fix codegen bug

paddle/cinn/backends/compiler.cc

@@ -37,6 +37,7 @@ PD_DECLARE_string(cinn_dump_group_lowered_func);
 PD_DECLARE_string(cinn_dump_group_source_code);
 PD_DECLARE_string(cinn_dump_group_ptx);
 PD_DECLARE_string(cinn_dump_group_instruction);
+PD_DECLARE_string(cinn_custom_code_path);
 
 namespace cinn {
 namespace backends {
@@ -267,6 +268,22 @@ void Compiler::BuildDefault(const Module& module) {
   });
 }
 
+std::string getFileContent(const std::string& filePath) {
+  std::ifstream file(filePath);
+
+  if (!file.is_open()) {
+    std::cerr << "Unable to open file: " << filePath << std::endl;
+    return "";
+  }
+
+  std::ostringstream contentStream;
+  contentStream << file.rdbuf();
+  std::string content = contentStream.str();
+
+  file.close();
+  return content;
+}
+
 void Compiler::CompileCudaModule(const Module& module,
                                  const std::string& code) {
 #ifdef CINN_WITH_CUDA
@@ -278,12 +295,17 @@ void Compiler::CompileCudaModule(const Module& module,
 
   VLOG(3) << "[CUDA] device module:\n" << device_module;
   std::string source_code;
-  if (code.empty()) {
+
+  if (!FLAGS_cinn_custom_code_path.empty()) {
+    std::string filePath = FLAGS_cinn_custom_code_path;
+    source_code = getFileContent(filePath);
+  } else if (code.empty()) {
     CodeGenCUDA_Dev codegen(target_);
     source_code = codegen.Compile(device_module);
   } else {
     source_code = code;
   }
+
   CHECK(!source_code.empty())
       << "Compile CUDA C code failed from device module:\n"
       << device_module;

paddle/cinn/ir/ir_printer.cc

@@ -94,10 +94,25 @@ void IrPrinter::Visit(const UIntImm *x) {
 namespace {
 template <typename T>
 bool isCloseEqualMaxValue(T value) {
-  T max_value = std::numeric_limits<T>::max();
+  T maxValue = std::numeric_limits<T>::max();
+  T minValue = std::numeric_limits<T>::lowest();
   T tol = std::numeric_limits<T>::denorm_min();
-  return (max_value - value) < tol;
+  return (maxValue - value) < tol || (value - minValue) < tol;
 }
+
+template <typename T>
+T truncateInfinity(T value) {
+  T maxValue = std::numeric_limits<T>::max();
+  T minValue = std::numeric_limits<T>::lowest();
+  if (value > maxValue) {
+    return maxValue;
+  }
+  if (value < minValue) {
+    return minValue;
+  }
+  return value;
+}
+
 }  // namespace
 
 void IrPrinter::Visit(const FloatImm *x) {
@@ -123,11 +138,12 @@ void IrPrinter::Visit(const FloatImm *x) {
       ss << static_cast<bfloat16>(x->value) << "f";
     }
   } else if (x->type().is_float(32)) {
-    if (isCloseEqualMaxValue<float>(x->value)) std::fesetround(FE_TOWARDZERO);
+    float v = truncateInfinity<float>(x->value);
+    if (isCloseEqualMaxValue<float>(v)) std::fesetround(FE_TOWARDZERO);
     ss << std::setprecision(std::numeric_limits<float>::max_digits10);
     ss << std::showpoint;
-    ss << x->value;
-    if (std::isfinite(x->value)) {
+    ss << v;
+    if (std::isfinite(v)) {
       ss << "f";
     }
   } else if (x->type().is_float(64)) {

paddle/cinn/runtime/flags.cc

@@ -207,6 +207,10 @@ PD_DEFINE_string(
     StringFromEnv("FLAGS_cinn_dump_group_instruction", ""),
     "Specify the path for dump instruction by group, which is used for debug.");
 
+PD_DEFINE_string(cinn_custom_code_path,
+                 StringFromEnv("FLAGS_cinn_custom_code_path", ""),
+                 "Specify custom code path for cinn.");
+
 PD_DEFINE_string(cinn_pass_visualize_dir,
                  StringFromEnv("FLAGS_cinn_pass_visualize_dir", ""),
                  "Specify the directory path of pass visualize file of graph, "

paddle/fluid/primitive/composite/composite.h

@@ -1039,20 +1039,27 @@ std::tuple<Tensor, Tensor> flatten_decomp(const Tensor& x,
 }
 
 template <typename T>
-Tensor clip_decomp(const Tensor& x, const Tensor& min_, const Tensor& max_) {
-  Tensor min_t = min_;
-  Tensor max_t = max_;
-  if (min_.dtype() == x.dtype()) {
-    min_t = cast<T>(min_, x.dtype());
+Tensor clip_decomp(const Tensor& x, const Tensor& min, const Tensor& max) {
+  auto min_reshape = min;
+  auto max_reshape = max;
+
+  if (has_dynamic_shape(x.shape())) {
+    min_reshape = backend::expand_with_tensor<T>(min, shape<T>(x));
+    max_reshape = backend::expand_with_tensor<T>(max, shape<T>(x));
+  } else {
+    min_reshape = expand<T>(min, x.shape());
+    max_reshape = expand<T>(max, x.shape());
   }
-  if (max_.dtype() == x.dtype()) {
-    max_t = cast<T>(max_, x.dtype());
+  if (min_reshape.dtype() != x.dtype()) {
+    min_reshape = cast<T>(min_reshape, x.dtype());
   }
-  if (x.size() == 0) {
-    min_t = reshape<T>(min_t, empty_shape);
-    max_t = reshape<T>(max_t, empty_shape);
+
+  if (max_reshape.dtype() != x.dtype()) {
+    max_reshape = cast<T>(max_reshape, x.dtype());
   }
-  return maximum<T>(minimum<T>(x, max_t), min_t);
+
+  auto ans = maximum<T>(minimum<T>(x, max_reshape), min_reshape);
+  return ans;
 }
 
 template <typename T>