#AVX512IFMA - Twitter Hashtag

over 3 years ago

@owainkenway 1/2 Going through a really cool paper from @enp1s0 on using int8 MatMull units to do DGEMM. An interesting note is that the technique used beats the HW DGEMM on NV Turing (RTX 2k series) amongst others using only the Integer units. Also thinking about using #AVX512IFMA ...

1

6

0

356

InstLatX64 @InstLatX64

over 4 years ago

This is my #AVX512F libdivide-based printing idea: not so dense and elegant as the #AVX512IFMA, but it works even on #KnightsLanding, #SkyLakeX, #SkyLakeSP, #CascadeLake, #CooperLake Free source: https://t.co/3WOTSwyyYV

InstLatX64's tweet photo. This is my #AVX512F libdivide-based printing idea: not so dense and elegant as the #AVX512IFMA, but it works even on #KnightsLanding, #SkyLakeX, #SkyLakeSP, #CascadeLake, #CooperLake
Free source:
https://t.co/3WOTSwyyYV https://t.co/QaHDfGLbKj

InstLatX64 @InstLatX64

over 4 years ago

This is my #AVX512IFMA printing idea: how can a VPMADD52LUQ + VPMADD52HUQ pair generate 8 decimal ascii digits? Free source: https://t.co/3WOTSwyyYV

InstLatX64's tweet photo. This is my #AVX512IFMA printing idea: how can a VPMADD52LUQ + VPMADD52HUQ pair generate 8 decimal ascii digits?
Free source:
https://t.co/3WOTSwyyYV https://t.co/yRvFLgmAky

0

9

0

1

0

1

12

1

0

InstLatX64 @InstLatX64

over 4 years ago

This is my #AVX512IFMA printing idea: how can a VPMADD52LUQ + VPMADD52HUQ pair generate 8 decimal ascii digits? Free source: https://t.co/3WOTSwyyYV

0

9

0

1

0

InstLatX64 @InstLatX64

over 4 years ago

I have implemented 16 versions of the decimal printing function to serve as reference work: - 32 / 64-bit - signed / unsigned - #AVX512IFMA / #AVX512F - intrinsic / assembly all with size- and negative-sign handling.

InstLatX64's tweet photo. I have implemented 16 versions of the decimal printing function to serve as reference work:
- 32 / 64-bit
- signed / unsigned
- #AVX512IFMA / #AVX512F
- intrinsic / assembly
all with size- and negative-sign handling. https://t.co/BWF5DLJcMr

InstLatX64 @InstLatX64

over 4 years ago

I have also implemented a libdivide-based #AVX512F version for the old #KnightsLanding, Cascade-, Cooper- and #SkylakeX, without #AVX512IFMA. It is not such a breakthrough as the IFMA version, but is still faster on SKX than everything else except the 40000B LUT.

InstLatX64's tweet photo. I have also implemented a libdivide-based #AVX512F version for the old #KnightsLanding, Cascade-, Cooper- and #SkylakeX, without #AVX512IFMA. It is not such a breakthrough as the IFMA version, but is still faster on SKX than everything else except the 40000B LUT. https://t.co/dTLjCm80Xf

0

5

0

11

2

0

InstLatX64 @InstLatX64

over 4 years ago

I have also implemented a libdivide-based #AVX512F version for the old #KnightsLanding, Cascade-, Cooper- and #SkylakeX, without #AVX512IFMA. It is not such a breakthrough as the IFMA version, but is still faster on SKX than everything else except the 40000B LUT.

InstLatX64 @InstLatX64

over 4 years ago

New #AVX512 levels can be useful for everyday tasks too, such as decimal printing. #AVX512IFMA version is not only faster, but it uses just a few cachelines instead of the 40000B LUT. #AMD # Zen4, #Intel #Cannon-, #Ice-, #Tiger-, #Rocket- or proper #AlderLake is required.

InstLatX64's tweet photo. New #AVX512 levels can be useful for everyday tasks too, such as decimal printing.
#AVX512IFMA version is not only faster, but it uses just a few cachelines instead of the 40000B LUT.
#AMD # Zen4, #Intel #Cannon-, #Ice-, #Tiger-, #Rocket- or proper #AlderLake is required. https://t.co/PtaekNTwLc

1

19

3

0

5

0

InstLatX64 @InstLatX64

over 4 years ago

New #AVX512 levels can be useful for everyday tasks too, such as decimal printing. #AVX512IFMA version is not only faster, but it uses just a few cachelines instead of the 40000B LUT. #AMD # Zen4, #Intel #Cannon-, #Ice-, #Tiger-, #Rocket- or proper #AlderLake is required.

Daniel Lemire

@lemire

almost 5 years ago

Converting integers to fix-digit representations quickly https://t.co/v112isaJLz With code from: @pkhuong @pshufb

2

35

12

7

0

1

19

3

0

InstLatX64 @InstLatX64

almost 7 years ago

This wikichip article reveals that Centaur CNS core supports #AVX512VBMI and #AVX512IFMA with 256b-wide ports. The ISA is on par with #Intel CannonLake https://t.co/e7VVZ1Bj3K

0

15

2

0

Top Tweets for #AVX512IFMA

Last Seen Hashtags on Sotwe

Trends for you

Most Popular Users