#[repr(u32)]pub enum rocblas_gemm_flags_ {
rocblas_gemm_flags_none,
rocblas_gemm_flags_pack_int8x4,
rocblas_gemm_flags_use_cu_efficiency,
rocblas_gemm_flags_fp16_alt_impl,
}
Expand description
\brief Control flags passed into gemm algorithms invoked by Tensile Host
Variants§
rocblas_gemm_flags_none
\brief Default empty flags
rocblas_gemm_flags_pack_int8x4
\brief Before ROCm 4.2, this flags is not implemented and rocblas uses packed-Int8x4 by default. After ROCm 4.2, set flag is neccesary if we want packed-Int8x4. Default (0x0) uses unpacked.
rocblas_gemm_flags_use_cu_efficiency
\brief Select the gemm problem with the highest efficiency per compute unit used. Useful for running multiple smaller problems simultaneously. This takes precedence over the performance metric set in rocblas_handle and currently only works for gemm_*_ex problems.
rocblas_gemm_flags_fp16_alt_impl
\brief Select an alternate implementation for the MI200 FP16 HPA (High Precision Accumulate) GEMM kernel utilizing the BF16 matrix instructions with reduced accuracy in cases where computation cannot tolerate the FP16 matrix instructions flushing subnormal FP16 input/output data to zero. See the “MI200 (gfx90a) Considerations” section for more details.
Trait Implementations§
source§impl Clone for rocblas_gemm_flags_
impl Clone for rocblas_gemm_flags_
source§fn clone(&self) -> rocblas_gemm_flags_
fn clone(&self) -> rocblas_gemm_flags_
1.0.0 · source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source
. Read moresource§impl Debug for rocblas_gemm_flags_
impl Debug for rocblas_gemm_flags_
source§impl Hash for rocblas_gemm_flags_
impl Hash for rocblas_gemm_flags_
source§impl PartialEq<rocblas_gemm_flags_> for rocblas_gemm_flags_
impl PartialEq<rocblas_gemm_flags_> for rocblas_gemm_flags_
source§fn eq(&self, other: &rocblas_gemm_flags_) -> bool
fn eq(&self, other: &rocblas_gemm_flags_) -> bool
self
and other
values to be equal, and is used
by ==
.