Skip to content

GL_ARTIFACT_INVESTIGATION.md

Latest commit

 

History

History
80 lines (57 loc) · 4.36 KB

GL_ARTIFACT_INVESTIGATION.md

File metadata and controls

80 lines (57 loc) · 4.36 KB

GL Texture Artifact Investigation

Symptoms

Misplaced geometry, dragged/stretched textures on the OpenGL ES renderer after porting from x1_box.

Fixes Applied (commit 427055fba3)

Fix 1: Clear Surface Bypasses GL State Cache

File: hw/xbox/nv2a/pgraph/gl/draw.c Problem: pgraph_gl_clear_surface() used raw glEnable/glDisable/glScissor calls, leaving the GL state cache stale. Next draw call skipped re-applying state. Fix: Invalidate scissor_test, scissor_rect, dither cache fields after clear.

Fix 2: Active Texture Unit Leak After Display Blit

File: hw/xbox/nv2a/pgraph/gl/display.c Problem: render_display_pvideo_overlay() set glActiveTexture(GL_TEXTURE0 + 1) but restore didn't reset to GL_TEXTURE0. Fix: Added glActiveTexture(GL_TEXTURE0) in restore path.

Fix 3: Missing MAG_FILTER on Display Textures

File: hw/xbox/nv2a/pgraph/gl/display.c Problem: Display buffer and pvideo textures set MIN_FILTER but not MAG_FILTER. Fix: Added glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR).

Fix 4: Full Cache Invalidation After Display Blit (commit 44ce0ee662)

File: hw/xbox/nv2a/pgraph/gl/display.c Problem: Display restore set glDisable(GL_BLEND/SCISSOR/DEPTH) without telling the cache. Fix: Added memset(&r->gl_cache, -1, sizeof(r->gl_cache)) after restore.

Verified NOT Bugs

NEON Loop Variable in texture.c

Lines 235-275: x = width inside for (x=0; x<width; x++). The NEON block processes all remaining pixels, writes them contiguously starting from out, then x = width exits the loop. The increment makes x = width+1, but x < width catches it. Correct.

Scissor Y-Flip

NV2A renders to FBO, not default framebuffer. FBO origin matches NV2A convention. Y-flip only needed at display presentation (done in display.c). Same as upstream xemu. Correct.

Vertex Attribute Buffer Offset

attrib_data_addr used as byte offset into GL_ARRAY_BUFFER, matching update_memory_buffer() upload range. Same as upstream xemu. Correct.

GL State Cache Desync

Both clear-surface (draw.c) and display-blit (display.c) paths now properly invalidate the cache. Confirmed in code review.

Remaining Differences from x1_box (Not Applied)

Material Color Lighting (vsh-ff.c)

x1_box restructured ambient/emission color computation:

  • Current: oD0 = ambient_source; oD0.rgb *= emission; oD0.rgb += emission_source
  • x1_box: oD0 = sceneAmbient * ambient_source; oD0.rgb += emission_source

Different math — would also affect VK. Not safe to apply without testing.

Push Constant/UBO Simplification (vsh.c)

x1_box removed vertex_push_offset and ubo_set handling. Would break VK descriptor set layout. Not applicable.

GLSL Float Literals (vsh-ff.c, vsh-prog.c, vsh.c)

All float literal fixes (1/length → 1.0/length, 2.0f → 2.0, etc.) are ALREADY APPLIED. Verified in code — all GLSL strings use proper float literals.

Potential Architectural Causes (Not Point Fixes)

CPU Format Conversions

GL uses android_surface_guest_to_rgba8() and android_texture_convert_to_rgba8() for BGRA→RGBA on Android. These introduce per-pixel CPU processing not present in VK (which uses native format support). Rounding/precision differences possible but unlikely to cause major geometry issues.

Surface Scaling

surface_copy_expand() / surface_copy_shrink() use scalar CPU code. VK handles scaling differently (GPU-side). Could cause visual differences but not geometry displacement.

Texture-from-Surface Slow Path

When a surface can't be aliased as a texture, GL falls back to CPU readback (glReadPixels) + format conversion + re-upload. This path loses any GPU-side state and may produce different results than VK's direct blit.

GL Renderer Maturity

The GL renderer is fundamentally less mature than VK. It lacks:

  • Surface aliasing optimizations (VK has check_surface_to_texture_compatiblity)
  • Deferred surface download coalescing
  • Multi-frame command buffer pipelining
  • Shader pipeline caching (VK has SPIR-V binary cache)

Next Steps

  1. Compare GL vs VK output on specific games with known artifacts
  2. Add GLSL shader dump logging to capture generated shaders for GL path
  3. Test if the material color lighting change (vsh-ff.c) improves GL without regressing VK
  4. Profile surface download frequency — if GL does more CPU readbacks, that's the bottleneck