Add test attribute in channelwise_quant op, test=develop (#27742)

* Add test attribute in channelwise_quant op, test=develop

paddle/fluid/operators/fake_quantize_op.cc

@@ -404,6 +404,10 @@ class FakeChannelWiseQuantizeAbsMaxOpMaker
                                 "the received is %d",
                                 bit_length));
         });
+    AddAttr<bool>("is_test",
+                  "(bool, default false) Set to true for inference only, false "
+                  "for training. Some layers may run faster when this is true.")
+        .SetDefault(false);
     AddComment(R"DOC(
 The scale of FakeChannelWiseQuantize operator is a vector.
 In detail, each channel of the input X has a scale value.

paddle/fluid/operators/fake_quantize_op.h

@@ -146,16 +146,19 @@ class FakeChannelWiseQuantizeAbsMaxKernel : public framework::OpKernel<T> {
 
     auto* out = context.Output<framework::Tensor>("Out");
     auto* out_scale = context.Output<framework::Tensor>("OutScale");
-    T* out_scale_data = out_scale->mutable_data<T>(context.GetPlace());
     out->mutable_data<T>(context.GetPlace());
 
     int bit_length = context.Attr<int>("bit_length");
     int bin_cnt = std::pow(2, bit_length - 1) - 1;
     int quant_axis = context.Attr<int>("quant_axis");
+    bool is_test = context.Attr<bool>("is_test");
 
     auto& dev_ctx = context.template device_context<DeviceContext>();
+    if (!is_test) {
+      T* out_scale_data = out_scale->mutable_data<T>(context.GetPlace());
       FindChannelAbsMaxFunctor<DeviceContext, T>()(dev_ctx, *in, quant_axis,
                                                    out_scale_data);
+    }
     ChannelClipAndFakeQuantFunctor<DeviceContext, T>()(
         dev_ctx, *in, *out_scale, bin_cnt, quant_axis, out);
   }

python/paddle/fluid/contrib/slim/quantization/quantization_pass.py

@@ -758,6 +758,7 @@ class QuantizationTransformPass(object):
             attrs={
                 'bit_length': quant_bits,
                 'quant_axis': quant_axis,
+                'is_test': self._is_test,
                 'op_role': core.op_proto_and_checker_maker.OpRole.Forward
             },
             inputs={'X': var_node},
@@ -1331,16 +1332,25 @@ class QuantizationFreezePass(object):
 
     def _quant(self, x, scale, num_bits, quant_axis):
         assert quant_axis in [0, 1], 'quant_axis should be 0 or 1 for now.'
+        bnt = (1 << (num_bits - 1)) - 1
+
+        def _clip(x, scale):
+            x[x > scale] = scale
+            x[x < -scale] = -scale
+            return x
+
         if isinstance(scale, list):
             for i, s in enumerate(scale):
                 if quant_axis == 0:
-                    x[i] = np.round(x[i] / s * ((1 << (num_bits - 1)) - 1))
+                    x[i] = _clip(x[i], s)
+                    x[i] = np.round(x[i] / s * bnt)
                 else:
-                    x[:, i] = np.round(x[:, i] / s * (
+                    x[:, i] = _clip(x[:, i], s)
-                        (1 << (num_bits - 1)) - 1))
+                    x[:, i] = np.round(x[:, i] / s * bnt)
-            return x
         else:
-            return np.round(x / scale * ((1 << (num_bits - 1)) - 1))
+            x = _clip(x, scale)
+            x = np.round(x / scale * bnt)
+        return x
 
 
 class ConvertToInt8Pass(object):