SSE2 optimization for a-rate Oscillator

Add SSE2 optimization for the main processing loop for the a-rate
Oscillator and for the computation in WaveDataForFundamentalFrequency.
Profiling showed WaveDataForFundamentalFrequency was taking 27% of the
total time.  The optimized version takes about half as long by computing
4 results at a time.

The failed test is caused by changing the linear interpolation formula
from (1-f)*x0 + f*x1 to the mathematically equivalent x0+f*(x1-x0).
However, this isn't exactly the same in floating-point.  The updated
formula has one less operation so it's a useful speed up.

Using a float virtual_read_index in the SIMD loop causes significant
loss of precision (50 dB or less in the osc sweep tests instead of
100+ dB), so we need to do a slightly more complex SIMD version with
double virtual_read_index.

WebAudio bench results without this CL:

TEST	μs	MIN	Q1	MEDIAN	Q3	MAX	MEAN	STDDEV
Oscillator.frequency-linear-a-rate	1884	1884	1948	1969	1997	2598	1987.72	94.13

With this CL:
TEST	μs	MIN	Q1	MEDIAN	Q3	MAX	MEAN	STDDEV
Oscillator.frequency-linear-a-rate	1348	1348	1397	1418	1453	2028	1474.74	151.59

We see about a 25% improvement in speed (based on the mean).

Bug: 1013118
Change-Id: I818834c64c3529b26467dbaa7623354f92b47b2a
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2264204Reviewed-by: Hongchan Choi <hongchan@chromium.org>
Reviewed-by: Dale Curtis <dalecurtis@chromium.org>
Commit-Queue: Raymond Toy <rtoy@chromium.org>
Cr-Commit-Position: refs/heads/master@{#789097}

third_party/blink/renderer/modules/webaudio/cpu/x86/oscillator_kernel_sse2.cc

@@ -154,4 +154,127 @@ std::tuple<int, double> OscillatorHandler::ProcessKRateVector(
   return std::make_tuple(k, virtual_read_index);
 }
 
+static __m128 WrapVirtualIndexVectorPd(__m128d x,
+                                       __m128d wave_size,
+                                       __m128d inv_wave_size) {
+  // Wrap the virtual index |x| to the range 0 to wave_size - 1.  This is done
+  // by computing x - floor(x/wave_size)*wave_size.
+  //
+  // But there's no SSE2 SIMD instruction for this, so we do it the following
+  // way.
+
+  // f = truncate(x/wave_size), truncating towards 0.
+  const __m128d r = _mm_mul_pd(x, inv_wave_size);
+  __m128i f = _mm_cvttpd_epi32(r);
+
+  // Note that if r >= 0, then f <= r. But if r < 0, then r <= f, with equality
+  // only if r is already an integer.  Hence if r < f, we want to subtract 1
+  // from f to get floor(r).
+
+  // cmplt(a,b) returns 0xffffffffffffffff (-1) if a < b and 0 if not.  So cmp
+  // is -1 or 0 depending on whether r < f, which is what we need to compute
+  // floor(r).
+  __m128d cmp = _mm_cmplt_pd(r, _mm_cvtepi32_pd(f));
+
+  // Take the low 32 bits of each 64-bit result and move them into the two
+  // lowest 32-bit fields.
+  cmp = _mm_shuffle_epi32(cmp, (2 << 2) | 0);
+
+  // This subtracts 1 if needed to get floor(r).
+  f = _mm_add_epi32(f, cmp);
+
+  // Convert back to float, and scale by wave_size.  And finally subtract that
+  // from x.
+  return _mm_sub_pd(x, _mm_mul_pd(_mm_cvtepi32_pd(f), wave_size));
+}
+
+double OscillatorHandler::ProcessARateVectorKernel(
+    float* dest_p,
+    double virtual_read_index,
+    const float* phase_increments,
+    unsigned periodic_wave_size,
+    const float* const lower_wave_data[4],
+    const float* const higher_wave_data[4],
+    const float table_interpolation_factor[4]) const {
+  // See the scalar version in oscillator_node.cc for the basic algorithm.
+  double inv_periodic_wave_size = 1.0 / periodic_wave_size;
+  unsigned read_index_mask = periodic_wave_size - 1;
+
+  // Accumulate the phase increments so we can set up the virtual read index
+  // vector appropriately.  This must be a double to preserve accuracy and
+  // to match the scalar version.
+  double incr_sum[4];
+  incr_sum[0] = phase_increments[0];
+  for (int m = 1; m < 4; ++m) {
+    incr_sum[m] = incr_sum[m - 1] + phase_increments[m];
+  }
+
+  // It's really important for accuracy that we use doubles instead of
+  // floats for the virtual_read_index.  Without this, we can only get some
+  // 30-50 dB in the sweep tests instead of 100+ dB.
+  __m128d v_read_index_hi = _mm_set_pd(virtual_read_index + incr_sum[2],
+                                       virtual_read_index + incr_sum[1]);
+  __m128d v_read_index_lo =
+      _mm_set_pd(virtual_read_index + incr_sum[0], virtual_read_index);
+
+  v_read_index_hi =
+      WrapVirtualIndexVectorPd(v_read_index_hi, _mm_set1_pd(periodic_wave_size),
+                               _mm_set1_pd(inv_periodic_wave_size));
+  v_read_index_lo =
+      WrapVirtualIndexVectorPd(v_read_index_lo, _mm_set1_pd(periodic_wave_size),
+                               _mm_set1_pd(inv_periodic_wave_size));
+
+  // Convert the virtual read index (parts) to an integer, and carefully
+  // merge them into one vector.
+  const __m128i v_read0 = _mm_movelh_ps(_mm_cvttpd_epi32(v_read_index_lo),
+                                        _mm_cvttpd_epi32(v_read_index_hi));
+
+  // Get index to next element being sure to wrap the index around if needed.
+  const __m128i v_read1 =
+      _mm_and_si128(_mm_add_epi32(v_read0, _mm_set1_epi32(1)),
+                    _mm_set1_epi32(read_index_mask));
+
+  float sample1_lower[4] __attribute__((aligned(16)));
+  float sample2_lower[4] __attribute__((aligned(16)));
+  float sample1_higher[4] __attribute__((aligned(16)));
+  float sample2_higher[4] __attribute__((aligned(16)));
+
+  const unsigned* read0 = reinterpret_cast<const unsigned*>(&v_read0);
+  const unsigned* read1 = reinterpret_cast<const unsigned*>(&v_read1);
+
+  for (int m = 0; m < 4; ++m) {
+    DCHECK_LT(read0[m], periodic_wave_size);
+    DCHECK_LT(read1[m], periodic_wave_size);
+
+    sample1_lower[m] = lower_wave_data[m][read0[m]];
+    sample2_lower[m] = lower_wave_data[m][read1[m]];
+    sample1_higher[m] = higher_wave_data[m][read0[m]];
+    sample2_higher[m] = higher_wave_data[m][read1[m]];
+  }
+
+  const __m128 v_factor =
+      _mm_sub_ps(_mm_movelh_ps(_mm_cvtpd_ps(v_read_index_lo),
+                               _mm_cvtpd_ps(v_read_index_hi)),
+                 _mm_cvtepi32_ps(v_read0));
+  const __m128 sample_higher =
+      _mm_add_ps(_mm_load_ps(sample1_higher),
+                 _mm_mul_ps(v_factor, _mm_sub_ps(_mm_load_ps(sample2_higher),
+                                                 _mm_load_ps(sample1_higher))));
+  const __m128 sample_lower =
+      _mm_add_ps(_mm_load_ps(sample1_lower),
+                 _mm_mul_ps(v_factor, _mm_sub_ps(_mm_load_ps(sample2_lower),
+                                                 _mm_load_ps(sample1_lower))));
+  const __m128 sample = _mm_add_ps(
+      sample_higher, _mm_mul_ps(_mm_load_ps(table_interpolation_factor),
+                                _mm_sub_ps(sample_lower, sample_higher)));
+
+  _mm_storeu_ps(dest_p, sample);
+
+  virtual_read_index += incr_sum[3];
+  virtual_read_index -=
+      floor(virtual_read_index * inv_periodic_wave_size) * periodic_wave_size;
+
+  return virtual_read_index;
+}
+
 }  // namespace blink

third_party/blink/renderer/modules/webaudio/oscillator_node.cc

@@ -358,7 +358,6 @@ static float DoInterpolation(double virtual_read_index,
 }
 
 #if !(defined(ARCH_CPU_X86_FAMILY) || defined(CPU_ARM_NEON))
-
 // Vector operations not supported, so there's nothing to do except return 0 and
 // virtual_read_index.  The scalar version will do the necessary processing.
 std::tuple<int, double> OscillatorHandler::ProcessKRateVector(
@@ -372,6 +371,65 @@ std::tuple<int, double> OscillatorHandler::ProcessKRateVector(
 }
 #endif
 
+#if !defined(ARCH_CPU_X86_FAMILY)
+double OscillatorHandler::ProcessARateVectorKernel(
+    float* dest_p,
+    double virtual_read_index,
+    const float* phase_increments,
+    unsigned periodic_wave_size,
+    const float* const lower_wave_data[4],
+    const float* const higher_wave_data[4],
+    const float table_interpolation_factor[4]) const {
+  double inv_periodic_wave_size = 1.0 / periodic_wave_size;
+  unsigned read_index_mask = periodic_wave_size - 1;
+
+  for (int m = 0; m < 4; ++m) {
+    unsigned read_index_0 = static_cast<unsigned>(virtual_read_index);
+
+    // Increment is fairly large, so we're doing no more than about 3
+    // points between each wave table entry. Assume linear
+    // interpolation between points is good enough.
+    unsigned read_index2 = read_index_0 + 1;
+
+    // Contain within valid range.
+    read_index_0 = read_index_0 & read_index_mask;
+    read_index2 = read_index2 & read_index_mask;
+
+    float sample1_lower = lower_wave_data[m][read_index_0];
+    float sample2_lower = lower_wave_data[m][read_index2];
+    float sample1_higher = higher_wave_data[m][read_index_0];
+    float sample2_higher = higher_wave_data[m][read_index2];
+
+    // Linearly interpolate within each table (lower and higher).
+    double interpolation_factor =
+        static_cast<float>(virtual_read_index) - read_index_0;
+    // Doing linear interpolation via x0 + f*(x1-x0) gives slightly
+    // different results from (1-f)*x0 + f*x1, but requires fewer
+    // operations.  This causes a very slight decrease in SNR (< 0.05 dB) in
+    // oscillator sweep tests.
+    float sample_higher =
+        sample1_higher +
+        interpolation_factor * (sample2_higher - sample1_higher);
+    float sample_lower =
+        sample1_lower + interpolation_factor * (sample2_lower - sample1_lower);
+
+    // Then interpolate between the two tables.
+    float sample = sample_higher + table_interpolation_factor[m] *
+                                       (sample_lower - sample_higher);
+
+    dest_p[m] = sample;
+
+    // Increment virtual read index and wrap virtualReadIndex into the range
+    // 0 -> periodicWaveSize.
+    virtual_read_index += phase_increments[m];
+    virtual_read_index -=
+        floor(virtual_read_index * inv_periodic_wave_size) * periodic_wave_size;
+  }
+
+  return virtual_read_index;
+}
+#endif
+
 double OscillatorHandler::ProcessKRateScalar(int start,
                                              int n,
                                              float* dest_p,
@@ -488,10 +546,75 @@ double OscillatorHandler::ProcessKRate(int n,
   return virtual_read_index;
 }
 
-double OscillatorHandler::ProcessARate(int n,
-                                       float* dest_p,
+std::tuple<int, double> OscillatorHandler::ProcessARateVector(
+    int n,
+    float* destination,
     double virtual_read_index,
-                                       float* phase_increments) const {
+    const float* phase_increments) const {
+  float rate_scale = periodic_wave_->RateScale();
+  float inv_rate_scale = 1 / rate_scale;
+  unsigned periodic_wave_size = periodic_wave_->PeriodicWaveSize();
+  double inv_periodic_wave_size = 1.0 / periodic_wave_size;
+  unsigned read_index_mask = periodic_wave_size - 1;
+
+  float* higher_wave_data[4];
+  float* lower_wave_data[4];
+  float table_interpolation_factor[4] __attribute__((aligned(16)));
+
+  int k = 0;
+  int n_loops = n / 4;
+
+  for (int loop = 0; loop < n_loops; ++loop, k += 4) {
+    bool is_big_increment = true;
+    float frequency[4];
+
+    for (int m = 0; m < 4; ++m) {
+      float phase_incr = phase_increments[k + m];
+      is_big_increment =
+          is_big_increment && (fabs(phase_incr) >= kInterpolate2Point);
+      frequency[m] = inv_rate_scale * phase_incr;
+    }
+
+    periodic_wave_->WaveDataForFundamentalFrequency(frequency, lower_wave_data,
+                                                    higher_wave_data,
+                                                    table_interpolation_factor);
+
+    // If all the phase increments are large enough, we can use linear
+    // interpolation with a possibly vectorized implementation.  If not, we need
+    // to call DoInterpolation to handle it correctly.
+    if (is_big_increment) {
+      virtual_read_index = ProcessARateVectorKernel(
+          destination + k, virtual_read_index, phase_increments + k,
+          periodic_wave_size, lower_wave_data, higher_wave_data,
+          table_interpolation_factor);
+    } else {
+      for (int m = 0; m < 4; ++m) {
+        float sample =
+            DoInterpolation(virtual_read_index, fabs(phase_increments[k + m]),
+                            read_index_mask, table_interpolation_factor[m],
+                            lower_wave_data[m], higher_wave_data[m]);
+
+        destination[k + m] = sample;
+
+        // Increment virtual read index and wrap virtualReadIndex into the range
+        // 0 -> periodicWaveSize.
+        virtual_read_index += phase_increments[k + m];
+        virtual_read_index -=
+            floor(virtual_read_index * inv_periodic_wave_size) *
+            periodic_wave_size;
+      }
+    }
+  }
+
+  return std::make_tuple(k, virtual_read_index);
+}
+
+double OscillatorHandler::ProcessARateScalar(
+    int k,
+    int n,
+    float* destination,
+    double virtual_read_index,
+    const float* phase_increments) const {
   float rate_scale = periodic_wave_->RateScale();
   float inv_rate_scale = 1 / rate_scale;
   unsigned periodic_wave_size = periodic_wave_->PeriodicWaveSize();
@@ -502,8 +625,8 @@ double OscillatorHandler::ProcessARate(int n,
   float* lower_wave_data = nullptr;
   float table_interpolation_factor = 0;
 
-  for (int k = 0; k < n; ++k) {
-    float incr = *phase_increments++;
+  for (int m = k; m < n; ++m) {
+    float incr = phase_increments[m];
 
     float frequency = inv_rate_scale * incr;
     periodic_wave_->WaveDataForFundamentalFrequency(frequency, lower_wave_data,
@@ -514,7 +637,7 @@ double OscillatorHandler::ProcessARate(int n,
                                    read_index_mask, table_interpolation_factor,
                                    lower_wave_data, higher_wave_data);
 
-    *dest_p++ = sample;
+    destination[m] = sample;
 
     // Increment virtual read index and wrap virtualReadIndex into the range
     // 0 -> periodicWaveSize.
@@ -526,6 +649,21 @@ double OscillatorHandler::ProcessARate(int n,
   return virtual_read_index;
 }
 
+double OscillatorHandler::ProcessARate(int n,
+                                       float* destination,
+                                       double virtual_read_index,
+                                       float* phase_increments) const {
+  int frames_processed = 0;
+
+  std::tie(frames_processed, virtual_read_index) =
+      ProcessARateVector(n, destination, virtual_read_index, phase_increments);
+
+  virtual_read_index = ProcessARateScalar(frames_processed, n, destination,
+                                          virtual_read_index, phase_increments);
+
+  return virtual_read_index;
+}
+
 void OscillatorHandler::Process(uint32_t frames_to_process) {
   AudioBus* output_bus = Output(0).Bus();
 

third_party/blink/renderer/modules/webaudio/oscillator_node.h

@@ -118,6 +118,63 @@ class OscillatorHandler final : public AudioScheduledSourceHandler {
                       double virtual_read_index,
                       float* phase_increments) const;
 
+  // Scalar version of ProcessARate().  Also handles any values not handled by
+  // the vector version.
+  //
+  //   k
+  //     start index for where to write the result (and read phase_increments)
+  //   n
+  //     total number of frames to process
+  //   destination
+  //     Array where the samples values are written
+  //   virtual_read_index
+  //     index into the wave data tables containing the waveform
+  //   phase_increments
+  //     phase change to use for each frame of output
+  //
+  // Returns the updated virtual_read_index.
+  double ProcessARateScalar(int k,
+                            int n,
+                            float* destination,
+                            double virtual_read_index,
+                            const float* phase_increments) const;
+
+  // Vector version of ProcessARate().  Returns the number of frames processed
+  // and the update virtual_read_index.
+  std::tuple<int, double> ProcessARateVector(
+      int n,
+      float* destination,
+      double virtual_read_index,
+      const float* phase_increments) const;
+
+  // Handles the linear interpolation in ProcessARateVector().
+  //
+  //   destination
+  //     Where the interpolated values are written.
+  //   virtual_read_index
+  //     index into the wave table data
+  //   phase_increments
+  //     phase increments array
+  //   periodic_wave_size
+  //     Length of the periodic wave stored in the wave tables
+  //   lower_wave_data
+  //     Array of the 4 lower wave table arrays
+  //   higher_wave_data
+  //     Array of the 4 higher wave table arrays
+  //   table_interpolation_factor
+  //     Array of linear interpolation factors to use between the lower and
+  //     higher wave tables.
+  //
+  // Returns the updated virtual_read_index
+  double ProcessARateVectorKernel(
+      float* destination,
+      double virtual_read_index,
+      const float* phase_increments,
+      unsigned periodic_wave_size,
+      const float* const lower_wave_data[4],
+      const float* const higher_wave_data[4],
+      const float table_interpolation_factor[4]) const;
+
   // One of the waveform types defined in the enum.
   uint8_t type_;
 

third_party/blink/renderer/modules/webaudio/periodic_wave.cc

@@ -29,6 +29,7 @@
 #include <algorithm>
 #include <memory>
 
+#include "build/build_config.h"
 #include "third_party/blink/renderer/bindings/modules/v8/v8_periodic_wave_options.h"
 #include "third_party/blink/renderer/modules/webaudio/base_audio_context.h"
 #include "third_party/blink/renderer/modules/webaudio/oscillator_node.h"
@@ -198,6 +199,90 @@ void PeriodicWave::WaveDataForFundamentalFrequency(
   table_interpolation_factor = pitch_range - range_index1;
 }
 
+#if defined(ARCH_CPU_X86_FAMILY)
+void PeriodicWave::WaveDataForFundamentalFrequency(
+    const float fundamental_frequency[4],
+    float* lower_wave_data[4],
+    float* higher_wave_data[4],
+    float table_interpolation_factor[4]) {
+  // Negative frequencies are allowed, in which case we alias to the positive
+  // frequency.  SSE2 doesn't have an fabs instruction, so just remove the sign
+  // bit of the float numbers, effecitvely taking the absolute value.
+  const __m128 frequency = _mm_and_ps(_mm_loadu_ps(fundamental_frequency),
+                                      _mm_set1_epi32(0x7fffffff));
+
+  // pos = 0xffffffff if freq > 0; otherwise 0
+  const __m128i pos = _mm_cmpgt_ps(frequency, _mm_set1_ps(0));
+
+  // Calculate the pitch range.
+  __m128 v_ratio =
+      _mm_div_ps(frequency, _mm_set1_ps(lowest_fundamental_frequency_));
+
+  // Set v_ratio to 0 if freq <= 0; otherwise keep the ratio.
+  v_ratio = _mm_and_ps(v_ratio, pos);
+
+  // If pos = 0, set value to 0.5 and 0 otherwise.  Or this into v_ratio so that
+  // v_ratio is 0.5 if freq <= 0.  Otherwise preserve v_ratio.
+  v_ratio = _mm_or_ps(v_ratio, _mm_andnot_ps(pos, _mm_set1_ps(0.5)));
+
+  const float* ratio = reinterpret_cast<float*>(&v_ratio);
+
+  float cents_above_lowest_frequency[4] __attribute__((aligned(16)));
+
+  for (int k = 0; k < 4; ++k) {
+    cents_above_lowest_frequency[k] = log2f(ratio[k]) * 1200;
+  }
+
+  __m128 v_pitch_range = _mm_add_ps(
+      _mm_set1_ps(1.0), _mm_div_ps(_mm_load_ps(cents_above_lowest_frequency),
+                                   _mm_set1_ps((cents_per_range_))));
+  v_pitch_range = _mm_max_ps(v_pitch_range, _mm_set1_ps(0.0));
+  v_pitch_range = _mm_min_ps(v_pitch_range, _mm_set1_ps(NumberOfRanges() - 1));
+
+  const __m128i v_index1 = _mm_cvttps_epi32(v_pitch_range);
+  __m128i v_index2 = _mm_add_epi32(v_index1, _mm_set1_epi32(1));
+
+  // SSE2 deosn't have _mm_min_epi32 (but SSE4.2 does).
+  //
+  // The following ought to work because the small integers for the index and
+  // number of ranges should look like tiny denormals that should compare in the
+  // same order as integers.  This doesn't work because we have flush-to-zero
+  // enabled.
+  //
+  //   __m128i v_range = _mm_set1_epi32(NumberOfRanges() - 1);
+  //  v_index2 = _mm_min_ps(v_index2, v_range);
+  //
+  // Instead we convert to float, take the min and convert back. No round off
+  // because the integers are small.
+  v_index2 = _mm_cvttps_epi32(
+      _mm_min_ps(_mm_cvtepi32_ps(v_index2), _mm_set1_ps(NumberOfRanges() - 1)));
+
+  const __m128 table_factor =
+      _mm_sub_ps(v_pitch_range, _mm_cvtepi32_ps(v_index1));
+  _mm_storeu_ps(table_interpolation_factor, table_factor);
+
+  const unsigned* range_index1 = reinterpret_cast<const unsigned*>(&v_index1);
+  const unsigned* range_index2 = reinterpret_cast<const unsigned*>(&v_index2);
+
+  for (int k = 0; k < 4; ++k) {
+    lower_wave_data[k] = band_limited_tables_[range_index2[k]]->Data();
+    higher_wave_data[k] = band_limited_tables_[range_index1[k]]->Data();
+  }
+}
+#else
+void PeriodicWave::WaveDataForFundamentalFrequency(
+    const float fundamental_frequency[4],
+    float* lower_wave_data[4],
+    float* higher_wave_data[4],
+    float table_interpolation_factor[4]) {
+  for (int k = 0; k < 4; ++k) {
+    WaveDataForFundamentalFrequency(fundamental_frequency[k],
+                                    lower_wave_data[k], higher_wave_data[k],
+                                    table_interpolation_factor[k]);
+  }
+}
+#endif
+
 unsigned PeriodicWave::NumberOfPartialsForRange(unsigned range_index) const {
   // Number of cents below nyquist where we cull partials.
   float cents_to_cull = range_index * cents_per_range_;

third_party/blink/renderer/modules/webaudio/periodic_wave.h

@@ -80,6 +80,15 @@ class PeriodicWave final : public ScriptWrappable {
                                        float*& higher_wave_data,
                                        float& table_interpolation_factor);
 
+  // Like the above, except we compute accept 4 frequencies at a time and return
+  // 4 lower/higher wave data tables and the 4 corresponding table interpolation
+  // factors.  Intended for use with the OscillatorNode for faster a-rate
+  // processing.
+  void WaveDataForFundamentalFrequency(const float fundamental_frequency[4],
+                                       float* lower_wave_data[4],
+                                       float* higher_wave_data[4],
+                                       float table_interpolation_factor[4]);
+
   // Returns the scalar multiplier to the oscillator frequency to calculate wave
   // buffer phase increment.
   float RateScale() const { return rate_scale_; }

third_party/blink/web_tests/webaudio/Oscillator/osc-sweep-snr-custom.html

@@ -22,7 +22,7 @@
             1, tester.sampleRate * tester.lengthInSeconds, tester.sampleRate);
 
         // The thresholds are experimentally determined.
-        tester.setThresholds({snr: 139.991, maxDiff: 2.2650e-6});
+        tester.setThresholds({snr: 139.98, maxDiff: 2.2650e-6});
         tester.runTest(
             context, 'custom', 'Custom Oscillator with Exponential Sweep', task,
             should);

third_party/blink/web_tests/webaudio/Oscillator/osc-sweep-snr-sawtooth.html

@@ -22,7 +22,7 @@
             1, tester.sampleRate * tester.lengthInSeconds, tester.sampleRate);
 
         // The thresholds are experimentally determined.
-        tester.setThresholds({snr: 134.39, maxDiff: 1.8925e-6});
+        tester.setThresholds({snr: 134.34, maxDiff: 1.8925e-6});
         tester.runTest(
             context, 'sawtooth', 'Sawtooth Oscillator with Exponential Sweep',
             task, should);

third_party/blink/web_tests/webaudio/Oscillator/osc-sweep-snr-triangle.html

@@ -22,7 +22,7 @@
             1, tester.sampleRate * tester.lengthInSeconds, tester.sampleRate);
 
         // The thresholds are experimentally determined.
-        tester.setThresholds({snr: 139.98, maxDiff: 2.8313e-6});
+        tester.setThresholds({snr: 139.88, maxDiff: 2.8313e-6});
         tester.runTest(
             context, 'triangle', 'Triangle Oscillator with Exponential Sweep ',
             task, should);