diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp | 67 |
1 files changed, 50 insertions, 17 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp index 4fdaab84bd..067d0bf258 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,52 +22,85 @@ * SOFTWARE. */ #pragma once +#ifdef ARM_COMPUTE_ENABLE_SVE -#ifdef __ARM_FEATURE_SVE - -#include <cstdint> #include "../std_transforms_sve.hpp" +#include "../performance_parameters.hpp" -namespace arm_gemm { +#define ARGLIST \ + const uint8_t *, const uint8_t *, \ + uint32_t *, int, int, int +namespace arm_gemm +{ // Actual kernel implementations -void sve_interleaved_u8u32_mmla_8x3VL(const uint8_t *, const uint8_t *, uint32_t *, int, int, int); +void sve_interleaved_u8u32_mmla_8x3VL( ARGLIST ); -class cls_sve_interleaved_u8u32_mmla_8x3VL { +class cls_sve_interleaved_u8u32_mmla_8x3VL +{ public: typedef uint8_t operand_type; typedef uint32_t result_type; - typedef void (*kern_type)(const uint8_t *, const uint8_t *, uint32_t *, int, int, int); + typedef void (*kern_type)( ARGLIST ); /* Kernel blocking parameters */ - static unsigned int out_width() + static constexpr unsigned int out_height() { - return get_vector_length<uint32_t>() * 3; + return 8; } - static unsigned int out_height() + static unsigned int out_width() { - return 8; + return get_vector_length<uint32_t>() * 3; } - static unsigned int k_unroll() + static constexpr unsigned int k_unroll() { return 8; } - // Use the standard fixed size transforms. + StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2> transforms = {}; StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2, true> transforms_quantized = {}; + template<typename T> + static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci) + { - kern_type kernel=sve_interleaved_u8u32_mmla_8x3VL; + if (std::is_same<T, uint32_t>::value) { + switch (ci->get_cpu_model()) { + default: + return { 61.97, 4.11, 7.93 }; + case CPUModel::A510: + return { 43.18, 3.57, 2.89 }; + case CPUModel::V1: + return { 123.47, 5.03, 11.76 }; + } + } + + + if (std::is_same<T, uint8_t>::value) { + switch (ci->get_cpu_model()) { + default: + return { 62.00, 4.08, 0.51 }; + case CPUModel::A510: + return { 38.02, 1.85, 0.28 }; + case CPUModel::V1: + return { 95.28, 7.99, 0.79 }; + } + } + return { 1.0 }; + } + + // Default to the generic kernel + kern_type kernel=sve_interleaved_u8u32_mmla_8x3VL; cls_sve_interleaved_u8u32_mmla_8x3VL(const CPUInfo *) { - } }; } // namespace arm_gemm -#endif // __ARM_FEATURE_SVE +#undef ARGLIST +#endif // ARM_COMPUTE_ENABLE_SVE |