123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348 |
- #ifndef _ARM_UTILS_HELIUM_H_
- #define _ARM_UTILS_HELIUM_H_
- #if defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEF) || defined(ARM_MATH_MVEI)
- #define INACTIVELANE 0
- #endif
- #if defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEF)
- __STATIC_FORCEINLINE float32_t vecAddAcrossF32Mve(float32x4_t in)
- {
- float32_t acc;
- acc = vgetq_lane(in, 0) + vgetq_lane(in, 1) +
- vgetq_lane(in, 2) + vgetq_lane(in, 3);
- return acc;
- }
- #define INVSQRT_MAGIC_F32 0x5f3759df
- #define INVSQRT_NEWTON_MVE_F32(invSqrt, xHalf, xStart)\
- { \
- float32x4_t tmp; \
- \
- \
- tmp = vmulq(xStart, xStart); \
- tmp = vmulq(tmp, xHalf); \
- \
- tmp = vsubq(vdupq_n_f32(1.5f), tmp); \
- \
- invSqrt = vmulq(tmp, xStart); \
- }
- #endif
- #if defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEI)
- #include "arm_common_tables.h"
- __STATIC_INLINE arm_status arm_mat_trans_32bit_2x2_mve(
- uint32_t * pDataSrc,
- uint32_t * pDataDest)
- {
- static const uint32x4_t vecOffs = { 0, 2, 1, 3 };
-
- uint32x4_t vecIn = vldrwq_u32((uint32_t const *)pDataSrc);
- vstrwq_scatter_shifted_offset_u32(pDataDest, vecOffs, vecIn);
- return (ARM_MATH_SUCCESS);
- }
- __STATIC_INLINE arm_status arm_mat_trans_32bit_3x3_mve(
- uint32_t * pDataSrc,
- uint32_t * pDataDest)
- {
- const uint32x4_t vecOffs1 = { 0, 3, 6, 1};
- const uint32x4_t vecOffs2 = { 4, 7, 2, 5};
-
- uint32x4_t vecIn1 = vldrwq_u32((uint32_t const *) pDataSrc);
- uint32x4_t vecIn2 = vldrwq_u32((uint32_t const *) &pDataSrc[4]);
- vstrwq_scatter_shifted_offset_u32(pDataDest, vecOffs1, vecIn1);
- vstrwq_scatter_shifted_offset_u32(pDataDest, vecOffs2, vecIn2);
- pDataDest[8] = pDataSrc[8];
- return (ARM_MATH_SUCCESS);
- }
- __STATIC_INLINE arm_status arm_mat_trans_32bit_4x4_mve(uint32_t * pDataSrc, uint32_t * pDataDest)
- {
-
- uint32x4x4_t vecIn;
- vecIn = vld4q((uint32_t const *) pDataSrc);
- vstrwq(pDataDest, vecIn.val[0]);
- pDataDest += 4;
- vstrwq(pDataDest, vecIn.val[1]);
- pDataDest += 4;
- vstrwq(pDataDest, vecIn.val[2]);
- pDataDest += 4;
- vstrwq(pDataDest, vecIn.val[3]);
- return (ARM_MATH_SUCCESS);
- }
- __STATIC_INLINE arm_status arm_mat_trans_32bit_generic_mve(
- uint16_t srcRows,
- uint16_t srcCols,
- uint32_t * pDataSrc,
- uint32_t * pDataDest)
- {
- uint32x4_t vecOffs;
- uint32_t i;
- uint32_t blkCnt;
- uint32_t const *pDataC;
- uint32_t *pDataDestR;
- uint32x4_t vecIn;
- vecOffs = vidupq_u32((uint32_t)0, 1);
- vecOffs = vecOffs * srcCols;
- i = srcCols;
- do
- {
- pDataC = (uint32_t const *) pDataSrc;
- pDataDestR = pDataDest;
- blkCnt = srcRows >> 2;
- while (blkCnt > 0U)
- {
- vecIn = vldrwq_gather_shifted_offset_u32(pDataC, vecOffs);
- vstrwq(pDataDestR, vecIn);
- pDataDestR += 4;
- pDataC = pDataC + srcCols * 4;
-
- blkCnt--;
- }
-
- blkCnt = srcRows & 3;
- if (blkCnt > 0U)
- {
- mve_pred16_t p0 = vctp32q(blkCnt);
- vecIn = vldrwq_gather_shifted_offset_u32(pDataC, vecOffs);
- vstrwq_p(pDataDestR, vecIn, p0);
- }
- pDataSrc += 1;
- pDataDest += srcRows;
- }
- while (--i);
- return (ARM_MATH_SUCCESS);
- }
- #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_FAST_SQRT_Q31_MVE)
- __STATIC_INLINE q31x4_t FAST_VSQRT_Q31(q31x4_t vecIn)
- {
- q63x2_t vecTmpLL;
- q31x4_t vecTmp0, vecTmp1;
- q31_t scale;
- q63_t tmp64;
- q31x4_t vecNrm, vecDst, vecIdx, vecSignBits;
- vecSignBits = vclsq(vecIn);
- vecSignBits = vbicq(vecSignBits, 1);
-
- vecNrm = vshlq(vecIn, vecSignBits);
-
- vecIdx = vecNrm >> 24;
- vecIdx = vecIdx << 1;
- vecTmp0 = vldrwq_gather_shifted_offset_s32(sqrtTable_Q31, vecIdx);
- vecIdx = vecIdx + 1;
- vecTmp1 = vldrwq_gather_shifted_offset_s32(sqrtTable_Q31, vecIdx);
- vecTmp1 = vqrdmulhq(vecTmp1, vecNrm);
- vecTmp0 = vecTmp0 - vecTmp1;
- vecTmp1 = vqrdmulhq(vecTmp0, vecTmp0);
- vecTmp1 = vqrdmulhq(vecNrm, vecTmp1);
- vecTmp1 = vdupq_n_s32(0x18000000) - vecTmp1;
- vecTmp0 = vqrdmulhq(vecTmp0, vecTmp1);
- vecTmpLL = vmullbq_int(vecNrm, vecTmp0);
-
- scale = 26 + (vecSignBits[0] >> 1);
- tmp64 = asrl(vecTmpLL[0], scale);
- vecDst[0] = (q31_t) tmp64;
- scale = 26 + (vecSignBits[2] >> 1);
- tmp64 = asrl(vecTmpLL[1], scale);
- vecDst[2] = (q31_t) tmp64;
- vecTmpLL = vmulltq_int(vecNrm, vecTmp0);
-
- scale = 26 + (vecSignBits[1] >> 1);
- tmp64 = asrl(vecTmpLL[0], scale);
- vecDst[1] = (q31_t) tmp64;
- scale = 26 + (vecSignBits[3] >> 1);
- tmp64 = asrl(vecTmpLL[1], scale);
- vecDst[3] = (q31_t) tmp64;
-
- vecDst = vdupq_m(vecDst, 0, vcmpltq_n_s32(vecIn, 0));
- return vecDst;
- }
- #endif
- #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_FAST_SQRT_Q15_MVE)
- __STATIC_INLINE q15x8_t FAST_VSQRT_Q15(q15x8_t vecIn)
- {
- q31x4_t vecTmpLev, vecTmpLodd, vecSignL;
- q15x8_t vecTmp0, vecTmp1;
- q15x8_t vecNrm, vecDst, vecIdx, vecSignBits;
- vecDst = vuninitializedq_s16();
- vecSignBits = vclsq(vecIn);
- vecSignBits = vbicq(vecSignBits, 1);
-
- vecNrm = vshlq(vecIn, vecSignBits);
- vecIdx = vecNrm >> 8;
- vecIdx = vecIdx << 1;
- vecTmp0 = vldrhq_gather_shifted_offset_s16(sqrtTable_Q15, vecIdx);
- vecIdx = vecIdx + 1;
- vecTmp1 = vldrhq_gather_shifted_offset_s16(sqrtTable_Q15, vecIdx);
- vecTmp1 = vqrdmulhq(vecTmp1, vecNrm);
- vecTmp0 = vecTmp0 - vecTmp1;
- vecTmp1 = vqrdmulhq(vecTmp0, vecTmp0);
- vecTmp1 = vqrdmulhq(vecNrm, vecTmp1);
- vecTmp1 = vdupq_n_s16(0x1800) - vecTmp1;
- vecTmp0 = vqrdmulhq(vecTmp0, vecTmp1);
- vecSignBits = vecSignBits >> 1;
- vecTmpLev = vmullbq_int(vecNrm, vecTmp0);
- vecTmpLodd = vmulltq_int(vecNrm, vecTmp0);
- vecTmp0 = vecSignBits + 10;
-
- vecTmp0 = -vecTmp0;
-
- vecSignL = vmovlbq(vecTmp0);
- vecTmpLev = vshlq(vecTmpLev, vecSignL);
-
- vecSignL = vmovltq(vecTmp0);
- vecTmpLodd = vshlq(vecTmpLodd, vecSignL);
-
- vecDst = vmovnbq_s32(vecDst, vecTmpLev);
- vecDst = vmovntq_s32(vecDst, vecTmpLodd);
-
- vecDst = vdupq_m(vecDst, 0, vcmpltq_n_s16(vecIn, 0));
- return vecDst;
- }
- #endif
- #endif
- #endif
|