| 27 | namespace cl { |
| 28 | |
| 29 | std::string GetCommonDefines(CalculationsPrecision precision) { |
| 30 | std::string result; |
| 31 | |
| 32 | switch (precision) { |
| 33 | case CalculationsPrecision::F32: |
| 34 | result += "#define ACCUM_FLT4 float4\n"; |
| 35 | result += "#define FLT float\n"; |
| 36 | result += "#define FLT2 float2\n"; |
| 37 | result += "#define FLT3 float3\n"; |
| 38 | result += "#define FLT4 float4\n"; |
| 39 | result += "#define TO_FLT4 convert_float4\n"; |
| 40 | result += "#define TO_ACCUM_TYPE convert_float4\n"; |
| 41 | result += "#define TO_ACCUM_FLT convert_float\n"; |
| 42 | result += "#define READ_IMAGE read_imagef\n"; |
| 43 | result += "#define WRITE_IMAGE write_imagef\n"; |
| 44 | break; |
| 45 | case CalculationsPrecision::F16: |
| 46 | result += "#pragma OPENCL EXTENSION cl_khr_fp16 : enable\n"; |
| 47 | result += "#define ACCUM_FLT4 half4\n"; |
| 48 | result += "#define FLT half\n"; |
| 49 | result += "#define FLT2 half2\n"; |
| 50 | result += "#define FLT3 half3\n"; |
| 51 | result += "#define FLT4 half4\n"; |
| 52 | result += "#define TO_FLT4 convert_half4\n"; |
| 53 | result += "#define TO_ACCUM_TYPE convert_half4\n"; |
| 54 | result += "#define TO_ACCUM_FLT convert_half\n"; |
| 55 | result += "#define READ_IMAGE read_imageh\n"; |
| 56 | result += "#define WRITE_IMAGE write_imageh\n"; |
| 57 | break; |
| 58 | case CalculationsPrecision::F32_F16: |
| 59 | result += "#pragma OPENCL EXTENSION cl_khr_fp16 : enable\n"; |
| 60 | result += "#define ACCUM_FLT4 float4\n"; |
| 61 | result += "#define FLT half\n"; |
| 62 | result += "#define FLT2 half2\n"; |
| 63 | result += "#define FLT3 half3\n"; |
| 64 | result += "#define FLT4 half4\n"; |
| 65 | result += "#define TO_FLT4 convert_half4\n"; |
| 66 | result += "#define TO_ACCUM_TYPE convert_float4\n"; |
| 67 | result += "#define TO_ACCUM_FLT convert_float\n"; |
| 68 | result += "#define READ_IMAGE read_imageh\n"; |
| 69 | result += "#define WRITE_IMAGE write_imageh\n"; |
| 70 | break; |
| 71 | } |
| 72 | |
| 73 | result += |
| 74 | "const sampler_t smp_edge = CLK_NORMALIZED_COORDS_FALSE | " |
| 75 | "CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;\n"; |
| 76 | result += |
| 77 | "const sampler_t smp_none = CLK_NORMALIZED_COORDS_FALSE | " |
| 78 | "CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;\n"; |
| 79 | result += |
| 80 | "const sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | " |
| 81 | "CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;\n"; |
| 82 | |
| 83 | return result; |
| 84 | } |
| 85 | |
| 86 | std::string GetGlobalAddressNoDeclaration(TensorStorageType storage_type, |
no outgoing calls
no test coverage detected