Merge pull request #62 from Cambricon/fix_conv_quant

fix(quant): fix quant compute error

lite/kernels/mlu/bridges/conv_op.cc

@@ -249,6 +249,10 @@ int ConvConverter(void* ctx, OpLite* op, KernelBase* kernel) {
     CNML_CALL(cnmlDestroyConvOpParam(&conv_param));
   } else {
     cnmlConvOpParam_t conv_param;
+    VLOG(5) << "conv param (" << input_var_name << ")"
+            << "stride: " << strides[0] << ',' << strides[1] << '\t'
+            << "dilations: " << dilations[0] << ',' << dilations[1] << '\t'
+            << "paddings: " << paddings[0] << ',' << paddings[2] << std::endl;
     CNML_CALL(cnmlCreateConvOpParam(&conv_param,
                                     strides[0],
                                     strides[1],
@@ -272,7 +276,7 @@ int ConvConverter(void* ctx, OpLite* op, KernelBase* kernel) {
     graph->SetComputingDataType(
         conv_op,
         filter_tensor->mlu_tensor(),
-        1 / *min_element(weight_scale.begin(), weight_scale.end()));
+        1 / *max_element(weight_scale.begin(), weight_scale.end()));
   }
   CNML_CALL(cnmlSetOperationComputingLayout(conv_op, CNML_NHWC));
   if (HasInputArg(op_info, scope, "Bias")) {

lite/kernels/mlu/bridges/fc_op.cc

@@ -157,7 +157,7 @@ int FCConverter(void* ctx, OpLite* op, KernelBase* kernel) {
   graph->SetComputingDataType(
       fc_op,
       w_tensor->mlu_tensor(),
-      1 / *min_element(weight_scale.begin(), weight_scale.end()));
+      1 / *max_element(weight_scale.begin(), weight_scale.end()));
 
   graph->FuseOp(fc_op);
   CNML_CALL(cnmlDestroyBaseOp(&fc_op));

lite/kernels/mlu/bridges/graph.h

@@ -215,8 +215,11 @@ class Graph {
                             float scale,
                             cnmlDataType_t data_type = CNML_DATA_INT8) {
     cnmlQuantizedParam_t quant_param;
-    CNML_CALL(
-        cnmlCreateQuantizedParam(&quant_param, scale2position(scale), 1, 0.0));
+    int pos = scale2position(scale);
+    auto cnml_scale = pow(2, pos) * scale;
+    VLOG(5) << "[cnml quantized param] pos: " << pos
+            << "\tscale: " << cnml_scale << std::endl;
+    CNML_CALL(cnmlCreateQuantizedParam(&quant_param, pos, cnml_scale, 0.0));
     CNML_CALL(
         cnmlSetOperationComputingDataType(op, tensor, data_type, quant_param));
     CNML_CALL(cnmlDestroyQuantizedParam(&quant_param));

lite/kernels/mlu/bridges/utility.cc

@@ -60,8 +60,6 @@ void transpose(float* input_data,
   }
 }
 
-int scale2position(float scale) { return static_cast<int>(-std::log2(scale)); }
-
 void dequant(float* dst, int8_t* src, size_t size, float scale) {
   for (size_t i = 0; i < size; ++i) {
     dst[i] = static_cast<float>(src[i]) * scale;

lite/kernels/mlu/bridges/utility.h

@@ -36,7 +36,9 @@ void transpose(float* input_data,
                float* output_data,
                std::vector<int> input_shape,
                std::vector<int> axis);
-int scale2position(float scale);
+
+inline int scale2position(float scale) { return std::floor(-std::log2(scale)); }
+
 void dequant(float* dst, int8_t* src, size_t size, float scale);
 
 void dequant(float* dst,