Add Simpatico as the compression subsystem

.codespell_words

@@ -3,4 +3,5 @@ ans
 foto
 fpr
 ist
+nd
 thirdparty

CMakeLists.txt

@@ -113,6 +113,12 @@ set(CUCASCADE_WARNINGS_AS_ERRORS
     CACHE BOOL "" FORCE)
 add_subdirectory(cucascade "${CMAKE_BINARY_DIR}/cucascade" EXCLUDE_FROM_ALL)
 
+# Simpatico GPU compression engine (Phase 1 integration) Tests disabled here —
+# run them via the standalone simpatico_codegen build. Benchmark harness
+# (compress_with_plan_benchmark) is built by default.
+add_subdirectory(src/compression/simpatico_codegen
+                 "${CMAKE_BINARY_DIR}/simpatico_codegen" EXCLUDE_FROM_ALL)
+
 if(VCPKG_BUILD AND TARGET CUDA::cudart_static)
   function(sirius_prefer_static_cudart target_name)
     foreach(_prop LINK_LIBRARIES INTERFACE_LINK_LIBRARIES)

pixi.toml

@@ -32,6 +32,7 @@ pkg-config = "*"
 mold = "*"
 libprotobuf = "*"
 rust = "*"
+cuda-nvrtc-dev = ">=12.9.86,<14"
 
 [feature.dev-libs.dependencies]
 librmm = "*"

src/compression/simpatico_codegen/.gitignore

@@ -0,0 +1,2 @@
+build/
+build_sys/

src/compression/simpatico_codegen/CMakeLists.txt

@@ -0,0 +1,185 @@
+cmake_minimum_required(VERSION 3.24)
+
+set(CMAKE_CUDA_ARCHITECTURES
+    native
+    CACHE STRING "CUDA architectures (native = detect host GPU)")
+
+project(simpatico_codegen LANGUAGES C CXX CUDA)
+
+set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CUDA_STANDARD 20)
+set(CMAKE_CUDA_STANDARD_REQUIRED ON)
+
+if(DEFINED ENV{CONDA_PREFIX})
+  set(CONDA_PREFIX
+      "$ENV{CONDA_PREFIX}"
+      CACHE PATH "Conda/pixi prefix")
+else()
+  message(FATAL_ERROR "CONDA_PREFIX must be set (run inside pixi shell)")
+endif()
+
+set(CMAKE_CUDA_COMPILER
+    "${CONDA_PREFIX}/bin/nvcc"
+    CACHE FILEPATH "Conda nvcc")
+
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64")
+  set(CUDA_TARGET_TRIPLE "sbsa-linux")
+else()
+  set(CUDA_TARGET_TRIPLE "x86_64-linux")
+endif()
+set(CUDA_TOOLKIT_INCLUDE
+    "${CONDA_PREFIX}/targets/${CUDA_TARGET_TRIPLE}/include")
+set(CUDA_CCCL_INCLUDE "${CUDA_TOOLKIT_INCLUDE}/cccl")
+
+include_directories(
+  ${CMAKE_CURRENT_SOURCE_DIR}/include ${CMAKE_CURRENT_SOURCE_DIR}/src/util
+  "${CONDA_PREFIX}/include" "${CONDA_PREFIX}/include/rapids"
+  "${CUDA_TOOLKIT_INCLUDE}")
+
+add_compile_definitions(LIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE)
+add_compile_options($<$<COMPILE_LANGUAGE:CUDA>:--expt-extended-lambda>)
+add_compile_options($<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>)
+add_link_options(-Wl,--enable-new-dtags)
+link_directories("${CONDA_PREFIX}/lib")
+
+add_library(
+  simpatico STATIC
+  # CUDA kernels: compact + operators + offsets
+  src/encode/compact.cu
+  src/bridge/offsets_cumsum.cu
+  src/operators/alp_compressor.cu
+  src/operators/alp_rd_compressor.cu
+  src/operators/ans_compressor.cu
+  src/operators/bitcomp_compressor.cu
+  src/operators/cascaded_compressor.cu
+  src/operators/snappy_compressor.cu
+  src/operators/lz4_compressor.cu
+  src/operators/deflate_compressor.cu
+  src/operators/bitjoin_bitextract.cu
+  src/operators/dictionary_compressor.cu
+  src/operators/for_compressor.cu
+  # Plain-CUDA JIT host
+  src/jit/fused_tree.cpp
+  src/jit/nvrtc_compiler.cpp
+  src/jit/kernel_cache.cpp
+  src/encode/jit/plain_compile.cpp
+  src/encode/jit/renderer.cpp
+  src/decode/jit/renderer.cpp
+  src/encode/compact.cpp
+  # Runtime: plan executor + bridge + API + rep glue
+  src/plan/plan_dsl.cpp
+  src/plan/plan_tree.cpp
+  src/plan/fusion.cpp
+  src/plan/operator_registry.cpp
+  src/plan/representation_factory.cpp
+  src/plan/column_copy.cpp
+  src/plan/bitjoin_layout.cpp
+  src/plan/compress.cpp
+  src/plan/decompress.cpp
+  src/rep/compact/bitpack.cpp
+  src/util/stream_pool.cpp
+  src/util/isoc23_compat.c
+  src/bridge/fused_tree_build.cpp
+  src/bridge/codegen_runtime.cpp
+  src/simpatico_codegen.cpp
+  src/api/compressed_table_io.cpp)
+set_target_properties(simpatico PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_compile_features(simpatico PUBLIC cxx_std_20)
+target_include_directories(
+  simpatico PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src "${CUDA_TOOLKIT_INCLUDE}"
+                    "${CUDA_CCCL_INCLUDE}")
+target_compile_definitions(
+  simpatico
+  PRIVATE "CODEGEN_JIT_CUDA_INCLUDE=\"${CUDA_TOOLKIT_INCLUDE}\""
+          "CODEGEN_JIT_CCCL_INCLUDE=\"${CUDA_CCCL_INCLUDE}\""
+          "CODEGEN_JIT_PROJECT_INCLUDE=\"${CMAKE_CURRENT_SOURCE_DIR}/include\"")
+target_link_libraries(
+  simpatico
+  PUBLIC cudf
+         rmm
+         cudart
+         nvcomp
+         nvrtc
+         cuda
+         dl)
+
+function(add_codegen_host_test name)
+  add_executable(${name} tests/${name}.cpp)
+  target_include_directories(${name} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)
+  target_link_libraries(
+    ${name}
+    PRIVATE simpatico
+            cudf
+            rmm
+            cudart
+            nvcomp
+            nvrtc
+            cuda
+            dl)
+  target_link_options(${name} PRIVATE -Wl,--defsym=__libc_single_threaded=0
+                      "-Wl,-rpath,${CONDA_PREFIX}/lib")
+endfunction()
+
+function(add_codegen_cuda_test name)
+  add_executable(${name} tests/${name}.cu)
+  target_include_directories(${name} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)
+  target_link_libraries(
+    ${name}
+    PRIVATE simpatico
+            cudf
+            rmm
+            cudart
+            nvcomp
+            nvrtc
+            cuda
+            dl)
+  target_link_options(${name} PRIVATE -Wl,--defsym=__libc_single_threaded=0
+                      "-Wl,-rpath,${CONDA_PREFIX}/lib")
+endfunction()
+
+add_codegen_host_test(test_compress_with_plan_roundtrip)
+add_codegen_host_test(test_compressed_table_io)
+add_codegen_host_test(test_leaf_describe)
+add_codegen_host_test(test_plan_tree)
+add_codegen_host_test(test_fused_tree_build)
+add_codegen_host_test(test_jit_kernel_cache_plain)
+add_codegen_host_test(test_jit_kernel_cache)
+add_codegen_host_test(test_shape_parity)
+add_codegen_cuda_test(test_encode_bitpack_compact)
+
+# Benchmark harness (not registered as ctest)
+add_executable(compress_with_plan_benchmark
+               bench/compress_with_plan_benchmark.cpp)
+target_include_directories(compress_with_plan_benchmark
+                           PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)
+target_link_libraries(
+  compress_with_plan_benchmark
+  PRIVATE simpatico
+          cudf
+          rmm
+          cudart
+          nvcomp
+          nvrtc
+          cuda
+          dl)
+target_link_options(
+  compress_with_plan_benchmark PRIVATE -Wl,--defsym=__libc_single_threaded=0
+  "-Wl,-rpath,${CONDA_PREFIX}/lib")
+
+enable_testing()
+add_test(NAME compress_with_plan_roundtrip
+         COMMAND test_compress_with_plan_roundtrip)
+add_test(NAME compressed_table_io COMMAND test_compressed_table_io)
+add_test(NAME leaf_describe COMMAND test_leaf_describe)
+add_test(NAME plan_tree COMMAND test_plan_tree)
+add_test(NAME fused_tree_build COMMAND test_fused_tree_build)
+add_test(NAME jit_kernel_cache_plain COMMAND test_jit_kernel_cache_plain)
+add_test(NAME jit_kernel_cache COMMAND test_jit_kernel_cache)
+add_test(NAME shape_parity COMMAND test_shape_parity)
+# Keep the ctest fast: depth 3 (1+3+9 = 13 shapes) still exercises two-level
+# nesting of every op. Run the binary directly (or override the env var) for the
+# full depth-4 sweep.
+set_tests_properties(shape_parity PROPERTIES ENVIRONMENT
+                                             "SIMPATICO_PARITY_DEPTH=3")
+add_test(NAME encode_bitpack_compact COMMAND test_encode_bitpack_compact)