diff options
| author | Mehmet Samet Duman <yongdohyun@projecttick.org> | 2026-04-02 19:56:09 +0300 |
|---|---|---|
| committer | Mehmet Samet Duman <yongdohyun@projecttick.org> | 2026-04-02 19:56:09 +0300 |
| commit | 7fb132859fda54aa96bc9dd46d302b343eeb5a02 (patch) | |
| tree | b43ae77d7451fb470a260c03349a1caf2846c5e5 /neozip/arch/loongarch/slide_hash_lsx.c | |
| parent | b1e34e861b5d732afe828d58aad2c638135061fd (diff) | |
| parent | c2712b8a345191f6ed79558c089777df94590087 (diff) | |
| download | Project-Tick-7fb132859fda54aa96bc9dd46d302b343eeb5a02.tar.gz Project-Tick-7fb132859fda54aa96bc9dd46d302b343eeb5a02.zip | |
Add 'neozip/' from commit 'c2712b8a345191f6ed79558c089777df94590087'
git-subtree-dir: neozip
git-subtree-mainline: b1e34e861b5d732afe828d58aad2c638135061fd
git-subtree-split: c2712b8a345191f6ed79558c089777df94590087
Diffstat (limited to 'neozip/arch/loongarch/slide_hash_lsx.c')
| -rw-r--r-- | neozip/arch/loongarch/slide_hash_lsx.c | 54 |
1 files changed, 54 insertions, 0 deletions
diff --git a/neozip/arch/loongarch/slide_hash_lsx.c b/neozip/arch/loongarch/slide_hash_lsx.c new file mode 100644 index 0000000000..f4c94ea70d --- /dev/null +++ b/neozip/arch/loongarch/slide_hash_lsx.c @@ -0,0 +1,54 @@ +/* + * LSX optimized hash slide, based on Intel SSE implementation + * + * Copyright (C) 2017 Intel Corporation + * Copyright (C) 2025 Vladislav Shchapov <vladislav@shchapov.ru> + * Authors: + * Arjan van de Ven <arjan@linux.intel.com> + * Jim Kukunas <james.t.kukunas@linux.intel.com> + * + * For conditions of distribution and use, see copyright notice in zlib.h + */ + +#ifdef LOONGARCH_LSX + +#include "zbuild.h" +#include "deflate.h" + +#include <lsxintrin.h> +#include <assert.h> + +static inline void slide_hash_chain(Pos *table, uint32_t entries, const __m128i wsize) { + table += entries; + table -= 16; + + /* ZALLOC allocates this pointer unless the user chose a custom allocator. + * Our alloc function is aligned to 64 byte boundaries */ + do { + __m128i value0, value1, result0, result1; + + value0 = __lsx_vld(table, 0); + value1 = __lsx_vld(table, 16); + result0 = __lsx_vssub_hu(value0, wsize); + result1 = __lsx_vssub_hu(value1, wsize); + __lsx_vst(result0, table, 0); + __lsx_vst(result1, table, 16); + + table -= 16; + entries -= 16; + } while (entries > 0); +} + +Z_INTERNAL void slide_hash_lsx(deflate_state *s) { + Assert(s->w_size <= UINT16_MAX, "w_size should fit in uint16_t"); + uint16_t wsize = (uint16_t)s->w_size; + const __m128i xmm_wsize = __lsx_vreplgr2vr_h((short)wsize); + + assert(((uintptr_t)s->head & 15) == 0); + assert(((uintptr_t)s->prev & 15) == 0); + + slide_hash_chain(s->head, HASH_SIZE, xmm_wsize); + slide_hash_chain(s->prev, wsize, xmm_wsize); +} + +#endif |
