Add apps/ggml: Halide reimplementation of GGML quant kernels - #9417

Draft

alexreinking wants to merge 26 commits into

alexreinking/approximationsfrom

alexreinking/ggml-app

Commits on Oct 2, 2026

Add Stage::distribute()

alexreinking
and
claude
committed
9354301
Add apps/ggml: Halide reimplementation of GGML quant kernels

alexreinking
and
claude
committed
5bef9db
apps/ggml: struct-type the symmetric codec block layout (Phase 3 pilot)

alexreinking
and
claude
committed
bf00ad3
apps/ggml: restore SDOT vec-dot via deep decode-chain inlining

alexreinking
and
claude
committed
9843d1d
apps/ggml: efficient SDOT schedule for repack gemv/gemm

alexreinking
and
claude
committed
ae84256
apps/ggml: struct-type the symmetric vec_dot weight operand

alexreinking
and
claude
committed
80275e9
apps/ggml: cast int8 codes straight to float in LinearDequant (enables SDOT)

alexreinking
and
claude
committed
6a3dfee
apps/ggml: bring q4_0/q8_0 vec_dot up to ggml-cpu speed

alexreinking
and
claude
committed
4fd828a
apps/ggml: take the affine vec_dots (q4_1, q5_1) to SDOT

alexreinking
and
claude
committed
4827a63
apps/ggml: propagate hoist_invariants()'s vector<Func> return through sdot_partial()

alexreinking
and
claude
committed
3b5688d
apps/ggml: notes for resuming the vec_dot performance work

alexreinking
and
claude
committed
b5633a5
apps/ggml: sever q4_1/q5_1 offset term to Q8_1's stored block sum

alexreinking
and
claude
committed
911da2f
apps/ggml: expand q5_x qh high bit via a compile-time LUT (0.51->0.63x)

alexreinking
and
claude
committed
2aa44ca
apps/ggml: q5_0/q5_1 no_asserts + register-resident codes (0.63->0.69x, 0.68->0.78x)

alexreinking
and
claude
committed
dceb4ce
apps/ggml: q5_0 vec_dot ABI wrapper on StackBuffer (166->155 ns)

alexreinking
and
claude
committed
dcb252b
Improve q5_0/q5_1 performance (note: breaks contracts)
alexreinking
committed
98cee49
Checkpoint
alexreinking
committed
9040687
Checkpoint
alexreinking
committed
f765bdd
Manual cleanup
alexreinking
committed
d40772a
Migrate apps/ggml to value-semantic Approximation handles

alexreinking
and
Copilot
committed
eca8b3c
Use simple Approximation unit forms in apps/ggml

alexreinking
and
Copilot
committed
5e367cc
Migrate apps/ggml to discovered Approximation intermediates

alexreinking
and
Copilot
committed
58623b7
Use ApproximationResult::is_stage_port() in the codec generator

alexreinking
and
Copilot
committed
33e72c8
Migrate apps/ggml to Parallel and app-local quantizers

alexreinking
and
Copilot
committed
9c688b6
Rename compute_offline to sever in ggml app

alexreinking
and
Copilot
committed
e11ae0c
Remove version qualifier from struct layout comment

alexreinking
and
Copilot
committed
3cba48a

Navigation Menu

Add apps/ggml: Halide reimplementation of GGML quant kernels - #9417

Add apps/ggml: Halide reimplementation of GGML quant kernels#9417
alexreinking wants to merge 26 commits into
alexreinking/approximationsfrom
alexreinking/ggml-app

Commits on Oct 2, 2026

Add Stage::distribute()

Add apps/ggml: Halide reimplementation of GGML quant kernels

apps/ggml: struct-type the symmetric codec block layout (Phase 3 pilot)

apps/ggml: restore SDOT vec-dot via deep decode-chain inlining

apps/ggml: efficient SDOT schedule for repack gemv/gemm

apps/ggml: struct-type the symmetric vec_dot weight operand

apps/ggml: cast int8 codes straight to float in LinearDequant (enables SDOT)

apps/ggml: bring q4_0/q8_0 vec_dot up to ggml-cpu speed

apps/ggml: take the affine vec_dots (q4_1, q5_1) to SDOT

apps/ggml: propagate hoist_invariants()'s vector<Func> return through sdot_partial()

apps/ggml: notes for resuming the vec_dot performance work

apps/ggml: sever q4_1/q5_1 offset term to Q8_1's stored block sum

apps/ggml: expand q5_x qh high bit via a compile-time LUT (0.51->0.63x)

apps/ggml: q5_0/q5_1 no_asserts + register-resident codes (0.63->0.69x, 0.68->0.78x)

apps/ggml: q5_0 vec_dot ABI wrapper on StackBuffer (166->155 ns)

Improve q5_0/q5_1 performance (note: breaks contracts)

Checkpoint

Checkpoint

Manual cleanup

Migrate apps/ggml to value-semantic Approximation handles

Use simple Approximation unit forms in apps/ggml

Migrate apps/ggml to discovered Approximation intermediates

Use ApproximationResult::is_stage_port() in the codec generator

Migrate apps/ggml to Parallel and app-local quantizers

Rename compute_offline to sever in ggml app

Remove version qualifier from struct layout comment