Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.
C
air_run is built for the phone and in the image: the slot can now be spent running
vcvkk19aa7a9cd4f4
diff --git a/docs/air.md b/docs/air.mdindex 7171472..17ba2ea 100644--- a/docs/air.md+++ b/docs/air.md@@ -442,11 +442,24 @@ the Mesa already on the phone rather than being built from nothing:screen (`sw_screen_create(null_sw_create())`); on the phone it wants the screen`osmesa_create_gpu_screen()` makes, and **no line below that changes**.-What is left to do on the phone is therefore small and known: export that screen from the port's-Mesa (one function, beside the `osmesa_surface_resource` the patch already exports), cross-build-`air_run.c` with `userland/build.sh`, and run it. What is *not* known until it is tried is-whether the kernel's GPU driver and the submit path behave for a draw that does not come from-CoreAnimation; the GLES path is the only consumer of that Mesa today.+**That is now built and in the image, and no export was needed.** `air_run` is built *inside* the+Mesa cross build (`AIR=1 tools/mk-mesa.sh`), so it links Gallium's own static libraries and opens+the Adreno itself — `drmOpenWithType("msm", …)` through `tools/mesa/redwood-drm`, then+`fd_drm_screen_create_renderonly()`, exactly as `osmesa_create_gpu_screen()` does. It needs+nothing exported from `libOSMesa` and does not load it: a 9.8 MB arm64 Mach-O against libSystem,+libz and IOKit alone. `userland/build.sh` puts it in `/redwood/bin`, and `rc.jobs.air` runs it.++The same source takes the GPU where there is one and falls back to software where there is not,+so it is the binary tested on this machine — not a variant of it. Its first line says which it+got (`air_run: clearFragment01 on softpipe` here), and that line must be read before any number+below it is believed. `AIR_RUN_SOFTWARE=1` forces software even on the phone, for comparing the+two on the same device.++What is *not* known until it is tried is whether the kernel's GPU driver and the submit path+behave for a draw that does not come from CoreAnimation: the GLES path is that driver's only+consumer today. `rc.jobs.air` asks with the GPU idle; `rc.jobs.air2` asks again after SpringBoard+is up and holding it. A refusal is a finding the whole port needs, not just this frontend, so+either answer is worth the run.## 6. Running itdiff --git a/tools/air/mk-host.sh b/tools/air/mk-host.shindex 70725cd..acc41ee 100755--- a/tools/air/mk-host.sh+++ b/tools/air/mk-host.sh@@ -70,11 +70,11 @@ if [ ! -d "$SRC" ] || [ "$(cat "$SRC/.air-patches" 2>/dev/null)" != "$PATCHES" ]open(p, 'w').write(s)PY- # the frontend itself, as a link: no copy to forget to refresh+ # the frontend itself, as a link: no copy to forget to refresh. The one+ # line of meson that reaches it comes from tools/mesa/air-frontend.patch,+ # applied above, so this build and the cross build get it the same way.rm -rf "$SRC/src/compiler/air"ln -s "$ROOT/tools/mesa/air" "$SRC/src/compiler/air"- grep -q "subdir('air')" "$SRC/src/compiler/meson.build" ||- sed -i "s|^subdir('spirv')|subdir('spirv')\nsubdir('air')|" "$SRC/src/compiler/meson.build"echo "$PATCHES" > "$SRC/.air-patches"fidiff --git a/tools/mesa/air-frontend.patch b/tools/mesa/air-frontend.patchnew file mode 100644index 0000000..eedf8b4--- /dev/null+++ b/tools/mesa/air-frontend.patch@@ -0,0 +1,18 @@+Build the AIR frontend as part of Mesa.++src/compiler/air is this repository's tools/mesa/air: the frontend that reads Apple's shader+bitcode and produces NIR (docs/air.md). tools/air/mk-host.sh links it there for the build on+this machine, tools/mk-mesa.sh copies it there for the cross build; this patch is only the one+line of meson that reaches it, so that both builds get it the same way and neither needs a sed.++--- a/src/compiler/meson.build++++ b/src/compiler/meson.build+@@ -63,6 +63,8 @@++ subdir('spirv')+++subdir('air')+++ if with_clc+ subdir('clc')+ endifdiff --git a/tools/mesa/air-ir3.patch b/tools/mesa/air-ir3.patchindex 3a113f2..823d288 100644--- a/tools/mesa/air-ir3.patch+++ b/tools/mesa/air-ir3.patch@@ -193,7 +193,7 @@ dimensions -- and hides on the rest, which is presumably why it survived.idep_vtn,idep_mesautil,],-@@ -318,6 +319,21 @@+@@ -318,6 +319,22 @@install : false,)@@ -204,10 +204,11 @@ dimensions -- and hides on the rest, which is presumably why it survived.+ 'air_run',+ files('../../../compiler/air/air_run.c'),+ include_directories : [inc_include, inc_src, inc_mapi, inc_mesa, inc_gallium,-+ inc_gallium_aux, inc_gallium_winsys_sw,++ inc_gallium_aux, inc_gallium_winsys_sw, inc_gallium_winsys,+ include_directories('../../drivers')],-+ dependencies : [dep_m, idep_air, idep_mesautil, driver_softpipe],-+ link_with : [libgallium, libgalliumvl_stub, libws_null],++ dependencies : [dep_m, idep_air, idep_mesautil, driver_softpipe, driver_freedreno,++ dep_libdrm],++ link_with : [libgallium, libgalliumvl_stub, libws_null, libfreedrenowinsys],+ build_by_default : false,+ install : false,+)diff --git a/tools/mesa/air/air_run.c b/tools/mesa/air/air_run.cindex bbb7596..93faea1 100644--- a/tools/mesa/air/air_run.c+++ b/tools/mesa/air/air_run.c@@ -37,14 +37,23 @@#include "pipe/p_state.h"#include "target-helpers/inline_sw_helper.h"#include "null/null_sw_winsys.h"+#ifdef GALLIUM_FREEDRENO+#include "freedreno/drm/freedreno_drm_public.h"+#include <xf86drm.h>+#endif#include "util/format/u_format.h"#include "util/u_inlines.h"+#include "util/u_debug.h"#include "util/u_memory.h"#ifdef __APPLE__#include <mach-o/dyld.h>#include <signal.h>-#include <ucontext.h>+/* <sys/ucontext.h>, not <ucontext.h>: the latter refuses to be included+ * without _XOPEN_SOURCE on the iOS SDK, and userland/src/inject.c takes the+ * same route for the same reason.+ */+#include <sys/ucontext.h>#endif#define WIDTH 4@@ -213,13 +222,27 @@ main(int argc, char **argv)name = lib.functions[pick].name;}- /* The screen. Here it is softpipe, drawing into memory with no GPU at all,- * which is the point: the whole path can be checked on this machine. On the- * phone it is the freedreno screen that the port's Mesa already makes- * (osmesa_create_gpu_screen in tools/mesa/freedreno-darwin.patch), and- * nothing below this line changes.+ /* The screen: the GPU's where the kernel has one to drive, softpipe+ * otherwise. The same source therefore runs on this machine (no msm+ * device, so it falls back and draws in software) and on the phone, where+ * it opens the Adreno exactly as the port's own Mesa does+ * (osmesa_create_gpu_screen in tools/mesa/freedreno-darwin.patch).+ * AIR_RUN_SOFTWARE=1 forces software even where a GPU is present.*/- struct pipe_screen *screen = sw_screen_create(null_sw_create());+ struct pipe_screen *screen = NULL;+#ifdef GALLIUM_FREEDRENO+ if (!debug_get_bool_option("AIR_RUN_SOFTWARE", false)) {+ int fd = drmOpenWithType("msm", NULL, DRM_NODE_RENDER);+ if (fd >= 0) {+ static struct pipe_screen_config config;+ screen = fd_drm_screen_create_renderonly(fd, NULL, &config);+ if (!screen)+ fprintf(stderr, "air_run: the msm device opened but gave no screen\n");+ }+ }+#endif+ if (!screen)+ screen = sw_screen_create(null_sw_create());if (!screen) {fprintf(stderr, "air_run: no gallium screen\n");return 1;diff --git a/tools/mk-mesa.sh b/tools/mk-mesa.shindex 61b1612..4e15eaf 100755--- a/tools/mk-mesa.sh+++ b/tools/mk-mesa.sh@@ -46,6 +46,11 @@ fi# ours, kept in the repository: always the current onesrm -rf mesa-$VERSION/src/redwood-drmcp -r "$ROOT/tools/mesa/redwood-drm" mesa-$VERSION/src/redwood-drm+# the AIR frontend (docs/air.md): the Metal runtime will call it in this+# library's own process, and tools/mesa/air-frontend.patch is the meson that+# reaches it. Copied rather than linked, because this tree is cross-built.+rm -rf mesa-$VERSION/src/compiler/air+cp -r "$ROOT/tools/mesa/air" mesa-$VERSION/src/compiler/airif [ ! -x venv/bin/python3 ]; thenpython3 -m venv venvvenv/bin/pip -q install mako pyyaml packaging@@ -83,6 +88,14 @@ BUILD=build-${DRIVERS//,/-}-Dlibunwind=disabled -Dglvnd=disabled -Dvideo-codecs= -Dtools= -Dgallium-vdpau=disabled -Dgallium-va=disabled \-Dgallium-xa=disabled -Dlmsensors=disabled -Dbuild-tests=false > meson-setup.logninja -C $BUILD > ninja.log+# air_run draws with a translated Metal shader and reads the pixels back; it is+# the only way to find out whether the Adreno takes work that did not come from+# the window server. Built only when asked, since it is a test program.+if [ "${AIR:-}" = 1 ]; then+ ninja -C $BUILD src/gallium/drivers/freedreno/air_run >> ninja.log+ cp $BUILD/src/gallium/drivers/freedreno/air_run out/air_run+ echo "mk-mesa.sh: $B/out/air_run ($(stat -c %s out/air_run) bytes)"+fillvm-strip -x -o out/libOSMesa.8.dylib $BUILD/src/gallium/targets/osmesa/libOSMesa.8.dylibinstall_name_tool -id /redwood/lib/libOSMesa.8.dylib out/libOSMesa.8.dylibdiff --git a/userland/build.sh b/userland/build.shindex 2837202..a47401a 100755--- a/userland/build.sh+++ b/userland/build.sh@@ -174,6 +174,12 @@ if [ -d $IOS/rootfs/usr/lib ]; thensrc/eagl.c src/gl_osmesa.c$DYN -framework CoreFoundation -framework IOKit $M/libOSMesa.8.dylib -o obj/gl_test src/gl_test.ccp $M/libOSMesa.8.dylib obj/+ # air_run draws with a Metal shader translated by the AIR frontend+ # and reads the pixels back (docs/air.md). It is built inside the+ # Mesa build, because it needs Gallium itself and not the gl*+ # entry points this library exports; AIR=1 tools/mk-mesa.sh makes+ # it, and it is copied in when it is there.+ [ -f $M/air_run ] && cp $M/air_run obj/air_runfificp obj/hello_dyn obj/svc_test obj/surf_test obj/rwtouch obj/mg_test root/bin/@@ -210,6 +216,7 @@ if [ -d $IOS/rootfs/usr/lib ]; thenif [ -n "$GL" ]; thencp obj/libOSMesa.8.dylib root/lib/cp obj/gl_test root/bin/+ [ -f obj/air_run ] && cp obj/air_run root/bin/[ "$GL" = test ] || cp obj/libredwood-gl.dylib root/lib/fi# a program is not started with a library that is not there: it would not start at alldiff --git a/userland/rc.jobs.air b/userland/rc.jobs.airnew file mode 100644index 0000000..8cd5478--- /dev/null+++ b/userland/rc.jobs.air@@ -0,0 +1,21 @@+# Commands run by the launchd job com.redwood.demo (see com.redwood.demo.plist).+#+# The first draw on this GPU by a shader Apple compiled for Metal (RC_JOBS=rc.jobs.air).+# /redwood/bin/air_run translates one function of a .metallib with the AIR frontend+# (docs/air.md), hands the NIR to Gallium as PIPE_SHADER_IR_NIR, draws one quad and prints+# the pixel it reads back. With the Adreno present it opens the msm device and ir3 compiles+# the shader; without one it falls back to softpipe and says so in its first line, so the+# line "air_run: <name> on FD643" (or similar) is what tells the two apart -- read it before+# believing any number below it.+#+# clearFragment01 is the shader M1 is stated against: four LLVM instructions, no air.*+# intrinsic, its colour read straight out of the constant buffer at index 14. The buffer is+# filled with (i * 7 + 13) & 0xff, so the four floats expected are the ones LLVM itself+# computes on a computer and softpipe draws there:+# 2.10170319e-18 0.178894654 1.48687698e+16 1.21394141e+33+# Anything else from this run is the finding, whatever it is.+#+# No wait for SpringBoard: this draws for itself and wants the GPU idle rather than busy with+# the window server. With gpu=1 the window server has it too, which is worth knowing about on+# purpose -- but the first run should be the simple one.+/redwood/bin/air_run /System/Library/Extensions/AppleMetalGLRenderer.bundle/default.metallib clearFragment01diff --git a/userland/rc.jobs.air2 b/userland/rc.jobs.air2new file mode 100644index 0000000..ddcb4ba--- /dev/null+++ b/userland/rc.jobs.air2@@ -0,0 +1,6 @@+# As rc.jobs.air, but after SpringBoard is up and drawing: does the kernel's GPU driver take+# work from a second process at all? That question is owed to the whole port, not to the AIR+# frontend, and a refusal here is as much of a result as a pixel.+/redwood/bin/sound_test springboard+/redwood/bin/rwtouch wait 25+/redwood/bin/air_run /System/Library/Extensions/AppleMetalGLRenderer.bundle/default.metallib clearFragment01