python and cutile in build not host

Author: divyegala

conda/recipes/libcuvs/recipe.yaml

@@ -72,6 +72,7 @@ cache:
       - ninja
       - if: cuda_major == "13"
         then:
+          - python
           - cutile-python
           - cuda-tileiras
       - ${{ stdlib("c") }}

cpp/cmake/modules/generate_cutile_kernels.cmake

@@ -36,7 +36,17 @@ function(_cutile_kernels_setup)
   set(multi_value)
   cmake_parse_arguments(_CUTILE "${options}" "${one_value}" "${multi_value}" ${ARGN})
 
-  find_package(Python3 REQUIRED COMPONENTS Interpreter)
+  if(DEFINED ENV{BUILD_PREFIX})
+    find_program(
+      _cutile_build_python
+      NAMES python3 python
+      PATHS "$ENV{BUILD_PREFIX}/bin"
+      NO_DEFAULT_PATH REQUIRED NO_CACHE
+    )
+    set(Python3_EXECUTABLE "${_cutile_build_python}")
+  else()
+    find_package(Python3 REQUIRED COMPONENTS Interpreter)
+  endif()
   find_package(CUDAToolkit REQUIRED)
 
   if(CUDAToolkit_VERSION VERSION_LESS 13.0)
@@ -68,6 +78,7 @@ function(_cutile_kernels_setup)
                   "Install cutile-python and cuda-tileiras (conda), or cuda-tile[tileiras] (pip)."
     )
   endif()
+  message(STATUS "Using cuTile Python: ${Python3_EXECUTABLE}")
 
   set_property(
     DIRECTORY

cpp/src/distance/detail/fused_distance_nn/cutile/fused_1nn_cutile_matrix.json

@@ -30,7 +30,7 @@
     ],
     "_tile": [
       {
-        "tile_m": 128,
+        "tile_m": 64,
         "tile_n": 128,
         "tile_k": 32
       }
@@ -63,12 +63,18 @@
       {
         "matrix_layout": "strict",
         "abi_abbrev": "strict",
-        "abi_tag": "cutile_abi_strict"
+        "abi_tag": "cutile_abi_strict",
+        "tile_m": 128,
+        "tile_n": 128,
+        "tile_k": 32
       },
       {
         "matrix_layout": "relaxed",
         "abi_abbrev": "relaxed",
-        "abi_tag": "cutile_abi_relaxed"
+        "abi_tag": "cutile_abi_relaxed",
+        "tile_m": 128,
+        "tile_n": 256,
+        "tile_k": 16
       }
     ],
     "_data": [
@@ -92,13 +98,6 @@
         "index_abbrev": "i32"
       }
     ],
-    "_tile": [
-      {
-        "tile_m": 128,
-        "tile_n": 128,
-        "tile_k": 32
-      }
-    ],
     "_export": [
       {
         "output_format": "cubin",
@@ -109,17 +108,51 @@
         "cc_major": 10,
         "cc_minor": 0,
         "arch_tag": "cutile_arch_10_0"
+      }
+    ]
+  },
+  {
+    "_abi": [
+      {
+        "matrix_layout": "strict",
+        "abi_abbrev": "strict",
+        "abi_tag": "cutile_abi_strict"
       },
       {
-        "output_format": "cubin",
-        "artifact_ext": "cubin",
-        "artifact_basename": "@data_type@_@index_abbrev@_@abi_abbrev@_@gpu_code@",
-        "register": "cubin",
-        "gpu_code": "sm_90",
-        "cc_major": 9,
-        "cc_minor": 0,
-        "arch_tag": "cutile_arch_9_0"
+        "matrix_layout": "relaxed",
+        "abi_abbrev": "relaxed",
+        "abi_tag": "cutile_abi_relaxed"
+      }
+    ],
+    "_data": [
+      {
+        "data_type": "half",
+        "data_abbrev": "h"
       },
+      {
+        "data_type": "float",
+        "data_abbrev": "f"
+      }
+    ],
+    "_metric": [
+      {
+        "metric": "runtime"
+      }
+    ],
+    "_index": [
+      {
+        "index_type": "int32",
+        "index_abbrev": "i32"
+      }
+    ],
+    "_tile": [
+      {
+        "tile_m": 128,
+        "tile_n": 128,
+        "tile_k": 32
+      }
+    ],
+    "_export": [
       {
         "output_format": "tileir_bytecode",
         "artifact_ext": "tilebc",
@@ -136,8 +169,8 @@
         "matrix_layout": "strict",
         "abi_abbrev": "strict",
         "abi_tag": "cutile_abi_strict",
-        "tile_m": 128,
-        "tile_n": 128,
+        "tile_m": 64,
+        "tile_n": 256,
         "tile_k": 32
       },
       {
@@ -146,7 +179,60 @@
         "abi_tag": "cutile_abi_relaxed",
         "tile_m": 128,
         "tile_n": 128,
+        "tile_k": 64
+      }
+    ],
+    "_data": [
+      {
+        "data_type": "half",
+        "data_abbrev": "h"
+      },
+      {
+        "data_type": "float",
+        "data_abbrev": "f"
+      }
+    ],
+    "_metric": [
+      {
+        "metric": "runtime"
+      }
+    ],
+    "_index": [
+      {
+        "index_type": "int32",
+        "index_abbrev": "i32"
+      }
+    ],
+    "_export": [
+      {
+        "output_format": "cubin",
+        "artifact_ext": "cubin",
+        "artifact_basename": "@data_type@_@index_abbrev@_@abi_abbrev@_@gpu_code@",
+        "register": "cubin",
+        "gpu_code": "sm_90",
+        "cc_major": 9,
+        "cc_minor": 0,
+        "arch_tag": "cutile_arch_9_0"
+      }
+    ]
+  },
+  {
+    "_abi": [
+      {
+        "matrix_layout": "strict",
+        "abi_abbrev": "strict",
+        "abi_tag": "cutile_abi_strict",
+        "tile_m": 64,
+        "tile_n": 128,
         "tile_k": 32
+      },
+      {
+        "matrix_layout": "relaxed",
+        "abi_abbrev": "relaxed",
+        "abi_tag": "cutile_abi_relaxed",
+        "tile_m": 64,
+        "tile_n": 128,
+        "tile_k": 64
       }
     ],
     "_data": [