trueno-gpu 0.4.29

Pure Rust PTX generation for NVIDIA CUDA - no LLVM, no nvcc
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

//! Batched Element-wise Kernels
//!
//! GPU kernels for batched element-wise operations in multi-head attention.
//!
//! - [`BatchedScaleKernel`]: Scale all elements by a scalar
//! - [`BatchedSoftmaxKernel`]: Row-wise softmax for attention

#![allow(clippy::similar_names)]

use crate::kernels::Kernel;
use crate::ptx::builder::{PtxArithmetic, PtxComparison, PtxControl};
use crate::ptx::{PtxKernel, PtxReg, PtxType};

/// Batched scale kernel: multiply all elements by a scalar
#[derive(Debug, Clone)]
pub struct BatchedScaleKernel {
    /// Total number of elements (batch * rows * cols)
    pub n: u32,
}

impl BatchedScaleKernel {
    /// Create a new batched scale kernel
    #[must_use]
    pub const fn new(n: u32) -> Self {
        Self { n }
    }
}

impl Kernel for BatchedScaleKernel {
    fn name(&self) -> &str {
        "batched_scale"
    }

    fn build_ptx(&self) -> PtxKernel {
        let total = self.n;

        PtxKernel::new("batched_scale")
            .param(PtxType::U64, "input_ptr")
            .param(PtxType::U64, "output_ptr")
            .param(PtxType::F32, "scale")
            .param(PtxType::U32, "n")
            .build(move |ctx| {
                let tid = ctx.special_reg(PtxReg::TidX);
                let ctaid = ctx.special_reg(PtxReg::CtaIdX);
                let ntid = ctx.special_reg(PtxReg::NtidX);
                let gid = ctx.mad_lo_u32(ctaid, ntid, tid);

                let total_reg = ctx.mov_u32_imm(total);
                let in_bounds = ctx.setp_lt_u32(gid, total_reg);
                ctx.branch_if_not(in_bounds, "exit");

                let input_ptr = ctx.load_param_u64("input_ptr");
                let output_ptr = ctx.load_param_u64("output_ptr");
                let scale = ctx.load_param_f32("scale");

                let four = ctx.mov_u32_imm(4);
                let offset = ctx.mul_wide_u32_reg(gid, four);
                let in_addr = ctx.add_u64(input_ptr, offset);
                let out_addr = ctx.add_u64(output_ptr, offset);

                let val = ctx.ld_global_f32(in_addr);
                let scaled = ctx.mul_f32(val, scale);
                ctx.st_global_f32(out_addr, scaled);

                ctx.label("exit");
                ctx.ret();
            })
    }
}

/// Batched softmax kernel: softmax for multiple independent rows
///
/// Uses warp shuffle for reduction. One warp per row.
#[derive(Debug, Clone)]
pub struct BatchedSoftmaxKernel {
    /// Total number of rows to process (batch * n_rows)
    pub total_rows: u32,
    /// Size of each row
    pub row_size: u32,
}

impl BatchedSoftmaxKernel {
    /// Create a new batched softmax kernel
    #[must_use]
    pub const fn new(total_rows: u32, row_size: u32) -> Self {
        Self { total_rows, row_size }
    }
}

impl Kernel for BatchedSoftmaxKernel {
    fn name(&self) -> &str {
        "batched_softmax"
    }

    fn build_ptx(&self) -> PtxKernel {
        let total_rows = self.total_rows;
        let row_size = self.row_size;

        PtxKernel::new("batched_softmax")
            .param(PtxType::U64, "input_ptr")
            .param(PtxType::U64, "output_ptr")
            .param(PtxType::U32, "total_rows")
            .param(PtxType::U32, "row_size")
            .shared_memory(72)
            .build(move |ctx| {
                let row_idx = ctx.special_reg(PtxReg::CtaIdX);
                let tid = ctx.special_reg(PtxReg::TidX);

                let total_rows_reg = ctx.mov_u32_imm(total_rows);
                let valid = ctx.setp_lt_u32(row_idx, total_rows_reg);
                ctx.branch_if_not(valid, "exit");

                let input_ptr = ctx.load_param_u64("input_ptr");
                let output_ptr = ctx.load_param_u64("output_ptr");
                let row_size_reg = ctx.mov_u32_imm(row_size);

                let row_offset = ctx.mul_wide_u32(row_idx, row_size * 4);
                let row_input_ptr = ctx.add_u64(input_ptr, row_offset);
                let row_output_ptr = ctx.add_u64(output_ptr, row_offset);

                let four = ctx.mov_u32_imm(4);
                let log2e = ctx.mov_f32_imm(std::f32::consts::LOG2_E);

                // Pass 1: Find max
                let local_max = ctx.mov_f32_imm(f32::NEG_INFINITY);
                let i_max = ctx.mov_u32_imm(0);
                ctx.add_u32_reg_inplace(i_max, tid);
                ctx.label("max_loop");
                let max_done = ctx.setp_ge_u32(i_max, row_size_reg);
                ctx.branch_if(max_done, "max_done");

                let offset = ctx.mul_wide_u32_reg(i_max, four);
                let addr = ctx.add_u64(row_input_ptr, offset);
                let val = ctx.ld_global_f32(addr);
                ctx.max_f32_inplace(local_max, val);
                ctx.add_u32_inplace(i_max, 32);
                ctx.branch("max_loop");

                ctx.label("max_done");

                // Warp reduce for max
                let tmp16 = ctx.shfl_down_f32(local_max, 16, 0xFFFF_FFFF);
                ctx.max_f32_inplace(local_max, tmp16);
                let tmp8 = ctx.shfl_down_f32(local_max, 8, 0xFFFF_FFFF);
                ctx.max_f32_inplace(local_max, tmp8);
                let tmp4 = ctx.shfl_down_f32(local_max, 4, 0xFFFF_FFFF);
                ctx.max_f32_inplace(local_max, tmp4);
                let tmp2 = ctx.shfl_down_f32(local_max, 2, 0xFFFF_FFFF);
                ctx.max_f32_inplace(local_max, tmp2);
                let tmp1 = ctx.shfl_down_f32(local_max, 1, 0xFFFF_FFFF);
                ctx.max_f32_inplace(local_max, tmp1);

                let row_max = ctx.shfl_idx_f32(local_max, 0, 0xFFFF_FFFF);

                // Pass 2: Sum of exp(x - max)
                let local_sum = ctx.mov_f32_imm(0.0);
                let i_sum = ctx.mov_u32_imm(0);
                ctx.add_u32_reg_inplace(i_sum, tid);
                ctx.label("sum_loop");
                let sum_done = ctx.setp_ge_u32(i_sum, row_size_reg);
                ctx.branch_if(sum_done, "sum_done");

                let offset = ctx.mul_wide_u32_reg(i_sum, four);
                let addr = ctx.add_u64(row_input_ptr, offset);
                let val = ctx.ld_global_f32(addr);
                let diff = ctx.sub_f32(val, row_max);
                let exp_arg = ctx.mul_f32(diff, log2e);
                let exp_val = ctx.ex2_f32(exp_arg);
                ctx.add_f32_inplace(local_sum, exp_val);
                ctx.add_u32_inplace(i_sum, 32);
                ctx.branch("sum_loop");

                ctx.label("sum_done");

                // Warp reduce for sum
                let stmp16 = ctx.shfl_down_f32(local_sum, 16, 0xFFFF_FFFF);
                ctx.add_f32_inplace(local_sum, stmp16);
                let stmp8 = ctx.shfl_down_f32(local_sum, 8, 0xFFFF_FFFF);
                ctx.add_f32_inplace(local_sum, stmp8);
                let stmp4 = ctx.shfl_down_f32(local_sum, 4, 0xFFFF_FFFF);
                ctx.add_f32_inplace(local_sum, stmp4);
                let stmp2 = ctx.shfl_down_f32(local_sum, 2, 0xFFFF_FFFF);
                ctx.add_f32_inplace(local_sum, stmp2);
                let stmp1 = ctx.shfl_down_f32(local_sum, 1, 0xFFFF_FFFF);
                ctx.add_f32_inplace(local_sum, stmp1);

                let row_sum = ctx.shfl_idx_f32(local_sum, 0, 0xFFFF_FFFF);

                // Pass 3: Write normalized values
                let i_write = ctx.mov_u32_imm(0);
                ctx.add_u32_reg_inplace(i_write, tid);
                ctx.label("write_loop");
                let write_done = ctx.setp_ge_u32(i_write, row_size_reg);
                ctx.branch_if(write_done, "exit");

                let offset = ctx.mul_wide_u32_reg(i_write, four);
                let in_addr = ctx.add_u64(row_input_ptr, offset);
                let out_addr = ctx.add_u64(row_output_ptr, offset);

                let val = ctx.ld_global_f32(in_addr);
                let diff = ctx.sub_f32(val, row_max);
                let exp_arg = ctx.mul_f32(diff, log2e);
                let exp_val = ctx.ex2_f32(exp_arg);
                let normalized = ctx.div_f32(exp_val, row_sum);
                ctx.st_global_f32(out_addr, normalized);

                ctx.add_u32_inplace(i_write, 32);
                ctx.branch("write_loop");

                ctx.label("exit");
                ctx.ret();
            })
    }
}