Skip to content

GitLab

Commit 454caac9 authored by peastman's avatar peastman
Browse files

Preliminary support for ARM/Android.

parent 51828eaa
Show whitespace changes
Inline Side-by-side
Showing with 722 additions and 316 deletions
openmmapi/include/openmm/internal/hardware.h
View file @ 454caac9
......@@ -46,10 +46,14 @@
#ifdef WIN32
#define NOMINMAX
#include <windows.h>
#else
#ifdef __ANDROID__
#include <cpu-features.h>
#else
#include <dlfcn.h>
#include <unistd.h>
#endif
#endif
#endif
static int getNumProcessors() {
......@@ -70,11 +74,15 @@ static int getNumProcessors() {
ncpu = 1;
return ncpu;
#else
#ifdef __ANDROID__
return android_getCpuCount();
#else
long nProcessorsOnline = sysconf(_SC_NPROCESSORS_ONLN);
if (nProcessorsOnline == -1)
return 1;
else
return (int) nProcessorsOnline;
#endif
#endif
#endif
}
......@@ -85,8 +93,9 @@ static int getNumProcessors() {
#ifdef _WIN32
#define cpuid __cpuid
#else
static void cpuid(int cpuInfo[4], int infoType){
#ifdef __LP64__
#ifndef __ANDROID__
static void cpuid(int cpuInfo[4], int infoType){
#ifdef __LP64__
__asm__ __volatile__ (
"cpuid":
"=a" (cpuInfo[0]),
......@@ -95,7 +104,7 @@ static void cpuid(int cpuInfo[4], int infoType){
"=d" (cpuInfo[3]) :
"a" (infoType)
);
#else
#else
__asm__ __volatile__ (
"pushl %%ebx\n"
"cpuid\n"
......@@ -107,8 +116,9 @@ static void cpuid(int cpuInfo[4], int infoType){
"=d" (cpuInfo[3]) :
"a" (infoType)
);
#endif
}
#endif
}
#endif
#endif
#endif // OPENMM_HARDWARE_H_
openmmapi/include/openmm/internal/vectorize.h
View file @ 454caac9
......@@ -9,7 +9,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2013 Stanford University and the Authors. *
* Portions copyright (c) 2014 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -32,255 +32,10 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
#include <smmintrin.h>
// This file defines classes and functions to simplify vectorizing code with SSE.
class ivec4;
/**
* A four element vector of floats.
*/
class fvec4 {
public:
__m128 val;
fvec4() {}
fvec4(float v) : val(_mm_set1_ps(v)) {}
fvec4(float v1, float v2, float v3, float v4) : val(_mm_set_ps(v4, v3, v2, v1)) {}
fvec4(__m128 v) : val(v) {}
fvec4(const float* v) : val(_mm_loadu_ps(v)) {}
operator __m128() const {
return val;
}
float operator[](int i) const {
float result[4];
store(result);
return result[i];
}
void store(float* v) const {
_mm_storeu_ps(v, val);
}
fvec4 operator+(const fvec4& other) const {
return _mm_add_ps(val, other);
}
fvec4 operator-(const fvec4& other) const {
return _mm_sub_ps(val, other);
}
fvec4 operator*(const fvec4& other) const {
return _mm_mul_ps(val, other);
}
fvec4 operator/(const fvec4& other) const {
return _mm_div_ps(val, other);
}
void operator+=(const fvec4& other) {
val = _mm_add_ps(val, other);
}
void operator-=(const fvec4& other) {
val = _mm_sub_ps(val, other);
}
void operator*=(const fvec4& other) {
val = _mm_mul_ps(val, other);
}
void operator/=(const fvec4& other) {
val = _mm_div_ps(val, other);
}
fvec4 operator-() const {
return _mm_sub_ps(_mm_set1_ps(0.0f), val);
}
fvec4 operator&(const fvec4& other) const {
return _mm_and_ps(val, other);
}
fvec4 operator|(const fvec4& other) const {
return _mm_or_ps(val, other);
}
fvec4 operator==(const fvec4& other) const {
return _mm_cmpeq_ps(val, other);
}
fvec4 operator!=(const fvec4& other) const {
return _mm_cmpneq_ps(val, other);
}
fvec4 operator>(const fvec4& other) const {
return _mm_cmpgt_ps(val, other);
}
fvec4 operator<(const fvec4& other) const {
return _mm_cmplt_ps(val, other);
}
fvec4 operator>=(const fvec4& other) const {
return _mm_cmpge_ps(val, other);
}
fvec4 operator<=(const fvec4& other) const {
return _mm_cmple_ps(val, other);
}
operator ivec4() const;
};
/**
* A four element vector of ints.
*/
class ivec4 {
public:
__m128i val;
ivec4() {}
ivec4(int v) : val(_mm_set1_epi32(v)) {}
ivec4(int v1, int v2, int v3, int v4) : val(_mm_set_epi32(v4, v3, v2, v1)) {}
ivec4(__m128i v) : val(v) {}
ivec4(const int* v) : val(_mm_loadu_si128((const __m128i*) v)) {}
operator __m128i() const {
return val;
}
int operator[](int i) const {
int result[4];
store(result);
return result[i];
}
void store(int* v) const {
_mm_storeu_si128((__m128i*) v, val);
}
ivec4 operator+(const ivec4& other) const {
return _mm_add_epi32(val, other);
}
ivec4 operator-(const ivec4& other) const {
return _mm_sub_epi32(val, other);
}
ivec4 operator*(const ivec4& other) const {
return _mm_mul_epi32(val, other);
}
void operator+=(const ivec4& other) {
val = _mm_add_epi32(val, other);
}
void operator-=(const ivec4& other) {
val = _mm_sub_epi32(val, other);
}
void operator*=(const ivec4& other) {
val = _mm_mul_epi32(val, other);
}
ivec4 operator-() const {
return _mm_sub_epi32(_mm_set1_epi32(0), val);
}
ivec4 operator&(const ivec4& other) const {
return _mm_and_si128(val, other);
}
ivec4 operator|(const ivec4& other) const {
return _mm_or_si128(val, other);
}
ivec4 operator==(const ivec4& other) const {
return _mm_cmpeq_epi32(val, other);
}
ivec4 operator!=(const ivec4& other) const {
return _mm_xor_si128(*this==other, _mm_set1_epi32(0xFFFFFFFF));
}
ivec4 operator>(const ivec4& other) const {
return _mm_cmpgt_epi32(val, other);
}
ivec4 operator<(const ivec4& other) const {
return _mm_cmplt_epi32(val, other);
}
ivec4 operator>=(const ivec4& other) const {
return _mm_xor_si128(_mm_cmplt_epi32(val, other), _mm_set1_epi32(0xFFFFFFFF));
}
ivec4 operator<=(const ivec4& other) const {
return _mm_xor_si128(_mm_cmpgt_epi32(val, other), _mm_set1_epi32(0xFFFFFFFF));
}
operator fvec4() const;
};
// Conversion operators.
inline fvec4::operator ivec4() const {
return _mm_cvttps_epi32(val);
}
inline ivec4::operator fvec4() const {
return _mm_cvtepi32_ps(val);
}
// Functions that operate on fvec4s.
static inline fvec4 floor(const fvec4& v) {
return fvec4(_mm_floor_ps(v.val));
}
static inline fvec4 ceil(const fvec4& v) {
return fvec4(_mm_ceil_ps(v.val));
}
static inline fvec4 round(const fvec4& v) {
return fvec4(_mm_round_ps(v.val, _MM_FROUND_TO_NEAREST_INT));
}
static inline fvec4 min(const fvec4& v1, const fvec4& v2) {
return fvec4(_mm_min_ps(v1.val, v2.val));
}
static inline fvec4 max(const fvec4& v1, const fvec4& v2) {
return fvec4(_mm_max_ps(v1.val, v2.val));
}
static inline fvec4 abs(const fvec4& v) {
static const __m128 mask = _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF));
return fvec4(_mm_and_ps(v.val, mask));
}
static inline fvec4 sqrt(const fvec4& v) {
return fvec4(_mm_sqrt_ps(v.val));
}
static inline float dot3(const fvec4& v1, const fvec4& v2) {
return _mm_cvtss_f32(_mm_dp_ps(v1, v2, 0x71));
}
static inline float dot4(const fvec4& v1, const fvec4& v2) {
return _mm_cvtss_f32(_mm_dp_ps(v1, v2, 0xF1));
}
static inline void transpose(fvec4& v1, fvec4& v2, fvec4& v3, fvec4& v4) {
_MM_TRANSPOSE4_PS(v1, v2, v3, v4);
}
// Functions that operate on ivec4s.
static inline ivec4 min(const ivec4& v1, const ivec4& v2) {
return ivec4(_mm_min_epi32(v1.val, v2.val));
}
static inline ivec4 max(const ivec4& v1, const ivec4& v2) {
return ivec4(_mm_max_epi32(v1.val, v2.val));
}
static inline ivec4 abs(const ivec4& v) {
return ivec4(_mm_abs_epi32(v.val));
}
static inline bool any(const ivec4& v) {
return !_mm_test_all_zeros(v, _mm_set1_epi32(0xFFFFFFFF));
}
// Mathematical operators involving a scalar and a vector.
static inline fvec4 operator+(float v1, const fvec4& v2) {
return fvec4(v1)+v2;
}
static inline fvec4 operator-(float v1, const fvec4& v2) {
return fvec4(v1)-v2;
}
static inline fvec4 operator*(float v1, const fvec4& v2) {
return fvec4(v1)*v2;
}
static inline fvec4 operator/(float v1, const fvec4& v2) {
return fvec4(v1)/v2;
}
// Operations for blending fvec4s based on an ivec4.
static inline fvec4 blend(const fvec4& v1, const fvec4& v2, const ivec4& mask) {
return fvec4(_mm_blendv_ps(v1.val, v2.val, _mm_castsi128_ps(mask.val)));
}
#if defined(__ANDROID__)
#include "vectorize_neon.h"
#else
#include "vectorize_sse.h"
#endif
#endif /*OPENMM_VECTORIZE_H_*/
openmmapi/include/openmm/internal/vectorize_neon.h 0 → 100644
View file @ 454caac9
#ifndef OPENMM_VECTORIZE_NEON_H_
#define OPENMM_VECTORIZE_NEON_H_
/* -------------------------------------------------------------------------- *
* OpenMM *
* -------------------------------------------------------------------------- *
* This is part of the OpenMM molecular simulation toolkit originating from *
* Simbios, the NIH National Center for Physics-Based Simulation of *
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2013-2014 Stanford University and the Authors. *
* Authors: Mateus Lima, Peter Eastman *
* Contributors: *
* *
* Permission is hereby granted, free of charge, to any person obtaining a *
* copy of this software and associated documentation files (the "Software"), *
* to deal in the Software without restriction, including without limitation *
* the rights to use, copy, modify, merge, publish, distribute, sublicense, *
* and/or sell copies of the Software, and to permit persons to whom the *
* Software is furnished to do so, subject to the following conditions: *
* *
* The above copyright notice and this permission notice shall be included in *
* all copies or substantial portions of the Software. *
* *
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR *
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, *
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL *
* THE AUTHORS, CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, *
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR *
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE *
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
#include <cpu-features.h>
#include <arm_neon.h>
#include <cmath>
typedef int int32_t;
// This file defines classes and functions to simplify vectorizing code with NEON.
class ivec4;
/**
* A four element vector of floats.
*/
class fvec4 {
public:
float32x4_t val;
fvec4() {}
fvec4(float v) : val(vdupq_n_f32(v)) {}
fvec4(float v1, float v2, float v3, float v4) {
float v[] = {v1, v2, v3, v4};
val = vld1q_f32(v);
}
fvec4(float32x4_t v) : val(v) {}
fvec4(const float* v) : val(vld1q_f32(v)) {}
operator float32x4_t() const {
return val;
}
float operator[](int i) const {
float result[4];
store(result);
return result[i];
}
void store(float* v) const {
vst1q_f32(v, val);
}
fvec4 operator+(const fvec4& other) const { // Tested OK
return vaddq_f32(val, other);
}
fvec4 operator-(const fvec4& other) const { // Tested OK
return vsubq_f32(val, other);
}
fvec4 operator*(const fvec4& other) const { // Tested OK
return vmulq_f32(val, other);
}
fvec4 operator/(const fvec4& other) const { // Tested OK
// NEON does not have a divide float-point operator, so we get the reciprocal and multiply.
float32x4_t reciprocal = vrecpeq_f32(other);
reciprocal = vmulq_f32(vrecpsq_f32(other, reciprocal), reciprocal);
reciprocal = vmulq_f32(vrecpsq_f32(other, reciprocal), reciprocal);
fvec4 result = vmulq_f32(val,reciprocal);
return result;
}
void operator+=(const fvec4& other) {
val = vaddq_f32(val, other);
}
void operator-=(const fvec4& other) {
val = vsubq_f32(val, other);
}
void operator*=(const fvec4& other) {
val = vmulq_f32(val, other);
}
void operator/=(const fvec4& other) {
val = val / other.val;
}
fvec4 operator-() const {
return vnegq_f32(val);
}
fvec4 operator&(const fvec4& other) const {
return vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(val), vreinterpretq_u32_f32(other)));
}
fvec4 operator|(const fvec4& other) const {
return vcvtq_f32_s32(vreinterpretq_s32_u32(vorrq_u32(vcvtq_u32_f32(val), vcvtq_u32_f32(other))));
}
fvec4 operator==(const fvec4& other) const {
return vcvtq_f32_s32(vreinterpretq_s32_u32(vceqq_f32(val, other)));
}
fvec4 operator!=(const fvec4& other) const {
return vcvtq_f32_s32(vreinterpretq_s32_u32(vmvnq_u32(vceqq_f32(val, other)))); // not(equals(val, other))
}
fvec4 operator>(const fvec4& other) const {
return vcvtq_f32_s32(vreinterpretq_s32_u32(vcgtq_f32(val, other)));
}
fvec4 operator<(const fvec4& other) const {
return vcvtq_f32_s32(vreinterpretq_s32_u32(vcltq_f32(val, other)));
}
fvec4 operator>=(const fvec4& other) const {
return vcvtq_f32_s32(vreinterpretq_s32_u32(vcgeq_f32(val, other)));
}
fvec4 operator<=(const fvec4& other) const {
return vcvtq_f32_s32(vreinterpretq_s32_u32(vcleq_f32(val, other)));
}
operator ivec4() const;
};
/**
* A four element vector of ints.
*/
class ivec4 {
public:
int32x4_t val;
ivec4() {}
ivec4(int v) : val(vdupq_n_s32(v)) {}
ivec4(int v1, int v2, int v3, int v4) {
int v[] = {v1, v2, v3, v4};
val = vld1q_s32(v);
}
ivec4(int32x4_t v) : val(v) {}
ivec4(const int* v) : val(vld1q_s32(v)) {}
operator int32x4_t() const {
return val;
}
int operator[](int i) const {
int result[4];
store(result);
return result[i];
}
void store(int* v) const {
vst1q_s32(v, val);
}
ivec4 operator+(const ivec4& other) const {
return vaddq_s32(val, other);
}
ivec4 operator-(const ivec4& other) const {
return vsubq_s32(val, other);
}
ivec4 operator*(const ivec4& other) const {
return vmulq_s32(val, other);
}
void operator+=(const ivec4& other) {
val = vaddq_s32(val, other);
}
void operator-=(const ivec4& other) {
val = vsubq_s32(val, other);
}
void operator*=(const ivec4& other) {
val = vmulq_s32(val, other);
}
ivec4 operator-() const {
return vnegq_s32(val);
}
ivec4 operator&(const ivec4& other) const { // Tested OK
return ivec4(vandq_s32(val, other));
}
ivec4 operator|(const ivec4& other) const {
return ivec4(vorrq_s32(val, other));
}
ivec4 operator==(const ivec4& other) const {
return ivec4(vreinterpretq_s32_u32(vceqq_s32(val, other)));
}
ivec4 operator!=(const ivec4& other) const { // OK
return ivec4(vreinterpretq_s32_u32(vmvnq_u32(vceqq_s32(val, other)))); // not(equal(val, other))
}
ivec4 operator>(const ivec4& other) const {
return ivec4(vreinterpretq_s32_u32(vcgtq_s32(val, other)));
}
ivec4 operator<(const ivec4& other) const {
return ivec4(vreinterpretq_s32_u32(vcltq_s32(val, other)));
}
ivec4 operator>=(const ivec4& other) const {
return ivec4(vreinterpretq_s32_u32(vcgeq_s32(val, other)));
}
ivec4 operator<=(const ivec4& other) const { // OK
return ivec4(vreinterpretq_s32_u32(vcleq_s32(val, other)));
}
operator fvec4() const;
};
// Conversion operators.
inline fvec4::operator ivec4() const {
return ivec4(vcvtq_s32_f32(val));
}
inline ivec4::operator fvec4() const {
return fvec4(vcvtq_f32_s32(val));
}
// Functions that operate on fvec4s.
static inline fvec4 floor(const fvec4& v) { // Tested: OK
fvec4 result = v + fvec4(0.5f);
result = (fvec4) ((ivec4) result);
return result;
}
static inline float roundToNearest(float num) {
return (num > 0.0f) ? std::floor(num + 0.5f) : std::ceil(num - 0.5f);
}
static inline fvec4 round(const fvec4& v) { // Tested: OK - Needs optimization
float aux[4];
vst1q_f32(aux, v);
return fvec4(roundToNearest(aux[0]), roundToNearest(aux[1]), roundToNearest(aux[2]), roundToNearest(aux[3]));
}
static inline fvec4 min(const fvec4& v1, const fvec4& v2) { // Tested OK
return fvec4(vminq_f32(v1.val, v2.val));
}
static inline fvec4 max(const fvec4& v1, const fvec4& v2) { // Tested OK
return fvec4(vmaxq_f32(v1.val, v2.val));
}
static inline fvec4 abs(const fvec4& v) { // Tested OK
return fvec4(vabdq_f32(v.val, fvec4(0.0)));
}
static inline fvec4 ceil(const fvec4& v) { // Tested OK
ivec4 intVersion = (ivec4) v;
fvec4 result = min((fvec4) (v > intVersion), fvec4(1.0f));
result += intVersion;
return result;
}
static inline fvec4 sqrt(const fvec4& v) {
float32x4_t recipSqrt = vrsqrteq_f32(v);
recipSqrt = vmulq_f32(recipSqrt, vrsqrtsq_f32(vmulq_f32(recipSqrt, v), recipSqrt));
recipSqrt = vmulq_f32(recipSqrt, vrsqrtsq_f32(vmulq_f32(recipSqrt, v), recipSqrt));
return vmulq_f32(v, recipSqrt);
}
static inline float dot3(const fvec4& v1, const fvec4& v2) { // Tested: OK
fvec4 result = v1 * v2;
float aux[4];
vst1q_f32(aux, result);
return aux[0] + aux[1] + aux[2]; // Ignore w component
}
static inline float dot4(const fvec4& v1, const fvec4& v2) { // Tested: OK
fvec4 result = v1 * v2;
float aux[4];
vst1q_f32(aux, result);
return aux[0] + aux[1] + aux[2] + aux[3];
}
static inline void transpose(fvec4& v1, fvec4& v2, fvec4& v3, fvec4& v4) { // Tested: OK
float aux1[4];
float aux2[4];
float aux3[4];
float aux4[4];
vst1q_f32(aux1, v1);
vst1q_f32(aux2, v2);
vst1q_f32(aux3, v3);
vst1q_f32(aux4, v4);
v1 = fvec4(aux1[0], aux2[0], aux3[0], aux4[0]);
v2 = fvec4(aux1[1], aux2[1], aux3[1], aux4[1]);
v3 = fvec4(aux1[2], aux2[2], aux3[2], aux4[2]);
v4 = fvec4(aux1[3], aux2[3], aux3[3], aux4[3]);
}
// Functions that operate on ivec4s.
static inline ivec4 min(const ivec4& v1, const ivec4& v2) { // Tested: not tested
ivec4 res = ivec4(vminq_s32(v1.val, v2.val));
return res;
}
static inline ivec4 max(const ivec4& v1, const ivec4& v2) { // Tested: not tested
ivec4 res = ivec4(vmaxq_s32(v1.val, v2.val));
return res;
}
static inline ivec4 abs(const ivec4& v) { // Tested: Not tested
ivec4 res = ivec4(vabdq_s32(v.val, ivec4(0)));
return res;
}
static inline bool any(const ivec4& v) { // Tested: OK
int result[4];
vst1q_s32(result, v);
return result[0] != 0 || result[1] != 0 || result[2] != 0 || result[3] != 0;
}
// Mathematical operators involving a scalar and a vector.
static inline fvec4 operator+(float v1, const fvec4& v2) {
return fvec4(v1)+v2;
}
static inline fvec4 operator-(float v1, const fvec4& v2) {
return fvec4(v1)-v2;
}
static inline fvec4 operator*(float v1, const fvec4& v2) {
return fvec4(v1)*v2;
}
static inline fvec4 operator/(float v1, const fvec4& v2) {
return fvec4(v1)/v2;
}
// Operations for blending fvec4s based on an ivec4.
static inline fvec4 blend(const fvec4& v1, const fvec4& v2, const ivec4& mask) { // Tested OK
return fvec4(vbslq_f32(vreinterpretq_u32_s32(mask.val), v2, v1));
}
#endif /*OPENMM_VECTORIZE_NEON_H_*/
openmmapi/include/openmm/internal/vectorize_sse.h 0 → 100644
View file @ 454caac9
#ifndef OPENMM_VECTORIZE_SSE_H_
#define OPENMM_VECTORIZE_SSE_H_
/* -------------------------------------------------------------------------- *
* OpenMM *
* -------------------------------------------------------------------------- *
* This is part of the OpenMM molecular simulation toolkit originating from *
* Simbios, the NIH National Center for Physics-Based Simulation of *
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2013 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
* Permission is hereby granted, free of charge, to any person obtaining a *
* copy of this software and associated documentation files (the "Software"), *
* to deal in the Software without restriction, including without limitation *
* the rights to use, copy, modify, merge, publish, distribute, sublicense, *
* and/or sell copies of the Software, and to permit persons to whom the *
* Software is furnished to do so, subject to the following conditions: *
* *
* The above copyright notice and this permission notice shall be included in *
* all copies or substantial portions of the Software. *
* *
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR *
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, *
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL *
* THE AUTHORS, CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, *
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR *
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE *
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
#include <smmintrin.h>
// This file defines classes and functions to simplify vectorizing code with SSE.
class ivec4;
/**
* A four element vector of floats.
*/
class fvec4 {
public:
__m128 val;
fvec4() {}
fvec4(float v) : val(_mm_set1_ps(v)) {}
fvec4(float v1, float v2, float v3, float v4) : val(_mm_set_ps(v4, v3, v2, v1)) {}
fvec4(__m128 v) : val(v) {}
fvec4(const float* v) : val(_mm_loadu_ps(v)) {}
operator __m128() const {
return val;
}
float operator[](int i) const {
float result[4];
store(result);
return result[i];
}
void store(float* v) const {
_mm_storeu_ps(v, val);
}
fvec4 operator+(const fvec4& other) const {
return _mm_add_ps(val, other);
}
fvec4 operator-(const fvec4& other) const {
return _mm_sub_ps(val, other);
}
fvec4 operator*(const fvec4& other) const {
return _mm_mul_ps(val, other);
}
fvec4 operator/(const fvec4& other) const {
return _mm_div_ps(val, other);
}
void operator+=(const fvec4& other) {
val = _mm_add_ps(val, other);
}
void operator-=(const fvec4& other) {
val = _mm_sub_ps(val, other);
}
void operator*=(const fvec4& other) {
val = _mm_mul_ps(val, other);
}
void operator/=(const fvec4& other) {
val = _mm_div_ps(val, other);
}
fvec4 operator-() const {
return _mm_sub_ps(_mm_set1_ps(0.0f), val);
}
fvec4 operator&(const fvec4& other) const {
return _mm_and_ps(val, other);
}
fvec4 operator|(const fvec4& other) const {
return _mm_or_ps(val, other);
}
fvec4 operator==(const fvec4& other) const {
return _mm_cmpeq_ps(val, other);
}
fvec4 operator!=(const fvec4& other) const {
return _mm_cmpneq_ps(val, other);
}
fvec4 operator>(const fvec4& other) const {
return _mm_cmpgt_ps(val, other);
}
fvec4 operator<(const fvec4& other) const {
return _mm_cmplt_ps(val, other);
}
fvec4 operator>=(const fvec4& other) const {
return _mm_cmpge_ps(val, other);
}
fvec4 operator<=(const fvec4& other) const {
return _mm_cmple_ps(val, other);
}
operator ivec4() const;
};
/**
* A four element vector of ints.
*/
class ivec4 {
public:
__m128i val;
ivec4() {}
ivec4(int v) : val(_mm_set1_epi32(v)) {}
ivec4(int v1, int v2, int v3, int v4) : val(_mm_set_epi32(v4, v3, v2, v1)) {}
ivec4(__m128i v) : val(v) {}
ivec4(const int* v) : val(_mm_loadu_si128((const __m128i*) v)) {}
operator __m128i() const {
return val;
}
int operator[](int i) const {
int result[4];
store(result);
return result[i];
}
void store(int* v) const {
_mm_storeu_si128((__m128i*) v, val);
}
ivec4 operator+(const ivec4& other) const {
return _mm_add_epi32(val, other);
}
ivec4 operator-(const ivec4& other) const {
return _mm_sub_epi32(val, other);
}
ivec4 operator*(const ivec4& other) const {
return _mm_mul_epi32(val, other);
}
void operator+=(const ivec4& other) {
val = _mm_add_epi32(val, other);
}
void operator-=(const ivec4& other) {
val = _mm_sub_epi32(val, other);
}
void operator*=(const ivec4& other) {
val = _mm_mul_epi32(val, other);
}
ivec4 operator-() const {
return _mm_sub_epi32(_mm_set1_epi32(0), val);
}
ivec4 operator&(const ivec4& other) const {
return _mm_and_si128(val, other);
}
ivec4 operator|(const ivec4& other) const {
return _mm_or_si128(val, other);
}
ivec4 operator==(const ivec4& other) const {
return _mm_cmpeq_epi32(val, other);
}
ivec4 operator!=(const ivec4& other) const {
return _mm_xor_si128(*this==other, _mm_set1_epi32(0xFFFFFFFF));
}
ivec4 operator>(const ivec4& other) const {
return _mm_cmpgt_epi32(val, other);
}
ivec4 operator<(const ivec4& other) const {
return _mm_cmplt_epi32(val, other);
}
ivec4 operator>=(const ivec4& other) const {
return _mm_xor_si128(_mm_cmplt_epi32(val, other), _mm_set1_epi32(0xFFFFFFFF));
}
ivec4 operator<=(const ivec4& other) const {
return _mm_xor_si128(_mm_cmpgt_epi32(val, other), _mm_set1_epi32(0xFFFFFFFF));
}
operator fvec4() const;
};
// Conversion operators.
inline fvec4::operator ivec4() const {
return _mm_cvttps_epi32(val);
}
inline ivec4::operator fvec4() const {
return _mm_cvtepi32_ps(val);
}
// Functions that operate on fvec4s.
static inline fvec4 floor(const fvec4& v) {
return fvec4(_mm_floor_ps(v.val));
}
static inline fvec4 ceil(const fvec4& v) {
return fvec4(_mm_ceil_ps(v.val));
}
static inline fvec4 round(const fvec4& v) {
return fvec4(_mm_round_ps(v.val, _MM_FROUND_TO_NEAREST_INT));
}
static inline fvec4 min(const fvec4& v1, const fvec4& v2) {
return fvec4(_mm_min_ps(v1.val, v2.val));
}
static inline fvec4 max(const fvec4& v1, const fvec4& v2) {
return fvec4(_mm_max_ps(v1.val, v2.val));
}
static inline fvec4 abs(const fvec4& v) {
static const __m128 mask = _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF));
return fvec4(_mm_and_ps(v.val, mask));
}
static inline fvec4 sqrt(const fvec4& v) {
return fvec4(_mm_sqrt_ps(v.val));
}
static inline float dot3(const fvec4& v1, const fvec4& v2) {
return _mm_cvtss_f32(_mm_dp_ps(v1, v2, 0x71));
}
static inline float dot4(const fvec4& v1, const fvec4& v2) {
return _mm_cvtss_f32(_mm_dp_ps(v1, v2, 0xF1));
}
static inline void transpose(fvec4& v1, fvec4& v2, fvec4& v3, fvec4& v4) {
_MM_TRANSPOSE4_PS(v1, v2, v3, v4);
}
// Functions that operate on ivec4s.
static inline ivec4 min(const ivec4& v1, const ivec4& v2) {
return ivec4(_mm_min_epi32(v1.val, v2.val));
}
static inline ivec4 max(const ivec4& v1, const ivec4& v2) {
return ivec4(_mm_max_epi32(v1.val, v2.val));
}
static inline ivec4 abs(const ivec4& v) {
return ivec4(_mm_abs_epi32(v.val));
}
static inline bool any(const ivec4& v) {
return !_mm_test_all_zeros(v, _mm_set1_epi32(0xFFFFFFFF));
}
// Mathematical operators involving a scalar and a vector.
static inline fvec4 operator+(float v1, const fvec4& v2) {
return fvec4(v1)+v2;
}
static inline fvec4 operator-(float v1, const fvec4& v2) {
return fvec4(v1)-v2;
}
static inline fvec4 operator*(float v1, const fvec4& v2) {
return fvec4(v1)*v2;
}
static inline fvec4 operator/(float v1, const fvec4& v2) {
return fvec4(v1)/v2;
}
// Operations for blending fvec4s based on an ivec4.
static inline fvec4 blend(const fvec4& v1, const fvec4& v2, const ivec4& mask) {
return fvec4(_mm_blendv_ps(v1.val, v2.val, _mm_castsi128_ps(mask.val)));
}
#endif /*OPENMM_VECTORIZE_SSE_H_*/
platforms/cpu/sharedTarget/CMakeLists.txt
View file @ 454caac9
......@@ -3,11 +3,15 @@ FOREACH(file ${SOURCE_FILES})
IF (MSVC)
SET_SOURCE_FILES_PROPERTIES(${file} PROPERTIES COMPILE_FLAGS "${EXTRA_COMPILE_FLAGS} /arch:AVX /D__AVX__")
ELSE (MSVC)
IF (NOT ANDROID)
SET_SOURCE_FILES_PROPERTIES(${file} PROPERTIES COMPILE_FLAGS "${EXTRA_COMPILE_FLAGS} -msse4.1 -mavx")
ENDIF (NOT ANDROID)
ENDIF (MSVC)
ELSE (file MATCHES ".*Vec8.*")
IF (NOT MSVC)
IF (NOT ANDROID)
SET_SOURCE_FILES_PROPERTIES(${file} PROPERTIES COMPILE_FLAGS "${EXTRA_COMPILE_FLAGS} -msse4.1")
ENDIF (NOT ANDROID)
ENDIF (NOT MSVC)
ENDIF (file MATCHES ".*Vec8.*")
ENDFOREACH(file)
......
platforms/cpu/src/CpuNeighborList.cpp
View file @ 454caac9
......@@ -37,7 +37,6 @@
#include <set>
#include <map>
#include <cmath>
#include <smmintrin.h>
using namespace std;
......
platforms/cpu/src/CpuNonbondedForceVec4.cpp
View file @ 454caac9
......@@ -103,7 +103,7 @@ void CpuNonbondedForceVec4::calculateBlockIxn(int blockIndex, float* forces, dou
dEdR = epsSig6*(12.0f*sig6 - 6.0f);
energy = epsSig6*(sig6-1.0f);
if (useSwitch) {
fvec4 t = (r>switchingDistance) & ((r-switchingDistance)*invSwitchingInterval);
fvec4 t = blend(0.0f, (r-switchingDistance)*invSwitchingInterval, r>switchingDistance);
fvec4 switchValue = 1+t*t*t*(-10.0f+t*(15.0f-t*6.0f));
fvec4 switchDeriv = t*t*(-30.0f+t*(60.0f-t*30.0f))*invSwitchingInterval;
dEdR = switchValue*dEdR - energy*switchDeriv*r;
......@@ -214,7 +214,7 @@ void CpuNonbondedForceVec4::calculateBlockEwaldIxn(int blockIndex, float* forces
dEdR = epsSig6*(12.0f*sig6 - 6.0f);
energy = epsSig6*(sig6-1.0f);
if (useSwitch) {
fvec4 t = (r>switchingDistance) & ((r-switchingDistance)*invSwitchingInterval);
fvec4 t = blend(0.0f, (r-switchingDistance)*invSwitchingInterval, r>switchingDistance);
fvec4 switchValue = 1+t*t*t*(-10.0f+t*(15.0f-t*6.0f));
fvec4 switchDeriv = t*t*(-30.0f+t*(60.0f-t*30.0f))*invSwitchingInterval;
dEdR = switchValue*dEdR - energy*switchDeriv*r;
......
platforms/cpu/src/CpuPlatform.cpp
View file @ 454caac9
......@@ -36,6 +36,7 @@
#include "ReferenceConstraints.h"
#include "openmm/internal/hardware.h"
#include <sstream>
#include <stdlib.h>
using namespace OpenMM;
using namespace std;
......@@ -93,8 +94,12 @@ bool CpuPlatform::supportsDoublePrecision() const {
}
bool CpuPlatform::isProcessorSupported() {
// Make sure the CPU supports SSE 4.1.
// Make sure the CPU supports SSE 4.1 or NEON.
#ifdef __ANDROID__
uint64_t features = android_getCpuFeatures();
return (features & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
#else
int cpuInfo[4];
cpuid(cpuInfo, 0);
if (cpuInfo[0] >= 1) {
......@@ -102,6 +107,7 @@ bool CpuPlatform::isProcessorSupported() {
return ((cpuInfo[2] & ((int) 1 << 19)) != 0);
}
return false;
#endif
}
void CpuPlatform::contextCreated(ContextImpl& context, const map<string, string>& properties) const {
......
platforms/opencl/src/OpenCLKernels.cpp
View file @ 454caac9
......@@ -4816,7 +4816,7 @@ void OpenCLIntegrateVariableVerletStepKernel::initialize(const System& system, c
kernel1 = cl::Kernel(program, "integrateVerletPart1");
kernel2 = cl::Kernel(program, "integrateVerletPart2");
selectSizeKernel = cl::Kernel(program, "selectVerletStepSize");
blockSize = min(min(256, system.getNumParticles()), (int) selectSizeKernel.getInfo<CL_KERNEL_WORK_GROUP_SIZE>());
blockSize = min(min(256, system.getNumParticles()), (int) selectSizeKernel.getWorkGroupInfo<CL_KERNEL_WORK_GROUP_SIZE>(cl.getDevice()));
}
double OpenCLIntegrateVariableVerletStepKernel::execute(ContextImpl& context, const VariableVerletIntegrator& integrator, double maxTime) {
......@@ -4926,7 +4926,7 @@ void OpenCLIntegrateVariableLangevinStepKernel::initialize(const System& system,
params = new OpenCLArray(cl, 3, cl.getUseDoublePrecision() || cl.getUseMixedPrecision() ? sizeof(cl_double) : sizeof(cl_float), "langevinParams");
blockSize = min(256, system.getNumParticles());
blockSize = max(blockSize, params->getSize());
blockSize = min(blockSize, (int) selectSizeKernel.getInfo<CL_KERNEL_WORK_GROUP_SIZE>());
blockSize = min(blockSize, (int) selectSizeKernel.getWorkGroupInfo<CL_KERNEL_WORK_GROUP_SIZE>(cl.getDevice()));
}
double OpenCLIntegrateVariableLangevinStepKernel::execute(ContextImpl& context, const VariableLangevinIntegrator& integrator, double maxTime) {
......
platforms/opencl/src/OpenCLPlatform.cpp
View file @ 454caac9
......@@ -32,6 +32,7 @@
#include "openmm/Context.h"
#include "openmm/System.h"
#include <algorithm>
#include <cctype>
#include <sstream>
#ifdef __APPLE__
#include "sys/sysctl.h"
......@@ -39,10 +40,7 @@
using namespace OpenMM;
using std::map;
using std::string;
using std::stringstream;
using std::vector;
using namespace std;
#ifdef OPENMM_OPENCL_BUILDING_STATIC_LIBRARY
extern "C" void registerOpenCLPlatform() {
......
plugins/cpupme/CMakeLists.txt
View file @ 454caac9
......@@ -57,7 +57,11 @@ ENDFOREACH(subdir)
INCLUDE_DIRECTORIES(BEFORE ${CMAKE_CURRENT_SOURCE_DIR}/src)
IF (NOT MSVC)
IF (ANDROID)
SET_SOURCE_FILES_PROPERTIES(${SOURCE_FILES} PROPERTIES COMPILE_FLAGS "")
ELSE (ANDROID)
SET_SOURCE_FILES_PROPERTIES(${SOURCE_FILES} PROPERTIES COMPILE_FLAGS "-msse4.1")
ENDIF (ANDROID)
ENDIF (NOT MSVC)
# Include FFTW related files.
......
plugins/cpupme/src/CpuPmeKernels.cpp
View file @ 454caac9
......@@ -569,13 +569,20 @@ double CpuCalcPmeReciprocalForceKernel::finishComputation(IO& io) {
}
bool CpuCalcPmeReciprocalForceKernel::isProcessorSupported() {
// Make sure the CPU supports SSE 4.1 or NEON.
#ifdef __ANDROID__
uint64_t features = android_getCpuFeatures();
return (features & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
#else
int cpuInfo[4];
cpuid(cpuInfo, 0);
if (cpuInfo[0] >= 1) {
cpuid(cpuInfo, 1);
return ((cpuInfo[2] & ((int) 1 << 19)) != 0); // Require SSE 4.1
return ((cpuInfo[2] & ((int) 1 << 19)) != 0);
}
return false;
#endif
}
int CpuCalcPmeReciprocalForceKernel::findFFTDimension(int minimum) {
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment