Merge pull request #438 from howjmay/vqdmull_high_n

feat: Add vqdmull_high_n_[s16|s32]
howjmay · Jul 23, 2024 · 05bb5f5 · 05bb5f5
2 parents d10196b + 3d5c529
commit 05bb5f5
Show file tree

Hide file tree

Showing 3 changed files with 44 additions and 6 deletions.
diff --git a/neon2rvv.h b/neon2rvv.h
@@ -10617,9 +10617,17 @@ FORCE_INLINE int64x2_t vqdmull_n_s32(int32x2_t a, int32_t b) {
   return __riscv_vlmul_trunc_v_i64m2_i64m1(__riscv_vsll_vx_i64m2(ab_mul, 1, 2));
 }
 
-// FORCE_INLINE int32x4_t vqdmull_high_n_s16(int16x8_t a, int16_t b);
+FORCE_INLINE int32x4_t vqdmull_high_n_s16(int16x8_t a, int16_t b) {
+  vint16m1_t a_high = __riscv_vslidedown_vx_i16m1(a, 4, 8);
+  vint32m2_t ab_mul = __riscv_vwmul_vx_i32m2(a_high, b, 4);
+  return __riscv_vlmul_trunc_v_i32m2_i32m1(__riscv_vsll_vx_i32m2(ab_mul, 1, 4));
+}
 
-// FORCE_INLINE int64x2_t vqdmull_high_n_s32(int32x4_t a, int32_t b);
+FORCE_INLINE int64x2_t vqdmull_high_n_s32(int32x4_t a, int32_t b) {
+  vint32m1_t a_high = __riscv_vslidedown_vx_i32m1(a, 2, 4);
+  vint64m2_t ab_mul = __riscv_vwmul_vx_i64m2(a_high, b, 2);
+  return __riscv_vlmul_trunc_v_i64m2_i64m1(__riscv_vsll_vx_i64m2(ab_mul, 1, 2));
+}
 
 FORCE_INLINE int16x8_t vqdmulhq_n_s16(int16x8_t a, int16_t b) {
   vint16m1_t b_dup = vdupq_n_s16(b);

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -33960,9 +33960,39 @@ result_t test_vqdmull_n_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #endif  // ENABLE_TEST_ALL
 }
 
-result_t test_vqdmull_high_n_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vqdmull_high_n_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const int16_t *_a = (int16_t *)impl.test_cases_int_pointer1;
+  const int16_t *_b = (int16_t *)impl.test_cases_int_pointer2;
+  int32_t _c[4];
+  int16x8_t a = vld1q_s16(_a);
+
+  for (int i = 0; i < 4; i++) {
+    _c[i] = sat_dmull(_a[i + 4], _b[0]);
+  }
+  int32x4_t c = vqdmull_high_n_s16(a, _b[0]);
+  return validate_int32(c, _c[0], _c[1], _c[2], _c[3]);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
-result_t test_vqdmull_high_n_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vqdmull_high_n_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const int32_t *_a = (int32_t *)impl.test_cases_int_pointer1;
+  const int32_t *_b = (int32_t *)impl.test_cases_int_pointer2;
+  int64_t _c[2];
+  int32x4_t a = vld1q_s32(_a);
+
+  for (int i = 0; i < 2; i++) {
+    _c[i] = sat_dmull(_a[i + 2], _b[0]);
+  }
+  int64x2_t c = vqdmull_high_n_s32(a, _b[0]);
+  return validate_int64(c, _c[0], _c[1]);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
 result_t test_vqdmulhq_n_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #ifdef ENABLE_TEST_ALL

diff --git a/tests/impl.h b/tests/impl.h
@@ -2197,8 +2197,8 @@
   _(vmull_high_n_u32)                                                            \
   _(vqdmull_n_s16)                                                               \
   _(vqdmull_n_s32)                                                               \
-  /*_(vqdmull_high_n_s16)                                                     */ \
-  /*_(vqdmull_high_n_s32)                                                     */ \
+  _(vqdmull_high_n_s16)                                                          \
+  _(vqdmull_high_n_s32)                                                          \
   _(vqdmulhq_n_s16)                                                              \
   _(vqdmulhq_n_s32)                                                              \
   _(vqdmulh_n_s16)                                                               \