mirrors/gnss-sdr
1
0
Fork 0
mirror of https://github.com/gnss-sdr/gnss-sdr synced 2025-11-04 17:23:20 +00:00
Code Issues Releases Wiki Activity

Implement 16ic multiply with RVV

Browse Source 
Implement multiplying two vectors of 16-bit complex numbers (where
each part, real and imaginary, is represented by 16 bits) by element
using RVV C intrinsics in `volk_gnsssdr_16ic_x2_multiply_16ic_rvv`.

Signed-off-by: Marcus Alagar <mvala079@gmail.com>
This commit is contained in:
Marcus Alagar
2025-08-09 12:16:54 -05:00
parent 15d828a897
commit b7940b1177
1 changed files with 57 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_multiply_16ic.h
View File

@@ -326,4 +326,61 @@ static inline void volk_gnsssdr_16ic_x2_multiply_16ic_neon(lv_16sc_t* out, const
}
#endif /* LV_HAVE_NEON*/
#ifdef LV_HAVE_RVV
#include <riscv_vector.h>
static inline void volk_gnsssdr_16ic_x2_multiply_16ic_rvv(lv_16sc_t* result, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
{
size_t n = num_points;
// Initialize pointers to keep track as stripmine
short* resPtr = (short*) result;
const short* aPtr = (const short*) in_a;
const short* bPtr = (const short*) in_b;
for (size_t vl; n > 0; n -= vl, resPtr += vl * 2, aPtr += vl * 2, bPtr += vl * 2)
{
// Record how many elements that will actually be processed
vl = __riscv_vsetvl_e16m4(n);
// Load aReal[0..vl), aImag[0..vl)
vint16m4x2_t aVal = __riscv_vlseg2e16_v_i16m4x2(aPtr, vl);
vint16m4_t aRealVal = __riscv_vget_v_i16m4x2_i16m4(aVal, 0);
vint16m4_t aImagVal = __riscv_vget_v_i16m4x2_i16m4(aVal, 1);
// Load bReal[0..vl), bImag[0..vl)
vint16m4x2_t bVal = __riscv_vlseg2e16_v_i16m4x2(bPtr, vl);
vint16m4_t bRealVal = __riscv_vget_v_i16m4x2_i16m4(bVal, 0);
vint16m4_t bImagVal = __riscv_vget_v_i16m4x2_i16m4(bVal, 1);
// resReal[i] = aReal[i] * bReal[i]
vint16m4_t resRealVal = __riscv_vmul_vv_i16m4(aRealVal, bRealVal, vl);
// resReal[i] = -(aImag[i] * bImag[i]) + resReal[i]
resRealVal = __riscv_vnmsac_vv_i16m4(resRealVal, aImagVal, bImagVal, vl);
// resImag[i] = aReal[i] * bImag[i]
vint16m4_t resImagVal = __riscv_vmul_vv_i16m4(aRealVal, bImagVal, vl);
// resImag[i] = (aImag[i] * bReal[i]) + resImag[i]
resImagVal = __riscv_vmacc_vv_i16m4(resImagVal, aImagVal, bRealVal, vl);
// Store resReal[0..vl), resImag[0..vl)
vint16m4x2_t resVal = __riscv_vset_v_i16m4_i16m4x2(
__riscv_vundefined_i16m4x2(), 0, resRealVal
);
resVal = __riscv_vset_v_i16m4_i16m4x2(resVal, 1, resImagVal);
__riscv_vsseg2e16_v_i16m4x2(resPtr, resVal, vl);
// In looping, decrement the number of
// elements left and increment the pointers
// by the numebr of elements processed,
// taking into account how the `vl` complex
// numbers are each stored as 2 `short`s
}
}
#endif /* LV_HAVE_RVV */
#endif /* INCLUDED_volk_gnsssdr_16ic_x2_multiply_16ic_H */