[cherry-pick] Fix multihead op bug. (#20783) (#21438)

The op should handle k=1024
Fix seq_len < warpsize error.

test=develop

Signed-off-by: zhaoyuchen <[email protected]>

Author: zhaoyuchen2018

paddle/fluid/operators/multihead_matmul_op.cc

@@ -134,7 +134,7 @@ MultiHeadMatMul Operator.
 This op is used for optimize multi head calculation in ernie model.
 Not suggest to use in other case except has same structure as ernie.
 
-Example of matrix multiplication with head_number of H
+Example of matrix multiplication with head_number of B
 - X: [B, M, K], Y: [B, K, N] => Out: [B, M, N]
 
 Both the input `Q` and `K` can carry the LoD (Level of Details) information,

paddle/fluid/operators/multihead_matmul_op.cu

@@ -28,10 +28,10 @@ namespace operators {
 #define WARP_SIZE 32
 
 template <typename T>
-__inline__ __device__ T warpReduceSum(T val) {
+__inline__ __device__ T warpReduceSum(T val, unsigned lane_mask) {
   for (int mask = HALF_WARP; mask > 0; mask >>= 1)
 #if __CUDA_ARCH__ >= 350 && CUDA_VERSION >= 9000
-    val += __shfl_xor_sync(FINAL_MASK, val, mask, warpSize);
+    val += __shfl_xor_sync(lane_mask, val, mask, warpSize);
 #else
     val += __shfl_xor(val, mask, warpSize);
 #endif
@@ -40,28 +40,30 @@ __inline__ __device__ T warpReduceSum(T val) {
 
 /* Calculate the sum of all elements in a block */
 template <typename T>
-__inline__ __device__ T blockReduceSum(T val) {
+__inline__ __device__ T blockReduceSum(T val, unsigned mask) {
   static __shared__ T shared[WARP_SIZE];
   int lane = threadIdx.x & 0x1f;
   int wid = threadIdx.x >> 5;
 
-  val = warpReduceSum<T>(val);
+  val = warpReduceSum<T>(val, mask);
 
   if (lane == 0) shared[wid] = val;
 
   __syncthreads();
 
-  val = (threadIdx.x < (blockDim.x >> 5)) ? shared[lane] : (T)(0.0f);
-  val = warpReduceSum<T>(val);
+  // align block_span to warpSize
+  int block_span = (blockDim.x + warpSize - 1) >> 5;
+  val = (threadIdx.x < block_span) ? shared[lane] : (T)(0.0f);
+  val = warpReduceSum<T>(val, mask);
 
   return val;
 }
 
 template <typename T>
-__inline__ __device__ T warpReduceMax(T val) {
+__inline__ __device__ T warpReduceMax(T val, unsigned lane_mask) {
   for (int mask = HALF_WARP; mask > 0; mask >>= 1)
 #if __CUDA_ARCH__ >= 350 && CUDA_VERSION >= 9000
-    val = max(val, __shfl_xor_sync(FINAL_MASK, val, mask, warpSize));
+    val = max(val, __shfl_xor_sync(lane_mask, val, mask, warpSize));
 #else
     val = max(val, __shfl_xor(val, mask, warpSize));
 #endif
@@ -70,19 +72,21 @@ __inline__ __device__ T warpReduceMax(T val) {
 
 /* Calculate the maximum of all elements in a block */
 template <typename T>
-__inline__ __device__ T blockReduceMax(T val) {
+__inline__ __device__ T blockReduceMax(T val, unsigned mask) {
   static __shared__ T shared[WARP_SIZE];
   int lane = threadIdx.x & 0x1f;
   int wid = threadIdx.x >> 5;
 
-  val = warpReduceMax(val);
+  val = warpReduceMax(val, mask);
 
   if (lane == 0) shared[wid] = val;
 
   __syncthreads();
 
-  val = (threadIdx.x < (blockDim.x >> 5)) ? shared[lane] : -1e10f;
-  val = warpReduceMax(val);
+  // align block_span to warpSize
+  int block_span = (blockDim.x + warpSize - 1) >> 5;
+  val = (threadIdx.x < block_span) ? shared[lane] : -1e10f;
+  val = warpReduceMax(val, mask);
 
   return val;
 }
@@ -190,7 +194,8 @@ template <typename T>
 __global__ void softmax_kernel_with_eltadd(T *qk_buf_, const T *bias_qk_,
                                            const int batch_size,
                                            const int head_num,
-                                           const int seq_len) {
+                                           const int seq_len,
+                                           const unsigned mask) {
   int seq_id = blockIdx.x % seq_len;
   int qk_offset = blockIdx.x * seq_len;
   int bias_offset = blockIdx.x % (head_num * seq_len) * seq_len;
@@ -202,13 +207,15 @@ __global__ void softmax_kernel_with_eltadd(T *qk_buf_, const T *bias_qk_,
                                        bias_qk_[threadIdx.x + bias_offset]))
                  : 0.0f;
   float tmp = threadIdx.x < seq_len ? static_cast<float>(qk) : -1e20f;
-  float max_val = blockReduceMax<float>(tmp);
+
+  float max_val = blockReduceMax<float>(tmp, mask);
+
   if (threadIdx.x == 0) s_max = max_val;
   __syncthreads();
 
   float qk_tmp =
       threadIdx.x < seq_len ? __expf(static_cast<float>(tmp - s_max)) : 0.0f;
-  float sum_val = blockReduceSum<float>(qk_tmp);
+  float sum_val = blockReduceSum<float>(qk_tmp, mask);
 
   if (threadIdx.x == 0) {
     s_sum = sum_val + 1e-6f;
@@ -258,8 +265,9 @@ void MatMulWithHeadQK(const platform::CUDADeviceContext &context, int head_num,
   int grid = m;
   int block = k;
 
+  unsigned mask = block < 32 ? (((unsigned)1 << block) - 1) : FINAL_MASK;
   softmax_kernel_with_eltadd<T><<<grid, block, 0, stream>>>(
-      qk_buf_, bias_qk, batch_size, head_num, seq_len);
+      qk_buf_, bias_qk, batch_size, head_num, seq_len, mask);
 }
 
 template <typename T>
@@ -331,7 +339,7 @@ void MultiHeadGPUCompute(const platform::CUDADeviceContext &dev_ctx,
   auto stream = dev_ctx.stream();
 
   int grid = m;
-  PADDLE_ENFORCE_LT(k, 1024,
+  PADDLE_ENFORCE_LE(k, 1024,
                     "Input head_number * size_per_head should <= 1024");
   int block = k <= 1024 ? k : 1024;
   add_QKV<T><<<grid, block, 0, stream>>>(Q, K, V, q_buf, k_buf, v_buf, bias_q,