コグノスケ

NEONを使ったCubeHashの素朴な実装

#if defined(__ARM_NEON__)
#  include <arm_neon.h>
#endif

//...

#define NEON_ROTL(x, n) do { \
		uint32x4_t mw0, mw1; \
		mw0 = vshlq_n_u32((x), (n)); \
		mw1 = vshrq_n_u32((x), 32 - (n)); \
		x = vorrq_u32(mw0, mw1); \
	} while (0);

#define NEON_SWP(a, b) do { \
		uint32x4_t mw; \
		mw = b; \
		b = a; \
		a = mw; \
	} while (0);

#define NEON_STEP5(x) do { \
		uint64x2_t mw; \
		mw = vreinterpretq_u64_u32((x)); \
		mw = vextq_u64(mw, mw, 1); \
		x = vreinterpretq_u32_u64(mw); \
	} while (0);

#define ROUND_ONE_NEON    do { \
		mxg = vaddq_u32(mx0, mxg); \
		mxk = vaddq_u32(mx4, mxk); \
		mxo = vaddq_u32(mx8, mxo); \
		mxs = vaddq_u32(mxc, mxs); \
		NEON_ROTL(mx0, 7); \
		NEON_ROTL(mx4, 7); \
		NEON_ROTL(mx8, 7); \
		NEON_ROTL(mxc, 7); \
		NEON_SWP(mx0, mx8); \
		NEON_SWP(mx4, mxc); \
		mx0 = veorq_u32(mx0, mxg); \
		mx4 = veorq_u32(mx4, mxk); \
		mx8 = veorq_u32(mx8, mxo); \
		mxc = veorq_u32(mxc, mxs); \
		NEON_STEP5(mxg); \
		NEON_STEP5(mxk); \
		NEON_STEP5(mxo); \
		NEON_STEP5(mxs); \
		mxg = vaddq_u32(mx0, mxg); \
		mxk = vaddq_u32(mx4, mxk); \
		mxo = vaddq_u32(mx8, mxo); \
		mxs = vaddq_u32(mxc, mxs); \
		NEON_ROTL(mx0, 11); \
		NEON_ROTL(mx4, 11); \
		NEON_ROTL(mx8, 11); \
		NEON_ROTL(mxc, 11); \
		NEON_SWP(mx0, mx4); \
		NEON_SWP(mx8, mxc); \
		mx0 = veorq_u32(mx0, mxg); \
		mx4 = veorq_u32(mx4, mxk); \
		mx8 = veorq_u32(mx8, mxo); \
		mxc = veorq_u32(mxc, mxs); \
		mxg = vrev64q_u32(mxg); \
		mxk = vrev64q_u32(mxk); \
		mxo = vrev64q_u32(mxo); \
		mxs = vrev64q_u32(mxs); \
	} while (0)

#define SIXTEEN_ROUNDS_NEON   do { \
		int j; \
		uint32x4_t mx0, mx4, mx8, mxc; \
		uint32x4_t mxg, mxk, mxo, mxs; \
		mx0 = vld1q_u32((void *)&x0); \
		mx4 = vld1q_u32((void *)&x4); \
		mx8 = vld1q_u32((void *)&x8); \
		mxc = vld1q_u32((void *)&xc); \
		mxg = vld1q_u32((void *)&xg); \
		mxk = vld1q_u32((void *)&xk); \
		mxo = vld1q_u32((void *)&xo); \
		mxs = vld1q_u32((void *)&xs); \
		for (j = 0; j < 16; j ++) { \
			ROUND_ONE_NEON; \
		} \
		vst1q_u32(&x0, mx0); \
		vst1q_u32(&x4, mx4); \
		vst1q_u32(&x8, mx8); \
		vst1q_u32(&xc, mxc); \
		vst1q_u32(&xg, mxg); \
		vst1q_u32(&xk, mxk); \
		vst1q_u32(&xo, mxo); \
		vst1q_u32(&xs, mxs); \
	} while (0)

//...

#if defined(__ARM_NEON__)
#  define ROUND_ONE    ROUND_ONE_NEON
#  define SIXTEEN_ROUNDS    SIXTEEN_ROUNDS_NEON
#else
#  define ROUND_ONE    ROUND_ONE_SLOW
#  define SIXTEEN_ROUNDS    SIXTEEN_ROUNDS_SLOW
#endif