Use native avx512 intrinsic to do int-to-float conversion if avilable

yumeyao · yumeyao · commit 844c60c8c82b · 2021-07-28T17:34:52.000+08:00
diff --git a/include/xsimd/math/xsimd_fp_manipulation.hpp b/include/xsimd/math/xsimd_fp_manipulation.hpp
@@ -77,14 +77,14 @@ namespace xsimd
     }
 
     // fexps are 8/11-bit numbers. so we can down-cast to int16_t/int32_t then do the job
-#if XSIMD_X86_INSTR_SET >= XSIMD_X86_AVX_VERSION
+#if XSIMD_X86_INSTR_SET >= XSIMD_X86_AVX_VERSION && XSIMD_X86_INSTR_SET < XSIMD_X86_AVX512_VERSION
     inline batch<double, 4> fexp_to_float(const batch<int64_t, 4>& x)
     {
         return _mm256_cvtepi32_pd(detail::xsimd_cvtepi64_epi32(x));
     }
 #endif
 
-#if XSIMD_X86_INSTR_SET >= XSIMD_X86_SSE2_VERSION
+#if XSIMD_X86_INSTR_SET >= XSIMD_X86_SSE2_VERSION && XSIMD_X86_INSTR_SET < XSIMD_X86_AVX512_VERSION
     inline batch<double, 2> fexp_to_float(const batch<int64_t, 2>& x)
     {
         return _mm_cvtepi32_pd(_mm_shuffle_epi32(x, _MM_SHUFFLE(0, 0, 2, 0)));

Original file line number	Diff line number	Diff line change
`@@ -77,14 +77,14 @@ namespace xsimd`
`77`	`77`	`}`
`78`	`78`
`79`	`79`	`// fexps are 8/11-bit numbers. so we can down-cast to int16_t/int32_t then do the job`
`80`		`-#if XSIMD_X86_INSTR_SET >= XSIMD_X86_AVX_VERSION`
	`80`	`+#if XSIMD_X86_INSTR_SET >= XSIMD_X86_AVX_VERSION && XSIMD_X86_INSTR_SET < XSIMD_X86_AVX512_VERSION`
`81`	`81`	`inline batch<double, 4> fexp_to_float(const batch<int64_t, 4>& x)`
`82`	`82`	`{`
`83`	`83`	`return _mm256_cvtepi32_pd(detail::xsimd_cvtepi64_epi32(x));`
`84`	`84`	`}`
`85`	`85`	`#endif`
`86`	`86`
`87`		`-#if XSIMD_X86_INSTR_SET >= XSIMD_X86_SSE2_VERSION`
	`87`	`+#if XSIMD_X86_INSTR_SET >= XSIMD_X86_SSE2_VERSION && XSIMD_X86_INSTR_SET < XSIMD_X86_AVX512_VERSION`
`88`	`88`	`inline batch<double, 2> fexp_to_float(const batch<int64_t, 2>& x)`
`89`	`89`	`{`
`90`	`90`	`return _mm_cvtepi32_pd(_mm_shuffle_epi32(x, _MM_SHUFFLE(0, 0, 2, 0)));`