From 0fd765dbe3be40cb920391bc10d08c02c7d5b039 Mon Sep 17 00:00:00 2001 From: zooko Date: Sun, 11 Nov 2007 22:34:44 +0530 Subject: [PATCH] zfec: do encoding within a fixed window of memory in order to be cache friendly darcs-hash:ecd8bb2a485674ce3178b9d9bdb9df107fa3411f --- zfec/zfec/fec.c | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-) diff --git a/zfec/zfec/fec.c b/zfec/zfec/fec.c index 796425e..2054697 100644 --- a/zfec/zfec/fec.c +++ b/zfec/zfec/fec.c @@ -75,6 +75,7 @@ static gf gf_mul_table[256][256]; #define gf_mul(x,y) gf_mul_table[x][y] #define USE_GF_MULC register gf * __gf_mulc_ + #define GF_MULC0(c) __gf_mulc_ = gf_mul_table[c] #define GF_ADDMULC(dst, x) dst ^= __gf_mulc_[x] @@ -470,19 +471,26 @@ fec_new(unsigned k, unsigned n) { return retval; } +/* To make sure that we stay within cache in the inner loops of fec_encode() + and fec_decode(). */ +#define STRIDE 1024 + void fec_encode(const fec_t* code, const gf*restrict const*restrict const src, gf*restrict const*restrict const fecs, const unsigned*restrict const block_nums, size_t num_block_nums, size_t sz) { unsigned char i, j; + size_t k; unsigned fecnum; - gf* p; + const gf* p; for (i=0; i= code->k); memset(fecs[i], 0, sz); p = &(code->enc_matrix[fecnum * code->k]); - for (j = 0; j < code->k; j++) - addmul(fecs[i], src[j], p[j], sz); +// DUFF ME + for (k = 0; k < sz; k += STRIDE) + for (j = 0; j < code->k; j++) + addmul(fecs[i]+k, src[j]+k, p[j], ((sz-k) < STRIDE)?(sz-k):STRIDE); } } -- 2.37.2