Replace per-capability libdevice bitcode with unified libdevice.10.bc

Makefile

@@ -152,7 +152,7 @@ POWERPC_LLVM_CONFIG_LIB=$(if $(WITH_POWERPC), powerpc, )
 
 PTX_CXX_FLAGS=$(if $(WITH_NVPTX), -DWITH_NVPTX, )
 PTX_LLVM_CONFIG_LIB=$(if $(WITH_NVPTX), nvptx, )
-PTX_DEVICE_INITIAL_MODULES=$(if $(WITH_NVPTX), libdevice.compute_20.10.bc libdevice.compute_30.10.bc libdevice.compute_35.10.bc, )
+PTX_DEVICE_INITIAL_MODULES=$(if $(WITH_NVPTX), $(BUILD_DIR)/initmod_ptx.libdevice_ll.o, )
 
 OPENCL_CXX_FLAGS=$(if $(WITH_OPENCL), -DWITH_OPENCL, )
 OPENCL_LLVM_CONFIG_LIB=$(if $(WITH_OPENCL), , )
@@ -972,7 +972,7 @@ INITIAL_MODULES = $(RUNTIME_CPP_COMPONENTS:%=$(BUILD_DIR)/initmod.%_32.o) \
                   $(HTML_TEMPLATE_FILES:%=$(BUILD_DIR)/html_template.%.o) \
                   $(BUILD_DIR)/initmod.inlined_c.o \
                   $(RUNTIME_LL_COMPONENTS:%=$(BUILD_DIR)/initmod.%_ll.o) \
-                  $(PTX_DEVICE_INITIAL_MODULES:libdevice.%.bc=$(BUILD_DIR)/initmod_ptx.%_ll.o)
+                  $(PTX_DEVICE_INITIAL_MODULES)
 
 TEST_DEPS = $(BIN_DIR)/libHalide.$(SHARED_EXT) $(INCLUDE_DIR)/Halide.h $(RUNTIME_EXPORTED_INCLUDES)
 ifneq (,$(WITH_EXCEPTIONS))
@@ -1237,8 +1237,8 @@ $(BUILD_DIR)/initmod.%_h.cpp: $(BIN_DIR)/binary2cpp $(SRC_DIR)/runtime/%.h
 $(BUILD_DIR)/initmod.inlined_c.cpp: $(BIN_DIR)/binary2cpp $(SRC_DIR)/runtime/halide_buffer_t.cpp
 	./$(BIN_DIR)/binary2cpp halide_internal_initmod_inlined_c < $(SRC_DIR)/runtime/halide_buffer_t.cpp > $@
 
-$(BUILD_DIR)/initmod_ptx.%_ll.cpp: $(BIN_DIR)/binary2cpp $(SRC_DIR)/runtime/nvidia_libdevice_bitcode/libdevice.%.bc
-	./$(BIN_DIR)/binary2cpp halide_internal_initmod_ptx_$(basename $*)_ll < $(SRC_DIR)/runtime/nvidia_libdevice_bitcode/libdevice.$*.bc > $@
+$(BUILD_DIR)/initmod_ptx.libdevice_ll.cpp: $(BIN_DIR)/binary2cpp $(SRC_DIR)/runtime/nvidia_libdevice_bitcode/libdevice.10.bc
+	./$(BIN_DIR)/binary2cpp halide_internal_initmod_ptx_libdevice_ll < $(SRC_DIR)/runtime/nvidia_libdevice_bitcode/libdevice.10.bc > $@
 
 $(BUILD_DIR)/c_template.%.cpp: $(BIN_DIR)/binary2cpp $(SRC_DIR)/%.template.cpp
 	./$(BIN_DIR)/binary2cpp halide_c_template_$* < $(SRC_DIR)/$*.template.cpp > $@

src/LLVM_Runtime_Linker.cpp

@@ -252,9 +252,7 @@ DECLARE_NO_INITMOD(windows_aarch64_cpu_features_arm)
 #endif  // WITH_AARCH64
 
 #ifdef WITH_NVPTX
-DECLARE_LL_INITMOD(ptx_compute_20)
-DECLARE_LL_INITMOD(ptx_compute_30)
-DECLARE_LL_INITMOD(ptx_compute_35)
+DECLARE_LL_INITMOD(ptx_libdevice)
 #endif  // WITH_NVPTX
 
 #if defined(WITH_D3D12) && defined(WITH_X86)
@@ -1412,22 +1410,7 @@ std::unique_ptr<llvm::Module> get_initial_module_for_ptx_device(Target target, l
     std::vector<std::unique_ptr<llvm::Module>> modules;
     modules.push_back(get_initmod_ptx_dev_ll(c));
 
-    std::unique_ptr<llvm::Module> module;
-
-    // This table is based on the guidance at:
-    // http://docs.nvidia.com/cuda/libdevice-users-guide/basic-usage.html#linking-with-libdevice
-    if (target.has_feature(Target::CUDACapability35)) {
-        module = get_initmod_ptx_compute_35_ll(c);
-    } else if (target.features_any_of({Target::CUDACapability32,
-                                       Target::CUDACapability50})) {
-        // For some reason sm_32 and sm_50 use libdevice 20
-        module = get_initmod_ptx_compute_20_ll(c);
-    } else if (target.has_feature(Target::CUDACapability30)) {
-        module = get_initmod_ptx_compute_30_ll(c);
-    } else {
-        module = get_initmod_ptx_compute_20_ll(c);
-    }
-    modules.push_back(std::move(module));
+    modules.push_back(get_initmod_ptx_libdevice_ll(c));
 
     link_modules(modules, target);
 

src/runtime/CMakeLists.txt

@@ -131,14 +131,6 @@ set(RUNTIME_LL
     # keep-sorted end
 )
 
-set(RUNTIME_BC
-    # keep-sorted start
-    compute_20
-    compute_30
-    compute_35
-    # keep-sorted end
-)
-
 set(RUNTIME_HEADER_FILES
     # keep-sorted start
     HalideBuffer.h
@@ -348,19 +340,16 @@ foreach (i IN LISTS RUNTIME_LL)
     target_sources(Halide_initmod PRIVATE ${INITMOD})
 endforeach ()
 
-foreach (i IN LISTS RUNTIME_BC)
-    set(INITMOD "_initmod_ptx_${i}.cpp")
-    set(RT_BC "${CMAKE_CURRENT_SOURCE_DIR}/nvidia_libdevice_bitcode/libdevice.${i}.10.bc")
-
-    add_custom_command(
-        OUTPUT "${INITMOD}"
-        COMMAND
-            binary2cpp "halide_internal_initmod_ptx_${i}_ll" < "$<SHELL_PATH:${RT_BC}>" > "${INITMOD}"
-        DEPENDS binary2cpp "${RT_BC}"
-        VERBATIM
-    )
-    target_sources(Halide_initmod PRIVATE ${INITMOD})
-endforeach ()
+set(RT_BC "${CMAKE_CURRENT_SOURCE_DIR}/nvidia_libdevice_bitcode/libdevice.10.bc")
+add_custom_command(
+    OUTPUT "_initmod_ptx_libdevice.cpp"
+    COMMAND
+        binary2cpp "halide_internal_initmod_ptx_libdevice_ll" < "$<SHELL_PATH:${RT_BC}>" >
+        "_initmod_ptx_libdevice.cpp"
+    DEPENDS binary2cpp "${RT_BC}"
+    VERBATIM
+)
+target_sources(Halide_initmod PRIVATE "_initmod_ptx_libdevice.cpp")
 
 add_custom_command(
     OUTPUT "_initmod_inlined_c.cpp"

src/runtime/nvidia_libdevice_bitcode/README.md

@@ -0,0 +1,26 @@
+# NVIDIA `libdevice` bitcode
+
+`libdevice.10.bc` is NVIDIA's device-side math library (`sin`, `exp`, `sqrt`,
+etc.), distributed as a single, architecture-generic LLVM bitcode module as part
+of the CUDA Toolkit. It is vendored here verbatim — Halide does not build it —
+and is linked into every GPU kernel module produced by Halide's CUDA/PTX device
+codegen. See:
+
+- `src/CodeGen_PTX_Dev.cpp` — `__nvvm_reflect` / `nvvm-reflect-ftz` handling
+  that libdevice relies on for its accuracy-vs-speed switches.
+- `src/LLVM_Runtime_Linker.cpp` — links this module into the PTX device initial
+  module.
+- `Makefile` / `src/runtime/CMakeLists.txt` — run the file through `binary2cpp`
+  to embed it as a byte array in the compiler binary.
+
+## Updating this file
+
+`libdevice.10.bc` lives at `nvvm/libdevice/libdevice.10.bc` inside any CUDA
+Toolkit install (e.g. `$CUDA_HOME/nvvm/libdevice/libdevice.10.bc`). To update
+the vendored copy, just copy that file over this one — no other files in this
+directory need to change. The current copy was taken from CUDA Toolkit 11.6.
+
+## License
+
+This file is distributed under, and its use governed by, NVIDIA's CUDA Toolkit
+End User License Agreement: https://docs.nvidia.com/cuda/eula