Then add the early-outs for NEON too.

hrydgard · Jun 13, 2023 · 0eb3702 · 0eb3702
1 parent 9647872
commit 0eb3702
Showing 1 changed file with 16 additions and 12 deletions.
diff --git a/GPU/Common/IndexGenerator.cpp b/GPU/Common/IndexGenerator.cpp
@@ -163,19 +163,23 @@ void IndexGenerator::AddStrip(int numVerts, bool clockwise) {
 	u16 *dst = inds_;
 	uint16x8_t offsets0 = vaddq_u16(ibase8, vld1q_u16(offsets));
 	vst1q_u16(dst, offsets0);
-	uint16x8_t offsets1 = vaddq_u16(ibase8, vld1q_u16(offsets + 8));
-	vst1q_u16(dst + 8, offsets1);
-	uint16x8_t offsets2 = vaddq_u16(ibase8, vld1q_u16(offsets + 16));
-	vst1q_u16(dst + 16, offsets2);
-	uint16x8_t increment = vdupq_n_u16(8);
-	for (int i = 1; i < numChunks; i++) {
-		dst += 3 * 8;
-		offsets0 = vaddq_u16(offsets0, increment);
-		offsets1 = vaddq_u16(offsets1, increment);
-		offsets2 = vaddq_u16(offsets2, increment);
-		vst1q_u16(dst, offsets0);
+	if (numTris > 2) {
+		uint16x8_t offsets1 = vaddq_u16(ibase8, vld1q_u16(offsets + 8));
 		vst1q_u16(dst + 8, offsets1);
-		vst1q_u16(dst + 16, offsets2);
+		if (numTris > 5) {
+			uint16x8_t offsets2 = vaddq_u16(ibase8, vld1q_u16(offsets + 16));
+			vst1q_u16(dst + 16, offsets2);
+			uint16x8_t increment = vdupq_n_u16(8);
+			for (int i = 1; i < numChunks; i++) {
+				dst += 3 * 8;
+				offsets0 = vaddq_u16(offsets0, increment);
+				offsets1 = vaddq_u16(offsets1, increment);
+				offsets2 = vaddq_u16(offsets2, increment);
+				vst1q_u16(dst, offsets0);
+				vst1q_u16(dst + 8, offsets1);
+				vst1q_u16(dst + 16, offsets2);
+			}
+		}
 	}
 	inds_ += numTris * 3;
 #else