search for: vld1_dup_f32

...s smart enough to figure that out). > + /* Just unroll the rest of the loop */ If you're not going to special case the last 2+1+1 samples, is there a measurable performance difference compared to simply looping? > + yi++; > + switch(len) { > + case 4: > + XX_2 = vld1_dup_f32(xi++); > + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); > + YY[0] = vld1q_f32(yi++); > + case 3: > + XX_2 = vld1_dup_f32(xi++); > + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); > + YY[0] = vld1q_f32(yi++); > + case 2: > + XX_2 = vld1_dup_f32...

Hi, Optimizes celt_pitch_xcorr for floating point. Changes from RFCv1: - Rebased on top of commit aad281878: Fix celt_pitch_xcorr_c signature. which got rid of ugly code around CELT_PITCH_XCORR_IMPL passing of "arch" parameter. - Unified with --enable-intrinsics used by x86 - Modified algorithm to be more in-line with algorithm in celt_pitch_xcorr_arm.s Viswanath Puttagunta

...MM, YEXT[1], vget_high_f32(XX[0]), 0); + YEXT[2] = vextq_f32(YY[0], YY[1], 3); + SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1); + + YY[0] = YY[1]; + len -= 4; + } + + /* Just unroll the rest of the loop */ + yi++; + switch(len) { + case 4: + XX_2 = vld1_dup_f32(xi++); + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); + YY[0] = vld1q_f32(yi++); + case 3: + XX_2 = vld1_dup_f32(xi++); + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); + YY[0] = vld1q_f32(yi++); + case 2: + XX_2 = vld1_dup_f32(xi++); + SUMM = vmlaq_lane_f32(SUM...

Hi, Optimizes celt_pitch_xcorr for floating point. Changes from RFCv2: - Changes recommended by Timothy for celt_neon_intr.c everything except, left the unrolled loop still unrolled - configure.ac - use AC_LINK_IFELSE instead of AC_COMPILE_IFELSE - Moved compile flags into Makefile.am - OPUS_ARM_NEON_INR --> typo --> OPUS_ARM_NEON_INTR Viswanath Puttagunta (1): armv7:

Hi, Optimizes celt_pitch_xcorr for ARM NEON floating point. Changes from RFCv3: - celt_neon_intr.c - removed warnings due to not having constant pointers - Put simpler loop to take care of corner cases. Unrolling using intrinsics was not really mapping well to what was done in celt_pitch_xcorr_arm.s - Makefile.am Removed explicit -O3 optimization - test_unit_mathops.c,

...Y[0], YY[1], 2); + SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[0]), 0); + YEXT[2] = vextq_f32(YY[0], YY[1], 3); + SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1); + + YY[0] = YY[1]; + len -= 4; + } + + yi++; + while (len > 1) { + XX_2 = vld1_dup_f32(xi++); + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); + YY[0]= vld1q_f32(yi++); + len--; + } + + if (len > 0) { + XX_2 = vld1_dup_f32(xi); + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); + } + + vst1q_f32(sum, SUMM); +} + +/* + * Function: xcorr_kernel_neon_float_...

...MM, YEXT[1], vget_high_f32(XX[0]), 0); + YEXT[2] = vextq_f32(YY[0], YY[1], 3); + SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1); + + YY[0] = YY[1]; + len -= 4; + } + + /* Just unroll the rest of the loop */ + yi++; + switch(len) { + case 4: + XX_2 = vld1_dup_f32(xi++); + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); + YY[0] = vld1q_f32(yi++); + case 3: + XX_2 = vld1_dup_f32(xi++); + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); + YY[0] = vld1q_f32(yi++); + case 2: + XX_2 = vld1_dup_f32(xi++); + SUMM = vmlaq_lane_f32(SUM...

From: Viswanath Puttagunta <viswanath.puttagunta at linaro.org> Hi, Optimizes celt_pitch_xcorr for floating point. Changes from RFCv1: - Rebased on top of commit aad281878: Fix celt_pitch_xcorr_c signature. which got rid of ugly code around CELT_PITCH_XCORR_IMPL passing of "arch" parameter. - Unified with --enable-intrinsics used by x86 - Modified algorithm to be more

..., YEXT[1], vget_high_f32(XX[0]), 0); > + YEXT[2] = vextq_f32(YY[0], YY[1], 3); > + SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1); > + > + YY[0] = YY[1]; > + len -= 4; > + } > + > + yi++; > + while (len > 1) { > + XX_2 = vld1_dup_f32(xi++); > + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); > + YY[0]= vld1q_f32(yi++); > + len--; > + } > + > + if (len > 0) { > + XX_2 = vld1_dup_f32(xi); > + SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); > + } Hi Timothy, After analyzing c...

...3 */ > + for (j = 0; j < cr; j++) { > + YY[0] = vld1q_f32(yi++); This load is always redundant in the first iteration, which is a bit unfortunate. > + XX_2 = vld1_lane_f32(xi++, XX_2, 0); Don't load a single lane when you don't need the value(s) in the other lane(s). Use vld1_dup_f32() instead. It's faster and breaks dependencies. > + SUMM[0] = vmlaq_lane_f32(SUMM[0], YY[0], XX_2, 0); > + } > + > + SUMM[0] = vaddq_f32(SUMM[0], SUMM[1]); > + SUMM[2] = vaddq_f32(SUMM[2], SUMM[3]); > + SUMM[0] = vaddq_f32(SUMM[0], SUMM[2]); > + > + vst1q_f32(sum, SUMM...

...ave to keep this scaling here). Ideally they'd be related since they refer to the same memory (e.g., scaled and scaledp or something). Also, float is _not_ compatible with float32_t (which is what vst1q_f32 takes) in all compiler versions. Please do not mix and match them. > + scale = vld1_dup_f32(&st->scale); Needs a (const float32_t *) cast. > + for (i = 0; i < N2; i++) { > + inq = vld1q_f32(in); > + in += 4; > + outq = vmulq_lane_f32(inq, scale, 0); > + vst1q_f32(out, outq); > + out += 4; > + } > + > + cfg->buffer =...

...YY[0] = vld1q_f32(yi++); > > This load is always redundant in the first iteration, which is a bit > unfortunate. > >> + XX_2 = vld1_lane_f32(xi++, XX_2, 0); > > Don't load a single lane when you don't need the value(s) in the other > lane(s). Use vld1_dup_f32() instead. It's faster and breaks dependencies. Will keep this in mind. Thanks. > >> + SUMM[0] = vmlaq_lane_f32(SUMM[0], YY[0], XX_2, 0); >> + } >> + >> + SUMM[0] = vaddq_f32(SUMM[0], SUMM[1]); >> + SUMM[2] = vaddq_f32(SUMM[2], SUMM[3]);...

...). > Ideally they'd be related since they refer to the same memory (e.g., scaled > and scaledp or something). > > Also, float is _not_ compatible with float32_t (which is what vst1q_f32 > takes) in all compiler versions. Please do not mix and match them. > >> + scale = vld1_dup_f32(&st->scale); > > > Needs a (const float32_t *) cast. > >> + for (i = 0; i < N2; i++) { >> + inq = vld1q_f32(in); >> + in += 4; >> + outq = vmulq_lane_f32(inq, scale, 0); >> + vst1q_f32(out, outq); >> + out += 4; &g...

Should call celt_inner_prod(). --- celt/bands.c | 7 ++++--- celt/bands.h | 2 +- celt/celt_encoder.c | 6 +++--- celt/pitch.c | 2 +- src/opus_multistream_encoder.c | 2 +- 5 files changed, 10 insertions(+), 9 deletions(-) diff --git a/celt/bands.c b/celt/bands.c index bbe8a4c..1ab24aa 100644 --- a/celt/bands.c +++ b/celt/bands.c

Hello, I received feedback from engineers working on NE10 [1] that it would be better to use NE10 [1] for FFT optimizations for opus use cases. However, these FFT patches are currently in review and haven't been integrated into NE10 yet. While the FFT functions in NE10 are getting baked, I wanted to optimize the celt_pitch_xcorr (floating point only) and use it to introduce ARM NEON

Hello opus-dev, I've been cooking up this patchset to integrate NE10 library into opus. Current patchset focuses on encode use case mainly effecting performance of clt_mdct_forward() and opus_fft() (for float only) Glad to report the following on Encode use case: (Measured on my Beaglebone Black Cortex-A8 board) - Performance improvement for encode use case ~= 12.34% (Based on time -p

...ACK; + int N2 = st->nfft >> 1; + float32x4_t inq, outq; + float32x2_t scale; + float *in = (float *)fin; + float *out; + int i; + ALLOC(temp, st->nfft, ne10_fft_cpx_float32_t); + ALLOC(tempin, st->nfft, ne10_fft_cpx_float32_t); + + out = (float *)tempin; + scale = vld1_dup_f32(&st->scale); + for (i = 0; i < N2; i++) { + inq = vld1q_f32(in); + in += 4; + outq = vmulq_lane_f32(inq, scale, 0); + vst1q_f32(out, outq); + out += 4; + } + + cfg->buffer = (ne10_fft_cpx_float32_t *)&temp[0]; + + ne10_fft_c2c_1d_float32_neon((ne10_...