diff options
Diffstat (limited to 'plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S')
| -rw-r--r-- | plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S | 1000 |
1 files changed, 812 insertions, 188 deletions
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 294685a..4e1e403 100644 --- a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S +++ b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S @@ -1,5 +1,6 @@ /* * Copyright (C) 2011 Gilead Kutnick "Exophase" <exophase@gmail.com> + * Copyright (C) 2012 GraÅžvydas Ignotas "notaz" <notasas@gmail.com> * * This program is free software; you can redistribute it and/or * modify it under the terms of the GNU General Public License as @@ -16,65 +17,13 @@ #define MAX_BLOCKS 64 #define MAX_BLOCKS_PER_ROW 128 -#define psx_gpu_test_mask_offset 0 -#define psx_gpu_uvrg_offset 16 -#define psx_gpu_uvrg_dx_offset 32 -#define psx_gpu_uvrg_dy_offset 48 -#define psx_gpu_u_block_span_offset 64 -#define psx_gpu_v_block_span_offset 80 -#define psx_gpu_r_block_span_offset 96 -#define psx_gpu_g_block_span_offset 112 -#define psx_gpu_b_block_span_offset 128 - -#define psx_gpu_b_dx_offset 132 - -#define psx_gpu_b_offset 144 -#define psx_gpu_b_dy_offset 148 -#define psx_gpu_triangle_area_offset 152 -#define psx_gpu_texture_window_settings_offset 156 -#define psx_gpu_current_texture_mask_offset 160 -#define psx_gpu_viewport_mask_offset 164 -#define psx_gpu_dirty_textures_4bpp_mask_offset 168 -#define psx_gpu_dirty_textures_8bpp_mask_offset 172 -#define psx_gpu_dirty_textures_8bpp_alternate_mask_offset 176 -#define psx_gpu_triangle_color_offset 180 -#define psx_gpu_dither_table_offset 184 -#define psx_gpu_render_block_handler_offset 200 -#define psx_gpu_texture_page_ptr_offset 204 -#define psx_gpu_texture_page_base_offset 208 -#define psx_gpu_clut_ptr_offset 212 -#define psx_gpu_vram_ptr_offset 216 - -#define psx_gpu_render_state_base_offset 220 -#define psx_gpu_render_state_offset 222 -#define psx_gpu_num_spans_offset 224 -#define psx_gpu_num_blocks_offset 226 -#define psx_gpu_offset_x_offset 228 -#define psx_gpu_offset_y_offset 230 -#define psx_gpu_clut_settings_offset 232 -#define psx_gpu_texture_settings_offset 234 -#define psx_gpu_viewport_start_x_offset 236 -#define psx_gpu_viewport_start_y_offset 238 -#define psx_gpu_viewport_end_x_offset 240 -#define psx_gpu_viewport_end_y_offset 242 -#define psx_gpu_mask_msb_offset 244 - -#define psx_gpu_triangle_winding_offset 246 -#define psx_gpu_display_area_draw_enable_offset 247 -#define psx_gpu_current_texture_page_offset 248 -#define psx_gpu_last_8bpp_texture_page_offset 249 -#define psx_gpu_texture_mask_width_offset 250 -#define psx_gpu_texture_mask_height_offset 251 -#define psx_gpu_texture_window_x_offset 252 -#define psx_gpu_texture_window_y_offset 253 -#define psx_gpu_primitive_type_offset 254 - -#define psx_gpu_reserved_a_offset 255 - -#define psx_gpu_blocks_offset 0x0100 -#define psx_gpu_span_uvrg_offset_offset 0x2100 -#define psx_gpu_span_edge_data_offset 0x4100 -#define psx_gpu_span_b_offset_offset 0x5100 +#define RENDER_STATE_MASK_EVALUATE 0x20 +#define RENDER_FLAGS_MODULATE_TEXELS 0x1 +#define RENDER_FLAGS_BLEND 0x2 + +#include "psx_gpu_offsets.h" + +#define psx_gpu_b_dx_offset (psx_gpu_b_block_span_offset + 4) #define edge_data_left_x_offset 0 #define edge_data_num_blocks_offset 2 @@ -238,6 +187,7 @@ #define uvrg_dx3l d6 #define uvrg_dx3h d7 +#define uvrgb_phase q13 .align 4 @@ -369,11 +319,16 @@ 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 @@ -382,7 +337,6 @@ 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 } @@ -391,7 +345,7 @@ 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 @@ -657,7 +611,7 @@ function(compute_all_gradients) \ vdup.u32 edge_shifts, temp; \ vsub.u32 heights_b, heights, c_0x01; \ - vshr.u32 height_reciprocals, edge_shifts, #12; \ + vshr.u32 height_reciprocals, edge_shifts, #10; \ \ vmla.s32 heights_b, x_starts, heights; \ vbic.u16 edge_shifts, #0xE0; \ @@ -682,8 +636,8 @@ function(compute_all_gradients) vsub.u32 heights_b, heights, c_0x01; \ sub height_b_alt, height_minor_b, #1; \ \ - vshr.u32 height_reciprocals, edge_shifts, #12; \ - lsr height_reciprocal_alt, edge_shift_alt, #12; \ + vshr.u32 height_reciprocals, edge_shifts, #10; \ + lsr height_reciprocal_alt, edge_shift_alt, #10; \ \ vmla.s32 heights_b, x_starts, heights; \ mla height_b_alt, height_minor_b, start_c, height_b_alt; \ @@ -1221,6 +1175,10 @@ function(setup_spans_up_down) ldrh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] add temp, temp, height_minor_b + + cmp temp, #MAX_SPANS + beq 5f + strh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] 2: @@ -1236,6 +1194,15 @@ function(setup_spans_up_down) setup_spans_prologue_b() bal 4b + 5: + // FIXME: overflow corner case + sub temp, temp, height_minor_b + bics height_minor_b, #3 + add temp, temp, height_minor_b + strh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] + bne 2b + bal 1b + .pool #undef span_uvrg_offset @@ -1393,7 +1360,7 @@ function(setup_blocks_shaded_textured_dithered_##swizzling##_indirect) \ add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ \ ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ \ cmp span_num_blocks, #0; \ beq 1f; \ @@ -1660,7 +1627,7 @@ function(setup_blocks_unshaded_textured_dithered_##swizzling##_indirect) \ add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ \ ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ \ cmp span_num_blocks, #0; \ beq 1f; \ @@ -1855,7 +1822,7 @@ function(setup_blocks_unshaded_untextured_undithered_unswizzled_indirect) ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ] ldrh y, [ span_edge_data, #edge_data_y_offset ] - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ] + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] cmp span_num_blocks, #0 beq 1f @@ -1975,7 +1942,7 @@ function(setup_blocks_unshaded_untextured_undithered_unswizzled_direct) ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ] ldrh y, [ span_edge_data, #edge_data_y_offset ] - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ] + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] cmp span_num_blocks, #0 beq 1f @@ -2162,7 +2129,7 @@ function(setup_blocks_shaded_untextured_##dithering##_unswizzled_indirect) \ add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ \ ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ \ cmp span_num_blocks, #0; \ beq 1f; \ @@ -2402,7 +2369,7 @@ function(setup_blocks_shaded_untextured_##dithering##_unswizzled_direct) \ add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ \ ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ \ cmp span_num_blocks, #0; \ beq 1f; \ @@ -3239,6 +3206,7 @@ function(shade_blocks_##shading##_textured_modulated_##dithering##_##target) \ shade_blocks_textured_modulated_load_bdm_##shading(); \ vshrn.u16 texels_b, texels, #7; \ \ + pld [ block_ptr_load_a ]; \ vmovn.u16 texels_r, texels; \ vmlal.u8 pixels, pixels_r_low, d64_1; \ \ @@ -3437,10 +3405,12 @@ function(shade_blocks_textured_unmodulated_direct) [ draw_mask_bits_ptr, :16 ], c_64 vbif.u16 fb_pixels, pixels, draw_mask_combined - vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 - sub fb_ptr_cmp, fb_ptr_next, fb_ptr + pld [ fb_ptr_next, #64 ] + add fb_ptr_cmp, fb_ptr_cmp, #14 + vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 + cmp fb_ptr_cmp, #28 bls 4f @@ -3799,11 +3769,15 @@ function(blend_blocks_textured_add_##mask_evaluate) \ vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ vand.u16 pixels_mg, pixels, d128_0x83E0; \ \ - vbit.u16 blend_pixels, fb_pixels, draw_mask; \ - vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ + sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ + pld [ fb_ptr_next, #64 ]; \ \ sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ + vbit.u16 blend_pixels, fb_pixels, draw_mask; \ + \ add fb_ptr_cmp, fb_ptr_cmp, #14; \ + vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ + \ cmp fb_ptr_cmp, #28; \ bls 2f; \ \ @@ -4456,6 +4430,12 @@ function(render_block_fill_body) #define draw_mask_fb_ptr_left d2 #define draw_mask_fb_ptr_right d3 +#define draw_mask_fb_ptr_left_a d2 +#define draw_mask_fb_ptr_left_b d3 +#define draw_mask_fb_ptr_right_a d10 +#define draw_mask_fb_ptr_right_b d11 +#define draw_masks_fb_ptrs2 q5 + #define clut_low_a d4 #define clut_low_b d5 #define clut_high_a d6 @@ -4467,37 +4447,24 @@ function(render_block_fill_body) #define clut_a q2 #define clut_b q3 -#define texels_low d10 -#define texels_high d11 - +#define texels_low d12 +#define texels_high d13 -setup_sprite_flush_blocks_single: - vpush { q1 - q4 } - - stmdb sp!, { r0 - r3, r12, r14 } - bl flush_render_block_buffer - ldmia sp!, { r0 - r3, r12, r14 } - - vpop { q1 - q4 } - - add block, psx_gpu, #psx_gpu_blocks_offset - - mov num_blocks, sub_tile_height - bx lr +#define texels_wide_low d14 +#define texels_wide_high d15 +#define texels_wide q7 -setup_sprite_flush_blocks_double: - vpush { q1 - q4 } +setup_sprite_flush_blocks: + vpush { q1 - q5 } stmdb sp!, { r0 - r3, r12, r14 } bl flush_render_block_buffer ldmia sp!, { r0 - r3, r12, r14 } - vpop { q1 - q4 } + vpop { q1 - q5 } add block, psx_gpu, #psx_gpu_blocks_offset - - mov num_blocks, sub_tile_height, lsl #1 bx lr @@ -4535,8 +4502,6 @@ setup_sprite_update_texture_8bpp_cache: blne setup_sprite_update_texture_8bpp_cache \ -#define setup_sprite_tile_setup_block_no(side, offset, texture_mode) \ - #define setup_sprite_block_count_single() \ sub_tile_height \ @@ -4547,7 +4512,8 @@ setup_sprite_update_texture_8bpp_cache: add num_blocks, num_blocks, setup_sprite_block_count_##type(); \ cmp num_blocks, #MAX_BLOCKS; \ \ - blgt setup_sprite_flush_blocks_##type \ + movgt num_blocks, setup_sprite_block_count_##type(); \ + blgt setup_sprite_flush_blocks \ #define setup_sprite_tile_full_4bpp(edge) \ @@ -4729,31 +4695,33 @@ setup_sprite_update_texture_8bpp_cache: #define setup_sprite_tile_column_edge_post_adjust_full(edge) \ -#define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode) \ +#define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode, \ + x4mode) \ mov sub_tile_height, column_data; \ - setup_sprite_tile_column_edge_pre_adjust_##edge_mode(edge); \ - setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ - setup_sprite_tile_column_edge_post_adjust_##edge_mode(edge) \ + setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ + setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ + setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ -#define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode) \ +#define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode, \ + x4mode) \ and sub_tile_height, column_data, #0xFF; \ mov tiles_remaining, column_data, lsr #16; \ - setup_sprite_tile_column_edge_pre_adjust_##edge_mode(edge); \ - setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ + setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ + setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ \ subs tiles_remaining, tiles_remaining, #1; \ beq 2f; \ \ 3: \ mov sub_tile_height, #16; \ - setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ + setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ subs tiles_remaining, tiles_remaining, #1; \ bne 3b; \ \ 2: \ uxtb sub_tile_height, column_data, ror #8; \ - setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ - setup_sprite_tile_column_edge_post_adjust_##edge_mode(edge) \ + setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ + setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ #define setup_sprite_column_data_single() \ @@ -4772,17 +4740,30 @@ setup_sprite_update_texture_8bpp_cache: \ orr column_data, column_data, height_rounded, lsl #8 \ -#define setup_sprite_tile_column_width_single(texture_mode, multi_height, \ - edge_mode, edge) \ - setup_sprite_##texture_mode##_single_##multi_height##_##edge_mode##_##edge: \ +#define setup_sprite_setup_left_draw_mask_fb_ptr() \ + vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ + vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ + +#define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column() \ + mov fb_ptr_advance_column, #32; \ + vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ + \ + sub fb_ptr_advance_column, height, lsl #11; \ + vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ + +#define setup_sprite_setup_right_draw_mask_fb_ptr() \ + vdup.u8 draw_mask_fb_ptr_left, block_masks[4]; \ + vdup.u8 draw_mask_fb_ptr_right, block_masks[5] \ + +#define setup_sprite_tile_column_width_single(tm, multi_height, edge_mode, \ + edge, x4mode) \ + setup_sprite_##tm##_single_##multi_height##_##edge_mode##_##edge##x4mode: \ setup_sprite_column_data_##multi_height(); \ vext.32 block_masks_shifted, block_masks, block_masks, #1; \ vorr.u32 block_masks, block_masks, block_masks_shifted; \ - vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ - vdup.u8 draw_mask_fb_ptr_right, block_masks[1]; \ + setup_sprite_setup_left_draw_mask_fb_ptr##x4mode(); \ \ - setup_sprite_tile_column_height_##multi_height(edge_mode, edge, \ - texture_mode); \ + setup_sprite_tile_column_height_##multi_height(edge_mode, edge, tm, x4mode); \ ldmia sp!, { r4 - r11, pc } \ #define setup_sprite_tiled_advance_column() \ @@ -4791,39 +4772,337 @@ setup_sprite_update_texture_8bpp_cache: subeq texture_offset_base, texture_offset_base, #(0x100 + 0xF00) \ #define setup_sprite_tile_column_width_multi(tm, multi_height, left_mode, \ - right_mode) \ - setup_sprite_##tm##_multi_##multi_height##_##left_mode##_##right_mode: \ + right_mode, x4mode) \ + setup_sprite_##tm##_multi_##multi_height##_##left_mode##_##right_mode##x4mode:\ setup_sprite_column_data_##multi_height(); \ - mov fb_ptr_advance_column, #32; \ \ - sub fb_ptr_advance_column, height, lsl #11; \ - vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ + setup_sprite_setup_left_draw_mask_fb_ptr_advance_column##x4mode(); \ \ - vdup.u8 draw_mask_fb_ptr_right, block_masks[1]; \ - setup_sprite_tile_column_height_##multi_height(left_mode, right, tm); \ + setup_sprite_tile_column_height_##multi_height(left_mode, right, tm, x4mode);\ \ subs tile_width, tile_width, #2; \ add fb_ptr, fb_ptr, fb_ptr_advance_column; \ \ - vmov.u8 draw_masks_fb_ptrs, #0; \ beq 1f; \ \ + vmov.u8 draw_masks_fb_ptrs, #0; \ + vmov.u8 draw_masks_fb_ptrs2, #0; \ + \ 0: \ setup_sprite_tiled_advance_column(); \ - setup_sprite_tile_column_height_##multi_height(full, none, tm); \ + setup_sprite_tile_column_height_##multi_height(full, none, tm, x4mode); \ add fb_ptr, fb_ptr, fb_ptr_advance_column; \ subs tile_width, tile_width, #1; \ bne 0b; \ \ 1: \ - vdup.u8 draw_mask_fb_ptr_left, block_masks[4]; \ - vdup.u8 draw_mask_fb_ptr_right, block_masks[5]; \ + setup_sprite_setup_right_draw_mask_fb_ptr##x4mode(); \ \ setup_sprite_tiled_advance_column(); \ - setup_sprite_tile_column_height_##multi_height(right_mode, left, tm); \ + setup_sprite_tile_column_height_##multi_height(right_mode, left, tm, x4mode);\ ldmia sp!, { r4 - r11, pc } \ +#define setup_sprite_offset_u_adjust() \ + +#define setup_sprite_get_left_block_mask() \ + and left_block_mask, left_block_mask, #0xFF \ + +#define setup_sprite_compare_left_block_mask() \ + cmp left_block_mask, #0xFF \ + +#define setup_sprite_get_right_block_mask() \ + uxtb right_block_mask, right_block_mask, ror #8 \ + +#define setup_sprite_compare_right_block_mask() \ + cmp right_block_mask, #0xFF \ + + + +/* 4x stuff */ +#define fb_ptr2 column_data + +#define setup_sprite_offset_u_adjust_4x() \ + sub fb_ptr, fb_ptr, offset_u, lsl #1; \ + lsl offset_u_right, #1; \ + lsl offset_u, #1; \ + add offset_u_right, #1 \ + +#define setup_sprite_get_left_block_mask_4x() \ + sxth left_block_mask, left_block_mask \ + +#define setup_sprite_compare_left_block_mask_4x() \ + cmp left_block_mask, #0xFFFFFFFF \ + +#define setup_sprite_get_right_block_mask_4x() \ + sxth right_block_mask, right_block_mask, ror #16 \ + +#define setup_sprite_compare_right_block_mask_4x() \ + cmp right_block_mask, #0xFFFFFFFF \ + + +#define widen_texels_16bpp(texels_) \ + vmov texels_wide_low, texels_; \ + vmov texels_wide_high, texels_; \ + vzip.16 texels_wide_low, texels_wide_high \ + +#define widen_texels_8bpp(texels_) \ + vmov texels_wide_low, texels_; \ + vmov texels_wide_high, texels_; \ + vzip.8 texels_wide_low, texels_wide_high \ + +#define write_block_16bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ + vst1.u32 { texels_ }, [ block_, :128 ]; \ + add block_, block_, #40; \ + \ + vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ + vst1.u32 { draw_mask_fb_ptr_ }, [ block_, :64 ]; \ + add block_, block_, #24 \ + +/* assumes 16-byte offset already added to block_ */ +#define write_block_8bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ + vst1.u32 { texels_ }, [ block_, :64 ]; \ + add block_, block_, #24; \ + \ + vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ + vst1.u32 { draw_mask_fb_ptr_ }, [ block_, :64 ]; \ + add block_, block_, #40 \ + +#define do_texture_block_16bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ + draw_mask_fb_ptr_b_) \ + widen_texels_16bpp(texels_low); \ + add fb_ptr_tmp, fb_ptr, #1024*2; \ + \ + write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr); \ + \ + write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ + widen_texels_16bpp(texels_high); \ + \ + add fb_ptr_tmp, fb_ptr, #8*2; \ + write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ + \ + add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ + write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ + +#define do_texture_block_8bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ + draw_mask_fb_ptr_b_) \ + widen_texels_8bpp(texels); \ + add fb_ptr_tmp, fb_ptr, #1024*2; \ + \ + write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr); \ + write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ + \ + add fb_ptr_tmp, fb_ptr, #8*2; \ + write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ + \ + add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ + write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ + + +#define setup_sprite_tiled_initialize_4bpp_4x() \ + ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ]; \ + vld1.u32 { clut_a, clut_b }, [ clut_ptr, :128 ]; \ + \ + vuzp.u8 clut_a, clut_b \ + +#define setup_sprite_tiled_initialize_8bpp_4x() \ + + +#define setup_sprite_block_count_single_4x() \ + sub_tile_height, lsl #2 \ + +#define setup_sprite_block_count_double_4x() \ + sub_tile_height, lsl #(1+2) \ + +#define setup_sprite_tile_full_4bpp_4x(edge) \ + setup_sprite_tile_add_blocks(double_4x); \ + str column_data, [sp, #-8]!; /* fb_ptr2 */ \ + \ + 4: \ + and texture_block_ptr, texture_offset, texture_mask; \ + pld [ fb_ptr ]; \ + \ + add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ + vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ + \ + add texture_block_ptr, texture_offset, #8; \ + vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ + \ + and texture_block_ptr, texture_block_ptr, texture_mask; \ + vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ + \ + vzip.8 texels_low, texels_high; \ + do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ + draw_mask_fb_ptr_left_b); \ + \ + add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ + pld [ fb_ptr, #2048 ]; \ + \ + vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ + add fb_ptr, fb_ptr, #16*2; \ + \ + vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ + vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ + \ + vzip.8 texels_low, texels_high; \ + do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ + draw_mask_fb_ptr_right_b); \ + \ + add texture_offset, texture_offset, #0x10; \ + add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ + \ + subs sub_tile_height, sub_tile_height, #1; \ + bne 4b; \ + \ + ldr column_data, [sp], #8; /* fb_ptr2 */ \ + add texture_offset, texture_offset, #0xF00; \ + strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ + + +#define setup_sprite_tile_half_4bpp_4x(edge) \ + setup_sprite_tile_add_blocks(single_4x); \ + str column_data, [sp, #-8]!; /* fb_ptr2 */ \ + \ + 4: \ + and texture_block_ptr, texture_offset, texture_mask; \ + pld [ fb_ptr ]; \ + \ + add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ + vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ + \ + add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ + vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ + \ + vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ + add texture_offset, texture_offset, #0x10; \ + \ + vzip.8 texels_low, texels_high; \ + do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ + draw_mask_fb_ptr_##edge##_b); \ + \ + pld [ fb_ptr, #2048 ]; \ + add fb_ptr, fb_ptr, #2048 * 2; \ + \ + subs sub_tile_height, sub_tile_height, #1; \ + bne 4b; \ + \ + ldr column_data, [sp], #8; /* fb_ptr2 */ \ + add texture_offset, texture_offset, #0xF00; \ + strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ + + +#define setup_sprite_tile_full_8bpp_4x(edge) \ + setup_sprite_tile_add_blocks(double_4x); \ + add block, block, #16; \ + str column_data, [sp, #-8]!; /* fb_ptr2 */ \ + \ + 4: \ + and texture_block_ptr, texture_offset, texture_mask; \ + pld [ fb_ptr ]; \ + \ + add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ + vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ + \ + add texture_block_ptr, texture_offset, #8; \ + do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ + draw_mask_fb_ptr_left_b); \ + \ + pld [ fb_ptr, #2048 ]; \ + and texture_block_ptr, texture_block_ptr, texture_mask; \ + \ + add fb_ptr, fb_ptr, #16*2; \ + add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ + \ + vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ + \ + do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ + draw_mask_fb_ptr_right_b); \ + \ + add texture_offset, texture_offset, #0x10; \ + add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ + \ + subs sub_tile_height, sub_tile_height, #1; \ + bne 4b; \ + \ + sub block, block, #16; \ + ldr column_data, [sp], #8; /* fb_ptr2 */ \ + add texture_offset, texture_offset, #0xF00; \ + strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ + + +#define setup_sprite_tile_half_8bpp_4x(edge) \ + setup_sprite_tile_add_blocks(single_4x); \ + add block, block, #16; \ + str column_data, [sp, #-8]!; /* fb_ptr2 */ \ + \ + 4: \ + and texture_block_ptr, texture_offset, texture_mask; \ + pld [ fb_ptr ]; \ + \ + add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ + vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ + \ + pld [ fb_ptr, #2048 ]; \ + do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ + draw_mask_fb_ptr_##edge##_b); \ + \ + add texture_offset, texture_offset, #0x10; \ + add fb_ptr, fb_ptr, #2048 * 2; \ + \ + subs sub_tile_height, sub_tile_height, #1; \ + bne 4b; \ + \ + sub block, block, #16; \ + ldr column_data, [sp], #8; /* fb_ptr2 */ \ + add texture_offset, texture_offset, #0xF00; \ + strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ + + +#define setup_sprite_tile_column_edge_pre_adjust_half_right_4x() \ + add texture_offset, texture_offset_base, #8; \ + add fb_ptr, fb_ptr, #16 * 2 \ + +#define setup_sprite_tile_column_edge_pre_adjust_half_left_4x() \ + mov texture_offset, texture_offset_base \ + +#define setup_sprite_tile_column_edge_pre_adjust_half_4x(edge) \ + setup_sprite_tile_column_edge_pre_adjust_half_##edge##_4x() \ + +#define setup_sprite_tile_column_edge_pre_adjust_full_4x(edge) \ + mov texture_offset, texture_offset_base \ + +#define setup_sprite_tile_column_edge_post_adjust_half_right_4x() \ + sub fb_ptr, fb_ptr, #16 * 2 \ + +#define setup_sprite_tile_column_edge_post_adjust_half_left_4x() \ + +#define setup_sprite_tile_column_edge_post_adjust_half_4x(edge) \ + setup_sprite_tile_column_edge_post_adjust_half_##edge##_4x() \ + +#define setup_sprite_tile_column_edge_post_adjust_full_4x(edge) \ + + +#define setup_sprite_setup_left_draw_mask_fb_ptr_4x() \ + vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ + vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ + vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ + vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ + +#define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column_4x() \ + mov fb_ptr_advance_column, #32 * 2; \ + vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ + vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ + sub fb_ptr_advance_column, height, lsl #11 + 1; \ + vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ + vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ + +#define setup_sprite_setup_right_draw_mask_fb_ptr_4x() \ + vdup.u8 draw_mask_fb_ptr_left_a, block_masks[4]; \ + vdup.u8 draw_mask_fb_ptr_left_b, block_masks[5]; \ + vdup.u8 draw_mask_fb_ptr_right_a, block_masks[6]; \ + vdup.u8 draw_mask_fb_ptr_right_b, block_masks[7] \ + + // r0: psx_gpu // r1: x // r2: y @@ -4833,34 +5112,48 @@ setup_sprite_update_texture_8bpp_cache: // [ sp + 8 ]: height // [ sp + 12 ]: color (unused) -#define setup_sprite_tiled_builder(texture_mode) \ - \ -setup_sprite_tile_column_width_multi(texture_mode, multi, full, full); \ -setup_sprite_tile_column_width_single(texture_mode, multi, full, none); \ -setup_sprite_tile_column_width_multi(texture_mode, single, full, full); \ -setup_sprite_tile_column_width_single(texture_mode, single, full, none); \ -setup_sprite_tile_column_width_multi(texture_mode, multi, half, full); \ -setup_sprite_tile_column_width_single(texture_mode, multi, half, right); \ -setup_sprite_tile_column_width_multi(texture_mode, single, half, full); \ -setup_sprite_tile_column_width_single(texture_mode, single, half, right); \ -setup_sprite_tile_column_width_multi(texture_mode, multi, full, half); \ -setup_sprite_tile_column_width_single(texture_mode, multi, half, left); \ -setup_sprite_tile_column_width_multi(texture_mode, single, full, half); \ -setup_sprite_tile_column_width_single(texture_mode, single, half, left); \ -setup_sprite_tile_column_width_multi(texture_mode, multi, half, half); \ -setup_sprite_tile_column_width_multi(texture_mode, single, half, half); \ +#define setup_sprite_tiled_builder(texture_mode, x4mode) \ + \ +setup_sprite_tile_column_width_multi(texture_mode, multi, full, full, \ + x4mode); \ +setup_sprite_tile_column_width_single(texture_mode, multi, full, none, \ + x4mode); \ +setup_sprite_tile_column_width_multi(texture_mode, single, full, full, \ + x4mode); \ +setup_sprite_tile_column_width_single(texture_mode, single, full, none, \ + x4mode); \ +setup_sprite_tile_column_width_multi(texture_mode, multi, half, full, \ + x4mode); \ +setup_sprite_tile_column_width_single(texture_mode, multi, half, right, \ + x4mode); \ +setup_sprite_tile_column_width_multi(texture_mode, single, half, full, \ + x4mode); \ +setup_sprite_tile_column_width_single(texture_mode, single, half, right, \ + x4mode); \ +setup_sprite_tile_column_width_multi(texture_mode, multi, full, half, \ + x4mode); \ +setup_sprite_tile_column_width_single(texture_mode, multi, half, left, \ + x4mode); \ +setup_sprite_tile_column_width_multi(texture_mode, single, full, half, \ + x4mode); \ +setup_sprite_tile_column_width_single(texture_mode, single, half, left, \ + x4mode); \ +setup_sprite_tile_column_width_multi(texture_mode, multi, half, half, \ + x4mode); \ +setup_sprite_tile_column_width_multi(texture_mode, single, half, half, \ + x4mode); \ \ .align 4; \ \ -function(setup_sprite_##texture_mode) \ +function(setup_sprite_##texture_mode##x4mode) \ stmdb sp!, { r4 - r11, r14 }; \ - setup_sprite_tiled_initialize_##texture_mode(); \ + setup_sprite_tiled_initialize_##texture_mode##x4mode(); \ \ ldr v, [ sp, #36 ]; \ and offset_u, u, #0xF; \ \ ldr width, [ sp, #40 ]; \ - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ \ ldr height, [ sp, #44 ]; \ add fb_ptr, fb_ptr, y, lsl #11; \ @@ -4883,11 +5176,13 @@ function(setup_sprite_##texture_mode) \ \ /* texture_offset_base = VH-UH-UL-00 */\ bfi texture_offset_base, u, #4, #8; \ - movw right_block_mask, #0xFFFE; \ + mov right_block_mask, #0xFFFFFFFE; \ + \ + setup_sprite_offset_u_adjust##x4mode(); \ \ /* texture_offset_base = VH-UH-VL-00 */\ bfi texture_offset_base, v, #4, #4; \ - movw left_block_mask, #0xFFFF; \ + mov left_block_mask, #0xFFFFFFFF; \ \ mov tile_height, height_rounded, lsr #4; \ mvn left_block_mask, left_block_mask, lsl offset_u; \ @@ -4907,16 +5202,16 @@ function(setup_sprite_##texture_mode) \ \ /* texture_mask = HH-WH-HL-WL */\ bfi texture_mask, texture_mask_rev, #8, #4; \ - and left_block_mask, left_block_mask, #0xFF; \ + setup_sprite_get_left_block_mask##x4mode(); \ \ mov control_mask, #0; \ - cmp left_block_mask, #0xFF; \ + setup_sprite_compare_left_block_mask##x4mode(); \ \ - uxtb right_block_mask, right_block_mask, ror #8; \ + setup_sprite_get_right_block_mask##x4mode(); \ orreq control_mask, control_mask, #0x4; \ \ ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ - cmp right_block_mask, #0xFF; \ + setup_sprite_compare_right_block_mask##x4mode(); \ \ orreq control_mask, control_mask, #0x8; \ cmp tile_width, #1; \ @@ -4931,25 +5226,31 @@ function(setup_sprite_##texture_mode) \ ldr pc, [ pc, control_mask, lsl #2 ]; \ nop; \ \ - .word setup_sprite_##texture_mode##_multi_multi_full_full; \ - .word setup_sprite_##texture_mode##_single_multi_full_none; \ - .word setup_sprite_##texture_mode##_multi_single_full_full; \ - .word setup_sprite_##texture_mode##_single_single_full_none; \ - .word setup_sprite_##texture_mode##_multi_multi_half_full; \ - .word setup_sprite_##texture_mode##_single_multi_half_right; \ - .word setup_sprite_##texture_mode##_multi_single_half_full; \ - .word setup_sprite_##texture_mode##_single_single_half_right; \ - .word setup_sprite_##texture_mode##_multi_multi_full_half; \ - .word setup_sprite_##texture_mode##_single_multi_half_left; \ - .word setup_sprite_##texture_mode##_multi_single_full_half; \ - .word setup_sprite_##texture_mode##_single_single_half_left; \ - .word setup_sprite_##texture_mode##_multi_multi_half_half; \ + .word setup_sprite_##texture_mode##_multi_multi_full_full##x4mode; \ + .word setup_sprite_##texture_mode##_single_multi_full_none##x4mode; \ + .word setup_sprite_##texture_mode##_multi_single_full_full##x4mode; \ + .word setup_sprite_##texture_mode##_single_single_full_none##x4mode; \ + .word setup_sprite_##texture_mode##_multi_multi_half_full##x4mode; \ + .word setup_sprite_##texture_mode##_single_multi_half_right##x4mode; \ + .word setup_sprite_##texture_mode##_multi_single_half_full##x4mode; \ + .word setup_sprite_##texture_mode##_single_single_half_right##x4mode; \ + .word setup_sprite_##texture_mode##_multi_multi_full_half##x4mode; \ + .word setup_sprite_##texture_mode##_single_multi_half_left##x4mode; \ + .word setup_sprite_##texture_mode##_multi_single_full_half##x4mode; \ + .word setup_sprite_##texture_mode##_single_single_half_left##x4mode; \ + .word setup_sprite_##texture_mode##_multi_multi_half_half##x4mode; \ .word 0x00000000; \ - .word setup_sprite_##texture_mode##_multi_single_half_half \ + .word setup_sprite_##texture_mode##_multi_single_half_half##x4mode; \ -setup_sprite_tiled_builder(4bpp); -setup_sprite_tiled_builder(8bpp); +setup_sprite_tiled_builder(4bpp,); +setup_sprite_tiled_builder(8bpp,); + +#undef draw_mask_fb_ptr_left +#undef draw_mask_fb_ptr_right + +setup_sprite_tiled_builder(4bpp, _4x); +setup_sprite_tiled_builder(8bpp, _4x); #undef block_ptr @@ -5038,6 +5339,12 @@ function(texture_sprite_blocks_8bpp) #undef texture_mask #undef num_blocks #undef texture_offset +#undef texels_low +#undef texels_high +#undef texels_wide_low +#undef texels_wide_high +#undef texels_wide +#undef fb_ptr2 #define psx_gpu r0 #define x r1 @@ -5049,6 +5356,7 @@ function(texture_sprite_blocks_8bpp) #define left_offset r8 #define width_rounded r9 #define right_width r10 + #define block_width r11 #define texture_offset_base r1 @@ -5059,6 +5367,7 @@ function(texture_sprite_blocks_8bpp) #define fb_ptr r7 #define texture_offset r8 #define blocks_remaining r9 +#define fb_ptr2 r10 #define fb_ptr_pitch r12 #define texture_block_ptr r14 @@ -5077,29 +5386,23 @@ function(texture_sprite_blocks_8bpp) #define draw_mask_fb_ptr d2 #define texels q2 +#define draw_mask_fb_ptr_a d2 +#define draw_mask_fb_ptr_b d3 +#define texels_low d4 +#define texels_high d5 +#define texels_wide_low d6 +#define texels_wide_high d7 +#define texels_wide q3 -setup_sprites_16bpp_flush_single: - vpush { d0 - d2 } - - stmdb sp!, { r0 - r3, r12, r14 } - bl flush_render_block_buffer - ldmia sp!, { r0 - r3, r12, r14 } - - vpop { d0 - d2 } - - add block, psx_gpu, #psx_gpu_blocks_offset - mov num_blocks, #1 - - bx lr -setup_sprites_16bpp_flush_row: - vpush { d0 - d2 } +setup_sprites_16bpp_flush: + vpush { d0 - d3 } stmdb sp!, { r0 - r3, r12, r14 } bl flush_render_block_buffer ldmia sp!, { r0 - r3, r12, r14 } - vpop { d0 - d2 } + vpop { d0 - d3 } add block, psx_gpu, #psx_gpu_blocks_offset mov num_blocks, block_width @@ -5108,7 +5411,7 @@ setup_sprites_16bpp_flush_row: function(setup_sprite_16bpp) stmdb sp!, { r4 - r11, r14 } - ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ] + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] ldr v, [ sp, #36 ] add fb_ptr, fb_ptr, y, lsl #11 @@ -5164,7 +5467,7 @@ function(setup_sprite_16bpp) 1: add num_blocks, num_blocks, #1 cmp num_blocks, #MAX_BLOCKS - blgt setup_sprites_16bpp_flush_single + blgt setup_sprites_16bpp_flush and texture_block_ptr, texture_offset_base, texture_mask subs height, height, #1 @@ -5193,7 +5496,7 @@ function(setup_sprite_16bpp) mov texture_offset, texture_offset_base cmp num_blocks, #MAX_BLOCKS - blgt setup_sprites_16bpp_flush_row + blgt setup_sprites_16bpp_flush add texture_offset_base, texture_offset_base, #2048 and texture_block_ptr, texture_offset, texture_mask @@ -5264,6 +5567,290 @@ function(setup_sprite_16bpp) ldmia sp!, { r4 - r11, pc } +// 4x version +// FIXME: duplicate code with normal version :( +#undef draw_mask_fb_ptr + +function(setup_sprite_16bpp_4x) + stmdb sp!, { r4 - r11, r14 } + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] + + ldr v, [ sp, #36 ] + add fb_ptr, fb_ptr, y, lsl #11 + + ldr width, [ sp, #40 ] + add fb_ptr, fb_ptr, x, lsl #1 + + ldr height, [ sp, #44 ] + and left_offset, u, #0x7 + + add texture_offset_base, u, u + add width_rounded, width, #7 + + add texture_offset_base, v, lsl #11 + movw left_mask_bits, #0xFFFF + + ldrb texture_mask_width, [ psx_gpu, #psx_gpu_texture_mask_width_offset ] + add width_rounded, width_rounded, left_offset + + lsl left_offset, #1 + + ldrb texture_mask_height, [ psx_gpu, #psx_gpu_texture_mask_height_offset ] + sub fb_ptr, fb_ptr, left_offset, lsl #1 + + add texture_mask, texture_mask_width, texture_mask_width + movw right_mask_bits, #0xFFFC + + and right_width, width_rounded, #0x7 + mvn left_mask_bits, left_mask_bits, lsl left_offset + + lsl right_width, #1 + + add texture_mask, texture_mask_height, lsl #11 + mov block_width, width_rounded, lsr #3 + + mov right_mask_bits, right_mask_bits, lsl right_width + movw fb_ptr_pitch, #(2048 + 16) * 2 + + sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4+1 + vmov block_masks, left_mask_bits, right_mask_bits + + ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] + add block, psx_gpu, #psx_gpu_blocks_offset + + bic texture_offset_base, texture_offset_base, #0xF + cmp block_width, #1 + + ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] + add block, block, num_blocks, lsl #6 + + lsl block_width, #2 + bne 0f + + vext.32 block_masks_shifted, block_masks, block_masks, #1 + vorr.u32 block_masks, block_masks, block_masks_shifted + vdup.u8 draw_mask_fb_ptr_a, block_masks[0] + vdup.u8 draw_mask_fb_ptr_b, block_masks[1] + + 1: + add num_blocks, num_blocks, block_width + cmp num_blocks, #MAX_BLOCKS + blgt setup_sprites_16bpp_flush + + and texture_block_ptr, texture_offset_base, texture_mask + subs height, height, #1 + + add texture_block_ptr, texture_page_ptr, texture_block_ptr + vld1.u32 { texels }, [ texture_block_ptr, :128 ] + + do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) + + add texture_offset_base, texture_offset_base, #2048 + add fb_ptr, fb_ptr, #2048*2 + strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] + bne 1b + + ldmia sp!, { r4 - r11, pc } + + 0: + add num_blocks, num_blocks, block_width + mov texture_offset, texture_offset_base + + vdup.u8 draw_mask_fb_ptr_a, block_masks[0] // left_mask_bits + vdup.u8 draw_mask_fb_ptr_b, block_masks[1] + + cmp num_blocks, #MAX_BLOCKS + blgt setup_sprites_16bpp_flush + + add texture_offset_base, texture_offset_base, #2048 + and texture_block_ptr, texture_offset, texture_mask + + add texture_block_ptr, texture_page_ptr, texture_block_ptr + vld1.u32 { texels }, [ texture_block_ptr, :128 ] + + do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) + + subs blocks_remaining, block_width, #2*4 + add texture_offset, texture_offset, #16 + + vmov.u8 draw_mask_fb_ptr_a, #0 + vmov.u8 draw_mask_fb_ptr_b, #0 + + add fb_ptr, fb_ptr, #16*2 + beq 2f + + 1: + and texture_block_ptr, texture_offset, texture_mask + subs blocks_remaining, blocks_remaining, #4 + + add texture_block_ptr, texture_page_ptr, texture_block_ptr + vld1.u32 { texels }, [ texture_block_ptr, :128 ] + + do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) + add texture_offset, texture_offset, #16 + + add fb_ptr, fb_ptr, #16*2 + bgt 1b + + 2: + vdup.u8 draw_mask_fb_ptr_a, block_masks[4] // right_mask_bits + vdup.u8 draw_mask_fb_ptr_b, block_masks[5] + + and texture_block_ptr, texture_offset, texture_mask + add texture_block_ptr, texture_page_ptr, texture_block_ptr + + vld1.u32 { texels }, [ texture_block_ptr, :128 ] + + do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) + subs height, height, #1 + + add fb_ptr, fb_ptr, fb_ptr_pitch + strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] + + bne 0b + + ldmia sp!, { r4 - r11, pc } + + +#undef width +#undef right_width +#undef right_mask_bits +#undef color +#undef height +#undef blocks_remaining +#undef colors +#undef right_mask +#undef test_mask +#undef draw_mask + +#define psx_gpu r0 +#define x r1 +#define y r2 +#define width r3 +#define right_width r5 +#define right_mask_bits r6 +#define fb_ptr r7 +#define color r8 +#define height r9 +#define fb_ptr_pitch r12 + +// referenced by setup_sprites_16bpp_flush +#define num_blocks r4 +#define block r5 +#define block_width r11 + +#define color_r r1 +#define color_g r2 +#define color_b r8 +#define blocks_remaining r6 + +#define colors q0 +#define right_mask q1 +#define test_mask q2 +#define draw_mask q2 +#define draw_mask_bits_fb_ptr d6 + + +.align 3 + +function(setup_sprite_untextured) + ldrh r12, [ psx_gpu, #psx_gpu_render_state_offset ] + tst r12, #(RENDER_STATE_MASK_EVALUATE | RENDER_FLAGS_MODULATE_TEXELS \ + | RENDER_FLAGS_BLEND) + beq setup_sprite_untextured_simple + + stmdb sp!, { r4 - r11, r14 } + + ldr width, [ sp, #40 ] + ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] + + ldr height, [ sp, #44 ] + add fb_ptr, fb_ptr, y, lsl #11 + + add fb_ptr, fb_ptr, x, lsl #1 + sub right_width, width, #1 + + ldr color, [ sp, #48 ] + and right_width, #7 + + add block_width, width, #7 + add right_width, #1 + + lsr block_width, #3 + mov right_mask_bits, #0xff + + sub fb_ptr_pitch, block_width, #1 + lsl right_mask_bits, right_width + + lsl fb_ptr_pitch, #3+1 + ubfx color_r, color, #3, #5 + + rsb fb_ptr_pitch, #1024*2 + ubfx color_g, color, #11, #5 + + vld1.u32 { test_mask }, [ psx_gpu, :128 ] + ubfx color_b, color, #19, #5 + + vdup.u16 right_mask, right_mask_bits + orr color, color_r, color_b, lsl #10 + + ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] + orr color, color, color_g, lsl #5 + + vtst.u16 right_mask, right_mask, test_mask + add block, psx_gpu, #psx_gpu_blocks_offset + + vdup.u16 colors, color + add block, block, num_blocks, lsl #6 + + +setup_sprite_untextured_height_loop: + add num_blocks, block_width + sub blocks_remaining, block_width, #1 + + cmp num_blocks, #MAX_BLOCKS + blgt setup_sprites_16bpp_flush + + cmp blocks_remaining, #0 + ble 1f + + vmov.u8 draw_mask, #0 /* zero_mask */ + vmov.u8 draw_mask_bits_fb_ptr, #0 + + 0: + vst1.u32 { draw_mask }, [ block, :128 ]! + subs blocks_remaining, #1 + + vst1.u32 { colors }, [ block, :128 ] + add block, block, #24 + + vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr + vst1.u32 { draw_mask_bits_fb_ptr }, [ block, :64 ] + + add block, block, #24 + add fb_ptr, #8*2 + bgt 0b + + 1: + vst1.u32 { right_mask }, [ block, :128 ]! + subs height, #1 + + vst1.u32 { colors }, [ block, :128 ] + add block, block, #24 + + vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr + vst1.u32 { draw_mask_bits_fb_ptr }, [ block, :64 ] + + add block, block, #24 + add fb_ptr, fb_ptr_pitch + + strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] + bgt setup_sprite_untextured_height_loop + + ldmia sp!, { r4 - r11, pc } + + + #undef texture_page_ptr #undef vram_ptr #undef dirty_textures_mask @@ -5461,3 +6048,40 @@ function(update_texture_8bpp_cache_slice) vpop { q0 - q3 } ldmia sp!, { r4 - r11, pc } + +/* void scale2x_tiles8(void *dst, const void *src, int w8, int h) */ +function(scale2x_tiles8) + push { r4, r14 } + + mov r4, r1 + add r12, r0, #1024*2 + mov r14, r2 + +0: + vld1.u16 { q0 }, [ r1, :128 ]! + vld1.u16 { q2 }, [ r1, :128 ]! + vmov q1, q0 + vmov q3, q2 + vzip.16 q0, q1 + vzip.16 q2, q3 + subs r14, #2 + vst1.u16 { q0, q1 }, [ r0, :128 ]! + vst1.u16 { q0, q1 }, [ r12, :128 ]! + blt 1f + vst1.u16 { q2, q3 }, [ r0, :128 ]! + vst1.u16 { q2, q3 }, [ r12, :128 ]! + bgt 0b +1: + subs r3, #1 + mov r14, r2 + add r0, #1024*2*2 + add r4, #1024*2 + sub r0, r2, lsl #4+1 + mov r1, r4 + add r12, r0, #1024*2 + bgt 0b + nop + + pop { r4, pc } + +// vim:filetype=armasm |