CMSIS-DSP: Corrected tail management bug in MVE version of arm_fir_f32.c

pull/19/head
GuentherMartin 6 years ago
parent 78d7e66bc9
commit f51e95f5a8

@ -188,7 +188,9 @@ static void arm_fir_f32_1_4_mve(const arm_fir_instance_f32 * S, const float32_t
blkCnt--; blkCnt--;
} }
blkCnt = blockSize & 3; blkCnt = blockSize & 3;
if (blkCnt > 0U)
{ {
mve_pred16_t p0 = vctp32q(blkCnt); mve_pred16_t p0 = vctp32q(blkCnt);
@ -217,6 +219,7 @@ static void arm_fir_f32_1_4_mve(const arm_fir_instance_f32 * S, const float32_t
*/ */
pTempSrc = &S->pState[blockSize]; pTempSrc = &S->pState[blockSize];
pTempDest = S->pState; pTempDest = S->pState;
blkCnt = numTaps >> 2; blkCnt = numTaps >> 2;
while (blkCnt > 0U) while (blkCnt > 0U)
{ {
@ -225,6 +228,7 @@ static void arm_fir_f32_1_4_mve(const arm_fir_instance_f32 * S, const float32_t
pTempDest += 4; pTempDest += 4;
blkCnt--; blkCnt--;
} }
blkCnt = numTaps & 3; blkCnt = numTaps & 3;
if (blkCnt > 0U) if (blkCnt > 0U)
{ {
@ -234,7 +238,6 @@ static void arm_fir_f32_1_4_mve(const arm_fir_instance_f32 * S, const float32_t
} }
static void arm_fir_f32_5_8_mve(const arm_fir_instance_f32 * S, const float32_t * pSrc, float32_t * pDst, uint32_t blockSize) static void arm_fir_f32_5_8_mve(const arm_fir_instance_f32 * S, const float32_t * pSrc, float32_t * pDst, uint32_t blockSize)
{ {
float32_t *pState = S->pState; /* State pointer */ float32_t *pState = S->pState; /* State pointer */
@ -314,9 +317,8 @@ static void arm_fir_f32_5_8_mve(const arm_fir_instance_f32 * S, const float32_t
} }
blkCnt = blockSize & 3; blkCnt = blockSize & 3;
if (blkCnt > 0) if (blkCnt > 0U)
{ {
mve_pred16_t p0 = vctp32q(blkCnt); mve_pred16_t p0 = vctp32q(blkCnt);
vstrwq_p_f32(pStateCur, vld1q(pTempSrc),p0); vstrwq_p_f32(pStateCur, vld1q(pTempSrc),p0);
@ -364,6 +366,7 @@ static void arm_fir_f32_5_8_mve(const arm_fir_instance_f32 * S, const float32_t
pTempDest += 4; pTempDest += 4;
blkCnt--; blkCnt--;
} }
blkCnt = numTaps & 3; blkCnt = numTaps & 3;
if (blkCnt > 0U) if (blkCnt > 0U)
{ {
@ -372,6 +375,7 @@ static void arm_fir_f32_5_8_mve(const arm_fir_instance_f32 * S, const float32_t
} }
} }
void arm_fir_f32( void arm_fir_f32(
const arm_fir_instance_f32 * S, const arm_fir_instance_f32 * S,
const float32_t * pSrc, const float32_t * pSrc,
@ -413,9 +417,6 @@ uint32_t blockSize)
} }
} }
if (blockSize >= 8) if (blockSize >= 8)
{ {
/* /*
@ -532,6 +533,7 @@ uint32_t blockSize)
} }
blkCnt = blockSize & 3; blkCnt = blockSize & 3;
if (blkCnt > 0U)
{ {
mve_pred16_t p0 = vctp32q(blkCnt); mve_pred16_t p0 = vctp32q(blkCnt);
int32_t i; int32_t i;
@ -628,7 +630,6 @@ uint32_t blockSize)
numCnt --; numCnt --;
} }
vstrwq_p_f32(pOutput, vecAcc0, p0); vstrwq_p_f32(pOutput, vecAcc0, p0);
} }
} }
@ -641,6 +642,7 @@ uint32_t blockSize)
uint32_t numTaps = S->numTaps; /* Number of filter coefficients in the filter */ uint32_t numTaps = S->numTaps; /* Number of filter coefficients in the filter */
uint32_t i, blkCnt; /* Loop counters */ uint32_t i, blkCnt; /* Loop counters */
pStateCurnt = &(S->pState[(numTaps - 1U)]); pStateCurnt = &(S->pState[(numTaps - 1U)]);
blkCnt = blockSize; blkCnt = blockSize;
while (blkCnt > 0U) while (blkCnt > 0U)
{ {
@ -668,7 +670,6 @@ uint32_t blockSize)
} }
/* Store result in destination buffer. */ /* Store result in destination buffer. */
*pDst++ = acc0; *pDst++ = acc0;
@ -694,6 +695,7 @@ uint32_t blockSize)
pTempDest += 4; pTempDest += 4;
blkCnt--; blkCnt--;
} }
blkCnt = numTaps & 3; blkCnt = numTaps & 3;
if (blkCnt > 0U) if (blkCnt > 0U)
{ {

Loading…
Cancel
Save