triton-windows 3.5.1.post21__cp313-cp313-win_amd64.whl

triton/tools/disasm.py

@@ -0,0 +1,143 @@
+# MIT License
+
+# Copyright (c) 2020 Da Yan @ HKUST
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+import functools
+import os
+import re
+import subprocess
+import tempfile
+
+FLINE_RE = re.compile(r'\s*/\*\w{4}\*/\s*([^;]*;)\s*/\* 0x(\w{16}) \*/\s*')
+SLINE_RE = re.compile(r'\s*/\* 0x(\w{16}) \*/\s*')
+FNAME_RE = re.compile(r'\s*Function : (\w+)\s*')
+BRA_RE = re.compile(r'(.*BRA(?:\.U)? )(0x\w+);')
+
+
+def parseCtrl(sline):
+    enc = int(SLINE_RE.match(sline).group(1), 16)
+    stall = (enc >> 41) & 0xf
+    yld = (enc >> 45) & 0x1
+    wrtdb = (enc >> 46) & 0x7
+    readb = (enc >> 49) & 0x7
+    watdb = (enc >> 52) & 0x3f
+
+    yld_str = 'Y' if yld == 0 else '-'
+    wrtdb_str = '-' if wrtdb == 7 else str(wrtdb)
+    readb_str = '-' if readb == 7 else str(readb)
+    watdb_str = '--' if watdb == 0 else f'{watdb:02d}'
+    return f'{watdb_str}:{readb_str}:{wrtdb_str}:{yld_str}:{stall:x}'
+
+
+def processSassLines(fline, sline, labels):
+    asm = FLINE_RE.match(fline).group(1)
+    # Remove tailing space
+    if asm.endswith(" ;"):
+        asm = asm[:-2] + ";"
+    ctrl = parseCtrl(sline)
+    # BRA target address
+    if BRA_RE.match(asm) is not None:
+        target = int(BRA_RE.match(asm).group(2), 16)
+        if target in labels:
+            pass
+        else:
+            labels[target] = len(labels)
+    return (f'{ctrl}', f'{asm}')
+
+
+@functools.lru_cache()
+def get_sass(cubin_asm, fun=None):
+    fd, path = tempfile.mkstemp()
+    try:
+        with open(fd, 'wb') as cubin:
+            cubin.write(cubin_asm)
+        sass = extract(path, fun)
+    finally:
+        os.remove(path)
+    return sass
+
+
+def path_to_cuobjdump():
+    from triton import knobs
+    return knobs.nvidia.cuobjdump.path
+
+
+def extract(file_path, fun):
+    cuobjdump = path_to_cuobjdump()
+    if fun is None:
+        sass_str = subprocess.check_output([cuobjdump, "-sass", file_path])
+    else:
+        sass_str = subprocess.check_output([cuobjdump, "-fun", fun, "-sass", file_path])
+    sass_lines = sass_str.splitlines()
+    line_idx = 0
+    while line_idx < len(sass_lines):
+        line = sass_lines[line_idx].decode()
+        # format:
+        # function : <function_name>
+        # .headerflags: ...
+        # /*0000*/ asmstr /*0x...*/
+        #                 /*0x...*/
+
+        # Looking for new function header (function: <name>)
+        while FNAME_RE.match(line) is None:
+            line_idx += 1
+            if line_idx < len(sass_lines):
+                line = sass_lines[line_idx].decode()
+            else:
+                return
+
+        fname = FNAME_RE.match(line).group(1)
+        ret = ''
+        ret += f'Function:{fname}\n'
+        line_idx += 2  # bypass .headerflags
+        line = sass_lines[line_idx].decode()
+        # Remapping address to label
+        labels = {}  # address -> label_idx
+        # store sass asm in buffer and them print them (for labels)
+        # (ctrl, asm)
+        asm_buffer = []
+        while FLINE_RE.match(line) is not None:
+            # First line (Offset ASM Encoding)
+            fline = sass_lines[line_idx].decode()
+            line_idx += 1
+            # Second line (Encoding)
+            sline = sass_lines[line_idx].decode()
+            line_idx += 1
+            asm_buffer.append(processSassLines(fline, sline, labels))
+            # peek the next line
+            line = sass_lines[line_idx].decode()
+        # Print sass
+        # label naming convention: LBB#i
+        for idx, (ctrl, asm) in enumerate(asm_buffer):
+            # Print label if this is BRA target
+            offset = idx * 16
+            if offset in labels:
+                label_name = f'LBB{labels[offset]}'
+                ret += f'{label_name}:\n'
+            ret += ctrl + '\t'
+            # if this is BRA, remap offset to label
+            if BRA_RE.match(asm):
+                target = int(BRA_RE.match(asm).group(2), 16)
+                target_name = f'LBB{labels[target]}'
+                asm = BRA_RE.sub(rf'\1{target_name};', asm)
+            ret += asm + '\n'
+        ret += '\n'
+        return ret

triton/tools/extra/cuda/compile.c

@@ -0,0 +1,70 @@
+/* clang-format off */
+#include <stdio.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <cuda.h>
+
+
+// helpers to check for cuda errors
+#define CUDA_CHECK(ans) {{\
+    gpuAssert((ans), __FILE__, __LINE__);\
+  }}\
+
+static inline void gpuAssert(CUresult code, const char *file, int line) {{
+  if (code != CUDA_SUCCESS) {{
+    const char *prefix = "Triton Error [CUDA]: ";
+    const char *str;
+    cuGetErrorString(code, &str);
+    char err[1024] = {{0}};
+    strcat(err, prefix);
+    strcat(err, str);
+    printf("%s\\n", err);
+    exit(code);
+  }}
+}}
+
+// globals
+#define CUBIN_NAME {kernel_name}_cubin
+CUmodule {kernel_name}_mod = NULL;
+CUfunction {kernel_name}_func = NULL;
+unsigned char CUBIN_NAME[{bin_size}] = {{ {bin_data} }};
+
+
+void unload_{kernel_name}(void) {{
+    CUDA_CHECK(cuModuleUnload({kernel_name}_mod));
+}}
+
+// TODO: some code duplication with `runtime/backend/cuda.c`
+void load_{kernel_name}() {{
+    int dev = 0;
+    void *bin = (void *)&CUBIN_NAME;
+    int shared = {shared};
+    CUDA_CHECK(cuModuleLoadData(&{kernel_name}_mod, bin));
+    CUDA_CHECK(cuModuleGetFunction(&{kernel_name}_func, {kernel_name}_mod, "{triton_kernel_name}"));
+    // set dynamic shared memory if necessary
+    int shared_optin;
+    CUDA_CHECK(cuDeviceGetAttribute(&shared_optin, CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN, dev));
+    if (shared > 49152 && shared_optin > 49152) {{
+      CUDA_CHECK(cuFuncSetCacheConfig({kernel_name}_func, CU_FUNC_CACHE_PREFER_SHARED));
+      CUDA_CHECK(cuFuncSetAttribute({kernel_name}_func, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, shared_optin))
+    }}
+}}
+
+/*
+{kernel_docstring}
+*/
+CUresult {kernel_name}(CUstream stream, {signature}) {{
+    if ({kernel_name}_func == NULL)
+       load_{kernel_name}();
+    unsigned int gX = {gridX};
+    unsigned int gY = {gridY};
+    unsigned int gZ = {gridZ};
+    CUdeviceptr global_scratch = 0;
+    CUdeviceptr profile_scratch = 0;
+    void *args[{num_args}] = {{ {arg_pointers} }};
+    // TODO: shared memory
+    if(gX * gY * gZ > 0)
+      return cuLaunchKernel({kernel_name}_func, gX, gY, gZ, {num_warps} * 32, 1, 1, {shared}, stream, args, NULL);
+    return (CUresult)NULL;
+}}

triton/tools/extra/cuda/compile.h

@@ -0,0 +1,14 @@
+#ifndef TT_KERNEL_INCLUDES
+#define TT_KERNEL_INCLUDES
+
+#include <cuda.h>
+#include <inttypes.h>
+#include <stdint.h>
+#include <stdio.h>
+
+#endif
+
+void unload_{kernel_name}(void);
+void load_{kernel_name}(void);
+// tt-linker: {kernel_name}:{full_signature}:{algo_info}
+CUresult{_placeholder} {kernel_name}(CUstream stream, {signature});

triton/tools/extra/hip/compile.cpp

@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+/* clang-format off */
+#include <stdio.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <hip/hip_runtime.h>
+
+// helpers to check for hip errors
+#define HIP_CHECK(ans) {{\
+    gpuAssert((ans), __FILE__, __LINE__);\
+  }}\
+
+static inline void gpuAssert(hipError_t code, const char *file, int line) {{
+  if (code != hipSuccess) {{
+    const char *prefix = "Triton Error [HIP]: ";
+    const char *str;
+    hipDrvGetErrorString(code, &str);
+    char err[1024] = {{0}};
+    strcat(err, prefix);
+    strcat(err, str);
+    printf("%s\\n", err);
+    exit(code);
+  }}
+}}
+
+// globals
+#define HSACO_NAME {kernel_name}_hsaco
+hipModule_t {kernel_name}_mod = nullptr;
+hipFunction_t {kernel_name}_func = nullptr;
+unsigned char HSACO_NAME[{bin_size}] = {{ {bin_data} }};
+
+
+void unload_{kernel_name}(void) {{
+    HIP_CHECK(hipModuleUnload({kernel_name}_mod));
+}}
+
+
+void load_{kernel_name}() {{
+    int dev = 0;
+    void *bin = (void *)&HSACO_NAME;
+    int shared = {shared};
+    HIP_CHECK(hipModuleLoadData(&{kernel_name}_mod, bin));
+    HIP_CHECK(hipModuleGetFunction(&{kernel_name}_func, {kernel_name}_mod, "{triton_kernel_name}"));
+}}
+
+/*
+{kernel_docstring}
+*/
+hipError_t {kernel_name}(hipStream_t stream, {signature}) {{
+    if ({kernel_name}_func == nullptr)
+       load_{kernel_name}();
+    unsigned int gX = {gridX};
+    unsigned int gY = {gridY};
+    unsigned int gZ = {gridZ};
+    hipDeviceptr_t global_scratch = 0;
+    hipDeviceptr_t profile_scratch = 0;
+    void *args[{num_args}] = {{ {arg_pointers} }};
+    // TODO: shared memory
+    if(gX * gY * gZ > 0)
+      return hipModuleLaunchKernel({kernel_name}_func, gX, gY, gZ, {num_warps} * warpSize, 1, 1, {shared}, stream, args, nullptr);
+    else
+      return hipErrorInvalidValue;
+}}

triton/tools/extra/hip/compile.h

@@ -0,0 +1,13 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <inttypes.h>
+#include <stdint.h>
+#include <stdio.h>
+
+void unload_{kernel_name}(void);
+void load_{kernel_name}(void);
+hipError_t{_placeholder} {kernel_name}(hipStream_t stream, {signature});

triton/tools/link.py

@@ -0,0 +1,322 @@
+from collections import defaultdict
+from pathlib import Path
+from typing import Sequence, Union
+
+from dataclasses import dataclass
+
+
+def _exists(x):
+    return x is not None
+
+
+class LinkerError(Exception):
+    pass
+
+
+@dataclass
+class KernelLinkerMeta:
+    orig_kernel_name: str
+    arg_names: Sequence[str]
+    arg_ctypes: Sequence[str]
+    sizes: Sequence[Union[int, None]]
+    sig_hash: str
+    triton_suffix: str
+    suffix: str
+    num_specs: int
+    """ number of specialized arguments """
+
+
+class HeaderParser:
+
+    def __init__(self) -> None:
+        import re
+
+        # [kernel_name, c signature]
+        self.linker_directives = re.compile("//[\\s]*tt-linker:[\\s]*([\\w]+):(.+):(.+)")
+        # [name, hash, suffix]
+        self.kernel_name = re.compile("^([\\w]+)_([\\w]+)_([\\w]+)$")
+        # [(type, name)]
+        self.c_sig = re.compile("[\\s]*(\\w+)\\s(\\w+)[,]?")
+        # [d|c]
+        self.arg_suffix = re.compile("[c,d]")
+
+        self.kernels = defaultdict(list)
+
+    def extract_linker_meta(self, header: str):
+        for ln in header.splitlines():
+            if ln.startswith("//"):
+                m = self.linker_directives.match(ln)
+                if _exists(m):
+                    ker_name, c_sig, algo_info = m.group(1), m.group(2), m.group(3)
+                    name, sig_hash, suffix = self._match_name(ker_name)
+                    c_types, arg_names = self._match_c_sig(c_sig)
+                    num_specs, sizes = self._match_suffix(suffix, c_sig)
+                    self._add_kernel(
+                        "_".join([name, algo_info]),
+                        KernelLinkerMeta(
+                            orig_kernel_name=name,
+                            arg_names=arg_names,
+                            arg_ctypes=c_types,
+                            sizes=sizes,
+                            sig_hash=sig_hash,
+                            triton_suffix=suffix,
+                            suffix=suffix,
+                            num_specs=num_specs,
+                        ),
+                    )
+
+    def _match_name(self, ker_name: str):
+        m = self.kernel_name.match(ker_name)
+        if _exists(m):
+            name, sig_hash, suffix = m.group(1), m.group(2), m.group(3)
+            return name, sig_hash, suffix
+        raise LinkerError(f"{ker_name} is not a valid kernel name")
+
+    def _match_c_sig(self, c_sig: str):
+        m = self.c_sig.findall(c_sig)
+        if len(m):
+            tys, args = [], []
+            for ty, arg_name in m:
+                tys.append(ty)
+                args.append(arg_name)
+            return tys, args
+
+        raise LinkerError(f"{c_sig} is not a valid argument signature")
+
+    def _match_suffix(self, suffix: str, c_sig: str):
+        args = c_sig.split(",")
+        s2i = {"c": 1, "d": 16}
+        num_specs = 0
+        sizes = []
+        # scan through suffix, first find the index,
+        # then see if it is followed by d or c
+        for i in range(len(args)):
+            pos = suffix.find(str(i))
+            if pos == -1:
+                raise LinkerError(f"{suffix} is not a valid kernel suffix")
+            pos += len(str(i))
+            if self.arg_suffix.match(suffix, pos):
+                num_specs += 1
+                sizes.extend([None] * (i - len(sizes)))
+                sizes.append(s2i[suffix[pos]])
+                pos += 1
+            if i < len(args) - 1:
+                suffix = suffix[pos:]
+            else:
+                sizes.extend([None] * (len(args) - len(sizes)))
+        return num_specs, sizes
+
+    def _add_kernel(self, name: str, ker: KernelLinkerMeta):
+        if name in self.kernels:
+            last: KernelLinkerMeta = self.kernels[name][-1]
+
+            for cur, new_ in zip(last.arg_ctypes, ker.arg_ctypes):
+                if cur != new_:
+                    raise LinkerError(
+                        f"Mismatched signature for kernel {name}: \n\texisting sig is: {','.join(last.arg_ctypes)}\n\tcurrent is: {','.join(ker.arg_ctypes)}"
+                    )
+
+        self.kernels[name].append(ker)
+
+
+def gen_signature_with_full_args(m):
+    return ", ".join([f"{ty} {arg}" for ty, arg in zip(m.arg_ctypes, m.arg_names)])
+
+
+def gen_signature(m):
+    arg_types = [ty for ty, hint in zip(m.arg_ctypes, m.sizes) if hint != 1]
+    arg_names = [arg for arg, hint in zip(m.arg_names, m.sizes) if hint != 1]
+    sig = ", ".join([f"{ty} {arg}" for ty, arg in zip(arg_types, arg_names)])
+    return sig
+
+
+# generate declarations of kernels with meta-parameter and constant values
+def make_algo_decls(name: str, metas: Sequence[KernelLinkerMeta]) -> str:
+    return f"""
+CUresult {name}(CUstream stream, {gen_signature_with_full_args(metas[-1])});
+void load_{name}();
+void unload_{name}();
+    """
+
+
+# generate declarations of kernels with meta-parameter and constant values
+def make_global_decl(meta: KernelLinkerMeta) -> str:
+    return f"""
+CUresult {meta.orig_kernel_name}_default(CUstream stream, {gen_signature_with_full_args(meta)});
+CUresult {meta.orig_kernel_name}(CUstream stream, {gen_signature_with_full_args(meta)}, int algo_id);
+void load_{meta.orig_kernel_name}();
+void unload_{meta.orig_kernel_name}();
+    """
+
+
+# generate dispatcher function for kernels with different meta-parameter and constant values
+def make_default_algo_kernel(meta: KernelLinkerMeta) -> str:
+    src = f"CUresult {meta.orig_kernel_name}_default(CUstream stream, {gen_signature_with_full_args(meta)}){{\n"
+    src += (f"  return {meta.orig_kernel_name}(stream, {', '.join(meta.arg_names)}, 0);\n")
+    src += "}\n"
+    return src
+
+
+# generate dispatcher function for kernels with different integer value hints
+def make_kernel_hints_dispatcher(name: str, metas: Sequence[KernelLinkerMeta]) -> str:
+    src = f"// launcher for: {name}\n"
+    for meta in sorted(metas, key=lambda m: -m.num_specs):
+        src += f"CUresult {meta.orig_kernel_name}_{meta.sig_hash}_{meta.suffix}(CUstream stream, {gen_signature(meta)});\n"
+    src += "\n"
+
+    src += (f"CUresult {name}(CUstream stream, {gen_signature_with_full_args(metas[-1])}){{")
+    src += "\n"
+    for meta in sorted(metas, key=lambda m: -m.num_specs):
+        cond_fn = (  #
+            lambda val, hint: f"({val} % {hint} == 0)"  #
+            if hint == 16  #
+            else f"({val} == {hint})"  #
+            if hint == 1  #
+            else None)
+        conds = " && ".join([  #
+            cond_fn(val, hint)  #
+            for val, hint in zip(meta.arg_names, meta.sizes)  #
+            if hint is not None
+        ])
+        src += (f"  if ({conds})\n" if any(meta.sizes) else "if (1)\n"
+                )  # Edge case where no specializations hence no dispatching required
+        arg_names = [arg for arg, hint in zip(meta.arg_names, meta.sizes) if hint != 1]
+        src += f"    return {meta.orig_kernel_name}_{meta.sig_hash}_{meta.suffix}(stream, {', '.join(arg_names)});\n"
+    src += "\n"
+    src += "  return CUDA_ERROR_INVALID_VALUE;\n"
+    src += "}\n"
+
+    for mode in ["load", "unload"]:
+        src += f"\n// {mode} for: {name}\n"
+        for meta in sorted(metas, key=lambda m: -m.num_specs):
+            src += f"void {mode}_{meta.orig_kernel_name}_{meta.sig_hash}_{meta.suffix}();\n"
+        src += f"void {mode}_{name}() {{"
+        src += "\n"
+        for meta in sorted(metas, key=lambda m: -m.num_specs):
+            src += (f"  {mode}_{meta.orig_kernel_name}_{meta.sig_hash}_{meta.suffix}();\n")
+        src += "}\n"
+    return src
+
+
+# generate dispatcher function for kernels with different meta-parameter and constant values
+def make_kernel_meta_const_dispatcher(meta: KernelLinkerMeta) -> str:
+    src = f"CUresult {meta.orig_kernel_name}(CUstream stream, {gen_signature_with_full_args(meta)}, int algo_id){{\n"
+    src += f"  assert (algo_id < (int)sizeof({meta.orig_kernel_name}_kernels));\n"
+    src += f"  return {meta.orig_kernel_name}_kernels[algo_id](stream, {', '.join(meta.arg_names)});\n"
+    src += "}\n"
+    return src
+
+
+# generate definition of function pointers of kernel dispatchers based on meta-parameter and constant values
+def make_func_pointers(names: str, meta: KernelLinkerMeta) -> str:
+    # the table of hint dispatchers
+    src = f"typedef CUresult (*kernel_func_t)(CUstream stream, {gen_signature_with_full_args(meta)});\n"
+    src += f"kernel_func_t {meta.orig_kernel_name}_kernels[] = {{\n"
+    for name in names:
+        src += f"  {name},\n"
+    src += "};\n"
+    return src
+
+
+# generate definition for load/unload functions for kernels with different meta-parameter and constant values
+def make_kernel_load_def(names: str, meta: KernelLinkerMeta) -> str:
+    src = ""
+    for mode in ["load", "unload"]:
+        src += f"void {mode}_{meta.orig_kernel_name}(void){{\n"
+        for name in names:
+            src += f"  {mode}_{name}();\n"
+        src += "}\n\n"
+    return src
+
+
+def make_get_num_algos_decl(meta: KernelLinkerMeta) -> str:
+    src = f"int {meta.orig_kernel_name}_get_num_algos(void);"
+    return src
+
+
+def make_get_num_algos_def(meta: KernelLinkerMeta) -> str:
+    src = f"int {meta.orig_kernel_name}_get_num_algos(void){{\n"
+    src += f"  return (int)(sizeof({meta.orig_kernel_name}_kernels) / sizeof({meta.orig_kernel_name}_kernels[0]));\n"
+    src += "}\n"
+    return src
+
+
+desc = """
+Triton ahead-of-time linker:
+
+This program takes in header files generated by compile.py, and generates a
+single entry-point responsible for dispatching the user's input to the right
+kernel given the specializations that were compiled.
+
+Example usage:
+python link.py /path/to/headers/*.h -o kernel_name
+"""
+
+if __name__ == "__main__":
+    from argparse import ArgumentParser
+
+    parser = ArgumentParser(description=desc)
+    parser.add_argument(
+        "headers",
+        nargs="+",
+        help="Paths to header files to link. Must include linker directive annotations (autogenerated by ttc)",
+    )
+    parser.add_argument("--out", "-o", type=Path, help="Out filename")
+    parser.add_argument(
+        "--prefix",
+        type=str,
+        default="",
+        help="String to prefix kernel dispatcher names",
+    )
+    args = parser.parse_args()
+
+    # metadata
+    parser = HeaderParser()
+    includes = []
+    for header in args.headers:
+        h_path = Path(header)
+        h_str = h_path.read_text()
+        includes.append(h_path.name)
+        parser.extract_linker_meta(h_str)
+
+    # generate headers
+    algo_decls = [make_algo_decls(name, meta) for name, meta in parser.kernels.items()]
+    meta_lists = [meta for name, meta in parser.kernels.items()]
+    meta = meta_lists[0][0]
+    get_num_algos_decl = make_get_num_algos_decl(meta)
+    global_decl = make_global_decl(meta)
+    with args.out.with_suffix(".h").open("w") as fp:
+        out = "#include <cuda.h>\n"
+        out += "\n".join(algo_decls)
+        out += "\n"
+        out += get_num_algos_decl
+        out += "\n"
+        out += global_decl
+        fp.write(out)
+
+    # generate source
+    defs = [make_kernel_hints_dispatcher(name, meta) for name, meta in parser.kernels.items()]
+    names = [name for name in parser.kernels.keys()]
+    func_pointers_def = make_func_pointers(names, meta)
+    meta_const_def = make_kernel_meta_const_dispatcher(meta)
+    load_unload_def = make_kernel_load_def(names, meta)
+    get_num_algos_def = make_get_num_algos_def(meta)
+    default_algo_kernel = make_default_algo_kernel(meta)
+    with args.out.with_suffix(".c").open("w") as fp:
+        out = ""
+        out += "#include <cuda.h>\n"
+        out += "#include <stdint.h>\n"
+        out += "#include <assert.h>\n"
+        out += "\n"
+        out += "\n".join(defs)
+        out += "\n"
+        out += func_pointers_def
+        out += "\n"
+        out += get_num_algos_def
+        out += "\n"
+        out += meta_const_def
+        out += "\n"
+        out += load_unload_def
+        out += "\n"
+        out += default_algo_kernel
+        fp.write(out)