Split WebGPU runtime into two variants (#7248 workaround)

Halide promises that you can crosscompile to *any* supported target from a 'stock' build of libHalide. Unfortunately, the initial landing of WebGPU support breaks that promise: we compile the webgpu runtime support (webgpu.cpp) with code that is predicated on `WITH_DAWN_NATIVE` (for Dawn vs Emscripten, respectively).

This means that if you build Halide with `WITH_DAWN_NATIVE` defined, you can *only* target Dawn with that build of Halide; similarly, if you build with `WITH_DAWN_NATIVE` not-defined, you can only target Emscripten. (Trying to use the 'wrong' version will produce link-time errors.)

For people who build everything from source, this isn't a big deal, but for people who just pull binary builds, this is a big problem.

This PR proposes a temporary workaround until the API discrepancies are resolved:
- Compile the existing webgpu.cpp runtime *both* ways
- in LLVM_Runtime_Linker.cpp, select the correct variant based on whether the Target is targeting wasm or not
- Profit!

This is a rather ugly hack, but it should hopefully be (relatively) temporary.

Makefile

@@ -830,7 +830,8 @@ RUNTIME_CPP_COMPONENTS = \
   trace_helper \
   tracing \
   wasm_cpu_features \
-  webgpu \
+  webgpu_dawn \
+  webgpu_emscripten \
   windows_clock \
   windows_cuda \
   windows_d3d12compute_arm \

src/LLVM_Runtime_Linker.cpp

@@ -144,7 +144,10 @@ DECLARE_CPP_INITMOD(timer_profiler)
 DECLARE_CPP_INITMOD(to_string)
 DECLARE_CPP_INITMOD(trace_helper)
 DECLARE_CPP_INITMOD(tracing)
-DECLARE_CPP_INITMOD(webgpu)
+// TODO(https://github.com/halide/Halide/issues/7248)
+// DECLARE_CPP_INITMOD(webgpu)
+DECLARE_CPP_INITMOD(webgpu_dawn)
+DECLARE_CPP_INITMOD(webgpu_emscripten)
 DECLARE_CPP_INITMOD(windows_clock)
 DECLARE_CPP_INITMOD(windows_cuda)
 DECLARE_CPP_INITMOD(windows_get_symbol)
@@ -1194,7 +1197,13 @@ std::unique_ptr<llvm::Module> get_initial_module_for_target(Target t, llvm::LLVM
                 // See https://github.com/halide/Halide/issues/7249
                 user_error << "WebGPU runtime not yet supported on Windows.\n";
             } else {
-                modules.push_back(get_initmod_webgpu(c, bits_64, debug));
+                // Select the right WebGPU runtime variant based on the Target's OS:
+                // if we are targeting wasm, use the Emscripten variant; in all other cases, use Dawn variant.
+                if (t.os == Target::WebAssemblyRuntime) {
+                    modules.push_back(get_initmod_webgpu_emscripten(c, bits_64, debug));
+                } else {
+                    modules.push_back(get_initmod_webgpu_dawn(c, bits_64, debug));
+                }
             }
         }
         if (t.arch != Target::Hexagon && t.has_feature(Target::HVX)) {

src/runtime/CMakeLists.txt

@@ -79,7 +79,10 @@ set(RUNTIME_CPP
     trace_helper
     tracing
     wasm_cpu_features
-    webgpu
+    # TODO(https://github.com/halide/Halide/issues/7248)
+    # webgpu
+    webgpu_dawn
+    webgpu_emscripten
     windows_clock
     windows_cuda
     windows_d3d12compute_arm
@@ -227,10 +230,6 @@ foreach (i IN LISTS RUNTIME_CPP)
         set(SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/${i}.cpp")
 
         set(RUNTIME_DEFINES -DCOMPILING_HALIDE_RUNTIME -DBITS_${j})
-        if (WEBGPU_NATIVE_LIB)
-            # TODO: Remove this once Emscripten and Dawn agree with each other.
-            set(RUNTIME_DEFINES -DWITH_DAWN_NATIVE ${RUNTIME_DEFINES})
-        endif ()
         set(RUNTIME_DEFINES_debug -g -DDEBUG_RUNTIME ${RUNTIME_DEFINES})
 
         foreach (SUFFIX IN ITEMS "" "_debug")

src/runtime/webgpu.cpp

@@ -7,6 +7,10 @@
 
 #include "mini_webgpu.h"
 
+#ifndef HALIDE_RUNTIME_WEBGPU_NATIVE_API
+#error "HALIDE_RUNTIME_WEBGPU_NATIVE_API must be defined"
+#endif
+
 namespace Halide {
 namespace Runtime {
 namespace Internal {
@@ -40,7 +44,7 @@ using namespace Halide::Runtime::Internal::WebGPU;
 extern "C" {
 // TODO: Remove all of this when wgpuInstanceProcessEvents() is supported.
 // See https://github.com/halide/Halide/issues/7248
-#ifdef WITH_DAWN_NATIVE
+#if HALIDE_RUNTIME_WEBGPU_NATIVE_API
 // Defined by Dawn, and used to yield execution to asynchronous commands.
 void wgpuDeviceTick(WGPUDevice);
 // From <unistd.h>, used to spin-lock while waiting for device initialization.
@@ -274,7 +278,7 @@ void request_adapter_callback(WGPURequestAdapterStatus status,
 
     // TODO: Enable for Emscripten when wgpuAdapterGetLimits is supported.
     // See https://github.com/halide/Halide/issues/7248
-#ifdef WITH_DAWN_NATIVE
+#if HALIDE_RUNTIME_WEBGPU_NATIVE_API
     WGPUSupportedLimits supportedLimits{};
     supportedLimits.nextInChain = nullptr;
     if (!wgpuAdapterGetLimits(adapter, &supportedLimits)) {
@@ -310,7 +314,7 @@ size_t round_up_to_multiple_of_4(size_t x) {
 WEAK int create_webgpu_context(void *user_context) {
     // TODO: Unify this when Emscripten implements wgpuCreateInstance().
     // See https://github.com/halide/Halide/issues/7248
-#ifdef WITH_DAWN_NATIVE
+#if HALIDE_RUNTIME_WEBGPU_NATIVE_API
     WGPUInstanceDescriptor desc{};
     desc.nextInChain = nullptr;
     global_instance = wgpuCreateInstance(&desc);
@@ -325,10 +329,10 @@ WEAK int create_webgpu_context(void *user_context) {
     while (!global_device && init_error_code == halide_error_code_success) {
         // TODO: Use wgpuInstanceProcessEvents() when it is supported.
         // See https://github.com/halide/Halide/issues/7248
-#ifndef WITH_DAWN_NATIVE
-        emscripten_sleep(10);
-#else
+#if HALIDE_RUNTIME_WEBGPU_NATIVE_API
         usleep(1000);
+#else
+        emscripten_sleep(10);
 #endif
     }
     if (init_error_code != halide_error_code_success) {
@@ -514,7 +518,7 @@ WEAK int halide_webgpu_device_release(void *user_context) {
 
             // TODO: Unify this when Emscripten supports wgpuInstanceRelease().
             // See https://github.com/halide/Halide/issues/7248
-#ifdef WITH_DAWN_NATIVE
+#if HALIDE_RUNTIME_WEBGPU_NATIVE_API
             wgpuInstanceRelease(instance);
             global_instance = nullptr;
 #endif

src/runtime/webgpu_dawn.cpp

@@ -0,0 +1,12 @@
+// TODO(https://github.com/halide/Halide/issues/7248):
+//
+// For now, we must build the webgpu runtime two ways:
+// - once for the native API (Dawn)
+// - once for the Emscripten API (Chrome)
+//
+// Once the API discrepancies are resolved we can
+// go back to building this in a sane manner, but for now,
+// we use this sad-but-effective approach.
+
+#define HALIDE_RUNTIME_WEBGPU_NATIVE_API 1
+#include "webgpu.cpp"

src/runtime/webgpu_emscripten.cpp

@@ -0,0 +1,12 @@
+// TODO(https://github.com/halide/Halide/issues/7248):
+//
+// For now, we must build the webgpu runtime two ways:
+// - once for the native API (Dawn)
+// - once for the Emscripten API (Chrome)
+//
+// Once the API discrepancies are resolved we can
+// go back to building this in a sane manner, but for now,
+// we use this sad-but-effective approach.
+
+#define HALIDE_RUNTIME_WEBGPU_NATIVE_API 0
+#include "webgpu.cpp"