diff options
| author |   Damjan Marion <damarion@cisco.com> | 2021-04-21 15:25:47 +0200 |
|---|---|---|
| committer |   Florin Coras <florin.coras@gmail.com> | 2021-04-21 18:30:52 +0000 |
| commit | 542088597886df774e63f841166721deeffef1c1 (patch) | |
| tree | 719589064b0a1d090398bfe39cbe5b448a96a5bf | |
| parent | 1c2297144a37ea6a42f5deec6e1613a297de3321 (diff) | |
buffers: vlib_get_buffers() with 512-bit SIMD
Type: improvement
Change-Id: Id8ce3ffc1299a38171b82a7082454412c840a40c
Signed-off-by: Damjan Marion <damarion@cisco.com>
| -rw-r--r-- | src/vlib/buffer_funcs.h | 51 | ||||
| -rw-r--r-- | src/vppinfra/vector_avx512.h | 1 |
2 files changed, 47 insertions, 5 deletions
diff --git a/src/vlib/buffer_funcs.h b/src/vlib/buffer_funcs.h index e4427d6c382..8ab9c4e53ad 100644 --- a/src/vlib/buffer_funcs.h +++ b/src/vlib/buffer_funcs.h @@ -235,20 +235,61 @@ vlib_get_buffers_with_offset (vlib_main_t * vm, u32 * bi, void **b, int count, i32 offset) { uword buffer_mem_start = vm->buffer_main->buffer_mem_start; -#ifdef CLIB_HAVE_VEC256 - u64x4 off = u64x4_splat (buffer_mem_start + offset); +#ifdef CLIB_HAVE_VEC512 + u64x8 of8 = u64x8_splat (buffer_mem_start + offset); + u64x4 off = u64x8_extract_lo (of8); /* if count is not const, compiler will not unroll while loop se we maintain two-in-parallel variant */ + while (count >= 32) + { + u64x8 b0 = u64x8_from_u32x8 (u32x8_load_unaligned (bi)); + u64x8 b1 = u64x8_from_u32x8 (u32x8_load_unaligned (bi + 8)); + u64x8 b2 = u64x8_from_u32x8 (u32x8_load_unaligned (bi + 16)); + u64x8 b3 = u64x8_from_u32x8 (u32x8_load_unaligned (bi + 24)); + /* shift and add to get vlib_buffer_t pointer */ + u64x8_store_unaligned ((b0 << CLIB_LOG2_CACHE_LINE_BYTES) + of8, b); + u64x8_store_unaligned ((b1 << CLIB_LOG2_CACHE_LINE_BYTES) + of8, b + 8); + u64x8_store_unaligned ((b2 << CLIB_LOG2_CACHE_LINE_BYTES) + of8, b + 16); + u64x8_store_unaligned ((b3 << CLIB_LOG2_CACHE_LINE_BYTES) + of8, b + 24); + b += 32; + bi += 32; + count -= 32; + } while (count >= 8) { + u64x8 b0 = u64x8_from_u32x8 (u32x8_load_unaligned (bi)); + /* shift and add to get vlib_buffer_t pointer */ + u64x8_store_unaligned ((b0 << CLIB_LOG2_CACHE_LINE_BYTES) + of8, b); + b += 8; + bi += 8; + count -= 8; + } +#elif defined CLIB_HAVE_VEC256 + u64x4 off = u64x4_splat (buffer_mem_start + offset); + /* if count is not const, compiler will not unroll while loop + se we maintain two-in-parallel variant */ + while (count >= 32) + { u64x4 b0 = u64x4_from_u32x4 (u32x4_load_unaligned (bi)); u64x4 b1 = u64x4_from_u32x4 (u32x4_load_unaligned (bi + 4)); + u64x4 b2 = u64x4_from_u32x4 (u32x4_load_unaligned (bi + 8)); + u64x4 b3 = u64x4_from_u32x4 (u32x4_load_unaligned (bi + 12)); + u64x4 b4 = u64x4_from_u32x4 (u32x4_load_unaligned (bi + 16)); + u64x4 b5 = u64x4_from_u32x4 (u32x4_load_unaligned (bi + 20)); + u64x4 b6 = u64x4_from_u32x4 (u32x4_load_unaligned (bi + 24)); + u64x4 b7 = u64x4_from_u32x4 (u32x4_load_unaligned (bi + 28)); /* shift and add to get vlib_buffer_t pointer */ u64x4_store_unaligned ((b0 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b); u64x4_store_unaligned ((b1 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 4); - b += 8; - bi += 8; - count -= 8; + u64x4_store_unaligned ((b2 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 8); + u64x4_store_unaligned ((b3 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 12); + u64x4_store_unaligned ((b4 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 16); + u64x4_store_unaligned ((b5 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 20); + u64x4_store_unaligned ((b6 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 24); + u64x4_store_unaligned ((b7 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 28); + b += 32; + bi += 32; + count -= 32; } #endif while (count >= 4) diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h index 3406a7ddb91..87999eee947 100644 --- a/src/vppinfra/vector_avx512.h +++ b/src/vppinfra/vector_avx512.h @@ -264,6 +264,7 @@ _ (u64x8, u8, epu64, _mm512, __m512i) _ (u16x16, u32x16, _mm512_cvtepi16_epi32, __m256i) _ (u32x16, u16x16, _mm512_cvtusepi32_epi16, __m512i) _ (u32x8, u16x8, _mm256_cvtusepi32_epi16, __m256i) +_ (u32x8, u64x8, _mm512_cvtepu32_epi64, __m256i) #undef _ #define _(vt, mt, bits, epi) \ |