From 0b02eb7712f1272fa7f38b0af41968b53af3b5a8 Mon Sep 17 00:00:00 2001
From: notaz
Date: Mon, 13 Aug 2012 00:02:23 +0300
Subject: add support for software-enhanced rendering

---
 plugins/gpulib/gpu.c     | 15 +++++++++++++++
 plugins/gpulib/gpu.h     |  4 ++++
 plugins/gpulib/vout_pl.c | 43 +++++++++++++++++++++++++++++++++----------
 3 files changed, 52 insertions(+), 10 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/gpu.c b/plugins/gpulib/gpu.c
index 46e92d1..462e301 100644
--- a/plugins/gpulib/gpu.c
+++ b/plugins/gpulib/gpu.c
@@ -9,6 +9,7 @@
  */
 
 #include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
 #include "gpu.h"
 
@@ -137,8 +138,21 @@ long GPUinit(void)
 {
   int ret;
   ret  = vout_init();
+
+  gpu.state.enhancement_available = 0;
   ret |= renderer_init();
 
+  if (gpu.state.enhancement_available) {
+    if (gpu.enhancement_bufer == NULL)
+      gpu.enhancement_bufer = malloc(2048 * 1024 * 2 + 1024 * 512 * 2);
+      if (gpu.enhancement_bufer == NULL)
+        gpu_log("OOM for enhancement buffer\n");
+  }
+  else if (gpu.enhancement_bufer != NULL) {
+    free(gpu.enhancement_bufer);
+    gpu.enhancement_bufer = NULL;
+  }
+
   gpu.state.frame_count = &gpu.zero;
   gpu.state.hcnt = &gpu.zero;
   gpu.frameskip.active = 0;
@@ -669,6 +683,7 @@ void GPUrearmedCallbacks(const struct rearmed_cbs *cbs)
   gpu.state.hcnt = cbs->gpu_hcnt;
   gpu.state.frame_count = cbs->gpu_frame_count;
   gpu.state.allow_interlace = cbs->gpu_neon.allow_interlace;
+  gpu.state.enhancement_enable = cbs->gpu_neon.enhancement_enable;
 
   if (cbs->pl_vout_set_raw_vram)
     cbs->pl_vout_set_raw_vram(gpu.vram);
diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index 1cbe38c..f514395 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -67,6 +67,9 @@ struct psx_gpu {
     uint32_t old_interlace:1;
     uint32_t allow_interlace:2;
     uint32_t blanked:1;
+    uint32_t enhancement_available:1;
+    uint32_t enhancement_enable:1;
+    uint32_t enhancement_active:1;
     uint32_t *frame_count;
     uint32_t *hcnt; /* hsync count */
     struct {
@@ -87,6 +90,7 @@ struct psx_gpu {
     uint32_t last_flip_frame;
     uint32_t pending_fill[3];
   } frameskip;
+  void *enhancement_bufer;
 };
 
 extern struct psx_gpu gpu;
diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 0bd1ecf..47c28f3 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -31,13 +31,25 @@ static void check_mode_change(void)
 {
   static uint32_t old_status;
   static int old_h;
+  int w = gpu.screen.hres;
+  int h = gpu.screen.h;
+
+  gpu.state.enhancement_active =
+    gpu.enhancement_bufer != NULL && gpu.state.enhancement_enable
+    && w <= 512 && h <= 256 && !gpu.status.rgb24;
+
+  if (gpu.state.enhancement_active) {
+    w *= 2;
+    h *= 2;
+  }
 
   // width|rgb24 change?
-  if ((gpu.status.reg ^ old_status) & ((7<<16)|(1<<21)) || gpu.screen.h != old_h)
+  if ((gpu.status.reg ^ old_status) & ((7<<16)|(1<<21)) || h != old_h)
   {
     old_status = gpu.status.reg;
-    old_h = gpu.screen.h;
-    screen_buf = cbs->pl_vout_set_mode(gpu.screen.hres, gpu.screen.h,
+    old_h = h;
+
+    screen_buf = cbs->pl_vout_set_mode(w, h,
       (gpu.status.rgb24 && !cbs->only_16bpp) ? 24 : 16);
   }
 }
@@ -50,6 +62,8 @@ static void blit(void)
   int h = gpu.screen.h;
   uint16_t *vram = gpu.vram;
   int stride = gpu.screen.hres;
+  int vram_stride = 1024;
+  int vram_mask = 1024 * 512 - 1;
   int fb_offs, doffs;
   uint8_t *dest;
 
@@ -57,7 +71,16 @@ static void blit(void)
   if (dest == NULL)
     return;
 
-  fb_offs = y * 1024 + x;
+  if (gpu.state.enhancement_active) {
+    vram = gpu.enhancement_bufer;
+    x *= 2;
+    y *= 2;
+    w *= 2;
+    h *= 2;
+    stride *= 2;
+    vram_mask = 1024 * 1024 - 1;
+  }
+  fb_offs = y * vram_stride + x;
 
   // only do centering, at least for now
   doffs = (stride - w) / 2 & ~1;
@@ -66,17 +89,17 @@ static void blit(void)
   {
     if (cbs->only_16bpp) {
       dest += doffs * 2;
-      for (; h-- > 0; dest += stride * 2, fb_offs += 1024)
+      for (; h-- > 0; dest += stride * 2, fb_offs += vram_stride)
       {
-        fb_offs &= 1024*512-1;
+        fb_offs &= vram_mask;
         bgr888_to_rgb565(dest, vram + fb_offs, w * 3);
       }
     }
     else {
       dest += (doffs / 8) * 24;
-      for (; h-- > 0; dest += stride * 3, fb_offs += 1024)
+      for (; h-- > 0; dest += stride * 3, fb_offs += vram_stride)
       {
-        fb_offs &= 1024*512-1;
+        fb_offs &= vram_mask;
         bgr888_to_rgb888(dest, vram + fb_offs, w * 3);
       }
     }
@@ -84,9 +107,9 @@ static void blit(void)
   else
   {
     dest += doffs * 2;
-    for (; h-- > 0; dest += stride * 2, fb_offs += 1024)
+    for (; h-- > 0; dest += stride * 2, fb_offs += vram_stride)
     {
-      fb_offs &= 1024*512-1;
+      fb_offs &= vram_mask;
       bgr555_to_rgb565(dest, vram + fb_offs, w * 2);
     }
   }
-- 
cgit v1.2.3


From f1359c5758c2e745b1cbec63e21445fa65f7cafe Mon Sep 17 00:00:00 2001
From: notaz
Date: Mon, 13 Aug 2012 02:53:21 +0300
Subject: psx_gpu: switch enhancement to 2048 width

otherwise games that position framebuffers horizontally corrupt the
display.
---
 plugins/gpulib/vout_pl.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 47c28f3..5131034 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -78,7 +78,8 @@ static void blit(void)
     w *= 2;
     h *= 2;
     stride *= 2;
-    vram_mask = 1024 * 1024 - 1;
+    vram_stride = 2048;
+    vram_mask = 2048 * 1024 - 1;
   }
   fb_offs = y * vram_stride + x;
 
-- 
cgit v1.2.3


From e929dec505f8d3692248fe0d42c84a37c994ad39 Mon Sep 17 00:00:00 2001
From: notaz
Date: Sat, 18 Aug 2012 02:37:49 +0300
Subject: psx_gpu: switch to 1024 width again.

---
 plugins/gpulib/gpu.c     | 16 ++--------------
 plugins/gpulib/gpu.h     |  5 +++--
 plugins/gpulib/vout_pl.c |  9 +++++----
 3 files changed, 10 insertions(+), 20 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/gpu.c b/plugins/gpulib/gpu.c
index 462e301..e133f07 100644
--- a/plugins/gpulib/gpu.c
+++ b/plugins/gpulib/gpu.c
@@ -9,7 +9,6 @@
  */
 
 #include <stdio.h>
-#include <stdlib.h>
 #include <string.h>
 #include "gpu.h"
 
@@ -138,21 +137,8 @@ long GPUinit(void)
 {
   int ret;
   ret  = vout_init();
-
-  gpu.state.enhancement_available = 0;
   ret |= renderer_init();
 
-  if (gpu.state.enhancement_available) {
-    if (gpu.enhancement_bufer == NULL)
-      gpu.enhancement_bufer = malloc(2048 * 1024 * 2 + 1024 * 512 * 2);
-      if (gpu.enhancement_bufer == NULL)
-        gpu_log("OOM for enhancement buffer\n");
-  }
-  else if (gpu.enhancement_bufer != NULL) {
-    free(gpu.enhancement_bufer);
-    gpu.enhancement_bufer = NULL;
-  }
-
   gpu.state.frame_count = &gpu.zero;
   gpu.state.hcnt = &gpu.zero;
   gpu.frameskip.active = 0;
@@ -164,6 +150,7 @@ long GPUinit(void)
 
 long GPUshutdown(void)
 {
+  renderer_finish();
   return vout_finish();
 }
 
@@ -221,6 +208,7 @@ void GPUwriteStatus(uint32_t data)
       gpu.screen.vres = vres[(gpu.status.reg >> 19) & 3];
       update_width();
       update_height();
+      renderer_notify_res_change();
       break;
     default:
       if ((cmd & 0xf0) == 0x10)
diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index f514395..5ad2a46 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -67,7 +67,6 @@ struct psx_gpu {
     uint32_t old_interlace:1;
     uint32_t allow_interlace:2;
     uint32_t blanked:1;
-    uint32_t enhancement_available:1;
     uint32_t enhancement_enable:1;
     uint32_t enhancement_active:1;
     uint32_t *frame_count;
@@ -90,7 +89,7 @@ struct psx_gpu {
     uint32_t last_flip_frame;
     uint32_t pending_fill[3];
   } frameskip;
-  void *enhancement_bufer;
+  uint16_t *enhancement_bufer;
 };
 
 extern struct psx_gpu gpu;
@@ -102,11 +101,13 @@ int do_cmd_list(uint32_t *list, int count, int *last_cmd);
 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_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);
 
 int  vout_init(void);
 int  vout_finish(void);
diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 5131034..cbd8034 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -68,18 +68,19 @@ static void blit(void)
   uint8_t *dest;
 
   dest = (uint8_t *)screen_buf;
-  if (dest == NULL)
+  if (dest == NULL || w == 0 || stride == 0)
     return;
 
   if (gpu.state.enhancement_active) {
-    vram = gpu.enhancement_bufer;
+    // this layout is gpu_neon specific..
+    vram = gpu.enhancement_bufer +
+      (x + 8) / stride * 1024 * 1024;
     x *= 2;
     y *= 2;
     w *= 2;
     h *= 2;
     stride *= 2;
-    vram_stride = 2048;
-    vram_mask = 2048 * 1024 - 1;
+    vram_mask = 1024 * 1024 - 1;
   }
   fb_offs = y * vram_stride + x;
 
-- 
cgit v1.2.3


From 50f9355a2338111d940ed408f52fe1defe4df23e Mon Sep 17 00:00:00 2001
From: notaz
Date: Sun, 19 Aug 2012 00:37:50 +0300
Subject: psx_gpu: start handling vram loads/moves for enhancement

---
 plugins/gpulib/vout_pl.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index cbd8034..5c74914 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -77,8 +77,8 @@ static void blit(void)
       (x + 8) / stride * 1024 * 1024;
     x *= 2;
     y *= 2;
-    w *= 2;
-    h *= 2;
+    w = (w - 2) * 2;
+    h = (h * 2) - 1;
     stride *= 2;
     vram_mask = 1024 * 1024 - 1;
   }
-- 
cgit v1.2.3


From 9ee0fd5b333039b1140d90f935aa9299825f1e42 Mon Sep 17 00:00:00 2001
From: notaz
Date: Sun, 19 Aug 2012 22:39:49 +0300
Subject: start mmap'ing vram, with hugetlb if possible

---
 plugins/gpulib/gpu.c | 44 ++++++++++++++++++++++++++++++++++++++++----
 plugins/gpulib/gpu.h |  5 +++--
 2 files changed, 43 insertions(+), 6 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/gpu.c b/plugins/gpulib/gpu.c
index e133f07..b61bff6 100644
--- a/plugins/gpulib/gpu.c
+++ b/plugins/gpulib/gpu.c
@@ -24,7 +24,7 @@
 //#define log_anomaly gpu_log
 #define log_anomaly(...)
 
-struct psx_gpu gpu __attribute__((aligned(2048)));
+struct psx_gpu gpu;
 
 static noinline int do_cmd_buffer(uint32_t *data, int count);
 static void finish_vram_transfer(int is_read);
@@ -133,6 +133,22 @@ static noinline void get_gpu_info(uint32_t data)
   }
 }
 
+// double, for overdraw guard
+#define VRAM_SIZE (1024 * 512 * 2 * 2)
+
+static int map_vram(void)
+{
+  gpu.vram = gpu.mmap(VRAM_SIZE);
+  if (gpu.vram != NULL) {
+    gpu.vram += 4096 / 2;
+    return 0;
+  }
+  else {
+    fprintf(stderr, "could not map vram, expect crashes\n");
+    return -1;
+  }
+}
+
 long GPUinit(void)
 {
   int ret;
@@ -145,13 +161,26 @@ long GPUinit(void)
   gpu.cmd_len = 0;
   do_reset();
 
+  if (gpu.mmap != NULL) {
+    if (map_vram() != 0)
+      ret = -1;
+  }
   return ret;
 }
 
 long GPUshutdown(void)
 {
+  long ret;
+
   renderer_finish();
-  return vout_finish();
+  ret = vout_finish();
+  if (gpu.vram != NULL) {
+    gpu.vram -= 4096 / 2;
+    gpu.munmap(gpu.vram, VRAM_SIZE);
+  }
+  gpu.vram = NULL;
+
+  return ret;
 }
 
 void GPUwriteStatus(uint32_t data)
@@ -584,13 +613,13 @@ long GPUfreeze(uint32_t type, struct GPUFreeze *freeze)
     case 1: // save
       if (gpu.cmd_len > 0)
         flush_cmd_buffer();
-      memcpy(freeze->psxVRam, gpu.vram, sizeof(gpu.vram));
+      memcpy(freeze->psxVRam, gpu.vram, 1024 * 512 * 2);
       memcpy(freeze->ulControl, gpu.regs, sizeof(gpu.regs));
       memcpy(freeze->ulControl + 0xe0, gpu.ex_regs, sizeof(gpu.ex_regs));
       freeze->ulStatus = gpu.status.reg;
       break;
     case 0: // load
-      memcpy(gpu.vram, freeze->psxVRam, sizeof(gpu.vram));
+      memcpy(gpu.vram, freeze->psxVRam, 1024 * 512 * 2);
       memcpy(gpu.regs, freeze->ulControl, sizeof(gpu.regs));
       memcpy(gpu.ex_regs, freeze->ulControl + 0xe0, sizeof(gpu.ex_regs));
       gpu.status.reg = freeze->ulStatus;
@@ -673,6 +702,13 @@ void GPUrearmedCallbacks(const struct rearmed_cbs *cbs)
   gpu.state.allow_interlace = cbs->gpu_neon.allow_interlace;
   gpu.state.enhancement_enable = cbs->gpu_neon.enhancement_enable;
 
+  gpu.mmap = cbs->mmap;
+  gpu.munmap = cbs->munmap;
+
+  // delayed vram mmap
+  if (gpu.vram == NULL)
+    map_vram();
+
   if (cbs->pl_vout_set_raw_vram)
     cbs->pl_vout_set_raw_vram(gpu.vram);
   renderer_set_config(cbs);
diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index 5ad2a46..78a8990 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -17,10 +17,9 @@ extern "C" {
 #define CMD_BUFFER_LEN          1024
 
 struct psx_gpu {
-  uint16_t vram[1024 * 512];
-  uint16_t guard[1024 * 512]; // overdraw guard
   uint32_t cmd_buffer[CMD_BUFFER_LEN];
   uint32_t regs[16];
+  uint16_t *vram;
   union {
     uint32_t reg;
     struct {
@@ -90,6 +89,8 @@ struct psx_gpu {
     uint32_t pending_fill[3];
   } frameskip;
   uint16_t *enhancement_bufer;
+  void *(*mmap)(unsigned int size);
+  void  (*munmap)(void *ptr, unsigned int size);
 };
 
 extern struct psx_gpu gpu;
-- 
cgit v1.2.3


From 06bc35c833797ce9f6f3287abf954f037bb12319 Mon Sep 17 00:00:00 2001
From: notaz
Date: Mon, 27 Aug 2012 02:04:01 +0300
Subject: various enhancement tweaks

---
 plugins/gpulib/vout_pl.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 5c74914..2116422 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -77,8 +77,8 @@ static void blit(void)
       (x + 8) / stride * 1024 * 1024;
     x *= 2;
     y *= 2;
-    w = (w - 2) * 2;
-    h = (h * 2) - 1;
+    w = w * 2;
+    h = h * 2;
     stride *= 2;
     vram_mask = 1024 * 1024 - 1;
   }
@@ -130,6 +130,7 @@ void vout_update(void)
 
 void vout_blank(void)
 {
+  gpu.state.enhancement_active = 0;
   check_mode_change();
   if (cbs->pl_vout_raw_flip == NULL) {
     int bytespp = gpu.status.rgb24 ? 3 : 2;
-- 
cgit v1.2.3


From 7956599fa5f666016f71870d9889748c97839041 Mon Sep 17 00:00:00 2001
From: notaz
Date: Mon, 22 Oct 2012 01:42:56 +0300
Subject: psx_gpu: select buffers differently

this handles weird drawing areas better
---
 plugins/gpulib/vout_pl.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 2116422..6e2764c 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -130,11 +130,15 @@ void vout_update(void)
 
 void vout_blank(void)
 {
-  gpu.state.enhancement_active = 0;
-  check_mode_change();
   if (cbs->pl_vout_raw_flip == NULL) {
+    int w = gpu.screen.hres;
+    int h = gpu.screen.h;
     int bytespp = gpu.status.rgb24 ? 3 : 2;
-    memset(screen_buf, 0, gpu.screen.hres * gpu.screen.h * bytespp);
+    if (gpu.state.enhancement_active) {
+      w *= 2;
+      h *= 2;
+    }
+    memset(screen_buf, 0, w * h * bytespp);
     screen_buf = cbs->pl_vout_flip();
   }
 }
-- 
cgit v1.2.3


From a8be0debff95f9b56af7c4c19eaacee782a09e28 Mon Sep 17 00:00:00 2001
From: notaz
Date: Tue, 23 Oct 2012 00:34:30 +0300
Subject: gpu: move enhacement logic out of vout_pl

---
 plugins/gpulib/gpu.h     |  3 ++-
 plugins/gpulib/vout_pl.c | 16 ++++------------
 2 files changed, 6 insertions(+), 13 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index 78a8990..ea5051e 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -88,7 +88,8 @@ struct psx_gpu {
     uint32_t last_flip_frame;
     uint32_t pending_fill[3];
   } frameskip;
-  uint16_t *enhancement_bufer;
+  uint16_t *(*get_enhancement_bufer)
+    (int *x, int *y, int *w, int *h, int *stride, int *mask);
   void *(*mmap)(unsigned int size);
   void  (*munmap)(void *ptr, unsigned int size);
 };
diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 6e2764c..9a84432 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -35,7 +35,7 @@ static void check_mode_change(void)
   int h = gpu.screen.h;
 
   gpu.state.enhancement_active =
-    gpu.enhancement_bufer != NULL && gpu.state.enhancement_enable
+    gpu.get_enhancement_bufer != NULL && gpu.state.enhancement_enable
     && w <= 512 && h <= 256 && !gpu.status.rgb24;
 
   if (gpu.state.enhancement_active) {
@@ -71,17 +71,9 @@ static void blit(void)
   if (dest == NULL || w == 0 || stride == 0)
     return;
 
-  if (gpu.state.enhancement_active) {
-    // this layout is gpu_neon specific..
-    vram = gpu.enhancement_bufer +
-      (x + 8) / stride * 1024 * 1024;
-    x *= 2;
-    y *= 2;
-    w = w * 2;
-    h = h * 2;
-    stride *= 2;
-    vram_mask = 1024 * 1024 - 1;
-  }
+  if (gpu.state.enhancement_active)
+    vram = gpu.get_enhancement_bufer(&x, &y, &w, &h, &stride, &vram_mask);
+
   fb_offs = y * vram_stride + x;
 
   // only do centering, at least for now
-- 
cgit v1.2.3


From fa56d36096cd4ab2b227ce2aa61c8404b8874689 Mon Sep 17 00:00:00 2001
From: notaz
Date: Mon, 29 Oct 2012 01:08:35 +0200
Subject: move blit to core, allow filtering while blitting

also adds libpicofe to pull filters from, and filter related UI stuff
---
 plugins/gpulib/gpu.h     |  2 +-
 plugins/gpulib/vout_pl.c | 91 ++++++++++++++----------------------------------
 2 files changed, 27 insertions(+), 66 deletions(-)

(limited to 'plugins/gpulib')

diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index ea5051e..d11f991 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -89,7 +89,7 @@ struct psx_gpu {
     uint32_t pending_fill[3];
   } frameskip;
   uint16_t *(*get_enhancement_bufer)
-    (int *x, int *y, int *w, int *h, int *stride, int *mask);
+    (int *x, int *y, int *w, int *h, int *vram_h);
   void *(*mmap)(unsigned int size);
   void  (*munmap)(void *ptr, unsigned int size);
 };
diff --git a/plugins/gpulib/vout_pl.c b/plugins/gpulib/vout_pl.c
index 9a84432..11307e2 100644
--- a/plugins/gpulib/vout_pl.c
+++ b/plugins/gpulib/vout_pl.c
@@ -15,7 +15,6 @@
 #include "../../frontend/plugin_lib.h"
 
 static const struct rearmed_cbs *cbs;
-static void *screen_buf;
 
 int vout_init(void)
 {
@@ -27,7 +26,7 @@ int vout_finish(void)
   return 0;
 }
 
-static void check_mode_change(void)
+static void check_mode_change(int force)
 {
   static uint32_t old_status;
   static int old_h;
@@ -44,95 +43,57 @@ static void check_mode_change(void)
   }
 
   // width|rgb24 change?
-  if ((gpu.status.reg ^ old_status) & ((7<<16)|(1<<21)) || h != old_h)
+  if (force || (gpu.status.reg ^ old_status) & ((7<<16)|(1<<21)) || h != old_h)
   {
     old_status = gpu.status.reg;
     old_h = h;
 
-    screen_buf = cbs->pl_vout_set_mode(w, h,
+    cbs->pl_vout_set_mode(w, h,
       (gpu.status.rgb24 && !cbs->only_16bpp) ? 24 : 16);
   }
 }
 
-static void blit(void)
+void vout_update(void)
 {
   int x = gpu.screen.x & ~1; // alignment needed by blitter
   int y = gpu.screen.y;
   int w = gpu.screen.w;
   int h = gpu.screen.h;
   uint16_t *vram = gpu.vram;
-  int stride = gpu.screen.hres;
-  int vram_stride = 1024;
-  int vram_mask = 1024 * 512 - 1;
-  int fb_offs, doffs;
-  uint8_t *dest;
-
-  dest = (uint8_t *)screen_buf;
-  if (dest == NULL || w == 0 || stride == 0)
+  int vram_h = 512;
+
+  if (w == 0 || h == 0)
     return;
 
+  check_mode_change(0);
   if (gpu.state.enhancement_active)
-    vram = gpu.get_enhancement_bufer(&x, &y, &w, &h, &stride, &vram_mask);
-
-  fb_offs = y * vram_stride + x;
+    vram = gpu.get_enhancement_bufer(&x, &y, &w, &h, &vram_h);
 
-  // only do centering, at least for now
-  doffs = (stride - w) / 2 & ~1;
-
-  if (gpu.status.rgb24)
-  {
-    if (cbs->only_16bpp) {
-      dest += doffs * 2;
-      for (; h-- > 0; dest += stride * 2, fb_offs += vram_stride)
-      {
-        fb_offs &= vram_mask;
-        bgr888_to_rgb565(dest, vram + fb_offs, w * 3);
-      }
-    }
-    else {
-      dest += (doffs / 8) * 24;
-      for (; h-- > 0; dest += stride * 3, fb_offs += vram_stride)
-      {
-        fb_offs &= vram_mask;
-        bgr888_to_rgb888(dest, vram + fb_offs, w * 3);
-      }
-    }
-  }
-  else
-  {
-    dest += doffs * 2;
-    for (; h-- > 0; dest += stride * 2, fb_offs += vram_stride)
-    {
-      fb_offs &= vram_mask;
-      bgr555_to_rgb565(dest, vram + fb_offs, w * 2);
+  if (y + h > vram_h) {
+    if (y + h - vram_h > h / 2) {
+      // wrap
+      y = 0;
+      h -= vram_h - y;
     }
+    else
+      // clip
+      h = vram_h - y;
   }
 
-  screen_buf = cbs->pl_vout_flip();
-}
+  vram += y * 1024 + x;
 
-void vout_update(void)
-{
-  check_mode_change();
-  if (cbs->pl_vout_raw_flip)
-    cbs->pl_vout_raw_flip(gpu.screen.x, gpu.screen.y);
-  else
-    blit();
+  cbs->pl_vout_flip(vram, 1024, gpu.status.rgb24, w, h);
 }
 
 void vout_blank(void)
 {
-  if (cbs->pl_vout_raw_flip == NULL) {
-    int w = gpu.screen.hres;
-    int h = gpu.screen.h;
-    int bytespp = gpu.status.rgb24 ? 3 : 2;
-    if (gpu.state.enhancement_active) {
-      w *= 2;
-      h *= 2;
-    }
-    memset(screen_buf, 0, w * h * bytespp);
-    screen_buf = cbs->pl_vout_flip();
+  int w = gpu.screen.hres;
+  int h = gpu.screen.h;
+  if (gpu.state.enhancement_active) {
+    w *= 2;
+    h *= 2;
   }
+  cbs->pl_vout_flip(NULL, 1024, gpu.status.rgb24, w, h);
 }
 
 long GPUopen(void **unused)
@@ -141,7 +102,7 @@ long GPUopen(void **unused)
   gpu.frameskip.frame_ready = 1;
 
   cbs->pl_vout_open();
-  screen_buf = cbs->pl_vout_flip();
+  check_mode_change(1);
   return 0;
 }
 
-- 
cgit v1.2.3