call custom cuda device function from the generated code -凯发k8网页登录

call custom cuda device function from the generated code

if you have highly optimized cuda^® code for certain subfunctions that you want to incorporate into your generated code, gpu coder™ extends the functionality to help you achieve this goal.

the external cuda function must use the __device__ qualifier to execute the function on the gpu device. these device functions are different from global functions (kernels) in that they can only be called from other device or global functions. therefore the coder.ceval calls to the device functions must be from within a loop that gets mapped to a kernel. for information on integrating cuda kernels with the generated code, see .

note

code generation fails if the loop containing the coder.ceval calls cannot be mapped to a kernel. see the troubleshooting topic in the gpu coder documentation to check for issues preventing kernel creation and their suggested workarounds. if your matlab^® code section contains unsupported functions, then you must remove the coder.ceval calls from such sections.

call `__usad4_wrap` cuda device function

the stereo disparity example measures the distance between two corresponding points in the left and the right image of a stereo pair. the stereodisparity_cuda_sample entry-point function calls the __usad4_wrap external device function by using the coder.ceval function.

%% modified algorithm for stereo disparity block matching
% in this implementation instead of finding shifted image ,indices are mapped 
% accordingly to save memory and some processing rgba column major packed 
% data is used as input for compatibility with cuda intrinsics. convolution
% is performed using separable filters (horizontal and then vertical)
function [out_disp] = stereodisparity_cuda_sample(img0,img1)
coder.cinclude('cuda_intrinsic.h');
% gpu code generation pragma
coder.gpu.kernelfun;
%% stereo disparity parameters
% win_rad is the radius of the window to be operated,min_disparity is the 
% minimum disparity level the search continues for, max_disparity is the maximum
% disparity level the search continues for.
win_rad = 8;
min_disparity = -16;
max_disparity = 0;
%% image dimensions for loop control
% the number of channels packed are 4 (rgba) so as nchannels are 4
[imgheight,imgwidth]=size(img0);
nchannels = 4;
imgheight = imgheight/nchannels;
%% to store the raw differences
diff_img = zeros([imgheight 2*win_rad,imgwidth 2*win_rad],'int32');
%to store the minimum cost
min_cost = zeros([imgheight,imgwidth],'int32');
min_cost(:,:) = 99999999;
% store the final disparity
out_disp = zeros([imgheight,imgwidth],'int16');
%% filters for aggregating the differences
% filter_h is the horizontal filter used in separable convolution
% filter_v is the vertical filter used in separable convolution which
% operates on the output of the row convolution
filt_h = ones([1 17],'int32');
filt_v = ones([17 1],'int32');
%% main loop that runs for all the disparity levels. this loop is currently
% expected to run on cpu.
for d=min_disparity:max_disparity
    
    % find the difference matrix for the current disparity level. expect
    % this to generate a kernel function.
    coder.gpu.kernel;
    for colidx=1:imgwidth 2*win_rad
        coder.gpu.kernel;
        for rowidx=1:imgheight 2*win_rad
            % row index calculation
            ind_h = rowidx - win_rad;
            
            % column indices calculation for left image
            ind_w1 = colidx - win_rad;
            
            % row indices calculation for right image
            ind_w2 = colidx   d - win_rad;
            
            % border clamping for row indices
            if ind_h <= 0
                ind_h = 1;
            end
            if ind_h > imgheight
                ind_h = imgheight;
            end
            
            % border clamping for column indices for left image
            if ind_w1 <= 0
                ind_w1 = 1;
            end
            if ind_w1 > imgwidth
                ind_w1 = imgwidth;
            end
            
            % border clamping for column indices for right image
            if ind_w2 <= 0
                ind_w2 = 1;
            end
            if ind_w2 > imgwidth
                ind_w2 = imgwidth;
            end
            
            % in this step, sum of absolute differences is performed
            % across four channels. this piece of code is suitable 
            % for replacement with sad intrinsics
            tdiff = int32(0);
            tdiff = coder.ceval('-gpudevicefcn', '__usad4_wrap', 
                    coder.rref(img0((ind_h-1)*(nchannels) 1,ind_w1)), 
                    coder.rref(img1((ind_h-1)*(nchannels) 1,ind_w2)));
            
            %store the sad cost into a matrix
            diff_img(rowidx,colidx) = tdiff;
        end
    end
    
    % aggregating the differences using separable convolution. expect this 
    % to generate two kernel using shared memory.the first kernel is the 
    % convolution with the horizontal kernel and second kernel operates on
    % its output the column wise convolution.
    cost_v = conv2(diff_img,filt_h,'valid');
    cost = conv2(cost_v,filt_v,'valid');
    
    % this part updates the min_cost matrix with by comparing the values
    % with current disparity level. expect to generate a kernel for this.
    for ll=1:imgwidth
        for kk=1:imgheight
            % load the cost
            temp_cost = int32(cost(kk,ll));
            
            % compare against the minimum cost available and store the
            % disparity value
            if min_cost(kk,ll) > temp_cost
                min_cost(kk,ll) = temp_cost;
                out_disp(kk,ll) = abs(d)   8;
            end
            
        end
    end
    
end
end

the definition for the __usad4_wrap is written in an external file cuda_intrinsic.h. the file is located in the same folder as the entry-point function.

__device__ unsigned int __usad4(unsigned int a, unsigned int b, unsigned int c=0)
{
    unsigned int result;
#if (__cuda_arch__ >= 300) // kepler (sm 3.x) supports a 4 vector sad simd
    asm("vabsdiff4.u32.u32.u32.add" " %0, %1, %2, %3;": "=r"(result):"r"(a),
    "r"(b), "r"(c));
#else // sm 2.0            // fermi  (sm 2.x) supports only 1 sad simd, 
                           // so there are 4 instructions
    asm("vabsdiff.u32.u32.u32.add" " %0, %1.b0, %2.b0, %3;": 
         "=r"(result):"r"(a), "r"(b), "r"(c));
    asm("vabsdiff.u32.u32.u32.add" " %0, %1.b1, %2.b1, %3;": 
         "=r"(result):"r"(a), "r"(b), "r"(result));
    asm("vabsdiff.u32.u32.u32.add" " %0, %1.b2, %2.b2, %3;": 
         "=r"(result):"r"(a), "r"(b), "r"(result));
    asm("vabsdiff.u32.u32.u32.add" " %0, %1.b3, %2.b3, %3;": 
         "=r"(result):"r"(a), "r"(b), "r"(result));
#endif
    return result;
}
__device__ unsigned int packbytes(const uint8_t *inbytes)
{
    unsigned int packed = inbytes[0] | (inbytes[1] << 8) | 
                    (inbytes[2] << 16) | (inbytes[3] << 24);
    return packed;
}
__device__ unsigned int __usad4_wrap(const uint8_t *a, const uint8_t *b)
{
    unsigned int x = packbytes(a);
    unsigned int y = packbytes(b);
    return __usad4(x, y);
}

generate cuda code

generate cuda code by creating a code configuration object. specify the location of the custom c files by setting custom code properties (custominclude) on configuration objects. the following is an example code generation script that points to the location of cuda_intrinsic.h file.

cfg = coder.gpuconfig('mex');
cfg.custominclude = pwd;
codegen -config cfg -args {imgrgb0, imgrgb1} stereodisparity_cuda_sample_intrinsic;

generated code

gpu coder creates four kernels. the following is a snippet of the generated cuda code.

e_stereodisparity_cuda_sample_i<<>>
                    (gpu_img1, gpu_img0, d, gpu_diff_img);*/
/*  aggregating the differences using separable convolution.*/ 
/*  expect this to generate two kernel using shared memory.*/
/*  the first kernel is the convolution with the horizontal kernel and*/
/*  second kernel operates on its output the column wise convolution. */
f_stereodisparity_cuda_sample_i<<>>
                    (gpu_diff_img, gpu_a);
g_stereodisparity_cuda_sample_i<<>>
                    (gpu_a, gpu_cost_v);
h_stereodisparity_cuda_sample_i<<>>
                    (gpu_a, gpu_cost_v);
/*  this part updates the min_cost matrix with by comparing the values */
/*  with current disparity level. expect to generate a kernel for this. */
i_stereodisparity_cuda_sample_i<<>>
                    (d, gpu_cost, gpu_out_disp, gpu_min_cost);

the e_stereodisparity_cuda_sample_i kernel is the one that calls the __usad4_wrap device function. the following is a snippet of e_stereodisparity_cuda_sample_i kernel code.

static __global__ __launch_bounds__(512, 1) void e_stereodisparity_cuda_sample_i
  (const uint8_t *img1, const uint8_t *img0, int32_t d, int32_t *diff_img)
{
  ...
    /*  in this step, sum of absolute differences is performed */
    /*  across four channels. this piece of code is suitable */
    /*  for replacement with sad intrinsics */
    temp_cost = __usad4_wrap(&img0[((ind_h - 1) << 2)   2132 * (ind_w1 - 1)],
      &img1[((ind_h - 1) << 2)   2132 * (temp_cost - 1)]);
    /* store the sad cost into a matrix */
    diff_img[rowidx   549 * colidx] = temp_cost;
  }
}

call custom cuda device function from the generated code -凯发k8网页登录