Add Zen5 tunings

animetosho · Aug 6, 2024 · 54fd976 · 54fd976
1 parent dc89435
commit 54fd976
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 5 deletions.
diff --git a/src/encoder_avx_base.h b/src/encoder_avx_base.h
@@ -217,12 +217,12 @@ HEDLEY_ALWAYS_INLINE void do_encode_avx2(int line_size, int* colOffset, const ui
 				// duplicate halves
 				data1A = _mm256_inserti128_si256(dataA, _mm256_castsi256_si128(dataA), 1);
 				data1B = _mm256_inserti128_si256(dataB, _mm256_castsi256_si128(dataB), 1);
-#if defined(__tune_znver2__) || defined(__tune_znver3__) || defined(__tune_znver4__)
-				data2A = _mm256_permute2x128_si256(dataA, dataA, 0x11);
-				data2B = _mm256_permute2x128_si256(dataB, dataB, 0x11);
-#else
+#if defined(__tune_znver1__) || defined(__tune_bdver4__)
 				data2A = _mm256_permute4x64_epi64(dataA, 0xee);
 				data2B = _mm256_permute4x64_epi64(dataB, 0xee);
+#else
+				data2A = _mm256_permute2x128_si256(dataA, dataA, 0x11);
+				data2B = _mm256_permute2x128_si256(dataB, dataB, 0x11);
 #endif
 
 				shuf1A = _mm256_load_si256(lookupsAVX2->shufExpand + m1);

diff --git a/src/encoder_sse_base.h b/src/encoder_sse_base.h
@@ -351,7 +351,7 @@ HEDLEY_ALWAYS_INLINE void do_encode_sse(int line_size, int* colOffset, const uin
 #if defined(__POPCNT__) && !defined(__tune_btver1__)
 				if(use_isa & ISA_FEATURE_POPCNT) {
 					shuf2Len = popcnt32(maskA) + 16;
-# if defined(__tune_znver4__) || defined(__tune_znver3__) || defined(__tune_znver2__) || defined(__tune_znver1__) || defined(__tune_btver2__)
+# if defined(__tune_znver6__) || defined(__tune_znver5__) || defined(__tune_znver4__) || defined(__tune_znver3__) || defined(__tune_znver2__) || defined(__tune_znver1__) || defined(__tune_btver2__)
 					shuf1Len = popcnt32(m1) + 8;
 					shuf3Len = popcnt32(m3) + shuf2Len + 8;
 # else