From 575fa168ed290b8cc2540eb3438abd92affda34b Mon Sep 17 00:00:00 2001
From: kris <kris.kennaway@gmail.com>
Date: Sun, 10 Jan 2021 22:12:14 +0000
Subject: [PATCH] Cythonize dither.apply

---
 dither.py        | 30 ++++++++++++++++++------------
 dither_apply.pyx | 39 +++++++++++++++++++++++++++++++++++++++
 2 files changed, 57 insertions(+), 12 deletions(-)
 create mode 100644 dither_apply.pyx

diff --git a/dither.py b/dither.py
index c7f8e17..9df0dfc 100644
--- a/dither.py
+++ b/dither.py
@@ -7,7 +7,8 @@ from typing import Tuple
 
 from PIL import Image
 import numpy as np
-
+import pyximport; pyximport.install(language_level=3)
+import dither_apply
 
 # TODO:
 # - only lookahead for 560px
@@ -281,23 +282,28 @@ class Dither:
 
     def apply(self, screen: Screen, image: np.ndarray, x: int, y: int,
               quant_error: np.ndarray, one_line=False):
-        error = self.PATTERN * quant_error.reshape((1, 1, 3))
         el, er, xl, xr = self.x_dither_bounds(screen, x)
         et, eb, yt, yb = self.y_dither_bounds(screen, y, one_line)
-
-        # We could avoid clipping here, i.e. allow RGB values to extend beyond
-        # 0..255 to capture a larger range of residual error.  This is faster
-        # but seems to reduce image quality.
-        # XXX extend image region to avoid need for boundary box clipping
-        image[yt:yb, xl:xr, :] = np.clip(
-            image[yt:yb, xl:xr, :] + error[et:eb, el:er, :], 0, 255)
+        return dither_apply.apply(self.PATTERN, el, er, xl, xr, et, eb, yt,
+                                  yb, image, quant_error)
+        # error = self.PATTERN * quant_error.reshape((1, 1, 3))
+        #
+        # # We could avoid clipping here, i.e. allow RGB values to extend beyond
+        # # 0..255 to capture a larger range of residual error.  This is faster
+        # # but seems to reduce image quality.
+        # # XXX extend image region to avoid need for boundary box clipping
+        # image[yt:yb, xl:xr, :] = np.clip(
+        #     image[yt:yb, xl:xr, :] + error[et:eb, el:er, :], 0, 255)
 
     def apply_one_line(self, screen: Screen, image: np.ndarray, x: int, y: int,
                        quant_error: np.ndarray):
-        error = self.PATTERN[0, :] * quant_error.reshape(1, 3)
         el, er, xl, xr = self.x_dither_bounds(screen, x)
-        image[y, xl:xr, :] = np.clip(
-            image[y, xl:xr, :] + error[el:er, :], 0, 255)
+        return dither_apply.apply_one_line(self.PATTERN, el, er, xl, xr, y,
+                                           image, quant_error)
+        # error = self.PATTERN[0, :] * quant_error.reshape(1, 3)
+        #
+        # image[y, xl:xr, :] = np.clip(
+        #     image[y, xl:xr, :] + error[el:er, :], 0, 255)
 
 
 class FloydSteinbergDither(Dither):
diff --git a/dither_apply.pyx b/dither_apply.pyx
new file mode 100644
index 0000000..2518d1d
--- /dev/null
+++ b/dither_apply.pyx
@@ -0,0 +1,39 @@
+# cython: infer_types=True
+
+cimport cython
+import numpy as np
+# from cython.parallel import prange
+from cython.view cimport array as cvarray
+
+
+cdef float clip(float a, float min_value, float max_value) nogil:
+    return min(max(a, min_value), max_value)
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+def apply_one_line(float[:, :, ::1] pattern, int el, int er, int xl, int xr, int y, float[:, :, ::1] image,
+                   float[::1] quant_error):
+    cdef int i, j
+    cdef float[:, ::1] error = cvarray(
+        shape=(pattern.shape[1], quant_error.shape[0]), itemsize=sizeof(float), format="f")
+
+    for i in range(pattern.shape[1]):
+        for j in range(quant_error.shape[0]):
+            error[i, j] = pattern[0, i, 0] * quant_error[j]
+
+    for i in range(xr - xl):
+        for j in range(3):
+            image[y, xl+i, j] = clip(image[y, xl + i, j] + error[el + i, j], 0, 255)
+
+
+# XXX cythonize
+def apply(pattern, int el, int er, int xl, int xr, int et, int eb, int yt, int yb, image, quant_error):
+    error = pattern * quant_error.reshape((1, 1, 3))
+
+    # We could avoid clipping here, i.e. allow RGB values to extend beyond
+    # 0..255 to capture a larger range of residual error.  This is faster
+    # but seems to reduce image quality.
+    # XXX extend image region to avoid need for boundary box clipping
+    image[yt:yb, xl:xr, :] = np.clip(
+        image[yt:yb, xl:xr, :] + error[et:eb, el:er, :], 0, 255)