[cherry-pick] fix Int8 conv compute error and support conv+hardswish fusion (#7657)

Author: chenjiaoAngel

lite/backends/arm/math/conv3x3_winograd_int8.cc

@@ -122,7 +122,8 @@ void conv_compute_2x2_3x3_int8(const int8_t* input,
       (int32_t*)(g_trans_remain_tmp_data + threads * 128);  // NOLINT
   auto act_type = act_param.active_type;
   int flag_act = 0;  // relu: 1, relu6: 2, leakey: 3
-  float alpha[4] = {0.f, 0.f, 0.f, 0.f};
+  float alpha[12] = {
+      0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
   bool flag_bias = (bias == nullptr) ? false : true;
   if (act_param.has_active) {
     if (act_type == lite_api::ActivationType::kRelu) {
@@ -141,6 +142,13 @@ void conv_compute_2x2_3x3_int8(const int8_t* input,
       alpha[1] = local_alpha;
       alpha[2] = local_alpha;
       alpha[3] = local_alpha;
+    } else if (act_type == lite_api::ActivationType::kHardSwish) {
+      flag_act = 4;
+      for (int i = 0; i < 4; i++) {
+        alpha[i] = 1.f / act_param.hard_swish_scale;
+        alpha[i + 4] = act_param.hard_swish_offset;
+        alpha[i + 8] = act_param.hard_swish_threshold;
+      }
     }
   }
   // begin compute
@@ -435,7 +443,8 @@ void conv_compute_4x4_3x3_int8(const int8_t* input,
       g_trans_tmp_output_data + threads * 192;
   auto act_type = act_param.active_type;
   int flag_act = 0;  // relu: 1, relu6: 2, leakey: 3
-  float alpha[4] = {0.f, 0.f, 0.f, 0.f};
+  float alpha[12] = {
+      0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
   bool flag_bias = (bias == nullptr) ? false : true;
   if (act_param.has_active) {
     if (act_type == lite_api::ActivationType::kRelu) {
@@ -454,6 +463,13 @@ void conv_compute_4x4_3x3_int8(const int8_t* input,
       alpha[1] = local_alpha;
       alpha[2] = local_alpha;
       alpha[3] = local_alpha;
+    } else if (act_type == lite_api::ActivationType::kHardSwish) {
+      flag_act = 4;
+      for (int i = 0; i < 4; i++) {
+        alpha[i] = 1.f / act_param.hard_swish_scale;
+        alpha[i + 4] = act_param.hard_swish_offset;
+        alpha[i + 8] = act_param.hard_swish_threshold;
+      }
     }
   }
   // begin compute

lite/backends/arm/math/conv3x3s1_direct_int8.cc

@@ -47,7 +47,8 @@ void conv_3x3s1_direct_int8(const int8_t* din,
   auto act_param = param.activation_param;
   auto act_type = act_param.active_type;
   int flag_act = 0;  // relu: 1, relu6: 2, leakey: 3
-  float alpha[4] = {0.f, 0.f, 0.f, 0.f};
+  float alpha[12] = {
+      0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
   if (act_param.has_active) {
     if (act_type == lite_api::ActivationType::kRelu) {
       flag_act = 1;
@@ -65,6 +66,13 @@ void conv_3x3s1_direct_int8(const int8_t* din,
       alpha[1] = local_alpha;
       alpha[2] = local_alpha;
       alpha[3] = local_alpha;
+    } else if (act_type == lite_api::ActivationType::kHardSwish) {
+      flag_act = 4;
+      for (int i = 0; i < 4; i++) {
+        alpha[i] = 1.f / act_param.hard_swish_scale;
+        alpha[i + 4] = act_param.hard_swish_offset;
+        alpha[i + 8] = act_param.hard_swish_threshold;
+      }
     }
   }
   int pad_h = paddings[0];

lite/backends/arm/math/conv3x3s2_direct_int8.cc

@@ -51,7 +51,8 @@ void conv_3x3s2_direct_int8(const int8_t* din,
   auto act_param = param.activation_param;
   auto act_type = act_param.active_type;
   int flag_act = 0;  // relu: 1, relu6: 2, leakey: 3
-  float alpha[4] = {0.f, 0.f, 0.f, 0.f};
+  float alpha[12] = {
+      0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
   if (act_param.has_active) {
     if (act_type == lite_api::ActivationType::kRelu) {
       flag_act = 1;
@@ -69,6 +70,13 @@ void conv_3x3s2_direct_int8(const int8_t* din,
       alpha[1] = local_alpha;
       alpha[2] = local_alpha;
       alpha[3] = local_alpha;
+    } else if (act_type == lite_api::ActivationType::kHardSwish) {
+      flag_act = 4;
+      for (int i = 0; i < 4; i++) {
+        alpha[i] = 1.f / act_param.hard_swish_scale;
+        alpha[i + 4] = act_param.hard_swish_offset;
+        alpha[i + 8] = act_param.hard_swish_threshold;
+      }
     }
   }
   int pad_h = paddings[0];
@@ -503,7 +511,8 @@ void conv_3x3s2_direct_int8(const int8_t* din,
   auto act_param = param.activation_param;
   auto act_type = act_param.active_type;
   int flag_act = 0;  // relu: 1, relu6: 2, leakey: 3
-  float alpha[4] = {0.f, 0.f, 0.f, 0.f};
+  float alpha[12] = {
+      0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
   if (act_param.has_active) {
     if (act_type == lite_api::ActivationType::kRelu) {
       flag_act = 1;
@@ -521,6 +530,13 @@ void conv_3x3s2_direct_int8(const int8_t* din,
       alpha[1] = local_alpha;
       alpha[2] = local_alpha;
       alpha[3] = local_alpha;
+    } else if (act_type == lite_api::ActivationType::kHardSwish) {
+      flag_act = 4;
+      for (int i = 0; i < 4; i++) {
+        alpha[i] = 1.f / act_param.hard_swish_scale;
+        alpha[i + 4] = act_param.hard_swish_offset;
+        alpha[i + 8] = act_param.hard_swish_threshold;
+      }
     }
   }
   int pad_h = paddings[0];