Use _mm256_loadu_si256 to make it work on g++-10

Raghuveer Devulapalli · Raghuveer Devulapalli · commit 353dc3c7ff23 · 2023-05-02T14:37:14.000-07:00
diff --git a/Makefile b/Makefile
@@ -1,4 +1,4 @@
-CXX		= g++-12
+CXX		?= g++-12
 SRCDIR		= ./src
 TESTDIR		= ./tests
 BENCHDIR	= ./benchmarks
diff --git a/src/avx512-64bit-common.h b/src/avx512-64bit-common.h
@@ -219,7 +219,7 @@ struct ymm_vector<uint32_t> {
     }
     static zmm_t loadu(void const *mem)
     {
-        return _mm256_loadu_epi32(mem);
+        return _mm256_loadu_si256((__m256i*) mem);
     }
     static zmm_t max(zmm_t x, zmm_t y)
     {
@@ -348,7 +348,7 @@ struct ymm_vector<int32_t> {
     }
     static zmm_t loadu(void const *mem)
     {
-        return _mm256_loadu_epi32(mem);
+        return _mm256_loadu_si256((__m256i*) mem);
     }
     static zmm_t max(zmm_t x, zmm_t y)
     {

Original file line number	Diff line number	Diff line change
`@@ -219,7 +219,7 @@ struct ymm_vector<uint32_t> {`
`219`	`219`	`}`
`220`	`220`	`static zmm_t loadu(void const *mem)`
`221`	`221`	`{`
`222`		`- return _mm256_loadu_epi32(mem);`
	`222`	`+ return _mm256_loadu_si256((__m256i*) mem);`
`223`	`223`	`}`
`224`	`224`	`static zmm_t max(zmm_t x, zmm_t y)`
`225`	`225`	`{`
`@@ -348,7 +348,7 @@ struct ymm_vector<int32_t> {`
`348`	`348`	`}`
`349`	`349`	`static zmm_t loadu(void const *mem)`
`350`	`350`	`{`
`351`		`- return _mm256_loadu_epi32(mem);`
	`351`	`+ return _mm256_loadu_si256((__m256i*) mem);`
`352`	`352`	`}`
`353`	`353`	`static zmm_t max(zmm_t x, zmm_t y)`
`354`	`354`	`{`