Go down to z14 to support older distros

Signed-off-by: Andreas Krebbel <[email protected]>
IBM · Nov 22, 2024 · 235f0d7 · 235f0d7
1 parent e976de7
commit 235f0d7
Show file tree

Hide file tree

Showing 2 changed files with 32 additions and 7 deletions.
diff --git a/config.zdnn b/config.zdnn
@@ -29,13 +29,13 @@ case "${target}" in
 	CFLAGS_INIT="-O3 -mzvector -Wall -std=gnu99 -fstack-protector-all ${CFLAGS_INIT:-}"
 	CFLAGS_QUOTE_INIT="-Wall" # Not needed on Linux. Just repeat an option to prevent it from being empty.
 	CFLAGS_OPT_EXPENSIVE="-funroll-loops"
-	CFLAGS="-O3 -march=z16 -mzvector -Wall -std=gnu99 -fstack-protector-all ${CFLAGS_OPT_EXPENSIVE} ${CFLAGS:-}"
+	CFLAGS="-O3 -march=z14 -mzvector -Wall -std=gnu99 -fstack-protector-all ${CFLAGS_OPT_EXPENSIVE} ${CFLAGS:-}"
 	CFLAGS_QUOTE="-Wall"
 	CFLAGS_DEBUG="-O0 -g3 ${CFLAGS_DEBUG:-}"
 	CFLAGS_SHARED="-fPIC ${CFLAGS_SHARED:-}"
 	CFLAGS_ASM="-Wa,-adhln -fno-asynchronous-unwind-tables ${CFLAGS_ASM:-}"
 	CFLAGS_NOSEARCH=""
-	CXXFLAGS="-O3 -march=z16 -Wall ${CXXFLAGS:-}"
+	CXXFLAGS="-O3 -march=z14 -Wall ${CXXFLAGS:-}"
 	CPP_SYMCHECK_FLAGS="-E -o zdnn.i"
 	SODIR="${SODIR:-lib}"
 	LIBNAME="${LIBNAME:-libzdnn}"

diff --git a/zdnn/zdnn_private.h b/zdnn/zdnn_private.h
@@ -589,13 +589,12 @@ void aiu_vec_lengthen_to_fp32(vec_int16 a, vec_float32 *out1,
 vec_int16 aiu_vec_convert_from_fp16(vec_int16 a);
 vec_int16 aiu_vec_convert_to_fp16(vec_int16 a);
 
-#ifdef __MVS__
+#if defined(__MVS__) || (defined(__ARCH__) && __ARCH__ < 14)
 #define VEC_ROUND_FROM_FP32(FP_HI, FP_LO)                                      \
-  aiu_vec_round_from_fp32(*(vec_float32 *)((void *)&(FP_HI)),                  \
-                          *(vec_float32 *)((void *)&(FP_LO)));
+  aiu_vec_round_from_fp32((vec_float32)(FP_HI), (vec_float32)(FP_LO));
 #define VEC_LENGTHEN_TO_FP32(IN, OUT_HI, OUT_LO)                               \
-  aiu_vec_lengthen_to_fp32((IN), (vec_float32 *)((void *)&(OUT_HI)),           \
-                           (vec_float32 *)((void *)&(OUT_LO)));
+  aiu_vec_lengthen_to_fp32((IN), (vec_float32 *)&(OUT_HI),                     \
+                           (vec_float32 *)&(OUT_LO));
 #else
 #define VEC_ROUND_FROM_FP32(FP_HI, FP_LO)                                      \
   (vec_int16) vec_round_from_fp32((FP_HI), (FP_LO), 0);
@@ -1542,6 +1541,32 @@ void dumpdata_ztensor(const zdnn_ztensor *ztensor, dump_mode mode,
 #define PADDED(x)                                                              \
   ((uint32_t)CEIL(x, AIU_2BYTE_CELLS_PER_STICK) * AIU_2BYTE_CELLS_PER_STICK)
 
+#if !defined(vec_float) || __ARCH__ < 13
+#undef vec_float
+#define vec_float(X)                                                           \
+  ({                                                                           \
+    __vector float out;                                                        \
+    /* vcefb\t%[out],%[in],0,0 */                                              \
+    __asm__(".insn vrr,0xe700000020c3,%[out],%[in],0,2,0,0"                    \
+            : [out] "=v"(out)                                                  \
+            : [in] "v"(X));                                                    \
+    out;                                                                       \
+  })
+#endif
+
+#if defined(__GNUC__) && __GNUC__ <= 7
+#undef vec_round
+#define vec_round(X)                                                           \
+  ({                                                                           \
+    __vector float out;                                                        \
+    /* vfisb %[out],%[in],4,4 */                                               \
+    __asm__(".insn vrr,0xe700000020c7,%[out],%[in],0,2,4,4"                    \
+            : [out] "=v"(out)                                                  \
+            : [in] "v"(X));                                                    \
+    out;                                                                       \
+  })
+#endif
+
 // -----------------------------------------------------------------------------
 // Private global variables
 // -----------------------------------------------------------------------------