avx2: Improve squared error computation

A further 10-20% improvement in overall performance with -w 9 (GCC 13/Haswell). Clean up a little.
hpjansson · Apr 21, 2024 · 5e9a6ea · 5e9a6ea
1 parent 301a715
commit 5e9a6ea
Showing 1 changed file with 16 additions and 27 deletions.
diff --git a/chafa/internal/chafa-avx2.c b/chafa/internal/chafa-avx2.c
@@ -29,11 +29,11 @@ calc_error_avx2 (const ChafaPixel *pixels, const ChafaColorPair *color_pair,
                  const guint32 *sym_mask_u32)
 {
     __m256i err_8x_u32 = { 0 };
-    const gint32 *e = (gint32 *) &err_8x_u32;
+    __m128i err_4x_u32;
     __m128i fg_4x_u32, bg_4x_u32;
     __m256i fg_4x_u64, bg_4x_u64;
-    const __m256i *pixels_8x_p = (const __m256i *) pixels;
-    const __m256i *sym_mask_8x_p = (const __m256i *) sym_mask_u32;
+    const __m128i *pixels_4x_p = (const __m128i *) pixels;
+    const __m128i *sym_mask_4x_p = (const __m128i *) sym_mask_u32;
     gint i;
 
     fg_4x_u32 = _mm_set1_epi32 (CHAFA_COLOR8_U32 (color_pair->colors [CHAFA_COLOR_PAIR_FG]));
@@ -42,41 +42,30 @@ calc_error_avx2 (const ChafaPixel *pixels, const ChafaColorPair *color_pair,
     bg_4x_u32 = _mm_set1_epi32 (CHAFA_COLOR8_U32 (color_pair->colors [CHAFA_COLOR_PAIR_BG]));
     bg_4x_u64 = _mm256_cvtepu8_epi16 (bg_4x_u32);
 
-    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;
+        __m128i pixels_4x, sym_mask_4x;
         __m256i p0, m0, fg0, bg0, d0;
-        __m256i p1, m1, fg1, bg1, d1;
 
-        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++;
+        pixels_4x = _mm_loadu_si128 (pixels_4x_p++);
+        sym_mask_4x = _mm_loadu_si128 (sym_mask_4x_p++);
 
-        p0 = _mm256_cvtepu8_epi16 (_mm256_extracti128_si256 (pixels_8x, 0));
-        m0 = _mm256_cvtepi8_epi16 (_mm256_extracti128_si256 (sym_mask_8x, 0));
+        p0 = _mm256_cvtepu8_epi16 (pixels_4x);
+        m0 = _mm256_cvtepi8_epi16 (sym_mask_4x);
         fg0 = _mm256_and_si256 (m0, _mm256_sub_epi16 (fg_4x_u64, p0));
         bg0 = _mm256_andnot_si256 (m0, _mm256_sub_epi16 (bg_4x_u64, p0));
         d0 = _mm256_or_si256 (fg0, bg0);
-        d0 = _mm256_mullo_epi16 (d0, d0);
-        d0 = _mm256_add_epi32 (_mm256_cvtepu16_epi32 (_mm256_extracti128_si256 (d0, 0)),
-                               _mm256_cvtepu16_epi32 (_mm256_extracti128_si256 (d0, 1)));
-
-        p1 = _mm256_cvtepu8_epi16 (_mm256_extracti128_si256 (pixels_8x, 1));
-        m1 = _mm256_cvtepi8_epi16 (_mm256_extracti128_si256 (sym_mask_8x, 1));
-        fg1 = _mm256_and_si256 (m1, _mm256_sub_epi16 (fg_4x_u64, p1));
-        bg1 = _mm256_andnot_si256 (m1, _mm256_sub_epi16 (bg_4x_u64, p1));
-        d1 = _mm256_or_si256 (fg1, bg1);
-        d1 = _mm256_mullo_epi16 (d1, d1);
-        d1 = _mm256_add_epi32 (_mm256_cvtepu16_epi32 (_mm256_extracti128_si256 (d1, 0)),
-                               _mm256_cvtepu16_epi32 (_mm256_extracti128_si256 (d1, 1)));
+        d0 = _mm256_madd_epi16 (d0, d0);
 
         err_8x_u32 = _mm256_add_epi32 (err_8x_u32, d0);
-        err_8x_u32 = _mm256_add_epi32 (err_8x_u32, d1);
     }
 
-    return e [0] + e [1] + e [2] + e [4] + e [5] + e [6];
+    err_4x_u32 = _mm_add_epi32 (_mm256_extracti128_si256 (err_8x_u32, 0),
+                                _mm256_extracti128_si256 (err_8x_u32, 1));
+    err_4x_u32 = _mm_hadd_epi32 (err_4x_u32, err_4x_u32);
+    err_4x_u32 = _mm_hadd_epi32 (err_4x_u32, err_4x_u32);
+
+    return _mm_extract_epi32 (err_4x_u32, 0);
 }
 
 void