Support VNNI on 256bit vectors

due to downclocking on current chips (tested up to cascade lake)
supporting avx512 and vnni512, it is better to use avx2 or vnni256
in multithreaded (in particular hyperthreaded) engine use.
In single threaded use, the picture is different.

gcc compilation for vnni256 requires a toolchain for gcc >= 9.

closes https://github.com/official-stockfish/Stockfish/pull/3038

No functional change

.travis.yml

@@ -77,8 +77,10 @@ script:
   # compile only for some more advanced architectures (might not run in travis)
   - make clean && make -j2 ARCH=x86-64-avx2 build
   - make clean && make -j2 ARCH=x86-64-bmi2 build
-  # needs gcc 10 to compile
-  - if [[ "$COMPILER" != "g++-8" ]]; then make clean && make -j2 ARCH=x86-64-avx512 build; fi
+  - make clean && make -j2 ARCH=x86-64-avx512 build
+  - make clean && make -j2 ARCH=x86-64-vnni512 build
+  # requires gcc 9 or higher
+  - if [[ "$COMPILER" != "g++-8" ]]; make clean && make -j2 ARCH=x86-64-vnni256 build; fi
 
   #
   # Check perft and reproducible search

src/Makefile

@@ -75,7 +75,8 @@ endif
 # sse41 = yes/no      --- -msse4.1         --- Use Intel Streaming SIMD Extensions 4.1
 # avx2 = yes/no       --- -mavx2           --- Use Intel Advanced Vector Extensions 2
 # avx512 = yes/no     --- -mavx512bw       --- Use Intel Advanced Vector Extensions 512
-# vnni = yes/no       --- -mavx512vnni     --- Use Intel Vector Neural Network Instructions 512
+# vnni256 = yes/no    --- -mavx512vnni     --- Use Intel Vector Neural Network Instructions 256
+# vnni512 = 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
@@ -102,7 +103,8 @@ ssse3 = no
 sse41 = no
 avx2 = no
 avx512 = no
-vnni = no
+vnni256 = no
+vnni512 = no
 neon = no
 ARCH = x86-64-modern
 STRIP = strip
@@ -192,7 +194,18 @@ ifeq ($(findstring -avx512,$(ARCH)),-avx512)
 	avx512 = yes
 endif
 
-ifeq ($(findstring -vnni,$(ARCH)),-vnni)
+ifeq ($(findstring -vnni256,$(ARCH)),-vnni256)
+	popcnt = yes
+	sse = yes
+	sse2 = yes
+	ssse3 = yes
+	sse41 = yes
+	avx2 = yes
+	pext = yes
+	vnni256 = yes
+endif
+
+ifeq ($(findstring -vnni512,$(ARCH)),-vnni512)
 	popcnt = yes
 	sse = yes
 	sse2 = yes
@@ -201,7 +214,7 @@ ifeq ($(findstring -vnni,$(ARCH)),-vnni)
 	avx2 = yes
 	pext = yes
 	avx512 = yes
-	vnni = yes
+	vnni512 = yes
 endif
 
 ifeq ($(sse),yes)
@@ -500,7 +513,14 @@ ifeq ($(avx512),yes)
 	endif
 endif
 
-ifeq ($(vnni),yes)
+ifeq ($(vnni256),yes)
+	CXXFLAGS += -DUSE_VNNI
+	ifeq ($(comp),$(filter $(comp),gcc clang mingw))
+		CXXFLAGS += -mavx512vnni -mavx512dq -mavx512vl -mprefer-vector-width=256
+	endif
+endif
+
+ifeq ($(vnni512),yes)
 	CXXFLAGS += -DUSE_VNNI
 	ifeq ($(comp),$(filter $(comp),gcc clang mingw))
 		CXXFLAGS += -mavx512vnni -mavx512dq -mavx512vl
@@ -623,7 +643,8 @@ help:
 	@echo ""
 	@echo "Supported archs:"
 	@echo ""
-	@echo "x86-64-vnni             > x86 64-bit with vnni support"
+	@echo "x86-64-vnni512          > x86 64-bit with vnni support 512bit wide"
+	@echo "x86-64-vnni256          > x86 64-bit with vnni support 256bit wide"
 	@echo "x86-64-avx512           > x86 64-bit with avx512 support"
 	@echo "x86-64-bmi2             > x86 64-bit with bmi2 support"
 	@echo "x86-64-avx2             > x86 64-bit with avx2 support"
@@ -767,7 +788,8 @@ config-sanity:
 	@echo "sse41: '$(sse41)'"
 	@echo "avx2: '$(avx2)'"
 	@echo "avx512: '$(avx512)'"
-	@echo "vnni: '$(vnni)'"
+	@echo "vnni256: '$(vnni256)'"
+	@echo "vnni512: '$(vnni512)'"
 	@echo "neon: '$(neon)'"
 	@echo ""
 	@echo "Flags:"
@@ -794,7 +816,8 @@ config-sanity:
 	@test "$(sse41)" = "yes" || test "$(sse41)" = "no"
 	@test "$(avx2)" = "yes" || test "$(avx2)" = "no"
 	@test "$(avx512)" = "yes" || test "$(avx512)" = "no"
-	@test "$(vnni)" = "yes" || test "$(vnni)" = "no"
+	@test "$(vnni256)" = "yes" || test "$(vnni256)" = "no"
+	@test "$(vnni512)" = "yes" || test "$(vnni512)" = "no"
 	@test "$(neon)" = "yes" || test "$(neon)" = "no"
 	@test "$(comp)" = "gcc" || test "$(comp)" = "icc" || test "$(comp)" = "mingw" || test "$(comp)" = "clang" \
 	|| test "$(comp)" = "armv7a-linux-androideabi16-clang"  || test "$(comp)" = "aarch64-linux-android21-clang"

src/nnue/layers/affine_transform.h

@@ -85,8 +85,10 @@ namespace Eval::NNUE::Layers {
 
   #elif defined(USE_AVX2)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
-      const __m256i kOnes = _mm256_set1_epi16(1);
       const auto input_vector = reinterpret_cast<const __m256i*>(input);
+  #if !defined(USE_VNNI)
+      const __m256i kOnes = _mm256_set1_epi16(1);
+  #endif
 
   #elif defined(USE_SSE2)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
@@ -145,9 +147,13 @@ namespace Eval::NNUE::Layers {
         __m256i sum = _mm256_setzero_si256();
         const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
         for (IndexType j = 0; j < kNumChunks; ++j) {
+  #if defined(USE_VNNI)
+          sum = _mm256_dpbusd_epi32(sum, _mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
+  #else
           __m256i product = _mm256_maddubs_epi16(_mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
           product = _mm256_madd_epi16(product, kOnes);
           sum = _mm256_add_epi32(sum, product);
+  #endif
         }
         __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
         sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));