[simd/v3i]: Refactor to use loops for computations with >8 lanes

titzer · Oct 15, 2023 · b578ee4 · b578ee4
1 parent 4cf612f
commit b578ee4
Showing 1 changed file with 34 additions and 33 deletions.
diff --git a/src/engine/V3Eval.v3 b/src/engine/V3Eval.v3
@@ -432,43 +432,44 @@ component V3Eval {
 		return ((u64.view(r1) << 32) | r0, (u64.view(r3) << 32) | r2);
 	}
 	private def do_vv_v_x8(a: (u64, u64), b: (u64, u64), f: (u16, u16) -> u16) -> (u64, u64) { // Performs an 8-lane binop
-		var r0 = f(u16.view(a.0), u16.view(b.0));
-		var r1 = f(u16.view(a.0 >> 16), u16.view(b.0 >> 16));
-		var r2 = f(u16.view(a.0 >> 32), u16.view(b.0 >> 32));
-		var r3 = f(u16.view(a.0 >> 48), u16.view(b.0 >> 48));
-		var r4 = f(u16.view(a.1), u16.view(b.1));
-		var r5 = f(u16.view(a.1 >> 16), u16.view(b.1 >> 16));
-		var r6 = f(u16.view(a.1 >> 32), u16.view(b.1 >> 32));
-		var r7 = f(u16.view(a.1 >> 48), u16.view(b.1 >> 48));
-		return ((u64.view(r3) << 48) | (u64.view(r2) << 32) | (u64.view(r1) << 16) | r0,
-			(u64.view(r7) << 48) | (u64.view(r6) << 32) | (u64.view(r5) << 16) | r4);
+		var result0: u64 = 0;
+		var result1: u64 = 0;
+
+		for (i < 4) {
+			var shift_amount: byte = byte.!(i * 16);
+
+			var r_a = u16.view((a.0 >> shift_amount) & 0xFFFF);
+			var r_b = u16.view((b.0 >> shift_amount) & 0xFFFF);
+			var res = f(r_a, r_b);
+			result0 |= (u64.view(res) << shift_amount);
+
+			r_a = u16.view((a.1 >> shift_amount) & 0xFFFF);
+			r_b = u16.view((b.1 >> shift_amount) & 0xFFFF);
+			res = f(r_a, r_b);
+			result1 |= (u64.view(res) << shift_amount);
+		}
+
+		return (result0, result1);
 	}
 	private def do_vv_v_x16(a: (u64, u64), b: (u64, u64), f: (u8, u8) -> u8) -> (u64, u64) { // Performs a 16-lane binop
-		var r0  = f(u8.view(a.0), u8.view(b.0));
-		var r1  = f(u8.view(a.0 >> 8), u8.view(b.0 >> 8));
-		var r2  = f(u8.view(a.0 >> 16), u8.view(b.0 >> 16));
-		var r3  = f(u8.view(a.0 >> 24), u8.view(b.0 >> 24));
-		var r4  = f(u8.view(a.0 >> 32), u8.view(b.0 >> 32));
-		var r5  = f(u8.view(a.0 >> 40), u8.view(b.0 >> 40));
-		var r6  = f(u8.view(a.0 >> 48), u8.view(b.0 >> 48));
-		var r7  = f(u8.view(a.0 >> 56), u8.view(b.0 >> 56));
-
-		var r8  = f(u8.view(a.1), u8.view(b.1));
-		var r9  = f(u8.view(a.1 >> 8), u8.view(b.1 >> 8));
-		var r10 = f(u8.view(a.1 >> 16), u8.view(b.1 >> 16));
-		var r11 = f(u8.view(a.1 >> 24), u8.view(b.1 >> 24));
-		var r12 = f(u8.view(a.1 >> 32), u8.view(b.1 >> 32));
-		var r13 = f(u8.view(a.1 >> 40), u8.view(b.1 >> 40));
-		var r14 = f(u8.view(a.1 >> 48), u8.view(b.1 >> 48));
-		var r15 = f(u8.view(a.1 >> 56), u8.view(b.1 >> 56));
+		var result0: u64 = 0;
+		var result1: u64 = 0;
 
-		return (
-			(u64.view(r7) << 56) | (u64.view(r6) << 48) | (u64.view(r5) << 40) | (u64.view(r4) << 32) |
-			(u64.view(r3) << 24) | (u64.view(r2) << 16) | (u64.view(r1) << 8) | r0,
+		for (i < 8) {
+			var shift_amount: byte = byte.!(i * 8);
+
+			var r_a = u8.view((a.0 >> shift_amount) & 0xFF);
+			var r_b = u8.view((b.0 >> shift_amount) & 0xFF);
+			var res = f(r_a, r_b);
+			result0 |= (u64.view(res) << shift_amount);
+
+			r_a = u8.view((a.1 >> shift_amount) & 0xFF);
+			r_b = u8.view((b.1 >> shift_amount) & 0xFF);
+			res = f(r_a, r_b);
+			result1 |= (u64.view(res) << shift_amount);
+		}
 
-			(u64.view(r15) << 56) | (u64.view(r14) << 48) | (u64.view(r13) << 40) | (u64.view(r12) << 32) |
-			(u64.view(r11) << 24) | (u64.view(r10) << 16) | (u64.view(r9) << 8) | r8
-		);
+		return (result0, result1);
 	}
 	private def canonf(a: float) -> float {
 		return if(a == a, a, float.nan);