mul.rn.f32 {numerator}, {val}, {clamped};"));
                b.raw_ptx(&format!("div.rn.f32 {result}, {numerator}, {six};"));
            } else {
                b.raw_ptx(&format!("add.rn.f64 {x_plus_3}, {val}, {three};"));
                b.raw_ptx(&format!("max.f64 {clamped_lo}, {x_plus_3}, {zero};"));
                b.raw_ptx(&format!("min.f64 {clamped}, {clamped_lo}, {six};"));
                b.raw_ptx(&format!("mul.rn.f64 {numerator}, {val}, {clamped};"));
                b.raw_ptx(&format!("div.rn.f64 {result}, {numerator}, {six};"));
            }
            result
        }
    }
}

/// Emits batch normalisation epilogue:
/// `val = gamma * (val - mean) / sqrt(var + eps) + beta`
///
/// Reads BN parameters from `bn_gamma`, `bn_beta`, `bn_mean`, `bn_var`
/// kernel params, indexed by `channel_reg`.
pub(super) fn emit_bn_epilogue(
    b: &mut oxicuda_ptx::builder::BodyBuilder<'_>,
    float_type: PtxType,
    elem_bytes: u32,
    val: &oxicuda_ptx::ir::Register,
    channel_reg: oxicuda_ptx::ir::Register,
) {
    b.comment("Batch Normalization epilogue");
    let gamma_ptr = b.load_param_u64("bn_gamma");
    let beta_ptr = b.load_param_u64("bn_beta");
    let mean_ptr = b.load_param_u64("bn_mean");
    let var_ptr = b.load_param_u64("bn_var");

    // Compute address offset for channel
    let ch64 = b.alloc_reg(PtxType::U64);
    let ch_off = b.alloc_reg(PtxType::U64);
    b.raw_ptx(&format!("cvt.u64.u32 {ch64}, {channel_reg};"));
    b.raw_ptx(&format!("mul.lo.u64 {ch_off}, {ch64}, {elem_bytes};"));

    // Load gamma, beta, mean, var for this channel
    let addr = b.alloc_reg(PtxType::U64);
    let gamma = b.alloc_reg(float_type);
    let beta = b.alloc_reg(float_type);
    let mean = b.alloc_reg(float_type);
    let var = b.alloc_reg(float_type);

    let ld_suffix = if float_type == PtxType::F32 {
        "f32"
    } else {
        "f64"
    };

    b.raw_ptx(&format!("add.u64 {addr}, {gamma_ptr}, {ch_off};"));
    b.raw_ptx(&format!("ld.global.{ld_suffix} {gamma}, [{addr}];"));

    b.raw_ptx(&format!("add.u64 {addr}, {beta_ptr}, {ch_off};"));
    b.raw_ptx(&format!("ld.global.{ld_suffix} {beta}, [{addr}];"));

    b.raw_ptx(&format!("add.u64 {addr}, {mean_ptr}, {ch_off};"));
    b.raw_ptx(&format!("ld.global.{ld_suffix} {mean}, [{addr}];"));

    b.raw_ptx(&format!("add.u64 {addr}, {var_ptr}, {ch_off};"));
    b.raw_ptx(&format!("ld.global.{ld_suffix} {var}, [{addr}];"));

    // eps = 1e-5
    let eps = alloc_float_const(b, float_type, 1e-5);

    // var_eps = var + eps
    let var_eps = alloc_float(b, float_type);
    // inv_std = rsqrt(var_eps)
    let inv_std = alloc_float(b, float_type);
    // x_minus_mean = val - mean
    let x_minus_mean = alloc_float(b, float_type);
    // normed = x_minus_mean * inv_std
    let normed = alloc_float(b, float_type);

    if float_type == PtxType::F32 {
        b.raw_ptx(&format!("add.rn.f32 {var_eps}, {var}, {eps};"));
        b.raw_ptx(&format!("rsqrt.approx.f32 {inv_std}, {var_eps};"));
        b.raw_ptx(&format!("sub.rn.f32 {x_minus_mean}, {val}, {mean};"));
        b.raw_ptx(&format!("mul.rn.f32 {normed}, {x_minus_mean}, {inv_std};"));
        // val = gamma * normed + beta
        b.raw_ptx(&format!("fma.rn.f32 {val}, {gamma}, {normed}, {beta};"));
    } else {
        b.raw_ptx(&format!("add.rn.f64 {var_eps}, {var}, {eps};"));
        // f64 doesn't have rsqrt, use rcp + sqrt
        let sqrt_val = alloc_float(b, float_type);
        b.raw_ptx(&format!("sqrt.rn.f64 {sqrt_val}, {var_eps};"));
        b.raw_ptx(&format!("rcp.rn.f64 {inv_std}, {sqrt_val};"));
        b.raw_ptx(&format!("sub.rn.f64 {x_minus_mean}, {val}, {mean};"));
        b.raw_ptx(&format!("mul.rn.f64 {normed}, {x_minus_mean}, {inv_std};"));
        b.raw_ptx(&format!("fma.rn.f64 {val}, {gamma}, {normed}, {beta};"));
    }
}

/// Allocates a zero-initialised float register.
pub(super) fn alloc_zero(
    b: &mut oxicuda_ptx::builder::BodyBuilder<'_>,
    float_type: PtxType,
) -> oxicuda_ptx::ir::Register {
    let r = b.alloc_reg(float_type);
    if float_type == PtxType::F32 {
        let bits = 0f32.to_bits();
        b.raw_ptx(&format!("mov.b32 {r}, 0F{bits:08X};"));
    } else {
        let bits = 0f64.to_bits();
        b.raw_ptx(&format!("mov.b64 {r}, 0D{bits:016X};"));
    }
    r
}

/// Allocates a float register initialised to a compile-time constant.
pub(super) fn alloc_float_const(
    b: &mut oxicuda_ptx::builder::BodyBuilder<'_>,
    float_type: PtxType,
    val: f64,
) -> oxicuda_ptx::ir::Register {
    let r = b.alloc_reg(float_type);
    if float_type == PtxType::F32 {
        let bits = (val as f32).to_bits();
        b.raw_ptx(&format!("mov.b32 {r}, 0F{bits:08X};"));
    } else {
        let bits = val.to_bits();
        b.raw_ptx(&format!("mov.b64 {r}, 0D{bits:016X};"));
    }
    r
}

/// Allocates an uninitialised float register.
pub(super) fn alloc_float(
    b: &mut oxicuda_ptx::builder::BodyBuilder<'_>,
    float_type: PtxType,
) -> oxicuda_ptx::ir::Register {
    b.alloc_reg(float_type)
}