!8040 [MS][LITE][GPU]fix bug in arithmetic broadcast C and Sigmoid

Merge pull request !8040 from chenzupeng/master-lite

mindspore/lite/src/runtime/kernel/opencl/cl/activation.cl

@@ -35,12 +35,27 @@ __kernel void Relu6(__read_only image2d_t input, __write_only image2d_t output,
   WRITE_IMAGE(output, (int2)(X, Y), in_c4);
 }
 
-__kernel void Sigmoid(__read_only image2d_t input, __write_only image2d_t output, const int2 img_shape) {
+__kernel void Sigmoid(__read_only image2d_t input, __write_only image2d_t output, const int2 img_shape, const int c4,
+                      const int last_c4) {
   int X = get_global_id(0);
   int Y = get_global_id(1);
   if (X >= img_shape.x || Y >= img_shape.y) return;
+  int C4 = X % c4;
   FLT4 in_c4 = READ_IMAGE(input, smp_zero, (int2)(X, Y));
-  in_c4 = (FLT4)(1.f) / ((FLT4)(1.f) + exp(-in_c4));
+  if (C4 < c4 - 1) {
+    in_c4 = (FLT4)(1.f) / ((FLT4)(1.f) + exp(-in_c4));
+  } else {
+    in_c4.x = (FLT)(1.f) / ((FLT)(1.f) + exp(-in_c4.x));
+    if (last_c4 > 1) {
+      in_c4.y = (FLT)(1.f) / ((FLT)(1.f) + exp(-in_c4.y));
+    }
+    if (last_c4 > 2) {
+      in_c4.z = (FLT)(1.f) / ((FLT)(1.f) + exp(-in_c4.z));
+    }
+    if (last_c4 > 3) {
+      in_c4.w = (FLT)(1.f) / ((FLT)(1.f) + exp(-in_c4.w));
+    }
+  }
   WRITE_IMAGE(output, (int2)(X, Y), in_c4);
 }
 

mindspore/lite/src/runtime/kernel/opencl/cl/arithmetic.cl

@@ -266,7 +266,7 @@ __kernel void ElementGreaterEqual_IMG(__read_only image2d_t input_a, __read_only
 
 __kernel void BroadcastNHWC4Add_IMG(__read_only image2d_t input_a, __read_only image2d_t input_b,
                                     __write_only image2d_t output, const int4 a_shape, const int4 b_shape,
-                                    const int4 output_shape, float act_min, float act_max) {
+                                    const int4 output_shape, const int broadcastC_flag, float act_min, float act_max) {
   int X = get_global_id(0);  // C4
   int Y = get_global_id(1);  // W
   int Z = get_global_id(2);  // H
@@ -281,14 +281,21 @@ __kernel void BroadcastNHWC4Add_IMG(__read_only image2d_t input_a, __read_only i
   int b_w = Y < b_shape.z ? Y : b_shape.z - 1;
   int b_h = Z < b_shape.y ? Z : b_shape.y - 1;
   FLT4 b = READ_IMAGE(input_b, smp_none, (int2)(b_w * b_shape.w + b_c, b_h));
-  FLT4 result = a + b;
+  FLT4 result;
+  if (broadcastC_flag == 0) {
+    result = a + b;
+  } else if (broadcastC_flag == 1) {
+    result = a.x + b;
+  } else {
+    result = a + b.x;
+  }
   result = clamp(result, (FLT)(act_min), (FLT)(act_max));
   WRITE_IMAGE(output, (int2)(Y * output_shape.w + X, Z), result);
 }
 
 __kernel void BroadcastNHWC4Sub_IMG(__read_only image2d_t input_a, __read_only image2d_t input_b,
                                     __write_only image2d_t output, const int4 a_shape, const int4 b_shape,
-                                    const int4 output_shape, float act_min, float act_max) {
+                                    const int4 output_shape, const int broadcastC_flag, float act_min, float act_max) {
   int X = get_global_id(0);  // C4
   int Y = get_global_id(1);  // W
   int Z = get_global_id(2);  // H
@@ -303,14 +310,21 @@ __kernel void BroadcastNHWC4Sub_IMG(__read_only image2d_t input_a, __read_only i
   int b_w = Y < b_shape.z ? Y : b_shape.z - 1;
   int b_h = Z < b_shape.y ? Z : b_shape.y - 1;
   FLT4 b = READ_IMAGE(input_b, smp_none, (int2)(b_w * b_shape.w + b_c, b_h));
-  FLT4 result = a - b;
+  FLT4 result;
+  if (broadcastC_flag == 0) {
+    result = a - b;
+  } else if (broadcastC_flag == 1) {
+    result = a.x - b;
+  } else {
+    result = a - b.x;
+  }
   result = clamp(result, (FLT)(act_min), (FLT)(act_max));
   WRITE_IMAGE(output, (int2)(Y * output_shape.w + X, Z), result);
 }
 
 __kernel void BroadcastNHWC4Mul_IMG(__read_only image2d_t input_a, __read_only image2d_t input_b,
                                     __write_only image2d_t output, const int4 a_shape, const int4 b_shape,
-                                    const int4 output_shape, float act_min, float act_max) {
+                                    const int4 output_shape, const int broadcastC_flag, float act_min, float act_max) {
   int X = get_global_id(0);  // C4
   int Y = get_global_id(1);  // W
   int Z = get_global_id(2);  // H
@@ -325,14 +339,21 @@ __kernel void BroadcastNHWC4Mul_IMG(__read_only image2d_t input_a, __read_only i
   int b_w = Y < b_shape.z ? Y : b_shape.z - 1;
   int b_h = Z < b_shape.y ? Z : b_shape.y - 1;
   FLT4 b = READ_IMAGE(input_b, smp_none, (int2)(b_w * b_shape.w + b_c, b_h));
-  FLT4 result = a * b;
+  FLT4 result;
+  if (broadcastC_flag == 0) {
+    result = a * b;
+  } else if (broadcastC_flag == 1) {
+    result = a.x * b;
+  } else {
+    result = a * b.x;
+  }
   result = clamp(result, (FLT)(act_min), (FLT)(act_max));
   WRITE_IMAGE(output, (int2)(Y * output_shape.w + X, Z), result);
 }
 
 __kernel void BroadcastNHWC4Div_IMG(__read_only image2d_t input_a, __read_only image2d_t input_b,
                                     __write_only image2d_t output, const int4 a_shape, const int4 b_shape,
-                                    const int4 output_shape, float act_min, float act_max) {
+                                    const int4 output_shape, const int broadcastC_flag, float act_min, float act_max) {
   int X = get_global_id(0);  // C4
   int Y = get_global_id(1);  // W
   int Z = get_global_id(2);  // H
@@ -347,7 +368,14 @@ __kernel void BroadcastNHWC4Div_IMG(__read_only image2d_t input_a, __read_only i
   int b_w = Y < b_shape.z ? Y : b_shape.z - 1;
   int b_h = Z < b_shape.y ? Z : b_shape.y - 1;
   FLT4 b = READ_IMAGE(input_b, smp_none, (int2)(b_w * b_shape.w + b_c, b_h));
-  FLT4 result = a / b;
+  FLT4 result;
+  if (broadcastC_flag == 0) {
+    result = a / b;
+  } else if (broadcastC_flag == 1) {
+    result = a.x / b;
+  } else {
+    result = a / b.x;
+  }
   result = clamp(result, (FLT)(act_min), (FLT)(act_max));
   WRITE_IMAGE(output, (int2)(Y * output_shape.w + X, Z), result);
 }

mindspore/lite/src/runtime/kernel/opencl/kernel/activation.cc

@@ -70,6 +70,12 @@ int ActivationOpenClKernel::SetArgs() {
   if (type_ == ActivationType_LEAKY_RELU) {
     ocl_runtime_->SetKernelArg(kernel_, arg_idx++, alpha_);
   }
+  if (type_ == ActivationType_SIGMOID) {
+    int c4 = outShape.Slice;
+    int last_c4 = outShape.C % 4 == 0 ? 4 : outShape.C % 4;
+    ocl_runtime_->SetKernelArg(kernel_, arg_idx++, c4);
+    ocl_runtime_->SetKernelArg(kernel_, arg_idx++, last_c4);
+  }
   return RET_OK;
 }
 

mindspore/lite/src/runtime/kernel/opencl/kernel/arithmetic.cc

@@ -148,6 +148,13 @@ int ArithmeticOpenCLKernel::SetArgs() {
     auto out_shape = GetNHWCShape(out_tensors_[0]->shape());
     cl_int4 output_shape{out_shape[0], out_shape[1], out_shape[2], UP_DIV(out_shape[3], C4NUM)};
     ocl_runtime_->SetKernelArg(kernel_, arg_idx++, output_shape);
+    int broadcastC_flag = 0;  // do not need broadcast in C4
+    if (inputs_nhwc_shapes_[0][3] == 1 && inputs_nhwc_shapes_[1][3] != 1) {
+      broadcastC_flag = 1;  // BroadCast C4 in input0
+    } else if (inputs_nhwc_shapes_[0][3] != 1 && inputs_nhwc_shapes_[1][3] == 1) {
+      broadcastC_flag = 2;  // BroadCast C4 in input1
+    }
+    ocl_runtime_->SetKernelArg(kernel_, arg_idx++, broadcastC_flag);
   } else {
     cl_int2 output_shape{static_cast<int>(global_size_[0]), static_cast<int>(global_size_[1])};
     ocl_runtime_->SetKernelArg(kernel_, arg_idx++, output_shape);