potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

neon: add dotprod dispatch implementation

Browse Source 
* read vector at runtime
     * add enum
This commit is contained in:
Tomoaki Teshima
2022-07-20 19:25:39 +09:00
parent 2a82467a6f
commit b3269b08a1
8 changed files with 106 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/include/opencv2/core/cv_cpu_dispatch.h
+4
View File
@@ -79,6 +79,10 @@
# endif
# define CV_FP16 1
#endif
#ifdef CV_CPU_COMPILE_NEON_DOTPROD
# include <arm_neon.h>
# define CV_NEON_DOT 1
#endif
#ifdef CV_CPU_COMPILE_AVX2
# include <immintrin.h>
# define CV_AVX2 1
modules/core/include/opencv2/core/cv_cpu_helper.h
+21
View File
@@ -420,6 +420,27 @@
#endif
#define __CV_CPU_DISPATCH_CHAIN_NEON(fn, args, mode, ...) CV_CPU_CALL_NEON(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_NEON_DOTPROD
# define CV_TRY_NEON_DOTPROD 1
# define CV_CPU_FORCE_NEON_DOTPROD 1
# define CV_CPU_HAS_SUPPORT_NEON_DOTPROD 1
# define CV_CPU_CALL_NEON_DOTPROD(fn, args) return (cpu_baseline::fn args)
# define CV_CPU_CALL_NEON_DOTPROD_(fn, args) return (opt_NEON_DOTPROD::fn args)
#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_NEON_DOTPROD
# define CV_TRY_NEON_DOTPROD 1
# define CV_CPU_FORCE_NEON_DOTPROD 0
# define CV_CPU_HAS_SUPPORT_NEON_DOTPROD (cv::checkHardwareSupport(CV_CPU_NEON_DOTPROD))
# define CV_CPU_CALL_NEON_DOTPROD(fn, args) if (CV_CPU_HAS_SUPPORT_NEON_DOTPROD) return (opt_NEON_DOTPROD::fn args)
# define CV_CPU_CALL_NEON_DOTPROD_(fn, args) if (CV_CPU_HAS_SUPPORT_NEON_DOTPROD) return (opt_NEON_DOTPROD::fn args)
#else
# define CV_TRY_NEON_DOTPROD 0
# define CV_CPU_FORCE_NEON_DOTPROD 0
# define CV_CPU_HAS_SUPPORT_NEON_DOTPROD 0
# define CV_CPU_CALL_NEON_DOTPROD(fn, args)
# define CV_CPU_CALL_NEON_DOTPROD_(fn, args)
#endif
#define __CV_CPU_DISPATCH_CHAIN_NEON_DOTPROD(fn, args, mode, ...) CV_CPU_CALL_NEON_DOTPROD(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_MSA
# define CV_TRY_MSA 1
# define CV_CPU_FORCE_MSA 1
modules/core/include/opencv2/core/cvdef.h
+2
View File
@@ -282,6 +282,7 @@ namespace cv {
#define CV_CPU_AVX_5124FMAPS 27
#define CV_CPU_NEON 100
#define CV_CPU_NEON_DOTPROD 101
#define CV_CPU_MSA 150
@@ -334,6 +335,7 @@ enum CpuFeatures {
CPU_AVX_5124FMAPS = 27,
CPU_NEON = 100,
CPU_NEON_DOTPROD = 101,
CPU_MSA = 150,
modules/core/include/opencv2/core/hal/intrin_neon.hpp
+30 -35
View File
@@ -78,8 +78,6 @@ CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
#define CV_NEON_AARCH64 0
#endif
// TODO
#define CV_NEON_DOT 0
//////////// Utils ////////////
@@ -667,11 +665,22 @@ inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64
}
// 8 >> 32
#ifdef CV_NEON_DOT
#define OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_OP(_Tpvec1, _Tpvec2, suffix) \
inline _Tpvec1 v_dotprod_expand(const _Tpvec2& a, const _Tpvec2& b) \
{ \
return _Tpvec1(vdotq_##suffix(vdupq_n_##suffix(0), a.val, b.val));\
} \
inline _Tpvec1 v_dotprod_expand(const _Tpvec2& a, const _Tpvec2& b, const _Tpvec1& c) \
{ \
return _Tpvec1(vdotq_##suffix(c.val, a.val, b.val)); \
}
OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_OP(v_uint32x4, v_uint8x16, u32)
OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_OP(v_int32x4, v_int8x16, s32)
#else
inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
{
#if CV_NEON_DOT
return v_uint32x4(vdotq_u32(vdupq_n_u32(0), a.val, b.val));
#else
const uint8x16_t zero = vreinterpretq_u8_u32(vdupq_n_u32(0));
const uint8x16_t mask = vreinterpretq_u8_u32(vdupq_n_u32(0x00FF00FF));
const uint16x8_t zero32 = vreinterpretq_u16_u32(vdupq_n_u32(0));
@@ -687,23 +696,15 @@ inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
uint32x4_t s1 = vaddq_u32(vshrq_n_u32(vreinterpretq_u32_u16(even), 16),
vshrq_n_u32(vreinterpretq_u32_u16(odd), 16));
return v_uint32x4(vaddq_u32(s0, s1));
#endif
}
inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b,
const v_uint32x4& c)
{
#if CV_NEON_DOT
return v_uint32x4(vdotq_u32(c.val, a.val, b.val));
#else
return v_dotprod_expand(a, b) + c;
#endif
}
inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
{
#if CV_NEON_DOT
return v_int32x4(vdotq_s32(vdupq_n_s32(0), a.val, b.val));
#else
int16x8_t p0 = vmull_s8(vget_low_s8(a.val), vget_low_s8(b.val));
int16x8_t p1 = vmull_s8(vget_high_s8(a.val), vget_high_s8(b.val));
int16x8_t uzp1, uzp2;
@@ -712,18 +713,13 @@ inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
int16x4_t uzpl1, uzpl2;
_v128_unzip(vget_low_s16(sum), vget_high_s16(sum), uzpl1, uzpl2);
return v_int32x4(vaddl_s16(uzpl1, uzpl2));
#endif
}
inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b,
const v_int32x4& c)
{
#if CV_NEON_DOT
return v_int32x4(vdotq_s32(c.val, a.val, b.val));
#else
return v_dotprod_expand(a, b) + c;
#endif
}
#endif
// 16 >> 64
inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
{
@@ -832,45 +828,44 @@ inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_
}
// 8 >> 32
#ifdef CV_NEON_DOT
#define OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_FAST_OP(_Tpvec1, _Tpvec2, suffix) \
inline _Tpvec1 v_dotprod_expand_fast(const _Tpvec2& a, const _Tpvec2& b) \
{ \
return v_dotprod_expand(a, b); \
} \
inline _Tpvec1 v_dotprod_expand_fast(const _Tpvec2& a, const _Tpvec2& b, const _Tpvec1& c) \
{ \
return v_dotprod_expand(a, b, c); \
}
OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_FAST_OP(v_uint32x4, v_uint8x16, u32)
OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_FAST_OP(v_int32x4, v_int8x16, s32)
#else
inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
{
#if CV_NEON_DOT
return v_uint32x4(vdotq_u32(vdupq_n_u32(0), a.val, b.val));
#else
uint16x8_t p0 = vmull_u8(vget_low_u8(a.val), vget_low_u8(b.val));
uint16x8_t p1 = vmull_u8(vget_high_u8(a.val), vget_high_u8(b.val));
uint32x4_t s0 = vaddl_u16(vget_low_u16(p0), vget_low_u16(p1));
uint32x4_t s1 = vaddl_u16(vget_high_u16(p0), vget_high_u16(p1));
return v_uint32x4(vaddq_u32(s0, s1));
#endif
}
inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
{
#if CV_NEON_DOT
return v_uint32x4(vdotq_u32(c.val, a.val, b.val));
#else
return v_dotprod_expand_fast(a, b) + c;
#endif
}
inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
{
#if CV_NEON_DOT
return v_int32x4(vdotq_s32(vdupq_n_s32(0), a.val, b.val));
#else
int16x8_t prod = vmull_s8(vget_low_s8(a.val), vget_low_s8(b.val));
prod = vmlal_s8(prod, vget_high_s8(a.val), vget_high_s8(b.val));
return v_int32x4(vaddl_s16(vget_low_s16(prod), vget_high_s16(prod)));
#endif
}
inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
{
#if CV_NEON_DOT
return v_int32x4(vdotq_s32(c.val, a.val, b.val));
#else
return v_dotprod_expand_fast(a, b) + c;
#endif
}
#endif
// 16 >> 64
inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)