diff options
author | lloyd <[email protected]> | 2009-10-28 19:50:06 +0000 |
---|---|---|
committer | lloyd <[email protected]> | 2009-10-28 19:50:06 +0000 |
commit | 185d85338562627aa4800436a3fe6efa11886351 (patch) | |
tree | a892d454d5d88e008624b60e3e88c037f312d770 /src/utils | |
parent | f5d4cf7509011669c25746e3b4c681b5ebfede79 (diff) |
Add an AltiVec SIMD_32 implementation. Tested and works for Serpent and XTEA
on a PowerPC 970 running Gentoo with GCC 4.3.4
Uses a GCC syntax for creating literal values instead of the Motorola
syntax [{1,2,3,4} instead of (1,2,3,4)].
In tests so far, this is much, much slower than either the standard scalar code,
or using the SIMD-in-scalar-registers code. It looks like for whatever reason
GCC is refusing to inline the function:
SIMD_Altivec(__vector unsigned int input) { reg = input; }
and calls it with a branch hundreds of times in each function. I don't know
if this is the entire reason it's slower, but it definitely can't be helping.
The code handles unaligned loads OK but assumes stores are to an aligned address.
This will fail drastically some day, and needs to be fixed to either use scalar
stores, which (most?) PPCs will handle (if slowly), or batch the loads and
stores so we can work across the loads. Considering the code so far loads 4
vectors of data in one go this would probably be a big win (and also for loads,
since instead of doing 8 loads for 4 registers only 5 are needed).
Diffstat (limited to 'src/utils')
-rw-r--r-- | src/utils/simd_32/simd_altivec.h | 178 |
1 files changed, 178 insertions, 0 deletions
diff --git a/src/utils/simd_32/simd_altivec.h b/src/utils/simd_32/simd_altivec.h new file mode 100644 index 000000000..d6aaa699d --- /dev/null +++ b/src/utils/simd_32/simd_altivec.h @@ -0,0 +1,178 @@ +/** +* Altivec SIMD +*/ + +#ifndef BOTAN_SIMD_ALTIVEC_H__ +#define BOTAN_SIMD_ALTIVEC_H__ + +#include <botan/loadstor.h> +#include <altivec.h> +#undef vector + +namespace Botan { + +class SIMD_Altivec + { + public: + + SIMD_Altivec(const u32bit B[4]) + { + reg = (__vector unsigned int){B[0], B[1], B[2], B[3]}; + } + + SIMD_Altivec(u32bit B0, u32bit B1, u32bit B2, u32bit B3) + { + reg = (__vector unsigned int){B0, B1, B2, B3}; + } + + SIMD_Altivec(u32bit B) + { + reg = (__vector unsigned int){B, B, B, B}; + } + + static SIMD_Altivec load_le(const void* in) + { + const u32bit* in_32 = static_cast<const u32bit*>(in); + + __vector unsigned int R0 = vec_ld(0, in_32); + __vector unsigned int R1 = vec_ld(12, in_32); + + __vector unsigned char perm = vec_lvsl(0, in_32); + + perm = vec_xor(perm, vec_splat_u8(3)); + + R0 = vec_perm(R0, R1, perm); + + return SIMD_Altivec(R0); + } + + static SIMD_Altivec load_be(const void* in) + { + const u32bit* in_32 = static_cast<const u32bit*>(in); + + __vector unsigned int R0 = vec_ld(0, in_32); + __vector unsigned int R1 = vec_ld(12, in_32); + + __vector unsigned char perm = vec_lvsl(0, in_32); + + R0 = vec_perm(R0, R1, perm); + + return SIMD_Altivec(R0); + } + + void store_le(byte out[]) const + { + u32bit* out_32 = reinterpret_cast<u32bit*>(out); + + __vector unsigned char perm = vec_lvsl(0, (int*)0); + + perm = vec_xor(perm, vec_splat_u8(3)); + + __vector unsigned int swapped = vec_perm(reg, reg, perm); + + vec_st(swapped, 0, out_32); + } + + void store_be(byte out[]) const + { + u32bit* out_32 = reinterpret_cast<u32bit*>(out); + vec_st(reg, 0, out_32); + } + + void rotate_left(u32bit rot) + { + __vector unsigned int rot_vec = + (__vector unsigned int){rot, rot, rot, rot}; + + reg = vec_rl(reg, rot_vec); + } + + void rotate_right(u32bit rot) + { + rotate_left(32 - rot); + } + + void operator+=(const SIMD_Altivec& other) + { + reg = vec_add(reg, other.reg); + } + + SIMD_Altivec operator+(const SIMD_Altivec& other) const + { + return vec_add(reg, other.reg); + } + + void operator-=(const SIMD_Altivec& other) + { + reg = vec_sub(reg, other.reg); + } + + SIMD_Altivec operator-(const SIMD_Altivec& other) const + { + return vec_sub(reg, other.reg); + } + + void operator^=(const SIMD_Altivec& other) + { + reg = vec_xor(reg, other.reg); + } + + SIMD_Altivec operator^(const SIMD_Altivec& other) const + { + return vec_xor(reg, other.reg); + } + + void operator|=(const SIMD_Altivec& other) + { + reg = vec_or(reg, other.reg); + } + + void operator&=(const SIMD_Altivec& other) + { + reg = vec_and(reg, other.reg); + } + + SIMD_Altivec operator<<(u32bit shift) const + { + __vector unsigned int shift_vec = + (__vector unsigned int){shift, shift, shift, shift}; + + return vec_sl(reg, shift_vec); + } + + SIMD_Altivec operator>>(u32bit shift) const + { + __vector unsigned int shift_vec = + (__vector unsigned int){shift, shift, shift, shift}; + + return vec_sr(reg, shift_vec); + } + + SIMD_Altivec operator~() const + { + return vec_nor(reg, reg); + } + + static void transpose(SIMD_Altivec& B0, SIMD_Altivec& B1, + SIMD_Altivec& B2, SIMD_Altivec& B3) + { + __vector unsigned int T0 = vec_mergeh(B0.reg, B2.reg); + __vector unsigned int T1 = vec_mergel(B0.reg, B2.reg); + __vector unsigned int T2 = vec_mergeh(B1.reg, B3.reg); + __vector unsigned int T3 = vec_mergel(B1.reg, B3.reg); + + B0.reg = vec_mergeh(T0, T2); + B1.reg = vec_mergel(T0, T2); + B2.reg = vec_mergeh(T1, T3); + B3.reg = vec_mergel(T1, T3); + } + + private: + SIMD_Altivec(__vector unsigned int input) { reg = input; } + + __vector unsigned int reg; + }; + +} + +#endif |