[Metal] fix bilinear_interp bug (#8285)

* fix_bilinear_interp * Update test_bilinear_interp_op.py
PaddlePaddle · Jan 18, 2022 · 37c3961 · 37c3961
1 parent 5f78558
commit 37c3961
Show file tree

Hide file tree

Showing 4 changed files with 80 additions and 49 deletions.
diff --git a/lite/backends/metal/metal_kernel/texture/BilinearInterp.metal b/lite/backends/metal/metal_kernel/texture/BilinearInterp.metal
@@ -32,8 +32,17 @@ kernel void bilinear_interp(texture2d_array<ftype, access::read> input[[texture(
     } else {
         ftype w = (gid.x + pm.align_delta) * pm.ratio_w - pm.align_delta;
         ftype h = (gid.y + pm.align_delta) * pm.ratio_h - pm.align_delta;
-        uint w0 = w, h0 = h;
-        uint w1 = w0 + 1, h1 = h0 + 1;
+        h = (h > 0) ? h : 0;
+        w = (w > 0) ? w : 0;
+        int w0 = (int)w;
+        int h0 = (int)h;
+        int w1 = w0 + 1, h1 = h0 + 1;
+        if (w0 < 0) {
+            w0 = 0;
+        }
+        if (h0 < 0) {
+            h0 = 0;
+        }
 
         ftype w1lambda = w - w0, h1lambda = h - h0;
         ftype w2lambda = 1.0 - w1lambda, h2lambda = 1.0 - h1lambda;

diff --git a/lite/kernels/metal/image_op/interp_image_compute.mm b/lite/kernels/metal/image_op/interp_image_compute.mm
@@ -185,11 +185,11 @@
             PRECISION(kFloat),
             DATALAYOUT(kMetalTexture2DArray))})
     .BindInput("OutSize",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("SizeTensor",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("Scale",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindOutput("Out",
         {LiteType::GetTensorTy(TARGET(kMetal),
             PRECISION(kFloat),
@@ -205,11 +205,11 @@
     .BindInput("X",
         {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kFP16), DATALAYOUT(kMetalTexture2DArray))})
     .BindInput("OutSize",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("SizeTensor",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("Scale",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindOutput("Out",
         {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kFP16), DATALAYOUT(kMetalTexture2DArray))})
     .Finalize();
@@ -225,11 +225,11 @@
             PRECISION(kFloat),
             DATALAYOUT(kMetalTexture2DArray))})
     .BindInput("OutSize",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("SizeTensor",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("Scale",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindOutput("Out",
         {LiteType::GetTensorTy(TARGET(kMetal),
             PRECISION(kFloat),
@@ -245,11 +245,11 @@
     .BindInput("X",
         {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kFP16), DATALAYOUT(kMetalTexture2DArray))})
     .BindInput("OutSize",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("SizeTensor",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindInput("Scale",
-        {LiteType::GetTensorTy(TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW))})
+        {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kInt32), DATALAYOUT(kNCHW))})
     .BindOutput("Out",
         {LiteType::GetTensorTy(TARGET(kMetal), PRECISION(kFP16), DATALAYOUT(kMetalTexture2DArray))})
     .Finalize();

diff --git a/lite/tests/unittest_py/op/test_bilinear_interp_op.py b/lite/tests/unittest_py/op/test_bilinear_interp_op.py
@@ -40,22 +40,29 @@ def __init__(self, *args, **kwargs):
             PrecisionType.FP32,
             DataLayoutType.NCHW,
             thread=[1, 4])
-        # opencl demo
-        # opencl has diff
-        # opencl_places = [
-        #     Place(TargetType.OpenCL, PrecisionType.FP16,
-        #           DataLayoutType.ImageDefault), Place(
-        #               TargetType.OpenCL, PrecisionType.FP16,
-        #               DataLayoutType.ImageFolder),
-        #     Place(TargetType.OpenCL, PrecisionType.FP32, DataLayoutType.NCHW),
-        #     Place(TargetType.OpenCL, PrecisionType.Any,
-        #           DataLayoutType.ImageDefault), Place(
-        #               TargetType.OpenCL, PrecisionType.Any,
-        #               DataLayoutType.ImageFolder),
-        #     Place(TargetType.OpenCL, PrecisionType.Any, DataLayoutType.NCHW),
-        #     Place(TargetType.Host, PrecisionType.FP32)
-        # ]
-        # self.enable_testing_on_place(places=opencl_places)
+        opencl_places = [
+            Place(TargetType.OpenCL, PrecisionType.FP16,
+                  DataLayoutType.ImageDefault), Place(
+                      TargetType.OpenCL, PrecisionType.FP16,
+                      DataLayoutType.ImageFolder),
+            Place(TargetType.OpenCL, PrecisionType.FP32, DataLayoutType.NCHW),
+            Place(TargetType.OpenCL, PrecisionType.Any,
+                  DataLayoutType.ImageDefault), Place(
+                      TargetType.OpenCL, PrecisionType.Any,
+                      DataLayoutType.ImageFolder),
+            Place(TargetType.OpenCL, PrecisionType.Any, DataLayoutType.NCHW),
+            Place(TargetType.Host, PrecisionType.FP32)
+        ]
+        self.enable_testing_on_place(places=opencl_places)
+        metal_places = [
+            Place(TargetType.Metal, PrecisionType.FP32,
+                  DataLayoutType.MetalTexture2DArray),
+            Place(TargetType.Metal, PrecisionType.FP16,
+                  DataLayoutType.MetalTexture2DArray),
+            Place(TargetType.ARM, PrecisionType.FP32),
+            Place(TargetType.Host, PrecisionType.FP32)
+        ]
+        self.enable_testing_on_place(places=metal_places)
 
     def is_program_valid(self,
                          program_config: ProgramConfig,
@@ -123,7 +130,11 @@ def generate_scale(*args, **kwargs):
         return program_config
 
     def sample_predictor_configs(self):
-        return self.get_predictor_configs(), ["bilinear_interp"], (1e-4, 1e-4)
+        atol, rtol = 1e-4, 1e-4
+        target_str = self.get_target()
+        if target_str == "Metal":
+            atol, rtol = 5e-1, 5e-1
+        return self.get_predictor_configs(), ["bilinear_interp"], (atol, rtol)
 
     def add_ignore_pass_case(self):
         pass

diff --git a/lite/tests/unittest_py/op/test_bilinear_interp_v2_op.py b/lite/tests/unittest_py/op/test_bilinear_interp_v2_op.py
@@ -40,22 +40,29 @@ def __init__(self, *args, **kwargs):
             PrecisionType.FP32,
             DataLayoutType.NCHW,
             thread=[1, 4])
-        # opencl demo
-        # opencl has diff
-        # opencl_places = [
-        #     Place(TargetType.OpenCL, PrecisionType.FP16,
-        #           DataLayoutType.ImageDefault), Place(
-        #               TargetType.OpenCL, PrecisionType.FP16,
-        #               DataLayoutType.ImageFolder),
-        #     Place(TargetType.OpenCL, PrecisionType.FP32, DataLayoutType.NCHW),
-        #     Place(TargetType.OpenCL, PrecisionType.Any,
-        #           DataLayoutType.ImageDefault), Place(
-        #               TargetType.OpenCL, PrecisionType.Any,
-        #               DataLayoutType.ImageFolder),
-        #     Place(TargetType.OpenCL, PrecisionType.Any, DataLayoutType.NCHW),
-        #     Place(TargetType.Host, PrecisionType.FP32)
-        # ]
-        # self.enable_testing_on_place(places=opencl_places)
+        opencl_places = [
+            Place(TargetType.OpenCL, PrecisionType.FP16,
+                  DataLayoutType.ImageDefault), Place(
+                      TargetType.OpenCL, PrecisionType.FP16,
+                      DataLayoutType.ImageFolder),
+            Place(TargetType.OpenCL, PrecisionType.FP32, DataLayoutType.NCHW),
+            Place(TargetType.OpenCL, PrecisionType.Any,
+                  DataLayoutType.ImageDefault), Place(
+                      TargetType.OpenCL, PrecisionType.Any,
+                      DataLayoutType.ImageFolder),
+            Place(TargetType.OpenCL, PrecisionType.Any, DataLayoutType.NCHW),
+            Place(TargetType.Host, PrecisionType.FP32)
+        ]
+        self.enable_testing_on_place(places=opencl_places)
+        metal_places = [
+            Place(TargetType.Metal, PrecisionType.FP32,
+                  DataLayoutType.MetalTexture2DArray),
+            Place(TargetType.Metal, PrecisionType.FP16,
+                  DataLayoutType.MetalTexture2DArray),
+            Place(TargetType.ARM, PrecisionType.FP32),
+            Place(TargetType.Host, PrecisionType.FP32)
+        ]
+        self.enable_testing_on_place(places=metal_places)
 
     def is_program_valid(self,
                          program_config: ProgramConfig,
@@ -128,8 +135,12 @@ def generate_scale(*args, **kwargs):
         return program_config
 
     def sample_predictor_configs(self):
-        return self.get_predictor_configs(), ["bilinear_interp_v2"], (1e-4,
-                                                                      1e-4)
+        atol, rtol = 1e-4, 1e-4
+        target_str = self.get_target()
+        if target_str == "Metal":
+            atol, rtol = 3e-1, 3e-1
+        return self.get_predictor_configs(), ["bilinear_interp_v2"], (atol,
+                                                                      rtol)
 
     def add_ignore_pass_case(self):
         pass