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

//! Fused GEMM + Bias + GELU kernel (WAPR-PERF-007)

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

use super::super::Kernel;

/// Fused GEMM + Bias + GELU kernel (WAPR-PERF-007)
///
/// Computes: output = GELU(A @ B + bias) in a single kernel launch.
/// Eliminates 3 kernel launches (GEMM, Bias, GELU) into 1.
///
/// For FFN in Whisper encoder:
/// - First linear: [seq, hidden] @ [hidden, intermediate] + bias -> GELU -> [seq, intermediate]
/// - This kernel handles that in ONE launch instead of THREE.
///
/// Grid: (N / block_size, M / block_size, 1)
/// Block: (block_size, block_size, 1) - typically 16x16
///
/// # Citations
///
/// - [Dao2022] FlashAttention: Fast and Memory-Efficient Exact Attention
/// - [Kwon2023] PagedAttention: Efficient LLM Serving
#[derive(Debug, Clone)]
pub struct FusedGemmBiasGeluKernel {
    /// M: Number of rows in A and C
    pub m: u32,
    /// N: Number of columns in B and C
    pub n: u32,
    /// K: Shared dimension (columns of A, rows of B)
    pub k: u32,
}

impl FusedGemmBiasGeluKernel {
    /// Create a new fused GEMM+Bias+GELU kernel
    #[must_use]
    pub fn new(m: u32, n: u32, k: u32) -> Self {
        Self { m, n, k }
    }
}

impl Kernel for FusedGemmBiasGeluKernel {
    fn name(&self) -> &str {
        "fused_gemm_bias_gelu"
    }

    fn build_ptx(&self) -> PtxKernel {
        let k_val = self.k;
        let n_val = self.n;

        PtxKernel::new("fused_gemm_bias_gelu")
            .param(PtxType::U64, "a_ptr") // Input matrix A [M, K]
            .param(PtxType::U64, "b_ptr") // Weight matrix B [K, N]
            .param(PtxType::U64, "bias_ptr") // Bias vector [N]
            .param(PtxType::U64, "c_ptr") // Output matrix C [M, N]
            .param(PtxType::U32, "m")
            .param(PtxType::U32, "n")
            .param(PtxType::U32, "k")
            .build(move |ctx| {
                // Calculate row and column from thread/block IDs
                // row = ctaid.y * ntid.y + tid.y
                // col = ctaid.x * ntid.x + tid.x
                let ctaid_y = ctx.special_reg(PtxReg::CtaIdY);
                let ntid_y = ctx.special_reg(PtxReg::NtidY);
                let tid_y = ctx.special_reg(PtxReg::TidY);
                let ctaid_x = ctx.special_reg(PtxReg::CtaIdX);
                let ntid_x = ctx.special_reg(PtxReg::NtidX);
                let tid_x = ctx.special_reg(PtxReg::TidX);

                let row = ctx.mad_lo_u32(ctaid_y, ntid_y, tid_y);
                let col = ctx.mad_lo_u32(ctaid_x, ntid_x, tid_x);

                // Load params
                let m_param = ctx.load_param_u32("m");
                let n_param = ctx.load_param_u32("n");
                let k_param = ctx.load_param_u32("k");

                // Bounds check
                let pred_m = ctx.setp_ge_u32(row, m_param);
                ctx.branch_if(pred_m, "exit");
                let pred_n = ctx.setp_ge_u32(col, n_param);
                ctx.branch_if(pred_n, "exit");

                // Load base pointers
                let a_ptr = ctx.load_param_u64("a_ptr");
                let b_ptr = ctx.load_param_u64("b_ptr");
                let bias_ptr = ctx.load_param_u64("bias_ptr");
                let c_ptr = ctx.load_param_u64("c_ptr");

                // Initialize accumulator
                let acc = ctx.mov_f32_imm(0.0);

                // Calculate base offset for A[row, 0]
                let row_offset = ctx.mul_wide_u32(row, k_val * 4);
                let a_row_ptr = ctx.add_u64(a_ptr, row_offset);

                // Calculate base offset for B[0, col]
                let col_offset = ctx.mul_wide_u32(col, 4);
                let b_col_base = ctx.add_u64(b_ptr, col_offset);

                // Loop over K dimension
                let i = ctx.mov_u32_imm(0);

                ctx.label("loop_k");
                let pred_k = ctx.setp_ge_u32(i, k_param);
                ctx.branch_if(pred_k, "loop_end");

                // Load A[row, i]
                let i_offset = ctx.mul_wide_u32(i, 4);
                let a_addr = ctx.add_u64(a_row_ptr, i_offset);
                let a_val = ctx.ld_global_f32(a_addr);

                // Load B[i, col]
                let b_row_offset = ctx.mul_wide_u32(i, n_val * 4);
                let b_addr = ctx.add_u64(b_col_base, b_row_offset);
                let b_val = ctx.ld_global_f32(b_addr);

                // acc += a_val * b_val
                ctx.fma_f32_inplace(acc, a_val, b_val);

                // i++
                ctx.add_u32_inplace(i, 1);
                ctx.branch("loop_k");

                ctx.label("loop_end");

                // Load and add bias[col]
                let bias_offset = ctx.mul_wide_u32(col, 4);
                let bias_addr = ctx.add_u64(bias_ptr, bias_offset);
                let bias_val = ctx.ld_global_f32(bias_addr);
                let acc_biased = ctx.add_f32(acc, bias_val);

                // ============================================
                // GELU approximation (fused in same kernel)
                // gelu(x) = 0.5 * x * (1 + tanh(sqrt(2/π) * (x + 0.044715 * x³)))
                // ============================================
                let x = acc_biased;

                // Constants
                let sqrt_2_pi = ctx.mov_f32_imm(0.797_884_6); // sqrt(2/π)
                let c = ctx.mov_f32_imm(0.044_715);
                let half = ctx.mov_f32_imm(0.5);
                let one = ctx.mov_f32_imm(1.0);
                let two = ctx.mov_f32_imm(2.0);
                let zero = ctx.mov_f32_imm(0.0);
                let log2_e = ctx.mov_f32_imm(std::f32::consts::LOG2_E);

                // x³
                let x2 = ctx.mul_f32(x, x);
                let x3 = ctx.mul_f32(x2, x);

                // 0.044715 * x³
                let cx3 = ctx.mul_f32(c, x3);

                // x + 0.044715 * x³
                let inner = ctx.add_f32(x, cx3);

                // sqrt(2/π) * (x + 0.044715 * x³)
                let scaled = ctx.mul_f32(sqrt_2_pi, inner);

                // tanh approximation: tanh(x) = 2*sigmoid(2x) - 1
                let two_x = ctx.mul_f32(two, scaled);
                let neg_two_x = ctx.sub_f32(zero, two_x);
                let scaled_exp = ctx.mul_f32(neg_two_x, log2_e);
                let exp_neg = ctx.ex2_f32(scaled_exp);
                let denom = ctx.add_f32(one, exp_neg);
                let sigmoid = ctx.div_f32(one, denom);
                let two_sigmoid = ctx.mul_f32(two, sigmoid);
                let tanh = ctx.sub_f32(two_sigmoid, one);

                // 1 + tanh(...)
                let one_plus_tanh = ctx.add_f32(one, tanh);

                // 0.5 * x
                let half_x = ctx.mul_f32(half, x);

                // result = 0.5 * x * (1 + tanh(...))
                let result = ctx.mul_f32(half_x, one_plus_tanh);

                // Store result: C[row, col]
                let c_row_offset = ctx.mul_wide_u32(row, n_val * 4);
                let c_row_ptr = ctx.add_u64(c_ptr, c_row_offset);
                let c_col_offset = ctx.mul_wide_u32(col, 4);
                let c_addr = ctx.add_u64(c_row_ptr, c_col_offset);
                ctx.st_global_f32(c_addr, result);

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