warp-lang 1.7.0__py3-none-manylinux_2_28_x86_64.whl

warp/jax_experimental/custom_call.py

@@ -0,0 +1,351 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import ctypes
+
+import warp as wp
+from warp.context import type_str
+from warp.jax import get_jax_device
+from warp.types import array_t, launch_bounds_t, strides_from_shape
+
+_jax_warp_p = None
+
+# Holder for the custom callback to keep it alive.
+_cc_callback = None
+_registered_kernels = [None]
+_registered_kernel_to_id = {}
+
+
+def jax_kernel(kernel, launch_dims=None):
+    """Create a Jax primitive from a Warp kernel.
+
+    NOTE: This is an experimental feature under development.
+
+    Args:
+        kernel: The Warp kernel to be wrapped.
+        launch_dims: Optional. Specify the kernel launch dimensions. If None,
+                     dimensions are inferred from the shape of the first argument.
+                     This option when set will specify the output dimensions.
+
+    Limitations:
+        - All kernel arguments must be contiguous arrays.
+        - Input arguments are followed by output arguments in the Warp kernel definition.
+        - There must be at least one input argument and at least one output argument.
+        - Only the CUDA backend is supported.
+    """
+
+    if _jax_warp_p is None:
+        # Create and register the primitive
+        _create_jax_warp_primitive()
+    if kernel not in _registered_kernel_to_id:
+        id = len(_registered_kernels)
+        _registered_kernels.append(kernel)
+        _registered_kernel_to_id[kernel] = id
+    else:
+        id = _registered_kernel_to_id[kernel]
+
+    def bind(*args):
+        return _jax_warp_p.bind(*args, kernel=id, launch_dims=launch_dims)
+
+    return bind
+
+
+def _warp_custom_callback(stream, buffers, opaque, opaque_len):
+    # The descriptor is the form
+    # <kernel-id>|<launch-dims>|<arg-dims-list>
+    # Example:  42|16,32|16,32;100;16,32
+    kernel_id_str, dim_str, args_str = opaque.decode().split("|")
+
+    # Get the kernel from the registry.
+    kernel_id = int(kernel_id_str)
+    kernel = _registered_kernels[kernel_id]
+
+    # Parse launch dimensions.
+    dims = [int(d) for d in dim_str.split(",")]
+    bounds = launch_bounds_t(dims)
+
+    # Parse arguments.
+    arg_strings = args_str.split(";")
+    num_args = len(arg_strings)
+    assert num_args == len(kernel.adj.args), "Incorrect number of arguments"
+
+    # First param is the launch bounds.
+    kernel_params = (ctypes.c_void_p * (1 + num_args))()
+    kernel_params[0] = ctypes.addressof(bounds)
+
+    # Parse array descriptors.
+    args = []
+    for i in range(num_args):
+        dtype = kernel.adj.args[i].type.dtype
+        shape = [int(d) for d in arg_strings[i].split(",")]
+        strides = strides_from_shape(shape, dtype)
+
+        arr = array_t(buffers[i], 0, len(shape), shape, strides)
+        args.append(arr)  # keep a reference
+        arg_ptr = ctypes.addressof(arr)
+
+        kernel_params[i + 1] = arg_ptr
+
+    # Get current device.
+    device = wp.device_from_jax(get_jax_device())
+
+    # Get kernel hooks.
+    # Note: module was loaded during jit lowering.
+    hooks = kernel.module.get_kernel_hooks(kernel, device)
+    assert hooks.forward, "Failed to find kernel entry point"
+
+    # Launch the kernel.
+    wp.context.runtime.core.cuda_launch_kernel(
+        device.context, hooks.forward, bounds.size, 0, 256, hooks.forward_smem_bytes, kernel_params, stream
+    )
+
+
+def _create_jax_warp_primitive():
+    from functools import reduce
+
+    import jax
+    from jax._src.interpreters import batching
+    from jax.interpreters import mlir
+    from jax.interpreters.mlir import ir
+    from jaxlib.hlo_helpers import custom_call
+
+    global _jax_warp_p
+    global _cc_callback
+
+    # Create and register the primitive.
+    # TODO add default implementation that calls the kernel via warp.
+    _jax_warp_p = jax.core.Primitive("jax_warp")
+    _jax_warp_p.multiple_results = True
+
+    # TODO Just launch the kernel directly, but make sure the argument
+    # shapes are massaged the same way as below so that vmap works.
+    def impl(*args):
+        raise Exception("Not implemented")
+
+    _jax_warp_p.def_impl(impl)
+
+    # Auto-batching. Make sure all the arguments are fully broadcasted
+    # so that Warp is not confused about dimensions.
+    def vectorized_multi_batcher(args, dims, **params):
+        # Figure out the number of outputs.
+        wp_kernel = _registered_kernels[params["kernel"]]
+        output_count = len(wp_kernel.adj.args) - len(args)
+        shape, dim = next((a.shape, d) for a, d in zip(args, dims) if d is not None)
+        size = shape[dim]
+        args = [batching.bdim_at_front(a, d, size) if len(a.shape) else a for a, d in zip(args, dims)]
+        # Create the batched primitive.
+        return _jax_warp_p.bind(*args, **params), [dims[0]] * output_count
+
+    batching.primitive_batchers[_jax_warp_p] = vectorized_multi_batcher
+
+    def get_vecmat_shape(warp_type):
+        if hasattr(warp_type.dtype, "_shape_"):
+            return warp_type.dtype._shape_
+        return []
+
+    def strip_vecmat_dimensions(warp_arg, actual_shape):
+        shape = get_vecmat_shape(warp_arg.type)
+        for i, s in enumerate(reversed(shape)):
+            item = actual_shape[-i - 1]
+            if s != item:
+                raise Exception(f"The vector/matrix shape for argument {warp_arg.label} does not match")
+        return actual_shape[: len(actual_shape) - len(shape)]
+
+    def collapse_into_leading_dimension(warp_arg, actual_shape):
+        if len(actual_shape) < warp_arg.type.ndim:
+            raise Exception(f"Argument {warp_arg.label} has too few non-matrix/vector dimensions")
+        index_rest = len(actual_shape) - warp_arg.type.ndim + 1
+        leading_size = reduce(lambda x, y: x * y, actual_shape[:index_rest])
+        return [leading_size] + actual_shape[index_rest:]
+
+    # Infer array dimensions from input type.
+    def infer_dimensions(warp_arg, actual_shape):
+        actual_shape = strip_vecmat_dimensions(warp_arg, actual_shape)
+        return collapse_into_leading_dimension(warp_arg, actual_shape)
+
+    def base_type_to_jax(warp_dtype):
+        if hasattr(warp_dtype, "_wp_scalar_type_"):
+            return wp.dtype_to_jax(warp_dtype._wp_scalar_type_)
+        return wp.dtype_to_jax(warp_dtype)
+
+    def base_type_to_jax_ir(warp_dtype):
+        warp_to_jax_dict = {
+            wp.float16: ir.F16Type.get(),
+            wp.float32: ir.F32Type.get(),
+            wp.float64: ir.F64Type.get(),
+            wp.int8: ir.IntegerType.get_signless(8),
+            wp.int16: ir.IntegerType.get_signless(16),
+            wp.int32: ir.IntegerType.get_signless(32),
+            wp.int64: ir.IntegerType.get_signless(64),
+            wp.uint8: ir.IntegerType.get_unsigned(8),
+            wp.uint16: ir.IntegerType.get_unsigned(16),
+            wp.uint32: ir.IntegerType.get_unsigned(32),
+            wp.uint64: ir.IntegerType.get_unsigned(64),
+        }
+        if hasattr(warp_dtype, "_wp_scalar_type_"):
+            warp_dtype = warp_dtype._wp_scalar_type_
+        jax_dtype = warp_to_jax_dict.get(warp_dtype)
+        if jax_dtype is None:
+            raise TypeError(f"Invalid or unsupported data type: {warp_dtype}")
+        return jax_dtype
+
+    def base_type_is_compatible(warp_type, jax_ir_type):
+        jax_ir_to_warp = {
+            "f16": wp.float16,
+            "f32": wp.float32,
+            "f64": wp.float64,
+            "i8": wp.int8,
+            "i16": wp.int16,
+            "i32": wp.int32,
+            "i64": wp.int64,
+            "ui8": wp.uint8,
+            "ui16": wp.uint16,
+            "ui32": wp.uint32,
+            "ui64": wp.uint64,
+        }
+        expected_warp_type = jax_ir_to_warp.get(str(jax_ir_type))
+        if expected_warp_type is not None:
+            if hasattr(warp_type, "_wp_scalar_type_"):
+                return warp_type._wp_scalar_type_ == expected_warp_type
+            else:
+                return warp_type == expected_warp_type
+        else:
+            raise TypeError(f"Invalid or unsupported data type: {jax_ir_type}")
+
+    # Abstract evaluation.
+    def jax_warp_abstract(*args, kernel=None, launch_dims=None):
+        wp_kernel = _registered_kernels[kernel]
+        # All the extra arguments to the warp kernel are outputs.
+        warp_outputs = [o.type for o in wp_kernel.adj.args[len(args) :]]
+
+        if launch_dims is None:
+            # Use the first input dimension to infer the output's dimensions if launch_dims is not provided
+            dims = strip_vecmat_dimensions(wp_kernel.adj.args[0], list(args[0].shape))
+        else:
+            dims = launch_dims
+
+        jax_outputs = []
+        for o in warp_outputs:
+            shape = list(dims) + list(get_vecmat_shape(o))
+            dtype = base_type_to_jax(o.dtype)
+            jax_outputs.append(jax.core.ShapedArray(shape, dtype))
+        return jax_outputs
+
+    _jax_warp_p.def_abstract_eval(jax_warp_abstract)
+
+    # Lowering to MLIR.
+
+    # Create python-land custom call target.
+    CCALLFUNC = ctypes.CFUNCTYPE(
+        ctypes.c_voidp, ctypes.c_void_p, ctypes.POINTER(ctypes.c_void_p), ctypes.c_char_p, ctypes.c_size_t
+    )
+    _cc_callback = CCALLFUNC(_warp_custom_callback)
+    ccall_address = ctypes.cast(_cc_callback, ctypes.c_void_p)
+
+    # Put the custom call into a capsule, as required by XLA.
+    PyCapsule_Destructor = ctypes.CFUNCTYPE(None, ctypes.py_object)
+    PyCapsule_New = ctypes.pythonapi.PyCapsule_New
+    PyCapsule_New.restype = ctypes.py_object
+    PyCapsule_New.argtypes = (ctypes.c_void_p, ctypes.c_char_p, PyCapsule_Destructor)
+    capsule = PyCapsule_New(ccall_address.value, b"xla._CUSTOM_CALL_TARGET", PyCapsule_Destructor(0))
+
+    # Register the callback in XLA.
+    jax.lib.xla_client.register_custom_call_target("warp_call", capsule, platform="gpu")
+
+    def default_layout(shape):
+        return range(len(shape) - 1, -1, -1)
+
+    def warp_call_lowering(ctx, *args, kernel=None, launch_dims=None):
+        if not kernel:
+            raise Exception("Unknown kernel id " + str(kernel))
+        wp_kernel = _registered_kernels[kernel]
+
+        # TODO This may not be necessary, but it is perhaps better not to be
+        # mucking with kernel loading while already running the workload.
+        module = wp_kernel.module
+        device = wp.device_from_jax(get_jax_device())
+        if not module.load(device):
+            raise Exception("Could not load kernel on device")
+
+        if launch_dims is None:
+            # Infer dimensions from the first input.
+            warp_arg0 = wp_kernel.adj.args[0]
+            actual_shape0 = ir.RankedTensorType(args[0].type).shape
+            dims = strip_vecmat_dimensions(warp_arg0, actual_shape0)
+            warp_dims = collapse_into_leading_dimension(warp_arg0, dims)
+        else:
+            dims = launch_dims
+            warp_dims = launch_dims
+        # Figure out the types and shapes of the input arrays.
+        arg_strings = []
+        operand_layouts = []
+        for actual, warg in zip(args, wp_kernel.adj.args):
+            wtype = warg.type
+            rtt = ir.RankedTensorType(actual.type)
+
+            if not isinstance(wtype, wp.array):
+                raise Exception("Only contiguous arrays are supported for Jax kernel arguments")
+
+            if not base_type_is_compatible(wtype.dtype, rtt.element_type):
+                raise TypeError(
+                    f"Incompatible data type for argument '{warg.label}', expected {type_str(wtype.dtype)}, got {rtt.element_type}"
+                )
+
+            # Infer array dimension (by removing the vector/matrix dimensions and
+            # collapsing the initial dimensions).
+            shape = infer_dimensions(warg, rtt.shape)
+
+            if len(shape) != wtype.ndim:
+                raise TypeError(f"Incompatible array dimensionality for argument '{warg.label}'")
+
+            arg_strings.append(",".join([str(d) for d in shape]))
+            operand_layouts.append(default_layout(rtt.shape))
+
+        # Figure out the types and shapes of the output arrays.
+        result_types = []
+        result_layouts = []
+        for warg in wp_kernel.adj.args[len(args) :]:
+            wtype = warg.type
+
+            if not isinstance(wtype, wp.array):
+                raise Exception("Only contiguous arrays are supported for Jax kernel arguments")
+
+            # Infer dimensions from the first input.
+            arg_strings.append(",".join([str(d) for d in warp_dims]))
+
+            result_shape = list(dims) + list(get_vecmat_shape(wtype))
+            result_types.append(ir.RankedTensorType.get(result_shape, base_type_to_jax_ir(wtype.dtype)))
+            result_layouts.append(default_layout(result_shape))
+
+        # Build opaque descriptor for callback.
+        shape_str = ",".join([str(d) for d in warp_dims])
+        args_str = ";".join(arg_strings)
+        descriptor = f"{kernel}|{shape_str}|{args_str}"
+
+        out = custom_call(
+            b"warp_call",
+            result_types=result_types,
+            operands=args,
+            backend_config=descriptor.encode("utf-8"),
+            operand_layouts=operand_layouts,
+            result_layouts=result_layouts,
+        ).results
+        return out
+
+    mlir.register_lowering(
+        _jax_warp_p,
+        warp_call_lowering,
+        platform="gpu",
+    )