quantize bias of the quantization parameters (#48749)

Summary:
Pull Request resolved: #48749

this change reverts D25179863 (55e225a) because in 1.0.0.14 this behavior got
reintroduced
we believe this was already working pre 1.0.0.9, then intel regressed which is
why we had to remove this quantization section, and in 1.0.0.14 they fixed it

Test Plan:
we tested ctr_instagram_5x which now passes with bitwise matching
hl475 will test the top6 models and if they match, we will use this point
to lock any further changes in the future

Reviewed By: venkatacrc

Differential Revision: D25283605

fbshipit-source-id: 33aa9af008c113d4d61e3461a44932b502bf42ea

caffe2/contrib/fakelowp/test/test_int8_ops_nnpi.py

@@ -8,7 +8,12 @@
 from caffe2.python.fakelowp.test_utils import print_test_debug_info
 import caffe2.python.serialized_test.serialized_test_util as serial
 
-core.GlobalInit(["caffe2", "--caffe2_log_level=-3", "--glow_global_fp16=1"])
+core.GlobalInit(["caffe2",
+                 "--caffe2_log_level=-3",
+                 "--glow_global_fp16=1",
+                 "--glow_clip_quant_range_to_fp16=1",
+                 "--glow_global_fp16_constants=1"
+                 ])
 
 
 class Int8OpsTest(serial.SerializedTestCase):

caffe2/quantization/server/fbgemm_pack_op.cc

@@ -207,6 +207,13 @@ void QuantizeConvBias(
         bias.data<int32_t>(), bias.data<int32_t>() + bias.numel());
   } else {
     const float* bdata = bias.data<float>();
+    vector<float> bdata_local;
+    if (use_fp16) {
+      bdata_local.resize(bias.numel());
+      fbgemm::RoundToFloat16(
+              bdata, bdata_local.data(), bias.numel(), false /* FLAGS_caffe2_fbgemm_fake_fp16_clamp */);
+      bdata = bdata_local.data();
+    }
     b_quantized.resize(bias.numel());
     for (int g = 0; g < filter_qparams.size(); ++g) {
       int i_begin = g * (M / filter_qparams.size());