Repository navigation
Conversation
github-actions
Bot
added
testing
|
Device: Arc Pro B70 (oneAPI 2026.1.1) I checked the bound algebra against the old continue filter: for s0 ≥ 1, [i_begin, i_end) selects precisely the same contributing inputs in the same ascending order, so accumulation order is unchanged - which is why byte-identical output is expected, not just close. Perf: your table reproduces. Op-level test-backend-ops perf matches within noise on all four shapes; end-to-end on the 4-layer chain at your 24-frame point I measure 29.33 s (base) → 32.4 ms (patch), vs your 28.64 s → 30.7 ms. CONV_TRANSPOSE_1D suite 116/116 on base and patch; the rest of my 110-op suite is unchanged vs base. One small ask: nothing upstream asserts s0 > 0 — ggml_conv_transpose_1d() checks p0 == 0 and d0 == 1 only. The old kernel happened to tolerate degenerate s0; the new one divides by it, so a graph built with s0 ≤ 0 is now UB. A one-line GGML_ASSERT(s0 > 0) (in ggml_sycl_op_conv_transpose_1d, or in ggml_conv_transpose_1d() so every backend benefits) would close that. Non-blocking - no in-tree caller uses s0 < 1. |
| test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {200000, 1, 1, 1})); | ||
| test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {200000, 16, 1, 1})); | ||
|
|
||
| // audio codec decoder upsampling stack (DAC-style, 4 latent frames in) |
There was a problem hiding this comment.
What does "stack" mean in this case?
There was a problem hiding this comment.
The drama 3 audio decoder stack, nothing of mine.
There was a problem hiding this comment.
This is not clear I think. Maybe put more context in the comment as to what the test data is from. The comment doesn't make much sense in its current form.
arthw
left a comment
arthw
left a comment
There was a problem hiding this comment.
It's good job!
I tested on B60.
Related UT cases are passed and the performance increased +170%.
Thank you!
|
@anantshri Please do not copy-paste post AI slop in PR comments. It is quite rude and also goes against our Contribution Guidelines https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy |
Used angelX to optimize SYCL/drama 3 coordination, results are as follows Each output reads only inputs with i*s0 <= idx < i*s0 + K, in the same ascending order as before. Arc Pro B70, oneAPI 2026.1, test-backend-ops perf, base 9b56ccf vs patch, median of 3 same-boot A/B/A rounds, A/A 1.000-1.001: | input [ne0, ne1] | kernel [K, OC, IC] | s0 | base us | patch us | speedup | |---|---|---|---|---|---| | [16, 1536] | [16, 768, 1536] | 8 | 3910.57 | 1348.59 | 2.90x | | [128, 768] | [16, 384, 768] | 8 | 37735.00 | 1665.26 | 22.66x | | [1024, 384] | [8, 192, 384] | 4 | 270371.75 | 1471.74 | 183.71x | | [4096, 192] | [4, 96, 192] | 2 | 534614.63 | 752.74 | 710.23x | Shapes are the drama 3 codec decoder upsampling stack, added as perf cases. Chained 4-layer stack at 1/4/24 input frames: base 66.3/847.4/28638.5 ms, patch 2.3/6.7/30.7 ms. Kernel numbers, not end-to-end TTS latency. Correctness: test-backend-ops -o CONV_TRANSPOSE_1D 116/116 on base and patch, and 116/116 rebased on 3f03b17. Seeded output dump (7.46 MB) byte-identical to base. Full test-backend-ops aborts on base and patch alike at the non-F32 SWIGLU_OAI assert (element_wise.cpp:1187, see ggml-org#30857). With only those cases excluded: 20145/20173 on base and patch, and every patch-run failure also fails on base.
bd3ec67 to
2a2e0fa
Compare
| test_cases.emplace_back(new test_conv_transpose_1d({16, 1536, 1, 1}, {16, 768, 1536, 1}, 8, 0, 1)); | ||
| test_cases.emplace_back(new test_conv_transpose_1d({128, 768, 1, 1}, {16, 384, 768, 1}, 8, 0, 1)); | ||
| test_cases.emplace_back(new test_conv_transpose_1d({1024, 384, 1, 1}, {8, 192, 384, 1}, 4, 0, 1)); | ||
| test_cases.emplace_back(new test_conv_transpose_1d({4096, 192, 1, 1}, {4, 96, 192, 1}, 2, 0, 1)); |
Overview
Used angelX to optimize SYCL/drama 3 coordination, results are as follows
Each output reads only inputs with is0 <= idx < is0 + K, in the same ascending order as before.
Arc Pro B70, oneAPI 2026.1, test-backend-ops perf, base 9b56ccf vs patch, median of 3 same-boot A/B/A rounds, A/A 1.000-1.001:
Shapes are the drama 3 codec decoder upsampling stack, added as perf cases. Chained 4-layer stack at 1/4/24 input frames: base 66.3/847.4/28638.5 ms, patch 2.3/6.7/30.7 ms. Kernel numbers, not end-to-end TTS latency.
Correctness: test-backend-ops -o CONV_TRANSPOSE_1D 116/116 on base and patch, and 116/116 rebased on 3f03b17. Seeded output dump (7.46 MB) byte-identical to base.
Full test-backend-ops aborts on base and patch alike at the non-F32 SWIGLU_OAI assert (element_wise.cpp:1187, see #30857). With only those cases excluded: 20145/20173 on base and patch, and every patch-run failure also fails on base.
Requirements