gpu_neon: rework buffer selection

[pcsx_rearmed.git] / plugins / gpu_neon / psx_gpu_if.c

diff --git a/plugins/gpu_neon/psx_gpu_if.c b/plugins/gpu_neon/psx_gpu_if.c
index 30faee2..69a2a1b 100644 (file)
--- a/plugins/gpu_neon/psx_gpu_if.c
+++ b/plugins/gpu_neon/psx_gpu_if.c
@@ -9,8 +9,11 @@
  */
 
 #include <stdio.h>
+#include <assert.h>
 #include <sys/mman.h>
 
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+
 extern const unsigned char cmd_lengths[256];
 #define command_lengths cmd_lengths
 
@@ -21,9 +24,12 @@ static int initialized;
 #define SET_Ex(r, v) \
   ex_regs[r] = v
 
+static __attribute__((noinline)) void
+sync_enhancement_buffers(int x, int y, int w, int h);
+
+#include "../gpulib/gpu.h"
 #include "psx_gpu/psx_gpu.c"
 #include "psx_gpu/psx_gpu_parse.c"
-#include "../gpulib/gpu.h"
 
 static psx_gpu_struct egpu __attribute__((aligned(256)));
 
@@ -110,8 +116,12 @@ static __attribute__((noinline)) void
 sync_enhancement_buffers(int x, int y, int w, int h)
 {
   const int step_x = 1024 / sizeof(egpu.enhancement_buf_by_x16);
+  int hres = egpu.saved_hres;
+  int x_buf, w1, s, fb_index;
   u16 *src, *dst;
-  int w1, fb_index;
+
+  if (egpu.enhancement_buf_ptr == NULL)
+    return;
 
   w += x & (step_x - 1);
   x &= ~(step_x - 1);
@@ -119,18 +129,29 @@ sync_enhancement_buffers(int x, int y, int w, int h)
   if (y + h > 512)
     h = 512 - y;
 
+  // find x_buf which is an offset into this enhancement_buf
+  fb_index = egpu.enhancement_buf_by_x16[x / step_x];
+  x_buf = x - egpu.enhancement_buf_start[fb_index];
+
   while (w > 0) {
     fb_index = egpu.enhancement_buf_by_x16[x / step_x];
-    for (w1 = 0; w > 0; w1++, w -= step_x)
+    for (w1 = 0; w > 0 && x_buf < hres; x_buf += step_x, w1++, w -= step_x)
       if (fb_index != egpu.enhancement_buf_by_x16[x / step_x + w1])
         break;
+    // skip further unneeded data, if any
+    for (s = 0; w > 0; s++, w -= step_x)
+      if (fb_index != egpu.enhancement_buf_by_x16[x / step_x + w1 + s])
+        break;
 
-    src = gpu.vram + y * 1024 + x;
-    dst = select_enhancement_buf_ptr(&egpu, x);
-    dst += (y * 1024 + x) * 2;
-    scale2x_tiles8(dst, src, w1 * step_x / 8, h);
+    if (w1 > 0) {
+      src = gpu.vram + y * 1024 + x;
+      dst = select_enhancement_buf_ptr(&egpu, x);
+      dst += (y * 1024 + x) * 2;
+      scale2x_tiles8(dst, src, w1 * step_x / 8, h);
+    }
 
-    x += w1 * step_x;
+    x += (w1 + s) * step_x;
+    x_buf = 0;
   }
 }
 
@@ -139,11 +160,18 @@ void renderer_sync_ecmds(uint32_t *ecmds)
   gpu_parse(&egpu, ecmds + 1, 6 * 4, NULL);
 }
 
-void renderer_update_caches(int x, int y, int w, int h)
+void renderer_update_caches(int x, int y, int w, int h, int state_changed)
 {
   update_texture_cache_region(&egpu, x, y, x + w - 1, y + h - 1);
-  if (gpu.state.enhancement_active && !(gpu.status & PSX_GPU_STATUS_RGB24))
+
+  if (gpu.state.enhancement_active) {
+    if (state_changed) {
+      egpu.saved_hres = 0;
+      renderer_notify_res_change();
+      return;
+    }
     sync_enhancement_buffers(x, y, w, h);
+  }
 }
 
 void renderer_flush_queues(void)
@@ -162,11 +190,44 @@ void renderer_set_interlace(int enable, int is_odd)
 
 void renderer_notify_res_change(void)
 {
-  // note: must keep it multiple of 8
-  if (egpu.enhancement_x_threshold != gpu.screen.hres)
+  renderer_notify_scanout_x_change(gpu.screen.src_x, gpu.screen.hres);
+}
+
+void renderer_notify_scanout_x_change(int x, int w)
+{
+  int hres = (w + 15) & ~15;
+  int max_bufs = ARRAY_SIZE(egpu.enhancement_scanout_x);
+  int need_update = 0;
+  int i;
+
+  if (!gpu.state.enhancement_active)
+    return;
+
+  assert(!(max_bufs & (max_bufs - 1)));
+  if (egpu.saved_hres != hres) {
+    for (i = 0; i < max_bufs; i++)
+      egpu.enhancement_scanout_x[i] = x;
+    need_update = 1;
+  }
+
+  if (egpu.enhancement_scanout_x[egpu.enhancement_scanout_select] != x)
   {
-    egpu.enhancement_x_threshold = gpu.screen.hres;
+    // maybe triple buffering?
+    for (i = 0; i < max_bufs; i++)
+      if (egpu.enhancement_scanout_x[i] == x)
+        break;
+    if (i == max_bufs)
+      need_update = 1;
+
+    egpu.enhancement_scanout_x[egpu.enhancement_scanout_select] = x;
+  }
+  egpu.enhancement_scanout_select++;
+  egpu.enhancement_scanout_select &= max_bufs - 1;
+  if (need_update)
+  {
+    egpu.saved_hres = hres;
     update_enhancement_buf_table_from_hres(&egpu);
+    sync_enhancement_buffers(0, 0, 1024, 512);
   }
 }
 
@@ -174,23 +235,18 @@ void renderer_notify_res_change(void)
 
 void renderer_set_config(const struct rearmed_cbs *cbs)
 {
-  static int enhancement_was_on;
-
-  disable_main_render = cbs->gpu_neon.enhancement_no_main;
-  if (egpu.enhancement_buf_ptr != NULL && cbs->gpu_neon.enhancement_enable
-      && !enhancement_was_on)
-  {
-    sync_enhancement_buffers(0, 0, 1024, 512);
-  }
-  enhancement_was_on = cbs->gpu_neon.enhancement_enable;
-
   if (!initialized) {
     initialize_psx_gpu(&egpu, gpu.vram);
     initialized = 1;
   }
-
-  if (gpu.mmap != NULL && egpu.enhancement_buf_ptr == NULL)
-    map_enhancement_buffer();
   if (cbs->pl_set_gpu_caps)
     cbs->pl_set_gpu_caps(GPU_CAP_SUPPORTS_2X);
+
+  disable_main_render = cbs->gpu_neon.enhancement_no_main;
+  if (gpu.state.enhancement_enable) {
+    if (gpu.mmap != NULL && egpu.enhancement_buf_ptr == NULL)
+      map_enhancement_buffer();
+  }
 }
+
+// vim:ts=2:sw=2:expandtab