As mentioned in the comments, numpy functions in general cannot be used from pycuda kernel code (or CUDA kernel code, or numba cuda kernels).
CUBLAS offers a batched matrix inversion function, but it is not currently exposed in either pyculib cublas interface or scikit-cuda cublas interface.
We could proceed to implement our own interface (e.g. using python ctypes), but since its known that the matrices to be inverted are 4x4, I thought the suggestion in the comments from talonmies was an interesting one. Referring to the answer here, there is a fairly concise C code to do a direct inversion of a 4x4 matrix.
What follows first is a realization of this in CUDA. The function inv4x4 is an adaptation of the previous code, allotting 16 threads per matrix (one per matrix element) and using that code as a model. Each thread is responsible for computing one result matrix element. First we will compare it to CUBLAS matinvBatched for performance:
$ cat t411.cu
#include <iostream>
#include <cublas_v2.h>
#include <cstdlib>
// 4x4 matrix inversion
// https://stackoverflow.com/questions/1148309/inverting-a-4x4-matrix
// assumes warp size is 32
// assumes block size is multiple of warp size
// therefore assumes number of matrices to be inverted (n) is even
// 16 threads per matrix to invert
const unsigned block_size = 256;
typedef float mt;
#include <time.h>
#include <sys/time.h>
#define USECPSEC 1000000ULL
long long dtime_usec(unsigned long long start){
timeval tv;
gettimeofday(&tv, 0);
return ((tv.tv_sec*USECPSEC)+tv.tv_usec)-start;
}
__device__ unsigned pat[3][16];
const unsigned hpat[3][16] = {
{ 0x0EB51FA5, 0x1EB10FA1, 0x0E711F61, 0x1A710B61, 0x1EB40FA4, 0x0EB01FA0, 0x1E700F60, 0x0A701B60, 0x0DB41F94, 0x1DB00F90, 0x0D701F50, 0x19700B50, 0x1DA40E94, 0x0DA01E90, 0x1D600E50, 0x09601A50},
{ 0x1E790F69, 0x0E391F29, 0x1E350F25, 0x0A351B25, 0x0E781F68, 0x1E380F28, 0x0E341F24, 0x1A340B24, 0x1D780F58, 0x0D381F18, 0x1D340F14, 0x09341B14, 0x0D681E58, 0x1D280E18, 0x0D241E14, 0x19240A14},
{ 0x0A7D1B6D, 0x1A3D0B2D, 0x063D172D, 0x16390729, 0x1A7C0B6C, 0x0A3C1B2C, 0x163C072C, 0x06381728, 0x097C1B5C, 0x193C0B1C, 0x053C171C, 0x15380718, 0x196C0A5C, 0x092C1A1C, 0x152C061C, 0x05281618}};
__device__ unsigned getoff(unsigned &off){
unsigned ret = off & 0x0F;
off = off >> 4;
return ret;
}
const unsigned tmsk = 0xFFFFFFFF;
// in-place is acceptable i.e. out == in)
// T = float or double only
template <typename T>
__global__ void inv4x4(const T * __restrict__ in, T * __restrict__ out, const size_t n){
__shared__ T si[block_size];
size_t idx = threadIdx.x+blockDim.x*blockIdx.x;
if (idx < n*16){
si[threadIdx.x] = in[idx];
unsigned lane = threadIdx.x & 15;
unsigned sibase = threadIdx.x & 0x03F0;
__syncwarp();
unsigned off = pat[0][lane];
T a,b;
a = si[sibase + getoff(off)];
a *= si[sibase + getoff(off)];
a *= si[sibase + getoff(off)];
if (!getoff(off)) a = -a;
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
off = pat[1][lane];
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
off = pat[2][lane];
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
T det = si[sibase + (lane>>2)]*a;
det += __shfl_down_sync(tmsk, det, 4, 16); // first add
det += __shfl_down_sync(tmsk, det, 8, 16); // second add
det = __shfl_sync(tmsk, det, 0, 16); // broadcast
out[idx] = a / det;
}
}
size_t nr = 2048;
int main(int argc, char *argv[]){
if (argc > 1) nr = atoi(argv[1]);
const mt m1[] = {1.0, 1.0, 1.0, 0.0, 0.0, 3.0, 1.0, 2.0, 2.0, 3.0, 1.0, 0.0, 1.0, 0.0, 2.0, 1.0};
const mt i1[] = {-3.0, -0.5, 1.5, 1.0, 1.0, 0.25, -0.25, -0.5, 3.0, 0.25, -1.25, -0.5, -3.0, 0.0, 1.0, 1.0};
const mt m2[] = {1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0};
const mt i2[] = {1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0};
mt *h_d, *d_d;
h_d = (mt *)malloc(nr*2*16*sizeof(mt));
cudaMalloc(&d_d, nr*2*16*sizeof(mt));
cudaMemcpyToSymbol(pat, hpat, 3*16*sizeof(unsigned));
for (int i = 0; i < nr; i++){
memcpy(h_d+i*16*2, m1, sizeof(m1));
memcpy(h_d+i*16*2+16, m2, sizeof(m2));}
cudaMemcpy(d_d, h_d, nr*2*16*sizeof(mt), cudaMemcpyHostToDevice);
long long t = dtime_usec(0);
inv4x4<<<nr*2*16/block_size, block_size>>>(d_d, d_d, nr*2);
cudaDeviceSynchronize();
t = dtime_usec(t);
cudaMemcpy(h_d, d_d, nr*2*16*sizeof(mt), cudaMemcpyDeviceToHost);
for (int i = 0; i < 2; i++){
for (int j = 0; j < 16; j++) std::cout << h_d[i*16 + j] << ",";
std::cout << std::endl;
for (int j = 0; j < 16; j++) std::cout << ((i==0)?i1[j]:i2[j]) << ",";
std::cout << std::endl;}
std::cout << "kernel time: " << t << " microseconds" << std::endl;
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) std::cout << cudaGetErrorString(err) << std::endl;
//cublas
for (int i = 0; i < nr; i++){
memcpy(h_d+i*16*2, m1, sizeof(m1));
memcpy(h_d+i*16*2+16, m2, sizeof(m2));}
cudaMemcpy(d_d, h_d, nr*2*16*sizeof(mt), cudaMemcpyHostToDevice);
cublasHandle_t h;
cublasStatus_t cs = cublasCreate(&h);
if (cs != CUBLAS_STATUS_SUCCESS) std::cout << "cublas create error" << std::endl;
mt **A, **Ai, *Aid, **Ap, **Aip;
A = (mt **)malloc(nr*2*sizeof(mt *));
Ai = (mt **)malloc(nr*2*sizeof(mt *));
cudaMalloc(&Aid, nr*2*16*sizeof(mt));
cudaMalloc(&Ap, nr*2*sizeof(mt *));
cudaMalloc(&Aip, nr*2*sizeof(mt *));
for (int i = 0; i < nr*2; i++) A[i] = d_d + 16*i;
for (int i = 0; i < nr*2; i++) Ai[i] = Aid + 16*i;
cudaMemcpy(Ap, A, nr*2*sizeof(mt *), cudaMemcpyHostToDevice);
cudaMemcpy(Aip, Ai, nr*2*sizeof(mt *), cudaMemcpyHostToDevice);
int *info;
cudaMalloc(&info, nr*2*sizeof(int));
t = dtime_usec(0);
cs = cublasSmatinvBatched(h, 4, Ap, 4, Aip, 4, info, nr*2);
if (cs != CUBLAS_STATUS_SUCCESS) std::cout << "cublas matinv error" << std::endl;
cudaDeviceSynchronize();
t = dtime_usec(t);
cudaMemcpy(h_d, Aid, nr*2*16*sizeof(mt), cudaMemcpyDeviceToHost);
for (int i = 0; i < 2; i++){
for (int j = 0; j < 16; j++) std::cout << h_d[i*16 + j] << ",";
std::cout << std::endl;
for (int j = 0; j < 16; j++) std::cout << ((i==0)?i1[j]:i2[j]) << ",";
std::cout << std::endl;}
std::cout << "cublas time: " << t << " microseconds" << std::endl;
err = cudaGetLastError();
if (err != cudaSuccess) std::cout << cudaGetErrorString(err) << std::endl;
return 0;
}
$ nvcc -o t411 t411.cu -lcublas
$ ./t411
-3,-0.5,1.5,1,1,0.25,-0.25,-0.5,3,0.25,-1.25,-0.5,-3,-0,1,1,
-3,-0.5,1.5,1,1,0.25,-0.25,-0.5,3,0.25,-1.25,-0.5,-3,0,1,1,
1,0,0,0,0,1,0,0,0,0,1,0,0,0,0,1,
1,0,0,0,0,1,0,0,0,0,1,0,0,0,0,1,
kernel time: 49 microseconds
-3,-0.5,1.5,1,1,0.25,-0.25,-0.5,3,0.25,-1.25,-0.5,-3,0,1,1,
-3,-0.5,1.5,1,1,0.25,-0.25,-0.5,3,0.25,-1.25,-0.5,-3,0,1,1,
1,0,0,0,0,1,0,0,0,0,1,0,0,0,0,1,
1,0,0,0,0,1,0,0,0,0,1,0,0,0,0,1,
cublas time: 95 microseconds
$
We see that the code appears to provide the correct result for 2 test matrices inverted, and the overall time to invert 4096 matrices on a Tesla P100 is about 50us and is about 2x faster than CUBLAS. Note that I have not exhaustively tested this code.
What follows next is a simple pycuda implementation of a similar function. Here, for simplicity we are just inverting 2 matrices:
$ cat t10.py
import numpy as np
import pycuda.driver as cuda
from pycuda.compiler import SourceModule
import pycuda.autoinit
# kernel
kernel = SourceModule("""
__device__ unsigned getoff(unsigned &off){
unsigned ret = off & 0x0F;
off = off >> 4;
return ret;
}
const int block_size = 256;
const unsigned tmsk = 0xFFFFFFFF;
// in-place is acceptable i.e. out == in)
// T = float or double only
typedef float T;
__global__ void inv4x4(const T * __restrict__ in, T * __restrict__ out, const size_t n, const unsigned * __restrict__ pat){
__shared__ T si[block_size];
size_t idx = threadIdx.x+blockDim.x*blockIdx.x;
if (idx < n*16){
si[threadIdx.x] = in[idx];
unsigned lane = threadIdx.x & 15;
unsigned sibase = threadIdx.x & 0x03F0;
__syncwarp();
unsigned off = pat[lane];
T a,b;
a = si[sibase + getoff(off)];
a *= si[sibase + getoff(off)];
a *= si[sibase + getoff(off)];
if (!getoff(off)) a = -a;
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
off = pat[lane+16];
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
off = pat[lane+32];
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
b = si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
b *= si[sibase + getoff(off)];
if (getoff(off)) a += b;
else a -=b;
T det = si[sibase + (lane>>2)]*a;
det += __shfl_down_sync(tmsk, det, 4, 16); // first add
det += __shfl_down_sync(tmsk, det, 8, 16); // second add
det = __shfl_sync(tmsk, det, 0, 16); // broadcast
out[idx] = a / det;
}
}
""")
# python function for inverting 4x4 matrices
# n should be an even number
def gpuinv4x4(inp, n):
# internal constants not to be modified
hpat = ( 0x0EB51FA5, 0x1EB10FA1, 0x0E711F61, 0x1A710B61, 0x1EB40FA4, 0x0EB01FA0, 0x1E700F60, 0x0A701B60, 0x0DB41F94, 0x1DB00F90, 0x0D701F50, 0x19700B50, 0x1DA40E94, 0x0DA01E90, 0x1D600E50, 0x09601A50, 0x1E790F69, 0x0E391F29, 0x1E350F25, 0x0A351B25, 0x0E781F68, 0x1E380F28, 0x0E341F24, 0x1A340B24, 0x1D780F58, 0x0D381F18, 0x1D340F14, 0x09341B14, 0x0D681E58, 0x1D280E18, 0x0D241E14, 0x19240A14, 0x0A7D1B6D, 0x1A3D0B2D, 0x063D172D, 0x16390729, 0x1A7C0B6C, 0x0A3C1B2C, 0x163C072C, 0x06381728, 0x097C1B5C, 0x193C0B1C, 0x053C171C, 0x15380718, 0x196C0A5C, 0x092C1A1C, 0x152C061C, 0x05281618)
# Convert parameters into numpy array
inpd = np.array(inp, dtype=np.float32)
hpatd = np.array(hpat, dtype=np.uint32)
output = np.empty((n*16), dtype= np.float32)
# Get kernel function
matinv4x4 = kernel.get_function("inv4x4")
# Define b
