Optimize AVX2 path in NNUE evaluation

always selecting AffineTransform specialization for small inputs.

A related patch was tested as

Initially tested as a simplification
STC https://tests.stockfishchess.org/tests/view/6317c3f437f41b13973d6dff
LLR: 2.95 (-2.94,2.94) <-1.75,0.25>
Total: 58072 W: 15619 L: 15425 D: 27028
Ptnml(0-2): 241, 6191, 15992, 6357, 255

Elo gain speedup test
STC https://tests.stockfishchess.org/tests/view/63181c1b37f41b13973d79dc
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 184496 W: 49922 L: 49401 D: 85173
Ptnml(0-2): 851, 19397, 51208, 19964, 828

and this patch gained in testing

speedup        = +0.0071
P(speedup > 0) =  1.0000
on CPU: 16 x AMD Ryzen 9 3950X

closes https://github.com/official-stockfish/Stockfish/pull/4158

No functional change

diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index 9a992608cc1a06a9808c4c96228a696bde3cc280..461a7b83eca829a9a45c5d215e11dd5991280af6 100644 (file)
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -25,7 +25,7 @@
 #include <algorithm>
 #include <type_traits>
 #include "../nnue_common.h"
-#include "../../simd.h"
+#include "simd.h"
 
 /*
   This file contains the definition for a fully connected layer (aka affine transform).
@@ -151,9 +151,15 @@ namespace Stockfish::Eval::NNUE::Layers {
   template <IndexType InDims, IndexType OutDims, typename Enabled = void>
   class AffineTransform;
 
+#if defined (USE_AVX512)
+  constexpr IndexType LargeInputSize = 2 * 64;
+#else
+  constexpr IndexType LargeInputSize = std::numeric_limits<IndexType>::max();
+#endif
+
   // A specialization for large inputs.
   template <IndexType InDims, IndexType OutDims>
-  class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) >= 2*64)>> {
+  class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) >= LargeInputSize)>> {
    public:
     // Input/output type
     using InputType = std::uint8_t;
@@ -170,7 +176,7 @@ namespace Stockfish::Eval::NNUE::Layers {
 
     using OutputBuffer = OutputType[PaddedOutputDimensions];
 
-    static_assert(PaddedInputDimensions >= 128, "Something went wrong. This specialization should not have been chosen.");
+    static_assert(PaddedInputDimensions >= LargeInputSize, "Something went wrong. This specialization should not have been chosen.");
 
 #if defined (USE_AVX512)
     static constexpr const IndexType InputSimdWidth = 64;
@@ -369,7 +375,7 @@ namespace Stockfish::Eval::NNUE::Layers {
   };
 
   template <IndexType InDims, IndexType OutDims>
-  class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) < 2*64)>> {
+  class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) < LargeInputSize)>> {
    public:
     // Input/output type
     // Input/output type
@@ -387,7 +393,7 @@ namespace Stockfish::Eval::NNUE::Layers {
 
     using OutputBuffer = OutputType[PaddedOutputDimensions];
 
-    static_assert(PaddedInputDimensions < 128, "Something went wrong. This specialization should not have been chosen.");
+    static_assert(PaddedInputDimensions < LargeInputSize, "Something went wrong. This specialization should not have been chosen.");
 
 #if defined (USE_SSSE3)
     static constexpr const IndexType OutputSimdWidth = SimdWidth / 4;

diff --git a/src/simd.h b/src/nnue/layers/simd.h
similarity index 100%
rename from src/simd.h
rename to src/nnue/layers/simd.h