[triton][PR] [triton][Fix] Back out "Revert D86123088", Fix old TMA API not supported issue

Summary:
Original commit changeset: 438f340d41ee

Original Phabricator Diff: D86533448

Import the fix from github relese/3.5.x branch
[release 3.5.x] Fix old TMA API not supported issue #653

Reviewed By: htyu, dshi7

Differential Revision: D86783598

fbshipit-source-id: e66895007c08e45b9ef8047949aa2f3bc8eee6ff

Author: agron911

third_party/nvidia/backend/driver.c

@@ -1,9 +1,15 @@
 #include "cuda.h"
 #include <dlfcn.h>
 #include <stdbool.h>
+#include <stdlib.h>
 #define PY_SSIZE_T_CLEAN
 #include <Python.h>
 
+typedef struct {
+  PyObject_HEAD;
+  _Alignas(128) CUtensorMap tensorMap;
+} PyCUtensorMapObject;
+
 // Raises a Python exception and returns false if code is not CUDA_SUCCESS.
 static bool gpuAssert(CUresult code, const char *file, int line) {
   if (code == CUDA_SUCCESS)
@@ -26,7 +32,7 @@ static bool gpuAssert(CUresult code, const char *file, int line) {
 #define CUDA_CHECK_AND_RETURN_NULL(ans)                                        \
   do {                                                                         \
     if (!gpuAssert((ans), __FILE__, __LINE__))                                 \
-      return NULL;                                                             \
+      goto cleanup;                                                            \
   } while (0)
 
 // To be used inside a Py_{BEGIN,END}_ALLOW_THREADS block.
@@ -44,7 +50,7 @@ static bool gpuAssert(CUresult code, const char *file, int line) {
     if ((funcPointer) == NULL) {                                               \
       (funcPointer) = (initializerFunction)();                                 \
       if ((funcPointer) == NULL) {                                             \
-        return NULL;                                                           \
+        goto cleanup;                                                          \
       }                                                                        \
     }                                                                          \
   } while (0)
@@ -87,6 +93,9 @@ static PyObject *getDeviceProperties(PyObject *self, PyObject *args) {
                        warp_size, "sm_clock_rate", sm_clock_rate,
                        "mem_clock_rate", mem_clock_rate, "mem_bus_width",
                        mem_bus_width);
+
+cleanup:
+  return NULL;
 }
 
 static PyObject *loadBinary(PyObject *self, PyObject *args) {
@@ -238,6 +247,9 @@ static PyObject *occupancyMaxActiveClusters(PyObject *self, PyObject *args) {
       cuOccupancyMaxActiveClusters(&maxActiveClusters, func, &config));
   Py_END_ALLOW_THREADS;
   return PyLong_FromLong(maxActiveClusters);
+
+cleanup:
+  return NULL;
 }
 
 static PyObject *setPrintfFifoSize(PyObject *self, PyObject *args) {
@@ -279,8 +291,43 @@ static PyObject *setPrintfFifoSize(PyObject *self, PyObject *args) {
   Py_RETURN_NONE;
 }
 
+static PyObject *PyCUtensorMap_alloc(PyTypeObject *type, Py_ssize_t n_items) {
+  PyCUtensorMapObject *self = NULL;
+  void *mem = NULL;
+  size_t size = type->tp_basicsize;
+
+  if (posix_memalign(&mem, 128, size) != 0) {
+    PyErr_NoMemory();
+    return NULL;
+  }
+
+  self = (PyCUtensorMapObject *)mem;
+  PyObject_INIT(self, type);
+  return (PyObject *)self;
+}
+
+static void PyCUtensorMap_dealloc(PyObject *self) {
+  Py_TYPE(self)->tp_free(self);
+}
+
+static void PyCUtensorMap_free(void *ptr) { free(ptr); }
+
+// clang-format off
+static PyTypeObject PyCUtensorMapType = {
+    PyVarObject_HEAD_INIT(NULL, 0)
+    .tp_name = "triton.backends.nvidia.PyCUtensorMap",
+    .tp_basicsize = sizeof(PyCUtensorMapObject),
+    .tp_itemsize = 0,
+    .tp_flags = Py_TPFLAGS_DEFAULT,
+    .tp_doc = "<PyCUtensorMap object>",
+    .tp_new = PyType_GenericNew,
+    .tp_alloc = PyCUtensorMap_alloc,
+    .tp_dealloc = (destructor)PyCUtensorMap_dealloc,
+    .tp_free = PyCUtensorMap_free,
+};
+// clang-format on
+
 static PyObject *fillTMADescriptor(PyObject *self, PyObject *args) {
-  unsigned long long desc_address;
   unsigned long long global_address;
   int swizzle;
   int elemSize;
@@ -290,16 +337,20 @@ static PyObject *fillTMADescriptor(PyObject *self, PyObject *args) {
   PyObject *strides;
   int padding;
 
-  if (!PyArg_ParseTuple(args, "KKiiiOOOi", &desc_address, &global_address,
-                        &swizzle, &elemSize, &elemType, &blockSize, &shape,
-                        &strides, &padding)) {
+  if (!PyArg_ParseTuple(args, "KiiiOOOi", &global_address, &swizzle, &elemSize,
+                        &elemType, &blockSize, &shape, &strides, &padding)) {
+    return NULL;
+  }
+
+  PyCUtensorMapObject *desc = (PyCUtensorMapObject *)PyObject_CallObject(
+      (PyObject *)&PyCUtensorMapType, NULL);
+  if (!desc) {
     return NULL;
   }
 
   PyObject *blockSizeFast = NULL;
   PyObject *shapeFast = NULL;
   PyObject *stridesFast = NULL;
-  PyObject *result = NULL;
 
   uint32_t blockSizeInt[5];
   uint64_t shapeInt[5];
@@ -370,17 +421,18 @@ static PyObject *fillTMADescriptor(PyObject *self, PyObject *args) {
   INITIALIZE_FUNCTION_POINTER_IF_NULL(cuTensorMapEncodeTiled,
                                       getCuTensorMapEncodeTiledHandle);
   CUDA_CHECK_AND_RETURN_NULL(cuTensorMapEncodeTiled(
-      (CUtensorMap *)desc_address, elemType, rank, (void *)global_address,
-      shapeInt, stridesLL, blockSizeInt, elementStrides,
-      CU_TENSOR_MAP_INTERLEAVE_NONE, swizzle,
-      CU_TENSOR_MAP_L2_PROMOTION_L2_128B, fill));
-  Py_RETURN_NONE;
+      &desc->tensorMap, elemType, rank, (void *)global_address, shapeInt,
+      stridesLL, blockSizeInt, elementStrides, CU_TENSOR_MAP_INTERLEAVE_NONE,
+      swizzle, CU_TENSOR_MAP_L2_PROMOTION_L2_128B, fill));
+
+  return (PyObject *)desc;
 
 cleanup:
   Py_XDECREF(blockSizeFast);
   Py_XDECREF(shapeFast);
   Py_XDECREF(stridesFast);
-  return result;
+  Py_XDECREF(desc);
+  return NULL;
 }
 
 // Simple helper to experiment creating TMA descriptors on the host.
@@ -426,6 +478,8 @@ static PyObject *fill1DTMADescriptor(PyObject *self, PyObject *args) {
       CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE));
   Py_INCREF(Py_None);
   return Py_None;
+cleanup:
+  return NULL;
 }
 
 // Simple helper to experiment creating TMA descriptors on the host.
@@ -490,6 +544,8 @@ static PyObject *fill2DTMADescriptor(PyObject *self, PyObject *args) {
       CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE));
   Py_INCREF(Py_None);
   return Py_None;
+cleanup:
+  return NULL;
 }
 
 // Simple helper to experiment creating TMA descriptors on the host.
@@ -545,6 +601,8 @@ static PyObject *fill1DTMADescriptorType(PyObject *self, PyObject *args) {
   Py_INCREF(Py_None);
 #endif
   return Py_None;
+cleanup:
+  return NULL;
 }
 
 // Simple helper to experiment creating TMA descriptors on the host.
@@ -619,6 +677,8 @@ static PyObject *fill2DTMADescriptorType(PyObject *self, PyObject *args) {
   Py_INCREF(Py_None);
 #endif
   return Py_None;
+cleanup:
+  return NULL;
 }
 
 static PyMethodDef ModuleMethods[] = {
@@ -651,12 +711,18 @@ static struct PyModuleDef ModuleDef = {PyModuleDef_HEAD_INIT, "cuda_utils",
                                        ModuleMethods};
 
 PyMODINIT_FUNC PyInit_cuda_utils(void) {
+  if (PyType_Ready(&PyCUtensorMapType) < 0) {
+    return NULL;
+  }
+
   PyObject *m = PyModule_Create(&ModuleDef);
   if (m == NULL) {
     return NULL;
   }
 
   PyModule_AddFunctions(m, ModuleMethods);
+  Py_INCREF(&PyCUtensorMapType);
+  PyModule_AddObject(m, "PyCUtensorMap", (PyObject *)&PyCUtensorMapType);
 
   return m;
 }