Add support for VNNI

Adds support for Vector Neural Network Instructions (avx512), as available on Intel Cascade Lake

The _mm512_dpbusd_epi32() intrinsic (vpdpbusd instruction) is taylor made for NNUE.

on a cascade lake CPU (AWS C5.24x.large, gcc 10) NNUE eval is at roughly 78% nps of classical
(single core test)

bench 1024 1 24 default depth:
target 	classical 	NNUE 	ratio
vnni 	2207232 	1725987 	78.20
avx512 	2216789 	1671734 	75.41
avx2 	2194006 	1611263 	73.44
modern 	2185001 	1352469 	61.90

closes https://github.com/official-stockfish/Stockfish/pull/2987

No functional change

diff --git a/src/Makefile b/src/Makefile
index e82b066b630c0688923f7de5066286566021356e..0804cdd51a9dcfe7bafc5feb15b11a4ed0455c4f 100644 (file)
--- a/src/Makefile
+++ b/src/Makefile
@@ -73,6 +73,7 @@ endif
 # avx2 = yes/no       --- -mavx2           --- Use Intel Advanced Vector Extensions 2
 # pext = yes/no       --- -DUSE_PEXT       --- Use pext x86_64 asm-instruction
 # avx512 = yes/no     --- -mavx512bw       --- Use Intel Advanced Vector Extensions 512
+# vnni = yes/no       --- -mavx512vnni     --- Use Intel Vector Neural Network Instructions 512
 # neon = yes/no       --- -DUSE_NEON       --- Use ARM SIMD architecture
 #
 # Note that Makefile is space sensitive, so when adding new architectures
@@ -93,6 +94,7 @@ sse41 = no
 avx2 = no
 pext = no
 avx512 = no
+vnni = no
 neon = no
 ARCH = x86-64-modern
 
@@ -190,6 +192,19 @@ ifeq ($(ARCH),x86-64-avx512)
        avx512 = yes
 endif
 
+ifeq ($(ARCH),x86-64-vnni)
+       arch = x86_64
+       prefetch = yes
+       popcnt = yes
+       sse = yes
+       ssse3 = yes
+       sse41 = yes
+       avx2 = yes
+       pext = yes
+       avx512 = yes
+       vnni = yes
+endif
+
 ifeq ($(ARCH),armv7)
        arch = armv7
        prefetch = yes
@@ -420,6 +435,13 @@ ifeq ($(avx512),yes)
        endif
 endif
 
+ifeq ($(vnni),yes)
+       CXXFLAGS += -DUSE_VNNI
+       ifeq ($(comp),$(filter $(comp),gcc clang mingw))
+               CXXFLAGS += -mavx512vnni -mavx512dq -mavx512vl
+       endif
+endif
+
 ifeq ($(sse41),yes)
        CXXFLAGS += -DUSE_SSE41
        ifeq ($(comp),$(filter $(comp),gcc clang mingw))
@@ -522,6 +544,7 @@ help:
        @echo ""
        @echo "Supported archs:"
        @echo ""
+       @echo "x86-64-vnni             > x86 64-bit with vnni support"
        @echo "x86-64-avx512           > x86 64-bit with avx512 support"
        @echo "x86-64-bmi2             > x86 64-bit with bmi2 support"
        @echo "x86-64-avx2             > x86 64-bit with avx2 support"
@@ -640,6 +663,7 @@ config-sanity:
        @echo "avx2: '$(avx2)'"
        @echo "pext: '$(pext)'"
        @echo "avx512: '$(avx512)'"
+       @echo "vnni: '$(vnni)'"
        @echo "neon: '$(neon)'"
        @echo ""
        @echo "Flags:"
@@ -664,6 +688,7 @@ config-sanity:
        @test "$(avx2)" = "yes" || test "$(avx2)" = "no"
        @test "$(pext)" = "yes" || test "$(pext)" = "no"
        @test "$(avx512)" = "yes" || test "$(avx512)" = "no"
+       @test "$(vnni)" = "yes" || test "$(vnni)" = "no"
        @test "$(neon)" = "yes" || test "$(neon)" = "no"
        @test "$(comp)" = "gcc" || test "$(comp)" = "icc" || test "$(comp)" = "mingw" || test "$(comp)" = "clang"
 

diff --git a/src/misc.cpp b/src/misc.cpp
index aeb3c912bf23a286e3c31028e66aaa5ed31938fc..ab52d30bb15bd449d4d6bcdff5681c7ced055efc 100644 (file)
--- a/src/misc.cpp
+++ b/src/misc.cpp
@@ -219,6 +219,9 @@ const std::string compiler_info() {
 
   compiler += "\nCompilation settings include: ";
   compiler += (Is64Bit ? " 64bit" : " 32bit");
+  #if defined(USE_VNNI)
+    compiler += " VNNI";
+  #endif
   #if defined(USE_AVX512)
     compiler += " AVX512";
   #endif

diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index 8d2acd1852eabfa9dce873b803040742a2da75f9..322e32402500029033caa5ddbb7123965adc45d7 100644 (file)
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -79,8 +79,10 @@ namespace Eval::NNUE::Layers {
 
   #if defined(USE_AVX512)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2);
-      const __m512i kOnes = _mm512_set1_epi16(1);
       const auto input_vector = reinterpret_cast<const __m512i*>(input);
+  #if !defined(USE_VNNI)
+      const __m512i kOnes = _mm512_set1_epi16(1);
+  #endif
 
   #elif defined(USE_AVX2)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
@@ -113,9 +115,13 @@ namespace Eval::NNUE::Layers {
         __m512i sum = _mm512_setzero_si512();
         const auto row = reinterpret_cast<const __m512i*>(&weights_[offset]);
         for (IndexType j = 0; j < kNumChunks; ++j) {
+  #if defined(USE_VNNI)
+            sum = _mm512_dpbusd_epi32(sum, _mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
+  #else
             __m512i product = _mm512_maddubs_epi16(_mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
             product = _mm512_madd_epi16(product, kOnes);
             sum = _mm512_add_epi32(sum, product);
+  #endif
         }
 
         // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks.
@@ -125,8 +131,14 @@ namespace Eval::NNUE::Layers {
         {
             const auto iv256  = reinterpret_cast<const __m256i*>(&input_vector[kNumChunks]);
             const auto row256 = reinterpret_cast<const __m256i*>(&row[kNumChunks]);
+  #if defined(USE_VNNI)
+            __m256i product256 = _mm256_dpbusd_epi32(
+                _mm512_castsi512_si256(sum), _mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
+            sum = _mm512_inserti32x8(sum, product256, 0);
+  #else
             __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
             sum = _mm512_add_epi32(sum, _mm512_cvtepi16_epi32(product256));
+  #endif
         }
         output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];