opelly/Stockfish
1
0
Fork 0
mirror of https://github.com/opelly27/Stockfish.git synced 2026-05-20 16:47:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add support for ARM dot product instructions

Browse Source 
The sdot instruction computes (and accumulates) a signed dot product,
which is quite handy for Stockfish's NNUE code. The instruction is
optional for Armv8.2 and Armv8.3, and mandatory for Armv8.4 and above.

The commit adds a new 'arm-dotprod' architecture with enabled dot
product support. It also enables dot product support for the existing
'apple-silicon' architecture, which is at least Armv8.5.

The following local speed test was performed on an Apple M1 with
ARCH=apple-silicon. I had to remove CPU pinning from the benchmark
script. However, the results were still consistent: Checking both
binaries against themselves reported a speedup of +0.0000 and +0.0005,
respectively.

```
Result of 100 runs
==================
base (...ish.037ef3e1) =    1917997  +/- 7152
test (...fish.dotprod) =    2159682  +/- 9066
diff                   =    +241684  +/- 2923

speedup        = +0.1260
P(speedup > 0) =  1.0000

CPU: 10 x arm
Hyperthreading: off
```

Fixes #4193

closes https://github.com/official-stockfish/Stockfish/pull/4400

No functional change
This commit is contained in:
Sebastian Buchwald
2023-02-21 22:18:17 +01:00
committed by Joost VandeVondele
parent 037ef3e18d
commit b4ad3a3c4b
3 changed files with 78 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/layers/affine_transform.h
+24
View File
@@ -72,6 +72,10 @@ namespace Stockfish::Eval::NNUE::Layers {
const __m64 Zeros = _mm_setzero_si64();
const auto inputVector = reinterpret_cast<const __m64*>(input);
# elif defined(USE_NEON_DOTPROD)
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
const auto inputVector = reinterpret_cast<const int8x16_t*>(input);
# elif defined(USE_NEON)
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
@@ -123,6 +127,14 @@ namespace Stockfish::Eval::NNUE::Layers {
sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
output[i] = _mm_cvtsi64_si32(sum);
# elif defined(USE_NEON_DOTPROD)
int32x4_t sum = {biases[i]};
const auto row = reinterpret_cast<const int8x16_t*>(&weights[offset]);
for (IndexType j = 0; j < NumChunks; ++j) {
sum = vdotq_s32(sum, inputVector[j], row[j]);
}
output[i] = vaddvq_s32(sum);
# elif defined(USE_NEON)
int32x4_t sum = {biases[i]};
const auto row = reinterpret_cast<const int8x8_t*>(&weights[offset]);
@@ -187,6 +199,9 @@ namespace Stockfish::Eval::NNUE::Layers {
#elif defined (USE_SSSE3)
static constexpr IndexType InputSimdWidth = 16;
static constexpr IndexType MaxNumOutputRegs = 8;
#elif defined (USE_NEON_DOTPROD)
static constexpr IndexType InputSimdWidth = 16;
static constexpr IndexType MaxNumOutputRegs = 8;
#elif defined (USE_NEON)
static constexpr IndexType InputSimdWidth = 8;
static constexpr IndexType MaxNumOutputRegs = 8;
@@ -292,6 +307,15 @@ namespace Stockfish::Eval::NNUE::Layers {
#define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
#define vec_hadd Simd::m128_hadd
#define vec_haddx4 Simd::m128_haddx4
#elif defined (USE_NEON_DOTPROD)
using acc_vec_t = int32x4_t;
using bias_vec_t = int32x4_t;
using weight_vec_t = int8x16_t;
using in_vec_t = int8x16_t;
#define vec_zero {0}
#define vec_add_dpbusd_32x2 Simd::dotprod_m128_add_dpbusd_epi32x2
#define vec_hadd Simd::neon_m128_hadd
#define vec_haddx4 Simd::neon_m128_haddx4
#elif defined (USE_NEON)
using acc_vec_t = int32x4_t;
using bias_vec_t = int32x4_t;
src/nnue/layers/simd.h
+13
View File
@@ -346,6 +346,19 @@ namespace Stockfish::Simd {
#endif
#if defined (USE_NEON_DOTPROD)
[[maybe_unused]] static void dotprod_m128_add_dpbusd_epi32x2(
int32x4_t& acc,
int8x16_t a0, int8x16_t b0,
int8x16_t a1, int8x16_t b1) {
acc = vdotq_s32(acc, a0, b0);
acc = vdotq_s32(acc, a1, b1);
}
#endif
#if defined (USE_NEON)
[[maybe_unused]] static int neon_m128_reduce_add_epi32(int32x4_t s) {