Avoid Ampere misaligment issue

CHANGELOG.md

@@ -41,6 +41,7 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 - Broken links to MNIST data sets
 
 ### Changed
+- Set REQUIRED_BIAS_ALIGNMENT = 16 in tensors/gpu/prod.cpp to avoid memory-misalignment on certain Ampere GPUs.
 - For BUILD_ARCH != native enable all intrinsics types by default, can be disabled like this: -DCOMPILE_AVX512=off
 - Moved FBGEMM pointer to commit c258054 for gcc 9.3+ fix
 - Change compile options a la -DCOMPILE_CUDA_SM35 to -DCOMPILE_KEPLER, -DCOMPILE_MAXWELL,

src/tensors/gpu/prod.cpp

@@ -22,7 +22,7 @@ namespace gpu {
 // It seems that the bias must be 8 byte aligned for the cublasLt epilogue to work. Therefore,
 // if the bias pointer is not 8 byte aligned, we do a normal matmul in cublasLt and invoke a 
 // custom epilogue kernel.
-static constexpr int REQUIRED_BIAS_ALIGNMENT = 8;  
+static constexpr int REQUIRED_BIAS_ALIGNMENT = 16; // @TODO: MJD: changed this to 16 to avoid alignment error on A100. Seems to work fine.
 
 // Used to set preferences for cublasLt to filter out algos if matrices to not meet default 256 byte alignment
 int getAlignmentUpTo256(const void *ptr) {