1
我看了DirectXMath Library的源代碼,發現執行XMVectorSetByIndex
和XMVectorSetX
是完全不一樣的。爲什麼不XMVectorSetX
只是返回XMVectorSetByIndex
(索引= 0)?index = 0時XMVectorSetByIndex和XMVectorSetX有什麼區別?
我看了DirectXMath Library的源代碼,發現執行XMVectorSetByIndex
和XMVectorSetX
是完全不一樣的。爲什麼不XMVectorSetX
只是返回XMVectorSetByIndex
(索引= 0)?index = 0時XMVectorSetByIndex和XMVectorSetX有什麼區別?
XMVectorSetX
實際上能夠使用SSE或ARM-NEON內在函數,而XMVectorSetByIndex
必須「溢出到內存」(即根本不是SIMD)。
// Set a single indexed floating point component
inline XMVECTOR XM_CALLCONV XMVectorSetByIndex(FXMVECTOR V, float f, size_t i)
{
assert(i < 4);
_Analysis_assume_(i < 4);
#if defined(_XM_NO_INTRINSICS_)
XMVECTOR U;
U = V;
U.vector4_f32[i] = f;
return U;
#elif defined(_XM_ARM_NEON_INTRINSICS_)
XMVECTOR U = V;
U.n128_f32[i] = f;
return U;
#elif defined(_XM_SSE_INTRINSICS_)
XMVECTOR U = V;
U.m128_f32[i] = f;
return U;
#endif
}
與
// Sets the X component of a vector to a passed floating point value
inline XMVECTOR XM_CALLCONV XMVectorSetX(FXMVECTOR V, float x)
{
#if defined(_XM_NO_INTRINSICS_)
XMVECTOR U;
U.vector4_f32[0] = x;
U.vector4_f32[1] = V.vector4_f32[1];
U.vector4_f32[2] = V.vector4_f32[2];
U.vector4_f32[3] = V.vector4_f32[3];
return U;
#elif defined(_XM_ARM_NEON_INTRINSICS_)
return vsetq_lane_f32(x,V,0);
#elif defined(_XM_SSE_INTRINSICS_)
XMVECTOR vResult = _mm_set_ss(x);
vResult = _mm_move_ss(V,vResult);
return vResult;
#endif
}
這是信息在XMVectorSetY
情況看,以及在那裏與/arch:AVX
或/arch:AVX2
它能夠使用SSE4指令_mm_insert_ps
否則它必須做的工作的公平位獲得SIMD代碼生成,而不必「泄漏到內存」。
inline XMVECTOR XM_CALLCONV XMVectorSetY(FXMVECTOR V, float y)
{
#if defined(_XM_NO_INTRINSICS_)
XMVECTOR U;
U.vector4_f32[0] = V.vector4_f32[0];
U.vector4_f32[1] = y;
U.vector4_f32[2] = V.vector4_f32[2];
U.vector4_f32[3] = V.vector4_f32[3];
return U;
#elif defined(_XM_ARM_NEON_INTRINSICS_)
return vsetq_lane_f32(y,V,1);
#elif defined(_XM_SSE4_INTRINSICS_)
XMVECTOR vResult = _mm_set_ss(y);
vResult = _mm_insert_ps(V, vResult, 0x10);
return vResult;
#elif defined(_XM_SSE_INTRINSICS_)
// Swap y and x
XMVECTOR vResult = XM_PERMUTE_PS(V,_MM_SHUFFLE(3,2,0,1));
// Convert input to vector
XMVECTOR vTemp = _mm_set_ss(y);
// Replace the x component
vResult = _mm_move_ss(vResult,vTemp);
// Swap y and x again
vResult = XM_PERMUTE_PS(vResult,_MM_SHUFFLE(3,2,0,1));
return vResult;
#endif
}
注意DirectXMath現已有GitHub。