[NPU] fix fp16 (PART II) (#40537)

[NPU] fix fp16 (PART II)

paddle/fluid/operators/conv_op_npu.cc

@@ -356,7 +356,7 @@ class NPUConvGradOpKernel : public framework::OpKernel<T> {
 
     auto stream = ctx.template device_context<NPUDeviceContext>().stream();
     if (filter_grad) {
-      filter_grad->mutable_data<T>(ctx.GetPlace());
+      filter_grad->mutable_data<float>(ctx.GetPlace());
       std::vector<int> filter_shape_vec = phi::vectorize<int>(filter->dims());
 
       const auto& runner = NpuOpRunner(

paddle/fluid/operators/pad_op_npu.cc

@@ -90,5 +90,5 @@ namespace plat = paddle::platform;
 REGISTER_OP_NPU_KERNEL(pad, ops::PadNPUKernel<plat::float16>,
                        ops::PadNPUKernel<float>, ops::PadNPUKernel<int>);
 
-REGISTER_OP_NPU_KERNEL(pad_grad, ops::PadNPUKernel<plat::float16>,
+REGISTER_OP_NPU_KERNEL(pad_grad, ops::PadGradNPUKernel<plat::float16>,
                        ops::PadGradNPUKernel<float>);

python/paddle/fluid/tests/unittests/npu/CMakeLists.txt

@@ -22,4 +22,5 @@ if (WITH_ASCEND_CL)
     set_tests_properties(test_conv2d_transpose_op_npu PROPERTIES TIMEOUT 200)
     set_tests_properties(test_conv2d_op_npu PROPERTIES TIMEOUT 300)
     set_tests_properties(test_matmulv2_op_npu PROPERTIES TIMEOUT 300)
+    set_tests_properties(test_elementwise_add_op_npu PROPERTIES TIMEOUT 200)
 endif()

python/paddle/fluid/tests/unittests/npu/test_conv2d_op_depthwise_conv_npu.py

@@ -132,9 +132,13 @@ class TestDepthwiseConvNPU(OpTest):
         self.check_output_with_place(self.place, atol=1e-2)
 
     def test_check_grad(self):
-        if self.dtype == np.float16:
-            return
         if self.dilations[0] == 1 and self.dilations[1] == 1:
+            if self.dtype == np.float16:
+                self.check_grad_with_place(
+                    self.place, {'Input', 'Filter'},
+                    'Output',
+                    max_relative_error=0.9)
+            else:
                 self.check_grad_with_place(
                     self.place, {'Input', 'Filter'},
                     'Output',
@@ -143,7 +147,12 @@ class TestDepthwiseConvNPU(OpTest):
 
     def test_check_grad_no_filter(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['Input'],
+                'Output',
+                no_grad_set=set(['Filter']),
+                max_relative_error=0.9)
+        else:
             self.check_grad_with_place(
                 self.place, ['Input'],
                 'Output',
@@ -152,9 +161,14 @@ class TestDepthwiseConvNPU(OpTest):
                 numeric_place=paddle.CPUPlace())
 
     def test_check_grad_no_input(self):
-        if self.dtype == np.float16:
-            return
         if self.dilations[0] == 1 and self.dilations[1] == 1:
+            if self.dtype == np.float16:
+                self.check_grad_with_place(
+                    self.place, ['Filter'],
+                    'Output',
+                    no_grad_set=set(['Input']),
+                    max_relative_error=0.9)
+            else:
                 self.check_grad_with_place(
                     self.place, ['Filter'],
                     'Output',
@@ -267,7 +281,11 @@ class TestDepthwiseConvNPU_Padding(OpTest):
 
     def test_check_grad(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, {'Input', 'Filter'},
+                'Output',
+                max_relative_error=1.2)
+        else:
             self.check_grad_with_place(
                 self.place, {'Input', 'Filter'},
                 'Output',
@@ -276,7 +294,12 @@ class TestDepthwiseConvNPU_Padding(OpTest):
 
     def test_check_grad_no_filter(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['Input'],
+                'Output',
+                max_relative_error=0.7,
+                no_grad_set=set(['Filter']))
+        else:
             self.check_grad_with_place(
                 self.place, ['Input'],
                 'Output',
@@ -286,7 +309,12 @@ class TestDepthwiseConvNPU_Padding(OpTest):
 
     def test_check_grad_no_input(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['Filter'],
+                'Output',
+                max_relative_error=0.8,
+                no_grad_set=set(['Input']))
+        else:
             self.check_grad_with_place(
                 self.place, ['Filter'],
                 'Output',

python/paddle/fluid/tests/unittests/npu/test_conv2d_op_npu.py

@@ -127,8 +127,6 @@ class TestConv2DOp(OpTest):
         self.check_output_with_place(fluid.NPUPlace(0), atol=1e-2)
 
     def test_check_grad(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(
             fluid.NPUPlace(0), {'Input', 'Filter'},
             'Output',
@@ -136,8 +134,6 @@ class TestConv2DOp(OpTest):
             numeric_place=paddle.CPUPlace())
 
     def test_check_grad_no_filter(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(
             fluid.NPUPlace(0), ['Input'],
             'Output',
@@ -146,8 +142,6 @@ class TestConv2DOp(OpTest):
             numeric_place=paddle.CPUPlace())
 
     def test_check_grad_no_input(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(
             fluid.NPUPlace(0), ['Filter'],
             'Output',
@@ -276,10 +270,13 @@ class TestConv2DOp_v2(OpTest):
     def set_npu(self):
         self.__class__.use_npu = True
 
+    def init_dtype(self):
+        self.dtype = np.float32
+
     def setUp(self):
         self.set_npu()
         self.op_type = "conv2d"
-        self.dtype = np.float32
+        self.init_dtype()
         self.init_kernel_type()
         self.init_group()
         self.init_dilation()
@@ -320,7 +317,11 @@ class TestConv2DOp_v2(OpTest):
 
     def test_check_grad(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                paddle.NPUPlace(0), {'Input', 'Filter'},
+                'Output',
+                max_relative_error=1.1)
+        else:
             self.check_grad_with_place(
                 paddle.NPUPlace(0), {'Input', 'Filter'},
                 'Output',
@@ -329,7 +330,12 @@ class TestConv2DOp_v2(OpTest):
 
     def test_check_grad_no_filter(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                paddle.NPUPlace(0), ['Input'],
+                'Output',
+                max_relative_error=0.99,
+                no_grad_set=set(['Filter']))
+        else:
             self.check_grad_with_place(
                 paddle.NPUPlace(0), ['Input'],
                 'Output',
@@ -339,7 +345,12 @@ class TestConv2DOp_v2(OpTest):
 
     def test_check_grad_no_input(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                paddle.NPUPlace(0), ['Filter'],
+                'Output',
+                max_relative_error=0.99,
+                no_grad_set=set(['Input']))
+        else:
             self.check_grad_with_place(
                 paddle.NPUPlace(0), ['Filter'],
                 'Output',

python/paddle/fluid/tests/unittests/npu/test_elementwise_add_op_npu.py

@@ -65,9 +65,16 @@ class TestElementwiseAddOp(OpTest):
         self.check_output_with_place(self.place)
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16 or self.dtype == np.int64:
+        if self.dtype == np.int64:
             return
 
+        if self.dtype == np.float16:
+            self.check_grad_with_place(
+                self.place,
+                ['X', 'Y'],
+                'Out',
+                max_relative_error=0.15, )
+        else:
             self.check_grad_with_place(
                 self.place,
                 ['X', 'Y'],
@@ -75,9 +82,17 @@ class TestElementwiseAddOp(OpTest):
                 max_relative_error=0.006, )
 
     def test_check_grad_ingore_x(self):
-        if self.dtype == np.float16 or self.dtype == np.int64:
+        if self.dtype == np.int64:
             return
 
+        if self.dtype == np.float16:
+            self.check_grad_with_place(
+                self.place,
+                ['Y'],
+                'Out',
+                no_grad_set=set("X"),
+                max_relative_error=0.92, )
+        else:
             self.check_grad_with_place(
                 self.place,
                 ['Y'],
@@ -86,9 +101,17 @@ class TestElementwiseAddOp(OpTest):
                 max_relative_error=0.006, )
 
     def test_check_grad_ingore_y(self):
-        if self.dtype == np.float16 or self.dtype == np.int64:
+        if self.dtype == np.int64:
             return
 
+        if self.dtype == np.float16:
+            self.check_grad_with_place(
+                self.place,
+                ['X'],
+                'Out',
+                no_grad_set=set("Y"),
+                max_relative_error=0.8, )
+        else:
             self.check_grad_with_place(
                 self.place,
                 ['X'],

python/paddle/fluid/tests/unittests/npu/test_elementwise_max_op_npu.py

@@ -116,19 +116,13 @@ class TestElementwiseMaxOp(OpTest):
         self.check_output_with_place(self.place)
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(self.place, ['X', 'Y'], 'Out')
 
     def test_check_grad_ingore_x(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(
             self.place, ['Y'], 'Out', no_grad_set=set("X"))
 
     def test_check_grad_ingore_y(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(
             self.place, ['X'], 'Out', no_grad_set=set("Y"))
 
@@ -213,15 +207,11 @@ class TestElementwiseMaxOp_broadcast_2(TestElementwiseMaxOp):
         self.out = np.maximum(self.x, self.y.reshape(1, 1, 100))
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16:
-            return
         dx, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
         self.check_grad_with_place(
             self.place, ['X', 'Y'], 'Out', user_defined_grads=[dx, dy])
 
     def test_check_grad_ingore_x(self):
-        if self.dtype == np.float16:
-            return
         _, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
         self.check_grad_with_place(
             self.place, ['Y'],
@@ -230,8 +220,6 @@ class TestElementwiseMaxOp_broadcast_2(TestElementwiseMaxOp):
             user_defined_grads=[dy])
 
     def test_check_grad_ingore_y(self):
-        if self.dtype == np.float16:
-            return
         dx, _ = ComputeGrad(self.x, self.y, self.out, self.axis)
         self.check_grad_with_place(
             self.place, ['X'],

python/paddle/fluid/tests/unittests/npu/test_elementwise_min_op_npu.py

@@ -64,8 +64,9 @@ class TestElementwiseMinOp(OpTest):
 
     def test_check_grad_normal(self):
         if self.dtype == np.float16:
-            return
-
+            self.check_grad_with_place(
+                self.place, ['X', 'Y'], 'Out', max_relative_error=0.5)
+        else:
             self.check_grad_with_place(
                 self.place,
                 ['X', 'Y'],
@@ -73,8 +74,12 @@ class TestElementwiseMinOp(OpTest):
 
     def test_check_grad_ingore_x(self):
         if self.dtype == np.float16:
-            return
-
+            self.check_grad_with_place(
+                self.place, ['Y'],
+                'Out',
+                no_grad_set=set("X"),
+                max_relative_error=0.9)
+        else:
             self.check_grad_with_place(
                 self.place,
                 ['Y'],
@@ -83,8 +88,12 @@ class TestElementwiseMinOp(OpTest):
 
     def test_check_grad_ingore_y(self):
         if self.dtype == np.float16:
-            return
-
+            self.check_grad_with_place(
+                self.place, ['X'],
+                'Out',
+                no_grad_set=set("Y"),
+                max_relative_error=0.1)
+        else:
             self.check_grad_with_place(
                 self.place,
                 ['X'],

python/paddle/fluid/tests/unittests/npu/test_elementwise_pow_op_npu.py

@@ -114,8 +114,6 @@ class TestElementwisePow(OpTest):
         self.out = np.power(self.x, self.y)
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16:
-            return
         dx, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
         self.check_grad_with_place(
             self.place, ['X', 'Y'], 'Out', user_defined_grads=[dx, dy])
@@ -184,8 +182,6 @@ class TestElementwisePowOp_broadcast_0(TestElementwisePow):
         self.out = np.power(self.x, self.y)
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16:
-            return
         dx, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
         self.check_grad_with_place(
             self.place, ['X', 'Y'], 'Out', user_defined_grads=[dx, dy])
@@ -218,8 +214,6 @@ class TestElementwisePowOp_broadcast_1(TestElementwisePow):
         self.out = np.power(self.x, self.y.reshape(1, 100, 1))
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16:
-            return
         dx, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
         self.check_grad_with_place(
             self.place, ['X', 'Y'], 'Out', user_defined_grads=[dx, dy])
@@ -252,8 +246,6 @@ class TestElementwisePowOp_broadcast_2(TestElementwisePow):
         self.out = np.power(self.x, self.y.reshape(100, 1, 1))
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16:
-            return
         dx, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
         self.check_grad_with_place(
             self.place, ['X', 'Y'], 'Out', user_defined_grads=[dx, dy])

python/paddle/fluid/tests/unittests/npu/test_expand_op_npu.py

@@ -34,7 +34,7 @@ class TestExpand(OpTest):
 
         self.init_dtype()
         np.random.seed(SEED)
-        x = np.random.randn(3, 1, 7).astype(self.dtype)
+        x = np.random.randn(30, 1, 7).astype(self.dtype)
         out = np.tile(x, [1, 10, 1])
 
         self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
@@ -50,12 +50,8 @@ class TestExpand(OpTest):
     def test_check_output(self):
         self.check_output_with_place(self.place)
 
-    # TODO(ascendrc): Add grad test
-    # def test_check_grad(self):
-    #     if self.dtype == np.float16:
-    #         return
-    #     self.check_grad(['X'], 'Out')
-    #
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
 
 
 class TestExpandV2(TestExpand):
@@ -66,7 +62,7 @@ class TestExpandV2(TestExpand):
 
         self.init_dtype()
         np.random.seed(SEED)
-        x = np.random.randn(3, 1, 7).astype(self.dtype)
+        x = np.random.randn(30, 1, 7).astype(self.dtype)
         out = np.tile(x, [1, 10, 1])
         expand_times = np.array([1, 10, 1]).astype(np.int32)
 
@@ -145,7 +141,7 @@ class TestExpand_expand_times_all_one(TestExpand):
 
         self.init_dtype()
         np.random.seed(SEED)
-        x = np.random.randn(3, 1, 7).astype(self.dtype)
+        x = np.random.randn(30, 1, 7).astype(self.dtype)
         out = np.tile(x, [1, 1, 1])
 
         self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}

python/paddle/fluid/tests/unittests/npu/test_huber_loss_op_npu.py

@@ -81,13 +81,9 @@ class TestHuberLossOp(OpTest):
         self.check_output_with_place(self.place)
 
     def test_check_grad_normal(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(self.place, ['X', 'Y'], 'Out')
 
     def test_check_grad_ingore_x(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(
             self.place, ['Y'],
             'Out',
@@ -95,8 +91,6 @@ class TestHuberLossOp(OpTest):
             no_grad_set=set("residual"))
 
     def test_check_grad_ingore_y(self):
-        if self.dtype == np.float16:
-            return
         self.check_grad_with_place(
             self.place, ['X'],
             'Out',

python/paddle/fluid/tests/unittests/npu/test_label_smooth_op_npu.py

@@ -78,7 +78,9 @@ class TestLabelSmoothOp(OpTest):
 
     def test_check_grad(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['X'], 'Out', max_relative_error=0.5)
+        else:
             self.check_grad_with_place(self.place, ['X'], 'Out')
 
 

python/paddle/fluid/tests/unittests/npu/test_leaky_relu_op_npu.py

@@ -63,7 +63,9 @@ class TestLeadyRelu(OpTest):
 
     def test_check_grad(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['X'], 'Out', max_relative_error=0.006)
+        else:
             self.check_grad_with_place(self.place, ['X'], 'Out')
 
 

python/paddle/fluid/tests/unittests/npu/test_log_softmax_op_npu.py

@@ -63,7 +63,11 @@ class TestLogSoftmaxNPUOp(OpTest):
 
     def test_check_grad(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['X'], ['Out'],
+                user_defined_grads=[self.x_grad],
+                max_relative_error=0.02)
+        else:
             self.check_grad_with_place(
                 self.place, ['X'], ['Out'], user_defined_grads=[self.x_grad])
 

python/paddle/fluid/tests/unittests/npu/test_lookup_table_v2_op_npu.py

@@ -77,7 +77,9 @@ class TestLookupTableV2(OpTest):
 
     def test_check_grad(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['W'], 'Out', max_relative_error=0.01)
+        else:
             self.check_grad_with_place(self.place, ['W'], 'Out')
 
 

python/paddle/fluid/tests/unittests/npu/test_nearest_interp_v2_op_npu.py

@@ -39,10 +39,11 @@ class TestNearestInterpOp(OpTest):
         self.set_npu()
         self.out_size = None
         self.actual_shape = None
+        self.init_dtype()
         self.data_layout = 'NCHW'
         self.init_test_case()
         self.op_type = "nearest_interp_v2"
-        input_np = np.random.random(self.input_shape).astype("float32")
+        input_np = np.random.random(self.input_shape).astype(self.dtype)
 
         if self.data_layout == "NCHW":
             in_h = self.input_shape[2]
@@ -95,8 +96,21 @@ class TestNearestInterpOp(OpTest):
         self.check_output_with_place(self.place)
 
     def test_check_grad(self):
+        if self.dtype == np.float16:
+            self.check_grad_with_place(
+                self.place, ['X'],
+                'Out',
+                in_place=True,
+                max_relative_error=0.02)
+        else:
             self.check_grad_with_place(
-            self.place, ['X'], 'Out', in_place=True, max_relative_error=0.006)
+                self.place, ['X'],
+                'Out',
+                in_place=True,
+                max_relative_error=0.006)
+
+    def init_dtype(self):
+        self.dtype = np.float32
 
     def init_test_case(self):
         self.interp_method = 'nearest'
@@ -108,6 +122,11 @@ class TestNearestInterpOp(OpTest):
         self.align_corners = False
 
 
+class TestNearestNeighborInterpFP16(TestNearestInterpOp):
+    def init_dtype(self):
+        self.dtype = np.float16
+
+
 class TestNearestNeighborInterpCase1(TestNearestInterpOp):
     def init_test_case(self):
         self.interp_method = 'nearest'

python/paddle/fluid/tests/unittests/npu/test_pad_op_npu.py

@@ -50,8 +50,9 @@ class TestPadOp(OpTest):
 
     def test_check_grad_normal(self):
         if self.dtype == np.float16:
-            return
-
+            self.check_grad_with_place(
+                self.place, ['X'], 'Out', max_relative_error=0.6)
+        else:
             self.check_grad_with_place(self.place, ['X'], 'Out')
 
     def set_npu(self):

python/paddle/fluid/tests/unittests/npu/test_relu_op_npu.py

@@ -34,11 +34,12 @@ class TestRelu(OpTest):
 
         self.init_dtype()
         np.random.seed(SEED)
-        x = np.random.rand(3, 2).astype(self.dtype)
-        out = x
 
-        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
-        self.attrs = {}
+        x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype)
+        # The same reason with TestAbs
+        x[np.abs(x) < 0.005] = 0.02
+        out = np.maximum(x, 0)
+        self.inputs = {'X': x}
         self.outputs = {'Out': out}
 
     def set_npu(self):
@@ -50,32 +51,18 @@ class TestRelu(OpTest):
     def test_check_output(self):
         self.check_output_with_place(self.place)
 
+    def test_check_grad(self):
+        if self.dtype == np.float16:
+            self.check_grad_with_place(
+                self.place, ['X'], 'Out', max_relative_error=0.006)
+        else:
+            self.check_grad_with_place(self.place, ['X'], 'Out')
 
-class TestReluFp16(OpTest):
-    def setUp(self):
-        self.set_npu()
-        self.op_type = "relu"
-        self.place = paddle.NPUPlace(0)
-
-        self.init_dtype()
-        np.random.seed(SEED)
-        x = np.random.rand(3, 2).astype(self.dtype)
-        out = x
-
-        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
-        self.attrs = {}
-        self.outputs = {'Out': out}
-
-    def set_npu(self):
-        self.__class__.use_npu = True
-        self.__class__.no_need_check_grad = True
 
+class TestReluFp16(TestRelu):
     def init_dtype(self):
         self.dtype = np.float16
 
-    def test_check_output(self):
-        self.check_output_with_place(self.place, atol=1e-5)
-
 
 class TestReluNeg(OpTest):
     def setUp(self):

python/paddle/fluid/tests/unittests/npu/test_slice_op_npu.py

@@ -58,11 +58,16 @@ class TestSliceOp(OpTest):
         self.place = paddle.NPUPlace(0)
 
     def test_check_output(self):
+        if self.dtype == np.float16:
+            self.check_output_with_place(self.place)
+        else:
             self.check_output_with_place(self.place)
 
     def test_check_grad_normal(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['Input'], 'Out', max_relative_error=0.02)
+        else:
             self.check_grad_with_place(self.place, ['Input'], 'Out')
 
 
@@ -347,7 +352,9 @@ class TestSliceOpDecsDim(OpTest):
 
     def test_check_grad_normal(self):
         if self.dtype == np.float16:
-            return
+            self.check_grad_with_place(
+                self.place, ['Input'], 'Out', max_relative_error=0.5)
+        else:
             self.check_grad_with_place(self.place, ['Input'], 'Out')