Merge branch 'Tencent:master' into arm-conv-unified-elempack-4

.github/workflows/release-python.yml

@@ -68,7 +68,7 @@ jobs:
         brew uninstall --ignore-dependencies libomp
 
     - name: Build wheels
-      uses: pypa/cibuildwheel@v2.15.0
+      uses: pypa/cibuildwheel@v2.16.2
       env:
         CIBW_ARCHS_MACOS: ${{ matrix.arch }}
         CIBW_ARCHS_LINUX: ${{ matrix.arch }}
@@ -124,7 +124,7 @@ jobs:
         platforms: all
 
     - name: Build wheels
-      uses: pypa/cibuildwheel@v2.15.0
+      uses: pypa/cibuildwheel@v2.16.2
       env:
         CIBW_ARCHS_LINUX: ${{ matrix.arch }}
         CIBW_BUILD: ${{ matrix.build }}

benchmark/README.md

@@ -1631,7 +1631,92 @@ cooling_down = 1
   vision_transformer  min = 6605.19  max = 6606.66  avg = 6605.73
           FastestDet  min =   52.11  max =   52.97  avg =   52.61
 ```
+### Raspberry Pi 5 Broadcom BCM2712, Cortex-A76 (ARMv8) (2.4GHz x 4)
+```
+pi@raspberrypi:~/ncnn/benchmark $ ./benchncnn 10 4 0 -1 1
+loop_count = 10
+num_threads = 4
+powersave = 0
+gpu_device = -1
+cooling_down = 1
+          squeezenet  min =    8.56  max =    8.65  avg =    8.61
+     squeezenet_int8  min =   11.65  max =   12.64  avg =   11.94
+           mobilenet  min =   11.32  max =   13.46  avg =   11.75
+      mobilenet_int8  min =   11.30  max =   11.60  avg =   11.45
+        mobilenet_v2  min =   13.57  max =   13.77  avg =   13.63
+        mobilenet_v3  min =    9.18  max =   10.52  avg =    9.48
+          shufflenet  min =    4.56  max =    6.19  avg =    5.98
+       shufflenet_v2  min =    5.04  max =    5.13  avg =    5.09
+             mnasnet  min =    8.27  max =    9.86  avg =    8.65
+     proxylessnasnet  min =    9.36  max =   11.18  avg =    9.62
+     efficientnet_b0  min =   14.77  max =   14.96  avg =   14.87
+   efficientnetv2_b0  min =   19.91  max =   20.11  avg =   19.99
+        regnety_400m  min =   11.91  max =   12.10  avg =   11.96
+           blazeface  min =    2.26  max =    2.29  avg =    2.28
+           googlenet  min =   32.80  max =   33.17  avg =   32.97
+      googlenet_int8  min =   32.63  max =   32.99  avg =   32.78
+            resnet18  min =   23.95  max =   24.21  avg =   24.12
+       resnet18_int8  min =   32.50  max =   32.79  avg =   32.68
+             alexnet  min =   25.31  max =   25.75  avg =   25.51
+               vgg16  min =  162.19  max =  165.08  avg =  163.75
+          vgg16_int8  min =  187.46  max =  191.21  avg =  189.09
+            resnet50  min =   55.95  max =   56.61  avg =   56.29
+       resnet50_int8  min =   73.34  max =   73.97  avg =   73.59
+      squeezenet_ssd  min =   40.48  max =   41.39  avg =   40.92
+ squeezenet_ssd_int8  min =   45.67  max =   46.35  avg =   46.06
+       mobilenet_ssd  min =   31.15  max =   31.73  avg =   31.48
+  mobilenet_ssd_int8  min =   31.09  max =   31.44  avg =   31.27
+      mobilenet_yolo  min =   71.51  max =   72.38  avg =   71.95
+  mobilenetv2_yolov3  min =   47.86  max =   48.41  avg =   48.04
+         yolov4-tiny  min =   55.95  max =   56.51  avg =   56.19
+           nanodet_m  min =   14.26  max =   14.68  avg =   14.48
+    yolo-fastest-1.1  min =    6.48  max =    8.10  avg =    7.30
+      yolo-fastestv2  min =    6.03  max =    7.33  avg =    7.04
+  vision_transformer  min =  613.62  max =  637.97  avg =  629.51
+          FastestDet  min =    6.53  max =    6.66  avg =    6.59
+pi@raspberrypi:~/ncnn/benchmark $ ./benchncnn 10 1 0 -1 1
+loop_count = 10
+num_threads = 1
+powersave = 0
+gpu_device = -1
+cooling_down = 1
+          squeezenet  min =   13.18  max =   13.27  avg =   13.22
+     squeezenet_int8  min =   15.69  max =   15.93  avg =   15.78
+           mobilenet  min =   21.42  max =   21.55  avg =   21.46
+      mobilenet_int8  min =   14.92  max =   20.91  avg =   17.34
+        mobilenet_v2  min =   18.56  max =   23.06  avg =   19.24
+        mobilenet_v3  min =   13.16  max =   13.33  avg =   13.25
+          shufflenet  min =    7.25  max =   11.14  avg =    8.43
+       shufflenet_v2  min =    7.17  max =   11.15  avg =    7.70
+             mnasnet  min =   13.89  max =   13.94  avg =   13.91
+     proxylessnasnet  min =   17.01  max =   17.26  avg =   17.07
+     efficientnet_b0  min =   26.19  max =   26.30  avg =   26.24
+   efficientnetv2_b0  min =   39.69  max =   40.12  avg =   39.97
+        regnety_400m  min =   17.30  max =   17.44  avg =   17.36
+           blazeface  min =    4.74  max =    4.78  avg =    4.76
+           googlenet  min =   57.64  max =   57.84  avg =   57.72
+      googlenet_int8  min =   55.80  max =   56.01  avg =   55.93
+            resnet18  min =   31.90  max =   32.09  avg =   32.00
+       resnet18_int8  min =   56.92  max =   57.16  avg =   57.01
+             alexnet  min =   39.84  max =   40.12  avg =   39.92
+               vgg16  min =  208.33  max =  211.06  avg =  209.64
+          vgg16_int8  min =  437.53  max =  440.55  avg =  439.35
+            resnet50  min =   95.75  max =   96.68  avg =   96.28
+       resnet50_int8  min =  116.80  max =  118.01  avg =  117.57
+      squeezenet_ssd  min =   47.75  max =   47.97  avg =   47.86
+ squeezenet_ssd_int8  min =   61.98  max =   62.90  avg =   62.47
+       mobilenet_ssd  min =   52.83  max =   53.39  avg =   53.07
+  mobilenet_ssd_int8  min =   46.15  max =   46.60  avg =   46.35
+      mobilenet_yolo  min =  117.68  max =  117.97  avg =  117.81
+  mobilenetv2_yolov3  min =   67.37  max =   67.67  avg =   67.48
+         yolov4-tiny  min =   73.85  max =   74.35  avg =   74.10
+           nanodet_m  min =   22.78  max =   23.33  avg =   22.96
+    yolo-fastest-1.1  min =    8.82  max =    8.91  avg =    8.87
+      yolo-fastestv2  min =    8.18  max =   11.42  avg =    8.59
+  vision_transformer  min = 1267.90  max = 1269.45  avg = 1268.82
+          FastestDet  min =    7.79  max =   11.14  avg =    9.03
 
+```
 ### Raspberry Pi Zero 2 W Broadcom BCM2710A1, Cortex-A53 (ARMv8) (1.0GHz x 4)
 
 ```

src/layer/memorydata.cpp

@@ -28,6 +28,7 @@ int MemoryData::load_param(const ParamDict& pd)
     h = pd.get(1, 0);
     d = pd.get(11, 0);
     c = pd.get(2, 0);
+    load_type = pd.get(21, 1);
 
     return 0;
 }
@@ -36,19 +37,19 @@ int MemoryData::load_model(const ModelBin& mb)
 {
     if (d != 0)
     {
-        data = mb.load(w, h, d, c, 1);
+        data = mb.load(w, h, d, c, load_type);
     }
     else if (c != 0)
     {
-        data = mb.load(w, h, c, 1);
+        data = mb.load(w, h, c, load_type);
     }
     else if (h != 0)
     {
-        data = mb.load(w, h, 1);
+        data = mb.load(w, h, load_type);
     }
     else if (w != 0)
     {
-        data = mb.load(w, 1);
+        data = mb.load(w, load_type);
     }
     else // 0 0 0
     {

src/layer/memorydata.h

@@ -35,6 +35,7 @@ class MemoryData : public Layer
     int h;
     int d;
     int c;
+    int load_type;
 
     Mat data;
 };

tests/test_convolution1d.cpp

@@ -77,7 +77,7 @@ static int test_convolution1d_0()
         const int s = kdsp[i][2];
         const int p = kdsp[i][3];
         const int b0 = i % 2;
-        const int b1 = 1 - b1;
+        const int b1 = 1 - b0;
 
         int ret = 0
                   || test_convolution1d(9, 1, 1, k, d, s, p, b0)

tools/pnnx/src/CMakeLists.txt

@@ -154,6 +154,7 @@ set(pnnx_pass_level2_SRCS
     pass_level2/F_mish.cpp
     pass_level2/F_normalize.cpp
     pass_level2/F_pad.cpp
+    pass_level2/F_pairwise_distance.cpp
     pass_level2/F_pixel_shuffle.cpp
     pass_level2/F_pixel_unshuffle.cpp
     pass_level2/F_prelu.cpp

tools/pnnx/src/pass_level2/F_pairwise_distance.cpp

@@ -0,0 +1,44 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "pass_level2.h"
+
+namespace pnnx {
+
+class F_pairwise_distance : public GraphRewriterPass
+{
+public:
+    const char* match_pattern_graph() const
+    {
+        return R"PNNXIR(7767517
+7 6
+pnnx.Input              input_0         0 1 x1
+pnnx.Input              input_1         0 1 x2
+prim::Constant          op_0     	0 1 p value=%p
+prim::Constant          op_1    	0 1 eps value=%eps
+prim::Constant          op_2    	0 1 keepdim value=%keepdim
+aten::pairwise_distance op_3     	5 1 x1 x2 p eps keepdim out
+pnnx.Output             output          1 0 out
+)PNNXIR";
+    }
+
+    const char* type_str() const
+    {
+        return "F.pairwise_distance";
+    }
+};
+
+REGISTER_GLOBAL_PNNX_GRAPH_REWRITER_PASS(F_pairwise_distance, 10)
+
+} // namespace pnnx

tools/pnnx/tests/CMakeLists.txt

@@ -54,6 +54,7 @@ pnnx_add_test(F_max_pool2d)
 pnnx_add_test(F_max_pool3d)
 pnnx_add_test(F_normalize)
 pnnx_add_test(F_pad)
+pnnx_add_test(F_pairwise_distance)
 pnnx_add_test(F_pixel_shuffle)
 pnnx_add_test(F_pixel_unshuffle)
 pnnx_add_test(F_prelu)

tools/pnnx/tests/test_F_pairwise_distance.py

@@ -0,0 +1,58 @@
+# Tencent is pleased to support the open source community by making ncnn available.
+#
+# Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
+#
+# Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+# in compliance with the License. You may obtain a copy of the License at
+#
+# https://opensource.org/licenses/BSD-3-Clause
+#
+# Unless required by applicable law or agreed to in writing, software distributed
+# under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+# CONDITIONS OF ANY KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, x, y):
+        z1 = F.pairwise_distance(x,y,p=1,keepdim=False)
+        z2 = F.pairwise_distance(x,y,p=2,keepdim=True)
+        z3 = F.pairwise_distance(x,y)
+        z4 = F.pairwise_distance(x,y,eps = 1e-3)
+        return z1,z2,z3,z4
+
+def test():
+    net = Model()
+    net.eval()
+
+    torch.manual_seed(0)
+    x = torch.rand(12, 128, 128)
+    y = torch.rand(12, 128, 128)
+
+    a0,a1,a2,a3 = net(x, y)
+
+    # export torchscript
+    mod = torch.jit.trace(net, (x, y))
+    mod.save("test_F_pairwise_distance.pt")
+
+    # torchscript to pnnx
+    import os
+    os.system("../src/pnnx test_F_pairwise_distance.pt inputshape=[12,128,128],[12,128,128]")
+
+    # pnnx inference
+    import test_F_pairwise_distance_pnnx
+    b0,b1,b2,b3 = test_F_pairwise_distance_pnnx.test_inference()
+
+    return torch.equal(a0,b0) and torch.equal(a1,b1) and torch.equal(a2,b2) and torch.equal(a3,b3)
+
+if __name__ == "__main__":
+    if test():
+        exit(0)
+    else:
+        exit(1)