mirror of
https://github.com/tangerineturtle1/NX2navi.git
synced 2026-11-19 02:09:57 -04:00
The task records were a fixed 4096 per submission (3 KiB each): a dense patch of grass that launches more task workgroups in one submission lost the rest silently. Their count is now a push constant (NX2NAVI_MESH_EMU_TASKS, default 16384) and the arena defaults to 256 MiB. The header counts what the stages asked for past the capacities too; before the counters start over they go to a readback ring, and the perf line reports the most one submission used of the vertex slots, indices and task records, and the mesh workgroups that did not fit. The first draws the emulation skips are logged with the reason. (cherry picked from commit 81cfd91bf468f13e4debf287d080f96637a4c382) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
42 lines
2.8 KiB
C
42 lines
2.8 KiB
C
#pragma once
|
|
// Mesh shader emulation for GPUs without VK_EXT_mesh_shader (NX2NAVI_MESH_EMULATION, vk_pipeline.c / vk_exec_draw.c):
|
|
// the translator's TaskEXT and MeshEXT modules (tools/xlate/MeshSpirv.cs) become GLCompute modules that write their
|
|
// vertices, triangles and task payloads into one GPU buffer (the arena), and a generated vertex shader reads the
|
|
// vertices back for the game's fragment shader. Plain C over SPIR-V words; no Vulkan types.
|
|
#include <stddef.h>
|
|
#include <stdint.h>
|
|
|
|
// The arena: one buffer, its address in the push constants of every emulation stage.
|
|
// MESH_EMU_DOFF: draw records, MESH_EMU_DWORDS words each: DrawIndexedIndirect {indexCount, instanceCount,
|
|
// MESH_EMU_TOFF: task records, MESH_EMU_TWORDS words each: mesh workgroup count, then the payload
|
|
// index area (uint32 indices), vertex area (16-byte slots: position, then each generic output) at offsets the
|
|
// replay passes in the push constants.
|
|
#define MESH_EMU_DOFF 256u
|
|
#define MESH_EMU_DWORDS 16u
|
|
#define MESH_EMU_DRECS 4096u
|
|
#define MESH_EMU_TOFF (MESH_EMU_DOFF + MESH_EMU_DRECS * MESH_EMU_DWORDS * 4u)
|
|
// task records: as many as the push constants say (the replay sizes them, NX2NAVI_MESH_EMU_TASKS); the index area
|
|
// follows them
|
|
// Push constants of every emulation stage (compute and the vertex shader), 32 bytes:
|
|
// u64 arena address; uint draw record (word index); uint index area (word index); uint index capacity;
|
|
// uint vertex area (16-byte slot index); uint vertex capacity (slots); uint task record capacity
|
|
// Header words 0-2 count what the stages asked for, beyond the capacities too (the replay logs the most per
|
|
// submission); word 3 counts mesh workgroups that did not fit.
|
|
#define MESH_EMU_PC_BYTES 32u
|
|
|
|
typedef struct MeshEmuInfo {
|
|
int task; // 1 TaskEXT, 0 MeshEXT
|
|
uint32_t threads; // LocalSize x
|
|
uint32_t max_v, max_p; // mesh: OutputVertices, OutputPrimitivesEXT
|
|
uint32_t payload_words; // the task payload's size (mesh: what it reads, 0 without a task stage)
|
|
uint32_t max_mesh; // task: the most mesh workgroups one task workgroup launches (dispatch width)
|
|
uint32_t slots; // mesh: 16-byte slots per vertex: position and each generic output
|
|
uint32_t n_generic; // mesh: generic outputs, in slot order 1..n
|
|
uint8_t generic_loc[32]; // their locations
|
|
} MeshEmuInfo;
|
|
|
|
// The compute module for a TaskEXT or MeshEXT module (malloc'd, *out_n words); NULL with why set when the module is not
|
|
// one the converter knows (it then stays unemulated: its draws are skipped as before).
|
|
uint32_t *mesh_emu_compute(const uint32_t *w, size_t n, MeshEmuInfo *info, size_t *out_n, const char **why);
|
|
|
|
// The vertex shader of an emulated mesh pipeline: position and the generic outputs from the vertex slots, as types[l]
|