add RVV optimization for ZSTD_row_getMatchMask

Co-authored-by: gong-flying <gongxiaofei24@iscas.ac.cn>

diff --git a/lib/compress/zstd_lazy.c b/lib/compress/zstd_lazy.c
index 272ebe0ece7d9ceecaaa22a7ce0f067aa483c99b..ca8fb4194ee8f0a31580a2169e5670466c10542e 100644 (file)
--- a/lib/compress/zstd_lazy.c
+++ b/lib/compress/zstd_lazy.c
@@ -1050,6 +1050,36 @@ ZSTD_row_getNEONMask(const U32 rowEntries, const BYTE* const src, const BYTE tag
     }
 }
 #endif
+#if defined(ZSTD_ARCH_RISCV_RVV) && (__riscv_xlen == 64)
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getRVVMask(int nbChunks, const BYTE* const src, const BYTE tag, const U32 head)
+{
+    U16 matches[4] = {0};
+    int i;
+    assert(nbChunks == 1 || nbChunks == 2 || nbChunks == 4);
+
+    size_t vl = __riscv_vsetvl_e8m1(16);
+
+    for (i = 0; i < nbChunks; i++) {
+        vuint8m1_t chunk = __riscv_vle8_v_u8m1((const uint8_t*)(src + 16 * i), vl);
+        vbool8_t equalMask = __riscv_vmseq_vx_u8m1_b8(chunk, tag, vl);
+
+        size_t vl_w = __riscv_vsetvl_e16m2(16);
+        vuint16m2_t one = __riscv_vmv_v_x_u16m2(1, vl_w);
+        vuint16m2_t indices = __riscv_vid_v_u16m2(vl_w);
+        vuint16m2_t powers_of_2 = __riscv_vsll_vv_u16m2(one, indices, vl_w);
+        vuint16m2_t zero = __riscv_vmv_v_x_u16m2(0, vl_w);
+        vuint16m2_t selected_bits = __riscv_vmerge_vvm_u16m2(zero, powers_of_2, equalMask, vl_w);
+        vuint16m1_t reduction = __riscv_vredor_vs_u16m2_u16m1(selected_bits, __riscv_vmv_s_x_u16m1(0, vl_w), vl_w);
+        matches[i] = __riscv_vmv_x_s_u16m1_u16(reduction);
+    }
+    
+    if (nbChunks == 1) return ZSTD_rotateRight_U16(matches[0], head);
+    if (nbChunks == 2) return ZSTD_rotateRight_U32((U32)matches[1] << 16 | (U32)matches[0], head);
+    assert(nbChunks == 4);
+    return ZSTD_rotateRight_U64((U64)matches[3] << 48 | (U64)matches[2] << 32 | (U64)matches[1] << 16 | (U64)matches[0], head);
+}
+#endif
 
 /* Returns a ZSTD_VecMask (U64) that has the nth group (determined by
  * ZSTD_row_matchMaskGroupWidth) of bits set to 1 if the newly-computed "tag"
@@ -1069,14 +1099,20 @@ ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 headGr
 
     return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, headGrouped);
 
-#else /* SW or NEON-LE */
+#elif defined(ZSTD_ARCH_RISCV_RVV) && (__riscv_xlen == 64)
+
+    return ZSTD_row_getRVVMask(rowEntries, src, tag, headGrouped);
+
+#else
 
-# if defined(ZSTD_ARCH_ARM_NEON)
+#if defined(ZSTD_ARCH_ARM_NEON)
   /* This NEON path only works for little endian - otherwise use SWAR below */
     if (MEM_isLittleEndian()) {
         return ZSTD_row_getNEONMask(rowEntries, src, tag, headGrouped);
     }
-# endif /* ZSTD_ARCH_ARM_NEON */
+
+
+#endif
     /* SWAR */
     {   const int chunkSize = sizeof(size_t);
         const size_t shiftAmount = ((chunkSize * 8) - chunkSize);