Rewriting reduce_all() helpers. They aren't actually used, but they are good to have around.
diff --git a/encoder/cppspmd_sse.h b/encoder/cppspmd_sse.h
index b39cb82..9a97eeb 100644
--- a/encoder/cppspmd_sse.h
+++ b/encoder/cppspmd_sse.h
@@ -1327,33 +1327,15 @@
CPPSPMD_FORCE_INLINE float reduce_add(vfloat v)
{
__m128 k3210 = _mm_castsi128_ps(blendv_mask_epi32(_mm_setzero_si128(), _mm_castps_si128(v.m_value), m_exec.m_mask));
-
-//#if CPPSPMD_SSE2
-#if 1
- // See https://stackoverflow.com/questions/6996764/fastest-way-to-do-horizontal-sse-vector-sum-or-other-reduction/35270026#35270026
- __m128 shuf = _mm_shuffle_ps(k3210, k3210, _MM_SHUFFLE(2, 3, 0, 1));
- __m128 sums = _mm_add_ps(k3210, shuf);
- shuf = _mm_movehl_ps(shuf, sums);
- sums = _mm_add_ss(sums, shuf);
- return _mm_cvtss_f32(sums);
-#else
- // This is pretty slow.
- __m128 a = _mm_hadd_ps(k3210, k3210);
- __m128 b = _mm_hadd_ps(a, a);
- return extractf_ps_x(b);
-#endif
+ __m128 temp = _mm_add_ps(_mm_shuffle_ps(k3210, k3210, _MM_SHUFFLE(0, 1, 2, 3)), k3210);
+ return _mm_cvtss_f32(_mm_add_ss(_mm_movehl_ps(temp, temp), temp));
}
-
+
CPPSPMD_FORCE_INLINE int reduce_add(vint v)
{
__m128i k3210 = blendv_mask_epi32(_mm_setzero_si128(), v.m_value, m_exec.m_mask);
-
- // See https://stackoverflow.com/questions/6996764/fastest-way-to-do-horizontal-sse-vector-sum-or-other-reduction/35270026#35270026
- __m128i shuf = _mm_shuffle_epi32(k3210, _MM_SHUFFLE(2, 3, 0, 1));
- __m128i sums = _mm_add_epi32(k3210, shuf);
- shuf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(shuf), _mm_castsi128_ps(sums)));
- sums = _mm_add_epi32(sums, shuf);
- return extract_x(sums);
+ __m128i temp = _mm_add_epi32(_mm_shuffle_epi32(k3210, _MM_SHUFFLE(0, 1, 2, 3)), k3210);
+ return extract_x(_mm_add_epi32(_mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(temp), _mm_castsi128_ps(temp))), temp));
}
#include "cppspmd_math_declares.h"
@@ -1686,6 +1668,12 @@
CPPSPMD_FORCE_INLINE vint undefined_vint() { return vint{ _mm_undefined_si128() }; }
CPPSPMD_FORCE_INLINE vfloat undefined_vfloat() { return vfloat{ _mm_undefined_ps() }; }
+CPPSPMD_FORCE_INLINE vint vint_lane_set(int v0, int v1, int v2, int v3) { return vint{ _mm_set_epi32(v3, v2, v1, v0) }; }
+CPPSPMD_FORCE_INLINE vfloat vfloat_lane_set(float v0, float v1, float v2, float v3) { return vfloat{ _mm_set_ps(v3, v2, v1, v0) }; }
+
+CPPSPMD_FORCE_INLINE vint vint_lane_set_r(int v3, int v2, int v1, int v0) { return vint{ _mm_set_epi32(v3, v2, v1, v0) }; }
+CPPSPMD_FORCE_INLINE vfloat vfloat_lane_set_r(float v3, float v2, float v1, float v0) { return vfloat{ _mm_set_ps(v3, v2, v1, v0) }; }
+
// control is an 8-bit immediate value containing 4 2-bit indices which shuffles the int32's in each 128-bit lane.
#define VINT_LANE_SHUFFLE_EPI32(a, control) vint(_mm_shuffle_epi32((a).m_value, control))
