================ @@ -144,6 +144,92 @@ double3 cos(double3); _HLSL_BUILTIN_ALIAS(__builtin_elementwise_cos) double4 cos(double4); +//===----------------------------------------------------------------------===// +// dot product builtins +//===----------------------------------------------------------------------===// +#ifdef __HLSL_ENABLE_16_BIT +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +half dot(half, half); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +half dot(half2, half2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +half dot(half3, half3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +half dot(half4, half4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int16_t dot(int16_t, int16_t); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int16_t dot(int16_t2, int16_t2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int16_t dot(int16_t3, int16_t3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int16_t dot(int16_t4, int16_t4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint16_t dot(uint16_t, uint16_t); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint16_t dot(uint16_t2, uint16_t2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint16_t dot(uint16_t3, uint16_t3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint16_t dot(uint16_t4, uint16_t4); +#endif + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +float dot(float, float); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +float dot(float2, float2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +float dot(float3, float3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +float dot(float4, float4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +double dot(double, double); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +double dot(double2, double2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +double dot(double3, double3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +double dot(double4, double4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int dot(int, int); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int dot(int2, int2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int dot(int3, int3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int dot(int4, int4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint dot(uint, uint); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint dot(uint2, uint2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint dot(uint3, uint3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint dot(uint4, uint4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int64_t dot(int64_t, int64_t); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int64_t dot(int64_t2, int64_t2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int64_t dot(int64_t3, int64_t3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +int64_t dot(int64_t4, int64_t4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint64_t dot(uint64_t, uint64_t); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint64_t dot(uint64_t2, uint64_t2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint64_t dot(uint64_t3, uint64_t3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_dot) +uint64_t dot(uint64_t4, uint64_t4); ---------------- dmpots wrote:
I am still not clear on exactly which overloads you are proposing that we support. > my preference is to not have overloads that don't lower to DXIL ops So that would mean no scalar overloads at all (since those lower to multiply) and no double overloads. But keep support for i64 and u64 vector overloads? As for silently generating less efficient code I am somewhat skeptical that there is GPU hardware that will accelerate 64-bit dot operations. My guess is those all get expanded by the target anyway. https://github.com/llvm/llvm-project/pull/81190 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
