GaussianBlur and Canny execution times are much longer on T-API

edit

Hello.

I've just started to learn OpenCV 3. I'm on OS X Yosemite. Here's my clinfo in the GPU part:

  Device Name                                     GeForce GT 330M
  Device Vendor                                   NVIDIA
  Device Vendor ID                                0x1022600
  Device Version                                  OpenCL 1.0 
  Driver Version                                  10.0.31 310.90.10.05b12
  Device OpenCL C Version                         OpenCL C 1.1 
  Device Type                                     GPU
  Device Profile                                  FULL_PROFILE
  Max compute units                               6
  Max clock frequency                             1100MHz
  Max work item dimensions                        3
  Max work item sizes                             512x512x64
  Max work group size                             512
  Preferred work group size multiple              32
  Preferred / native vector sizes                 
    char                                                 1 / 1       
    short                                                1 / 1       
    int                                                  1 / 1       
    long                                                 1 / 1       
    half                                                 0 / 0        (n/a)
    float                                                1 / 1       
    double                                               0 / 0        (n/a)
  Half-precision Floating-point support           (n/a)
  Single-precision Floating-point support         (core)
    Denormals                                     No
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               No
    Support is emulated in software               No
    Correctly-rounded divide and sqrt operations  No
  Double-precision Floating-point support         (n/a)
  Address bits                                    32, Little-Endian
  Global memory size                              268435456 (256MiB)
  Error Correction support                        No
  Max memory allocation                           134217728 (128MiB)
  Unified memory for Host and Device              No
  Minimum alignment for any data type             128 bytes
  Alignment of base address                       1024 bits (128 bytes)
  Global Memory cache type                        None
  Image support                                   Yes
    Max number of samplers per kernel             16
    Max 2D image size                             4096x4096 pixels
    Max 3D image size                             2048x2048x2048 pixels
    Max number of read image args                 128
    Max number of write image args                8
  Local memory type                               Local
  Local memory size                               16384 (16KiB)
  Max constant buffer size                        65536 (64KiB)
  Max number of constant args                     9
  Max size of kernel argument                     4352 (4.25KiB)
  Queue properties                                
    Out-of-order execution                        No
    Profiling                                     Yes
  Profiling timer resolution                      1000ns
  Execution capabilities                          
    Run OpenCL kernels                            Yes
    Run native kernels                            No
  Device Available                                Yes
  Compiler Available                              Yes
  Device Extensions                               cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event cl_khr_byte_addressable_store cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics

I wrote a little program to test T-API and it turns out that GaussianBlur and Canny take much much longer time to execute on T-API. Here's the code. It loads image and applies these two filter without and with T-API:

double totalTime = 0;
int64 start = getTickCount();
cvtColor(image, gray, COLOR_BGR2GRAY);
double timeMs = (getTickCount() - start) / getTickFrequency() * 1000;
totalTime += timeMs;
cout << "cvtColor ms [" << timeMs<< "]" << endl;

start = getTickCount();
GaussianBlur(gray, gray, Size(7, 7), 1.5);
timeMs = (getTickCount() - start) / getTickFrequency() * 1000;
totalTime += timeMs;
cout << "GaussianBlur ms [" << timeMs<< "]" << endl;

start = getTickCount();
Canny(gray, gray, 0, 50);
timeMs = (getTickCount() - start) / getTickFrequency() * 1000;
totalTime += timeMs;
cout << "Canny ms [" << timeMs<< "]" << endl;
cout << "= Total [" <<  totalTime << "]" << endl;

// TAPI
cout << endl << "TAPI results" << endl;
totalTime = 0;
UMat uimage;
UMat ugray;
imread(argv[1], CV_LOAD_IMAGE_COLOR).copyTo(uimage);

start = getTickCount();
cvtColor(uimage, ugray, COLOR_BGR2GRAY);
timeMs = (getTickCount() - start) / getTickFrequency() * 1000;
totalTime += timeMs;
cout << "TAPI cvtColor ms [" << timeMs<< "]" << endl;

start = getTickCount();
GaussianBlur(ugray, ugray, Size(7, 7), 1.5);
timeMs = (getTickCount() - start) / getTickFrequency() * 1000;
totalTime += timeMs;
cout << "TAPI GaussianBlur ms [" << timeMs<< "]" << endl;

start = getTickCount();
Canny(ugray, ugray, 0, 50);
timeMs = (getTickCount() - start) / getTickFrequency ...