Vulkan: make half-variant GLSL PRECISION configurable

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam

[jgibson2]

Summary

Adds 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.

Why

GLSL 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 conv2d_half / conv2d_pw_tiled_half workloads in our depth model — 289 → 256 ms p50 end-to-end). On Adreno it is typically ignored (harmless).

Tradeoff

mediump relaxes minimum precision guarantees for fp16 shader math. The driver may use fp16 ALUs for the multiply portion of FMAs (often with fp32 accumulators, driver-dependent). That makes this an accuracy tradeoff — kept off by default so this cannot silently affect any existing build.

Wiring

• backends/vulkan/runtime/gen_vulkan_spv.py — new --fp16-precision CLI flag, forwarded into SPVGenerator.create_shader_params().
• backends/vulkan/cmake/ShaderLibrary.cmake — forwards the cmake variable into the Python invocation when set.
• tools/cmake/preset/default.cmake — declares EXECUTORCH_VULKAN_FP16_PRECISION as an overridable STRING option, default "highp".
• scripts/build_android_library.sh — reads EXECUTORCH_VULKAN_FP16_PRECISION from env, defaults to "highp".

Test plan

• Default build (-DEXECUTORCH_VULKAN_FP16_PRECISION unset or highp) emits shaders byte-identical to pre-patch.
• With -DEXECUTORCH_VULKAN_FP16_PRECISION=mediump, conv2d_half.glsl emits #define PRECISION mediump while conv2d_float.glsl stays #define PRECISION highp.
• End-to-end Pixel 9 latency on a polycam depth model improves from 289 ms → 256 ms p50 (~11%).
• Output accuracy comparison vs highp on a representative input set (up to the user of the flag to validate per model — which is why the default is highp).

🤖 Generated with Claude Code

[jgibson2]

Closing because this is merged into polycam