diff --git a/src/fdr/teddy.c b/src/fdr/teddy.c index fd149016..f28d0a50 100644 --- a/src/fdr/teddy.c +++ b/src/fdr/teddy.c @@ -74,7 +74,294 @@ const u8 ALIGN_DIRECTIVE p_mask_arr[17][32] = { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00} }; -#if defined(__AVX2__) // reinforced teddy +#define CONF_CHUNK_64(chunk, bucket, off, reason, conf_fn) \ +do { \ + if (unlikely(chunk != ones_u64a)) { \ + chunk = ~chunk; \ + conf_fn(&chunk, bucket, off, confBase, reason, a, ptr, \ + &control, &last_match); \ + CHECK_HWLM_TERMINATE_MATCHING; \ + } \ +} while(0) + +#define CONF_CHUNK_32(chunk, bucket, off, reason, conf_fn) \ +do { \ + if (unlikely(chunk != ones_u32a)) { \ + chunk = ~chunk; \ + conf_fn(&chunk, bucket, off, confBase, reason, a, ptr, \ + &control, &last_match); \ + CHECK_HWLM_TERMINATE_MATCHING; \ + } \ +} while(0) + +#if defined(HAVE_AVX512) // AVX512 reinforced teddy + +#ifdef ARCH_64_BIT +#define CONFIRM_TEDDY(var, bucket, offset, reason, conf_fn) \ +do { \ + if (unlikely(diff512(var, ones512()))) { \ + m128 p128_0 = extract128from512(var, 0); \ + m128 p128_1 = extract128from512(var, 1); \ + m128 p128_2 = extract128from512(var, 2); \ + m128 p128_3 = extract128from512(var, 3); \ + u64a part1 = movq(p128_0); \ + u64a part2 = movq(rshiftbyte_m128(p128_0, 8)); \ + u64a part3 = movq(p128_1); \ + u64a part4 = movq(rshiftbyte_m128(p128_1, 8)); \ + u64a part5 = movq(p128_2); \ + u64a part6 = movq(rshiftbyte_m128(p128_2, 8)); \ + u64a part7 = movq(p128_3); \ + u64a part8 = movq(rshiftbyte_m128(p128_3, 8)); \ + CONF_CHUNK_64(part1, bucket, offset, reason, conf_fn); \ + CONF_CHUNK_64(part2, bucket, offset + 8, reason, conf_fn); \ + CONF_CHUNK_64(part3, bucket, offset + 16, reason, conf_fn); \ + CONF_CHUNK_64(part4, bucket, offset + 24, reason, conf_fn); \ + CONF_CHUNK_64(part5, bucket, offset + 32, reason, conf_fn); \ + CONF_CHUNK_64(part6, bucket, offset + 40, reason, conf_fn); \ + CONF_CHUNK_64(part7, bucket, offset + 48, reason, conf_fn); \ + CONF_CHUNK_64(part8, bucket, offset + 56, reason, conf_fn); \ + } \ +} while(0) +#else +#define CONFIRM_TEDDY(var, bucket, offset, reason, conf_fn) \ +do { \ + if (unlikely(diff512(var, ones512()))) { \ + m128 p128_0 = extract128from512(var, 0); \ + m128 p128_1 = extract128from512(var, 1); \ + m128 p128_2 = extract128from512(var, 2); \ + m128 p128_3 = extract128from512(var, 3); \ + u32 part1 = movd(p128_0); \ + u32 part2 = movd(rshiftbyte_m128(p128_0, 4)); \ + u32 part3 = movd(rshiftbyte_m128(p128_0, 8)); \ + u32 part4 = movd(rshiftbyte_m128(p128_0, 12)); \ + u32 part5 = movd(p128_1); \ + u32 part6 = movd(rshiftbyte_m128(p128_1, 4)); \ + u32 part7 = movd(rshiftbyte_m128(p128_1, 8)); \ + u32 part8 = movd(rshiftbyte_m128(p128_1, 12)); \ + u32 part9 = movd(p128_2); \ + u32 part10 = movd(rshiftbyte_m128(p128_2, 4)); \ + u32 part11 = movd(rshiftbyte_m128(p128_2, 8)); \ + u32 part12 = movd(rshiftbyte_m128(p128_2, 12)); \ + u32 part13 = movd(p128_3); \ + u32 part14 = movd(rshiftbyte_m128(p128_3, 4)); \ + u32 part15 = movd(rshiftbyte_m128(p128_3, 8)); \ + u32 part16 = movd(rshiftbyte_m128(p128_3, 12)); \ + CONF_CHUNK_32(part1, bucket, offset, reason, conf_fn); \ + CONF_CHUNK_32(part2, bucket, offset + 4, reason, conf_fn); \ + CONF_CHUNK_32(part3, bucket, offset + 8, reason, conf_fn); \ + CONF_CHUNK_32(part4, bucket, offset + 12, reason, conf_fn); \ + CONF_CHUNK_32(part5, bucket, offset + 16, reason, conf_fn); \ + CONF_CHUNK_32(part6, bucket, offset + 20, reason, conf_fn); \ + CONF_CHUNK_32(part7, bucket, offset + 24, reason, conf_fn); \ + CONF_CHUNK_32(part8, bucket, offset + 28, reason, conf_fn); \ + CONF_CHUNK_32(part9, bucket, offset + 32, reason, conf_fn); \ + CONF_CHUNK_32(part10, bucket, offset + 36, reason, conf_fn); \ + CONF_CHUNK_32(part11, bucket, offset + 40, reason, conf_fn); \ + CONF_CHUNK_32(part12, bucket, offset + 44, reason, conf_fn); \ + CONF_CHUNK_32(part13, bucket, offset + 48, reason, conf_fn); \ + CONF_CHUNK_32(part14, bucket, offset + 52, reason, conf_fn); \ + CONF_CHUNK_32(part15, bucket, offset + 56, reason, conf_fn); \ + CONF_CHUNK_32(part16, bucket, offset + 60, reason, conf_fn); \ + } \ +} while(0) +#endif + +#define PREP_SHUF_MASK_NO_REINFORCEMENT(val) \ + m512 lo = and512(val, *lo_mask); \ + m512 hi = and512(rshift64_m512(val, 4), *lo_mask) + +#define PREP_SHUF_MASK \ + PREP_SHUF_MASK_NO_REINFORCEMENT(load512(ptr)); \ + *c_16 = *(ptr + 15); \ + *c_32 = *(ptr + 31); \ + *c_48 = *(ptr + 47); \ + m512 r_msk = set512_64(0ULL, r_msk_base[*c_48], 0ULL, r_msk_base[*c_32],\ + 0ULL, r_msk_base[*c_16], 0ULL, r_msk_base[*c_0]);\ + *c_0 = *(ptr + 63) + +#define SHIFT_OR_M1 \ + or512(pshufb_m512(dup_mask[0], lo), pshufb_m512(dup_mask[1], hi)) + +#define SHIFT_OR_M2 \ + or512(lshift128_m512(or512(pshufb_m512(dup_mask[2], lo), \ + pshufb_m512(dup_mask[3], hi)), \ + 1), SHIFT_OR_M1) + +#define SHIFT_OR_M3 \ + or512(lshift128_m512(or512(pshufb_m512(dup_mask[4], lo), \ + pshufb_m512(dup_mask[5], hi)), \ + 2), SHIFT_OR_M2) + +#define SHIFT_OR_M4 \ + or512(lshift128_m512(or512(pshufb_m512(dup_mask[6], lo), \ + pshufb_m512(dup_mask[7], hi)), \ + 3), SHIFT_OR_M3) + +static really_inline +m512 prep_conf_teddy_no_reinforcement_m1(const m512 *lo_mask, + const m512 *dup_mask, + const m512 val) { + PREP_SHUF_MASK_NO_REINFORCEMENT(val); + return SHIFT_OR_M1; +} + +static really_inline +m512 prep_conf_teddy_no_reinforcement_m2(const m512 *lo_mask, + const m512 *dup_mask, + const m512 val) { + PREP_SHUF_MASK_NO_REINFORCEMENT(val); + return SHIFT_OR_M2; +} + +static really_inline +m512 prep_conf_teddy_no_reinforcement_m3(const m512 *lo_mask, + const m512 *dup_mask, + const m512 val) { + PREP_SHUF_MASK_NO_REINFORCEMENT(val); + return SHIFT_OR_M3; +} + +static really_inline +m512 prep_conf_teddy_no_reinforcement_m4(const m512 *lo_mask, + const m512 *dup_mask, + const m512 val) { + PREP_SHUF_MASK_NO_REINFORCEMENT(val); + return SHIFT_OR_M4; +} + +static really_inline +m512 prep_conf_teddy_m1(const m512 *lo_mask, const m512 *dup_mask, + const u8 *ptr, const u64a *r_msk_base, + u32 *c_0, u32 *c_16, u32 *c_32, u32 *c_48) { + PREP_SHUF_MASK; + return or512(SHIFT_OR_M1, r_msk); +} + +static really_inline +m512 prep_conf_teddy_m2(const m512 *lo_mask, const m512 *dup_mask, + const u8 *ptr, const u64a *r_msk_base, + u32 *c_0, u32 *c_16, u32 *c_32, u32 *c_48) { + PREP_SHUF_MASK; + return or512(SHIFT_OR_M2, r_msk); +} + +static really_inline +m512 prep_conf_teddy_m3(const m512 *lo_mask, const m512 *dup_mask, + const u8 *ptr, const u64a *r_msk_base, + u32 *c_0, u32 *c_16, u32 *c_32, u32 *c_48) { + PREP_SHUF_MASK; + return or512(SHIFT_OR_M3, r_msk); +} + +static really_inline +m512 prep_conf_teddy_m4(const m512 *lo_mask, const m512 *dup_mask, + const u8 *ptr, const u64a *r_msk_base, + u32 *c_0, u32 *c_16, u32 *c_32, u32 *c_48) { + PREP_SHUF_MASK; + return or512(SHIFT_OR_M4, r_msk); +} + +#define PREP_CONF_FN_NO_REINFORCEMENT(val, n) \ + prep_conf_teddy_no_reinforcement_m##n(&lo_mask, dup_mask, val) + +#define PREP_CONF_FN(ptr, n) \ + prep_conf_teddy_m##n(&lo_mask, dup_mask, ptr, r_msk_base, \ + &c_0, &c_16, &c_32, &c_48) + +#define PREPARE_MASKS_1 \ + dup_mask[0] = set4x128(maskBase[0]); \ + dup_mask[1] = set4x128(maskBase[1]); + +#define PREPARE_MASKS_2 \ + PREPARE_MASKS_1 \ + dup_mask[2] = set4x128(maskBase[2]); \ + dup_mask[3] = set4x128(maskBase[3]); + +#define PREPARE_MASKS_3 \ + PREPARE_MASKS_2 \ + dup_mask[4] = set4x128(maskBase[4]); \ + dup_mask[5] = set4x128(maskBase[5]); + +#define PREPARE_MASKS_4 \ + PREPARE_MASKS_3 \ + dup_mask[6] = set4x128(maskBase[6]); \ + dup_mask[7] = set4x128(maskBase[7]); + +#define PREPARE_MASKS(n) \ + m512 lo_mask = set64x8(0xf); \ + m512 dup_mask[n * 2]; \ + PREPARE_MASKS_##n + +#define FDR_EXEC_TEDDY(fdr, a, control, n_msk, conf_fn) \ +do { \ + const u8 *buf_end = a->buf + a->len; \ + const u8 *ptr = a->buf + a->start_offset; \ + u32 floodBackoff = FLOOD_BACKOFF_START; \ + const u8 *tryFloodDetect = a->firstFloodDetect; \ + u32 last_match = (u32)-1; \ + const struct Teddy *teddy = (const struct Teddy *)fdr; \ + const size_t iterBytes = 128; \ + DEBUG_PRINTF("params: buf %p len %zu start_offset %zu\n", \ + a->buf, a->len, a->start_offset); \ + \ + const m128 *maskBase = getMaskBase(teddy); \ + PREPARE_MASKS(n_msk); \ + const u32 *confBase = getConfBase(teddy); \ + \ + const u64a *r_msk_base = getReinforcedMaskBase(teddy, n_msk); \ + u32 c_0 = 0x100; \ + u32 c_16 = 0x100; \ + u32 c_32 = 0x100; \ + u32 c_48 = 0x100; \ + const u8 *mainStart = ROUNDUP_PTR(ptr, 64); \ + DEBUG_PRINTF("derive: ptr: %p mainstart %p\n", ptr, mainStart); \ + if (ptr < mainStart) { \ + ptr = mainStart - 64; \ + m512 p_mask; \ + m512 val_0 = vectoredLoad512(&p_mask, ptr, a->start_offset, \ + a->buf, buf_end, \ + a->buf_history, a->len_history, n_msk); \ + m512 r_0 = PREP_CONF_FN_NO_REINFORCEMENT(val_0, n_msk); \ + r_0 = or512(r_0, p_mask); \ + CONFIRM_TEDDY(r_0, 8, 0, VECTORING, conf_fn); \ + ptr += 64; \ + } \ + \ + if (ptr + 64 <= buf_end) { \ + m512 r_0 = PREP_CONF_FN(ptr, n_msk); \ + CONFIRM_TEDDY(r_0, 8, 0, VECTORING, conf_fn); \ + ptr += 64; \ + } \ + \ + for (; ptr + iterBytes <= buf_end; ptr += iterBytes) { \ + __builtin_prefetch(ptr + (iterBytes * 4)); \ + CHECK_FLOOD; \ + m512 r_0 = PREP_CONF_FN(ptr, n_msk); \ + CONFIRM_TEDDY(r_0, 8, 0, NOT_CAUTIOUS, conf_fn); \ + m512 r_1 = PREP_CONF_FN(ptr + 64, n_msk); \ + CONFIRM_TEDDY(r_1, 8, 64, NOT_CAUTIOUS, conf_fn); \ + } \ + \ + if (ptr + 64 <= buf_end) { \ + m512 r_0 = PREP_CONF_FN(ptr, n_msk); \ + CONFIRM_TEDDY(r_0, 8, 0, NOT_CAUTIOUS, conf_fn); \ + ptr += 64; \ + } \ + \ + assert(ptr + 64 > buf_end); \ + if (ptr < buf_end) { \ + m512 p_mask; \ + m512 val_0 = vectoredLoad512(&p_mask, ptr, 0, ptr, buf_end, \ + a->buf_history, a->len_history, n_msk); \ + m512 r_0 = PREP_CONF_FN_NO_REINFORCEMENT(val_0, n_msk); \ + r_0 = or512(r_0, p_mask); \ + CONFIRM_TEDDY(r_0, 8, 0, VECTORING, conf_fn); \ + } \ + \ + return HWLM_SUCCESS; \ +} while(0) + +#elif defined(HAVE_AVX2) // not HAVE_AVX512 but HAVE_AVX2 reinforced teddy #ifdef ARCH_64_BIT #define CONFIRM_TEDDY(var, bucket, offset, reason, conf_fn) \ @@ -86,30 +373,10 @@ do { \ u64a part2 = movq(rshiftbyte_m128(lo, 8)); \ u64a part3 = movq(hi); \ u64a part4 = movq(rshiftbyte_m128(hi, 8)); \ - if (unlikely(part1 != ones_u64a)) { \ - part1 = ~part1; \ - conf_fn(&part1, bucket, offset, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part2 != ones_u64a)) { \ - part2 = ~part2; \ - conf_fn(&part2, bucket, offset + 8, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part3 != ones_u64a)) { \ - part3 = ~part3; \ - conf_fn(&part3, bucket, offset + 16, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part4 != ones_u64a)) { \ - part4 = ~part4; \ - conf_fn(&part4, bucket, offset + 24, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ + CONF_CHUNK_64(part1, bucket, offset, reason, conf_fn); \ + CONF_CHUNK_64(part2, bucket, offset + 8, reason, conf_fn); \ + CONF_CHUNK_64(part3, bucket, offset + 16, reason, conf_fn); \ + CONF_CHUNK_64(part4, bucket, offset + 24, reason, conf_fn); \ } \ } while(0) #else @@ -126,54 +393,14 @@ do { \ u32 part6 = movd(rshiftbyte_m128(hi, 4)); \ u32 part7 = movd(rshiftbyte_m128(hi, 8)); \ u32 part8 = movd(rshiftbyte_m128(hi, 12)); \ - if (unlikely(part1 != ones_u32)) { \ - part1 = ~part1; \ - conf_fn(&part1, bucket, offset, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part2 != ones_u32)) { \ - part2 = ~part2; \ - conf_fn(&part2, bucket, offset + 4, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part3 != ones_u32)) { \ - part3 = ~part3; \ - conf_fn(&part3, bucket, offset + 8, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part4 != ones_u32)) { \ - part4 = ~part4; \ - conf_fn(&part4, bucket, offset + 12, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part5 != ones_u32)) { \ - part5 = ~part5; \ - conf_fn(&part5, bucket, offset + 16, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part6 != ones_u32)) { \ - part6 = ~part6; \ - conf_fn(&part6, bucket, offset + 20, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part7 != ones_u32)) { \ - part7 = ~part7; \ - conf_fn(&part7, bucket, offset + 24, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part8 != ones_u32)) { \ - part8 = ~part8; \ - conf_fn(&part8, bucket, offset + 28, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ + CONF_CHUNK_32(part1, bucket, offset, reason, conf_fn); \ + CONF_CHUNK_32(part2, bucket, offset + 4, reason, conf_fn); \ + CONF_CHUNK_32(part3, bucket, offset + 8, reason, conf_fn); \ + CONF_CHUNK_32(part4, bucket, offset + 12, reason, conf_fn); \ + CONF_CHUNK_32(part5, bucket, offset + 16, reason, conf_fn); \ + CONF_CHUNK_32(part6, bucket, offset + 20, reason, conf_fn); \ + CONF_CHUNK_32(part7, bucket, offset + 24, reason, conf_fn); \ + CONF_CHUNK_32(part8, bucket, offset + 28, reason, conf_fn); \ } \ } while(0) #endif @@ -270,121 +497,6 @@ m256 prep_conf_teddy_m4(const m256 *lo_mask, const m256 *dup_mask, return or256(SHIFT_OR_M4, r_msk); } -#else // not defined __AVX2__ - -#ifdef ARCH_64_BIT -#define CONFIRM_TEDDY(var, bucket, offset, reason, conf_fn) \ -do { \ - if (unlikely(diff128(var, ones128()))) { \ - u64a lo = movq(var); \ - u64a hi = movq(rshiftbyte_m128(var, 8)); \ - if (unlikely(lo != ones_u64a)) { \ - lo = ~lo; \ - conf_fn(&lo, bucket, offset, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(hi != ones_u64a)) { \ - hi = ~hi; \ - conf_fn(&hi, bucket, offset + 8, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - } \ -} while(0) -#else -#define CONFIRM_TEDDY(var, bucket, offset, reason, conf_fn) \ -do { \ - if (unlikely(diff128(var, ones128()))) { \ - u32 part1 = movd(var); \ - u32 part2 = movd(rshiftbyte_m128(var, 4)); \ - u32 part3 = movd(rshiftbyte_m128(var, 8)); \ - u32 part4 = movd(rshiftbyte_m128(var, 12)); \ - if (unlikely(part1 != ones_u32)) { \ - part1 = ~part1; \ - conf_fn(&part1, bucket, offset, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part2 != ones_u32)) { \ - part2 = ~part2; \ - conf_fn(&part2, bucket, offset + 4, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part3 != ones_u32)) { \ - part3 = ~part3; \ - conf_fn(&part3, bucket, offset + 8, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part4 != ones_u32)) { \ - part4 = ~part4; \ - conf_fn(&part4, bucket, offset + 12, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - } \ -} while(0) -#endif - -static really_inline -m128 prep_conf_teddy_m1(const m128 *maskBase, m128 val) { - m128 mask = set16x8(0xf); - m128 lo = and128(val, mask); - m128 hi = and128(rshift64_m128(val, 4), mask); - return or128(pshufb_m128(maskBase[0 * 2], lo), - pshufb_m128(maskBase[0 * 2 + 1], hi)); -} - -static really_inline -m128 prep_conf_teddy_m2(const m128 *maskBase, m128 *old_1, m128 val) { - m128 mask = set16x8(0xf); - m128 lo = and128(val, mask); - m128 hi = and128(rshift64_m128(val, 4), mask); - m128 r = prep_conf_teddy_m1(maskBase, val); - - m128 res_1 = or128(pshufb_m128(maskBase[1 * 2], lo), - pshufb_m128(maskBase[1 * 2 + 1], hi)); - m128 res_shifted_1 = palignr(res_1, *old_1, 16 - 1); - *old_1 = res_1; - return or128(r, res_shifted_1); -} - -static really_inline -m128 prep_conf_teddy_m3(const m128 *maskBase, m128 *old_1, m128 *old_2, - m128 val) { - m128 mask = set16x8(0xf); - m128 lo = and128(val, mask); - m128 hi = and128(rshift64_m128(val, 4), mask); - m128 r = prep_conf_teddy_m2(maskBase, old_1, val); - - m128 res_2 = or128(pshufb_m128(maskBase[2 * 2], lo), - pshufb_m128(maskBase[2 * 2 + 1], hi)); - m128 res_shifted_2 = palignr(res_2, *old_2, 16 - 2); - *old_2 = res_2; - return or128(r, res_shifted_2); -} - -static really_inline -m128 prep_conf_teddy_m4(const m128 *maskBase, m128 *old_1, m128 *old_2, - m128 *old_3, m128 val) { - m128 mask = set16x8(0xf); - m128 lo = and128(val, mask); - m128 hi = and128(rshift64_m128(val, 4), mask); - m128 r = prep_conf_teddy_m3(maskBase, old_1, old_2, val); - - m128 res_3 = or128(pshufb_m128(maskBase[3 * 2], lo), - pshufb_m128(maskBase[3 * 2 + 1], hi)); - m128 res_shifted_3 = palignr(res_3, *old_3, 16 - 3); - *old_3 = res_3; - return or128(r, res_shifted_3); -} - -#endif // __AVX2__ - -#if defined(__AVX2__) // reinforced teddy - #define PREP_CONF_FN_NO_REINFORCEMENT(val, n) \ prep_conf_teddy_no_reinforcement_m##n(&lo_mask, dup_mask, val) @@ -415,42 +527,6 @@ m128 prep_conf_teddy_m4(const m128 *maskBase, m128 *old_1, m128 *old_2, m256 dup_mask[n * 2]; \ PREPARE_MASKS_##n -#else // not defined __AVX2__ - -#define FDR_EXEC_TEDDY_RES_OLD_1 - -#define FDR_EXEC_TEDDY_RES_OLD_2 \ - m128 res_old_1 = zeroes128(); - -#define FDR_EXEC_TEDDY_RES_OLD_3 \ - m128 res_old_1 = zeroes128(); \ - m128 res_old_2 = zeroes128(); - -#define FDR_EXEC_TEDDY_RES_OLD_4 \ - m128 res_old_1 = zeroes128(); \ - m128 res_old_2 = zeroes128(); \ - m128 res_old_3 = zeroes128(); - -#define FDR_EXEC_TEDDY_RES_OLD(n) FDR_EXEC_TEDDY_RES_OLD_##n - -#define PREP_CONF_FN_1(mask_base, val) \ - prep_conf_teddy_m1(mask_base, val) - -#define PREP_CONF_FN_2(mask_base, val) \ - prep_conf_teddy_m2(mask_base, &res_old_1, val) - -#define PREP_CONF_FN_3(mask_base, val) \ - prep_conf_teddy_m3(mask_base, &res_old_1, &res_old_2, val) - -#define PREP_CONF_FN_4(mask_base, val) \ - prep_conf_teddy_m4(mask_base, &res_old_1, &res_old_2, &res_old_3, val) - -#define PREP_CONF_FN(mask_base, val, n) \ - PREP_CONF_FN_##n(mask_base, val) -#endif // __AVX2__ - - -#if defined(__AVX2__) // reinforced teddy #define FDR_EXEC_TEDDY(fdr, a, control, n_msk, conf_fn) \ do { \ const u8 *buf_end = a->buf + a->len; \ @@ -517,7 +593,119 @@ do { \ \ return HWLM_SUCCESS; \ } while(0) -#else // not defined __AVX2__ + +#else // not defined HAVE_AVX2 + +#ifdef ARCH_64_BIT +#define CONFIRM_TEDDY(var, bucket, offset, reason, conf_fn) \ +do { \ + if (unlikely(diff128(var, ones128()))) { \ + u64a lo = movq(var); \ + u64a hi = movq(rshiftbyte_m128(var, 8)); \ + CONF_CHUNK_64(lo, bucket, offset, reason, conf_fn); \ + CONF_CHUNK_64(hi, bucket, offset + 8, reason, conf_fn); \ + } \ +} while(0) +#else +#define CONFIRM_TEDDY(var, bucket, offset, reason, conf_fn) \ +do { \ + if (unlikely(diff128(var, ones128()))) { \ + u32 part1 = movd(var); \ + u32 part2 = movd(rshiftbyte_m128(var, 4)); \ + u32 part3 = movd(rshiftbyte_m128(var, 8)); \ + u32 part4 = movd(rshiftbyte_m128(var, 12)); \ + CONF_CHUNK_32(part1, bucket, offset, reason, conf_fn); \ + CONF_CHUNK_32(part2, bucket, offset + 4, reason, conf_fn); \ + CONF_CHUNK_32(part3, bucket, offset + 8, reason, conf_fn); \ + CONF_CHUNK_32(part4, bucket, offset + 12, reason, conf_fn); \ + } \ +} while(0) +#endif + +static really_inline +m128 prep_conf_teddy_m1(const m128 *maskBase, m128 val) { + m128 mask = set16x8(0xf); + m128 lo = and128(val, mask); + m128 hi = and128(rshift64_m128(val, 4), mask); + return or128(pshufb_m128(maskBase[0 * 2], lo), + pshufb_m128(maskBase[0 * 2 + 1], hi)); +} + +static really_inline +m128 prep_conf_teddy_m2(const m128 *maskBase, m128 *old_1, m128 val) { + m128 mask = set16x8(0xf); + m128 lo = and128(val, mask); + m128 hi = and128(rshift64_m128(val, 4), mask); + m128 r = prep_conf_teddy_m1(maskBase, val); + + m128 res_1 = or128(pshufb_m128(maskBase[1 * 2], lo), + pshufb_m128(maskBase[1 * 2 + 1], hi)); + m128 res_shifted_1 = palignr(res_1, *old_1, 16 - 1); + *old_1 = res_1; + return or128(r, res_shifted_1); +} + +static really_inline +m128 prep_conf_teddy_m3(const m128 *maskBase, m128 *old_1, m128 *old_2, + m128 val) { + m128 mask = set16x8(0xf); + m128 lo = and128(val, mask); + m128 hi = and128(rshift64_m128(val, 4), mask); + m128 r = prep_conf_teddy_m2(maskBase, old_1, val); + + m128 res_2 = or128(pshufb_m128(maskBase[2 * 2], lo), + pshufb_m128(maskBase[2 * 2 + 1], hi)); + m128 res_shifted_2 = palignr(res_2, *old_2, 16 - 2); + *old_2 = res_2; + return or128(r, res_shifted_2); +} + +static really_inline +m128 prep_conf_teddy_m4(const m128 *maskBase, m128 *old_1, m128 *old_2, + m128 *old_3, m128 val) { + m128 mask = set16x8(0xf); + m128 lo = and128(val, mask); + m128 hi = and128(rshift64_m128(val, 4), mask); + m128 r = prep_conf_teddy_m3(maskBase, old_1, old_2, val); + + m128 res_3 = or128(pshufb_m128(maskBase[3 * 2], lo), + pshufb_m128(maskBase[3 * 2 + 1], hi)); + m128 res_shifted_3 = palignr(res_3, *old_3, 16 - 3); + *old_3 = res_3; + return or128(r, res_shifted_3); +} + +#define FDR_EXEC_TEDDY_RES_OLD_1 + +#define FDR_EXEC_TEDDY_RES_OLD_2 \ + m128 res_old_1 = zeroes128(); + +#define FDR_EXEC_TEDDY_RES_OLD_3 \ + m128 res_old_1 = zeroes128(); \ + m128 res_old_2 = zeroes128(); + +#define FDR_EXEC_TEDDY_RES_OLD_4 \ + m128 res_old_1 = zeroes128(); \ + m128 res_old_2 = zeroes128(); \ + m128 res_old_3 = zeroes128(); + +#define FDR_EXEC_TEDDY_RES_OLD(n) FDR_EXEC_TEDDY_RES_OLD_##n + +#define PREP_CONF_FN_1(mask_base, val) \ + prep_conf_teddy_m1(mask_base, val) + +#define PREP_CONF_FN_2(mask_base, val) \ + prep_conf_teddy_m2(mask_base, &res_old_1, val) + +#define PREP_CONF_FN_3(mask_base, val) \ + prep_conf_teddy_m3(mask_base, &res_old_1, &res_old_2, val) + +#define PREP_CONF_FN_4(mask_base, val) \ + prep_conf_teddy_m4(mask_base, &res_old_1, &res_old_2, &res_old_3, val) + +#define PREP_CONF_FN(mask_base, val, n) \ + PREP_CONF_FN_##n(mask_base, val) + #define FDR_EXEC_TEDDY(fdr, a, control, n_msk, conf_fn) \ do { \ const u8 *buf_end = a->buf + a->len; \ @@ -581,7 +769,8 @@ do { \ \ return HWLM_SUCCESS; \ } while(0) -#endif // __AVX2__ + +#endif // HAVE_AVX2 HAVE_AVX512 hwlm_error_t fdr_exec_teddy_msks1(const struct FDR *fdr, const struct FDR_Runtime_Args *a, diff --git a/src/fdr/teddy_avx2.c b/src/fdr/teddy_avx2.c index 11ea0f8e..4091963c 100644 --- a/src/fdr/teddy_avx2.c +++ b/src/fdr/teddy_avx2.c @@ -109,6 +109,31 @@ const u8 ALIGN_AVX_DIRECTIVE p_mask_arr256[33][64] = { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00} }; +#define CONF_FAT_CHUNK_64(chunk, bucket, off, reason, conf_fn) \ +do { \ + if (unlikely(chunk != ones_u64a)) { \ + chunk = ~chunk; \ + conf_fn(&chunk, bucket, off, confBase, reason, a, ptr, \ + &control, &last_match); \ + CHECK_HWLM_TERMINATE_MATCHING; \ + } \ +} while(0) + +#define CONF_FAT_CHUNK_32(chunk, bucket, off, reason, conf_fn) \ +do { \ + if (unlikely(chunk != ones_u32a)) { \ + chunk = ~chunk; \ + conf_fn(&chunk, bucket, off, confBase, reason, a, ptr, \ + &control, &last_match); \ + CHECK_HWLM_TERMINATE_MATCHING; \ + } \ +} while(0) + +static really_inline +const m256 *getMaskBase_avx2(const struct Teddy *teddy) { + return (const m256 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy))); +} + #ifdef ARCH_64_BIT #define CONFIRM_FAT_TEDDY(var, bucket, offset, reason, conf_fn) \ do { \ @@ -120,30 +145,10 @@ do { \ r = interleave256hi(var, swap); \ u64a part3 = extractlow64from256(r); \ u64a part4 = extract64from256(r, 1); \ - if (unlikely(part1 != ones_u64a)) { \ - part1 = ~part1; \ - conf_fn(&part1, bucket, offset, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part2 != ones_u64a)) { \ - part2 = ~part2; \ - conf_fn(&part2, bucket, offset + 4, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part3 != ones_u64a)) { \ - part3 = ~part3; \ - conf_fn(&part3, bucket, offset + 8, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part4 != ones_u64a)) { \ - part4 = ~part4; \ - conf_fn(&part4, bucket, offset + 12, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ + CONF_FAT_CHUNK_64(part1, bucket, offset, reason, conf_fn); \ + CONF_FAT_CHUNK_64(part2, bucket, offset + 4, reason, conf_fn); \ + CONF_FAT_CHUNK_64(part3, bucket, offset + 8, reason, conf_fn); \ + CONF_FAT_CHUNK_64(part4, bucket, offset + 12, reason, conf_fn); \ } \ } while(0) #else @@ -161,53 +166,14 @@ do { \ u32 part6 = extract32from256(r, 1); \ u32 part7 = extract32from256(r, 2); \ u32 part8 = extract32from256(r, 3); \ - if (unlikely(part1 != ones_u32)) { \ - part1 = ~part1; \ - conf_fn(&part1, bucket, offset, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part2 != ones_u32)) { \ - part2 = ~part2; \ - conf_fn(&part2, bucket, offset + 2, confBase, reason, a, ptr, \ - &control, &last_match); \ - } \ - if (unlikely(part3 != ones_u32)) { \ - part3 = ~part3; \ - conf_fn(&part3, bucket, offset + 4, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part4 != ones_u32)) { \ - part4 = ~part4; \ - conf_fn(&part4, bucket, offset + 6, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part5 != ones_u32)) { \ - part5 = ~part5; \ - conf_fn(&part5, bucket, offset + 8, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part6 != ones_u32)) { \ - part6 = ~part6; \ - conf_fn(&part6, bucket, offset + 10, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part7 != ones_u32)) { \ - part7 = ~part7; \ - conf_fn(&part7, bucket, offset + 12, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ - if (unlikely(part8 != ones_u32)) { \ - part8 = ~part8; \ - conf_fn(&part8, bucket, offset + 14, confBase, reason, a, ptr, \ - &control, &last_match); \ - CHECK_HWLM_TERMINATE_MATCHING; \ - } \ + CONF_FAT_CHUNK_32(part1, bucket, offset, reason, conf_fn); \ + CONF_FAT_CHUNK_32(part2, bucket, offset + 2, reason, conf_fn); \ + CONF_FAT_CHUNK_32(part3, bucket, offset + 4, reason, conf_fn); \ + CONF_FAT_CHUNK_32(part4, bucket, offset + 6, reason, conf_fn); \ + CONF_FAT_CHUNK_32(part5, bucket, offset + 8, reason, conf_fn); \ + CONF_FAT_CHUNK_32(part6, bucket, offset + 10, reason, conf_fn); \ + CONF_FAT_CHUNK_32(part7, bucket, offset + 12, reason, conf_fn); \ + CONF_FAT_CHUNK_32(part8, bucket, offset + 14, reason, conf_fn); \ } \ } while(0) #endif @@ -277,11 +243,6 @@ m256 prep_conf_fat_teddy_m4(const m256 *maskBase, m256 *old_1, m256 *old_2, return or256(r, res_shifted_3); } -static really_inline -const m256 *getMaskBase_avx2(const struct Teddy *teddy) { - return (const m256 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy))); -} - #define FDR_EXEC_FAT_TEDDY_RES_OLD_1 \ do { \ } while(0) diff --git a/src/fdr/teddy_runtime_common.h b/src/fdr/teddy_runtime_common.h index c1333964..6b809cce 100644 --- a/src/fdr/teddy_runtime_common.h +++ b/src/fdr/teddy_runtime_common.h @@ -41,7 +41,7 @@ #include "util/uniform_ops.h" extern const u8 ALIGN_DIRECTIVE p_mask_arr[17][32]; -#if defined(__AVX2__) +#if defined(HAVE_AVX2) extern const u8 ALIGN_DIRECTIVE p_mask_arr256[33][64]; #endif @@ -123,7 +123,7 @@ void copyRuntBlock128(u8 *dst, const u8 *src, size_t len) { // for start zone, see below // lo ptr hi hi // |----------|-------|----------------|............| -// start 0 start+offset end(<=16) +// -start 0 -start+offset MIN(avail,16) // p_mask ffff..ff0000...........00ffff.......... // ptr < lo: // only start zone. @@ -182,7 +182,7 @@ m128 vectoredLoad128(m128 *p_mask, const u8 *ptr, const size_t start_offset, return u.val128; } -#if defined(__AVX2__) +#if defined(HAVE_AVX2) /* * \brief Copy a block of [0,31] bytes efficiently. * @@ -251,7 +251,7 @@ void copyRuntBlock256(u8 *dst, const u8 *src, size_t len) { // for start zone, see below // lo ptr hi hi // |----------|-------|----------------|............| -// start 0 start+offset end(<=32) +// -start 0 -start+offset MIN(avail,32) // p_mask ffff..ff0000...........00ffff.......... // ptr < lo: // only start zone. @@ -309,7 +309,77 @@ m256 vectoredLoad256(m256 *p_mask, const u8 *ptr, const size_t start_offset, return u.val256; } -#endif // __AVX2__ +#endif // HAVE_AVX2 + +#if defined(HAVE_AVX512) +// Note: p_mask is an output param that initialises a poison mask. +// u64a k = ones_u64a << n' >> m'; // m' < n' +// *p_mask = set_mask_m512(~k); +// means p_mask is consist of: +// (n' - m') poison bytes "0xff" at the beginning, +// followed by (64 - n') valid bytes "0x00", +// then followed by the rest m' poison bytes "0xff". +// ptr >= lo: +// no history. +// for end/short zone, ptr==lo and start_offset==0 +// for start zone, see below +// lo ptr hi hi +// |----------|-------|----------------|............| +// -start 0 -start+offset MIN(avail,64) +// p_mask ffff..ff0000...........00ffff.......... +// ptr < lo: +// only start zone. +// history +// ptr lo hi hi +// |----------|-------|----------------|............| +// 0 start start+offset end(<=64) +// p_mask ffff.....ffffff..ff0000...........00ffff.......... +static really_inline +m512 vectoredLoad512(m512 *p_mask, const u8 *ptr, const size_t start_offset, + const u8 *lo, const u8 *hi, const u8 *hbuf, size_t hlen, + const u32 nMasks) { + m512 val; + + uintptr_t copy_start; + uintptr_t copy_len; + + if (ptr >= lo) { // short/end/start zone + uintptr_t start = (uintptr_t)(ptr - lo); + uintptr_t avail = (uintptr_t)(hi - ptr); + if (avail >= 64) { + assert(start_offset - start <= 64); + u64a k = ones_u64a << (start_offset - start); + *p_mask = set_mask_m512(~k); + return loadu512(ptr); + } + assert(start_offset - start <= avail); + u64a k = ones_u64a << (64 - avail + start_offset - start) + >> (64 - avail); + *p_mask = set_mask_m512(~k); + copy_start = 0; + copy_len = avail; + } else { //start zone + uintptr_t need = MIN((uintptr_t)(lo - ptr), + MIN(hlen, nMasks - 1)); + uintptr_t start = (uintptr_t)(lo - ptr); + u64a j = 0x7fffffffffffffffULL >> (63 - need) << (start - need); + val = loadu_maskz_m512(j, &hbuf[hlen - start]); + uintptr_t end = MIN(64, (uintptr_t)(hi - ptr)); + assert(start + start_offset <= end); + u64a k = ones_u64a << (64 - end + start + start_offset) >> (64 - end); + *p_mask = set_mask_m512(~k); + copy_start = start; + copy_len = end - start; + } + + assert(copy_len < 64); + assert(copy_len > 0); + u64a j = ones_u64a >> (64 - copy_len) << copy_start; + val = loadu_mask_m512(val, j, ptr); + + return val; +} +#endif // HAVE_AVX512 static really_inline u64a getConfVal(const struct FDR_Runtime_Args *a, const u8 *ptr, u32 byte, diff --git a/src/util/simd_utils.h b/src/util/simd_utils.h index 047cdbab..8c469d16 100644 --- a/src/util/simd_utils.h +++ b/src/util/simd_utils.h @@ -755,6 +755,10 @@ m256 combine2x128(m128 hi, m128 lo) { } #endif //AVX2 +#if defined(HAVE_AVX512) +#define extract128from512(a, imm) _mm512_extracti32x4_epi32(a, imm) +#endif + /**** **** 384-bit Primitives ****/ @@ -969,6 +973,13 @@ m512 set8x64(u64a a) { return _mm512_set1_epi64(a); } +static really_inline +m512 set512_64(u64a hi_3, u64a hi_2, u64a hi_1, u64a hi_0, + u64a lo_3, u64a lo_2, u64a lo_1, u64a lo_0) { + return _mm512_set_epi64(hi_3, hi_2, hi_1, hi_0, + lo_3, lo_2, lo_1, lo_0); +} + static really_inline m512 set4x128(m128 a) { return _mm512_broadcast_i32x4(a); @@ -1059,6 +1070,7 @@ m512 lshift64_m512(m512 a, unsigned b) { #if defined(HAVE_AVX512) #define rshift64_m512(a, b) _mm512_srli_epi64((a), (b)) #define rshift128_m512(a, count_immed) _mm512_bsrli_epi128(a, count_immed) +#define lshift128_m512(a, count_immed) _mm512_bslli_epi128(a, count_immed) #endif #if !defined(_MM_CMPINT_NE) @@ -1169,6 +1181,11 @@ static really_inline m512 loadu_mask_m512(m512 src, __mmask64 k, const void *ptr) { return _mm512_mask_loadu_epi8(src, k, ptr); } + +static really_inline +m512 set_mask_m512(__mmask64 k) { + return _mm512_movm_epi8(k); +} #endif // packed unaligned store of first N bytes