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

[avr-crypto-lib.git] / keccak / keccak-asm.S

diff --git a/keccak/keccak-asm.S b/keccak/keccak-asm.S
index 6887a755d573fded11f9ec6e73d72a7907c706f9..9ccd79e0d8c9d3d7c6dafdb0c09b119620276347 100644 (file)
--- a/keccak/keccak-asm.S
+++ b/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];
@@ -37,18 +39,101 @@
        }
 */
 
-.global keccak_theta
-keccak_theta:
-       movw r30, r24
-       movw r26, r22
+/*********************************************
+ * theta_2a
+ *********************************************
+       input:
+               r24:r25 = a ; uint64_t a[5][5]
+               X = b       ; uint64_t *b
+       output:
+               a[0..4][0] ^= b
+               r20 = 0
+               r21 = XX
+               r22 = XX
+               r24:r25 += 8
+               X += 8
+               Z = r24:r25 + 7 + 4 * 40
+*/
+theta_2a:
+       ldi r20, 8
+10:
+       movw ZL, r24
+       ld  r21, X+
+       .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
+       ret
 
-;      .irp offset, 0, 1, 2, 3, 4
+/*********************************************
+ * 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
+       movw r26, r22 ; X = b
        ldi r19, 5
 10:
        ldi r20, 8
 20:
-
        ld  r22, Z
        adiw ZL, 40
        ld  r21, Z
@@ -62,10 +147,8 @@ keccak_theta:
        adiw ZL, 40
        ld  r21, Z
        eor r22, r21
-
        adiw r24, 1
        movw r30, r24
-
        st X+, r22
        dec r20
        brne 20b
@@ -73,6 +156,57 @@ keccak_theta:
        adiw XL, 8 * 4
        dec r19
        brne 10b
-;      .endr
+/*
+       for(i = 0; i < 5; ++i){
+               for(j = 0; j < 5; ++j){
+                       a[j][i] ^= b[(4 + i) % 5][0];
+               }
+       }
+
+*/
+/* 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)
+       sbci XH, hi8(4 * 5 * 8 + 8)
+       rcall theta_2a
+/* a[0..4][2]{0..7} ^= b[1][0]{0..7} */
+       adiw XL, 4 * 8
+       rcall theta_2a
+/* a[0..4][3]{0..7} ^= b[2][0]{0..7} */
+       adiw XL, 4 * 8
+       rcall theta_2a
+/* a[0..4][4]{0..7} ^= b[3][0]{0..7} */
+       adiw XL, 4 * 8
+       rcall theta_2a
+/*
+       for(i = 0; i < 5; ++i){
+        for(j = 0; j < 5; ++j){
+            a[j][i] ^= rotate64_1bit_left(b[(i + 1) % 5][0]);
+        }
+    }
+*/
+/* a[0..4][0]{0..7} ^= rol(b[1][0]{0..7}) */
+       subi r24, lo8(5 * 8 - 1)
+       sbci r25, hi8(5 * 8 - 1)
+       subi XL, lo8(2 * 5 * 8 + 8)
+       sbci XH, hi8(2 * 5 * 8 + 8)
+       rcall theta_2b
+/* a[0..4][1]{0..7} ^= rol(b[2][0]{0..7}) */
+       adiw XL, 4 * 8
+       rcall theta_2b
+/* a[0..4][21]{0..7} ^= rol(b[3][0]{0..7}) */
+       adiw XL, 4 * 8
+       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