36 #include <OpenCL/cl.h>
48 #include "opgrad_kernel.cl.h"
54 void *
dx,
void *
dy,
void *
dz,
58 void *
w3,
int *nel,
int *lx) {
64 const size_t global_item_size = 256 * (*nel);
65 const size_t local_item_size = 256;
71 cl_kernel kernel = clCreateKernel(opgrad_program, \
72 STR(opgrad_kernel_lx##LX), &err); \
75 CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *) &ux)); \
76 CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *) &uy)); \
77 CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *) &uz)); \
78 CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), (void *) &u)); \
79 CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), (void *) &dx)); \
80 CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem), (void *) &dy)); \
81 CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), (void *) &dz)); \
82 CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_mem), (void *) &drdx)); \
83 CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_mem), (void *) &dsdx)); \
84 CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_mem), (void *) &dtdx)); \
85 CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_mem), (void *) &drdy)); \
86 CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_mem), (void *) &dsdy)); \
87 CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_mem), (void *) &dtdy)); \
88 CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_mem), (void *) &drdz)); \
89 CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_mem), (void *) &dsdz)); \
90 CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_mem), (void *) &dtdz)); \
91 CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_mem), (void *) &w3)); \
93 CL_CHECK(clEnqueueNDRangeKernel((cl_command_queue) glb_cmd_queue, \
94 kernel, 1, NULL, &global_item_size, \
95 &local_item_size,0, NULL, NULL)); \
117 fprintf(stderr, __FILE__
": size not supported: %d\n", *lx);
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ drdy
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ drdz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dsdz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dsdy
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dtdy
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ u
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ drdx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dtdz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dsdx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dtdx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dy
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ w3
__global__ void T *__restrict__ uy
__global__ void T *__restrict__ T *__restrict__ uz
void opencl_kernel_jit(const char *kernel, cl_program *program)
void opencl_opgrad(void *ux, void *uy, void *uz, void *u, void *dx, void *dy, void *dz, void *drdx, void *dsdx, void *dtdx, void *drdy, void *dsdy, void *dtdy, void *drdz, void *dsdz, void *dtdz, void *w3, int *nel, int *lx)