diff options
author | Henning Baldersheim <balder@yahoo-inc.com> | 2019-04-01 07:38:39 +0000 |
---|---|---|
committer | Henning Baldersheim <balder@yahoo-inc.com> | 2019-04-01 07:38:39 +0000 |
commit | ca2e83ed3f8eef6e69c6110ee32d1c59d5d19c2a (patch) | |
tree | 2063bd3e6db381bdc0f3b159f6e6bbc4bb5e83d3 /vespalib | |
parent | aaf479240531f08548bb236ade6f99b272481051 (diff) |
Revert unintended change.
Diffstat (limited to 'vespalib')
-rw-r--r-- | vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp | 5 | ||||
-rw-r--r-- | vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp | 12 |
2 files changed, 0 insertions, 17 deletions
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp index 4295f9850c5..1bf7ea1c44c 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp +++ b/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp @@ -10,11 +10,6 @@ template <typename ACCUM, typename T, size_t UNROLL> ACCUM multiplyAdd(const T * a, const T * b, size_t sz) { -#if 1 - for (int i(0); i < 16; i++) { - __builtin_prefetch(&b[(4+i)*(64/sizeof(T))], 0, 0); - } -#endif ACCUM partial[UNROLL]; for (size_t i(0); i < UNROLL; i++) { partial[i] = 0; diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp index 209de4666a2..f135de52e5a 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp +++ b/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp @@ -16,20 +16,12 @@ bool validAlignment16(const void * a, const void * b) { } -#define PREFETCH_DISTANCE 16 -#define L1_DIST 4 float Sse2Accelrator::dotProduct(const float * af, const float * bf, size_t sz) const { if ( ! validAlignment16(af, bf)) { return GenericAccelrator::dotProduct(af, bf, sz); } - -#if PREFETCH_DISTANCE > 0 - for (int i(0); i < PREFETCH_DISTANCE; i++) { - __builtin_prefetch(&bf[i*16], 0, 0); - } -#endif typedef float v4sf __attribute__ ((vector_size (16))); const size_t ChunkSize(16); const size_t VectorsPerChunk(ChunkSize/4); @@ -39,10 +31,6 @@ Sse2Accelrator::dotProduct(const float * af, const float * bf, size_t sz) const const size_t numChunks(sz/ChunkSize); for (size_t i(0); i < numChunks; i++) { -#if PREFETCH_DISTANCE > 0 - __builtin_prefetch(&bf[(i+PREFETCH_DISTANCE)*16], 0, 0); - __builtin_prefetch(&af[((i+L1_DIST)&0xf)*16], 0, 3); -#endif for (size_t j(0); j < VectorsPerChunk; j++) { partial[j] += a[VectorsPerChunk*i+j] * b[VectorsPerChunk*i+j]; } |