SDWA source extraction and destination insertion are independent. Cover the reported word1-to-word0 preserve form and byte-source partial destinations in compute tests. Verify 64 distinct compute buffers survive resource tracking and binding layout, and remove the obsolete CFG rejection.
(cherry picked from commit 9e74df7dc0)
* shader: support V_CMPX_EQ_U16
Decode gfx11 VOPC 0xba and lower it as an unsigned 16-bit equality
compare with CMPX EXEC-mask semantics.
🤖 Generated with Codebuff 🤖
Co-Authored-By: Codebuff <noreply@codebuff.com>
* shader: correct CMPX EQ U16 runtime expectations
Remove the added CFG tests and correct two EXEC expectations in the compute regression case.
---------
Co-authored-by: Matthew Hurley <171001439+theantipopau@users.noreply.github.com>
Co-authored-by: Codebuff <noreply@codebuff.com>
Co-authored-by: nmzik <Nmzik@mail.ru>
* graphics: hoist loop-invariant checks out of draw/attrib hot paths
DrawIndirectMulti recomputed the index_type_and_size switch on every
sub-draw of a multi-draw-indirect call, and ShaderApplyAttribSemantics
re-checked the debug-dump config flag on every vertex attribute. Both
values are invariant across their respective loops, so compute them
once before iterating instead.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* graphics: cache vertex buffer descriptor size in AcquireVertexBuffers
VertexBufferDescriptorSize was computed twice per vertex-buffer slot
on every draw call: once while collecting guest ranges, and again
while rebuilding slot bindings. Both calls use identical inputs, so
cache the first pass's result and reuse it in the second.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Preserve live lane predicates through RDNA2 whole-quad expansion
Keep the source-live predicate when WQM expands a scalar mask, so entry masks remain known live without an instruction pre-scan. Preserve scalar mask values, SCC, half writes, and raw-source behavior.
Cover entry pixel masks, cross-half copies, and scalar/high-word sources with existing shader regression suites.
Co-authored-by: nmzik <Nmzik@mail.ru>
Co-authored-by: Douglas <douglasproducoes21@gmail.com>
Decode GFX10 DS opcode 0xa0 as DS_WRITE_B8_D16_HI (next to
DS_WRITE_B16_D16_HI at 0xa1) and lower it through the existing 8-bit
shared-memory write. The stored byte is bits 23:16 of the data VGPR,
selected with SDWA BYTE_2 on src1.
Add captured-instruction decode coverage (from GTA V, PPSA04264) and a
runtime regression that writes one byte of a distinctive source value
into a pre-filled LDS dword.
nmzik