gemm-vega64

Implement asm gemm on vega64 for 4096x4096 fp32 matrix

Results

v10 is the highest performing kernel

Card	Shader Clock	Memory Clock	TFLOPs (fp32)	TFLOPs Peak (fp32)	Efficiency
MI25	1500Mhz	945Mhz	10.6	12.3	86%
RXVega64	1630Mhz	945Mhz	11.2	13.35	84%
VegaFE

As the code is not written in pure isa, compiler tries to move instructions around causing drop in performance.

Name		Name	Last commit message	Last commit date
Latest commit History 82 Commits
docs		docs
opencl		opencl
xpmts		xpmts
.gitignore		.gitignore
README.md		README.md
_config.yml		_config.yml
bload.s		bload.s
datatypes.h		datatypes.h
dims.h		dims.h
dpp_helper.h		dpp_helper.h
fmac.cpp		fmac.cpp
gemm_opt.txt		gemm_opt.txt
gload.s		gload.s
global_ops.h		global_ops.h
h2gemm_v10.cpp		h2gemm_v10.cpp
id.py		id.py
mixgemm_v10.cpp		mixgemm_v10.cpp
outer_product.h		outer_product.h
sgemm.cpp		sgemm.cpp
sgemm_block_v1.cpp		sgemm_block_v1.cpp
sgemm_module.cpp		sgemm_module.cpp
sgemm_v10.cpp		sgemm_v10.cpp
sgemm_v11.cpp		sgemm_v11.cpp
sgemm_v12.cpp		sgemm_v12.cpp
sgemm_v13.cpp		sgemm_v13.cpp
sgemm_v2.cpp		sgemm_v2.cpp
sgemm_v3.cpp		sgemm_v3.cpp
sgemm_v4.cpp		sgemm_v4.cpp
sgemm_v5.cpp		sgemm_v5.cpp
sgemm_v6.cpp		sgemm_v6.cpp
sgemm_v7.cpp		sgemm_v7.cpp
sgemm_v8.cpp		sgemm_v8.cpp
sgemm_v9.cpp		sgemm_v9.cpp
shared_ops.h		shared_ops.h