[FA2] Update flash-attn-mma shared-qkv🎉 (#168)

* Update flash_attn_mma_share_qkv.cu * Update flash_attn_mma_share_kv.cu * Update flash_attn_mma_share_kv.cu * Update flash_attn_mma_share_qkv.cu * Update flash_attn_mma.py * Update README.md * Update README.md * Update README.md
DefTruth · Dec 19, 2024 · 9324ddf · 9324ddf
1 parent db8b8e8
commit 9324ddf
Show file tree

Hide file tree

Showing 5 changed files with 49 additions and 43 deletions.
diff --git a/README.md b/README.md
@@ -119,7 +119,7 @@ flash_attn_mma_stages_split_q_kernel(half* Q, // [B, H, N, D]
                                      int QKV_seqlen);
 ```
 
-- 📚 Split Q + Shared KV SMEM (Faster+)
+- 📚 Split Q + Shared KV SMEM (**1/2 SRAM** vs FA2)
 <div id="mma-share-kv"></div>  
 
 ```C++
@@ -131,7 +131,7 @@ flash_attn_mma_stages_split_q_shared_kv_kernel(half* Q,
                                                half* O, 
                                                int QKV_seqlen);
 ```
-- 📚 Split Q + Fully Shared QKV SMEM (Faster++)
+- 📚 Split Q + Fully Shared QKV SMEM (**1/4 SRAM** vs FA2)
 
 <div id="mma-share-qkv"></div>  
 

diff --git a/kernels/flash-attn/README.md b/kernels/flash-attn/README.md
@@ -93,7 +93,7 @@ flash_attn_mma_stages_split_q_kernel(half* Q, // [B, H, N, D]
                                      int QKV_seqlen);
 ```
 
-- 📚 Split Q + Shared KV SMEM (Faster+)
+- 📚 Split Q + Shared KV SMEM (**1/2 SRAM** vs FA2)
 <div id="mma-share-kv"></div>  
 
 ```C++
@@ -105,7 +105,7 @@ flash_attn_mma_stages_split_q_shared_kv_kernel(half* Q,
                                                half* O, 
                                                int QKV_seqlen);
 ```
-- 📚 Split Q + Fully Shared QKV SMEM (Faster++)
+- 📚 Split Q + Fully Shared QKV SMEM (**1/4 SRAM** vs FA2)
 
 <div id="mma-share-qkv"></div>  
 
@@ -119,6 +119,7 @@ flash_attn_mma_stages_split_q_shared_qkv_kernel(half* Q,
                                                 int QKV_seqlen);
 ```
 
+
 ## 📖 Prerequisites
 <div id="prerequisites"></div>  
 

diff --git a/kernels/flash-attn/flash_attn_mma.py b/kernels/flash-attn/flash_attn_mma.py
@@ -45,10 +45,11 @@ def get_args():
     parser.add_argument("--N", type=int, default=None)
     parser.add_argument("--D", type=int, default=None)
     parser.add_argument("--seed", type=int, default=None)
+    parser.add_argument("--sleep", type=float, default=0.05)
     parser.add_argument("--debug", action="store_true")
     parser.add_argument("--verbose", '--v', action="store_true")
-    parser.add_argument("--warmup", type=int, default=1)
-    parser.add_argument("--iters", type=int, default=5)
+    parser.add_argument("--warmup", "--w", type=int, default=1)
+    parser.add_argument("--iters", "--i", type=int, default=5)
     parser.add_argument("--range-k", '--gk', action="store_true")
     return parser.parse_args()
 
@@ -178,7 +179,7 @@ def run_benchmark(perf_func: callable,
     print(f"{out_info:>30}: {out_val}, time:{mean_time:<.6f}ms, TFLOPS:{TFLOPS:<6.2f}")
     if show_all: 
         print(out)
-    time.sleep(0.05)
+    time.sleep(args.sleep)
     torch.cuda.synchronize()
     return out.clone(), mean_time
 

diff --git a/kernels/flash-attn/mma/flash_attn_mma_share_kv.cu b/kernels/flash-attn/mma/flash_attn_mma_share_kv.cu
@@ -213,38 +213,8 @@ flash_attn_mma_stages_split_q_shared_kv_kernel(half* Q,
   for (int tile_K_seqlen = 0; tile_K_seqlen < Tc; ++tile_K_seqlen) { 
     // TODO: process last tile_K_seqlen ? pad to multiple of 8.
 
-    // <Prefetch Q s2r>: Load Q tile from smem -> regs, before Q@K^T.
-    if constexpr (kCanPrefetchQs2r) {
-      // Wait Q ready and let K copy async, then prefetch Q from smem -> regs.
-      // NOTE: we only need to load Q once from smem -> regs, and then reuse it.
-      if (tile_K_seqlen == 0) {
-        CP_ASYNC_WAIT_GROUP(0); 
-        __syncthreads(); 
-
-        #pragma unroll
-        for (int tile_K_d = 0; tile_K_d < (kHeadDim / kMmaAtomK); ++tile_K_d) {
-          // Allocate R_Q[(kHeadDim / kMmaAtomK)][1][4], e.g R_Q[4][1][4] 16 regs. 
-          // By the way, we have to reduce R_Z to 0 regs and reuse R_Q for collective store.
-          // Then we can load Q from smem only once and reuse it for <loop over K seqlen>
-          // processes. This will reduce large io-access for Q smem while N is large.
-          #pragma unroll
-          for (int i = 0; i < kWarpTileSeqLenQ; ++i) { // Q[Br,d]=[M,K]
-            int warp_smem_Q_Br = warp_QP * (kMmaAtomM * kWarpTileSeqLenQ) + i * kMmaAtomM;
-            int lane_smem_Q_Br = warp_smem_Q_Br + lane_id % 16; // 0~15
-            int lane_smem_Q_d  = tile_K_d * kMmaAtomK + (lane_id / 16) * 8; // 0,8
-            uint32_t lane_smem_Q_ptr = (
-                smem_Q_base_ptr + (lane_smem_Q_Br * (kHeadDim + kPad) + 
-                                   lane_smem_Q_d) * sizeof(half)
-            );
-            LDMATRIX_X4(R_Q[tile_K_d][i][0], R_Q[tile_K_d][i][1], 
-                        R_Q[tile_K_d][i][2], R_Q[tile_K_d][i][3], 
-                        lane_smem_Q_ptr); // now, R_Q[1/2/4/8][1][4]
-          }
-        }
-      } // end if tile_K_seqlen == 0
-    } // end if kCanPrefetchQs2r
-
-    // Load K tile from gmem -> smem, always use smem part 0.
+    // Load K tile from gmem -> smem, always use smem part 0, send g2s 
+    // memory issues before Prefetch Q s2r to enable time overlap.
     if constexpr (kCanPrefetchKVg2s) {
       if (tile_K_seqlen == 0) {
         load_gmem_K_Bc_offset = tile_K_seqlen * Bc; // e.g (0~3)*64=(0,64,128,192,...)
@@ -301,6 +271,38 @@ flash_attn_mma_stages_split_q_shared_kv_kernel(half* Q,
       __syncthreads(); 
     }
 
+    // <Prefetch Q s2r>: Load Q tile from smem -> regs, before Q@K^T.
+    if constexpr (kCanPrefetchQs2r) {
+      // Wait Q ready and let K copy async, then prefetch Q from smem -> regs.
+      // NOTE: we only need to load Q once from smem -> regs, and then reuse it.
+      if (tile_K_seqlen == 0) {
+        CP_ASYNC_WAIT_GROUP(0); 
+        __syncthreads(); 
+
+        #pragma unroll
+        for (int tile_K_d = 0; tile_K_d < (kHeadDim / kMmaAtomK); ++tile_K_d) {
+          // Allocate R_Q[(kHeadDim / kMmaAtomK)][1][4], e.g R_Q[4][1][4] 16 regs. 
+          // By the way, we have to reduce R_Z to 0 regs and reuse R_Q for collective store.
+          // Then we can load Q from smem only once and reuse it for <loop over K seqlen>
+          // processes. This will reduce large io-access for Q smem while N is large.
+          #pragma unroll
+          for (int i = 0; i < kWarpTileSeqLenQ; ++i) { // Q[Br,d]=[M,K]
+            int warp_smem_Q_Br = warp_QP * (kMmaAtomM * kWarpTileSeqLenQ) + i * kMmaAtomM;
+            int lane_smem_Q_Br = warp_smem_Q_Br + lane_id % 16; // 0~15
+            int lane_smem_Q_d  = tile_K_d * kMmaAtomK + (lane_id / 16) * 8; // 0,8
+            uint32_t lane_smem_Q_ptr = (
+                smem_Q_base_ptr + (lane_smem_Q_Br * (kHeadDim + kPad) + 
+                                   lane_smem_Q_d) * sizeof(half)
+            );
+            LDMATRIX_X4(R_Q[tile_K_d][i][0], R_Q[tile_K_d][i][1], 
+                        R_Q[tile_K_d][i][2], R_Q[tile_K_d][i][3], 
+                        lane_smem_Q_ptr); // now, R_Q[1/2/4/8][1][4]
+          }
+        }
+        __syncthreads(); // wait all warps ready.
+      } // end if tile_K_seqlen == 0
+    } // end if kCanPrefetchQs2r
+
     // <loop over K d>: tile_K_d, kMmaAtomK = 16, K_tile_d[kMmaAtomK,Bc]
     // Matmul with NN layout, Q row major, K row major. 
     // S_tile[Br,Bc]=Q_tile[Br,d]@K[d,Bc]

diff --git a/kernels/flash-attn/mma/flash_attn_mma_share_qkv.cu b/kernels/flash-attn/mma/flash_attn_mma_share_qkv.cu
@@ -240,20 +240,22 @@ flash_attn_mma_stages_split_q_shared_qkv_kernel(half* Q,
                         lane_smem_Q_ptr); // now, R_Q[1/2/4/8][1][4]
           }
         }
+        __syncthreads(); // wait all warps ready.
       } // end if tile_K_seqlen == 0
     } // end if kCanPrefetchQs2r
 
-    // Load K tile from gmem -> smem, always use smem part 0.
+    // Load K tile from gmem -> smem, always use smem part 0. 
+    // must after prefetch Q s2r in order to reuse Q smem.
     if constexpr (kCanPrefetchKVg2s) {
       if (tile_K_seqlen == 0) {
         load_gmem_K_Bc_offset = tile_K_seqlen * Bc; // e.g (0~3)*64=(0,64,128,192,...)
         int load_gmem_K_d  = load_smem_K_d; // load K^T [d,Bc] from [d,seqlen]
         int load_gmem_K_Bc = load_gmem_K_Bc_offset + load_smem_K_Bc; // < seqlen
         int load_gmem_K_addr = (K_gmem_offset + load_gmem_K_d * QKV_seqlen + load_gmem_K_Bc);
         uint32_t load_smem_K_ptr = (
-        smem_K_base_ptr + (kPrefetchKg2sSmemId * KV_tile_size + 
-                           load_smem_K_d * (Bc + kPad) + 
-                           load_smem_K_Bc) * sizeof(half));
+          smem_K_base_ptr + (kPrefetchKg2sSmemId * KV_tile_size + 
+                             load_smem_K_d * (Bc + kPad) + 
+                             load_smem_K_Bc) * sizeof(half));
         #pragma unroll
         for (int i = 0; i < (Bc / (kNumThreads / kHeadDim)); i += 8) {
         CP_ASYNC_CG(load_smem_K_ptr + i * 2, &K[load_gmem_K_addr + i], 16);