diff --git a/docs/reference/tables/CUDA_Device_API_supported_by_HIP.md b/docs/reference/tables/CUDA_Device_API_supported_by_HIP.md index 0f871305..b1af586e 100644 --- a/docs/reference/tables/CUDA_Device_API_supported_by_HIP.md +++ b/docs/reference/tables/CUDA_Device_API_supported_by_HIP.md @@ -19,7 +19,7 @@ |`__ballot`| | | | |`__ballot`|1.6.0| | | | | |`__ballot_sync`|9.0| | | |`__ballot_sync`|6.2.0| | | | | |`__bfloat1622float2`|11.0| | | |`__bfloat1622float2`|5.7.0| | | | | -|`__bfloat162bfloat162`|11.0| | | |`__bfloat162bfloat162`|5.7.0| | | | | +|`__bfloat162bfloat162`|11.0| |12.2| |`__bfloat162bfloat162`|5.7.0| | | | | |`__bfloat162char_rz`|12.2| | | | | | | | | | |`__bfloat162float`|11.0| | | |`__bfloat162float`|5.7.0| | | | | |`__bfloat162int_rd`|11.0| | | | | | | | | | @@ -264,12 +264,12 @@ |`__hgtu`| | | | |`__hgtu`|1.9.0| | | | | |`__hgtu2`| | | | |`__hgtu2`|1.9.0| | | | | |`__hgtu2_mask`|12.0| | | | | | | | | | -|`__high2bfloat16`|11.0| | | |`__high2bfloat16`|5.7.0| | | | | +|`__high2bfloat16`|11.0| |12.2| |`__high2bfloat16`|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| | | | | -|`__highs2bfloat162`|11.0| | | |`__highs2bfloat162`|5.7.0| | | | | +|`__highs2bfloat162`|11.0| |12.2| |`__highs2bfloat162`|5.7.0| | | | | |`__highs2half2`| | | | |`__highs2half2`|1.6.0| | | | | |`__hiloint2double`| | | | |`__hiloint2double`|1.6.0| | | | | |`__hisinf`| | | | |`__hisinf`|1.6.0| | | | | @@ -367,7 +367,7 @@ |`__low2half`| | | | |`__low2half`|1.6.0| | | | | |`__low2half2`| | | | |`__low2half2`|1.6.0| | | | | |`__lowhigh2highlow`| | | | |`__lowhigh2highlow`|1.6.0| | | | | -|`__lows2bfloat162`|11.0| | | |`__lows2bfloat162`|5.7.0| | | | | +|`__lows2bfloat162`|11.0| |12.2| |`__lows2bfloat162`|5.7.0| | | | | |`__lows2half2`| | | | |`__lows2half2`|1.6.0| | | | | |`__match_all_sync`|9.0| | | |`__match_all_sync`|6.2.0| | | | | |`__match_any_sync`|9.0| | | |`__match_any_sync`|6.2.0| | | | | diff --git a/src/CUDA2HIP_Device_functions.cpp b/src/CUDA2HIP_Device_functions.cpp index 394bcea1..ac453c96 100644 --- a/src/CUDA2HIP_Device_functions.cpp +++ b/src/CUDA2HIP_Device_functions.cpp @@ -1541,6 +1541,10 @@ const std::map CUDA_DEVICE_FUNCTION_CHA {"__bfloat16_as_ushort", {CUDA_122}}, {"__short_as_bfloat16", {CUDA_122}}, {"__ushort_as_bfloat16", {CUDA_122}}, + {"__bfloat162bfloat162", {CUDA_122}}, + {"__lows2bfloat162", {CUDA_122}}, + {"__highs2bfloat162", {CUDA_122}}, + {"__high2bfloat16", {CUDA_122}}, }; const std::map CUDA_DEVICE_FUNCTION_API_SECTION_MAP { 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 index fb1b962b..eaa11326 100644 --- a/tests/unit_tests/synthetic/libraries/cudevice2hipdevice_before_11080_after_12011.cu +++ b/tests/unit_tests/synthetic/libraries/cudevice2hipdevice_before_11080_after_12011.cu @@ -103,12 +103,12 @@ int main() { // 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); + _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); + _bf16 = __ushort_as_bfloat16(ushi); #endif #endif