OMXResult omxSP_FFTFwd_RToCCS_F32_Sfs(const OMX_F32 *pSrc, OMX_F32 *pDst,
const OMXFFTSpec_R_F32 *pFFTSpec) {
// Input must be 32 byte aligned
if (!pSrc || !pDst || (OMX_INT)pSrc & 31 || (OMX_INT)pDst & 31)
return OMX_Sts_BadArgErr;
OMX_INT n;
OMX_INT n_by_2;
OMX_INT n_by_4;
const OMX_F32 *twiddle;
OMX_F32 *buf;
const X86FFTSpec_R_FC32 *pFFTStruct = (const X86FFTSpec_R_FC32*) pFFTSpec;
n = pFFTStruct->N;
// This is to handle the case of order == 1.
if (n == 2) {
pDst[0] = (pSrc[0] + pSrc[1]);
pDst[1] = 0.0f;
pDst[2] = (pSrc[0] - pSrc[1]);
pDst[3] = 0.0f;
return OMX_Sts_NoErr;
}
n_by_2 = n >> 1;
n_by_4 = n >> 2;
buf = pFFTStruct->pBuf1;
twiddle = pFFTStruct->pTwiddle;
if(n_by_2 >= 16) {
buf = x86SP_F32_radix4_kernel_OutOfPlace_sse(
pSrc,
pFFTStruct->pBuf2,
buf,
twiddle,
n_by_2,
1);
} else {
buf = x86SP_F32_radix2_kernel_OutOfPlace(
pSrc,
pFFTStruct->pBuf2,
buf,
twiddle,
n_by_2,
1);
}
if(n >= 8)
RevbinPermuteFwdSse(buf, pDst, twiddle, n);
else
RevbinPermuteFwd(buf, pDst, twiddle, n);
return OMX_Sts_NoErr;
}
OMXResult omxSP_FFTInv_CCSToR_F32_Sfs(const OMX_F32 *pSrc, OMX_F32 *pDst,
const OMXFFTSpec_R_F32 *pFFTSpec) {
OMX_INT n;
OMX_INT n_by_2;
OMX_INT n_by_4;
OMX_INT i;
const OMX_F32 *twiddle;
OMX_F32 *buf;
OMX_F32 *in = (OMX_F32*) pSrc;
const X86FFTSpec_R_FC32 *pFFTStruct = (const X86FFTSpec_R_FC32*) pFFTSpec;
// Input must be 32 byte aligned
if (!pSrc || !pDst || (const uintptr_t)pSrc & 31 || (uintptr_t)pDst & 31)
return OMX_Sts_BadArgErr;
n = pFFTStruct->N;
// This is to handle the case of order == 1.
if (n == 2) {
pDst[0] = (pSrc[0] + pSrc[2]) / 2;
pDst[1] = (pSrc[0] - pSrc[2]) / 2;
return OMX_Sts_NoErr;
}
n_by_2 = n >> 1;
n_by_4 = n >> 2;
buf = pFFTStruct->pBuf1;
twiddle = pFFTStruct->pTwiddle;
if (n < 8)
RevbinPermuteInv(in, buf, twiddle, n);
else
RevbinPermuteInvSse(in, buf, twiddle, n);
if (n_by_2 < 16) {
buf = x86SP_F32_radix2_kernel_OutOfPlace(
buf,
pFFTStruct->pBuf2,
buf,
twiddle,
n_by_2,
0);
} else {
buf = x86SP_F32_radix4_kernel_OutOfPlace_sse(
buf,
pFFTStruct->pBuf2,
buf,
twiddle,
n_by_2,
0);
}
// Scale the result by 1/n.
// It contains a scaling factor of 1/2 in
// RevbinPermuteInv/RevbinPermuteInvSse.
OMX_F32 factor = 1.0f / n;
if (n < 8) {
for (i = 0; i < n_by_2; i++) {
pDst[i << 1] = buf[i] * factor;
pDst[(i << 1) + 1] = buf[i + n_by_2] * factor;
}
} else {
OMX_F32 *base;
OMX_F32 *dst;
VC temp0;
VC temp1;
__m128 mFactor = _mm_load1_ps(&factor);
// Two things are done in this loop:
// 1 Get the result scaled; 2 Change the format from split to interleaved.
for (i = 0; i < n_by_2; i += 4) {
base = buf + i;
dst = pDst + (i << 1);
VC_LOAD_SPLIT(&temp0, base, n_by_2);
VC_MUL_F(&temp1, &temp0, mFactor);
VC_STORE_INTERLEAVE(dst, &temp1);
}
}
return OMX_Sts_NoErr;
}