////////
// User kernel host side function gets called to execute the user kernel node.
// Perform element-wise consine function on input tensor to produce output tensor.
//
// TODO:********
// 1. Get fixed-point position and dimensions of input and output tensors.
// Note that both input and output tensors have same dimensions.
// 2. Access input and output tensor object data using vxMapTensorPatch API.
// 3. Perform element-wise cosine function using fixed-point position.
// 4. Use vxUnmapTensorPatch API to give the data buffers control back to OpenVX framework.
vx_status VX_CALLBACK tensor_cos_host_side_function( vx_node node, const vx_reference * refs, vx_uint32 num )
{
// Get fixed-point position and dimensions of input and output tensors.
// Note that both input and output tensors have same dimensions.
vx_tensor input = ( vx_tensor ) refs[0];
vx_tensor output = ( vx_tensor ) refs[1];
vx_size num_of_dims;
vx_size dims[4] = { 1, 1, 1, 1 };
vx_uint8 input_fixed_point_pos;
vx_uint8 output_fixed_point_pos;
ERROR_CHECK_STATUS( vxQueryTensor( input, VX_TENSOR_NUMBER_OF_DIMS, &num_of_dims, sizeof( num_of_dims ) ) );
ERROR_CHECK_STATUS( vxQueryTensor( input, VX_TENSOR_DIMS, &dims, num_of_dims * sizeof(vx_size) ) );
ERROR_CHECK_STATUS( vxQueryTensor( input, VX_TENSOR_FIXED_POINT_POSITION, &input_fixed_point_pos, sizeof( input_fixed_point_pos ) ) );
ERROR_CHECK_STATUS( vxQueryTensor( output, VX_TENSOR_FIXED_POINT_POSITION, &output_fixed_point_pos, sizeof( output_fixed_point_pos ) ) );
// Access input and output tensor object data using vxMapTensorPatch API.
vx_size zeros[4] = { 0 };
vx_map_id map_input, map_output;
vx_uint8 * buf_input, * buf_output;
vx_size stride_input[4] = { 0 };
vx_size stride_output[4] = { 0 };
ERROR_CHECK_STATUS( vxMapTensorPatch( input,
num_of_dims, zeros, dims,
&map_input, stride_input,
(void **)&buf_input, VX_READ_ONLY, VX_MEMORY_TYPE_HOST, 0 ) );
ERROR_CHECK_STATUS( vxMapTensorPatch( output,
num_of_dims, zeros, dims,
&map_output, stride_output,
(void **)&buf_output, VX_READ_ONLY, VX_MEMORY_TYPE_HOST, 0 ) );
// Perform element-wise cosine function using fixed-point position.
vx_float32 input_to_float_multiplier = 1.0f / (vx_float32)(1 << input_fixed_point_pos);
vx_float32 output_to_int16_multiplier = (vx_float32)(1 << output_fixed_point_pos);
for( vx_size dim3 = 0; dim3 < dims[3]; dim3++)
{
for( vx_size dim2 = 0; dim2 < dims[2]; dim2++)
{
for( vx_size dim1 = 0; dim1 < dims[1]; dim1++)
{
const vx_int16 * ibuf = (const vx_int16 *) (buf_input +
dim3 * stride_input[3] +
dim2 * stride_input[2] +
dim1 * stride_input[1] );
vx_int16 * obuf = (vx_int16 *) (buf_output +
dim3 * stride_output[3] +
dim2 * stride_output[2] +
dim1 * stride_output[1] );
for( vx_size dim0 = 0; dim0 < dims[0]; dim0++)
{
// no saturation done here
vx_int16 ivalue = ibuf[dim0];
vx_int16 ovalue = (vx_int16)(cosf((vx_float32)ivalue * input_to_float_multiplier) * output_to_int16_multiplier + 0.5f);
obuf[dim0] = ovalue;
}
}
}
}
// Use vxUnmapTensorPatch API to give the data buffers control back to OpenVX framework.
ERROR_CHECK_STATUS( vxUnmapTensorPatch( input, map_input ) );
ERROR_CHECK_STATUS( vxUnmapTensorPatch( output, map_output ) );
return VX_SUCCESS;
}
////////
// main() has all the OpenVX application code for this exercise.
// Command-line usage:
// % solution_exercise3 [<video-sequence>|<camera-device-number>]
// When neither video sequence nor camera device number is specified,
// it defaults to the video sequence in "PETS09-S1-L1-View001.avi".
int main( int argc, char * argv[] )
{
// Get default video sequence when nothing is specified on command-line and
// instantiate OpenCV GUI module for reading input RGB images and displaying
// the image with OpenVX results
const char * video_sequence = argv[1];
CGuiModule gui( video_sequence );
// Try grab first video frame from the sequence using cv::VideoCapture
// and check if video frame is available
if( !gui.Grab() )
{
printf( "ERROR: input has no video\n" );
return 1;
}
////////
// Set the application configuration parameters. Note that input video
// sequence is an 8-bit RGB image with dimensions given by gui.GetWidth()
// and gui.GetHeight(). The parameters for the tensors are:
// tensor_dims - 3 dimensions of tensor [3 x <width> x <height>]
// tensor_input_fixed_point_pos - fixed-point position for input tensor
// tensor_output_fixed_point_pos - fixed-point position for output tensor
vx_uint32 width = gui.GetWidth();
vx_uint32 height = gui.GetHeight();
vx_size tensor_dims[3] = { width, height, 3 }; // 3 channels (RGB)
vx_uint8 tensor_input_fixed_point_pos = 5; // input[-128..127] will be mapped to -4..3.96875
vx_uint8 tensor_output_fixed_point_pos = 7; // output[-1..1] will be mapped to -128 to 128
////////
// Create the OpenVX context and make sure returned context is valid and
// register the log_callback to receive messages from OpenVX framework.
vx_context context = vxCreateContext();
ERROR_CHECK_OBJECT( context );
vxRegisterLogCallback( context, log_callback, vx_false_e );
////////
// Register user kernels with the context.
//
// TODO:********
// 1. Register user kernel with context by calling your implementation of "registerUserKernel()".
ERROR_CHECK_STATUS( registerUserKernel( context ) );
////////
// Create OpenVX tensor objects for input and output
//
// TODO:********
// 1. Create tensor objects using tensor_dims, tensor_input_fixed_point_pos, and
// tensor_output_fixed_point_pos
vx_tensor input_tensor = vxCreateTensor( context, 3, tensor_dims, VX_TYPE_INT16, tensor_input_fixed_point_pos );
vx_tensor output_tensor = vxCreateTensor( context, 3, tensor_dims, VX_TYPE_INT16, tensor_output_fixed_point_pos );
ERROR_CHECK_OBJECT( input_tensor );
ERROR_CHECK_OBJECT( output_tensor );
////////
// Create, build, and verify the graph with user kernel node.
//
// TODO:********
// 1. Build a graph with just one node created using userTensorCosNode()
vx_graph graph = vxCreateGraph( context );
ERROR_CHECK_OBJECT( graph );
vx_node cos_node = userTensorCosNode( graph, input_tensor, output_tensor );
ERROR_CHECK_OBJECT( cos_node );
ERROR_CHECK_STATUS( vxReleaseNode( &cos_node ) );
ERROR_CHECK_STATUS( vxVerifyGraph( graph ) );
////////
// Process the video sequence frame by frame until the end of sequence or aborted.
cv::Mat bgrMatForOutputDisplay( height, width, CV_8UC3 );
for( int frame_index = 0; !gui.AbortRequested(); frame_index++ )
{
////////
// Copy input RGB frame from OpenCV into input_tensor with UINT8 to Q10.5 (INT16) conversion.
// In order to do this, vxMapTensorPatch API (see "vx_ext_amd.h").
//
// TODO:********
// 1. Use vxMapTensorPatch API for access to input tensor object for writing
// 2. Copy UINT8 data from OpenCV RGB image to tensor object
// 3. Use vxUnmapTensorPatch API to return control of buffer back to framework
vx_uint8 * cv_rgb_image_buffer = gui.GetBuffer();
vx_size rgb_stride = gui.GetStride();
vx_size zeros[3] = { 0 };
vx_size tensor_stride[3];
vx_map_id map_id;
vx_uint8 * buf;
ERROR_CHECK_STATUS( vxMapTensorPatch( input_tensor,
3, zeros, tensor_dims,
&map_id, tensor_stride,
(void **)&buf, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST, 0 ) );
for( vx_size c = 0; c < 3; c++ )
{
for( vx_size y = 0; y < height; y++ )
{
const vx_uint8 * img = cv_rgb_image_buffer + y * rgb_stride + c;
vx_int16 * inp = (vx_int16 *)(buf + y * tensor_stride[1] + c * tensor_stride[2]);
for( vx_size x = 0; x < width; x++ )
{
// convert 0..255 to Q10.5 [-4..3.96875 range] fixed-point format
inp[x] = (vx_int16)img[x * 3] - 128;
}
}
}
ERROR_CHECK_STATUS( vxUnmapTensorPatch( input_tensor, map_id ) );
////////
// Now that input tensor is ready, just run the graph.
//
// TODO:********
// 1. Call vxProcessGraph to execute the tensor_cos kernel in graph
ERROR_CHECK_STATUS( vxProcessGraph( graph ) );
////////
// Display the output tensor object as RGB image
//
// TODO:********
// 1. Use vxMapTensorPatch API for access to output tensor object for reading
// 2. Copy tensor object data into OpenCV RGB image
// 3. Use vxUnmapTensorPatch API to return control of buffer back to framework
ERROR_CHECK_STATUS( vxMapTensorPatch( output_tensor,
3, zeros, tensor_dims,
&map_id, tensor_stride,
(void **)&buf, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST, 0 ) );
vx_uint8 * cv_bgr_image_buffer = bgrMatForOutputDisplay.data;
vx_size bgr_stride = bgrMatForOutputDisplay.step;
for( vx_size c = 0; c < 3; c++ )
{
for( vx_size y = 0; y < height; y++ )
{
const vx_int16 * out = (const vx_int16 *)(buf + y * tensor_stride[1] + c * tensor_stride[2]);
vx_uint8 * img = cv_bgr_image_buffer + y * bgr_stride + (2 - c); // (2 - c) for RGB to BGR conversion
for( vx_size x = 0; x < width; x++ )
{
// scale convert Q8.7 [-1..1 range] fixed-point format to 0..255 with saturation
vx_int16 value = out[x] + 128;
value = value > 255 ? 255 : value; // saturation needed
img[x * 3] = (vx_uint8)value;
}
}
}
#if ENABLE_DISPLAY
cv::imshow( "Cosine", bgrMatForOutputDisplay );
#endif
ERROR_CHECK_STATUS( vxUnmapTensorPatch( output_tensor, map_id ) );
////////
// Display the results and grab the next input RGB frame for the next iteration.
char text[128];
sprintf( text, "Keyboard ESC/Q-Quit SPACE-Pause [FRAME %d] [fixed_point_pos input:%d output:%d]", frame_index, tensor_input_fixed_point_pos, tensor_output_fixed_point_pos );
gui.DrawText( 0, 16, text );
gui.Show();
if( !gui.Grab() )
{
// Terminate the processing loop if the end of sequence is detected.
gui.WaitForKey();
break;
}
}
////////
// To release an OpenVX object, you need to call vxRelease<Object> API which takes a pointer to the object.
// If the release operation is successful, the OpenVX framework will reset the object to NULL.
//
// TODO:****
// 1. Release graph and tensor objects
ERROR_CHECK_STATUS( vxReleaseGraph( &graph ) );
ERROR_CHECK_STATUS( vxReleaseTensor( &input_tensor ) );
ERROR_CHECK_STATUS( vxReleaseTensor( &output_tensor ) );
ERROR_CHECK_STATUS( vxReleaseContext( &context ) );
return 0;
}
int CVxParamTensor::CompareFrame(int frameNumber)
{
// check if there is no user request to compare
if (m_fileNameCompare.length() < 1) return 0;
// reading data from reference file
char fileName[MAX_FILE_NAME_LENGTH]; sprintf(fileName, m_fileNameCompare.c_str(), frameNumber);
if(!_stricmp(fileName + strlen(fileName) - 4, ".dat")) {
ReportError("ERROR: read from .dat files not supported: %s\n", fileName);
}
FILE * fp = fopen(fileName, m_compareFileIsBinary ? "rb" : "r");
if (!fp) {
ReportError("ERROR: Unable to open: %s\n", fileName);
}
if (fread(m_data, 1, m_size, fp) != m_size)
ReportError("ERROR: not enough data (%d bytes) in %s\n", (vx_uint32)m_size, fileName);
fclose(fp);
// compare
vx_map_id map_id;
vx_size stride[MAX_TENSOR_DIMENSIONS];
vx_uint8 * ptr;
vx_status status = vxMapTensorPatch(m_tensor, m_num_of_dims, nullptr, nullptr, &map_id, stride, (void **)&ptr, VX_READ_ONLY, VX_MEMORY_TYPE_HOST, 0);
if (status != VX_SUCCESS)
ReportError("ERROR: vxMapTensorPatch: read failed (%d)\n", status);
bool mismatchDetected = false;
if (m_data_type == VX_TYPE_INT16) {
vx_int32 maxError = 0;
vx_int64 sumError = 0;
for (vx_size d3 = 0; d3 < m_dims[3]; d3++) {
for (vx_size d2 = 0; d2 < m_dims[2]; d2++) {
for (vx_size d1 = 0; d1 < m_dims[1]; d1++) {
vx_size roffset = m_stride[3] * d3 + m_stride[2] * d2 + m_stride[1] * d1;
vx_size doffset = stride[3] * d3 + stride[2] * d2 + stride[1] * d1;
const vx_int16 * buf1 = (const vx_int16 *)(((vx_uint8 *)ptr) + doffset);
const vx_int16 * buf2 = (const vx_int16 *)(m_data + roffset);
for (vx_size d0 = 0; d0 < m_dims[0]; d0++) {
vx_int32 v1 = buf1[d0];
vx_int32 v2 = buf2[d0];
vx_int32 d = v1 - v2;
d = (d < 0) ? -d : d;
maxError = (d > maxError) ? d : maxError;
sumError += d * d;
}
}
}
}
vx_size count = m_dims[0] * m_dims[1] * m_dims[2] * m_dims[3];
float avgError = (float)sumError / (float)count;
mismatchDetected = true;
if (((float)maxError <= m_maxErrorLimit) && ((float)avgError <= m_avgErrorLimit))
mismatchDetected = false;
if (mismatchDetected)
printf("ERROR: tensor COMPARE MISMATCHED [max-err: %d] [avg-err: %.6f] for %s with frame#%d of %s\n", maxError, avgError, GetVxObjectName(), frameNumber, fileName);
else if (m_verbose)
printf("OK: tensor COMPARE MATCHED [max-err: %d] [avg-err: %.6f] for %s with frame#%d of %s\n", maxError, avgError, GetVxObjectName(), frameNumber, fileName);
}
else if (m_data_type == VX_TYPE_FLOAT32) {
vx_float32 maxError = 0;
vx_float64 sumError = 0;
for (vx_size d3 = 0; d3 < m_dims[3]; d3++) {
for (vx_size d2 = 0; d2 < m_dims[2]; d2++) {
for (vx_size d1 = 0; d1 < m_dims[1]; d1++) {
vx_size roffset = m_stride[3] * d3 + m_stride[2] * d2 + m_stride[1] * d1;
vx_size doffset = stride[3] * d3 + stride[2] * d2 + stride[1] * d1;
const vx_float32 * buf1 = (const vx_float32 *)(((vx_uint8 *)ptr) + doffset);
const vx_float32 * buf2 = (const vx_float32 *)(m_data + roffset);
for (vx_size d0 = 0; d0 < m_dims[0]; d0++) {
vx_float32 v1 = buf1[d0];
vx_float32 v2 = buf2[d0];
vx_float32 d = v1 - v2;
d = (d < 0) ? -d : d;
maxError = (d > maxError) ? d : maxError;
sumError += d * d;
}
}
}
}
vx_size count = m_dims[0] * m_dims[1] * m_dims[2] * m_dims[3];
float avgError = (float)sumError / (float)count;
mismatchDetected = true;
if ((maxError <= m_maxErrorLimit) && (avgError <= m_avgErrorLimit))
mismatchDetected = false;
if (mismatchDetected)
printf("ERROR: tensor COMPARE MISMATCHED [max-err: %.6f] [avg-err: %.6f] for %s with frame#%d of %s\n", maxError, avgError, GetVxObjectName(), frameNumber, fileName);
else if (m_verbose)
printf("OK: tensor COMPARE MATCHED [max-err: %.6f] [avg-err: %.6f] for %s with frame#%d of %s\n", maxError, avgError, GetVxObjectName(), frameNumber, fileName);
}
else if (m_data_type == VX_TYPE_FLOAT16) {
vx_float32 maxError = 0;
vx_float64 sumError = 0;
for (vx_size d3 = 0; d3 < m_dims[3]; d3++) {
for (vx_size d2 = 0; d2 < m_dims[2]; d2++) {
for (vx_size d1 = 0; d1 < m_dims[1]; d1++) {
vx_size roffset = m_stride[3] * d3 + m_stride[2] * d2 + m_stride[1] * d1;
vx_size doffset = stride[3] * d3 + stride[2] * d2 + stride[1] * d1;
const vx_uint16 * buf1 = (const vx_uint16 *)(((vx_uint8 *)ptr) + doffset);
const vx_uint16 * buf2 = (const vx_uint16 *)(m_data + roffset);
for (vx_size d0 = 0; d0 < m_dims[0]; d0++) {
vx_uint16 h1 = buf1[d0];
vx_uint16 h2 = buf2[d0];
vx_uint32 d1 = ((h1 & 0x8000) << 16) | (((h1 & 0x7c00) + 0x1c000) << 13) | ((h1 & 0x03ff) << 13);
vx_uint32 d2 = ((h2 & 0x8000) << 16) | (((h2 & 0x7c00) + 0x1c000) << 13) | ((h2 & 0x03ff) << 13);
vx_float32 v1 = *(float *)&d1;
vx_float32 v2 = *(float *)&d2;
vx_float32 d = v1 - v2;
d = (d < 0) ? -d : d;
maxError = (d > maxError) ? d : maxError;
sumError += d * d;
}
}
}
}
vx_size count = m_dims[0] * m_dims[1] * m_dims[2] * m_dims[3];
float avgError = (float)sumError / (float)count;
mismatchDetected = true;
if ((maxError <= m_maxErrorLimit) && (avgError <= m_avgErrorLimit))
mismatchDetected = false;
if (mismatchDetected)
printf("ERROR: tensor COMPARE MISMATCHED [max-err: %.6f] [avg-err: %.6f] for %s with frame#%d of %s\n", maxError, avgError, GetVxObjectName(), frameNumber, fileName);
else if (m_verbose)
printf("OK: tensor COMPARE MATCHED [max-err: %.6f] [avg-err: %.6f] for %s with frame#%d of %s\n", maxError, avgError, GetVxObjectName(), frameNumber, fileName);
}
else {
for (vx_size d3 = 0; d3 < m_dims[3]; d3++) {
for (vx_size d2 = 0; d2 < m_dims[2]; d2++) {
for (vx_size d1 = 0; d1 < m_dims[1]; d1++) {
vx_size roffset = m_stride[3] * d3 + m_stride[2] * d2 + m_stride[1] * d1;
vx_size doffset = stride[3] * d3 + stride[2] * d2 + stride[1] * d1;
if (memcpy(((vx_uint8 *)ptr) + doffset, m_data + roffset, stride[0] * m_dims[0])) {
mismatchDetected = true;
break;
}
}
if (mismatchDetected)
break;
}
if (mismatchDetected)
break;
}
if (mismatchDetected)
printf("ERROR: tensor COMPARE MISMATCHED for %s with frame#%d of %s\n", GetVxObjectName(), frameNumber, fileName);
else if (m_verbose)
printf("OK: tensor COMPARE MATCHED for %s with frame#%d of %s\n", GetVxObjectName(), frameNumber, fileName);
}
status = vxUnmapTensorPatch(m_tensor, map_id);
if (status != VX_SUCCESS)
ReportError("ERROR: vxUnmapTensorPatch: read failed (%d)\n", status);
// report error if mismatched
if (mismatchDetected) {
m_compareCountMismatches++;
if (!m_discardCompareErrors) return -1;
}
else {
m_compareCountMatches++;
}
return 0;
}