Added proper bounds checking to SIMD wide-exp function, as it was blo…

…wing up in the exp(-hnu/kT) term for continua at very short wavelengths (< 1 nm). This is very unlikely to have affected results, as it produces nans as soon as it explodes

Source/SimdImpl_AVX2FMA.cpp

@@ -57,6 +57,24 @@ inline __m256d pow2n(const __m256d n) {
     __m256d d = _mm256_castsi256_pd(c);   // bit-cast back to double
     return d;
 }
+
+inline __m256d abs_pd_avx(__m256d xIn)
+{
+    unsigned int data[2] = {0xFFFFFFFFu, 0x7FFFFFFFu};
+    __m256d mask = _mm256_broadcast_sd((double*)data);
+    return _mm256_and_pd(xIn, mask);
+}
+
+inline __m256i finite_mask_avx(__m256d x)
+{
+    __m256i i = _mm256_castpd_si256(x);
+    __m256i iShift = _mm256_sll_epi64(i, _mm_cvtsi64_si128(1));
+    __m256i exp_val = _mm256_set1_epi64x(0xFFE0000000000000);
+    __m256i result = ~_mm256_cmpeq_epi64(_mm256_and_si256(iShift, exp_val), exp_val);
+    return result;
+}
+
+
 // NOTE(cmo): AVX impl of exp_pd, based on Agner Fog's vector class
 // https://github.com/vectorclass/version2/blob/master/vectormath_exp.h
 // The implementation here, based on a classic Taylor series, rather than a
@@ -98,7 +116,24 @@ inline __m256d exp_pd_avx(__m256d xIn)
     __m256d n2 = pow2n(r);
     z = _mm256_mul_pd(_mm256_add_pd(z, _mm256_set1_pd(1.0)), n2);
 
-    // TODO(cmo): Probably should have some of the nan/inf error handling code.
+    // NOTE(cmo): Error/edge-case handling code. The previous warning was prophetic.
+    // abs(xIn) < xMax
+    __m256i mask1 = _mm256_castpd_si256(_mm256_cmp_pd(abs_pd_avx(xIn), _mm256_set1_pd(xMax), 1));
+    __m256i mask2 = finite_mask_avx(xIn);
+    __m256i mask = mask1 & mask2;
+    // if all mask is set, then exit normally.
+    if (_mm256_testc_si256(mask, _mm256_set1_epi64x(-1)) != 0)
+        return z;
+
+    __m256d maskd = _mm256_castsi256_pd(mask);
+    __m256d inputSign = _mm256_and_pd(xIn, _mm256_set1_pd(-0.0));
+    __m256d inf = _mm256_castsi256_pd(_mm256_set1_epi64x(0x7FF0000000000000));
+    r = _mm256_blendv_pd(inf, _mm256_set1_pd(0.0), inputSign); // values for over/underflow/inf
+    z = _mm256_blendv_pd(r, z, maskd); // +/- underflow
+
+    __m256d nan_mask = _mm256_cmp_pd(xIn, xIn, 3); // check for unordered comparison, i.e. a value is nan
+    z = _mm256_blendv_pd(z, xIn, nan_mask); // set output to nan if input is nan
+
     return z;
 }
 

Source/SimdImpl_AVX512.cpp

@@ -57,6 +57,23 @@ inline __m512d pow2n(const __m512d n) {
     __m512d d = _mm512_castsi512_pd(c);   // bit-cast back to double
     return d;
 }
+
+inline __m512d abs_pd_avx512(__m512d xIn)
+{
+    unsigned int data[2] = {0xFFFFFFFFu, 0x7FFFFFFFu};
+    __m512d mask = _mm512_broadcast_f64x4(_mm256_broadcast_sd(((double*)data)));
+    return _mm512_and_pd(xIn, mask);
+}
+
+inline __mmask8 finite_mask_avx512(__m512d x)
+{
+    __m512i i = _mm512_castpd_si512(x);
+    __m512i iShift = _mm512_sll_epi64(i, _mm_cvtsi64_si128(1));
+    __m512i exp_val = _mm512_set1_epi64(0xFFE0000000000000);
+    __mmask8 result = ~_mm512_cmpeq_epi64_mask(_mm512_and_si512(iShift, exp_val), exp_val);
+    return result;
+}
+
 // NOTE(cmo): AVX impl of exp_pd, based on Agner Fog's vector class
 // https://github.com/vectorclass/version2/blob/master/vectormath_exp.h
 // The implementation here, based on a classic Taylor series, rather than a
@@ -98,7 +115,23 @@ inline __m512d exp_pd_avx512(__m512d xIn)
     __m512d n2 = pow2n(r);
     z = _mm512_mul_pd(_mm512_add_pd(z, _mm512_set1_pd(1.0)), n2);
 
-    // TODO(cmo): Probably should have some of the nan/inf error handling code.
+    // NOTE(cmo): Error/edge-case handling code. The previous warning was prophetic.
+    // abs(xIn) < xMax
+    __mmask8 mask1 = _mm512_cmp_pd_mask(abs_pd_avx512(xIn), _mm512_set1_pd(xMax), 1);
+    __mmask8 mask2 = finite_mask_avx512(xIn);
+    __mmask8 mask = mask1 & mask2;
+    // if all mask is set, then exit normally.
+    if (mask == 255)
+        return z;
+
+    // __m256d maskd = _mm256_castsi256_pd(mask);
+    __mmask8 inputNegative = _mm512_cmp_pd_mask(xIn, _mm512_set1_pd(0.0), 1);
+    __m512d inf = _mm512_castsi512_pd(_mm512_set1_epi64(0x7FF0000000000000));
+    r = _mm512_mask_blend_pd(inputNegative, inf, _mm512_set1_pd(0.0)); // values for over/underflow/inf
+    z = _mm512_mask_blend_pd(mask, r, z); // +/- underflow
+
+    __mmask8 nan_mask = _mm512_cmp_pd_mask(xIn, xIn, 3); // check for unordered comparison, i.e. a value is nan
+    z = _mm512_mask_blend_pd(nan_mask, z, xIn); // set output to nan if input is nan
     return z;
 }