From: notaz <notasas@gmail.com>
Date: Sun, 20 Aug 2023 19:38:03 +0000 (+0300)
Subject: gpu_neon: rework buffer selection
X-Git-Tag: r24l~230
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=3b7b00650a53bc493c7da7d69ad54e1a25111ebf;p=pcsx_rearmed.git

gpu_neon: rework buffer selection

to fix MGS codec and maybe more
---

diff --git a/plugins/dfxvideo/gpulib_if.c b/plugins/dfxvideo/gpulib_if.c
index e883b86d..ee1f68a4 100644
--- a/plugins/dfxvideo/gpulib_if.c
+++ b/plugins/dfxvideo/gpulib_if.c
@@ -315,6 +315,10 @@ void renderer_notify_res_change(void)
 {
 }
 
+void renderer_notify_scanout_x_change(int x, int w)
+{
+}
+
 extern const unsigned char cmd_lengths[256];
 
 int do_cmd_list(uint32_t *list, int list_len, int *last_cmd)
@@ -430,7 +434,7 @@ void renderer_sync_ecmds(uint32_t *ecmds_)
   cmdSTP((unsigned char *)&ecmds[6]);
 }
 
-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)
 {
 }
 
diff --git a/plugins/gpu-gles/gpulib_if.c b/plugins/gpu-gles/gpulib_if.c
index 6ffd42eb..24488e9e 100644
--- a/plugins/gpu-gles/gpulib_if.c
+++ b/plugins/gpu-gles/gpulib_if.c
@@ -514,6 +514,10 @@ void renderer_notify_res_change(void)
 {
 }
 
+void renderer_notify_scanout_x_change(int x, int w)
+{
+}
+
 extern const unsigned char cmd_lengths[256];
 
 // XXX: mostly dupe code from soft peops
@@ -622,7 +626,7 @@ void renderer_sync_ecmds(uint32_t *ecmds)
   cmdSTP((unsigned char *)&ecmds[6]);
 }
 
-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)
 {
  VRAMWrite.x = x;
  VRAMWrite.y = y;
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.c b/plugins/gpu_neon/psx_gpu/psx_gpu.c
index 85cf89fa..e252d04e 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c
@@ -4963,7 +4963,7 @@ void initialize_psx_gpu(psx_gpu_struct *psx_gpu, u16 *vram)
 
   psx_gpu->primitive_type = PRIMITIVE_TYPE_UNKNOWN;
 
-  psx_gpu->enhancement_x_threshold = 256;
+  psx_gpu->saved_hres = 256;
 }
 
 u64 get_us(void)
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.h b/plugins/gpu_neon/psx_gpu/psx_gpu.h
index e365b8a2..88e40ac4 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.h
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.h
@@ -185,18 +185,22 @@ typedef struct
   u32 *reciprocal_table_ptr;
 
   // enhancement stuff
-  u16 *enhancement_buf_ptr;
-  u16 *enhancement_current_buf_ptr;
-  u32 enhancement_x_threshold;
+  u16 *enhancement_buf_ptr;          // main alloc
+  u16 *enhancement_current_buf_ptr;  // offset into above, 4 bufs
+  u32 saved_hres;
   s16 saved_viewport_start_x;
   s16 saved_viewport_start_y;
   s16 saved_viewport_end_x;
   s16 saved_viewport_end_y;
-  u8 enhancement_buf_by_x16[64];
+  u8  enhancement_buf_by_x16[64];    // 0-3 specifying which buf
+  u16 enhancement_buf_start[4];      // x pos where buf[n] begins
+
+  u16 enhancement_scanout_x[4];
+  u16 enhancement_scanout_select;
 
   // Align up to 64 byte boundary to keep the upcoming buffers cache line
   // aligned, also make reachable with single immediate addition
-  u8 reserved_a[160];
+  u8 reserved_a[142];
 
   // 8KB
   block_struct blocks[MAX_BLOCKS_PER_ROW];
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c b/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
index 942b3d30..d7ec3409 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
@@ -1,6 +1,10 @@
+#define select_enhancement_buf_index(psx_gpu, x) \
+  ((psx_gpu)->enhancement_buf_by_x16[(u32)(x) / \
+    (1024u / sizeof((psx_gpu)->enhancement_buf_by_x16))])
+
 #define select_enhancement_buf_ptr(psx_gpu, x) \
   ((psx_gpu)->enhancement_buf_ptr + \
-   ((psx_gpu)->enhancement_buf_by_x16[(x) / 16] << 20))
+    (select_enhancement_buf_index(psx_gpu, x) << 20))
 
 #if !defined(NEON_BUILD) || defined(SIMD_BUILD)
 
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c b/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
index 37622580..32c32fdb 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
@@ -250,10 +250,9 @@ static void do_fill(psx_gpu_struct *psx_gpu, u32 x, u32 y,
 #define SET_Ex(r, v)
 #endif
 
-vertex_struct vertexes[4] __attribute__((aligned(32)));
-
 u32 gpu_parse(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_command)
 {
+  vertex_struct vertexes[4] __attribute__((aligned(16))) = {};
   u32 current_command = 0, command_length;
 
   u32 *list_start = list;
@@ -790,26 +789,61 @@ breakloop:
 
 #define ENH_BUF_TABLE_STEP (1024 / sizeof(psx_gpu->enhancement_buf_by_x16))
 
+static int is_new_scanout(psx_gpu_struct *psx_gpu, int x)
+{
+  int i, scanout_x;
+  for (i = 0; i < ARRAY_SIZE(psx_gpu->enhancement_scanout_x); i++)
+  {
+    scanout_x = psx_gpu->enhancement_scanout_x[i];
+    if (x <= scanout_x && scanout_x < x + ENH_BUF_TABLE_STEP)
+    {
+      if (x != scanout_x)
+        log_anomaly("unaligned scanout x: %d,%d\n", scanout_x, x);
+      return 1;
+    }
+  }
+  return 0;
+}
+
 static void update_enhancement_buf_table_from_hres(psx_gpu_struct *psx_gpu)
 {
-  u32 b, x, s;
+  u32 b, x;
 
   b = 0;
-  s = psx_gpu->enhancement_x_threshold;
-  for (x = 0; x < sizeof(psx_gpu->enhancement_buf_by_x16); x++)
+  psx_gpu->enhancement_buf_by_x16[0] = b;
+  psx_gpu->enhancement_buf_start[0] = 0;
+  for (x = 1; x < sizeof(psx_gpu->enhancement_buf_by_x16); x++)
   {
-    if (b < 3 && x * ENH_BUF_TABLE_STEP >= s)
-    {
-      s += psx_gpu->enhancement_x_threshold;
+    if (b < 3 && is_new_scanout(psx_gpu, x * ENH_BUF_TABLE_STEP)) {
       b++;
+      psx_gpu->enhancement_buf_start[b] = x * ENH_BUF_TABLE_STEP;
     }
+
     psx_gpu->enhancement_buf_by_x16[x] = b;
   }
+#if 0
+  printf("buf_by_x16:\n");
+  for (b = 0; b < 3; b++) {
+    int first = -1, count = 0;
+    for (x = 0; x < sizeof(psx_gpu->enhancement_buf_by_x16); x++) {
+      if (psx_gpu->enhancement_buf_by_x16[x] == b) {
+        if (first < 0) first = x;
+        count++;
+      }
+    }
+    if (count) {
+      assert(first * ENH_BUF_TABLE_STEP == psx_gpu->enhancement_buf_start[b]);
+      printf("%d: %3zd-%zd\n", b, first * ENH_BUF_TABLE_STEP,
+          (first + count) * ENH_BUF_TABLE_STEP);
+    }
+  }
+#endif
 }
 
 static void update_enhancement_buf_table_from_x(psx_gpu_struct *psx_gpu,
  u32 x0, u32 len)
 {
+#if 0
   u32 x, b;
 
   for (x = x0, b = 0; x >= len; b++)
@@ -819,6 +853,7 @@ static void update_enhancement_buf_table_from_x(psx_gpu_struct *psx_gpu,
 
   memset(psx_gpu->enhancement_buf_by_x16 + x0 / ENH_BUF_TABLE_STEP,
    b, (len + ENH_BUF_TABLE_STEP - 1) / ENH_BUF_TABLE_STEP);
+#endif
 }
 
 #define select_enhancement_buf(psx_gpu) \
@@ -844,30 +879,27 @@ static void update_enhancement_buf_table_from_x(psx_gpu_struct *psx_gpu,
 }
 
 #define shift_vertices3(v) { \
-  v[0]->x *= 2; \
-  v[0]->y *= 2; \
-  v[1]->x *= 2; \
-  v[1]->y *= 2; \
-  v[2]->x *= 2; \
-  v[2]->y *= 2; \
+  v[0]->x <<= 1; \
+  v[0]->y <<= 1; \
+  v[1]->x <<= 1; \
+  v[1]->y <<= 1; \
+  v[2]->x <<= 1; \
+  v[2]->y <<= 1; \
 }
 
 #define unshift_vertices3(v) { \
-  v[0]->x /= 2; \
-  v[0]->y /= 2; \
-  v[1]->x /= 2; \
-  v[1]->y /= 2; \
-  v[2]->x /= 2; \
-  v[2]->y /= 2; \
+  v[0]->x >>= 1; \
+  v[0]->y >>= 1; \
+  v[1]->x >>= 1; \
+  v[1]->y >>= 1; \
+  v[2]->x >>= 1; \
+  v[2]->y >>= 1; \
 }
 
 #define shift_triangle_area() \
   psx_gpu->triangle_area *= 4
 
-extern void scale2x_tiles8(void *dst, const void *src, int w8, int h);
-
 #ifndef NEON_BUILD
-// TODO?
 void scale2x_tiles8(void *dst, const void *src, int w8, int h)
 {
   uint16_t* d = (uint16_t*)dst;
@@ -938,6 +970,16 @@ void scale2x_tiles8(void *dst, const void *src, int w8, int h)
 
 static int disable_main_render;
 
+static int check_enhanced_range(psx_gpu_struct *psx_gpu, int x, int x_end)
+{
+  // simple reject to avoid oveflowing the 1024 width
+  // (assume some offscreen render-to-texture thing)
+  if (x >= (int)(psx_gpu->saved_viewport_start_x + 512))
+    return 0;
+
+  return 1;
+}
+
 static void do_triangle_enhanced(psx_gpu_struct *psx_gpu,
  vertex_struct *vertexes, u32 current_command)
 {
@@ -949,37 +991,22 @@ static void do_triangle_enhanced(psx_gpu_struct *psx_gpu,
   if (!disable_main_render)
     render_triangle_p(psx_gpu, vertex_ptrs, current_command);
 
+  if (!check_enhanced_range(psx_gpu, vertex_ptrs[0]->x, vertex_ptrs[2]->x))
+    return;
+
   enhancement_enable();
   shift_vertices3(vertex_ptrs);
   shift_triangle_area();
   render_triangle_p(psx_gpu, vertex_ptrs, current_command);
+  unshift_vertices3(vertex_ptrs);
 }
 
 static void do_quad_enhanced(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
  u32 current_command)
 {
-  vertex_struct *vertex_ptrs[3];
-
-  if (prepare_triangle(psx_gpu, vertexes, vertex_ptrs)) {
-    if (!disable_main_render)
-      render_triangle_p(psx_gpu, vertex_ptrs, current_command);
-
-    enhancement_enable();
-    shift_vertices3(vertex_ptrs);
-    shift_triangle_area();
-    render_triangle_p(psx_gpu, vertex_ptrs, current_command);
-    unshift_vertices3(vertex_ptrs);
-  }
+  do_triangle_enhanced(psx_gpu, vertexes, current_command);
   enhancement_disable();
-  if (prepare_triangle(psx_gpu, &vertexes[1], vertex_ptrs)) {
-    if (!disable_main_render)
-      render_triangle_p(psx_gpu, vertex_ptrs, current_command);
-
-    enhancement_enable();
-    shift_vertices3(vertex_ptrs);
-    shift_triangle_area();
-    render_triangle_p(psx_gpu, vertex_ptrs, current_command);
-  }
+  do_triangle_enhanced(psx_gpu, &vertexes[1], current_command);
 }
 
 #if 0
@@ -1062,6 +1089,7 @@ static void do_sprite_enhanced(psx_gpu_struct *psx_gpu, int x, int y,
 u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
  u32 *last_command)
 {
+  vertex_struct vertexes[4] __attribute__((aligned(16))) = {};
   u32 current_command = 0, command_length;
 
   u32 *list_start = list;
@@ -1097,12 +1125,24 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         u32 width = list_s16[4] & 0x3FF;
         u32 height = list_s16[5] & 0x1FF;
         u32 color = list[0] & 0xFFFFFF;
+        u32 i1, i2;
 
         x &= ~0xF;
         width = ((width + 0xF) & ~0xF);
+        if (width == 0 || height == 0)
+          break;
 
         do_fill(psx_gpu, x, y, width, height, color);
 
+        i1 = select_enhancement_buf_index(psx_gpu, x);
+        i2 = select_enhancement_buf_index(psx_gpu, x + width - 1);
+        if (i1 != i2) {
+          sync_enhancement_buffers(x, y, width, height);
+          break;
+        }
+        if (x >= psx_gpu->enhancement_buf_start[i1] + psx_gpu->saved_hres)
+          break;
+
         psx_gpu->vram_out_ptr = select_enhancement_buf_ptr(psx_gpu, x);
         x *= 2;
         y *= 2;
@@ -1353,7 +1393,9 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         u32 height = list_s16[5] & 0x1FF;
 
         render_sprite(psx_gpu, x, y, 0, 0, width, height, current_command, list[0]);
-        do_sprite_enhanced(psx_gpu, x, y, 0, 0, width, height, list[0]);
+
+        if (check_enhanced_range(psx_gpu, x, x + width))
+          do_sprite_enhanced(psx_gpu, x, y, 0, 0, width, height, list[0]);
         break;
       }
   
@@ -1370,7 +1412,9 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
 
         render_sprite(psx_gpu, x, y, u, v, width, height,
          current_command, list[0]);
-        do_sprite_enhanced(psx_gpu, x, y, u, v, width, height, list[0]);
+
+        if (check_enhanced_range(psx_gpu, x, x + width))
+          do_sprite_enhanced(psx_gpu, x, y, u, v, width, height, list[0]);
         break;
       }
   
@@ -1383,7 +1427,9 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         s32 y = sign_extend_11bit(list_s16[3] + psx_gpu->offset_y);
 
         render_sprite(psx_gpu, x, y, 0, 0, 1, 1, current_command, list[0]);
-        do_sprite_enhanced(psx_gpu, x, y, 0, 0, 1, 1, list[0]);
+
+        if (check_enhanced_range(psx_gpu, x, x + 1))
+          do_sprite_enhanced(psx_gpu, x, y, 0, 0, 1, 1, list[0]);
         break;
       }
   
@@ -1396,7 +1442,9 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         s32 y = sign_extend_11bit(list_s16[3] + psx_gpu->offset_y);
 
         render_sprite(psx_gpu, x, y, 0, 0, 8, 8, current_command, list[0]);
-        do_sprite_enhanced(psx_gpu, x, y, 0, 0, 8, 8, list[0]);
+
+        if (check_enhanced_range(psx_gpu, x, x + 8))
+          do_sprite_enhanced(psx_gpu, x, y, 0, 0, 8, 8, list[0]);
         break;
       }
   
@@ -1414,7 +1462,9 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
 
         render_sprite(psx_gpu, x, y, u, v, 8, 8,
          current_command, list[0]);
-        do_sprite_enhanced(psx_gpu, x, y, u, v, 8, 8, list[0]);
+
+        if (check_enhanced_range(psx_gpu, x, x + 8))
+          do_sprite_enhanced(psx_gpu, x, y, u, v, 8, 8, list[0]);
         break;
       }
   
@@ -1427,7 +1477,9 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         s32 y = sign_extend_11bit(list_s16[3] + psx_gpu->offset_y);
 
         render_sprite(psx_gpu, x, y, 0, 0, 16, 16, current_command, list[0]);
-        do_sprite_enhanced(psx_gpu, x, y, 0, 0, 16, 16, list[0]);
+
+        if (check_enhanced_range(psx_gpu, x, x + 16))
+          do_sprite_enhanced(psx_gpu, x, y, 0, 0, 16, 16, list[0]);
         break;
       }
   
@@ -1444,7 +1496,9 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         set_clut(psx_gpu, list_s16[5]);
 
         render_sprite(psx_gpu, x, y, u, v, 16, 16, current_command, list[0]);
-        do_sprite_enhanced(psx_gpu, x, y, u, v, 16, 16, list[0]);
+
+        if (check_enhanced_range(psx_gpu, x, x + 16))
+          do_sprite_enhanced(psx_gpu, x, y, u, v, 16, 16, list[0]);
         break;
       }
   
@@ -1456,21 +1510,12 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         u32 dy = list_s16[5] & 0x1FF;
         u32 w = ((list_s16[6] - 1) & 0x3FF) + 1;
         u32 h = ((list_s16[7] - 1) & 0x1FF) + 1;
-        u16 *buf;
 
         if (sx == dx && sy == dy && psx_gpu->mask_msb == 0)
           break;
 
         render_block_move(psx_gpu, sx, sy, dx, dy, w, h);
-        if (dy + h > 512)
-          h = 512 - dy;
-        sx = sx & ~7; // FIXME?
-        dx = dx * 2 & ~7;
-        dy *= 2;
-        w = (w + 7) / 8;
-        buf = select_enhancement_buf_ptr(psx_gpu, dx / 2);
-        scale2x_tiles8(buf + dy * 1024 + dx,
-          psx_gpu->vram_ptr + sy * 1024 + sx, w, h);
+        sync_enhancement_buffers(dx, dy, w, h);
         break;
       }
  
@@ -1542,7 +1587,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         psx_gpu->saved_viewport_start_y = viewport_start_y;
 
         w = (u32)psx_gpu->viewport_end_x - (u32)viewport_start_x + 1;
-        d = psx_gpu->enhancement_x_threshold - w;
+        d = psx_gpu->saved_hres - w;
         if(-16 <= d && d <= 16)
         {
           update_enhancement_buf_table_from_x(psx_gpu,
@@ -1579,7 +1624,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size,
         psx_gpu->saved_viewport_end_y = viewport_end_y;
 
         w = (u32)viewport_end_x - (u32)psx_gpu->viewport_start_x + 1;
-        d = psx_gpu->enhancement_x_threshold - w;
+        d = psx_gpu->saved_hres - w;
         if(-16 <= d && d <= 16)
         {
           update_enhancement_buf_table_from_x(psx_gpu,
diff --git a/plugins/gpu_neon/psx_gpu_if.c b/plugins/gpu_neon/psx_gpu_if.c
index fe3a7ecd..868f7b72 100644
--- a/plugins/gpu_neon/psx_gpu_if.c
+++ b/plugins/gpu_neon/psx_gpu_if.c
@@ -9,12 +9,9 @@
  */
 
 #include <stdio.h>
+#include <assert.h>
 
-#ifdef _WIN32
-#include <mman.h>
-#else
-#include <sys/mman.h>
-#endif
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 
 extern const unsigned char cmd_lengths[256];
 #define command_lengths cmd_lengths
@@ -26,9 +23,12 @@ 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)));
 
@@ -115,8 +115,12 @@ 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);
@@ -124,18 +128,29 @@ 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;
   }
 }
 
@@ -144,11 +159,18 @@ 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)
@@ -167,11 +189,44 @@ 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)
+  {
+    // 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.enhancement_x_threshold = gpu.screen.hres;
+    egpu.saved_hres = hres;
     update_enhancement_buf_table_from_hres(&egpu);
+    sync_enhancement_buffers(0, 0, 1024, 512);
   }
 }
 
@@ -179,23 +234,10 @@ 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);
   
@@ -212,10 +254,19 @@ void renderer_set_config(const struct rearmed_cbs *cbs)
     egpu.dither_table[3] = dither_table_row(3, -1, 2, -2); 
   }
 
+  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();
+  }
 }
+
 void renderer_sync(void)
 {
 }
+
 void renderer_notify_update_lace(int updated)
 {
 }
+
+// vim:ts=2:sw=2:expandtab
diff --git a/plugins/gpu_unai/gpulib_if.cpp b/plugins/gpu_unai/gpulib_if.cpp
index 0066ce03..710a7137 100644
--- a/plugins/gpu_unai/gpulib_if.cpp
+++ b/plugins/gpu_unai/gpulib_if.cpp
@@ -317,6 +317,10 @@ void renderer_notify_res_change(void)
   */
 }
 
+void renderer_notify_scanout_x_change(int x, int w)
+{
+}
+
 #ifdef USE_GPULIB
 // Handles GP0 draw settings commands 0xE1...0xE6
 static void gpuGP0Cmd_0xEx(gpu_unai_t &gpu_unai, u32 cmd_word)
@@ -786,7 +790,7 @@ void renderer_sync_ecmds(uint32_t *ecmds)
   do_cmd_list(&ecmds[1], 6, &dummy);
 }
 
-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)
 {
 }
 
diff --git a/plugins/gpulib/gpu.c b/plugins/gpulib/gpu.c
index 6c811d8d..40a6bd85 100644
--- a/plugins/gpulib/gpu.c
+++ b/plugins/gpulib/gpu.c
@@ -26,13 +26,8 @@
 #define noinline
 #endif
 
-#define gpu_log(fmt, ...) \
-  printf("%d:%03d: " fmt, *gpu.state.frame_count, *gpu.state.hcnt, ##__VA_ARGS__)
-
 //#define log_io gpu_log
 #define log_io(...)
-//#define log_anomaly gpu_log
-#define log_anomaly(...)
 
 struct psx_gpu gpu;
 
@@ -67,6 +62,7 @@ static noinline void do_reset(void)
   gpu.screen.hres = gpu.screen.w = 256;
   gpu.screen.vres = gpu.screen.h = 240;
   gpu.screen.x = gpu.screen.y = 0;
+  renderer_notify_res_change();
 }
 
 static noinline void update_width(void)
@@ -295,9 +291,11 @@ long GPUinit(void)
   ret  = vout_init();
   ret |= renderer_init();
 
+  memset(&gpu.state, 0, sizeof(gpu.state));
+  memset(&gpu.frameskip, 0, sizeof(gpu.frameskip));
+  gpu.zero = 0;
   gpu.state.frame_count = &gpu.zero;
   gpu.state.hcnt = &gpu.zero;
-  gpu.frameskip.active = 0;
   gpu.cmd_len = 0;
   do_reset();
 
@@ -361,6 +359,7 @@ void GPUwriteStatus(uint32_t data)
     case 0x05:
       gpu.screen.src_x = data & 0x3ff;
       gpu.screen.src_y = (data >> 10) & 0x1ff;
+      renderer_notify_scanout_x_change(gpu.screen.src_x, gpu.screen.hres);
       if (gpu.frameskip.set) {
         decide_frameskip_allow(gpu.ex_regs[3]);
         if (gpu.frameskip.last_flip_frame != *gpu.state.frame_count) {
@@ -510,7 +509,7 @@ static void finish_vram_transfer(int is_read)
     gpu.status &= ~PSX_GPU_STATUS_IMG;
   else
     renderer_update_caches(gpu.dma_start.x, gpu.dma_start.y,
-                           gpu.dma_start.w, gpu.dma_start.h);
+                           gpu.dma_start.w, gpu.dma_start.h, 0);
 }
 
 static noinline int do_cmd_list_skip(uint32_t *data, int count, int *last_cmd)
@@ -819,7 +818,7 @@ long GPUfreeze(uint32_t type, struct GPUFreeze *freeze)
         GPUwriteStatus((i << 24) | (gpu.regs[i] ^ 1));
       }
       renderer_sync_ecmds(gpu.ex_regs);
-      renderer_update_caches(0, 0, 1024, 512);
+      renderer_update_caches(0, 0, 1024, 512, 1);
       break;
   }
 
@@ -856,6 +855,9 @@ void GPUupdateLace(void)
   }
 
   vout_update();
+  if (gpu.state.enhancement_active && !gpu.state.enhancement_was_active)
+    renderer_update_caches(0, 0, 1024, 512, 1);
+  gpu.state.enhancement_was_active = gpu.state.enhancement_active;
   gpu.state.fb_dirty = 0;
   gpu.state.blanked = 0;
   renderer_notify_update_lace(1);
diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index 28f10fa4..bb62d597 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -13,6 +13,12 @@
 
 #include <stdint.h>
 
+#define gpu_log(fmt, ...) \
+  printf("%d:%03d: " fmt, *gpu.state.frame_count, *gpu.state.hcnt, ##__VA_ARGS__)
+
+//#define log_anomaly gpu_log
+#define log_anomaly(...)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -69,6 +75,7 @@ struct psx_gpu {
     uint32_t blanked:1;
     uint32_t enhancement_enable:1;
     uint32_t enhancement_active:1;
+    uint32_t enhancement_was_active:1;
     uint32_t downscale_enable:1;
     uint32_t downscale_active:1;
     uint32_t dims_changed:1;
@@ -118,13 +125,14 @@ struct rearmed_cbs;
 int  renderer_init(void);
 void renderer_finish(void);
 void renderer_sync_ecmds(uint32_t * ecmds);
-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);
 void renderer_flush_queues(void);
 void renderer_set_interlace(int enable, int is_odd);
 void renderer_set_config(const struct rearmed_cbs *config);
 void renderer_notify_res_change(void);
 void renderer_notify_update_lace(int updated);
 void renderer_sync(void);
+void renderer_notify_scanout_x_change(int x, int w);
 
 int  vout_init(void);
 int  vout_finish(void);
diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 3bd1d05b..4ba8e31c 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -84,8 +84,11 @@ void vout_update(void)
 
   check_mode_change(0);
   if (gpu.state.enhancement_active) {
+    if (!gpu.state.enhancement_was_active)
+      return; // buffer not ready yet
     vram = gpu.get_enhancement_bufer(&src_x, &src_y, &w, &h, &vram_h);
     x *= 2; y *= 2;
+    src_x2 *= 2;
   }
 
   if (gpu.state.downscale_active)