39 #ifndef PCL_GPU_KINFU_CUDA_UTILS_HPP_
40 #define PCL_GPU_KINFU_CUDA_UTILS_HPP_
51 __device__ __host__ __forceinline__
void swap ( T& a, T& b )
56 __device__ __forceinline__
float
57 dot(
const float3& v1,
const float3& v2)
59 return v1.x * v2.x + v1.y*v2.y + v1.z*v2.z;
62 __device__ __forceinline__ float3&
65 vec.x += v; vec.y += v; vec.z += v;
return vec;
68 __device__ __forceinline__ float3
71 return make_float3(v1.x + v2.x, v1.y + v2.y, v1.z + v2.z);
74 __device__ __forceinline__ float3&
77 vec.x *= v; vec.y *= v; vec.z *= v;
return vec;
80 __device__ __forceinline__ float3
83 return make_float3(v1.x - v2.x, v1.y - v2.y, v1.z - v2.z);
86 __device__ __forceinline__ float3
89 return make_float3(v1.x * v, v1.y * v, v1.z * v);
92 __device__ __forceinline__
float
95 return sqrt(
dot(v, v));
98 __device__ __forceinline__ float3
101 return v * rsqrt(
dot(v, v));
104 __device__ __host__ __forceinline__ float3
105 cross(
const float3& v1,
const float3& v2)
107 return make_float3(v1.y * v2.z - v1.z * v2.y, v1.z * v2.x - v1.x * v2.z, v1.x * v2.y - v1.y * v2.x);
110 __device__ __forceinline__
void computeRoots2(
const float& b,
const float& c, float3& roots)
113 float d = b * b - 4.f * c;
119 roots.z = 0.5f * (b + sd);
120 roots.y = 0.5f * (b - sd);
123 __device__ __forceinline__
void
126 if ( std::abs(c0) < std::numeric_limits<float>::epsilon())
132 const float s_inv3 = 1.f/3.f;
133 const float s_sqrt3 = sqrtf(3.f);
136 float c2_over_3 = c2 * s_inv3;
137 float a_over_3 = (c1 - c2*c2_over_3)*s_inv3;
141 float half_b = 0.5f * (c0 + c2_over_3 * (2.f * c2_over_3 * c2_over_3 - c1));
143 float q = half_b * half_b + a_over_3 * a_over_3 * a_over_3;
148 float rho = sqrtf(-a_over_3);
149 float theta = std::atan2 (sqrtf (-q), half_b)*s_inv3;
150 float cos_theta = __cosf (theta);
151 float sin_theta = __sinf (theta);
152 roots.x = c2_over_3 + 2.f * rho * cos_theta;
153 roots.y = c2_over_3 - rho * (cos_theta + s_sqrt3 * sin_theta);
154 roots.z = c2_over_3 - rho * (cos_theta - s_sqrt3 * sin_theta);
157 if (roots.x >= roots.y)
158 swap(roots.x, roots.y);
160 if (roots.y >= roots.z)
162 swap(roots.y, roots.z);
164 if (roots.x >= roots.y)
165 swap (roots.x, roots.y);
179 __device__ __host__ __forceinline__ float3&
operator[](
int i) {
return data[i]; }
180 __device__ __host__ __forceinline__
const float3&
operator[](
int i)
const {
return data[i]; }
186 static __forceinline__ __device__ float3
197 if(!isMuchSmallerThan(src.x, src.z) || !isMuchSmallerThan(src.y, src.z))
199 float invnm = rsqrtf(src.x*src.x + src.y*src.y);
200 perp.x = -src.y * invnm;
201 perp.y = src.x * invnm;
210 float invnm = rsqrtf(src.z * src.z + src.y * src.y);
212 perp.y = -src.z * invnm;
213 perp.z = src.y * invnm;
219 __device__ __forceinline__
220 Eigen33(
volatile float* mat_pkg_arg) : mat_pkg(mat_pkg_arg) {}
221 __device__ __forceinline__
void
227 float max01 = fmaxf( std::abs(mat_pkg[0]), std::abs(mat_pkg[1]) );
228 float max23 = fmaxf( std::abs(mat_pkg[2]), std::abs(mat_pkg[3]) );
229 float max45 = fmaxf( std::abs(mat_pkg[4]), std::abs(mat_pkg[5]) );
230 float m0123 = fmaxf( max01, max23);
231 float scale = fmaxf( max45, m0123);
233 if (scale <= std::numeric_limits<float>::min())
246 float c0 = m00() * m11() * m22()
247 + 2.f * m01() * m02() * m12()
248 - m00() * m12() * m12()
249 - m11() * m02() * m02()
250 - m22() * m01() * m01();
251 float c1 = m00() * m11() -
257 float c2 = m00() + m11() + m22();
261 if(evals.z - evals.x <= std::numeric_limits<float>::epsilon())
263 evecs[0] = make_float3(1.f, 0.f, 0.f);
264 evecs[1] = make_float3(0.f, 1.f, 0.f);
265 evecs[2] = make_float3(0.f, 0.f, 1.f);
267 else if (evals.y - evals.x <= std::numeric_limits<float>::epsilon() )
270 tmp[0] = row0(); tmp[1] = row1(); tmp[2] = row2();
271 tmp[0].x -= evals.z; tmp[1].y -= evals.z; tmp[2].z -= evals.z;
273 vec_tmp[0] =
cross(tmp[0], tmp[1]);
274 vec_tmp[1] =
cross(tmp[0], tmp[2]);
275 vec_tmp[2] =
cross(tmp[1], tmp[2]);
277 float len1 =
dot (vec_tmp[0], vec_tmp[0]);
278 float len2 =
dot (vec_tmp[1], vec_tmp[1]);
279 float len3 =
dot (vec_tmp[2], vec_tmp[2]);
281 if (len1 >= len2 && len1 >= len3)
283 evecs[2] = vec_tmp[0] * rsqrtf (len1);
285 else if (len2 >= len1 && len2 >= len3)
287 evecs[2] = vec_tmp[1] * rsqrtf (len2);
291 evecs[2] = vec_tmp[2] * rsqrtf (len3);
295 evecs[0] =
cross(evecs[1], evecs[2]);
297 else if (evals.z - evals.y <= std::numeric_limits<float>::epsilon() )
300 tmp[0] = row0(); tmp[1] = row1(); tmp[2] = row2();
301 tmp[0].x -= evals.x; tmp[1].y -= evals.x; tmp[2].z -= evals.x;
303 vec_tmp[0] =
cross(tmp[0], tmp[1]);
304 vec_tmp[1] =
cross(tmp[0], tmp[2]);
305 vec_tmp[2] =
cross(tmp[1], tmp[2]);
307 float len1 =
dot(vec_tmp[0], vec_tmp[0]);
308 float len2 =
dot(vec_tmp[1], vec_tmp[1]);
309 float len3 =
dot(vec_tmp[2], vec_tmp[2]);
311 if (len1 >= len2 && len1 >= len3)
313 evecs[0] = vec_tmp[0] * rsqrtf(len1);
315 else if (len2 >= len1 && len2 >= len3)
317 evecs[0] = vec_tmp[1] * rsqrtf(len2);
321 evecs[0] = vec_tmp[2] * rsqrtf(len3);
325 evecs[2] =
cross(evecs[0], evecs[1]);
330 tmp[0] = row0(); tmp[1] = row1(); tmp[2] = row2();
331 tmp[0].x -= evals.z; tmp[1].y -= evals.z; tmp[2].z -= evals.z;
333 vec_tmp[0] =
cross(tmp[0], tmp[1]);
334 vec_tmp[1] =
cross(tmp[0], tmp[2]);
335 vec_tmp[2] =
cross(tmp[1], tmp[2]);
337 float len1 =
dot(vec_tmp[0], vec_tmp[0]);
338 float len2 =
dot(vec_tmp[1], vec_tmp[1]);
339 float len3 =
dot(vec_tmp[2], vec_tmp[2]);
343 unsigned int min_el = 2;
344 unsigned int max_el = 2;
345 if (len1 >= len2 && len1 >= len3)
348 evecs[2] = vec_tmp[0] * rsqrtf (len1);
350 else if (len2 >= len1 && len2 >= len3)
353 evecs[2] = vec_tmp[1] * rsqrtf (len2);
358 evecs[2] = vec_tmp[2] * rsqrtf (len3);
361 tmp[0] = row0(); tmp[1] = row1(); tmp[2] = row2();
362 tmp[0].x -= evals.y; tmp[1].y -= evals.y; tmp[2].z -= evals.y;
364 vec_tmp[0] =
cross(tmp[0], tmp[1]);
365 vec_tmp[1] =
cross(tmp[0], tmp[2]);
366 vec_tmp[2] =
cross(tmp[1], tmp[2]);
368 len1 =
dot(vec_tmp[0], vec_tmp[0]);
369 len2 =
dot(vec_tmp[1], vec_tmp[1]);
370 len3 =
dot(vec_tmp[2], vec_tmp[2]);
372 if (len1 >= len2 && len1 >= len3)
375 evecs[1] = vec_tmp[0] * rsqrtf (len1);
376 min_el = len1 <= mmax[min_el] ? 1 : min_el;
377 max_el = len1 > mmax[max_el] ? 1 : max_el;
379 else if (len2 >= len1 && len2 >= len3)
382 evecs[1] = vec_tmp[1] * rsqrtf (len2);
383 min_el = len2 <= mmax[min_el] ? 1 : min_el;
384 max_el = len2 > mmax[max_el] ? 1 : max_el;
389 evecs[1] = vec_tmp[2] * rsqrtf (len3);
390 min_el = len3 <= mmax[min_el] ? 1 : min_el;
391 max_el = len3 > mmax[max_el] ? 1 : max_el;
394 tmp[0] = row0(); tmp[1] = row1(); tmp[2] = row2();
395 tmp[0].x -= evals.x; tmp[1].y -= evals.x; tmp[2].z -= evals.x;
397 vec_tmp[0] =
cross(tmp[0], tmp[1]);
398 vec_tmp[1] =
cross(tmp[0], tmp[2]);
399 vec_tmp[2] =
cross(tmp[1], tmp[2]);
401 len1 =
dot (vec_tmp[0], vec_tmp[0]);
402 len2 =
dot (vec_tmp[1], vec_tmp[1]);
403 len3 =
dot (vec_tmp[2], vec_tmp[2]);
406 if (len1 >= len2 && len1 >= len3)
409 evecs[0] = vec_tmp[0] * rsqrtf (len1);
410 min_el = len3 <= mmax[min_el] ? 0 : min_el;
411 max_el = len3 > mmax[max_el] ? 0 : max_el;
413 else if (len2 >= len1 && len2 >= len3)
416 evecs[0] = vec_tmp[1] * rsqrtf (len2);
417 min_el = len3 <= mmax[min_el] ? 0 : min_el;
418 max_el = len3 > mmax[max_el] ? 0 : max_el;
423 evecs[0] = vec_tmp[2] * rsqrtf (len3);
424 min_el = len3 <= mmax[min_el] ? 0 : min_el;
425 max_el = len3 > mmax[max_el] ? 0 : max_el;
428 unsigned mid_el = 3 - min_el - max_el;
429 evecs[min_el] =
normalized(
cross( evecs[(min_el+1) % 3], evecs[(min_el+2) % 3] ) );
430 evecs[mid_el] =
normalized(
cross( evecs[(mid_el+1) % 3], evecs[(mid_el+2) % 3] ) );
436 volatile float* mat_pkg;
438 __device__ __forceinline__
float m00()
const {
return mat_pkg[0]; }
439 __device__ __forceinline__
float m01()
const {
return mat_pkg[1]; }
440 __device__ __forceinline__
float m02()
const {
return mat_pkg[2]; }
441 __device__ __forceinline__
float m10()
const {
return mat_pkg[1]; }
442 __device__ __forceinline__
float m11()
const {
return mat_pkg[3]; }
443 __device__ __forceinline__
float m12()
const {
return mat_pkg[4]; }
444 __device__ __forceinline__
float m20()
const {
return mat_pkg[2]; }
445 __device__ __forceinline__
float m21()
const {
return mat_pkg[4]; }
446 __device__ __forceinline__
float m22()
const {
return mat_pkg[5]; }
448 __device__ __forceinline__ float3 row0()
const {
return make_float3( m00(), m01(), m02() ); }
449 __device__ __forceinline__ float3 row1()
const {
return make_float3( m10(), m11(), m12() ); }
450 __device__ __forceinline__ float3 row2()
const {
return make_float3( m20(), m21(), m22() ); }
452 __device__ __forceinline__
static bool isMuchSmallerThan (
float x,
float y)
455 const float prec_sqr = std::numeric_limits<float>::epsilon() * std::numeric_limits<float>::epsilon();
456 return x * x <= prec_sqr * y * y;
462 static __device__ __forceinline__
unsigned int stride()
464 return blockDim.x * blockDim.y * blockDim.z;
467 static __device__ __forceinline__
int
470 return threadIdx.z * blockDim.x * blockDim.y + threadIdx.y * blockDim.x + threadIdx.x;
473 template<
int CTA_SIZE,
typename T,
class BinOp>
474 static __device__ __forceinline__
void reduce(
volatile T* buffer, BinOp op)
479 if (CTA_SIZE >= 1024) {
if (tid < 512) buffer[tid] = val = op(val, buffer[tid + 512]); __syncthreads(); }
480 if (CTA_SIZE >= 512) {
if (tid < 256) buffer[tid] = val = op(val, buffer[tid + 256]); __syncthreads(); }
481 if (CTA_SIZE >= 256) {
if (tid < 128) buffer[tid] = val = op(val, buffer[tid + 128]); __syncthreads(); }
482 if (CTA_SIZE >= 128) {
if (tid < 64) buffer[tid] = val = op(val, buffer[tid + 64]); __syncthreads(); }
486 if (CTA_SIZE >= 64) { buffer[tid] = val = op(val, buffer[tid + 32]); }
487 if (CTA_SIZE >= 32) { buffer[tid] = val = op(val, buffer[tid + 16]); }
488 if (CTA_SIZE >= 16) { buffer[tid] = val = op(val, buffer[tid + 8]); }
489 if (CTA_SIZE >= 8) { buffer[tid] = val = op(val, buffer[tid + 4]); }
490 if (CTA_SIZE >= 4) { buffer[tid] = val = op(val, buffer[tid + 2]); }
491 if (CTA_SIZE >= 2) { buffer[tid] = val = op(val, buffer[tid + 1]); }
495 template<
int CTA_SIZE,
typename T,
class BinOp>
496 static __device__ __forceinline__ T
reduce(
volatile T* buffer, T init, BinOp op)
499 T val = buffer[tid] = init;
502 if (CTA_SIZE >= 1024) {
if (tid < 512) buffer[tid] = val = op(val, buffer[tid + 512]); __syncthreads(); }
503 if (CTA_SIZE >= 512) {
if (tid < 256) buffer[tid] = val = op(val, buffer[tid + 256]); __syncthreads(); }
504 if (CTA_SIZE >= 256) {
if (tid < 128) buffer[tid] = val = op(val, buffer[tid + 128]); __syncthreads(); }
505 if (CTA_SIZE >= 128) {
if (tid < 64) buffer[tid] = val = op(val, buffer[tid + 64]); __syncthreads(); }
509 if (CTA_SIZE >= 64) { buffer[tid] = val = op(val, buffer[tid + 32]); }
510 if (CTA_SIZE >= 32) { buffer[tid] = val = op(val, buffer[tid + 16]); }
511 if (CTA_SIZE >= 16) { buffer[tid] = val = op(val, buffer[tid + 8]); }
512 if (CTA_SIZE >= 8) { buffer[tid] = val = op(val, buffer[tid + 4]); }
513 if (CTA_SIZE >= 4) { buffer[tid] = val = op(val, buffer[tid + 2]); }
514 if (CTA_SIZE >= 2) { buffer[tid] = val = op(val, buffer[tid + 1]); }
531 static __device__ __forceinline__
unsigned int
535 asm(
"mov.u32 %0, %laneid;" :
"=r"(ret) );
539 static __device__ __forceinline__
unsigned int id()
541 int tid = threadIdx.z * blockDim.x * blockDim.y + threadIdx.y * blockDim.x + threadIdx.x;
545 static __device__ __forceinline__
549 asm(
"mov.u32 %0, %lanemask_lt;" :
"=r"(ret) );
562 static __device__ __forceinline__
int
565 const unsigned int lane = tid & 31;
569 int partial = ptr[tid];
571 ptr[tid] = partial = partial + ptr[tid + 16];
572 ptr[tid] = partial = partial + ptr[tid + 8];
573 ptr[tid] = partial = partial + ptr[tid + 4];
574 ptr[tid] = partial = partial + ptr[tid + 2];
575 ptr[tid] = partial = partial + ptr[tid + 1];
577 return ptr[tid - lane];
580 static __forceinline__ __device__
int
581 Ballot(
int predicate,
volatile int* cta_buffer)
583 #if CUDA_VERSION >= 9000
585 return __ballot_sync (__activemask (), predicate);
588 return __ballot(predicate);
592 static __forceinline__ __device__
bool
593 All(
int predicate,
volatile int* cta_buffer)
595 #if CUDA_VERSION >= 9000
597 return __all_sync (__activemask (), predicate);
600 return __all(predicate);
__device__ __host__ __forceinline__ void swap(T &a, T &b)
static __device__ __forceinline__ unsigned int stride()
static __device__ __forceinline__ void reduce(volatile T *buffer, BinOp op)
__device__ __forceinline__ float3 & operator+=(float3 &vec, const float &v)
static __device__ __forceinline__ T reduce(volatile T *buffer, T init, BinOp op)
__device__ __forceinline__ float3 operator-(const float3 &v1, const float3 &v2)
__device__ __forceinline__ float3 normalized(const float3 &v)
__device__ __host__ __forceinline__ const float3 & operator[](int i) const
__device__ __forceinline__ float3 operator*(const Mat33 &m, const float3 &vec)
__device__ __host__ __forceinline__ float3 cross(const float3 &v1, const float3 &v2)
static __device__ __forceinline__ int binaryExclScan(int ballot_mask)
static __device__ __forceinline__ unsigned int id()
__device__ __forceinline__ void compute(Mat33 &tmp, Mat33 &vec_tmp, Mat33 &evecs, float3 &evals)
static __device__ __forceinline__ int laneMaskLt()
__device__ __forceinline__ float dot(const float3 &v1, const float3 &v2)
static __forceinline__ __device__ bool All(int predicate, volatile int *cta_buffer)
static __forceinline__ __device__ int Ballot(int predicate, volatile int *cta_buffer)
static __forceinline__ __device__ float3 unitOrthogonal(const float3 &src)
__device__ __forceinline__ float3 & operator*=(float3 &vec, const float &v)
__device__ __forceinline__ void computeRoots2(const float &b, const float &c, float3 &roots)
__device__ __forceinline__ float3 operator+(const float3 &v1, const float3 &v2)
static __device__ __forceinline__ int flattenedThreadId()
static __device__ __forceinline__ unsigned int laneId()
Returns the warp lane ID of the calling thread.
__device__ __host__ __forceinline__ float norm(const float3 &v1, const float3 &v2)
__device__ __forceinline__ Eigen33(volatile float *mat_pkg_arg)
__device__ __host__ __forceinline__ float3 & operator[](int i)
__device__ __forceinline__ void computeRoots3(float c0, float c1, float c2, float3 &roots)
static __device__ __forceinline__ int warp_reduce(volatile int *ptr, const unsigned int tid)