schroedinger/schrocuda.c

   1
   2 #ifdef HAVE_CONFIG_H
   3 #include "config.h"
   4 #endif
   5
   6 #include <schroedinger/schro.h>
   7 #include <cuda_runtime_api.h>
   8
   9 #include <schroedinger/schrogpuframe.h>
  10 #include <schroedinger/schrogpumotion.h>
  11
  12 void
  13 schro_cuda_init (void)
  14 {
  15   int n;
  16   int i;
  17   cudaError_t ret;
  18
  19   ret = cudaGetDeviceCount (&n);
  20   SCHRO_DEBUG ("cudaGetDeviceCount returned %d", ret);
  21
  22   SCHRO_DEBUG ("CUDA devices %d", n);
  23
  24   for (i = 0; i < n; i++) {
  25     struct cudaDeviceProp prop;
  26
  27     cudaGetDeviceProperties (&prop, i);
  28     SCHRO_DEBUG ("CUDA props: %d: %d.%d mem=%d %s", i,
  29         prop.major, prop.minor, prop.totalGlobalMem, prop.name);
  30   }
  31 }
  32
  33
  34 static void *
  35 schro_cuda_alloc (int size)
  36 {
  37   void *ptr;
  38   int ret;
  39
  40   SCHRO_DEBUG ("domain is %d", schro_async_get_exec_domain ());
  41   //SCHRO_ASSERT(schro_async_get_exec_domain () == SCHRO_EXEC_DOMAIN_CUDA);
  42
  43   ret = cudaMalloc (&ptr, size);
  44
  45   return ptr;
  46 }
  47
  48 static void
  49 schro_cuda_free (void *ptr, int size)
  50 {
  51   //SCHRO_ASSERT(schro_async_get_exec_domain () == SCHRO_EXEC_DOMAIN_CUDA);
  52
  53   if (schro_async_get_exec_domain () == SCHRO_EXEC_DOMAIN_CUDA) {
  54     cudaFree (ptr);
  55   } else {
  56     SCHRO_ERROR ("Freeing CUDA memory outside CUDA thread.");
  57   }
  58 }
  59
  60 SchroMemoryDomain *
  61 schro_memory_domain_new_cuda (void)
  62 {
  63   SchroMemoryDomain *domain;
  64
  65   domain = schro_memory_domain_new ();
  66   domain->flags = SCHRO_MEMORY_DOMAIN_CUDA;
  67   domain->alloc = schro_cuda_alloc;
  68   domain->free = schro_cuda_free;
  69
  70   return domain;
  71 }
  72
  73
  74 void
  75 schro_motion_render_cuda (SchroMotion * motion, SchroFrame * dest)
  76 {
  77   SchroGPUMotion *gpumotion;
  78
  79   gpumotion = schro_gpumotion_new (0);
  80
  81   schro_gpumotion_init (gpumotion, motion);
  82
  83   schro_gpumotion_copy (gpumotion, motion);
  84   schro_gpumotion_render (gpumotion, motion, dest);
  85
  86   schro_gpumotion_free (gpumotion);
  87 }
  88
  89 void
  90 schro_frame_inverse_iwt_transform_cuda (SchroFrame * frame,
  91     SchroFrame * transform_frame, SchroParams * params)
  92 {
  93   schro_frame_to_gpu (frame, transform_frame);
  94
  95   schro_gpuframe_inverse_iwt_transform (frame, params);
  96 }