def elementwise_add(op: Sgemm) -> List[Kernel]:
A = op.inputs["A"]
B = op.inputs["B"]
C = op.outputs["C"]
assert ChannelMode.get_mode(A) == ChannelMode.get_mode(B)
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"A": A,
"B": B,
"s_c": texture_stride(C),
"d_C": [op.M, op.N],
"s_C": [op.N, 1],
"d_a": texture_shape(A),
"s_a": texture_stride(A),
"s_A": [op.K, 1] if op.transpose_A else [1, op.M],
"d_b": texture_shape(B),
"s_b": texture_stride(B),
"s_B": [op.N, 1] if op.transpose_B else [1, op.K],
"K": op.K
})
source = generate_template(mode=ChannelMode.get_mode(A), K=op.K)
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, C)
return [kernel]
def elementwise_add(op: Tanh) -> List[Kernel]:
x0 = op.inputs["x0"]
y = op.outputs["y"]
shapes, strides = optimize_loop_structure([x0, y], y)
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"X0": x0,
"s_y": texture_stride(y),
"d_Y": shapes[y],
"s_Y": strides[y],
"d_x0": texture_shape(x0),
"s_x0": texture_stride(x0),
"d_X0": shapes[x0],
"s_X0": strides[x0],
})
source = template_R if ChannelMode.get(
y) == ChannelModeEnum.R else template_RGBA
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def convert_rgba_to_r(op: ConvertRGBAtoR) -> List[Kernel]:
x0 = op.inputs["x0"]
y = op.outputs["y"]
assert ChannelMode.get(x0) == ChannelModeEnum.RGBA
assert ChannelMode.get(y) == ChannelModeEnum.R
if x0.order != y.order:
raise NotImplementedError
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"X0": x0,
"s_y": texture_stride(y),
"d_x0": texture_shape(x0),
"s_x0": texture_stride(x0),
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def _generate_template(op: Reduce, reduction_size: int, shapes: Dict[Variable, Sequence[int]], strides: Dict[Variable, Sequence[int]]):
x = op.inputs["x"]
y = op.outputs["y"]
params = []
for key, callable in _registered_items[op.__class__].parameters.items():
value = callable(op)
params.append(GlobalDeclarationNode(VType.Float if isinstance(value, float) else VType.Int, key, value=value, with_value=True))
return KernelCode([f"""
void main() {{
ivec4 variable_position_y = """,
convert_position("gl_FragCoord.yx", texture_shape(y)[:2], texture_stride(y)[:2], shapes[y], strides[y]), f""";
ivec4 variable_position_x = mod(variable_position_y, """, ivec(shapes[x]), f""");
const int n_x = {reduction_size};
float y;
""", params, f"""
""", _registered_items[op.__class__].pre_reduction_snippet, f"""
for (int i_x = 0; i_x < {reduction_size}; i_x++) {{
variable_position_x.w = i_x;
float x = texture2D(""", x, ", ",
convert_coord(f"variable_position_x", shapes[x], strides[x], texture_shape(x)[:2][::-1],
texture_stride(x)[:2][::-1]), f""").r;
{{
""", _registered_items[op.__class__].body_snippet, f"""
}}
}}
""", _registered_items[op.__class__].post_reduction_snippet, f"""
gl_FragColor.r = y;
}}
"""], name=op.__class__.__name__)
def average_pooling_2d(op: Unpooling2D) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
assert x.order == OrderNHWC
assert y.order == OrderNHWC
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"X": x,
"s_y": texture_stride(y),
"d_Y": y.shape,
"s_Y": y.stride,
"d_x": texture_shape(x),
"s_x": texture_stride(x),
"s_X": x.stride,
"C1": x.shape_dict[Axis.C],
"H1": x.shape_dict[Axis.H],
"W1": x.shape_dict[Axis.W],
"SH": op.parameters["stride"][0],
"SW": op.parameters["stride"][1],
"PH": op.parameters["padding"][0],
"PW": op.parameters["padding"][1],
})
source = generate_template(ksize=op.parameters["ksize"])
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def convert_rgba_to_r(op: ConvertRGBAtoR) -> List[Kernel]:
x0 = op.inputs["x0"]
y = op.outputs["y"]
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"X0": x0,
"s_y": texture_stride(y),
"d_x0": texture_shape(x0),
"s_x0": texture_stride(x0),
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
source,
name_injector.name,
uniform_injector.samplers,
uniform_injector.uniforms,
y
)
return [kernel]
def reinterpret_axis(op: ReinterpretAxis) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"X": x,
"s_y": texture_stride(y),
"d_x": texture_shape(x),
"s_x": texture_stride(x),
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
source,
name_injector.name,
uniform_injector.samplers,
uniform_injector.uniforms,
y
)
return [kernel]
def convert_rgba_to_r(op: ConvertRGBAtoR) -> List[Kernel]:
x = op.inputs["x0"]
y = op.outputs["y"]
assert ChannelMode.get(x) == ChannelModeEnum.RGBA
assert ChannelMode.get(y) == ChannelModeEnum.R
assert x.order == y.order
# noinspection PyUnresolvedReferences
inv_x_shape = [
np.double(1) / np.double(v) for v in texture_shape(x)[:2][::-1]
]
code = KernelCode([
"""
void main() {
ivec3 texture_position_x = """,
convert_position("gl_FragCoord.yx",
texture_shape(y)[:2],
texture_stride(y)[:2], texture_shape(x),
texture_stride(x)), """;
vec2 texture_coord_x = (vec2(texture_position_x.yx) + 0.5) * """,
vec2(inv_x_shape), """;
vec4 x = texture2D(""", x, """, texture_coord_x);
gl_FragColor.r = texture_position_x.z == 0 ? x.r :
texture_position_x.z == 1 ? x.g :
texture_position_x.z == 2 ? x.b :
x.a;
}
"""
],
name=op.__class__.__name__)
source = code.generate()
return [Kernel(source, code.name, code.samplers, code.uniforms, y)]
def space2depth(op: Space2Depth) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
r = op.parameters['r']
assert x.order == OrderNHWC
assert y.order == OrderNHWC
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"X": x,
"s_y": texture_stride(y),
"d_Y": y.shape,
"s_Y": y.stride,
"d_x": texture_shape(x),
"s_x": texture_stride(x),
"d_X": x.shape,
"s_X": x.stride,
"r": r,
"C1": x.shape_dict[Axis.C],
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def elementwise_add(op: ClippedRelu) -> List[Kernel]:
x0 = op.inputs["x0"]
y = op.outputs["y"]
shapes, strides = optimize_loop_structure([x0, y], y)
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"X0": x0,
"s_y": texture_stride(y),
"d_Y": shapes[y],
"s_Y": strides[y],
"d_x0": texture_shape(x0),
"s_x0": texture_stride(x0),
"d_X0": shapes[x0],
"s_X0": strides[x0],
"cap": op.parameters["cap"]
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def _generate_template_convert_position(op: Elementwise,
shapes: Dict[Variable, Sequence[int]],
strides: Dict[Variable,
Sequence[int]]):
load_nodes = []
y = op.outputs["y"]
for k, v in op.inputs.items():
if shapes[v] == shapes[y]:
load_nodes += [
f"float {k} = texture2D(", v, ", ",
convert_coord(f"variable_position_y", shapes[v], strides[v],
texture_shape(v)[:2][::-1],
texture_stride(v)[:2][::-1]), ").r;\n"
]
else:
load_nodes += [
f"ivec4 variable_position_{k} = mod(variable_position_y, ",
ivec(shapes[v]), f");\n"
]
load_nodes += [
f"float {k} = texture2D(", v, ", ",
convert_coord(f"variable_position_{k}", shapes[v], strides[v],
texture_shape(v)[:2][::-1],
texture_stride(v)[:2][::-1]), ").r;\n"
]
for key, callable in _registered_items[op.__class__].parameters.items():
value = callable(op)
load_nodes.append(
GlobalDeclarationNode(
VType.Float if isinstance(value, float) else VType.Int,
key,
value=value,
with_value=True))
return KernelCode([
"""
void main() {
float y;
ivec4 variable_position_y = """,
convert_position("gl_FragCoord.yx",
texture_shape(y)[:2],
texture_stride(y)[:2], shapes[y], strides[y]),
""";
""", load_nodes, _registered_items[op.__class__].code, """
gl_FragColor = vec4(y, 0, 0, 0);
}
"""
],
name=op.__class__.__name__)
def convert_r_to_rgba(op: ConvertRtoRGBA) -> List[Kernel]:
x = op.inputs["x0"]
y = op.outputs["y"]
assert ChannelMode.get(x) == ChannelModeEnum.R
assert ChannelMode.get(y) == ChannelModeEnum.RGBA
orders, shape_dicts = simplify_orders([x, y])
shapes = {v: [shape_dicts[v][a] for a in orders[v].axes] for v in [x, y]}
strides = {
v:
[mul(shapes[v][orders[v].axes_dict[a] + 1:]) for a in orders[v].axes]
for v in [x, y]
}
stride_dicts = {v: AxisKeyDict(orders[v].axes, strides[v]) for v in [x, y]}
# Change x's shapes and strides order to same as y's order
shapes[x] = [
shape_dicts[x][a] if a in orders[x].axes else 1 for a in orders[y].axes
]
strides[x] = [
stride_dicts[x][a] if a in orders[x].axes else 1
for a in orders[y].axes
]
# Padding shapes and strides to 4D
if orders[y].ndim > 4:
raise NotImplementedError(f"Too large number of dimension: {y}")
for v in [x, y]:
shape = shapes[v]
stride = strides[v]
while len(shape) < 4:
stride.append(1)
shape.append(1)
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"sampler_x": x,
"texture_stride_y": texture_stride(y),
"variable_shape_y": shapes[y],
"variable_stride_y": strides[y],
"texture_shape_x": texture_shape(x),
"texture_stride_x": texture_stride(x),
"variable_shape_x": shapes[x],
"variable_stride_x": strides[x],
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def partial_im2col(op: PartialIm2Col) -> List[Kernel]:
im = op.inputs["im"]
cols = [op.outputs[f"col{i}"] for i in range(len(op.outputs))]
sections = [0] + op.sections
axis = op.axis
kernels = []
for i, col in enumerate(cols):
assert im.order == col.order == OrderNHWC
assert ChannelMode.get(im) == ChannelModeEnum.R
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
offset = [sections[i] if a == axis else 0 for a in col.order.axes]
uniform_injector.register({
"sampler_im": im,
"texture_stride_col": texture_stride(col),
"variable_shape_col": col.shape,
"variable_stride_col": col.stride,
"offset_col": offset,
"texture_shape_im": texture_shape(im),
"texture_stride_im": texture_stride(im),
"variable_shape_im": im.shape,
"variable_stride_im": im.stride,
"C1": im.shape_dict[Axis.C],
"H1": im.shape_dict[Axis.H],
"W1": im.shape_dict[Axis.W],
"KH": op.KH,
"KW": op.KW,
"DH": op.DH,
"DW": op.DW,
"SH": op.SH,
"SW": op.SW,
"PH": op.PH,
"PW": op.PW,
})
source = template_R if ChannelMode.get(col) == ChannelModeEnum.R else template_RGBA
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
source,
name_injector.name,
uniform_injector.samplers,
uniform_injector.uniforms,
col
)
kernels.append(kernel)
return kernels
def split_axis(op: SplitAxis) -> List[Kernel]:
x = op.inputs["x"]
ys = [op.outputs[f"y{i}"] for i in range(len(op.outputs))]
sections = [0] + op.sections
axis = op.axis
kernels = []
for i, y in enumerate(ys):
assert x.order.check_same_axes(y.order)
assert ChannelMode.get(x) == ChannelMode.get(y) == ChannelModeEnum.R
if x.ndim > 4:
# simplify orders
orders, shape_dicts = simplify_orders([x, y], keep_axes=[axis])
shapes = {
v: [shape_dicts[v][a] for a in order.axes]
for v, order in orders.items()
}
strides = {
v: [mul(shapes[v][i + 1:]) for i in range(order.ndim)]
for v, order in orders.items()
}
else:
orders = {y: y.order, x: x.order}
shapes = {y: y.shape, x: x.shape}
strides = {y: y.stride, x: x.stride}
code = KernelCode([
f"""
void main() {{
""",
Type.Ivec.get_name(shapes[x]), f""" variable_position_x = """,
change_order(
convert_position("gl_FragCoord.yx",
texture_shape(y)[:2],
texture_stride(y)[:2], shapes[y], strides[y]),
orders[y], orders[x]), f""";
variable_position_x[{orders[x].axes_dict[axis]}] += {sections[i]};
gl_FragColor.r = texture2D(""", x, ",",
convert_coord("variable_position_x", shapes[x], strides[x],
texture_shape(x)[:2][::-1],
texture_stride(x)[:2][::-1]), f""").r;
}}
"""
],
name=op.__class__.__name__)
source = code.generate()
kernels.append(
Kernel(source, code.name, code.samplers, code.uniforms, y))
return kernels
def concat(op: Concat) -> List[Kernel]:
assert len(op.inputs) == 2
x0 = op.inputs["x0"]
x1 = op.inputs["x1"]
y = op.outputs["y"]
axis = op.axis
assert x0.order.check_same_axes(y.order)
assert x1.order.check_same_axes(y.order)
assert ChannelMode.get(x0) == ChannelMode.get(x1) == ChannelMode.get(y)
if x0.ndim > 4 or x1.ndim > 4:
# simplify orders
orders, shape_dicts = simplify_orders([x0, x1, y], keep_axes=[axis])
shapes = {v: [shape_dicts[v][a] for a in order.axes] for v, order in orders.items()}
strides = {v: [mul(shapes[v][i + 1:]) for i in range(order.ndim)] for v, order in orders.items()}
else:
orders = {y: y.order, x0: x0.order, x1: x1.order}
shape_dicts = {y: y.shape_dict, x0: x0.shape_dict, x1: x1.shape_dict}
shapes = {y: y.shape, x0: x0.shape, x1: x1.shape}
strides = {y: y.stride, x0: x0.stride, x1: x1.stride}
code = KernelCode([f"""
void main() {{
""", Type.Ivec.get_name(shapes[x0]), f""" variable_position_x0 = """, change_order(
convert_position("gl_FragCoord.yx", texture_shape(y)[:2], texture_stride(y)[:2], shapes[y], strides[y]),
orders[y], orders[x0]
), f""";
""", Type.Ivec.get_name(shapes[x1]), f""" variable_position_x1 = """, change_order(
convert_position("gl_FragCoord.yx", texture_shape(y)[:2], texture_stride(y)[:2], shapes[y], strides[y]),
orders[y], orders[x1]
), f""";
variable_position_x1[{orders[x1].axes_dict[axis]}] -= {x0.shape_dict[axis]};
gl_FragColor.r = (
(variable_position_x0[{orders[x0].axes_dict[axis]}] >= {shape_dicts[x0][axis]})
? texture2D(""", x1, ",", convert_coord("variable_position_x1", shapes[x1], strides[x1], texture_shape(x1)[:2][::-1],
texture_stride(x1)[:2][::-1]), f""")
: texture2D(""", x0, ",", convert_coord("variable_position_x0", shapes[x0], strides[x0], texture_shape(x0)[:2][::-1],
texture_stride(x0)[:2][::-1]), f""")
).r;
}}
"""], name=op.__class__.__name__)
source = code.generate()
return [Kernel(
source,
code.name,
code.samplers,
code.uniforms,
y
)]
def elementwise_kernel(op: Elementwise):
xs = list(op.inputs.values())
y = op.outputs["y"]
shapes, strides = _optimize_loop_structure(xs + [y], y)
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"texture_stride_y": texture_stride(y),
"variable_shape_y": shapes[y],
"variable_stride_y": strides[y]
})
for k, v in op.inputs.items():
uniform_injector.register({
f"sampler_{k}": v,
f"texture_shape_{k}": texture_shape(v),
f"texture_stride_{k}": texture_stride(v),
f"variable_shape_{k}": shapes[v],
f"variable_stride_{k}": strides[v],
})
for name, callable in _registered_items[op.__class__].parameters.items():
uniform_injector.register({name: callable(op)})
if all([
x.shape == y.shape and x.order == y.order
and texture_shape(x) == texture_shape(y) for x in xs
]):
# For all variables, not only element position (=logical position), pixel position (=actual position) is also same.
# Therefore computing logical position is no need.
source = _generate_template_no_convert_position(op)
else:
# Computing logical position is required.
source = _generate_template_convert_position(op)
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def split_axis(op: SplitAxis) -> List[Kernel]:
x = op.inputs["x"]
ys = [op.outputs[f"y{i}"] for i in range(len(op.outputs))]
sections = [0] + op.sections
axis = op.axis
kernels = []
for i, y in enumerate(ys):
assert x.order.check_same_axes(y.order)
assert ChannelMode.get(x) == ChannelMode.get(y) == ChannelModeEnum.R
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
offset = [sections[i] if a == axis else 0 for a in y.order.axes]
uniform_injector.register({
"sampler_x":
x,
"texture_stride_y":
texture_stride(y),
"variable_shape_y":
_pad_to_4d(y.shape),
"variable_stride_y":
_pad_to_4d(y.stride),
"texture_shape_x":
texture_shape(x),
"texture_stride_x":
texture_stride(x),
"variable_shape_x":
_pad_to_4d([x.shape_dict[a] for a in y.order.axes]),
"variable_stride_x":
_pad_to_4d([x.stride_dict[a] for a in y.order.axes]),
"offset":
_pad_to_4d(offset, 0)
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
kernels.append(kernel)
return kernels
def col2im(op: Col2Im) -> List[Kernel]:
col = op.inputs["col"]
im = op.outputs["im"]
assert col.order == OrderNHWC
assert im.order == OrderNHWC
assert ChannelMode.get(col) == ChannelModeEnum.R
assert ChannelMode.get(im) == ChannelModeEnum.R
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"col": col,
"s_im": texture_stride(im),
"d_Im": im.shape,
"s_Im": im.stride,
"d_col": texture_shape(col),
"s_col": texture_stride(col),
"d_Col": col.shape,
"s_Col": col.stride,
"H2": col.shape_dict[Axis.H],
"W2": col.shape_dict[Axis.W],
"C1": im.shape_dict[Axis.C],
"SH": op.SH,
"SW": op.SW,
"PH": op.PH,
"PW": op.PW,
})
source = generate_template(op)
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
source,
name_injector.name,
uniform_injector.samplers,
uniform_injector.uniforms,
im
)
return [kernel]
def col2im(op: Col2Im) -> List[Kernel]:
col = op.inputs["col"]
im = op.outputs["im"]
assert col.order.check_same_axes(
Order([Axis.N, Axis.H, Axis.W, Axis.KH, Axis.KW, Axis.C]))
assert col.order.axes_dict[Axis.KH] + 2 == col.order.axes_dict[
Axis.KW] + 1 == col.order.axes_dict[Axis.C] == 5
assert im.order.check_same_axes(OrderNHWC)
assert ChannelMode.get(col) == ChannelModeEnum.R
assert ChannelMode.get(im) == ChannelModeEnum.R
col_shape = col.shape[0:3] + (mul(col.shape[3:6]), )
col_stride = [mul(col_shape[i + 1:]) for i in range(len(col_shape))]
col_order = Order(col.order.axes[0:3] + (Axis.C, ))
code = KernelCode([
"""
void main() {
ivec4 variable_position_im = """,
change_order(get_output_position(im), im.order, OrderNHWC), f""";
int n = variable_position_im.x;
int h1 = variable_position_im.y;
int w1 = variable_position_im.z;
int c1 = variable_position_im.w;
float sum = 0.0;
for (int kh = 0; kh < {op.KH}; kh++) {{
int h2 = (h1 + {op.PH} - kh) / {op.SH};
if (mod(h1 + {op.PH} - kh, {op.SH}) != 0 || h2 < 0 || h2 >= {col.shape_dict[Axis.H]}) continue;
for (int kw = 0; kw < {op.KW}; kw++) {{
int w2 = (w1 + {op.PW} - kw) / {op.SW};
if (mod(w1 + {op.PW} - kw, {op.SW}) != 0 || w2 < 0 || w2 >= {col.shape_dict[Axis.W]}) continue;
int khkwc1 = (kh * {op.KW} + kw) * {im.shape_dict[Axis.C]} + c1;
sum += texture2D(""", col, ",",
convert_coord(
change_order("vec4(n, h2, w2, khkwc1)", OrderNHWC, col_order),
col_shape, col_stride,
texture_shape(col)[:2][::-1],
texture_stride(col)[:2][::-1]), """).r;
}
}
gl_FragColor.r = sum;
}
"""
],
name=op.__class__.__name__)
source = code.generate()
return [Kernel(source, code.name, code.samplers, code.uniforms, im)]
def im2col(op: Im2Col) -> List[Kernel]:
im = op.inputs["im"]
col = op.outputs["col"]
assert im.order == OrderNHWC
assert col.order == OrderNHWC
assert ChannelMode.get(im) == ChannelModeEnum.R
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"im": im,
"s_col": texture_stride(col),
"d_Col": col.shape,
"s_Col": col.stride,
"d_im": texture_shape(im),
"s_im": texture_stride(im),
"d_Im": im.shape,
"s_Im": im.stride,
"C1": im.shape_dict[Axis.C],
"H1": im.shape_dict[Axis.H],
"W1": im.shape_dict[Axis.W],
"KH": op.KH,
"KW": op.KW,
"DH": op.DH,
"DW": op.DW,
"SH": op.SH,
"SW": op.SW,
"PH": op.PH,
"PW": op.PW,
})
source = template_R if ChannelMode.get(
col) == ChannelModeEnum.R else template_RGBA
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, col)
return [kernel]
def convert_r_to_rgba(op: ConvertRtoRGBA) -> List[Kernel]:
x = op.inputs["x0"]
y = op.outputs["y"]
assert ChannelMode.get(x) == ChannelModeEnum.R
assert ChannelMode.get(y) == ChannelModeEnum.RGBA
assert x.order == y.order
shape_x = texture_shape(x)
stride_x = texture_stride(x)
shape_y = texture_shape(y)
stride_y = texture_stride(y)
code = KernelCode([
"""
void main() {
float y0 = texture2D(""", x, ", ",
convert_coord("ivec3(gl_FragCoord.y, gl_FragCoord.x, 0)", shape_y,
stride_y, shape_x, stride_x), """.yx).r;
float y1 = texture2D(""", x, ", ",
convert_coord("ivec3(gl_FragCoord.y, gl_FragCoord.x, 1)", shape_y,
stride_y, shape_x, stride_x), """.yx).r;
float y2 = texture2D(""", x, ", ",
convert_coord("ivec3(gl_FragCoord.y, gl_FragCoord.x, 2)", shape_y,
stride_y, shape_x, stride_x), """.yx).r;
float y3 = texture2D(""", x, ", ",
convert_coord("ivec3(gl_FragCoord.y, gl_FragCoord.x, 3)", shape_y,
stride_y, shape_x, stride_x), """.yx).r;
gl_FragColor = vec4(y0, y1, y2, y3);
}
"""
],
name=op.__class__.__name__)
source = code.generate()
return [Kernel(source, code.name, code.samplers, code.uniforms, y)]
def tensordot(op: Tensordot) -> List[Kernel]:
A = op.inputs["A"]
B = op.inputs["B"]
C = op.outputs["C"]
axes = op.axes
assert ChannelMode.get(A) == ChannelMode.get(B)
assert ChannelMode.get(C) == ChannelModeEnum.R
# Reduced axes must be located on inside of input variables.
assert A.order.axes[-len(axes[0]):] == axes[0]
assert B.order.axes[-len(axes[1]):] == axes[1]
# output variable's axes order must be as [*a_remained_axes, *b_remained_axes]
assert C.order.axes[:A.ndim - len(axes[0])] == A.order.axes[:-len(axes[0])]
assert C.order.axes[-(B.ndim -
len(axes[1])):] == B.order.axes[:-len(axes[1])]
assert C.ndim == A.ndim - len(axes[0]) + B.ndim - len(axes[1])
K = mul(A.shape[-len(axes[0]):])
M = A.size // K
N = B.size // K
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"A": A,
"B": B,
"s_c": texture_stride(C),
"d_C": [M, N],
"s_C": [N, 1],
"d_a": texture_shape(A),
"d_b": texture_shape(B),
"K": K
})
source = generate_template(mode=ChannelMode.get(A), reduction_size=K)
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, C)
return [kernel]
def concat(op: Concat) -> List[Kernel]:
xs = [op.inputs[f"x{i}"] for i in range(len(op.inputs) - 1)]
workspace = op.inputs["workspace"]
y = op.outputs["y"]
axis = op.axis
kernels = []
# noinspection PyUnresolvedReferences
inv_texture_shape_y = [
float(np.double(1.0) / np.double(v))
for v in texture_shape(y)[:2][::-1]
]
# noinspection PyUnresolvedReferences
inv_texture_shape_workspace = [
float(np.double(1.0) / np.double(v))
for v in texture_shape(workspace)[:2][::-1]
]
sections = [0]
for x in xs:
sections.append(sections[-1] + x.shape_dict[axis])
for i, x in enumerate(xs):
assert x.order.check_same_axes(y.order)
assert ChannelMode.get(x) == ChannelMode.get(y)
if x.ndim > 4:
# simplify orders
orders, shape_dicts = simplify_orders([x, y], keep_axes=[axis])
shapes = {
v: [shape_dicts[v][a] for a in order.axes]
for v, order in orders.items()
}
strides = {
v: [mul(shapes[v][i + 1:]) for i in range(order.ndim)]
for v, order in orders.items()
}
else:
orders = {y: y.order, x: x.order}
shape_dicts = {y: y.shape_dict, x: x.shape_dict}
shapes = {y: y.shape, x: x.shape}
strides = {y: y.stride, x: x.stride}
# copy xs[i] or workspace's value into y
code1 = KernelCode([
f"""
void main() {{
""",
Type.Ivec.get_name(shapes[x]), f""" variable_position_x = """,
change_order(
convert_position("gl_FragCoord.yx",
texture_shape(y)[:2],
texture_stride(y)[:2], shapes[y], strides[y]),
orders[y], orders[x]), f""";
variable_position_x[{orders[x].axes_dict[axis]}] -= {sections[i]};
gl_FragColor.r = (
variable_position_x[{orders[x].axes_dict[axis]}] < 0 || variable_position_x[{orders[x].axes_dict[axis]}] >= {shape_dicts[x][axis]}
)
? texture2D(""", workspace, """, gl_FragCoord.xy * """,
inv_texture_shape_workspace, """).r
: texture2D(""", x, ",",
convert_coord("variable_position_x", shapes[x], strides[x],
texture_shape(x)[:2][::-1],
texture_stride(x)[:2][::-1]), f""").r;
}}
"""
],
name="Concat_copy_to_y")
# copy y's value into workspace
code2 = KernelCode([
"""
void main() { gl_FragColor = texture2D(""", y, """, gl_FragCoord.xy * """,
inv_texture_shape_y, """); }
"""
],
name="Concat_escape_to_ws")
source1 = code1.generate()
source2 = code2.generate()
kernels += [
Kernel(source1, code1.name, code1.samplers, code1.uniforms, y),
Kernel(source2, code2.name, code2.samplers, code2.uniforms,
workspace)
]
return kernels
def slice_handler(op: Slice) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
assert ChannelMode.get(x) == ChannelMode.get(y) == ChannelModeEnum.R
x_shape = []
x_stride = []
x_index_offset = 0
y_shape = []
y_stride = []
x_stride_dict = x.stride_dict
y_shape_dict = y.shape_dict
y_stride_dict = y.stride_dict
x_axes = list(x.order.axes)
# reduce number of axis
flag_removed = False
merge_target = None # type: Axis
for axis in reversed(x.order.axes):
if not isinstance(op.indices[axis], slice):
flag_removed = False
merge_target = None
continue
index = normalize_slice(op.indices[axis], x.shape_dict[axis])
if index.start != 0 or index.stop != x.shape_dict[
axis] or index.step != 1:
flag_removed = False
merge_target = None
continue
# This axis is not changed, so it can be simplified
if flag_removed == True:
del x_stride_dict[axis]
x_axes.remove(axis)
del y_stride_dict[axis]
y_shape_dict[merge_target] *= y_shape_dict[axis]
del y_shape_dict[axis]
else:
flag_removed = True
merge_target = axis
for axis in x_axes:
if isinstance(op.indices[axis], slice):
index = normalize_slice(op.indices[axis], x.shape_dict[axis])
x_shape.append(y_shape_dict[axis])
x_stride.append(x_stride_dict[axis] * index.step)
x_index_offset += x_stride_dict[axis] * index.start
y_shape.append(y_shape_dict[axis])
y_stride.append(y_stride_dict[axis])
elif isinstance(op.indices[axis], int):
x_index_offset += x_stride_dict[axis] * op.indices[axis]
if len(y_shape) == 1:
y_shape.append(0)
y_stride.append(1)
x_stride.append(0)
x_shape.append(0)
code = KernelCode([
"""
void main() {
gl_FragColor.r = texture2D(""", x, ", (",
convert_coord(
ExpressionNode([
convert_position("gl_FragCoord.yx",
texture_shape(y)[:2],
texture_stride(y)[:2], y_shape, y_stride)
]), x_shape, x_stride,
texture_shape(x)[:2],
texture_stride(x)[:2], x_index_offset), """).yx).r;
}
"""
],
name="Slice")
source = code.generate()
return [Kernel(source, code.name, code.samplers, code.uniforms, y)]
def reshape(op: Reshape) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
in_order = op.parameters["in_order"]
out_order = op.parameters["out_order"]
dummy_y = Variable(y.shape, y.order).change_order(out_order)
orders_y_dy, shapes_y_dy = simplify_orders([y, dummy_y])
if orders_y_dy[y] == orders_y_dy[dummy_y]:
order = Order([None] * 4)
shape = factorize(y.size)
stride = [mul(shape[i + 1:]) for i in range(4)]
dummy_y = Variable(y.shape, y.order)
shapes_y_dy = {y: shape, dummy_y: shape}
strides_y_dy = {y: stride, dummy_y: stride}
orders_y_dy = {y: order, dummy_y: order}
else:
shapes_y_dy = {v: [shapes_y_dy[v][a] for a in orders_y_dy[v].axes] for v in [y, dummy_y]}
strides_y_dy = {v: [mul(shapes_y_dy[v][i + 1:]) for i in range(orders_y_dy[v].ndim)] for v in [y, dummy_y]}
dummy_x = Variable(x.shape, x.order).change_order(in_order)
orders_x_dx, shapes_x_dx = simplify_orders([x, dummy_x])
if orders_x_dx[x] == orders_x_dx[dummy_x]:
order = Order([None] * 4)
shape = factorize(x.size)
stride = [mul(shape[i + 1:]) for i in range(4)]
dummy_x = Variable(x.shape, x.order)
shapes_x_dx = {x: shape, dummy_x: shape}
strides_x_dx = {x: stride, dummy_x: stride}
orders_x_dx = {x: order, dummy_x: order}
else:
shapes_x_dx = {v: [shapes_x_dx[v][a] for a in orders_x_dx[v].axes] for v in [x, dummy_x]}
strides_x_dx = {v: [mul(shapes_x_dx[v][i + 1:]) for i in range(orders_x_dx[v].ndim)] for v in [x, dummy_x]}
# FIXME: optimize
# y -{change_order}-> dummy_y -{convert_position}-> dummy_x -{change_order}-> x
code = KernelCode([f"""
void main() {{
gl_FragColor.r = texture2D(""", x, """,""", convert_coord(
change_order(
convert_position(
change_order(
convert_position("gl_FragCoord.yx", texture_shape(y)[:2], texture_stride(y)[:2], shapes_y_dy[y], strides_y_dy[y]),
orders_y_dy[y], orders_y_dy[dummy_y]
),
shapes_y_dy[dummy_y], strides_y_dy[dummy_y], shapes_x_dx[dummy_x], strides_x_dx[dummy_x]
),
orders_x_dx[dummy_x], orders_x_dx[x]
),
shapes_x_dx[x], strides_x_dx[x], texture_shape(x)[:2][::-1], texture_stride(x)[:2][::-1]
), f""").r;
}}
"""], name=op.__class__.__name__)
source = code.generate()
return [Kernel(
source,
code.name,
code.samplers,
code.uniforms,
y
)]
def tensordot(op: Tensordot) -> List[Kernel]:
A = op.inputs["A"]
B = op.inputs["B"]
C = op.outputs["C"]
axes = op.axes
assert ChannelMode.get(A) == ChannelMode.get(B)
assert ChannelMode.get(C) == ChannelModeEnum.R
# Reduced axes must be located on inside of input variables.
assert A.order.axes[-len(axes[0]):] == axes[0]
assert B.order.axes[-len(axes[1]):] == axes[1]
# output variable's axes order must be as [*a_remained_axes, *b_remained_axes]
assert C.order.axes[:A.ndim - len(axes[0])] == A.order.axes[:-len(axes[0])]
assert C.order.axes[-(B.ndim -
len(axes[1])):] == B.order.axes[:-len(axes[1])]
assert C.ndim == A.ndim - len(axes[0]) + B.ndim - len(axes[1])
K = mul(A.shape[-len(axes[0]):])
M = A.size // K
N = B.size // K
if ChannelMode.get(A) == ChannelModeEnum.R:
code = KernelCode([
f"""
void main() {{
ivec2 variable_position_c = """,
convert_position("gl_FragCoord.yx",
texture_shape(C)[:2],
texture_stride(C)[:2], [M, N], [N, 1]), f""";
int m = variable_position_c.x;
int n = variable_position_c.y;
float v = 0.0;
for (int k = 0; k < {int(K)}; k++) {{
float v_a = texture2D(""", A, f""", (vec2(k, m) + 0.5) * """,
vec([1.0 / K, 1.0 / M]), f""").r;
float v_b = texture2D(""", B, f""", (vec2(k, n) + 0.5) * """,
vec([1.0 / K, 1.0 / N]), f""").r;
v += v_a * v_b;
}}
gl_FragColor.r = v;
}}
"""
],
name="Tensordot_R")
elif ChannelMode.get(A) == ChannelModeEnum.RGBA:
code = KernelCode([
f"""
void main() {{
ivec2 variable_position_c = """,
convert_position("gl_FragCoord.yx",
texture_shape(C)[:2],
texture_stride(C)[:2], [M, N], [N, 1]), f""";
int m = variable_position_c.x;
int n = variable_position_c.y;
float v = 0.0;
for (int k = 0; k < {int(K // 4)}; k++) {{
vec4 v_a = texture2D(""", A, f""", (vec2(k, m) + 0.5) * """,
vec([1.0 / (K // 4), 1.0 / M]), f""");
vec4 v_b = texture2D(""", B, f""", (vec2(k, n) + 0.5) * """,
vec([1.0 / (K // 4), 1.0 / N]), f""");
v += dot(v_a, v_b);
}}
gl_FragColor.r = v;
}}
"""
],
name="Tensordot_RGBA")
else:
raise NotImplementedError
source = code.generate()
kernel = Kernel(source, code.name, code.samplers, code.uniforms, C)
return [kernel]
def concat(op: Concat) -> List[Kernel]:
xs = [op.inputs[f"x{i}"] for i in range(len(op.inputs) - 1)]
workspace = op.inputs["workspace"]
y = op.outputs["y"]
axis = op.axis
kernels = []
sections = [0]
for x in xs[1:]:
sections.append(sections[-1] + x.shape_dict[axis])
for i, x in enumerate(xs):
assert x.order.check_same_axes(y.order)
assert ChannelMode.get(x) == ChannelMode.get(y)
offset = [sections[i] if a == axis else 0 for a in y.order.axes]
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"sampler_x": x,
"sampler_workspace": workspace,
"texture_shape_workspace": texture_shape(workspace),
"texture_stride_y": texture_stride(y),
"variable_shape_y": _pad_to_4d(y.shape),
"variable_stride_y": _pad_to_4d(y.stride),
"texture_shape_x": texture_shape(x),
"texture_stride_x": texture_stride(x),
"variable_shape_x": _pad_to_4d([x.shape_dict[a] for a in y.order.axes]),
"variable_stride_x": _pad_to_4d([x.stride_dict[a] for a in y.order.axes]),
"offset": _pad_to_4d(offset, 0)
})
source = template
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
source,
name_injector.name,
uniform_injector.samplers,
uniform_injector.uniforms,
y
)
kernels.append(kernel)
name_injector2 = KernelNameInjector(op)
uniform_injector2 = UniformInjector()
uniform_injector2.register({
"sampler_y": y,
"texture_shape_y": texture_shape(y),
})
source2 = template2
source2 = uniform_injector2.inject(source2)
source2 = name_injector2.inject(source2)
kernel2 = Kernel(
source2,
name_injector2.name,
uniform_injector2.samplers,
uniform_injector2.uniforms,
workspace
)
kernels.append(kernel2)
return kernels
def reduce_kernel(op: Reduce):
x = op.inputs["x"]
y = op.outputs["y"]
axis = op.axis
orders, shape_dicts = simplify_orders([x, y], keep_axes=[axis])
# Padding shapes and strides to 4D
if orders[y].ndim > 4:
raise NotImplementedError(f"Too large number of dimension: {y}")
shapes = {v: [shape_dicts[v][a] for a in orders[v].axes] for v in [x, y]}
strides = {
v:
[mul(shapes[v][orders[v].axes_dict[a] + 1:]) for a in orders[v].axes]
for v in [x, y]
}
stride_dicts = {v: AxisKeyDict(orders[v].axes, strides[v]) for v in [x, y]}
# Change x's shapes and strides order to same as y's order
x_virtual_shape = [
shape_dicts[x][a] if a in orders[x].axes else 1 for a in orders[y].axes
]
x_virtual_stride = [
stride_dicts[x][a] if a in orders[x].axes else 1
for a in orders[y].axes
]
while len(x_virtual_shape) < 3:
x_virtual_stride.append(1)
x_virtual_shape.append(stride_dicts[x][axis])
x_virtual_shape.append(shape_dicts[x][axis])
x_virtual_stride.append(stride_dicts[x][axis])
y_virtual_shape = shapes[y]
y_virtual_stride = strides[y]
while len(y_virtual_shape) < 4:
y_virtual_stride.append(1)
y_virtual_shape.append(1)
name_injector = KernelNameInjector(op)
uniform_injector = UniformInjector()
uniform_injector.register({
"texture_stride_y": texture_stride(y),
"variable_shape_y": y_virtual_shape,
"variable_stride_y": y_virtual_stride,
f"sampler_x": x,
f"texture_shape_x": texture_shape(x),
f"texture_stride_x": texture_stride(x),
f"variable_shape_x": x_virtual_shape,
f"variable_stride_x": x_virtual_stride,
})
for name, callable in _registered_items[op.__class__].parameters.items():
uniform_injector.register({name: callable(op)})
# Computing logical position is required.
source = _generate_template_convert_position(
op, reduction_size=shape_dicts[x][axis])
source = uniform_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(source, name_injector.name, uniform_injector.samplers,
uniform_injector.uniforms, y)
return [kernel]
def im2col(op: Im2Col) -> List[Kernel]:
im = op.inputs["im"]
col = op.outputs["col"]
H1 = im.shape_dict[Axis.H]
W1 = im.shape_dict[Axis.W]
C1 = im.shape_dict[Axis.C]
assert col.order.check_same_axes(
Order([Axis.N, Axis.H, Axis.W, Axis.KH, Axis.KW, Axis.C]))
assert col.order.axes_dict[Axis.KH] + 2 == col.order.axes_dict[
Axis.KW] + 1 == col.order.axes_dict[Axis.C] == 5
assert im.order.check_same_axes(OrderNHWC)
assert ChannelMode.get(im) == ChannelModeEnum.R
col_shape = col.shape[0:3] + (mul(col.shape[3:6]), )
col_stride = [mul(col_shape[i + 1:]) for i in range(len(col_shape))]
col_order = Order(col.order.axes[0:3] + (Axis.C, ))
if ChannelMode.get(col) == ChannelModeEnum.R:
code = KernelCode([
"""
void main() {
ivec4 variable_position_col = """,
change_order(
convert_position("gl_FragCoord.yx",
texture_shape(col)[:2],
texture_stride(col)[:2], col_shape,
col_stride), col_order, OrderNHWC), f""";
int n = variable_position_col.x;
int h2 = variable_position_col.y;
int w2 = variable_position_col.z;
int khkwc1 = variable_position_col.w;
int kh = khkwc1 / {C1} / {op.KW};
int kw = khkwc1 / {C1} - kh * {op.KW};
int c1 = khkwc1 - (kh * {op.KW} + kw) * {C1};
int h1 = h2 * {op.SH} - {op.PH} + kh * {op.DH};
int w1 = w2 * {op.SW} - {op.PW} + kw * {op.DW};
if (h1 < 0 || h1 >= {H1} || w1 < 0 || w1 >= {W1}) {{
gl_FragColor.r = 0.0;
}} else {{
gl_FragColor.r = """,
texel_fetch(
im, change_order("vec4(n, h1, w1, c1)", OrderNHWC, im.order)),
f""".r;
}}
}}
"""
],
name="Im2Col_R")
elif ChannelMode.get(col) == ChannelModeEnum.RGBA:
code = KernelCode([
"""
void main() {
ivec4 variable_position_col = """,
change_order(
convert_position("gl_FragCoord.yx",
texture_shape(col)[:2],
texture_stride(col)[:2], col_shape,
col_stride), col_order, OrderNHWC), f""";
int n = variable_position_col.x;
int h2 = variable_position_col.y;
int w2 = variable_position_col.z;
int khkwc1 = variable_position_col.w;
int kh = khkwc1 / {C1} / {op.KW};
int kw = khkwc1 / {C1} - kh * {op.KW};
int c1 = khkwc1 - (kh * {op.KW} + kw) * {C1};
int h1 = h2 * {op.SH} - {op.PH} + kh * {op.DH};
int w1 = w2 * {op.SW} - {op.PW} + kw * {op.DW};
if (h1 < 0 || h1 >= {H1} || w1 < 0 || w1 >= {W1}) {{
gl_FragColor = vec4(0.0, 0.0, 0.0, 0.0);
}} else {{
gl_FragColor.r = """,
texel_fetch(
im, change_order("vec4(n, h1, w1, c1 + 0)", OrderNHWC,
im.order)), f""".r;
gl_FragColor.g = """,
texel_fetch(
im, change_order("vec4(n, h1, w1, c1 + 1)", OrderNHWC,
im.order)), f""".r;
gl_FragColor.b = """,
texel_fetch(
im, change_order("vec4(n, h1, w1, c1 + 2)", OrderNHWC,
im.order)), f""".r;
gl_FragColor.a = """,
texel_fetch(
im, change_order("vec4(n, h1, w1, c1 + 3)", OrderNHWC,
im.order)), f""".r;
}}
}}
"""
],
name="Im2Col_RGBA")
else:
raise NotImplementedError
source = code.generate()
return [Kernel(source, code.name, code.samplers, code.uniforms, col)]
