fix embedding_with_eltwise_add_xpu (#55354)

paddle/phi/kernels/fusion/xpu/embedding_with_eltwise_add_xpu_kernel.cc

@@ -19,9 +19,56 @@
 namespace phi {
 namespace fusion {
 
-template <typename T, typename Context>
-void EmbeddingWithEltwiseAddXpuKernel(
-    const Context& ctx,
+namespace {
+template <typename T>
+void FillSeqLod(int batch_size, int max_seq_len, const T* mask, int* seq_lod) {
+  for (int batch_idx = 0; batch_idx < batch_size; batch_idx++) {
+    int cur_batch_seq_len = 0;
+    for (int seq_idx = 0; seq_idx < max_seq_len; seq_idx++) {
+      int mask_idx = batch_idx * max_seq_len + seq_idx;
+      if (mask[mask_idx] > 0) {
+        cur_batch_seq_len++;
+      } else {
+        break;
+      }
+    }
+    PADDLE_ENFORCE_GT(
+        cur_batch_seq_len,
+        0,
+        errors::PreconditionNotMet(
+            "cur_batch_seq_len should be greater than 0, but got %d.",
+            cur_batch_seq_len));
+    seq_lod[batch_idx + 1] = seq_lod[batch_idx] + cur_batch_seq_len;
+  }
+}
+
+template <>
+void FillSeqLod<float>(int batch_size,
+                       int max_seq_len,
+                       const float* mask,
+                       int* seq_lod) {
+  for (int batch_idx = 0; batch_idx < batch_size; batch_idx++) {
+    int cur_batch_seq_len = 0;
+    for (int seq_idx = 0; seq_idx < max_seq_len; seq_idx++) {
+      int mask_idx = batch_idx * max_seq_len + seq_idx;
+      if (fabs(mask[mask_idx]) > 1e-5) {
+        cur_batch_seq_len++;
+      } else {
+        break;
+      }
+    }
+    PADDLE_ENFORCE_GT(
+        cur_batch_seq_len,
+        0,
+        errors::PreconditionNotMet(
+            "cur_batch_seq_len should be greater than 0, but got %d.",
+            cur_batch_seq_len));
+    seq_lod[batch_idx + 1] = seq_lod[batch_idx] + cur_batch_seq_len;
+  }
+}
+
+template <typename TT, typename TID, typename Context>
+void MultiEmbeddingKernel(const Context& ctx,
                           const std::vector<const DenseTensor*>& ids,
                           const std::vector<const DenseTensor*>& tables,
                           const paddle::optional<DenseTensor>& mask,
@@ -29,10 +76,10 @@ void EmbeddingWithEltwiseAddXpuKernel(
                           DenseTensor* out,
                           DenseTensor* seq_lod,
                           DenseTensor* max_seq_len) {
-  using XPUType = typename XPUTypeTrait<T>::Type;
-  int emb_dim = tables[0]->dims()[1];
-  std::vector<int> table_lens;
-  std::vector<const float*> arg_tables;
+  using XPUType = typename XPUTypeTrait<TT>::Type;
+  int64_t emb_dim = tables[0]->dims()[1];
+  std::vector<TID> table_lens;
+  std::vector<const XPUType*> arg_tables;
   for (auto* table : tables) {
     auto& table_dims = table->dims();
     PADDLE_ENFORCE_EQ(
@@ -49,20 +96,7 @@ void EmbeddingWithEltwiseAddXpuKernel(
             table_dims[1],
             emb_dim));
     table_lens.push_back(table_dims[0]);
-    if (std::is_same<T, phi::dtype::float16>::value) {
-      DenseTensor table_data_fp32_t;
-      ctx.template Alloc<float>(&table_data_fp32_t,
-                                table->numel() * sizeof(float));
-      int r = xpu::cast<XPUType, float>(
-          ctx.x_context(),
-          reinterpret_cast<const XPUType*>(table->data<T>()),
-          table_data_fp32_t.data<float>(),
-          table->numel());
-      PADDLE_ENFORCE_XDNN_SUCCESS(r, "cast");
-      arg_tables.push_back(table_data_fp32_t.data<float>());
-    } else {
-      arg_tables.push_back(table->data<float>());
-    }
+    arg_tables.push_back(reinterpret_cast<const XPUType*>(table->data<TT>()));
   }
 
   int emb_layer_num = ids.size();
@@ -78,118 +112,81 @@ void EmbeddingWithEltwiseAddXpuKernel(
   auto& id_dims = ids[0]->dims();
   int batch_size = id_dims[0];
   int max_seq_len_value = id_dims[1];
-  int ids_len = id_dims[0] * id_dims[1];
-  std::vector<std::vector<int>> int_ids(emb_layer_num,
-                                        std::vector<int>(ids_len, 0));
-  std::vector<xpu::VectorParam<int>> arg_ids;
   auto* mask_tensor = mask.get_ptr();
   if (mask_tensor != nullptr) {
-    auto mask_dtype = mask_tensor->dtype();
-    PADDLE_ENFORCE(
-        mask_dtype == phi::DataType::INT64 ||
-            mask_dtype == phi::DataType::FLOAT32,
-        errors::InvalidArgument(
-            "The data type of mask should be int64 or float32, but got %s.",
-            DataTypeToString(mask_dtype)));
     max_seq_len->Resize({1});
     ctx.template HostAlloc<int>(max_seq_len)[0] = max_seq_len_value;
 
     seq_lod->Resize({batch_size + 1});
     int* seq_lod_data = ctx.template HostAlloc<int>(seq_lod);
     seq_lod_data[0] = 0;
-    for (int batch_idx = 0; batch_idx < batch_size; batch_idx++) {
-      int cur_batch_seq_len = 0;
-      for (int seq_idx = 0; seq_idx < max_seq_len_value; seq_idx++) {
-        int mask_idx = batch_idx * max_seq_len_value + seq_idx;
-        if ((mask_dtype == phi::DataType::INT64 &&
-             mask->data<int64_t>()[mask_idx] > 0) ||
-            (mask_dtype == phi::DataType::FLOAT32 &&
-             fabs(mask->data<float>()[mask_idx]) > 1e-5)) {
-          cur_batch_seq_len++;
-        } else {
+    switch (mask_tensor->dtype()) {
+      case DataType::FLOAT32:
+        FillSeqLod(batch_size,
+                   max_seq_len_value,
+                   mask_tensor->data<float>(),
+                   seq_lod_data);
+        break;
+      case DataType::INT64:
+        FillSeqLod(batch_size,
+                   max_seq_len_value,
+                   mask_tensor->data<int64_t>(),
+                   seq_lod_data);
+        break;
+      default:
+        PADDLE_THROW(
+            phi::errors::Unimplemented("Only support mask data type is int64 "
+                                       "or float, not support %s now.",
+                                       DataTypeToString(mask_tensor->dtype())));
         break;
-        }
-      }
-      PADDLE_ENFORCE_GT(
-          cur_batch_seq_len,
-          0,
-          errors::PreconditionNotMet(
-              "cur_batch_seq_len should be greater than 0, but got %d.",
-              cur_batch_seq_len));
-      seq_lod_data[batch_idx + 1] = seq_lod_data[batch_idx] + cur_batch_seq_len;
     }
     out->Resize({batch_size, seq_lod_data[batch_size], emb_dim});
-
-    for (int i = 0; i < emb_layer_num; i++) {
-      if (ids[i]->dtype() == DataType::INT64) {
-        auto* ids_data = ids[i]->data<int64_t>();
-        for (int batch_idx = 0; batch_idx < batch_size; batch_idx++) {
-          for (int j = 0;
-               j < seq_lod_data[batch_idx + 1] - seq_lod_data[batch_idx];
-               j++) {
-            int_ids[i][seq_lod_data[batch_idx] + j] =
-                ids_data[batch_idx * max_seq_len_value + j];
-          }
   }
-      } else {
-        auto* ids_data = ids[i]->data<int>();
-        for (int batch_idx = 0; batch_idx < batch_size; batch_idx++) {
-          for (int j = 0;
-               j < seq_lod_data[batch_idx + 1] - seq_lod_data[batch_idx];
-               j++) {
-            int_ids[i][seq_lod_data[batch_idx] + j] =
-                ids_data[batch_idx * max_seq_len_value + j];
-          }
-        }
-      }
-      arg_ids.push_back(
-          xpu::VectorParam<int>{int_ids[i].data(), ids_len, nullptr});
-    }
-  } else {
+
+  int ids_len = id_dims[0] * id_dims[1];
+  std::vector<xpu::VectorParam<TID>> arg_ids;
   for (int i = 0; i < emb_layer_num; i++) {
-      for (int j = 0; j < ids_len; j++) {
-        if (ids[i]->dtype() == phi::DataType::INT64) {
-          int_ids[i][j] = static_cast<int>(ids[i]->data<int64_t>()[j]);
-        } else if (ids[i]->dtype() == phi::DataType::INT32) {
-          int_ids[i][j] = ids[i]->data<int>()[j];
-        }
-      }
     arg_ids.push_back(
-          xpu::VectorParam<int>{int_ids[i].data(), ids_len, nullptr});
-    }
+        xpu::VectorParam<TID>{ids[i]->data<TID>(), ids_len, nullptr});
   }
 
-  ctx.template Alloc<T>(out);
-  if (std::is_same<T, phi::dtype::float16>::value) {
-    DenseTensor out_fp32_t;
-    ctx.template Alloc<float>(&out_fp32_t, out->numel() * sizeof(float));
-    int r = xpu::multi_embedding_fusion<float, float, int>(
+  int r = xpu::multi_embedding_fusion<XPUType, XPUType, TID>(
       ctx.x_context(),
-        arg_tables, /* tables */
-        out_fp32_t.data<float>(),
+      arg_tables,
+      reinterpret_cast<XPUType*>(ctx.template Alloc<TT>(out)),
       arg_ids,
       table_lens,
       emb_dim,
       std::vector<float>(table_lens.size(), 1.0f),
-        std::vector<int>(table_lens.size(), padding_idx));
+      std::vector<TID>(table_lens.size(), padding_idx));
   PADDLE_ENFORCE_XDNN_SUCCESS(r, "embedding_with_eltwise_add_xpu");
+}
+}  // namespace
 
-    r = xpu::cast(ctx.x_context(),
-                  out_fp32_t.data<float>(),
-                  reinterpret_cast<XPUTypeFP16*>(out->data<T>()),
-                  out->numel());
-    PADDLE_ENFORCE_XDNN_SUCCESS(r, "cast");
-  } else {
-    int r = xpu::multi_embedding_fusion<float, float, int>(
-        ctx.x_context(),
-        arg_tables, /* tables */
-        out->data<float>(),
-        arg_ids,
-        table_lens,
-        emb_dim,
-        std::vector<float>(table_lens.size(), 1.0f),
-        std::vector<int>(table_lens.size(), padding_idx));
-    PADDLE_ENFORCE_XDNN_SUCCESS(r, "embedding_with_eltwise_add_xpu");
+template <typename T, typename Context>
+void EmbeddingWithEltwiseAddXpuKernel(
+    const Context& ctx,
+    const std::vector<const DenseTensor*>& ids,
+    const std::vector<const DenseTensor*>& tables,
+    const paddle::optional<DenseTensor>& mask,
+    int64_t padding_idx,
+    DenseTensor* out,
+    DenseTensor* seq_lod,
+    DenseTensor* max_seq_len) {
+  switch (ids[0]->dtype()) {
+    case DataType::INT32:
+      MultiEmbeddingKernel<T, int, Context>(
+          ctx, ids, tables, mask, padding_idx, out, seq_lod, max_seq_len);
+      break;
+    case DataType::INT64:
+      MultiEmbeddingKernel<T, int64_t, Context>(
+          ctx, ids, tables, mask, padding_idx, out, seq_lod, max_seq_len);
+      break;
+    default:
+      PADDLE_THROW(phi::errors::Unimplemented(
+          "Only support ids data type is int64 or int32, not support %s now.",
+          DataTypeToString(ids[0]->dtype())));
+      break;
   }
 }