def test_detection_type_bools(self):
glsl = """
#version 450
#extension GL_ARB_separate_shader_objects : enable
layout({layout}, binding = 0) buffer Buffer {{
bool var1;
bool[720][1280] var2;
bvec2 var3;
bvec3 var4;
bvec4 var5;
bvec3[5] var6;
}};
void main() {{}}
"""
for layout in (Layout.STD140, Layout.STD430):
# the vulkan spir-v compiler turns bools into uints
expected_definition = Struct([
Scalar.uint(),
Array(Scalar.uint(), (720, 1280), layout),
Vector.uvec2(),
Vector.uvec3(),
Vector.uvec4(),
Array(Vector.uvec3(), 5, layout)
], layout)
shader = self.shader_from_txt(glsl.format(layout=layout), verbose=False)
shader.inspect()
detected_definition, _ = shader.code.get_block(0)
equal = expected_definition.compare(detected_definition, quiet=True)
self.assertTrue(equal)
def test_pass_through_array_of_vectors(self):
glsl_template = """
#version 450
#extension GL_ARB_separate_shader_objects : enable
layout(local_size_x=1, local_size_y=1, local_size_z=1) in;
layout({layout_in}, binding = 0) buffer BufferA {{
{dtype}[720][1280] imageIn;
}};
layout({layout_out}, binding = 1) buffer BufferB {{
{dtype}[720][1280] imageOut;
}};
void main() {{
vec3 pixel = gl_GlobalInvocationID;
int h = int(pixel.x);
int w = int(pixel.y);
imageOut[h][w] = imageIn[h][w];
}}
"""
rng = np.random.RandomState(123)
w = 1280
h = 720
for layout_in, layout_out, n, dtype in itertools.product(
self.LAYOUTS, self.LAYOUTS, range(2, 5), DataType.ALL):
vector = Vector(n, dtype)
im = rng.randint(0, 255,
size=(h, w,
n)).astype(vector.scalar.numpy_dtype())
glsl = glsl_template.format(
**{
"layout_in": self.LAYOUT_MAP[layout_in],
"layout_out": self.LAYOUT_MAP[layout_out],
"dtype": vector.glsl_dtype()
})
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
cache_in = ByteCache(shader.code.get_block_definition(0))
cache_in["imageIn"] = im
bytez_in = cache_in.definition.to_bytes(cache_in.get_as_dict())
cache_out = ByteCache(shader.code.get_block_definition(1))
bytez_out_count = cache_out.definition.size()
bytez_out = self.run_compiled_program(shader,
bytez_in,
bytez_out_count,
groups=(h, w, 1))
cache_out.set_from_dict(cache_out.definition.from_bytes(bytez_out))
self.assertTrue((cache_out["imageOut"] == im).all())
def test_scalars_and_vectors(self):
rng = np.random.RandomState(123)
variables = [
Scalar.uint(),
Scalar.int(),
Scalar.float(),
Scalar.double()
]
variables += [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
containers_std140 = [Struct(variables, Layout.STD140)]
containers_std430 = [Struct(variables, Layout.STD430)]
for _ in range(5):
containers_std140.append(
Struct(rng.permutation(variables), Layout.STD140))
containers_std430.append(
Struct(rng.permutation(variables), Layout.STD430))
for container in containers_std140:
self.run_test(container, [], BufferUsage.STORAGE_BUFFER)
self.run_test(container, [], BufferUsage.UNIFORM_BUFFER)
for container in containers_std430:
self.run_test(container, [], BufferUsage.STORAGE_BUFFER)
def test_detection_layout_stdxxx_ssbo(self):
variables = [
Scalar.uint(),
Scalar.int(),
Scalar.float(),
Scalar.double()
]
variables += [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
binding = 0
usage = BufferUsage.STORAGE_BUFFER
glsl_std140 = self.build_glsl_program(
((Struct(variables, Layout.STD140), binding, usage), ))
glsl_std430 = self.build_glsl_program(
((Struct(variables, Layout.STD430), binding, usage), ))
glsls = [glsl_std140, glsl_std430]
for glsl in glsls:
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
definition, detected_usage = shader.get_block(binding)
self.assertEqual(detected_usage, usage)
self.assertEqual(definition.layout, Layout.STDXXX)
def test_detection_type_nested_with_structs(self):
rng = np.random.RandomState(321)
simple = [Scalar.uint(), Scalar.int(), Scalar.float(), Scalar.double()]
simple += [Vector(n, dtype) for n, dtype in itertools.product(range(2, 5), DataType.ALL)]
for layout, _ in itertools.product([Layout.STD140, Layout.STD430], range(5)):
matrices = [Matrix(n, m, dtype, layout) for n, m, dtype in
itertools.product(range(2, 5), range(2, 5), [DataType.FLOAT, DataType.DOUBLE])]
simple_and_matrices = simple + matrices
struct = Struct(rng.choice(simple_and_matrices, size=3, replace=False), layout, type_name="SomeStruct")
structs = [struct]
for _ in range(4):
members = [structs[-1]] + rng.choice(simple_and_matrices, size=2, replace=False).tolist()
structs.append(Struct(rng.permutation(members), layout, type_name="SomeStruct{}".format(len(structs))))
container = structs[-1]
structs = structs[:-1]
glsl = self.build_glsl_program(((container, 0, BufferUsage.STORAGE_BUFFER),), structs)
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
definition, _ = shader.code.get_block(0)
self.assertTrue(container.compare(definition, quiet=True))
def test_detection_layout_stdxxx_ubo(self):
variables = [
Scalar.uint(),
Scalar.int(),
Scalar.float(),
Scalar.double()
]
variables += [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
binding = 0
usage = BufferUsage.UNIFORM_BUFFER
glsl = self.build_glsl_program(
((Struct(variables, Layout.STD140), binding, usage), ))
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
definition, detected_usage = shader.get_block(binding)
self.assertEqual(detected_usage, usage)
self.assertEqual(
definition.layout,
Layout.STD140) # uniform buffer objects can not use std430
def test_detection_type_arrays(self):
rng = np.random.RandomState(321)
variables = [
Scalar.uint(),
Scalar.int(),
Scalar.float(),
Scalar.double()
]
variables += [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
for definition, layout, _ in itertools.product(
variables, [Layout.STD140, Layout.STD430], range(3)):
container = Struct(
[Array(definition, Random.shape(rng, 3, 5), layout)], layout)
glsl = self.build_glsl_program(
((container, 0, BufferUsage.STORAGE_BUFFER), ))
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
detected_definition, _ = shader.get_block(0)
self.assertTrue(container.compare(detected_definition, quiet=True))
if isinstance(definition, Vector):
if definition.length(
) < 3 and definition.dtype != DataType.DOUBLE:
self.assertEqual(detected_definition.layout, layout)
def test_bytecode_parsing_without_gpu(self):
with self.env_backup():
del os.environ["VULKAN_SDK"]
import lava as lv
from lava.api.bytecode.logical import ByteCode
from lava.api.bytecode.physical import ByteCodeData
from lava.api.bytes import Vector, Scalar, Struct
from lava.api.constants.spirv import Layout
self.assertTrue(not lv.initialized())
glsl = """
#version 450
#extension GL_ARB_separate_shader_objects : enable
layout(local_size_x=1, local_size_y=1, local_size_z=1) in;
layout(std430, binding = 0) readonly buffer bufIn {
vec3 var1;
};
layout(std430, binding = 1) writeonly buffer bufOut {
float var2;
};
void main() {
var2 = var1.x + var1.y + var1.z;
}
"""
path_shader = write_to_temp_file(glsl, suffix=".comp")
path_shader_spirv = lv.compile_glsl(path_shader, verbose=True)
with self.env_backup():
del os.environ["VULKAN_SDK"]
self.assertTrue(not lv.initialized())
byte_code_data = ByteCodeData.from_file(path_shader_spirv)
byte_code = ByteCode(byte_code_data, None)
quiet = True
container0 = Struct([Vector.vec3()], Layout.STD430)
container1 = Struct([Scalar.float()], Layout.STD430)
self.assertTrue(
container0.compare(byte_code.get_block_definition(0),
quiet=quiet))
self.assertTrue(
container1.compare(byte_code.get_block_definition(1),
quiet=quiet))
os.remove(path_shader)
os.remove(path_shader_spirv)
def test_manually(self):
# byte cache test
buffer_usage = BufferUsage.STORAGE_BUFFER
buffer_layout = Layout.STD430
buffer_order = Order.ROW_MAJOR
struct1 = Struct([Vector.vec3(), Vector.ivec2()],
buffer_layout,
member_names=["a", "b"],
type_name="structB")
struct2 = Struct(
[Scalar.double(), Scalar.double(), struct1],
buffer_layout,
type_name="structC")
structs = [struct1, struct2]
variables = [
Scalar.uint(),
Array(Vector.vec2(), (5, 2, 3), buffer_layout),
Array(Scalar.float(), 5, buffer_layout),
struct2, # this struct needs padding at the end
Scalar.uint(),
Array(struct1, (2, 3), buffer_layout)
]
container = Struct(variables, buffer_layout, type_name="block")
cache = ByteCache(container)
print("")
print("")
pprint.pprint(cache.values)
print(cache[-1][0][0]["a"])
print("")
print("")
pprint.pprint(cache)
print(cache[-1][0][0])
print("")
print("")
pprint.pprint(cache.get_as_dict())
def test_arb_example_std140(self):
layout = Layout.STD430
struct_a = Struct(
[
Scalar.int(),
Vector.uvec2() # actually bvec2
],
layout,
type_name="structA")
struct_b = Struct([
Vector.uvec3(),
Vector.vec2(),
Array(Scalar.float(), 2, layout),
Vector.vec2(),
Array(Matrix(3, 3, DataType.FLOAT, layout), 2, layout)
],
layout,
type_name="structB")
container = Struct([
Scalar.float(),
Vector.vec2(),
Vector.vec3(), struct_a,
Scalar.float(),
Array(Scalar.float(), 2, layout),
Matrix(2, 3, DataType.FLOAT, layout),
Array(struct_b, 2, layout)
], layout)
self.run_test(container, [struct_a, struct_b],
BufferUsage.STORAGE_BUFFER)
def test_detection_binding(self):
container = Struct([Scalar.int(), Vector.vec3()], Layout.STD140)
for binding, usage in itertools.product([0, 1, 2, 3, 4, 99, 512], [BufferUsage.UNIFORM_BUFFER, BufferUsage.STORAGE_BUFFER]):
glsl = self.build_glsl_program(((container, binding, usage),))
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
detected_definition, detected_usage = shader.code.get_block(binding)
self.assertEqual(detected_usage, usage)
equal = container.compare(detected_definition, quiet=True)
self.assertTrue(equal)
def test_struct_shared_between_different_layouts(self):
glsl = """
#version 450
#extension GL_ARB_separate_shader_objects : enable
struct Shared {
int var1;
double var2;
};
layout(std140, binding = 0) buffer BufferA {
uvec2 varA1;
Shared varA2; // expected offset 16
};
layout(std430, binding = 1) buffer BufferB {
uvec2 varB1;
Shared varB2; // expected offset 8
};
void main() {}
"""
shared_std140 = Struct([Scalar.int(), Scalar.double()], Layout.STD140)
shared_std430 = Struct([Scalar.int(), Scalar.double()], Layout.STD430)
container_std140 = Struct([Vector.uvec2(), shared_std140],
Layout.STD140)
container_std430 = Struct([Vector.uvec2(), shared_std430],
Layout.STD430)
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
definition0, _ = shader.get_block(0)
definition1, _ = shader.get_block(1)
self.assertTrue(container_std140.compare(definition0, quiet=True))
self.assertFalse(container_std140.compare(definition1, quiet=True))
self.assertFalse(container_std430.compare(definition0, quiet=True))
self.assertTrue(container_std430.compare(definition1, quiet=True))
def test_array_of_vectors(self):
rng = np.random.RandomState(123)
vector_types = [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
for definition, layout, _ in itertools.product(
vector_types, [Layout.STD140, Layout.STD430], range(5)):
container = Struct(
[Array(definition, Random.shape(rng, 3, 5), layout)], layout)
self.run_test(container, [], BufferUsage.STORAGE_BUFFER)
if layout == Layout.STD140:
self.run_test(container, [], BufferUsage.UNIFORM_BUFFER)
def test_manual(self):
buffer_usage = BufferUsage.STORAGE_BUFFER
buffer_layout = Layout.STD140
buffer_order = Order.ROW_MAJOR
structA = Struct([Vector.ivec2(), Scalar.double()],
buffer_layout,
member_names=["a", "b"],
type_name="structA")
structB = Struct([Scalar.uint(), Scalar.double()],
buffer_layout,
type_name="structB")
structC = Struct([structB, Vector.ivec2()],
buffer_layout,
type_name="structC")
structs = [structA, structB, structC]
variables = [
Vector.vec3(),
Vector.ivec4(),
Array(structC, 2, buffer_layout),
Vector.ivec4(),
Scalar.uint(),
Array(Scalar.double(), (5, 2), buffer_layout),
Scalar.int(),
Array(Vector.vec4(), (2, 3, 4), buffer_layout),
Vector.dvec2(), structA
]
container = Struct(variables, buffer_layout, type_name="block")
print(container)
glsl = self.build_glsl_program(container, structs, buffer_usage)
# print(glsl)
values, array_expected = self.build_values(container.definitions)
array_expected = np.array(array_expected, dtype=np.float32)
bytez_input = container.to_bytes(values)
bytez_output = self.run_program(glsl,
bytez_input,
array_expected.nbytes,
usage_input=buffer_usage)
array = np.frombuffer(bytez_output, dtype=array_expected.dtype)
print(array_expected)
print(array)
print("equal", ((array_expected - array) == 0).all())
def test_array_of_structs(self):
rng = np.random.RandomState(123)
simple = [Scalar.uint(), Scalar.int(), Scalar.float(), Scalar.double()]
simple += [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
for layout, _ in itertools.product([Layout.STD140, Layout.STD430],
range(5)):
struct = Struct(rng.choice(simple, size=3, replace=False),
layout,
type_name="SomeStruct")
array = Array(struct, Random.shape(rng, 3, 5), layout)
container = Struct([array], layout)
self.run_test(container, [struct], BufferUsage.STORAGE_BUFFER)
if layout == Layout.STD140:
self.run_test(container, [struct], BufferUsage.UNIFORM_BUFFER)
def test_nested_with_arrays_of_structs(self):
rng = np.random.RandomState(123)
simple = [Scalar.uint(), Scalar.int(), Scalar.float(), Scalar.double()]
simple += [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
for layout, _ in itertools.product([Layout.STD140, Layout.STD430],
range(10)):
matrices = [
Matrix(n, m, dtype, layout)
for n, m, dtype in itertools.product(range(2, 5), range(
2, 5), [DataType.FLOAT, DataType.DOUBLE])
]
simple_and_matrices = simple + matrices
struct = Struct(rng.choice(simple_and_matrices,
size=4,
replace=False),
layout,
type_name="SomeStruct")
structs = [struct]
arrays = [Array(struct, Random.shape(rng, 2, 3), layout)]
for _ in range(2):
members = [arrays[-1]] + rng.choice(
simple_and_matrices, size=3, replace=False).tolist()
structs.append(
Struct(rng.permutation(members),
layout,
type_name="SomeStruct{}".format(len(structs))))
arrays.append(
Array(structs[-1], Random.shape(rng, 2, 3), layout))
container = structs[-1]
structs = structs[:-1]
self.run_test(container, structs, BufferUsage.STORAGE_BUFFER)
if layout == Layout.STD140:
self.run_test(container, structs, BufferUsage.UNIFORM_BUFFER)
def inspect(self):
self.block_data = self.byte_code.find_blocks()
self.definitions_block = {}
defs_scalar = {
index: Scalar.of(dtype)
for index, dtype in self.byte_code.types_scalar.items()
}
defs_vector = {
index: Vector(n, dtype)
for index, (dtype, n) in self.byte_code.types_vector.items()
}
defs_array = {}
defs_struct = {}
for binding in self.get_bindings():
index, _ = self.get_block_index(binding)
self.deduce_definition(index, defs_scalar, defs_vector, defs_array,
defs_struct, index)
self.definitions_block[index] = defs_struct[index]
self.deduce_layout(binding)
def test_scalars_and_vectors_and_matrices(self):
rng = np.random.RandomState(123)
variables = [
Scalar.uint(),
Scalar.int(),
Scalar.float(),
Scalar.double()
]
variables += [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
matrix_combinations = itertools.product(range(2, 5), range(
2, 5), [DataType.FLOAT, DataType.DOUBLE])
variables_std140 = variables + [
Matrix(n, m, dtype, Layout.STD140)
for n, m, dtype in matrix_combinations
]
variables_std430 = variables + [
Matrix(n, m, dtype, Layout.STD430)
for n, m, dtype in matrix_combinations
]
containers_std140 = [Struct(variables_std140, Layout.STD140)]
containers_std430 = [Struct(variables_std430, Layout.STD430)]
for _ in range(5):
containers_std140.append(
Struct(rng.permutation(variables_std140), Layout.STD140))
containers_std430.append(
Struct(rng.permutation(variables_std430), Layout.STD430))
for container in containers_std140 + containers_std430:
self.run_test(container, [], BufferUsage.STORAGE_BUFFER)
def test_manually(self):
buffer_usage = BufferUsage.STORAGE_BUFFER
buffer_layout = Layout.STD140
buffer_order = Order.ROW_MAJOR
structA = Struct([Vector.ivec2(), Scalar.double()],
buffer_layout,
member_names=["a", "b"],
type_name="structA")
structB = Struct([Scalar.uint(), Scalar.double()],
buffer_layout,
type_name="structB")
structC = Struct([structB, Vector.ivec2()],
buffer_layout,
type_name="structC")
structs = [structA, structB, structC]
variables = [
Vector.vec3(),
Vector.ivec4(),
Array(structC, 2, buffer_layout),
Vector.ivec4(),
Scalar.uint(),
Array(Scalar.double(), (5, 2), buffer_layout),
Scalar.int(),
Array(Vector.vec4(), (2, 3, 4), buffer_layout),
Vector.dvec2(), structA
]
container = Struct(variables, buffer_layout, type_name="block")
print(container)
glsl = self.build_glsl_program(container, structs, buffer_usage)
# print(glsl)
values_expected, array = self.build_values(container.definitions)
array = np.array(array, dtype=np.float32)
bytez_in = array.tobytes() # std430
bytez_out = self.run_program(glsl, bytez_in, container.size())
values = container.from_bytes(bytez_out)
register = {}
steps = {Scalar: 0, Vector: 0, Array: 0, Struct: 0}
for struct in structs + [container]:
self.build_register(register, struct, steps)
values_ftd = self.format_values(container, values, register)
values_expected_ftd = self.format_values(container, values_expected,
register)
print("")
self.print_formatted_values(values_ftd)
print("")
self.print_formatted_values(values_expected_ftd)
print("")
print("")
print(values_ftd == values_expected_ftd)
def test_pass_through_struct(self):
glsl_template = """
#version 450
#extension GL_ARB_separate_shader_objects : enable
layout(local_size_x=1, local_size_y=1, local_size_z=1) in;
{struct_glsl}
layout({layout_in}, binding = 0) buffer BufferA {{
{dtype} structIn;
}};
layout({layout_out}, binding = 1) buffer BufferB {{
{dtype} structOut;
}};
void main() {{
structOut = structIn;
}}
"""
rng = np.random.RandomState(123)
scalars = [Scalar.of(dtype) for dtype in DataType.ALL]
vectors = [
Vector(n, dtype)
for n, dtype in itertools.product(range(2, 5), DataType.ALL)
]
type_name = "SomeStruct"
for layout_in, layout_out in itertools.product(self.LAYOUTS,
self.LAYOUTS):
members = rng.permutation(scalars + scalars + vectors + vectors)
glsl_struct = ["struct {} {{".format(type_name)]
for i, member in enumerate(members):
glsl_struct.append("{} member{};".format(
member.glsl_dtype(), i))
glsl_struct.append("};")
glsl = glsl_template.format(
**{
"layout_in": self.LAYOUT_MAP[layout_in],
"layout_out": self.LAYOUT_MAP[layout_out],
"struct_glsl": "\n".join(glsl_struct),
"dtype": type_name
})
shader = self.shader_from_txt(glsl, verbose=False)
shader.inspect()
cache_in = ByteCache(shader.code.get_block_definition(0))
for i, member in enumerate(members):
if isinstance(member, Scalar):
value = member.numpy_dtype()(1.)
elif isinstance(member, Vector):
value = np.ones(member.length(),
member.scalar.numpy_dtype())
else:
value = None
cache_in["structIn"][i] = value
bytez_in = cache_in.definition.to_bytes(cache_in.get_as_dict())
cache_out = ByteCache(shader.code.get_block_definition(1))
bytez_out_count = cache_out.definition.size()
bytez_out = self.run_compiled_program(shader, bytez_in,
bytez_out_count)
cache_out.set_from_dict(cache_out.definition.from_bytes(bytez_out))
for i, member in enumerate(members):
a = cache_in["structIn"][i]
b = cache_out["structOut"][i]
if isinstance(member, Vector):
a = a.tolist()
b = b.tolist()
self.assertEqual(a, b)