bench(amd): add Autolykos v2 throughput benchmark by yuzi-co · Pull Request #173 · luminousmining/miner

yuzi-co · 2026-06-11T21:56:03Z

What

Adds an in-tree AMD throughput benchmark for Autolykos v2 — the last
AMD-mined algorithm without one. It reuses the production resolver kernels
(DAG build, search, verify) byte-for-byte via the kernel generator, so it
measures exactly what the miner runs.

The DAG is built once (untimed setup); the two hot-loop kernels —
autolykos_v2_search and autolykos_v2_verify — are timed independently,
with the grid expressed in nonce terms so the report reads nonces/s.

Changes

sources/benchmark/amd/autolykos_v2.cpp — runAmdAutolykos() (new)
sources/benchmark/workflow.{hpp,cpp} — declare + call it in runAmd()
sources/benchmark/config.json — amd.autolykos_v2 entry (disabled by default, like its siblings)
sources/benchmark/config.cpp — register autolykos_v2 in the AMD defaults (parity with NVIDIA)
documentation/BENCHMARK.md — source-tree / call-graph diagrams, kernel-table row, RX 9070 XT reference results

sources/benchmark/amd/CMakeLists.txt globs *.cpp, and the .cl files
auto-stage into bin/kernel/autolykos/ via the existing ALL target — no
build-system changes required.

Verification

Built for AMD and run on an RX 9070 XT:

Kernel	Hashrate
`autolykos_v2_search`	~1.05 GH/s
`autolykos_v2_verify`	~69 MH/s

Issue

Part of #89 — completes AMD → AutolykosV2. (AMD Ethash, still unchecked
there, already landed in #142; details in an issue comment.)

Closes the last coverage gap in the AMD benchmark suite: autolykos_v2 was the only mined AMD algorithm with no in-tree benchmark, while NVIDIA already had one. Rescued from the stale bench/amd-ethash-progpow-ab branch (the only unmerged content on it) and normalized to the sibling-benchmark conventions. The benchmark reuses the production resolver kernels (DAG build, search, verify) byte-for-byte via the kernel generator, so it measures exactly what the miner runs. The DAG is built once (untimed) and the two hot-loop kernels are timed independently; the grid is expressed in nonce terms so the report reads nonces/s. - sources/benchmark/amd/autolykos_v2.cpp: runAmdAutolykos() (new) - workflow.{hpp,cpp}: declare + call it in runAmd() - config.json: amd.autolykos_v2 entry, disabled by default like its siblings - config.cpp: register autolykos_v2 in the AMD defaults (parity with NVIDIA) Verified on RX 9070 XT: autolykos_v2_search ~1.05 GH/s, autolykos_v2_verify ~69 MH/s.

Add autolykos_v2 to the benchmark source-tree and runAmd() call-graph diagrams, a row in the kernel coverage table, and an AMD RX 9070 XT reference-results section (search ~1.05 GH/s, verify ~69 MH/s).

The AMD autolykos_v2 benchmark assembled the live production kernels (kernel/autolykos/*.cl). Move to self-contained snapshots under benchmark/opencl/autolykos_v2/, matching the kheavyhash/ethash convention: - autolykos_v2_lm1.cl - search throughput kernel (entry renamed) - autolykos_v2_verify.cl - autolykos_v2_dag.cl - untimed DAG fill Each is a frozen copy of the production assembly (result + rotate_byte + blake2b [+ blake2b_compress] + var_global + search/verify/dag), so the benchmark never tracks live mining-kernel edits. Register the files in the opencl copy target, point the driver at kernel/autolykos_v2/*, and rename the search stage autolykos_v2_search -> autolykos_v2_lm1 (config.json + docs). Verified on RX 9070 XT (gfx1201): autolykos_v2_lm1 ~1.03 GH, verify ~65 MH.

luminousmining · 2026-06-12T13:00:11Z

 | Ethash | 2 (lm_0, lm_1) | barrier vs sub_group_barrier |
 | ProgPOW | 2 (lm_0, lm_1) | barrier vs sub_group_barrier |
 | kHeavyHash | 6 (lm0–lm5) | ALU-bound; `v_dot4` matmul + keccak midstate ([study](reasearch_and_development/kheavyhash/amd.md)) |
+| Autolykos V2 | 2 (search, verify) | production Ergo kernels — DAG-bound throughput coverage |


I should see search_lm and verify_lm.

luminousmining · 2026-06-12T13:00:27Z

+| Kernel                | Hashrate (steady) | Notes |
+|-----------------------|-------------------|-------|
+| `autolykos_v2_lm1`    | **~1.05 GH** | blake2b prehash over the DAG → BHashes; tracks the GPU boost clock, so it varies run to run |
+| `autolykos_v2_verify` | ~69 MH | final blake2b + boundary test over the full grid |


luminousmining · 2026-06-12T13:00:35Z

+
+| Kernel                | Hashrate (steady) | Notes |
+|-----------------------|-------------------|-------|
+| `autolykos_v2_lm1`    | **~1.05 GH** | blake2b prehash over the DAG → BHashes; tracks the GPU boost clock, so it varies run to run |


ssearch_lm0

luminousmining · 2026-06-12T13:01:20Z

+    0x1F83D9ABFB41BD6B, 0x5BE0CD19137E2179
+};
+__kernel
+void autolykos_v2_verify(


autolykos_v2_verify -> autolykos_v2_verify_lm0

luminousmining · 2026-06-12T13:02:55Z

+    ulong nonces[4];
+} t_result;
+inline
+ulong rol_u64(


already exist in rotate_byte.cl

luminousmining · 2026-06-12T13:11:34Z

+
+
+inline
+ulong ror_u64(


already exisst in rotate_byte.cl

luminousmining · 2026-06-12T13:11:38Z

+
+
+inline
+uint rol_u32(uint x, uint n)


already exisst in rotate_byte.cl

luminousmining · 2026-06-12T13:11:41Z

+
+
+inline
+uint ror_u32(uint x, uint n)


already exisst in rotate_byte.cl

luminousmining · 2026-06-12T13:11:48Z

+
+
+inline
+uint bswap32(uint const x)


move to rotate_byte.cl

…, rename kernels - Drop inlined rol/ror/bswap helpers from the dag/search/verify kernels; the driver now prepends the shared kernel/common/rotate_byte.cl, matching the production resolver's assembly order. - Verify kernel uses the shared kernel/common/result.cl t_result struct (driver appends it); the local copy and the unused struct in the search kernel are removed. - Uppercase the function-style macros and the blake2b IV constant: reverseBytesInt -> REVERSE_BYTES_INT, fn_Add -> FN_ADD, ivals -> IVALS. - Rename the timed kernels and their files to the *_lm0 convention: autolykos_v2_lm1{.cl,} -> autolykos_v2_search_lm0, autolykos_v2_verify{.cl,} -> autolykos_v2_verify_lm0. Driver setKernelName/appendFile/runStage labels, the opencl CMakeLists copy list and BENCHMARK.md updated to match.

yuzi-co · 2026-06-12T13:58:01Z

Pushed 99dce0d addressing the review.

Common code dedup

Removed the inlined rol_u64/ror_u64/rol_u32/ror_u32/bswap32 helpers from the dag, search and verify kernels. The driver now prepends kernel/common/rotate_byte.cl to each KernelGenerator, same as ResolverAmdAutolykosV2 does.
The verify kernel now uses the shared kernel/common/result.cl t_result struct (driver appends it before the kernel body). The local struct copy is gone, and the unused struct in the search kernel was dropped too.

Uppercase

reverseBytesInt → REVERSE_BYTES_INT, fn_Add → FN_ADD, ivals → IVALS.

Renames — one note: your suggested targets had a few typos (autolykoss, ssearch, earch) and mixed lm0/lm1, so I normalised to the *_lm0 convention used by the other algos. Please confirm these are what you meant:

file autolykos_v2_lm1.cl → autolykos_v2_search_lm0.cl, kernel autolykos_v2_search_lm0
file autolykos_v2_verify.cl → autolykos_v2_verify_lm0.cl, kernel autolykos_v2_verify_lm0
autolykos_v2_dag.cl kept its name (untimed setup, not a timed lm variant)

Driver setKernelName/appendFile/runStage labels, the benchmark/opencl/CMakeLists.txt copy list, and BENCHMARK.md (variants column + per-kernel table) all updated to match.

Happy to switch the names if you prefer a different scheme.

luminousmining · 2026-06-13T23:23:36Z

+
+
+inline
+void blake2b_compress(ulong *h, const ulong *m, ulong t, ulong f)


const order long const* const m

luminousmining · 2026-06-13T23:24:13Z

+
+
+inline
+void blake2b_compress(ulong *h, const ulong *m, ulong t, ulong f)


t and f must be const

luminousmining · 2026-06-13T23:24:28Z

+
+
+inline
+void blake2b_compress(ulong *h, const ulong *m, ulong t, ulong f)


long* const h

luminousmining · 2026-06-13T23:25:29Z

+}
+__kernel
+void autolykos_v2_build_dag(
+    __global uint* restrict dag,


__global uint* restrict dag, -> __global uint* const restrict dag, ?

luminousmining · 2026-06-13T23:26:00Z

+    //  Hash constant message
+    //====================================================================//
+    ulong ctr = 0;
+    for (int x = 1; x < 16; ++x, ++ctr)


luminousmining · 2026-06-13T23:33:48Z

    ${OUT_COMMON}/kheavyhash/kHeavyHash_lm3.cl
    ${OUT_COMMON}/kheavyhash/kHeavyHash_lm4.cl
    ${OUT_COMMON}/kheavyhash/kHeavyHash_lm5.cl
+    ${OUT_COMMON}/autolykos_v2/autolykos_v2_dag.cl


Alphabetical order

luminousmining · 2026-06-13T23:33:52Z

    kheavyhash/kHeavyHash_lm3.cl
    kheavyhash/kHeavyHash_lm4.cl
    kheavyhash/kHeavyHash_lm5.cl
+    autolykos_v2/autolykos_v2_dag.cl


Alphabetical order

…64 helper, unroll attributes - blake2b_compress: const-correct signature (ulong* const h, ulong const* const m, ulong const t/f); build_dag dag arg is now __global uint* const restrict - add bswap64 to common/opencl/rotate_byte.cl and use it (plus existing bswap32) in the DAG kernel instead of inlined as_ulong/as_uchar8 byte-swaps - DAG loop counters renamed to i; add the missing unroll hint on the message-init loop - replace #pragma unroll / unroll N with __attribute__((opencl_unroll_hint[(N)])) across dag/search/verify kernels (house convention) - 2 blank lines before each __kernel - CMakeLists: order autolykos_v2 entries alphabetically (after common/)

yuzi-co · 2026-06-14T00:42:42Z

Pushed 3a543e2 addressing the latest review.

const-correctness

blake2b_compress signature is now void blake2b_compress(ulong* const h, ulong const* const m, ulong const t, ulong const f).
autolykos_v2_build_dag takes __global uint* const restrict dag.

bswap helpers

Added bswap64 to sources/common/opencl/rotate_byte.cl (alongside the existing bswap32). The DAG kernel now calls bswap64(...)/bswap32(...) instead of the inlined as_ulong(as_uchar8(...).s76543210) / as_uint(as_uchar4(...).s3210) byte-swaps. rotate_byte.cl is already prepended to all three kernels, so the helper is in scope everywhere.

loop counters / unroll

DAG loop counters renamed to i; added the missing unroll hint on the message-init loop.
Replaced every #pragma unroll / #pragma unroll N with __attribute__((opencl_unroll_hint)) / __attribute__((opencl_unroll_hint(N))) across the dag/search/verify kernels.

layout

2 blank lines before each __kernel.

CMakeLists

Ordered the autolykos_v2 entries alphabetically (placed after the common/ block) in both BENCHMARK_OPENCL_FILES and the copy_benchmark_opencl DEPENDS list. I left the pre-existing progpow/kheavyhash order untouched since it's outside this PR — happy to fully sort the list if you'd prefer.

One open question: the verify/search B2B_* macros and REVERSE_BYTES_INT still use inline byte ops rather than the common helpers — I kept those as-is to keep these frozen snapshots byte-identical to the production assembly, but I can route them through rotate_byte.cl too if you'd like consistency there.

luminousmining · 2026-06-14T03:58:11Z

If you can create common functions that can be used in other kernels and that are not "specific" to AutolykosV2, then I would be happy to have those functions in rotate_byte.cl.

If you are referring to:

#define B2B_G(v, a, b, c, d, x, y)

This part is too specific to AutolykosV2 and should therefore be kept exclusively for AutolykosV2.

The search and verify benchmark kernels carried a local REVERSE_BYTES_INT(input, output) macro that is byte-for-byte identical to the bswap32 helper already provided by common/opencl/rotate_byte.cl (both reverse the four bytes of a uint). rotate_byte.cl is already prepended to every autolykos_v2 benchmark kernel, so route the call sites through bswap32 and drop the duplicate macro from both files. The B2B_* blake2b macros remain AutolykosV2-specific. No functional change: GPU-verified on an RX 9070 XT (search ~1.49 GH/s, verify ~69 MH/s), kernels compile and run unchanged.

The kernels were renamed to autolykos_v2_search_lm0 / autolykos_v2_verify_lm0, but the AMD benchmark config still referenced the pre-rename names (autolykos_v2_lm1 / autolykos_v2_verify). isKernelEnabled therefore matched nothing and the autolykos_v2 table rendered empty. Update the config to the current kernel names so the benchmark times the kernels the driver actually builds.

yuzi-co · 2026-06-14T07:54:52Z

Done — pushed 5ce5b34.

Generic helper → rotate_byte.cl
The only generic byte op left in the search/verify kernels was REVERSE_BYTES_INT(input, output), which is byte-for-byte identical to the bswap32 already in common/opencl/rotate_byte.cl (both reverse the 4 bytes of a uint). Since rotate_byte.cl is already prepended to every autolykos_v2 benchmark kernel, I routed all call sites through bswap32 and dropped the duplicate macro from both files.

B2B_G / B2B_* — kept AutolykosV2-specific as you asked; they stay in the kernels.

Also fixed: the AMD benchmark config.json still referenced the pre-rename kernel names (autolykos_v2_lm1 / autolykos_v2_verify), so isKernelEnabled matched nothing and the autolykos table rendered empty. Updated it to autolykos_v2_search_lm0 / autolykos_v2_verify_lm0.

GPU re-verified on an RX 9070 XT (gfx1201): autolykos_v2_search_lm0 ~1.49 GH/s, autolykos_v2_verify_lm0 ~69 MH/s — kernels compile and run unchanged.

yuzi-co mentioned this pull request Jun 11, 2026

Benchmark #89

Closed

6 tasks

yuzi-co force-pushed the bench/amd-autolykos-v2 branch from bdb4b05 to 5d62eb0 Compare June 12, 2026 05:36

luminousmining reviewed Jun 12, 2026

View reviewed changes

Comment thread sources/benchmark/amd/autolykos_v2.cpp

yuzi-co added 3 commits June 12, 2026 12:52

docs(benchmark): document the AMD Autolykos v2 benchmark

03aab41

Add autolykos_v2 to the benchmark source-tree and runAmd() call-graph diagrams, a row in the kernel coverage table, and an AMD RX 9070 XT reference-results section (search ~1.05 GH/s, verify ~69 MH/s).

yuzi-co force-pushed the bench/amd-autolykos-v2 branch from 5d62eb0 to dd65791 Compare June 12, 2026 10:10

luminousmining reviewed Jun 12, 2026

View reviewed changes

luminousmining reviewed Jun 13, 2026

View reviewed changes

yuzi-co added 2 commits June 14, 2026 10:54

luminousmining merged commit 665718d into luminousmining:main Jun 14, 2026
12 checks passed

yuzi-co deleted the bench/amd-autolykos-v2 branch June 19, 2026 12:40



		inline
		void blake2b_compress(ulong h, const ulong m, ulong t, ulong f)

Conversation

yuzi-co commented Jun 11, 2026

What

Changes

Verification

Issue

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

yuzi-co commented Jun 12, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yuzi-co commented Jun 14, 2026

Uh oh!

luminousmining commented Jun 14, 2026

Uh oh!

yuzi-co commented Jun 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants