[xpu] fix xpu unittests: cast, transpose, pool, layer_norm, elementwise, multiclass_nms (#4259)

lite/kernels/xpu/bridges/cast_op.cc

@@ -88,7 +88,8 @@ int CastConverter(void* ctx, OpLite* op, KernelBase* kernel) {
   // Cast node
   graph->Add(
       out_name,
-      graph->builder_.CreateCast(*x_node->data(), CvtPrecisionType(out_ptype)));
+      graph->builder_.CreateCast(*x_node->data(), CvtPrecisionType(out_ptype)),
+      PrecisionType(out_ptype));
 
   return SUCCESS;
 }

lite/kernels/xpu/subgraph_compute.cc

@@ -54,6 +54,7 @@ bool SubgraphEngine::BuildDeviceProgram() {
       return false;
     }
   }
+
   // Collect the input and output nodes of the XPU IR graph
   std::vector<xtcl::xExpr*> device_inodes;
   std::vector<xtcl::xExpr*> device_onodes;
@@ -62,16 +63,20 @@ bool SubgraphEngine::BuildDeviceProgram() {
     CHECK(graph.Get(input_names_[i])->is_data());
     device_inodes.push_back(graph.Get(input_names_[i])->data().get());
   }
-  std::vector<std::string> valid_output_names;
   for (size_t i = 0; i < output_names_.size(); i++) {
     if (graph.Has(output_names_[i])) {
       device_onodes.push_back(graph.Get(output_names_[i])->data().get());
-      valid_output_names.push_back(output_names_[i]);
+    } else {
+      // update output_names_ and origin_otensors because some outputs may be
+      // useless
+      output_names_.erase(output_names_.begin() + i);
+      origin_otensors_.erase(origin_otensors_.begin() + i);
+      i--;
     }
   }
-  // update output_names_ because some outputs may be useless
-  output_names_ = valid_output_names;
   CHECK_GT(output_names_.size(), 0);
+  CHECK_EQ(output_names_.size(), origin_otensors_.size());
+
   // Build the XPU IR graph to the XPU runtime for inference
   device_program_ = lite::xpu::Device::Global().Build(
       &graph.builder_, &graph.params_, &device_onodes);

lite/tests/kernels/cast_compute_test.cc

@@ -135,8 +135,8 @@ TEST(Cast, precision) {
   float abs_error = 2e-5;
 #if defined(LITE_WITH_ARM)
   place = TARGET(kARM);
-// #elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
-//   place = TARGET(kXPU);
+#elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
+  place = TARGET(kXPU);
 #elif defined(LITE_WITH_HUAWEI_ASCEND_NPU)
   place = TARGET(kHuaweiAscendNPU);
   abs_error = 1e-2;  // precision_mode default is force_fp16

lite/tests/kernels/elementwise_compute_test.cc

@@ -182,6 +182,12 @@ void TestElt(Place place,
              std::vector<int64_t> y_shape,
              int axis,
              std::string act_type = "") {
+#if defined(LITE_WITH_XPU)
+  if ((y_shape.size() != 1 && x_shape.size() != y_shape.size()) ||
+      elt_type != std::string("add") || !act_type.empty()) {
+    return;
+  }
+#endif
   std::unique_ptr<arena::TestCase> tester(new ElementwiseComputeTester(
       place, "def", elt_type, x_shape, y_shape, axis, act_type));
   arena::Arena arena(std::move(tester), place, abs_error);
@@ -231,8 +237,8 @@ TEST(Elementwise, precision) {
   abs_error = 1e-2;  // precision_mode default is force_fp16
 #elif defined(LITE_WITH_ARM)
   place = TARGET(kARM);
-// #elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
-//   place = TARGET(kXPU);
+#elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
+  place = TARGET(kXPU);
 #else
   return;
 #endif

lite/tests/kernels/layer_norm_compute_test.cc

@@ -156,8 +156,8 @@ TEST(LayerNorm, precision) {
 #elif defined(LITE_WITH_ARM)
   place = TARGET(kARM);
   abs_error = 6e-5;
-// #elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
-//   place = TARGET(kXPU);
+#elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
+  place = TARGET(kXPU);
 #else
   return;
 #endif

lite/tests/kernels/multiclass_nms_compute_test.cc

@@ -478,8 +478,6 @@ TEST(multiclass_nms, precision) {
   Place place;
 #if defined(LITE_WITH_ARM)
   place = TARGET(kHost);
-// #elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
-//   place = TARGET(kXPU);
 #else
   return;
 #endif

lite/tests/kernels/pool_compute_test.cc

@@ -313,6 +313,7 @@ void TestPoolPaddings(Place place, float abs_error = 2e-5) {
   for (auto pooling_type : {"max", "avg"}) {
     TestPoolHelper(
         place, abs_error, {2, 3, 6, 7}, pooling_type, {1, 1}, {0, 0}, {2, 2});
+#if !defined(LITE_WITH_XPU)
     TestPoolHelper(
         place, abs_error, {2, 3, 6, 7}, pooling_type, {1, 1}, {1, 1}, {2, 2});
     TestPoolHelper(place,
@@ -336,6 +337,7 @@ void TestPoolPaddings(Place place, float abs_error = 2e-5) {
                    {1, 1},
                    {1, 0, 0, 1},
                    {2, 2});
+#endif
   }
 }
 
@@ -349,6 +351,7 @@ void TestPoolKsize(Place place, float abs_error = 2e-5) {
                      {1, 1},
                      {0, 0},
                      {ksize, ksize});
+#if !defined(LITE_WITH_XPU)
       TestPoolHelper(place,
                      abs_error,
                      {2, 3, 6, 7},
@@ -356,12 +359,16 @@ void TestPoolKsize(Place place, float abs_error = 2e-5) {
                      {2, 2},
                      {1, 1},
                      {ksize, ksize});
+#endif
     }
   }
 }
 
 void TestPoolCeilMode(Place place, float abs_error = 2e-5) {
   for (auto pooling_type : {"max", "avg"}) {
+#if defined(LITE_WITH_XPU)
+    if (pooling_type == std::string("max")) continue;
+#endif
     TestPoolHelper(place,
                    abs_error,
                    {2, 3, 6, 6},
@@ -384,8 +391,8 @@ TEST(Pool, precision) {
 #elif defined(LITE_WITH_HUAWEI_ASCEND_NPU)
   place = TARGET(kHuaweiAscendNPU);
   abs_error = 1e-2;  // precision_mode default is force_fp16
-// #elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)  // NOLINT
-//   place = TARGET(kXPU);
+#elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)  // NOLINT
+  place = TARGET(kXPU);
 #else
   return;
 #endif

lite/tests/kernels/transpose_compute_test.cc

@@ -127,7 +127,12 @@ class TransposeComputeTester : public arena::TestCase {
 
 void TestTranspose2D(Place place, float abs_error) {
   DDim x_dims{{4, 5}};
-  std::vector<std::vector<int>> axes{{0, 1}, {1, 0}};
+  std::vector<std::vector<int>> axes {
+#if !defined(LITE_WITH_XPU)
+    {0, 1},
+#endif
+        {1, 0},
+  };
   for (auto axis : axes) {
     std::unique_ptr<arena::TestCase> tester(
         new TransposeComputeTester(place, "def", x_dims, axis));
@@ -138,8 +143,12 @@ void TestTranspose2D(Place place, float abs_error) {
 
 void TestTranspose3D(Place place, float abs_error) {
   DDim x_dims{{3, 4, 5}};
-  std::vector<std::vector<int>> axes{
-      {0, 1, 2}, {0, 2, 1}, {1, 0, 2}, {2, 1, 0}};
+  std::vector<std::vector<int>> axes {
+#if !defined(LITE_WITH_XPU)
+    {0, 1, 2},
+#endif
+        {0, 2, 1}, {1, 0, 2}, {2, 1, 0},
+  };
   for (auto axis : axes) {
     std::unique_ptr<arena::TestCase> tester(
         new TransposeComputeTester(place, "def", x_dims, axis));
@@ -150,8 +159,12 @@ void TestTranspose3D(Place place, float abs_error) {
 
 void TestTranspose4D(Place place, float abs_error) {
   DDim x_dims{{2, 3, 4, 5}};
-  std::vector<std::vector<int>> axes{
-      {0, 1, 2, 3}, {0, 1, 3, 2}, {0, 2, 1, 3}, {3, 1, 2, 0}, {3, 1, 0, 2}};
+  std::vector<std::vector<int>> axes {
+#if !defined(LITE_WITH_XPU)
+    {0, 1, 2, 3}, {0, 1, 3, 2}, {0, 2, 1, 3}, {3, 1, 2, 0}, {3, 1, 0, 2},
+#endif
+        {0, 2, 3, 1}, {0, 3, 1, 2},
+  };
   for (auto axis : axes) {
     std::unique_ptr<arena::TestCase> tester(
         new TransposeComputeTester(place, "def", x_dims, axis));
@@ -170,8 +183,8 @@ TEST(Transpose, precision) {
 #elif defined(LITE_WITH_HUAWEI_ASCEND_NPU)
   place = TARGET(kHuaweiAscendNPU);
   abs_error = 1e-2;  // precision_mode default is force_fp16
-// #elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)  // NOLINT
-//   place = TARGET(kXPU);
+#elif defined(LITE_WITH_XPU) && defined(LITE_WITH_XTCL)
+  place = TARGET(kXPU);
 #else
   return;
 #endif