[PHI decoupling] move softmax from fluid to phi and remove cpu_vec.h …

…in fluid (#48970)

paddle/fluid/inference/api/analysis_config.cc

@@ -19,9 +19,9 @@
 #include "paddle/fluid/inference/api/paddle_analysis_config.h"
 #include "paddle/fluid/inference/api/paddle_pass_builder.h"
 #include "paddle/fluid/inference/utils/table_printer.h"
-#include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/device/gpu/gpu_info.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 #include "paddle/utils/string/split.h"
 
 #ifdef PADDLE_WITH_TENSORRT
@@ -624,10 +624,11 @@ void AnalysisConfig::EnableMkldnnQuantizer() {
 
 void AnalysisConfig::EnableMkldnnBfloat16() {
 #ifdef PADDLE_WITH_MKLDNN
-  if (platform::MayIUse(platform::cpu_isa_t::avx512_core)) {
+  if (phi::backends::cpu::MayIUse(phi::backends::cpu::cpu_isa_t::avx512_core)) {
     use_mkldnn_bfloat16_ = true;
     LOG(INFO) << "Hardware support for BFLOAT16"
-              << (platform::MayIUse(platform::cpu_isa_t::avx512_bf16)
+              << (phi::backends::cpu::MayIUse(
+                      phi::backends::cpu::cpu_isa_t::avx512_bf16)
                       ? " is enabled"
                       : " is disabled. Simulation will be used");
   } else {

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -29,7 +29,7 @@
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/tests/api/tester_helper.h"
 #include "paddle/fluid/inference/utils/io_utils.h"
-#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 
 DEFINE_string(dirname, "", "dirname to tests.");
 
@@ -327,7 +327,7 @@ TEST(AnalysisPredictor, bf16_gpu_pass_strategy) {
   config.EnableUseGpu(100, 0);
   config.EnableMkldnnBfloat16();
 #ifdef PADDLE_WITH_MKLDNN
-  if (platform::MayIUse(platform::cpu_isa_t::avx512_core))
+  if (phi::backends::cpu::MayIUse(phi::backends::cpu::cpu_isa_t::avx512_core))
     ASSERT_EQ(config.mkldnn_bfloat16_enabled(), true);
   else
     ASSERT_EQ(config.mkldnn_bfloat16_enabled(), false);

paddle/fluid/inference/api/onnxruntime_predictor_tester.cc

@@ -27,7 +27,7 @@
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/tests/api/tester_helper.h"
 #include "paddle/fluid/inference/utils/io_utils.h"
-#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 
 DEFINE_string(dirname, "", "dirname to tests.");
 

paddle/fluid/inference/tests/api/analyzer_bfloat16_image_classification_tester.cc

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/inference/api/paddle_analysis_config.h"
 #include "paddle/fluid/inference/tests/api/tester_helper.h"
-#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 
 DEFINE_bool(enable_mkldnn, true, "Enable MKLDNN");
 
@@ -47,7 +47,7 @@ TEST(Analyzer_bfloat16_image_classification, bfloat16) {
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInputs(&input_slots_all);
   if (FLAGS_enable_mkldnn && FLAGS_enable_bf16 &&
-      platform::MayIUse(platform::cpu_isa_t::avx512_bf16)) {
+      phi::backends::cpu::MayIUse(phi::backends::cpu::cpu_isa_t::avx512_bf16)) {
     b_cfg.EnableMkldnnBfloat16();
   } else {
     FLAGS_enable_bf16 = false;

paddle/fluid/memory/allocation/buddy_allocator.h

@@ -27,9 +27,9 @@ limitations under the License. */
 
 #include "paddle/fluid/memory/allocation/memory_block.h"
 #include "paddle/fluid/memory/allocation/system_allocator.h"
-#include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/device/gpu/gpu_info.h"
 #include "paddle/fluid/platform/device/npu/npu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 
 namespace paddle {
 namespace memory {

paddle/fluid/memory/allocation/naive_best_fit_allocator.cc

@@ -78,8 +78,8 @@ BuddyAllocator *GetCPUBuddyAllocator() {
   std::call_once(init_flag, []() {
     a = new detail::BuddyAllocator(
         std::unique_ptr<detail::SystemAllocator>(new detail::CPUAllocator),
-        platform::CpuMinChunkSize(),
-        platform::CpuMaxChunkSize());
+        phi::backends::cpu::CpuMinChunkSize(),
+        phi::backends::cpu::CpuMaxChunkSize());
   });
 
   return a;
@@ -290,8 +290,8 @@ BuddyAllocator *GetNPUPinnedBuddyAllocator() {
   std::call_once(init_flag, []() {
     ba = new BuddyAllocator(std::unique_ptr<detail::SystemAllocator>(
                                 new detail::NPUPinnedAllocator),
-                            platform::NPUPinnedMinChunkSize(),
-                            platform::NPUPinnedMaxChunkSize());
+                            phi::backends::cpu::NPUPinnedMinChunkSize(),
+                            phi::backends::cpu::NPUPinnedMaxChunkSize());
   });
 
   return ba;
@@ -562,8 +562,8 @@ BuddyAllocator *GetCUDAPinnedBuddyAllocator() {
   std::call_once(init_flag, []() {
     ba = new BuddyAllocator(std::unique_ptr<detail::SystemAllocator>(
                                 new detail::CUDAPinnedAllocator),
-                            platform::CUDAPinnedMinChunkSize(),
-                            platform::CUDAPinnedMaxChunkSize());
+                            phi::backends::cpu::CUDAPinnedMinChunkSize(),
+                            phi::backends::cpu::CUDAPinnedMaxChunkSize());
   });
 
   return ba;

paddle/fluid/memory/allocation/system_allocator.cc

@@ -28,10 +28,10 @@ limitations under the License. */
 #endif
 #include "gflags/gflags.h"
 #include "paddle/fluid/memory/allocation/allocator.h"
-#include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/device/gpu/gpu_info.h"
 #include "paddle/fluid/platform/device/npu/npu_info.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 #ifdef PADDLE_WITH_MLU
 #include "paddle/fluid/platform/device/mlu/mlu_info.h"
 #endif
@@ -206,7 +206,7 @@ void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) {
   // of host pinned allocation. Allocates too much would reduce
   // the amount of memory available to the underlying system for paging.
   size_t usable =
-      paddle::platform::CUDAPinnedMaxAllocSize() - cuda_pinnd_alloc_size_;
+      phi::backends::cpu::CUDAPinnedMaxAllocSize() - cuda_pinnd_alloc_size_;
 
   if (size > usable) {
     LOG(WARNING) << "Cannot malloc " << size / 1024.0 / 1024.0
@@ -362,7 +362,7 @@ void* NPUPinnedAllocator::Alloc(size_t* index, size_t size) {
   if (size <= 0) return nullptr;
 
   size_t usable =
-      paddle::platform::NPUPinnedMaxAllocSize() - npu_pinnd_alloc_size_;
+      phi::backends::cpu::NPUPinnedMaxAllocSize() - npu_pinnd_alloc_size_;
 
   if (size > usable) {
     LOG(WARNING) << "Cannot malloc " << size / 1024.0 / 1024.0

paddle/fluid/memory/pinned_memory_test.cu

@@ -18,9 +18,9 @@ limitations under the License. */
 #include "paddle/fluid/memory/allocation/memory_block.h"
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/memory/memory.h"
-#include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/device/gpu/gpu_info.h"
 #include "paddle/fluid/platform/place.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 
 // This unit test is an example comparing the performance between using pinned
 // memory and not. In general, using pinned memory will be faster.

paddle/fluid/operators/attention_lstm_op.cc

@@ -16,7 +16,7 @@ limitations under the License. */
 
 #include <string>
 
-#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 #include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/cpu_vec.h"
 #include "paddle/phi/kernels/funcs/fc_functor.h"
@@ -315,10 +315,10 @@ use lstm_x_t as input and compute as standard LSTM.
 template <typename T>
 inline void bias_relu(const int n, const T* x, const T* bias, T* y) {
   if (bias) {
-    phi::funcs::vec_add_bias<T, platform::avx>(n, *bias, x, y);
-    phi::funcs::vec_relu<T, platform::avx>(n, y, y);
+    phi::funcs::vec_add_bias<T, phi::backends::cpu::avx>(n, *bias, x, y);
+    phi::funcs::vec_relu<T, phi::backends::cpu::avx>(n, y, y);
   } else {
-    phi::funcs::vec_relu<T, platform::avx>(n, x, y);
+    phi::funcs::vec_relu<T, phi::backends::cpu::avx>(n, x, y);
   }
 }
 
@@ -329,8 +329,9 @@ inline void vec_softmax(const int n, const T* x, T* y) {
   for (int i = 1; i < n; ++i) {
     scalar = scalar < x[i] ? x[i] : scalar;
   }
-  phi::funcs::vec_add_bias<T, platform::avx>(n, -scalar, x, y);  // sub
-  phi::funcs::vec_exp<T>(n, y, y);                               // exp
+  phi::funcs::vec_add_bias<T, phi::backends::cpu::avx>(
+      n, -scalar, x, y);            // sub
+  phi::funcs::vec_exp<T>(n, y, y);  // exp
   // sum
   scalar = T(0);
   for (int i = 0; i < n; ++i) {
@@ -393,13 +394,13 @@ class AttentionLSTMKernel : public framework::OpKernel<T> {
     auto& act_gate_str = ctx.Attr<std::string>("gate_activation");
     auto& act_cell_str = ctx.Attr<std::string>("cell_activation");
     auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");
-    if (platform::MayIUse(platform::avx)) {
-      phi::funcs::VecActivations<T, platform::avx> act_functor;
+    if (phi::backends::cpu::MayIUse(phi::backends::cpu::avx)) {
+      phi::funcs::VecActivations<T, phi::backends::cpu::avx> act_functor;
       act_gate = act_functor(act_gate_str);
       act_cell = act_functor(act_cell_str);
       act_cand = act_functor(act_cand_str);
     } else {
-      phi::funcs::VecActivations<T, platform::isa_any> act_functor;
+      phi::funcs::VecActivations<T, phi::backends::cpu::isa_any> act_functor;
       act_gate = act_functor(act_gate_str);
       act_cell = act_functor(act_cell_str);
       act_cand = act_functor(act_cand_str);

paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.cu

@@ -13,13 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.h"
-#include "paddle/fluid/operators/math/softmax_impl.h"
+
+#include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/device/gpu/nccl_helper.h"
 #include "paddle/fluid/string/string_helper.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/kernels/funcs/axis_utils.h"
 #include "paddle/phi/kernels/funcs/cross_entropy.h"
+#include "paddle/phi/kernels/funcs/softmax_impl.h"
 
 namespace paddle {
 namespace operators {
@@ -129,15 +131,15 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
     softmax_2d.ShareDataWith(*softmax).Resize({N, D});
     loss_2d.ShareDataWith(*loss).Resize({N, 1});
 
-    auto eigen_logits = math::EigenMatrix<T>::From(logits_2d);
-    auto eigen_softmax = math::EigenMatrix<T>::From(softmax_2d);
+    auto eigen_logits = phi::funcs::EigenMatrix<T>::From(logits_2d);
+    auto eigen_softmax = phi::funcs::EigenMatrix<T>::From(softmax_2d);
 
     // step 1, obtain logit_max
     phi::DenseTensor logits_max;
     logits_max = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);
     void* logits_max_buff = logits_max.mutable_data<T>(place);
 
-    auto eigen_logits_max = math::EigenMatrix<T>::From(logits_max);
+    auto eigen_logits_max = phi::funcs::EigenMatrix<T>::From(logits_max);
     Eigen::DSizes<int, 1> along_axis(1);
     eigen_logits_max.device(*dev_ctx.eigen_device()) =
         eigen_logits.maximum(along_axis);
@@ -158,7 +160,7 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
     eigen_softmax.device(*dev_ctx.eigen_device()) =
         (eigen_logits -
          eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class))
-            .unaryExpr(math::ValueClip<T>());
+            .unaryExpr(phi::funcs::ValueClip<T>());
 
     // step 3, obtain predict target
     phi::DenseTensor predicted_logits;
@@ -217,7 +219,8 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
     sum_exp_logits = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);
     void* sum_exp_logits_buff = sum_exp_logits.mutable_data<T>(place);
 
-    auto eigen_sum_exp_logits = math::EigenMatrix<T>::From(sum_exp_logits);
+    auto eigen_sum_exp_logits =
+        phi::funcs::EigenMatrix<T>::From(sum_exp_logits);
     eigen_sum_exp_logits.device(*dev_ctx.eigen_device()) =
         eigen_softmax.sum(along_axis);
 
@@ -231,8 +234,9 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
         comm->comm(),
         stream));
 
-    auto eigen_loss = math::EigenMatrix<T>::From(loss_2d);
-    auto eigen_predicted_logits = math::EigenMatrix<T>::From(predicted_logits);
+    auto eigen_loss = phi::funcs::EigenMatrix<T>::From(loss_2d);
+    auto eigen_predicted_logits =
+        phi::funcs::EigenMatrix<T>::From(predicted_logits);
 
     eigen_loss.device(*dev_ctx.eigen_device()) =
         (eigen_sum_exp_logits.log().unaryExpr(phi::funcs::TolerableValue<T>()) -
@@ -281,14 +285,14 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
     softmax_2d.ShareDataWith(*softmax).Resize({N, D});
     loss_2d.ShareDataWith(*loss).Resize({N, 1});
 
-    auto eigen_logits = math::EigenMatrix<T>::From(logits_2d);
-    auto eigen_softmax = math::EigenMatrix<T>::From(softmax_2d);
+    auto eigen_logits = phi::funcs::EigenMatrix<T>::From(logits_2d);
+    auto eigen_softmax = phi::funcs::EigenMatrix<T>::From(softmax_2d);
 
     // step 1, obtain logit_max
     phi::DenseTensor logits_max;
     logits_max = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);
 
-    auto eigen_logits_max = math::EigenMatrix<T>::From(logits_max);
+    auto eigen_logits_max = phi::funcs::EigenMatrix<T>::From(logits_max);
     Eigen::DSizes<int, 1> along_axis(1);
     eigen_logits_max.device(*dev_ctx.eigen_device()) =
         eigen_logits.maximum(along_axis);
@@ -304,7 +308,7 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
     eigen_softmax.device(*dev_ctx.eigen_device()) =
         (eigen_logits -
          eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class))
-            .unaryExpr(math::ValueClip<T>());
+            .unaryExpr(phi::funcs::ValueClip<T>());
 
     // step 3, obtain predict target
     phi::DenseTensor predicted_logits;
@@ -357,7 +361,8 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
     sum_exp_logits = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);
     void* sum_exp_logits_buff = sum_exp_logits.mutable_data<T>(place);
 
-    auto eigen_sum_exp_logits = math::EigenMatrix<T>::From(sum_exp_logits);
+    auto eigen_sum_exp_logits =
+        phi::funcs::EigenMatrix<T>::From(sum_exp_logits);
     eigen_sum_exp_logits.device(*dev_ctx.eigen_device()) =
         eigen_softmax.sum(along_axis);
 
@@ -366,8 +371,9 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
     opts.reduce_op = distributed::ReduceOp::SUM;
     pg->AllReduce(in_out, in_out, opts)->Synchronize();
 
-    auto eigen_loss = math::EigenMatrix<T>::From(loss_2d);
-    auto eigen_predicted_logits = math::EigenMatrix<T>::From(predicted_logits);
+    auto eigen_loss = phi::funcs::EigenMatrix<T>::From(loss_2d);
+    auto eigen_predicted_logits =
+        phi::funcs::EigenMatrix<T>::From(predicted_logits);
 
     eigen_loss.device(*dev_ctx.eigen_device()) =
         (eigen_sum_exp_logits.log().unaryExpr(phi::funcs::TolerableValue<T>()) -

paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.h

@@ -22,9 +22,9 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/softmax.h"
 #include "paddle/phi/api/include/tensor.h"
 #include "paddle/phi/kernels/funcs/cross_entropy.h"
+#include "paddle/phi/kernels/funcs/softmax.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/elementwise/elementwise_mul_op.h

@@ -17,7 +17,7 @@ limitations under the License. */
 #include <string>
 
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
-#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 #include "paddle/phi/kernels/elementwise_kernel.h"
 
 namespace paddle {

paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc

@@ -16,7 +16,7 @@ limitations under the License. */
 
 #include <string>
 
-#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 #include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/cpu_vec.h"
 #include "paddle/phi/kernels/funcs/sequence2batch.h"
@@ -278,13 +278,13 @@ class FusedEmbeddingFCLSTMKernel : public framework::OpKernel<T> {
   auto& act_gate_str = ctx.Attr<std::string>("gate_activation");             \
   auto& act_cell_str = ctx.Attr<std::string>("cell_activation");             \
   auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");        \
-  if (platform::MayIUse(platform::avx)) {                                    \
-    phi::funcs::VecActivations<T, platform::avx> act_functor;                \
+  if (phi::backends::cpu::MayIUse(phi::backends::cpu::avx)) {                \
+    phi::funcs::VecActivations<T, phi::backends::cpu::avx> act_functor;      \
     act_gate = act_functor(act_gate_str);                                    \
     act_cell = act_functor(act_cell_str);                                    \
     act_cand = act_functor(act_cand_str);                                    \
   } else {                                                                   \
-    phi::funcs::VecActivations<T, platform::isa_any> act_functor;            \
+    phi::funcs::VecActivations<T, phi::backends::cpu::isa_any> act_functor;  \
     act_gate = act_functor(act_gate_str);                                    \
     act_cell = act_functor(act_cell_str);                                    \
     act_cand = act_functor(act_cand_str);                                    \

paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc

@@ -16,7 +16,7 @@ limitations under the License. */
 
 #include <string>
 
-#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/phi/backends/cpu/cpu_info.h"
 #include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/cpu_vec.h"
 #include "paddle/phi/kernels/funcs/fc_functor.h"
@@ -225,11 +225,11 @@ class FusionSeqExpandConcatFCOpKernel : public framework::OpKernel<T> {
 
     std::function<void(const int, const T*, T*)> fc_act;
     auto& fc_act_str = ctx.Attr<std::string>("fc_activation");
-    if (platform::MayIUse(platform::avx)) {
-      phi::funcs::VecActivations<T, platform::avx> act_functor;
+    if (phi::backends::cpu::MayIUse(phi::backends::cpu::avx)) {
+      phi::funcs::VecActivations<T, phi::backends::cpu::avx> act_functor;
       fc_act = act_functor(fc_act_str);
     } else {
-      phi::funcs::VecActivations<T, platform::isa_any> act_functor;
+      phi::funcs::VecActivations<T, phi::backends::cpu::isa_any> act_functor;
       fc_act = act_functor(fc_act_str);
     }