warp-lang 0.9.0__py3-none-win_amd64.whl → 0.11.0__py3-none-win_amd64.whl

warp/native/warp.h

@@ -32,10 +32,12 @@ extern "C"
     WP_API void* alloc_host(size_t s);
     WP_API void* alloc_pinned(size_t s);
     WP_API void* alloc_device(void* context, size_t s);
+    WP_API void* alloc_temp_device(void* context, size_t s);
 
     WP_API void free_host(void* ptr);
     WP_API void free_pinned(void* ptr);
     WP_API void free_device(void* context, void* ptr);
+    WP_API void free_temp_device(void* context, void* ptr);
 
     // all memcpys are performed asynchronously
     WP_API void memcpy_h2h(void* dest, void* src, size_t n);
@@ -49,24 +51,24 @@ extern "C"
     WP_API void memset_device(void* context, void* dest, int value, size_t n);
     
     // takes srcsize bytes starting at src and repeats them n times at dst (writes srcsize * n bytes in total):
-    WP_API void memtile_host(void* dest, void *src, size_t srcsize, size_t n);
-    WP_API void memtile_device(void* context, void* dest, void *src, size_t srcsize, size_t n);
+    WP_API void memtile_host(void* dest, const void* src, size_t srcsize, size_t n);
+    WP_API void memtile_device(void* context, void* dest, const void* src, size_t srcsize, size_t n);
 
-	WP_API uint64_t bvh_create_host(wp::vec3* lowers, wp::vec3* uppers, int num_bounds);
+	WP_API uint64_t bvh_create_host(wp::vec3* lowers, wp::vec3* uppers, int num_items);
 	WP_API void bvh_destroy_host(uint64_t id);
     WP_API void bvh_refit_host(uint64_t id);
 
-	WP_API uint64_t bvh_create_device(void* context, wp::vec3* lowers, wp::vec3* uppers, int num_bounds);
+	WP_API uint64_t bvh_create_device(void* context, wp::vec3* lowers, wp::vec3* uppers, int num_items);
 	WP_API void bvh_destroy_device(uint64_t id);
     WP_API void bvh_refit_device(uint64_t id);
 
     // create a user-accessible copy of the mesh, it is the 
     // users responsibility to keep-alive the points/tris data for the duration of the mesh lifetime
-	WP_API uint64_t mesh_create_host(wp::array_t<wp::vec3> points, wp::array_t<wp::vec3> velocities, wp::array_t<int> tris, int num_points, int num_tris);
+	WP_API uint64_t mesh_create_host(wp::array_t<wp::vec3> points, wp::array_t<wp::vec3> velocities, wp::array_t<int> tris, int num_points, int num_tris, int support_winding_number);
 	WP_API void mesh_destroy_host(uint64_t id);
     WP_API void mesh_refit_host(uint64_t id);
 
-	WP_API uint64_t mesh_create_device(void* context, wp::array_t<wp::vec3> points, wp::array_t<wp::vec3> velocities, wp::array_t<int> tris, int num_points, int num_tris);
+	WP_API uint64_t mesh_create_device(void* context, wp::array_t<wp::vec3> points, wp::array_t<wp::vec3> velocities, wp::array_t<int> tris, int num_points, int num_tris, int support_winding_number);
 	WP_API void mesh_destroy_device(uint64_t id);
     WP_API void mesh_refit_device(uint64_t id);
 
@@ -107,11 +109,19 @@ extern "C"
     WP_API size_t array_copy_host(void* dst, void* src, int dst_type, int src_type, int elem_size);
     WP_API size_t array_copy_device(void* context, void* dst, void* src, int dst_type, int src_type, int elem_size);
 
-    WP_API void array_inner_host(uint64_t a, uint64_t b, uint64_t out, int len);
-    WP_API void array_sum_host(uint64_t a, uint64_t out, int len);
+    // generic fill for non-contiguous arrays
+    WP_API void array_fill_host(void* arr, int arr_type, const void* value, int value_size);
+    WP_API void array_fill_device(void* context, void* arr, int arr_type, const void* value, int value_size);
 
-    WP_API void array_inner_device(uint64_t a, uint64_t b, uint64_t out, int len);
-    WP_API void array_sum_device(uint64_t a, uint64_t out, int len);
+    WP_API void array_inner_float_host(uint64_t a, uint64_t b, uint64_t out, int count, int stride_a, int stride_b, int type_len);
+    WP_API void array_inner_double_host(uint64_t a, uint64_t b, uint64_t out, int count, int stride_a, int stride_b, int type_len);
+    WP_API void array_inner_float_device(uint64_t a, uint64_t b, uint64_t out, int count, int stride_a, int stride_b, int type_len);
+    WP_API void array_inner_double_device(uint64_t a, uint64_t b, uint64_t out, int count, int stride_a, int stride_b, int type_len);
+
+    WP_API void array_sum_float_device(uint64_t a, uint64_t out, int count, int stride, int type_len);
+    WP_API void array_sum_float_host(uint64_t a, uint64_t out, int count, int stride, int type_len);
+    WP_API void array_sum_double_host(uint64_t a, uint64_t out, int count, int stride, int type_len);
+    WP_API void array_sum_double_device(uint64_t a, uint64_t out, int count, int stride, int type_len);
 
     WP_API void array_scan_int_host(uint64_t in, uint64_t out, int len, bool inclusive);
     WP_API void array_scan_float_host(uint64_t in, uint64_t out, int len, bool inclusive);
@@ -119,8 +129,92 @@ extern "C"
     WP_API void array_scan_int_device(uint64_t in, uint64_t out, int len, bool inclusive);
     WP_API void array_scan_float_device(uint64_t in, uint64_t out, int len, bool inclusive);
 
+    WP_API void radix_sort_pairs_int_host(uint64_t keys, uint64_t values, int n);
+    WP_API void radix_sort_pairs_int_device(uint64_t keys, uint64_t values, int n);
+
+    WP_API void runlength_encode_int_host(uint64_t values, uint64_t run_values, uint64_t run_lengths, uint64_t run_count, int n);
+    WP_API void runlength_encode_int_device(uint64_t values, uint64_t run_values, uint64_t run_lengths, uint64_t run_count, int n);
+
+    WP_API int bsr_matrix_from_triplets_float_host(
+        int rows_per_block,
+        int cols_per_block,
+        int row_count,
+        int nnz,
+        uint64_t tpl_rows,
+        uint64_t tpl_columns,
+        uint64_t tpl_values,
+        uint64_t bsr_offsets,
+        uint64_t bsr_columns,
+        uint64_t bsr_values);
+    WP_API int bsr_matrix_from_triplets_double_host(
+        int rows_per_block,
+        int cols_per_block,
+        int row_count,
+        int nnz,
+        uint64_t tpl_rows,
+        uint64_t tpl_columns,
+        uint64_t tpl_values,
+        uint64_t bsr_offsets,
+        uint64_t bsr_columns,
+        uint64_t bsr_values);
+
+    WP_API int bsr_matrix_from_triplets_float_device(
+        int rows_per_block,
+        int cols_per_block,
+        int row_count,
+        int nnz,
+        uint64_t tpl_rows,
+        uint64_t tpl_columns,
+        uint64_t tpl_values,
+        uint64_t bsr_offsets,
+        uint64_t bsr_columns,
+        uint64_t bsr_values);
+    WP_API int bsr_matrix_from_triplets_double_device(
+        int rows_per_block,
+        int cols_per_block,
+        int row_count,
+        int nnz,
+        uint64_t tpl_rows,
+        uint64_t tpl_columns,
+        uint64_t tpl_values,
+        uint64_t bsr_offsets,
+        uint64_t bsr_columns,
+        uint64_t bsr_values);
+
+    WP_API void bsr_transpose_float_host(int rows_per_block, int cols_per_block,
+        int row_count, int col_count, int nnz,
+        uint64_t bsr_offsets, uint64_t bsr_columns,
+        uint64_t bsr_values,
+        uint64_t transposed_bsr_offsets,
+        uint64_t transposed_bsr_columns,
+        uint64_t transposed_bsr_values);
+    WP_API void bsr_transpose_double_host(int rows_per_block, int cols_per_block,
+        int row_count, int col_count, int nnz,
+        uint64_t bsr_offsets, uint64_t bsr_columns,
+        uint64_t bsr_values,
+        uint64_t transposed_bsr_offsets,
+        uint64_t transposed_bsr_columns,
+        uint64_t transposed_bsr_values);
+
+    WP_API void bsr_transpose_float_device(int rows_per_block, int cols_per_block,
+        int row_count, int col_count, int nnz,
+        uint64_t bsr_offsets, uint64_t bsr_columns,
+        uint64_t bsr_values,
+        uint64_t transposed_bsr_offsets,
+        uint64_t transposed_bsr_columns,
+        uint64_t transposed_bsr_values);
+    WP_API void bsr_transpose_double_device(int rows_per_block, int cols_per_block,
+        int row_count, int col_count, int nnz,
+        uint64_t bsr_offsets, uint64_t bsr_columns,
+        uint64_t bsr_values,
+        uint64_t transposed_bsr_offsets,
+        uint64_t transposed_bsr_columns,
+        uint64_t transposed_bsr_values);
+
+
     WP_API int cuda_driver_version();   // CUDA driver version
     WP_API int cuda_toolkit_version();  // CUDA Toolkit version used to build Warp
+    WP_API bool cuda_driver_is_initialized();
 
     WP_API int nvrtc_supported_arch_count();
     WP_API void nvrtc_supported_archs(int* archs);
@@ -130,7 +224,12 @@ extern "C"
     WP_API void cuda_device_primary_context_release(int ordinal);
     WP_API const char* cuda_device_get_name(int ordinal);
     WP_API int cuda_device_get_arch(int ordinal);
+    WP_API void cuda_device_get_uuid(int ordinal, char uuid[16]);
+    WP_API int cuda_device_get_pci_domain_id(int ordinal);
+    WP_API int cuda_device_get_pci_bus_id(int ordinal);
+    WP_API int cuda_device_get_pci_device_id(int ordinal);
     WP_API int cuda_device_is_uva(int ordinal);
+    WP_API int cuda_device_is_memory_pool_supported(int ordinal);
 
     WP_API void* cuda_context_get_current();
     WP_API void cuda_context_set_current(void* context);
@@ -140,6 +239,7 @@ extern "C"
     WP_API void cuda_context_destroy(void* context);
     WP_API int cuda_context_get_device_ordinal(void* context);
     WP_API int cuda_context_is_primary(void* context);
+    WP_API int cuda_context_is_memory_pool_supported(void* context);
     WP_API void* cuda_context_get_stream(void* context);
     WP_API void cuda_context_set_stream(void* context, void* stream);
     WP_API int cuda_context_can_access_peer(void* context, void* peer_context);
@@ -172,7 +272,7 @@ extern "C"
     WP_API void* cuda_load_module(void* context, const char* ptx);
     WP_API void cuda_unload_module(void* context, void* module);
     WP_API void* cuda_get_kernel(void* context, void* module, const char* name);
-    WP_API size_t cuda_launch_kernel(void* context, void* kernel, size_t dim, void** args);
+    WP_API size_t cuda_launch_kernel(void* context, void* kernel, size_t dim, int max_blocks, void** args);
 
     WP_API void cuda_set_context_restore_policy(bool always_restore);
     WP_API int cuda_get_context_restore_policy();
@@ -184,4 +284,4 @@ extern "C"
     WP_API void cuda_graphics_unregister_resource(void* context, void* resource);
 
 } // extern "C"
- 
+ 

warp/optim/__init__.py

@@ -6,3 +6,4 @@
 # license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 from .adam import Adam
+from .sgd import SGD