From f35fa998a6851a28f37865f32743855bea11d5a7 Mon Sep 17 00:00:00 2001 From: "Unknown W. Brackets" Date: Sun, 30 Jul 2017 11:46:08 -0700 Subject: [PATCH] Core: Handle 20-bit volumes in SIMD optimizations. It was previously only accounting for 17 bits of volume, by mistake. This was introduced in v1.0.0. May help #9379. --- Core/Util/AudioFormat.cpp | 24 ++++++++++++++------ Core/Util/AudioFormatNEON.cpp | 41 +++++++++++++++++++---------------- 2 files changed, 39 insertions(+), 26 deletions(-) diff --git a/Core/Util/AudioFormat.cpp b/Core/Util/AudioFormat.cpp index 4588726d13..24a76ed313 100644 --- a/Core/Util/AudioFormat.cpp +++ b/Core/Util/AudioFormat.cpp @@ -27,7 +27,7 @@ void AdjustVolumeBlockStandard(s16 *out, s16 *in, size_t size, int leftVol, int rightVol) { #ifdef _M_SSE - if (leftVol <= 0x7fff && rightVol <= 0x7fff) { + if (leftVol <= 0x7fff && -leftVol <= 0x8000 && rightVol <= 0x7fff && -rightVol <= 0x8000) { __m128i volume = _mm_set_epi16(leftVol, rightVol, leftVol, rightVol, leftVol, rightVol, leftVol, rightVol); while (size >= 16) { __m128i indata1 = _mm_loadu_si128((__m128i *)in); @@ -39,15 +39,25 @@ void AdjustVolumeBlockStandard(s16 *out, s16 *in, size_t size, int leftVol, int size -= 16; } } else { - // We have to shift inside the loop to avoid the signed multiply issue. - leftVol >>= 1; - rightVol >>= 1; - __m128i volume = _mm_set_epi16(leftVol, rightVol, leftVol, rightVol, leftVol, rightVol, leftVol, rightVol); + // We have to shift inside the loop to avoid the signed 16-bit multiply issue. + int leftShift = 0; + int leftVol16 = leftVol; + while (leftVol16 > 0x7fff || -leftVol16 > 0x8000) { + ++leftShift; + leftVol16 >>= 1; + } + int rightShift = 0; + int rightVol16 = rightVol; + while (rightVol16 > 0x7fff || -rightVol16 > 0x8000) { + ++rightShift; + rightVol16 >>= 1; + } + __m128i volume = _mm_set_epi16(leftVol16, rightVol16, leftVol16, rightVol16, leftVol16, rightVol16, leftVol16, rightVol16); while (size >= 16) { __m128i indata1 = _mm_loadu_si128((__m128i *)in); __m128i indata2 = _mm_loadu_si128((__m128i *)(in + 8)); - _mm_storeu_si128((__m128i *)out, _mm_slli_epi16(_mm_mulhi_epi16(indata1, volume), 1)); - _mm_storeu_si128((__m128i *)(out + 8), _mm_slli_epi16(_mm_mulhi_epi16(indata2, volume), 1)); + _mm_storeu_si128((__m128i *)out, _mm_slli_epi16(_mm_mulhi_epi16(indata1, volume), leftShift)); + _mm_storeu_si128((__m128i *)(out + 8), _mm_slli_epi16(_mm_mulhi_epi16(indata2, volume), rightShift)); in += 16; out += 16; size -= 16; diff --git a/Core/Util/AudioFormatNEON.cpp b/Core/Util/AudioFormatNEON.cpp index 0cb2a3e14a..90e262ac0e 100644 --- a/Core/Util/AudioFormatNEON.cpp +++ b/Core/Util/AudioFormatNEON.cpp @@ -30,29 +30,32 @@ static s16 MEMORY_ALIGNED16(volumeValues[4]) = {}; void AdjustVolumeBlockNEON(s16 *out, s16 *in, size_t size, int leftVol, int rightVol) { - volumeValues[0] = leftVol >> 1; - volumeValues[1] = rightVol >> 1; - volumeValues[2] = leftVol >> 1; - volumeValues[3] = rightVol >> 1; + if (leftVol <= 0xFFFF && -leftVol <= 0x10000 && rightVol <= 0xFFFF && -rightVol <= 0x10000) { + // Note: vqshrn_n_s32 takes a const argument, so we always go with 1 here, 15 there. + volumeValues[0] = leftVol >> 1; + volumeValues[1] = rightVol >> 1; + volumeValues[2] = leftVol >> 1; + volumeValues[3] = rightVol >> 1; - const int16x4_t vol = vld1_s16(volumeValues); + const int16x4_t vol = vld1_s16(volumeValues); - while (size >= 16) { - int16x8_t indata1 = vld1q_s16(in); - int16x8_t indata2 = vld1q_s16(in + 8); + while (size >= 16) { + int16x8_t indata1 = vld1q_s16(in); + int16x8_t indata2 = vld1q_s16(in + 8); - int32x4_t outh1 = vmull_s16(vget_high_s16(indata1), vol); - int32x4_t outh2 = vmull_s16(vget_high_s16(indata2), vol); - int32x4_t outl1 = vmull_s16(vget_low_s16(indata1), vol); - int32x4_t outl2 = vmull_s16(vget_low_s16(indata2), vol); + int32x4_t outh1 = vmull_s16(vget_high_s16(indata1), vol); + int32x4_t outh2 = vmull_s16(vget_high_s16(indata2), vol); + int32x4_t outl1 = vmull_s16(vget_low_s16(indata1), vol); + int32x4_t outl2 = vmull_s16(vget_low_s16(indata2), vol); - int16x8_t outdata1 = vcombine_s16(vqshrn_n_s32(outl1, 15), vqshrn_n_s32(outh1, 15)); - int16x8_t outdata2 = vcombine_s16(vqshrn_n_s32(outl2, 15), vqshrn_n_s32(outh2, 15)); - vst1q_s16(out, outdata1); - vst1q_s16(out + 8, outdata2); - in += 16; - out += 16; - size -= 16; + int16x8_t outdata1 = vcombine_s16(vqshrn_n_s32(outl1, 15), vqshrn_n_s32(outh1, 15)); + int16x8_t outdata2 = vcombine_s16(vqshrn_n_s32(outl2, 15), vqshrn_n_s32(outh2, 15)); + vst1q_s16(out, outdata1); + vst1q_s16(out + 8, outdata2); + in += 16; + out += 16; + size -= 16; + } } for (size_t i = 0; i < size; i += 2) {