Integrate bitpack mappings into LLAMA

CMakeLists.txt

@@ -71,9 +71,9 @@ if (LLAMA_BUILD_EXAMPLES)
 	add_subdirectory("examples/viewcopy")
 	add_subdirectory("examples/bufferguard")
 	add_subdirectory("examples/raycast")
-	add_subdirectory("examples/bitpack")
 	add_subdirectory("examples/bytesplit")
-	add_subdirectory("examples/floatpack")
+	add_subdirectory("examples/bitpackint")
+	add_subdirectory("examples/bitpackfloat")
 
 	# alpaka examples
 	find_package(alpaka 0.7.0 QUIET)

examples/floatpack/CMakeLists.txt → examples/bitpackfloat/CMakeLists.txt

@@ -1,9 +1,9 @@
 cmake_minimum_required (VERSION 3.15)
-project(llama-floatpack CXX)
+project(llama-bitpackfloat CXX)
 
 if (NOT TARGET llama::llama)
 	find_package(llama REQUIRED)
 endif()
-add_executable(${PROJECT_NAME} floatpack.cpp)
+add_executable(${PROJECT_NAME} bitpackfloat.cpp)
 target_compile_features(${PROJECT_NAME} PRIVATE cxx_std_17)
 target_link_libraries(${PROJECT_NAME} PRIVATE llama::llama)

examples/bitpackfloat/bitpackfloat.cpp

@@ -0,0 +1,54 @@
+#include <fmt/core.h>
+#include <llama/llama.hpp>
+#include <random>
+
+// clang-format off
+namespace tag
+{
+    struct X{};
+    struct Y{};
+} // namespace tag
+
+using Vector = llama::Record<
+    llama::Field<tag::X, double>,
+    llama::Field<tag::Y, double>
+>;
+// clang-format on
+
+auto main() -> int
+{
+    constexpr auto N = 100;
+    constexpr auto exponentBits = 5;
+    constexpr auto mantissaBits = 13;
+    const auto mapping
+        = llama::mapping::BitPackedFloatSoA{exponentBits, mantissaBits, llama::ArrayExtents<llama::dyn>{N}, Vector{}};
+
+    auto view = llama::allocView(mapping);
+
+    boost::mp11::mp_for_each<boost::mp11::mp_iota_c<decltype(mapping)::blobCount>>(
+        [&](auto ic)
+        {
+            fmt::print(
+                "Blob {}: {} bytes (uncompressed {} bytes)\n",
+                ic,
+                mapping.blobSize(ic),
+                N * sizeof(llama::GetType<Vector, llama::RecordCoord<decltype(ic)::value>>));
+        });
+
+    std::default_random_engine engine;
+    std::uniform_real_distribution dist{0.0f, 100.0f};
+
+    // view(0)(tag::X{}) = -123.456789f;
+    // float f = view(0)(tag::X{});
+    // fmt::print("{}", f);
+
+    for(std::size_t i = 0; i < N; i++)
+    {
+        const auto v = dist(engine);
+        view(i)(tag::X{}) = v;
+        view(i)(tag::Y{}) = -v;
+
+        fmt::print("{:11} -> {:11}\n", v, static_cast<float>(view(i)(tag::X{})));
+        fmt::print("{:11} -> {:11}\n", -v, static_cast<float>(view(i)(tag::Y{})));
+    }
+}

examples/bitpack/CMakeLists.txt → examples/bitpackint/CMakeLists.txt

@@ -1,9 +1,9 @@
 cmake_minimum_required (VERSION 3.15)
-project(llama-bitpack CXX)
+project(llama-bitpackint CXX)
 
 if (NOT TARGET llama::llama)
 	find_package(llama REQUIRED)
 endif()
-add_executable(${PROJECT_NAME} bitpack.cpp)
+add_executable(${PROJECT_NAME} bitpackint.cpp)
 target_compile_features(${PROJECT_NAME} PRIVATE cxx_std_17)
 target_link_libraries(${PROJECT_NAME} PRIVATE llama::llama)

examples/bitpackint/bitpackint.cpp

@@ -0,0 +1,66 @@
+#include <cstdint>
+#include <fmt/core.h>
+#include <llama/llama.hpp>
+
+// clang-format off
+namespace tag
+{
+    struct X{};
+    struct Y{};
+    struct Z{};
+} // namespace tag
+
+using Vector = llama::Record<
+    llama::Field<tag::X, std::uint16_t>,
+    llama::Field<tag::Y, std::int32_t>,
+    llama::Field<tag::Z, std::uint64_t>
+>;
+// clang-format on
+
+auto main() -> int
+{
+    constexpr auto N = 128;
+    constexpr auto bits = 7;
+    const auto mapping = llama::mapping::BitPackedIntSoA<llama::ArrayExtentsDynamic<1>, Vector>{bits, {N}};
+
+    auto view = llama::allocView(mapping);
+
+    boost::mp11::mp_for_each<boost::mp11::mp_iota_c<decltype(mapping)::blobCount>>(
+        [&](auto ic)
+        {
+            fmt::print(
+                "Blob {}: {} bytes (uncompressed {} bytes)\n",
+                ic,
+                mapping.blobSize(ic),
+                N * sizeof(llama::GetType<Vector, llama::RecordCoord<decltype(ic)::value>>));
+        });
+
+    for(std::size_t i = 0; i < N; i++)
+    {
+        view(i)(tag::X{}) = i;
+        view(i)(tag::Y{}) = -static_cast<std::int32_t>(i); // cut-off of sign bits after -64
+        view(i)(tag::Z{}) = i * 2; // exceeds bits
+    }
+
+    fmt::print("Bitpacked initial:\n");
+    for(std::size_t i = 0; i < N; i++)
+        fmt::print("[{}, {}, {}]\n", view(i)(tag::X{}), view(i)(tag::Y{}), view(i)(tag::Z{}));
+
+    // extract into a view of full size integers
+    auto viewExtracted
+        = llama::allocViewUninitialized(llama::mapping::AoS<llama::ArrayExtents<llama::dyn>, Vector>{{N}});
+    llama::copy(view, viewExtracted);
+    if(!std::equal(view.begin(), view.end(), viewExtracted.begin(), viewExtracted.end()))
+        fmt::print("ERROR: unpacked view is different\n");
+
+    // compute something on the extracted view
+    for(std::size_t i = 0; i < N; i++)
+        viewExtracted(i) = viewExtracted(i) % 10;
+
+    // compress back
+    llama::copy(viewExtracted, view);
+
+    fmt::print("Bitpacked after % 10:\n");
+    for(std::size_t i = 0; i < N; i++)
+        fmt::print("[{}, {}, {}]\n", view(i)(tag::X{}), view(i)(tag::Y{}), view(i)(tag::Z{}));
+}