trueno-gpu 0.4.33

//! Fused Transformer Kernels (PMAT-PERF-009)
//!
//! Implements fused operations for transformer inference to reduce kernel launch overhead:
//! - FusedQKVKernel: Q/K/V projection in single kernel (3x reduction in launches)
//! - FusedGateUpKernel: Gate+Up FFN with SwiGLU activation (2x reduction)
//!
//! # Five-Whys Root Cause (PMAT-PERF-009)
//!
//! ```text
//! Why 1: Why is decode throughput 131 tok/s vs 400 tok/s target?
//! → 280+ kernel launches per token (10+ per layer × 28 layers)
//!
//! Why 2: Why so many kernel launches?
//! → Q, K, V computed as 3 separate GEMV operations
//!
//! Why 3: Why separate operations?
//! → Original implementation didn't consider launch overhead
//!
//! Why 4: Why does launch overhead matter?
//! → GPU kernel launch: ~5-10µs, 280 launches = 1.4-2.8ms overhead/token
//!
//! Why 5: ROOT CAUSE
//! → Kernel launch overhead (2.8ms) exceeds compute time for small batch decode
//! → FIX: Fuse Q/K/V into single kernel, reducing launches by 2/3
//! ```

// Allow similar names for related variables (wq/wk/wv, shfl_q/shfl_k/shfl_v, etc.)
// Allow unused_assignments/unused_mut because PTX branch semantics aren't tracked by Rust
#![allow(clippy::similar_names, unused_assignments, unused_mut)]

mod gemm_bias_gelu;

pub use gemm_bias_gelu::FusedGemmBiasGeluKernel;

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

use super::Kernel;

/// Fused Q/K/V projection kernel
///
/// Computes Q, K, V projections in a single kernel launch:
/// - Q = x @ W_q^T (hidden_size → hidden_size)
/// - K = x @ W_k^T (hidden_size → kv_dim)
/// - V = x @ W_v^T (hidden_size → kv_dim)
///
/// Grid: (max(hidden_size, kv_dim), 1, 1)
/// Block: (32, 1, 1) - one warp per output element
#[derive(Debug, Clone)]
pub struct FusedQKVKernel {
    /// Hidden dimension size
    pub hidden_size: usize,
    /// KV dimension (may differ for GQA)
    pub kv_dim: usize,
}

impl FusedQKVKernel {
    /// Create a new fused QKV kernel.
    pub fn new(hidden_size: usize, kv_dim: usize) -> Self {
        Self { hidden_size, kv_dim }
    }
}

impl Kernel for FusedQKVKernel {
    fn name(&self) -> &str {
        "fused_qkv_gemv"
    }

    fn build_ptx(&self) -> PtxKernel {
        let hidden = self.hidden_size as u32;
        let kv = self.kv_dim as u32;

        PtxKernel::new("fused_qkv_gemv")
            // Parameters: x, W_q, W_k, W_v, out_q, out_k, out_v
            .param(PtxType::U64, "x_ptr")
            .param(PtxType::U64, "wq_ptr")
            .param(PtxType::U64, "wk_ptr")
            .param(PtxType::U64, "wv_ptr")
            .param(PtxType::U64, "out_q_ptr")
            .param(PtxType::U64, "out_k_ptr")
            .param(PtxType::U64, "out_v_ptr")
            .build(move |ctx| {
                // Get thread/block IDs
                let tid = ctx.special_reg(PtxReg::TidX);
                let row = ctx.special_reg(PtxReg::CtaIdX);

                // lane = tid & 31
                let lane = ctx.and_u32_imm(tid, 31);

                // Load constants
                let hidden_size = ctx.mov_u32_imm(hidden);
                let kv_dim_val = ctx.mov_u32_imm(kv);

                // Initialize accumulators
                let mut acc_q = ctx.mov_f32_imm(0.0);
                let mut acc_k = ctx.mov_f32_imm(0.0);
                let mut acc_v = ctx.mov_f32_imm(0.0);

                // Load base pointers
                let x_ptr = ctx.load_param_u64("x_ptr");
                let wq_ptr = ctx.load_param_u64("wq_ptr");
                let wk_ptr = ctx.load_param_u64("wk_ptr");
                let wv_ptr = ctx.load_param_u64("wv_ptr");

                // k = lane (start offset)
                let mut k = lane;

                // Main loop: stride by 32 (warp size)
                ctx.label("loop_start");
                let pred_exit = ctx.setp_ge_u32(k, hidden_size);
                ctx.branch_if(pred_exit, "loop_end");

                // Load x[k]
                let offset_k = ctx.mul_wide_u32(k, 4);
                let x_addr = ctx.add_u64(x_ptr, offset_k);
                let x_val = ctx.ld_global_f32(x_addr);

                // Compute weight offset: row * hidden + k
                let row_offset = ctx.mul_u32_reg(row, hidden_size);
                let weight_idx = ctx.add_u32_reg(row_offset, k);
                let weight_byte_offset = ctx.mul_wide_u32(weight_idx, 4);

                // Load and accumulate Q
                let wq_addr = ctx.add_u64(wq_ptr, weight_byte_offset);
                let wq_val = ctx.ld_global_f32(wq_addr);
                acc_q = ctx.fma_f32(x_val, wq_val, acc_q);

                // Load and accumulate K
                let wk_addr = ctx.add_u64(wk_ptr, weight_byte_offset);
                let wk_val = ctx.ld_global_f32(wk_addr);
                acc_k = ctx.fma_f32(x_val, wk_val, acc_k);

                // Load and accumulate V
                let wv_addr = ctx.add_u64(wv_ptr, weight_byte_offset);
                let wv_val = ctx.ld_global_f32(wv_addr);
                acc_v = ctx.fma_f32(x_val, wv_val, acc_v);

                // k += 32 (must be in-place to update loop variable)
                ctx.add_u32_inplace(k, 32);
                ctx.branch("loop_start");

                ctx.label("loop_end");

                // Warp reduction for all accumulators
                // acc_q
                let shfl_q_16 = ctx.shfl_down_f32(acc_q, 16, 0xFFFF_FFFF);
                acc_q = ctx.add_f32(acc_q, shfl_q_16);
                let shfl_q_8 = ctx.shfl_down_f32(acc_q, 8, 0xFFFF_FFFF);
                acc_q = ctx.add_f32(acc_q, shfl_q_8);
                let shfl_q_4 = ctx.shfl_down_f32(acc_q, 4, 0xFFFF_FFFF);
                acc_q = ctx.add_f32(acc_q, shfl_q_4);
                let shfl_q_2 = ctx.shfl_down_f32(acc_q, 2, 0xFFFF_FFFF);
                acc_q = ctx.add_f32(acc_q, shfl_q_2);
                let shfl_q_1 = ctx.shfl_down_f32(acc_q, 1, 0xFFFF_FFFF);
                acc_q = ctx.add_f32(acc_q, shfl_q_1);

                // acc_k
                let shfl_k_16 = ctx.shfl_down_f32(acc_k, 16, 0xFFFF_FFFF);
                acc_k = ctx.add_f32(acc_k, shfl_k_16);
                let shfl_k_8 = ctx.shfl_down_f32(acc_k, 8, 0xFFFF_FFFF);
                acc_k = ctx.add_f32(acc_k, shfl_k_8);
                let shfl_k_4 = ctx.shfl_down_f32(acc_k, 4, 0xFFFF_FFFF);
                acc_k = ctx.add_f32(acc_k, shfl_k_4);
                let shfl_k_2 = ctx.shfl_down_f32(acc_k, 2, 0xFFFF_FFFF);
                acc_k = ctx.add_f32(acc_k, shfl_k_2);
                let shfl_k_1 = ctx.shfl_down_f32(acc_k, 1, 0xFFFF_FFFF);
                acc_k = ctx.add_f32(acc_k, shfl_k_1);

                // acc_v
                let shfl_v_16 = ctx.shfl_down_f32(acc_v, 16, 0xFFFF_FFFF);
                acc_v = ctx.add_f32(acc_v, shfl_v_16);
                let shfl_v_8 = ctx.shfl_down_f32(acc_v, 8, 0xFFFF_FFFF);
                acc_v = ctx.add_f32(acc_v, shfl_v_8);
                let shfl_v_4 = ctx.shfl_down_f32(acc_v, 4, 0xFFFF_FFFF);
                acc_v = ctx.add_f32(acc_v, shfl_v_4);
                let shfl_v_2 = ctx.shfl_down_f32(acc_v, 2, 0xFFFF_FFFF);
                acc_v = ctx.add_f32(acc_v, shfl_v_2);
                let shfl_v_1 = ctx.shfl_down_f32(acc_v, 1, 0xFFFF_FFFF);
                acc_v = ctx.add_f32(acc_v, shfl_v_1);

                // Lane 0 stores results (skip if lane != 0)
                let zero = ctx.mov_u32_imm(0);
                let is_lane0 = ctx.setp_eq_u32(lane, zero);
                ctx.branch_if_not(is_lane0, "done");

                // Store outputs
                let out_q_ptr = ctx.load_param_u64("out_q_ptr");
                let out_k_ptr = ctx.load_param_u64("out_k_ptr");
                let out_v_ptr = ctx.load_param_u64("out_v_ptr");

                let row_byte_offset = ctx.mul_wide_u32(row, 4);

                // Store Q (always for row < hidden_size, which is ensured by grid size)
                let out_q_addr = ctx.add_u64(out_q_ptr, row_byte_offset);
                ctx.st_global_f32(out_q_addr, acc_q);

                // Store K/V only if row < kv_dim
                let pred_kv = ctx.setp_lt_u32(row, kv_dim_val);
                ctx.branch_if_not(pred_kv, "done");

                let out_k_addr = ctx.add_u64(out_k_ptr, row_byte_offset);
                ctx.st_global_f32(out_k_addr, acc_k);

                let out_v_addr = ctx.add_u64(out_v_ptr, row_byte_offset);
                ctx.st_global_f32(out_v_addr, acc_v);

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

/// Fused Gate+Up FFN kernel with SwiGLU activation
///
/// Computes: output = SiLU(gate) * up
/// Where:
/// - gate = x @ W_gate^T
/// - up = x @ W_up^T
/// - SiLU(x) = x * sigmoid(x)
///
/// Grid: (intermediate_size, 1, 1)
/// Block: (32, 1, 1) - one warp per output element
#[derive(Debug, Clone)]
pub struct FusedGateUpKernel {
    /// Hidden dimension size
    pub hidden_size: usize,
    /// Intermediate FFN dimension
    pub intermediate_size: usize,
}

impl FusedGateUpKernel {
    /// Create a new fused gate+up kernel.
    pub fn new(hidden_size: usize, intermediate_size: usize) -> Self {
        Self { hidden_size, intermediate_size }
    }
}

impl Kernel for FusedGateUpKernel {
    fn name(&self) -> &str {
        "fused_gate_up_swiglu"
    }

    fn build_ptx(&self) -> PtxKernel {
        let hidden = self.hidden_size as u32;

        PtxKernel::new("fused_gate_up_swiglu")
            // Parameters: x, W_gate, W_up, output
            .param(PtxType::U64, "x_ptr")
            .param(PtxType::U64, "wg_ptr")
            .param(PtxType::U64, "wu_ptr")
            .param(PtxType::U64, "out_ptr")
            .build(move |ctx| {
                // Get thread/block IDs
                let tid = ctx.special_reg(PtxReg::TidX);
                let row = ctx.special_reg(PtxReg::CtaIdX);
                let lane = ctx.and_u32_imm(tid, 31);
                let hidden_size = ctx.mov_u32_imm(hidden);

                // Initialize accumulators
                let mut acc_gate = ctx.mov_f32_imm(0.0);
                let mut acc_up = ctx.mov_f32_imm(0.0);

                // Load base pointers
                let x_ptr = ctx.load_param_u64("x_ptr");
                let wg_ptr = ctx.load_param_u64("wg_ptr");
                let wu_ptr = ctx.load_param_u64("wu_ptr");

                // k = lane
                let mut k = lane;

                // Main loop
                ctx.label("loop_start");
                let pred_exit = ctx.setp_ge_u32(k, hidden_size);
                ctx.branch_if(pred_exit, "loop_end");

                // Load x[k]
                let offset_k = ctx.mul_wide_u32(k, 4);
                let x_addr = ctx.add_u64(x_ptr, offset_k);
                let x_val = ctx.ld_global_f32(x_addr);

                // Weight offset: row * hidden + k
                let row_offset = ctx.mul_u32_reg(row, hidden_size);
                let weight_idx = ctx.add_u32_reg(row_offset, k);
                let weight_byte_offset = ctx.mul_wide_u32(weight_idx, 4);

                // Load W_gate and accumulate
                let wg_addr = ctx.add_u64(wg_ptr, weight_byte_offset);
                let wg_val = ctx.ld_global_f32(wg_addr);
                acc_gate = ctx.fma_f32(x_val, wg_val, acc_gate);

                // Load W_up and accumulate
                let wu_addr = ctx.add_u64(wu_ptr, weight_byte_offset);
                let wu_val = ctx.ld_global_f32(wu_addr);
                acc_up = ctx.fma_f32(x_val, wu_val, acc_up);

                // k += 32 (must be in-place to update loop variable)
                ctx.add_u32_inplace(k, 32);
                ctx.branch("loop_start");

                ctx.label("loop_end");

                // Warp reduction for acc_gate
                let shfl_g_16 = ctx.shfl_down_f32(acc_gate, 16, 0xFFFF_FFFF);
                acc_gate = ctx.add_f32(acc_gate, shfl_g_16);
                let shfl_g_8 = ctx.shfl_down_f32(acc_gate, 8, 0xFFFF_FFFF);
                acc_gate = ctx.add_f32(acc_gate, shfl_g_8);
                let shfl_g_4 = ctx.shfl_down_f32(acc_gate, 4, 0xFFFF_FFFF);
                acc_gate = ctx.add_f32(acc_gate, shfl_g_4);
                let shfl_g_2 = ctx.shfl_down_f32(acc_gate, 2, 0xFFFF_FFFF);
                acc_gate = ctx.add_f32(acc_gate, shfl_g_2);
                let shfl_g_1 = ctx.shfl_down_f32(acc_gate, 1, 0xFFFF_FFFF);
                acc_gate = ctx.add_f32(acc_gate, shfl_g_1);

                // Warp reduction for acc_up
                let shfl_u_16 = ctx.shfl_down_f32(acc_up, 16, 0xFFFF_FFFF);
                acc_up = ctx.add_f32(acc_up, shfl_u_16);
                let shfl_u_8 = ctx.shfl_down_f32(acc_up, 8, 0xFFFF_FFFF);
                acc_up = ctx.add_f32(acc_up, shfl_u_8);
                let shfl_u_4 = ctx.shfl_down_f32(acc_up, 4, 0xFFFF_FFFF);
                acc_up = ctx.add_f32(acc_up, shfl_u_4);
                let shfl_u_2 = ctx.shfl_down_f32(acc_up, 2, 0xFFFF_FFFF);
                acc_up = ctx.add_f32(acc_up, shfl_u_2);
                let shfl_u_1 = ctx.shfl_down_f32(acc_up, 1, 0xFFFF_FFFF);
                acc_up = ctx.add_f32(acc_up, shfl_u_1);

                // Lane 0 computes SiLU and stores (skip if lane != 0)
                let zero = ctx.mov_u32_imm(0);
                let is_lane0 = ctx.setp_eq_u32(lane, zero);
                ctx.branch_if_not(is_lane0, "done");

                // SiLU(gate) = gate * sigmoid(gate) = gate / (1 + exp(-gate))
                // exp(-x) = 2^(-x * log2(e))
                let neg_gate = ctx.neg_f32(acc_gate);
                let log2_e = ctx.mov_f32_imm(std::f32::consts::LOG2_E);
                let scaled = ctx.mul_f32(neg_gate, log2_e);
                let exp_val = ctx.ex2_f32(scaled);
                let one = ctx.mov_f32_imm(1.0);
                let one_plus_exp = ctx.add_f32(one, exp_val);
                let sigmoid = ctx.rcp_f32(one_plus_exp);
                let silu = ctx.mul_f32(acc_gate, sigmoid);

                // output = SiLU(gate) * up
                let output = ctx.mul_f32(silu, acc_up);

                // Store output[row]
                let out_ptr = ctx.load_param_u64("out_ptr");
                let row_byte_offset = ctx.mul_wide_u32(row, 4);
                let out_addr = ctx.add_u64(out_ptr, row_byte_offset);
                ctx.st_global_f32(out_addr, output);

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

#[cfg(test)]
mod tests;