[vlc-devel] [PATCH 01/16] block cipher: improve performance
Serg Chernyavskiy
glenvt18 at gmail.com
Tue Jul 7 02:50:27 CEST 2015
No, it doesn't. The input is assumed unaligned. testbitslice.c takes care
of generating packets with random offsets and sizes. All dvbcsa_bs_word
buffers are "naturally" aligned by a compiler. Unaligned access is handled
by inline functions dvbcsa_xor_XX, dvbcsa_copy_XX, dvbcsa_load/store_XX
(dvbcsa_pv.h) and it does have noticeable effect on performance. As for the
scratch buffers, the only restriction is that scratch2 must be at least 16
bit aligned (for patch #7). All bs words in the current implementation are
at least 32 bit aligned. BTW, neon uint64x2_t words are 64 bit aligned by a
compiler, not 128. In other words, as long as something like uint8x16_t is
not used as a bs word for scratch2 it's OK. I also tried to manually align
buffers at 128 bit forcing gcc generate vld1 @128 instructions and didn't
notice any performance changes on cortex-a7.
>Does this add latency?
No, actually, the scratch buffers remove latency. This patch could as well
be titled "fix block cipher performance on modern x86 architectures".
Doesn't have that much effect on ARM.
2015-07-06 19:08 GMT+03:00 Jean-Baptiste Kempf <jb at videolan.org>:
> LGTM.
> Does this add a special requirement to the input, like alignment?
> Does this add latency?
>
> On 26 Jun, glenvt18 wrote :
> > 1. Use a dvbcsa_bs_world_t x 8 scratch buffer between the output of sbox
> and the input of
> > the permute logic. It reduces the dependency between sbox output and
> permute logic input.
> > 2x speed-up of CSA on x86 with both uint64 and sse2.
> >
> > 2. Lookup 4 values at a time (a, b, c, d) - reduces dependency on a read
> value.
> >
> > 3. Use separate scratch buffers for reading and writing.
> >
> > Changes 2 and 3 add about 5% depending on the architecture.
> > ---
> > src/dvbcsa_bs_block.c | 96
> +++++++++++++++++++++++++++++++++------------------
> > 1 file changed, 63 insertions(+), 33 deletions(-)
> >
> > diff --git a/src/dvbcsa_bs_block.c b/src/dvbcsa_bs_block.c
> > index 58f6ca9..e4f532c 100644
> > --- a/src/dvbcsa_bs_block.c
> > +++ b/src/dvbcsa_bs_block.c
> > @@ -31,32 +31,48 @@
> > DVBCSA_INLINE static inline void
> > dvbcsa_bs_block_decrypt_register (const dvbcsa_bs_word_t *block,
> dvbcsa_bs_word_t *r)
> > {
> > - int i, j, g;
> > + dvbcsa_bs_word_t scratch1[8];
> > + dvbcsa_bs_word_t scratch2[8];
> > + int i, j, g;
> >
> > r += 8 * 56;
> >
> > - // loop over kk[55]..kk[0]
> > + /* loop over kk[55]..kk[0] */
> > for (i = 55; i >= 0; i--)
> > {
> > dvbcsa_bs_word_t *r6_N = r + 8 * 6;
> >
> > - r -= 8; /* virtual shift of registers */
> > + r -= 8; /* virtual shift of registers */
> >
> > for (g = 0; g < 8; g++)
> > - {
> > - union {
> > - dvbcsa_bs_word_t so;
> > - uint8_t si[BS_BATCH_BYTES];
> > - } u;
> > + scratch1[g] = BS_XOR(block[i], r6_N[g]);
> > +
> > + /* sbox */
> > + {
> > + uint8_t *p1, *p2;
> > + uint8_t a, b, c, d;
> > +
> > + p1 = (uint8_t *)scratch1;
> > + p2 = (uint8_t *)scratch2;
> > + for (j = 0; j < BS_BATCH_BYTES * 8; j += 4)
> > + {
> > + a = p1[j + 0];
> > + b = p1[j + 1];
> > + c = p1[j + 2];
> > + d = p1[j + 3];
> > + p2[j + 0] = dvbcsa_block_sbox[a];
> > + p2[j + 1] = dvbcsa_block_sbox[b];
> > + p2[j + 2] = dvbcsa_block_sbox[c];
> > + p2[j + 3] = dvbcsa_block_sbox[d];
> > + }
> > + }
> >
> > - u.so = BS_XOR(block[i], r6_N[g]);
> > -
> > - for (j = 0; j < BS_BATCH_BYTES; j++)
> > - u.si[j] = dvbcsa_block_sbox[u.si[j]];
> > -
> > - dvbcsa_bs_word_t sbox_out = u.so;
> > + for (g = 0; g < 8; g++)
> > + {
> > + dvbcsa_bs_word_t sbox_out = scratch2[g];
> > + dvbcsa_bs_word_t w;
> >
> > - // bit permutation
> > + /* bit permutation */
> >
> > dvbcsa_bs_word_t in = BS_OR(
> > BS_OR(
> > @@ -67,8 +83,7 @@ dvbcsa_bs_block_decrypt_register (const
> dvbcsa_bs_word_t *block, dvbcsa_bs_word_
> > BS_OR( BS_SHR (BS_AND
> (sbox_out, BS_VAL8(40)), 6),
> > BS_SHR (BS_AND
> (sbox_out, BS_VAL8(80)), 4)));
> >
> > - dvbcsa_bs_word_t w = BS_XOR(r[8 * 8 + g], sbox_out);
> > -
> > + w = BS_XOR(r[8 * 8 + g], sbox_out);
> > r[8 * 0 + g] = w;
> > BS_XOREQ(r[8 * 2 + g], w);
> > BS_XOREQ(r[8 * 3 + g], w);
> > @@ -116,30 +131,46 @@ void dvbcsa_bs_block_decrypt_batch(const struct
> dvbcsa_bs_key_s *key,
> > DVBCSA_INLINE static inline void
> > dvbcsa_bs_block_encrypt_register (const dvbcsa_bs_word_t *block,
> dvbcsa_bs_word_t *r)
> > {
> > - int i, j, g;
> > + dvbcsa_bs_word_t scratch1[8];
> > + dvbcsa_bs_word_t scratch2[8];
> > + int i, j, g;
> >
> > - // loop over kk[55]..kk[0]
> > + /* loop over kk[55]..kk[0] */
> > for (i = 0; i < 56; i++)
> > {
> > dvbcsa_bs_word_t *r7_N = r + 8 * 7;
> >
> > - r += 8; /* virtual shift of registers */
> > + r += 8; /* virtual shift of registers */
> >
> > for (g = 0; g < 8; g++)
> > - {
> > - union {
> > - dvbcsa_bs_word_t so;
> > - uint8_t si[BS_BATCH_BYTES];
> > - } u;
> > + scratch1[g] = BS_XOR(block[i], r7_N[g]);
> > +
> > + /* sbox */
> > + {
> > + uint8_t *p1, *p2;
> > + uint8_t a, b, c, d;
> > +
> > + p1 = (uint8_t *)scratch1;
> > + p2 = (uint8_t *)scratch2;
> > + for (j = 0; j < BS_BATCH_BYTES * 8; j += 4)
> > + {
> > + a = p1[j + 0];
> > + b = p1[j + 1];
> > + c = p1[j + 2];
> > + d = p1[j + 3];
> > + p2[j + 0] = dvbcsa_block_sbox[a];
> > + p2[j + 1] = dvbcsa_block_sbox[b];
> > + p2[j + 2] = dvbcsa_block_sbox[c];
> > + p2[j + 3] = dvbcsa_block_sbox[d];
> > + }
> > + }
> >
> > - u.so = BS_XOR(block[i], r7_N[g]);
> > -
> > - for (j = 0; j < BS_BATCH_BYTES; j++)
> > - u.si[j] = dvbcsa_block_sbox[u.si[j]];
> > -
> > - dvbcsa_bs_word_t sbox_out = u.so;
> > + for (g = 0; g < 8; g++)
> > + {
> > + dvbcsa_bs_word_t sbox_out = scratch2[g];
> > + dvbcsa_bs_word_t w = r[-8 * 1 + g];
> >
> > - // bit permutation
> > + /* bit permutation */
> >
> > dvbcsa_bs_word_t in = BS_OR(
> > BS_OR(
> > @@ -150,7 +181,6 @@ dvbcsa_bs_block_encrypt_register (const
> dvbcsa_bs_word_t *block, dvbcsa_bs_word_
> > BS_OR( BS_SHR (BS_AND
> (sbox_out, BS_VAL8(40)), 6),
> > BS_SHR (BS_AND
> (sbox_out, BS_VAL8(80)), 4)));
> >
> > - dvbcsa_bs_word_t w = r[-8 * 1 + g];
> >
> > r[8 * 7 + g] = BS_XOR(w, sbox_out);
> > BS_XOREQ(r[8 * 1 + g], w);
> > --
> > 1.9.1
> >
> > _______________________________________________
> > vlc-devel mailing list
> > To unsubscribe or modify your subscription options:
> > https://mailman.videolan.org/listinfo/vlc-devel
>
> --
> With my kindest regards,
>
> --
> Jean-Baptiste Kempf
> http://www.jbkempf.com/ - +33 672 704 734
> Sent from my Electronic Device
> _______________________________________________
> vlc-devel mailing list
> To unsubscribe or modify your subscription options:
> https://mailman.videolan.org/listinfo/vlc-devel
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/vlc-devel/attachments/20150707/50d630bf/attachment.html>
More information about the vlc-devel
mailing list