discard riscv weight memory (#3874)

* discard riscv innerproduct weight
* drop riscv conv convdw weight
* drop riscv deconv deconvdw weight

src/layer/arm/convolutiondepthwise_arm.cpp

@@ -130,6 +130,11 @@ int ConvolutionDepthWise_arm::create_pipeline(const Option& opt)
 
             ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
 
+            if (opt.lightmode)
+            {
+                weight_data.release();
+            }
+
             return 0;
         }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
@@ -153,6 +158,11 @@ int ConvolutionDepthWise_arm::create_pipeline(const Option& opt)
                 ncnn::cast_float32_to_bfloat16(weight_data, weight_data_tm, opt);
             }
 
+            if (opt.lightmode)
+            {
+                weight_data.release();
+            }
+
             return 0;
         }
 #endif // NCNN_BF16
@@ -163,8 +173,6 @@ int ConvolutionDepthWise_arm::create_pipeline(const Option& opt)
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, group);
             convert_packing(weight_data_r2, weight_data_tm, 4, opt);
-
-            return 0;
         }
 #endif // __ARM_NEON
 
@@ -173,24 +181,32 @@ int ConvolutionDepthWise_arm::create_pipeline(const Option& opt)
             if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
                 weight_data_tm = weight_data;
-                return 0;
             }
-            if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
+            else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
                 weight_data_tm = weight_data;
-                return 0;
             }
-            if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
+            else if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
                 weight_data_tm = weight_data;
-                return 0;
             }
-            if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
+            else if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
                 weight_data_tm = weight_data;
-                return 0;
+            }
+            else
+            {
+                // group convolution
+                create_group_ops(opt);
             }
         }
+
+        if (opt.lightmode)
+        {
+            weight_data.release();
+        }
+
+        return 0;
     }
 
     // group convolution
@@ -1591,6 +1607,11 @@ int ConvolutionDepthWise_arm::create_pipeline_int8_arm(const Option& opt)
             weight_data_tm = weight_data;
         }
 
+        if (opt.lightmode)
+        {
+            weight_data.release();
+        }
+
         return 0;
     }
 

src/layer/mips/convolutiondepthwise_mips.cpp

@@ -83,6 +83,11 @@ int ConvolutionDepthWise_mips::create_pipeline(const Option& opt)
             weight_data_tm = weight_data;
         }
 
+        if (opt.lightmode)
+        {
+            weight_data.release();
+        }
+
         return 0;
     }
 

src/layer/riscv/convolution_riscv.cpp

@@ -88,7 +88,7 @@ Convolution_riscv::Convolution_riscv()
     activation = 0;
 }
 
-static void convolution_transform_kernel_packed_rvv(const Mat& weight_data, Mat& weight_data_packed, int num_input, int num_output, int kernel_w, int kernel_h, int elempack, int out_elempack)
+static void convolution_transform_kernel_packed_rvv(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, int kernel_w, int kernel_h, int elempack, int out_elempack)
 {
     const int maxk = kernel_w * kernel_h;
 
@@ -97,11 +97,11 @@ static void convolution_transform_kernel_packed_rvv(const Mat& weight_data, Mat&
     {
         Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);
 
-        weight_data_packed.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)4u * elempack * out_elempack, elempack * out_elempack);
+        weight_data_tm.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)4u * elempack * out_elempack, elempack * out_elempack);
 
         for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
         {
-            float* g00 = weight_data_packed.channel(q / out_elempack);
+            float* g00 = weight_data_tm.channel(q / out_elempack);
 
             for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
             {
@@ -131,6 +131,14 @@ int Convolution_riscv::create_pipeline(const Option& opt)
 
     activation = create_activation_layer(activation_type, activation_params, opt);
 
+#if NCNN_INT8
+    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
+    {
+        // TODO implement int8
+        return 0;
+    }
+#endif
+
 #if __riscv_vector && __riscv_zfh
     if (opt.use_fp16_storage)
     {
@@ -166,14 +174,14 @@ int Convolution_riscv::create_pipeline(const Option& opt)
         }
         else
         {
-            convolution_transform_kernel_packed_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
+            convolution_transform_kernel_packed_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
         }
     }
 
     // pack1ton
     if (elempack == 1 && out_elempack == packn)
     {
-        convolution_transform_kernel_packed_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
+        convolution_transform_kernel_packed_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
     }
 
     // packnto1
@@ -181,19 +189,19 @@ int Convolution_riscv::create_pipeline(const Option& opt)
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else
         {
-            convolution_transform_kernel_packed_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
+            convolution_transform_kernel_packed_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
         }
     }
 #endif // __riscv_vector
@@ -203,7 +211,7 @@ int Convolution_riscv::create_pipeline(const Option& opt)
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            convolution_im2col_sgemm_transform_kernel_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else if (opt.use_winograd_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
@@ -218,10 +226,19 @@ int Convolution_riscv::create_pipeline(const Option& opt)
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_transform_kernel_rvv(weight_data, weight_data_packed, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
+        }
+        else
+        {
+            weight_data_tm = weight_data;
         }
     }
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -240,7 +257,7 @@ int Convolution_riscv::destroy_pipeline(const Option& opt)
 int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
 {
 #if NCNN_INT8
-    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
+    if (opt.use_int8_inference && int8_scale_term)
     {
         Mat bottom_blob_unpacked = bottom_blob;
         if (bottom_blob.elempack != 1)
@@ -266,9 +283,44 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     }
 #endif
 
-    if (bottom_blob.dims != 3)
+    // flattened blob, implement as InnerProduct
+    if (bottom_blob.dims == 1 && kernel_w == 1 && kernel_h == 1)
     {
-        return Convolution::forward(bottom_blob, top_blob, opt);
+        Mat bottom_blob_3d;
+        if (bottom_blob.elemsize % 16 == 0)
+        {
+            bottom_blob_3d = bottom_blob;
+            bottom_blob_3d.dims = 3;
+            bottom_blob_3d.w = 1;
+            bottom_blob_3d.h = 1;
+            bottom_blob_3d.c = bottom_blob.w;
+            bottom_blob_3d.cstep = 1;
+        }
+        else
+        {
+            bottom_blob_3d = bottom_blob.reshape(1, 1, bottom_blob.w, opt.workspace_allocator);
+        }
+
+        Mat top_blob_3d;
+        int ret = forward(bottom_blob_3d, top_blob_3d, opt);
+        if (ret != 0)
+            return ret;
+
+        if (top_blob_3d.elemsize % 16 == 0)
+        {
+            top_blob = top_blob_3d;
+            top_blob.dims = 1;
+            top_blob.w = top_blob_3d.c;
+            top_blob.h = 1;
+            top_blob.c = 1;
+            bottom_blob_3d.cstep = top_blob_3d.c;
+        }
+        else
+        {
+            top_blob = top_blob_3d.reshape(top_blob_3d.c, opt.blob_allocator);
+        }
+
+        return 0;
     }
 
     int elembits = bottom_blob.elembits();
@@ -328,7 +380,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv1x1s1_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -337,7 +389,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv1x1s2_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv1x1s2_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -363,7 +415,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -372,7 +424,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else
         {
-            convolution_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
@@ -380,7 +432,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv1x1s1_sgemm_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -389,7 +441,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv3x3s1_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv3x3s1_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -398,7 +450,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv3x3s2_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv3x3s2_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -407,7 +459,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (kernel_w == 7 && kernel_h == 7 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv7x7s2_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv7x7s2_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -416,7 +468,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -425,7 +477,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else
         {
-            convolution_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
@@ -433,7 +485,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv1x1s1_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -442,7 +494,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv1x1s2_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv1x1s2_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -451,7 +503,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -460,7 +512,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else
         {
-            convolution_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 #endif // __riscv_vector
@@ -469,7 +521,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+            conv1x1s1_sgemm_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
             if (activation)
             {
@@ -494,7 +546,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -541,7 +593,7 @@ int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
                             sum = bias_data[p];
                         }
 
-                        const float* kptr = (const float*)weight_data + maxk * channels * p;
+                        const float* kptr = (const float*)weight_data_tm + maxk * channels * p;
 
                         // channels
                         for (int q = 0; q < channels; q++)
@@ -666,7 +718,7 @@ int Convolution_riscv::forward(const std::vector<Mat>& bottom_blobs, std::vector
 }
 
 #if __riscv_vector && __riscv_zfh
-static void convolution_transform_kernel_packed_fp16s_rvv(const Mat& weight_data, Mat& weight_data_fp16, int num_input, int num_output, int kernel_w, int kernel_h, int elempack, int out_elempack)
+static void convolution_transform_kernel_packed_fp16s_rvv(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, int kernel_w, int kernel_h, int elempack, int out_elempack)
 {
     const int maxk = kernel_w * kernel_h;
 
@@ -675,11 +727,11 @@ static void convolution_transform_kernel_packed_fp16s_rvv(const Mat& weight_data
     {
         Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);
 
-        weight_data_fp16.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)2u * elempack * out_elempack, elempack * out_elempack);
+        weight_data_tm.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)2u * elempack * out_elempack, elempack * out_elempack);
 
         for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
         {
-            __fp16* g00 = weight_data_fp16.channel(q / out_elempack);
+            __fp16* g00 = weight_data_tm.channel(q / out_elempack);
 
             for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
             {
@@ -728,14 +780,14 @@ int Convolution_riscv::create_pipeline_fp16s(const Option& opt)
         }
         else
         {
-            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
+            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
         }
     }
 
     // pack1ton
     if (elempack == 1 && out_elempack == packn)
     {
-        convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
+        convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
     }
 
     // packnto1
@@ -743,19 +795,19 @@ int Convolution_riscv::create_pipeline_fp16s(const Option& opt)
     {
         if (opt.use_fp16_arithmetic && kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else if (opt.use_fp16_arithmetic && kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else if (opt.use_fp16_arithmetic && opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else
         {
-            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
+            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
         }
     }
 
@@ -764,15 +816,15 @@ int Convolution_riscv::create_pipeline_fp16s(const Option& opt)
     {
         if (opt.use_fp16_arithmetic && kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            convolution_im2col_sgemm_transform_kernel_fp16sa_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else if (opt.use_fp16_arithmetic && opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_transform_kernel_fp16sa_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h);
+            convolution_im2col_sgemm_transform_kernel_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
         }
         else
         {
-            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_fp16, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
+            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
         }
     }
 
@@ -781,6 +833,11 @@ int Convolution_riscv::create_pipeline_fp16s(const Option& opt)
         ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
     }
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -818,28 +875,28 @@ int Convolution_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, cons
     if (elempack == packn && out_elempack == packn)
     {
         {
-            convolution_packn_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_packn_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == 1 && out_elempack == packn)
     {
         {
-            convolution_pack1ton_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_pack1ton_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == packn && out_elempack == 1)
     {
         {
-            convolution_packnto1_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_packnto1_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == 1 && out_elempack == 1)
     {
         {
-            convolution_fp16s(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_fp16s(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
@@ -884,7 +941,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv1x1s1_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -893,7 +950,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv1x1s2_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv1x1s2_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -919,7 +976,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -928,7 +985,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else
         {
-            convolution_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
@@ -936,7 +993,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv1x1s1_sgemm_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -945,7 +1002,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv3x3s1_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv3x3s1_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -954,7 +1011,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv3x3s2_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv3x3s2_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -963,7 +1020,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (kernel_w == 7 && kernel_h == 7 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv7x7s2_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv7x7s2_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -972,7 +1029,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -981,7 +1038,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else
         {
-            convolution_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
@@ -989,7 +1046,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv1x1s1_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -998,7 +1055,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
         {
-            conv1x1s2_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv1x1s2_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -1007,7 +1064,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -1016,7 +1073,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else
         {
-            convolution_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
@@ -1024,7 +1081,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
     {
         if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
         {
-            conv1x1s1_sgemm_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+            conv1x1s1_sgemm_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
             if (activation)
             {
@@ -1033,7 +1090,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else if (opt.use_sgemm_convolution)
         {
-            convolution_im2col_sgemm_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
+            convolution_im2col_sgemm_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
 
             if (activation)
             {
@@ -1042,7 +1099,7 @@ int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
         }
         else
         {
-            convolution_fp16s(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            convolution_fp16s(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 

src/layer/riscv/convolution_riscv.h

@@ -41,14 +41,12 @@ protected:
 public:
     Layer* activation;
 
-    // packn
-    Mat weight_data_packed;
+    Mat weight_data_tm;
     Mat weight_winograd23_data;
     Mat weight_winograd43_data;
     Mat weight_winograd63_data;
 
     // fp16
-    Mat weight_data_fp16;
     Mat bias_data_fp16;
 };
 

src/layer/riscv/convolutiondepthwise_riscv.cpp

@@ -61,6 +61,14 @@ int ConvolutionDepthWise_riscv::create_pipeline(const Option& opt)
 
     activation = create_activation_layer(activation_type, activation_params, opt);
 
+#if NCNN_INT8
+    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
+    {
+        // TODO implement int8
+        return 0;
+    }
+#endif
+
 #if __riscv_vector && __riscv_zfh
     if (opt.use_fp16_storage)
     {
@@ -91,12 +99,18 @@ int ConvolutionDepthWise_riscv::create_pipeline(const Option& opt)
         if (elempack == packn)
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, group);
-            convert_packing(weight_data_r2, weight_data_packed, packn, opt);
+            convert_packing(weight_data_r2, weight_data_tm, packn, opt);
         }
 #endif // __riscv_vector
 
         if (elempack == 1)
         {
+            weight_data_tm = weight_data;
+        }
+
+        if (opt.lightmode)
+        {
+            weight_data.release();
         }
 
         return 0;
@@ -105,6 +119,11 @@ int ConvolutionDepthWise_riscv::create_pipeline(const Option& opt)
     // group convolution
     create_group_ops(opt);
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -126,7 +145,7 @@ int ConvolutionDepthWise_riscv::create_group_ops(const Option& opt)
 
     for (int g = 0; g < group; g++)
     {
-        Mat weight_data_g = weight_data.range(maxk * channels_g * num_output_g * g, maxk * channels_g * num_output_g);
+        Mat weight_data_g = weight_data.range(maxk * channels_g * num_output_g * g, maxk * channels_g * num_output_g).clone();
         Mat bias_data_g;
         if (bias_term)
             bias_data_g = bias_data.range(num_output_g * g, num_output_g);
@@ -227,7 +246,7 @@ int ConvolutionDepthWise_riscv::destroy_pipeline(const Option& opt)
 int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
 {
 #if NCNN_INT8
-    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
+    if (opt.use_int8_inference && int8_scale_term)
     {
         Mat bottom_blob_unpacked = bottom_blob;
         if (bottom_blob.elempack != 1)
@@ -310,7 +329,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
         {
             if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
-                convdw3x3s1_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+                convdw3x3s1_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
                 if (activation)
                 {
@@ -319,7 +338,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
             }
             else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
-                convdw3x3s2_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+                convdw3x3s2_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
                 if (activation)
                 {
@@ -328,7 +347,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
             }
             else if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
-                convdw5x5s1_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+                convdw5x5s1_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
                 if (activation)
                 {
@@ -337,7 +356,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
             }
             else if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
-                convdw5x5s2_packn_rvv(bottom_blob_bordered, top_blob, weight_data_packed, bias_data, opt);
+                convdw5x5s2_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
                 if (activation)
                 {
@@ -371,7 +390,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
                 for (int g = 0; g < channels; g++)
                 {
                     float* outptr = top_blob.channel(g);
-                    const float* kptr = (const float*)weight_data_packed + maxk * g * packn;
+                    const float* kptr = (const float*)weight_data_tm + maxk * g * packn;
                     const Mat m = bottom_blob_bordered.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -410,7 +429,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
         {
             if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
-                convdw3x3s1_rvv(bottom_blob_bordered, top_blob, weight_data, bias_data, opt);
+                convdw3x3s1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
                 if (activation)
                 {
@@ -419,7 +438,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
             }
             else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
-                convdw3x3s2_rvv(bottom_blob_bordered, top_blob, weight_data, bias_data, opt);
+                convdw3x3s2_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
 
                 if (activation)
                 {
@@ -453,7 +472,7 @@ int ConvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob, c
                 for (int g = 0; g < group; g++)
                 {
                     float* outptr = top_blob.channel(g);
-                    const float* kptr = (const float*)weight_data + maxk * g;
+                    const float* kptr = (const float*)weight_data_tm + maxk * g;
                     const Mat m = bottom_blob_bordered.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -663,22 +682,32 @@ int ConvolutionDepthWise_riscv::create_pipeline_fp16s(const Option& opt)
             Mat weight_data_r2_packed;
             convert_packing(weight_data_r2, weight_data_r2_packed, packn, opt);
 
-            ncnn::cast_float32_to_float16(weight_data_r2_packed, weight_data_fp16, opt);
+            ncnn::cast_float32_to_float16(weight_data_r2_packed, weight_data_tm, opt);
         }
 
         if (elempack == 1)
         {
-            ncnn::cast_float32_to_float16(weight_data, weight_data_fp16, opt);
+            ncnn::cast_float32_to_float16(weight_data, weight_data_tm, opt);
         }
 
         ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
 
+        if (opt.lightmode)
+        {
+            weight_data.release();
+        }
+
         return 0;
     }
 
     // group convolution
     create_group_ops(opt);
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -744,7 +773,7 @@ int ConvolutionDepthWise_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_b
                 for (int g = 0; g < channels; g++)
                 {
                     __fp16* outptr = top_blob.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g * packn;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g * packn;
                     const Mat m = bottom_blob_bordered.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -806,7 +835,7 @@ int ConvolutionDepthWise_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_b
                 for (int g = 0; g < group; g++)
                 {
                     __fp16* outptr = top_blob.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g;
                     const Mat m = bottom_blob_bordered.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -930,7 +959,7 @@ int ConvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_
         {
             if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
-                convdw3x3s1_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+                convdw3x3s1_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
                 if (activation)
                 {
@@ -939,7 +968,7 @@ int ConvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_
             }
             else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
-                convdw3x3s2_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+                convdw3x3s2_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
                 if (activation)
                 {
@@ -948,7 +977,7 @@ int ConvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_
             }
             else if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
-                convdw5x5s1_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+                convdw5x5s1_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
                 if (activation)
                 {
@@ -957,7 +986,7 @@ int ConvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_
             }
             else if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
-                convdw5x5s2_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_fp16, bias_data_fp16, opt);
+                convdw5x5s2_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
 
                 if (activation)
                 {
@@ -991,7 +1020,7 @@ int ConvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_
                 for (int g = 0; g < channels; g++)
                 {
                     __fp16* outptr = top_blob.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g * packn;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g * packn;
                     const Mat m = bottom_blob_bordered.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -1053,7 +1082,7 @@ int ConvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_
                 for (int g = 0; g < group; g++)
                 {
                     __fp16* outptr = top_blob.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g;
                     const Mat m = bottom_blob_bordered.channel(g);
 
                     for (int i = 0; i < outh; i++)

src/layer/riscv/convolutiondepthwise_riscv.h

@@ -43,11 +43,9 @@ public:
     Layer* activation;
     std::vector<ncnn::Layer*> group_ops;
 
-    // packing
-    Mat weight_data_packed;
+    Mat weight_data_tm;
 
     // fp16
-    Mat weight_data_fp16;
     Mat bias_data_fp16;
 };
 

src/layer/riscv/deconvolution_riscv.cpp

@@ -100,11 +100,11 @@ int Deconvolution_riscv::create_pipeline(const Option& opt)
     {
         Mat weight_data_r2 = weight_data_transposed.reshape(maxk, num_input, num_output);
 
-        weight_data_packed.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)4u * elempack * out_elempack, elempack * out_elempack);
+        weight_data_tm.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)4u * elempack * out_elempack, elempack * out_elempack);
 
         for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
         {
-            float* g00 = weight_data_packed.channel(q / out_elempack);
+            float* g00 = weight_data_tm.channel(q / out_elempack);
 
             for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
             {
@@ -148,6 +148,11 @@ int Deconvolution_riscv::create_pipeline(const Option& opt)
     {
     }
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -218,21 +223,21 @@ int Deconvolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Op
     if (elempack == packn && out_elempack == packn)
     {
         {
-            deconvolution_packn_rvv(bottom_blob, top_blob_bordered, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_packn_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == 1 && out_elempack == packn)
     {
         {
-            deconvolution_pack1ton_rvv(bottom_blob, top_blob_bordered, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_pack1ton_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == packn && out_elempack == 1)
     {
         {
-            deconvolution_packnto1_rvv(bottom_blob, top_blob_bordered, weight_data_packed, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_packnto1_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 #endif // __riscv_vector
@@ -257,7 +262,7 @@ int Deconvolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Op
                             sum = bias_data[p];
                         }
 
-                        const float* kptr = (const float*)weight_data_packed.channel(p);
+                        const float* kptr = (const float*)weight_data_tm.channel(p);
 
                         // channels
                         for (int q = 0; q < channels; q++)
@@ -356,11 +361,11 @@ int Deconvolution_riscv::create_pipeline_fp16s(const Option& opt)
     {
         Mat weight_data_r2 = weight_data_transposed.reshape(maxk, num_input, num_output);
 
-        weight_data_fp16.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)2u * elempack * out_elempack, elempack * out_elempack);
+        weight_data_tm.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)2u * elempack * out_elempack, elempack * out_elempack);
 
         for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
         {
-            __fp16* g00 = weight_data_fp16.channel(q / out_elempack);
+            __fp16* g00 = weight_data_tm.channel(q / out_elempack);
 
             for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
             {
@@ -404,6 +409,11 @@ int Deconvolution_riscv::create_pipeline_fp16s(const Option& opt)
 
     ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -446,28 +456,28 @@ int Deconvolution_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, co
     if (elempack == packn && out_elempack == packn)
     {
         {
-            deconvolution_packn_fp16s_rvv(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_packn_fp16s_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == 1 && out_elempack == packn)
     {
         {
-            deconvolution_pack1ton_fp16s_rvv(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_pack1ton_fp16s_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == packn && out_elempack == 1)
     {
         {
-            deconvolution_packnto1_fp16s_rvv(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_packnto1_fp16s_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == 1 && out_elempack == 1)
     {
         {
-            deconvolution_fp16s(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_fp16s(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
@@ -517,28 +527,28 @@ int Deconvolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, c
     if (elempack == packn && out_elempack == packn)
     {
         {
-            deconvolution_packn_fp16sa_rvv(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_packn_fp16sa_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == 1 && out_elempack == packn)
     {
         {
-            deconvolution_pack1ton_fp16sa_rvv(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_pack1ton_fp16sa_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == packn && out_elempack == 1)
     {
         {
-            deconvolution_packnto1_fp16sa_rvv(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_packnto1_fp16sa_rvv(bottom_blob, top_blob_bordered, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 
     if (elempack == 1 && out_elempack == 1)
     {
         {
-            deconvolution_fp16s(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
+            deconvolution_fp16s(bottom_blob, top_blob_bordered, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
         }
     }
 

src/layer/riscv/deconvolution_riscv.h

@@ -37,11 +37,9 @@ protected:
 #endif
 
 public:
-    // packn
-    Mat weight_data_packed;
+    Mat weight_data_tm;
 
     // fp16
-    Mat weight_data_fp16;
     Mat bias_data_fp16;
 };
 

src/layer/riscv/deconvolutiondepthwise_riscv.cpp

@@ -88,13 +88,18 @@ int DeconvolutionDepthWise_riscv::create_pipeline(const Option& opt)
         if (elempack == packn)
         {
             Mat weight_data_r2 = weight_data_transposed.reshape(maxk, group);
-            convert_packing(weight_data_r2, weight_data_packed, packn, opt);
+            convert_packing(weight_data_r2, weight_data_tm, packn, opt);
         }
 #endif // __riscv_vector
 
         if (elempack == 1)
         {
-            weight_data_packed = weight_data_transposed;
+            weight_data_tm = weight_data_transposed;
+        }
+
+        if (opt.lightmode)
+        {
+            weight_data.release();
         }
 
         return 0;
@@ -103,6 +108,11 @@ int DeconvolutionDepthWise_riscv::create_pipeline(const Option& opt)
     // group convolution
     create_group_ops(opt);
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -124,7 +134,7 @@ int DeconvolutionDepthWise_riscv::create_group_ops(const Option& opt)
 
     for (int g = 0; g < group; g++)
     {
-        Mat weight_data_g = weight_data.range(maxk * channels_g * num_output_g * g, maxk * channels_g * num_output_g);
+        Mat weight_data_g = weight_data.range(maxk * channels_g * num_output_g * g, maxk * channels_g * num_output_g).clone();
         Mat bias_data_g;
         if (bias_term)
             bias_data_g = bias_data.range(num_output_g * g, num_output_g);
@@ -256,7 +266,7 @@ int DeconvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob,
                 for (int g = 0; g < channels; g++)
                 {
                     float* outptr = top_blob_bordered.channel(g);
-                    const float* kptr = (const float*)weight_data_packed + maxk * g * packn;
+                    const float* kptr = (const float*)weight_data_tm + maxk * g * packn;
                     const Mat m = bottom_blob.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -318,7 +328,7 @@ int DeconvolutionDepthWise_riscv::forward(const Mat& bottom_blob, Mat& top_blob,
             for (int g = 0; g < channels; g++)
             {
                 float* outptr = top_blob_bordered.channel(g);
-                const float* kptr = (const float*)weight_data_packed + maxk * g;
+                const float* kptr = (const float*)weight_data_tm + maxk * g;
                 const Mat m = bottom_blob.channel(g);
 
                 for (int i = 0; i < outh; i++)
@@ -480,22 +490,32 @@ int DeconvolutionDepthWise_riscv::create_pipeline_fp16s(const Option& opt)
             Mat weight_data_r2_packed;
             convert_packing(weight_data_r2, weight_data_r2_packed, packn, opt);
 
-            ncnn::cast_float32_to_float16(weight_data_r2_packed, weight_data_fp16, opt);
+            ncnn::cast_float32_to_float16(weight_data_r2_packed, weight_data_tm, opt);
         }
 
         if (elempack == 1)
         {
-            ncnn::cast_float32_to_float16(weight_data_transposed, weight_data_fp16, opt);
+            ncnn::cast_float32_to_float16(weight_data_transposed, weight_data_tm, opt);
         }
 
         ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
 
+        if (opt.lightmode)
+        {
+            weight_data.release();
+        }
+
         return 0;
     }
 
     // group convolution
     create_group_ops(opt);
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -543,7 +563,7 @@ int DeconvolutionDepthWise_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top
                 for (int g = 0; g < channels; g++)
                 {
                     __fp16* outptr = top_blob_bordered.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g * packn;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g * packn;
                     const Mat m = bottom_blob.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -605,7 +625,7 @@ int DeconvolutionDepthWise_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top
                 for (int g = 0; g < channels; g++)
                 {
                     __fp16* outptr = top_blob_bordered.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g;
                     const Mat m = bottom_blob.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -764,7 +784,7 @@ int DeconvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& to
                 for (int g = 0; g < channels; g++)
                 {
                     __fp16* outptr = top_blob_bordered.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g * packn;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g * packn;
                     const Mat m = bottom_blob.channel(g);
 
                     for (int i = 0; i < outh; i++)
@@ -826,7 +846,7 @@ int DeconvolutionDepthWise_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& to
                 for (int g = 0; g < channels; g++)
                 {
                     __fp16* outptr = top_blob_bordered.channel(g);
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * g;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + maxk * g;
                     const Mat m = bottom_blob.channel(g);
 
                     for (int i = 0; i < outh; i++)

src/layer/riscv/deconvolutiondepthwise_riscv.h

@@ -40,11 +40,9 @@ protected:
 public:
     std::vector<ncnn::Layer*> group_ops;
 
-    // packing
-    Mat weight_data_packed;
+    Mat weight_data_tm;
 
     // fp16
-    Mat weight_data_fp16;
     Mat bias_data_fp16;
 };
 

src/layer/riscv/innerproduct_riscv.cpp

@@ -43,8 +43,6 @@ InnerProduct_riscv::InnerProduct_riscv()
 
 int InnerProduct_riscv::create_pipeline(const Option& opt)
 {
-#if __riscv_vector
-    if (opt.use_packing_layout || opt.use_int8_inference)
     {
         flatten = ncnn::create_layer(ncnn::LayerType::Flatten);
 
@@ -54,7 +52,14 @@ int InnerProduct_riscv::create_pipeline(const Option& opt)
 
         flatten->create_pipeline(opt);
     }
-#endif // __riscv_vector
+
+#if NCNN_INT8
+    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
+    {
+        // TODO implement int8
+        return 0;
+    }
+#endif
 
 #if __riscv_vector && __riscv_zfh
     if (opt.use_fp16_storage)
@@ -63,6 +68,53 @@ int InnerProduct_riscv::create_pipeline(const Option& opt)
     }
 #endif
 
+    int out_elempack = 1;
+
+#if __riscv_vector
+    const int packn = csrr_vlenb() / 4;
+
+    const int num_input = weight_data_size / num_output;
+
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % packn == 0 ? packn : 1;
+    }
+
+    if (out_elempack == packn)
+    {
+        // src = inch-outch
+        // dst = packn-inch-outch/packn
+        {
+            Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
+
+            weight_data_tm.create(num_input, num_output / packn, (size_t)4u * packn, packn);
+
+            for (int q = 0; q + (packn - 1) < num_output; q += packn)
+            {
+                float* g0 = weight_data_tm.row(q / packn);
+
+                for (int p = 0; p < num_input; p++)
+                {
+                    for (int j = 0; j < packn; j++)
+                    {
+                        *g0++ = weight_data_r2.row(q + j)[p];
+                    }
+                }
+            }
+        }
+    }
+#endif // __riscv_vector
+
+    if (out_elempack == 1)
+    {
+        weight_data_tm = weight_data;
+    }
+
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -81,7 +133,7 @@ int InnerProduct_riscv::destroy_pipeline(const Option& opt)
 int InnerProduct_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
 {
 #if NCNN_INT8
-    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
+    if (opt.use_int8_inference && int8_scale_term)
     {
         Mat bottom_blob_unpacked = bottom_blob;
         if (bottom_blob.elempack != 1)
@@ -136,20 +188,104 @@ int InnerProduct_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opt
         if (top_blob.empty())
             return -100;
 
+        int num_output_elempack = 1;
+#if __riscv_vector
+        if (opt.use_packing_layout)
+        {
+            num_output_elempack = num_output % packn == 0 ? packn : 1;
+        }
+#endif
+
         #pragma omp parallel for num_threads(opt.num_threads)
         for (int j = 0; j < h; j++)
         {
 #if __riscv_vector
-            if (elempack == packn)
+            if (elempack == packn && num_output_elempack == packn)
             {
+                const word_type vl = vsetvl_e32m1(packn);
+
+                float* outptr = top_blob.row(j);
+
+                for (int p = 0; p < num_output / num_output_elempack; p++)
+                {
+                    for (int l = 0; l < packn; l++)
+                    {
+                        const float* kptr = weight_data_tm.row(p) + l;
+                        const float* m = bottom_blob.row(j);
+
+                        vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
+
+                        if (bias_term)
+                        {
+                            _sum = vfmv_v_f_f32m1(bias_data[p * packn + l], vl);
+                        }
+
+                        int n = num_input;
+                        while (n > 0)
+                        {
+                            vfloat32m1_t _val = vle32_v_f32m1(m, vl);
+                            _sum = vfmacc_vf_f32m1(_sum, *kptr, _val, vl);
+
+                            m += packn;
+                            kptr += packn;
+                            n -= 1;
+                        }
+
+                        _sum = activation_ps(_sum, activation_type, activation_params, vl);
+
+                        vse32_v_f32m1(outptr, _sum, vl);
+                        outptr += packn;
+                    }
+                }
+            }
+
+            if (elempack == 1 && num_output_elempack == packn)
+            {
+                const word_type vl = vsetvl_e32m1(packn);
+
+                float* outptr = top_blob.row(j);
+
+                for (int p = 0; p < num_output / num_output_elempack; p++)
+                {
+                    const float* kptr = weight_data_tm.row(p);
+                    const float* m = bottom_blob.row(j);
+
+                    vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
+
+                    if (bias_term)
+                    {
+                        _sum = vle32_v_f32m1((const float*)bias_data + p * packn, vl);
+                    }
+
+                    int n = num_input;
+                    while (n > 0)
+                    {
+                        vfloat32m1_t _w = vle32_v_f32m1(kptr, vl);
+                        _sum = vfmacc_vf_f32m1(_sum, *m, _w, vl);
+
+                        m += 1;
+                        kptr += packn;
+                        n -= 1;
+                    }
+
+                    _sum = activation_ps(_sum, activation_type, activation_params, vl);
+
+                    vse32_v_f32m1(outptr, _sum, vl);
+                    outptr += packn;
+                }
+            }
+
+            if (elempack == packn && num_output_elempack == 1)
+            {
+                const word_type vl = vsetvl_e32m1(packn);
+
                 float* outptr = top_blob.row(j);
 
                 for (int p = 0; p < num_output; p++)
                 {
-                    const float* kptr = (const float*)weight_data + num_input * p;
+                    const float* kptr = (const float*)weight_data_tm + num_input * p;
                     const float* m = bottom_blob.row(j);
 
-                    const word_type vl = vsetvl_e32m1(packn);
                     vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
 
                     if (bias_term)
@@ -176,13 +312,13 @@ int InnerProduct_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opt
             }
 #endif // __riscv_vector
 
-            if (elempack == 1)
+            if (elempack == 1 && num_output_elempack == 1)
             {
                 float* outptr = top_blob.row(j);
 
                 for (int p = 0; p < num_output; p++)
                 {
-                    const float* kptr = (const float*)weight_data + num_input * p;
+                    const float* kptr = (const float*)weight_data_tm + num_input * p;
                     const float* m = bottom_blob.row(j);
 
                     float sum = 0.f;
@@ -208,69 +344,93 @@ int InnerProduct_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opt
         return 0;
     }
 
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-    int size = w * h;
-
-#if __riscv_vector
-    if (elempack == packn)
+    // flatten
+    Mat bottom_blob_flattened = bottom_blob;
+    if (bottom_blob.dims != 1)
     {
-        // flatten
-        Mat bottom_blob_flattened = bottom_blob;
-        if (bottom_blob.dims != 1)
-        {
-            Option opt_flatten = opt;
-            opt_flatten.blob_allocator = opt.workspace_allocator;
+        Option opt_flatten = opt;
+        opt_flatten.blob_allocator = opt.workspace_allocator;
 
-            flatten->forward(bottom_blob, bottom_blob_flattened, opt_flatten);
-        }
+        flatten->forward(bottom_blob, bottom_blob_flattened, opt_flatten);
+    }
 
-        // pack1
-        {
-            bottom_blob_flattened.w *= bottom_blob_flattened.elempack;
-            bottom_blob_flattened.cstep = bottom_blob_flattened.w;
-            bottom_blob_flattened.elemsize = 4u;
-            bottom_blob_flattened.elempack = 1;
-        }
+    size_t elemsize = bottom_blob_flattened.elemsize;
+    int elempack = bottom_blob_flattened.elempack;
 
-        return forward(bottom_blob_flattened, top_blob, opt);
+    int out_elempack = 1;
+#if __riscv_vector
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % packn == 0 ? packn : 1;
     }
 #endif
+    size_t out_elemsize = elemsize / elempack * out_elempack;
 
-    top_blob.create(num_output, elemsize, opt.blob_allocator);
+    top_blob.create(num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
     if (top_blob.empty())
         return -100;
 
-    const float* weight_data_ptr = weight_data;
-
 #if __riscv_vector
-    int nn_num_output = num_output / packn;
-    int remain_num_output_start = nn_num_output * packn;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_num_output; pp++)
+    if (out_elempack == packn)
     {
-        int p = pp * packn;
+        #pragma omp parallel for num_threads(opt.num_threads)
+        for (int p = 0; p < num_output / out_elempack; p++)
+        {
+            const word_type vl = vsetvl_e32m1(packn);
+            vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
 
-        const word_type vl = vsetvl_e32m1(packn);
-        vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
+            if (bias_term)
+            {
+                _sum = vle32_v_f32m1((const float*)bias_data + p * packn, vl);
+            }
 
-        if (bias_term)
-        {
-            _sum = vle32_v_f32m1((const float*)bias_data + p, vl);
+            const float* kptr = weight_data_tm.row(p);
+
+            const float* sptr = bottom_blob_flattened;
+
+            int n = num_input;
+            while (n > 0)
+            {
+                vfloat32m1_t _w = vle32_v_f32m1(kptr, vl);
+                _sum = vfmacc_vf_f32m1(_sum, *sptr, _w, vl);
+
+                sptr += 1;
+                kptr += packn;
+                n -= 1;
+            }
+
+            _sum = activation_ps(_sum, activation_type, activation_params, vl);
+
+            float* outptr = top_blob;
+            vse32_v_f32m1(outptr + p * packn, _sum, vl);
         }
+    }
+#endif // __riscv_vector
 
-        const float* w = weight_data_ptr + num_input * p;
+    if (out_elempack == 1)
+    {
+#if __riscv_vector
+        int nn_num_output = num_output / packn;
+        int remain_num_output_start = nn_num_output * packn;
 
-        // channels
-        for (int q = 0; q < channels; q++)
+        #pragma omp parallel for num_threads(opt.num_threads)
+        for (int pp = 0; pp < nn_num_output; pp++)
         {
-            const float* m = bottom_blob.channel(q);
+            int p = pp * packn;
 
-            int n = size;
+            const word_type vl = vsetvl_e32m1(packn);
+            vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
+
+            if (bias_term)
+            {
+                _sum = vle32_v_f32m1((const float*)bias_data + p, vl);
+            }
+
+            const float* w = (const float*)weight_data_tm + num_input * p;
+
+            const float* m = bottom_blob_flattened;
+
+            int n = num_input;
             while (n > 0)
             {
                 vfloat32m1_t _w = vlse32_v_f32m1(w, num_input * sizeof(float), vl);
@@ -281,45 +441,41 @@ int InnerProduct_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opt
                 w += 1;
                 n -= 1;
             }
-        }
 
-        _sum = activation_ps(_sum, activation_type, activation_params, vl);
+            _sum = activation_ps(_sum, activation_type, activation_params, vl);
 
-        vse32_v_f32m1((float*)top_blob + p, _sum, vl);
-    }
+            vse32_v_f32m1((float*)top_blob + p, _sum, vl);
+        }
 #else // __riscv_vector
-    int nn_num_output = num_output / 4;
-    int remain_num_output_start = nn_num_output * 4;
+        int nn_num_output = num_output / 4;
+        int remain_num_output_start = nn_num_output * 4;
 
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_num_output; pp++)
-    {
-        int p = pp * 4;
+        #pragma omp parallel for num_threads(opt.num_threads)
+        for (int pp = 0; pp < nn_num_output; pp++)
+        {
+            int p = pp * 4;
 
-        float sum0 = 0.f;
-        float sum1 = 0.f;
-        float sum2 = 0.f;
-        float sum3 = 0.f;
+            float sum0 = 0.f;
+            float sum1 = 0.f;
+            float sum2 = 0.f;
+            float sum3 = 0.f;
 
-        if (bias_term)
-        {
-            sum0 = bias_data[p];
-            sum1 = bias_data[p + 1];
-            sum2 = bias_data[p + 2];
-            sum3 = bias_data[p + 3];
-        }
+            if (bias_term)
+            {
+                sum0 = bias_data[p];
+                sum1 = bias_data[p + 1];
+                sum2 = bias_data[p + 2];
+                sum3 = bias_data[p + 3];
+            }
 
-        const float* w0 = weight_data_ptr + num_input * p;
-        const float* w1 = weight_data_ptr + num_input * (p + 1);
-        const float* w2 = weight_data_ptr + num_input * (p + 2);
-        const float* w3 = weight_data_ptr + num_input * (p + 3);
+            const float* w0 = (const float*)weight_data_tm + num_input * p;
+            const float* w1 = (const float*)weight_data_tm + num_input * (p + 1);
+            const float* w2 = (const float*)weight_data_tm + num_input * (p + 2);
+            const float* w3 = (const float*)weight_data_tm + num_input * (p + 3);
 
-        // channels
-        for (int q = 0; q < channels; q++)
-        {
-            const float* m = bottom_blob.channel(q);
+            const float* m = bottom_blob_flattened;
 
-            for (int i = 0; i < size; i++)
+            for (int i = 0; i < num_input; i++)
             {
                 sum0 += *m * *w0;
                 sum1 += *m * *w1;
@@ -332,48 +488,43 @@ int InnerProduct_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Opt
                 w2++;
                 w3++;
             }
-        }
 
-        sum0 = activation_ss(sum0, activation_type, activation_params);
-        sum1 = activation_ss(sum1, activation_type, activation_params);
-        sum2 = activation_ss(sum2, activation_type, activation_params);
-        sum3 = activation_ss(sum3, activation_type, activation_params);
+            sum0 = activation_ss(sum0, activation_type, activation_params);
+            sum1 = activation_ss(sum1, activation_type, activation_params);
+            sum2 = activation_ss(sum2, activation_type, activation_params);
+            sum3 = activation_ss(sum3, activation_type, activation_params);
 
-        top_blob[p] = sum0;
-        top_blob[p + 1] = sum1;
-        top_blob[p + 2] = sum2;
-        top_blob[p + 3] = sum3;
-    }
+            top_blob[p] = sum0;
+            top_blob[p + 1] = sum1;
+            top_blob[p + 2] = sum2;
+            top_blob[p + 3] = sum3;
+        }
 #endif // __riscv_vector
 
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_num_output_start; p < num_output; p++)
-    {
-        float sum = 0.f;
+        #pragma omp parallel for num_threads(opt.num_threads)
+        for (int p = remain_num_output_start; p < num_output; p++)
+        {
+            float sum = 0.f;
 
-        if (bias_term)
-            sum = bias_data[p];
+            if (bias_term)
+                sum = bias_data[p];
 
-        const float* w = weight_data_ptr + num_input * p;
+            const float* w = (const float*)weight_data_tm + num_input * p;
 
-        // channels
-        for (int q = 0; q < channels; q++)
-        {
-            const float* m = bottom_blob.channel(q);
+            const float* m = bottom_blob_flattened;
 
-            for (int i = 0; i < size; i++)
+            for (int i = 0; i < num_input; i++)
             {
                 sum += *m * *w;
 
                 m++;
                 w++;
             }
-        }
 
-        sum = activation_ss(sum, activation_type, activation_params);
+            sum = activation_ss(sum, activation_type, activation_params);
 
-        top_blob[p] = sum;
+            top_blob[p] = sum;
+        }
     }
 
     return 0;
@@ -398,11 +549,11 @@ int InnerProduct_riscv::create_pipeline_fp16s(const Option& opt)
     {
         Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
 
-        weight_data_fp16.create(num_input, num_output / out_elempack, (size_t)2u * out_elempack, out_elempack);
+        weight_data_tm.create(num_input, num_output / out_elempack, (size_t)2u * out_elempack, out_elempack);
 
         for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
         {
-            __fp16* g0 = weight_data_fp16.row<__fp16>(q / out_elempack);
+            __fp16* g0 = weight_data_tm.row<__fp16>(q / out_elempack);
 
             for (int p = 0; p < num_input; p++)
             {
@@ -416,6 +567,11 @@ int InnerProduct_riscv::create_pipeline_fp16s(const Option& opt)
 
     ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
 
+    if (opt.lightmode)
+    {
+        weight_data.release();
+    }
+
     return 0;
 }
 
@@ -451,7 +607,7 @@ int InnerProduct_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, con
                 {
                     for (int l = 0; l < packn; l++)
                     {
-                        const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p * packn + l;
+                        const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p * packn + l;
                         const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                         vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
@@ -489,7 +645,7 @@ int InnerProduct_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, con
 
                 for (int p = 0; p < num_output / num_output_elempack; p++)
                 {
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p * packn;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p * packn;
                     const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                     vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
@@ -526,7 +682,7 @@ int InnerProduct_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, con
 
                 for (int p = 0; p < num_output; p++)
                 {
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p;
                     const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                     vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
@@ -561,7 +717,7 @@ int InnerProduct_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, con
 
                 for (int p = 0; p < num_output; p++)
                 {
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p;
                     const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                     float sum = 0.f;
@@ -621,7 +777,7 @@ int InnerProduct_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, con
                 _sum = vle32_v_f32m2((const float*)bias_data + p * packn, vl);
             }
 
-            const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
+            const __fp16* kptr = weight_data_tm.row<const __fp16>(p);
 
             const __fp16* sptr = bottom_blob_flattened;
 
@@ -655,7 +811,7 @@ int InnerProduct_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, con
             if (bias_term)
                 sum = bias_data[p];
 
-            const __fp16* kptr = weight_data_fp16.row<__fp16>(p);
+            const __fp16* kptr = weight_data_tm.row<__fp16>(p);
 
             const __fp16* sptr = bottom_blob_flattened;
 
@@ -713,7 +869,7 @@ int InnerProduct_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, co
                 {
                     for (int l = 0; l < packn; l++)
                     {
-                        const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p * packn + l;
+                        const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p * packn + l;
                         const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                         vfloat16m1_t _sum = vfmv_v_f_f16m1((__fp16)0.f, vl);
@@ -751,7 +907,7 @@ int InnerProduct_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, co
 
                 for (int p = 0; p < num_output / num_output_elempack; p++)
                 {
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p * packn;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p * packn;
                     const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                     vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
@@ -788,7 +944,7 @@ int InnerProduct_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, co
 
                 for (int p = 0; p < num_output; p++)
                 {
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p;
                     const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                     vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
@@ -823,7 +979,7 @@ int InnerProduct_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, co
 
                 for (int p = 0; p < num_output; p++)
                 {
-                    const __fp16* kptr = (const __fp16*)weight_data_fp16 + num_input * p;
+                    const __fp16* kptr = (const __fp16*)weight_data_tm + num_input * p;
                     const __fp16* m = bottom_blob.row<const __fp16>(j);
 
                     float sum = 0.f;
@@ -883,7 +1039,7 @@ int InnerProduct_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, co
                 _sum = vle16_v_f16m1((const __fp16*)bias_data_fp16 + p * packn, vl);
             }
 
-            const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
+            const __fp16* kptr = weight_data_tm.row<const __fp16>(p);
 
             const __fp16* sptr = bottom_blob_flattened;
 
@@ -917,7 +1073,7 @@ int InnerProduct_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, co
             if (bias_term)
                 sum = bias_data[p];
 
-            const __fp16* kptr = weight_data_fp16.row<__fp16>(p);
+            const __fp16* kptr = weight_data_tm.row<__fp16>(p);
 
             const __fp16* sptr = bottom_blob_flattened;
 

src/layer/riscv/innerproduct_riscv.h

@@ -39,8 +39,9 @@ protected:
 public:
     Layer* flatten;
 
+    Mat weight_data_tm;
+
     // fp16
-    Mat weight_data_fp16;
     Mat bias_data_fp16;
 };
 

src/layer/x86/convolutiondepthwise_x86.cpp

@@ -96,8 +96,6 @@ int ConvolutionDepthWise_x86::create_pipeline(const Option& opt)
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, group);
             convert_packing(weight_data_r2, weight_data_tm, 16, opt);
-
-            return 0;
         }
 #endif // __AVX512F__
 
@@ -106,8 +104,6 @@ int ConvolutionDepthWise_x86::create_pipeline(const Option& opt)
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, group);
             convert_packing(weight_data_r2, weight_data_tm, 8, opt);
-
-            return 0;
         }
 #endif // __AVX__
 
@@ -116,8 +112,6 @@ int ConvolutionDepthWise_x86::create_pipeline(const Option& opt)
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, group);
             convert_packing(weight_data_r2, weight_data_tm, 4, opt);
-
-            return 0;
         }
 #endif // __SSE2__
 
@@ -127,14 +121,23 @@ int ConvolutionDepthWise_x86::create_pipeline(const Option& opt)
             if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
             {
                 weight_data_tm = weight_data;
-                return 0;
             }
-            if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
+            else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
             {
                 weight_data_tm = weight_data;
-                return 0;
+            }
+            else
+            {
+                create_group_ops(opt);
             }
         }
+
+        if (opt.lightmode)
+        {
+            weight_data.release();
+        }
+
+        return 0;
     }
 
     // group convolution