add half gemm support#68

Draft

cayrols wants to merge 25 commits into

icl-utk-edu:masterfrom

cayrols:seb/gemm.add_hgemm_support

Commits on Dec 14, 2023

hgemm: add half definition and interfaces
cayrols
committed
hgemm: add hgemm tester; conversion routine from and to half; Nvidia support for now.
cayrols
committed
utils: move utils into test/cuda; Add missing NVCC flag; Clean utils header.
cayrols
committed
utils: move utils into test/cuda; Add missing NVCC flag; Clean utils header.
cayrols
committed
half: Add hipify from the cuda utils; update the compilation chain to detect hip.
cayrols
committed
Replace the definition of half from __half to _float16; rename blas::half into blas::float16
cayrols
committed
Rename test/utils.cuh into test/utils.hh
cayrols
committed
Add hip files that got generated from cuda files.
cayrols
committed
hgemm: Add hip support for hgemm
cayrols
committed
hgemm: fix compilation after cleaning.
cayrols
committed
test: gemm change the bound by removing sqrt.
cayrols
committed
test: fix scalar type used to get the flop count in half gemm.
cayrols
committed
hgemm: Use a class float16 instead of an alias.
cayrols
committed
config: fix gpu_backend name issue.
cayrols
committed
hgemm: fix compilation issue.
cayrols
committed
TMP: Add explicit compilation flag for reproducer purpose.
cayrols
committed
hgemm: Search _Float16 support from compiler; If so, the macro BLAS_USE_ISO_FLOAT16 is defined.
cayrols
committed
hgemm: add CPU support through MKL.
cayrols
committed
hgemm: add CPU test with cblas wrapper; add cast_onto_device util that cast from and to float16.
cayrols
committed
float16: update configure search and macro definition.
cayrols
committed
hgemm: Fake casting in cublas_wrapper through pointer casting.
cayrols
committed
test: in hgemm, add cpu casting support, remove cast_onto_device routine; Fix compilation issue.
cayrols
committed
float16: add casting routines from/to fp16 to/from fp32.
cayrols
committed
float16: add missing config file.
cayrols
committed
hgemm: enable CPU hgemm only when MKL is provided.
cayrols
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

add half gemm support#68

add half gemm support#68
cayrols wants to merge 25 commits into
icl-utk-edu:masterfrom
cayrols:seb/gemm.add_hgemm_support

Commits on Dec 14, 2023

hgemm: add half definition and interfaces

hgemm: add hgemm tester; conversion routine from and to half; Nvidia support for now.

utils: move utils into test/cuda; Add missing NVCC flag; Clean utils header.

utils: move utils into test/cuda; Add missing NVCC flag; Clean utils header.

half: Add hipify from the cuda utils; update the compilation chain to detect hip.

Replace the definition of half from __half to _float16; rename blas::half into blas::float16

Rename test/utils.cuh into test/utils.hh

Add hip files that got generated from cuda files.

hgemm: Add hip support for hgemm

hgemm: fix compilation after cleaning.

test: gemm change the bound by removing sqrt.

test: fix scalar type used to get the flop count in half gemm.

hgemm: Use a class float16 instead of an alias.

config: fix gpu_backend name issue.

hgemm: fix compilation issue.

TMP: Add explicit compilation flag for reproducer purpose.

hgemm: Search _Float16 support from compiler; If so, the macro BLAS_USE_ISO_FLOAT16 is defined.

hgemm: add CPU support through MKL.

hgemm: add CPU test with cblas wrapper; add cast_onto_device util that cast from and to float16.

float16: update configure search and macro definition.

hgemm: Fake casting in cublas_wrapper through pointer casting.

test: in hgemm, add cpu casting support, remove cast_onto_device routine; Fix compilation issue.

float16: add casting routines from/to fp16 to/from fp32.

float16: add missing config file.

hgemm: enable CPU hgemm only when MKL is provided.