Merge pull request #1846 from emankov/HIPIFY

[HIPIFY][Device][#1796][tests][fix] Fix device test failures on CUDA 11.8.0, 12.0.x, and 12.1.x
ROCm · Jan 28, 2025 · ac6ce7e · ac6ce7e
2 parents cb2c328 + 27f88e1
commit ac6ce7e
Show file tree

Hide file tree

Showing 6 changed files with 89 additions and 65 deletions.
diff --git a/docs/reference/tables/CUDA_Device_API_supported_by_HIP.md b/docs/reference/tables/CUDA_Device_API_supported_by_HIP.md
@@ -47,8 +47,8 @@
 |`__bfloat162ushort_rn`|11.0| | | | | | | | | |
 |`__bfloat162ushort_ru`|11.0| | | | | | | | | |
 |`__bfloat162ushort_rz`|11.0| | | | | | | | | |
-|`__bfloat16_as_short`|11.0| | | |`__bfloat16_as_short`|5.7.0| | | | |
-|`__bfloat16_as_ushort`|11.0| | | |`__bfloat16_as_ushort`|5.7.0| | | | |
+|`__bfloat16_as_short`|11.0| |12.2| |`__bfloat16_as_short`|5.7.0| | | | |
+|`__bfloat16_as_ushort`|11.0| |12.2| |`__bfloat16_as_ushort`|5.7.0| | | | |
 |`__brev`| | | | |`__brev`|1.6.0| | | | |
 |`__brevll`| | | | |`__brevll`|1.6.0| | | | |
 |`__brkpt`| | | | | | | | | | |
@@ -224,7 +224,7 @@
 |`__half2ushort_rz`| | | | |`__half2ushort_rz`|1.6.0| | | | |
 |`__half_as_short`| | | | |`__half_as_short`|1.6.0| | | | |
 |`__half_as_ushort`| | | | |`__half_as_ushort`|1.6.0| | | | |
-|`__halves2bfloat162`|11.0| | | |`__halves2bfloat162`|5.7.0| | | | |
+|`__halves2bfloat162`|11.0| |12.2| |`__halves2bfloat162`|5.7.0| | | | |
 |`__halves2half2`| | | | |`__halves2half2`|1.6.0| | | | |
 |`__hbeq2`| | | | |`__hbeq2`|1.6.0| | | | |
 |`__hbequ2`| | | | |`__hbequ2`|1.9.0| | | | |
@@ -265,7 +265,7 @@
 |`__hgtu2`| | | | |`__hgtu2`|1.9.0| | | | |
 |`__hgtu2_mask`|12.0| | | | | | | | | |
 |`__high2bfloat16`|11.0| | | |`__high2bfloat16`|5.7.0| | | | |
-|`__high2bfloat162`|11.0| | | |`__high2bfloat162`|5.7.0| | | | |
+|`__high2bfloat162`|11.0| |12.2| |`__high2bfloat162`|5.7.0| | | | |
 |`__high2float`| | | | |`__high2float`|1.6.0| | | | |
 |`__high2half`| | | | |`__high2half`|1.6.0| | | | |
 |`__high2half2`| | | | |`__high2half2`|1.6.0| | | | |
@@ -288,11 +288,11 @@
 |`__hltu2`| | | | |`__hltu2`|1.9.0| | | | |
 |`__hltu2_mask`|12.0| | | | | | | | | |
 |`__hmax`|11.0| | | |`__hmax`|5.5.0| | | | |
-|`__hmax2`|11.0| | | |`__hmax2`|5.7.0| | | | |
+|`__hmax2`|11.0| |12.2| |`__hmax2`|5.7.0| | | | |
 |`__hmax2_nan`|11.0| | | | | | | | | |
 |`__hmax_nan`|11.0| | | |`__hmax_nan`|5.5.0| | | | |
 |`__hmin`|11.0| | | |`__hmin`|5.5.0| | | | |
-|`__hmin2`|11.0| | | |`__hmin2`|5.7.0| | | | |
+|`__hmin2`|11.0| |12.2| |`__hmin2`|5.7.0| | | | |
 |`__hmin2_nan`|11.0| | | | | | | | | |
 |`__hmin_nan`|11.0| | | |`__hmin_nan`|5.5.0| | | | |
 |`__hmul`| | | | |`__hmul`|1.6.0| | | | |
@@ -361,8 +361,8 @@
 |`__log2f`| | | | |`__log2f`|1.6.0| | | | |
 |`__logf`| | | | |`__logf`|1.6.0| | | | |
 |`__longlong_as_double`| | | | |`__longlong_as_double`|1.6.0| | | | |
-|`__low2bfloat16`|11.0| | | |`__low2bfloat16`|5.7.0| | | | |
-|`__low2bfloat162`|11.0| | | |`__low2bfloat162`|5.7.0| | | | |
+|`__low2bfloat16`|11.0| |12.2| |`__low2bfloat16`|5.7.0| | | | |
+|`__low2bfloat162`|11.0| |12.2| |`__low2bfloat162`|5.7.0| | | | |
 |`__low2float`| | | | |`__low2float`|1.6.0| | | | |
 |`__low2half`| | | | |`__low2half`|1.6.0| | | | |
 |`__low2half2`| | | | |`__low2half2`|1.6.0| | | | |
@@ -411,7 +411,7 @@
 |`__short2half_rn`| | | | |`__short2half_rn`|1.6.0| | | | |
 |`__short2half_ru`| | | | |`__short2half_ru`|1.6.0| | | | |
 |`__short2half_rz`| | | | |`__short2half_rz`|1.6.0| | | | |
-|`__short_as_bfloat16`|11.0| | | |`__short_as_bfloat16`|5.7.0| | | | |
+|`__short_as_bfloat16`|11.0| |12.2| |`__short_as_bfloat16`|5.7.0| | | | |
 |`__short_as_half`| | | | |`__short_as_half`|1.9.0| | | | |
 |`__signbit`| | | | | | | | | | |
 |`__signbitf`| | | | | | | | | | |
@@ -475,7 +475,7 @@
 |`__ushort2half_rn`| | | | |`__ushort2half_rn`|1.6.0| | | | |
 |`__ushort2half_ru`| | | | |`__ushort2half_ru`|1.6.0| | | | |
 |`__ushort2half_rz`| | | | |`__ushort2half_rz`|1.6.0| | | | |
-|`__ushort_as_bfloat16`|11.0| | | |`__ushort_as_bfloat16`|5.7.0| | | | |
+|`__ushort_as_bfloat16`|11.0| |12.2| |`__ushort_as_bfloat16`|5.7.0| | | | |
 |`__ushort_as_half`| | | | |`__ushort_as_half`|1.6.0| | | | |
 |`__vabs2`| | | | | | | | | | |
 |`__vabs4`| | | | | | | | | | |

diff --git a/src/CUDA2HIP.h b/src/CUDA2HIP.h
@@ -161,6 +161,7 @@ extern const std::map<llvm::StringRef, hipAPIversions> HIP_CAFFE2_TYPE_NAME_VER_
 extern const std::map<llvm::StringRef, hipAPIversions> HIP_CAFFE2_FUNCTION_VER_MAP;
 extern const std::map<llvm::StringRef, hipAPIversions> HIP_DEVICE_TYPE_NAME_VER_MAP;
 extern const std::map<llvm::StringRef, hipAPIversions> HIP_DEVICE_FUNCTION_VER_MAP;
+extern const std::map<llvm::StringRef, cudaAPIChangedVersions> CUDA_DEVICE_FUNCTION_CHANGED_VER_MAP;
 extern const std::map<llvm::StringRef, hipAPIversions> HIP_CUB_TYPE_NAME_VER_MAP;
 extern const std::map<llvm::StringRef, hipAPIversions> HIP_CUB_FUNCTION_VER_MAP;
 extern const std::map<llvm::StringRef, hipAPIversions> HIP_RTC_TYPE_NAME_VER_MAP;

diff --git a/src/CUDA2HIP_Device_functions.cpp b/src/CUDA2HIP_Device_functions.cpp
@@ -1530,6 +1530,19 @@ const std::map<llvm::StringRef, hipAPIversions> HIP_DEVICE_FUNCTION_VER_MAP {
   {"__float22bfloat162_rn",             {HIP_5070, HIP_0,    HIP_0   }},
 };
 
+const std::map<llvm::StringRef, cudaAPIChangedVersions> CUDA_DEVICE_FUNCTION_CHANGED_VER_MAP {
+  {"__hmax2",                           {CUDA_122}},
+  {"__hmin2",                           {CUDA_122}},
+  {"__low2bfloat16",                    {CUDA_122}},
+  {"__halves2bfloat162",                {CUDA_122}},
+  {"__low2bfloat162",                   {CUDA_122}},
+  {"__high2bfloat162",                  {CUDA_122}},
+  {"__bfloat16_as_short",               {CUDA_122}},
+  {"__bfloat16_as_ushort",              {CUDA_122}},
+  {"__short_as_bfloat16",               {CUDA_122}},
+  {"__ushort_as_bfloat16",              {CUDA_122}},
+};
+
 const std::map<unsigned int, llvm::StringRef> CUDA_DEVICE_FUNCTION_API_SECTION_MAP {
   {1, "Device Functions"},
   {2, "Device Types"},

diff --git a/src/CUDA2HIP_Doc.cpp b/src/CUDA2HIP_Doc.cpp
@@ -801,6 +801,7 @@ namespace doc {
       const typeMap &getTypes() const override { return CUDA_DEVICE_TYPE_NAME_MAP; }
       const versionMap &getFunctionVersions() const override { return CUDA_DEVICE_FUNCTION_VER_MAP; }
       const hipVersionMap &getHipFunctionVersions() const override { return HIP_DEVICE_FUNCTION_VER_MAP; }
+      const cudaChangedVersionMap &getCudaChangedFunctionVersions() const override { return CUDA_DEVICE_FUNCTION_CHANGED_VER_MAP; }
       const versionMap &getTypeVersions() const override { return CUDA_DEVICE_TYPE_NAME_VER_MAP; }
       const hipVersionMap &getHipTypeVersions() const override { return HIP_DEVICE_TYPE_NAME_VER_MAP; }
       const string &getName() const override { return sCUDEVICE; }

diff --git a/tests/unit_tests/synthetic/libraries/cudevice2hipdevice.cu b/tests/unit_tests/synthetic/libraries/cudevice2hipdevice.cu
@@ -69,56 +69,6 @@ int main() {
   // CHECK: f2 = __bfloat1622float2(bf162);
   f2 = __bfloat1622float2(bf162);
 
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __hmax2(const __nv_bfloat162 a, const __nv_bfloat162 b);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hmax2(const __hip_bfloat162 a, const __hip_bfloat162 b);
-  // CHECK: bf162 = __hmax2(bf162a, bf162b);
-  bf162 = __hmax2(bf162a, bf162b);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __hmin2(const __nv_bfloat162 a, const __nv_bfloat162 b);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hmin2(const __hip_bfloat162 a, const __hip_bfloat162 b);
-  // CHECK: bf162 = __hmin2(bf162a, bf162b);
-  bf162 = __hmin2(bf162a, bf162b);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat16 __low2bfloat16(const __nv_bfloat162 a);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __low2bfloat16(const __hip_bfloat162 a);
-  // CHECK: _bf16 = __low2bfloat16(bf162a);
-  _bf16 = __low2bfloat16(bf162a);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __halves2bfloat162(const __nv_bfloat16 a, const __nv_bfloat16 b);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __halves2bfloat162(const __hip_bfloat16 a, const __hip_bfloat16 b);
-  // CHECK: bf162 = __halves2bfloat162(bf16a, bf16b);
-  bf162 = __halves2bfloat162(bf16a, bf16b);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __low2bfloat162(const __nv_bfloat162 a);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __low2bfloat162(const __hip_bfloat162 a);
-  // CHECK: bf162 = __low2bfloat162(bf162a);
-  bf162 = __low2bfloat162(bf162a);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __high2bfloat162(const __nv_bfloat162 a);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __high2bfloat162(const __hip_bfloat162 a);
-  // CHECK: bf162 = __high2bfloat162(bf162a);
-  bf162 = __high2bfloat162(bf162a);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ short int __bfloat16_as_short(const __nv_bfloat16 h);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ short int __bfloat16_as_short(const __hip_bfloat16 h);
-  // CHECK: shi = __bfloat16_as_short(_bf16);
-  shi = __bfloat16_as_short(_bf16);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ unsigned short int __bfloat16_as_ushort(const __nv_bfloat16 h);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ unsigned short int __bfloat16_as_ushort(const __hip_bfloat16 h);
-  // CHECK: ushi = __bfloat16_as_ushort(_bf16);
-  ushi = __bfloat16_as_ushort(_bf16);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat16 __short_as_bfloat16(const short int i);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __short_as_bfloat16(const short int a);
-  // CHECK: _bf16 = __short_as_bfloat16(shi);
-  _bf16 = __short_as_bfloat16(shi);
-
-  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat16 __ushort_as_bfloat16(const unsigned short int i);
-  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __ushort_as_bfloat16(const unsigned short int a);
-  // CHECK: _bf16 = __ushort_as_bfloat16(ushi);
-  _bf16 = __ushort_as_bfloat16(ushi);
-
   // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __float22bfloat162_rn(const float2 a);
   // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __float22bfloat162_rn(const float2 a);
   // CHECK: bf162 = __float22bfloat162_rn(f2);

diff --git a/tests/unit_tests/synthetic/libraries/cudevice2hipdevice_before_11080_after_12011.cu b/tests/unit_tests/synthetic/libraries/cudevice2hipdevice_before_11080_after_12011.cu
@@ -15,6 +15,8 @@ int main() {
   double dx = 0.0f;
   float fa = 0.0f;
   float fx = 0.0f;
+  short int shi = 0;
+  unsigned short int ushi = 0;
   double2 d2 = { 0.0f, 0.0f };
   float2 f2 = { 0.0f, 0.0f };
   __half_raw hrx = { 0 };
@@ -23,18 +25,25 @@ int main() {
 #if CUDA_VERSION >= 11000
   // CHECK: __hip_bfloat16 bf16 = { 0 };
   __nv_bfloat16 bf16 = { 0 };
+  // CHECK: __hip_bfloat16 _bf16 = { 0.0f };
+  // CHECK-NEXT: __hip_bfloat16 bf16a = { 0.0f };
+  // CHECK-NEXT: __hip_bfloat16 bf16b = { 0.0f };
+  __nv_bfloat16 _bf16 = { 0.0f };
+  __nv_bfloat16 bf16a = { 0.0f };
+  __nv_bfloat16 bf16b = { 0.0f };
+
   // CHECK: __hip_bfloat162 bf162 = { 0, 0 };
   // CHECK-NEXT: __hip_bfloat162 bf162a = { 0, 0 };
   // CHECK-NEXT: __hip_bfloat162 bf162b = { 0, 0 };
   __nv_bfloat162 bf162 = { 0, 0 };
   __nv_bfloat162 bf162a = { 0, 0 };
   __nv_bfloat162 bf162b = { 0, 0 };
 
-#if CUDA_VERSION < 11080 || CUDA_VERSION >= 12000
+#if CUDA_VERSION < 11080 || CUDA_VERSION >= 12020
   // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __bfloat162bfloat162(const __nv_bfloat16 a);
   // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __bfloat162bfloat162(const __hip_bfloat16 a);
-  // CHECK: bf162 = __bfloat162bfloat162(bf16);
-  bf162 = __bfloat162bfloat162(bf16);
+  // CHECK: bf162 = __bfloat162bfloat162(_bf16);
+  bf162 = __bfloat162bfloat162(_bf16);
 
   // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __lows2bfloat162(const __nv_bfloat162 a, const __nv_bfloat162 b);
   // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __lows2bfloat162(const __hip_bfloat162 a, const __hip_bfloat162 b);
@@ -48,8 +57,58 @@ int main() {
 
   // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat16 __high2bfloat16(const __nv_bfloat162 a);
   // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __high2bfloat16(const __hip_bfloat162 a);
-  // CHECK: bf16 = __high2bfloat16(bf162a);
-  bf16 = __high2bfloat16(bf162a);
+  // CHECK: _bf16 = __high2bfloat16(bf162a);
+  _bf16 = __high2bfloat16(bf162a);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __hmax2(const __nv_bfloat162 a, const __nv_bfloat162 b);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hmax2(const __hip_bfloat162 a, const __hip_bfloat162 b);
+  // CHECK: bf162 = __hmax2(bf162a, bf162b);
+  bf162 = __hmax2(bf162a, bf162b);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __hmin2(const __nv_bfloat162 a, const __nv_bfloat162 b);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hmin2(const __hip_bfloat162 a, const __hip_bfloat162 b);
+  // CHECK: bf162 = __hmin2(bf162a, bf162b);
+  bf162 = __hmin2(bf162a, bf162b);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat16 __low2bfloat16(const __nv_bfloat162 a);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __low2bfloat16(const __hip_bfloat162 a);
+  // CHECK: _bf16 = __low2bfloat16(bf162a);
+  _bf16 = __low2bfloat16(bf162a);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __halves2bfloat162(const __nv_bfloat16 a, const __nv_bfloat16 b);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __halves2bfloat162(const __hip_bfloat16 a, const __hip_bfloat16 b);
+  // CHECK: bf162 = __halves2bfloat162(bf16a, bf16b);
+  bf162 = __halves2bfloat162(bf16a, bf16b);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __low2bfloat162(const __nv_bfloat162 a);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __low2bfloat162(const __hip_bfloat162 a);
+  // CHECK: bf162 = __low2bfloat162(bf162a);
+  bf162 = __low2bfloat162(bf162a);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat162 __high2bfloat162(const __nv_bfloat162 a);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __high2bfloat162(const __hip_bfloat162 a);
+  // CHECK: bf162 = __high2bfloat162(bf162a);
+  bf162 = __high2bfloat162(bf162a);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ short int __bfloat16_as_short(const __nv_bfloat16 h);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ short int __bfloat16_as_short(const __hip_bfloat16 h);
+  // CHECK: shi = __bfloat16_as_short(_bf16);
+  shi = __bfloat16_as_short(_bf16);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ unsigned short int __bfloat16_as_ushort(const __nv_bfloat16 h);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ unsigned short int __bfloat16_as_ushort(const __hip_bfloat16 h);
+  // CHECK: ushi = __bfloat16_as_ushort(_bf16);
+  ushi = __bfloat16_as_ushort(_bf16);
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat16 __short_as_bfloat16(const short int i);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __short_as_bfloat16(const short int a);
+  // CHECK: _bf16 = __short_as_bfloat16(shi);
+  _bf16 = __short_as_bfloat16(shi); 
+
+  // CUDA: __CUDA_HOSTDEVICE_BF16_DECL__ __nv_bfloat16 __ushort_as_bfloat16(const unsigned short int i);
+  // HIP: __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __ushort_as_bfloat16(const unsigned short int a);
+  // CHECK: _bf16 = __ushort_as_bfloat16(ushi);
+  _bf16 = __ushort_as_bfloat16(ushi); 
 #endif
 
 #endif