diff options
Diffstat (limited to 'src/vppinfra/vector')
| -rw-r--r-- | src/vppinfra/vector/count_equal.h | 238 |
1 files changed, 238 insertions, 0 deletions
diff --git a/src/vppinfra/vector/count_equal.h b/src/vppinfra/vector/count_equal.h new file mode 100644 index 00000000000..98770cff7c0 --- /dev/null +++ b/src/vppinfra/vector/count_equal.h @@ -0,0 +1,238 @@ +/* SPDX-License-Identifier: Apache-2.0 + * Copyright(c) 2021 Cisco Systems, Inc. + */ + +#ifndef included_vector_count_equal_h +#define included_vector_count_equal_h +#include <vppinfra/clib.h> + +static_always_inline uword +clib_count_equal_u64 (u64 *data, uword max_count) +{ + uword count; + u64 first; + + if (max_count <= 1) + return max_count; + if (data[0] != data[1]) + return 1; + + count = 0; + first = data[0]; + +#if defined(CLIB_HAVE_VEC256) + u64x4 splat = u64x4_splat (first); + while (count + 3 < max_count) + { + u64 bmp; + bmp = u8x32_msb_mask ((u8x32) (u64x4_load_unaligned (data) == splat)); + if (bmp != 0xffffffff) + { + count += count_trailing_zeros (~bmp) / 8; + return count; + } + + data += 4; + count += 4; + } +#else + count += 2; + data += 2; + while (count + 3 < max_count && ((data[0] ^ first) | (data[1] ^ first) | + (data[2] ^ first) | (data[3] ^ first)) == 0) + { + data += 4; + count += 4; + } +#endif + while (count < max_count && (data[0] == first)) + { + data += 1; + count += 1; + } + return count; +} + +static_always_inline uword +clib_count_equal_u32 (u32 *data, uword max_count) +{ + uword count; + u32 first; + + if (max_count <= 1) + return max_count; + if (data[0] != data[1]) + return 1; + + count = 0; + first = data[0]; + +#if defined(CLIB_HAVE_VEC256) + u32x8 splat = u32x8_splat (first); + while (count + 7 < max_count) + { + u64 bmp; + bmp = u8x32_msb_mask ((u8x32) (u32x8_load_unaligned (data) == splat)); + if (bmp != 0xffffffff) + { + count += count_trailing_zeros (~bmp) / 4; + return count; + } + + data += 8; + count += 8; + } +#elif defined(CLIB_HAVE_VEC128) && defined(CLIB_HAVE_VEC128_MSB_MASK) + u32x4 splat = u32x4_splat (first); + while (count + 3 < max_count) + { + u64 bmp; + bmp = u8x16_msb_mask ((u8x16) (u32x4_load_unaligned (data) == splat)); + if (bmp != 0xffff) + { + count += count_trailing_zeros (~bmp) / 4; + return count; + } + + data += 4; + count += 4; + } +#else + count += 2; + data += 2; + while (count + 3 < max_count && ((data[0] ^ first) | (data[1] ^ first) | + (data[2] ^ first) | (data[3] ^ first)) == 0) + { + data += 4; + count += 4; + } +#endif + while (count < max_count && (data[0] == first)) + { + data += 1; + count += 1; + } + return count; +} + +static_always_inline uword +clib_count_equal_u16 (u16 *data, uword max_count) +{ + uword count; + u16 first; + + if (max_count <= 1) + return max_count; + if (data[0] != data[1]) + return 1; + + count = 0; + first = data[0]; + +#if defined(CLIB_HAVE_VEC256) + u16x16 splat = u16x16_splat (first); + while (count + 15 < max_count) + { + u64 bmp; + bmp = u8x32_msb_mask ((u8x32) (u16x16_load_unaligned (data) == splat)); + if (bmp != 0xffffffff) + { + count += count_trailing_zeros (~bmp) / 2; + return count; + } + + data += 16; + count += 16; + } +#elif defined(CLIB_HAVE_VEC128) && defined(CLIB_HAVE_VEC128_MSB_MASK) + u16x8 splat = u16x8_splat (first); + while (count + 7 < max_count) + { + u64 bmp; + bmp = u8x16_msb_mask ((u8x16) (u16x8_load_unaligned (data) == splat)); + if (bmp != 0xffff) + { + count += count_trailing_zeros (~bmp) / 2; + return count; + } + + data += 8; + count += 8; + } +#else + count += 2; + data += 2; + while (count + 3 < max_count && ((data[0] ^ first) | (data[1] ^ first) | + (data[2] ^ first) | (data[3] ^ first)) == 0) + { + data += 4; + count += 4; + } +#endif + while (count < max_count && (data[0] == first)) + { + data += 1; + count += 1; + } + return count; +} + +static_always_inline uword +clib_count_equal_u8 (u8 *data, uword max_count) +{ + uword count; + u8 first; + + if (max_count <= 1) + return max_count; + if (data[0] != data[1]) + return 1; + + count = 0; + first = data[0]; + +#if defined(CLIB_HAVE_VEC256) + u8x32 splat = u8x32_splat (first); + while (count + 31 < max_count) + { + u64 bmp; + bmp = u8x32_msb_mask ((u8x32) (u8x32_load_unaligned (data) == splat)); + if (bmp != 0xffffffff) + return max_count; + + data += 32; + count += 32; + } +#elif defined(CLIB_HAVE_VEC128) && defined(CLIB_HAVE_VEC128_MSB_MASK) + u8x16 splat = u8x16_splat (first); + while (count + 15 < max_count) + { + u64 bmp; + bmp = u8x16_msb_mask ((u8x16) (u8x16_load_unaligned (data) == splat)); + if (bmp != 0xffff) + { + count += count_trailing_zeros (~bmp); + return count; + } + + data += 16; + count += 16; + } +#else + count += 2; + data += 2; + while (count + 3 < max_count && ((data[0] ^ first) | (data[1] ^ first) | + (data[2] ^ first) | (data[3] ^ first)) == 0) + { + data += 4; + count += 4; + } +#endif + while (count < max_count && (data[0] == first)) + { + data += 1; + count += 1; + } + return count; +} +#endif |