def transpose8x3(xmm_rows):
assert isinstance(xmm_rows, list) and len(xmm_rows) == 8 and all(
isinstance(xmm_row, XMMRegister) for xmm_row in xmm_rows)
# xmm_rows[0] = ( 0.0, g02, g01, g00 )
# xmm_rows[1] = ( 0.0, g12, g11, g10 )
# xmm_rows[2] = ( 0.0, g22, g21, g20 )
# xmm_rows[3] = ( 0.0, g32, g31, g30 )
# xmm_rows[4] = ( 0.0, g42, g41, g40 )
# xmm_rows[5] = ( 0.0, g52, g51, g50 )
# xmm_rows[6] = ( 0.0, g62, g61, g60 )
# xmm_rows[7] = ( 0.0, g72, g71, g70 )
ymm_rows = [YMMRegister() for _ in range(4)]
VINSERTF128(ymm_rows[0], xmm_rows[0].as_ymm, xmm_rows[4], 1)
VINSERTF128(ymm_rows[1], xmm_rows[1].as_ymm, xmm_rows[5], 1)
VINSERTF128(ymm_rows[2], xmm_rows[2].as_ymm, xmm_rows[6], 1)
VINSERTF128(ymm_rows[3], xmm_rows[3].as_ymm, xmm_rows[7], 1)
# ymm_rows[0] = ( 0.0, g42, g41, g40, 0.0, g02, g01, g00 )
# ymm_rows[1] = ( 0.0, g52, g51, g50, 0.0, g12, g11, g10 )
# ymm_rows[2] = ( 0.0, g62, g61, g60, 0.0, g22, g21, g20 )
# ymm_rows[3] = ( 0.0, g72, g71, g70, 0.0, g32, g31, g30 )
ymm_new_rows = [YMMRegister() for _ in range(4)]
VUNPCKLPS(ymm_new_rows[0], ymm_rows[0], ymm_rows[1])
VUNPCKHPS(ymm_new_rows[1], ymm_rows[0], ymm_rows[1])
VUNPCKLPS(ymm_new_rows[2], ymm_rows[2], ymm_rows[3])
VUNPCKHPS(ymm_new_rows[3], ymm_rows[2], ymm_rows[3])
for ymm_row, ymm_new_row in zip(ymm_rows, ymm_new_rows):
SWAP.REGISTERS(ymm_row, ymm_new_row)
# ymm_rows[0] = ( g51, g41, g50, g40, g11, g01, g10, g00 )
# ymm_rows[1] = ( 0.0, 0.0, g52, g42, 0.0, 0.0, g12, g02 )
# ymm_rows[2] = ( g71, g61, g70, g60, g31, g21, g30, g20 )
# ymm_rows[3] = ( 0.0, 0.0, g72, g62, 0.0, 0.0, g32, g22 )
# ymm_rows[0] = ( g70, g60, g50, g40, g30, g20, g10, g00 )
# ymm_rows[2] = ( g71, g61, g51, g41, g31, g21, g11, g01 )
transpose2x2x2x64(ymm_rows[0], ymm_rows[2])
# ymm_rows[1] = ( g72, g62, g52, g42, g32, g22, g12, g02 )
VUNPCKLPD(ymm_rows[1], ymm_rows[1], ymm_rows[3])
SWAP.REGISTERS(ymm_rows[1], ymm_rows[2])
return ymm_rows[0:3]
def transpose8x3(xmm_rows):
assert isinstance(xmm_rows, list) and len(xmm_rows) == 8 and all(isinstance(xmm_row, XMMRegister) for xmm_row in xmm_rows)
# xmm_rows[0] = ( 0.0, g02, g01, g00 )
# xmm_rows[1] = ( 0.0, g12, g11, g10 )
# xmm_rows[2] = ( 0.0, g22, g21, g20 )
# xmm_rows[3] = ( 0.0, g32, g31, g30 )
# xmm_rows[4] = ( 0.0, g42, g41, g40 )
# xmm_rows[5] = ( 0.0, g52, g51, g50 )
# xmm_rows[6] = ( 0.0, g62, g61, g60 )
# xmm_rows[7] = ( 0.0, g72, g71, g70 )
ymm_rows = [YMMRegister() for _ in range(4)]
VINSERTF128(ymm_rows[0], xmm_rows[0].as_ymm, xmm_rows[4], 1)
VINSERTF128(ymm_rows[1], xmm_rows[1].as_ymm, xmm_rows[5], 1)
VINSERTF128(ymm_rows[2], xmm_rows[2].as_ymm, xmm_rows[6], 1)
VINSERTF128(ymm_rows[3], xmm_rows[3].as_ymm, xmm_rows[7], 1)
# ymm_rows[0] = ( 0.0, g42, g41, g40, 0.0, g02, g01, g00 )
# ymm_rows[1] = ( 0.0, g52, g51, g50, 0.0, g12, g11, g10 )
# ymm_rows[2] = ( 0.0, g62, g61, g60, 0.0, g22, g21, g20 )
# ymm_rows[3] = ( 0.0, g72, g71, g70, 0.0, g32, g31, g30 )
ymm_new_rows = [YMMRegister() for _ in range(4)]
VUNPCKLPS(ymm_new_rows[0], ymm_rows[0], ymm_rows[1])
VUNPCKHPS(ymm_new_rows[1], ymm_rows[0], ymm_rows[1])
VUNPCKLPS(ymm_new_rows[2], ymm_rows[2], ymm_rows[3])
VUNPCKHPS(ymm_new_rows[3], ymm_rows[2], ymm_rows[3])
for ymm_row, ymm_new_row in zip(ymm_rows, ymm_new_rows):
SWAP.REGISTERS(ymm_row, ymm_new_row)
# ymm_rows[0] = ( g51, g41, g50, g40, g11, g01, g10, g00 )
# ymm_rows[1] = ( 0.0, 0.0, g52, g42, 0.0, 0.0, g12, g02 )
# ymm_rows[2] = ( g71, g61, g70, g60, g31, g21, g30, g20 )
# ymm_rows[3] = ( 0.0, 0.0, g72, g62, 0.0, 0.0, g32, g22 )
# ymm_rows[0] = ( g70, g60, g50, g40, g30, g20, g10, g00 )
# ymm_rows[2] = ( g71, g61, g51, g41, g31, g21, g11, g01 )
transpose2x2x2x64(ymm_rows[0], ymm_rows[2])
# ymm_rows[1] = ( g72, g62, g52, g42, g32, g22, g12, g02 )
VUNPCKLPD(ymm_rows[1], ymm_rows[1], ymm_rows[3])
SWAP.REGISTERS(ymm_rows[1], ymm_rows[2])
return ymm_rows[0:3]
def transpose6x8(ymm_rows):
assert isinstance(ymm_rows, list) and len(ymm_rows) == 6 and all(
isinstance(ymm_row, YMMRegister) for ymm_row in ymm_rows)
# ymm_rows[0] = ( g07, g06, g05, g04, g03, g02, g01, g00 )
# ymm_rows[1] = ( g17, g16, g15, g14, g13, g12, g11, g10 )
# ymm_rows[2] = ( g27, g26, g25, g24, g23, g22, g21, g20 )
# ymm_rows[3] = ( g37, g36, g35, g34, g33, g32, g31, g30 )
# ymm_rows[4] = ( g47, g46, g45, g44, g43, g42, g41, g40 )
# ymm_rows[5] = ( g57, g56, g55, g54, g53, g52, g51, g50 )
for ymm_even_row, ymm_odd_row in zip(ymm_rows[0::2], ymm_rows[1::2]):
ymm_temp = YMMRegister()
VUNPCKLPS(ymm_temp, ymm_even_row, ymm_odd_row)
VUNPCKHPS(ymm_odd_row, ymm_even_row, ymm_odd_row)
SWAP.REGISTERS(ymm_even_row, ymm_temp)
# ymm_rows[0] = ( g15, g05, g14, g04, g11, g01, g10, g00 )
# ymm_rows[1] = ( g17, g07, g16, g06, g13, g03, g12, g02 )
# ymm_rows[2] = ( g35, g25, g34, g24, g31, g21, g30, g20 )
# ymm_rows[3] = ( g37, g27, g36, g26, g33, g23, g32, g22 )
# ymm_rows[4] = ( g55, g45, g54, g44, g51, g41, g50, g40 )
# ymm_rows[5] = ( g57, g47, g56, g46, g53, g43, g52, g42 )
ymm_zero_rows = [YMMRegister(), YMMRegister()]
for ymm_zero in ymm_zero_rows:
VXORPS(ymm_zero, ymm_zero, ymm_zero)
ymm_rows += ymm_zero_rows
# ymm_rows[6] = ( 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 )
# ymm_rows[7] = ( 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 )
transpose2x2x2x64(ymm_rows[0], ymm_rows[2])
transpose2x2x2x64(ymm_rows[1], ymm_rows[3])
transpose2x2x2x64(ymm_rows[4], ymm_rows[6])
transpose2x2x2x64(ymm_rows[5], ymm_rows[7])
# ymm_rows[0] = ( g34, g24, g14, g04, g30, g20, g10, g00 )
# ymm_rows[1] = ( g36, g26, g16, g06, g32, g22, g12, g02 )
# ymm_rows[2] = ( g35, g25, g15, g05, g31, g21, g11, g01 )
# ymm_rows[3] = ( g37, g27, g17, g07, g33, g23, g13, g03 )
# ymm_rows[4] = ( 0.0, 0.0, g54, g44, 0.0, 0.0, g50, g40 )
# ymm_rows[5] = ( 0.0, 0.0, g56, g46, 0.0, 0.0, g52, g42 )
# ymm_rows[6] = ( 0.0, 0.0, g55, g45, 0.0, 0.0, g51, g41 )
# ymm_rows[7] = ( 0.0, 0.0, g57, g47, 0.0, 0.0, g53, g43 )
transpose2x2x128(ymm_rows[0], ymm_rows[4])
transpose2x2x128(ymm_rows[1], ymm_rows[5])
transpose2x2x128(ymm_rows[2], ymm_rows[6])
transpose2x2x128(ymm_rows[3], ymm_rows[7])
SWAP.REGISTERS(ymm_rows[1], ymm_rows[2])
SWAP.REGISTERS(ymm_rows[5], ymm_rows[6])
return ymm_rows
def transpose6x8(ymm_rows):
assert isinstance(ymm_rows, list) and len(ymm_rows) == 6 and all(isinstance(ymm_row, YMMRegister) for ymm_row in ymm_rows)
# ymm_rows[0] = ( g07, g06, g05, g04, g03, g02, g01, g00 )
# ymm_rows[1] = ( g17, g16, g15, g14, g13, g12, g11, g10 )
# ymm_rows[2] = ( g27, g26, g25, g24, g23, g22, g21, g20 )
# ymm_rows[3] = ( g37, g36, g35, g34, g33, g32, g31, g30 )
# ymm_rows[4] = ( g47, g46, g45, g44, g43, g42, g41, g40 )
# ymm_rows[5] = ( g57, g56, g55, g54, g53, g52, g51, g50 )
for ymm_even_row, ymm_odd_row in zip(ymm_rows[0::2], ymm_rows[1::2]):
ymm_temp = YMMRegister()
VUNPCKLPS(ymm_temp, ymm_even_row, ymm_odd_row)
VUNPCKHPS(ymm_odd_row, ymm_even_row, ymm_odd_row)
SWAP.REGISTERS(ymm_even_row, ymm_temp)
# ymm_rows[0] = ( g15, g05, g14, g04, g11, g01, g10, g00 )
# ymm_rows[1] = ( g17, g07, g16, g06, g13, g03, g12, g02 )
# ymm_rows[2] = ( g35, g25, g34, g24, g31, g21, g30, g20 )
# ymm_rows[3] = ( g37, g27, g36, g26, g33, g23, g32, g22 )
# ymm_rows[4] = ( g55, g45, g54, g44, g51, g41, g50, g40 )
# ymm_rows[5] = ( g57, g47, g56, g46, g53, g43, g52, g42 )
ymm_zero_rows = [YMMRegister(), YMMRegister()]
for ymm_zero in ymm_zero_rows:
VXORPS(ymm_zero, ymm_zero, ymm_zero)
ymm_rows += ymm_zero_rows
# ymm_rows[6] = ( 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 )
# ymm_rows[7] = ( 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 )
transpose2x2x2x64(ymm_rows[0], ymm_rows[2])
transpose2x2x2x64(ymm_rows[1], ymm_rows[3])
transpose2x2x2x64(ymm_rows[4], ymm_rows[6])
transpose2x2x2x64(ymm_rows[5], ymm_rows[7])
# ymm_rows[0] = ( g34, g24, g14, g04, g30, g20, g10, g00 )
# ymm_rows[1] = ( g36, g26, g16, g06, g32, g22, g12, g02 )
# ymm_rows[2] = ( g35, g25, g15, g05, g31, g21, g11, g01 )
# ymm_rows[3] = ( g37, g27, g17, g07, g33, g23, g13, g03 )
# ymm_rows[4] = ( 0.0, 0.0, g54, g44, 0.0, 0.0, g50, g40 )
# ymm_rows[5] = ( 0.0, 0.0, g56, g46, 0.0, 0.0, g52, g42 )
# ymm_rows[6] = ( 0.0, 0.0, g55, g45, 0.0, 0.0, g51, g41 )
# ymm_rows[7] = ( 0.0, 0.0, g57, g47, 0.0, 0.0, g53, g43 )
transpose2x2x128(ymm_rows[0], ymm_rows[4])
transpose2x2x128(ymm_rows[1], ymm_rows[5])
transpose2x2x128(ymm_rows[2], ymm_rows[6])
transpose2x2x128(ymm_rows[3], ymm_rows[7])
SWAP.REGISTERS(ymm_rows[1], ymm_rows[2])
SWAP.REGISTERS(ymm_rows[5], ymm_rows[6])
return ymm_rows
def transpose8x8(ymm_rows):
assert isinstance(ymm_rows, list) and len(ymm_rows) == 8 and all(
isinstance(ymm_row, YMMRegister) for ymm_row in ymm_rows)
# ymm_rows[0] = ( g07, g06, g05, g04, g03, g02, g01, g00 )
# ymm_rows[1] = ( g17, g16, g15, g14, g13, g12, g11, g10 )
# ymm_rows[2] = ( g27, g26, g25, g24, g23, g22, g21, g20 )
# ymm_rows[3] = ( g37, g36, g35, g34, g33, g32, g31, g30 )
# ymm_rows[4] = ( g47, g46, g45, g44, g43, g42, g41, g40 )
# ymm_rows[5] = ( g57, g56, g55, g54, g53, g52, g51, g50 )
# ymm_rows[6] = ( g67, g66, g65, g64, g63, g62, g61, g60 )
# ymm_rows[7] = ( g77, g76, g75, g74, g73, g72, g71, g70 )
for ymm_even_row, ymm_odd_row in zip(ymm_rows[0::2], ymm_rows[1::2]):
ymm_temp = YMMRegister()
VUNPCKLPS(ymm_temp, ymm_even_row, ymm_odd_row)
VUNPCKHPS(ymm_odd_row, ymm_even_row, ymm_odd_row)
SWAP.REGISTERS(ymm_even_row, ymm_temp)
# ymm_rows[0] = ( g15, g05, g14, g04, g11, g01, g10, g00 )
# ymm_rows[1] = ( g17, g07, g16, g06, g13, g03, g12, g02 )
# ymm_rows[2] = ( g35, g25, g34, g24, g31, g21, g30, g20 )
# ymm_rows[3] = ( g37, g27, g36, g26, g33, g23, g32, g22 )
# ymm_rows[4] = ( g55, g45, g54, g44, g51, g41, g50, g40 )
# ymm_rows[5] = ( g57, g47, g56, g46, g53, g43, g52, g42 )
# ymm_rows[6] = ( g75, g65, g74, g64, g71, g61, g70, g60 )
# ymm_rows[7] = ( g77, g67, g76, g66, g73, g63, g72, g62 )
transpose2x2x2x64(ymm_rows[0], ymm_rows[2])
transpose2x2x2x64(ymm_rows[1], ymm_rows[3])
transpose2x2x2x64(ymm_rows[4], ymm_rows[6])
transpose2x2x2x64(ymm_rows[5], ymm_rows[7])
# ymm_rows[0] = ( g34, g24, g14, g04, g30, g20, g10, g00 )
# ymm_rows[1] = ( g36, g26, g16, g06, g32, g22, g12, g02 )
# ymm_rows[2] = ( g35, g25, g15, g05, g31, g21, g11, g01 )
# ymm_rows[3] = ( g37, g27, g17, g07, g33, g23, g13, g03 )
# ymm_rows[4] = ( g74, g64, g54, g44, g70, g60, g50, g40 )
# ymm_rows[5] = ( g76, g66, g56, g46, g72, g62, g52, g42 )
# ymm_rows[6] = ( g75, g65, g55, g45, g71, g61, g51, g41 )
# ymm_rows[7] = ( g77, g67, g57, g47, g73, g63, g53, g43 )
transpose2x2x128(ymm_rows[0], ymm_rows[4])
transpose2x2x128(ymm_rows[1], ymm_rows[5])
transpose2x2x128(ymm_rows[2], ymm_rows[6])
transpose2x2x128(ymm_rows[3], ymm_rows[7])
SWAP.REGISTERS(ymm_rows[1], ymm_rows[2])
SWAP.REGISTERS(ymm_rows[5], ymm_rows[6])
def transpose8x8(ymm_rows):
assert isinstance(ymm_rows, list) and len(ymm_rows) == 8 and all(isinstance(ymm_row, YMMRegister) for ymm_row in ymm_rows)
# ymm_rows[0] = ( g07, g06, g05, g04, g03, g02, g01, g00 )
# ymm_rows[1] = ( g17, g16, g15, g14, g13, g12, g11, g10 )
# ymm_rows[2] = ( g27, g26, g25, g24, g23, g22, g21, g20 )
# ymm_rows[3] = ( g37, g36, g35, g34, g33, g32, g31, g30 )
# ymm_rows[4] = ( g47, g46, g45, g44, g43, g42, g41, g40 )
# ymm_rows[5] = ( g57, g56, g55, g54, g53, g52, g51, g50 )
# ymm_rows[6] = ( g67, g66, g65, g64, g63, g62, g61, g60 )
# ymm_rows[7] = ( g77, g76, g75, g74, g73, g72, g71, g70 )
for ymm_even_row, ymm_odd_row in zip(ymm_rows[0::2], ymm_rows[1::2]):
ymm_temp = YMMRegister()
VUNPCKLPS(ymm_temp, ymm_even_row, ymm_odd_row)
VUNPCKHPS(ymm_odd_row, ymm_even_row, ymm_odd_row)
SWAP.REGISTERS(ymm_even_row, ymm_temp)
# ymm_rows[0] = ( g15, g05, g14, g04, g11, g01, g10, g00 )
# ymm_rows[1] = ( g17, g07, g16, g06, g13, g03, g12, g02 )
# ymm_rows[2] = ( g35, g25, g34, g24, g31, g21, g30, g20 )
# ymm_rows[3] = ( g37, g27, g36, g26, g33, g23, g32, g22 )
# ymm_rows[4] = ( g55, g45, g54, g44, g51, g41, g50, g40 )
# ymm_rows[5] = ( g57, g47, g56, g46, g53, g43, g52, g42 )
# ymm_rows[6] = ( g75, g65, g74, g64, g71, g61, g70, g60 )
# ymm_rows[7] = ( g77, g67, g76, g66, g73, g63, g72, g62 )
transpose2x2x2x64(ymm_rows[0], ymm_rows[2])
transpose2x2x2x64(ymm_rows[1], ymm_rows[3])
transpose2x2x2x64(ymm_rows[4], ymm_rows[6])
transpose2x2x2x64(ymm_rows[5], ymm_rows[7])
# ymm_rows[0] = ( g34, g24, g14, g04, g30, g20, g10, g00 )
# ymm_rows[1] = ( g36, g26, g16, g06, g32, g22, g12, g02 )
# ymm_rows[2] = ( g35, g25, g15, g05, g31, g21, g11, g01 )
# ymm_rows[3] = ( g37, g27, g17, g07, g33, g23, g13, g03 )
# ymm_rows[4] = ( g74, g64, g54, g44, g70, g60, g50, g40 )
# ymm_rows[5] = ( g76, g66, g56, g46, g72, g62, g52, g42 )
# ymm_rows[6] = ( g75, g65, g55, g45, g71, g61, g51, g41 )
# ymm_rows[7] = ( g77, g67, g57, g47, g73, g63, g53, g43 )
transpose2x2x128(ymm_rows[0], ymm_rows[4])
transpose2x2x128(ymm_rows[1], ymm_rows[5])
transpose2x2x128(ymm_rows[2], ymm_rows[6])
transpose2x2x128(ymm_rows[3], ymm_rows[7])
SWAP.REGISTERS(ymm_rows[1], ymm_rows[2])
SWAP.REGISTERS(ymm_rows[5], ymm_rows[6])
def ifft16_within_rows(ymm_real_rows, ymm_imag_rows, bit_reversal=True):
if isinstance(ymm_real_rows, tuple) and isinstance(ymm_imag_rows, tuple):
return ifft16_within_rows([ymm_real_rows], [ymm_imag_rows])
assert isinstance(ymm_real_rows, list) and all(
isinstance(ymm_real, tuple) and all(
isinstance(ymm, YMMRegister) for ymm in ymm_real)
for ymm_real in ymm_real_rows)
assert isinstance(ymm_imag_rows, list) and all(
isinstance(ymm_imag, tuple) and all(
isinstance(ymm, YMMRegister) for ymm in ymm_imag)
for ymm_imag in ymm_imag_rows)
if bit_reversal:
# Bit reversal
# w[0] = x0 x8 x4 x12 x2 x10 x6 x14
# w[1] = x1 x9 x5 x13 x3 x11 x7 x15
ymm_bit_reversal_mask = YMMRegister()
VMOVDQA(ymm_bit_reversal_mask,
Constant.uint32x8(0, 2, 4, 6, 1, 3, 5, 7))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
for i in range(2):
VPERMPS(ymm_real[i], ymm_bit_reversal_mask, ymm_real[i])
VPERMPS(ymm_imag[i], ymm_bit_reversal_mask, ymm_imag[i])
# 8x FFT2: Butterfly
# w[0] = x0 x4 x2 x6 x8 x12 x10 x14
# w[1] = x1 x5 x3 x7 x9 x13 x11 x15
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
ymm_new_real = YMMRegister(), YMMRegister()
VUNPCKLPS(ymm_new_real[0], ymm_real[0], ymm_real[1])
VUNPCKHPS(ymm_new_real[1], ymm_real[0], ymm_imag[1])
ymm_new_imag = YMMRegister(), YMMRegister()
VUNPCKLPS(ymm_new_imag[0], ymm_imag[0], ymm_imag[1])
VUNPCKHPS(ymm_new_imag[1], ymm_imag[0], ymm_real[1])
SWAP.REGISTERS(ymm_imag[0], ymm_new_imag[0])
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag[1])
SWAP.REGISTERS(ymm_real[0], ymm_new_real[0])
SWAP.REGISTERS(ymm_real[1], ymm_new_real[1])
# w[0] = x0 x1 x4 x5 x8 x9 x12 x13
# w[1] = x2 x3 x6 x7 x10 x11 x14 x15
# 4x FFT4: Butterfly and multiplication by twiddle factors
ymm_fft4_twiddle_factor = YMMRegister()
VMOVAPS(ymm_fft4_twiddle_factor,
Constant.float32x8(+1.0, -1.0, +1.0, -1.0, +1.0, -1.0, +1.0, -1.0))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1], scale_b=ymm_fft4_twiddle_factor)
butterfly(ymm_imag[0], ymm_imag[1])
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x2x64(ymm_real[0], ymm_real[1])
transpose2x2x2x64(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x2 x3 x8 x9 x10 x11
# w[1] = x4 x5 x6 x7 x12 x13 x14 x15
# 2x FFT8: Multiplication by twiddle factors
ymm_fft8_cos_twiddle_factor, ymm_fft8_sin_twiddle_factor = YMMRegister(
), YMMRegister()
VMOVAPS(ymm_fft8_cos_twiddle_factor,
Constant.float32x8(*(cos_npi_over_4 * 2)))
VMOVAPS(ymm_fft8_sin_twiddle_factor,
Constant.float32x8(*(sin_npi_over_4 * 2)))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft8_cos_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft8_cos_twiddle_factor)
VFNMADD231PS(ymm_new_real1, ymm_imag[1], ymm_fft8_sin_twiddle_factor)
VFMADD231PS(ymm_new_imag1, ymm_real[1], ymm_fft8_sin_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# 2x FFT8: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x128(ymm_real[0], ymm_real[1])
transpose2x2x128(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x2 x3 x4 x5 x6 x7
# w[1] = x8 x9 x10 x11 x12 x13 x14 x15
# FFT16: Multiplication by twiddle factors and scale
scale_factor = 0.0625
ymm_fft16_cos_scale_twiddle_factor, ymm_fft16_sin_scale_twiddle_factor = YMMRegister(
), YMMRegister()
VMOVAPS(
ymm_fft16_cos_scale_twiddle_factor,
Constant.float32x8(*[cos * scale_factor for cos in cos_npi_over_8]))
VMOVAPS(
ymm_fft16_sin_scale_twiddle_factor,
Constant.float32x8(*[sin * scale_factor for sin in sin_npi_over_8]))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft16_cos_scale_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft16_cos_scale_twiddle_factor)
VFNMADD231PS(ymm_new_real1, ymm_imag[1],
ymm_fft16_sin_scale_twiddle_factor)
VFMADD231PS(ymm_new_imag1, ymm_real[1],
ymm_fft16_sin_scale_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# FFT16: Butterfly and scale
ymm_scale_factor = YMMRegister()
VMOVAPS(ymm_scale_factor, Constant.float32x8(scale_factor))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1], scale_a=ymm_scale_factor)
butterfly(ymm_imag[0], ymm_imag[1], scale_a=ymm_scale_factor)
def fft16_within_rows(ymm_real_rows, ymm_imag_rows, bit_reversal=True):
if isinstance(ymm_real_rows, tuple) and isinstance(ymm_imag_rows, tuple):
return fft16_within_rows([ymm_real_rows], [ymm_imag_rows])
assert isinstance(ymm_real_rows, list) and all(
isinstance(ymm_real, tuple) and all(
isinstance(ymm, YMMRegister) for ymm in ymm_real)
for ymm_real in ymm_real_rows)
assert isinstance(ymm_imag_rows, list) and all(
isinstance(ymm_imag, tuple) and all(
isinstance(ymm, YMMRegister) for ymm in ymm_imag)
for ymm_imag in ymm_imag_rows)
# FFT16: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
# FFT16: Multiplication by twiddle factors
ymm_fft16_cos_twiddle_factor, ymm_fft16_sin_twiddle_factor = YMMRegister(
), YMMRegister()
VMOVAPS(ymm_fft16_cos_twiddle_factor, Constant.float32x8(*cos_npi_over_8))
VMOVAPS(ymm_fft16_sin_twiddle_factor, Constant.float32x8(*sin_npi_over_8))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft16_cos_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft16_cos_twiddle_factor)
VFMADD231PS(ymm_new_real1, ymm_imag[1], ymm_fft16_sin_twiddle_factor)
VFNMADD231PS(ymm_new_imag1, ymm_real[1], ymm_fft16_sin_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# 2x FFT8: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x128(ymm_real[0], ymm_real[1])
transpose2x2x128(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x2 x3 x8 x9 x10 x11
# w[1] = x4 x5 x6 x7 x12 x13 x14 x15
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
# 2x FFT8: Multiplication by twiddle factors
ymm_fft8_cos_twiddle_factor, ymm_fft8_sin_twiddle_factor = YMMRegister(
), YMMRegister()
VMOVAPS(ymm_fft8_cos_twiddle_factor,
Constant.float32x8(*(cos_npi_over_4 * 2)))
VMOVAPS(ymm_fft8_sin_twiddle_factor,
Constant.float32x8(*(sin_npi_over_4 * 2)))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft8_cos_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft8_cos_twiddle_factor)
VFMADD231PS(ymm_new_real1, ymm_imag[1], ymm_fft8_sin_twiddle_factor)
VFNMADD231PS(ymm_new_imag1, ymm_real[1], ymm_fft8_sin_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# 4x FFT4: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x2x64(ymm_real[0], ymm_real[1])
transpose2x2x2x64(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x4 x5 x8 x9 x12 x13
# w[1] = x2 x3 x6 x7 x10 x11 x14 x15
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
# 4x FFT4: Multiplication by twiddle factors and 8x FFT2: Butterfly
ymm_fft4_twiddle_factor = YMMRegister()
VMOVAPS(ymm_fft4_twiddle_factor,
Constant.float32x8(+1.0, +1.0, -1.0, -1.0, +1.0, +1.0, -1.0, -1.0))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real = YMMRegister(), YMMRegister()
VSHUFPS(ymm_new_real[0], ymm_real[0], ymm_real[1],
_MM_SHUFFLE(2, 0, 2, 0))
VSHUFPS(ymm_new_real[1], ymm_real[0], ymm_imag[1],
_MM_SHUFFLE(3, 1, 3, 1))
butterfly(ymm_new_real[0], ymm_new_real[1])
ymm_new_imag = YMMRegister(), YMMRegister()
VSHUFPS(ymm_new_imag[0], ymm_imag[0], ymm_imag[1],
_MM_SHUFFLE(2, 0, 2, 0))
VSHUFPS(ymm_new_imag[1], ymm_imag[0], ymm_real[1],
_MM_SHUFFLE(3, 1, 3, 1))
butterfly(ymm_new_imag[0],
ymm_new_imag[1],
scale_b=ymm_fft4_twiddle_factor)
SWAP.REGISTERS(ymm_real[0], ymm_new_real[0])
SWAP.REGISTERS(ymm_real[1], ymm_new_real[1])
SWAP.REGISTERS(ymm_imag[0], ymm_new_imag[0])
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag[1])
# w[0] = x0 x4 x2 x6 x8 x12 x10 x14
# w[1] = x1 x5 x3 x7 x9 x11 x13 x15
if bit_reversal:
# Bit reversal
ymm_bit_reversal_mask = YMMRegister()
VMOVDQA(ymm_bit_reversal_mask,
Constant.uint32x8(0, 4, 1, 5, 2, 6, 3, 7))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
for i in range(2):
VPERMPS(ymm_real[i], ymm_bit_reversal_mask, ymm_real[i])
VPERMPS(ymm_imag[i], ymm_bit_reversal_mask, ymm_imag[i])
def ifft16_within_rows(ymm_real_rows, ymm_imag_rows, bit_reversal=True):
if isinstance(ymm_real_rows, tuple) and isinstance(ymm_imag_rows, tuple):
return ifft16_within_rows([ymm_real_rows], [ymm_imag_rows])
assert isinstance(ymm_real_rows, list) and all(isinstance(ymm_real, tuple) and all(isinstance(ymm, YMMRegister) for ymm in ymm_real) for ymm_real in ymm_real_rows)
assert isinstance(ymm_imag_rows, list) and all(isinstance(ymm_imag, tuple) and all(isinstance(ymm, YMMRegister) for ymm in ymm_imag) for ymm_imag in ymm_imag_rows)
if bit_reversal:
# Bit reversal
# w[0] = x0 x8 x4 x12 x2 x10 x6 x14
# w[1] = x1 x9 x5 x13 x3 x11 x7 x15
ymm_bit_reversal_mask = YMMRegister()
VMOVDQA(ymm_bit_reversal_mask, Constant.uint32x8(0, 2, 4, 6, 1, 3, 5, 7))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
for i in range(2):
VPERMPS(ymm_real[i], ymm_bit_reversal_mask, ymm_real[i])
VPERMPS(ymm_imag[i], ymm_bit_reversal_mask, ymm_imag[i])
# 8x FFT2: Butterfly
# w[0] = x0 x4 x2 x6 x8 x12 x10 x14
# w[1] = x1 x5 x3 x7 x9 x13 x11 x15
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
ymm_new_real = YMMRegister(), YMMRegister()
VUNPCKLPS(ymm_new_real[0], ymm_real[0], ymm_real[1])
VUNPCKHPS(ymm_new_real[1], ymm_real[0], ymm_imag[1])
ymm_new_imag = YMMRegister(), YMMRegister()
VUNPCKLPS(ymm_new_imag[0], ymm_imag[0], ymm_imag[1])
VUNPCKHPS(ymm_new_imag[1], ymm_imag[0], ymm_real[1])
SWAP.REGISTERS(ymm_imag[0], ymm_new_imag[0])
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag[1])
SWAP.REGISTERS(ymm_real[0], ymm_new_real[0])
SWAP.REGISTERS(ymm_real[1], ymm_new_real[1])
# w[0] = x0 x1 x4 x5 x8 x9 x12 x13
# w[1] = x2 x3 x6 x7 x10 x11 x14 x15
# 4x FFT4: Butterfly and multiplication by twiddle factors
ymm_fft4_twiddle_factor = YMMRegister()
VMOVAPS(ymm_fft4_twiddle_factor, Constant.float32x8(+1.0, -1.0, +1.0, -1.0, +1.0, -1.0, +1.0, -1.0))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1], scale_b=ymm_fft4_twiddle_factor)
butterfly(ymm_imag[0], ymm_imag[1])
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x2x64(ymm_real[0], ymm_real[1])
transpose2x2x2x64(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x2 x3 x8 x9 x10 x11
# w[1] = x4 x5 x6 x7 x12 x13 x14 x15
# 2x FFT8: Multiplication by twiddle factors
ymm_fft8_cos_twiddle_factor, ymm_fft8_sin_twiddle_factor = YMMRegister(), YMMRegister()
VMOVAPS(ymm_fft8_cos_twiddle_factor, Constant.float32x8(*(cos_npi_over_4 * 2)))
VMOVAPS(ymm_fft8_sin_twiddle_factor, Constant.float32x8(*(sin_npi_over_4 * 2)))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft8_cos_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft8_cos_twiddle_factor)
VFNMADD231PS(ymm_new_real1, ymm_imag[1], ymm_fft8_sin_twiddle_factor)
VFMADD231PS(ymm_new_imag1, ymm_real[1], ymm_fft8_sin_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# 2x FFT8: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x128(ymm_real[0], ymm_real[1])
transpose2x2x128(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x2 x3 x4 x5 x6 x7
# w[1] = x8 x9 x10 x11 x12 x13 x14 x15
# FFT16: Multiplication by twiddle factors and scale
scale_factor = 0.0625
ymm_fft16_cos_scale_twiddle_factor, ymm_fft16_sin_scale_twiddle_factor = YMMRegister(), YMMRegister()
VMOVAPS(ymm_fft16_cos_scale_twiddle_factor, Constant.float32x8(*[cos * scale_factor for cos in cos_npi_over_8]))
VMOVAPS(ymm_fft16_sin_scale_twiddle_factor, Constant.float32x8(*[sin * scale_factor for sin in sin_npi_over_8]))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft16_cos_scale_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft16_cos_scale_twiddle_factor)
VFNMADD231PS(ymm_new_real1, ymm_imag[1], ymm_fft16_sin_scale_twiddle_factor)
VFMADD231PS(ymm_new_imag1, ymm_real[1], ymm_fft16_sin_scale_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# FFT16: Butterfly and scale
ymm_scale_factor = YMMRegister()
VMOVAPS(ymm_scale_factor, Constant.float32x8(scale_factor))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1], scale_a=ymm_scale_factor)
butterfly(ymm_imag[0], ymm_imag[1], scale_a=ymm_scale_factor)
def fft16_within_rows(ymm_real_rows, ymm_imag_rows, bit_reversal=True):
if isinstance(ymm_real_rows, tuple) and isinstance(ymm_imag_rows, tuple):
return fft16_within_rows([ymm_real_rows], [ymm_imag_rows])
assert isinstance(ymm_real_rows, list) and all(isinstance(ymm_real, tuple) and all(isinstance(ymm, YMMRegister) for ymm in ymm_real) for ymm_real in ymm_real_rows)
assert isinstance(ymm_imag_rows, list) and all(isinstance(ymm_imag, tuple) and all(isinstance(ymm, YMMRegister) for ymm in ymm_imag) for ymm_imag in ymm_imag_rows)
# FFT16: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
# FFT16: Multiplication by twiddle factors
ymm_fft16_cos_twiddle_factor, ymm_fft16_sin_twiddle_factor = YMMRegister(), YMMRegister()
VMOVAPS(ymm_fft16_cos_twiddle_factor, Constant.float32x8(*cos_npi_over_8))
VMOVAPS(ymm_fft16_sin_twiddle_factor, Constant.float32x8(*sin_npi_over_8))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft16_cos_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft16_cos_twiddle_factor)
VFMADD231PS(ymm_new_real1, ymm_imag[1], ymm_fft16_sin_twiddle_factor)
VFNMADD231PS(ymm_new_imag1, ymm_real[1], ymm_fft16_sin_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# 2x FFT8: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x128(ymm_real[0], ymm_real[1])
transpose2x2x128(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x2 x3 x8 x9 x10 x11
# w[1] = x4 x5 x6 x7 x12 x13 x14 x15
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
# 2x FFT8: Multiplication by twiddle factors
ymm_fft8_cos_twiddle_factor, ymm_fft8_sin_twiddle_factor = YMMRegister(), YMMRegister()
VMOVAPS(ymm_fft8_cos_twiddle_factor, Constant.float32x8(*(cos_npi_over_4 * 2)))
VMOVAPS(ymm_fft8_sin_twiddle_factor, Constant.float32x8(*(sin_npi_over_4 * 2)))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real1, ymm_new_imag1 = YMMRegister(), YMMRegister()
VMULPS(ymm_new_real1, ymm_real[1], ymm_fft8_cos_twiddle_factor)
VMULPS(ymm_new_imag1, ymm_imag[1], ymm_fft8_cos_twiddle_factor)
VFMADD231PS(ymm_new_real1, ymm_imag[1], ymm_fft8_sin_twiddle_factor)
VFNMADD231PS(ymm_new_imag1, ymm_real[1], ymm_fft8_sin_twiddle_factor)
SWAP.REGISTERS(ymm_real[1], ymm_new_real1)
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag1)
# 4x FFT4: Butterfly
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
transpose2x2x2x64(ymm_real[0], ymm_real[1])
transpose2x2x2x64(ymm_imag[0], ymm_imag[1])
# w[0] = x0 x1 x4 x5 x8 x9 x12 x13
# w[1] = x2 x3 x6 x7 x10 x11 x14 x15
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
butterfly(ymm_real[0], ymm_real[1])
butterfly(ymm_imag[0], ymm_imag[1])
# 4x FFT4: Multiplication by twiddle factors and 8x FFT2: Butterfly
ymm_fft4_twiddle_factor = YMMRegister()
VMOVAPS(ymm_fft4_twiddle_factor, Constant.float32x8(+1.0, +1.0, -1.0, -1.0, +1.0, +1.0, -1.0, -1.0))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
ymm_new_real = YMMRegister(), YMMRegister()
VSHUFPS(ymm_new_real[0], ymm_real[0], ymm_real[1], _MM_SHUFFLE(2, 0, 2, 0))
VSHUFPS(ymm_new_real[1], ymm_real[0], ymm_imag[1], _MM_SHUFFLE(3, 1, 3, 1))
butterfly(ymm_new_real[0], ymm_new_real[1])
ymm_new_imag = YMMRegister(), YMMRegister()
VSHUFPS(ymm_new_imag[0], ymm_imag[0], ymm_imag[1], _MM_SHUFFLE(2, 0, 2, 0))
VSHUFPS(ymm_new_imag[1], ymm_imag[0], ymm_real[1], _MM_SHUFFLE(3, 1, 3, 1))
butterfly(ymm_new_imag[0], ymm_new_imag[1], scale_b=ymm_fft4_twiddle_factor)
SWAP.REGISTERS(ymm_real[0], ymm_new_real[0])
SWAP.REGISTERS(ymm_real[1], ymm_new_real[1])
SWAP.REGISTERS(ymm_imag[0], ymm_new_imag[0])
SWAP.REGISTERS(ymm_imag[1], ymm_new_imag[1])
# w[0] = x0 x4 x2 x6 x8 x12 x10 x14
# w[1] = x1 x5 x3 x7 x9 x11 x13 x15
if bit_reversal:
# Bit reversal
ymm_bit_reversal_mask = YMMRegister()
VMOVDQA(ymm_bit_reversal_mask, Constant.uint32x8(0, 4, 1, 5, 2, 6, 3, 7))
for ymm_real, ymm_imag in zip(ymm_real_rows, ymm_imag_rows):
for i in range(2):
VPERMPS(ymm_real[i], ymm_bit_reversal_mask, ymm_real[i])
VPERMPS(ymm_imag[i], ymm_bit_reversal_mask, ymm_imag[i])