Merge pull request #72 from howjmay/rcp

feat: Add _mm_rcp_[ps|ss]
pattonkan · Jan 27, 2024 · b0befa1 · b0befa1
2 parents 33badd5 + 029ee45
commit b0befa1
Show file tree

Hide file tree

Showing 2 changed files with 35 additions and 26 deletions.
diff --git a/sse2rvv.h b/sse2rvv.h
@@ -2572,9 +2572,18 @@ FORCE_INLINE __m64 _m_pshufw(__m64 a, int imm8) {
   return _mm_shuffle_pi16(a, imm8);
 }
 
-// FORCE_INLINE __m128 _mm_rcp_ps (__m128 a) {}
+FORCE_INLINE __m128 _mm_rcp_ps(__m128 a) {
+  // TODO add high precision mode
+  vfloat32m1_t _a = vreinterpretq_m128_f32(a);
+  return vreinterpretq_f32_m128(__riscv_vfrec7_v_f32m1(_a, 4));
+}
 
-// FORCE_INLINE __m128 _mm_rcp_ss (__m128 a) {}
+FORCE_INLINE __m128 _mm_rcp_ss(__m128 a) {
+  // TODO add high precision mode
+  vfloat32m1_t _a = vreinterpretq_m128_f32(a);
+  vfloat32m1_t recip = __riscv_vfrec7_v_f32m1(_a, 4);
+  return vreinterpretq_f32_m128(__riscv_vslideup_vx_f32m1(_a, recip, 0, 1));
+}
 
 // FORCE_INLINE __m128d _mm_round_pd (__m128d a, int rounding) {}
 

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -3017,35 +3017,35 @@ result_t test_m_pshufw(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
 }
 
 result_t test_mm_rcp_ps(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   const float *_a = impl.test_cases_float_pointer1;
-  //   float dx = 1.0f / _a[0];
-  //   float dy = 1.0f / _a[1];
-  //   float dz = 1.0f / _a[2];
-  //   float dw = 1.0f / _a[3];
-  //
-  //   __m128 a = load_m128(_a);
-  //   __m128 c = _mm_rcp_ps(a);
-  //   return validate_float_error(c, dx, dy, dz, dw, 0.001f);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  const float *_a = impl.test_cases_float_pointer1;
+  float _c[4];
+  for (int i = 0; i < 4; i++) {
+    _c[i] = 1.0f / _a[i];
+  }
+
+  __m128 a = load_m128(_a);
+  __m128 c = _mm_rcp_ps(a);
+  return validate_float_error(c, _c[0], _c[1], _c[2], _c[3], 0.01f);
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_rcp_ss(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   const float *_a = impl.test_cases_float_pointer1;
-  //
-  //   float dx = 1.0f / _a[0];
-  //   float dy = _a[1];
-  //   float dz = _a[2];
-  //   float dw = _a[3];
-  //   __m128 a = load_m128(_a);
-  //   __m128 c = _mm_rcp_ss(a);
-  //   return validate_float_error(c, dx, dy, dz, dw, 0.001f);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  const float *_a = impl.test_cases_float_pointer1;
+  float _c[4];
+  _c[0] = 1.0f / _a[0];
+  _c[1] = _a[1];
+  _c[2] = _a[2];
+  _c[3] = _a[3];
+  __m128 a = load_m128(_a);
+  __m128 c = _mm_rcp_ss(a);
+  return validate_float_error(c, _c[0], _c[1], _c[2], _c[3], 0.01f);
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_rsqrt_ps(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {