diff options
Diffstat (limited to 'modules/remote-sycl/benchmarks/kernel_mixed_precision.cpp')
-rw-r--r-- | modules/remote-sycl/benchmarks/kernel_mixed_precision.cpp | 13 |
1 files changed, 7 insertions, 6 deletions
diff --git a/modules/remote-sycl/benchmarks/kernel_mixed_precision.cpp b/modules/remote-sycl/benchmarks/kernel_mixed_precision.cpp index 0ac9756..c17c137 100644 --- a/modules/remote-sycl/benchmarks/kernel_mixed_precision.cpp +++ b/modules/remote-sycl/benchmarks/kernel_mixed_precision.cpp @@ -6,33 +6,34 @@ saw::interface<sch::MixedPrecisionBenchmarkInterface, saw::encode::Native, saw:: * Mixed */ [&](saw::data<sch::MixedArray, saw::encode::Native, saw::rmt::Sycl>& in, cl::sycl::queue* cmd) -> saw::error_or<void> { - + uint64_t in_size = in.size(); + mixed_ev = cmd->submit([&](cl::sycl::handler& h){ auto acc_buff = in.template access<cl::sycl::access::mode::read_write>(h); - h.parallel_for(cl::sycl::range<1>(in.size()), [=] (cl::sycl::id<1> it){ + h.parallel_for(cl::sycl::range<1>(in_size), [=] (cl::sycl::id<1> it){ acc_buff[0u].at(it[0u]) = acc_buff[0u].at(it[0u]) * saw::data<sch::Float64>{2.0}; }); }); return saw::void_t{}; }, [&](saw::data<sch::Float64Array, saw::encode::Native, saw::rmt::Sycl>& in, cl::sycl::queue* cmd) -> saw::error_or<void> { - + uint64_t in_size = in.size(); float64_ev = cmd->submit([&](cl::sycl::handler& h){ auto acc_buff = in.template access<cl::sycl::access::mode::read_write>(h); - h.parallel_for(cl::sycl::range<1>(in.size()), [=] (cl::sycl::id<1> it){ + h.parallel_for(cl::sycl::range<1>(in_size), [=] (cl::sycl::id<1> it){ acc_buff[0u].at(it[0u]) = acc_buff[0u].at(it[0u]) * saw::data<sch::Float64>{2.0}; }); }); return saw::void_t{}; }, [&](saw::data<sch::Float32Array, saw::encode::Native, saw::rmt::Sycl>& in, cl::sycl::queue* cmd) -> saw::error_or<void> { - + uint64_t in_size = in.size(); float32_ev = cmd->submit([&](cl::sycl::handler& h){ auto acc_buff = in.template access<cl::sycl::access::mode::read_write>(h); - h.parallel_for(cl::sycl::range<1>(in.size()), [=] (cl::sycl::id<1> it){ + h.parallel_for(cl::sycl::range<1>(in_size), [=] (cl::sycl::id<1> it){ acc_buff[0u].at(it[0u]) = acc_buff[0u].at(it[0u]) * saw::data<sch::Float32>{2.0f}; }); }); |