X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=plugins%2Fgpu_neon%2Fpsx_gpu%2Fpsx_gpu_arm_neon.S;h=11a11b1e104fb6db8fc4f8388ea8e5c81836c079;hp=0dc7ecee0183c2b7b32b65de0821453be6854ad1;hb=05740673341dbd61b23085c8cf1303f632e44f87;hpb=a1bef51562faf42f76de0c79b9fab38a23490d7b

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
index 0dc7ecee..11a11b1e 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
@@ -4337,102 +4337,52 @@ function(warmup)
 
   bx lr
 
+#undef vram_ptr
 #undef color
-#undef y
+#undef width
 #undef height
-
-#define psx_gpu                                           r0
-#define color                                             r1
-#define x                                                 r2
-#define y                                                 r3
+#undef pitch
 
 #define vram_ptr                                          r0
-#define width                                             r3
-#define height                                            r12
-
-#define parameter_width_offset                            0
-#define parameter_height_offset                           4
+#define color                                             r1
+#define width                                             r2
+#define height                                            r3
 
-#define color_r                                           r14
-#define color_g                                           r4
-#define color_b                                           r5  
+#define pitch                                             r1
 
-#define left_unaligned                                    r14
-#define right_unaligned                                   r4
-#define pitch                                             r5
-#define num_unaligned                                     r2
-#define num_width                                         r6
+#define num_width                                         r12
 
-#undef colors
+#undef colors_a
+#undef colors_b
 
-#define colors                                            q0
+#define colors_a                                          q0
+#define colors_b                                          q1
 
 .align 3
 
 function(render_block_fill_body)
-  ldr vram_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]
-  ldr height, [ sp, #parameter_height_offset ]
-
-  add vram_ptr, vram_ptr, y, lsl #11
-  ldr width, [ sp, #parameter_width_offset ]
-
-  add vram_ptr, vram_ptr, x, lsl #1
-  stmdb sp!, { r4 - r6, r14 }
-
-  ubfx color_r, color, #3, #5
-  ubfx color_g, color, #11, #5
-
-  ubfx color_b, color, #19, #5
-  orr color, color_r, color_g, lsl #5
-
-  orr color, color, color_b, lsl #10
-  add left_unaligned, x, #0x7
-
-  bic left_unaligned, left_unaligned, #0x7
-  vdup.u16 colors, color
-
-  sub left_unaligned, left_unaligned, x
+  vdup.u16 colors_a, color
   mov pitch, #2048
 
+  vmov colors_b, colors_a
   sub pitch, pitch, width, lsl #1
-  sub width, width, left_unaligned
 
-  and right_unaligned, width, #0x7
-  bic width, width, #0x7
+  mov num_width, width
 
- 0:
-  mov num_width, width, lsr #3
-
-  movs num_unaligned, left_unaligned
-  beq 2f
-
- 1:
-  strh color, [ vram_ptr ], #2
-
-  subs num_unaligned, num_unaligned, #1
-  bne 1b
-
- 2:
-  vst1.u32 { colors }, [ vram_ptr, :128 ]!
-  subs num_width, num_width, #1
-  bne 2b
+ 0:  
+  vst1.u32 { colors_a, colors_b }, [ vram_ptr, :256 ]!
 
-  movs num_unaligned, right_unaligned
-  beq 4f
-
- 3:
-  strh color, [ vram_ptr ], #2
-  
-  subs num_unaligned, num_unaligned, #1
-  bne 3b
+  subs num_width, num_width, #2
+  bne 0b
 
- 4:
   add vram_ptr, vram_ptr, pitch
+  mov num_width, width
+
   subs height, height, #1
   bne 0b
-  
-  ldmia sp!, { r4 - r6, pc }
 
+  bx lr
+ 
 
 #undef x
 #undef y