LayerNormInt8QuantizeFakeNNPI fix to match ICEREF. (#47140)

Summary:
Pull Request resolved: #47140

LayerNorm + Int8Quantize fix to match ICEREF.

(Note: this ignores all push blocking failures!)

Test Plan:
buck test --debug //caffe2/caffe2/contrib/fakelowp/test:test_layernorm_nnpi_fp16nnpi -- test_fused_ln_quantize --print-passing-details

https://internalfb.com/intern/testinfra/testrun/7881299371969005

Reviewed By: hyuen

Differential Revision: D24659904

fbshipit-source-id: 026d1a1f69a68eca662a39752af5ab0756bace2d

caffe2/contrib/fakelowp/layernorm_fp16_fake_op.h

@@ -189,7 +189,7 @@ class LayerNormFakeFp16Op final : public Operator<CPUContext> {
       int Nout = X.numel();
 
       std::vector<float> inv_scalev(Nout, inv_scale);
-      std::vector<float> offsetv(Nout, Y_offset - 128.0);
+      std::vector<float> offsetv(Nout, Y_offset);
       uint8_t* Y_uint8_data = Y_int8->t.template mutable_data<uint8_t>();
 
       fake_fp16::fma_fp16(Nout, Y_fp16.data(), inv_scalev.data(), offsetv.data());
@@ -200,7 +200,6 @@ class LayerNormFakeFp16Op final : public Operator<CPUContext> {
       for (int i = 0; i < Nout; i++) {
         float halfRes = offsetv[i];
         halfRes = round(halfRes);
-        halfRes = halfRes + 128.0;
         if (std::isinf(halfRes)) {
           if (halfRes > 0) {
             halfRes = qmax;

caffe2/contrib/fakelowp/test/test_layernorm_nnpi_fp16.py

@@ -1,8 +1,3 @@
-
-
-
-
-
 import numpy as np
 import caffe2.python.fakelowp.init_shared_libs  # noqa
 from caffe2.proto import caffe2_pb2