Adjust the bsinc filter table packing

Now each scale's filter and phase delta are interleaved for each phase index,
followed by the scale and scale+phase delta for each phase index. This ensures
no holes in the filter coefficients for the fast bsinc resampler for a given
run, while keeping the scale deltas in the same vicinity for the non-fast bsinc
resampler.

4 files changed, 29 insertions, 41 deletions

diff --git a/core/bsinc_tables.cpp b/core/bsinc_tables.cpp
index 315e1448..334822f5 100644
--- a/core/bsinc_tables.cpp
+++ b/core/bsinc_tables.cpp
@@ -189,16 +189,16 @@ struct BSincFilterArray {
         }
 
         size_t idx{0};
-        for(size_t si{0};si < BSincScaleCount-1;++si)
+        for(size_t si{0};si < BSincScaleCount;++si)
         {
             const size_t m{((hdr.a[si]*2) + 3) & ~3u};
             const size_t o{(BSincPointsMax-m) / 2};
 
+            /* Write out each phase index's filter and phase delta for this
+             * quality scale.
+             */
             for(size_t pi{0};pi < BSincPhaseCount;++pi)
             {
-                /* Write out the filter. Also calculate and write out the phase
-                 * and scale deltas.
-                 */
                 for(size_t i{0};i < m;++i)
                     mTable[idx++] = static_cast<float>(filter[si][pi][o+i]);
 
@@ -210,11 +210,22 @@ struct BSincFilterArray {
                     const double phDelta{filter[si][pi+1][o+i] - filter[si][pi][o+i]};
                     mTable[idx++] = static_cast<float>(phDelta);
                 }
-
+            }
+            /* Calculate and write out each phase index's filter quality scale
+             * deltas. The last scale index doesn't have any scale or scale-
+             * phase deltas.
+             */
+            if(si == BSincScaleCount-1)
+            {
+                for(size_t i{0};i < BSincPhaseCount*m*2;++i)
+                    mTable[idx++] = 0.0f;
+            }
+            else for(size_t pi{0};pi < BSincPhaseCount;++pi)
+            {
                 /* Linear interpolation between scales is also simplified.
                  *
-                 * Given a difference in points between scales, the destination
-                 * points will be 0, thus: x = a + f (-a)
+                 * Given a difference in the number of points between scales,
+                 * the destination points will be 0, thus: x = a + f (-a)
                  */
                 for(size_t i{0};i < m;++i)
                 {
@@ -223,8 +234,8 @@ struct BSincFilterArray {
                 }
 
                 /* This last simplification is done to complete the bilinear
-                 * equation for the combination of phase and scale.
-                 */
+                * equation for the combination of phase and scale.
+                */
                 for(size_t i{0};i < m;++i)
                 {
                     const double spDelta{(filter[si+1][pi+1][o+i] - filter[si+1][pi][o+i]) -
@@ -233,29 +244,6 @@ struct BSincFilterArray {
                 }
             }
         }
-        {
-            /* The last scale index doesn't have any scale or scale-phase
-             * deltas.
-             */
-            constexpr size_t si{BSincScaleCount-1};
-            const size_t m{((hdr.a[si]*2) + 3) & ~3u};
-            const size_t o{(BSincPointsMax-m) / 2};
-
-            for(size_t pi{0};pi < BSincPhaseCount;++pi)
-            {
-                for(size_t i{0};i < m;++i)
-                    mTable[idx++] = static_cast<float>(filter[si][pi][o+i]);
-                for(size_t i{0};i < m;++i)
-                {
-                    const double phDelta{filter[si][pi+1][o+i] - filter[si][pi][o+i]};
-                    mTable[idx++] = static_cast<float>(phDelta);
-                }
-                for(size_t i{0};i < m;++i)
-                    mTable[idx++] = 0.0f;
-                for(size_t i{0};i < m;++i)
-                    mTable[idx++] = 0.0f;
-            }
-        }
         assert(idx == hdr.total_size);
     }
 };
diff --git a/core/mixer/mixer_c.cpp b/core/mixer/mixer_c.cpp
index 8c2c60c5..f82f7dd1 100644
--- a/core/mixer/mixer_c.cpp
+++ b/core/mixer/mixer_c.cpp
@@ -38,9 +38,9 @@ inline float do_bsinc(const InterpState &istate, const float *RESTRICT vals, con
     const uint pi{frac >> FracPhaseBitDiff};
     const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
 
-    const float *RESTRICT fil{istate.bsinc.filter + m*pi*4};
+    const float *RESTRICT fil{istate.bsinc.filter + m*pi*2};
     const float *RESTRICT phd{fil + m};
-    const float *RESTRICT scd{phd + m};
+    const float *RESTRICT scd{fil + BSincPhaseCount*2*m};
     const float *RESTRICT spd{scd + m};
 
     // Apply the scale and phase interpolated filter.
@@ -58,7 +58,7 @@ inline float do_fastbsinc(const InterpState &istate, const float *RESTRICT vals,
     const uint pi{frac >> FracPhaseBitDiff};
     const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
 
-    const float *RESTRICT fil{istate.bsinc.filter + m*pi*4};
+    const float *RESTRICT fil{istate.bsinc.filter + m*pi*2};
     const float *RESTRICT phd{fil + m};
 
     // Apply the phase interpolated filter.
diff --git a/core/mixer/mixer_neon.cpp b/core/mixer/mixer_neon.cpp
index 8cd7e610..a3afdc6b 100644
--- a/core/mixer/mixer_neon.cpp
+++ b/core/mixer/mixer_neon.cpp
@@ -131,9 +131,9 @@ float *Resample_<BSincTag,NEONTag>(const InterpState *state, float *RESTRICT src
         float32x4_t r4{vdupq_n_f32(0.0f)};
         {
             const float32x4_t pf4{vdupq_n_f32(pf)};
-            const float *RESTRICT fil{filter + m*pi*4};
+            const float *RESTRICT fil{filter + m*pi*2};
             const float *RESTRICT phd{fil + m};
-            const float *RESTRICT scd{phd + m};
+            const float *RESTRICT scd{fil + BSincPhaseCount*2*m};
             const float *RESTRICT spd{scd + m};
             size_t td{m >> 2};
             size_t j{0u};
@@ -177,7 +177,7 @@ float *Resample_<FastBSincTag,NEONTag>(const InterpState *state, float *RESTRICT
         float32x4_t r4{vdupq_n_f32(0.0f)};
         {
             const float32x4_t pf4{vdupq_n_f32(pf)};
-            const float *RESTRICT fil{filter + m*pi*4};
+            const float *RESTRICT fil{filter + m*pi*2};
             const float *RESTRICT phd{fil + m};
             size_t td{m >> 2};
             size_t j{0u};
diff --git a/core/mixer/mixer_sse.cpp b/core/mixer/mixer_sse.cpp
index ba53d0a5..2a11601c 100644
--- a/core/mixer/mixer_sse.cpp
+++ b/core/mixer/mixer_sse.cpp
@@ -96,9 +96,9 @@ float *Resample_<BSincTag,SSETag>(const InterpState *state, float *RESTRICT src,
         __m128 r4{_mm_setzero_ps()};
         {
             const __m128 pf4{_mm_set1_ps(pf)};
-            const float *RESTRICT fil{filter + m*pi*4};
+            const float *RESTRICT fil{filter + m*pi*2};
             const float *RESTRICT phd{fil + m};
-            const float *RESTRICT scd{phd + m};
+            const float *RESTRICT scd{fil + BSincPhaseCount*2*m};
             const float *RESTRICT spd{scd + m};
             size_t td{m >> 2};
             size_t j{0u};
@@ -143,7 +143,7 @@ float *Resample_<FastBSincTag,SSETag>(const InterpState *state, float *RESTRICT
         __m128 r4{_mm_setzero_ps()};
         {
             const __m128 pf4{_mm_set1_ps(pf)};
-            const float *RESTRICT fil{filter + m*pi*4};
+            const float *RESTRICT fil{filter + m*pi*2};
             const float *RESTRICT phd{fil + m};
             size_t td{m >> 2};
             size_t j{0u};