def dpnp_matrix_power_impl(a, n):
dpnp_lowering.ensure_dpnp("matrix_power")
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels.cpp#L42
Function declaration:
void dpnp_matmul_c(void* array1_in, void* array2_in, void* result1, size_t size_m,
size_t size_n, size_t size_k)
"""
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a, n):
if n < 0:
raise ValueError(
"n < 0 is not supported for np.linalg.matrix_power(a, n)")
if n == 0:
if PRINT_DEBUG:
print("dpnp implementation")
return np.identity(a.shape[0], a.dtype)
result = a
for idx in range(0, n - 1):
result = numba_dppy.dpnp.matmul(result, a)
return result
return dpnp_impl
def dpnp_det_impl(a):
name = "det"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels_linalg.cpp#L83
Function declaration:
void custom_det_c(void* array1_in, void* result1, size_t* shape, size_t ndim)
"""
sig = signature(ret_type, types.voidptr, types.voidptr, types.voidptr,
types.intp)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
n = a.shape[-1]
if a.shape[-2] != n:
raise ValueError("Input array must be square.")
dpnp_ext._check_finite_matrix(a)
if a.ndim == 2:
out = np.empty((1, ), dtype=a.dtype)
out[0] = -4
else:
out = np.empty(a.shape[:-2], dtype=a.dtype)
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(a_usm, out_usm, a.shapeptr, a.ndim)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([out.size, a.size])
if PRINT_DEBUG:
print("dpnp implementation")
if a.ndim == 2:
return out[0]
else:
return out
return dpnp_impl
def dpnp_cumprod_impl(a):
name = "cumprod"
dpnp_lowering.ensure_dpnp(name)
res_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.5.1/dpnp/backend/kernels/dpnp_krnl_mathematical.cpp#L110
Function declaration:
void dpnp_cumprod_c(void* array1_in, void* result1, size_t size)
"""
sig = signature(res_type, types.voidptr, types.voidptr, types.intp)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
if a.dtype == types.Integer:
ret_dtype = np.int64
else:
ret_dtype = a.dtype
def dpnp_impl(a):
out = np.arange(0, a.size, 1, ret_dtype)
common_impl(a, out, dpnp_func, PRINT_DEBUG)
return out
return dpnp_impl
def dpnp_matrix_rank_impl(M, tol=None, hermitian=False):
name = "matrix_rank"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels_linalg.cpp#L186
Function declaration:
void custom_matrix_rank_c(void* array1_in, void* result1, size_t* shape, size_t ndim)
"""
sig = signature(ret_type, types.voidptr, types.voidptr, types.voidptr,
types.intp)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [M.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(M, tol=None, hermitian=False):
if tol is not None:
raise ValueError(
"tol is not supported for np.linalg.matrix_rank(M)")
if hermitian:
raise ValueError(
"hermitian is not supported for np.linalg.matrix_rank(M)")
if M.ndim > 2:
raise ValueError(
"np.linalg.matrix_rank(M) is only supported on 1 or 2-d arrays"
)
out = np.empty(1, dtype=M.dtype)
sycl_queue = dpctl_functions.get_current_queue()
M_usm = dpctl_functions.malloc_shared(M.size * M.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, M_usm, M.ctypes,
M.size * M.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(M_usm, out_usm, M.shapeptr, M.ndim)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(M_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([out.size, M.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out[0]
return dpnp_impl
def dpnp_eigvals_impl(a):
dpnp_lowering.ensure_dpnp("eigvals")
def dpnp_impl(a):
eigval, eigvec = numba_dppy.dpnp.eig(a)
return eigval
return dpnp_impl
def dpnp_all_impl(a):
name = "all"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.6.2/dpnp/backend/kernels/dpnp_krnl_logic.cpp#L36
Function declaration:
void dpnp_all_c(const void* array1_in, void* result1, const size_t size)
"""
sig = signature(ret_type, types.voidptr, types.voidptr, types.intp)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
if a.size == 0:
return True
out = np.empty(1, dtype=np.bool_)
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(a_usm, out_usm, a.size)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
# TODO: sometimes all() returns ndarray
return out[0]
return dpnp_impl
def dpnp_sort_impl(a):
name = "sort"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.5.0/dpnp/backend/kernels/dpnp_krnl_sorting.cpp#L90
Function declaration:
void dpnp_sort_c(void* array1_in, void* result1, size_t size)
"""
sig = signature(ret_type, types.voidptr, types.voidptr, types.intp)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
res_dtype = a.dtype
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
if a.size == 0:
raise ValueError("Passed Empty array")
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out = np.arange(0, a.size, 1, res_dtype)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(a_usm, out_usm, a.size)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out
return dpnp_impl
def dpnp_argmin_impl(a):
name = "argmin"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels_searching.cpp#L56
Function declaration:
void custom_argmin_c(void* array1_in, void* result1, size_t size)
"""
sig = signature(ret_type, types.voidptr, types.voidptr, types.intp)
dpnp_func = dpnp_ext.dpnp_func(
"dpnp_" + name, [a.dtype.name, np.dtype(np.int64).name], sig)
res_dtype = np.int64
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
if a.size == 0:
raise ValueError("Passed Empty array")
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out = np.empty(1, dtype=res_dtype)
out_usm = dpctl_functions.malloc_shared(out.itemsize, sycl_queue)
dpnp_func(a_usm, out_usm, a.size)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out[0]
return dpnp_impl
def dpnp_vdot_impl(a, b):
dpnp_lowering.ensure_dpnp("vdot")
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels.cpp#L118
Function declaration:
void dpnp_dot_c(void* array1_in, void* array2_in, void* result1, size_t size)
"""
def dpnp_impl(a, b):
return numba_dppy.dpnp.dot(np.ravel(a), np.ravel(b))
return dpnp_impl
def dpnp_sum_impl(a):
name = "sum"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.6.1dev/dpnp/backend/kernels/dpnp_krnl_reduction.cpp#L59
Function declaration:
void dpnp_sum_c(void* result_out,
const void* input_in,
const size_t* input_shape,
const size_t input_shape_ndim,
const long* axes,
const size_t axes_ndim,
const void* initial,
const long* where)
"""
sig = signature(
ret_type,
types.voidptr, # void* result_out,
types.voidptr, # const void* input_in,
types.voidptr, # const size_t* input_shape,
types.intp, # const size_t input_shape_ndim,
types.voidptr, # const long* axes,
types.intp, # const size_t axes_ndim,
types.voidptr, # const void* initial,
types.voidptr, # const long* where)
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
out = np.empty(1, dtype=a.dtype)
common_impl(a, out, dpnp_func, PRINT_DEBUG)
return out[0]
return dpnp_impl
def dpnp_nanprod_impl(a):
name = "nanprod"
dpnp_lowering.ensure_dpnp(name)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
a_ravel = a.ravel()
a_ravel_copy = np.copy(a_ravel)
for i in range(len(a_ravel_copy)):
if np.isnan(a_ravel_copy[i]):
a_ravel_copy[i] = 1
result = numba_dppy.dpnp.prod(a_ravel_copy)
dpnp_ext._dummy_liveness_func([a.size, a_ravel_copy.size])
if PRINT_DEBUG:
print("dpnp implementation")
return result
return dpnp_impl
def dpnp_multi_dot_impl(arrays):
dpnp_lowering.ensure_dpnp("multi_dot")
print_debug = dpnp_lowering.DEBUG
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels.cpp#L118
Function declaration:
void dpnp_dot_c(void* array1_in, void* array2_in, void* result1, size_t size)
"""
def dpnp_impl(arrays):
n = len(arrays)
result = arrays[0]
for idx in range(1, n):
result = numba_dppy.dpnp.dot(result, arrays[idx])
if print_debug:
print("dpnp implementation")
return result
return dpnp_impl
def dpnp_take_impl(a, ind):
name = "take"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/ca6eb1b8fc561957402b6f258529f862c4a8f945/dpnp/backend/kernels/dpnp_krnl_indexing.cpp#L479
Function declaration:
void dpnp_take_c(void* array1_in, const size_t array1_size, void* indices1, void* result1, size_t size)
"""
sig = signature(
ret_type,
types.voidptr,
types.intp,
types.voidptr,
types.voidptr,
types.intp,
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
res_dtype = a.dtype
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a, ind):
if a.size == 0:
raise ValueError("Passed Empty array")
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
ind_usm = dpctl_functions.malloc_shared(ind.size * ind.itemsize,
sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, ind_usm, ind.ctypes,
ind.size * ind.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out = np.arange(0, ind.size, 1, res_dtype).reshape(ind.shape)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(a_usm, a.size * a.itemsize, ind_usm, out_usm, ind.size)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(ind_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, ind.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out
return dpnp_impl
def dpnp_cov_impl(a):
name = "cov"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels_statistics.cpp#L51
Function declaration:
void custom_cov_c(void* array1_in, void* result1, size_t nrows, size_t ncols)
"""
sig = signature(
ret_type, types.voidptr, types.voidptr, types.intp, types.intp
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
res_dtype = np.float64
copy_input_to_double = True
if a.dtype == types.float64:
copy_input_to_double = False
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
if a.size == 0:
raise ValueError("Passed Empty array")
sycl_queue = dpctl_functions.get_current_queue()
""" We have to pass a array in double precision to DpNp """
if copy_input_to_double:
a_copy_in_double = a.astype(np.float64)
else:
a_copy_in_double = a
a_usm = dpctl_functions.malloc_shared(
a_copy_in_double.size * a_copy_in_double.itemsize, sycl_queue
)
event = dpctl_functions.queue_memcpy(
sycl_queue,
a_usm,
a_copy_in_double.ctypes,
a_copy_in_double.size * a_copy_in_double.itemsize,
)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
if a.ndim == 2:
rows = a.shape[0]
cols = a.shape[1]
out = np.empty((rows, rows), dtype=res_dtype)
elif a.ndim == 1:
rows = 1
cols = a.shape[0]
out = np.empty(rows, dtype=res_dtype)
out_usm = dpctl_functions.malloc_shared(
out.size * out.itemsize, sycl_queue
)
dpnp_func(a_usm, out_usm, rows, cols)
event = dpctl_functions.queue_memcpy(
sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a_copy_in_double.size, a.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
if a.ndim == 2:
return out
elif a.ndim == 1:
return out[0]
return dpnp_impl
def dpnp_amax_impl(a):
name = "max"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/e389248c709531b181be8bf33b1a270fca812a92/dpnp/backend/kernels/dpnp_krnl_statistics.cpp#L149
Function declaration:
void dpnp_max_c(void* array1_in, void* result1, const size_t result_size, const size_t* shape, size_t ndim, const size_t* axis, size_t naxis)
We are using void * in case of size_t * as Numba currently does not have
any type to represent size_t *. Since, both the types are pointers,
if the compiler allows there should not be any mismatch in the size of
the container to hold different types of pointer.
"""
sig = signature(
ret_type,
types.voidptr,
types.voidptr,
types.intp,
types.voidptr,
types.intp,
types.voidptr,
types.intp,
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
if a.size == 0:
raise ValueError("Passed Empty array")
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(
sycl_queue, a_usm, a.ctypes, a.size * a.itemsize
)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out_usm = dpctl_functions.malloc_shared(a.itemsize, sycl_queue)
axis, naxis = 0, 0
dpnp_func(
a_usm, out_usm, a.size * a.itemsize, a.shapeptr, a.ndim, axis, naxis
)
out = np.empty(1, dtype=a.dtype)
event = dpctl_functions.queue_memcpy(
sycl_queue, out.ctypes, out_usm, out.size * out.itemsize
)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out[0]
return dpnp_impl
def dpnp_diagonal_impl(a, offset=0):
name = "diagonal"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/e389248c709531b181be8bf33b1a270fca812a92/dpnp/backend/kernels/dpnp_krnl_indexing.cpp#L39
Function declaration:
void dpnp_diagonal_c(
void* array1_in, const size_t input1_size, void* result1, const size_t offset, size_t* shape, size_t* res_shape, const size_t res_ndim)
"""
sig = signature(
ret_type,
types.voidptr,
types.intp,
types.voidptr,
types.intp,
types.voidptr,
types.voidptr,
types.intp,
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
function_text = f"""\
def tuplizer(a):
return ({", ".join(f"a[{i}]" for i in range(a.ndim - 1))})
"""
locals = {}
exec(function_text, globals(), locals)
tuplizer = register_jitable(locals["tuplizer"])
def dpnp_impl(a, offset=0):
if a.size == 0:
raise ValueError("Passed Empty array")
n = min(a.shape[0], a.shape[1])
res_shape = np.zeros(a.ndim - 1, dtype=np.int64)
if a.ndim > 2:
for i in range(a.ndim - 2):
res_shape[i] = a.shape[i + 2]
if (n + offset) > a.shape[1]:
res_shape[-1] = a.shape[1] - offset
elif (n + offset) > a.shape[0]:
res_shape[-1] = a.shape[0]
else:
res_shape[-1] = n + offset
shape = tuplizer(res_shape)
out = np.empty(shape, dtype=a.dtype)
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(
a_usm,
a.size * a.itemsize,
out_usm,
offset,
a.shapeptr,
out.shapeptr,
out.ndim,
)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out
return dpnp_impl
def dpnp_partition_impl(a, kth):
name = "partition"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.6.2/dpnp/backend/kernels/dpnp_krnl_sorting.cpp#L90
Function declaration:
void dpnp_partition_c(
void* array1_in, void* array2_in, void* result1, const size_t kth, const size_t* shape_, const size_t ndim)
"""
sig = signature(
ret_type,
types.voidptr,
types.voidptr,
types.voidptr,
types.intp,
types.voidptr,
types.intp,
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a, kth):
if a.size == 0:
raise ValueError("Passed Empty array")
kth_ = kth if kth >= 0 else (a.ndim + kth)
arr2 = numba_dppy.dpnp.copy(a)
out = np.empty(a.shape, dtype=a.dtype)
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
arr2_usm = dpctl_functions.malloc_shared(arr2.size * arr2.itemsize,
sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, arr2_usm, arr2.ctypes,
arr2.size * arr2.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(a_usm, arr2_usm, out_usm, kth_, a.shapeptr, a.ndim)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(arr2_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, arr2.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out
return dpnp_impl
def dpnp_eig_impl(a):
name = "eig"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.4.0/dpnp/backend/custom_kernels.cpp#L180
Function declaration:
void dpnp_eig_c(const void* array_in, void* result1, void* result2, size_t size)
"""
sig = signature(ret_type, types.voidptr, types.voidptr, types.voidptr,
types.intp)
dpnp_eig = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
res_dtype = np.float64
if a.dtype == types.float32:
res_dtype = np.float32
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a):
n = a.shape[-1]
if a.shape[-2] != n:
msg = "Last 2 dimensions of the array must be square."
raise ValueError(msg)
dpnp_ext._check_finite_matrix(a)
wr = np.empty(n, dtype=res_dtype)
vr = np.empty((n, n), dtype=res_dtype)
if n == 0:
return (wr, vr)
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
wr_usm = dpctl_functions.malloc_shared(wr.size * wr.itemsize,
sycl_queue)
vr_usm = dpctl_functions.malloc_shared(vr.size * vr.itemsize,
sycl_queue)
dpnp_eig(a_usm, wr_usm, vr_usm, n)
event = dpctl_functions.queue_memcpy(sycl_queue, wr.ctypes, wr_usm,
wr.size * wr.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
event = dpctl_functions.queue_memcpy(sycl_queue, vr.ctypes, vr_usm,
vr.size * vr.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(wr_usm, sycl_queue)
dpctl_functions.free_with_queue(vr_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([wr.size, vr.size])
if PRINT_DEBUG:
print("dpnp implementation")
return (wr, vr)
return dpnp_impl
def dpnp_repeat_impl(a, repeats):
name = "repeat"
dpnp_lowering.ensure_dpnp(name)
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/0.6.2/dpnp/backend/kernels/dpnp_krnl_manipulation.cpp#L46
Function declaration:
void dpnp_repeat_c(const void* array1_in, void* result1, const size_t repeats, const size_t size)
"""
sig = signature(
ret_type,
types.voidptr,
types.voidptr,
types.intp,
types.intp,
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_" + name, [a.dtype.name, "NONE"], sig)
PRINT_DEBUG = dpnp_lowering.DEBUG
def dpnp_impl(a, repeats):
if a.size == 0:
raise ValueError("Passed Empty array")
if a.ndim >= 2:
raise ValueError("Not supported in dpnp")
new_size = a.size * repeats
out = np.zeros(new_size, dtype=a.dtype)
sycl_queue = dpctl_functions.get_current_queue()
a_usm = dpctl_functions.malloc_shared(a.size * a.itemsize, sycl_queue)
event = dpctl_functions.queue_memcpy(sycl_queue, a_usm, a.ctypes,
a.size * a.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
out_usm = dpctl_functions.malloc_shared(out.size * out.itemsize,
sycl_queue)
dpnp_func(a_usm, out_usm, repeats, a.size)
event = dpctl_functions.queue_memcpy(sycl_queue, out.ctypes, out_usm,
out.size * out.itemsize)
dpctl_functions.event_wait(event)
dpctl_functions.event_delete(event)
dpctl_functions.free_with_queue(a_usm, sycl_queue)
dpctl_functions.free_with_queue(out_usm, sycl_queue)
dpnp_ext._dummy_liveness_func([a.size, out.size])
if PRINT_DEBUG:
print("dpnp implementation")
return out
return dpnp_impl
def dpnp_dot_impl(a, b):
dpnp_lowering.ensure_dpnp("dot")
ret_type = types.void
"""
dpnp source:
https://github.com/IntelPython/dpnp/blame/67a101c90cf253cfe9b9ba80ac397811ce94edee/dpnp/backend/kernels/dpnp_krnl_common.cpp#L322
Function declaration:
void dpnp_matmul_c(void* result_out,
const size_t result_size,
const size_t result_ndim,
const size_t* result_shape,
const size_t* result_strides,
const void* input1_in,
const size_t input1_size,
const size_t input1_ndim,
const size_t* input1_shape,
const size_t* input1_strides,
const void* input2_in,
const size_t input2_size,
const size_t input2_ndim,
const size_t* input2_shape,
const size_t* input2_strides)
"""
sig = signature(
ret_type,
types.voidptr,
types.intp,
types.intp,
types.voidptr,
types.voidptr,
types.voidptr,
types.intp,
types.intp,
types.voidptr,
types.voidptr,
types.voidptr,
types.intp,
types.intp,
types.voidptr,
types.voidptr,
)
res_dtype = get_res_dtype(a, b)
PRINT_DEBUG = dpnp_lowering.DEBUG
ndims = [a.ndim, b.ndim]
if ndims == [2, 2]:
dpnp_func = dpnp_ext.dpnp_func("dpnp_matmul", [a.dtype.name, "NONE"],
sig)
def dot_2_mm(a, b):
m, k = a.shape
_k, n = b.shape
if _k != k:
raise ValueError("Incompatible array sizes for np.dot(a, b)")
out = np.empty((m, n), dtype=res_dtype)
common_matmul_impl(dpnp_func, a, b, out, m, n, k, PRINT_DEBUG)
return out
return dot_2_mm
elif ndims == [2, 1]:
dpnp_func = dpnp_ext.dpnp_func("dpnp_matmul", [a.dtype.name, "NONE"],
sig)
def dot_2_mv(a, b):
m, k = a.shape
(_n, ) = b.shape
n = 1
if _n != k:
raise ValueError("Incompatible array sizes for np.dot(a, b)")
out = np.empty((m, ), dtype=res_dtype)
common_matmul_impl(dpnp_func, a, b, out, m, n, k, PRINT_DEBUG)
return out
return dot_2_mv
elif ndims == [1, 2]:
dpnp_func = dpnp_ext.dpnp_func("dpnp_matmul", [a.dtype.name, "NONE"],
sig)
def dot_2_vm(a, b):
(m, ) = a.shape
k, n = b.shape
if m != k:
raise ValueError("Incompatible array sizes for np.dot(a, b)")
out = np.empty((n, ), dtype=res_dtype)
common_matmul_impl(dpnp_func, a, b, out, m, n, k, PRINT_DEBUG)
return out
return dot_2_vm
elif ndims == [1, 1]:
"""
dpnp source:
https://github.com/IntelPython/dpnp/blob/67a101c90cf253cfe9b9ba80ac397811ce94edee/dpnp/backend/kernels/dpnp_krnl_common.cpp#L79
Function declaration:
void dpnp_dot_c(void* result_out,
const size_t result_size,
const size_t result_ndim,
const size_t* result_shape,
const size_t* result_strides,
const void* input1_in,
const size_t input1_size,
const size_t input1_ndim,
const size_t* input1_shape,
const size_t* input1_strides,
const void* input2_in,
const size_t input2_size,
const size_t input2_ndim,
const size_t* input2_shape,
const size_t* input2_strides)
"""
sig = signature(
ret_type,
types.voidptr,
types.intp,
types.intp,
types.voidptr,
types.voidptr,
types.voidptr,
types.intp,
types.intp,
types.voidptr,
types.voidptr,
types.voidptr,
types.intp,
types.intp,
types.voidptr,
types.voidptr,
)
dpnp_func = dpnp_ext.dpnp_func("dpnp_dot", [a.dtype.name, "NONE"], sig)
def dot_2_vv(a, b):
(m, ) = a.shape
(n, ) = b.shape
if m != n:
raise ValueError("Incompatible array sizes for np.dot(a, b)")
out = np.empty(1, dtype=res_dtype)
common_dot_impl(dpnp_func, a, b, out, m, PRINT_DEBUG)
return out[0]
return dot_2_vv
else:
assert 0
