progress in asm version of keccak (theta now in asm)

keccak/keccak-asm.S

@@ -28,6 +28,8 @@
 .nolist
 #include "avr-asm-macros.S"
 .list
+
+.equ __zero_reg__, 1
 /*
 	void keccak_theta (uint64_t *a, uint64_t *b){
 	// uint64_t b[5][5];
@@ -37,8 +39,9 @@
   	}
 */
 
-theta_2a:
+/*********************************************
-/*
+ * theta_2a
+ *********************************************
 	input:
 		r24:r25 = a ; uint64_t a[5][5]
 		X = b       ; uint64_t *b
@@ -51,6 +54,7 @@ theta_2a:
 		X += 8
 		Z = r24:r25 + 7 + 4 * 40
 */
+theta_2a:
 	ldi r20, 8
 10:
 	movw ZL, r24
@@ -68,6 +72,60 @@ theta_2a:
 	brne 10b
 	ret
 
+/*********************************************
+ * theta_2b
+ *********************************************
+	input:
+		r24:r25 = a+1 ; uint64_t a[5][5]
+		X = b       ; uint64_t *b
+	output:
+		a[0..4][0] ^= rol(b,1)
+		r19 = XX
+		r20 = 0
+		r21 = XX
+		r22 = XX
+		r24:r25 += 8
+		X += 8
+		Z = r24:r25 + 7 + 4 * 40
+*/
+theta_2b:
+	ldi r20, 7
+	ld r19, X+
+	lsl r19
+	rol __zero_reg__
+10:
+	movw ZL, r24
+	ld  r21, X+
+	ror __zero_reg__
+	rol r21
+	rol __zero_reg__
+	.irp r, 0, 1, 2, 3, 4
+		ld  r22, Z
+		eor r22, r21
+		st  Z, r22
+	.if \r != 4
+		adiw ZL, 40
+	.endif
+	.endr
+	adiw r24, 1
+	dec r20
+	brne 10b
+	add r19, __zero_reg__
+	sbiw r24, 8
+	movw ZL, r24
+	.irp r, 0, 1, 2, 3, 4
+		ld  r22, Z
+		eor r22, r19
+		st  Z, r22
+	.if \r != 4
+		adiw ZL, 40
+	.endif
+	.endr
+	adiw r24, 9
+	clr __zero_reg__
+	ret
+
+
 .global keccak_theta
 keccak_theta:
 	movw r30, r24 ; Z = a
@@ -104,17 +162,11 @@ keccak_theta:
 			a[j][i] ^= b[(4 + i) % 5][0];
 		}
 	}
-	for(i = 0; i < 5; ++i){
-        for(j = 0; j < 5; ++j){
-            a[j][i] ^= rotate64_1bit_left(b[(i + 1) % 5][0]);
-        }
-    }
 
 */
-	sbiw XL, 5 * 8
-
-	sbiw r24, 40
 /* a[0..4][0]{0..7} ^= b[4][0]{0..7} */
+	sbiw XL, 5 * 8
+	sbiw r24, 40
 	rcall theta_2a
 /* a[0..4][1]{0..7} ^= b[0][0]{0..7} */
 	subi XL, lo8(4 * 5 * 8 + 8)
@@ -129,25 +181,32 @@ keccak_theta:
 /* a[0..4][4]{0..7} ^= b[3][0]{0..7} */
 	adiw XL, 4 * 8
 	rcall theta_2a
-
+/*
-	ret
+	for(i = 0; i < 5; ++i){
-
+        for(j = 0; j < 5; ++j){
-	ldi r20, 8
+            a[j][i] ^= rotate64_1bit_left(b[(i + 1) % 5][0]);
-
+        }
-10:
+    }
-	movw ZL, r24
+*/
-	ld  r21, X+
+/* a[0..4][0]{0..7} ^= rol(b[1][0]{0..7}) */
-	.irp r, 0, 1, 2, 3, 4
+	subi r24, lo8(5 * 8 - 1)
-		ld  r22, Z
+	sbci r25, hi8(5 * 8 - 1)
-		eor r22, r21
+	subi XL, lo8(2 * 5 * 8 + 8)
-		st  Z, r22
+	sbci XH, hi8(2 * 5 * 8 + 8)
-	.if \r != 4
+	rcall theta_2b
-		adiw ZL, 40
+/* a[0..4][1]{0..7} ^= rol(b[2][0]{0..7}) */
-	.endif
+	adiw XL, 4 * 8
-	.endr
+	rcall theta_2b
-	adiw r24, 1
+/* a[0..4][21]{0..7} ^= rol(b[3][0]{0..7}) */
-	dec r20
+	adiw XL, 4 * 8
-	brne 10b
+	rcall theta_2b
+/* a[0..4][3]{0..7} ^= rol(b[4][0]{0..7}) */
+	adiw XL, 4 * 8
+	rcall theta_2b
+/* a[0..4][4]{0..7} ^= rol(b[0][0]{0..7}) */
+	subi XL, lo8(4 * 5 * 8 + 8)
+	sbci XH, hi8(4 * 5 * 8 + 8)
+	rcall theta_2b
 
 
 	ret

keccak/keccak-stub.c

@@ -30,7 +30,7 @@
 #  undef DEBUG
 #endif
 
-#define DEBUG 1
+#define DEBUG 0
 
 #if DEBUG
 #include "cli.h"
@@ -107,20 +107,7 @@ void keccak_round(uint64_t a[5][5], uint8_t rci){
 			uint8_t v8[8];
 		} t;
 	/* theta */
-#if DEBUG & 0
-    printf_P(PSTR("\npre-theta(1):"));
-    keccak_dump_state(b);
-#endif
 	keccak_theta((uint64_t*)a, (uint64_t*)b);
-#if DEBUG
-    printf_P(PSTR("\npost-theta(1):"));
-	keccak_dump_state(b);
-#endif
-	for(i = 0; i < 5; ++i){
-        for(j = 0; j < 5; ++j){
-            a[j][i] ^= rotate64_1bit_left(b[(i + 1) % 5][0]);
-        }
-    }
 #if DEBUG
 	cli_putstr_P(PSTR("\r\nAfter theta:"));
 	keccak_dump_state(a);