def gen_kernels(self, runtime, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w, dil_d, dil_h, dil_w):
self.I = TensorDescriptionWrapper(self.I, len(self.I.shape))
self.F = TensorDescriptionWrapper(self.F, len(self.F.shape))
self.O = TensorDescriptionWrapper(self.O, len(self.O.shape))
self.flex_entry_I = self.I.flex_entry()
self.flex_entry_F = self.F.flex_entry()
self.flex_entry_O = self.O.flex_entry()
vec_size = 4 if self.dtype.itemsize == 4 else 8
assert N % 32 == 0, "N dim must be multiple of 32"
assert K % vec_size == 0, "K dim must be multiple of %d" % vec_size
if self.dtype.type == "flex":
clss = "fconv"
else:
raise TypeError("Type not supported.")
self.C = C
self.K = K
self.M = M
self.P = P
self.Q = Q
self.NCK = (N, C, K)
self.TRS = (T, R, S)
self.DHW = (D, H, W)
self.MPQ = (M, P, Q)
self.padding = (pad_d, pad_h, pad_w)
self.strides = (str_d, str_h, str_w)
self.all_params = (N, C, K, D, H, W, T, R, S, pad_d, pad_h, pad_w, str_d, str_h, str_w)
self.dimI = (C, D, H, W, N)
self.dimF = (C, T, R, S, K)
self.dimF = (K, T, R, S, C)
self.dimO = (K, M, P, Q, N)
self.dimI2 = (C * D * H * W, N)
self.dimF2 = (C * T * R * S, K)
self.dimF2t = (K, C * T * R * S)
self.dimO2 = (K * M * P * Q, N)
self.dimS = (K, 1)
self.sizeI = reduce(mul, self.dimI, 1)
self.sizeF = reduce(mul, self.dimF, 1)
self.sizeO = reduce(mul, self.dimO, 1)
self.nOut = reduce(mul, self.MPQ, 1) * K
# precompute some multiplications for fast constant memory access
WN = W * N
HWN = H * WN
DHWN = D * HWN
RS = R * S
RST = T * RS
CRST = C * RST
KRST = K * RST
PQ = P * Q
PQM = M * PQ
QN = Q * N
PQN = P * QN
MPQN = M * PQN
if CRST > 2**16:
assert CRST < 2**16, "Integer division is faster with 16bit numerators"
# precompute the magic numbers and shift amounts for integer division
magic_PQ = _magic64(PQ)
magic_Q = _magic64(Q)
magic_RS = _magic32(RST + 32, RS)
magic_S = _magic32(RS + 32, S)
# flop count for benchmarking
self.flops = PQM * K * N * CRST * 2.0
tile_N = 128 if N > 64 else 64
grid_N = _grid_dim(tile_N, N)
tiles_CK = (128, 64, 32) if tile_N == 128 else (128, 64)
# FPROP #
self.fprop_kernels = kernel_specs.xprop_conv_kernels(
clss, "fprop", "K", tile_N, grid_N, K, tiles_CK, PQM, RST,
_flatten([N, K, D, H, W, WN, HWN, DHWN,
C, KRST, RST, RS, magic_RS, S, magic_S,
pad_d, pad_h, pad_w, str_d, str_h, str_w,
Q, PQ, QN, PQN, MPQN, magic_Q, magic_PQ]))
# shared lookup table size
self.fprop_lut_size = RST * 4 * 2
# Set to 5 for the current T1000 HW config
self.trunc_rows = 32
flags = self.trunc_rows << 8
self.kernels = []
for kernel in self.fprop_kernels:
# TODO: Populate alpha and beta parameters (in a separate loop!).
# alpha (used to be params[6]) will be multiplied with
self.kernels.append([
kernel_specs.get_kernel(kernel[0]), kernel[1], kernel[2], None,
0, self.O, self.I, self.F, 1.0, 0.0, flags,
kernel[3]] + kernel[4])
for kernel in self.kernels:
kernel.extend((FlexPtrDescription(self.flex_entry_O), 1.0))
kernel[10] &= 0xfffffffe # Enable output flag
# record output flex id for autoflex
self.output_flex_ids = [self.flex_entry_O.flex_id]
def gen_kernels(self, runtime, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w, dil_d, dil_h, dil_w):
self.E = TensorDescriptionWrapper(self.E, len(self.E.shape))
self.F = TensorDescriptionWrapper(self.F, len(self.F.shape))
self.O = TensorDescriptionWrapper(self.O, len(self.O.shape))
self.flex_entry_E = self.E.flex_entry()
self.flex_entry_F = self.F.flex_entry()
self.flex_entry_O = self.O.flex_entry()
F_size = int(np.prod(self.F.shape) * 2)
O_size = int(np.prod(self.O.shape) * 2)
vec_size = 4 if self.dtype.itemsize == 4 else 8
assert N % 32 == 0, "N dim must be multiple of 32"
assert K % vec_size == 0, "K dim must be multiple of %d" % vec_size
if self.dtype.type == "flex":
clss = "fconv"
else:
raise TypeError("Type not supported.")
self.C = C
self.K = K
self.M = M
self.P = P
self.Q = Q
self.NCK = (N, C, K)
self.TRS = (T, R, S)
self.DHW = (D, H, W)
self.MPQ = (M, P, Q)
self.padding = (pad_d, pad_h, pad_w)
self.strides = (str_d, str_h, str_w)
self.all_params = (N, C, K, D, H, W, T, R, S, pad_d, pad_h, pad_w, str_d, str_h, str_w)
self.dimI = (C, D, H, W, N)
self.dimF = (C, T, R, S, K)
self.dimFb = (K, T, R, S, C)
self.dimO = (K, M, P, Q, N)
self.dimI2 = (C * D * H * W, N)
self.dimF2 = (C * T * R * S, K)
self.dimF2t = (K, C * T * R * S)
self.dimO2 = (K * M * P * Q, N)
self.dimS = (K, 1)
self.sizeI = reduce(mul, self.dimI, 1)
self.sizeF = reduce(mul, self.dimF, 1)
self.sizeO = reduce(mul, self.dimO, 1)
self.nOut = reduce(mul, self.MPQ, 1) * K
# precompute some multiplications for fast constant memory access
HW = H * W
DHW = D * HW
WN = W * N
HWN = H * WN
DHWN = D * HWN
RS = R * S
RST = T * RS
CRST = C * RST
PQ = P * Q
PQM = M * PQ
QN = Q * N
PQN = P * QN
MPQN = M * PQN
if CRST > 2**16:
assert CRST < 2**16, "Integer division is faster with 16bit numerators"
# precompute the magic numbers and shift amounts for integer division
magic_HW = _magic64(HW)
magic_W = _magic64(W)
magic_PQ = _magic64(PQ)
magic_Q = _magic64(Q)
magic_RST = _magic32(CRST, RST)
magic_RS = _magic32(RST + 32, RS)
magic_S = _magic32(RS + 32, S)
magic_str_w = _magic32(W + S, str_w)
magic_str_h = _magic32(H + R, str_h)
magic_str_d = _magic32(D + T, str_d)
# flop count for benchmarking
self.flops = PQM * K * N * CRST * 2.0
tile_N = 128 if N > 64 else 64
grid_N = _grid_dim(tile_N, N)
tiles_CK = (128, 64, 32) if tile_N == 128 else (128, 64)
# BPROP #
if C < 16 or C % vec_size != 0:
# special kernel for deconv into first layer
kernel_name = "%s_bprop_C1_N64" % clss
grid = (PQM, _grid_dim(32, CRST), _grid_dim(64, N))
block = (32, 1, 1)
self.bprop_kernels = [[kernel_name, grid, block, 0, _flatten([
N, K, D, H, W, WN, HWN, DHWN,
C, CRST, RST, magic_RST, RS, magic_RS, S, magic_S,
pad_d, pad_h, pad_w, str_d, str_h, str_w,
Q, PQ, QN, PQN, MPQN, magic_Q, magic_PQ,
CRST * 8 * self.dtype.itemsize, MPQN * 8 * self.dtype.itemsize])]]
# generate the kernel args for transpose CRST,K => K,CRST
self.shuffle_args = [CRST, K]
gridX = (K >> 5) + (K & 31 != 0)
gridY = (CRST >> 5) + (CRST & 31 != 0)
self.shuffle_grid = (gridX, gridY, 1)
self.shuffle_block = (32, 8, 1)
self.bprop_zero = self.sizeI * self.dtype.itemsize
self.bprop_lut_size = 0
else:
self.bprop_kernels = kernel_specs.xprop_conv_kernels(
clss, "bprop", "C", tile_N, grid_N, C, tiles_CK, DHW, RST, _flatten([
N, C, M, P, Q, QN, PQN, MPQN,
K, CRST, RST, RS, magic_RS, S, magic_S,
pad_d, pad_h, pad_w, str_d, str_h, str_w,
W, HW, WN, HWN, DHWN, magic_W, magic_HW,
R, T, magic_str_w, magic_str_h, magic_str_d]))
# generate the kernel args for dim shuffling CRSTK => KRSTC
self.shuffle_args = _flatten([
RST * K, RS * K, S * K, K,
RST * C, RS * C, S * C, C,
RS, magic_RS, S, magic_S])
gridX = (K >> 5) + (K & 31 != 0)
gridY = (C >> 5) + (C & 31 != 0)
self.shuffle_grid = (gridX, gridY, RST)
self.shuffle_block = (32, 8, 1)
self.bprop_zero = 0
self.bprop_lut_size = RST * 4 * 2
# Set to 5 for the current T1000 HW config
self.trunc_rows = 32
flags = self.trunc_rows << 8
# Must dim shuffle filter data for bprop kernel
F_data = ScratchBufferWrapper(F_size, 0, runtime)
if self.bprop_zero:
Out = ScratchBufferWrapper(O_size, F_size, runtime)
shuffle_kernel = _get_transpose_kernel(self.dtype)
else:
Out = self.O
# can point to transpose or dimshuffle kernel
shuffle_kernel = _get_shuffle_kernel(self.dtype)
shuffle_args = [self.shuffle_grid, self.shuffle_block, None,
F_data, self.F] + self.shuffle_args
shuffle_kernel = [shuffle_kernel] + shuffle_args
# Have to zero output buffer and use type conversion for kernel using atomics
if self.bprop_zero:
shape = [int(np.prod(self.O.shape[:-1])), self.O.shape[-1]]
convert_kernel = _prepare_convert_kernel(Out, "f2", self.O, shape,
FlexPtrDescription(self.flex_entry_O))
self.convert_out = True
else:
self.convert_out = False
self.kernels = []
for kernel in self.bprop_kernels:
# TODO: Populate alpha and beta parameters (in a separate loop!).
# alpha (used to be params[6]) will be multiplied with
self.kernels.append([
kernel_specs.get_kernel(kernel[0]), kernel[1], kernel[2], None,
0, Out, self.E, F_data, 1.0, 0.0, flags, kernel[3]] + kernel[4])
for kernel in self.kernels:
kernel.extend((FlexPtrDescription(self.flex_entry_O), 1.0))
kernel[10] &= 0xfffffffe # Enable output flag
self.kernels = [shuffle_kernel] + self.kernels
if self.convert_out:
self.kernels.append(convert_kernel)
# record output flex id for autoflex
self.output_flex_ids = [self.flex_entry_O.flex_id]