Optimized a0 pure kernel for AMD

This commit is contained in:
jsteube
2017-08-23 13:40:22 +02:00
parent a5659d5619
commit 0783289e2f
76 changed files with 311 additions and 296 deletions

View File

@@ -60,7 +60,7 @@ __kernel void m04500_mxx (__global pw_t *pws, __constant const kernel_rule_t *ru
* base
*/
pw_t pw = pws[gid];
COPY_PW (pws[gid]);
/**
* loop
@@ -68,7 +68,7 @@ __kernel void m04500_mxx (__global pw_t *pws, __constant const kernel_rule_t *ru
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
{
pw_t tmp = pw;
pw_t tmp = PASTE_PW;
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);
@@ -169,7 +169,7 @@ __kernel void m04500_sxx (__global pw_t *pws, __constant const kernel_rule_t *ru
* base
*/
pw_t pw = pws[gid];
COPY_PW (pws[gid]);
/**
* loop
@@ -177,7 +177,7 @@ __kernel void m04500_sxx (__global pw_t *pws, __constant const kernel_rule_t *ru
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
{
pw_t tmp = pw;
pw_t tmp = PASTE_PW;
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);