kate/BadFish
1
0
Fork 0
mirror of https://github.com/sockspls/badfish synced 2025-05-01 01:03:09 +00:00
Code Issues Releases Wiki Activity

Optimize AVX2 path in NNUE evaluation

Browse source 
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
This commit is contained in:
mstembera 2022-09-06 15:02:35 -07:00 • committed by Joost VandeVondele
parent 1591e5ac3b
commit 82bb21dc7a
2 changed files with 11 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

16
src/nnue/layers/affine_transform.h
View file

@ -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;

0
src/simd.h → src/nnue/layers/simd.h
View file