perf: group-split register-blocking gqa_prefill kernel

- one query row per group of G=8 lanes, each owning HEAD_DIM/G dims of qreg[]/acc[] in registers
- removes full 32-lane warp_reduce_sum; S dot reduces over only G lanes
- templated on <HEAD_DIM,G,ROWS,P_BC>, block=(G,ROWS)=(8,32)
- per-group shuffle mask so causal loop-bound divergence doesn't deadlock the shuffle
- update pure-C test to the templated launch

csrc/kernels/gqa_prefill_attn.cu

@@ -43,11 +43,25 @@ torch::Tensor gqa_prefill_attn(
     auto O = torch::empty_like(q);
     p.o = (bf16*)O.data_ptr();
 
-    dim3 grid((p.q_len + Br - 1) / Br, p.q_head, p.batch);
+    constexpr int G = 8, ROWS = 32, P_BC = 32;
-    dim3 block(32, Br, 1);
+    dim3 grid((p.q_len + ROWS - 1) / ROWS, p.q_head, p.batch);
-    size_t smem = 2 * Bc * p.head_dim * sizeof(bf16);
+    dim3 block(G, ROWS, 1);
+    size_t smem = 2 * P_BC * p.head_dim * sizeof(bf16);
 
-    gqa_prefill_attn_kernel<<<grid, block, smem>>>(p);
+    switch (p.head_dim) {
+        case 64:
+            gqa_prefill_attn_kernel_t<64, G, ROWS, P_BC><<<grid, block, smem>>>(p);
+            break;
+        case 128:
+            gqa_prefill_attn_kernel_t<128, G, ROWS, P_BC><<<grid, block, smem>>>(p);
+            break;
+        case 256:
+            gqa_prefill_attn_kernel_t<256, G, ROWS, P_BC><<<grid, block, smem>>>(p);
+            break;
+        default:
+            TORCH_CHECK(false, "prefill: unsupported head_dim ", p.head_dim,
+                        " (supported: 64,128,256)");
+    }
     return O;
 }
 

csrc/kernels/gqa_prefill_attn.cuh

@@ -1,44 +1,72 @@
 #pragma once
 #include "gqa_common.cuh"
 
-__global__ void gqa_prefill_attn_kernel(GQAParams p) {
+// v9: group-split register blocking. G threads cooperate on one query row,
+// each owning HEAD_DIM/G dims of qreg[]/acc[]. Small per-thread footprint keeps
+// occupancy high; the S dot product is reduced across the G-lane group with a
+// short shuffle chain (log2(G) shuffles) instead of a full 32-lane warp reduce.
+// Online (per-kv) softmax — cheap because acc[] is only HEAD_DIM/G long.
+// Templated on <HEAD_DIM, G, ROWS, P_BC>. Block = (G, ROWS). G power-of-two,
+// G*ROWS a multiple of 32 with groups warp-aligned.
+
+template <int G>
+__device__ __forceinline__ float group_reduce_sum(float v, unsigned mask) {
+#pragma unroll
+    for (int o = G / 2; o > 0; o >>= 1)
+        v += __shfl_xor_sync(mask, v, o);
+    return v;
+}
+
+template <int HEAD_DIM, int G, int ROWS, int P_BC>
+__global__ void gqa_prefill_attn_kernel_t(GQAParams p) {
+    constexpr int DPT = HEAD_DIM / G;
+
     int q_tile = blockIdx.x;
     int q_head = blockIdx.y;
     int batch  = blockIdx.z;
-    int q_row  = q_tile * Br + threadIdx.y;
+    int gpos   = threadIdx.x;  // 0..G-1  (which d-chunk)
-    int d_part = threadIdx.x;
+    int row    = threadIdx.y;  // 0..ROWS-1
-    int dpw    = p.head_dim >> 5;
+    int q_row  = q_tile * ROWS + row;
 
     int kv_head = q_head / (p.q_head / p.kv_head);
 
-    float qs[8] = {0};
-    if (q_row < p.q_len) {
-        int q_off = (((batch * p.q_head + q_head) * p.q_len + q_row) * p.head_dim) + d_part * dpw;
-        for (int i = 0; i < dpw; i++)
-            qs[i] = __bfloat162float(p.q[q_off + i]) * p.scale;
-    }
-
-    int kv_base = ((batch * p.kv_head + kv_head) * p.kv_len) * p.head_dim;
-
     extern __shared__ __align__(16) bf16 smem[];
     bf16* sK = smem;
-    bf16* sV = smem + Bc * p.head_dim;
+    bf16* sV = sK + P_BC * HEAD_DIM;
 
-    float m = -FLT_MAX, l = 0.0f, acc[8] = {0};
+    float qreg[DPT];
+    if (q_row < p.q_len) {
+        int q_off = ((batch * p.q_head + q_head) * p.q_len + q_row) * HEAD_DIM + gpos * DPT;
+#pragma unroll
+        for (int i = 0; i < DPT; i++)
+            qreg[i] = __bfloat162float(p.q[q_off + i]) * p.scale;
+    }
 
-    int tiles = (p.kv_len + Bc - 1) / Bc;
+    float m = -FLT_MAX, l = 0.0f;
-    int tt = blockDim.x * blockDim.y;
+    float acc[DPT];
+#pragma unroll
+    for (int i = 0; i < DPT; i++)
+        acc[i] = 0.0f;
+
+    int kv_base = ((batch * p.kv_head + kv_head) * p.kv_len) * HEAD_DIM;
+    int tiles   = (p.kv_len + P_BC - 1) / P_BC;
+    int tt      = G * ROWS;
+    int lid     = row * G + gpos;
+
+    // per-group shuffle mask: only the G lanes of this row's group participate,
+    // so causal masking (differing loop bounds across rows in a warp) is safe.
+    int lane_in_warp = lid & 31;
+    unsigned gmask = (G == 32) ? 0xFFFFFFFFu
+                               : (((1u << G) - 1u) << (lane_in_warp & ~(G - 1)));
 
     for (int ti = 0; ti < tiles; ti++) {
-        int kv0  = ti * Bc;
+        int kv0  = ti * P_BC;
-        int tlen = min(Bc, p.kv_len - kv0);
+        int tlen = min(P_BC, p.kv_len - kv0);
 
-        for (int i = threadIdx.y * blockDim.x + threadIdx.x;
+        for (int i = lid; i < tlen * HEAD_DIM; i += tt) {
-             i < tlen * p.head_dim; i += tt) {
+            int gidx = kv_base + (kv0 + i / HEAD_DIM) * HEAD_DIM + (i % HEAD_DIM);
-            int r = i / p.head_dim, c = i % p.head_dim, idx = r * p.head_dim + c;
+            sK[i] = p.k[gidx];
-            int g_off = kv_base + (kv0 + r) * p.head_dim + c;
+            sV[i] = p.v[gidx];
-            sK[idx] = p.k[g_off];
-            sV[idx] = p.v[g_off];
         }
         __syncthreads();
 
@@ -52,30 +80,35 @@ __global__ void gqa_prefill_attn_kernel(GQAParams p) {
         }
 
         for (int s = 0; s < lim; s++) {
-            float dot = 0.0f;
+            const bf16* kr = sK + s * HEAD_DIM + gpos * DPT;
-            for (int i = 0; i < dpw; i++)
+            float part = 0.0f;
-                dot += qs[i] * __bfloat162float(sK[s * p.head_dim + d_part * dpw + i]);
+#pragma unroll
-            dot = warp_reduce_sum(dot);
+            for (int i = 0; i < DPT; i++)
+                part = fmaf(qreg[i], __bfloat162float(kr[i]), part);
+            float dot = group_reduce_sum<G>(part, gmask);
 
             if (p.use_mask && p.mask && !p.mask[batch * p.kv_len + kv0 + s])
                 dot = -FLT_MAX;
 
             float nm = fmaxf(m, dot);
-            float al = expf(m - nm);
+            float al = __expf(m - nm);
-            float be = expf(dot - nm);
+            float be = __expf(dot - nm);
             l = l * al + be;
 
-            for (int i = 0; i < dpw; i++)
+            const bf16* vr = sV + s * HEAD_DIM + gpos * DPT;
-                acc[i] = acc[i] * al + __bfloat162float(sV[s * p.head_dim + d_part * dpw + i]) * be;
+#pragma unroll
+            for (int i = 0; i < DPT; i++)
+                acc[i] = fmaf(__bfloat162float(vr[i]), be, acc[i] * al);
             m = nm;
         }
         __syncthreads();
     }
 
     if (q_row < p.q_len) {
-        int o_off = (((batch * p.q_head + q_head) * p.q_len + q_row) * p.head_dim) + d_part * dpw;
+        int o_off = ((batch * p.q_head + q_head) * p.q_len + q_row) * HEAD_DIM + gpos * DPT;
         float rl = (l > 1e-10f) ? (1.0f / l) : 0.0f;
-        for (int i = 0; i < dpw; i++)
+#pragma unroll
+        for (int i = 0; i < DPT; i++)
             p.o[o_off + i] = __float2bfloat16(acc[i] * rl);
     }
 }

csrc/tests/gqa_prefill_test.cu

@@ -87,14 +87,20 @@ int main() {
         p.scale=1.0f/sqrtf((float)D);
         p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
 
-        dim3 grid((ql+Br-1)/Br, Hq, B);
+        constexpr int G=8, ROWS=32, P_BC=32;
-        dim3 block(32, Br, 1);
+        dim3 grid((ql+ROWS-1)/ROWS, Hq, B);
-        size_t smem=2*Bc*D*sizeof(bf16);
+        dim3 block(G, ROWS, 1);
+        size_t smem=2*P_BC*D*sizeof(bf16);
         printf("grid=(%d,%d,%d) block=(%d,%d,%d) smem=%zu\n",
                grid.x,grid.y,grid.z, block.x,block.y,block.z, smem);
 
         double t0=now_ms();
-        gqa_prefill_attn_kernel<<<grid,block,smem>>>(p);
+        switch (D) {
+            case 32:  gqa_prefill_attn_kernel_t<32, G,ROWS,P_BC><<<grid,block,smem>>>(p); break;
+            case 64:  gqa_prefill_attn_kernel_t<64, G,ROWS,P_BC><<<grid,block,smem>>>(p); break;
+            case 128: gqa_prefill_attn_kernel_t<128,G,ROWS,P_BC><<<grid,block,smem>>>(p); break;
+            default: printf("unsupported D=%d\n",D); return 1;
+        }
         cudaDeviceSynchronize();
         double kms=now_ms()-t0;
         cudaError_t err=cudaGetLastError();