xmos-ai-tools 1.3.2.dev294__py3-none-macosx_10_15_universal2.whl → 1.3.2.dev314__py3-none-macosx_10_15_universal2.whl

xmos_ai_tools/runtime/buildfiles/aitoolslib.cmake

@@ -5,7 +5,7 @@ set(XMOS_AITOOLSLIB_DEFINITIONS
 		"NO_INTERPRETER"
     )
 
-if("${APP_BUILD_ARCH}" STREQUAL xs3a OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL XCORE_XS3A)
+if("${APP_BUILD_ARCH}" STREQUAL xs3a OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL XCORE_XS3A OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL XCORE_XS)
 	set(XMOS_AITOOLSLIB_LIBRARIES "${CMAKE_CURRENT_LIST_DIR}/../lib/libxtflitemicro.a")
 else()
 	set(XMOS_AITOOLSLIB_LIBRARIES "${CMAKE_CURRENT_LIST_DIR}/../lib/libhost_xtflitemicro.a")

xmos_ai_tools/runtime/include/lib_nn/api/nn_arch.h

@@ -0,0 +1,17 @@
+#pragma once
+
+#include "nn_api.h"
+
+typedef enum {
+  TARGET_ARCH_XS3A = 0,
+  TARGET_ARCH_VX4A = 1,
+} nn_target_arch_t;
+
+extern nn_target_arch_t NN_ARCH;
+
+C_API void SetNNTargetArch(nn_target_arch_t arch);
+
+typedef enum {
+  VLMUL_SHR_XS3A = 14,
+  VLMUL_SHR_VX4A = 15,
+} nn_vlmul_shr_t;

xmos_ai_tools/runtime/include/lib_nn/api/nn_op_helper.h

@@ -106,9 +106,15 @@ static inline unsigned smin(const unsigned a, const unsigned b) {
 
 static inline int ceil_log2(uint32_t a) {
   if (a == 0) return -1;
-#if defined(__XS3A__)
+#ifdef __xcore__
   unsigned x;
-  asm("clz %0, %1" : "=r"(x) : "r"(a));
+  #ifdef __XS3A__
+    asm("clz %0, %1" : "=r"(x) : "r"(a));
+  #endif
+  #ifdef __VX4A__
+    asm("xm.clz %0, %1" : "=r"(x) : "r"(a));
+  #endif
+
   unsigned y = 31 - x;
 
   //  clz(1) = 31 -> 31-31 = 0 -> 2^0 = 1

xmos_ai_tools/runtime/include/lib_nn/api/vpu_sim.h

@@ -28,6 +28,7 @@ typedef enum {
   MODE_S32 = 0x00,
   MODE_S16 = 0x100,
   MODE_S8 = 0x200,
+  MODE_S16x8 = 0x400,
 } vector_mode;
 
 C_API

xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_shared_config.h

@@ -2,6 +2,8 @@
 #ifndef XCORE_SHARED_CONFIG_H_
 #define XCORE_SHARED_CONFIG_H_
 
+#include "lib_nn/api/nn_arch.h"
+
 namespace shared_config {
 
 // This string is used as a key to store the shared config
@@ -19,6 +21,8 @@ struct tensor_info_t {
 // The metadata struct must be aligned to 16 bytes
 // We cannot use alignas(16) yet in xcore
 struct xcore_metadata_t {
+  // Target arch can be XS3A = 0, or VX4A = 1
+  nn_target_arch_t target_arch;
   // Versions of libraries used to build the model
   uint32_t lib_nn_major_version;
   uint32_t lib_nn_minor_version;

xmos_ai_tools/runtime/include/lib_tflite_micro/src/thread_call.h

@@ -2,6 +2,11 @@
 #define __micro_thread_library_h__
 
 #include <stdint.h>
+#ifdef __xcore__
+#include <xcore/parallel.h>
+#else
+typedef unsigned synchronizer_t;
+#endif
 
 #ifdef __cplusplus
 extern "C" {
@@ -21,8 +26,6 @@ typedef struct { // THIS STRUCT MUST BE IN SYNC WITH ASSEMBLY CODE.
     uint32_t id[4];         // Actual IDs
   } thread_ids;             // ids of at most 4 threads - live during invoke
   uint32_t synchroniser;    // synchroniser for threads - live during invoke
-  uint32_t nstackwords;     // nstackwords per stack   - live after load model
-  void *UNSAFE stacks;      // pointer to top of stack - live after load model
 } thread_info_t;
 
 
@@ -31,56 +34,15 @@ typedef struct { // THIS STRUCT MUST BE IN SYNC WITH ASSEMBLY CODE.
 typedef void (*thread_function_pointer_t)(void * arg0, void * arg1, void * arg2);
 struct inference_engine;
 
-/** Function that creates threads, then calls a interp_invoke_internal,
- * then destroys threads
- * This function creates four threads for a total of five threads.
- * other versions of the functions create 3, 2, 1, or 0 threads.
- *
- * \param   ie     Pointer to the inference object to be passed to
- *                 interp_invoke_internal
- * \param   ptr    Pointer to a thread_info block. The thread-ids will
- *                 be stored in this block, and a stack pointer is expected
- *                 in this block:
- *                     ptr[0] [out] thread-id-0    (versions with fewer threads
- *                     ptr[1] [out] thread-id-1     will only use the first few
- *                     ptr[2] [out] thread-id-2     slots)
- *                     ptr[3] [out] thread-id-3
- *                     ptr[4] [out] synchroniser-id
- *                     ptr[5] [in]  top of stacks
- *                     ptr[6] [in]  number of words per stack
+/** Function that runs the client task
  */
-int thread_invoke_5(struct inference_engine *ie, thread_info_t *ptr);
-int thread_invoke_4(struct inference_engine *ie, thread_info_t *ptr);
-int thread_invoke_3(struct inference_engine *ie, thread_info_t *ptr);
-int thread_invoke_2(struct inference_engine *ie, thread_info_t *ptr);
-int thread_invoke_1(struct inference_engine *ie, thread_info_t *ptr);
+void thread_client(thread_info_t *ptr, int n);
 
-/** Function that creates threads.
- * This function creates four threads for a total of five threads.
- * other versions of the functions create 3, 2, 1, or 0 threads.
- *
- * \param   ptr    Pointer to a thread_info block. The thread-ids will
- *                 be stored in this block, and a stack pointer is expected
- *                 in this block:
- *                     ptr[0] [out] thread-id-0    (versions with fewer threads
- *                     ptr[1] [out] thread-id-1     will only use the first few
- *                     ptr[2] [out] thread-id-2     slots)
- *                     ptr[3] [out] thread-id-3
- *                     ptr[4] [out] synchroniser-id
- *                     ptr[5] [in]  top of stacks
- *                     ptr[6] [in]  number of words per stack
- */
-void thread_init_5(thread_info_t *ptr);
-void thread_init_4(thread_info_t *ptr);
-void thread_init_3(thread_info_t *ptr);
-void thread_init_2(thread_info_t *ptr);
-void thread_init_1(thread_info_t *ptr);
-/** Function that destroys threads. Must be called from the same function that
- * called an _init_ above.
- *
- * \param   ptr    Pointer to a thread_info block.
+/** Function that runs the client task
  */
-void thread_destroy(thread_info_t *ptr);
+static inline void thread_store_sync(thread_info_t *ptr, uint32_t s) {
+  ptr->synchroniser = s;
+}
 
 /** Function that sets up parameters for one of the client threads
  * This particular one passes the second and third arguments to the thread.
@@ -94,7 +56,21 @@ void thread_destroy(thread_info_t *ptr);
  * \param arg2      Third argument for the thread function
  * \param thread_id The thread_id to initialise; one of ptr[0]..ptr[3] above
  */
-void thread_variable_setup(void * arg1, void * arg2, uint32_t thread_id);
+#ifdef __xcore__
+static inline void thread_variable_setup(void * arg1, void * arg2, uint32_t thread_id) {
+#ifdef __VX4A__
+    asm volatile("xm.tsetr %0, 11, %1" :: "r" (thread_id), "r" (arg1));
+    asm volatile("xm.tsetr %0, 12, %1" :: "r" (thread_id), "r" (arg2));
+    asm volatile("xm.tsetr %0, 24, %1" :: "r" (thread_id), "r" (1));
+#else
+    asm volatile("set t[%0]:r1, %1" :: "r" (thread_id), "r" (arg1));
+    asm volatile("set t[%0]:r2, %1" :: "r" (thread_id), "r" (arg2));
+    asm volatile("set t[%0]:r10, %1" :: "r" (thread_id), "r" (1));
+#endif
+}
+#else
+extern void thread_variable_setup(void * arg1, void * arg2, uint32_t thread_id);
+#endif
 
 /** Function that starts all thread functions and runs them until completion.
  * It is assumed that the variable parts have been set up per thread.

{xmos_ai_tools-1.3.2.dev294.dist-info → xmos_ai_tools-1.3.2.dev314.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xmos_ai_tools
-Version: 1.3.2.dev294
+Version: 1.3.2.dev314
 Summary: XMOS AI Tools
 Home-page: https://github.com/xmos/ai_tools
 Author: XMOS
@@ -41,4 +41,4 @@ Dynamic: summary
 Documentation
 -------------
 
-Click [here](https://github.com/xmos/ai_tools/blob/113f04f7d5485ed2f43037ae48bc5d8772acd831/README.md) for documentation on using xmos-ai-tools to deploy AI models on xcore.
+Click [here](https://github.com/xmos/ai_tools/blob/e358ca58e3fa05f0132d216e1dfefac098d2bb63/README.md) for documentation on using xmos-ai-tools to deploy AI models on xcore.

{xmos_ai_tools-1.3.2.dev294.dist-info → xmos_ai_tools-1.3.2.dev314.dist-info}/RECORD

@@ -1,7 +1,7 @@
 xmos_ai_tools/__init__.py,sha256=AEb1P1AVQeYBcBsImZNpg46juG7y-0EQU-eRVczLUgc,172
 xmos_ai_tools/io_server/__init__.py,sha256=Wr8MHyIxhY80hnvwUH0AsE3BojqIt-AwbYBj-02HDiM,4801
 xmos_ai_tools/runtime/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-xmos_ai_tools/runtime/buildfiles/aitoolslib.cmake,sha256=7xlnDRNVsEIxu929Nt887drbV104t_Nsye1Rv_FpasA,478
+xmos_ai_tools/runtime/buildfiles/aitoolslib.cmake,sha256=yxN1_9nhMo2PVQ3A7uroPjBudDB_Kx6_DcJ5eik5W50,525
 xmos_ai_tools/runtime/buildfiles/aitoolslib.make,sha256=D4GFs74BsWTuQT6R_yxXwR4LLH0pzyFiAXnGpe5hntI,253
 xmos_ai_tools/runtime/include/flash_server.h,sha256=4OQaauv9maSV9pFnxmf6R1LL_TzgnFdXeVPPVWSKPhU,3322
 xmos_ai_tools/runtime/include/ioserver.h,sha256=gEVBhuVLjwi9c00vHUKDEByibl6GVtSUpl5ePAk5GiE,1045
@@ -45,12 +45,13 @@ xmos_ai_tools/runtime/include/lib_nn/api/expand_8_to_16.h,sha256=7-HFoUFgiOjgyGl
 xmos_ai_tools/runtime/include/lib_nn/api/multiply_int16.h,sha256=gBgXPvMYhZ9ZIqKthWKSeWmLPwhFCy6eKKmj0f1M2JQ,1551
 xmos_ai_tools/runtime/include/lib_nn/api/multiply_int16_transform.h,sha256=Vaycg971CZkR9wHO1k3-DxHxp026hgDYqyfZI_YNSy4,2454
 xmos_ai_tools/runtime/include/lib_nn/api/nn_api.h,sha256=UhLAFqf245v5XdcAohZOInOcbsRqJSHGCGqI1iBmu3U,219
+xmos_ai_tools/runtime/include/lib_nn/api/nn_arch.h,sha256=RRRL3AOAQxiFR0qwQaqDvNMnWsmpF4Jb_N5kud2Q0jo,283
 xmos_ai_tools/runtime/include/lib_nn/api/nn_bin_types.h,sha256=DTFX2ecfzwl-_tHlqTR2xypgK7oxmvVWc58eJqJK2rE,325
 xmos_ai_tools/runtime/include/lib_nn/api/nn_config.h,sha256=KFoA5zyOTTHi6x-qDuIWkF8wAz4Gc4BdOBgUzOArOD0,11558
 xmos_ai_tools/runtime/include/lib_nn/api/nn_conv2d_structs.h,sha256=Kfyq3bD1aLcSnWTB_cijU_D5ThGol5utJ1k-WsjqeMY,2268
 xmos_ai_tools/runtime/include/lib_nn/api/nn_image.h,sha256=gz7NyxmZACZiDED0wT7px2qjEEMcUwfa5TjeOUmYzL4,514
 xmos_ai_tools/runtime/include/lib_nn/api/nn_layers.h,sha256=iM2hvfUp_yPK0cp0Bt0heRK97moepModnsNIbTsu4-s,10817
-xmos_ai_tools/runtime/include/lib_nn/api/nn_op_helper.h,sha256=r8FexVvHL0MGjZl8hnshO42pS3PlWk-tTuxA14FoGvY,3722
+xmos_ai_tools/runtime/include/lib_nn/api/nn_op_helper.h,sha256=xmOZZFZ6quj7mc2MFvz-nMtAcIy088OrnVKy08qxKHg,3819
 xmos_ai_tools/runtime/include/lib_nn/api/nn_op_utils.h,sha256=zdCGBeyiHx4-T1hZ4Cjf-QTtXwIyWp4xzuI1dfOjHPA,4948
 xmos_ai_tools/runtime/include/lib_nn/api/nn_operator.h,sha256=Nj4jiXVHlnFjVBtV7mCGulTzSyNwCIpitz8Ao9kSIsU,337
 xmos_ai_tools/runtime/include/lib_nn/api/nn_pooling.h,sha256=DUJNIhtabaMJvP4m46BICDNY2nCSXWKdfE0AUu7VhaY,24456
@@ -66,7 +67,7 @@ xmos_ai_tools/runtime/include/lib_nn/api/quantize_int16_transform.h,sha256=Rg2XO
 xmos_ai_tools/runtime/include/lib_nn/api/version.h,sha256=e6-GRd2_R0vtq6Jps57nM3F0Zt37ZpClL7u3Er0W1ZI,303
 xmos_ai_tools/runtime/include/lib_nn/api/vpu_memmove_word_aligned.h,sha256=-OgMM0JMG1kIlL6YSiEg-6o6t43t1LUDwAS_inW-r_8,579
 xmos_ai_tools/runtime/include/lib_nn/api/vpu_memset_256.h,sha256=u71fu0EldAXMiMfQ-il9pN9ZXUkrcSp7kiMcJRnESRE,1967
-xmos_ai_tools/runtime/include/lib_nn/api/vpu_sim.h,sha256=3S1_2eSaEj2q8nhFuRiCefQhwTeYpfothuFu0fC54_s,3811
+xmos_ai_tools/runtime/include/lib_nn/api/vpu_sim.h,sha256=0ujuOUFwKd8YYVgVAvBzFlxLsND_EBLRDhuevrzvU7Y,3833
 xmos_ai_tools/runtime/include/lib_nn/api/xs3_vpu.h,sha256=xuEpNayaSWeEIoL1--EfQAXkr4TidwDASFYgbshJ-Mw,6227
 xmos_ai_tools/runtime/include/lib_nn/api/xs3a_registers.h,sha256=5i6Tue1VplRI2VlIes-gaD2BXOKreuUit51OYTsbgFs,139774
 xmos_ai_tools/runtime/include/lib_nn/src/asm/asm_constants.h,sha256=VB8pk_H4vI4S18xYfPFCOGZoLG6S0FYW5uVfG3u9glU,1018
@@ -78,8 +79,8 @@ xmos_ai_tools/runtime/include/lib_tflite_micro/api/memory_parallel_transport.h,s
 xmos_ai_tools/runtime/include/lib_tflite_micro/api/version.h,sha256=ImFxGU2PzLZ3TzGiB1V2Ghgd2yh0qWR7UI7ao-MrG6w,318
 xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_config.h,sha256=gywR3hBJe05GTLkcdJss0bmk3dzsWivXK6U9FF7Pfm0,589
 xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_device_memory.h,sha256=bLWcRDNrzClLh8_eR3XRRz3sA2pEAzzxGLDoPTsNp8A,1917
-xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_shared_config.h,sha256=bcpjAc7bfGrxovRkBZIr3hu4vpayX-TdhjCNrYXzQik,1451
-xmos_ai_tools/runtime/include/lib_tflite_micro/src/thread_call.h,sha256=KNG-3gNWMSonzYTGbk4L0kT8WU9JD0bOzlk8naz7JO0,4965
+xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_shared_config.h,sha256=EdqTecMZIdSk1ZWUE7rdl6ilXe4-vbFiwwcf0Y124c4,1562
+xmos_ai_tools/runtime/include/lib_tflite_micro/src/thread_call.h,sha256=YobSgoLTISGz3o7mz-ZXiLNlJWnko5R44NmYh8SvxkQ,3206
 xmos_ai_tools/runtime/include/lib_tflite_micro/src/tflite-xcore-kernels/conv2d_float.h,sha256=XfVVWAKhbv6nByOzO1j6WA4oQ4RTX3PvYx22f-_3JFA,7681
 xmos_ai_tools/runtime/include/lib_tflite_micro/src/tflite-xcore-kernels/xcore_custom_options.h,sha256=lC4Tw1Pxxg3zOXRdqNNtokuU-_cX9TTkYmGLe47-9dQ,630
 xmos_ai_tools/runtime/include/lib_tflite_micro/src/tflite-xcore-kernels/xcore_error_reporter.h,sha256=_NIzvBYMqlwJexYESP5t5JXpxYTt-ZKq-1AdqAB9-Sc,812
@@ -379,17 +380,17 @@ xmos_ai_tools/runtime/include/tensorflow/lite/micro/tools/benchmarking/metrics.h
 xmos_ai_tools/runtime/include/tensorflow/lite/micro/tools/benchmarking/op_resolver.h,sha256=g0dl9tzUqngiINvjBlqDclFqvkC85MC4kbU13vE4OkY,6071
 xmos_ai_tools/runtime/include/tensorflow/lite/schema/schema_generated.h,sha256=hSumwEqc-LmMHG82KKl1yPLiVMNbLj_q19-VRPCR7ng,1093175
 xmos_ai_tools/runtime/include/tensorflow/lite/schema/schema_utils.h,sha256=tkHMDPARjIqppYCVInIowwdHxjNP3pfSS9O7vx-ODeo,1333
-xmos_ai_tools/runtime/lib/libhost_xtflitemicro.a,sha256=-DhzLYrSOKlBtk7BbC8KMQ5mxIgUhTtaLN2yjSyBK_g,2470136
-xmos_ai_tools/runtime/lib/libxtflitemicro.a,sha256=KqyxIQSTcpcoQxwDO-N9rwU_xgghlSgO6Y0YlxPU00s,71877202
+xmos_ai_tools/runtime/lib/libhost_xtflitemicro.a,sha256=zNVlVeITK-V46GeccRWRBh-e33bLq1zJr4jpj6b-q1U,2479336
+xmos_ai_tools/runtime/lib/libxtflitemicro.a,sha256=SQU77BXngIHnTu1q9o6_no3cvvMoL93yCSy8cBj9F8E,72202704
 xmos_ai_tools/xformer/__init__.py,sha256=a4K06oiSu2TYVwGvzwMDGGejPUYTvW9-5Uw6SfuFCX4,1903
 xmos_ai_tools/xformer/flash.py,sha256=MG4coi_Lvvg-oQmw1pomJD8eeOH4gAMjixjBFvO2BCk,6376
 xmos_ai_tools/xinterpreters/__init__.py,sha256=PFRB9VxOLKaA--j2ZvWGcmesv2C6uNYqJ_kBam68aUI,50
 xmos_ai_tools/xinterpreters/exceptions.py,sha256=HOjADxHYMPI9mN0YIbWxtw9hSeL2B6XWWwqtGtyJdVs,577
 xmos_ai_tools/xinterpreters/host_interpreter.py,sha256=ZdvzXUnAdnCcSU-dNQvq6ad2MDk9PZqu9UyNoBgBYKI,25755
-xmos_ai_tools/xinterpreters/libs/macos/xtflm_python.1.0.1.dylib,sha256=tKwL8dJLDWfttTdKQzAONMlWPzOxUVUZcyQIzck__DU,2359448
-xmos_ai_tools/xinterpreters/libs/macos/xtflm_python.dylib,sha256=tKwL8dJLDWfttTdKQzAONMlWPzOxUVUZcyQIzck__DU,2359448
-xmos_ai_tools-1.3.2.dev294.data/data/bin/xcore-opt,sha256=09Pzk2TD7TzzkS73oZC2ZOUSC1HYPMNSaZZl4WKTuhE,286566280
-xmos_ai_tools-1.3.2.dev294.dist-info/METADATA,sha256=pN1sKQ1zAibNroaAd_Q28YHt31IKllRirM1gCCcHi5k,1574
-xmos_ai_tools-1.3.2.dev294.dist-info/WHEEL,sha256=zptREcKej0ZTXp5pemGtOgadxZp76B2ndJnv7YH9sHw,112
-xmos_ai_tools-1.3.2.dev294.dist-info/top_level.txt,sha256=YWegea73ll3tMlRWRdHJemUy2VOuEYDdOIaffxu_eF0,14
-xmos_ai_tools-1.3.2.dev294.dist-info/RECORD,,
+xmos_ai_tools/xinterpreters/libs/macos/xtflm_python.1.0.1.dylib,sha256=mQp_HoqoSMCxV3Wj1MVVSjtoobePEkDJxgXolJmJsPA,2359560
+xmos_ai_tools/xinterpreters/libs/macos/xtflm_python.dylib,sha256=mQp_HoqoSMCxV3Wj1MVVSjtoobePEkDJxgXolJmJsPA,2359560
+xmos_ai_tools-1.3.2.dev314.data/data/bin/xcore-opt,sha256=F9lfM3nO9moykySibqzSvqf0fR_dl-SOCiumIEa2nVw,286686296
+xmos_ai_tools-1.3.2.dev314.dist-info/METADATA,sha256=4PARxhHjjMdc_jnNIG63WgAEbsQ_UfR3ww_QpZllsZ8,1574
+xmos_ai_tools-1.3.2.dev314.dist-info/WHEEL,sha256=98VLpmTXZiW_X2RJ0T4x1A5tW4Utj3RcaNGBZZbavbg,112
+xmos_ai_tools-1.3.2.dev314.dist-info/top_level.txt,sha256=YWegea73ll3tMlRWRdHJemUy2VOuEYDdOIaffxu_eF0,14
+xmos_ai_tools-1.3.2.dev314.dist-info/RECORD,,

{xmos_ai_tools-1.3.2.dev294.dist-info → xmos_ai_tools-1.3.2.dev314.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (78.1.0)
+Generator: setuptools (79.0.1)
 Root-Is-Purelib: false
 Tag: py3-none-macosx_10_15_universal2