[XPU] Stop supporting xpu conv autotune config with paddlelite c api. (#9316)

Author: zealoct

lite/api/paddle_api.cc

@@ -626,8 +626,12 @@ void CxxConfig::set_xpu_multi_encoder_method(const std::string &precision,
 void CxxConfig::set_xpu_conv_autotune(bool autotune,
                                       const std::string &autotune_file) {
 #ifdef LITE_WITH_XPU
-  lite::TargetWrapperXPU::conv_autotune = autotune;
-  lite::TargetWrapperXPU::conv_autotune_file = autotune_file;
+  LOG(WARNING)
+      << "This function "
+         "'set_xpu_conv_autotune' is deprecated, "
+         "if you want to use autotune, please refer to "
+         "http://agroup.baidu.com/share/md/f9233d84df11452488a1fdd4f859647f";
+
 #else
   LOG(WARNING) << "The invoking of the function "
                   "'set_xpu_conv_autotune' is ignored, please "

lite/api/paddle_api.h

@@ -472,6 +472,8 @@ class LITE_API CxxConfig : public ConfigBase {
 
   void set_xpu_gm_workspace_method(size_t gm_size);
 
+  // **DEPRECATED**, use environ variable to enable autotune
+  // check http://agroup.baidu.com/share/md/f9233d84df11452488a1fdd4f859647f
   void set_xpu_conv_autotune(bool autotune = true,
                              const std::string& autotune_file = "");
 

lite/backends/xpu/target_wrapper.cc

@@ -165,9 +165,6 @@ LITE_THREAD_LOCAL std::shared_ptr<void> TargetWrapperXPU::xpu_stream_{nullptr};
 LITE_THREAD_LOCAL std::string
     TargetWrapperXPU::multi_encoder_precision;  // NOLINT
 LITE_THREAD_LOCAL bool TargetWrapperXPU::multi_encoder_adaptive_seqlen{false};
-// conv autotune config
-LITE_THREAD_LOCAL bool TargetWrapperXPU::conv_autotune{false};
-LITE_THREAD_LOCAL std::string TargetWrapperXPU::conv_autotune_file;
 // l3 cache config
 LITE_THREAD_LOCAL bool TargetWrapperXPU::need_l3_mutex{false};
 LITE_THREAD_LOCAL size_t TargetWrapperXPU::local_l3_size{

lite/backends/xpu/target_wrapper.h

@@ -111,14 +111,6 @@ class TargetWrapper<TARGET(kXPU)> {
         quantizer_.reset(new XPUQuantizer());
       }
       CHECK(quantizer_.get());
-      if (conv_autotune) {
-        tls_raw_ctx_->_xpu1_conv_selector.set_autotune_loop(true);
-        tls_raw_ctx_->_xpu1_conv_selector.set_inference_mode(true);
-      }
-      if (!conv_autotune_file.empty()) {
-        tls_raw_ctx_->_xpu1_conv_selector.set_autotune_file(
-            conv_autotune_file.c_str());
-      }
       int devid = -1;
       uint64_t max_l3_size = 0;
       XPU_CALL(xpu_current_device(&devid));
@@ -173,9 +165,6 @@ class TargetWrapper<TARGET(kXPU)> {
   // multi encoder config
   static LITE_THREAD_LOCAL std::string multi_encoder_precision;  // NOLINT
   static LITE_THREAD_LOCAL bool multi_encoder_adaptive_seqlen;
-  // conv autotune config
-  static LITE_THREAD_LOCAL bool conv_autotune;
-  static LITE_THREAD_LOCAL std::string conv_autotune_file;  // NOLINT
   // l3 cache config
   static LITE_THREAD_LOCAL bool need_l3_mutex;    // model level l3 size
   static LITE_THREAD_LOCAL size_t local_l3_size;  // model level l3 size

lite/tests/api/test_resnet50_fp32_baidu_xpu.cc

@@ -14,6 +14,7 @@
 
 #include <gflags/gflags.h>
 #include <gtest/gtest.h>
+#include <stdlib.h>
 #include <vector>
 #include "lite/api/paddle_api.h"
 #include "lite/api/paddle_use_kernels.h"
@@ -35,13 +36,13 @@ namespace paddle {
 namespace lite {
 
 TEST(resnet50, test_resnet50_fp32_baidu_xpu) {
+  setenv("XPU_CONV_AUTOTUNE", "5", 1);
   lite_api::CxxConfig config;
   config.set_model_dir(FLAGS_model_dir);
   config.set_valid_places({lite_api::Place{TARGET(kXPU), PRECISION(kFloat)},
                            lite_api::Place{TARGET(kX86), PRECISION(kFloat)},
                            lite_api::Place{TARGET(kHost), PRECISION(kFloat)}});
   config.set_xpu_l3_cache_method(16773120, false);
-  config.set_xpu_conv_autotune(true);
   auto predictor = lite_api::CreatePaddlePredictor(config);
 
   std::string raw_data_dir = FLAGS_data_dir + std::string("/raw_data");