Merge pull request #2299 from Xreki/support_scalar_kernels
Centralize the use of simd intrinsic and implement scalar kernels.gangliao-patch-1
Yiqun Liu
8 years ago
committed by
GitHub
commit
e36e24d49c
@ -0,0 +1,50 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#ifndef HL_CPU_SCALAR_CUH_
|
||||
#define HL_CPU_SCALAR_CUH_
|
||||
|
||||
#define VECTOR_SIMD false
|
||||
#define VECTOR_SET hl_vec_set
|
||||
|
||||
#ifndef PADDLE_TYPE_DOUBLE
|
||||
/* size of float */
|
||||
#define VECTOR_SIZE 4
|
||||
#else
|
||||
/* size of double */
|
||||
#define VECTOR_SIZE 8
|
||||
#endif
|
||||
|
||||
typedef real vecType;
|
||||
|
||||
/* Consider a real as a vector */
|
||||
#define VECTOR_LEN 1
|
||||
|
||||
template <class Agg>
|
||||
inline real hl_agg_op(Agg agg, vecType mm) {
|
||||
return mm;
|
||||
}
|
||||
|
||||
INLINE real hl_vec_set(const real r) {
|
||||
return r;
|
||||
}
|
||||
|
||||
INLINE real hl_vec_classification_error(const real a,
|
||||
const real b,
|
||||
const real p,
|
||||
const real r) {
|
||||
return ((a > p) == (b > p)) ? 0.0f : 1.0f;
|
||||
}
|
||||
|
||||
#endif // HL_CPU_SCALAR_CUH_
|
||||
@ -0,0 +1,73 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#ifndef HL_CPU_SIMD_NEON_CUH_
|
||||
#define HL_CPU_SIMD_NEON_CUH_
|
||||
|
||||
#include <arm_neon.h>
|
||||
|
||||
#define VECTOR_SIMD true
|
||||
#define VECTOR_SIZE 16
|
||||
#define VECTOR_SET hl_vec_set
|
||||
|
||||
#ifndef PADDLE_TYPE_DOUBLE
|
||||
|
||||
typedef float32x4_t vecType;
|
||||
|
||||
/* number of float in vector */
|
||||
#define VECTOR_LEN 4
|
||||
|
||||
template <class Agg>
|
||||
inline real hl_agg_op(Agg agg, vecType mm) {
|
||||
float32x4_t rev = vrev64q_f32(mm);
|
||||
float32x4_t tmp1 = agg.vecOp(rev, rev);
|
||||
float32x2_t lo = vget_high_f32(rev);
|
||||
float32x2_t hi = vget_low_f32(rev);
|
||||
float32x4_t tmp2 = vcombine_f32(hi, lo);
|
||||
float32x4_t ret = agg.vecOp(tmp1, tmp2);
|
||||
|
||||
return vgetq_lane_f32(ret, 0);
|
||||
}
|
||||
|
||||
inline float32x4_t hl_vec_set(const real f) {
|
||||
return vdupq_n_f32(f);
|
||||
}
|
||||
|
||||
inline float32x4_t hl_vec_classification_error(const float32x4_t a,
|
||||
const float32x4_t b,
|
||||
const float32x4_t p,
|
||||
const float32x4_t r) {
|
||||
uint32x4_t tmp1 = vcgtq_f32(a, p);
|
||||
uint32x4_t tmp2 = vcgtq_f32(b, p);
|
||||
uint32x4_t tmp3 = veorq_u32(tmp1, tmp2);
|
||||
return vcvtq_f32_u32(vandq_u32(tmp3, vcvtq_u32_f32(r)));
|
||||
}
|
||||
|
||||
#else
|
||||
|
||||
#ifdef __aarch64__
|
||||
typedef float64x2_t vecType;
|
||||
|
||||
/* number of float in vector */
|
||||
#define VECTOR_LEN 2
|
||||
#define VECTOR_SET vdupq_n_f64
|
||||
|
||||
#error To be implemented
|
||||
#else
|
||||
#error NEON instructions does not support double precision
|
||||
#endif // __aarch64__
|
||||
|
||||
#endif
|
||||
|
||||
#endif // HL_CPU_SIMD_NEON_CUH_
|
||||
@ -0,0 +1,94 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#ifndef HL_CPU_SIMD_SSE_CUH_
|
||||
#define HL_CPU_SIMD_SSE_CUH_
|
||||
|
||||
#include <mmintrin.h>
|
||||
#include <xmmintrin.h>
|
||||
#include <emmintrin.h>
|
||||
|
||||
#define VECTOR_SIMD true
|
||||
#define VECTOR_SIZE 16
|
||||
#define VECTOR_SET hl_vec_set
|
||||
|
||||
#ifndef PADDLE_TYPE_DOUBLE
|
||||
|
||||
typedef __m128 vecType;
|
||||
|
||||
/* number of float in vector */
|
||||
#define VECTOR_LEN 4
|
||||
|
||||
template <class Agg>
|
||||
inline real hl_agg_op(Agg agg, vecType mm) {
|
||||
__m128 lo = _mm_unpacklo_ps(mm, mm);
|
||||
__m128 hi = _mm_unpackhi_ps(mm, mm);
|
||||
__m128 tmp1 = agg.vecOp(lo, hi);
|
||||
__m128 tmp2 = _mm_movehl_ps(tmp1, tmp1);
|
||||
__m128 ret = agg.vecOp(tmp1, tmp2);
|
||||
|
||||
return _mm_cvtss_f32(ret);
|
||||
}
|
||||
|
||||
inline __m128 hl_vec_set(const real f) {
|
||||
return _mm_set_ps1(f);
|
||||
}
|
||||
|
||||
inline __m128 hl_vec_classification_error(const __m128 a,
|
||||
const __m128 b,
|
||||
const __m128 p,
|
||||
const __m128 r) {
|
||||
__m128 tmp1 = _mm_cmpgt_ps(a, p);
|
||||
__m128 tmp2 = _mm_cmpgt_ps(b, p);
|
||||
__m128 tmp3 = _mm_xor_ps(tmp1, tmp2);
|
||||
return _mm_and_ps(tmp3, r);
|
||||
}
|
||||
|
||||
#else
|
||||
|
||||
typedef __m128d vecType;
|
||||
|
||||
/* number of double in vector */
|
||||
#define VECTOR_LEN 2
|
||||
|
||||
template <class Agg>
|
||||
inline real hl_agg_op(Agg agg, vecType mm) {
|
||||
__m128d lo = _mm_unpacklo_pd(mm, mm);
|
||||
__m128d hi = _mm_unpackhi_pd(mm, mm);
|
||||
__m128d ret = agg.vecOp(lo, hi);
|
||||
|
||||
return _mm_cvtsd_f64(ret);
|
||||
}
|
||||
|
||||
inline __m128d hl_vec_set(const real d) {
|
||||
#if defined(__APPLE__) || defined(__OSX__)
|
||||
return _mm_set1_pd(d);
|
||||
#else
|
||||
return _mm_set_pd1(d);
|
||||
#endif
|
||||
}
|
||||
|
||||
inline __m128d hl_vec_classification_error(const __m128d a,
|
||||
const __m128d b,
|
||||
const __m128d p,
|
||||
const __m128d r) {
|
||||
__m128d tmp1 = _mm_cmpgt_pd(a, p);
|
||||
__m128d tmp2 = _mm_cmpgt_pd(b, p);
|
||||
__m128d tmp3 = _mm_xor_pd(tmp1, tmp2);
|
||||
return _mm_and_pd(tmp3, r);
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
#endif // HL_CPU_SIMD_SSE_CUH_
|
||||
@ -0,0 +1,153 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#ifndef HL_MATRIX_BASE_DETAIL_CUH_
|
||||
#define HL_MATRIX_BASE_DETAIL_CUH_
|
||||
|
||||
#include "hl_matrix_type.cuh"
|
||||
#include "hl_tensor_ops.h"
|
||||
|
||||
namespace aggregate {
|
||||
class SSESum {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::add<vecType>()(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMax {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::max<vecType>()(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMin {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::min<vecType>()(a, b);
|
||||
}
|
||||
};
|
||||
} // namespace aggregate
|
||||
|
||||
namespace base {
|
||||
namespace unary {
|
||||
class SSEIdentity {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a) const {
|
||||
return a;
|
||||
}
|
||||
};
|
||||
} // namespace unary
|
||||
|
||||
namespace binary {
|
||||
class SSEAdd {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::add<vecType>()(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEAdd2 {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
const real p1;
|
||||
const real p2;
|
||||
vecType mp1;
|
||||
vecType mp2;
|
||||
|
||||
public:
|
||||
SSEAdd2(const real s1, const real s2) : p1(s1), p2(s2) {
|
||||
mp1 = hl_vec_set(p1);
|
||||
mp2 = hl_vec_set(p2);
|
||||
}
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::add_scale<vecType>(mp1, mp2)(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSESub {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::sub<vecType>()(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMul {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::mul<vecType>()(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEDiv {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hppl::binary::div<vecType>()(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSESquaredDiff {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
vecType tmp = hppl::binary::sub<vecType>()(a, b);
|
||||
return hppl::binary::mul<vecType>()(tmp, tmp);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEFirst {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return a;
|
||||
}
|
||||
};
|
||||
|
||||
class SSESecond {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return b;
|
||||
}
|
||||
};
|
||||
|
||||
class SSEClassificationError {
|
||||
public:
|
||||
static const bool sse = VECTOR_SIMD;
|
||||
const real p;
|
||||
vecType mp;
|
||||
vecType result;
|
||||
|
||||
public:
|
||||
explicit SSEClassificationError(const real s) : p(s) {
|
||||
mp = hl_vec_set(p);
|
||||
result = hl_vec_set(1.0f);
|
||||
}
|
||||
INLINE vecType vecOp(const vecType a, const vecType b) const {
|
||||
return hl_vec_classification_error(a, b, mp, result);
|
||||
}
|
||||
};
|
||||
} // namespace binary
|
||||
} // namespace base
|
||||
|
||||
#endif /* HL_MATRIX_BASE_DETAIL_CUH_ */
|
||||
@ -1,161 +0,0 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
|
||||
#ifndef HL_MATRIX_BASE_NEON_CUH_
|
||||
#define HL_MATRIX_BASE_NEON_CUH_
|
||||
|
||||
namespace aggregate {
|
||||
class SSESum {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return vaddq_f32(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMax {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return vmaxq_f32(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMin {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return vminq_f32(a, b);
|
||||
}
|
||||
};
|
||||
} // namespace aggregate
|
||||
|
||||
namespace base {
|
||||
namespace unary {
|
||||
class SSEIdentity {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a) const {
|
||||
return a;
|
||||
}
|
||||
};
|
||||
} // namespace unary
|
||||
|
||||
namespace binary {
|
||||
class SSEAdd {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return vaddq_f32(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEAdd2 {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
const real p1;
|
||||
const real p2;
|
||||
float32x4_t mp1;
|
||||
float32x4_t mp2;
|
||||
|
||||
public:
|
||||
SSEAdd2(const real s1, const real s2) : p1(s1), p2(s2) {
|
||||
mp1 = vdupq_n_f32(p1);
|
||||
mp2 = vdupq_n_f32(p2);
|
||||
}
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
float32x4_t tmp1, tmp2;
|
||||
tmp1 = vmulq_f32(mp1, a);
|
||||
tmp2 = vmulq_f32(mp2, b);
|
||||
return vaddq_f32(tmp1, tmp2);
|
||||
}
|
||||
};
|
||||
|
||||
class SSESub {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return vsubq_f32(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMul {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return vmulq_f32(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEDiv {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
float32x4_t tmp;
|
||||
tmp = vrecpeq_f32(b);
|
||||
return vmulq_f32(a, tmp);
|
||||
}
|
||||
};
|
||||
|
||||
class SSESquaredDiff {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
float32x4_t tmp;
|
||||
tmp = vsubq_f32(a, b);
|
||||
return vmulq_f32(tmp, tmp);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEFirst {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return a;
|
||||
}
|
||||
};
|
||||
|
||||
class SSESecond {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
return b;
|
||||
}
|
||||
};
|
||||
|
||||
class SSEClassificationError {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
const real p;
|
||||
float32x4_t mp;
|
||||
uint32x4_t result;
|
||||
|
||||
public:
|
||||
explicit SSEClassificationError(const real s) : p(s) {
|
||||
mp = vdupq_n_f32(p);
|
||||
result = vdupq_n_u32(1);
|
||||
}
|
||||
// TODO: to be check
|
||||
INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
|
||||
uint32x4_t tmp1 = vcgtq_f32(a, mp);
|
||||
uint32x4_t tmp2 = vcgtq_f32(b, mp);
|
||||
uint32x4_t tmp3 = veorq_u32(tmp1, tmp2);
|
||||
return vcvtq_f32_u32(vandq_u32(tmp3, result));
|
||||
}
|
||||
};
|
||||
} // namespace binary
|
||||
} // namespace base
|
||||
|
||||
#endif /* HL_MATRIX_BASE_NEON_CUH_ */
|
||||
@ -1,211 +0,0 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
|
||||
#ifndef HL_MATRIX_BASE_SSE_CUH_
|
||||
#define HL_MATRIX_BASE_SSE_CUH_
|
||||
|
||||
namespace aggregate {
|
||||
class SSESum {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_add_ps(a, b);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_add_pd(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMax {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_max_ps(a, b);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_max_pd(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMin {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_min_ps(a, b);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_min_pd(a, b);
|
||||
}
|
||||
};
|
||||
} // namespace aggregate
|
||||
|
||||
namespace base {
|
||||
namespace unary {
|
||||
class SSEIdentity {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a) const {
|
||||
return a;
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a) const {
|
||||
return a;
|
||||
}
|
||||
};
|
||||
} // namespace unary
|
||||
|
||||
namespace binary {
|
||||
class SSEAdd {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_add_ps(a, b);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_add_pd(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEAdd2 {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
const real p1;
|
||||
const real p2;
|
||||
union {__m128 f; __m128d d;} mp1;
|
||||
union {__m128 f; __m128d d;} mp2;
|
||||
|
||||
public:
|
||||
SSEAdd2(const real s1, const real s2) : p1(s1), p2(s2) {
|
||||
if (sizeof(real) == sizeof(float)) {
|
||||
mp1.f = _mm_set1_ps(p1);
|
||||
mp2.f = _mm_set1_ps(p2);
|
||||
} else {
|
||||
mp1.d = _mm_set1_pd(p1);
|
||||
mp2.d = _mm_set1_pd(p2);
|
||||
}
|
||||
}
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
__m128 tmp1, tmp2;
|
||||
tmp1 = _mm_mul_ps(mp1.f, a);
|
||||
tmp2 = _mm_mul_ps(mp2.f, b);
|
||||
return _mm_add_ps(tmp1, tmp2);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
__m128d tmp1, tmp2;
|
||||
tmp1 = _mm_mul_pd(mp1.d, a);
|
||||
tmp2 = _mm_mul_pd(mp2.d, b);
|
||||
return _mm_add_pd(tmp1, tmp2);
|
||||
}
|
||||
};
|
||||
|
||||
class SSESub {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_sub_ps(a, b);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_sub_pd(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEMul {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_mul_ps(a, b);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_mul_pd(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSEDiv {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_div_ps(a, b);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_div_pd(a, b);
|
||||
}
|
||||
};
|
||||
|
||||
class SSESquaredDiff {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return _mm_mul_ps(_mm_sub_ps(a, b), _mm_sub_ps(a, b));
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return _mm_mul_pd(_mm_sub_pd(a, b), _mm_sub_pd(a, b));
|
||||
}
|
||||
};
|
||||
|
||||
class SSEFirst {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return a;
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return a;
|
||||
}
|
||||
};
|
||||
|
||||
class SSESecond {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
return b;
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
return b;
|
||||
}
|
||||
};
|
||||
|
||||
class SSEClassificationError {
|
||||
public:
|
||||
static const bool sse = true;
|
||||
const real p;
|
||||
union {__m128 f; __m128d d;} mp;
|
||||
union {__m128 f; __m128d d;} result;
|
||||
|
||||
public:
|
||||
explicit SSEClassificationError(const real s) : p(s) {
|
||||
if (sizeof(real) == sizeof(float)) {
|
||||
mp.f = _mm_set1_ps(p);
|
||||
result.f = _mm_set1_ps(1.0f);
|
||||
} else {
|
||||
mp.d = _mm_set1_pd(p);
|
||||
result.d = _mm_set1_pd(1.0);
|
||||
}
|
||||
}
|
||||
INLINE __m128 vecOp(const __m128 a, const __m128 b) const {
|
||||
__m128 tmp1 = _mm_cmpgt_ps(a, mp.f);
|
||||
__m128 tmp2 = _mm_cmpgt_ps(b, mp.f);
|
||||
__m128 tmp3 = _mm_xor_ps(tmp1, tmp2);
|
||||
return _mm_and_ps(tmp3, result.f);
|
||||
}
|
||||
INLINE __m128d vecOp(const __m128d a, const __m128d b) const {
|
||||
__m128d tmp1 = _mm_cmpgt_pd(a, mp.d);
|
||||
__m128d tmp2 = _mm_cmpgt_pd(b, mp.d);
|
||||
__m128d tmp3 = _mm_xor_pd(tmp1, tmp2);
|
||||
return _mm_and_pd(tmp3, result.d);
|
||||
}
|
||||
};
|
||||
} // namespace binary
|
||||
} // namespace base
|
||||
|
||||
#endif /* HL_MATRIX_BASE_SSE_CUH_ */
|
||||
File diff suppressed because it is too large
Load Diff
Loading…
Reference in new issue