Fix wrong method in some cases

dabnn/layers/BinConv.cpp

@@ -25,14 +25,16 @@ BinConv::BinConv(NetCP net, const std::string &name, css input, css weight,
       stride_h(stride_h),
       stride_w(stride_w) {
     auto &mat_map = net.lock()->mat_map_;
-    const auto binaized_name = "binaized_for_" + output + "_cal";
-    if (mat_map.find(binaized_name) == mat_map.end()) {
-        auto &input_mat = *mat_map[input];
-        mat_map[binaized_name] =
-            std::make_shared<Mat>(input_mat.h, input_mat.w, input_mat.elem_c,
-                                  DataType::Bit, binaized_name);
+    if (method() == Method::DIRECT_CONV || method() == Method::BCONV_NAIVE) {
+        const auto binaized_name = "binaized_for_" + output + "_cal";
+        if (mat_map.find(binaized_name) == mat_map.end()) {
+            auto &input_mat = *mat_map[input];
+            mat_map[binaized_name] = std::make_shared<Mat>(
+                input_mat.h, input_mat.w, input_mat.elem_c, DataType::Bit,
+                binaized_name);
+        }
+        binarized_mat = mat(binaized_name);
     }
-    binarized_mat = mat(binaized_name);
 
     const auto pad_name = "pad_for_" + output + "_cal";
     if (mat_map.find(pad_name) == mat_map.end()) {
@@ -43,18 +45,17 @@ BinConv::BinConv(NetCP net, const std::string &name, css input, css weight,
     }
     padded_mat = mat(pad_name);
 
-    const auto col_mat_name = "col_for_" + output + "_cal";
-    if (mat_map.find(col_mat_name) == mat_map.end()) {
-        const auto len =
-            output_mat->h * output_mat->w *
-            align_to(weight_mat->h * weight_mat->w * input_mat->elem_c, 128);
-        mat_map[col_mat_name] =
-            std::make_shared<Mat>(1, 1, len, bnn::DataType::Bit);
-    }
-    col_mat = mat(col_mat_name);
-
-    if (net.lock()->optimize && !direct_conv_compatible() &&
-        gemm_compatible()) {
+    if (method() == Method::BGEMM || method() == Method::BGEMM_NAIVE) {
+        const auto col_mat_name = "col_for_" + output + "_cal";
+        if (mat_map.find(col_mat_name) == mat_map.end()) {
+            const auto len =
+                output_mat->h * output_mat->w *
+                align_to(weight_mat->h * weight_mat->w * input_mat->elem_c,
+                         128);
+            mat_map[col_mat_name] =
+                std::make_shared<Mat>(1, 1, len, bnn::DataType::Bit);
+        }
+        col_mat = mat(col_mat_name);
         const auto trans_weight_mat_name = "trans_" + weight;
         // transpose the weight for bgemm
         const int m = weight_mat->n;
@@ -76,6 +77,24 @@ BinConv::BinConv(NetCP net, const std::string &name, css input, css weight,
     }
 }
 
+BinConv::Method BinConv::method() const {
+    if (net_.lock()->optimize) {
+        if (direct_conv_compatible()) {
+            return Method::DIRECT_CONV;
+        } else if (gemm_compatible()) {
+            return Method::BGEMM;
+        } else {
+            return Method::BCONV_NAIVE;
+        }
+    } else {
+        if (weight_mat->c == 1) {
+            return Method::BCONV_NAIVE;
+        } else {
+            return Method::BGEMM_NAIVE;
+        }
+    }
+}
+
 bool BinConv::direct_conv_compatible() const {
 #ifdef __aarch64__
     if (weight_mat->h == 3 && weight_mat->w == 3 && input_mat->elem_c == 64 &&
@@ -121,12 +140,14 @@ bool BinConv::gemm_compatible() const {
 }
 
 void BinConv::forward_impl() const {
-    if (net_.lock()->optimize) {
-        if (direct_conv_compatible()) {
+    switch (method()) {
+        case Method::DIRECT_CONV: {
             pack_mat(*input_mat, *binarized_mat);
             pad(*binarized_mat, pad_h, pad_w, *padded_mat);
             bconv_3x3(*padded_mat, *weight_mat, *output_mat, stride_h);
-        } else if (gemm_compatible()) {
+            break;
+        }
+        case Method::BGEMM: {
             output_mat->fill<float>(0.f);
 
             bnn::fused_binarize_im2col(*input_mat, weight_mat->h, weight_mat->w,
@@ -139,17 +160,31 @@ void BinConv::forward_impl() const {
             bgemm(m, n, k, static_cast<uint64_t *>(transposed_weight_mat->data),
                   m, static_cast<uint64_t *>(col_mat->data), k,
                   static_cast<float *>(output_mat->data), m);
-        } else {
+            break;
+        }
+        case Method::BGEMM_NAIVE: {
+            output_mat->fill<float>(0.f);
+
+            bnn::fused_binarize_im2col(*input_mat, weight_mat->h, weight_mat->w,
+                                       pad_h, pad_w, stride_h, stride_w, 1, 1,
+                                       *col_mat);
+
+            const int m = weight_mat->n;
+            const int n = output_mat->h * output_mat->w;
+            const int k = weight_mat->total() / weight_mat->n;
+            bgemm_naive(m, n, k,
+                        static_cast<uint64_t *>(transposed_weight_mat->data), m,
+                        static_cast<uint64_t *>(col_mat->data), k,
+                        static_cast<float *>(output_mat->data), m);
+            break;
+        }
+        case Method::BCONV_NAIVE: {
             pack_mat(*input_mat, *binarized_mat);
             baseline_bconv(*binarized_mat, *weight_mat, weight_mat->h,
                            weight_mat->w, pad_h, pad_w, stride_h, stride_w, 1,
                            1, output_mat->c, *output_mat);
+            break;
         }
-    } else {
-        pack_mat(*input_mat, *binarized_mat);
-        baseline_bconv(*binarized_mat, *weight_mat, weight_mat->h,
-                       weight_mat->w, pad_h, pad_w, stride_h, stride_w, 1, 1,
-                       output_mat->c, *output_mat);
     }
 }
 

dabnn/layers/BinConv.h

@@ -26,8 +26,15 @@ class BinConv : public Layer {
     virtual std::string to_str() const;
 
    private:
+    enum Method {
+        DIRECT_CONV = 0,
+        BGEMM,
+        BCONV_NAIVE,
+        BGEMM_NAIVE
+    };
     bool direct_conv_compatible() const;
     bool gemm_compatible() const;
+    Method method() const;
 };
 }  // namespace bnn