psx_gpu: use different uvrgb phase for enhancement

6 files changed, 59 insertions, 50 deletions

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.c b/plugins/gpu_neon/psx_gpu/psx_gpu.c
index 1385e2e..28ebcf5 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c
@@ -567,7 +567,7 @@ void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
 
   vec_4x32u uvrg_base;
   vec_4x32u b_base;
-  vec_4x32u const_0x8000;
+  vec_4x32u uvrgb_phase;
 
   vec_4x16s d0_a_d3_c, d0_b, d0_c;
   vec_4x16s d1_a, d1_b, d1_c_d2_a;
@@ -596,12 +596,12 @@ void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
   setup_gradient_calculation_input(1, b);
   setup_gradient_calculation_input(2, c);
 
-  dup_4x32b(const_0x8000, 0x8000);
+  dup_4x32b(uvrgb_phase, psx_gpu->uvrgb_phase);
   shl_long_4x16b(uvrg_base, x0_a_y0_c, 16);
   shl_long_4x16b(b_base, x0_b, 16);
 
-  add_4x32b(uvrg_base, uvrg_base, const_0x8000);
-  add_4x32b(b_base, b_base, const_0x8000);
+  add_4x32b(uvrg_base, uvrg_base, uvrgb_phase);
+  add_4x32b(b_base, b_base, uvrgb_phase);
 
   // Can probably pair these, but it'll require careful register allocation
   sub_4x16b(d0_a_d3_c, x1_a_y1_c, x0_a_y0_c);
@@ -4632,6 +4632,7 @@ void initialize_psx_gpu(psx_gpu_struct *psx_gpu, u16 *vram)
   psx_gpu->render_state = 0;
   psx_gpu->render_state_base = 0;
   psx_gpu->num_blocks = 0;
+  psx_gpu->uvrgb_phase = 0x8000;
 
   psx_gpu->vram_ptr = vram;
   psx_gpu->vram_out_ptr = vram;
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.h b/plugins/gpu_neon/psx_gpu/psx_gpu.h
index 71b99cd..7252dc2 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.h
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.h
@@ -123,7 +123,6 @@ typedef struct
   vec_4x32u g_block_span;
   vec_4x32u b_block_span;
 
-  // 76 bytes
   u32 b;
   u32 b_dy;
 
@@ -139,6 +138,8 @@ typedef struct
   u32 triangle_color;
   u32 dither_table[4];
 
+  u32 uvrgb_phase;
+
   struct render_block_handler_struct *render_block_handler;
   void *texture_page_ptr;
   void *texture_page_base;
@@ -146,19 +147,12 @@ typedef struct
   u16 *vram_ptr;
   u16 *vram_out_ptr;
 
-  // 26 bytes
   u16 render_state_base;
   u16 render_state;
 
   u16 num_spans;
   u16 num_blocks;
 
-  s16 offset_x;
-  s16 offset_y;
-
-  u16 clut_settings;
-  u16 texture_settings;
-
   s16 viewport_start_x;
   s16 viewport_start_y;
   s16 viewport_end_x;
@@ -166,7 +160,6 @@ typedef struct
 
   u16 mask_msb;
 
-  // 8 bytes
   u8 triangle_winding;
 
   u8 display_area_draw_enable;
@@ -182,6 +175,12 @@ typedef struct
   u8 primitive_type;
   u8 render_mode;
 
+  s16 offset_x;
+  s16 offset_y;
+
+  u16 clut_settings;
+  u16 texture_settings;
+
   // enhancement stuff
   u16 *enhancement_buf_ptr;
   s16 saved_viewport_start_x;
@@ -191,7 +190,7 @@ typedef struct
 
   // Align up to 64 byte boundary to keep the upcoming buffers cache line
   // aligned, also make reachable with single immediate addition
-  u8 reserved_a[240];
+  u8 reserved_a[236];
 
   // 8KB
   block_struct blocks[MAX_BLOCKS_PER_ROW];
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 6393e15..3239412 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
@@ -182,6 +182,7 @@
 #define uvrg_dx3l                                         d6
 #define uvrg_dx3h                                         d7
 
+#define uvrgb_phase                                       q13
 
 .align 4
 
@@ -313,11 +314,16 @@ function(compute_all_gradients)
   vmull.s16 ga_uvrg_y, d0_b, d1_b
   rsbmi ga_bx, ga_bx, #0
 
+  @ r12 = psx_gpu->uvrgb_phase
+  ldr r12, [ psx_gpu, #psx_gpu_uvrgb_phase_offset ]
+
   vmlsl.s16 ga_uvrg_y, d2_b, d3_b
   movs gs_by, ga_by, asr #31
 
   vshr.u64 d0, d30, #22
-  mov b_base, b0, lsl #16
+  add b_base, r12, b0, lsl #16
+
+  vdup.u32 uvrgb_phase, r12
 
   rsbmi ga_by, ga_by, #0
   vclt.s32 gs_uvrg_x, ga_uvrg_x, #0  @ gs_uvrg_x = ga_uvrg_x < 0
@@ -326,7 +332,6 @@ function(compute_all_gradients)
   ldrb r12, [ psx_gpu, #psx_gpu_triangle_winding_offset ]
   vclt.s32 gs_uvrg_y, ga_uvrg_y, #0  @ gs_uvrg_y = ga_uvrg_y < 0
 
-  add b_base, b_base, #0x8000
   rsb r12, r12, #0                   @ r12 = -(triangle->winding)
 
   vdup.u32 w_mask, r12               @ w_mask = { -w, -w, -w, -w }
@@ -335,7 +340,7 @@ function(compute_all_gradients)
   vshll.u16 uvrg_base, uvrg0, #16    @ uvrg_base = uvrg0 << 16
   vdup.u32 r_shift, r14              @ r_shift = { shift, shift, shift, shift }
 
-  vorr.u32 uvrg_base, #0x8000
+  vadd.u32 uvrg_base, uvrgb_phase
   vabs.s32 ga_uvrg_x, ga_uvrg_x      @ ga_uvrg_x = abs(ga_uvrg_x)
 
   vmov area_r_s, s0                  @ area_r_s = triangle_reciprocal
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_offsets.h b/plugins/gpu_neon/psx_gpu/psx_gpu_offsets.h
index 7ebf7db..1307891 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_offsets.h
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_offsets.h
@@ -18,35 +18,36 @@
 #define psx_gpu_dirty_textures_8bpp_alternate_mask_offset 0xb0
 #define psx_gpu_triangle_color_offset                     0xb4
 #define psx_gpu_dither_table_offset                       0xb8
-#define psx_gpu_render_block_handler_offset               0xc8
-#define psx_gpu_texture_page_ptr_offset                   0xcc
-#define psx_gpu_texture_page_base_offset                  0xd0
-#define psx_gpu_clut_ptr_offset                           0xd4
-#define psx_gpu_vram_ptr_offset                           0xd8
-#define psx_gpu_vram_out_ptr_offset                       0xdc
-#define psx_gpu_render_state_base_offset                  0xe0
-#define psx_gpu_render_state_offset                       0xe2
-#define psx_gpu_num_spans_offset                          0xe4
-#define psx_gpu_num_blocks_offset                         0xe6
-#define psx_gpu_offset_x_offset                           0xe8
-#define psx_gpu_offset_y_offset                           0xea
-#define psx_gpu_clut_settings_offset                      0xec
-#define psx_gpu_texture_settings_offset                   0xee
-#define psx_gpu_viewport_start_x_offset                   0xf0
-#define psx_gpu_viewport_start_y_offset                   0xf2
-#define psx_gpu_viewport_end_x_offset                     0xf4
-#define psx_gpu_viewport_end_y_offset                     0xf6
-#define psx_gpu_mask_msb_offset                           0xf8
-#define psx_gpu_triangle_winding_offset                   0xfa
-#define psx_gpu_display_area_draw_enable_offset           0xfb
-#define psx_gpu_current_texture_page_offset               0xfc
-#define psx_gpu_last_8bpp_texture_page_offset             0xfd
-#define psx_gpu_texture_mask_width_offset                 0xfe
-#define psx_gpu_texture_mask_height_offset                0xff
-#define psx_gpu_texture_window_x_offset                   0x100
-#define psx_gpu_texture_window_y_offset                   0x101
-#define psx_gpu_primitive_type_offset                     0x102
-#define psx_gpu_render_mode_offset                        0x103
+#define psx_gpu_uvrgb_phase_offset                        0xc8
+#define psx_gpu_render_block_handler_offset               0xcc
+#define psx_gpu_texture_page_ptr_offset                   0xd0
+#define psx_gpu_texture_page_base_offset                  0xd4
+#define psx_gpu_clut_ptr_offset                           0xd8
+#define psx_gpu_vram_ptr_offset                           0xdc
+#define psx_gpu_vram_out_ptr_offset                       0xe0
+#define psx_gpu_render_state_base_offset                  0xe4
+#define psx_gpu_render_state_offset                       0xe6
+#define psx_gpu_num_spans_offset                          0xe8
+#define psx_gpu_num_blocks_offset                         0xea
+#define psx_gpu_viewport_start_x_offset                   0xec
+#define psx_gpu_viewport_start_y_offset                   0xee
+#define psx_gpu_viewport_end_x_offset                     0xf0
+#define psx_gpu_viewport_end_y_offset                     0xf2
+#define psx_gpu_mask_msb_offset                           0xf4
+#define psx_gpu_triangle_winding_offset                   0xf6
+#define psx_gpu_display_area_draw_enable_offset           0xf7
+#define psx_gpu_current_texture_page_offset               0xf8
+#define psx_gpu_last_8bpp_texture_page_offset             0xf9
+#define psx_gpu_texture_mask_width_offset                 0xfa
+#define psx_gpu_texture_mask_height_offset                0xfb
+#define psx_gpu_texture_window_x_offset                   0xfc
+#define psx_gpu_texture_window_y_offset                   0xfd
+#define psx_gpu_primitive_type_offset                     0xfe
+#define psx_gpu_render_mode_offset                        0xff
+#define psx_gpu_offset_x_offset                           0x100
+#define psx_gpu_offset_y_offset                           0x102
+#define psx_gpu_clut_settings_offset                      0x104
+#define psx_gpu_texture_settings_offset                   0x106
 #define psx_gpu_blocks_offset                             0x200
 #define psx_gpu_span_uvrg_offset_offset                   0x2200
 #define psx_gpu_span_edge_data_offset                     0x4200
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_offsets_update.c b/plugins/gpu_neon/psx_gpu/psx_gpu_offsets_update.c
index ff74f34..5adfb75 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_offsets_update.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_offsets_update.c
@@ -43,6 +43,7 @@ int main()
 	WRITE_OFFSET(f, dirty_textures_8bpp_alternate_mask);
 	WRITE_OFFSET(f, triangle_color);
 	WRITE_OFFSET(f, dither_table);
+	WRITE_OFFSET(f, uvrgb_phase);
 	WRITE_OFFSET(f, render_block_handler);
 	WRITE_OFFSET(f, texture_page_ptr);
 	WRITE_OFFSET(f, texture_page_base);
@@ -53,10 +54,6 @@ int main()
 	WRITE_OFFSET(f, render_state);
 	WRITE_OFFSET(f, num_spans);
 	WRITE_OFFSET(f, num_blocks);
-	WRITE_OFFSET(f, offset_x);
-	WRITE_OFFSET(f, offset_y);
-	WRITE_OFFSET(f, clut_settings);
-	WRITE_OFFSET(f, texture_settings);
 	WRITE_OFFSET(f, viewport_start_x);
 	WRITE_OFFSET(f, viewport_start_y);
 	WRITE_OFFSET(f, viewport_end_x);
@@ -72,6 +69,10 @@ int main()
 	WRITE_OFFSET(f, texture_window_y);
 	WRITE_OFFSET(f, primitive_type);
 	WRITE_OFFSET(f, render_mode);
+	WRITE_OFFSET(f, offset_x);
+	WRITE_OFFSET(f, offset_y);
+	WRITE_OFFSET(f, clut_settings);
+	WRITE_OFFSET(f, texture_settings);
 	WRITE_OFFSET(f, blocks);
 	WRITE_OFFSET(f, span_uvrg_offset);
 	WRITE_OFFSET(f, span_edge_data);
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c b/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
index 2e1c0b8..1c449ea 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
@@ -757,6 +757,7 @@ breakloop:
   psx_gpu->viewport_end_x = psx_gpu->saved_viewport_end_x; \
   psx_gpu->viewport_end_y = psx_gpu->saved_viewport_end_y; \
   psx_gpu->render_mode &= ~RENDER_DOUBLE_MODE; \
+  psx_gpu->uvrgb_phase = 0x8000; \
 }
 
 #define enhancement_enable() { \
@@ -766,6 +767,7 @@ breakloop:
   psx_gpu->viewport_end_x = psx_gpu->saved_viewport_end_x * 2; \
   psx_gpu->viewport_end_y = psx_gpu->saved_viewport_end_y * 2; \
   psx_gpu->render_mode |= RENDER_DOUBLE_MODE; \
+  psx_gpu->uvrgb_phase = 0x1000; \
 }
 
 #define shift_vertices3(v) { \