def test_matrix_multiply():
shapeX = (500, 700)
shapeY = (700, 1000)
shapeZ = (500, 1000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, False, shapeY, False, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(x, y), z, rtol=1e-5)
shapeX = (1000, 500)
shapeY = (2000, 500)
shapeZ = (1000, 2000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, False, shapeY, True, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(x, np.transpose(y)), z, rtol=1e-5)
shapeX = (500, 1000)
shapeY = (500, 2000)
shapeZ = (1000, 2000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, True, shapeY, False, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(np.transpose(x), y), z, rtol=1e-5)
shapeX = (500, 1000)
shapeY = (2000, 500)
shapeZ = (1000, 2000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, True, shapeY, True, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(np.transpose(x), np.transpose(y)), z, rtol=1e-5)
def test_matrix_multiply():
shapeX = (500, 700)
shapeY = (700, 1000)
shapeZ = (500, 1000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, False, shapeY, False, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(x, y), z, rtol=1e-5)
shapeX = (1000, 500)
shapeY = (2000, 500)
shapeZ = (1000, 2000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, False, shapeY, True, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(x, np.transpose(y)), z, rtol=1e-5)
shapeX = (500, 1000)
shapeY = (500, 2000)
shapeZ = (1000, 2000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, True, shapeY, False, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(np.transpose(x), y), z, rtol=1e-5)
shapeX = (500, 1000)
shapeY = (2000, 500)
shapeZ = (1000, 2000)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, True, shapeY, True, tgt, tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(np.transpose(x), np.transpose(y)), z, rtol=1e-5)
def test_GEMM_performance():
pf_tgt_host = "llvm"
pf_tgt = "llvm -mcpu=core-avx2"
dtype = "float32"
pf_ctx = tvm.context(tgt, 0)
M = 1000
N = 256
K = 784
transpose_A = False
transpose_B = False
gemm_func = tvm_op.make_matrix_mul(
(M, K), transpose_A, (K, N), transpose_B, pf_tgt, pf_tgt_host, "GEMM")
# Random generated tensor for testing
a = tvm.nd.array(np.random.rand(M, K).astype(dtype), pf_ctx)
b = tvm.nd.array(np.random.rand(K, N).astype(dtype), pf_ctx)
c = tvm.nd.array(np.zeros((M, N), dtype=dtype), ctx)
np_repeat = 100
np_runing_time = timeit.timeit(setup='import numpy as np\n'
'M = ' + str(M) + '\n'
'K = ' + str(K) + '\n'
'N = ' + str(N) + '\n'
'dtype = "float32"\n'
'a = np.random.rand(M, K).astype(dtype)\n'
'b = np.random.rand(K, N).astype(dtype)\n',
stmt='answer = np.dot(a, b)',
number=np_repeat)
print("Numpy running time: %f" % (np_runing_time / np_repeat))
answer = np.dot(a.asnumpy(), b.asnumpy())
gemm_func(a, b, c)
tvm.testing.assert_allclose(c.asnumpy(), answer, rtol=1e-5)
evaluator = gemm_func.time_evaluator(gemm_func.entry_name,
pf_ctx,
number=1)
print('Baseline: %f' % evaluator(a, b, c).mean)
def test_matrix_multiply_time():
M = 512
K = 2048
N = 4096
shapeX = (M, K)
shapeY = (K, N)
shapeZ = (M, N)
x = np.random.uniform(0, 10, size=shapeX).astype(dtype)
y = np.random.uniform(0, 10, size=shapeY).astype(dtype)
z = np.zeros(shapeZ).astype(dtype)
arr_x = tvm.nd.array(x, ctx=ctx)
arr_y = tvm.nd.array(y, ctx=ctx)
arr_z = tvm.nd.array(z, ctx=ctx)
matrix_mul = tvm_op.make_matrix_mul(shapeX, False, shapeY, False, tgt,
tgt_host, "matrix_mul")
matrix_mul(arr_x, arr_y, arr_z)
z = arr_z.asnumpy()
np.testing.assert_allclose(np.dot(x, y), z, rtol=1e-5)
np_repeat = 100
np_runing_time = timeit.timeit(
setup='import numpy\n'
'M = ' + str(M) + '\n'
'K = ' + str(K) + '\n'
'N = ' + str(N) + '\n'
'dtype = "float32"\n'
'a = numpy.random.rand(M, K).astype(dtype)\n'
'b = numpy.random.rand(K, N).astype(dtype)\n',
stmt='answer = numpy.dot(a, b)',
number=np_repeat)
print("Numpy running time: %f" % (np_runing_time / np_repeat))
evaluator = matrix_mul.time_evaluator(matrix_mul.entry_name,
ctx,
number=10)
print('Opt1: %f' % evaluator(arr_x, arr_y, arr_z).mean)