pp

Author: nihui

src/layer/arm/gemm_arm.cpp

@@ -4357,6 +4357,8 @@ int Gemm_arm::create_pipeline(const Option& opt)
         {
             int C_elempack = constantM % 4 == 0 ? 4 : 1;
             convert_packing(C_data, CT_data, C_elempack, opt);
+            if (CT_data.empty())
+                return -100;
         }
 #endif // __ARM_NEON
 
@@ -4365,6 +4367,8 @@ int Gemm_arm::create_pipeline(const Option& opt)
         {
             Mat C2;
             C2.create_like(CT_data);
+            if (C2.empty())
+                return -100;
 
             const int size = CT_data.total() * CT_data.elempack;
             for (int i = 0; i < size; i++)
@@ -5079,6 +5083,8 @@ int Gemm_arm::create_pipeline_bf16s(const Option& opt)
         {
             int C_elempack = constantM % 4 == 0 ? 4 : 1;
             convert_packing(C_data, CT_data, C_elempack, opt);
+            if (CT_data.empty())
+                return -100;
         }
 #endif // __ARM_NEON
 
@@ -5087,6 +5093,8 @@ int Gemm_arm::create_pipeline_bf16s(const Option& opt)
         {
             Mat C2;
             C2.create_like(CT_data);
+            if (C2.empty())
+                return -100;
 
             const int size = CT_data.total() * CT_data.elempack;
             for (int i = 0; i < size; i++)
@@ -5210,6 +5218,8 @@ int Gemm_arm::forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Ma
             {
                 Mat CT_data;
                 CT_data.create_like(C, opt.workspace_allocator);
+                if (CT_data.empty())
+                    return -100;
 
                 const int size = C.total() * C.elempack;
                 for (int i = 0; i < size; i++)

src/layer/arm/gemm_arm_asimdhp.cpp

@@ -2801,13 +2801,17 @@ int Gemm_arm::create_pipeline_fp16sa(const Option& opt)
     if (constantC && constant_broadcast_type_C != -1)
     {
         cast_float32_to_float16(C_data, CT_data, opt);
+        if (CT_data.empty())
+            return -100;
 
         if (constant_broadcast_type_C == 3 && opt.use_packing_layout)
         {
             int C_elempack = constantM % 8 == 0 ? 8 : constantM % 4 == 0 ? 4 : 1;
             Mat tmp;
             convert_packing(CT_data, tmp, C_elempack, opt);
             CT_data = tmp;
+            if (CT_data.empty())
+                return -100;
         }
 
         // pre-multiply C with beta

src/layer/arm/gemm_arm_vfpv4.cpp

@@ -498,6 +498,8 @@ int Gemm_arm::create_pipeline_fp16s(const Option& opt)
         {
             int C_elempack = constantM % 4 == 0 ? 4 : 1;
             convert_packing(C_data, CT_data, C_elempack, opt);
+            if (CT_data.empty())
+                return -100;
         }
 #endif // __ARM_NEON
 
@@ -506,6 +508,8 @@ int Gemm_arm::create_pipeline_fp16s(const Option& opt)
         {
             Mat C2;
             C2.create_like(CT_data);
+            if (C2.empty())
+                return -100;
 
             const int size = CT_data.total() * CT_data.elempack;
             for (int i = 0; i < size; i++)
@@ -622,13 +626,17 @@ int Gemm_arm::forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Ma
                 Mat CT_data;
                 cast_float16_to_float32(C, CT_data);
                 C = CT_data;
+                if (C.empty())
+                    return -100;
             }
 
             // pre-multiply C with beta
             if (beta != 1.f)
             {
                 Mat CT_data;
                 CT_data.create_like(C, opt.workspace_allocator);
+                if (CT_data.empty())
+                    return -100;
 
                 const int size = C.total() * C.elempack;
                 for (int i = 0; i < size; i++)

src/layer/x86/gemm_x86.cpp

@@ -7500,6 +7500,8 @@ int Gemm_x86::create_pipeline(const Option& opt)
             int C_elempack = constantM % 4 == 0 ? 4 : 1;
 #endif
             convert_packing(C_data, CT_data, C_elempack, opt);
+            if (CT_data.empty())
+                return -100;
         }
 #endif // __SSE2__
 
@@ -7508,6 +7510,8 @@ int Gemm_x86::create_pipeline(const Option& opt)
         {
             Mat C2;
             C2.create_like(CT_data);
+            if (C2.empty())
+                return -100;
 
             const int size = CT_data.total() * CT_data.elempack;
             for (int i = 0; i < size; i++)