test_cublas3 Namespace Reference

Functions
def	test_float_gemm ()
def	test_float_symm ()
def	test_float_dgmm ()
def	test_float_syrk ()
def	test_float_syrkx ()
def	test_float_geam ()
def	test_double_gemm ()
def	test_double_symm ()
def	test_double_dgmm ()
def	test_double_syrk ()
def	test_double_syrkx ()
def	test_double_geam ()

Variables
int	dec = 4
int	prec = 10**-dec
int	sizem = 128
int	sizen = 256
int	sizek = 512
	seed = np.int32(time.perf_counter() * 1e3)
	c = ch.context.get_instance()

Function Documentation

test_double_dgmm()

def test_cublas3.test_double_dgmm

(

)

Definition at line 447 of file test_cublas3.py.

def test_double_dgmm():
    #function dgmm
    #testing: C=A.d
    # C,A matrices, d vector (diagonal matrix as a vector)
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_double(c, np.random.randn(sizek, sizek))
    Vectx = ch.obj_double(c, np.random.randn(sizek))
 
    # matA.random_host(seed, 'U')
    # Vectx.random_host(seed + 2, 'U')
    A = np.array(matA)
    x = np.array(Vectx)
 
    #matrices product
    t1 = time.perf_counter()
    matC = matA.dgmm(Vectx)
    t2 = time.perf_counter()
    C = A * x
    t3 = time.perf_counter()
 
    print("")
    print("test dgmm:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    npt.assert_almost_equal(C, res, decimal=2 * dec)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
    #     d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
    # print(res.item(M))
    # print(C.item(M))
 
    # npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=2*dec)
 
 

test_double_geam()

def test_cublas3.test_double_geam

(

)

Definition at line 596 of file test_cublas3.py.

def test_double_geam():
 
    #function geam
    #testing: C=A.B
    #A,B matrices
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_double(c, np.random.randn(sizem, sizen))
    matB = ch.obj_double(c, np.random.randn(sizem, sizen))
 
    # matA.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
 
    A = np.array(matA)
    B = np.array(matB)
 
    #matrices product
    t1 = time.perf_counter()
    C = A + B
    t2 = time.perf_counter()
    matC = matA.geam(matB, beta=1)
    t3 = time.perf_counter()
 
    print("")
    print("test geam:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #testing result
 
    npt.assert_array_almost_equal(C, np.array(matC), decimal=2 * dec)

test_double_gemm()

def test_cublas3.test_double_gemm

(

)

Definition at line 326 of file test_cublas3.py.

def test_double_gemm():
 
    # function gemm
    #testing: C=A.B+C
    #A,B,C matrices
 
    #generating random matrices A,B,C and associated CarmaObj
 
    # np.random.seed(seed)
    # A = A.dot(A.T)
    # B = B.dot(B.T)
 
    matA = ch.obj_double(c, np.random.randn(sizem, sizek))
    matAT = ch.obj_double(c, np.random.randn(sizek, sizem))
    matB = ch.obj_double(c, np.random.randn(sizek, sizen))
    matBT = ch.obj_double(c, np.random.randn(sizen, sizek))
    matC = ch.obj_double(c, np.random.randn(sizem, sizen))
    matC2 = ch.obj_double(c, np.random.randn(sizem, sizen))
    matC3 = ch.obj_double(c, np.random.randn(sizem, sizen))
 
    # matA.random_host(seed, 'U')
    # matAT.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
    # matBT.random_host(seed + 2, 'U')
    # matC.random_host(seed + 3, 'U')
    # matC2.random_host(seed + 3, 'U')
    # matC3.random_host(seed + 3, 'U')
 
    A = np.array(matA)
    AT = np.array(matAT)
    B = np.array(matB)
    BT = np.array(matBT)
    C = np.array(matC)
    C2 = np.array(matC2)
    C3 = np.array(matC3)
 
    alpha = 2
    beta = 1
 
    #matrices product
    matA.gemm(matB, 'n', 'n', alpha, matC, beta)
    matAT.gemm(matB, 't', 'n', alpha, matC2, beta)
    matAT.gemm(matBT, 't', 't', alpha, matC3, beta)
    matC4 = matA.gemm(matB, 'n', 'n', alpha)
    matC5 = matAT.gemm(matB, 't', 'n', alpha)
    matC6 = matAT.gemm(matBT, 't', 't', alpha)
 
    C = alpha * A.dot(B) + beta * C
    C2 = alpha * AT.T.dot(B) + beta * C2
    C3 = alpha * AT.T.dot(BT.T) + beta * C3
    C4 = alpha * A.dot(B)
    C5 = alpha * AT.T.dot(B)
    C6 = alpha * AT.T.dot(BT.T)
 
    npt.assert_array_almost_equal(C, np.array(matC), decimal=2 * dec - 1)
    npt.assert_array_almost_equal(C2, np.array(matC2), decimal=2 * dec - 1)
    npt.assert_array_almost_equal(C3, np.array(matC3), decimal=2 * dec - 1)
    npt.assert_array_almost_equal(C4, np.array(matC4), decimal=2 * dec - 1)
    npt.assert_array_almost_equal(C5, np.array(matC5), decimal=2 * dec - 1)
    npt.assert_array_almost_equal(C6, np.array(matC6), decimal=2 * dec - 1)
 
 

test_double_symm()

def test_cublas3.test_double_symm

(

)

Definition at line 388 of file test_cublas3.py.

def test_double_symm():
    #function symm
    #testing: C=A.B+C
    #A ssymetric matrix, B,C matrices
    A = np.random.randn(sizek, sizek)
    B = np.random.randn(sizek, sizen)
    C = np.random.randn(sizek, sizen)
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_double(c, A)  #np.random.randn(sizek, sizek)))
    matB = ch.obj_double(c, B)  #np.random.randn(sizek, sizen)))
    matC = ch.obj_double(c, C)  #np.random.randn(sizek, sizen)))
 
    # matA.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
    # matC.random_host(seed + 3, 'U')
 
    #A symetric
    A = np.array(matA)
    A = (A + A.T) / 2
    matA.host2device(A)
    B = np.array(matB)
    C = np.array(matC)
 
    #matrices multiplication
    t1 = time.perf_counter()
    matA.symm(matB, 1, matC, 1)
    t2 = time.perf_counter()
    C = A.dot(B) + C
    t3 = time.perf_counter()
 
    matC2 = matA.symm(matB)
    C2 = A.dot(B)
 
    print("")
    print("test symm:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    npt.assert_almost_equal(C, res, decimal=2 * dec - 1)
    npt.assert_almost_equal(C2, np.array(matC2), decimal=2 * dec - 1)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
 
 
# d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
# print(res.item(M))
# print(C.item(M))
 
# npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=2*dec)
 
 

test_double_syrk()

def test_cublas3.test_double_syrk

(

)

Definition at line 490 of file test_cublas3.py.

def test_double_syrk():
    #function syrk
    #testing: C=A.transpose(A)+C
    #A matrix, C symetric matrix
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_double(c, np.random.randn(sizen, sizek))
    matC = ch.obj_double(c, np.random.randn(sizen, sizen))
 
    # matA.random_host(seed, 'U')
    # matC.random_host(seed + 2, 'U')
 
    A = np.array(matA)
    C = np.array(matC)
    #syrk: C matrix is symetric
    C = (C + C.T) / 2
 
    matC.host2device(C)
 
    #matrices product
    t1 = time.perf_counter()
    matA.syrk(matC=matC, beta=1)
    t2 = time.perf_counter()
    C = A.dot(A.T) + C
    t3 = time.perf_counter()
 
    print("")
    print("test syrk:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    #only upper triangle is computed
    C = np.triu(C).flatten()
    res = np.triu(res).flatten()
    npt.assert_almost_equal(C, res, decimal=2 * dec - 1)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
    #     d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
    # print(res.item(M))
    # print(C.item(M))
 
    # npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=2*dec)
 
 

test_double_syrkx()

def test_cublas3.test_double_syrkx

(

)

Definition at line 541 of file test_cublas3.py.

def test_double_syrkx():
    #function syrkx
    #testing: C=A.transpose(B)+C
    #A matrix, C symetric matrix
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_double(c, np.random.randn(sizen, sizek))
    matB = ch.obj_double(c, np.random.randn(sizen, sizek))
    matC = ch.obj_double(c, np.random.randn(sizen, sizen))
 
    # matA.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
    # matC.random_host(seed + 3, 'U')
 
    A = np.array(matA)
    B = np.array(matB)
    C = np.array(matC)
 
    #C is symetric
    C = np.dot(C, C.T)
 
    matC.host2device(C)
 
    #matrices product
    t1 = time.perf_counter()
    matA.syrkx(matB, alpha=1, matC=matC, beta=1)
    t2 = time.perf_counter()
    C = A.dot(B.T) + C.T
    t3 = time.perf_counter()
 
    print("")
    print("test syrkx:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    #only upper triangle is computed
    res = np.triu(res)
    C = np.triu(C)
    npt.assert_almost_equal(C, res, decimal=2 * dec - 1)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
    #     d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
    # print(res.item(M))
    # print(C.item(M))
 
    # npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=2*dec)
 
 

test_float_dgmm()

def test_cublas3.test_float_dgmm

(

)

Definition at line 143 of file test_cublas3.py.

def test_float_dgmm():
    #function dgmm
    #testing: C=A.d
    # C,A matrices, d vector (diagonal matrix as a vector)
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_float(c, np.random.randn(sizek, sizek))
    Vectx = ch.obj_float(c, np.random.randn(sizek))
 
    # matA.random_host(seed, 'U')
    # Vectx.random_host(seed + 2, 'U')
    A = np.array(matA)
    x = np.array(Vectx)
 
    #matrices product
    t1 = time.perf_counter()
    matC = matA.dgmm(Vectx)
    t2 = time.perf_counter()
    C = A * x
    t3 = time.perf_counter()
 
    print("")
    print("test dgmm:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    npt.assert_almost_equal(C, res, decimal=dec)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
    #     d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
    # print(res.item(M))
    # print(C.item(M))
 
    # npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=dec)
 
 

test_float_geam()

def test_cublas3.test_float_geam

(

)

Definition at line 292 of file test_cublas3.py.

def test_float_geam():
 
    #function geam
    #testing: C=A.B
    #A,B matrices
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_float(c, np.random.randn(sizem, sizen))
    matB = ch.obj_float(c, np.random.randn(sizem, sizen))
 
    # matA.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
 
    A = np.array(matA)
    B = np.array(matB)
 
    #matrices product
    t1 = time.perf_counter()
    C = A + B
    t2 = time.perf_counter()
    matC = matA.geam(matB, beta=1)
    t3 = time.perf_counter()
 
    print("")
    print("test geam:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #testing result
 
    npt.assert_array_almost_equal(C, np.array(matC), decimal=dec)
 
 

test_float_gemm()

def test_cublas3.test_float_gemm

(

)

Definition at line 22 of file test_cublas3.py.

def test_float_gemm():
 
    # function gemm
    #testing: C=A.B+C
    #A,B,C matrices
 
    #generating random matrices A,B,C and associated CarmaObj
 
    # np.random.seed(seed)
    # A = A.dot(A.T)
    # B = B.dot(B.T)
 
    matA = ch.obj_float(c, np.random.randn(sizem, sizek))
    matAT = ch.obj_float(c, np.random.randn(sizek, sizem))
    matB = ch.obj_float(c, np.random.randn(sizek, sizen))
    matBT = ch.obj_float(c, np.random.randn(sizen, sizek))
    matC = ch.obj_float(c, np.random.randn(sizem, sizen))
    matC2 = ch.obj_float(c, np.random.randn(sizem, sizen))
    matC3 = ch.obj_float(c, np.random.randn(sizem, sizen))
 
    # matA.random_host(seed, 'U')
    # matAT.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
    # matBT.random_host(seed + 2, 'U')
    # matC.random_host(seed + 3, 'U')
    # matC2.random_host(seed + 3, 'U')
    # matC3.random_host(seed + 3, 'U')
 
    A = np.array(matA)
    AT = np.array(matAT)
    B = np.array(matB)
    BT = np.array(matBT)
    C = np.array(matC)
    C2 = np.array(matC2)
    C3 = np.array(matC3)
 
    alpha = 2
    beta = 1
 
    #matrices product
    matA.gemm(matB, 'n', 'n', alpha, matC, beta)
    matAT.gemm(matB, 't', 'n', alpha, matC2, beta)
    matAT.gemm(matBT, 't', 't', alpha, matC3, beta)
    matC4 = matA.gemm(matB, 'n', 'n', alpha)
    matC5 = matAT.gemm(matB, 't', 'n', alpha)
    matC6 = matAT.gemm(matBT, 't', 't', alpha)
 
    C = alpha * A.dot(B) + beta * C
    C2 = alpha * AT.T.dot(B) + beta * C2
    C3 = alpha * AT.T.dot(BT.T) + beta * C3
    C4 = alpha * A.dot(B)
    C5 = alpha * AT.T.dot(B)
    C6 = alpha * AT.T.dot(BT.T)
 
    npt.assert_array_almost_equal(C, np.array(matC), decimal=dec - 1)
    npt.assert_array_almost_equal(C2, np.array(matC2), decimal=dec - 1)
    npt.assert_array_almost_equal(C3, np.array(matC3), decimal=dec - 1)
    npt.assert_array_almost_equal(C4, np.array(matC4), decimal=dec - 1)
    npt.assert_array_almost_equal(C5, np.array(matC5), decimal=dec - 1)
    npt.assert_array_almost_equal(C6, np.array(matC6), decimal=dec - 1)
 
 

test_float_symm()

def test_cublas3.test_float_symm

(

)

Definition at line 84 of file test_cublas3.py.

def test_float_symm():
    #function symm
    #testing: C=A.B+C
    #A ssymetric matrix, B,C matrices
    A = np.random.randn(sizek, sizek)
    B = np.random.randn(sizek, sizen)
    C = np.random.randn(sizek, sizen)
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_float(c, A)  #np.random.randn(sizek, sizek)))
    matB = ch.obj_float(c, B)  #np.random.randn(sizek, sizen)))
    matC = ch.obj_float(c, C)  #np.random.randn(sizek, sizen)))
 
    # matA.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
    # matC.random_host(seed + 3, 'U')
 
    #A symetric
    A = np.array(matA)
    A = (A + A.T) / 2
    matA.host2device(A)
    B = np.array(matB)
    C = np.array(matC)
 
    #matrices multiplication
    t1 = time.perf_counter()
    matA.symm(matB, 1, matC, 1)
    t2 = time.perf_counter()
    C = A.dot(B) + C
    t3 = time.perf_counter()
 
    matC2 = matA.symm(matB)
    C2 = A.dot(B)
 
    print("")
    print("test symm:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    npt.assert_almost_equal(C, res, decimal=dec - 1)
    npt.assert_almost_equal(C2, np.array(matC2), decimal=dec - 1)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
 
 
# d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
# print(res.item(M))
# print(C.item(M))
 
# npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=dec)
 
 

test_float_syrk()

def test_cublas3.test_float_syrk

(

)

Definition at line 186 of file test_cublas3.py.

def test_float_syrk():
    #function syrk
    #testing: C=A.transpose(A)+C
    #A matrix, C symetric matrix
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_float(c, np.random.randn(sizen, sizek))
    matC = ch.obj_float(c, np.random.randn(sizen, sizen))
 
    # matA.random_host(seed, 'U')
    # matC.random_host(seed + 2, 'U')
 
    A = np.array(matA)
    C = np.array(matC)
    #syrk: C matrix is symetric
    C = (C + C.T) / 2
 
    matC.host2device(C)
 
    #matrices product
    t1 = time.perf_counter()
    matA.syrk(matC=matC, beta=1)
    t2 = time.perf_counter()
    C = A.dot(A.T) + C
    t3 = time.perf_counter()
 
    print("")
    print("test syrk:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    #only upper triangle is computed
    C = np.triu(C).flatten()
    res = np.triu(res).flatten()
    npt.assert_almost_equal(C, res, decimal=dec - 1)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
    #     d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
    # print(res.item(M))
    # print(C.item(M))
 
    # npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=dec)
 
 

test_float_syrkx()

def test_cublas3.test_float_syrkx

(

)

Definition at line 237 of file test_cublas3.py.

def test_float_syrkx():
    #function syrkx
    #testing: C=A.transpose(B)+C
    #A matrix, C symetric matrix
 
    #generating random matrices and associated CarmaObj
    matA = ch.obj_float(c, np.random.randn(sizen, sizek))
    matB = ch.obj_float(c, np.random.randn(sizen, sizek))
    matC = ch.obj_float(c, np.random.randn(sizen, sizen))
 
    # matA.random_host(seed, 'U')
    # matB.random_host(seed + 2, 'U')
    # matC.random_host(seed + 3, 'U')
 
    A = np.array(matA)
    B = np.array(matB)
    C = np.array(matC)
 
    #C is symetric
    C = np.dot(C, C.T)
 
    matC.host2device(C)
 
    #matrices product
    t1 = time.perf_counter()
    matA.syrkx(matB, alpha=1, matC=matC, beta=1)
    t2 = time.perf_counter()
    C = A.dot(B.T) + C.T
    t3 = time.perf_counter()
 
    print("")
    print("test syrkx:")
    print("execution time (s)")
    print("python: ", t3 - t2)
    print("carma : ", t2 - t1)
 
    #test results
    res = np.array(matC)
 
    #only upper triangle is computed
    res = np.triu(res)
    C = np.triu(C)
    npt.assert_almost_equal(C, res, decimal=dec - 1)
 
    # M = np.argmax(np.abs(res - C))
    # d = 5
    # if (0 < np.abs(C.item(M))):
    #     d = 10**np.ceil(np.log10(np.abs(C.item(M))))
 
    # print(res.item(M))
    # print(C.item(M))
 
    # npt.assert_almost_equal(C.item(M) / d, res.item(M) / d, decimal=dec)
 
 

Variable Documentation

c

test_cublas3.c = ch.context.get_instance()

Definition at line 19 of file test_cublas3.py.

dec

int test_cublas3.dec = 4

Definition at line 6 of file test_cublas3.py.

prec

int test_cublas3.prec = 10**-dec

Definition at line 7 of file test_cublas3.py.

seed

test_cublas3.seed = np.int32(time.perf_counter() * 1e3)

Definition at line 13 of file test_cublas3.py.

sizek

int test_cublas3.sizek = 512

Definition at line 11 of file test_cublas3.py.

sizem

int test_cublas3.sizem = 128

Definition at line 9 of file test_cublas3.py.

sizen

int test_cublas3.sizen = 256

Definition at line 10 of file test_cublas3.py.