From 1e75d637065f4c0969b2d83f4f4cbe96ea24ebc9 Mon Sep 17 00:00:00 2001
From: itzpr3d4t0r <103119829+itzpr3d4t0r@users.noreply.github.com>
Date: Sun, 15 Oct 2023 11:14:05 +0200
Subject: [PATCH] removed a comment, simplified bpp calculations

---
 src_c/simd_fill.h              |  2 --
 src_c/simd_surface_fill_avx2.c | 13 +++++--------
 2 files changed, 5 insertions(+), 10 deletions(-)

diff --git a/src_c/simd_fill.h b/src_c/simd_fill.h
index 36d98afb09..05c30d095a 100644
--- a/src_c/simd_fill.h
+++ b/src_c/simd_fill.h
@@ -4,8 +4,6 @@
 int
 _pg_has_avx2();
 
-// SSE2 functions
-
 // AVX2 functions
 int
 surface_fill_blend_add_avx2(SDL_Surface *surface, SDL_Rect *rect,
diff --git a/src_c/simd_surface_fill_avx2.c b/src_c/simd_surface_fill_avx2.c
index aedfaffb11..101a20a2a0 100644
--- a/src_c/simd_surface_fill_avx2.c
+++ b/src_c/simd_surface_fill_avx2.c
@@ -29,19 +29,16 @@ _pg_has_avx2()
 #define SETUP_AVX2_FILLER(COLOR_PROCESS_CODE)                                 \
     /* initialize surface data */                                             \
     int width = rect->w, height = rect->h;                                    \
-    int bpp = surface->format->BytesPerPixel;                                 \
-    int skip = (surface->pitch - width * bpp) >> 2;                           \
-    int pxl_skip = bpp >> 2;                                                  \
+    int skip = surface->pitch / 4 - width;                                    \
     /* indicates the number of pixels that can't be processed in 8-pixel      \
      * blocks */                                                              \
     int pxl_excess = width % 8;                                               \
     /* indicates the number of 8-pixel blocks that can be processed */        \
     int n_iters_8 = width / 8;                                                \
-    int excess_skip = pxl_excess * pxl_skip, block_skip = pxl_skip * 8;       \
     int i;                                                                    \
     /* load pixel data */                                                     \
-    Uint32 *pixels = (Uint32 *)surface->pixels +                              \
-                     rect->y * (surface->pitch >> 2) + rect->x * pxl_skip;    \
+    Uint32 *pixels =                                                          \
+        (Uint32 *)surface->pixels + rect->y * (surface->pitch / 4) + rect->x; \
                                                                               \
     __m256i mm256_dst;                                                        \
     __m256i mask =                                                            \
@@ -69,7 +66,7 @@ _pg_has_avx2()
             /* store 8 pixels */                                    \
             _mm256_storeu_si256((__m256i *)pixels, mm256_dst);      \
                                                                     \
-            pixels += block_skip;                                   \
+            pixels += 8;                                            \
         }                                                           \
                                                                     \
         if (pxl_excess) {                                           \
@@ -81,7 +78,7 @@ _pg_has_avx2()
             /* store up to 7 pixels */                              \
             _mm256_maskstore_epi32((int *)pixels, mask, mm256_dst); \
                                                                     \
-            pixels += excess_skip;                                  \
+            pixels += pxl_excess;                                   \
         }                                                           \
                                                                     \
         pixels += skip;                                             \