avx2: Improve mean color extractor

Smaller loads perform slightly better than load + extract. Tested
on GCC 13/Haswell.

Cut down on code verbosity.

chafa/internal/chafa-avx2.c

@@ -83,51 +83,36 @@ void
 calc_colors_avx2 (const ChafaPixel *pixels, ChafaColorAccum *accums_out,
                   const guint32 *sym_mask_u32)
 {
-    const __m256i *pixels_8x_p = (const __m256i *) pixels;
-    const __m256i *sym_mask_8x_p = (const __m256i *) sym_mask_u32;
-    __m256i accum_fg [2] = { { 0 }, { 0 } };
-    __m256i accum_bg [2] = { { 0 }, { 0 } };
+    const __m128i *pixels_4x_p = (const __m128i *) pixels;
+    const __m128i *sym_mask_4x_p = (const __m128i *) sym_mask_u32;
+    __m256i accum_fg = { 0 };
+    __m256i accum_bg = { 0 };
     __m128i accum_fg_128;
     __m128i accum_bg_128;
     gint i;
 
-    for (i = 0; i < CHAFA_SYMBOL_N_PIXELS / 8; i++)
+    for (i = 0; i < CHAFA_SYMBOL_N_PIXELS / 4; i++)
     {
-        __m256i pixels_8x, sym_mask_8x;
-        __m256i p0, fg0, bg0;
-        __m256i p1, fg1, bg1;
-
-        pixels_8x = _mm256_loadu_si256 (pixels_8x_p);
-        pixels_8x_p++;
-
-        sym_mask_8x = _mm256_loadu_si256 (sym_mask_8x_p);
-        sym_mask_8x_p++;
+        __m128i pixels_4x, sym_mask_4x;
 
-        p0 = _mm256_andnot_si256 (sym_mask_8x, pixels_8x);
-        p1 = _mm256_and_si256 (sym_mask_8x, pixels_8x);
+        pixels_4x = _mm_loadu_si128 (pixels_4x_p++);
+        sym_mask_4x = _mm_loadu_si128 (sym_mask_4x_p++);
 
-        fg0 = _mm256_cvtepu8_epi16 (_mm256_extracti128_si256 (p0, 0));
-        fg1 = _mm256_cvtepu8_epi16 (_mm256_extracti128_si256 (p0, 1));
-        accum_fg [0] = _mm256_add_epi16 (accum_fg [0], fg0);
-        accum_fg [1] = _mm256_add_epi16 (accum_fg [1], fg1);
-
-        bg0 = _mm256_cvtepu8_epi16 (_mm256_extracti128_si256 (p1, 0));
-        bg1 = _mm256_cvtepu8_epi16 (_mm256_extracti128_si256 (p1, 1));
-        accum_bg [0] = _mm256_add_epi16 (accum_bg [0], bg0);
-        accum_bg [1] = _mm256_add_epi16 (accum_bg [1], bg1);
+        accum_fg = _mm256_add_epi16 (accum_fg,
+            _mm256_cvtepu8_epi16 (_mm_and_si128 (sym_mask_4x, pixels_4x)));
+        accum_bg = _mm256_add_epi16 (accum_bg,
+            _mm256_cvtepu8_epi16 (_mm_andnot_si128 (sym_mask_4x, pixels_4x)));
     }
 
-    accum_fg [0] = _mm256_add_epi16 (accum_fg [0], accum_fg [1]);
-    accum_fg_128 = _mm_add_epi16 (_mm256_extracti128_si256 (accum_fg [0], 0),
-                                  _mm256_extracti128_si256 (accum_fg [0], 1));
+    accum_bg_128 = _mm_add_epi16 (_mm256_extracti128_si256 (accum_bg, 0),
+                                  _mm256_extracti128_si256 (accum_bg, 1));
     ((guint64 *) accums_out) [0] =
-        (guint64) _mm_extract_epi64 (accum_fg_128, 0)
-        + (guint64) _mm_extract_epi64 (accum_fg_128, 1);
-
-    accum_bg [0] = _mm256_add_epi16 (accum_bg [0], accum_bg [1]);
-    accum_bg_128 = _mm_add_epi16 (_mm256_extracti128_si256 (accum_bg [0], 0),
-                                  _mm256_extracti128_si256 (accum_bg [0], 1));
-    ((guint64 *) accums_out) [1] =
         (guint64) _mm_extract_epi64 (accum_bg_128, 0)
         + (guint64) _mm_extract_epi64 (accum_bg_128, 1);
+
+    accum_fg_128 = _mm_add_epi16 (_mm256_extracti128_si256 (accum_fg, 0),
+                                  _mm256_extracti128_si256 (accum_fg, 1));
+    ((guint64 *) accums_out) [1] =
+        (guint64) _mm_extract_epi64 (accum_fg_128, 0)
+        + (guint64) _mm_extract_epi64 (accum_fg_128, 1);
 }