201 lines
7.8 KiB
Plaintext
201 lines
7.8 KiB
Plaintext
#pragma once
|
|
#include <cfloat>
|
|
#include <cuda_bf16.h>
|
|
#include "attn_common.h"
|
|
#include "attn_mma_utils.cuh"
|
|
|
|
using bf16 = __nv_bfloat16;
|
|
|
|
// Split-K (FlashDecoding) tensor-core decode via GQA head-packing.
|
|
//
|
|
// Decode has q_len == 1, so S = q @ K^T is a GEMV per head — no tensor-core work
|
|
// on its own. But GQA gives us G = q_head / kv_head query heads that all share
|
|
// one kv_head. We pack those G heads into the M=16 rows of mma.sync.m16n8k16,
|
|
// turning G independent GEMVs into a single GEMM that reuses each loaded K/V tile
|
|
// across all G heads (K/V load is the decode bottleneck, so the reuse is the win,
|
|
// not the flops). The KV sequence is partitioned across gridDim.z blocks so that
|
|
// a decode with only batch*kv_head independent tasks can fill all SMs. Each
|
|
// (batch, kv_head, split) block computes an UN-normalised partial (Oacc, m, l)
|
|
// over its KV slice; the combine kernel below reduces across splits. Fixes the
|
|
// "grid too small" bottleneck (0.04 waves/SM → many blocks) for long-context,
|
|
// small-batch decode.
|
|
//
|
|
// Partial layout (float, contiguous):
|
|
// o_part : [batch, q_head, num_splits, HEAD_DIM]
|
|
// ml_part: [batch, q_head, num_splits, 2] (m, l)
|
|
//
|
|
// Optimizations:
|
|
// - cp.async global→shared for K/V (bypasses registers, cuts instruction count)
|
|
// - XOR swizzle (swiz_col): LD=HEAD_DIM, zero waste, no bank conflicts
|
|
// - Q loaded directly from global into mma A-operand registers (no sQ staging,
|
|
// no prologue syncwarp) — frees shared memory for double-buffering
|
|
// - Double-buffered KV (STAGES=2): next tile's cp.async overlaps current
|
|
// tile's MMA compute — hides global load latency / boosts bandwidth
|
|
// utilization for small-batch (low-occupancy) decode
|
|
// - Predicated cp.async (cp_async_16_pred) for full AND partial tiles on one
|
|
// uniform path — eliminates the scalar fallback branch
|
|
//
|
|
// Smem footprint (BC=32): STAGES=2 → 2*(sK+sV) = 2*2*32*HEAD_DIM*2 bytes.
|
|
// D=128: 16 KB (fits 48 KB static cap). D=256: 32 KB (also fits).
|
|
// STAGES=1 fallback (4/8 KB) for smem-constrained configs.
|
|
template <int HEAD_DIM, int BC, int STAGES = 2>
|
|
__global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
|
|
constexpr int BR = 16;
|
|
constexpr int KD = HEAD_DIM / 16;
|
|
constexpr int NC8 = BC / 8;
|
|
constexpr int KT2 = BC / 16;
|
|
constexpr int DN8 = HEAD_DIM / 8;
|
|
constexpr int LD = HEAD_DIM;
|
|
constexpr int SWIZ_MASK = (HEAD_DIM >= 64) ? 7 : (HEAD_DIM / 8 - 1);
|
|
constexpr int VEC = 8;
|
|
constexpr int TOTAL = BC * HEAD_DIM;
|
|
|
|
const int lane = threadIdx.x;
|
|
const int gid = lane >> 2;
|
|
const int tid4 = lane & 3;
|
|
|
|
const int kv_head = blockIdx.x;
|
|
const int batch = blockIdx.y;
|
|
const int split = blockIdx.z;
|
|
const int G = p.q_head / p.kv_head;
|
|
const int q_head0 = kv_head * G;
|
|
|
|
// Double-buffered shared memory for K/V (no sQ needed — Q goes direct
|
|
// from global to registers).
|
|
__shared__ __align__(16) bf16 sK[STAGES * BC * LD];
|
|
__shared__ __align__(16) bf16 sV[STAGES * BC * LD];
|
|
|
|
// ---- Load Q directly from global into mma A-operand registers ----
|
|
// Same layout as prefill: frag[0]/[2] = row gid, frag[1]/[3] = row gid+8
|
|
// cols kt*16 + tid4*2 + {0,1} / +{8,9}. pau[0]=cols c,c+1; pau[4]=c+8,c+9.
|
|
const int q_base = (batch * p.q_head + q_head0) * HEAD_DIM;
|
|
const int qra = gid;
|
|
const int qrb = gid + 8;
|
|
const bool va = qra < G, vb = qrb < G;
|
|
unsigned Qa[KD][4];
|
|
#pragma unroll
|
|
for (int kt = 0; kt < KD; kt++) {
|
|
int c = kt * 16 + tid4 * 2;
|
|
const unsigned* pau = reinterpret_cast<const unsigned*>(
|
|
&p.q[q_base + qra * HEAD_DIM + c]);
|
|
const unsigned* pbu = reinterpret_cast<const unsigned*>(
|
|
&p.q[q_base + qrb * HEAD_DIM + c]);
|
|
Qa[kt][0] = va ? pau[0] : 0u;
|
|
Qa[kt][1] = vb ? pbu[0] : 0u;
|
|
Qa[kt][2] = va ? pau[4] : 0u;
|
|
Qa[kt][3] = vb ? pbu[4] : 0u;
|
|
}
|
|
|
|
float Oacc[DN8][4];
|
|
#pragma unroll
|
|
for (int j = 0; j < DN8; j++)
|
|
Oacc[j][0] = Oacc[j][1] = Oacc[j][2] = Oacc[j][3] = 0.0f;
|
|
float m0 = -FLT_MAX, m1 = -FLT_MAX, l0 = 0.0f, l1 = 0.0f;
|
|
|
|
const int kv_base = (batch * p.kv_head + kv_head) * p.kv_len * HEAD_DIM;
|
|
const int mask_base = batch * p.kv_len;
|
|
const int tiles_total = (p.kv_len + BC - 1) / BC;
|
|
const int tiles_per_split = (tiles_total + p.num_splits - 1) / p.num_splits;
|
|
const int ti_begin = split * tiles_per_split;
|
|
const int ti_end = min(tiles_total, ti_begin + tiles_per_split);
|
|
const int has_mask = p.use_mask && p.mask;
|
|
|
|
// ---- Load tile lambda: predicated cp.async, unified full/partial ----
|
|
auto load_tile = [&](int ti, int buf) {
|
|
int kv0 = ti * BC;
|
|
bf16* dK = sK + buf * BC * LD;
|
|
bf16* dV = sV + buf * BC * LD;
|
|
#pragma unroll
|
|
for (int i = lane * VEC; i < TOTAL; i += 32 * VEC) {
|
|
int r = i / HEAD_DIM, d = i % HEAD_DIM;
|
|
int kc = kv0 + r;
|
|
bool valid = kc < p.kv_len;
|
|
int off = r * LD + swiz_col(d, r, SWIZ_MASK);
|
|
cp_async_16_pred(&dK[off], &p.k[kv_base + kc * HEAD_DIM + d], valid);
|
|
cp_async_16_pred(&dV[off], &p.v[kv_base + kc * HEAD_DIM + d], valid);
|
|
}
|
|
cp_async_commit();
|
|
};
|
|
|
|
// ---- Prologue: issue first tile load ----
|
|
if (ti_begin < ti_end) {
|
|
load_tile(ti_begin, 0);
|
|
}
|
|
|
|
for (int ti = ti_begin; ti < ti_end; ti++) {
|
|
constexpr int BUF_MASK = (STAGES > 1) ? (STAGES - 1) : 0;
|
|
int buf = (ti - ti_begin) & BUF_MASK;
|
|
|
|
// Wait for current tile, then issue next tile's prefetch (overlaps
|
|
// with this tile's compute). Single syncwarp covers both hazards.
|
|
// When STAGES==1, no prefetch — load happens at end of prior iter.
|
|
cp_async_wait_group<0>();
|
|
__syncwarp();
|
|
if constexpr (STAGES > 1) {
|
|
if (ti + 1 < ti_end)
|
|
load_tile(ti + 1, (ti + 1 - ti_begin) & BUF_MASK);
|
|
}
|
|
|
|
const bf16* bK = sK + buf * BC * LD;
|
|
const bf16* bV = sV + buf * BC * LD;
|
|
int kv0 = ti * BC;
|
|
|
|
float Sacc[NC8][4];
|
|
mma_compute_scores<KD, NC8>(Qa, bK, LD, SWIZ_MASK, lane, Sacc);
|
|
|
|
#pragma unroll
|
|
for (int n8 = 0; n8 < NC8; n8++)
|
|
Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
|
|
Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
|
|
|
|
int maxc = p.is_causal ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
|
|
mma_softmax_tile<NC8, DN8>(kv0, maxc, maxc,
|
|
mask_base, p.mask, has_mask,
|
|
Sacc, Oacc, m0, m1, l0, l1, lane);
|
|
|
|
mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);
|
|
__syncwarp();
|
|
|
|
if constexpr (STAGES == 1) {
|
|
if (ti + 1 < ti_end)
|
|
load_tile(ti + 1, 0);
|
|
}
|
|
}
|
|
|
|
// ---- write UN-normalised partials for this split ----
|
|
auto split_slot = [&](int h) -> size_t {
|
|
size_t bh = (size_t)batch * p.q_head + h;
|
|
return bh * p.num_splits + split;
|
|
};
|
|
#pragma unroll
|
|
for (int dn8 = 0; dn8 < DN8; dn8++) {
|
|
int d = dn8 * 8 + 2 * tid4;
|
|
int r0 = gid, r1 = gid + 8;
|
|
if (r0 < G) {
|
|
int h = q_head0 + r0;
|
|
float* op = p.o_part + split_slot(h) * HEAD_DIM;
|
|
op[d] = Oacc[dn8][0];
|
|
op[d + 1] = Oacc[dn8][1];
|
|
}
|
|
if (r1 < G) {
|
|
int h = q_head0 + r1;
|
|
float* op = p.o_part + split_slot(h) * HEAD_DIM;
|
|
op[d] = Oacc[dn8][2];
|
|
op[d + 1] = Oacc[dn8][3];
|
|
}
|
|
}
|
|
if (tid4 == 0) {
|
|
int r0 = gid, r1 = gid + 8;
|
|
if (r0 < G) {
|
|
int h = q_head0 + r0;
|
|
float* mp = p.ml_part + split_slot(h) * 2;
|
|
mp[0] = m0; mp[1] = l0;
|
|
}
|
|
if (r1 < G) {
|
|
int h = q_head0 + r1;
|
|
float* mp = p.ml_part + split_slot(h) * 2;
|
|
mp[0] = m1; mp[1] = l1;
|
|
}
|
|
}
|
|
}
|