From 4c53ff459595c9ddc05002ee1847313127175b5f Mon Sep 17 00:00:00 2001 From: Damjan Marion Date: Thu, 28 Oct 2021 23:03:04 +0200 Subject: vppinfra: vectorized index to pointer function Type: improvement Change-Id: I05e1a8fa31761b113355123429d72da18881d4b0 Signed-off-by: Damjan Marion --- src/vlib/buffer_funcs.h | 121 ++++++++++++------------------------------------ 1 file changed, 29 insertions(+), 92 deletions(-) (limited to 'src/vlib/buffer_funcs.h') diff --git a/src/vlib/buffer_funcs.h b/src/vlib/buffer_funcs.h index 77964fde821..30fe23443ab 100644 --- a/src/vlib/buffer_funcs.h +++ b/src/vlib/buffer_funcs.h @@ -42,6 +42,7 @@ #include #include +#include #include #include #include @@ -201,102 +202,38 @@ vlib_buffer_pool_get_default_for_numa (vlib_main_t * vm, u32 numa_node) @param offset - (i32) offset applied to each pointer */ static_always_inline void -vlib_get_buffers_with_offset (vlib_main_t * vm, u32 * bi, void **b, int count, +vlib_get_buffers_with_offset (vlib_main_t *vm, u32 *bi, void **b, u32 count, i32 offset) { uword buffer_mem_start = vm->buffer_main->buffer_mem_start; -#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); - 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) - { -#ifdef CLIB_HAVE_VEC256 - u64x4 b0 = u64x4_from_u32x4 (u32x4_load_unaligned (bi)); - /* shift and add to get vlib_buffer_t pointer */ - u64x4_store_unaligned ((b0 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b); -#elif defined (CLIB_HAVE_VEC128) - u64x2 off = u64x2_splat (buffer_mem_start + offset); - u32x4 bi4 = u32x4_load_unaligned (bi); - u64x2 b0 = u64x2_from_u32x4 ((u32x4) bi4); -#if defined (__aarch64__) - u64x2 b1 = u64x2_from_u32x4_high ((u32x4) bi4); -#else - bi4 = u32x4_shuffle (bi4, 2, 3, 0, 1); - u64x2 b1 = u64x2_from_u32x4 ((u32x4) bi4); -#endif - u64x2_store_unaligned ((b0 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b); - u64x2_store_unaligned ((b1 << CLIB_LOG2_CACHE_LINE_BYTES) + off, b + 2); -#else - b[0] = vlib_buffer_ptr_from_index (buffer_mem_start, bi[0], offset); - b[1] = vlib_buffer_ptr_from_index (buffer_mem_start, bi[1], offset); - b[2] = vlib_buffer_ptr_from_index (buffer_mem_start, bi[2], offset); - b[3] = vlib_buffer_ptr_from_index (buffer_mem_start, bi[3], offset); -#endif - b += 4; - bi += 4; - count -= 4; - } - while (count) + void *base = (void *) (buffer_mem_start + offset); + int objsize = __builtin_object_size (b, 0); + const int sh = CLIB_LOG2_CACHE_LINE_BYTES; + + if (COMPILE_TIME_CONST (count) == 0 && objsize >= 64 * sizeof (b[0]) && + (objsize & ((8 * sizeof (b[0])) - 1)) == 0) { - b[0] = vlib_buffer_ptr_from_index (buffer_mem_start, bi[0], offset); - b += 1; - bi += 1; - count -= 1; + u32 n = round_pow2 (count, 8); + ASSERT (objsize >= count); + CLIB_ASSUME (objsize >= count); + while (n >= 64) + { + clib_index_to_ptr_u32 (bi, base, sh, b, 64); + b += 64; + bi += 64; + n -= 64; + } + + while (n) + { + clib_index_to_ptr_u32 (bi, base, sh, b, 8); + b += 8; + bi += 8; + n -= 8; + } } + else + clib_index_to_ptr_u32 (bi, base, sh, b, count); } /** \brief Translate array of buffer indices into buffer pointers @@ -308,7 +245,7 @@ vlib_get_buffers_with_offset (vlib_main_t * vm, u32 * bi, void **b, int count, */ static_always_inline void -vlib_get_buffers (vlib_main_t * vm, u32 * bi, vlib_buffer_t ** b, int count) +vlib_get_buffers (vlib_main_t *vm, u32 *bi, vlib_buffer_t **b, u32 count) { vlib_get_buffers_with_offset (vm, bi, (void **) b, count, 0); } -- cgit 1.2.3-korg