vtgl

Commit: 73e7573e49b63dbf70c55427fa3dfb5df6102554
Parent: c20911f58d50f6a446c0c4959fc36a7b97018483
Author: Randy Palamar
Date:   Wed, 30 Oct 2024 06:10:46 -0600

simd mem_copy

Diffstat:
M
extern/stb_truetype.h
 | 
6
++----
M
terminal.c
 | 
2
+-
M
util.c
 | 
16
+++++++++++-----

3 files changed, 14 insertions(+), 10 deletions(-)
diff --git a/extern/stb_truetype.h b/extern/stb_truetype.h
@@ -1467,10 +1467,8 @@ static int stbtt__GetGlyphShapeTT(Arena *a, const stbtt_fontinfo *info, int glyp
             // Append vertices.
             tmp = alloc(a, stbtt_vertex, num_vertices + comp_num_verts);
             if (num_vertices > 0 && vertices)
-               mem_copy((s8){.len = num_vertices * sizeof(stbtt_vertex), .data = (u8 *)vertices},
-		        (s8){.len = (num_vertices + comp_num_verts) * sizeof(stbtt_vertex), .data = (u8 *)tmp});
-            mem_copy((s8){.len = comp_num_verts * sizeof(stbtt_vertex), .data = (u8 *)comp_verts},
-                     (s8){.len = comp_num_verts * sizeof(stbtt_vertex), .data = (u8 *)(tmp + num_vertices)});
+               mem_copy(vertices, tmp, num_vertices * sizeof(stbtt_vertex));
+            mem_copy(comp_verts, tmp + num_vertices, comp_num_verts * sizeof(stbtt_vertex));
             vertices = tmp;
             num_vertices += comp_num_verts;
          }
diff --git a/terminal.c b/terminal.c
@@ -1469,7 +1469,7 @@ handle_input(Term *t, Arena a, s8 raw)
 					feed_line(&tv->lines, old, t->cursor.style);
 					TermView *nv = t->views + t->view_idx;
 					size nstart  = nv->log.widx;
-					mem_copy(raw, (s8){nv->log.cap, nv->log.buf + nstart});
+					mem_copy(raw.data, nv->log.buf + nstart, raw.len);
 					commit_to_rb(tv, -raw.len);
 					commit_to_rb(nv,  raw.len);
 					raw.data = nv->log.buf + nstart;
diff --git a/util.c b/util.c
@@ -53,12 +53,18 @@ normalize_range(Range r)
 }
 
 static void
-mem_copy(s8 src, s8 dest)
+mem_copy(void *src, void *dest, size len)
 {
-	ASSERT(src.len > 0 && dest.len > 0);
-	ASSERT(dest.len >= src.len);
-	for (size i = 0; i < src.len; i++)
-		dest.data[i] = src.data[i];
+	ASSERT(len >= 0);
+	u8 *s = src, *d = dest;
+#if defined(__AVX512BW__)
+	/* TODO: aligned load/store and comparison */
+	for (; len >= 64; len -= 64, s += 64, d += 64)
+		_mm512_storeu_epi8(d, _mm512_loadu_epi8(s));
+#endif
+	for (; len >= 16; len -= 16, s += 16, d += 16)
+		_mm_storeu_si128((__m128i *)d, _mm_loadu_si128((__m128i*)s));
+	for (; len; len--) *d++ = *s++;
 }
 
 static void *