diff --git a/NativeAdrenoMetalPlugin.mm b/NativeAdrenoMetalPlugin.mm index 7f39766..d15fa4f 100644 --- a/NativeAdrenoMetalPlugin.mm +++ b/NativeAdrenoMetalPlugin.mm @@ -344,13 +344,14 @@ static id deviceForPort(unsigned port) { } // Turnip chooses each pass's rendering: tile memory (GMEM) or system memory, by its own rules (a pass with a memoryless // attachment or a feedback loop in tile memory, a pass of few draws in system memory, the others by the bandwidth each would - // take), with concurrent binning. A pass in system memory has no tile load and store, which is what a screen-size pass that + // take). A pass in system memory has no tile load and store, which is what a screen-size pass that // draws on a small part of its target otherwise pays for whole. The colour order of GMEM loads and stores is the image's own // swap on the memory side (patches/mesa/turnip-gmem-memory-side-swap.patch). The RM row of the memory log counts both kinds; // tu-debug= (for instance gmem,nocb: every pass in tile memory, no concurrent binning) overrides this. // Concurrent binning is a Turnip option that is off unless asked for (tu_allow_concurrent_binning, read from the environment - // by Mesa's option parser): with it the BV thread bins a pass while BR still renders the one before. nocb in tu-debug= turns it off. - setenv("tu_allow_concurrent_binning", "true", 0); + // by Mesa's option parser): with it the BV thread bins a pass while BR still renders the one before. tu-cb=1 turns it on. + // It stays at Mesa's default, off: what it gains here was never measured. + if (bootArgSet("tu-cb=1")) setenv("tu_allow_concurrent_binning", "true", 0); // diag=1 turns on every diagnostic of the GPU stack in this process: the frontend's frame trace, Turnip's pass stamps and // submission ring, the runtime's stage ring, and in the render server the status and telemetry lines. Without it none of them runs. if (bootArgSet("diag=1") || bootArgSet("gpu-trace=1")) setenv("SHENG_DIAG", "1", 1);