Adapt spconv for DTK SIMT fallback

Add a DTK-specific kernel filter path for running spconv through the
DTK CUDA compatibility layer on BW100. The recommended `dtk_simt`
filter keeps only SIMT kernels, forces SIMT params to static codegen,
and removes Volta/Turing/Ampere TensorOp, int8, and NVRTC paths from
the active kernel set.

Add `dtk_tensorop` as a separate non-default adaptation entry point for
future Ampere TensorOp work. This keeps static non-int8 Ampere TensorOp
params while still excluding Volta/Turing, int8, and NVRTC paths.

Allow fp16 workloads to use SIMT fallback when `SPCONV_DTK_KERNEL_FILTER`
is set to `dtk_simt`. This updates both the Python tuner and generated
C++ ConvTunerSimple logic so fp16 no longer depends on currently
unsupported TensorOp paths on DTK.

Add `SPCONV_FORCE_CUDA_ARCH` to keep runtime dispatch aligned with the
compiled arch list, and keep the BW100 path explicit with `9.3`.

Adjust DTK build/runtime compatibility:
- reuse the guarded editable-install state during constants setup
- skip the Linux Thrust `-fno-gnu-unique` flag under the DTK inline-PTX
  compatibility path
- add launch bounds to helper kernels that are launched with 1024
  threads/block

This leaves full TensorOp, int8, fp8, and NVRTC support out of the
recommended DTK path. Those remain future adaptation work.

spconv/algo.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import contextlib
+import os
 import time
 from enum import Enum
 from threading import Lock
@@ -728,7 +729,8 @@ class SimpleConv:
                 if (desp.tensorop[0] > 0 and inp.dtype == tv.float32 
                         and weight.dtype == tv.float32 and out.dtype == tv.float32):
                     continue
-            if arch >= (7, 0) and is_fp16:
+            allow_dtk_fp16_simt = os.getenv("SPCONV_DTK_KERNEL_FILTER", "").lower() == "dtk_simt"
+            if arch >= (7, 0) and is_fp16 and not allow_dtk_fp16_simt:
                 if desp.algo == GemmAlgo.Simt:
                     continue
                 if use_f32_as_accum:

spconv/constants.py

@@ -60,7 +60,7 @@ SPCONV_BWD_SPLITK = list(map(int, os.getenv("SPCONV_BWD_SPLITK", "1,2,4,8,16,32,
 SPCONV_NVRTC_MODE = NVRTCMode.ConstantMemory
 SPCONV_DEBUG_NVRTC_KERNELS = False
 
-SPCONV_DEBUG_CPP_ONLY = project_is_editable(PACKAGE_NAME)
+SPCONV_DEBUG_CPP_ONLY = EDITABLE_INSTALLED
 
 
 class AllocKeys:

spconv/core.py

@@ -11,6 +11,7 @@
 # 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 os
 from enum import Enum
 from cumm.gemm.main import gen_shuffle_params_v2 as gen_shuffle_params, GemmAlgoParams
 from cumm.gemm import kernel
@@ -1338,6 +1339,65 @@ if not SPCONV_INT8_DEBUG:
                         int8_inference=True),
     ])
 
+def _is_ampere_param(param):
+    return param.algo == GemmAlgo.Ampere
+
+
+def _is_simt_param(param):
+    return param.algo == GemmAlgo.Simt
+
+
+def _is_static_param(param):
+    return not getattr(param, "is_nvrtc", False)
+
+
+def _is_non_int8_param(param):
+    return not getattr(param, "int8_inference", False)
+
+
+def _is_ampere_no_int8_static_param(param):
+    return _is_ampere_param(param) and _is_static_param(param) and _is_non_int8_param(param)
+
+
+def _filter_params(params, predicate):
+    return [param for param in params if predicate(param)]
+
+
+def _force_static_params(params, predicate):
+    for param in params:
+        if predicate(param):
+            param.is_nvrtc = False
+    return params
+
+
+def _clear_turing_volta():
+    global SHUFFLE_TURING_PARAMS, SHUFFLE_VOLTA_PARAMS
+    global IMPLGEMM_TURING_PARAMS, IMPLGEMM_VOLTA_PARAMS
+    SHUFFLE_TURING_PARAMS = []
+    SHUFFLE_VOLTA_PARAMS = []
+    IMPLGEMM_TURING_PARAMS = []
+    IMPLGEMM_VOLTA_PARAMS = []
+
+
+_DTK_KERNEL_FILTER = os.getenv("SPCONV_DTK_KERNEL_FILTER", "").lower()
+if _DTK_KERNEL_FILTER == "dtk_simt":
+    SHUFFLE_SIMT_PARAMS = _force_static_params(SHUFFLE_SIMT_PARAMS, _is_simt_param)
+    SHUFFLE_AMPERE_PARAMS = []
+    IMPLGEMM_SIMT_PARAMS = _force_static_params(IMPLGEMM_SIMT_PARAMS, _is_simt_param)
+    IMPLGEMM_AMPERE_PARAMS = []
+    _clear_turing_volta()
+elif _DTK_KERNEL_FILTER == "dtk_tensorop":
+    SHUFFLE_SIMT_PARAMS = _force_static_params(SHUFFLE_SIMT_PARAMS, _is_simt_param)
+    SHUFFLE_AMPERE_PARAMS = _filter_params(SHUFFLE_AMPERE_PARAMS, _is_ampere_no_int8_static_param)
+    IMPLGEMM_SIMT_PARAMS = _force_static_params(IMPLGEMM_SIMT_PARAMS, _is_simt_param)
+    IMPLGEMM_AMPERE_PARAMS = _filter_params(IMPLGEMM_AMPERE_PARAMS, _is_ampere_no_int8_static_param)
+    _clear_turing_volta()
+elif _DTK_KERNEL_FILTER == "":
+    pass
+else:
+    raise ValueError(f"unknown SPCONV_DTK_KERNEL_FILTER: {_DTK_KERNEL_FILTER}")
+
+
 ALL_NATIVE_PARAMS = SHUFFLE_SIMT_PARAMS + SHUFFLE_TURING_PARAMS + SHUFFLE_VOLTA_PARAMS + SHUFFLE_AMPERE_PARAMS
 
 ALL_IMPGEMM_PARAMS = IMPLGEMM_SIMT_PARAMS + IMPLGEMM_TURING_PARAMS + IMPLGEMM_VOLTA_PARAMS + IMPLGEMM_AMPERE_PARAMS

spconv/csrc/sparse/all.py

@@ -36,7 +36,7 @@ class CustomThrustLib(pccm.Class):
         super().__init__()
         self.add_dependency(ThrustLib)
         # https://github.com/NVIDIA/thrust/issues/1401#issuecomment-806403746
-        if compat.InLinux:
+        if compat.InLinux and os.getenv("CUMM_DTK_DISABLE_INLINE_PTX", "0") != "1":
             self.build_meta.add_public_cflags("nvcc", "-Xcompiler -fno-gnu-unique", "-Xcompiler -fvisibility=hidden")
 
 

spconv/csrc/sparse/convops.py

+import os
 from typing import Optional
 import pccm
 from cumm.common import GemmBasicHost, NlohmannJson, TensorView
@@ -1058,7 +1059,8 @@ class ConvTunerSimple(pccm.ParameterizedClass):
                     continue;
                 }}
             }}
-            if (arch >= std::make_tuple(7, 0) && is_fp16){{
+            bool allow_dtk_fp16_simt = {pccm.boolean(os.getenv("SPCONV_DTK_KERNEL_FILTER", "").lower() == "dtk_simt")};
+            if (arch >= std::make_tuple(7, 0) && is_fp16 && !allow_dtk_fp16_simt){{
                 // skip simt fp16 kernels if we have tensor core
                 if (desp.algo == {pccm.literal(GemmAlgo.Simt.value)}){{
                     continue;

spconv/csrc/sparse/indices.py

@@ -35,6 +35,7 @@ class CudaCommonKernel(pccm.ParameterizedClass):
     @pccm.cuda.cuda_global_function
     def arange_kernel(self):
         code = pccm.FunctionCode()
+        code.add_pre_attr("__launch_bounds__(1024)")
         code.targ("T")
         code.arg("data", f"T*")
         code.arg("size", f"int")
@@ -48,6 +49,7 @@ class CudaCommonKernel(pccm.ParameterizedClass):
     @pccm.cuda.cuda_global_function
     def fill_kernel(self):
         code = pccm.FunctionCode()
+        code.add_pre_attr("__launch_bounds__(1024)")
         code.targ("T")
         code.arg("data", f"T*")
         code.arg("val", f"T")
@@ -62,6 +64,7 @@ class CudaCommonKernel(pccm.ParameterizedClass):
     @pccm.cuda.cuda_global_function
     def maximum_value_kernel(self):
         code = pccm.FunctionCode()
+        code.add_pre_attr("__launch_bounds__(1024)")
         code.targ("T")
         code.arg("data", f"T*")
         code.arg("val", f"T")
@@ -723,6 +726,7 @@ class SparseConvIndicesKernel(pccm.ParameterizedClass):
     @pccm.cuda.cuda_global_function
     def build_subm_conv_hash_table(self):
         code = pccm.FunctionCode()
+        code.add_pre_attr("__launch_bounds__(1024)")
         code.targ("TTable")
         code.targ("TLayoutNPQ")
 
@@ -806,6 +810,7 @@ class SparseConvIndicesKernel(pccm.ParameterizedClass):
     @pccm.cuda.cuda_global_function
     def calc_subm_conv_indices_mask(self):
         code = pccm.FunctionCode()
+        code.add_pre_attr("__launch_bounds__(1024)")
         code.targ("TTable")
         code.targ("TConvLocIter")
         code.arg("loc_iter", f"TConvLocIter")  # [N, ndim + 1]

spconv/pytorch/cppcore.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from cumm import tensorview as tv
+import os
 import torch
 from typing import Dict, Optional, List, Union
 from spconv.constants import AllocKeys
@@ -100,7 +101,12 @@ def get_current_stream():
 
 
 def get_arch():
-    arch = torch.cuda.get_device_capability()
+    force_arch = os.getenv("SPCONV_FORCE_CUDA_ARCH", "")
+    if force_arch:
+        force_arch = force_arch.replace(".", "")
+        arch = (int(force_arch[:-1]), int(force_arch[-1]))
+    else:
+        arch = torch.cuda.get_device_capability()
     if not CompileInfo.arch_is_compatible(arch) and not CompileInfo.algo_can_use_ptx((0, 0), arch):
         warnings.warn(
             f"[WARNING]your gpu arch {arch} isn't compiled in prebuilt, "