ccc7c9b86d
* py-horovod: fix compilation of ~cuda * Rewrite py-horovod with only 3 variants * Add upstream patch to workaround compilation issue
53 lines
2.1 KiB
Diff
53 lines
2.1 KiB
Diff
From 717e72f91f02d1dc3c859719ef1d804b10f88017 Mon Sep 17 00:00:00 2001
|
|
From: Nicolas V Castet <nvcastet@us.ibm.com>
|
|
Date: Mon, 30 Mar 2020 12:47:50 -0500
|
|
Subject: [PATCH] Add extra preprocessor guard for FMA optimization
|
|
|
|
Fixes #1832
|
|
|
|
Signed-off-by: Nicolas V Castet <nvcastet@us.ibm.com>
|
|
---
|
|
horovod/common/ops/adasum/adasum.h | 8 ++++----
|
|
1 file changed, 4 insertions(+), 4 deletions(-)
|
|
|
|
diff --git a/horovod/common/ops/adasum/adasum.h b/horovod/common/ops/adasum/adasum.h
|
|
index 0330f5850..876f7f12b 100644
|
|
--- a/horovod/common/ops/adasum/adasum.h
|
|
+++ b/horovod/common/ops/adasum/adasum.h
|
|
@@ -19,7 +19,7 @@
|
|
#include <cstring>
|
|
#include <float.h>
|
|
|
|
-#if __AVX__ && __F16C__
|
|
+#if __AVX__ && __F16C__ && __FMA__
|
|
#include <emmintrin.h>
|
|
#include <immintrin.h>
|
|
#endif
|
|
@@ -104,7 +104,7 @@ template <typename Communicator_type> class Adasum {
|
|
int count, double& dotProduct,
|
|
double& anormsq, double& bnormsq,
|
|
int layerid) {
|
|
-#if __AVX__ && __F16C__
|
|
+#if __AVX__ && __F16C__ && __FMA__
|
|
if (horovod_datatype == DataType::HOROVOD_FLOAT16) {
|
|
ComputeDotAndNormSqrdsfp16((uint16_t*)a, (uint16_t*)b, count, dotProduct,
|
|
anormsq, bnormsq, layerid);
|
|
@@ -125,7 +125,7 @@ template <typename Communicator_type> class Adasum {
|
|
double acoeff, void* __restrict__ a,
|
|
double bcoeff, void* __restrict__ b,
|
|
int layerid) {
|
|
-#if __AVX__ && __F16C__
|
|
+#if __AVX__ && __F16C__ && __FMA__
|
|
if (horovod_datatype == DataType::HOROVOD_FLOAT16) {
|
|
ScaledAddfp16(count, acoeff, (uint16_t*)a, bcoeff, (uint16_t*)b, layerid);
|
|
} else
|
|
@@ -425,7 +425,7 @@ template <typename Communicator_type> class Adasum {
|
|
}
|
|
|
|
|
|
-#if __AVX__ && __F16C__
|
|
+#if __AVX__ && __F16C__ && __FMA__
|
|
inline void ComputeDotAndNormSqrdsfp16(const uint16_t* __restrict__ a,
|
|
const uint16_t* __restrict__ b,
|
|
int len, double& dotProduct,
|