This library intercepts function calls for cuBLAS DGEMM functions and executes ozIMMU instead
git clone https://github.com/enp1s0/ozIMMU --recursive
cd ozIMMU
mkdir build
cd build
cmake ..
make -j4
- Set an environmental variable to hijack the function calls
export LD_PRELOAD=/path/to/ozIMMU/build/libozimmu.so
- Set an environmental variable to choose the compute mode
export OZIMMU_COMPUTE_MODE=fp64_int8_9
The supported compute modes are here.
- Execute the application
Mode | Tensor Core type | Num splits | |
---|---|---|---|
dgemm | -- | -- | Disable hijacking |
sgemm | -- | -- | Use SGEMM internally |
fp64_int8_3 | Int8 TC | 3 | |
fp64_int8_4 | Int8 TC | 4 | |
fp64_int8_5 | Int8 TC | 5 | |
fp64_int8_6 | Int8 TC | 6 | |
fp64_int8_7 | Int8 TC | 7 | |
fp64_int8_8 | Int8 TC | 8 | |
fp64_int8_9 | Int8 TC | 9 | |
fp64_int8_10 | Int8 TC | 10 | |
fp64_int8_11 | Int8 TC | 11 | |
fp64_int8_12 | Int8 TC | 12 | |
fp64_int8_13 | Int8 TC | 13 | |
fp64_int8_14 | Int8 TC | 14 | |
fp64_int8_15 | Int8 TC | 15 | |
fp64_int8_16 | Int8 TC | 16 | |
fp64_int8_17 | Int8 TC | 17 | |
fp64_int8_18 | Int8 TC | 18 | |
fp64_int8_auto | Int8 TC | AUTO | fp64_int8_3..18 / dgemm |
# Show info log
export OZIMMU_INFO=1
# Show error and warning log
export OZIMMU_ERROR=1
# Show CULiP ( https://github.com/enp1s0/CULiP ) log
export OZIMMU_ENABLE_CULIP_PROFILING=1
# Choose malloc mode
export OZIMMU_MALLOC_ASYNC=1
# Set AUTO mode mantissa loss threshold
export OZIMMU_AUTO_AVG_MANTISSA_LOSS_THRESHOLD=1.5
# Set ozIMMU intercept threshold.
export OZIMMU_INTERCEPT_THRESHOLD_M=128
export OZIMMU_INTERCEPT_THRESHOLD_N=128
export OZIMMU_INTERCEPT_THRESHOLD_K=128
# The ozIMMU gemm function is executed if `m`, `n`, and `k` are larger or equal to `OZIMMU_INTERCEPT_THRESHOLD_M`, `N`, and `K`.
# Otherwise, the original cuBLAS function is executed.
@article{ootomo2024dgemm,
author = {Hiroyuki Ootomo and Katsuhisa Ozaki and Rio Yokota},
title = {DGEMM on integer matrix multiplication unit},
journal = {The International Journal of High Performance Computing Applications},
year = {2024},
doi = {10.1177/10943420241239588},
URL = {https://doi.org/10.1177/10943420241239588}
}
MIT