diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp | 44 |
1 files changed, 25 insertions, 19 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp b/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp index 2b3e170a3b..59591935cd 100644 --- a/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp +++ b/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020 Arm Limited. + * Copyright (c) 2020-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -28,7 +28,9 @@ #include "interleave_indirect.hpp" #include "bfloat.hpp" +#if !defined(_WIN64) && !defined(__OpenBSD__) #include <alloca.h> +#endif /* !defined(_WIN64) && !defined(__OpenBSD__) */ #include <algorithm> #include <cstddef> @@ -292,21 +294,21 @@ void Interleave(TOut *out, const TIn *in, size_t in_stride, const unsigned int y /* AArch32 */ #ifdef __arm__ /* FP32 */ -/* NEON implementation (height 6) */ +/* Arm® Neon™ implementation (height 6) */ template void IndirectInterleave<6, 1, VLType::None>(float *, const float * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<6, 1, VLType::None>(float *, const float *, size_t, const convolver<float> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<6, 1, VLType::None>(float *, const float *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); /* FP16 */ #if __ARM_FP16_ARGS -/* NEON implementation using FP32 kernel (height 6) */ +/* Arm® Neon™ implementation using FP32 kernel (height 6) */ template void IndirectInterleave<6, 1, VLType::None>(float *, const __fp16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<6, 1, VLType::None>(float *, const __fp16 *, size_t, const convolver<__fp16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<6, 1, VLType::None>(float *, const __fp16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); #endif /* __ARM_FP16_ARGS */ /* BF16 */ -/* NEON implementation using FP32 kernel */ +/* Arm® Neon™ implementation using FP32 kernel */ template void IndirectInterleave<6, 1, VLType::None>(float *, const bfloat16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<6, 1, VLType::None>(float *, const bfloat16 *, size_t, const convolver<bfloat16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<6, 1, VLType::None>(float *, const bfloat16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); @@ -314,34 +316,33 @@ template void Interleave<6, 1, VLType::None>(float *, const bfloat16 *, size_t, /* AArch64 */ #ifdef __aarch64__ -/* FP64 */ -/* NEON/SVE implementation (height 8) */ -template void IndirectInterleave<8, 1, VLType::None>(double *, const double * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -template void ConvolutionInterleave<8, 1, VLType::None>(double *, const double *, size_t, const convolver<double> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -template void Interleave<8, 1, VLType::None>(double *, const double *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); - /* FP32 */ -/* NEON/SVE implementation (height 8) */ +/* Arm® Neon™/SVE implementation (height 8) */ template void IndirectInterleave<8, 1, VLType::None>(float *, const float * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<8, 1, VLType::None>(float *, const float *, size_t, const convolver<float> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 1, VLType::None>(float *, const float *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); +#if defined(ARM_COMPUTE_ENABLE_SVE) && defined(ARM_COMPUTE_ENABLE_SVEF32MM) /* FMMLA */ template void IndirectInterleave<8, 2, VLType::None>(float *, const float * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<8, 2, VLType::None>(float *, const float *, size_t, const convolver<float> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 2, VLType::None>(float *, const float *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); +#endif // ARM_COMPUTE_ENABLE_SVE && ARM_COMPUTE_ENABLE_SVEF32MM /* FP16 */ +#if defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) template void IndirectInterleave<8, 1, VLType::None>(__fp16 *, const __fp16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<8, 1, VLType::None>(__fp16 *, const __fp16 *, size_t, const convolver<__fp16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 1, VLType::None>(__fp16 *, const __fp16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); +#endif // FP16_KERNELS ar __ARM_FEATURE_FP16_VECTOR_ARITHMETIC template void IndirectInterleave<8, 1, VLType::None>(float *, const __fp16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<8, 1, VLType::None>(float *, const __fp16 *, size_t, const convolver<__fp16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 1, VLType::None>(float *, const __fp16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); /* BF16 */ -/* NEON/SVE BFDOT */ +/* Arm® Neon™/SVE BFDOT */ +#ifdef ARM_COMPUTE_ENABLE_BF16 template void IndirectInterleave<8, 2, VLType::None>(bfloat16 *, const bfloat16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<8, 2, VLType::None>(bfloat16 *, const bfloat16 *, size_t, const convolver<bfloat16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 2, VLType::None>(bfloat16 *, const bfloat16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); @@ -350,7 +351,12 @@ template void IndirectInterleave<8, 4, VLType::None>(bfloat16 *, const bfloat16 template void ConvolutionInterleave<8, 4, VLType::None>(bfloat16 *, const bfloat16 *, size_t, const convolver<bfloat16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 4, VLType::None>(bfloat16 *, const bfloat16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -/* NEON/SVE using FP32 kernel */ +template void IndirectInterleave<8, 4, VLType::None>(bfloat16 *, const float * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); +template void ConvolutionInterleave<8, 4, VLType::None>(bfloat16 *, const float *, size_t, const convolver<float> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); +template void Interleave<8, 4, VLType::None>(bfloat16 *, const float *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); +#endif // ARM_COMPUTE_ENABLE_BF16 + +/* Arm® Neon™/SVE using FP32 kernel */ template void IndirectInterleave<8, 1, VLType::None>(float *, const bfloat16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<8, 1, VLType::None>(float *, const bfloat16 *, size_t, const convolver<bfloat16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 1, VLType::None>(float *, const bfloat16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); @@ -365,12 +371,12 @@ template void ConvolutionInterleave<8, 1, VLType::None>(uint16_t *, const uint16 template void Interleave<8, 1, VLType::None>(uint16_t *, const uint16_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); /* INT8 */ -/* NEON SMLA/SMLAL (height 4, block 16) */ +/* Arm® Neon™ SMLA/SMLAL (height 4, block 16) */ template void IndirectInterleave<4, 16, VLType::None>(int8_t *, const int8_t * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<4, 16, VLType::None>(int8_t *, const int8_t *, size_t, const convolver<int8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<4, 16, VLType::None>(int8_t *, const int8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -/* NEON SDOT (height 8, block 4) */ +/* Arm® Neon™ SDOT (height 8, block 4) */ template void IndirectInterleave<8, 4, VLType::None>(int8_t *, const int8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t); template void ConvolutionInterleave<8, 4, VLType::None>(int8_t *, const int8_t *, size_t, const convolver<int8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 4, VLType::None>(int8_t *, const int8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); @@ -380,17 +386,17 @@ template void IndirectInterleave<8, 8, VLType::None>(int8_t *, const int8_t * co template void ConvolutionInterleave<8, 8, VLType::None>(int8_t *, const int8_t *, size_t, const convolver<int8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 8, VLType::None>(int8_t *, const int8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -/* NEON SDOT (height 8, block 1) */ +/* Arm® Neon™ SDOT (height 8, block 1) */ template void IndirectInterleave<8, 1, VLType::None>(int16_t *, const int8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t); template void ConvolutionInterleave<8, 1, VLType::None>(int16_t *, const int8_t *, size_t, const convolver<int8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 1, VLType::None>(int16_t *, const int8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -/* NEON SMLA/SMLAL (height 4, block 16) */ +/* Arm® Neon™ SMLA/SMLAL (height 4, block 16) */ template void IndirectInterleave<4, 16, VLType::None>(uint8_t *, const uint8_t * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void ConvolutionInterleave<4, 16, VLType::None>(uint8_t *, const uint8_t *, size_t, const convolver<uint8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<4, 16, VLType::None>(uint8_t *, const uint8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -/* NEON SDOT (height 8, block 4) */ +/* Arm® Neon™ SDOT (height 8, block 4) */ template void IndirectInterleave<8, 4, VLType::None>(uint8_t *, const uint8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t); template void ConvolutionInterleave<8, 4, VLType::None>(uint8_t *, const uint8_t *, size_t, const convolver<uint8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 4, VLType::None>(uint8_t *, const uint8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); @@ -400,7 +406,7 @@ template void IndirectInterleave<8, 8, VLType::None>(uint8_t *, const uint8_t * template void ConvolutionInterleave<8, 8, VLType::None>(uint8_t *, const uint8_t *, size_t, const convolver<uint8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 8, VLType::None>(uint8_t *, const uint8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); -/* NEON 16-bit (height 8, block 1) */ +/* Arm® Neon™ 16-bit (height 8, block 1) */ template void IndirectInterleave<8, 1, VLType::None>(uint16_t *, const uint8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t); template void ConvolutionInterleave<8, 1, VLType::None>(uint16_t *, const uint8_t *, size_t, const convolver<uint8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); template void Interleave<8, 1, VLType::None>(uint16_t *, const uint8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t); |