gpu_neon: fix wrong mask bit for fills

Fixes notaz/pcsx_rearmed#344

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.c b/plugins/gpu_neon/psx_gpu/psx_gpu.c
index b671a75..a59e9cd 100644 (file)
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c
@@ -4810,8 +4810,7 @@ void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
   u32 r = color & 0xFF;
   u32 g = (color >> 8) & 0xFF;
   u32 b = (color >> 16) & 0xFF;
-  u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
-   psx_gpu->mask_msb;
+  u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10);
   u32 color_32bpp = color_16bpp | (color_16bpp << 16);
 
   u32 *vram_ptr = (u32 *)(psx_gpu->vram_out_ptr + x + (y * 1024));
@@ -4863,8 +4862,7 @@ void render_block_fill_enh(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
   u32 r = color & 0xFF;
   u32 g = (color >> 8) & 0xFF;
   u32 b = (color >> 16) & 0xFF;
-  u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
-   psx_gpu->mask_msb;
+  u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10);
   u32 color_32bpp = color_16bpp | (color_16bpp << 16);
 
   u32 *vram_ptr = (u32 *)(psx_gpu->vram_out_ptr + x + (y * 1024));

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 ffbea04..1ba562b 100644 (file)
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
@@ -4386,51 +4386,6 @@ function(warmup)
 
 #undef vram_ptr
 #undef color
-#undef width
-#undef height
-#undef pitch
-
-#define vram_ptr                                          r0
-#define color                                             r1
-#define width                                             r2
-#define height                                            r3
-
-#define pitch                                             r1
-
-#define num_width                                         r12
-
-#undef colors_a
-#undef colors_b
-
-#define colors_a                                          q0
-#define colors_b                                          q1
-
-.align 3
-
-function(render_block_fill_body)
-  vdup.u16 colors_a, color
-  mov pitch, #2048
-
-  vmov colors_b, colors_a
-  sub pitch, pitch, width, lsl #1
-
-  mov num_width, width
-
- 0:  
-  vst1.u32 { colors_a, colors_b }, [vram_ptr, :256]!
-
-  subs num_width, num_width, #16
-  bne 0b
-
-  add vram_ptr, vram_ptr, pitch
-  mov num_width, width
-
-  subs height, height, #1
-  bne 0b
-
-  bx lr
- 
-
 #undef x
 #undef y
 #undef width
@@ -4523,6 +4478,7 @@ function(render_block_fill_body)
 #define texels_wide_high                                  d15
 #define texels_wide                                       q7
 
+.align 3
 
 setup_sprite_flush_blocks:
   vpush { q1 - q5 }