| 1 | #include <stdio.h>
|
|---|
| 2 | #include <stdlib.h>
|
|---|
| 3 | #include <string.h>
|
|---|
| 4 |
|
|---|
| 5 | /*
|
|---|
| 6 | If increase MAX_DATA_SIZE, nothing happens if DATA_SIZE is given a concrete value
|
|---|
| 7 | If I increase DATA_SIZE to 2, I get
|
|---|
| 8 | Computed '2/2' correct values!
|
|---|
| 9 | Computed '1/2' correct values!
|
|---|
| 10 |
|
|---|
| 11 | If I make num_devices >= data_size, I get a num/num correct values, like the default
|
|---|
| 12 | data = 3, devices = 1
|
|---|
| 13 | Computed '3/3' correct values!
|
|---|
| 14 | Computed '2/3' correct values!
|
|---|
| 15 | Computed '2/3' correct values!
|
|---|
| 16 | Computed '1/3' correct values!
|
|---|
| 17 |
|
|---|
| 18 | data = 3, devices = 2
|
|---|
| 19 | Computed '3/3' correct values!
|
|---|
| 20 | Computed '2/3' correct values!
|
|---|
| 21 |
|
|---|
| 22 | data = 3, devices = 3
|
|---|
| 23 | Computed '3/3' correct values!
|
|---|
| 24 | */
|
|---|
| 25 | $input int DATA_SIZE;
|
|---|
| 26 | $input int NUM_DEVICES;
|
|---|
| 27 | $input int MAX_DATA_SIZE;
|
|---|
| 28 | $input int MAX_NUM_DEVICES;
|
|---|
| 29 | $assume 0 < DATA_SIZE && DATA_SIZE < MAX_DATA_SIZE;
|
|---|
| 30 | $assume 0 < NUM_DEVICES && NUM_DEVICES < MAX_NUM_DEVICES;
|
|---|
| 31 | $gbarrier gbarrier = $gbarrier_create($here, NUM_DEVICES);
|
|---|
| 32 | //struct goes here
|
|---|
| 33 | typedef struct
|
|---|
| 34 | {
|
|---|
| 35 | int device_id;
|
|---|
| 36 | //other variables
|
|---|
| 37 | float * input;
|
|---|
| 38 | float * output;
|
|---|
| 39 | int count;
|
|---|
| 40 | }process;
|
|---|
| 41 |
|
|---|
| 42 | //kernel goes here
|
|---|
| 43 | void square(int device_id, float* input, float* output, const unsigned int count)
|
|---|
| 44 | {
|
|---|
| 45 | //int i = get_global_id(0);
|
|---|
| 46 | int i = device_id;
|
|---|
| 47 | if (i < count)
|
|---|
| 48 | {
|
|---|
| 49 | output[i] = input[i] * input[i];
|
|---|
| 50 | }
|
|---|
| 51 | }
|
|---|
| 52 | /*
|
|---|
| 53 | Note that the original lines were "__kernel void square( \n" \
|
|---|
| 54 | " __global float* input, \n" \
|
|---|
| 55 | " __global float* output, \n" \
|
|---|
| 56 | " const unsigned int count) \n" \
|
|---|
| 57 |
|
|---|
| 58 | Any parser must take note of and don't input \n, "", or \ as is
|
|---|
| 59 | __global float * input, __global float * output, int count;
|
|---|
| 60 | */
|
|---|
| 61 |
|
|---|
| 62 |
|
|---|
| 63 | int main(int argc, char** argv)
|
|---|
| 64 | {
|
|---|
| 65 | //get the number from clGetDeviceIDs 3rd parameter
|
|---|
| 66 | //int num_devices = 1;
|
|---|
| 67 |
|
|---|
| 68 |
|
|---|
| 69 | //variables from __kernel come here
|
|---|
| 70 | float * input;
|
|---|
| 71 | float * output;
|
|---|
| 72 | int count;
|
|---|
| 73 |
|
|---|
| 74 |
|
|---|
| 75 | //from the code before
|
|---|
| 76 | float data[DATA_SIZE]; // original data set given to device
|
|---|
| 77 | float results[DATA_SIZE]; // results returned from device
|
|---|
| 78 | int correct; // number of correct results returned
|
|---|
| 79 | //handle the definitions being put in different places
|
|---|
| 80 |
|
|---|
| 81 | int i = 0;
|
|---|
| 82 | count = DATA_SIZE; //count defined here
|
|---|
| 83 | for(i = 0; i < count; i++)
|
|---|
| 84 | {
|
|---|
| 85 | data[i] = rand() / (float)RAND_MAX;
|
|---|
| 86 | }
|
|---|
| 87 |
|
|---|
| 88 |
|
|---|
| 89 |
|
|---|
| 90 | //comes from clCreateBuffer
|
|---|
| 91 | input = (float *) malloc(sizeof(float) * count);
|
|---|
| 92 | output = (float *) malloc(sizeof(float) * count);
|
|---|
| 93 |
|
|---|
| 94 | //Possibly keep a list of variables, with a flag for whether they are init or not
|
|---|
| 95 | //Not init, malloc one from what is found in
|
|---|
| 96 | //output = clCreateBuffer(context, CL_MEM_WRITE_ONLY, sizeof(float) * count, NULL, NULL);
|
|---|
| 97 |
|
|---|
| 98 | input = data; //input defined here
|
|---|
| 99 | //came from clEnqueueWriteBuffer rather than the start of code
|
|---|
| 100 |
|
|---|
| 101 |
|
|---|
| 102 | //Phase after this is the definitions
|
|---|
| 103 | process now[NUM_DEVICES];
|
|---|
| 104 | //put device_ids
|
|---|
| 105 | for(int i = 0; i < NUM_DEVICES; i++)
|
|---|
| 106 | {
|
|---|
| 107 | now[i].device_id = i;
|
|---|
| 108 | //other variables
|
|---|
| 109 | now[i].input = input;
|
|---|
| 110 | now[i].output = output;
|
|---|
| 111 | now[i].count = count;
|
|---|
| 112 | }
|
|---|
| 113 |
|
|---|
| 114 | //spawns processes according to devices, uses the struct for inputs
|
|---|
| 115 | $proc procs[NUM_DEVICES];
|
|---|
| 116 | for(int i = 0; i < NUM_DEVICES; i++)
|
|---|
| 117 | {
|
|---|
| 118 | procs[i] = $spawn square(now[i].device_id, now[i].input, now[i].output, now[i].count);
|
|---|
| 119 | }
|
|---|
| 120 |
|
|---|
| 121 | for(int i = 0; i < NUM_DEVICES; i++)
|
|---|
| 122 | {
|
|---|
| 123 | $wait(procs[i]);
|
|---|
| 124 | }
|
|---|
| 125 |
|
|---|
| 126 | //$barrier barrier = $barrier_create($here, gbarrier, now[i].device_id);
|
|---|
| 127 | //$barrier_call(barrier);
|
|---|
| 128 | //$barrier_destroy(barrier);
|
|---|
| 129 |
|
|---|
| 130 |
|
|---|
| 131 |
|
|---|
| 132 | $gbarrier_destroy(gbarrier);
|
|---|
| 133 |
|
|---|
| 134 | //use the information from clEnqueueReadBuffer
|
|---|
| 135 | //may have to alter later
|
|---|
| 136 | for(int i = 0; i < count; i++)
|
|---|
| 137 | {
|
|---|
| 138 | results[i] = output[i];
|
|---|
| 139 | }
|
|---|
| 140 |
|
|---|
| 141 | correct = 0;
|
|---|
| 142 | for(i = 0; i < count; i++)
|
|---|
| 143 | {
|
|---|
| 144 | if(results[i] == data[i] * data[i])
|
|---|
| 145 | {
|
|---|
| 146 | correct++;
|
|---|
| 147 | }
|
|---|
| 148 | }
|
|---|
| 149 |
|
|---|
| 150 | // Print a brief summary detailing the results
|
|---|
| 151 | //
|
|---|
| 152 | printf("Computed '%d/%d' correct values!\n", correct, count);
|
|---|
| 153 |
|
|---|
| 154 |
|
|---|
| 155 | return 0;
|
|---|
| 156 | }
|
|---|