Repository navigation
truncation error fro simd vector size > 1 #209
Description
diff --git a/madspace/src/cpu/simd_x86_256.hpp b/madspace/src/cpu/simd_x86_256.hpp
index 6c2815e25..fab1f9752 100644
--- a/madspace/src/cpu/simd_x86_256.hpp
+++ b/madspace/src/cpu/simd_x86_256.hpp
@@ -28,7 +28,7 @@ struct IVec {
IVec() = default;
IVec(__m256i _v) : v(_v) {};
IVec(int _v) : v(_mm256_set1_epi64x(_v)) {};
- explicit IVec(__m256d _v) { v = _mm256_cvtepi32_epi64(_mm256_cvtpd_epi32(_v)); }
+ explicit IVec(__m256d _v) { v = _mm256_cvtepi32_epi64(_mm256_cvttpd_epi32(_v)); }
operator __m256i() { return v; }
IVec operator+=(IVec _v) {
v = _mm256_add_epi64(v, _v);
diff --git a/madspace/src/cpu/simd_x86_512.hpp b/madspace/src/cpu/simd_x86_512.hpp
index 1a7c97a98..74808d83c 100644
--- a/madspace/src/cpu/simd_x86_512.hpp
+++ b/madspace/src/cpu/simd_x86_512.hpp
@@ -21,7 +21,7 @@ struct IVec {
IVec() = default;
IVec(__m512i _v) : v(_v) {};
IVec(int _v) : v(_mm512_set1_epi64(_v)) {};
- explicit IVec(__m512d _v) { v = _mm512_cvtepi32_epi64(_mm512_cvtpd_epi32(_v)); }
+ explicit IVec(__m512d _v) { v = _mm512_cvtepi32_epi64(_mm512_cvttpd_epi32(_v)); }
operator __m512i() { return v; }
IVec operator+=(IVec _v) {
v = _mm512_add_epi64(v, _v);
It looks like when using simd vector size with avx2, the numbers are rounded instead of being truncated
It doesn't correspond to scalar library case e.g. https://github.com/MadGraphTeam/MadGraph7/blob/main/madspace/src/kernels/discrete.hpp#L14
And I think above patch is the fix using cvttpd instead of cvtpd?
i ran p p > ve ve~ j and scalar was giving 838.7pb, simd_vector_size=4 was crashing before the change above with and gives me the same numerical result 838.7pb
Metadata
Metadata
Assignees
Labels
No labels
Activity
theoheimel commentedon Oct 7, 2026
Thanks for pointing this out! This is in madspace SIMD mode (i.e., only affects phase space sampling) which is currently still experimental and has several bugs. I'm working on it now in #222 . Note that the SIMD mode for the matrix elements (accessible through the
cpu_modeparameter) is fully working, so we'll also make this distinction more clear in the run card.more SIMD bugfixes, including the one proposed in #209
sihyunjeon commentedon Oct 7, 2026
@theoheimel thanks for covering this :) closing since it's addressed in your PR