gpu: rework dma vs busy timing

[pcsx_rearmed.git] / plugins / dfxvideo / gpulib_if.c

diff --git a/plugins/dfxvideo/gpulib_if.c b/plugins/dfxvideo/gpulib_if.c
index 20383ab..af35f3c 100644 (file)
--- a/plugins/dfxvideo/gpulib_if.c
+++ b/plugins/dfxvideo/gpulib_if.c
@@ -306,12 +306,13 @@ void renderer_notify_scanout_change(int x, int y)
 #include "../gpulib/gpu_timing.h"
 extern const unsigned char cmd_lengths[256];
 
-int do_cmd_list(uint32_t *list, int list_len, int *cpu_cycles_out, int *last_cmd)
+int do_cmd_list(uint32_t *list, int list_len,
+ int *cycles_sum_out, int *cycles_last, int *last_cmd)
 {
+  int cpu_cycles_sum = 0, cpu_cycles = *cycles_last;
   unsigned int cmd = 0, len;
   uint32_t *list_start = list;
   uint32_t *list_end = list + list_len;
-  u32 cpu_cycles = 0;
 
   for (; list < list_end; list += 1 + len)
   {
@@ -341,7 +342,7 @@ int do_cmd_list(uint32_t *list, int list_len, int *cpu_cycles_out, int *last_cmd
 
         while(1)
         {
-          cpu_cycles += gput_line(0);
+          gput_sum(cpu_cycles_sum, cpu_cycles, gput_line(0));
 
           if(list_position >= list_end) {
             cmd = -1;
@@ -366,7 +367,7 @@ int do_cmd_list(uint32_t *list, int list_len, int *cpu_cycles_out, int *last_cmd
 
         while(1)
         {
-          cpu_cycles += gput_line(0);
+          gput_sum(cpu_cycles_sum, cpu_cycles, gput_line(0));
 
           if(list_position >= list_end) {
             cmd = -1;
@@ -387,8 +388,8 @@ int do_cmd_list(uint32_t *list, int list_len, int *cpu_cycles_out, int *last_cmd
 #ifdef TEST
       case 0xA0:          //  sys -> vid
       {
-        u32 load_width = LE2HOST32(slist[4]);
-        u32 load_height = LE2HOST32(slist[5]);
+        u32 load_width = LE2HOST16(slist[4]);
+        u32 load_height = LE2HOST16(slist[5]);
         u32 load_size = load_width * load_height;
 
         len += load_size / 2;
@@ -398,32 +399,32 @@ int do_cmd_list(uint32_t *list, int list_len, int *cpu_cycles_out, int *last_cmd
 
       // timing
       case 0x02:
-        cpu_cycles += gput_fill(LE2HOST32(slist[4]) & 0x3ff,
-            LE2HOST32(slist[5]) & 0x1ff);
+        gput_sum(cpu_cycles_sum, cpu_cycles,
+            gput_fill(LE2HOST16(slist[4]) & 0x3ff, LE2HOST16(slist[5]) & 0x1ff));
         break;
-      case 0x20 ... 0x23: cpu_cycles += gput_poly_base();    break;
-      case 0x24 ... 0x27: cpu_cycles += gput_poly_base_t();  break;
-      case 0x28 ... 0x2B: cpu_cycles += gput_quad_base();    break;
-      case 0x2C ... 0x2F: cpu_cycles += gput_quad_base_t();  break;
-      case 0x30 ... 0x33: cpu_cycles += gput_poly_base_g();  break;
-      case 0x34 ... 0x37: cpu_cycles += gput_poly_base_gt(); break;
-      case 0x38 ... 0x3B: cpu_cycles += gput_quad_base_g();  break;
-      case 0x3C ... 0x3F: cpu_cycles += gput_quad_base_gt(); break;
-      case 0x40 ... 0x47: cpu_cycles += gput_line(0);        break;
-      case 0x50 ... 0x57: cpu_cycles += gput_line(0);        break;
+      case 0x20 ... 0x23: gput_sum(cpu_cycles_sum, cpu_cycles, gput_poly_base());    break;
+      case 0x24 ... 0x27: gput_sum(cpu_cycles_sum, cpu_cycles, gput_poly_base_t());  break;
+      case 0x28 ... 0x2B: gput_sum(cpu_cycles_sum, cpu_cycles, gput_quad_base());    break;
+      case 0x2C ... 0x2F: gput_sum(cpu_cycles_sum, cpu_cycles, gput_quad_base_t());  break;
+      case 0x30 ... 0x33: gput_sum(cpu_cycles_sum, cpu_cycles, gput_poly_base_g());  break;
+      case 0x34 ... 0x37: gput_sum(cpu_cycles_sum, cpu_cycles, gput_poly_base_gt()); break;
+      case 0x38 ... 0x3B: gput_sum(cpu_cycles_sum, cpu_cycles, gput_quad_base_g());  break;
+      case 0x3C ... 0x3F: gput_sum(cpu_cycles_sum, cpu_cycles, gput_quad_base_gt()); break;
+      case 0x40 ... 0x47: gput_sum(cpu_cycles_sum, cpu_cycles, gput_line(0));        break;
+      case 0x50 ... 0x57: gput_sum(cpu_cycles_sum, cpu_cycles, gput_line(0));        break;
       case 0x60 ... 0x63:
-        cpu_cycles += gput_sprite(LE2HOST32(slist[4]) & 0x3ff,
-            LE2HOST32(slist[5]) & 0x1ff);
+        gput_sum(cpu_cycles_sum, cpu_cycles,
+            gput_sprite(LE2HOST16(slist[4]) & 0x3ff, LE2HOST16(slist[5]) & 0x1ff));
         break;
       case 0x64 ... 0x67:
-        cpu_cycles += gput_sprite(LE2HOST32(slist[6]) & 0x3ff,
-            LE2HOST32(slist[7]) & 0x1ff);
+        gput_sum(cpu_cycles_sum, cpu_cycles,
+            gput_sprite(LE2HOST16(slist[6]) & 0x3ff, LE2HOST16(slist[7]) & 0x1ff));
         break;
-      case 0x68 ... 0x6B: cpu_cycles += gput_sprite(1, 1);   break;
+      case 0x68 ... 0x6B: gput_sum(cpu_cycles_sum, cpu_cycles, gput_sprite(1, 1));   break;
       case 0x70 ... 0x73:
-      case 0x74 ... 0x77: cpu_cycles += gput_sprite(8, 8);   break;
+      case 0x74 ... 0x77: gput_sum(cpu_cycles_sum, cpu_cycles, gput_sprite(8, 8));   break;
       case 0x78 ... 0x7B:
-      case 0x7C ... 0x7F: cpu_cycles += gput_sprite(16, 16); break;
+      case 0x7C ... 0x7F: gput_sum(cpu_cycles_sum, cpu_cycles, gput_sprite(16, 16)); break;
     }
   }
 
@@ -431,7 +432,8 @@ breakloop:
   gpu.ex_regs[1] &= ~0x1ff;
   gpu.ex_regs[1] |= lGPUstatusRet & 0x1ff;
 
-  *cpu_cycles_out += cpu_cycles;
+  *cycles_sum_out += cpu_cycles_sum;
+  *cycles_last = cpu_cycles;
   *last_cmd = cmd;
   return list - list_start;
 }