Merge pull request numpy#48 from howjmay/vst2_s8

howjmay · web-flow · commit ebf8a79a672d · 2023-11-07T11:45:56.000+01:00
feat: Add vst2_s8
diff --git a/neon2rvv.h b/neon2rvv.h
@@ -2822,7 +2822,7 @@ FORCE_INLINE int8x8x2_t vld2_s8(const int8_t *__a) { return __riscv_vlseg2e8_v_i
 
 // FORCE_INLINE uint64x1x2_t vld2_dup_u64(const uint64_t *__a);
 
-// FORCE_INLINE void vst2_s8(int8_t *__a, int8x8x2_t __b);
+FORCE_INLINE void vst2_s8(int8_t *__a, int8x8x2_t __b) { return __riscv_vsseg2e8_v_i8mf2x2(__a, __b, 8); }
 
 // FORCE_INLINE void vst2_s16(int16_t *__a, int16x4x2_t __b);
 
diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -3701,7 +3701,14 @@ result_t test_vld2_dup_s64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { retu
 
 result_t test_vld2_dup_u64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
 
-result_t test_vst2_s8(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vst2_s8(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+  const int8_t *_a = (int8_t *)impl.test_cases_int_pointer1;
+  int8_t _b[16];
+  int8x8x2_t a = vld2_s8(_a);
+  vst2_s8(_b, a);
+  return validate_int8(a, _b[0], _b[2], _b[4], _b[6], _b[8], _b[10], _b[12], _b[14], _b[1], _b[3], _b[5], _b[7], _b[9],
+                       _b[11], _b[13], _b[15]);
+}
 
 result_t test_vst2_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
 
