buffer_c16_t TranslateByFSOver4AndDecimateBy2CIC3::execute(const buffer_c8_t& src, const buffer_c16_t& dst) {
/* Translates incoming complex<int8_t> samples by -fs/4,
* decimates by two using a non-recursive third-order CIC filter.
*/
/* Derivation of algorithm:
* Original CIC filter (decimating by two):
* D_I0 = i3 * 1 + i2 * 3 + i1 * 3 + i0 * 1
* D_Q0 = q3 * 1 + q2 * 3 + q1 * 3 + q0 * 1
*
* D_I1 = i5 * 1 + i4 * 3 + i3 * 3 + i2 * 1
* D_Q1 = q5 * 1 + q4 * 3 + q3 * 3 + q2 * 1
*
* Translate -fs/4, phased 180 degrees, accomplished by complex multiplication
* of complex length-4 sequence:
*
* Substitute:
* i0 = -i0, q0 = -q0
* i1 = -q1, q1 = i1
* i2 = i2, q2 = q2
* i3 = q3, q3 = -i3
* i4 = -i4, q4 = -q4
* i5 = -q5, q5 = i5
*
* Resulting taps (with decimation by 2, four samples in, two samples out):
* D_I0 = q3 * 1 + i2 * 3 + -q1 * 3 + -i0 * 1
* D_Q0 = -i3 * 1 + q2 * 3 + i1 * 3 + -q0 * 1
*
* D_I1 = -q5 * 1 + -i4 * 3 + q3 * 3 + i2 * 1
* D_Q1 = i5 * 1 + -q4 * 3 + -i3 * 3 + q2 * 1
*/
// 6 cycles per complex input sample, not including loop overhead.
uint32_t q1_i0 = _q1_i0;
uint32_t q0_i1 = _q0_i1;
/* 3:1 Scaled by 32 to normalize output to +/-32768-ish. */
constexpr uint32_t scale_factor = 32;
const uint32_t k_3_1 = 0x00030001 * scale_factor;
uint32_t* src_p = reinterpret_cast<uint32_t*>(&src.p[0]);
uint32_t* const src_end = reinterpret_cast<uint32_t*>(&src.p[src.count]);
uint32_t* dst_p = reinterpret_cast<uint32_t*>(&dst.p[0]);
while(src_p < src_end) {
const uint32_t q3_i3_q2_i2 = *(src_p++); // 3
const uint32_t q5_i5_q4_i4 = *(src_p++);
const uint32_t i2_i3 = __SXTB16(q3_i3_q2_i2, 16); // 1: (q3_i3_q2_i2 ror 16)[23:16]:(q3_i3_q2_i2 ror 16)[7:0]
const uint32_t q3_q2 = __SXTB16(q3_i3_q2_i2, 8); // 1: (q3_i3_q2_i2 ror 8)[23:16]:(q3_i3_q2_i2 ror 8)[7:0]
const uint32_t i2_q3 = __PKHTB(i2_i3, q3_q2, 16); // 1: Rn[31:16]:(Rm>>16)[15:0]
const uint32_t i3_q2 = __PKHBT(q3_q2, i2_i3, 16); // 1:(Rm<<16)[31:16]:Rn[15:0]
// D_I0 = 3 * (i2 - q1) + (q3 - i0)
const uint32_t i2_m_q1_q3_m_i0 = __QSUB16(i2_q3, q1_i0); // 1: Rn[31:16]-Rm[31:16]:Rn[15:0]-Rm[15:0]
const uint32_t d_i0 = __SMUAD(k_3_1, i2_m_q1_q3_m_i0); // 1: Rm[15:0]*Rs[15:0]+Rm[31:16]*Rs[31:16]
// D_Q0 = 3 * (q2 + i1) - (i3 + q0)
const uint32_t i3_p_q0_q2_p_i1 = __QADD16(i3_q2, q0_i1); // 1: Rn[31:16]+Rm[31:16]:Rn[15:0]+Rm[15:0]
const uint32_t d_q0 = __SMUSDX(i3_p_q0_q2_p_i1, k_3_1); // 1: Rm[15:0]*Rs[31:16]–Rm[31:16]*RsX[15:0]
const uint32_t d_q0_i0 = __PKHBT(d_i0, d_q0, 16); // 1: (Rm<<16)[31:16]:Rn[15:0]
const uint32_t i5_i4 = __SXTB16(q5_i5_q4_i4, 0); // 1: (q5_i5_q4_i4 ror 0)[23:16]:(q5_i5_q4_i4 ror 0)[7:0]
const uint32_t q4_q5 = __SXTB16(q5_i5_q4_i4, 24); // 1: (q5_i5_q4_i4 ror 24)[23:16]:(q5_i5_q4_i4 ror 24)[7:0]
const uint32_t q4_i5 = __PKHTB(q4_q5, i5_i4, 16); // 1: Rn[31:16]:(Rm>>16)[15:0]
const uint32_t q5_i4 = __PKHBT(i5_i4, q4_q5, 16); // 1: (Rm<<16)[31:16]:Rn[15:0]
// D_I1 = (i2 - q5) + 3 * (q3 - i4)
const uint32_t i2_m_q5_q3_m_i4 = __QSUB16(i2_q3, q5_i4); // 1: Rn[31:16]-Rm[31:16]:Rn[15:0]-Rm[15:0]
const uint32_t d_i1 = __SMUADX(i2_m_q5_q3_m_i4, k_3_1); // 1: Rm[15:0]*Rs[31:16]+Rm[31:16]*Rs[15:0]
// D_Q1 = (i5 + q2) - 3 * (q4 + i3)
const uint32_t q4_p_i3_i5_p_q2 = __QADD16(q4_i5, i3_q2); // 1: Rn[31:16]+Rm[31:16]:Rn[15:0]+Rm[15:0]
const uint32_t d_q1 = __SMUSD(k_3_1, q4_p_i3_i5_p_q2); // 1: Rm[15:0]*Rs[15:0]–Rm[31:16]*Rs[31:16]
const uint32_t d_q1_i1 = __PKHBT(d_i1, d_q1, 16); // 1: (Rm<<16)[31:16]:Rn[15:0]
*(dst_p++) = d_q0_i0; // 3
*(dst_p++) = d_q1_i1;
q1_i0 = q5_i4;
q0_i1 = q4_i5;
}
_q1_i0 = q1_i0;
_q0_i1 = q0_i1;
return { dst.p, src.count / 2, src.sampling_rate / 2 };
}
void arm_biquad_cascade_df1_fast_q15(
const arm_biquad_casd_df1_inst_q15 * S,
q15_t * pSrc,
q15_t * pDst,
uint32_t blockSize)
{
q15_t *pIn = pSrc; /* Source pointer */
q15_t *pOut = pDst; /* Destination pointer */
q31_t in; /* Temporary variable to hold input value */
q31_t out; /* Temporary variable to hold output value */
q31_t b0; /* Temporary variable to hold bo value */
q31_t b1, a1; /* Filter coefficients */
q31_t state_in, state_out; /* Filter state variables */
q31_t acc0; /* Accumulator */
int32_t shift = (int32_t) (15 - S->postShift); /* Post shift */
q15_t *pState = S->pState; /* State pointer */
q15_t *pCoeffs = S->pCoeffs; /* Coefficient pointer */
q31_t *pState_q31; /* 32-bit state pointer for SIMD implementation */
uint32_t sample, stage = S->numStages; /* Stage loop counter */
do {
/* Initialize state pointer of type q31 */
pState_q31 = (q31_t *) (pState);
/* Read the b0 and 0 coefficients using SIMD */
b0 = *__SIMD32(pCoeffs)++;
/* Read the b1 and b2 coefficients using SIMD */
b1 = *__SIMD32(pCoeffs)++;
/* Read the a1 and a2 coefficients using SIMD */
a1 = *__SIMD32(pCoeffs)++;
/* Read the input state values from the state buffer: x[n-1], x[n-2] */
state_in = (q31_t) (*pState_q31++);
/* Read the output state values from the state buffer: y[n-1], y[n-2] */
state_out = (q31_t) (*pState_q31);
/* Apply loop unrolling and compute 2 output values simultaneously. */
/* The variables acc0 ... acc3 hold output values that are being computed:
*
* acc0 = b0 * x[n] + b1 * x[n-1] + b2 * x[n-2] + a1 * y[n-1] + a2 * y[n-2]
* acc0 = b0 * x[n] + b1 * x[n-1] + b2 * x[n-2] + a1 * y[n-1] + a2 * y[n-2]
*/
sample = blockSize >> 1u;
/* First part of the processing with loop unrolling. Compute 2 outputs at a time.
** a second loop below computes the remaining 1 sample. */
while(sample > 0u) {
/* Read the input */
in = *__SIMD32(pIn)++;
/* out = b0 * x[n] + 0 * 0 */
out = __SMUAD(b0, in);
/* acc0 = b1 * x[n-1] + acc0 += b2 * x[n-2] + out */
acc0 = __SMLAD(b1, state_in, out);
/* acc0 += a1 * y[n-1] + acc0 += a2 * y[n-2] */
acc0 = __SMLAD(a1, state_out, acc0);
/* The result is converted from 3.29 to 1.31 and then saturation is applied */
out = __SSAT((acc0 >> shift), 16);
/* Every time after the output is computed state should be updated. */
/* The states should be updated as: */
/* Xn2 = Xn1 */
/* Xn1 = Xn */
/* Yn2 = Yn1 */
/* Yn1 = acc0 */
/* x[n-N], x[n-N-1] are packed together to make state_in of type q31 */
/* y[n-N], y[n-N-1] are packed together to make state_out of type q31 */
#ifndef ARM_MATH_BIG_ENDIAN
state_in = __PKHBT(in, state_in, 16);
state_out = __PKHBT(out, state_out, 16);
#else
state_in = __PKHBT(state_in >> 16, (in >> 16), 16);
state_out = __PKHBT(state_out >> 16, (out), 16);
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */
/* out = b0 * x[n] + 0 * 0 */
out = __SMUADX(b0, in);
/* acc0 = b1 * x[n-1] + acc0 += b2 * x[n-2] + out */
acc0 = __SMLAD(b1, state_in, out);
/* acc0 += a1 * y[n-1] + acc0 += a2 * y[n-2] */
acc0 = __SMLAD(a1, state_out, acc0);
/* The result is converted from 3.29 to 1.31 and then saturation is applied */
out = __SSAT((acc0 >> shift), 16);
/* Store the output in the destination buffer. */
#ifndef ARM_MATH_BIG_ENDIAN
*__SIMD32(pOut)++ = __PKHBT(state_out, out, 16);
#else
*__SIMD32(pOut)++ = __PKHBT(out, state_out >> 16, 16);
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */
/* Every time after the output is computed state should be updated. */
/* The states should be updated as: */
/* Xn2 = Xn1 */
/* Xn1 = Xn */
/* Yn2 = Yn1 */
/* Yn1 = acc0 */
/* x[n-N], x[n-N-1] are packed together to make state_in of type q31 */
/* y[n-N], y[n-N-1] are packed together to make state_out of type q31 */
#ifndef ARM_MATH_BIG_ENDIAN
state_in = __PKHBT(in >> 16, state_in, 16);
state_out = __PKHBT(out, state_out, 16);
#else
state_in = __PKHBT(state_in >> 16, in, 16);
state_out = __PKHBT(state_out >> 16, out, 16);
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */
/* Decrement the loop counter */
sample--;
}
/* If the blockSize is not a multiple of 2, compute any remaining output samples here.
** No loop unrolling is used. */
if((blockSize & 0x1u) != 0u) {
/* Read the input */
in = *pIn++;
/* out = b0 * x[n] + 0 * 0 */
#ifndef ARM_MATH_BIG_ENDIAN
out = __SMUAD(b0, in);
#else
out = __SMUADX(b0, in);
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */
/* acc0 = b1 * x[n-1] + acc0 += b2 * x[n-2] + out */
acc0 = __SMLAD(b1, state_in, out);
/* acc0 += a1 * y[n-1] + acc0 += a2 * y[n-2] */
acc0 = __SMLAD(a1, state_out, acc0);
/* The result is converted from 3.29 to 1.31 and then saturation is applied */
out = __SSAT((acc0 >> shift), 16);
/* Store the output in the destination buffer. */
*pOut++ = (q15_t) out;
/* Every time after the output is computed state should be updated. */
/* The states should be updated as: */
/* Xn2 = Xn1 */
/* Xn1 = Xn */
/* Yn2 = Yn1 */
/* Yn1 = acc0 */
/* x[n-N], x[n-N-1] are packed together to make state_in of type q31 */
/* y[n-N], y[n-N-1] are packed together to make state_out of type q31 */
#ifndef ARM_MATH_BIG_ENDIAN
state_in = __PKHBT(in, state_in, 16);
state_out = __PKHBT(out, state_out, 16);
#else
state_in = __PKHBT(state_in >> 16, in, 16);
state_out = __PKHBT(state_out >> 16, out, 16);
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */
}
/* The first stage goes from the input buffer to the output buffer. */
/* Subsequent (numStages - 1) occur in-place in the output buffer */
pIn = pDst;
/* Reset the output pointer */
pOut = pDst;
/* Store the updated state variables back into the state array */
*__SIMD32(pState)++ = state_in;
*__SIMD32(pState)++ = state_out;
/* Decrement the loop counter */
stage--;
} while(stage > 0u);
/**
\brief Test case: TC_CoreSimd_ParMul16
\details
- Check Parallel 16-bit multiplication:
__SMLAD
__SMLADX
__SMLALD
__SMLALDX
__SMLSD
__SMLSDX
__SMLSLD
__SMLSLDX
__SMUAD
__SMUADX
__SMUSD
__SMUSDX
*/
void TC_CoreSimd_ParMul16 (void) {
#if ((defined (__ARM_ARCH_7EM__ ) && (__ARM_ARCH_7EM__ == 1)) || \
(defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)) )
volatile int32_t op1_s32, op2_s32, op3_s32;
volatile int32_t res_s32;
volatile int64_t op1_s64;
volatile int64_t res_s64;
/* --- __SMLAD Test ---------------------------------------------- */
op1_s32 = 0x00030002;
op2_s32 = 0x00050004;
op3_s32 = 0x20000000;
res_s32 = __SMLAD(op1_s32, op2_s32, op3_s32);
ASSERT_TRUE(res_s32 == 0x20000017);
/* --- __SMLADX Test ---------------------------------------------- */
op1_s32 = 0x00030002;
op2_s32 = 0x00050004;
op3_s32 = 0x00000800;
res_s32 = __SMLADX(op1_s32, op2_s32, op3_s32);
ASSERT_TRUE(res_s32 == 0x00000816);
/* --- __SMLALD Test ---------------------------------------------- */
op1_s32 = 0x00030002;
op2_s32 = 0x00050004;
op1_s64 = 0x00000000200000000LL;
res_s64 = __SMLALD(op1_s32, op2_s32, op1_s64);
ASSERT_TRUE(res_s64 == 0x0000000200000017LL);
/* --- __SMLALDX Test ---------------------------------------------- */
op1_s32 = 0x00030002;
op2_s32 = 0x00050004;
op1_s64 = 0x00000000200000000LL;
res_s64 = __SMLALDX(op1_s32, op2_s32, op1_s64);
ASSERT_TRUE(res_s64 == 0x0000000200000016LL);
/* --- __SMLSD Test ---------------------------------------------- */
op1_s32 = 0x00030006;
op2_s32 = 0x00050004;
op3_s32 = 0x00000800;
res_s32 = __SMLSD(op1_s32, op2_s32, op3_s32);
ASSERT_TRUE(res_s32 == 0x00000809);
/* --- __SMLSDX Test ---------------------------------------------- */
op1_s32 = 0x00030002;
op2_s32 = 0x00050004;
op3_s32 = 0x00000800;
res_s32 = __SMLSDX(op1_s32, op2_s32, op3_s32);
ASSERT_TRUE(res_s32 == 0x000007FE);
/* --- __SMLSLD Test ---------------------------------------------- */
op1_s32 = 0x00030006;
op2_s32 = 0x00050004;
op1_s64 = 0x00000000200000000LL;
res_s64 = __SMLSLD(op1_s32, op2_s32, op1_s64);
ASSERT_TRUE(res_s64 == 0x0000000200000009LL);
/* --- __SMLSLDX Test ---------------------------------------------- */
op1_s32 = 0x00030006;
op2_s32 = 0x00050004;
op1_s64 = 0x00000000200000000LL;
res_s64 = __SMLSLDX(op1_s32, op2_s32, op1_s64);
ASSERT_TRUE(res_s64 == 0x0000000200000012LL);
/* --- __SMUAD Test ---------------------------------------------- */
op1_s32 = 0x00030001;
op2_s32 = 0x00040002;
res_s32 = __SMUAD(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == 0x0000000E);
op1_s32 = (int32_t)0xFFFDFFFF;
op2_s32 = (int32_t)0x00040002;
res_s32 = __SMUAD(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == (int32_t)0xFFFFFFF2);
/* --- __SMUADX Test ---------------------------------------------- */
op1_s32 = 0x00030001;
op2_s32 = 0x00040002;
res_s32 = __SMUADX(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == 0x0000000A);
op1_s32 = (int32_t)0xFFFDFFFF;
op2_s32 = (int32_t)0x00040002;
res_s32 = __SMUADX(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == (int32_t)0xFFFFFFF6);
/* --- __SMUSD Test ---------------------------------------------- */
op1_s32 = (int32_t)0x00030001;
op2_s32 = (int32_t)0x00040002;
res_s32 = __SMUSD(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == (int32_t)0xFFFFFFF6);
op1_s32 = (int32_t)0xFFFDFFFF;
op2_s32 = (int32_t)0x00040002;
res_s32 = __SMUSD(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == 0x0000000A);
/* --- __SMUSDX Test ---------------------------------------------- */
op1_s32 = 0x00030001;
op2_s32 = 0x00040002;
res_s32 = __SMUSDX(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == (int32_t)0xFFFFFFFE);
op1_s32 = (int32_t)0xFFFDFFFF;
op2_s32 = (int32_t)0x00040002;
res_s32 = __SMUSDX(op1_s32,op2_s32);
ASSERT_TRUE(res_s32 == (int32_t)0x00000002);
#endif
}
