Merge branch 'release/2.2' into paddle2.2_cherrypick

cmake/operators.cmake

@@ -185,6 +185,7 @@ function(op_library TARGET)
         list(REMOVE_ITEM hip_srcs "cholesky_op.cu")
         list(REMOVE_ITEM hip_srcs "matrix_rank_op.cu")
         list(REMOVE_ITEM hip_srcs "svd_op.cu")
+        list(REMOVE_ITEM hip_srcs "qr_op.cu")
         list(REMOVE_ITEM hip_srcs "eigh_op.cu")
         list(REMOVE_ITEM hip_srcs "multinomial_op.cu")
         list(REMOVE_ITEM hip_srcs "decode_jpeg_op.cu")
@@ -216,7 +217,7 @@ function(op_library TARGET)
 "fusion_transpose_flatten_concat_op" "fusion_conv_inception_op"
 "sync_batch_norm_op" "sparse_attention_op"  "dgc_op" "fused_fc_elementwise_layernorm_op"
 "skip_layernorm_op" "multihead_matmul_op" "fusion_group_op" "fused_bn_activation_op" "fused_embedding_eltwise_layernorm_op" "fusion_gru_op" "fusion_lstm_op"
-"fused_bn_add_activation_op")
+"fused_bn_add_activation_op" "fused_attention_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()

paddle/fluid/framework/channel.h

@@ -157,7 +157,19 @@ class ChannelObject {
     p.resize(finished);
     return finished;
   }
+  // read once only
+  size_t ReadOnce(std::vector<T>& p, size_t size) {  // NOLINT
+    if (size == 0) {
+      return 0;
+    }
+    std::unique_lock<std::mutex> lock(mutex_);
+    p.resize(size);
+    size_t finished = Read(size, &p[0], lock, true);
+    p.resize(finished);
+    Notify();
 
+    return finished;
+  }
   size_t ReadAll(std::vector<T>& p) {  // NOLINT
     p.clear();
     size_t finished = 0;
@@ -241,17 +253,21 @@ class ChannelObject {
     return !closed_;
   }
 
-  size_t Read(size_t n, T* p, std::unique_lock<std::mutex>& lock) {  // NOLINT
+  size_t Read(size_t n, T* p, std::unique_lock<std::mutex>& lock,  // NOLINT
+              bool once = false) {                                 // NOLINT
     size_t finished = 0;
     CHECK(n <= MaxCapacity() - reading_count_);
     reading_count_ += n;
     while (finished < n && WaitForRead(lock)) {
-      size_t m = std::min(n - finished, data_.size());
+      size_t m = (std::min)(n - finished, data_.size());
       for (size_t i = 0; i < m; i++) {
         p[finished++] = std::move(data_.front());
         data_.pop_front();
       }
       reading_count_ -= m;
+      if (once && m > 0) {
+        break;
+      }
     }
     reading_count_ -= n - finished;
     return finished;