Add comments to explain inv-variance in batchnorm forward and backward

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp

@@ -529,6 +529,7 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
             auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
                 resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
 
+            // calculate inv-variance as 1/sqrt(epsilon+variance)
             static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
                 welford_var_thread_buf(I) =
                     type_convert<AccDataType>(1.0f) / sqrt(epsilon + welford_var_thread_buf[I]);

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp

@@ -359,7 +359,7 @@ struct GridwiseWelfordSecondHalfReduceFirstHalf
                                       welford_count_thread_buf(I));
             });
 
-            // calculate inv-variance from variance, stored in place
+            // calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
             static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
                 welford_var_thread_buf(I) =
                     type_convert<AccDataType>(1.0) / sqrt(welford_var_thread_buf[I] + epsilon);

include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp

@@ -410,6 +410,7 @@ struct GridwiseBatchNormBackwardWithBlockwiseWelford
                 BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
             });
 
+            // calculate inv-variance as 1/sqrt(epsilon+variance)
             static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
                 inv_var_thread_buf(I) =
                     type_convert<AccDataType>(1.0) / sqrt(var_thread_buf[I] + epsilon);

include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp

@@ -441,6 +441,7 @@ struct GridwiseBatchNormForwardWithBlockwiseWelford
             auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
                 resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
 
+            // calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
             static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
                 var_thread_buf(I) =
                     type_convert<AccDataType>(1.0f) / sqrt(epsilon + var_thread_buf[I]);