Elide identity VRAM copies without draining deferred scenes

runtime/include/gpu_gl_renderer.h

@@ -571,7 +571,7 @@ int gl_renderer_perf_aggregate(int wide_filter, double out[18]);
 /* Bounded benchmark attribution. CPU spans require PSX_GL_CPU_PERF=1;
  * reading drains the render queue, so read outside the measured interval. */
 typedef struct {
-    uint64_t counts[9];
+    uint64_t counts[10];
     uint64_t calls[8], ticks[8], max_ticks[8], frequency;
     uint64_t prior_ticks[8], prior_max_ticks[8];
 } GlRenderWorkStats;

runtime/src/gpu_gl_renderer.c

@@ -5410,8 +5410,28 @@ static void gpu_fill(int x,int y,int w,int h,uint16_t c) {
 /* VRAM->VRAM copy: blit the source region to the scratch texture (resolves
  * overlap), then draw it back at the destination through the BLIT program so
  * mask set/check and the stencil mirror apply, exactly like sw_copy_rect. */
+static uint64_t s_identity_copies;
+static int identity_copy_fast(void) {
+    static int enabled = -1;
+    if (enabled < 0) {
+        const char *e = getenv("PSX_GL_IDENTITY_COPY_FAST");
+        enabled = !(e && e[0] == '0');
+    }
+    return enabled;
+}
 static void gpu_copy_rect(int sx,int sy,int dx,int dy,int w,int h) {
     if (w <= 0 || h <= 0) return;
+    if (identity_copy_fast() && sx == dx && sy == dy && !s_mask_set &&
+        sx >= 0 && sy >= 0 && sx < VRAM_W && sy < VRAM_H) {
+        /* No pixel can change. Retain the existing primitive batch boundary
+         * (PGXP's two-pass depth policy is batch-sensitive), but don't drain
+         * deferred high-resolution scenes or stage identical pixels.
+         * The guest GPU command and its scheduled device timing still run. */
+        flush_flat_batch();
+        flush_tex_batch();
+        s_identity_copies++;
+        return;
+    }
     if (gl_cpu_mode()) {
         static int seen[32][6], used;
         int key[6] = {sx,sy,dx,dy,w,h}, i;
@@ -6903,6 +6923,7 @@ void gl_renderer_work_stats(GlRenderWorkStats *out) {
     out->counts[4] = s_hq_center_pruned;
     out->counts[5] = s_tc_hits; out->counts[6] = s_tc_misses; out->counts[7] = s_tc_evictions;
     out->counts[8] = s_hiw_copy_blits;
+    out->counts[9] = s_identity_copies;
 }
 void gl_renderer_shutdown(void) {
     gl_renderer_render_thread_stop();   /* drains; the context comes back to this thread */

runtime/src/main.cpp

@@ -5081,8 +5081,8 @@ static void input_replay_frontend_tick(uint64_t frame) {
             f = std::fopen(path.c_str(), "w");
             if (!f) throw std::runtime_error("cannot write input replay render report");
             std::fputs("stat,count,total_us,max_us\n", f);
-            const char *counts[9] = {"queue_uploads", "replay_draws", "bloom_fused", "coverage_depth", "centre_pruned", "texel_cache_hits", "texel_cache_misses", "texel_cache_evictions", "hiw_copy_blits"};
-            for (int i=0; i<9; ++i) std::fprintf(f, "%s,%llu,0,0\n", counts[i], (unsigned long long)stats.counts[i]);
+            const char *counts[10] = {"queue_uploads", "replay_draws", "bloom_fused", "coverage_depth", "centre_pruned", "texel_cache_hits", "texel_cache_misses", "texel_cache_evictions", "hiw_copy_blits", "identity_copies"};
+            for (int i=0; i<10; ++i) std::fprintf(f, "%s,%llu,0,0\n", counts[i], (unsigned long long)stats.counts[i]);
             const char *spans[8] = {"depth_restore", "queue_replay", "vram_copy", "interp_capture", "interp_interval", "centre_copy", "depth_gpu_sample", "colour_gpu_sample"};
             double us = stats.frequency ? 1e6 / stats.frequency : 0;
             for (int i=0; i<8; ++i) std::fprintf(f, "%s,%llu,%.3f,%.3f\n", spans[i],

runtime/tests/run_gl_render_thread.py

@@ -157,7 +157,7 @@ def main():
                 ok = False
             runs[threaded] = p
             if args.wide_present:
-                for margin, coverage, state, upload, bloom, cache, layers, copy in ((0,0,0,0,0,0,256,1), (0,0,1,1,1,1,256,1), (0,1,1,1,1,1,256,1), (1,0,1,1,1,1,256,1), (1,1,1,0,1,1,256,1), (1,1,1,1,0,1,256,1), (1,1,1,1,1,0,256,1), (1,1,1,1,1,1,4,1), (1,1,1,1,1,1,256,0)):
+                for margin, coverage, state, upload, bloom, cache, layers, copy, identity in ((0,0,0,0,0,0,256,1,1), (0,0,1,1,1,1,256,1,1), (0,1,1,1,1,1,256,1,1), (1,0,1,1,1,1,256,1,1), (1,1,1,0,1,1,256,1,1), (1,1,1,1,0,1,256,1,1), (1,1,1,1,1,0,256,1,1), (1,1,1,1,1,1,4,1,1), (1,1,1,1,1,1,256,0,1), (1,1,1,1,1,1,256,1,0)):
                     reference = run([probe, s, threaded, args.frames, *mode],
                                     {"PSX_GL_WIDE_MARGIN_SCISSOR": str(margin),
                                      "PSX_GL_COVERAGE_FAST": str(coverage),
@@ -166,7 +166,8 @@ def main():
                                      "PSX_GL_BLOOM_FUSE_EXTEND": str(bloom),
                                      "PSX_GL_TEXEL_CACHE": str(cache),
                                      "PSX_GL_TEXEL_CACHE_LAYERS": str(layers),
-                                     "PSX_GL_HIW_COPY_FAST": str(copy)})
+                                     "PSX_GL_HIW_COPY_FAST": str(copy),
+                                     "PSX_GL_IDENTITY_COPY_FAST": str(identity)})
                     baseline = parse(reference.stdout)
                     if reference.returncode or baseline["failures"] != 0 or any(baseline[k] != p[k] for k in KEYS):
                         print(f"FAIL scale {s}: margin={margin} coverage={coverage} changed rendered output", baseline, p)

runtime/tests/test_gl_render_thread.c

@@ -247,6 +247,10 @@ static void frame(int f, int buf_x) {
     gr_draw_flat_rect(576, 64, 32, 32, (uint16_t)(0x03e0 + f));
     gr_draw_textured_rect(buf_x + 100, 200, 32, 32, 0, 64, 512, 256, 0x0009);
     gr_copy_rect(buf_x + 10, 10, buf_x + 14, 13, 80, 50);
+    /* Identity copies preserve colour/mask unless set-mask adds bit15. */
+    mask_bits(f % 4 == 2, 0);
+    gr_copy_rect(buf_x + 14, 13, buf_x + 14, 13, 80, 50);
+    mask_bits(0, 0);
     if (quality_scene && f % 3 == 1) {
         /* Consume the copied bit15 mask before any later framebuffer clear. */
         mask_bits(0, 1);