support uint8 and int8 allreduce in tensorflow

update CHANGELOG
fix int8, uint8 tests when averaging in allreduce
extend uint8 and int8 allreduce tests to xla and process sets

Signed-off-by: Vignesh Kothapalli <k.vignesh1420@gmail.com>

CHANGELOG.md

@@ -8,6 +8,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
+- Added support for `int8, uint8` allreduce and grouped_allreduce in tensorflow. ([#3649](https://github.com/horovod/horovod/pull/3649))
 - Added support for batched memory copies in GPUAllgather. ([#3590](https://github.com/horovod/horovod/pull/3590))
 - Added support for batched memory copies in GPUReducescatter. ([#3621](https://github.com/horovod/horovod/pull/3621))
 - Added `hvd.grouped_allgather()` and `hvd.grouped_reducescatter()` operations. ([#3594](https://github.com/horovod/horovod/pull/3594))

horovod/tensorflow/mpi_ops.cc

@@ -510,7 +510,7 @@ REGISTER_KERNEL_BUILDER(Name("HorovodAllreduce").Device(DEVICE_GPU),
 #endif
 
 REGISTER_OP("HorovodAllreduce")
-    .Attr("T: {int32, int64, float16, float32, float64}")
+    .Attr("T: {uint8, int8, int32, int64, float16, float32, float64}")
     .Attr("reduce_op: int")
     .Attr("prescale_factor: float")
     .Attr("postscale_factor: float")
@@ -647,7 +647,7 @@ REGISTER_KERNEL_BUILDER(Name("HorovodGroupedAllreduce").Device(DEVICE_GPU),
 #endif
 
 REGISTER_OP("HorovodGroupedAllreduce")
-    .Attr("T: {int32, int64, float16, float32, float64}")
+    .Attr("T: {uint8, int8, int32, int64, float16, float32, float64}")
     .Attr("reduce_op: int")
     .Attr("prescale_factor: float")
     .Attr("postscale_factor: float")