From WikiChip
AVX-512 Vector Neural Network Instructions (VNNI) - x86
< x86
Revision as of 18:00, 13 May 2019 by David (talk | contribs) (Microarchitecture support)

AVX-512 Vector Neural Network Instructions (AVX512 VNNI) is an x86 extension, part of the AVX-512, designed to accelerate convolutional neural network-based algorithms.

Overview

The AVX512 VNNI x86 extension extends AVX-512 Foundation by introducing four new instructions for accelerating inner convolutional neural network loops.

  • VPDPBUSD - Multiplies the individual bytes (8-bit) of the first source operand by the corresponding bytes (8-bit) of the second source operand, producing intermediate word (16-bit) results which are summed and accumulated in the double word (32-bit) of the destination operand.
    • VPDPBUSDS - Same as above except on intermediate sum overflow which saturates to 0x7FFF_FFFF/0x8000_0000 for positive/negative numbers.
  • VPDPWSSD - Multiplies the individual words (16-bit) of the first source operand by the corresponding word (16-bit) of the second source operand, producing intermediate word results which are summed and accumulated in the double word (32-bit) of the destination operand.
    • VPDPWSSDS - Same as above except on intermediate sum overflow which saturates to 0x7FFF_FFFF/0x8000_0000 for positive/negative numbers.

Motivation

The major motivation behind the AVX512 VNNI extension is the observation that many tight convolutional neural network loops require the repeated multiplication of two 16-bit values or two 8-bit values and accumulate the result to a 32-bit accumulator. Using the foundation AVX-512, for 16-bit, this is possible using two instructions - VPMADDWD which is used to multiply two 16-bit pairs and add them together followed a VPADDD which adds the accumulate value.

vnni-vpdpwssd.svg

Likewise, for 8-bit values, three instructions are needed - VPMADDUBSW which is used to multiply two 8-bit pairs and add them together, followed by a VPMADDWD with the value 1 in order to simply up-convert the 16-bit values to 32-bit values, followed by the VPADDD instruction which adds the result to an accumulator.

vnni-vpdpbusd.svg

To address those two common operations, two new instructions were added (as well as two saturated versions): VPDPBUSD fuses VPMADDUBSW, VPMADDWD, and VPADDD and VPDPWSSD fuses VPMADDWD and VPADDD.

vnni-vpdpbusd-i.svg vnni-vpdpwssd-i.svg

Detection

CPUID Instruction Set
Input Output
EAX=07H, ECX=0 ECX[bit 11] AVX512VNNI

Microarchitecture support

Instructions Introduction
AVX512VNNI Cascade Lake (server)
Ice Lake (client)

Intrinsic functions

# vpdpbusd
__m512i _mm512_dpbusd_epi32 (__m512i src, __m512i a, __m512i b)
__m512i _mm512_mask_dpbusd_epi32 (__m512i src, __mmask16 k, __m512i a, __m512i b)
__m512i _mm512_maskz_dpbusd_epi32 (__mmask16 k, __m512i src, __m512i a, __m512i b)
# vpdpbusds
__m512i _mm512_dpbusds_epi32 (__m512i src, __m512i a, __m512i b)
__m512i _mm512_mask_dpbusds_epi32 (__m512i src, __mmask16 k, __m512i a, __m512i b)
__m512i _mm512_maskz_dpbusds_epi32 (__mmask16 k, __m512i src, __m512i a, __m512i b)
# vpdpwssd
__m512i _mm512_dpwssd_epi32 (__m512i src, __m512i a, __m512i b)
__m512i _mm512_mask_dpwssd_epi32 (__m512i src, __mmask16 k, __m512i a, __m512i b)
__m512i _mm512_maskz_dpwssd_epi32 (__mmask16 k, __m512i src, __m512i a, __m512i b)
# vpdpwssds
__m512i _mm512_dpwssds_epi32 (__m512i src, __m512i a, __m512i b)
__m512i _mm512_mask_dpwssds_epi32 (__m512i src, __mmask16 k, __m512i a, __m512i b)
__m512i _mm512_maskz_dpwssds_epi32 (__mmask16 k, __m512i src, __m512i a, __m512i b)

Bibliography