gpu_neon: start calculating cycles

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

diff --git a/plugins/gpu_neon/gpu.c b/plugins/gpu_neon/gpu.c
index 3f21fa9..4fc2897 100644 (file)
--- a/plugins/gpu_neon/gpu.c
+++ b/plugins/gpu_neon/gpu.c
@@ -24,7 +24,7 @@
 //#define log_anomaly gpu_log
 #define log_anomaly(...)
 
-struct psx_gpu gpu __attribute__((aligned(64)));
+struct psx_gpu gpu __attribute__((aligned(2048)));
 
 static noinline void do_reset(void)
 {
@@ -272,8 +272,10 @@ static void start_vram_transfer(uint32_t pos_word, uint32_t size_word, int is_re
 
   if (is_read)
     gpu.status.img = 1;
-  else
+  else {
+    renderer_flush_queues();
     renderer_invalidate_caches(gpu.dma.x, gpu.dma.y, gpu.dma.w, gpu.dma.h);
+  }
 
   log_io("start_vram_transfer %c (%d, %d) %dx%d\n", is_read ? 'r' : 'w',
     gpu.dma.x, gpu.dma.y, gpu.dma.w, gpu.dma.h);
@@ -357,7 +359,7 @@ static int check_cmd(uint32_t *data, int count)
   return count - pos;
 }
 
-static void flush_cmd_buffer(void)
+void flush_cmd_buffer(void)
 {
   int left = check_cmd(gpu.cmd_buffer, gpu.cmd_len);
   if (left > 0)
@@ -392,7 +394,7 @@ long GPUdmaChain(uint32_t *rambase, uint32_t start_addr)
   uint32_t addr, *list;
   uint32_t *llist_entry = NULL;
   int len, left, count;
-  long dma_words = 0;
+  long cpu_cycles = 0;
 
   if (unlikely(gpu.cmd_len > 0))
     flush_cmd_buffer();
@@ -400,7 +402,7 @@ long GPUdmaChain(uint32_t *rambase, uint32_t start_addr)
   // ff7 sends it's main list twice, detect this
   if (*gpu.state.frame_count == gpu.state.last_list.frame &&
       *gpu.state.hcnt - gpu.state.last_list.hcnt <= 1 &&
-       gpu.state.last_list.words > 1024)
+       gpu.state.last_list.cycles > 2048)
   {
     llist_entry = rambase + (gpu.state.last_list.addr & 0x1fffff) / 4;
     *llist_entry |= 0x800000;
@@ -413,7 +415,9 @@ long GPUdmaChain(uint32_t *rambase, uint32_t start_addr)
     list = rambase + (addr & 0x1fffff) / 4;
     len = list[0] >> 24;
     addr = list[0] & 0xffffff;
-    dma_words += 1 + len;
+    cpu_cycles += 10;
+    if (len > 0)
+      cpu_cycles += 5 + len;
 
     log_io(".chain %08x #%d\n", (list - rambase) * 4, len);
 
@@ -444,10 +448,10 @@ long GPUdmaChain(uint32_t *rambase, uint32_t start_addr)
 
   gpu.state.last_list.frame = *gpu.state.frame_count;
   gpu.state.last_list.hcnt = *gpu.state.hcnt;
-  gpu.state.last_list.words = dma_words;
+  gpu.state.last_list.cycles = cpu_cycles;
   gpu.state.last_list.addr = start_addr;
 
-  return dma_words;
+  return cpu_cycles;
 }
 
 void GPUreadDataMem(uint32_t *mem, int count)