- constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
- const __m128i Zeros = _mm_setzero_si128();
- const auto inputVector = reinterpret_cast<const __m128i*>(input);
-
-# elif defined(USE_MMX)
- constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 8;
- 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);
-# endif
-
- for (IndexType i = 0; i < OutputDimensions; ++i) {
- const IndexType offset = i * PaddedInputDimensions;
-
-# if defined(USE_SSE2)
- __m128i sumLo = _mm_cvtsi32_si128(biases[i]);
- __m128i sumHi = Zeros;
- const auto row = reinterpret_cast<const __m128i*>(&weights[offset]);
- for (IndexType j = 0; j < NumChunks; ++j) {
- __m128i row_j = _mm_load_si128(&row[j]);
- __m128i input_j = _mm_load_si128(&inputVector[j]);
- __m128i extendedRowLo = _mm_srai_epi16(_mm_unpacklo_epi8(row_j, row_j), 8);
- __m128i extendedRowHi = _mm_srai_epi16(_mm_unpackhi_epi8(row_j, row_j), 8);
- __m128i extendedInputLo = _mm_unpacklo_epi8(input_j, Zeros);
- __m128i extendedInputHi = _mm_unpackhi_epi8(input_j, Zeros);
- __m128i productLo = _mm_madd_epi16(extendedRowLo, extendedInputLo);
- __m128i productHi = _mm_madd_epi16(extendedRowHi, extendedInputHi);
- sumLo = _mm_add_epi32(sumLo, productLo);
- sumHi = _mm_add_epi32(sumHi, productHi);
- }
- __m128i sum = _mm_add_epi32(sumLo, sumHi);
- __m128i sumHigh_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
- sum = _mm_add_epi32(sum, sumHigh_64);
- __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
- sum = _mm_add_epi32(sum, sum_second_32);
- output[i] = _mm_cvtsi128_si32(sum);
-
-# elif defined(USE_MMX)
- __m64 sumLo = _mm_cvtsi32_si64(biases[i]);
- __m64 sumHi = Zeros;
- const auto row = reinterpret_cast<const __m64*>(&weights[offset]);
- for (IndexType j = 0; j < NumChunks; ++j) {
- __m64 row_j = row[j];
- __m64 input_j = inputVector[j];
- __m64 extendedRowLo = _mm_srai_pi16(_mm_unpacklo_pi8(row_j, row_j), 8);
- __m64 extendedRowHi = _mm_srai_pi16(_mm_unpackhi_pi8(row_j, row_j), 8);
- __m64 extendedInputLo = _mm_unpacklo_pi8(input_j, Zeros);
- __m64 extendedInputHi = _mm_unpackhi_pi8(input_j, Zeros);
- __m64 productLo = _mm_madd_pi16(extendedRowLo, extendedInputLo);
- __m64 productHi = _mm_madd_pi16(extendedRowHi, extendedInputHi);
- sumLo = _mm_add_pi32(sumLo, productLo);
- sumHi = _mm_add_pi32(sumHi, productHi);
- }
- __m64 sum = _mm_add_pi32(sumLo, sumHi);
- 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]);
- for (IndexType j = 0; j < NumChunks; ++j) {
- int16x8_t product = vmull_s8(inputVector[j * 2], row[j * 2]);
- product = vmlal_s8(product, inputVector[j * 2 + 1], row[j * 2 + 1]);
- sum = vpadalq_s16(sum, product);
- }
- 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;