Vulkan: make half-variant GLSL PRECISION configurable

Add EXECUTORCH_VULKAN_FP16_PRECISION cmake option (default "highp" —
upstream-identical behavior). When set to "mediump" (or "lowp"),
gen_vulkan_spv.py overrides the PRECISION define in shader variants where
DTYPE=half.

GLSL's mediump is a driver hint that fp16 ALUs may be used for arithmetic.
On Mali GPUs this is typically honored and gives measurable conv speedups
(~15% on Pixel 9 Mali G715 for this repo's conv2d_half / conv2d_pw_tiled_half
workloads); on Adreno it is typically ignored (harmless).

Accuracy tradeoff: mediump relaxes minimum precision guarantees for fp16
shader math. Kept off by default so this cannot silently affect any
existing build. Exposed via scripts/build_android_library.sh env var for
Android users.

Wiring:
 * backends/vulkan/runtime/gen_vulkan_spv.py: --fp16-precision CLI flag,
   forwarded into SPVGenerator.create_shader_params().
 * backends/vulkan/cmake/ShaderLibrary.cmake: forward cmake var into the
   python invocation when set.
 * tools/cmake/preset/default.cmake: declare overridable STRING option.
 * scripts/build_android_library.sh: read EXECUTORCH_VULKAN_FP16_PRECISION
   from env, default "highp".

Author: jgibson2

backends/vulkan/cmake/ShaderLibrary.cmake

@@ -60,6 +60,14 @@ function(gen_vulkan_shader_lib_cpp shaders_path)
     )
   endif()
 
+  # Allow overriding GLSL PRECISION for fp16 shader variants. Empty / unset
+  # keeps upstream default (`highp`). Accepted values: highp, mediump, lowp.
+  if(EXECUTORCH_VULKAN_FP16_PRECISION)
+    list(APPEND GEN_SPV_ARGS "--fp16-precision"
+         "${EXECUTORCH_VULKAN_FP16_PRECISION}"
+    )
+  endif()
+
   # Ninja cannot expand wildcards (*) in DEPENDS lists.
   file(GLOB VULKAN_SHADERS "${shaders_path}/*.glsl" "${shaders_path}/*.glslh"
        "${shaders_path}/*.yaml" "${shaders_path}/*.h"

backends/vulkan/runtime/gen_vulkan_spv.py

@@ -663,6 +663,7 @@ def __init__(
         glslc_path: Optional[str],
         glslc_flags: str = "",
         replace_u16vecn: bool = False,
+        fp16_precision: str = "highp",
     ) -> None:
         if isinstance(src_dir_paths, str):
             self.src_dir_paths = [src_dir_paths]
@@ -678,6 +679,7 @@ def __init__(
         if "-Os" in self.glslc_flags_no_opt:
             self.glslc_flags_no_opt.remove("-Os")
         self.replace_u16vecn = replace_u16vecn
+        self.fp16_precision = fp16_precision
 
         self.src_files: Dict[str, str] = {}
         self.template_yaml_files: List[str] = []
@@ -857,6 +859,17 @@ def create_shader_params(
         for key, value in variant_params.items():
             shader_params[key] = value
 
+        # Optionally override PRECISION for half-precision variants. GLSL
+        # `mediump` is a hint the driver may use fp16 ALUs for arithmetic.
+        # On Mali GPUs it's typically honored; on Adreno it's typically
+        # ignored (harmless). Default is `highp` to match upstream behavior.
+        if (
+            self.fp16_precision != "highp"
+            and shader_params.get("DTYPE") == "half"
+            and shader_params.get("PRECISION") == "highp"
+        ):
+            shader_params["PRECISION"] = self.fp16_precision
+
         return shader_params
 
     def constructOutputMap(self) -> None:
@@ -1488,6 +1501,16 @@ def main(argv: List[str]) -> int:
         default=-1,
         help="Number of threads for shader compilation. -1 (default) uses all available CPU cores, 1 uses sequential compilation.",
     )
+    parser.add_argument(
+        "--fp16-precision",
+        choices=["highp", "mediump", "lowp"],
+        default="highp",
+        help=(
+            "GLSL PRECISION qualifier for DTYPE=half shader variants. "
+            "`mediump` lets drivers (notably Mali) use fp16 ALUs for arithmetic. "
+            "Default `highp` matches upstream behavior. Ignored on fp32 variants."
+        ),
+    )
     options = parser.parse_args()
 
     env = DEFAULT_ENV
@@ -1520,6 +1543,7 @@ def main(argv: List[str]) -> int:
         options.glslc_path,
         glslc_flags=glslc_flags_str,
         replace_u16vecn=options.replace_u16vecn,
+        fp16_precision=options.fp16_precision,
     )
     output_spv_files = shader_generator.generateSPV(
         options.output_path,

scripts/build_android_library.sh

@@ -50,6 +50,7 @@ build_android_native_library() {
     -DEXECUTORCH_BUILD_QNN="${EXECUTORCH_BUILD_QNN}" \
     -DQNN_SDK_ROOT="${QNN_SDK_ROOT}" \
     -DEXECUTORCH_BUILD_VULKAN="${EXECUTORCH_BUILD_VULKAN}" \
+    -DEXECUTORCH_VULKAN_FP16_PRECISION="${EXECUTORCH_VULKAN_FP16_PRECISION:-highp}" \
     -DXNNPACK_ENABLE_ARM_SME2="${XNNPACK_ENABLE_ARM_SME2}" \
     -DSUPPORT_REGEX_LOOKAHEAD=ON \
     -DCMAKE_BUILD_TYPE="${EXECUTORCH_CMAKE_BUILD_TYPE}" \

tools/cmake/preset/default.cmake

@@ -168,6 +168,12 @@ define_overridable_option(
 define_overridable_option(
   EXECUTORCH_BUILD_VULKAN "Build the Vulkan backend" BOOL OFF
 )
+define_overridable_option(
+  EXECUTORCH_VULKAN_FP16_PRECISION
+  "GLSL PRECISION for Vulkan half-precision shader variants. Accepted values: highp, mediump, lowp. `mediump` lets Mali drivers use fp16 ALUs; ignored on Adreno. Default `highp` matches upstream."
+  STRING
+  highp
+)
 define_overridable_option(
   EXECUTORCH_BUILD_PORTABLE_OPS "Build portable_ops library" BOOL ON
 )