break;
case CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF:
- buf.as_scalar<cl_uint>() = 0;
+ buf.as_scalar<cl_uint>() = dev.has_halves() ? 8 : 0;
break;
case CL_DEVICE_MAX_CLOCK_FREQUENCY:
buf.as_scalar<cl_uint>() = 128;
break;
+ case CL_DEVICE_HALF_FP_CONFIG:
+ // This is the "mandated minimum half precision floating-point
+ // capability" for OpenCL 1.x.
+ buf.as_scalar<cl_device_fp_config>() =
+ CL_FP_INF_NAN | CL_FP_ROUND_TO_NEAREST;
+ break;
+
case CL_DEVICE_SINGLE_FP_CONFIG:
// This is the "mandated minimum single precision floating-point
// capability" for OpenCL 1.1. In OpenCL 1.2, nothing is required for
" cl_khr_local_int32_base_atomics"
" cl_khr_local_int32_extended_atomics"
" cl_khr_byte_addressable_store"
- + std::string(dev.has_doubles() ? " cl_khr_fp64" : "");
+ + std::string(dev.has_doubles() ? " cl_khr_fp64" : "")
+ + std::string(dev.has_halves() ? " cl_khr_fp16" : "");
break;
case CL_DEVICE_PLATFORM:
break;
case CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF:
- buf.as_scalar<cl_uint>() = 0;
+ buf.as_scalar<cl_uint>() = dev.has_halves() ? 8 : 0;
break;
case CL_DEVICE_OPENCL_C_VERSION:
return pipe->get_param(pipe, PIPE_CAP_DOUBLES);
}
+bool
+device::has_halves() const {
+ return pipe->get_shader_param(pipe, PIPE_SHADER_COMPUTE,
+ PIPE_SHADER_CAP_FP16);
+}
+
bool
device::has_unified_memory() const {
return pipe->get_param(pipe, PIPE_CAP_UMA);
cl_uint max_compute_units() const;
bool image_support() const;
bool has_doubles() const;
+ bool has_halves() const;
bool has_unified_memory() const;
cl_uint mem_base_addr_align() const;