Add support for SSSE3-only compiles

For Core 2 Duo.

To compile:
make ARCH=x86-64 ssse3=yes nnue

No observable difference in speed to SSE4.1 on my machine.

src/Makefile

@@ -387,12 +387,20 @@ ifeq ($(avx2),yes)
 endif
 
 ifeq ($(sse41),yes)
+	ssse3 = yes
 	CXXFLAGS += -DUSE_SSE41
 	ifeq ($(comp),$(filter $(comp),gcc clang mingw msys2))
 		CXXFLAGS += -msse4.1
 	endif
 endif
 
+ifeq ($(ssse3),yes)
+	CXXFLAGS += -DUSE_SSSE3
+	ifeq ($(comp),$(filter $(comp),gcc clang mingw msys2))
+		CXXFLAGS += -mssse3
+	endif
+endif
+
 ifeq ($(arch),x86_64)
 	CXXFLAGS += -DUSE_SSE2
 endif

src/eval/nnue/layers/affine_transform.h

@@ -86,7 +86,7 @@ class AffineTransform {
     constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
     const __m256i kOnes = _mm256_set1_epi16(1);
     const auto input_vector = reinterpret_cast<const __m256i*>(input);
-#elif defined(USE_SSE41)
+#elif defined(USE_SSSE3)
     constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
     const __m128i kOnes = _mm_set1_epi16(1);
     const auto input_vector = reinterpret_cast<const __m128i*>(input);
@@ -118,7 +118,7 @@ class AffineTransform {
       const __m128i lo = _mm256_extracti128_si256(sum, 0);
       const __m128i hi = _mm256_extracti128_si256(sum, 1);
       output[i] = _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi);
-#elif defined(USE_SSE41)
+#elif defined(USE_SSSE3)
       __m128i sum = _mm_cvtsi32_si128(biases_[i]);
       const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
       for (IndexType j = 0; j < kNumChunks; ++j) {

src/eval/nnue/layers/clipped_relu.h

@@ -110,9 +110,12 @@ class ClippedReLU {
           _mm256_packs_epi16(words0, words1), kZero), kOffsets));
     }
     constexpr IndexType kStart = kNumChunks * kSimdWidth;
-#elif defined(USE_SSE41)
+#elif defined(USE_SSSE3)
     constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
     const __m128i kZero = _mm_setzero_si128();
+#ifndef USE_SSE41
+    const __m128i k0x80s = _mm_set1_epi8(-128);
+#endif
     const auto in = reinterpret_cast<const __m128i*>(input);
     const auto out = reinterpret_cast<__m128i*>(output);
     for (IndexType i = 0; i < kNumChunks; ++i) {
@@ -122,8 +125,14 @@ class ClippedReLU {
       const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
           _mm_load_si128(&in[i * 4 + 2]),
           _mm_load_si128(&in[i * 4 + 3])), kWeightScaleBits);
-      _mm_store_si128(&out[i], _mm_max_epi8(
+      const __m128i packedbytes = _mm_packs_epi16(words0, words1);
-          _mm_packs_epi16(words0, words1), kZero));
+      _mm_store_si128(&out[i], 
+#ifdef USE_SSE41
+        _mm_max_epi8(packedbytes, kZero)
+#else
+        _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
+#endif
+      );
     }
     constexpr IndexType kStart = kNumChunks * kSimdWidth;
 #elif defined(IS_ARM)

src/eval/nnue/nnue_common.h

@@ -9,6 +9,8 @@
 #include <immintrin.h>
 #elif defined(USE_SSE41)
 #include <smmintrin.h>
+#elif defined(USE_SSSE3)
+#include <tmmintrin.h>
 #elif defined(USE_SSE2)
 #include <emmintrin.h>
 #endif

src/eval/nnue/nnue_feature_transformer.h

@@ -87,9 +87,12 @@ class FeatureTransformer {
     constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
     constexpr int kControl = 0b11011000;
     const __m256i kZero = _mm256_setzero_si256();
-#elif defined(USE_SSE41)
+#elif defined(USE_SSSE3)
     constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
     const __m128i kZero = _mm_setzero_si128();
+#ifndef USE_SSE41
+    const __m128i k0x80s = _mm_set1_epi8(-128);
+#endif
 #elif defined(IS_ARM)
     constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
     const int8x8_t kZero = {0};
@@ -133,7 +136,7 @@ class FeatureTransformer {
         (&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
             _mm256_packs_epi16(sum0, sum1), kZero), kControl));
       }
-#elif defined(USE_SSE41)
+#elif defined(USE_SSSE3)
       auto out = reinterpret_cast<__m128i*>(&output[offset]);
       for (IndexType j = 0; j < kNumChunks; ++j) {
         __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
@@ -146,8 +149,15 @@ class FeatureTransformer {
           sum1 = _mm_add_epi16(sum1, reinterpret_cast<const __m128i*>(
               accumulation[perspectives[p]][i])[j * 2 + 1]);
         }
-        _mm_store_si128(&out[j], _mm_max_epi8(
+  	const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
-            _mm_packs_epi16(sum0, sum1), kZero));
+ 
+        _mm_store_si128(&out[j],
+#ifdef USE_SSE41
+          _mm_max_epi8(packedbytes, kZero)
+#else
+          _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
+#endif
+        );
       }
 #elif defined(IS_ARM)
       const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);