alexreinking/ggml app - #9417

Draft

alexreinking wants to merge 22 commits into

alexreinking/struct-typesfrom

alexreinking/ggml-app

Commits on Sep 15, 2026

hoist_invariants(): return one Func per accumulator, not a Tuple

alexreinking
and
claude
committed
0a59d1a
Add Stage::distribute()

alexreinking
and
claude
committed
9cca9fa
Add Pipeline::compute_offline() directive

alexreinking
and
claude
committed
d72cc09
Add Func::approximate_by() directive and Approximation combinators

alexreinking
and
claude
committed
87640c2
Add apps/ggml: Halide reimplementation of GGML quant kernels

alexreinking
and
claude
committed
11819fc
apps/ggml: struct-type the symmetric codec block layout (Phase 3 pilot)

alexreinking
and
claude
committed
730217e
apps/ggml: restore SDOT vec-dot via deep decode-chain inlining

alexreinking
and
claude
committed
ca5aa76
apps/ggml: efficient SDOT schedule for repack gemv/gemm

alexreinking
and
claude
committed
c49e222
apps/ggml: struct-type the symmetric vec_dot weight operand

alexreinking
and
claude
committed
e216435
apps/ggml: cast int8 codes straight to float in LinearDequant (enables SDOT)

alexreinking
and
claude
committed
355efef
apps/ggml: bring q4_0/q8_0 vec_dot up to ggml-cpu speed

alexreinking
and
claude
committed
ec0ab5c
apps/ggml: take the affine vec_dots (q4_1, q5_1) to SDOT

alexreinking
and
claude
committed
7f432b3
apps/ggml: propagate hoist_invariants()'s vector<Func> return through sdot_partial()

alexreinking
and
claude
committed
9790cd9
apps/ggml: notes for resuming the vec_dot performance work

alexreinking
and
claude
committed
779aecb
apps/ggml: sever q4_1/q5_1 offset term to Q8_1's stored block sum

alexreinking
and
claude
committed
900e7f2
apps/ggml: expand q5_x qh high bit via a compile-time LUT (0.51->0.63x)

alexreinking
and
claude
committed
9e1d63c
apps/ggml: q5_0/q5_1 no_asserts + register-resident codes (0.63->0.69x, 0.68->0.78x)

alexreinking
and
claude
committed
f181f90
apps/ggml: q5_0 vec_dot ABI wrapper on StackBuffer (166->155 ns)

alexreinking
and
claude
committed
c650997
Improve q5_0/q5_1 performance (note: breaks contracts)
alexreinking
committed
761799c
Checkpoint
alexreinking
committed
8b47850
Checkpoint
alexreinking
committed
c0e3f5e
Manual cleanup
alexreinking
committed
4ad2f5b

Navigation Menu

alexreinking/ggml app - #9417

alexreinking/ggml app#9417
alexreinking wants to merge 22 commits into
alexreinking/struct-typesfrom
alexreinking/ggml-app

Commits on Sep 15, 2026

hoist_invariants(): return one Func per accumulator, not a Tuple

Add Stage::distribute()

Add Pipeline::compute_offline() directive

Add Func::approximate_by() directive and Approximation combinators

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