Merge pull request #198 from sterrettm2/gcc_8

Raghuveer Devulapalli · web-flow · commit eafc913dbe74 · 2025-04-21T20:38:59.000-07:00
Avoid use of _mm512_set_epi16 for reversing 16-bit vectors
diff --git a/src/avx512-16bit-qsort.hpp b/src/avx512-16bit-qsort.hpp
@@ -175,7 +175,8 @@ struct zmm_vector<float16> {
     }
     static reg_t reverse(reg_t zmm)
     {
-        const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);
+        constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};
+        const auto rev_index = _mm512_loadu_si512(arr);
         return permutexvar(rev_index, zmm);
     }
     static reg_t sort_vec(reg_t x)
@@ -320,7 +321,8 @@ struct zmm_vector<int16_t> {
     }
     static reg_t reverse(reg_t zmm)
     {
-        const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);
+        constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};
+        const auto rev_index = _mm512_loadu_si512(arr);
         return permutexvar(rev_index, zmm);
     }
     static reg_t sort_vec(reg_t x)
@@ -462,7 +464,8 @@ struct zmm_vector<uint16_t> {
     }
     static reg_t reverse(reg_t zmm)
     {
-        const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);
+        constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};
+        const auto rev_index = _mm512_loadu_si512(arr);
         return permutexvar(rev_index, zmm);
     }
     static reg_t sort_vec(reg_t x)
diff --git a/src/avx512fp16-16bit-qsort.hpp b/src/avx512fp16-16bit-qsort.hpp
@@ -139,7 +139,8 @@ struct zmm_vector<_Float16> {
     }
     static reg_t reverse(reg_t zmm)
     {
-        const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);
+        constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};
+        const auto rev_index = _mm512_loadu_si512(arr);
         return permutexvar(rev_index, zmm);
     }
     static reg_t sort_vec(reg_t x)
diff --git a/src/xss-common-includes.h b/src/xss-common-includes.h
@@ -79,8 +79,8 @@
 #define NETWORK_REVERSE_16LANES \
     0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
 #define NETWORK_REVERSE_32LANES \
-    0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, \
-            21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+    31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, \
+            13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
 
 #if defined(XSS_USE_OPENMP) && defined(_OPENMP)
 #define XSS_COMPILE_OPENMP

Original file line number	Diff line number	Diff line change
`@@ -175,7 +175,8 @@ struct zmm_vector<float16> {`
`175`	`175`	`}`
`176`	`176`	`static reg_t reverse(reg_t zmm)`
`177`	`177`	`{`
`178`		`- const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);`
	`178`	`+ constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};`
	`179`	`+ const auto rev_index = _mm512_loadu_si512(arr);`
`179`	`180`	`return permutexvar(rev_index, zmm);`
`180`	`181`	`}`
`181`	`182`	`static reg_t sort_vec(reg_t x)`
`@@ -320,7 +321,8 @@ struct zmm_vector<int16_t> {`
`320`	`321`	`}`
`321`	`322`	`static reg_t reverse(reg_t zmm)`
`322`	`323`	`{`
`323`		`- const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);`
	`324`	`+ constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};`
	`325`	`+ const auto rev_index = _mm512_loadu_si512(arr);`
`324`	`326`	`return permutexvar(rev_index, zmm);`
`325`	`327`	`}`
`326`	`328`	`static reg_t sort_vec(reg_t x)`
`@@ -462,7 +464,8 @@ struct zmm_vector<uint16_t> {`
`462`	`464`	`}`
`463`	`465`	`static reg_t reverse(reg_t zmm)`
`464`	`466`	`{`
`465`		`- const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);`
	`467`	`+ constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};`
	`468`	`+ const auto rev_index = _mm512_loadu_si512(arr);`
`466`	`469`	`return permutexvar(rev_index, zmm);`
`467`	`470`	`}`
`468`	`471`	`static reg_t sort_vec(reg_t x)`
Original file line number	Diff line number	Diff line change
`@@ -139,7 +139,8 @@ struct zmm_vector<_Float16> {`
`139`	`139`	`}`
`140`	`140`	`static reg_t reverse(reg_t zmm)`
`141`	`141`	`{`
`142`		`- const auto rev_index = _mm512_set_epi16(NETWORK_REVERSE_32LANES);`
	`142`	`+ constexpr static uint16_t arr[] = {NETWORK_REVERSE_32LANES};`
	`143`	`+ const auto rev_index = _mm512_loadu_si512(arr);`
`143`	`144`	`return permutexvar(rev_index, zmm);`
`144`	`145`	`}`
`145`	`146`	`static reg_t sort_vec(reg_t x)`