Repository navigation
D3D12: analytic-filter ports of the fused output kernels (post_w64.co), byte for byte against HIP - #145
Merged
Conversation
…), byte for byte against HIP post.hlsl WG_LZ_ANALYTIC: w64/post_tf_f32_p12_diet_vsit and w64/post_tf4_dl_f32_p12_diet_vsit of post_w64.co (the build the runtime loads when the model file has no output-filter weights, e.g. Model L): FP32 Lanczos constants, FP32 values between the passes, no weight arguments. The original's DXILs are unchanged. post_tf_d3d12_check.py: these ports against HIP post_w64.co with l_d3d12_ops_check.py's engine (raw bytes, sentinels, slack, negative control), the original port against post_w64__orig.co as the Model M control; the harness takes a kinds table and a HIP block. Refs #89 Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
…ed build) The analytic-filter ports declare the same kernel names as the original filter's with co=post_w64.co; every declaration line is still checked for the 256 lanes and the cap. Refs #89 Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
phantoslayer21
marked this pull request as ready for review
November 14, 2026 04:44
phantoslayer21
added a commit
that referenced
this pull request
Nov 14, 2026
…well The direct-resource route required the "original" output-filter build of the wave64 fused post (post_w64__orig.co) with the model file's filter weights. With the analytic-filter ports on main (#145, #146; keyed by name and code object, #143) it also takes post_w64.co, the build a model file without those weights uses, as Model L's: the check now refuses only quality5 / quality9, a post outside wave64, or the original filter without weights. test_l_runtime_frame.py: the D3D12 tests use the declared ports as they are (both filter builds of Model M's fused post are declared now); a test that the analytic build of the fused pack12 post is among them. Refs #89 Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
phantoslayer21
added a commit
that referenced
this pull request
Nov 14, 2026
…well The direct-resource route required the "original" output-filter build of the wave64 fused post (post_w64__orig.co) with the model file's filter weights. With the analytic-filter ports on main (#145, #146; keyed by name and code object, #143) it also takes post_w64.co, the build a model file without those weights uses, as Model L's: the check now refuses only quality5 / quality9, a post outside wave64, or the original filter without weights. test_l_runtime_frame.py: the D3D12 tests use the declared ports as they are (both filter builds of Model M's fused post are declared now); a test that the analytic build of the fused pack12 post is among them. Refs #89 Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
phantoslayer21
added a commit
that referenced
this pull request
Nov 14, 2026
…well The direct-resource route required the "original" output-filter build of the wave64 fused post (post_w64__orig.co) with the model file's filter weights. With the analytic-filter ports on main (#145, #146; keyed by name and code object, #143) it also takes post_w64.co, the build a model file without those weights uses, as Model L's: the check now refuses only quality5 / quality9, a post outside wave64, or the original filter without weights. test_l_runtime_frame.py: the D3D12 tests use the declared ports as they are (both filter builds of Model M's fused post are declared now); a test that the analytic build of the fused pack12 post is among them. Refs #89 Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Refs #89
#143 (ports keyed by name and code object) has merged; this branch is current with main.
Model L's model file has no output-filter weights, so the runtime loads
post_w64.co(the analytic filter) underw64. D3D12 had ports only of the original filter's build (post_w64__orig.co). This PR adds the analytic build's ports of the fused output kernels, byte for byte against HIPpost_w64.co.What
clean/dx12/kernels/post.hlsl,WG_LZ_ANALYTIC = 1:w64/post_tf_f32_p12_diet_vsitandw64/post_tf4_dl_f32_p12_diet_vsitwithco=post_w64.co; the same names as the original's ports, and the device runs the port of the loaded build;_directforms are not in this PR (see Not tested).clean/tests/test_post_tf_geometry.py: the tf4 declarations are compared as a set of names (a name is declared once per build it ports); every declaration line is still checked for 256 lanes and the cap.clean/dx12/tests/post_tf_d3d12_check.py(new): the check below, onclean/l/runner/l_d3d12_ops_check.py's engine. That engine now takes a kinds table and a HIP block size: the tf4 port has no HIP kernel of its own.GPU record
post_tf_d3d12_check.py --run --seeds 3, one GPU slot, this branch's DXILs,kernel_replay.exe, andpost_w64.co/post_w64__orig.cobuilt asclean/runner/build.pybuilds them.w64/post_tf_f32_p12_diet_vsit(analytic) vs HIPpost_tf_f32_p12_diet_vsit(post_w64.co)w64/post_tf4_dl_f32_p12_diet_vsit(analytic) vs HIPpost_tf_f32_p12_diet_vsit/post_tf_dl_f32_p12w64/post_tf_f32_p12_diet_vsit(original) vspost_w64__orig.cow64, and the device picks the port by it, so the run also exercises D3D12: ports keyed by (name, code object), so one kernel name can have ports of several builds #143's selection.w64_post_tf_f32_p12_diet_vsit_s0_4again on D3D12 with half its workgroups. Caught, 8,899,584 differing bytes.post.hlsl(w64/post_tf_*,pmo/post_tf*_mo*, the_tex,_directand tf4 forms) plusnative_fill,exposure_auto_dietandl_coef, built from main (cc0a6df) and from this branch: SHA-256 identical, 13 of 13. A manifest of this branch holds these entries unchanged plus the two analytic entries.Tested
Not tested
_directforms of the analytic ports (the game route): they follow in their own PR. The direct form's output table (FP16-indexed) does not apply to an FP32 filter result.post_w64.co.