trueno-gpu 0.4.11

//! Basic Q4_K GEMV Kernel (PAR-003)
//!
//! Optimized for M=1 matmuls (token generation critical path):
//! y = W * x where W is (N×K) in Q4_K format, x is (K), y is (N)
//!
//! Strategy:
//! - One warp (32 threads) per output element
//! - Each thread processes K/32 super-block-elements sequentially
//! - Dequantizes Q4_K weights on-the-fly (no intermediate buffer)
//! - Warp shuffle reduce for final sum
//!
//! Memory bandwidth: 144 bytes per 256 values = 0.5625 bytes/value (vs 4 bytes for f32)
//! This is 7.1x more memory efficient than dequantize+GEMV approach.

use crate::kernels::quantize::{Kernel, Q4K_SUPER_BLOCK_BYTES, Q4K_SUPER_BLOCK_SIZE};
use crate::ptx::builder::{PtxArithmetic, PtxComparison, PtxControl};
use crate::ptx::{PtxKernel, PtxReg, PtxType};

/// Q4_K quantized GEMV kernel for M=1 decode throughput (PAR-003)
///
/// This kernel is optimized for the critical path of LLM token generation
/// where each new token requires M=1 matrix-vector multiplies through all layers.
///
/// # Performance
///
/// - Memory: Reads packed Q4_K directly (0.5625 bytes/value vs 4 bytes for f32)
/// - Compute: Fused dequant+multiply avoids intermediate buffer
/// - Reduction: Warp shuffle for fast parallel reduction
///
/// # Launch Configuration
///
/// - Grid: N blocks (one per output element)
/// - Block: 32 threads (one warp)
/// - No shared memory required
#[derive(Debug, Clone)]
pub struct Q4KGemvKernel {
    /// K dimension (input dimension, must be multiple of 256)
    pub k: u32,
    /// N dimension (output dimension)
    pub n: u32,
}

impl Q4KGemvKernel {
    /// Create a new Q4_K GEMV kernel for y = W * x
    ///
    /// # Arguments
    /// * `k` - Input vector length / weight matrix rows (must be multiple of 256)
    /// * `n` - Output vector length / weight matrix columns
    #[must_use]
    pub fn new(k: u32, n: u32) -> Self {
        Self { k, n }
    }

    /// Get number of super-blocks per row (ceiling division)
    #[must_use]
    pub const fn num_super_blocks_per_row(&self) -> u32 {
        // CRITICAL: GGUF uses ceiling division for super-block count
        (self.k + Q4K_SUPER_BLOCK_SIZE - 1) / Q4K_SUPER_BLOCK_SIZE
    }
}

impl Kernel for Q4KGemvKernel {
    fn name(&self) -> &str {
        "q4k_gemv_warp_reduce"
    }

    fn build_ptx(&self) -> PtxKernel {
        // No shared memory needed - each warp works independently
        PtxKernel::new("q4k_gemv_warp_reduce")
            .param(PtxType::U64, "y_ptr") // Output vector (N)
            .param(PtxType::U64, "w_ptr") // Q4_K weights (N × K/256 super-blocks)
            .param(PtxType::U64, "x_ptr") // Input vector (K)
            .param(PtxType::U32, "k_dim") // K dimension
            .param(PtxType::U32, "n_dim") // N dimension
            .build(|ctx| {
                // Block = 32 threads (one warp), grid = N blocks
                // Each block computes one output element y[block_id]

                let block_id = ctx.special_reg(PtxReg::CtaIdX);
                let thread_id = ctx.special_reg(PtxReg::TidX);

                // Bounds check: if block_id >= n_dim, return
                let n_dim = ctx.load_param_u32("n_dim");
                let oob = ctx.setp_ge_u32(block_id, n_dim);
                ctx.branch_if(oob, "exit");

                // Load parameters
                let k_dim = ctx.load_param_u32("k_dim");
                let y_ptr = ctx.load_param_u64("y_ptr");
                let w_ptr = ctx.load_param_u64("w_ptr");
                let x_ptr = ctx.load_param_u64("x_ptr");

                // Initialize accumulator for this output element
                let acc = ctx.mov_f32_imm(0.0);

                // Calculate number of super-blocks per row: ceil(K / 256)
                // CRITICAL: GGUF uses ceiling division for super-block count
                // e.g., K=5504 requires (5504+255)/256 = 22 super-blocks, not 21
                let k_rounded = ctx.add_u32(k_dim, Q4K_SUPER_BLOCK_SIZE - 1);
                let num_super_blocks = ctx.div_u32(k_rounded, Q4K_SUPER_BLOCK_SIZE);

                // Calculate base address for this row's Q4_K data
                // row_addr = w_ptr + block_id * num_super_blocks * 144
                let sb_bytes = ctx.mov_u32_imm(Q4K_SUPER_BLOCK_BYTES);
                let row_bytes = ctx.mul_u32_reg(num_super_blocks, sb_bytes);
                let row_offset = ctx.mul_wide_u32_reg(block_id, row_bytes);
                let row_base = ctx.add_u64(w_ptr, row_offset);

                // Loop over super-blocks
                let sb_idx = ctx.mov_u32_imm(0);

                ctx.label("sb_loop");
                let sb_done = ctx.setp_ge_u32(sb_idx, num_super_blocks);
                ctx.branch_if(sb_done, "sb_loop_end");

                // Calculate super-block address
                let sb_offset = ctx.mul_wide_u32(sb_idx, Q4K_SUPER_BLOCK_BYTES);
                let sb_addr = ctx.add_u64(row_base, sb_offset);

                // Load d (f16 at offset 0)
                let d_f16 = ctx.ld_global_f16(sb_addr);
                let d = ctx.cvt_f32_f16(d_f16);

                // Load dmin (f16 at offset 2)
                let two = ctx.mov_u64_imm(2);
                let dmin_addr = ctx.add_u64(sb_addr, two);
                let dmin_f16 = ctx.ld_global_f16(dmin_addr);
                let dmin = ctx.cvt_f32_f16(dmin_f16);

                // scales base = sb_addr + 4
                let four_64 = ctx.mov_u64_imm(4);
                let scales_base = ctx.add_u64(sb_addr, four_64);

                // Load all 12 scale bytes into registers for reuse
                // This avoids repeated global loads
                let s0 = ctx.ld_global_u8(scales_base);
                let s0_32 = ctx.cvt_u32_u8(s0);
                let one_64 = ctx.mov_u64_imm(1);
                let s1_addr = ctx.add_u64(scales_base, one_64);
                let s1 = ctx.ld_global_u8(s1_addr);
                let s1_32 = ctx.cvt_u32_u8(s1);
                let two_64 = ctx.mov_u64_imm(2);
                let s2_addr = ctx.add_u64(scales_base, two_64);
                let s2 = ctx.ld_global_u8(s2_addr);
                let s2_32 = ctx.cvt_u32_u8(s2);
                let three_64 = ctx.mov_u64_imm(3);
                let s3_addr = ctx.add_u64(scales_base, three_64);
                let s3 = ctx.ld_global_u8(s3_addr);
                let s3_32 = ctx.cvt_u32_u8(s3);
                let four_64b = ctx.mov_u64_imm(4);
                let s4_addr = ctx.add_u64(scales_base, four_64b);
                let s4 = ctx.ld_global_u8(s4_addr);
                let s4_32 = ctx.cvt_u32_u8(s4);
                let five_64 = ctx.mov_u64_imm(5);
                let s5_addr = ctx.add_u64(scales_base, five_64);
                let s5 = ctx.ld_global_u8(s5_addr);
                let s5_32 = ctx.cvt_u32_u8(s5);
                let six_64 = ctx.mov_u64_imm(6);
                let s6_addr = ctx.add_u64(scales_base, six_64);
                let s6 = ctx.ld_global_u8(s6_addr);
                let s6_32 = ctx.cvt_u32_u8(s6);
                let seven_64 = ctx.mov_u64_imm(7);
                let s7_addr = ctx.add_u64(scales_base, seven_64);
                let s7 = ctx.ld_global_u8(s7_addr);
                let s7_32 = ctx.cvt_u32_u8(s7);
                let eight_64 = ctx.mov_u64_imm(8);
                let s8_addr = ctx.add_u64(scales_base, eight_64);
                let s8 = ctx.ld_global_u8(s8_addr);
                let s8_32 = ctx.cvt_u32_u8(s8);
                let nine_64 = ctx.mov_u64_imm(9);
                let s9_addr = ctx.add_u64(scales_base, nine_64);
                let s9 = ctx.ld_global_u8(s9_addr);
                let s9_32 = ctx.cvt_u32_u8(s9);
                let ten_64 = ctx.mov_u64_imm(10);
                let s10_addr = ctx.add_u64(scales_base, ten_64);
                let s10 = ctx.ld_global_u8(s10_addr);
                let s10_32 = ctx.cvt_u32_u8(s10);
                let eleven_64 = ctx.mov_u64_imm(11);
                let s11_addr = ctx.add_u64(scales_base, eleven_64);
                let s11 = ctx.ld_global_u8(s11_addr);
                let s11_32 = ctx.cvt_u32_u8(s11);

                // Constants
                let mask_6bit = ctx.mov_u32_imm(0x3F);
                let mask_4bit = ctx.mov_u32_imm(0x0F);
                let four = ctx.mov_u32_imm(4);
                let six = ctx.mov_u32_imm(6);

                // Extract scale/min for all 8 blocks using get_scale_min_k4 logic:
                // Blocks 0-3: scale = scales[j] & 63, min = scales[j+4] & 63
                // Blocks 4-7: scale = (scales[j+4] & 0xF) | ((scales[j-4] >> 6) << 4)
                //             min = (scales[j+4] >> 4) | ((scales[j] >> 6) << 4)

                // Block 0: scale = s0 & 63, min = s4 & 63
                let scale0 = ctx.and_u32(s0_32, mask_6bit);
                let min0 = ctx.and_u32(s4_32, mask_6bit);
                let scale0_f = ctx.cvt_f32_u32(scale0);
                let min0_f = ctx.cvt_f32_u32(min0);

                // Block 1: scale = s1 & 63, min = s5 & 63
                let scale1 = ctx.and_u32(s1_32, mask_6bit);
                let min1 = ctx.and_u32(s5_32, mask_6bit);
                let scale1_f = ctx.cvt_f32_u32(scale1);
                let min1_f = ctx.cvt_f32_u32(min1);

                // Block 2: scale = s2 & 63, min = s6 & 63
                let scale2 = ctx.and_u32(s2_32, mask_6bit);
                let min2 = ctx.and_u32(s6_32, mask_6bit);
                let scale2_f = ctx.cvt_f32_u32(scale2);
                let min2_f = ctx.cvt_f32_u32(min2);

                // Block 3: scale = s3 & 63, min = s7 & 63
                let scale3 = ctx.and_u32(s3_32, mask_6bit);
                let min3 = ctx.and_u32(s7_32, mask_6bit);
                let scale3_f = ctx.cvt_f32_u32(scale3);
                let min3_f = ctx.cvt_f32_u32(min3);

                // Block 4: scale = (s8 & 0xF) | ((s0 >> 6) << 4)
                //          min = (s8 >> 4) | ((s4 >> 6) << 4)
                let s8_lo = ctx.and_u32(s8_32, mask_4bit);
                let s0_hi = ctx.shr_u32(s0_32, six);
                let s0_hi_shifted = ctx.shl_u32(s0_hi, four);
                let scale4 = ctx.or_u32(s8_lo, s0_hi_shifted);
                let s8_hi = ctx.shr_u32(s8_32, four);
                let s4_hi = ctx.shr_u32(s4_32, six);
                let s4_hi_shifted = ctx.shl_u32(s4_hi, four);
                let min4 = ctx.or_u32(s8_hi, s4_hi_shifted);
                let scale4_f = ctx.cvt_f32_u32(scale4);
                let min4_f = ctx.cvt_f32_u32(min4);

                // Block 5: scale = (s9 & 0xF) | ((s1 >> 6) << 4)
                //          min = (s9 >> 4) | ((s5 >> 6) << 4)
                let s9_lo = ctx.and_u32(s9_32, mask_4bit);
                let s1_hi = ctx.shr_u32(s1_32, six);
                let s1_hi_shifted = ctx.shl_u32(s1_hi, four);
                let scale5 = ctx.or_u32(s9_lo, s1_hi_shifted);
                let s9_hi = ctx.shr_u32(s9_32, four);
                let s5_hi = ctx.shr_u32(s5_32, six);
                let s5_hi_shifted = ctx.shl_u32(s5_hi, four);
                let min5 = ctx.or_u32(s9_hi, s5_hi_shifted);
                let scale5_f = ctx.cvt_f32_u32(scale5);
                let min5_f = ctx.cvt_f32_u32(min5);

                // Block 6: scale = (s10 & 0xF) | ((s2 >> 6) << 4)
                //          min = (s10 >> 4) | ((s6 >> 6) << 4)
                let s10_lo = ctx.and_u32(s10_32, mask_4bit);
                let s2_hi = ctx.shr_u32(s2_32, six);
                let s2_hi_shifted = ctx.shl_u32(s2_hi, four);
                let scale6 = ctx.or_u32(s10_lo, s2_hi_shifted);
                let s10_hi = ctx.shr_u32(s10_32, four);
                let s6_hi = ctx.shr_u32(s6_32, six);
                let s6_hi_shifted = ctx.shl_u32(s6_hi, four);
                let min6 = ctx.or_u32(s10_hi, s6_hi_shifted);
                let scale6_f = ctx.cvt_f32_u32(scale6);
                let min6_f = ctx.cvt_f32_u32(min6);

                // Block 7: scale = (s11 & 0xF) | ((s3 >> 6) << 4)
                //          min = (s11 >> 4) | ((s7 >> 6) << 4)
                let s11_lo = ctx.and_u32(s11_32, mask_4bit);
                let s3_hi = ctx.shr_u32(s3_32, six);
                let s3_hi_shifted = ctx.shl_u32(s3_hi, four);
                let scale7 = ctx.or_u32(s11_lo, s3_hi_shifted);
                let s11_hi = ctx.shr_u32(s11_32, four);
                let s7_hi = ctx.shr_u32(s7_32, six);
                let s7_hi_shifted = ctx.shl_u32(s7_hi, four);
                let min7 = ctx.or_u32(s11_hi, s7_hi_shifted);
                let scale7_f = ctx.cvt_f32_u32(scale7);
                let min7_f = ctx.cvt_f32_u32(min7);

                // Precompute d*scale and dmin*min for each block
                let ds0 = ctx.mul_f32(d, scale0_f);
                let dm0 = ctx.mul_f32(dmin, min0_f);
                let ds1 = ctx.mul_f32(d, scale1_f);
                let dm1 = ctx.mul_f32(dmin, min1_f);
                let ds2 = ctx.mul_f32(d, scale2_f);
                let dm2 = ctx.mul_f32(dmin, min2_f);
                let ds3 = ctx.mul_f32(d, scale3_f);
                let dm3 = ctx.mul_f32(dmin, min3_f);
                let ds4 = ctx.mul_f32(d, scale4_f);
                let dm4 = ctx.mul_f32(dmin, min4_f);
                let ds5 = ctx.mul_f32(d, scale5_f);
                let dm5 = ctx.mul_f32(dmin, min5_f);
                let ds6 = ctx.mul_f32(d, scale6_f);
                let dm6 = ctx.mul_f32(dmin, min6_f);
                let ds7 = ctx.mul_f32(d, scale7_f);
                let dm7 = ctx.mul_f32(dmin, min7_f);

                // qs base = sb_addr + 16
                let sixteen_64 = ctx.mov_u64_imm(16);
                let qs_base = ctx.add_u64(sb_addr, sixteen_64);

                // Each thread handles 8 values (256 values / 32 threads = 8 per thread)
                // Thread t handles values: t, t+32, t+64, t+96, t+128, t+160, t+192, t+224
                // These correspond to blocks: 0, 1, 2, 3, 4, 5, 6, 7
                let thread_partial = ctx.mov_f32_imm(0.0);

                // Process 8 values per thread (unrolled with known block index)
                let offsets_and_blocks: [(u32, u32); 8] = [
                    (0, 0),
                    (32, 1),
                    (64, 2),
                    (96, 3),
                    (128, 4),
                    (160, 5),
                    (192, 6),
                    (224, 7),
                ];

                for (offset, block_idx) in offsets_and_blocks {
                    // Get precomputed d*scale and dmin*min for this block
                    let (ds, dm) = match block_idx {
                        0 => (ds0, dm0),
                        1 => (ds1, dm1),
                        2 => (ds2, dm2),
                        3 => (ds3, dm3),
                        4 => (ds4, dm4),
                        5 => (ds5, dm5),
                        6 => (ds6, dm6),
                        _ => (ds7, dm7),
                    };

                    // Calculate value index within super-block
                    let offset_reg = ctx.mov_u32_imm(offset);
                    let val_idx = ctx.add_u32_reg(thread_id, offset_reg);

                    // Load 4-bit quantized value from qs (128 bytes for 256 values)
                    // qs layout: values are in 64-value chunks
                    //   chunk 0 (32 bytes): values 0-31 low nibbles, values 32-63 high nibbles
                    //   chunk 1 (32 bytes): values 64-95 low nibbles, values 96-127 high nibbles
                    //   etc.
                    let chunk_idx = ctx.div_u32(val_idx, 64);
                    let val_in_chunk = ctx.rem_u32(val_idx, 64);
                    let byte_in_chunk = ctx.rem_u32(val_in_chunk, 32);

                    // qs byte address = qs_base + chunk_idx * 32 + byte_in_chunk
                    let chunk_offset = ctx.mul_u32(chunk_idx, 32);
                    let qs_byte_offset = ctx.add_u32_reg(chunk_offset, byte_in_chunk);
                    let qs_byte_offset_64 = ctx.cvt_u64_u32(qs_byte_offset);
                    let qs_addr = ctx.add_u64(qs_base, qs_byte_offset_64);
                    let packed = ctx.ld_global_u8(qs_addr);
                    let packed_32 = ctx.cvt_u32_u8(packed);

                    // Extract nibble (low or high)
                    let mask_4bit_q = ctx.mov_u32_imm(0xF);
                    let four_q = ctx.mov_u32_imm(4);
                    let val_in_chunk_div_32 = ctx.div_u32(val_in_chunk, 32);
                    let shift_amount = ctx.mul_u32_reg(val_in_chunk_div_32, four_q);
                    let shifted = ctx.shr_u32(packed_32, shift_amount);
                    let quant = ctx.and_u32(shifted, mask_4bit_q);

                    // Dequantize: val = d*scale*quant - dmin*min
                    let quant_f32 = ctx.cvt_f32_u32(quant);
                    let scaled = ctx.mul_f32(ds, quant_f32);
                    let dequant = ctx.sub_f32(scaled, dm);

                    // Load activation x[sb_idx * 256 + val_idx]
                    let sb_k_base = ctx.mul_u32(sb_idx, Q4K_SUPER_BLOCK_SIZE);
                    let x_idx = ctx.add_u32_reg(sb_k_base, val_idx);
                    let x_idx_64 = ctx.cvt_u64_u32(x_idx);
                    let x_bytes = ctx.mul_u64(x_idx_64, 4);
                    let x_addr = ctx.add_u64(x_ptr, x_bytes);
                    let x_val = ctx.ld_global_f32(x_addr);

                    // Accumulate: thread_partial += x_val * dequant
                    ctx.fma_f32_inplace(thread_partial, x_val, dequant);
                }

                // Add thread's partial sum to accumulator
                ctx.add_f32_inplace(acc, thread_partial);

                // Next super-block
                ctx.add_u32_inplace(sb_idx, 1);
                ctx.branch("sb_loop");

                ctx.label("sb_loop_end");

                // Warp shuffle reduce: sum all 32 thread partials
                let tmp16 = ctx.shfl_down_f32(acc, 16, 0xFFFF_FFFF);
                ctx.add_f32_inplace(acc, tmp16);

                let tmp8 = ctx.shfl_down_f32(acc, 8, 0xFFFF_FFFF);
                ctx.add_f32_inplace(acc, tmp8);

                let tmp4 = ctx.shfl_down_f32(acc, 4, 0xFFFF_FFFF);
                ctx.add_f32_inplace(acc, tmp4);

                let tmp2 = ctx.shfl_down_f32(acc, 2, 0xFFFF_FFFF);
                ctx.add_f32_inplace(acc, tmp2);

                let tmp1 = ctx.shfl_down_f32(acc, 1, 0xFFFF_FFFF);
                ctx.add_f32_inplace(acc, tmp1);

                // Only thread 0 writes the result
                let one_u32 = ctx.mov_u32_imm(1);
                let is_thread0 = ctx.setp_lt_u32(thread_id, one_u32);
                ctx.branch_if_not(is_thread0, "exit");

                // Store y[block_id] = acc
                let y_offset = ctx.mul_wide_u32(block_id, 4);
                let y_addr = ctx.add_u64(y_ptr, y_offset);
                ctx.st_global_f32(y_addr, acc);

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