!5614 [MS][LITE][GPU]fix transpose softmax reshape bug

Merge pull request !5614 from chenzupeng/master-lite

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

@@ -1,11 +1,14 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 __constant sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
-__kernel void reshape(__read_only image2d_t src_data, __write_only image2d_t dst_data, int4 size) {
+__kernel void reshape(__read_only image2d_t src_data, __write_only image2d_t dst_data, int4 size, int4 size_out) {
   int X = get_global_id(0);
   int Y = get_global_id(1);
   int Z = get_global_id(2);
-  if (X >= size.x || Y >= size.y || Z >= size.z) {
+  if (X >= size_out.x || Y >= size_out.y || Z >= size_out.z) {
     return;
   }
-  WRITE_IMAGE(dst_data, (int2)(Y * size.z + Z, X), READ_IMAGE(src_data, smp_zero, (int2)(Y * size.z + Z, X)));
+  int out_index = X * size_out.y + Y;
+  int ih = out_index / size.y;
+  int iw = out_index % size.y;
+  WRITE_IMAGE(dst_data, (int2)(Y * size.z + Z, X), READ_IMAGE(src_data, smp_zero, (int2)(iw * size.z + Z, ih)));
 }

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

@@ -1,16 +1,21 @@
-__kernel void SoftMax_BUF(__global float4 *input, __global float4 *output, const int4 input_shape) {
-  int X = get_global_id(0);
-  int Y = get_global_id(1);
+#ifdef cl_khr_fp16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#endif
+__constant sampler_t smp_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
+__constant sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
+__kernel void SoftMax_BUF(__read_only image2d_t input, __global FLT4 *output, const int4 input_shape) {
+  int X = get_global_id(0);  // H
+  int Y = get_global_id(1);  // W
   int H = input_shape.x;
   int W = input_shape.y;
   int C = input_shape.z;
   int S = input_shape.w;
 
-  if (X >= W || Y >= H) return;
+  if (X >= H || Y >= W) return;
 
-  float sum = 0.0f;
+  FLT sum = 0.0f;
   for (int d = 0; d < S; ++d) {
-    float4 t = input[(Y * W + X * H) * C + d];
+    FLT4 t = READ_IMAGE(input, smp_zero, (int2)(Y * S + d, X));
     sum += exp(t.x);
     if (d * 4 + 1 < C) sum += exp(t.y);
     if (d * 4 + 2 < C) sum += exp(t.z);
@@ -18,15 +23,17 @@ __kernel void SoftMax_BUF(__global float4 *input, __global float4 *output, const
   }
 
   for (int d = 0; d < S; ++d) {
-    float4 t = input[(Y * W + X * H) * C + d];
+    FLT4 t = READ_IMAGE(input, smp_zero, (int2)(Y * S + d, X));
     t = exp(t) / sum;
-    float4 result = convert_float4(t);
-    output[(Y * W + X * H) * C + d] = result;
+    __global FLT *output_flt = (__global FLT *)output;
+    output_flt += (X * W + Y) * C + 4 * d;
+    output_flt[0] = t.x;
+    if (d * 4 + 1 < C) output_flt[1] += t.y;
+    if (d * 4 + 2 < C) output_flt[2] += t.z;
+    if (d * 4 + 3 < C) output_flt[3] += t.w;
   }
 }
 
-__constant sampler_t smp_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
-
 __kernel void SoftMax_IMG(__read_only image2d_t input, __write_only image2d_t output, const int4 input_shape) {
   int X = get_global_id(0);
   int Y = get_global_id(1);
@@ -34,7 +41,7 @@ __kernel void SoftMax_IMG(__read_only image2d_t input, __write_only image2d_t ou
 
   float sum = 0.0f;
   for (int d = 0; d < input_shape.w; ++d) {
-    float4 t = read_imagef(input, smp_none, (int2)(Y * input_shape.w + d, X));
+    FLT4 t = READ_IMAGE(input, smp_none, (int2)(Y * input_shape.w + d, X));
     sum += exp(t.x);
     if (d * 4 + 1 < input_shape.z) sum += exp(t.y);
     if (d * 4 + 2 < input_shape.z) sum += exp(t.z);
@@ -42,9 +49,112 @@ __kernel void SoftMax_IMG(__read_only image2d_t input, __write_only image2d_t ou
   }
 
   for (int d = 0; d < input_shape.w; ++d) {
-    float4 t = read_imagef(input, smp_none, (int2)(Y * input_shape.w + d, X));
+    FLT4 t = READ_IMAGE(input, smp_none, (int2)(Y * input_shape.w + d, X));
     t = exp(t) / sum;
-    float4 result = convert_float4(t);
-    write_imagef(output, (int2)(Y * input_shape.w + d, X), result);
+    FLT4 result = TO_FLT4(t);
+    WRITE_IMAGE(output, (int2)(Y * input_shape.w + d, X), result);
+  }
+}
+
+__kernel void SoftMax1x1_IMG(__read_only image2d_t input, __write_only image2d_t output, const FLT4 mask,
+                             const int slices, const int slices_x32) {
+  int tid = get_local_id(0);
+  int slices_count = 0;
+  int offset = 0;
+  FLT sum = 0.0f;
+  do {
+    int z = offset + tid;
+    if (z < slices) {
+      FLT4 mask_temp = z == slices - 1 ? mask : (FLT4)(1.0f);
+      FLT4 src = READ_IMAGE(input, smp_none, (int2)(0, 0));
+      sum += dot(mask_temp, exp(src));
+      offset += 32;
+    }
+    slices_count++;
+  } while (slices_count < slices_x32);
+
+  __local FLT4 tmp[8];
+  __local FLT *tmpx1 = (__local FLT *)tmp;
+  tmpx1[tid] = sum;
+  barrier(CLK_LOCAL_MEM_FENCE);
+  if (tid == 0) {
+    sum = dot((FLT4)(1.0f), tmp[0]);
+    sum += dot((FLT4)(1.0f), tmp[1]);
+    sum += dot((FLT4)(1.0f), tmp[2]);
+    sum += dot((FLT4)(1.0f), tmp[3]);
+    sum += dot((FLT4)(1.0f), tmp[4]);
+    sum += dot((FLT4)(1.0f), tmp[5]);
+    sum += dot((FLT4)(1.0f), tmp[6]);
+    sum += dot((FLT4)(1.0f), tmp[7]);
+    tmpx1[0] = 1.0f / sum;
+  }
+  barrier(CLK_LOCAL_MEM_FENCE);
+  sum = tmpx1[0];
+
+  offset = 0;
+  slices_count = 0;
+  do {
+    int z = offset + tid;
+    if (z < slices) {
+      FLT4 res = TO_FLT4(exp(READ_IMAGE(input, smp_none, (int2)(0, 0))) * sum);
+      WRITE_IMAGE(output, (int2)(0, 0), res);
+      offset += 32;
+    }
+    slices_count++;
+  } while (slices_count < slices_x32);
+}
+
+__kernel void SoftMax1x1_BUF(__read_only image2d_t input, __global FLT4 *output, const float4 mask, const int slices,
+                             const int slices_x32) {
+  int tid = get_local_id(0);
+  FLT sum = 0.0f;
+  for (size_t i = tid; i < slices - 1; i += 32) {
+    FLT4 src = READ_IMAGE(input, smp_zero, (int2)(i, 0));
+    sum += dot((FLT4)(1.0f), exp(src));
+  }
+  if ((slices - 1) % 32 == tid) {
+    FLT4 src = READ_IMAGE(input, smp_zero, (int2)(slices - 1, 0));
+
+    sum += dot(TO_FLT4(mask), exp(src));
+  }
+
+  __local FLT4 tmp[8];
+  __local FLT *tmpx1 = (__local FLT *)tmp;
+  tmpx1[tid] = sum;
+  barrier(CLK_LOCAL_MEM_FENCE);
+  if (tid == 0) {
+    sum = dot((FLT4)(1.0f), tmp[0]);
+    sum += dot((FLT4)(1.0f), tmp[1]);
+    sum += dot((FLT4)(1.0f), tmp[2]);
+    sum += dot((FLT4)(1.0f), tmp[3]);
+    sum += dot((FLT4)(1.0f), tmp[4]);
+    sum += dot((FLT4)(1.0f), tmp[5]);
+    sum += dot((FLT4)(1.0f), tmp[6]);
+    sum += dot((FLT4)(1.0f), tmp[7]);
+    tmpx1[0] = 1.0f / sum;
+  }
+  barrier(CLK_LOCAL_MEM_FENCE);
+  sum = tmpx1[0];
+  for (size_t i = tid; i < slices - 1; i += 32) {
+    FLT4 result = READ_IMAGE(input, smp_zero, (int2)(i, 0));
+    result = exp(result) * sum;
+    output[i] = result;
+  }
+  if ((slices - 1) % 32 == tid) {
+    FLT4 result = READ_IMAGE(input, smp_zero, (int2)(slices - 1, 0));
+    result = exp(result) * sum;
+    __global FLT4 *remain_ptr4 = output;
+    remain_ptr4 += slices - 1;
+    __global FLT *remain_ptr = (__global FLT *)remain_ptr4;
+    remain_ptr[0] = result.x;
+    if (mask.y > 0.f) {
+      remain_ptr[1] = result.y;
+    }
+    if (mask.z > 0.f) {
+      remain_ptr[2] = result.z;
+    }
+    if (mask.w > 0.f) {
+      remain_ptr[3] = result.w;
+    }
   }
 }

mindspore/lite/src/runtime/kernel/opencl/cl/softmax1x1.cl

@@ -1,104 +0,0 @@
-__constant sampler_t smp_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
-__constant sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
-// what is mask and args.slices_x32
-__kernel void SoftMax1x1_IMG(__read_only image2d_t input, __write_only image2d_t output, const float4 mask,
-                             const int slices, const int slices_x32) {
-  int tid = get_local_id(0);
-  int slices_count = 0;
-  int offset = 0;
-  float sum = 0.0f;
-  do {
-    int z = offset + tid;
-    if (z < slices) {
-      float4 mask_temp = z == slices - 1 ? mask : (float4)(1.0f);
-      float4 src = read_imagef(input, smp_none, (int2)(0, 0));
-      sum += dot(mask_temp, exp(src));
-      offset += 32;
-    }
-    slices_count++;
-  } while (slices_count < slices_x32);
-
-  __local float4 tmp[8];
-  __local float *tmpx1 = (__local float *)tmp;
-  tmpx1[tid] = sum;
-  barrier(CLK_LOCAL_MEM_FENCE);
-  if (tid == 0) {
-    sum = dot((float4)(1.0f), tmp[0]);
-    sum += dot((float4)(1.0f), tmp[1]);
-    sum += dot((float4)(1.0f), tmp[2]);
-    sum += dot((float4)(1.0f), tmp[3]);
-    sum += dot((float4)(1.0f), tmp[4]);
-    sum += dot((float4)(1.0f), tmp[5]);
-    sum += dot((float4)(1.0f), tmp[6]);
-    sum += dot((float4)(1.0f), tmp[7]);
-    tmpx1[0] = 1.0f / sum;
-  }
-  barrier(CLK_LOCAL_MEM_FENCE);
-  sum = tmpx1[0];
-
-  offset = 0;
-  slices_count = 0;
-  do {
-    int z = offset + tid;
-    if (z < slices) {
-      float4 res = convert_float4(exp(read_imagef(input, smp_none, (int2)(0, 0))) * sum);
-      write_imagef(output, (int2)(0, 0), res);
-      offset += 32;
-    }
-    slices_count++;
-  } while (slices_count < slices_x32);
-}
-
-__kernel void SoftMax1x1_BUF(__read_only image2d_t input, __global float4 *output, const float4 mask, const int slices,
-                             const int slices_x32) {
-  int tid = get_local_id(0);
-  float sum = 0.0f;
-  for (size_t i = tid; i < slices - 1; i += 32) {
-    float4 src = read_imagef(input, smp_zero, (int2)(i, 0));
-    sum += dot((float4)(1.0f), exp(src));
-  }
-  if ((slices - 1) % 32 == tid) {
-    float4 src = read_imagef(input, smp_zero, (int2)(slices - 1, 0));
-    sum += dot(mask, exp(src));
-  }
-
-  __local float4 tmp[8];
-  __local float *tmpx1 = (__local float *)tmp;
-  tmpx1[tid] = sum;
-  barrier(CLK_LOCAL_MEM_FENCE);
-  if (tid == 0) {
-    sum = dot((float4)(1.0f), tmp[0]);
-    sum += dot((float4)(1.0f), tmp[1]);
-    sum += dot((float4)(1.0f), tmp[2]);
-    sum += dot((float4)(1.0f), tmp[3]);
-    sum += dot((float4)(1.0f), tmp[4]);
-    sum += dot((float4)(1.0f), tmp[5]);
-    sum += dot((float4)(1.0f), tmp[6]);
-    sum += dot((float4)(1.0f), tmp[7]);
-    tmpx1[0] = 1.0f / sum;
-  }
-  barrier(CLK_LOCAL_MEM_FENCE);
-  sum = tmpx1[0];
-  for (size_t i = tid; i < slices - 1; i += 32) {
-    float4 result = read_imagef(input, smp_zero, (int2)(i, 0));
-    result = exp(result) * sum;
-    output[i] = result;
-  }
-  if ((slices - 1) % 32 == tid) {
-    float4 result = read_imagef(input, smp_zero, (int2)(slices - 1, 0));
-    result = exp(result) * sum;
-    __global float4 *remain_ptr4 = output;
-    remain_ptr4 += slices - 1;
-    __global float *remain_ptr = (__global float *)remain_ptr4;
-    remain_ptr[0] = result.x;
-    if (mask.y > 0.f) {
-      remain_ptr[1] = result.y;
-    }
-    if (mask.z > 0.f) {
-      remain_ptr[2] = result.z;
-    }
-    if (mask.w > 0.f) {
-      remain_ptr[3] = result.w;
-    }
-  }
-}

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

@@ -1,4 +1,6 @@
+#ifdef cl_khr_fp16
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#endif
 __constant sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
 __kernel void transpose_IMG(__read_only image2d_t src_data, __write_only image2d_t dst_data, int2 HW, int2 C) {
   int X = get_global_id(0);
@@ -41,7 +43,7 @@ __kernel void transpose_IMG(__read_only image2d_t src_data, __write_only image2d
   WRITE_IMAGE(dst_data, (int2)(X, 4 * Y + 3), result[3]);
 }
 
-__kernel void transpose_BUF(__read_only image2d_t src_data, global FLT4 *dst_data, int2 HW, int2 C) {
+__kernel void transpose_BUF(__read_only image2d_t src_data, global FLT4 *dst_data, int2 HW, int2 C, int W) {
   int X = get_global_id(0);
   int Y = get_global_id(1);
   if (X >= HW.y || Y >= C.y) {
@@ -52,10 +54,10 @@ __kernel void transpose_BUF(__read_only image2d_t src_data, global FLT4 *dst_dat
   result[1] = (FLT4)(0.0f);
   result[2] = (FLT4)(0.0f);
   result[3] = (FLT4)(0.0f);
-  FLT4 x0 = READ_IMAGE(src_data, smp_zero, (int2)(Y, 4 * X));
-  FLT4 x1 = READ_IMAGE(src_data, smp_zero, (int2)(Y, 4 * X + 1));
-  FLT4 x2 = READ_IMAGE(src_data, smp_zero, (int2)(Y, 4 * X + 2));
-  FLT4 x3 = READ_IMAGE(src_data, smp_zero, (int2)(Y, 4 * X + 3));
+  FLT4 x0 = READ_IMAGE(src_data, smp_zero, (int2)((4 * X) % W * C.y + Y, (4 * X) / W));
+  FLT4 x1 = READ_IMAGE(src_data, smp_zero, (int2)((4 * X + 1) % W * C.y + Y, (4 * X + 1) / W));
+  FLT4 x2 = READ_IMAGE(src_data, smp_zero, (int2)((4 * X + 2) % W * C.y + Y, (4 * X + 2) / W));
+  FLT4 x3 = READ_IMAGE(src_data, smp_zero, (int2)((4 * X + 3) % W * C.y + Y, (4 * X + 3) / W));
   result[0].x = x0.x;
   result[0].y = x1.x;
   result[0].z = x2.x;

mindspore/lite/src/runtime/kernel/opencl/kernel/pooling2d.cc

@@ -65,7 +65,8 @@ int PoolingOpenCLKernel::Init() {
   kernel_ = ocl_runtime->GetKernelFromBinary(kernel_name);
 #else
   if (out_mem_type_ == OpenCLMemType::BUF) {
-    kernel_name += "_BUF";
+    MS_LOG(ERROR) << "buffer output not support yet.";
+    return RET_ERROR;
   } else {
     kernel_name += "_IMG";
   }

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

@@ -68,10 +68,16 @@ int ReshapeOpenCLKernel::ReSize() { return RET_OK; }
 
 int ReshapeOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size) {
   size_t im_dst_x, im_dst_y;
-  std::vector<int> shapex = in_tensors_[0]->shape();
-  int h = shapex[1];
-  int w = shapex[2];
-  int c = shapex[3];
+  std::vector<int> shapex = out_tensors_[0]->shape();
+  int h, w, c;
+  if (shapex.size() == 2) {
+    h = w = 1;
+    c = shapex[1];
+  } else {
+    h = shapex[1];
+    w = shapex[2];
+    c = shapex[3];
+  }
   im_dst_x = w * UP_DIV(c, C4NUM);
   im_dst_y = h;
   size_t img_dtype = CL_FLOAT;
@@ -91,13 +97,23 @@ int ReshapeOpenCLKernel::Run() {
   int w = shapex[2];
   int c = shapex[3];
   int c4 = UP_DIV(c, C4NUM);
+  int oh, ow;
+  if (out_tensors_[0]->shape().size() == 2) {
+    oh = ow = 1;
+  } else {
+    oh = out_tensors_[0]->shape()[1];
+    ow = out_tensors_[0]->shape()[2];
+  }
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   std::vector<size_t> local = {};
-  std::vector<size_t> global = {(size_t)h, (size_t)w, (size_t)c4};
+  std::vector<size_t> global = {(size_t)oh, (size_t)ow, (size_t)c4};
   cl_int4 size = {h, w, c4, 1};
-  ocl_runtime->SetKernelArg(kernel_, 0, in_tensors_[0]->Data());
-  ocl_runtime->SetKernelArg(kernel_, 1, out_tensors_[0]->Data());
-  ocl_runtime->SetKernelArg(kernel_, 2, size);
+  cl_int4 size_out = {oh, ow, c4, 1};
+  int arg_idx = 0;
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, in_tensors_[0]->Data());
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data());
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, size);
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, size_out);
   ocl_runtime->RunKernel(kernel_, global, local, nullptr);
   return RET_OK;
 }

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

@@ -23,7 +23,6 @@
 #include "src/runtime/kernel/opencl/utils.h"
 #ifndef PROGRAM_WITH_IL
 #include "src/runtime/kernel/opencl/cl/softmax.cl.inc"
-#include "src/runtime/kernel/opencl/cl/softmax1x1.cl.inc"
 #endif
 
 using mindspore::kernel::KERNEL_ARCH::kGPU;
@@ -42,8 +41,8 @@ std::vector<float> SoftmaxOpenCLKernel::GetMaskForLastChannel(int channels) {
 }
 
 int SoftmaxOpenCLKernel::InitGlobalSize() {
-  const size_t global_x = out_tensors_[0]->Height();
-  const size_t global_y = out_tensors_[0]->Width();
+  const size_t global_x = out_tensors_[0]->shape()[1];
+  const size_t global_y = out_tensors_[0]->shape()[2];
   const size_t global_z = 1;
   global_size_ = {global_x, global_y, global_z};
   return lite::RET_OK;
@@ -74,11 +73,10 @@ int SoftmaxOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size)
     im_dst_x = out_tensors_[0]->Width() * CO4;
     im_dst_y = out_tensors_[0]->Height();
   }
-#ifdef ENABLE_FP16
-  size_t img_dtype = CL_HALF_FLOAT;
-#else
   size_t img_dtype = CL_FLOAT;
-#endif
+  if (enable_fp16_) {
+    img_dtype = CL_HALF_FLOAT;
+  }
   img_size->clear();
   std::vector<size_t> vec{im_dst_x, im_dst_y, img_dtype};
   *img_size = vec;
@@ -90,27 +88,28 @@ int SoftmaxOpenCLKernel::Init() {
   std::string program_name = "SoftMax";
   std::string source = softmax_source;
   runtime_ = lite::opencl::OpenCLRuntime::GetInstance();
+  enable_fp16_ = runtime_->GetFp16Enable();
   // framework not set this param yet! just use default.
-  if (parameter_->axis_ == -1) {
-    parameter_->axis_ = 1;
-  }
-  if (in_tensors_[0]->shape().size() == 4 && parameter_->axis_ == 3) {
+  if (in_tensors_[0]->shape().size() == 4) {
     // support 4d tensor
     onexone_flag_ = false;
-  } else if (in_tensors_[0]->shape().size() == 2 && parameter_->axis_ == 1) {
+  } else if (in_tensors_[0]->shape().size() == 2) {
     // support 2d tensor
     kernel_name += "1x1";
     program_name += "1x1";
-    source = softmax1x1_source;
     onexone_flag_ = true;
   } else {
-    MS_LOG(EXCEPTION) << "Init `Softmax` kernel failed: Unsupported axis: " << parameter_->axis_;
+    MS_LOG(ERROR) << "Init `Softmax` kernel failed: Unsupported shape size: " << in_tensors_[0]->shape().size();
+    return RET_ERROR;
   }
 #ifdef PROGRAM_WITH_IL
   kernel_ = ocl_runtime->GetKernelFromBinary(kernel_name);
 #else
   if (!is_image_out_) {
     out_mem_type_ = OpenCLMemType::BUF;
+  } else {
+    MS_LOG(ERROR) << "image2d output not support yet.";
+    return RET_ERROR;
   }
   if (out_mem_type_ == OpenCLMemType::BUF) {
     kernel_name += "_BUF";
@@ -124,12 +123,23 @@ int SoftmaxOpenCLKernel::Init() {
   runtime_->BuildKernel(kernel_, program_name, kernel_name, build_options);
 #endif
   in_ori_format_ = in_tensors_[0]->GetFormat();
-  in_tensors_[0]->SetFormat(schema::Format_NHWC4);
   out_ori_format_ = out_tensors_[0]->GetFormat();
-  out_tensors_[0]->SetFormat(schema::Format_NHWC4);
-  if (!is_image_out_) {
-    out_ori_format_ = schema::Format_NC;
-    out_tensors_[0]->SetFormat(schema::Format_NC);
+  if (in_tensors_[0]->shape().size() == 2) {
+    in_tensors_[0]->SetFormat(schema::Format_NC4);
+  } else {
+    in_tensors_[0]->SetFormat(schema::Format_NHWC4);
+  }
+
+  if (is_image_out_) {
+    if (out_tensors_[0]->shape().size() == 2) {
+      out_ori_format_ = schema::Format_NC;
+      out_tensors_[0]->SetFormat(schema::Format_NC4);
+    } else {
+      out_ori_format_ = schema::Format_NHWC;
+      out_tensors_[0]->SetFormat(schema::Format_NHWC4);
+    }
+  } else {
+    out_tensors_[0]->SetFormat(out_ori_format_);
   }
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
   return lite::RET_OK;
@@ -147,17 +157,25 @@ int SoftmaxOpenCLKernel::Run() {
     cl_float4 mask = {mask_[0], mask_[1], mask_[2], mask_[3]};
 
     runtime_->SetKernelArg(kernel_, arg_idx++, in_tensors_[0]->Data());
-    runtime_->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data());
+    if (is_image_out_) {
+      runtime_->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data());
+    } else {
+      runtime_->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data(), lite::opencl::MemType::BUF);
+    }
     runtime_->SetKernelArg(kernel_, arg_idx++, mask);
     runtime_->SetKernelArg(kernel_, arg_idx++, slices);
     runtime_->SetKernelArg(kernel_, arg_idx, slices_x32);
     SetWorkGroupSize1x1();
   } else {
-    int slices = UP_DIV(out_tensors_[0]->Channel(), C4NUM);
-    cl_int4 input_shape = {in_tensors_[0]->Height(), in_tensors_[0]->Width(), in_tensors_[0]->Channel(), slices};
+    int slices = UP_DIV(out_tensors_[0]->shape()[3], C4NUM);
+    cl_int4 input_shape = {in_tensors_[0]->shape()[1], in_tensors_[0]->shape()[2], in_tensors_[0]->shape()[3], slices};
 
     runtime_->SetKernelArg(kernel_, arg_idx++, in_tensors_[0]->Data());
-    runtime_->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data());
+    if (is_image_out_) {
+      runtime_->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data());
+    } else {
+      runtime_->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data(), lite::opencl::MemType::BUF);
+    }
     runtime_->SetKernelArg(kernel_, arg_idx, input_shape);
     SetWorkGroupSize();
   }
@@ -193,4 +211,5 @@ kernel::LiteKernel *OpenCLSoftMaxKernelCreator(const std::vector<lite::tensor::T
 }
 
 REG_KERNEL(kGPU, kNumberTypeFloat32, PrimitiveType_SoftMax, OpenCLSoftMaxKernelCreator)
+REG_KERNEL(kGPU, kNumberTypeFloat16, PrimitiveType_SoftMax, OpenCLSoftMaxKernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/opencl/kernel/softmax.h

@@ -52,6 +52,7 @@ class SoftmaxOpenCLKernel : public OpenCLKernel {
   std::vector<size_t> local_size_;
   std::vector<size_t> global_size_;
   bool is_image_out_{false};
+  bool enable_fp16_{false};
 };
 
 }  // namespace mindspore::kernel

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

@@ -119,11 +119,9 @@ int ToFormatOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size
     im_dst_x = w * UP_DIV(c, C4NUM);
     im_dst_y = h;
   } else if (out_tensors_[0]->GetFormat() == schema::Format_NC4) {
-    const int h = 1;
-    const int w = 1;
     int c = shapex[1];
-    im_dst_x = w * UP_DIV(c, C4NUM);
-    im_dst_y = h;
+    im_dst_x = UP_DIV(c, C4NUM);
+    im_dst_y = 1;
   } else {
     MS_LOG(ERROR) << "Unsupported format. " << out_tensors_[0]->GetFormat();
     return RET_ERROR;

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

@@ -69,7 +69,7 @@ int TransposeOpenCLKernel::ReSize() { return RET_OK; }
 
 int TransposeOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size) {
   size_t im_dst_x, im_dst_y;
-  im_dst_x = UP_DIV(out_tensors_[0]->Height() * out_tensors_[0]->Width(), C4NUM);
+  im_dst_x = out_tensors_[0]->Height() * UP_DIV(out_tensors_[0]->Width(), C4NUM);
   im_dst_y = out_tensors_[0]->Channel();
   size_t img_dtype = CL_FLOAT;
   if (enable_fp16_) {
@@ -96,10 +96,12 @@ int TransposeOpenCLKernel::Run() {
 
   cl_int2 HW = {h * w, hw4};
   cl_int2 C = {c, c4};
-  ocl_runtime->SetKernelArg(kernel_, 0, in_tensors_[0]->Data());
-  ocl_runtime->SetKernelArg(kernel_, 1, out_tensors_[0]->Data());
-  ocl_runtime->SetKernelArg(kernel_, 2, HW);
-  ocl_runtime->SetKernelArg(kernel_, 3, C);
+  int arg_idx = 0;
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, in_tensors_[0]->Data());
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data());
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, HW);
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, C);
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, w);
   ocl_runtime->RunKernel(kernel_, global, local, nullptr);
   return RET_OK;
 }

mindspore/lite/test/ut/src/runtime/kernel/opencl/reshape_tests.cc

@@ -86,14 +86,14 @@ void RunTestCaseReshape(const std::vector<int> &shape, void *input_data, void *o
   inputs[0]->SetData(nullptr);
   outputs[0]->SetData(nullptr);
 
-  MS_LOG(INFO) << "Test ReshapeFp32 passed";
+  MS_LOG(INFO) << "Test Reshape passed";
   lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 TEST_F(TestReshapeOpenCL, ReshapeFp32) {
   int c = 7;
   std::vector<int> shape = {c};
-  std::vector<float> input_data = {0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f};
+  std::vector<float> input_data = {0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
   std::vector<float> output_data = {0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
 
   RunTestCaseReshape(shape, input_data.data(), output_data.data(), false);
@@ -102,7 +102,7 @@ TEST_F(TestReshapeOpenCL, ReshapeFp32) {
 TEST_F(TestReshapeOpenCL, ReshapeFp16) {
   int c = 7;
   std::vector<int> shape = {c};
-  std::vector<float16_t> input_data = {0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f};
+  std::vector<float16_t> input_data = {0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
   std::vector<float16_t> output_data = {0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
 
   RunTestCaseReshape(shape, input_data.data(), output_data.data(), true);

mindspore/lite/test/ut/src/runtime/kernel/opencl/softmax_tests.cc

@@ -17,94 +17,134 @@
 #include <memory>
 #include "mindspore/core/utils/log_adapter.h"
 #include "common/common_test.h"
+#include "mindspore/lite/src/common/file_utils.h"
 #include "mindspore/lite/src/runtime/opencl/opencl_runtime.h"
 #include "mindspore/lite/src/runtime/kernel/opencl/subgraph_opencl_kernel.h"
 #include "mindspore/lite/src/runtime/kernel/opencl/kernel/softmax.h"
 #include "mindspore/lite/test/ut/src/runtime/kernel/opencl/utils_tests.h"
 
 namespace mindspore {
+class TestSoftmaxOpenCL : public mindspore::CommonTest {
+ public:
+  TestSoftmaxOpenCL() {}
+};
 
-class TestSoftmaxOpenCL : public mindspore::CommonTest {};
-
-void RunTestCase(std::vector<int> input_shape, std::vector<int> output_shape, std::string input_file,
-                 std::string expect_file, SoftmaxParameter *param, schema::Format format) {
+void RunTestCaseSoftmax(const std::vector<int> &shape, void *input_data, void *output_data, bool enable_fp16) {
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   ocl_runtime->Init();
+  size_t dtype_size = sizeof(float);
+  if (enable_fp16) {
+    ocl_runtime->SetFp16Enable(true);
+    dtype_size = sizeof(float16_t);
+  }
   auto allocator = ocl_runtime->GetAllocator();
-
-  // define tensor
-  MS_LOG(INFO) << "defineTensor";
-  auto data_type = kNumberTypeFloat32;
-  auto tensorType = schema::NodeType_ValueNode;
-  auto input_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, format, tensorType);
-  auto output_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, output_shape, format, tensorType);
-  if (input_tensor == nullptr) {
-    MS_LOG(ERROR) << "input tensor null";
+  int n, h, w, c;
+  bool is_2d = false;
+  if (shape.size() == 2) {
+    is_2d = true;
+    h = w = 1;
+    n = shape[0];
+    c = shape[1];
+  } else {
+    n = shape[0];
+    h = shape[1];
+    w = shape[2];
+    c = shape[3];
+  }
+  std::vector<int> input_shape = {n, h, w, c};
+  if (is_2d) {
+    input_shape = {n, c};
+  }
+  auto input_format = is_2d ? schema::Format_NC : schema::Format_NHWC;
+  auto input_dtype = enable_fp16 ? kNumberTypeFloat16 : kNumberTypeFloat32;
+  auto tensor_x_ptr = std::make_unique<lite::tensor::Tensor>(TypeId(input_dtype), input_shape, input_format);
+  auto tensor_x = tensor_x_ptr.get();
+  if (tensor_x == nullptr) {
+    MS_LOG(ERROR) << "tensor_x create error.";
     return;
   }
-  if (output_tensor == nullptr) {
-    MS_LOG(ERROR) << "output tensor null";
+  auto tensor_out_ptr = std::make_unique<lite::tensor::Tensor>(TypeId(input_dtype), input_shape, input_format);
+  auto tensor_out = tensor_out_ptr.get();
+  if (tensor_out == nullptr) {
+    MS_LOG(ERROR) << "tensor_out create error.";
     return;
   }
-  std::vector<lite::tensor::Tensor *> inputs{input_tensor};
-  std::vector<lite::tensor::Tensor *> outputs{output_tensor};
-
-  // run
-  MS_LOG(INFO) << "NewOpenCLKernel";
-  auto *kernel = new kernel::SoftmaxOpenCLKernel(reinterpret_cast<OpParameter *>(param), inputs, outputs);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel null";
+  std::vector<lite::tensor::Tensor *> inputs{tensor_x};
+  std::vector<lite::tensor::Tensor *> outputs{tensor_out};
+  auto arith_kernel_ptr = std::make_unique<kernel::SoftmaxOpenCLKernel>(nullptr, inputs, outputs);
+  auto arith_kernel = arith_kernel_ptr.get();
+  if (arith_kernel == nullptr) {
+    MS_LOG(ERROR) << "arith_kernel create error.";
     return;
   }
-  MS_LOG(INFO) << "KernelInit";
-  kernel->Init();
+  arith_kernel->Init();
 
-  std::vector<kernel::LiteKernel *> kernels{kernel};
   inputs[0]->MallocData(allocator);
-  auto *pGraph = new (std::nothrow) kernel::SubGraphOpenCLKernel(inputs, outputs, kernels, kernels, kernels);
+
+  std::vector<kernel::LiteKernel *> kernels{arith_kernel};
+  auto pGraph_ptr = std::make_unique<kernel::SubGraphOpenCLKernel>(inputs, outputs, kernels, kernels, kernels);
+  auto pGraph = pGraph_ptr.get();
   if (pGraph == nullptr) {
-    MS_LOG(ERROR) << "pGraph null";
+    MS_LOG(ERROR) << "pGraph create error.";
     return;
   }
-  MS_LOG(INFO) << "pGraphinit";
   pGraph->Init();
-
-  // load data
-  MS_LOG(INFO) << "load data1";
-  LoadTestData(input_tensor->Data(), input_tensor->Size(), input_file);
-  auto *input_data = reinterpret_cast<float *>(input_tensor->Data());
-  printf("\ninput[0:10]:");
-  for (int i = 0; i < 10; i++) {
-    printf("[%d]:%.3f ", i, input_data[i]);
-  }
-  printf("\n\n");
-
-  MS_LOG(INFO) << "Run";
+  memcpy(inputs[0]->Data(), input_data, inputs[0]->ElementsNum() * dtype_size);
   pGraph->Run();
 
-  MS_LOG(INFO) << "compare result";
-  CompareOutput(output_tensor, expect_file, static_cast<float>(1e-5));
-  for (auto tensor : inputs) {
-    delete tensor;
+  if (enable_fp16) {
+    CompareOutput(outputs[0]->Data(), output_data, outputs[0]->ElementsNum(), static_cast<float16_t>(1e-3), 2e-2);
+  } else {
+    CompareOutput(outputs[0]->Data(), output_data, outputs[0]->ElementsNum(), static_cast<float>(1e-5));
   }
-  for (auto tensor : outputs) {
-    delete tensor;
-  }
-  delete kernel;
-  delete pGraph;
+  inputs[0]->SetData(nullptr);
+  outputs[0]->SetData(nullptr);
+
+  MS_LOG(INFO) << "Test Softmax passed";
   lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
-TEST_F(TestSoftmaxOpenCL, Softmax_1) {
-  std::vector<int> input_shape = {1, 2, 2, 8};
-  std::vector<int> output_shape = {1, 2, 2, 8};
-  std::string input_file = "softmax_in.bin";
-  std::string expect_file = "softmax_out.bin";
-  auto param = new (std::nothrow) SoftmaxParameter;
-  param->axis_ = 3;
-  schema::Format format = schema::Format_NHWC4;
+TEST_F(TestSoftmaxOpenCL, Softmax2DFp32) {
+  int n = 1;
+  int c = 10;
+  std::vector<int> shape = {n, c};
+  std::vector<float> input_data = {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f};
+  std::vector<float> output_data = {0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f};
+
+  RunTestCaseSoftmax(shape, input_data.data(), output_data.data(), false);
+}
+
+TEST_F(TestSoftmaxOpenCL, Softmax2DFp16) {
+  int n = 1;
+  int c = 10;
+  std::vector<int> shape = {n, c};
+  std::vector<float16_t> input_data = {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f};
+  std::vector<float16_t> output_data = {0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f, 0.1f};
 
-  RunTestCase(input_shape, output_shape, input_file, expect_file, param, format);
+  RunTestCaseSoftmax(shape, input_data.data(), output_data.data(), true);
 }
 
+TEST_F(TestSoftmaxOpenCL, Softmax4DFp32) {
+  int n = 1;
+  int h = 2;
+  int w = 1;
+  int c = 5;
+  std::vector<int> shape = {n, h, w, c};
+  std::vector<float> input_data = {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f};
+  std::vector<float> output_data = {0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f};
+
+  RunTestCaseSoftmax(shape, input_data.data(), output_data.data(), false);
+}
+
+TEST_F(TestSoftmaxOpenCL, Softmax4DFp16) {
+  int n = 1;
+  int h = 2;
+  int w = 1;
+  int c = 5;
+  std::vector<int> shape = {n, h, w, c};
+  std::vector<float16_t> input_data = {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f};
+  std::vector<float16_t> output_data = {0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f, 0.2f};
+
+  RunTestCaseSoftmax(shape, input_data.data(), output_data.data(), true);
+}
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/transpose_tests.cc

@@ -117,8 +117,8 @@ TEST_F(TestTransposeOpenCL, TransposeFp32) {
 }
 
 TEST_F(TestTransposeOpenCL, TransposeFp16) {
-  int h = 4;
-  int w = 1;
+  int h = 2;
+  int w = 2;
   int c = 3;
   std::vector<int> shape = {h, w, c};
   std::vector<float16_t> input_data = {0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f, 11.0f};