/*
* compute d = r
* descr[0] = d, descr[1] = r
*/
void cpu_codelet_func_2(void *descr[], STARPU_ATTRIBUTE_UNUSED void *arg)
{
/* simply copy r into d */
uint32_t nx = STARPU_VECTOR_GET_NX(descr[0]);
size_t elemsize = STARPU_VECTOR_GET_ELEMSIZE(descr[0]);
STARPU_ASSERT(STARPU_VECTOR_GET_NX(descr[0]) == STARPU_VECTOR_GET_NX(descr[1]));
STARPU_ASSERT(STARPU_VECTOR_GET_ELEMSIZE(descr[0]) == STARPU_VECTOR_GET_ELEMSIZE(descr[1]));
float *src = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
float *dst = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
memcpy(dst, src, nx*elemsize);
}
void advance_eye_paths(void* buffers[], void* args_orig) {
const timeval start_time = my_WallClockTime();
// cl_args
const starpu_args args;
unsigned iteration;
starpu_codelet_unpack_args(args_orig, &args, &iteration);
// buffers
// hit point static info
HitPointPosition* const hit_points = reinterpret_cast<HitPointPosition* const>(STARPU_VECTOR_GET_PTR(buffers[0]));
// eye paths
EyePath* const eye_paths = reinterpret_cast<EyePath* const>(STARPU_VECTOR_GET_PTR(buffers[1]));
const unsigned eye_paths_count = STARPU_VECTOR_GET_NX(buffers[1]);
// seed buffer
Seed* const seed_buffer = reinterpret_cast<Seed* const>(STARPU_VECTOR_GET_PTR(buffers[2]));
advance_eye_paths_impl(hit_points, // hit_points_count,
eye_paths, eye_paths_count,
seed_buffer, // seed_buffer_count,
args.cpu_scene,
args.config->max_eye_path_depth,
starpu_combined_worker_get_size());
const timeval end_time = my_WallClockTime();
task_info("CPU", 0, starpu_combined_worker_get_size(), iteration, start_time, end_time, "(3) advance_eye_paths");
}
/* This kernel takes a buffer and scales it by a constant factor */
void vector_scal_cpu(void *buffers[], void *cl_arg)
{
unsigned i;
float *factor = cl_arg;
/*
* The "buffers" array matches the task->handles array: for instance
* task->handles[0] is a handle that corresponds to a data with
* vector "interface", so that the first entry of the array in the
* codelet is a pointer to a structure describing such a vector (ie.
* struct starpu_vector_interface *). Here, we therefore manipulate
* the buffers[0] element as a vector: nx gives the number of elements
* in the array, ptr gives the location of the array (that was possibly
* migrated/replicated), and elemsize gives the size of each elements.
*/
struct starpu_vector_interface *vector = buffers[0];
/* length of the vector */
unsigned n = STARPU_VECTOR_GET_NX(vector);
/* get a pointer to the local copy of the vector : note that we have to
* cast it in (float *) since a vector could contain any type of
* elements so that the .ptr field is actually a uintptr_t */
float *val = (float *)STARPU_VECTOR_GET_PTR(vector);
/* scale the vector */
for (i = 0; i < n; i++)
val[i] *= *factor;
}
void task_region_h(void *buffers[], void *_args)
{
void **args = _args;
struct starpu_vector_interface *_vector = buffers[0];
int nx = STARPU_VECTOR_GET_NX(_vector);
int elemsize = STARPU_VECTOR_GET_ELEMSIZE(_vector);
int slice_base = STARPU_VECTOR_GET_SLICE_BASE(_vector);
int *v = (int *)STARPU_VECTOR_GET_PTR(_vector);
int f = (int)(intptr_t)args[0];
int imin = (int)(intptr_t)args[1];
int imax = (int)(intptr_t)args[2];
int i;
assert(elemsize == sizeof(v[0]));
printf("depth 2 task, entry: vector ptr = %p, slice_base = %d, imin = %d, imax = %d\n", v, slice_base, imin, imax);
for (i = imin; i < imax; i++)
{
assert(i-slice_base>=0);
assert(i-slice_base<NX);
(v-slice_base)[i] += f;
}
printf("depth 2 task ending\n");
}
void cublas_codelet_func_5(void *descr[], void *arg)
{
float dot;
struct cg_problem *pb = arg;
float *vecd, *vecq;
uint32_t size;
/* get the vector */
vecd = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
vecq = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
STARPU_ASSERT(STARPU_VECTOR_GET_NX(descr[0]) == STARPU_VECTOR_GET_NX(descr[1]));
size = STARPU_VECTOR_GET_NX(descr[0]);
dot = cublasSdot (size, vecd, 1, vecq, 1);
pb->alpha = pb->delta_new / dot;
}
/* functions/codelet to fill the bufferss*/
void fill_tmp_buffer(void *buffers[], void *cl_arg)
{
int *tmp = (int *) STARPU_VECTOR_GET_PTR(buffers[0]);
int nx = STARPU_VECTOR_GET_NX(buffers[0]);
int i;
for (i=0; i<nx; i++)
tmp[i]=nx+i;
}
static void memset_cuda(void *descr[], void *arg)
{
STARPU_SKIP_IF_VALGRIND;
int *ptr = (int *)STARPU_VECTOR_GET_PTR(descr[0]);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
cudaMemsetAsync(ptr, 42, n * sizeof(*ptr), starpu_cuda_get_local_stream());
}
void memset_cpu(void *descr[], void *arg)
{
STARPU_SKIP_IF_VALGRIND;
int *ptr = (int *)STARPU_VECTOR_GET_PTR(descr[0]);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
memset(ptr, 42, n * sizeof(*ptr));
}
void read_ghost(void *buffers[], void *cl_arg)
{
int *tmp = (int *) STARPU_VECTOR_GET_PTR(buffers[0]);
int nx=STARPU_VECTOR_GET_NX(buffers[0]);
int i;
for(i=0; i<nx; i++)
{
assert(tmp[i]==nx+i);
}
}
void axpy_gpu(void *descr[], STARPU_ATTRIBUTE_UNUSED void *arg)
{
TYPE alpha = *((TYPE *)arg);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
TYPE *block_x = (TYPE *)STARPU_VECTOR_GET_PTR(descr[0]);
TYPE *block_y = (TYPE *)STARPU_VECTOR_GET_PTR(descr[1]);
CUBLASAXPY((int)n, alpha, block_x, 1, block_y, 1);
}
//! [To be included. You should update doxygen if you see this text.]
void scal_cpu_func(void *buffers[], void *_args)
{
unsigned i;
float *factor = _args;
struct starpu_vector_interface *vector = buffers[0];
unsigned n = STARPU_VECTOR_GET_NX(vector);
float *val = (float *)STARPU_VECTOR_GET_PTR(vector);
#pragma omp parallel for num_threads(starpu_combined_worker_get_size())
for (i = 0; i < n; i++)
val[i] *= *factor;
}
void task_region_g(void *buffers[], void *args)
{
struct starpu_vector_interface *_vector_1 = buffers[0];
int nx1 = STARPU_VECTOR_GET_NX(_vector_1);
int *v1 = (int *)STARPU_VECTOR_GET_PTR(_vector_1);
struct starpu_vector_interface *_vector_2 = buffers[1];
int nx2 = STARPU_VECTOR_GET_NX(_vector_2);
int *v2 = (int *)STARPU_VECTOR_GET_PTR(_vector_2);
int f = (int)(intptr_t)args;
STARPU_ASSERT(nx1 == nx2);
printf("depth 1 task, entry: vector_1 ptr = %p\n", v1);
printf("depth 1 task, entry: vector_2 ptr = %p\n", v2);
printf("depth 1 task, entry: f = %d\n", f);
fprintf(stderr, "cudaMemcpy: -->\n");
cudaMemcpy(v2,v1,nx1*sizeof(*_vector_1), cudaMemcpyDeviceToDevice);
fprintf(stderr, "cudaMemcpy: <--\n");
}
void cpu_codelet_func_6(void *descr[], void *arg)
{
struct cg_problem *pb = arg;
float *vecx, *vecd;
uint32_t size;
/* get the vector */
vecx = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
vecd = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
size = STARPU_VECTOR_GET_NX(descr[0]);
STARPU_SAXPY(size, pb->alpha, vecd, 1, vecx, 1);
}
void cublas_codelet_func_7(void *descr[], void *arg)
{
struct cg_problem *pb = arg;
float *vecr, *vecq;
uint32_t size;
/* get the vector */
vecr = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
vecq = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
size = STARPU_VECTOR_GET_NX(descr[0]);
cublasSaxpy (size, -pb->alpha, vecq, 1, vecr, 1);
}
void cublas_codelet_func_3(void *descr[], void *arg)
{
struct cg_problem *pb = arg;
float dot;
float *vec;
uint32_t size;
/* get the vector */
vec = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
size = STARPU_VECTOR_GET_NX(descr[0]);
dot = cublasSdot (size, vec, 1, vec, 1);
pb->delta_new = dot;
pb->delta_0 = dot;
}
/*
* Dot product codelet
*/
void dot_cpu_func(void *descr[], void *cl_arg)
{
long int *local_x = (long int *)STARPU_VECTOR_GET_PTR(descr[0]);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
long int *dot = (long int *)STARPU_VARIABLE_GET_PTR(descr[1]);
//FPRINTF_MPI(stderr, "Before dot=%ld (adding %d elements...)\n", *dot, n);
unsigned i;
for (i = 0; i < n; i++)
{
//FPRINTF_MPI(stderr, "Adding %ld\n", local_x[i]);
*dot += local_x[i];
}
//FPRINTF_MPI(stderr, "After dot=%ld\n", *dot);
}
void axpy_opencl(void *buffers[], void *_args)
{
TYPE *alpha = _args;
int id, devid;
cl_int err;
cl_kernel kernel;
cl_command_queue queue;
cl_event event;
unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
cl_mem x = (cl_mem) STARPU_VECTOR_GET_DEV_HANDLE(buffers[0]);
cl_mem y = (cl_mem) STARPU_VECTOR_GET_DEV_HANDLE(buffers[1]);
id = starpu_worker_get_id();
devid = starpu_worker_get_devid(id);
err = starpu_opencl_load_kernel(&kernel, &queue, &opencl_program, "_axpy_opencl", devid);
if (err != CL_SUCCESS)
STARPU_OPENCL_REPORT_ERROR(err);
err = clSetKernelArg(kernel, 0, sizeof(x), &x);
err|= clSetKernelArg(kernel, 1, sizeof(y), &y);
err|= clSetKernelArg(kernel, 2, sizeof(n), &n);
err|= clSetKernelArg(kernel, 3, sizeof(*alpha), alpha);
if (err)
STARPU_OPENCL_REPORT_ERROR(err);
{
size_t global=n;
size_t local;
size_t s;
cl_device_id device;
starpu_opencl_get_device(devid, &device);
err = clGetKernelWorkGroupInfo (kernel, device, CL_KERNEL_WORK_GROUP_SIZE, sizeof(local), &local, &s);
if (err != CL_SUCCESS)
STARPU_OPENCL_REPORT_ERROR(err);
if (local > global)
local=global;
err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 0, NULL, NULL);
if (err != CL_SUCCESS)
STARPU_OPENCL_REPORT_ERROR(err);
}
starpu_opencl_release_kernel(kernel);
}
void cublas_codelet_func_8(void *descr[], void *arg)
{
float dot;
struct cg_problem *pb = arg;
float *vecr;
uint32_t size;
/* get the vector */
vecr = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
size = STARPU_VECTOR_GET_NX(descr[0]);
dot = cublasSdot (size, vecr, 1, vecr, 1);
pb->delta_old = pb->delta_new;
pb->delta_new = dot;
pb->beta = pb->delta_new/pb->delta_old;
}
void cublas_codelet_func_9(void *descr[], void *arg)
{
struct cg_problem *pb = arg;
float *vecd, *vecr;
uint32_t size;
/* get the vector */
vecd = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
vecr = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
size = STARPU_VECTOR_GET_NX(descr[0]);
/* d = beta d */
cublasSscal(size, pb->beta, vecd, 1);
/* d = r + d */
cublasSaxpy (size, 1.0f, vecr, 1, vecd, 1);
}
void cpu_f(void *descr[], STARPU_ATTRIBUTE_UNUSED void *_args)
{
STARPU_SKIP_IF_VALGRIND;
unsigned *v = (unsigned *)STARPU_VECTOR_GET_PTR(descr[0]);
unsigned *tmp = (unsigned *)STARPU_VECTOR_GET_PTR(descr[1]);
unsigned nx = STARPU_VECTOR_GET_NX(descr[0]);
size_t elemsize = STARPU_VECTOR_GET_ELEMSIZE(descr[0]);
memcpy(tmp, v, nx*elemsize);
unsigned i;
for (i = 0; i < nx; i++)
{
v[i] = tmp[i] + 1;
}
}
void dot_cpu_func(void *descr[], void *cl_arg)
{
float *local_x = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
float *local_y = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
DOT_TYPE *dot = (DOT_TYPE *)STARPU_VARIABLE_GET_PTR(descr[2]);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
DOT_TYPE local_dot = 0.0;
unsigned i;
for (i = 0; i < n; i++)
{
local_dot += (DOT_TYPE)local_x[i]*(DOT_TYPE)local_y[i];
}
*dot = *dot + local_dot;
}
void cpu_codelet_func_3(void *descr[], void *arg)
{
struct cg_problem *pb = arg;
float dot;
float *vec;
int size;
/* get the vector */
vec = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
size = (int)STARPU_VECTOR_GET_NX(descr[0]);
dot = STARPU_SDOT(size, vec, 1, vec, 1);
fprintf(stderr, "func 3 : DOT = %f\n", dot);
pb->delta_new = dot;
pb->delta_0 = dot;
}
void dot_cuda_func(void *descr[], void *cl_arg)
{
DOT_TYPE current_dot;
DOT_TYPE local_dot;
float *local_x = (float *)STARPU_VECTOR_GET_PTR(descr[0]);
float *local_y = (float *)STARPU_VECTOR_GET_PTR(descr[1]);
DOT_TYPE *dot = (DOT_TYPE *)STARPU_VARIABLE_GET_PTR(descr[2]);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
cudaMemcpyAsync(¤t_dot, dot, sizeof(DOT_TYPE), cudaMemcpyDeviceToHost, starpu_cuda_get_local_stream());
cudaStreamSynchronize(starpu_cuda_get_local_stream());
local_dot = (DOT_TYPE)cublasSdot(n, local_x, 1, local_y, 1);
/* FPRINTF(stderr, "current_dot %f local dot %f -> %f\n", current_dot, local_dot, current_dot + local_dot); */
current_dot += local_dot;
cudaMemcpyAsync(dot, ¤t_dot, sizeof(DOT_TYPE), cudaMemcpyHostToDevice, starpu_cuda_get_local_stream());
cudaStreamSynchronize(starpu_cuda_get_local_stream());
}
void vector_scal_opencl(void *buffers[], void *_args)
{
float *factor = _args;
int id, devid, err;
cl_kernel kernel;
cl_command_queue queue;
cl_event event;
/* length of the vector */
unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
/* OpenCL copy of the vector pointer */
cl_mem val = (cl_mem) STARPU_VECTOR_GET_DEV_HANDLE(buffers[0]);
id = starpu_worker_get_id();
devid = starpu_worker_get_devid(id);
err = starpu_opencl_load_kernel(&kernel, &queue, &programs,
"vector_mult_opencl", devid); /* Name of the codelet defined above */
if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
err = clSetKernelArg(kernel, 0, sizeof(val), &val);
err |= clSetKernelArg(kernel, 1, sizeof(n), &n);
err |= clSetKernelArg(kernel, 2, sizeof(*factor), factor);
if (err) STARPU_OPENCL_REPORT_ERROR(err);
{
size_t global=1;
size_t local=1;
err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 0, NULL, &event);
if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);
}
clFinish(queue);
starpu_opencl_collect_stats(event);
clReleaseEvent(event);
starpu_opencl_release_kernel(kernel);
}
void minmax_cpu_func(void *descr[], void *cl_arg)
{
/* The array containing the values */
TYPE *local_array = (TYPE *)STARPU_VECTOR_GET_PTR(descr[0]);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
TYPE *minmax = (TYPE *)STARPU_VARIABLE_GET_PTR(descr[1]);
TYPE local_min = minmax[0];
TYPE local_max = minmax[1];
/* Compute the min and the max elements in the array */
unsigned i;
for (i = 0; i < n; i++)
{
TYPE val = local_array[i];
local_min = STARPU_MIN(local_min, val);
local_max = STARPU_MAX(local_max, val);
}
minmax[0] = local_min;
minmax[1] = local_max;
}
#define CUBLASAXPY cublasSaxpy
#define N (16*1024*1024)
#define NBLOCKS 8
TYPE *vec_x, *vec_y;
/* descriptors for StarPU */
starpu_data_handle handle_y, handle_x;
void axpy_cpu(void *descr[], __attribute__((unused)) void *arg)
{
TYPE alpha = *((TYPE *)arg);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
TYPE *block_x = (TYPE *)STARPU_VECTOR_GET_PTR(descr[0]);
TYPE *block_y = (TYPE *)STARPU_VECTOR_GET_PTR(descr[1]);
AXPY((int)n, alpha, block_x, 1, block_y, 1);
}
#ifdef STARPU_USE_CUDA
void axpy_gpu(void *descr[], __attribute__((unused)) void *arg)
{
TYPE alpha = *((TYPE *)arg);
unsigned n = STARPU_VECTOR_GET_NX(descr[0]);
TYPE *block_x = (TYPE *)STARPU_VECTOR_GET_PTR(descr[0]);
void task_region_g(void *buffers[], void *args)
{
struct starpu_vector_interface *_vector = buffers[0];
int nx = STARPU_VECTOR_GET_NX(_vector);
int *v = (int *)STARPU_VECTOR_GET_PTR(_vector);
int f = (int)(intptr_t)args;
printf("depth 1 task, entry: vector ptr = %p\n", v);
{
int i;
for (i = 0; i < nx; i++)
{
v[i] += f;
}
}
{
const int half_nx = nx/2;
starpu_data_handle_t vector_slice_1_handle;
starpu_vector_data_register(&vector_slice_1_handle, STARPU_MAIN_RAM, (uintptr_t)&v[0], half_nx, sizeof(v[0]));
printf("depth 1 task, block 1: vector_slice_1_handle = %p\n", vector_slice_1_handle);
starpu_data_handle_t vector_slice_2_handle;
starpu_vector_data_register(&vector_slice_2_handle, STARPU_MAIN_RAM, (uintptr_t)&v[half_nx], nx-half_nx, sizeof(v[0]));
/* set slice base */
starpu_omp_vector_annotate(vector_slice_2_handle, half_nx);
printf("depth 1 task, block 1: vector_slice_2_handle = %p\n", vector_slice_2_handle);
}
void *cl_arg_1[3];
void *cl_arg_2[3];
{
struct starpu_omp_task_region_attr attr;
const int half_nx = nx/2;
int i;
starpu_data_handle_t vector_slice_1_handle = starpu_data_lookup(&v[0]);
printf("depth 1 task, block 2: vector_slice_1_handle = %p\n", vector_slice_1_handle);
starpu_data_handle_t vector_slice_2_handle = starpu_data_lookup(&v[half_nx]);
printf("depth 1 task, block 2: vector_slice_2_handle = %p\n", vector_slice_2_handle);
memset(&attr, 0, sizeof(attr));
attr.cl.cpu_funcs[0] = task_region_h;
attr.cl.where = STARPU_CPU;
attr.cl.nbuffers = 1;
attr.cl.modes[0] = STARPU_RW;
attr.cl_arg_size = 3*sizeof(void *);
attr.cl_arg_free = 0;
attr.if_clause = 1;
attr.final_clause = 0;
attr.untied_clause = 1;
attr.mergeable_clause = 0;
i = 0;
cl_arg_1[0] = (void *)(intptr_t)i++;
cl_arg_1[1] = (void *)(intptr_t)0;
cl_arg_1[2] = (void *)(intptr_t)half_nx;
attr.cl_arg = cl_arg_1;
attr.handles = &vector_slice_1_handle;
starpu_omp_task_region(&attr);
cl_arg_2[0] = (void *)(intptr_t)i++;
cl_arg_2[1] = (void *)(intptr_t)half_nx;
cl_arg_2[2] = (void *)(intptr_t)nx;
attr.cl_arg = cl_arg_2;
attr.handles = &vector_slice_2_handle;
starpu_omp_task_region(&attr);
}
starpu_omp_taskwait();
}
#include <stdlib.h>
#define NLOOPS 1000
#define VECTORSIZE 1024
static starpu_data_handle v_handle;
/*
* Memset
*/
#ifdef STARPU_USE_CUDA
static void cuda_memset_codelet(void *descr[], __attribute__ ((unused)) void *_args)
{
char *buf = (char *)STARPU_VECTOR_GET_PTR(descr[0]);
unsigned length = STARPU_VECTOR_GET_NX(descr[0]);
cudaMemset(buf, 42, length);
cudaThreadSynchronize();
}
#endif
static void cpu_memset_codelet(void *descr[], __attribute__ ((unused)) void *_args)
{
char *buf = (char *)STARPU_VECTOR_GET_PTR(descr[0]);
unsigned length = STARPU_VECTOR_GET_NX(descr[0]);
memset(buf, 42, length);
}
static starpu_codelet memset_cl = {