Skip to content

D
dlib
Commit 5f0c1a1c authored by Evgeniy Fominov's avatar Evgeniy Fominov Committed by Davis E. King
Browse files
Options

ARM NEON SIMD support (#564) 

* added neon functions

* TK1 tests passed

* code cleanup

* Re-tested on TK1

* improve simd4i

* fixed simd4f_bool neon

* fix simd8i sse operator <=

* restored rsqrt/VSX

* fixed simd4f/neon/reciprocal_sqrt
parent 63f4f73b
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 93 additions and 7 deletions
dlib/simd/simd4f.h
View file @ 5f0c1a1c
This diff is collapsed.
dlib/simd/simd4i.h
View file @ 5f0c1a1c
......@@ -54,9 +54,9 @@ namespace dlib
vector bool int b;
signed int x[4];
} v4i;
v4i x;
public:
inline simd4i() : x{0,0,0,0} { }
inline simd4i(const simd4i& v) : x(v.x) { }
......@@ -81,13 +81,56 @@ namespace dlib
inline void load(const int32* ptr) { x.v = vec_vsx_ld(0, ptr); }
inline void store(int32* ptr) const { vec_vsx_st(x.v, 0, ptr); }
struct rawarray
{
v4i v;
};
inline simd4i(const rawarray& a) : x{a.v} { }
};
#elif defined(DLIB_HAVE_NEON)
class simd4i
{
public:
typedef int32 type;
inline simd4i() {}
inline simd4i(int32 f) { x = vdupq_n_s32(f); }
inline simd4i(int32 r0, int32 r1, int32 r2, int32 r3)
{
int32 __attribute__((aligned(16))) data[4] = { r0, r1, r2, r3 };
x = vld1q_s32(data);
}
inline simd4i(const int32x4_t& val):x(val) {}
inline simd4i(const uint32x4_t& val):x((int32x4_t)val) {}
inline simd4i& operator=(const int32x4_t& val)
{
x = val;
return *this;
}
inline operator int32x4_t() const { return x; }
inline operator uint32x4_t() const { return (uint32x4_t)x; }
inline void load_aligned(const type* ptr) { x = vld1q_s32(ptr); }
inline void store_aligned(type* ptr) const { vst1q_s32(ptr, x); }
inline void load(const type* ptr) { x = vld1q_s32(ptr); }
inline void store(type* ptr) const { vst1q_s32(ptr, x); }
inline unsigned int size() const { return 4; }
inline int32 operator[](unsigned int idx) const
{
int32 temp[4];
store(temp);
return temp[idx];
}
private:
int32x4_t x;
};
#else
......@@ -165,6 +208,8 @@ namespace dlib
return _mm_add_epi32(lhs, rhs);
#elif defined(DLIB_HAVE_VSX)
return vec_add(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return vaddq_s32(lhs, rhs);
#else
return simd4i(lhs[0]+rhs[0],
lhs[1]+rhs[1],
......@@ -183,6 +228,8 @@ namespace dlib
return _mm_sub_epi32(lhs, rhs);
#elif defined(DLIB_HAVE_VSX)
return vec_sub(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return vsubq_s32(lhs, rhs);
#else
return simd4i(lhs[0]-rhs[0],
lhs[1]-rhs[1],
......@@ -210,6 +257,8 @@ namespace dlib
vector int a = lhs(), b = rhs();
asm("vmuluwm %0, %0, %1\n\t" : "+&v" (a) : "v" (b) );
return simd4i(a);
#elif defined(DLIB_HAVE_NEON)
return vmulq_s32(lhs, rhs);
#else
return simd4i(lhs[0]*rhs[0],
lhs[1]*rhs[1],
......@@ -228,6 +277,8 @@ namespace dlib
return _mm_and_si128(lhs, rhs);
#elif defined(DLIB_HAVE_VSX)
return vec_and(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return vandq_s32(lhs, rhs);
#else
return simd4i(lhs[0]&rhs[0],
lhs[1]&rhs[1],
......@@ -246,6 +297,8 @@ namespace dlib
return _mm_or_si128(lhs, rhs);
#elif defined(DLIB_HAVE_VSX)
return vec_or(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return vorrq_s32(lhs, rhs);
#else
return simd4i(lhs[0]|rhs[0],
lhs[1]|rhs[1],
......@@ -264,6 +317,8 @@ namespace dlib
return _mm_xor_si128(lhs, rhs);
#elif defined(DLIB_HAVE_VSX)
return vec_xor(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return veorq_s32(lhs, rhs);
#else
return simd4i(lhs[0]^rhs[0],
lhs[1]^rhs[1],
......@@ -282,6 +337,8 @@ namespace dlib
return _mm_xor_si128(lhs, _mm_set1_epi32(0xFFFFFFFF));
#elif defined(DLIB_HAVE_VSX)
return vec_xor(lhs(), vec_splats(~0));
#elif defined(DLIB_HAVE_NEON)
return vmvnq_s32(lhs);
#else
return simd4i(~lhs[0],
~lhs[1],
......@@ -298,6 +355,8 @@ namespace dlib
return _mm_sll_epi32(lhs,_mm_cvtsi32_si128(rhs));
#elif defined(DLIB_HAVE_VSX)
return vec_sl(lhs(), vec_splats((uint32_t)rhs));
#elif defined(DLIB_HAVE_NEON)
return vshlq_s32(lhs, simd4i(rhs));
#else
return simd4i(lhs[0]<<rhs,
lhs[1]<<rhs,
......@@ -316,6 +375,12 @@ namespace dlib
return _mm_sra_epi32(lhs,_mm_cvtsi32_si128(rhs));
#elif defined(DLIB_HAVE_VSX)
return vec_sr(lhs(), vec_splats((uint32_t)rhs));
#elif defined(DLIB_HAVE_NEON)
int32 _lhs[4]; lhs.store(_lhs);
return simd4i(_lhs[0]>>rhs,
_lhs[1]>>rhs,
_lhs[2]>>rhs,
_lhs[3]>>rhs);
#else
return simd4i(lhs[0]>>rhs,
lhs[1]>>rhs,
......@@ -334,6 +399,8 @@ namespace dlib
return _mm_cmpeq_epi32(lhs, rhs);
#elif defined(DLIB_HAVE_VSX)
return vec_cmpeq(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return (int32x4_t)vceqq_s32(lhs,rhs);
#else
return simd4i(lhs[0]==rhs[0] ? 0xFFFFFFFF : 0,
lhs[1]==rhs[1] ? 0xFFFFFFFF : 0,
......@@ -346,7 +413,7 @@ namespace dlib
inline simd4i operator!= (const simd4i& lhs, const simd4i& rhs)
{
#if defined(DLIB_HAVE_SSE2) || defined(DLIB_HAVE_VSX)
#if defined(DLIB_HAVE_SSE2) || defined(DLIB_HAVE_VSX) || defined(DLIB_HAVE_NEON)
return ~(lhs==rhs);
#else
return simd4i(lhs[0]!=rhs[0] ? 0xFFFFFFFF : 0,
......@@ -364,6 +431,8 @@ namespace dlib
return _mm_cmplt_epi32(lhs, rhs);
#elif defined(DLIB_HAVE_VSX)
return vec_cmplt(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return (int32x4_t)vcltq_s32(lhs, rhs);
#else
return simd4i(lhs[0]<rhs[0] ? 0xFFFFFFFF : 0,
lhs[1]<rhs[1] ? 0xFFFFFFFF : 0,
......@@ -385,6 +454,8 @@ namespace dlib
{
#ifdef DLIB_HAVE_SSE2
return ~(lhs > rhs);
#elif defined(DLIB_HAVE_NEON)
return (int32x4_t)vcleq_s32(lhs, rhs);
#else
return simd4i(lhs[0]<=rhs[0] ? 0xFFFFFFFF : 0,
lhs[1]<=rhs[1] ? 0xFFFFFFFF : 0,
......@@ -415,6 +486,8 @@ namespace dlib
std::min(_lhs[3],_rhs[3]));
#elif defined(DLIB_HAVE_VSX)
return vec_min(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return (int32x4_t)vminq_s32(lhs, rhs);
#else
return simd4i(std::min(lhs[0],rhs[0]),
std::min(lhs[1],rhs[1]),
......@@ -438,6 +511,8 @@ namespace dlib
std::max(_lhs[3],_rhs[3]));
#elif defined(DLIB_HAVE_VSX)
return vec_max(lhs(), rhs());
#elif defined(DLIB_HAVE_NEON)
return vmaxq_s32(lhs, rhs);
#else
return simd4i(std::max(lhs[0],rhs[0]),
std::max(lhs[1],rhs[1]),
......@@ -458,6 +533,9 @@ namespace dlib
int32 temp[4];
item.store(temp);
return temp[0]+temp[1]+temp[2]+temp[3];
#elif defined(DLIB_HAVE_NEON)
int32x2_t r = vadd_s32(vget_high_s32(item), vget_low_s32(item));
return vget_lane_s32(vpadd_s32(r, r), 0);
#else
return item[0]+item[1]+item[2]+item[3];
#endif
......@@ -474,6 +552,8 @@ namespace dlib
return ((cmp&a) | _mm_andnot_si128(cmp,b));
#elif defined(DLIB_HAVE_VSX)
return vec_sel(b(), a(), cmp.to_bool());
#elif defined(DLIB_HAVE_NEON)
return vbslq_s32(cmp, a, b);
#else
return ((cmp&a) | (~cmp&b));
#endif
......
dlib/simd/simd_check.h
View file @ 5f0c1a1c
......@@ -66,7 +66,11 @@
#define DLIB_HAVE_POWER_VEC
#endif
#endif
#ifdef __ARM_NEON
#ifndef DLIB_HAVE_NEON
#define DLIB_HAVE_NEON
#endif
#endif
#endif
#endif
......@@ -97,7 +101,9 @@
#include <immintrin.h> // AVX
// #include <avx2intrin.h>
#endif
#ifdef DLIB_HAVE_NEON
#include <arm_neon.h> // ARM NEON
#endif
#endif // DLIB_SIMd_CHECK_Hh_
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment