From a6825a427a51da815a805d66ce65c98de282d89b Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 13 Sep 2018 12:24:03 +0100 Subject: COMPMID-1540 Implement YOLOLayer on NEON Change-Id: Ice28996959dc666fff5e8ae486c1ff8093db083f Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148367 Reviewed-by: Georgios Pinitas Tested-by: bsgcomp --- arm_compute/core/NEON/wrapper/intrinsics/add.h | 68 ++++++++++++++++++++++ arm_compute/core/NEON/wrapper/intrinsics/exp.h | 47 +++++++++++++++ .../core/NEON/wrapper/intrinsics/intrinsics.h | 6 ++ arm_compute/core/NEON/wrapper/intrinsics/inv.h | 53 +++++++++++++++++ arm_compute/core/NEON/wrapper/intrinsics/mla.h | 58 ++++++++++++++++++ arm_compute/core/NEON/wrapper/intrinsics/mul.h | 58 ++++++++++++++++++ arm_compute/core/NEON/wrapper/intrinsics/neg.h | 65 +++++++++++++++++++++ 7 files changed, 355 insertions(+) create mode 100644 arm_compute/core/NEON/wrapper/intrinsics/add.h create mode 100644 arm_compute/core/NEON/wrapper/intrinsics/exp.h create mode 100644 arm_compute/core/NEON/wrapper/intrinsics/inv.h create mode 100644 arm_compute/core/NEON/wrapper/intrinsics/mla.h create mode 100644 arm_compute/core/NEON/wrapper/intrinsics/mul.h create mode 100644 arm_compute/core/NEON/wrapper/intrinsics/neg.h (limited to 'arm_compute/core/NEON/wrapper/intrinsics') diff --git a/arm_compute/core/NEON/wrapper/intrinsics/add.h b/arm_compute/core/NEON/wrapper/intrinsics/add.h new file mode 100644 index 0000000000..da730f133c --- /dev/null +++ b/arm_compute/core/NEON/wrapper/intrinsics/add.h @@ -0,0 +1,68 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_WRAPPER_ADD_H__ +#define __ARM_COMPUTE_WRAPPER_ADD_H__ + +#include + +namespace arm_compute +{ +namespace wrapper +{ +#define VADD_IMPL(stype, vtype, prefix, postfix) \ + inline vtype vadd(const vtype &a, const vtype &b) \ + { \ + return prefix##_##postfix(a, b); \ + } + +VADD_IMPL(uint8x8_t, uint8x8_t, vadd, u8) +VADD_IMPL(int8x8_t, int8x8_t, vadd, s8) +VADD_IMPL(uint16x4_t, uint16x4_t, vadd, u16) +VADD_IMPL(int16x4_t, int16x4_t, vadd, s16) +VADD_IMPL(uint32x2_t, uint32x2_t, vadd, u32) +VADD_IMPL(int32x2_t, int32x2_t, vadd, s32) +VADD_IMPL(uint64x1_t, uint64x1_t, vadd, u64) +VADD_IMPL(int64x1_t, int64x1_t, vadd, s64) +VADD_IMPL(float32x2_t, float32x2_t, vadd, f32) +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +VADD_IMPL(float16x4_t, float16x4_t, vadd, f16) +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +VADD_IMPL(uint8x16_t, uint8x16_t, vaddq, u8) +VADD_IMPL(int8x16_t, int8x16_t, vaddq, s8) +VADD_IMPL(uint16x8_t, uint16x8_t, vaddq, u16) +VADD_IMPL(int16x8_t, int16x8_t, vaddq, s16) +VADD_IMPL(uint32x4_t, uint32x4_t, vaddq, u32) +VADD_IMPL(int32x4_t, int32x4_t, vaddq, s32) +VADD_IMPL(uint64x2_t, uint64x2_t, vaddq, u64) +VADD_IMPL(int64x2_t, int64x2_t, vaddq, s64) +VADD_IMPL(float32x4_t, float32x4_t, vaddq, f32) +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +VADD_IMPL(float16x8_t, float16x8_t, vaddq, f16) +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#undef VADD_IMPL +} // namespace wrapper +} // namespace arm_compute +#endif /* __ARM_COMPUTE_WRAPPER_ADD_H__ */ diff --git a/arm_compute/core/NEON/wrapper/intrinsics/exp.h b/arm_compute/core/NEON/wrapper/intrinsics/exp.h new file mode 100644 index 0000000000..85aa9c739a --- /dev/null +++ b/arm_compute/core/NEON/wrapper/intrinsics/exp.h @@ -0,0 +1,47 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_WRAPPER_EXP_H__ +#define __ARM_COMPUTE_WRAPPER_EXP_H__ + +#include "arm_compute/core/NEON/NEMath.h" +#include + +namespace arm_compute +{ +namespace wrapper +{ +#define VEXPQ_IMPL(vtype, postfix) \ + inline vtype vexpq(const vtype &a) \ + { \ + return vexpq_##postfix(a); \ + } + +VEXPQ_IMPL(float32x4_t, f32) +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +VEXPQ_IMPL(float16x8_t, f16) +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#undef VEXPQ_IMPL +} // namespace wrapper +} // namespace arm_compute +#endif /* __ARM_COMPUTE_WRAPPER_EXP_H__ */ diff --git a/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h b/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h index b302b366cd..58bfba9645 100644 --- a/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h +++ b/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h @@ -24,11 +24,17 @@ #ifndef __ARM_COMPUTE_WRAPPER_INTRINSICS_H__ #define __ARM_COMPUTE_WRAPPER_INTRINSICS_H__ +#include "arm_compute/core/NEON/wrapper/intrinsics/add.h" #include "arm_compute/core/NEON/wrapper/intrinsics/and.h" #include "arm_compute/core/NEON/wrapper/intrinsics/dup_n.h" +#include "arm_compute/core/NEON/wrapper/intrinsics/exp.h" +#include "arm_compute/core/NEON/wrapper/intrinsics/inv.h" #include "arm_compute/core/NEON/wrapper/intrinsics/load.h" #include "arm_compute/core/NEON/wrapper/intrinsics/max.h" #include "arm_compute/core/NEON/wrapper/intrinsics/min.h" +#include "arm_compute/core/NEON/wrapper/intrinsics/mla.h" +#include "arm_compute/core/NEON/wrapper/intrinsics/mul.h" +#include "arm_compute/core/NEON/wrapper/intrinsics/neg.h" #include "arm_compute/core/NEON/wrapper/intrinsics/store.h" #endif /* __ARM_COMPUTE_WRAPPER_INTRINSICS_H__ */ diff --git a/arm_compute/core/NEON/wrapper/intrinsics/inv.h b/arm_compute/core/NEON/wrapper/intrinsics/inv.h new file mode 100644 index 0000000000..a86a9d4671 --- /dev/null +++ b/arm_compute/core/NEON/wrapper/intrinsics/inv.h @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_WRAPPER_INV_H__ +#define __ARM_COMPUTE_WRAPPER_INV_H__ + +#include "arm_compute/core/NEON/NEMath.h" +#include + +namespace arm_compute +{ +namespace wrapper +{ +#define VINV_IMPL(vtype, prefix, postfix) \ + inline vtype vinv(const vtype &a) \ + { \ + return prefix##_##postfix(a); \ + } + +VINV_IMPL(float32x2_t, vinv, f32) +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +VINV_IMPL(float16x4_t, vinv, f16) +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +VINV_IMPL(float32x4_t, vinvq, f32) +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +VINV_IMPL(float16x8_t, vinvq, f16) +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#undef VINV_IMPL +} // namespace wrapper +} // namespace arm_compute +#endif /* __ARM_COMPUTE_WRAPPER_INV_H__ */ diff --git a/arm_compute/core/NEON/wrapper/intrinsics/mla.h b/arm_compute/core/NEON/wrapper/intrinsics/mla.h new file mode 100644 index 0000000000..32a650b57f --- /dev/null +++ b/arm_compute/core/NEON/wrapper/intrinsics/mla.h @@ -0,0 +1,58 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_WRAPPER_MLA_H__ +#define __ARM_COMPUTE_WRAPPER_MLA_H__ + +#include + +namespace arm_compute +{ +namespace wrapper +{ +#define VMLA_IMPL(stype, vtype, prefix, postfix) \ + inline vtype vmla(const vtype &a, const vtype &b, const vtype &c) \ + { \ + return prefix##_##postfix(a, b, c); \ + } + +VMLA_IMPL(uint8x8_t, uint8x8_t, vmla, u8) +VMLA_IMPL(int8x8_t, int8x8_t, vmla, s8) +VMLA_IMPL(uint16x4_t, uint16x4_t, vmla, u16) +VMLA_IMPL(int16x4_t, int16x4_t, vmla, s16) +VMLA_IMPL(uint32x2_t, uint32x2_t, vmla, u32) +VMLA_IMPL(int32x2_t, int32x2_t, vmla, s32) +VMLA_IMPL(float32x2_t, float32x2_t, vmla, f32) + +VMLA_IMPL(uint8x16_t, uint8x16_t, vmlaq, u8) +VMLA_IMPL(int8x16_t, int8x16_t, vmlaq, s8) +VMLA_IMPL(uint16x8_t, uint16x8_t, vmlaq, u16) +VMLA_IMPL(int16x8_t, int16x8_t, vmlaq, s16) +VMLA_IMPL(uint32x4_t, uint32x4_t, vmlaq, u32) +VMLA_IMPL(int32x4_t, int32x4_t, vmlaq, s32) +VMLA_IMPL(float32x4_t, float32x4_t, vmlaq, f32) + +#undef VMLA_IMPL +} // namespace wrapper +} // namespace arm_compute +#endif /* __ARM_COMPUTE_WRAPPER_MLA_H__ */ diff --git a/arm_compute/core/NEON/wrapper/intrinsics/mul.h b/arm_compute/core/NEON/wrapper/intrinsics/mul.h new file mode 100644 index 0000000000..c1908fc7b3 --- /dev/null +++ b/arm_compute/core/NEON/wrapper/intrinsics/mul.h @@ -0,0 +1,58 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_WRAPPER_MUL_H__ +#define __ARM_COMPUTE_WRAPPER_MUL_H__ + +#include + +namespace arm_compute +{ +namespace wrapper +{ +#define VMUL_IMPL(stype, vtype, prefix, postfix) \ + inline vtype vmul(const vtype &a, const vtype &b) \ + { \ + return prefix##_##postfix(a, b); \ + } + +VMUL_IMPL(uint8x8_t, uint8x8_t, vmul, u8) +VMUL_IMPL(int8x8_t, int8x8_t, vmul, s8) +VMUL_IMPL(uint16x4_t, uint16x4_t, vmul, u16) +VMUL_IMPL(int16x4_t, int16x4_t, vmul, s16) +VMUL_IMPL(uint32x2_t, uint32x2_t, vmul, u32) +VMUL_IMPL(int32x2_t, int32x2_t, vmul, s32) +VMUL_IMPL(float32x2_t, float32x2_t, vmul, f32) + +VMUL_IMPL(uint8_t, uint8x16_t, vmulq, u8) +VMUL_IMPL(int8_t, int8x16_t, vmulq, s8) +VMUL_IMPL(uint16_t, uint16x8_t, vmulq, u16) +VMUL_IMPL(int16_t, int16x8_t, vmulq, s16) +VMUL_IMPL(uint32_t, uint32x4_t, vmulq, u32) +VMUL_IMPL(int32_t, int32x4_t, vmulq, s32) +VMUL_IMPL(float32x4_t, float32x4_t, vmulq, f32) + +#undef VMUL_IMPL +} // namespace wrapper +} // namespace arm_compute +#endif /* __ARM_COMPUTE_WRAPPER_MUL_H__ */ diff --git a/arm_compute/core/NEON/wrapper/intrinsics/neg.h b/arm_compute/core/NEON/wrapper/intrinsics/neg.h new file mode 100644 index 0000000000..0ea1d429fe --- /dev/null +++ b/arm_compute/core/NEON/wrapper/intrinsics/neg.h @@ -0,0 +1,65 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_WRAPPER_NEG_H__ +#define __ARM_COMPUTE_WRAPPER_NEG_H__ + +#include + +namespace arm_compute +{ +namespace wrapper +{ +#define VNEG_IMPL(vtype, postfix) \ + inline vtype vneg(const vtype &a) \ + { \ + return vneg_##postfix(a); \ + } + +VNEG_IMPL(int8x8_t, s8) +VNEG_IMPL(int16x4_t, s16) +VNEG_IMPL(int32x2_t, s32) +VNEG_IMPL(float32x2_t, f32) +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +VNEG_IMPL(float16x4_t, f16) +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#undef VNEG_IMPL +#define VNEGQ_IMPL(vtype, postfix) \ + inline vtype vnegq(const vtype &a) \ + { \ + return vnegq_##postfix(a); \ + } + +VNEGQ_IMPL(int8x16_t, s8) +VNEGQ_IMPL(int16x8_t, s16) +VNEGQ_IMPL(int32x4_t, s32) +VNEGQ_IMPL(float32x4_t, f32) +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +VNEGQ_IMPL(float16x8_t, f16) +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#undef VNEGQ_IMPL +} // namespace wrapper +} // namespace arm_compute +#endif /* __ARM_COMPUTE_WRAPPER_NEG_H__ */ -- cgit v1.2.1