- /*
- 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]);
- }
- }
- }
-
-# 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
-
- return output;
- }
-
- private:
- using BiasType = OutputType;
- using WeightType = std::int8_t;
-
- alignas(CacheLineSize) BiasType biases[OutputDimensions];
- alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
- };
-
- // A specialization for small 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
- // 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 (for small inputs) should not have been chosen.");
-
- // 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;
- }
-
- static IndexType get_weight_index_scrambled(IndexType i)