filter_kernels: add ewa_lanczos4sharpest

Mainly for vo_gpu_next.

DOCS/interface-changes.rst

@@ -58,6 +58,7 @@ Interface changes
     - remove `--tone-mapping-mode`
     - change `--subs-fallback-forced` so that it works alongside `--slang`
     - add `--icc-3dlut-size=auto` and make it the default
+    - add `--scale=ewa_lanczos4sharpest`
  --- mpv 0.36.0 ---
     - add `--target-contrast`
     - Target luminance value is now also applied when ICC profile is used.

DOCS/man/options.rst

@@ -5278,6 +5278,12 @@ them.
         A slightly sharpened version of ewa_lanczos. If your hardware can run
         it, this is probably what you should use by default.
 
+    ``ewa_lanczos4sharpest``
+        Very sharp scaler, but also slightly slower than ``ewa_lanczossharp``.
+        Prone to ringing, so it's recommended to combine this with an
+        anti-ringing shader. On ``--vo=gpu-next``, setting this filter enables
+        built-in anti-ringing, so no extra action needs to be taken.
+
     ``mitchell``
         Mitchell-Netravali. The ``B`` and ``C`` parameters can be set with
         ``--scale-param1`` and ``--scale-param2``. This filter is very good at

video/out/filter_kernels.c

@@ -363,6 +363,7 @@ const struct filter_window mp_filter_windows[] = {
 };
 
 #define JINC_R3 3.2383154841662362
+#define JINC_R4 4.2410628637960699
 
 const struct filter_kernel mp_filter_kernels[] = {
     // Spline filters
@@ -382,6 +383,11 @@ const struct filter_kernel mp_filter_kernels[] = {
     // preserve horizontal/vertical lines)
     {{"ewa_lanczossharp", JINC_R3, jinc, .blur = 0.9812505837223707, .resizable = true},
         .polar = true, .window = "jinc"},
+    // Similar to the above, but sharpened substantially to the point of
+    // minimizing the total impulse response error on an integer grid. Tends
+    // to preserve hash patterns well. Very sharp but rings a lot.
+    {{"ewa_lanczos4sharpest", JINC_R4, jinc, .blur = 0.8845120932605005, .resizable = true},
+        .polar = true, .window = "jinc"},
     // Similar to the above, but softened instead, to make even/odd integer
     // contributions exactly symmetrical. Designed to smooth out hash patterns.
     {{"ewa_lanczossoft", JINC_R3, jinc, .blur = 1.0164667662867047, .resizable = true},