Correct conversion between float and fixed formats
This CL changes the way that values are converted
between fixed and floating point to
-Avoid the former asymmetric conversion causing
nonlinear distortions.
-Reduce the complexity.
Bug: webrtc:6594
Change-Id: I64d0cc31c5d16f397686a59a062cfbc4b336d94d
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/132783
Reviewed-by: Henrik Lundin <henrik.lundin@webrtc.org>
Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
Commit-Queue: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#28867}
diff --git a/common_audio/audio_util.cc b/common_audio/audio_util.cc
index 735ba5f..eb132ca 100644
--- a/common_audio/audio_util.cc
+++ b/common_audio/audio_util.cc
@@ -12,11 +12,6 @@
namespace webrtc {
-void FloatToS16(const float* src, size_t size, int16_t* dest) {
- for (size_t i = 0; i < size; ++i)
- dest[i] = FloatToS16(src[i]);
-}
-
void S16ToFloat(const int16_t* src, size_t size, float* dest) {
for (size_t i = 0; i < size; ++i)
dest[i] = S16ToFloat(src[i]);
diff --git a/common_audio/audio_util_unittest.cc b/common_audio/audio_util_unittest.cc
index cf85a2d..a215a12 100644
--- a/common_audio/audio_util_unittest.cc
+++ b/common_audio/audio_util_unittest.cc
@@ -31,25 +31,6 @@
}
}
-TEST(AudioUtilTest, FloatToS16) {
- static constexpr float kInput[] = {0.f,
- 0.4f / 32767.f,
- 0.6f / 32767.f,
- -0.4f / 32768.f,
- -0.6f / 32768.f,
- 1.f,
- -1.f,
- 1.1f,
- -1.1f};
- static constexpr int16_t kReference[] = {0, 0, 1, 0, -1,
- 32767, -32768, 32767, -32768};
- static constexpr size_t kSize = arraysize(kInput);
- static_assert(arraysize(kReference) == kSize, "");
- int16_t output[kSize];
- FloatToS16(kInput, kSize, output);
- ExpectArraysEq(kReference, output, kSize);
-}
-
TEST(AudioUtilTest, S16ToFloat) {
static constexpr int16_t kInput[] = {0, 1, -1, 16384, -16384, 32767, -32768};
static constexpr float kReference[] = {
@@ -74,16 +55,16 @@
TEST(AudioUtilTest, FloatToFloatS16) {
static constexpr float kInput[] = {0.f,
- 0.4f / 32767.f,
- 0.6f / 32767.f,
+ 0.4f / 32768.f,
+ 0.6f / 32768.f,
-0.4f / 32768.f,
-0.6f / 32768.f,
1.f,
-1.f,
- 1.1f,
- -1.1f};
+ 1.f,
+ -1.f};
static constexpr float kReference[] = {
- 0.f, 0.4f, 0.6f, -0.4f, -0.6f, 32767.f, -32768.f, 36043.7f, -36044.8f};
+ 0.f, 0.4f, 0.6f, -0.4f, -0.6f, 32768.f, -32768.f, 32768.f, -32768.f};
static constexpr size_t kSize = arraysize(kInput);
static_assert(arraysize(kReference) == kSize, "");
float output[kSize];
@@ -92,17 +73,17 @@
}
TEST(AudioUtilTest, FloatS16ToFloat) {
- static constexpr float kInput[] = {
- 0.f, 0.4f, 0.6f, -0.4f, -0.6f, 32767.f, -32768.f, 36043.7f, -36044.8f};
+ static constexpr float kInput[] = {0.f, 0.4f, 0.6f, -0.4f, -0.6f,
+ 32767.f, -32768.f, 32767.f, -32768.f};
static constexpr float kReference[] = {0.f,
- 0.4f / 32767.f,
- 0.6f / 32767.f,
+ 0.4f / 32768.f,
+ 0.6f / 32768.f,
-0.4f / 32768.f,
-0.6f / 32768.f,
1.f,
-1.f,
- 1.1f,
- -1.1f};
+ 1.f,
+ -1.f};
static constexpr size_t kSize = arraysize(kInput);
static_assert(arraysize(kReference) == kSize, "");
float output[kSize];
diff --git a/common_audio/include/audio_util.h b/common_audio/include/audio_util.h
index 50c9cf2..255abcc 100644
--- a/common_audio/include/audio_util.h
+++ b/common_audio/include/audio_util.h
@@ -27,45 +27,35 @@
// The conversion functions use the following naming convention:
// S16: int16_t [-32768, 32767]
// Float: float [-1.0, 1.0]
-// FloatS16: float [-32768.0, 32767.0]
+// FloatS16: float [-32768.0, 32768.0]
// Dbfs: float [-20.0*log(10, 32768), 0] = [-90.3, 0]
// The ratio conversion functions use this naming convention:
// Ratio: float (0, +inf)
// Db: float (-inf, +inf)
-static inline int16_t FloatToS16(float v) {
- if (v > 0)
- return v >= 1 ? limits_int16::max()
- : static_cast<int16_t>(v * limits_int16::max() + 0.5f);
- return v <= -1 ? limits_int16::min()
- : static_cast<int16_t>(-v * limits_int16::min() - 0.5f);
-}
-
static inline float S16ToFloat(int16_t v) {
- static const float kMaxInt16Inverse = 1.f / limits_int16::max();
- static const float kMinInt16Inverse = 1.f / limits_int16::min();
- return v * (v > 0 ? kMaxInt16Inverse : -kMinInt16Inverse);
+ constexpr float kScaling = 1.f / 32768.f;
+ return v * kScaling;
}
static inline int16_t FloatS16ToS16(float v) {
- static const float kMaxRound = limits_int16::max() - 0.5f;
- static const float kMinRound = limits_int16::min() + 0.5f;
- if (v > 0)
- return v >= kMaxRound ? limits_int16::max()
- : static_cast<int16_t>(v + 0.5f);
- return v <= kMinRound ? limits_int16::min() : static_cast<int16_t>(v - 0.5f);
+ v = std::min(v, 32767.f);
+ v = std::max(v, -32768.f);
+ return static_cast<int16_t>(v + std::copysign(0.5f, v));
}
static inline float FloatToFloatS16(float v) {
- return v * (v > 0 ? limits_int16::max() : -limits_int16::min());
+ RTC_DCHECK_LE(v, 1.f);
+ RTC_DCHECK_GE(v, -1.f);
+ return v * 32768.f;
}
static inline float FloatS16ToFloat(float v) {
- static const float kMaxInt16Inverse = 1.f / limits_int16::max();
- static const float kMinInt16Inverse = 1.f / limits_int16::min();
- return v * (v > 0 ? kMaxInt16Inverse : -kMinInt16Inverse);
+ RTC_DCHECK_LE(v, 32768.f);
+ RTC_DCHECK_GE(v, -32768.f);
+ constexpr float kScaling = 1.f / 32768.f;
+ return v * kScaling;
}
-void FloatToS16(const float* src, size_t size, int16_t* dest);
void S16ToFloat(const int16_t* src, size_t size, float* dest);
void FloatS16ToS16(const float* src, size_t size, int16_t* dest);
void FloatToFloatS16(const float* src, size_t size, float* dest);
diff --git a/common_audio/resampler/push_sinc_resampler_unittest.cc b/common_audio/resampler/push_sinc_resampler_unittest.cc
index 1a25a8c..f9943b3 100644
--- a/common_audio/resampler/push_sinc_resampler_unittest.cc
+++ b/common_audio/resampler/push_sinc_resampler_unittest.cc
@@ -33,6 +33,18 @@
return 20 * std::log10(x);
}
+void FloatToS16(const float* src, size_t size, int16_t* dest) {
+ for (size_t i = 0; i < size; ++i) {
+ RTC_DCHECK_GE(32767.f, src[i]);
+ RTC_DCHECK_LE(-32768.f, src[i]);
+ if (src[i] >= 1.f)
+ dest[i] = 32767;
+ if (src[i] <= -1.f)
+ dest[i] = -32768;
+ dest[i] = static_cast<int16_t>(src[i] * 32767.5f);
+ }
+}
+
} // namespace
class PushSincResamplerTest : public ::testing::TestWithParam<
@@ -322,14 +334,14 @@
::testing::make_tuple(32000, 16000, -18.48, -28.59),
::testing::make_tuple(44100, 16000, -19.30, -19.67),
::testing::make_tuple(48000, 16000, -19.81, -18.11),
- ::testing::make_tuple(96000, 16000, -20.95, -10.96),
+ ::testing::make_tuple(96000, 16000, -20.95, -10.9596),
// To 32 kHz
::testing::make_tuple(8000, 32000, kResamplingRMSError, -70.30),
::testing::make_tuple(16000, 32000, kResamplingRMSError, -75.51),
::testing::make_tuple(32000, 32000, kResamplingRMSError, -75.51),
- ::testing::make_tuple(44100, 32000, -16.44, -51.10),
- ::testing::make_tuple(48000, 32000, -16.90, -44.03),
+ ::testing::make_tuple(44100, 32000, -16.44, -51.0349),
+ ::testing::make_tuple(48000, 32000, -16.90, -43.9967),
::testing::make_tuple(96000, 32000, -19.61, -18.04),
::testing::make_tuple(192000, 32000, -21.02, -10.94)));
diff --git a/modules/audio_processing/echo_cancellation_bit_exact_unittest.cc b/modules/audio_processing/echo_cancellation_bit_exact_unittest.cc
index 80f36a8..d44483c 100644
--- a/modules/audio_processing/echo_cancellation_bit_exact_unittest.cc
+++ b/modules/audio_processing/echo_cancellation_bit_exact_unittest.cc
@@ -324,8 +324,8 @@
DISABLED_Stereo32kHz_HighLevel_NoDrift_StreamDelay0) {
#endif
#if defined(WEBRTC_MAC)
- const float kOutputReference[] = {-0.000458f, 0.000244f, 0.000153f,
- -0.000458f, 0.000244f, 0.000153f};
+ const float kOutputReference[] = {-0.000458f, 0.000214f, 0.000122f,
+ -0.000458f, 0.000214f, 0.000122f};
#else
const float kOutputReference[] = {-0.000427f, 0.000183f, 0.000183f,
-0.000427f, 0.000183f, 0.000183f};
diff --git a/modules/audio_processing/low_cut_filter_unittest.cc b/modules/audio_processing/low_cut_filter_unittest.cc
index ea4fb67..fb950da 100644
--- a/modules/audio_processing/low_cut_filter_unittest.cc
+++ b/modules/audio_processing/low_cut_filter_unittest.cc
@@ -110,9 +110,9 @@
0.349705f, 0.173054f, 0.016750f, -0.415957f, -0.461001f, -0.557111f,
0.738711f, 0.275720f};
- const float kReference[] = {0.142277f, -0.418518f, -0.028229f, -0.102112f,
- 0.141270f, 0.137791f, 0.124577f, -0.088715f,
- -0.142273f, -0.125885f, 0.266640f, -0.468079f};
+ const float kReference[] = {0.142273f, -0.418518f, -0.028229f, -0.102112f,
+ 0.141266f, 0.137787f, 0.124573f, -0.088715f,
+ -0.142273f, -0.125885f, 0.266663f, -0.468109f};
RunBitexactnessTest(
8000, 1, CreateVector(rtc::ArrayView<const float>(kReferenceInput)),
@@ -162,9 +162,9 @@
0.036391f, -0.408991f, 0.369330f, 0.399785f, -0.471419f, 0.551138f,
-0.307569f, 0.064315f, 0.311605f, 0.041736f, 0.650943f, 0.780496f};
- const float kReference[] = {-0.173553f, -0.265778f, 0.158757f, -0.259399f,
- -0.176361f, 0.192877f, 0.056825f, 0.171453f,
- 0.050752f, -0.194580f, -0.208679f, 0.153722f};
+ const float kReference[] = {-0.173584f, -0.265778f, 0.158783f, -0.259430f,
+ -0.176361f, 0.192841f, 0.056854f, 0.171448f,
+ 0.050751f, -0.194580f, -0.208710f, 0.153717f};
RunBitexactnessTest(
8000, 1, CreateVector(rtc::ArrayView<const float>(kReferenceInput)),
@@ -202,10 +202,10 @@
0.768778f, -0.122021f, 0.563445f, -0.703070f};
const float kReference[] = {
- 0.733329f, 0.084109f, 0.072695f, 0.566210f, -1.000000f, 0.652120f,
- -0.297424f, -0.964020f, 0.438551f, -0.698364f, -0.654449f, 0.266243f,
- 0.454115f, 0.684774f, -0.586823f, -0.747345f, -0.503021f, -0.222961f,
- -0.314972f, 0.907224f, -0.796265f, 0.284280f, -0.533417f, 0.773980f};
+ 0.733307f, 0.084106f, 0.072693f, 0.566193f, -1.000000f, 0.652130f,
+ -0.297424f, -0.964020f, 0.438568f, -0.698364f, -0.654449f, 0.266205f,
+ 0.454102f, 0.684784f, -0.586823f, -0.747375f, -0.503021f, -0.222961f,
+ -0.314972f, 0.907196f, -0.796295f, 0.284271f, -0.533417f, 0.773956f};
RunBitexactnessTest(
8000, 2, CreateVector(rtc::ArrayView<const float>(kReferenceInput)),
@@ -296,10 +296,10 @@
-0.034654f, -0.743470f, -0.494178f, 0.767923f, -0.607446f, -0.757293f};
const float kReference[] = {
- -0.544495f, 0.264199f, 0.647938f, 0.565569f, 0.496231f, 0.271340f,
- 0.519944f, 0.318094f, -0.792999f, 0.733421f, -1.000000f, 0.103977f,
- 0.981719f, 0.314859f, 0.476882f, 0.514267f, -0.196381f, -0.425781f,
- -0.783081f, 0.101108f, 0.419782f, -0.291718f, 0.183355f, -0.332489f};
+ -0.544525f, 0.264221f, 0.647919f, 0.565552f, 0.496185f, 0.271332f,
+ 0.519958f, 0.318085f, -0.792999f, 0.733429f, -1.000000f, 0.103973f,
+ 0.981720f, 0.314850f, 0.476837f, 0.514252f, -0.196411f, -0.425812f,
+ -0.783112f, 0.101105f, 0.419739f, -0.291718f, 0.183350f, -0.332489f};
RunBitexactnessTest(
8000, 2, CreateVector(rtc::ArrayView<const float>(kReferenceInput)),
@@ -669,10 +669,10 @@
0.462557f, 0.807713f, -0.095536f, -0.858625f, -0.517444f, 0.463730f};
const float kReference[] = {
- -0.816528f, 0.085421f, 0.739647f, -0.922089f, 0.669301f, -0.048187f,
- -0.290039f, -0.818085f, -0.596008f, -0.177826f, -0.002197f, -0.350647f,
- -0.064301f, 0.337291f, -0.621765f, 0.115909f, 0.311899f, -0.915924f,
- 0.020478f, 0.836055f, -0.714020f, -0.037140f, 0.391125f, -0.340118f};
+ -0.816559f, 0.085419f, 0.739655f, -0.922089f, 0.669312f, -0.048218f,
+ -0.290039f, -0.818085f, -0.596039f, -0.177856f, -0.002197f, -0.350647f,
+ -0.064331f, 0.337280f, -0.621765f, 0.115906f, 0.311890f, -0.915924f,
+ 0.020477f, 0.836029f, -0.714020f, -0.037140f, 0.391113f, -0.340118f};
RunBitexactnessTest(
16000, 2, CreateVector(rtc::ArrayView<const float>(kReferenceInput)),
diff --git a/modules/audio_processing/noise_suppression_unittest.cc b/modules/audio_processing/noise_suppression_unittest.cc
index b207485..29aae8b 100644
--- a/modules/audio_processing/noise_suppression_unittest.cc
+++ b/modules/audio_processing/noise_suppression_unittest.cc
@@ -106,10 +106,10 @@
7677.521973f};
const float kOutputReference[] = {0.003510f, 0.004517f, 0.004669f};
#else
- const float kSpeechProbabilityReference = 0.73421317f;
- const float kNoiseEstimateReference[] = {1175.266113f, 3289.305908f,
- 7532.991211f};
- const float kOutputReference[] = {0.003263f, 0.004402f, 0.004537f};
+ const float kSpeechProbabilityReference = 0.73650402f;
+ const float kNoiseEstimateReference[] = {1176.856812f, 3287.490967f,
+ 7525.964844f};
+ const float kOutputReference[] = {0.003306f, 0.004442f, 0.004574f};
#endif
RunBitexactnessTest(8000, 1, NoiseSuppression::Level::kLow,
@@ -129,9 +129,9 @@
14367.499023f};
const float kOutputReference[] = {0.003449f, 0.004334f, 0.004303f};
#else
- const float kSpeechProbabilityReference = 0.71672988f;
- const float kNoiseEstimateReference[] = {2151.313965f, 6509.765137f,
- 15658.848633f};
+ const float kSpeechProbabilityReference = 0.71743423f;
+ const float kNoiseEstimateReference[] = {2179.853027f, 6507.995117f,
+ 15652.758789f};
const float kOutputReference[] = {0.003574f, 0.004494f, 0.004499f};
#endif
@@ -166,18 +166,18 @@
TEST(NoiseSuppresionBitExactnessTest, Mono48kHzLow) {
#if defined(WEBRTC_ARCH_ARM64)
const float kSpeechProbabilityReference = -4.0f;
- const float kNoiseEstimateReference[] = {2564.605713f, 6213.656250f,
- 13372.284180f};
- const float kOutputReference[] = {-0.013185f, -0.012769f, -0.012023f};
+ const float kNoiseEstimateReference[] = {2135.292480f, 6692.695801f,
+ 14647.632812f};
+ const float kOutputReference[] = {-0.012738f, -0.012312f, -0.011576f};
#elif defined(WEBRTC_ARCH_ARM)
const float kSpeechProbabilityReference = -4.0f;
- const float kNoiseEstimateReference[] = {2564.605713f, 6213.656250f,
- 13372.284180f};
- const float kOutputReference[] = {-0.013185f, -0.012769f, -0.012023f};
+ const float kNoiseEstimateReference[] = {2135.292480f, 6692.695801f,
+ 14647.632812f};
+ const float kOutputReference[] = {-0.012738f, -0.012312f, -0.011576f};
#else
- const float kSpeechProbabilityReference = 0.70645678f;
- const float kNoiseEstimateReference[] = {2168.783203f, 6902.895508f,
- 13190.677734f};
+ const float kSpeechProbabilityReference = 0.70737761f;
+ const float kNoiseEstimateReference[] = {2187.394043f, 6913.306641f,
+ 13182.945312f};
const float kOutputReference[] = {-0.013062f, -0.012657f, -0.011934f};
#endif
@@ -200,9 +200,9 @@
const float kOutputReference[] = {-0.011108f, -0.007904f, -0.012390f,
-0.002472f, 0.000916f, -0.003235f};
#else
- const float kSpeechProbabilityReference = 0.67230678f;
- const float kNoiseEstimateReference[] = {9771.250000f, 11329.377930f,
- 10503.052734f};
+ const float kSpeechProbabilityReference = 0.67285913f;
+ const float kNoiseEstimateReference[] = {9753.257812f, 11515.603516f,
+ 10503.309570f};
const float kOutputReference[] = {-0.011459f, -0.008110f, -0.012728f,
-0.002399f, 0.001018f, -0.003189f};
#endif
@@ -224,9 +224,9 @@
16726.523438f};
const float kOutputReference[] = {0.004669f, 0.005615f, 0.005585f};
#else
- const float kSpeechProbabilityReference = 0.70897013f;
- const float kNoiseEstimateReference[] = {2171.490723f, 6553.567871f,
- 15626.562500f};
+ const float kSpeechProbabilityReference = 0.70916927f;
+ const float kNoiseEstimateReference[] = {2172.830566f, 6552.661133f,
+ 15624.025391f};
const float kOutputReference[] = {0.004513f, 0.005590f, 0.005614f};
#endif
@@ -247,9 +247,9 @@
16920.960938f};
const float kOutputReference[] = {0.004547f, 0.005432f, 0.005402f};
#else
- const float kSpeechProbabilityReference = 0.70106733f;
- const float kNoiseEstimateReference[] = {2224.968506f, 6712.025879f,
- 15785.087891f};
+ const float kSpeechProbabilityReference = 0.70104003f;
+ const float kNoiseEstimateReference[] = {2225.081055f, 6711.529785f,
+ 15785.949219};
const float kOutputReference[] = {0.004394f, 0.005406f, 0.005416f};
#endif
@@ -270,9 +270,9 @@
14365.744141f};
const float kOutputReference[] = {0.004273f, 0.005127f, 0.005188f};
#else
- const float kSpeechProbabilityReference = 0.70281971f;
- const float kNoiseEstimateReference[] = {2254.347900f, 6723.699707f,
- 15771.625977f};
+ const float kSpeechProbabilityReference = 0.70290041f;
+ const float kNoiseEstimateReference[] = {2254.921875f, 6723.172852f,
+ 15770.559570f};
const float kOutputReference[] = {0.004321f, 0.005247f, 0.005263f};
#endif
