Fix bugs in multi-threads sgemm and fuse conv/add/bn/relu

src/framework/cl/cl_tensor.h

@@ -137,14 +137,18 @@ class CLTensor : TensorBase {
         : ptr_(clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                               size, reinterpret_cast<void *>(input), NULL)),
           size_(size),
+          capatity_(size),
           type_(type),
+          context_(context),
           command_queue_(command_queue) {}
 
     PlaceholderImpl(size_t size, std::type_index type, cl_context context,
                     cl_command_queue command_queue)
         : ptr_(clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, NULL)),
           size_(size),
+          capatity_(size),
           type_(type),
+          context_(context),
           command_queue_(command_queue) {}
 
     virtual size_t size() const { return size_; }
@@ -155,13 +159,25 @@ class CLTensor : TensorBase {
 
     virtual void set_type(std::type_index type) { type_ = type; }
 
+    virtual void resize(size_t size) {
+      if (size > capatity_) {
+        capatity_ = size;
+        ptr_.reset(
+            clCreateBuffer(context_, CL_MEM_READ_WRITE, capatity_, NULL, NULL));
+      }
+      size_ = size;
+    }
+
     std::unique_ptr<_cl_mem, CLMemDeleter> ptr_;
 
     size_t size_;
 
+    size_t capatity_;
+
     /* the current type of memory */
     std::type_index type_;
 
+    cl_context context_;
     cl_command_queue command_queue_;
   };
 };

src/framework/context.h

@@ -68,7 +68,8 @@ struct CPUContext {
 };
 
 inline void set_global_num_threads(int threads) {
-  CPUContext::Context()->set_num_threads(threads);
+  // CPUContext::Context()->set_num_threads(threads);
+  CPUContext::Context()->num_threads = threads;
 }
 
 inline int get_global_num_threads() {

src/operators/kernel/arm/convolution/conv_add_bn_relu_kernel.cpp

@@ -30,12 +30,14 @@ bool ConvAddBNReluKernel<CPU, float>::Init(
   const Tensor *variance = param->InputVariance();
   const Tensor *scale = param->InputScale();
   const Tensor *bias = param->InputBias();
+  const Tensor *bias1 = param->Bias();
   const float epsilon = param->Epsilon();
 
   auto mean_ptr = mean->data<float>();
   auto variance_ptr = variance->data<float>();
   auto scale_ptr = scale->data<float>();
   auto bias_ptr = bias->data<float>();
+  auto bias1_ptr = bias1->data<float>();
 
   const int C = mean->numel();
   float inv_std_ptr[C];
@@ -52,7 +54,8 @@ bool ConvAddBNReluKernel<CPU, float>::Init(
   auto new_bias_ptr = new_bias->mutable_data<float>({C});
   for (int i = 0; i < C; i++) {
     new_scale_ptr[i] = inv_std_ptr[i] * scale_ptr[i];
-    new_bias_ptr[i] = bias_ptr[i] - mean_ptr[i] * inv_std_ptr[i] * scale_ptr[i];
+    new_bias_ptr[i] = bias_ptr[i] + (bias1_ptr[i] - mean_ptr[i]) *
+                                        inv_std_ptr[i] * scale_ptr[i];
   }
   param->SetNewScale(new_scale);
   param->SetNewBias(new_bias);

src/operators/math/gemm/executor.h

@@ -107,8 +107,8 @@ class GemmExecutor : public Executor {
     //  gettimeofday(&tv_begin,NULL);
     if (M_ > N_) {
       int nblock = CeilDiv(N_, Strategy::out_width()) * Strategy::out_width();
-      lhs_worksize_ = sizeof(Itype) * lhs_tile_num_ * K_;
-      rhs_worksize_ = sizeof(Itype) * K_ * nblock * num_threads_;
+      lhs_worksize_ = sizeof(Itype) * lhs_tile_num_ * K_ * num_threads_;
+      rhs_worksize_ = sizeof(Itype) * K_ * nblock;
       out_worksize_ = sizeof(Otype) * lhs_tile_num_ * nblock * num_threads_;
       ldc_ = nblock;
     } else {
@@ -133,7 +133,7 @@ class GemmExecutor : public Executor {
     if (M_ > N_) {
       strategy_.pack_rhs(K_, N_, B, ldb, rhs_workspace_, true);
 
-      #pragma omp parallel for if (M_ > 128)
+      #pragma omp parallel for
       for (int lhs_block = 0; lhs_block < M_; lhs_block += lhs_tile_num_) {
         int lhs_range = std::min(M_ - lhs_block, lhs_tile_num_);
 #ifdef _OPENMP
@@ -165,7 +165,7 @@ class GemmExecutor : public Executor {
     } else {
       strategy_.pack_lhs(M_, K_, A, lda, lhs_workspace_, true);
 
-      #pragma omp parallel for if (N_ > 128)
+      #pragma omp parallel for
       for (int rhs_block = 0; rhs_block < N_; rhs_block += rhs_tile_num_) {
         int rhs_range = std::min(N_ - rhs_block, rhs_tile_num_);
 #ifdef _OPENMP