[cherry-pick] fix int16 store in fc-bias bug and add sa8155/sa8195 devices support (#8955)

* [int16_store_pass]fix bias on fc using int16-type to store error (#8918)

* [device_info] add sa8155 and sa8195 info support (#8937)

Author: chenjiaoAngel

lite/api/light_api.cc

@@ -316,7 +316,14 @@ void LightPredictor::DequantizeWeight() {
             } else if (op_type == "fc" || op_type == "mul" ||
                        op_type == "lookup_table") {
               int64_t chin = input_tensor->dims()[0];
-              int64_t chout = input_tensor->dims()[1];
+              int64_t chout = input_tensor->numel() / chin;
+              if (input_tensor->dims().size() > 1) {
+                chout = input_tensor->dims()[1];
+              } else {
+                // swap
+                chout = chin;
+                chin = 1;
+              }
               CHECK_EQ(scale_list.size(), chout);
               if (quantize_weight_bits == 8) {
                 const int8_t* int_data = tmp_tensor.data<int8_t>();

lite/core/device_info.cc

@@ -706,8 +706,21 @@ bool DeviceInfo::SetCPUInfoByName() {
     big_core_ids_ = {4, 5, 6, 7};
     little_core_ids_ = {0, 1, 2, 3};
     cluster_ids_ = {1, 1, 1, 1, 0, 0, 0, 0};
-    SetArchInfo(2, kA76, kA55);
-    SetCacheInfo(0, 2, 192 * 1024, 256 * 1024);
+    SetArchInfo(3, kGold_Prime, kGold, kSilver);
+    SetCacheInfo(0, 3, 512 * 1024, 256 * 1024, 128 * 1024);
+    SetCacheInfo(1, 3, 512 * 1024, 256 * 1024, 128 * 1024);
+    SetCacheInfo(2, 1, 2 * 1024 * 1024);
+    SetFP16Info(1, 1);
+    SetDotInfo(2, 1, 1);
+    return true;
+  } else if (dev_name_.find("SA8195") != std::string::npos) {  // sa8195
+    core_num_ = 8;
+    core_ids_ = {0, 1, 2, 3, 4, 5, 6, 7};
+    big_core_ids_ = {4, 5, 6, 7};
+    little_core_ids_ = {0, 1, 2, 3};
+    cluster_ids_ = {1, 1, 1, 1, 0, 0, 0, 0};
+    SetArchInfo(2, kGold_Prime, kSilver);
+    SetCacheInfo(0, 2, 512 * 1024, 128 * 1024);
     SetCacheInfo(1, 2, 512 * 1024, 128 * 1024);
     SetCacheInfo(2, 1, 4 * 1024 * 1024);
     SetFP16Info(1, 1);

lite/core/device_info.h

@@ -49,6 +49,9 @@ typedef enum {
   kA76 = 76,
   kA77 = 77,
   kA78 = 78,
+  kGold = 79,
+  kGold_Prime = 80,
+  kSilver = 81,
   kARMArch_UNKOWN = -1
 } ARMArch;
 

lite/core/optimizer/mir/fusion/flatten_fc_fuser.cc

@@ -97,6 +97,8 @@ cpp::OpDesc FlattenFcFuser::GenOpDesc(const key2nodes_t& matched) {
   op_desc.SetOutput("Out", {matched.at("fc_out")->arg()->name});
   auto in_num_col_dim = op_desc.GetAttr<int>("in_num_col_dims");
   op_desc.SetAttr("in_num_col_dims", in_num_col_dim);
+  std::string op_type = "mul";
+  op_desc.SetAttr("op_type", op_type);
   return op_desc;
 }
 

lite/core/optimizer/mir/post_quant_dynamic_pass.cc

@@ -29,7 +29,7 @@ const std::vector<std::string> PostQuantDynamicPass::quant_axis1_ops = {
     "fc", "mul", "matmul", "matmul_v2", "lookup_table"};
 
 std::vector<std::string> PostQuantDynamicPass::quant_ops = {
-    "conv2d", "mul", "lookup_table"};
+    "fc", "conv2d", "mul", "matmul", "matmul_v2", "lookup_table"};
 
 static bool abs_compare(float a, float b) {
   return std::fabs(a) < std::fabs(b);
@@ -202,6 +202,9 @@ void PostQuantDynamicPass::Apply(const std::unique_ptr<SSAGraph>& graph) {
         auto iter =
             std::find(quant_axis1_ops.begin(), quant_axis1_ops.end(), op_type);
         int quant_axis = iter != quant_axis1_ops.end() ? 1 : 0;
+        if (weight->dims().size() == 1) {
+          quant_axis = 0;
+        }
         PostQuantDynamicPerChannel(
             op_info, weight, weight_name, quant_axis, quant_bits);
       }