| #include "edge-impulse-sdk/dsp/config.hpp" |
| #if EIDSP_LOAD_CMSIS_DSP_SOURCES |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| #include "edge-impulse-sdk/CMSIS/DSP/Include/dsp/statistics_functions_f16.h" |
|
|
| #if defined(ARM_FLOAT16_SUPPORTED) |
|
|
|
|
| #if (defined(ARM_MATH_NEON) || defined(ARM_MATH_MVEF)) && !defined(ARM_MATH_AUTOVECTORIZE) |
| #include <limits.h> |
| #endif |
|
|
|
|
| |
| |
| |
|
|
| |
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| #if defined(ARM_MATH_MVE_FLOAT16) && !defined(ARM_MATH_AUTOVECTORIZE) |
|
|
| #include "edge-impulse-sdk/CMSIS/DSP/Include/arm_helium_utils.h" |
| void arm_absmin_f16( |
| const float16_t * pSrc, |
| uint32_t blockSize, |
| float16_t * pResult, |
| uint32_t * pIndex) |
| { |
| uint16_t blkCnt; |
| f16x8_t vecSrc; |
| float16_t const *pSrcVec; |
| f16x8_t curExtremValVec = vdupq_n_f16(F16_ABSMAX); |
| float16_t minValue = F16_ABSMAX; |
| uint16_t idx = blockSize; |
| uint16x8_t indexVec; |
| uint16x8_t curExtremIdxVec; |
| mve_pred16_t p0; |
|
|
|
|
| indexVec = vidupq_u16((uint32_t)0, 1); |
| curExtremIdxVec = vdupq_n_u16(0); |
|
|
| pSrcVec = (float16_t const *) pSrc; |
| blkCnt = blockSize >> 3; |
| while (blkCnt > 0U) |
| { |
| vecSrc = vldrhq_f16(pSrcVec); |
| pSrcVec += 8; |
| vecSrc = vabsq(vecSrc); |
| |
| |
| |
| |
| p0 = vcmpleq(vecSrc, curExtremValVec); |
| curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); |
| curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); |
|
|
| indexVec = indexVec + 8; |
| |
| |
| |
| blkCnt--; |
| } |
| |
| |
| |
| |
| blkCnt = blockSize & 7; |
| if (blkCnt > 0U) |
| { |
| p0 = vctp16q(blkCnt); |
|
|
| vecSrc = vldrhq_f16(pSrcVec); |
| pSrcVec += 8; |
| vecSrc = vabsq(vecSrc); |
| |
| |
| |
| |
| p0 = vcmpleq_m(vecSrc, curExtremValVec, p0); |
| curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); |
| curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); |
| } |
| |
| |
| |
| minValue = vminnmvq(minValue, curExtremValVec); |
| |
| |
| |
| p0 = vcmpleq(curExtremValVec, minValue); |
| indexVec = vpselq(curExtremIdxVec, vdupq_n_u16(blockSize), p0); |
| |
| |
| |
| idx = vminvq(idx, indexVec); |
| |
| |
| |
| *pIndex = idx; |
| *pResult = minValue; |
| } |
|
|
| #else |
| #if defined(ARM_MATH_LOOPUNROLL) |
| void arm_absmin_f16( |
| const float16_t * pSrc, |
| uint32_t blockSize, |
| float16_t * pResult, |
| uint32_t * pIndex) |
| { |
| float16_t cur_absmin, out; \ |
| uint32_t blkCnt, outIndex; \ |
| uint32_t index; \ |
| \ |
| \ |
| outIndex = 0U; \ |
| \ |
| out = *pSrc++; \ |
| out = ((_Float16)out > 0.0f16) ? out : -(_Float16)out; \ |
| \ |
| index = 0U; \ |
| \ |
| \ |
| blkCnt = (blockSize - 1U) >> 2U; \ |
| \ |
| while (blkCnt > 0U) \ |
| { \ |
| \ |
| cur_absmin = *pSrc++; \ |
| cur_absmin = ((_Float16)cur_absmin > 0.0f16) ? cur_absmin : -(_Float16)cur_absmin; \ |
| \ |
| if ((_Float16)cur_absmin < (_Float16)out) \ |
| { \ |
| \ |
| out = cur_absmin; \ |
| outIndex = index + 1U; \ |
| } \ |
| \ |
| cur_absmin = *pSrc++; \ |
| cur_absmin = ((_Float16)cur_absmin > 0.0f16) ? cur_absmin : -(_Float16)cur_absmin; \ |
| if ((_Float16)cur_absmin < (_Float16)out) \ |
| { \ |
| out = cur_absmin; \ |
| outIndex = index + 2U; \ |
| } \ |
| \ |
| cur_absmin = *pSrc++; \ |
| cur_absmin = ((_Float16)cur_absmin > 0.0f16) ? cur_absmin : -(_Float16)cur_absmin; \ |
| if ((_Float16)cur_absmin < (_Float16)out) \ |
| { \ |
| out = cur_absmin; \ |
| outIndex = index + 3U; \ |
| } \ |
| \ |
| cur_absmin = *pSrc++; \ |
| cur_absmin = ((_Float16)cur_absmin > 0.0f16) ? cur_absmin : -(_Float16)cur_absmin; \ |
| if ((_Float16)cur_absmin < (_Float16)out) \ |
| { \ |
| out = cur_absmin; \ |
| outIndex = index + 4U; \ |
| } \ |
| \ |
| index += 4U; \ |
| \ |
| \ |
| blkCnt--; \ |
| } \ |
| \ |
| \ |
| blkCnt = (blockSize - 1U) % 4U; \ |
| \ |
| \ |
| while (blkCnt > 0U) \ |
| { \ |
| cur_absmin = *pSrc++; \ |
| cur_absmin = ((_Float16)cur_absmin > 0.0f16) ? cur_absmin : -(_Float16)cur_absmin; \ |
| if ((_Float16)cur_absmin < (_Float16)out) \ |
| { \ |
| out = cur_absmin; \ |
| outIndex = blockSize - blkCnt; \ |
| } \ |
| \ |
| \ |
| blkCnt--; \ |
| } \ |
| \ |
| \ |
| *pResult = out; \ |
| *pIndex = outIndex; |
| } |
| #else |
| void arm_absmin_f16( |
| const float16_t * pSrc, |
| uint32_t blockSize, |
| float16_t * pResult, |
| uint32_t * pIndex) |
| { |
| float16_t minVal, out; |
| uint32_t blkCnt, outIndex; |
|
|
| |
| outIndex = 0U; |
|
|
| |
| out = (_Float16)fabsf((float32_t)*pSrc++); |
|
|
| |
| blkCnt = (blockSize - 1U); |
|
|
| while (blkCnt > 0U) |
| { |
| |
| minVal = (_Float16)fabsf((float32_t)*pSrc++); |
|
|
| |
| if ((_Float16)out > (_Float16)minVal) |
| { |
| |
| out = minVal; |
| outIndex = blockSize - blkCnt; |
| } |
|
|
| |
| blkCnt--; |
| } |
|
|
| |
| *pResult = out; |
| *pIndex = outIndex; |
| } |
| #endif |
| #endif |
| |
| |
| |
|
|
| #endif |
|
|
|
|
| #endif |
|
|