replace "_avx2" with "_fat".

This commit is contained in:
Chang, Harry 2017-07-24 15:09:17 +08:00 committed by Matthew Barr
parent 8da2d13baa
commit 2b1d3383aa
3 changed files with 62 additions and 62 deletions

View File

@ -803,14 +803,14 @@ static const FDRFUNCTYPE funcs[] = {
fdr_engine_exec, fdr_engine_exec,
NULL, /* old: fast teddy */ NULL, /* old: fast teddy */
NULL, /* old: fast teddy */ NULL, /* old: fast teddy */
ONLY_AVX2(fdr_exec_teddy_avx2_msks1_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks1),
ONLY_AVX2(fdr_exec_teddy_avx2_msks1_pck_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks1_pck),
ONLY_AVX2(fdr_exec_teddy_avx2_msks2_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks2),
ONLY_AVX2(fdr_exec_teddy_avx2_msks2_pck_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks2_pck),
ONLY_AVX2(fdr_exec_teddy_avx2_msks3_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks3),
ONLY_AVX2(fdr_exec_teddy_avx2_msks3_pck_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks3_pck),
ONLY_AVX2(fdr_exec_teddy_avx2_msks4_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks4),
ONLY_AVX2(fdr_exec_teddy_avx2_msks4_pck_fat), ONLY_AVX2(fdr_exec_fat_teddy_msks4_pck),
fdr_exec_teddy_msks1, fdr_exec_teddy_msks1,
fdr_exec_teddy_msks1_pck, fdr_exec_teddy_msks1_pck,
fdr_exec_teddy_msks2, fdr_exec_teddy_msks2,

View File

@ -73,35 +73,35 @@ hwlm_error_t fdr_exec_teddy_msks4_pck(const struct FDR *fdr,
#if defined(HAVE_AVX2) #if defined(HAVE_AVX2)
hwlm_error_t fdr_exec_teddy_avx2_msks1_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks1(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);
hwlm_error_t fdr_exec_teddy_avx2_msks1_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks1_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);
hwlm_error_t fdr_exec_teddy_avx2_msks2_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks2(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);
hwlm_error_t fdr_exec_teddy_avx2_msks2_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks2_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);
hwlm_error_t fdr_exec_teddy_avx2_msks3_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks3(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);
hwlm_error_t fdr_exec_teddy_avx2_msks3_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks3_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);
hwlm_error_t fdr_exec_teddy_avx2_msks4_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks4(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);
hwlm_error_t fdr_exec_teddy_avx2_msks4_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks4_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control); hwlm_group_t control);

View File

@ -130,15 +130,15 @@ do { \
} while(0) } while(0)
static really_inline static really_inline
const m256 *getMaskBase_avx2(const struct Teddy *teddy) { const m256 *getMaskBase_fat(const struct Teddy *teddy) {
return (const m256 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy))); return (const m256 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy)));
} }
#if defined(HAVE_AVX512) #if defined(HAVE_AVX512)
static really_inline static really_inline
const u64a *getReinforcedMaskBase_avx2(const struct Teddy *teddy, u8 numMask) { const u64a *getReinforcedMaskBase_fat(const struct Teddy *teddy, u8 numMask) {
return (const u64a *)((const u8 *)getMaskBase_avx2(teddy) return (const u64a *)((const u8 *)getMaskBase_fat(teddy)
+ ROUNDUP_CL(2 * numMask * sizeof(m256))); + ROUNDUP_CL(2 * numMask * sizeof(m256)));
} }
@ -370,11 +370,11 @@ do { \
DEBUG_PRINTF("params: buf %p len %zu start_offset %zu\n", \ DEBUG_PRINTF("params: buf %p len %zu start_offset %zu\n", \
a->buf, a->len, a->start_offset); \ a->buf, a->len, a->start_offset); \
\ \
const m256 *maskBase = getMaskBase_avx2(teddy); \ const m256 *maskBase = getMaskBase_fat(teddy); \
PREPARE_FAT_MASKS(n_msk); \ PREPARE_FAT_MASKS(n_msk); \
const u32 *confBase = getConfBase(teddy); \ const u32 *confBase = getConfBase(teddy); \
\ \
const u64a *r_msk_base_lo = getReinforcedMaskBase_avx2(teddy, n_msk); \ const u64a *r_msk_base_lo = getReinforcedMaskBase_fat(teddy, n_msk); \
const u64a *r_msk_base_hi = r_msk_base_lo + (N_CHARS + 1); \ const u64a *r_msk_base_hi = r_msk_base_lo + (N_CHARS + 1); \
u32 c_0 = 0x100; \ u32 c_0 = 0x100; \
u32 c_16 = 0x100; \ u32 c_16 = 0x100; \
@ -582,7 +582,7 @@ do { \
DEBUG_PRINTF("params: buf %p len %zu start_offset %zu\n", \ DEBUG_PRINTF("params: buf %p len %zu start_offset %zu\n", \
a->buf, a->len, a->start_offset); \ a->buf, a->len, a->start_offset); \
\ \
const m256 *maskBase = getMaskBase_avx2(teddy); \ const m256 *maskBase = getMaskBase_fat(teddy); \
const u32 *confBase = getConfBase(teddy); \ const u32 *confBase = getConfBase(teddy); \
\ \
FDR_EXEC_FAT_TEDDY_RES_OLD(n_msk); \ FDR_EXEC_FAT_TEDDY_RES_OLD(n_msk); \
@ -638,49 +638,49 @@ do { \
#endif // HAVE_AVX512 #endif // HAVE_AVX512
hwlm_error_t fdr_exec_teddy_avx2_msks1_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks1(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 1, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 1, do_confWithBit_teddy);
} }
hwlm_error_t fdr_exec_teddy_avx2_msks1_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks1_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 1, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 1, do_confWithBit_teddy);
} }
hwlm_error_t fdr_exec_teddy_avx2_msks2_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks2(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 2, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 2, do_confWithBit_teddy);
} }
hwlm_error_t fdr_exec_teddy_avx2_msks2_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks2_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 2, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 2, do_confWithBit_teddy);
} }
hwlm_error_t fdr_exec_teddy_avx2_msks3_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks3(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 3, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 3, do_confWithBit_teddy);
} }
hwlm_error_t fdr_exec_teddy_avx2_msks3_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks3_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 3, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 3, do_confWithBit_teddy);
} }
hwlm_error_t fdr_exec_teddy_avx2_msks4_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks4(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 4, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 4, do_confWithBit_teddy);
} }
hwlm_error_t fdr_exec_teddy_avx2_msks4_pck_fat(const struct FDR *fdr, hwlm_error_t fdr_exec_fat_teddy_msks4_pck(const struct FDR *fdr,
const struct FDR_Runtime_Args *a, const struct FDR_Runtime_Args *a,
hwlm_group_t control) { hwlm_group_t control) {
FDR_EXEC_FAT_TEDDY(fdr, a, control, 4, do_confWithBit_teddy); FDR_EXEC_FAT_TEDDY(fdr, a, control, 4, do_confWithBit_teddy);