#ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
#define NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
+#include <cstdint>
#include <iostream>
-#include <algorithm>
-#include <type_traits>
+
#include "../nnue_common.h"
#include "simd.h"
/*
This file contains the definition for a fully connected layer (aka affine transform).
- Two approaches are employed, depending on the sizes of the transform.
-
- Approach 1:
- - used when the PaddedInputDimensions >= 128
- - uses AVX512 if possible
- - processes inputs in batches of 2*InputSimdWidth
- - so in batches of 128 for AVX512
- - the weight blocks of size InputSimdWidth are transposed such that
- access is sequential
- - N columns of the weight matrix are processed a time, where N
- depends on the architecture (the amount of registers)
- - accumulate + hadd is used
-
- Approach 2:
- - used when the PaddedInputDimensions < 128
- - does not use AVX512
+
- expected use-case is for when PaddedInputDimensions == 32 and InputDimensions <= 32.
- that's why AVX512 is hard to implement
- expected use-case is small layers
- - not optimized as well as the approach 1
- inputs are processed in chunks of 4, weights are respectively transposed
- accumulation happens directly to int32s
*/
namespace Stockfish::Eval::NNUE::Layers {
// Fallback implementation for older/other architectures.
-// Identical for both approaches. Requires the input to be padded to at least 16 values.
+// Requires the input to be padded to at least 16 values.
#if !defined(USE_SSSE3)
template <IndexType InputDimensions, IndexType PaddedInputDimensions, IndexType OutputDimensions>
static void affine_transform_non_ssse3(std::int32_t* output, const std::int8_t* weights, const std::int32_t* biases, const std::uint8_t* input)
{
+# if defined(USE_SSE2) || defined(USE_MMX) || defined(USE_NEON_DOTPROD) || defined(USE_NEON)
# if defined(USE_SSE2)
// At least a multiple of 16, with SSE2.
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
}
output[i] = sum[0] + sum[1] + sum[2] + sum[3];
-# else
- std::int32_t sum = biases[i];
- for (IndexType j = 0; j < InputDimensions; ++j) {
- sum += weights[offset + j] * input[j];
- }
- output[i] = sum;
# endif
}
# if defined(USE_MMX)
_mm_empty();
# endif
- }
-#endif
-
- 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) >= LargeInputSize)>> {
- public:
- // Input/output type
- using InputType = std::uint8_t;
- using OutputType = std::int32_t;
-
- // Number of input/output dimensions
- static constexpr IndexType InputDimensions = InDims;
- static constexpr IndexType OutputDimensions = OutDims;
-
- static constexpr IndexType PaddedInputDimensions =
- ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
- static constexpr IndexType PaddedOutputDimensions =
- ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
-
- using OutputBuffer = OutputType[PaddedOutputDimensions];
-
- static_assert(PaddedInputDimensions >= LargeInputSize, "Something went wrong. This specialization should not have been chosen.");
-
-#if defined (USE_AVX512)
- static constexpr IndexType InputSimdWidth = 64;
- static constexpr IndexType MaxNumOutputRegs = 16;
-#elif defined (USE_AVX2)
- static constexpr IndexType InputSimdWidth = 32;
- static constexpr IndexType MaxNumOutputRegs = 8;
-#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;
-#else
- // The fallback implementation will not have permuted weights.
- // We define these to avoid a lot of ifdefs later.
- static constexpr IndexType InputSimdWidth = 1;
- static constexpr IndexType MaxNumOutputRegs = 1;
-#endif
-
- // A big block is a region in the weight matrix of the size [PaddedInputDimensions, NumOutputRegs].
- // A small block is a region of size [InputSimdWidth, 1]
-
- static constexpr IndexType NumOutputRegs = std::min(MaxNumOutputRegs, OutputDimensions);
- static constexpr IndexType SmallBlockSize = InputSimdWidth;
- static constexpr IndexType BigBlockSize = NumOutputRegs * PaddedInputDimensions;
- static constexpr IndexType NumSmallBlocksInBigBlock = BigBlockSize / SmallBlockSize;
- static constexpr IndexType NumSmallBlocksPerOutput = PaddedInputDimensions / SmallBlockSize;
- static constexpr IndexType NumBigBlocks = OutputDimensions / NumOutputRegs;
-
- static_assert(OutputDimensions % NumOutputRegs == 0);
-
- // Hash value embedded in the evaluation file
- static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
- std::uint32_t hashValue = 0xCC03DAE4u;
- hashValue += OutputDimensions;
- hashValue ^= prevHash >> 1;
- hashValue ^= prevHash << 31;
- return hashValue;
- }
-
- /*
- Transposes the small blocks within a block.
- Effectively means that weights can be traversed sequentially during inference.
- */
- static IndexType get_weight_index(IndexType i)
- {
- const IndexType smallBlock = (i / SmallBlockSize) % NumSmallBlocksInBigBlock;
- const IndexType smallBlockCol = smallBlock / NumSmallBlocksPerOutput;
- const IndexType smallBlockRow = smallBlock % NumSmallBlocksPerOutput;
- const IndexType bigBlock = i / BigBlockSize;
- const IndexType rest = i % SmallBlockSize;
-
- const IndexType idx =
- bigBlock * BigBlockSize
- + smallBlockRow * SmallBlockSize * NumOutputRegs
- + smallBlockCol * SmallBlockSize
- + rest;
-
- return idx;
- }
-
- // Read network parameters
- bool read_parameters(std::istream& stream) {
- read_little_endian<BiasType>(stream, biases, OutputDimensions);
-
- for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
- weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
-
- return !stream.fail();
- }
-
- // Write network parameters
- bool write_parameters(std::ostream& stream) const {
- write_little_endian<BiasType>(stream, biases, OutputDimensions);
-
- for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
- write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
-
- return !stream.fail();
- }
-
- // Forward propagation
- const OutputType* propagate(
- const InputType* input, OutputType* output) const {
-
-#if defined (USE_AVX512)
- using acc_vec_t = __m512i;
- using bias_vec_t = __m128i;
- using weight_vec_t = __m512i;
- using in_vec_t = __m512i;
- #define vec_zero _mm512_setzero_si512()
- #define vec_add_dpbusd_32x2 Simd::m512_add_dpbusd_epi32x2
- #define vec_hadd Simd::m512_hadd
- #define vec_haddx4 Simd::m512_haddx4
-#elif defined (USE_AVX2)
- using acc_vec_t = __m256i;
- using bias_vec_t = __m128i;
- using weight_vec_t = __m256i;
- using in_vec_t = __m256i;
- #define vec_zero _mm256_setzero_si256()
- #define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
- #define vec_hadd Simd::m256_hadd
- #define vec_haddx4 Simd::m256_haddx4
-#elif defined (USE_SSSE3)
- using acc_vec_t = __m128i;
- using bias_vec_t = __m128i;
- using weight_vec_t = __m128i;
- using in_vec_t = __m128i;
- #define vec_zero _mm_setzero_si128()
- #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;
- using weight_vec_t = int8x8_t;
- using in_vec_t = int8x8_t;
- #define vec_zero {0}
- #define vec_add_dpbusd_32x2 Simd::neon_m128_add_dpbusd_epi32x2
- #define vec_hadd Simd::neon_m128_hadd
- #define vec_haddx4 Simd::neon_m128_haddx4
-#endif
-
-#if defined (USE_SSSE3) || defined (USE_NEON)
- const in_vec_t* invec = reinterpret_cast<const in_vec_t*>(input);
-
- // Perform accumulation to registers for each big block
- for (IndexType bigBlock = 0; bigBlock < NumBigBlocks; ++bigBlock)
- {
- acc_vec_t acc[NumOutputRegs] = { vec_zero };
-
- // Each big block has NumOutputRegs small blocks in each "row", one per register.
- // We process two small blocks at a time to save on one addition without VNNI.
- for (IndexType smallBlock = 0; smallBlock < NumSmallBlocksPerOutput; smallBlock += 2)
- {
- const weight_vec_t* weightvec =
- reinterpret_cast<const weight_vec_t*>(
- weights
- + bigBlock * BigBlockSize
- + smallBlock * SmallBlockSize * NumOutputRegs);
- const in_vec_t in0 = invec[smallBlock + 0];
- const in_vec_t in1 = invec[smallBlock + 1];
-
- for (IndexType k = 0; k < NumOutputRegs; ++k)
- vec_add_dpbusd_32x2(acc[k], in0, weightvec[k], in1, weightvec[k + NumOutputRegs]);
- }
-
- // Horizontally add all accumulators.
- if constexpr (NumOutputRegs % 4 == 0)
- {
- bias_vec_t* outputvec = reinterpret_cast<bias_vec_t*>(output);
- const bias_vec_t* biasvec = reinterpret_cast<const bias_vec_t*>(biases);
-
- for (IndexType k = 0; k < NumOutputRegs; k += 4)
- {
- const IndexType idx = (bigBlock * NumOutputRegs + k) / 4;
- outputvec[idx] = vec_haddx4(acc[k+0], acc[k+1], acc[k+2], acc[k+3], biasvec[idx]);
- }
- }
- else
- {
- for (IndexType k = 0; k < NumOutputRegs; ++k)
- {
- const IndexType idx = (bigBlock * NumOutputRegs + k);
- output[idx] = vec_hadd(acc[k], biases[idx]);
- }
- }
+# else
+ std::memcpy(output, biases, sizeof(std::int32_t) * OutputDimensions);
+
+ // Traverse weights in transpose order to take advantage of input sparsity
+ for (IndexType i = 0; i < InputDimensions; ++i)
+ if (input[i]) {
+ const std::int8_t* w = &weights[i];
+ const int in = input[i];
+ for (IndexType j = 0; j < OutputDimensions; ++j)
+ output[j] += w[j * PaddedInputDimensions] * in;
}
-
-# undef vec_zero
-# undef vec_add_dpbusd_32x2
-# undef vec_hadd
-# undef vec_haddx4
-#else
- // Use old implementation for the other architectures.
- affine_transform_non_ssse3<
- InputDimensions,
- PaddedInputDimensions,
- OutputDimensions>(output, weights, biases, input);
-
+# endif
+ }
#endif
- return output;
- }
-
- private:
- using BiasType = OutputType;
- using WeightType = std::int8_t;
-
- alignas(CacheLineSize) BiasType biases[OutputDimensions];
- alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
- };
-
template <IndexType InDims, IndexType OutDims>
- class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) < LargeInputSize)>> {
+ class AffineTransform {
public:
- // Input/output type
// Input/output type
using InputType = std::uint8_t;
using OutputType = std::int32_t;
using OutputBuffer = OutputType[PaddedOutputDimensions];
- static_assert(PaddedInputDimensions < LargeInputSize, "Something went wrong. This specialization should not have been chosen.");
-
-#if defined (USE_SSSE3)
- static constexpr IndexType OutputSimdWidth = SimdWidth / 4;
- static constexpr IndexType InputSimdWidth = SimdWidth;
-#endif
-
// Hash value embedded in the evaluation file
static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
std::uint32_t hashValue = 0xCC03DAE4u;
return hashValue;
}
- static IndexType get_weight_index_scrambled(IndexType i)
+ static constexpr IndexType get_weight_index_scrambled(IndexType i)
{
return
(i / 4) % (PaddedInputDimensions / 4) * OutputDimensions * 4 +
i % 4;
}
- static IndexType get_weight_index(IndexType i)
+ static constexpr IndexType get_weight_index(IndexType i)
{
#if defined (USE_SSSE3)
return get_weight_index_scrambled(i);
return !stream.fail();
}
// Forward propagation
- const OutputType* propagate(
+ void propagate(
const InputType* input, OutputType* output) const {
-#if defined (USE_AVX2)
+#if defined (USE_SSSE3)
+
+ if constexpr (OutputDimensions > 1)
+ {
+
+#if defined (USE_AVX512)
+ using vec_t = __m512i;
+ #define vec_setzero _mm512_setzero_si512
+ #define vec_set_32 _mm512_set1_epi32
+ #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
+ #define vec_add_dpbusd_32x2 Simd::m512_add_dpbusd_epi32x2
+ #define vec_hadd Simd::m512_hadd
+#elif defined (USE_AVX2)
using vec_t = __m256i;
#define vec_setzero _mm256_setzero_si256
#define vec_set_32 _mm256_set1_epi32
#define vec_hadd Simd::m128_hadd
#endif
-#if defined (USE_SSSE3)
- const auto inputVector = reinterpret_cast<const vec_t*>(input);
+ static constexpr IndexType OutputSimdWidth = sizeof(vec_t) / sizeof(OutputType);
- static_assert(OutputDimensions % OutputSimdWidth == 0 || OutputDimensions == 1);
+ static_assert(OutputDimensions % OutputSimdWidth == 0);
- if constexpr (OutputDimensions % OutputSimdWidth == 0)
- {
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 4;
constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
vec_t* outptr = reinterpret_cast<vec_t*>(output);
for (IndexType k = 0; k < NumRegs; ++k)
outptr[k] = acc[k];
+
+# undef vec_setzero
+# undef vec_set_32
+# undef vec_add_dpbusd_32
+# undef vec_add_dpbusd_32x2
+# undef vec_hadd
+
}
else if constexpr (OutputDimensions == 1)
{
- constexpr IndexType NumChunks = PaddedInputDimensions / SimdWidth;
+
+// We cannot use AVX512 for the last layer because there's only 32 inputs and the buffer is not padded to 64 elements.
+#if defined (USE_AVX2)
+ using vec_t = __m256i;
+ #define vec_setzero _mm256_setzero_si256
+ #define vec_set_32 _mm256_set1_epi32
+ #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
+ #define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
+ #define vec_hadd Simd::m256_hadd
+#elif defined (USE_SSSE3)
+ using vec_t = __m128i;
+ #define vec_setzero _mm_setzero_si128
+ #define vec_set_32 _mm_set1_epi32
+ #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
+ #define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
+ #define vec_hadd Simd::m128_hadd
+#endif
+
+ const auto inputVector = reinterpret_cast<const vec_t*>(input);
+
+ static constexpr IndexType InputSimdWidth = sizeof(vec_t) / sizeof(InputType);
+
+ static_assert(PaddedInputDimensions % InputSimdWidth == 0);
+
+ constexpr IndexType NumChunks = PaddedInputDimensions / InputSimdWidth;
vec_t sum0 = vec_setzero();
const auto row0 = reinterpret_cast<const vec_t*>(&weights[0]);
- for (int j = 0; j < (int)NumChunks; ++j)
+ for (int j = 0; j < int(NumChunks); ++j)
{
const vec_t in = inputVector[j];
vec_add_dpbusd_32(sum0, in, row0[j]);
}
output[0] = vec_hadd(sum0, biases[0]);
- }
# undef vec_setzero
# undef vec_set_32
# undef vec_add_dpbusd_32
# undef vec_add_dpbusd_32x2
# undef vec_hadd
+
+ }
#else
// Use old implementation for the other architectures.
affine_transform_non_ssse3<
PaddedInputDimensions,
OutputDimensions>(output, weights, biases, input);
#endif
-
- return output;
}
private: