code format

Author: tsingmicro-public-e

third_party/tsingmicro/backend/compiler.py

@@ -32,6 +32,7 @@ def _get_libc_root() -> str:
         raise Exception("LIB_C_ROOT is not set.")
     return path
 
+
 def _get_core_dialects_to_mk_pass_arg() -> str:
     value = os.getenv("PRECISION_MODE", "0").strip()
     if value not in ("0", "1", "2"):
@@ -50,7 +51,8 @@ def compile_accelerator(src, metadata, o_path):
     if cache_path is None:
         with tempfile.TemporaryDirectory() as tmpdir:
             dst_path = os.path.join(tmpdir, f"{name}.so")
-            xuantie_dir=txda_tools.get_tx8_deps_path("rcs1fw-rtt/tool/rcsfw-xuantie-sdk/Xuantie-900-gcc-elf-newlib-x86_64-V2.8.0")
+            xuantie_dir = txda_tools.get_tx8_deps_path(
+                "rcs1fw-rtt/tool/rcsfw-xuantie-sdk/Xuantie-900-gcc-elf-newlib-x86_64-V2.8.0")
             gcc_path = os.path.join(xuantie_dir, "bin", "riscv64-unknown-elf-gcc")
             libc_lib = os.path.join(xuantie_dir, "riscv64-unknown-elf", "lib", "rv64imfdc", "lp64d")
             libgcc_lib = os.path.join(xuantie_dir, "lib", "gcc", "riscv64-unknown-elf", "10.4.0", "rv64imfdc", "lp64d")
@@ -118,20 +120,11 @@ def _ttir_to_coreir(mod, num_stages=2):
         pipeline_flag = f"--mk-pipeline=num-stages={num_stages}"
 
         args = [
-            triton_opt_path,
-            src_path,
-            "--triton-to-core-dialects",
-            "--tle-to-mk",
-            "--dsa-memory-to-core",
-            "--linalg-tiling",
-            f"{coreir_to_mk_mode}",
-            "--linalg-fusion",
-            "--legalize-tensor-form-loops",
-            "--one-shot-bufferize",
-            "--convert-bufferization-to-memref",
-            "--materialize-strided-linalg-inputs",
-            "--cse",
-            "--canonicalize"]
+            triton_opt_path, src_path, "--triton-to-core-dialects", "--tle-to-mk", "--dsa-memory-to-core",
+            "--linalg-tiling", f"{coreir_to_mk_mode}", "--linalg-fusion", "--legalize-tensor-form-loops",
+            "--one-shot-bufferize", "--convert-bufferization-to-memref", "--materialize-strided-linalg-inputs", "--cse",
+            "--canonicalize"
+        ]
 
         if os.getenv("TRITON_PIPELINE", "1") == "1":
             args.append(pipeline_flag)
@@ -191,7 +184,9 @@ def _coreir_to_txir(mod):
         triton_opt_path = txda_tools.get_tsm_opt_path()
         txda_tools.dump_ir_if_needed([src_path])
 
-        args = [triton_opt_path, src_path,
+        args = [
+            triton_opt_path,
+            src_path,
             "--spmd-allocate-shared-memory",
             "--expand-strided-metadata",
             "--lower-affine",  # convert affine.load to memref.load, need exec before tx81-to-llvm since we will support spm offset to memref.load
@@ -231,10 +226,9 @@ def _txir_to_llir(mod, metadata):
         args = [
             triton_opt_path, src_path,
             # Use tx81-memref-to-llvm to replace "--finalize-memref-to-llvm".
-            "--tx81-memref-to-llvm", "--addr-to-llvm", "--convert-scf-to-cf",
-            "--expand-strided-metadata",
-            "--convert-math-to-llvm",
-            "--convert-math-to-libm", "--convert-cf-to-llvm",  # need exec before "convert-func-to-llvm"
+            "--tx81-memref-to-llvm", "--addr-to-llvm", "--convert-scf-to-cf", "--expand-strided-metadata",
+            "--convert-math-to-llvm", "--convert-math-to-libm",
+            "--convert-cf-to-llvm",  # need exec before "convert-func-to-llvm"
             "--convert-func-to-llvm",  # need exec before "kernel-arg-buffer", otherwise un-rank memref will translate to int(rank) + ptr
             # Other unconverted memref ops, eg: memref.global from scan op conversion
             "--finalize-memref-to-llvm"
@@ -257,8 +251,7 @@ def _txir_to_llir(mod, metadata):
             tx81_to_llvm = "--tx81-to-llvm=gather-scatter-async=true"
 
         args += [
-            tx81_to_llvm,
-            "--convert-arith-to-llvm",  # need exec last since arith.const conversion
+            tx81_to_llvm, "--convert-arith-to-llvm",  # need exec last since arith.const conversion
             # Remove all unrealized casts created
             "--reconcile-unrealized-casts", "--canonicalize", "--export-kernel-symbols", "-o", llvmir_path
         ]
@@ -434,6 +427,7 @@ def hash(self):
 
 
 class TXDABackend(BaseBackend):
+
     @staticmethod
     def supports_target(target: GPUTarget):
         return target.backend == 'txda'

third_party/tsingmicro/backend/driver.py

@@ -1187,6 +1187,7 @@ def get_device_properties(self, *args):
 _last_launch_res = None
 _launch_counter = 0
 
+
 # Launch cross compiled runtime program on controller
 class TXDALauncher(object):
 
@@ -1211,8 +1212,7 @@ def __call__(self, gridX, gridY, gridZ, stream, function, *args, **kwargs):
         logger.info(f"{self.func_name} launch card:{device_id} count:{_launch_counter} begin")
         launchRes = self.launch(device_id, self.metadata.so_key, self.metadata.kernel_path, self.func_name,
                                 txda_tools.is_dump_args_profile(), txda_tools.get_dump_dir(), log_level,
-                                _launch_counter,
-                                gridX, gridY, gridZ, stream, function, *args, **kwargs)
+                                _launch_counter, gridX, gridY, gridZ, stream, function, *args, **kwargs)
         logger.info(f"{self.func_name} launch card:{device_id} count:{_launch_counter} end")
         if launchRes.res != 0:
             logger.error(f"launch error code:{launchRes.res}")

third_party/tsingmicro/backend/include/tsm_prof_query.h

@@ -1,8 +1,9 @@
 // Copyright (c) 2026 Tsingmicro CORPORATION
 //
-// Query device kernel timing (AP / kcore) after txLaunch* + txStreamSynchronize.
-// Requires libtsm-api-log-tracing.so listed in ROCP_TOOL_LIBRARIES.
-// Resolve symbols via dlopen(ROCP_TOOL_LIBRARIES) + dlsym — not dlsym(RTLD_DEFAULT).
+// Query device kernel timing (AP / kcore) after txLaunch* +
+// txStreamSynchronize. Requires libtsm-api-log-tracing.so listed in
+// ROCP_TOOL_LIBRARIES. Resolve symbols via dlopen(ROCP_TOOL_LIBRARIES) + dlsym
+// — not dlsym(RTLD_DEFAULT).
 //
 #pragma once
 
@@ -12,7 +13,7 @@
 extern "C" {
 #endif
 
-#define TSM_PROF_API_OP_MAX   64
+#define TSM_PROF_API_OP_MAX 64
 #define TSM_PROF_API_ARGS_MAX 512
 
 typedef struct tsm_prof_timing_t {
@@ -25,23 +26,26 @@ typedef struct tsm_prof_timing_t {
   uint64_t ap_end;
   uint64_t kcore_start;
   uint64_t kcore_end;
-  char     api_op[TSM_PROF_API_OP_MAX];
-  char     api_args[TSM_PROF_API_ARGS_MAX];
-  char     api_result[64];
-  int      dev_id;
-  int      ready;
+  char api_op[TSM_PROF_API_OP_MAX];
+  char api_args[TSM_PROF_API_ARGS_MAX];
+  char api_result[64];
+  int dev_id;
+  int ready;
 } tsm_prof_timing_t;
 
-/* Start a profiling scope; call immediately before the target traced API. Returns 0 on failure. */
+/* Start a profiling scope; call immediately before the target traced API.
+ * Returns 0 on failure. */
 uint64_t tsm_prof_begin(void);
 
 /*
- * Query device timing for token. Call after txStreamSynchronize (same thread as tsm_prof_begin).
- * The next traced TSM_RUNTIME_API EXIT after begin binds token to that API's corid.
- * Returns 0 on success, -1 invalid args/unknown token, -2 corid not bound yet, -3 timeout.
- * timeout_ms: 0 = poll once, >0 = wait up to N ms for kernel notify.
+ * Query device timing for token. Call after txStreamSynchronize (same thread as
+ * tsm_prof_begin). The next traced TSM_RUNTIME_API EXIT after begin binds token
+ * to that API's corid. Returns 0 on success, -1 invalid args/unknown token, -2
+ * corid not bound yet, -3 timeout. timeout_ms: 0 = poll once, >0 = wait up to N
+ * ms for kernel notify.
  */
-int tsm_prof_get_timing(uint64_t token, tsm_prof_timing_t* out, uint32_t timeout_ms);
+int tsm_prof_get_timing(uint64_t token, tsm_prof_timing_t *out,
+                        uint32_t timeout_ms);
 
 #ifdef __cplusplus
 }

third_party/tsingmicro/backend/include/tsm_profiler.h

@@ -14,40 +14,42 @@
 #pragma once
 
 #include "tsm_prof_query.h"
-#include <dlfcn.h>
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
+#include <dlfcn.h>
 #include <string>
 
-static void* g_prof_tool_handle = nullptr;
+static void *g_prof_tool_handle = nullptr;
 static uint64_t (*resolve_tsm_prof_begin)(void) = nullptr;
-static int (*resolve_tsm_prof_get_timing)(uint64_t, tsm_prof_timing_t*, uint32_t) = nullptr;
+static int (*resolve_tsm_prof_get_timing)(uint64_t, tsm_prof_timing_t *,
+                                          uint32_t) = nullptr;
 
-inline void* open_prof_tool_from_env() {
+inline void *open_prof_tool_from_env() {
   if (g_prof_tool_handle)
     return g_prof_tool_handle;
 
-  const char* env = std::getenv("ROCP_TOOL_LIBRARIES");
+  const char *env = std::getenv("ROCP_TOOL_LIBRARIES");
   if (!env || env[0] == '\0')
     return nullptr;
 
   std::string paths{env};
-  size_t      start = 0;
+  size_t start = 0;
   while (start < paths.size()) {
     const size_t end = paths.find(':', start);
-    const auto   path =
-        (end == std::string::npos) ? paths.substr(start) : paths.substr(start, end - start);
+    const auto path = (end == std::string::npos)
+                          ? paths.substr(start)
+                          : paths.substr(start, end - start);
     start = (end == std::string::npos) ? paths.size() : end + 1;
 
     if (path.empty())
       continue;
 
-    void* handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_NOLOAD);
+    void *handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_NOLOAD);
     if (!handle)
       handle = dlopen(path.c_str(), RTLD_LAZY);
     if (handle) {
-      void* sym = dlsym(handle, "tsm_prof_begin");
+      void *sym = dlsym(handle, "tsm_prof_begin");
       if (sym != nullptr) {
         g_prof_tool_handle = handle;
         return handle;
@@ -58,14 +60,16 @@ inline void* open_prof_tool_from_env() {
 }
 
 inline bool resolve_prof_query_api() {
-  void* handle = open_prof_tool_from_env();
+  void *handle = open_prof_tool_from_env();
   if (!handle)
     return false;
 
   if (!resolve_tsm_prof_begin)
-    resolve_tsm_prof_begin = reinterpret_cast<uint64_t (*)(void)>(dlsym(handle, "tsm_prof_begin"));
+    resolve_tsm_prof_begin =
+        reinterpret_cast<uint64_t (*)(void)>(dlsym(handle, "tsm_prof_begin"));
   if (!resolve_tsm_prof_get_timing)
     resolve_tsm_prof_get_timing =
-        reinterpret_cast<int (*)(uint64_t, tsm_prof_timing_t*, uint32_t)>(dlsym(handle, "tsm_prof_get_timing"));
+        reinterpret_cast<int (*)(uint64_t, tsm_prof_timing_t *, uint32_t)>(
+            dlsym(handle, "tsm_prof_get_timing"));
   return resolve_tsm_prof_begin && resolve_tsm_prof_get_timing;
 }