ChafaWorkCell: Add AVX2 implementation for mean color pair extraction

It's about twice as fast as the MMX impl. in my tests. This makes a big difference with -w 9. The improvement is much smaller in the popcount-accelerated paths.
hpjansson · Apr 21, 2024 · 0e7d5fb · 0e7d5fb
1 parent dc17c01
commit 0e7d5fb
Show file tree

Hide file tree

Showing 3 changed files with 60 additions and 1 deletion.
diff --git a/chafa/internal/chafa-avx2.c b/chafa/internal/chafa-avx2.c
@@ -78,3 +78,56 @@ calc_error_avx2 (const ChafaPixel *pixels, const ChafaColorPair *color_pair,
 
     return e [0] + e [1] + e [2] + e [4] + e [5] + e [6];
 }
+
+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 } };
+    __m128i accum_fg_128;
+    __m128i accum_bg_128;
+    gint i;
+
+    for (i = 0; i < CHAFA_SYMBOL_N_PIXELS / 8; 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++;
+
+        p0 = _mm256_andnot_si256 (sym_mask_8x, pixels_8x);
+        p1 = _mm256_and_si256 (sym_mask_8x, pixels_8x);
+
+        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 [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));
+    ((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);
+}
diff --git a/chafa/internal/chafa-private.h b/chafa/internal/chafa-private.h
@@ -217,6 +217,8 @@ gint calc_error_sse41 (const ChafaPixel *pixels, const ChafaColorPair *color_pai
 #ifdef HAVE_AVX2_INTRINSICS
 gint calc_error_avx2 (const ChafaPixel *pixels, const ChafaColorPair *color_pair,
                       const guint32 *sym_mask_u32) G_GNUC_PURE;
+void calc_colors_avx2 (const ChafaPixel *pixels, ChafaColorAccum *accums_out,
+                       const guint32 *sym_mask_u32);
 #endif
 
 #if defined(HAVE_POPCNT64_INTRINSICS) || defined(HAVE_POPCNT32_INTRINSICS)

diff --git a/chafa/internal/chafa-work-cell.c b/chafa/internal/chafa-work-cell.c
@@ -83,7 +83,11 @@ chafa_work_cell_get_mean_colors_for_symbol (const ChafaWorkCell *wcell, const Ch
     const guint8 *covp = (guint8 *) &sym->coverage [0];
     ChafaColorAccum accums [2] = { 0 };
 
-#ifdef HAVE_MMX_INTRINSICS
+#ifdef HAVE_AVX2_INTRINSICS
+    if (chafa_have_avx2 ())
+        calc_colors_avx2 (wcell->pixels, accums, sym->mask_u32);
+    else
+#elif HAVE_MMX_INTRINSICS
     if (chafa_have_mmx ())
         calc_colors_mmx (wcell->pixels, accums, covp);
     else