device_wrapper.txx 5.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229
  1. /**
  2. * \file device_wrapper.txx
  3. * \author Dhairya Malhotra, dhairya.malhotra@gmail.com
  4. * \date 6-5-2013
  5. * \brief This file contains implementation of DeviceWrapper.
  6. */
  7. #include <vector.hpp>
  8. #include <device_wrapper.hpp>
  9. namespace pvfmm{
  10. namespace DeviceWrapper{
  11. // MIC functions
  12. inline uintptr_t alloc_device_mic(char* dev_handle, size_t len){
  13. assert(dev_handle!=NULL);
  14. uintptr_t dev_ptr=(uintptr_t)NULL;
  15. #ifdef __INTEL_OFFLOAD
  16. #pragma offload target(mic:0) nocopy( dev_handle: length(len) ALLOC) out(dev_ptr)
  17. #endif
  18. {dev_ptr=(uintptr_t)dev_handle;}
  19. return dev_ptr;
  20. }
  21. inline void free_device_mic(char* dev_handle, uintptr_t dev_ptr){
  22. #ifdef __INTEL_OFFLOAD
  23. #pragma offload target(mic:0) in( dev_handle: length(0) FREE)
  24. {
  25. assert(dev_ptr==(uintptr_t)dev_handle);
  26. }
  27. #endif
  28. }
  29. inline int host2device_mic(char* host_ptr, char* dev_handle, uintptr_t dev_ptr, size_t len){
  30. #ifdef __INTEL_OFFLOAD
  31. int wait_lock_idx=MIC_Lock::curr_lock();
  32. int lock_idx=MIC_Lock::get_lock();
  33. if(dev_handle==host_ptr){
  34. #pragma offload target(mic:0) in( dev_handle : length(len) REUSE ) signal(&MIC_Lock::lock_vec[lock_idx])
  35. {
  36. assert(dev_ptr==(uintptr_t)dev_handle);
  37. MIC_Lock::wait_lock(wait_lock_idx);
  38. MIC_Lock::release_lock(lock_idx);
  39. }
  40. }else{
  41. #pragma offload target(mic:0) in(host_ptr [0:len] : into ( dev_handle[0:len]) REUSE ) signal(&MIC_Lock::lock_vec[lock_idx])
  42. {
  43. assert(dev_ptr==(uintptr_t)dev_handle);
  44. MIC_Lock::wait_lock(wait_lock_idx);
  45. MIC_Lock::release_lock(lock_idx);
  46. }
  47. }
  48. #ifndef __MIC_ASYNCH__ // Wait
  49. #pragma offload target(mic:0)
  50. {MIC_Lock::wait_lock(lock_idx);}
  51. #endif
  52. return lock_idx;
  53. #endif
  54. return -1;
  55. }
  56. inline int device2host_mic(char* dev_handle, uintptr_t dev_ptr, char* host_ptr, size_t len){
  57. #ifdef __INTEL_OFFLOAD
  58. int wait_lock_idx=MIC_Lock::curr_lock();
  59. int lock_idx=MIC_Lock::get_lock();
  60. if(dev_handle==host_ptr){
  61. #pragma offload target(mic:0) out( dev_handle : length(len) REUSE ) signal(&MIC_Lock::lock_vec[lock_idx])
  62. {
  63. assert(dev_ptr==(uintptr_t)dev_handle);
  64. MIC_Lock::wait_lock(wait_lock_idx);
  65. MIC_Lock::release_lock(lock_idx);
  66. }
  67. }else{
  68. #pragma offload target(mic:0) out( dev_handle[0:len] : into (host_ptr [0:len]) REUSE ) signal(&MIC_Lock::lock_vec[lock_idx])
  69. {
  70. assert(dev_ptr==(uintptr_t)dev_handle);
  71. MIC_Lock::wait_lock(wait_lock_idx);
  72. MIC_Lock::release_lock(lock_idx);
  73. }
  74. }
  75. #ifndef __MIC_ASYNCH__ // Wait
  76. MIC_Lock::wait_lock(lock_idx);
  77. #endif
  78. return lock_idx;
  79. #endif
  80. return -1;
  81. }
  82. inline void wait_mic(int lock_idx){
  83. #ifdef __INTEL_OFFLOAD
  84. MIC_Lock::wait_lock(lock_idx);
  85. #endif
  86. }
  87. // Wrapper functions
  88. inline uintptr_t alloc_device(char* dev_handle, size_t len){
  89. #ifdef __INTEL_OFFLOAD
  90. return alloc_device_mic(dev_handle,len);
  91. #else
  92. uintptr_t dev_ptr=(uintptr_t)NULL;
  93. {dev_ptr=(uintptr_t)dev_handle;}
  94. return dev_ptr;
  95. #endif
  96. }
  97. inline void free_device(char* dev_handle, uintptr_t dev_ptr){
  98. #ifdef __INTEL_OFFLOAD
  99. free_device_mic(dev_handle,dev_ptr);
  100. #else
  101. ;
  102. #endif
  103. }
  104. inline int host2device(char* host_ptr, char* dev_handle, uintptr_t dev_ptr, size_t len){
  105. int lock_idx=-1;
  106. #ifdef __INTEL_OFFLOAD
  107. lock_idx=host2device_mic(host_ptr,dev_handle,dev_ptr,len);
  108. #else
  109. ;
  110. #endif
  111. return lock_idx;
  112. }
  113. inline int device2host(char* dev_handle, uintptr_t dev_ptr, char* host_ptr, size_t len){
  114. int lock_idx=-1;
  115. #ifdef __INTEL_OFFLOAD
  116. lock_idx=device2host_mic(dev_handle,dev_ptr, host_ptr, len);
  117. #else
  118. ;
  119. #endif
  120. return lock_idx;
  121. }
  122. inline void wait(int lock_idx){
  123. #ifdef __INTEL_OFFLOAD
  124. wait_mic(lock_idx);
  125. #else
  126. ;
  127. #endif
  128. }
  129. }
  130. // Implementation of MIC_Lock
  131. #ifdef __MIC__
  132. #define have_mic 1
  133. #else
  134. #define have_mic 0
  135. #endif
  136. inline void MIC_Lock::init(){
  137. if(have_mic) abort();// Cannot be called from MIC.
  138. lock_idx=0;
  139. lock_vec.Resize(NUM_LOCKS);
  140. lock_vec.SetZero();
  141. lock_vec_=lock_vec.AllocDevice(false);
  142. {for(size_t i=0;i<NUM_LOCKS;i++) lock_vec [i]=1;}
  143. #ifdef __INTEL_OFFLOAD
  144. #pragma offload target(mic:0)
  145. {for(size_t i=0;i<NUM_LOCKS;i++) lock_vec_[i]=1;}
  146. #endif
  147. }
  148. inline int MIC_Lock::get_lock(){
  149. if(have_mic) abort();// Cannot be called from MIC.
  150. int idx;
  151. #pragma omp critical
  152. {
  153. if(lock_idx==NUM_LOCKS-1){
  154. int wait_lock_idx=-1;
  155. wait_lock_idx=MIC_Lock::curr_lock();
  156. MIC_Lock::wait_lock(wait_lock_idx);
  157. #ifdef __INTEL_OFFLOAD
  158. #pragma offload target(mic:0)
  159. {MIC_Lock::wait_lock(wait_lock_idx);}
  160. #endif
  161. MIC_Lock::init();
  162. }
  163. idx=lock_idx;
  164. lock_idx++;
  165. assert(lock_idx<NUM_LOCKS);
  166. }
  167. return idx;
  168. }
  169. inline int MIC_Lock::curr_lock(){
  170. if(have_mic) abort();// Cannot be called from MIC.
  171. return lock_idx-1;
  172. }
  173. inline void MIC_Lock::release_lock(int idx){ // Only call from inside an offload section
  174. #ifdef __MIC__
  175. if(idx>=0) lock_vec_[idx]=0;
  176. #endif
  177. }
  178. inline void MIC_Lock::wait_lock(int idx){
  179. #ifdef __MIC__
  180. if(idx>=0) while(lock_vec_[idx]==1){
  181. _mm_delay_32(8192);
  182. }
  183. #else
  184. if(idx<0 || lock_vec[idx]==0) return;
  185. if(lock_vec[idx]==2){
  186. while(lock_vec[idx]==2);
  187. return;
  188. }
  189. lock_vec[idx]=2;
  190. #ifdef __INTEL_OFFLOAD
  191. #pragma offload_wait target(mic:0) wait(&lock_vec[idx])
  192. #endif
  193. lock_vec[idx]=0;
  194. #endif
  195. }
  196. Vector<char> MIC_Lock::lock_vec;
  197. Vector<char>::Device MIC_Lock::lock_vec_;
  198. int MIC_Lock::lock_idx;
  199. }//end namespace