[Sparse] optimize sparse attention (#44743)

paddle/fluid/platform/dynload/cusparse.h

@@ -56,6 +56,7 @@ CUSPARSE_ROUTINE_EACH(PLATFORM_DECLARE_DYNAMIC_LOAD_CUSPARSE_WRAP)
 
 #if CUDA_VERSION >= 11030
 #define CUSPARSE_ROUTINE_EACH_R2(__macro) \
+  __macro(cusparseSpMM_preprocess);       \
   __macro(cusparseSDDMM_bufferSize);      \
   __macro(cusparseSDDMM_preprocess);      \
   __macro(cusparseSDDMM);

paddle/phi/backends/dynload/cusparse.h

@@ -68,6 +68,7 @@ CUSPARSE_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUSPARSE_WRAP)
 
 #if CUDA_VERSION >= 11030
 #define CUSPARSE_ROUTINE_EACH_R2(__macro) \
+  __macro(cusparseSpMM_preprocess);       \
   __macro(cusparseSDDMM_bufferSize);      \
   __macro(cusparseSDDMM_preprocess);      \
   __macro(cusparseSDDMM);

paddle/phi/kernels/funcs/sparse/sparse_blas_impl.cu.h

@@ -48,6 +48,15 @@ inline cusparseOperation_t GetTransposeOperation(const bool trans) {
   }
 }
 
+inline cusparseSpMMAlg_t GetSpMMAlgorithm(const SparseCsrTensor& x) {
+  // TODO(zhouwei): will change to 'CUSPARSE_SPMM_CSR_ALG2' when support batch
+  return CUSPARSE_SPMM_CSR_ALG2;
+}
+
+inline cusparseSpMMAlg_t GetSpMMAlgorithm(const SparseCooTensor& x) {
+  return CUSPARSE_SPMM_ALG_DEFAULT;
+}
+
 /************* SPARSE MATRIX DESCRIPTOR (COO/CSR) ************/
 
 template <typename T, typename IntT>
@@ -324,7 +333,7 @@ void SparseBlas<phi::GPUContext>::SPMM(bool transa,
                                           &beta,
                                           out_descriptor.descriptor(),
                                           gpu_type,
-                                          CUSPARSE_SPMM_ALG_DEFAULT,
+                                          GetSpMMAlgorithm(mat_a),
                                           &buffer_size);
   });
 
@@ -341,7 +350,7 @@ void SparseBlas<phi::GPUContext>::SPMM(bool transa,
                                &beta,
                                out_descriptor.descriptor(),
                                gpu_type,
-                               CUSPARSE_SPMM_ALG_DEFAULT,
+                               GetSpMMAlgorithm(mat_a),
                                tmp_buffer_ptr);
   });
 }

paddle/phi/kernels/sparse/gpu/fused_attention_grad_kernel.cu

@@ -43,21 +43,14 @@ __global__ void AttnSoftmaxGpuGradKernel(const int64_t* out_crows,
   int row_nnz = static_cast<int>(out_crows[crow_idx + 1] - out_crows[crow_idx]);
   if (row_nnz == 0) return;
 
-  int kIteration = (row_nnz + WARP_SIZE - 1) / WARP_SIZE;
-  T mul_result = 0;
-  for (int i = 0; i < kIteration; ++i) {
-    int idx = threadIdx.x + i * WARP_SIZE;
-    if (idx >= row_nnz) break;
-
-    mul_result += out_values[row_first + idx] * dout_values[row_first + idx];
+  T mul = 0;
+  for (int idx = threadIdx.x; idx < row_nnz; idx += blockDim.x) {
+    mul += out_values[row_first + idx] * dout_values[row_first + idx];
   }
-  T sum = phi::funcs::warpReduceSum<T>(mul_result, 0xFFFFFFFF);
-
-  for (int i = 0; i < kIteration; ++i) {
-    int idx = threadIdx.x + i * WARP_SIZE;
-    if (idx >= row_nnz) break;
+  T mul_sum = phi::funcs::warpReduceSum<T>(mul, 0xFFFFFFFF);
 
-    dx_values[row_first + idx] = (dout_values[row_first + idx] - sum) *
+  for (int idx = threadIdx.x; idx < row_nnz; idx += blockDim.x) {
+    dx_values[row_first + idx] = (dout_values[row_first + idx] - mul_sum) *
                                  out_values[row_first + idx] / scale;
   }
 }
@@ -96,8 +89,8 @@ void FusedAttentionCsrGradKernel(const Context& dev_ctx,
   int N = q_dim[q_rank - 1];
   int batch_nnz = softmax.nnz() / batch_num;
 
-  dim3 grid((total_row_num + 3) / 4);
-  dim3 block(WARP_SIZE, 4);
+  dim3 grid((total_row_num + 7) / 8);
+  dim3 block(WARP_SIZE, 8);
 
   AttnSoftmaxGpuGradKernel<T><<<grid, block, 0, dev_ctx.stream()>>>(
       softmax.non_zero_crows().data<int64_t>(),

paddle/phi/kernels/sparse/gpu/fused_attention_kernel.cu

@@ -26,30 +26,7 @@ limitations under the License. */
 namespace phi {
 namespace sparse {
 
-#define PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, size, HINT, ...) \
-  case size: {                                                 \
-    constexpr int HINT = size;                                 \
-    __VA_ARGS__();                                             \
-    break;                                                     \
-  }
-
-#define VISIT_ATTN_SFOTMAX(SIZE, NAME, ...)                                 \
-  [&] {                                                                     \
-    const auto& __size__ = SIZE;                                            \
-    switch (__size__) {                                                     \
-      PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, 1, KBufferSize, __VA_ARGS__)    \
-      PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, 2, KBufferSize, __VA_ARGS__)    \
-      PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, 3, KBufferSize, __VA_ARGS__)    \
-      PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, 4, KBufferSize, __VA_ARGS__)    \
-      PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, 8, KBufferSize, __VA_ARGS__)    \
-      PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, 12, KBufferSize, __VA_ARGS__)   \
-      PRIVATE_CASE_VISIT_ATTN_SOFTMAX(NAME, 16, KBufferSize, __VA_ARGS__)   \
-      default:                                                              \
-        PD_THROW("function " #NAME " is not implemented for columns>512 "); \
-    }                                                                       \
-  }()
-
-template <typename T, int BufferSize>
+template <typename T>
 __global__ void AttnSoftmaxGpuKernel(const int64_t* x_crows,
                                      const int64_t* x_cols,
                                      const T* x_values,
@@ -58,7 +35,6 @@ __global__ void AttnSoftmaxGpuKernel(const int64_t* x_crows,
                                      T* out_values,
                                      int M,
                                      int total_row_num,
-                                     float scale,
                                      int num_heads,
                                      int batch_nnz) {
   // out = exp(x-x_max) / sum(exp(x-x_max))
@@ -72,17 +48,10 @@ __global__ void AttnSoftmaxGpuKernel(const int64_t* x_crows,
   int row_nnz = static_cast<int>(x_crows[crow_idx + 1] - x_crows[crow_idx]);
   if (row_nnz == 0) return;
 
-  T buffer[BufferSize] = {0};
-  int kIteration = (row_nnz + WARP_SIZE - 1) / WARP_SIZE;
-
   T max_val = -std::numeric_limits<T>::infinity();
-  for (int i = 0; i < kIteration; ++i) {
+  for (int idx = threadIdx.x; idx < row_nnz; idx += blockDim.x) {
     bool mask = false;
-    int idx = threadIdx.x + i * WARP_SIZE;
-    if (idx >= row_nnz) break;
-
     int col_idx = static_cast<int>(x_cols[row_first + idx]);
-
     if (kp_mask != nullptr &&
         kp_mask[(cur_batch / num_heads) * M + col_idx] == 0) {
       mask = true;
@@ -92,37 +61,30 @@ __global__ void AttnSoftmaxGpuKernel(const int64_t* x_crows,
     }
 
     if (!mask) {
-      buffer[i] = x_values[row_first + idx] / scale;
-      if (buffer[i] > max_val) {
-        max_val = buffer[i];
+      T val = x_values[row_first + idx];
+      if (val > max_val) {
+        max_val = val;
       }
+      out_values[row_first + idx] = val;
+    } else {
+      // Note corner case: when all elements of the row are masked, result
+      // may be wrong because of exp('-inf' - '-inf'), just ignore now.
+      out_values[row_first + idx] = -std::numeric_limits<T>::infinity();
     }
   }
   T row_max_val = phi::funcs::warpReduceMax<T>(max_val, 0xFFFFFFFF);
 
-  auto functor = phi::funcs::CudaExpFunctor<T>();
   T exp_sum = 0;
-  for (int i = 0; i < kIteration; ++i) {
-    int idx = threadIdx.x + i * WARP_SIZE;
-    if (idx >= row_nnz) break;
-
-    if (buffer[i]) {
-      T exp = functor(buffer[i] - row_max_val);
+  for (int idx = threadIdx.x; idx < row_nnz; idx += blockDim.x) {
+    auto functor = phi::funcs::CudaExpFunctor<T>();
+    T exp = functor(out_values[row_first + idx] - row_max_val);
     exp_sum += exp;
-      buffer[i] = exp;
-    }
+    out_values[row_first + idx] = exp;
   }
   T row_exp_sum = phi::funcs::warpReduceSum<T>(exp_sum, 0xFFFFFFFF);
 
-  for (int i = 0; i < kIteration; ++i) {
-    int idx = threadIdx.x + i * WARP_SIZE;
-    if (idx >= row_nnz) break;
-
-    if (buffer[i]) {
-      out_values[row_first + idx] = buffer[i] / row_exp_sum;
-    } else {
-      out_values[row_first + idx] = static_cast<T>(0);
-    }
+  for (int idx = threadIdx.x; idx < row_nnz; idx += blockDim.x) {
+    out_values[row_first + idx] = out_values[row_first + idx] / row_exp_sum;
   }
 }
 
@@ -219,35 +181,25 @@ void FusedAttentionCsrKernel(
                           "shape of 'attn_mask' must be [seq_len, seq_len]"));
   }
 
-  /* Step1: SDD Matmul, reuse */
+  /* Step1: SDD Matmul, reuse matmul */
   SparseCsrTensor sdd_result;
   EmptyLikeCsrKernel<T, Context>(dev_ctx, sparse_mask, &sdd_result);
   auto sparse_blas = phi::funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
   sparse_blas.SDDMM(false,
                     true,
-                    static_cast<T>(1),
+                    static_cast<T>(1 / std::sqrt(N)),
                     query,
                     key,
                     static_cast<T>(0),
                     &sdd_result);
 
-  /* Step2: Softmax with kp_mask/attn_mask, manualy not reuse */
   EmptyLikeCsrKernel<T, Context>(dev_ctx, sdd_result, softmax);
 
-  int buffer_size;
-  if (M < 128) {
-    buffer_size = (M + 32 - 1) / 32;
-  } else {
-    buffer_size = ((M + 128 - 1) / 128) * 4;
-  }
-
-  dim3 grid((total_row_num + 3) / 4);
-  dim3 block(WARP_SIZE, 4);
+  dim3 grid((total_row_num + 7) / 8);
+  dim3 block(WARP_SIZE, 8);
 
   int batch_nnz = sdd_result.nnz() / batch_num;
-
-  VISIT_ATTN_SFOTMAX(buffer_size, "AttnSoftmaxGpuKernel", [&] {
-    AttnSoftmaxGpuKernel<T, KBufferSize><<<grid, block, 0, dev_ctx.stream()>>>(
+  AttnSoftmaxGpuKernel<T><<<grid, block, 0, dev_ctx.stream()>>>(
       sdd_result.non_zero_crows().data<int64_t>(),
       sdd_result.non_zero_cols().data<int64_t>(),
       sdd_result.non_zero_elements().data<T>(),
@@ -256,12 +208,9 @@ void FusedAttentionCsrKernel(
       softmax->mutable_non_zero_elements()->data<T>(),
       M,
       total_row_num,
-        std::sqrt(N),
       q_dim[1],
       batch_nnz);
-  });
 
-  /* Step3: DSD Matmul, reuse */
   softmax->set_dims(phi::make_ddim({q_dim[0], q_dim[1], q_dim[2], q_dim[2]}));
   MatmulCsrDenseKernel<T, Context>(dev_ctx, *softmax, value, out);
 #else

python/paddle/fluid/tests/unittests/test_sparse_fused_attention_op.py

@@ -37,7 +37,7 @@ def get_cuda_version():
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda() or get_cuda_version() < 11070,
-    "core is not compiled with CUDA and cuda version need larger than or equal to 11.3"
+    "core is not compiled with CUDA and cuda version need larger than or equal to 11.7"
 )
 class TestSparseAttentionAPI1(unittest.TestCase):