int main(int argc, char *argv[])
{
int ret = 0;
if (argc < 2)
{
fprintf(stderr, "Usage: transalign_killer [--cldev=x.y] <input file>\n");
fprintf(stderr, " --cldev=x.y: x specifies the platform index, y the device index.\n");
return 1;
}
long seq_length;
char *sequence = load_text(argv[argc - 1], &seq_length);
if (!sequence)
return 1;
seq_length--; // Cut final 0 byte
// FIXME: All the following code relies on seq_length being a multiple of BASE.
long round_seq_length = round_up_to_power_of_two(seq_length, BASE_EXP);
long res_length = 0;
for (long len = round_seq_length / BASE; len; len /= BASE)
res_length += len;
// Use some random index to be searched for here
unsigned letter_index = seq_length / 2;
// Select an OpenCL device
cl_device_id dev = select_device(argc - 1, argv);
if (!dev)
return 1;
// Initialize the OpenCL st...ack
cl_context ctx = clCreateContext(NULL, 1, &dev, NULL, NULL, NULL);
cl_command_queue queue = clCreateCommandQueue(ctx, dev, 0, NULL);
// Load the OpenCL kernesl
char *prog_src = load_text("trans.cl", NULL);
if (!prog_src)
return 1;
cl_program prog = clCreateProgramWithSource(ctx, 1, (const char **)&prog_src, NULL, NULL);
free(prog_src);
// Build them
clBuildProgram(prog, 0, NULL, NULL, NULL, NULL);
cl_kernel k_iadd = clCreateKernel(prog, "k_iadd", NULL); // initial addition
cl_kernel k_cadd = clCreateKernel(prog, "k_cadd", NULL); // consecutive addition
assert(k_iadd);
assert(k_cadd);
// Create the result buffer
unsigned *result = malloc(res_length * sizeof(unsigned));
cl_mem result_gpu = clCreateBuffer(ctx, CL_MEM_READ_WRITE | HOST_PTR_POLICY, res_length * sizeof(unsigned), result, NULL);
clock_start();
/*** START OF ROCKET SCIENCE LEVEL RUNTIME-TIME INTENSIVE STUFF ***/
// Bandwidth intensive stuff goes here
// Copy the sequence to the video memory (or, generally speaking, the OpenCL device)
cl_mem seq_gpu = clCreateBuffer(ctx, CL_MEM_READ_WRITE | HOST_PTR_POLICY, seq_length * sizeof(char), sequence, NULL);
long bw1_time = clock_delta();
// GPU intensive stuff goes here
/**
* First, transform every - and \0 into a 0 and every other character into a
* 1. Then, add consecutive fields (BASE fields) together and store them at
* the beginning of the result buffer.
*/
clSetKernelArg(k_iadd, 0, sizeof(result_gpu), &result_gpu);
clSetKernelArg(k_iadd, 1, sizeof(seq_gpu), &seq_gpu);
clSetKernelArg(k_iadd, 2, sizeof(unsigned), &(unsigned){seq_length});
void clock_usdelay(unsigned int micros)
{
uint32_t start = clock_micros();
while (clock_delta(clock_micros(), start) < micros);
}
