add opencl tensor product contracted
This commit is contained in:
@@ -0,0 +1,156 @@
|
|||||||
|
#include "tensor_t/cl_tensor_t.h"
|
||||||
|
|
||||||
|
|
||||||
|
#define MAX_SOURCE_SIZE (0x100000)
|
||||||
|
|
||||||
|
|
||||||
|
#define CL_GEN_FUNC_TENSOR(type)\
|
||||||
|
tensor_##type* CREATE_CL_TENSOR_##type(dimension *dim){\
|
||||||
|
tensor_##type *r_tens=malloc(sizeof(tensor_##type));\
|
||||||
|
updateRankDim(dim);\
|
||||||
|
r_tens->dim = dim;\
|
||||||
|
r_tens->x = malloc(sizeof(type)*dim->rank);\
|
||||||
|
return r_tens;\
|
||||||
|
}\
|
||||||
|
\
|
||||||
|
\
|
||||||
|
void cl_tensorProd_##type(tensor_##type **MM, tensor_##type *M0, tensor_##type *M1) { \
|
||||||
|
dimension *dd; \
|
||||||
|
add_dimension(&dd, M0->dim, M1->dim); \
|
||||||
|
(*MM)=CREATE_TENSOR_##type(dd); \
|
||||||
|
tensor_##type *M = *MM; \
|
||||||
|
size_t m_idx;\
|
||||||
|
for(size_t i=0; i<M0->dim->rank; ++i){\
|
||||||
|
for(size_t j=0; j<M1->dim->rank; ++j){\
|
||||||
|
m_idx= i*M1->dim->rank + j ;\
|
||||||
|
M->x[m_idx]=M0->x[i]*M1->x[j];\
|
||||||
|
/*printf("[%ld|%ld:(%ld,%ld)]",x_idx++,m_idx,i,j);*/\
|
||||||
|
}\
|
||||||
|
}\
|
||||||
|
} \
|
||||||
|
\
|
||||||
|
/* M[x0,x1,x3..xn] X M[y0,y1,y3..ym] = M[z0,z1...zp] (deep = l > 0) /exists 1<= l<...<l=n / xl = y0,x{l+1}=y1, x{n}=yl et zi=xi i<n-l et zj=y{j-(n-l)} j>=n-l alor p=n+m-2l\
|
||||||
|
M[x0,x1,x3..xl x{l+1}...xn] X M[xn,x{n-1},x{n-2}...xl y{l+1} ..ym] = M[x0,x1..xly{l+1}...y{n+m-2l}] (deep = l > 0)\
|
||||||
|
M[[i][j]]=sum_{[k]}M0[[i][k]]*M[[k][j]]*/\
|
||||||
|
\
|
||||||
|
void cl_tensorContractnProd_##type(tensor_##type** MM, tensor_##type *M0, tensor_##type *M1, size_t contractionNumber) {\
|
||||||
|
\
|
||||||
|
size_t len0 = M0->dim->size - contractionNumber;\
|
||||||
|
size_t len1 = M1->dim->size - contractionNumber;\
|
||||||
|
\
|
||||||
|
size_t* tsub0 = malloc(sizeof(size_t) *len0);\
|
||||||
|
size_t* tsub1 = malloc(sizeof(size_t) *len1);\
|
||||||
|
size_t* tDk1 = malloc(sizeof(size_t) *contractionNumber);\
|
||||||
|
size_t* tDk0 = malloc(sizeof(size_t) *contractionNumber);\
|
||||||
|
subArray(tsub0, M0->dim->perm, 0, len0, 0);\
|
||||||
|
subArray(tsub1, M1->dim->perm, 0, len1, contractionNumber);\
|
||||||
|
subArray(tDk1, M1->dim->perm, 0, contractionNumber, 0);\
|
||||||
|
subArray(tDk0, M0->dim->perm, 0, contractionNumber, len0);\
|
||||||
|
dimension *dSub0 = init_dim(tsub0, len0);\
|
||||||
|
dimension *dSub1 = init_dim(tsub1, len1);\
|
||||||
|
dimension *dM1 = init_dim(tDk1, contractionNumber);\
|
||||||
|
dimension *dM0 = init_dim(tDk0, contractionNumber);\
|
||||||
|
dimension *dM;\
|
||||||
|
min_dimension(&dM, dM0, dM1);\
|
||||||
|
\
|
||||||
|
dimension *dd;\
|
||||||
|
add_dimension(&dd, dSub0, dSub1);\
|
||||||
|
updateRankDim(dd);\
|
||||||
|
*MM = CREATE_TENSOR_##type(dd);\
|
||||||
|
tensor_##type *M= *MM;\
|
||||||
|
\
|
||||||
|
\
|
||||||
|
/* Load the kernel source code into the array source_str*/ \
|
||||||
|
FILE *fp; \
|
||||||
|
char *source_str; \
|
||||||
|
size_t source_size; \
|
||||||
|
\
|
||||||
|
fp = fopen("/media/fanasina/corsair480/progr_/ytest/y_PROJECT/tensor_t/src/tensor_t/kernel_ProdContractnTensor.cl", "r"); \
|
||||||
|
if (!fp) { \
|
||||||
|
perror("kernel_ProdContractnTensor.cl");\
|
||||||
|
fprintf(stderr, "Failed to load kernel. \n"); \
|
||||||
|
exit(1); \
|
||||||
|
} \
|
||||||
|
source_str = (char*)malloc(MAX_SOURCE_SIZE); \
|
||||||
|
source_size = fread( source_str, 1, MAX_SOURCE_SIZE, fp); \
|
||||||
|
fclose( fp ); \
|
||||||
|
\
|
||||||
|
/*/ Get platform and device information */ \
|
||||||
|
cl_platform_id platform_id = NULL; \
|
||||||
|
cl_device_id device_id = NULL; \
|
||||||
|
cl_uint ret_num_devices; \
|
||||||
|
cl_uint ret_num_platforms; \
|
||||||
|
cl_int ret = clGetPlatformIDs(1, &platform_id, &ret_num_platforms); \
|
||||||
|
ret = clGetDeviceIDs( platform_id, CL_DEVICE_TYPE_DEFAULT, 1, \
|
||||||
|
&device_id, &ret_num_devices); \
|
||||||
|
\
|
||||||
|
/*/ Create an OpenCL context */ \
|
||||||
|
cl_context context = clCreateContext( NULL, 1, &device_id, NULL, NULL, &ret); \
|
||||||
|
\
|
||||||
|
/*/ Create a command queue */ \
|
||||||
|
cl_command_queue command_queue = clCreateCommandQueue(context, device_id, 0, &ret); \
|
||||||
|
\
|
||||||
|
\
|
||||||
|
/*/ Create memory buffers on the device for each vector */ \
|
||||||
|
cl_mem M0_mem_obj = clCreateBuffer(context, CL_MEM_READ_ONLY, \
|
||||||
|
M0->dim->rank * sizeof(type), NULL, &ret); \
|
||||||
|
cl_mem M1_mem_obj = clCreateBuffer(context, CL_MEM_READ_ONLY, \
|
||||||
|
M1->dim->rank * sizeof(type), NULL, &ret); \
|
||||||
|
cl_mem M_mem_obj = clCreateBuffer(context, CL_MEM_WRITE_ONLY, \
|
||||||
|
M->dim->rank * sizeof(type), NULL, &ret); \
|
||||||
|
\
|
||||||
|
/*/ Copy the lists A and B to their respective memory buffers */ \
|
||||||
|
ret = clEnqueueWriteBuffer(command_queue, M0_mem_obj, CL_TRUE, 0, \
|
||||||
|
M0->dim->rank * sizeof(type), M0->x, 0, NULL, NULL); \
|
||||||
|
ret = clEnqueueWriteBuffer(command_queue, M1_mem_obj, CL_TRUE, 0, \
|
||||||
|
M1->dim->rank * sizeof(type), M1->x, 0, NULL, NULL); \
|
||||||
|
\
|
||||||
|
/*/ Create a program from the kernel source */ \
|
||||||
|
cl_program program = clCreateProgramWithSource(context, 1, \
|
||||||
|
(const char **)&source_str, (const size_t *)&source_size, &ret); \
|
||||||
|
\
|
||||||
|
printf("log 0\n");\
|
||||||
|
/*/ Build the program */ \
|
||||||
|
ret = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL); \
|
||||||
|
size_t len; clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, NULL, NULL, &len);\
|
||||||
|
char *log = malloc(sizeof(char)*len);\
|
||||||
|
clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, len, log, NULL);\
|
||||||
|
printf("log: %s \n",log);\
|
||||||
|
\
|
||||||
|
/*/ Create the OpenCL kernel */ \
|
||||||
|
char func_cl_name[250]; sprintf(func_cl_name,"prodContractnTensorLin_%s", #type); \
|
||||||
|
printf("cl_func_type = %s\n",func_cl_name); \
|
||||||
|
cl_kernel kernel = clCreateKernel(program, func_cl_name, &ret); \
|
||||||
|
\
|
||||||
|
/*/ Set the arguments of the kernel */ \
|
||||||
|
ret = clSetKernelArg(kernel, 0, sizeof(size_t), (void *)&(dSub1->rank)); \
|
||||||
|
ret = clSetKernelArg(kernel, 1, sizeof(size_t), (void *)&(dM->rank)); \
|
||||||
|
ret = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&M0_mem_obj); \
|
||||||
|
ret = clSetKernelArg(kernel, 3, sizeof(cl_mem), (void *)&M1_mem_obj); \
|
||||||
|
ret = clSetKernelArg(kernel, 4, sizeof(cl_mem), (void *)&M_mem_obj); \
|
||||||
|
\
|
||||||
|
/*/ Execute the OpenCL kernel on the list */ \
|
||||||
|
size_t global_item_size = M->dim->rank; /*/ Process the entire lists */ \
|
||||||
|
size_t local_item_size = 1; /*64;*/ /*/ Divide work items into groups of 64 */ \
|
||||||
|
ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, \
|
||||||
|
&global_item_size, &local_item_size, 0, NULL, NULL); \
|
||||||
|
\
|
||||||
|
/*/ Read the memory buffer Mx on the device to the local variable M->x */ \
|
||||||
|
ret = clEnqueueReadBuffer(command_queue, M_mem_obj, CL_TRUE, 0, \
|
||||||
|
M->dim->rank * sizeof(type), M->x, 0, NULL, NULL); \
|
||||||
|
\
|
||||||
|
/*/ Clean up */ \
|
||||||
|
ret = clFlush(command_queue); \
|
||||||
|
ret = clFinish(command_queue); \
|
||||||
|
ret = clReleaseKernel(kernel); \
|
||||||
|
ret = clReleaseProgram(program); \
|
||||||
|
ret = clReleaseMemObject(M0_mem_obj); \
|
||||||
|
ret = clReleaseMemObject(M1_mem_obj); \
|
||||||
|
ret = clReleaseMemObject(M_mem_obj); \
|
||||||
|
ret = clReleaseCommandQueue(command_queue); \
|
||||||
|
ret = clReleaseContext(context); \
|
||||||
|
} \
|
||||||
|
|
||||||
|
|
||||||
|
CL_GEN_FUNC_TENSOR(TYPE_FLOAT);
|
||||||
|
CL_GEN_FUNC_TENSOR(TYPE_DOUBLE);
|
||||||
@@ -0,0 +1,24 @@
|
|||||||
|
#ifndef __CL_TENSOR_T__H__
|
||||||
|
#define __CL_TENSOR_T__H__
|
||||||
|
|
||||||
|
#include <stdio.h>
|
||||||
|
#include <stdlib.h>
|
||||||
|
|
||||||
|
#ifdef __APPLE__
|
||||||
|
#include <OpenCL/opencl.h>
|
||||||
|
#else
|
||||||
|
#include <CL/cl.h>
|
||||||
|
#endif
|
||||||
|
|
||||||
|
#include "tensor_t/tensor_t.h"
|
||||||
|
|
||||||
|
#define CL_GENERATE_TENSOR_TYPE(type) \
|
||||||
|
void cl_tensorProd_##type(tensor_##type **MM, tensor_##type *M0, tensor_##type *M1); \
|
||||||
|
void cl_tensorContractnProdNotOpt_##type(tensor_##type **MM, tensor_##type *M0, tensor_##type *M1, size_t contractionNumber); \
|
||||||
|
|
||||||
|
|
||||||
|
CL_GENERATE_TENSOR_TYPE(TYPE_FLOAT);
|
||||||
|
CL_GENERATE_TENSOR_TYPE(TYPE_DOUBLE);
|
||||||
|
|
||||||
|
|
||||||
|
#endif /* __CL_TENSOR_T__H__ */
|
||||||
@@ -0,0 +1,36 @@
|
|||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
__kernel void prodContractnTensorLin_TYPE_FLOAT(long unsigned int dSubRank, long unsigned int dMRank, __global const float *M0x , __global const float *M1x, __global float *Mx ){
|
||||||
|
|
||||||
|
//Get the index of the current element to be processed
|
||||||
|
size_t i = get_global_id(0);
|
||||||
|
size_t k, a0_id, a1_id, n0_id, n1_id;
|
||||||
|
a0_id = i / dSubRank;
|
||||||
|
a1_id = i % dSubRank;
|
||||||
|
Mx[i] = 0;
|
||||||
|
for (k = 0; k < dMRank; k++) {
|
||||||
|
n0_id = a0_id * dMRank + k;
|
||||||
|
n1_id = a1_id + dSubRank * k;
|
||||||
|
Mx[i] += M0x[n0_id] * M1x[n1_id];
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
__kernel void prodContractnTensorLin_TYPE_DOUBLE(long unsigned int dSubRank, long unsigned int dMRank, __global const double *M0x , __global const double *M1x, __global double *Mx ){
|
||||||
|
|
||||||
|
//Get the index of the current element to be processed
|
||||||
|
size_t i = get_global_id(0);
|
||||||
|
size_t k, a0_id, a1_id, n0_id, n1_id;
|
||||||
|
a0_id = i / dSubRank;
|
||||||
|
a1_id = i % dSubRank;
|
||||||
|
Mx[i] = 0;
|
||||||
|
for (k = 0; k < dMRank; k++) {
|
||||||
|
n0_id = a0_id * dMRank + k;
|
||||||
|
n1_id = a1_id + dSubRank * k;
|
||||||
|
Mx[i] += M0x[n0_id] * M1x[n1_id];
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
Reference in New Issue
Block a user