NPU Driver 1.10.0 release unified 2024WW44 (#57)

New driver and test code snapshot in 1.10.0 version.
Updated compiler to npu_ud_2024_44_rc1.
Updated firmware to ci_tag_ud202444_vpu_rc_20241025_1830 tag.

Signed-off-by: Schulfer, Magdalena <[email protected]>
Co-authored-by: Pawlak, Jakub <[email protected]>
Co-authored-by: Pradzynski, Krystian <[email protected]>
Co-authored-by: Tabaka, Mateusz <[email protected]>
Co-authored-by: Wludzik, Jozef <[email protected]>

Author: 5 people

CMakeLists.txt

@@ -10,7 +10,7 @@ if (${CMAKE_VERSION} VERSION_GREATER_EQUAL "3.24")
     cmake_policy(SET CMP0135 NEW)
 endif()
 
-set(STACK_VERSION 1.8.0 CACHE STRING "Main project version")
+set(STACK_VERSION 1.10.0 CACHE STRING "Main project version")
 project(npu-linux-driver VERSION ${STACK_VERSION})
 
 set(BUILD_NUMBER "dev-0" CACHE STRING "Build number composed of name and unique number used as driver version")

cmake/compiler_flags.cmake

@@ -12,7 +12,7 @@
 
 add_compile_options(
   # Compiler warnings
-  -Wall -Wextra -Werror -Wno-unused-parameter
+  -Wall -Wextra -Werror -Wno-unused-parameter -Wvla
   # Control flow integrity
   $<$<CONFIG:Release>:-fcf-protection=full>
   # Format string defence

cmake/packaging/setup.cmake

@@ -42,7 +42,8 @@ if (EXISTS "/etc/debian_version")
 elseif(EXISTS "/etc/portage")
     set(PACKAGE_TYPE "ebuild")
 else()
-    message(FATAL_ERROR "Unable to detect package type for this system")
+    message(WARNING "Unable to detect package type for this system")
+    return()
 endif()
 
 include(${CMAKE_CURRENT_LIST_DIR}/generators/${PACKAGE_TYPE}.cmake)

compiler/compiler_source.cmake

@@ -17,10 +17,10 @@ endif()
 include(ExternalProject)
 
 # OpenVINO + NPU Plugin package options
-set(OPENVINO_REVISION 7edb05f29487cbf5cc6a7d7ae0a8e228aac763b0)
-set(VPUX_PLUGIN_REVISION 1a83394af6430a8a4a29dfc6faf50ed26594e387)
-set(VPUX_PLUGIN_RELEASE npu_ud_2024_36_rc1)
-set(OPENCV_REVISION 5dc1b39e4c9dfb3339e0b910f7d824a02474ceed)
+set(OPENVINO_REVISION 0ebff040fd22daa37612a82fdf930ffce4ebb099)
+set(VPUX_PLUGIN_REVISION 4f89b7c000d98bb8f4bf5bb058967bbd02834caa)
+set(VPUX_PLUGIN_RELEASE npu_ud_2024_44_rc1)
+set(OPENCV_REVISION 78195bc3dfe20b96e721ae8b32d0aa3491755e78)
 
 # Directories
 set(OPENVINO_PREFIX_DIR "${CMAKE_BINARY_DIR}/third_party/openvino")

compiler/include/npu_driver_compiler.h

@@ -22,8 +22,8 @@
 extern "C" {
 #endif
 
-#define VCL_COMPILER_VERSION_MAJOR 5
-#define VCL_COMPILER_VERSION_MINOR 8
+#define VCL_COMPILER_VERSION_MAJOR 6
+#define VCL_COMPILER_VERSION_MINOR 1
 #define VCL_PROFILING_VERSION_MAJOR 2
 #define VCL_PROFILING_VERSION_MINOR 0
 
@@ -69,7 +69,7 @@ typedef struct __vcl_log_handle_t* vcl_log_handle_t;
 
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Defines type of requested data.
-/// Must be in sync with \b _ze_graph_profiling_type_t
+/// Must be in sync with _ze_graph_profiling_type_t
 typedef enum __vcl_profiling_request_type_t {
     VCL_PROFILING_LAYER_LEVEL = 0x1,
     VCL_PROFILING_TASK_LEVEL = 0x2,
@@ -120,10 +120,9 @@ typedef struct __vcl_profiling_properties_t {
 typedef enum __vcl_platform_t {
     VCL_PLATFORM_UNKNOWN = -1,
 
-    VCL_PLATFORM_VPU3700,  ///< VPU3700
-    VCL_PLATFORM_VPU3720,  ///< NPU3720
-    VCL_PLATFORM_VPU4000,  ///< NPU4000
-
+    VCL_PLATFORM_VPU3700 = 0,  ///< VPU3700
+    VCL_PLATFORM_VPU3720 = 1,  ///< VPU3720
+    VCL_PLATFORM_VPU4000 = 2,  ///< VPU4000
 } vcl_platform_t;
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -164,6 +163,23 @@ typedef struct __vcl_executable_desc_t {
     uint64_t optionsSize;  ///< Size of options
 } vcl_executable_desc_t;
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Defines query description to be passed during query network creation
+///
+///        Format of modelIRData (defined in L0 adaptor):
+///        1. API version : vcl_version_info_t
+///        2. Num of data elements (now only xml + weights = 2) : uint32_t
+///        3. Size of data 1 (xml) : uint64_t
+///        4. Data 1 : $2 bytes
+///        5. Size of data 2 (weights) : uint64_t
+///        6. Data 2 : $4 bytes
+typedef struct __vcl_query_desc_t {
+    const uint8_t* modelIRData;
+    uint64_t modelIRSize;  ///< Size of modelIRData
+    const char* options;   ///< Compiler config options
+    uint64_t optionsSize;  ///< Size of options
+} vcl_query_desc_t;
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Defines input that is required to create profiling handler
 typedef struct __vcl_profiling_input_t {
@@ -196,8 +212,8 @@ VCL_APIEXPORT vcl_result_t VCL_APICALL vclCompilerGetProperties(vcl_compiler_han
 
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Create an querynetwork object and return the handle
-VCL_APIEXPORT vcl_result_t VCL_APICALL vclQueryNetworkCreate(vcl_compiler_handle_t compiler, uint8_t* modelIR,
-                                                             uint64_t modelIRSize, vcl_query_handle_t* query);
+VCL_APIEXPORT vcl_result_t VCL_APICALL vclQueryNetworkCreate(vcl_compiler_handle_t compiler, vcl_query_desc_t desc,
+                                                             vcl_query_handle_t* query);
 
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Retrieve result of query network
@@ -214,6 +230,16 @@ VCL_APIEXPORT vcl_result_t VCL_APICALL vclQueryNetworkDestroy(vcl_query_handle_t
 VCL_APIEXPORT vcl_result_t VCL_APICALL vclExecutableCreate(vcl_compiler_handle_t compiler, vcl_executable_desc_t desc,
                                                            vcl_executable_handle_t* executable);
 
+typedef struct __vcl_allocator_t {
+    uint8_t* (*allocate)(uint64_t);
+    void (*deallocate)(uint8_t*);
+} vcl_allocator_t;
+
+VCL_APIEXPORT vcl_result_t VCL_APICALL vclAllocatedExecutableCreate(vcl_compiler_handle_t compiler,
+                                                                    vcl_executable_desc_t desc,
+                                                                    vcl_allocator_t const* allocator,
+                                                                    uint8_t** blobBuffer, uint64_t* blobSize);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Destroys the executable and releases the cached blob.
 VCL_APIEXPORT vcl_result_t VCL_APICALL vclExecutableDestroy(vcl_executable_handle_t executable);

firmware/bin/vpu_37xx_v0.0.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d0639a72240b89d37e827d89edb3f13f0c995a5b7d8a28e5e48ca326bf4effad
-size 2485040
+oid sha256:2600c098d22d49531edd263c6720e9e385d68ffbd1c87f36d7db86dbf381328a
+size 2428000

firmware/bin/vpu_40xx_v0.0.bin

@@ -62,10 +62,11 @@ typedef enum {
 
 // Number of dimensions for the dynamic task transfer
 typedef enum {
-    DMA_DYN_NUM_DIM_DISABLED = 0, // 1D dynamic task transfer
-    DMA_DYN_NUM_DIM_2D,           // 2D dynamic task transfer
-    DMA_DYN_NUM_DIM_3D,           // 3D dynamic task transfer
-    DMA_DYN_NUM_DIM_MAX
+    DMA_DYN_DIM_DISABLED = 0, // No dynamic task dimensionality
+    DMA_DYN_DIM_2D,           // Enable DESCRIPTOR.*_DIM_SIZE[1]/DESCRIPTOR.*_LIST_SIZE to be dynamic
+    DMA_DYN_DIM_3D,           // Enable DESCRIPTOR.*_DIM_SIZE[2] to be dynamic
+    DMA_DYN_DIM_2D_3D,        // Enable both 2D and 3D dynamic task dimensionality
+    DMA_DYN_DIM_MAX
 } DmaDynamicDimensions;
 
 // Burst Length Encoding
@@ -178,6 +179,12 @@ typedef enum {
     DMA_MODE_MAX,
 } DmaJobMode;
 
+typedef enum {
+    DMA_CTRG_0,
+    DMA_CTRG_1,
+    DMA_CTRG_MAX,
+} DmaCtrgEnum;
+
 #pragma pack(push, 1)
 
 #define DMA_L2CACHE_ALIGNMENT (32) // Descriptors must be 32-byte aligned
@@ -387,7 +394,9 @@ typedef struct ALIGN_DMA(DMA_L2CACHE_ALIGNMENT) {
     uint16_t task_dyn_id;          // Dynamic task phase ID
     uint16_t rsvd8;                // Reserved
     uint32_t task_dyn_addr;        // Dynamic Task address
-    uint64_t pad[2];               // Padding to make all descriptors 32-Byte aligned
+    uint32_t ptr_wr_addr;          // Address used to write the Task Descriptor pointer
+    uint32_t rsvd9;                // Reserved
+    uint64_t pad[1];               // Padding to make all descriptors 32-Byte aligned
 } DmaDescriptor;
 
 static_assert(sizeof(DmaDescriptor) == 192, "DmaDescriptor size != 192");