restruct repository to ytest, create shared library for ytest
This commit is contained in:
@@ -0,0 +1,493 @@
|
||||
/*#include <cuda.h>
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
#include "cuda.h"
|
||||
#include "cuda_runtime.h"
|
||||
*/
|
||||
|
||||
#include "d_tensCuda.h"
|
||||
//#include "index.h"
|
||||
#include <stdio.h>
|
||||
|
||||
////////////////////////////////////////////////////////
|
||||
|
||||
//1D grid of 1D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_1D_1D() {
|
||||
return blockIdx.x * blockDim.x + threadIdx.x;
|
||||
}
|
||||
//1D grid of 2D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_1D_2D() {
|
||||
return blockIdx.x * blockDim.x * blockDim.y
|
||||
+ threadIdx.y * blockDim.x + threadIdx.x;
|
||||
}
|
||||
//1D grid of 3D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_1D_3D() {
|
||||
return blockIdx.x * blockDim.x * blockDim.y * blockDim.z
|
||||
+ threadIdx.z * blockDim.y * blockDim.x
|
||||
+ threadIdx.y * blockDim.x + threadIdx.x;
|
||||
}
|
||||
//2D grid of 1D blocks
|
||||
__device__ int d_getGlobalIdx_2D_1D() {
|
||||
int blockId
|
||||
= blockIdx.y * gridDim.x + blockIdx.x;
|
||||
int threadId = blockId * blockDim.x + threadIdx.x;
|
||||
return threadId;
|
||||
}
|
||||
//2D grid of 2D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_2D_2D() {
|
||||
int blockId = blockIdx.x + blockIdx.y * gridDim.x;
|
||||
int threadId = blockId * (blockDim.x * blockDim.y)
|
||||
+ (threadIdx.y * blockDim.x) + threadIdx.x;
|
||||
return threadId;
|
||||
}
|
||||
//2D grid of 3D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_2D_3D() {
|
||||
int blockId = blockIdx.x + blockIdx.y * gridDim.x;
|
||||
int threadId = blockId * (blockDim.x * blockDim.y * blockDim.z)
|
||||
+ (threadIdx.z * (blockDim.x * blockDim.y))
|
||||
+ (threadIdx.y * blockDim.x) + threadIdx.x;
|
||||
return threadId;
|
||||
}
|
||||
//3D grid of 1D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_3D_1D() {
|
||||
int blockId = blockIdx.x + blockIdx.y * gridDim.x
|
||||
+ gridDim.x * gridDim.y * blockIdx.z;
|
||||
int threadId = blockId * blockDim.x + threadIdx.x;
|
||||
return threadId;
|
||||
}
|
||||
//3D grid of 2D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_3D_2D() {
|
||||
int blockId = blockIdx.x + blockIdx.y * gridDim.x
|
||||
+ gridDim.x * gridDim.y * blockIdx.z;
|
||||
int threadId = blockId * (blockDim.x * blockDim.y)
|
||||
+ (threadIdx.y * blockDim.x) + threadIdx.x;
|
||||
return threadId;
|
||||
}
|
||||
//3D grid of 3D blocks
|
||||
__device__
|
||||
int d_getGlobalIdx_3D_3D() {
|
||||
int blockId = blockIdx.x + blockIdx.y * gridDim.x
|
||||
+ gridDim.x * gridDim.y * blockIdx.z;
|
||||
int threadId = blockId * (blockDim.x * blockDim.y * blockDim.z)
|
||||
+ (threadIdx.z * (blockDim.x * blockDim.y))
|
||||
+ (threadIdx.y * blockDim.x) + threadIdx.x;
|
||||
return threadId;
|
||||
}
|
||||
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
__device__ void d_LinearToCoordEnd(int* ret, size_t lin, int* dim, int rank, size_t size) {
|
||||
size_t sm = lin;
|
||||
size_t pp = size;
|
||||
for (int i = rank - 1;i > 0; --i) {
|
||||
pp /= dim[i];
|
||||
ret[i] = sm / pp;
|
||||
sm %= pp;
|
||||
}
|
||||
ret[0] = sm;
|
||||
}
|
||||
|
||||
__device__ size_t d_CoordToLinearEnd(int* coo, int* dim, int rank) {
|
||||
size_t pp = 1;
|
||||
size_t sm = 0;
|
||||
for (int i = 0; i < rank; ++i) {
|
||||
sm += (coo[i] * pp);
|
||||
pp *= dim[i];
|
||||
}
|
||||
return sm;
|
||||
}
|
||||
|
||||
__device__ size_t d_CoordToLinear(int* coo, int* dim, int rank) {
|
||||
size_t pp = 1;
|
||||
size_t sm = 0;
|
||||
for (int i = rank - 1; i >= 0; --i) {
|
||||
sm += (coo[i] * pp);
|
||||
pp *= dim[i];
|
||||
}
|
||||
return sm;
|
||||
}
|
||||
|
||||
|
||||
|
||||
__device__ void d_LinearToCoord(int* ret, size_t lin, int* dim, int rank, size_t size) {
|
||||
size_t sm = lin;
|
||||
size_t pp = size;
|
||||
for (int i = 0; i < rank - 1; ++i) {
|
||||
pp /= dim[i];
|
||||
ret[i] = sm / pp;
|
||||
sm %= pp;
|
||||
}
|
||||
ret[rank - 1] = sm;
|
||||
}
|
||||
/*__device__ void d_LinearToSplitSubrankLimSz(size_t& part0, size_t& part1, size_t lin, int* dim, int rank, int rankA, size_t size, size_t sizeA) {
|
||||
size_t sm = lin;
|
||||
size_t pp = size;
|
||||
size_t s = 0;
|
||||
size_t p = sizeA;
|
||||
int ret;// = new int[rank];
|
||||
for (int i = 0; i < rank; ++i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
p /= dim[i];
|
||||
s += ret * p;
|
||||
|
||||
sm %= pp;
|
||||
if (i == rankA - 1) {
|
||||
part0 = s;
|
||||
s = 0;
|
||||
p = size / sizeA;
|
||||
}
|
||||
|
||||
}
|
||||
part1 = s;
|
||||
|
||||
}*/
|
||||
__device__ void d_LinearToSplitSubrankLimSz(size_t& part0, size_t& part1, size_t lin, int* dim, int rank, int rankA, size_t size, size_t sizeA) {
|
||||
size_t sm = lin;
|
||||
size_t pp = size;
|
||||
size_t s = 0;
|
||||
size_t p = sizeA;
|
||||
int ret;// = new int[rank];
|
||||
int i;
|
||||
for (i = 0; i < rankA; ++i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
p /= dim[i];
|
||||
s += ret * p;
|
||||
|
||||
sm %= pp;
|
||||
|
||||
}
|
||||
part0 = s;
|
||||
s = 0;
|
||||
p = size / sizeA;//sizeB
|
||||
for (; i < rank; ++i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
p /= dim[i];
|
||||
s += ret * p;
|
||||
|
||||
sm %= pp;
|
||||
|
||||
}
|
||||
|
||||
part1 = s;
|
||||
|
||||
}
|
||||
__device__ void d_LinearToSplitSubrankLimSzEnd(size_t& part0, size_t& part1, size_t lin, int* dim, int rank, int rankA, size_t size, size_t sizeA) {
|
||||
size_t sm = lin;
|
||||
size_t pp = size;
|
||||
size_t s = 0;
|
||||
size_t p = sizeA;
|
||||
int ret;// = new int[rank];
|
||||
for (int i = rank - 1; i >= 0; --i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
p /= dim[i];
|
||||
s += ret * p;
|
||||
|
||||
sm %= pp;
|
||||
if (i == rankA) {
|
||||
part1 = s;
|
||||
s = 0;
|
||||
p = size / sizeA;
|
||||
}
|
||||
|
||||
}
|
||||
part0 = s;
|
||||
|
||||
}
|
||||
|
||||
|
||||
__device__ void d_subArray(int* dst, int* src, int debDst, int finDst, int debSrc) {
|
||||
for (int i = debDst; i < finDst; i++) {
|
||||
dst[i] = src[i + debSrc];
|
||||
}
|
||||
}
|
||||
|
||||
template<typename T>
|
||||
__global__ void d_prodTensor(T* C, int* dimC, int rankC, size_t size, T* A, int* dimA, int rankA, size_t sizeA, T* B, int* dimB, int rankB) {
|
||||
size_t lin0, lin1;
|
||||
|
||||
size_t i = threadIdx.x + blockIdx.x * blockDim.x;
|
||||
if (i < size) {
|
||||
d_LinearToSplitSubrankLimSz(lin0, lin1, i, dimC, rankC, rankA, size, sizeA);
|
||||
|
||||
C[i] = A[lin0] * B[lin1];
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
template __global__ void d_prodTensor<float>(float* C, int* dimC, int rankC, size_t size, float* A, int* dimA, int rankA, size_t sizeA, float* B, int* dimB, int rankB);
|
||||
|
||||
template<typename T>
|
||||
__global__ void d_prodTensorEnd(T* C, int* dimC, int rankC, size_t size, T* A, int* dimA, int rankA, size_t sizeA, T* B, int* dimB, int rankB) {
|
||||
size_t lin0, lin1;
|
||||
|
||||
size_t i = threadIdx.x + blockIdx.x * blockDim.x;
|
||||
if (i < size) {
|
||||
d_LinearToSplitSubrankLimSzEnd(lin0, lin1, i, dimC, rankC, rankA, size, sizeA);
|
||||
|
||||
C[i] = A[lin0] * B[lin1];
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
template __global__ void d_prodTensorEnd<float>(float* C, int* dimC, int rankC, size_t size, float* A, int* dimA, int rankA, size_t sizeA, float* B, int* dimB, int rankB);
|
||||
|
||||
__device__ void d_minReverse(int* dim, int& rank, const int* dim0, int rank0, const int* dim1, int rank1, bool& rev) {
|
||||
if (rank0 > rank1) {
|
||||
rank = rank1;
|
||||
for (int i = 0; i < rank1; ++i) dim[i] = dim1[i];
|
||||
rev = true;
|
||||
}
|
||||
else if (rank0 < rank1) {
|
||||
rank = rank0;
|
||||
for (int i = 0; i < rank1; ++i) dim[i] = dim0[i];
|
||||
rev = false;
|
||||
}
|
||||
else {// rank0 == rank1
|
||||
rank = rank0;
|
||||
for (int i = 0; i < rank0; i++) {
|
||||
if (dim[i] > dim1[rank1 - 1 - i]) dim[i] = dim1[rank1 - 1 - i];
|
||||
else dim[i] = dim0[i];
|
||||
}
|
||||
rev = false;
|
||||
}
|
||||
}
|
||||
|
||||
__device__ void d_reverseArray(int* arr, int sz) {
|
||||
int* tmp;
|
||||
//tmp = (int*)malloc(sz * sizeof(int));
|
||||
|
||||
tmp = new int[sz];
|
||||
if (tmp == NULL) {
|
||||
size_t limit = 0;
|
||||
cudaDeviceGetLimit(&limit, cudaLimitStackSize);
|
||||
printf("cudaLimitStackSize: %u | %d (%d) %d | \n", (unsigned)limit, blockIdx.x, blockDim.x, threadIdx.x);
|
||||
cudaDeviceGetLimit(&limit, cudaLimitPrintfFifoSize);
|
||||
printf("cudaLimitPrintfFifoSize: %u | %d (%d) %d | \n", (unsigned)limit, blockIdx.x, blockDim.x, threadIdx.x);
|
||||
cudaDeviceGetLimit(&limit, cudaLimitMallocHeapSize);
|
||||
printf("cudaLimitMallocHeapSize: %u | %d (%d) %d | \n", (unsigned)limit, blockIdx.x, blockDim.x, threadIdx.x);
|
||||
|
||||
printf("error Allocation in tmp = (int*)malloc(sz * sizeof(int)); | | ");
|
||||
}int i = 0;
|
||||
for (; i < sz / 2; i++) {
|
||||
tmp[i] = arr[i];
|
||||
arr[i] = arr[sz - 1 - i];
|
||||
}
|
||||
for (; i < sz; i++) {
|
||||
arr[i] = tmp[sz - 1 - i];
|
||||
}
|
||||
//free(tmp);
|
||||
delete[]tmp;
|
||||
}
|
||||
|
||||
__device__ int d_min(int a, int b) {
|
||||
if (a < b) return a;
|
||||
return b;
|
||||
}
|
||||
|
||||
__device__ void d_concatArray(int* dst, int* src0, int* src1, int debDst, int debSrc0, int finSrc0, int debSrc1, int finSrc1) {
|
||||
int i = debDst;
|
||||
for (int j = debSrc0; j < finSrc0; j++) {
|
||||
dst[i++] = src0[j];
|
||||
}
|
||||
for (int j = debSrc1; j < finSrc1; j++) {
|
||||
dst[i++] = src1[j];
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
__device__ void d_ConcatLinearToSplitSubrankLimSz(size_t& part0, size_t& part1, size_t lin, int* dim, int rank, int rankA, int rankB, size_t size, size_t sizeA, size_t sizeB, int* dM, int dMrank, size_t dMsize, int ind) {
|
||||
size_t sm = lin;
|
||||
size_t pp = size;
|
||||
size_t s = 0;
|
||||
size_t p = sizeA;
|
||||
//size_t sz_dA = sizeA / dMsize;
|
||||
int rankdA = rankA - dMrank;
|
||||
|
||||
int ret;
|
||||
int i;
|
||||
for (i = 0; i < rankdA; ++i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
p /= dim[i];
|
||||
s += ret * p;
|
||||
sm %= pp;
|
||||
}
|
||||
size_t s1 = 0;
|
||||
|
||||
size_t pb = sizeB / dMsize;
|
||||
for (; i < rank; ++i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
pb /= dim[i];
|
||||
s1 += ret * pb;
|
||||
sm %= pp;
|
||||
}
|
||||
|
||||
size_t smd = ind;
|
||||
size_t ppb = dMsize;
|
||||
//size_t pb = size / sz_dA;
|
||||
pb = sizeB;
|
||||
p = dMsize;
|
||||
for (int j = 0;j < dMrank;j++) {
|
||||
ppb /= dM[j];
|
||||
ret = smd / ppb;
|
||||
p /= dM[j];
|
||||
s += ret * p;
|
||||
pb /= dM[j];
|
||||
s1 += ret * pb;
|
||||
smd %= ppb;
|
||||
}
|
||||
//pp = size / sz_dA;
|
||||
part0 = s;
|
||||
part1 = s1;
|
||||
}
|
||||
|
||||
__device__ void d_SplitLineardToSubrank(size_t& part0, size_t& part1, size_t lin, int* dim, int rank, int rankA, int rankB, size_t size, size_t sizeA, size_t sizeB, int* dM, int dMrank, size_t dMsize) {
|
||||
size_t sm = lin;
|
||||
size_t pp = size;
|
||||
size_t s = 0;
|
||||
size_t p = sizeA;
|
||||
//size_t sz_dA = sizeA / dMsize;
|
||||
int rankdA = rankA - dMrank;
|
||||
|
||||
int ret;
|
||||
int i;
|
||||
for (i = 0; i < rankdA; ++i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
p /= dim[i];
|
||||
s += ret * p;
|
||||
sm %= pp;
|
||||
}
|
||||
size_t s1 = 0;
|
||||
|
||||
size_t pb = sizeB / dMsize;
|
||||
for (; i < rank; ++i) {
|
||||
pp /= dim[i];
|
||||
ret = sm / pp;
|
||||
pb /= dim[i];
|
||||
s1 += ret * pb;
|
||||
sm %= pp;
|
||||
}
|
||||
part0 = s;
|
||||
part1 = s1;
|
||||
}
|
||||
|
||||
|
||||
__device__ void d_UnionConcatLinearSplitedSubrank(size_t& part0, size_t& part1, size_t p0, size_t p1, size_t size, size_t sizeB, int* dM, int dMrank, size_t dMsize, int ind) {
|
||||
size_t s = p0;
|
||||
size_t s1 = p1;
|
||||
int ret;
|
||||
size_t smd = ind;
|
||||
size_t ppb = dMsize;
|
||||
//size_t pb = size / sz_dA;
|
||||
size_t pb = sizeB;
|
||||
size_t p = dMsize;
|
||||
for (int j = 0;j < dMrank;j++) {
|
||||
ppb /= dM[j];
|
||||
ret = smd / ppb;
|
||||
p /= dM[j];
|
||||
s += ret * p;
|
||||
pb /= dM[j];
|
||||
s1 += ret * pb;
|
||||
smd %= ppb;
|
||||
}
|
||||
//pp = size / sz_dA;
|
||||
part0 = s;
|
||||
part1 = s1;
|
||||
}
|
||||
|
||||
template<typename T>
|
||||
__global__ void d_TensorContractnReverseProd(T* C, int* dimC, int rankC, size_t sizeC, T* A, int rankA, size_t sizeA, T* B, int rankB, size_t sizeB, int* dM, int dMrank, size_t dMsize) {
|
||||
|
||||
size_t p0, p1;
|
||||
size_t lin0, lin1;
|
||||
|
||||
|
||||
//size_t i = threadIdx.x + blockIdx.x * blockDim.x;
|
||||
size_t i = d_getGlobalIdx_1D_1D();
|
||||
|
||||
if (i < sizeC) {
|
||||
|
||||
d_SplitLineardToSubrank(p0, p1, i, dimC, rankC, rankA, rankB, sizeC, sizeA, sizeB, dM, dMrank, dMsize);
|
||||
|
||||
C[i] = 0;
|
||||
for (size_t k = 0; k < dMsize; k++) {
|
||||
|
||||
d_UnionConcatLinearSplitedSubrank(lin0, lin1, p0, p1, sizeC, sizeB, dM, dMrank, dMsize, k);
|
||||
|
||||
//d_ConcatLinearToSplitSubrankLimSz(lin0, lin1, i, dimC, rankC, rankA, rankB, sizeC, sizeA, sizeB, dM, dMrank, dMsize, k);
|
||||
|
||||
C[i] += A[lin0] * B[lin1];
|
||||
}
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
template
|
||||
__global__ void d_TensorContractnReverseProd<float>(float* C, int* dimC, int rankC, size_t size, float* A, int rankA, size_t sizeA, float* B, int rankB, size_t sizeB, int* dM, int dMrank, size_t dMsize);
|
||||
|
||||
__device__ void d_LinearTransformCoord(size_t& dst, size_t src, int* inversePerm, size_t sizeA, int rankDst, int rankSrc, int* dDst, int* dSrc) {
|
||||
size_t sm = src;
|
||||
size_t pp = sizeA;
|
||||
size_t s = 0;
|
||||
size_t p = 1;
|
||||
int ret;// = new int[rank];
|
||||
int i, j;
|
||||
for (i = 0; i < rankSrc; ++i) {
|
||||
pp /= dSrc[i];
|
||||
ret = sm / pp;
|
||||
p = 1;
|
||||
for (j = inversePerm[i] + 1; j < rankDst;j++) {
|
||||
p *= dDst[j];
|
||||
}
|
||||
s += ret * p;
|
||||
|
||||
sm %= pp;
|
||||
|
||||
}
|
||||
dst = s;
|
||||
if (s > sizeA) printf("I have a problem in LinearTransformCoord: s:%ld siez:%ld \n", s, sizeA);
|
||||
|
||||
}
|
||||
|
||||
template<typename T>
|
||||
__global__ void d_PermLinearTransformCoord(T* C, int* dimC, int rankC, size_t sizeC, T* A, int* dimA, int rankA, size_t sizeA, int* invPerm) {
|
||||
|
||||
//size_t i = threadIdx.x + blockIdx.x * blockDim.x;
|
||||
size_t i = d_getGlobalIdx_1D_1D();
|
||||
|
||||
if (i < sizeC) {
|
||||
//printf("<i:%*ld ", 3, i);
|
||||
|
||||
size_t img = 0;
|
||||
//printf("<i:%*ld, img:%*ld\n", 3, i, 3, img);
|
||||
d_LinearTransformCoord(img, i, invPerm, sizeA, rankC, rankA, dimC, dimA);
|
||||
//img = d_LinearTransformCoord(i, invPerm, sizeC, dimC, dimA, rankC);
|
||||
|
||||
if (img < sizeC)
|
||||
C[img] = A[i];
|
||||
else {
|
||||
printf("something wrong in device: i:%ld , s:%ld\n", i, img);
|
||||
}
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
template
|
||||
__global__ void d_PermLinearTransformCoord<float>(float* C, int* dimC, int rankC, size_t size, float* A, int* dimA, int rankA, size_t sizeA, int* invPerm);
|
||||
|
||||
Reference in New Issue
Block a user