lpython
lpython copied to clipboard
lcompilers
Unhandled Exception in Array code
(different but similar to #2479 )
import numpy
from lpython import (i16, i32, ccallback, c_p_pointer, Pointer, u64, CPtr, i64)
######## ALL THE LINES WITH EIGHT COMMENT MARKS ARE THE ONES WE NEED TO
######## BRING UP! AS IT STANDS, THIS CODE WORKS IN LPYTHON MAIN AS OF 4
######## FEBRUARY 2024.
# https://numpy.org/devdocs/reference/typing.html
######## from numpy.typing import NDArray
# plan for 30 Jan 2024 --
# step 0: comment out this code and ./build_baryon.sh to run on APU
# emulator; or ./run_full_emulation.sh to run in CPython.
# step 1: side-by-side numpy implementation in full-emulation
# - get there line-by-line
# = focus on gvml_add_u16 first
def numpy_side_by_side(n: i32, m: i32, l: i32, M1: i32, M2: i32,
A: CPtr, B: CPtr, C: CPtr) -> \
None: ######## NDArray[numpy.int16]:
VR_SIZE: i32 = 32_768
# In the primary example, n = 15, m = 3, l = 32_768,
# M1 = 1, M2 = 5
# source GSI L4 arrays
pA_nm: Pointer[i16[:]] = c_p_pointer(A, i16[:], array([n * m]))
pB_ml: Pointer[i16[:]] = c_p_pointer(B, i16[:], array([m * l]))
# source numpy arrays
######## A_nm: NDArray[numpy.int16] = numpy.zeros((n, m), dtype=numpy.int16)
######## for row in range(n):
######## A_nm[row,:] = pA_nm[(row * m):((row + 1) * m)]
A_nm: Array[i16, n, m]
row : i32
for row in range(n):
A_nm[row,:] = pA_nm[(row * m):((row + 1) * m)]
######## B_ml: NDArray[numpy.int16] = numpy.zeros((m, l), dtype=numpy.int16)
######## for row in range(m):
######## B_ml[row,:] = pB_ml[(row * l):((row + 1) * l)]
# # destination numpy array
######## C_nl: NDArray[numpy.int16] = numpy.zeros((n, l), dtype=numpy.int16)
# destination GSI L4 array
pC_nl: Pointer[i16[:]] = c_p_pointer(C, i16[:], array([n * l]))
# First, accumulate outer product without blocking. This is
# the code we would -ultimately- like to compile. Notice that
# all GSI-specific L1, L4, MMB are hidden.
k: i32
######## for k in range(0, m):
######## C_nl += numpy.outer(A_nm[:,k], B_ml[k,:])
######## pass
# expect
# [[ 5 8 11 ... 20 23 26],
# [ 8 14 20 ... 38 44 50],
# [11 20 29 ... 56 65 74], ...
#
# [ 8 14 20 ... 38 44 50],
# [11 20 29 ... 56 65 74],
# [14 26 38 ... 74 86 98]]
set_breakpoint_here_and_inspect_C_nl : i32 = 0
# Second, with explicit blocking. This is a stepping-stone
# for our back-end. Notice that L1 and MMB are hidden.
# T_1l: NDArray[numpy.int16] = numpy.zeros((1, l), dtype=numpy.int16)
# B_1l: NDArray[numpy.int16] = numpy.zeros((1, l), dtype=numpy.int16)
A_ik: i16
jj: i32
ii: i32
i: i32
for jj in range(0, l, VR_SIZE): # each VR-col chunk in B and C
for ii in range(0, n, M2): # each M2 block in A cols and B rows
for i in range(0, M2): # zero-out rows of C
######## C_nl[i + ii, :] = 0
pass
for k in range(0, m): # rows of B
# B_1l[0, :] = B_ml[k, :]
for i in range(0, M2):
######## A_ik = A_nm[i + ii, k]
# broadcast a single element of A
# T_1l[0, :] = A_ik
# pointwise (Hadamard) product:
# T_1l[0, :] = np.multiply(B_1l[0, :], T_1l[0, :])
# C_nl[i + ii, :] += T_1l[0, :]
# optimization without the temporaries
######## C_nl[i + ii, :] += B_ml[k, :] * A_ik
pass
set_breakpoint_here_and_inspect_C_nl = 0
######## return C_nl
def main():
n : i32 = 15
m : i32 = 3
l : i32 = 32_768
M1 : i32 = 1
M2 : i32 = 5
A_l4 : CPtr
B_l4 : CPtr
C_l4 : CPtr
numpy_side_by_side(n, m, l, M1, M2, A_l4, B_l4, C_l4)
print ("hello, world!")
if __name__ == "__main__":
main()
└─(10:50:28 on vector-backend ✖ ✭)──> lpython ../ISSUES/UNHANDLED-EXCEPTIONS/Issue2480.py 1 ↵ ──(Mon,Feb05)─┘
Traceback (most recent call last):
File "/Users/brian/Dropbox/Mac/Documents/GitHub/lpython/src/bin/lpython.cpp", line 1872
err = compile_python_to_object_file(arg_file, tmp_o, runtime_library_dir,
File "/Users/brian/Dropbox/Mac/Documents/GitHub/lpython/src/bin/lpython.cpp", line 824
res = fe.get_llvm3(*asr, pass_manager, diagnostics, infile);
File "/Users/brian/Dropbox/Mac/Documents/GitHub/lpython/src/lpython/python_evaluator.cpp", line 71
run_fn, infile);
File "/Users/brian/Dropbox/Mac/Documents/GitHub/lpython/src/libasr/codegen/asr_to_llvm.cpp", line 9276
pass_manager.apply_passes(al, &asr, co.po, diagnostics);
File "/Users/brian/Documents/GitHub/lpython/src/libasr/pass/pass_manager.h", line 299
apply_passes(al, asr, _passes, pass_options, diagnostics);
File "/Users/brian/Documents/GitHub/lpython/src/libasr/pass/pass_manager.h", line 160
_passes_db[passes[i]](al, *asr, pass_options);
File "/Users/brian/Dropbox/Mac/Documents/GitHub/lpython/src/libasr/pass/array_op.cpp", line 1910
u.visit_TranslationUnit(unit);
File "../libasr/asr.h", line 5277
File "../libasr/asr.h", line 5060
File "../libasr/asr.h", line 4774
File "../libasr/pass/pass_utils.h", line 317
File "../libasr/asr.h", line 5290
File "../libasr/asr.h", line 5060
File "../libasr/asr.h", line 4775
File "../libasr/pass/pass_utils.h", line 298
File "../libasr/asr.h", line 5303
File "../libasr/asr.h", line 5077
File "../libasr/asr.h", line 4805
File "../libasr/asr.h", line 5441
File "../libasr/asr.h", line 5077
File "../libasr/asr.h", line 5077
Binary file "/usr/lib/system/libsystem_platform.dylib", local address: 0x18046da23
Segfault: Signal SIGSEGV (segmentation fault) received
(lp) ┌─(~/Documents/GitHub/lpython/integration_tests)───────────────────────────────────────(brian@MacBook-Pro:s001)─┐
It's probably some array operations. We have made good improvements, so we should sync ASR from LC and LFortran, but probably there are more bugs.
