u32 span_pixels = 0;
u32 span_pixel_blocks = 0;
-u32 span_pixel_blocks_unaligned = 0;
u32 spans = 0;
u32 triangles = 0;
u32 sprites = 0;
u32 texel_blocks_16bpp = 0;
u32 texel_blocks_untextured = 0;
u32 blend_blocks = 0;
-u32 untextured_pixels = 0;
-u32 blend_pixels = 0;
-u32 transparent_pixels = 0;
u32 render_buffer_flushes = 0;
u32 state_changes = 0;
u32 left_split_triangles = 0;
u32 clipped_triangles = 0;
u32 zero_block_spans = 0;
u32 texture_cache_loads = 0;
-u32 false_modulated_triangles = 0;
-u32 false_modulated_sprites = 0;
+u32 false_modulated_blocks = 0;
u32 reciprocal_table[512];
{
u32 mask = texture_region_mask(x1, y1, x2, y2) &
psx_gpu->viewport_mask;
+
psx_gpu->dirty_textures_4bpp_mask |= mask;
psx_gpu->dirty_textures_8bpp_mask |= mask;
psx_gpu->dirty_textures_8bpp_alternate_mask |= mask;
return mask;
}
+void update_texture_cache_region(psx_gpu_struct *psx_gpu, u32 x1, u32 y1,
+ u32 x2, u32 y2)
+{
+ u32 mask = texture_region_mask(x1, y1, x2, y2);
+ u32 texture_page;
+ u8 *texture_page_ptr;
+ u16 *vram_ptr;
+ u32 texel_block;
+ u32 sub_x, sub_y;
+
+ psx_gpu->dirty_textures_8bpp_mask |= mask;
+ psx_gpu->dirty_textures_8bpp_alternate_mask |= mask;
+
+ if ((psx_gpu->dirty_textures_4bpp_mask & mask) == 0 &&
+ (x1 & 3) == 0 && (y1 & 15) == 0 && x2 - x1 < 4 && y2 - y1 < 16)
+ {
+ texture_page = ((x1 / 64) & 15) + (y1 / 256) * 16;
+ texture_page_ptr = psx_gpu->texture_4bpp_cache[texture_page];
+ texture_page_ptr += (x1 / 4 & 15) * 16*16 + (y1 / 16 & 15) * 16*16*16;
+ vram_ptr = psx_gpu->vram_ptr + x1 + y1 * 1024;
+ sub_x = 4;
+ sub_y = 16;
+
+ while(sub_y)
+ {
+ while(sub_x)
+ {
+ texel_block = *vram_ptr;
+
+ texture_page_ptr[0] = texel_block & 0xF;
+ texture_page_ptr[1] = (texel_block >> 4) & 0xF;
+ texture_page_ptr[2] = (texel_block >> 8) & 0xF;
+ texture_page_ptr[3] = texel_block >> 12;
+
+ vram_ptr++;
+ texture_page_ptr += 4;
+
+ sub_x--;
+ }
+
+ vram_ptr -= 4;
+ sub_x = 4;
+
+ sub_y--;
+ vram_ptr += 1024;
+ }
+ }
+ else
+ {
+ psx_gpu->dirty_textures_4bpp_mask |= mask;
+ }
+}
void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
u32 texture_page);
void update_texture_4bpp_cache(psx_gpu_struct *psx_gpu)
{
u32 current_texture_page = psx_gpu->current_texture_page;
- u8 *texture_page_ptr = psx_gpu->texture_page_ptr;
+ u8 *texture_page_ptr = psx_gpu->texture_page_base;
u16 *vram_ptr = psx_gpu->vram_ptr;
u32 texel_block;
while(sub_x)
{
texel_block = *vram_ptr;
+
texture_page_ptr[0] = texel_block & 0xF;
texture_page_ptr[1] = (texel_block >> 4) & 0xF;
texture_page_ptr[2] = (texel_block >> 8) & 0xF;
void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
u32 texture_page)
{
- u16 *texture_page_ptr = psx_gpu->texture_page_ptr;
+ u16 *texture_page_ptr = psx_gpu->texture_page_base;
u16 *vram_ptr = psx_gpu->vram_ptr;
u32 tile_x, tile_y;
void flush_render_block_buffer(psx_gpu_struct *psx_gpu)
{
+ if((psx_gpu->interlace_mode & RENDER_INTERLACE_ENABLED) &&
+ (psx_gpu->primitive_type == PRIMITIVE_TYPE_SPRITE))
+ {
+ u32 num_blocks_dest = 0;
+ block_struct *block_src = psx_gpu->blocks;
+ block_struct *block_dest = psx_gpu->blocks;
+
+ u16 *vram_ptr = psx_gpu->vram_ptr;
+ u32 i;
+
+ if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+ {
+ for(i = 0; i < psx_gpu->num_blocks; i++)
+ {
+ u32 fb_offset = (u32)((u8 *)block_src->fb_ptr - (u8 *)vram_ptr);
+ if(fb_offset & (1 << 11))
+ {
+ *block_dest = *block_src;
+ num_blocks_dest++;
+ block_dest++;
+ }
+ block_src++;
+ }
+ }
+ else
+ {
+ for(i = 0; i < psx_gpu->num_blocks; i++)
+ {
+ u32 fb_offset = (u32)((u8 *)block_src->fb_ptr - (u8 *)vram_ptr);
+ if((fb_offset & (1 << 11)) == 0)
+ {
+ *block_dest = *block_src;
+ num_blocks_dest++;
+ block_dest++;
+ }
+ block_src++;
+ }
+ }
+
+ psx_gpu->num_blocks = num_blocks_dest;
+ }
+
if(psx_gpu->num_blocks)
{
render_block_handler_struct *render_block_handler =
render_block_handler->shade_blocks(psx_gpu);
render_block_handler->blend_blocks(psx_gpu);
+#ifdef PROFILE
span_pixel_blocks += psx_gpu->num_blocks;
render_buffer_flushes++;
+#endif
psx_gpu->num_blocks = 0;
}
} \
#define setup_blocks_add_blocks_direct() \
+ texel_blocks_untextured += span_num_blocks; \
+ span_pixel_blocks += span_num_blocks \
#define setup_blocks_builder(shading, texturing, dithering, sw, target) \
s32 pixel_span = span_num_blocks * 8; \
pixel_span -= __builtin_popcount(span_edge_data->right_mask & 0xFF); \
span_pixels += pixel_span; \
- span_pixel_blocks_unaligned += (pixel_span + 7) / 8; \
\
span_num_blocks--; \
while(span_num_blocks) \
} \
-#define shade_blocks_textured_modulated_shaded_primitive_load() \
+#define shade_blocks_textured_false_modulated_check_dithered(target) \
+ if(psx_gpu->triangle_color == 0x808080) \
+ { \
+ false_modulated_blocks += num_blocks; \
+ } \
+
+#define shade_blocks_textured_false_modulated_check_undithered(target) \
+ if(psx_gpu->triangle_color == 0x808080) \
+ { \
+ \
+ shade_blocks_textured_unmodulated_##target(psx_gpu); \
+ false_modulated_blocks += num_blocks; \
+ return; \
+ } \
+
+
+#define shade_blocks_textured_modulated_shaded_primitive_load(dithering, \
+ target) \
-#define shade_blocks_textured_modulated_unshaded_primitive_load() \
+#define shade_blocks_textured_modulated_unshaded_primitive_load(dithering, \
+ target) \
{ \
u32 color = psx_gpu->triangle_color; \
dup_8x8b(colors_r, color); \
dup_8x8b(colors_g, color >> 8); \
dup_8x8b(colors_b, color >> 16); \
- if(psx_gpu->triangle_color == 0x808080) \
- false_modulated_triangles++; \
+ shade_blocks_textured_false_modulated_check_##dithering(target); \
} \
#define shade_blocks_textured_modulated_shaded_block_load() \
\
dup_8x16b(d128_0x8000, 0x8000); \
\
- shade_blocks_textured_modulated_##shading##_primitive_load(); \
+ shade_blocks_textured_modulated_##shading##_primitive_load(dithering, \
+ target); \
\
while(num_blocks) \
{ \
void shade_blocks_unshaded_textured_modulated_undithered_indirect(psx_gpu_struct
*psx_gpu);
+void shade_blocks_textured_unmodulated_indirect(psx_gpu_struct *psx_gpu);
+void shade_blocks_textured_unmodulated_direct(psx_gpu_struct *psx_gpu);
+
#ifndef NEON_BUILD
shade_blocks_textured_modulated_builder(shaded, dithered, direct);
} \
} \
-void shade_blocks_textured_unmodulated_indirect(psx_gpu_struct *psx_gpu);
-void shade_blocks_textured_unmodulated_direct(psx_gpu_struct *psx_gpu);
+#define shade_blocks_textured_unmodulated_dithered_builder(target) \
+void shade_blocks_textured_unmodulated_dithered_##target(psx_gpu_struct \
+ *psx_gpu) \
+{ \
+ block_struct *block = psx_gpu->blocks; \
+ u32 num_blocks = psx_gpu->num_blocks; \
+ vec_8x16u draw_mask; \
+ vec_8x16u test_mask = psx_gpu->test_mask; \
+ u32 draw_mask_bits; \
+ \
+ vec_8x16u pixels; \
+ shade_blocks_load_msb_mask_##target(); \
+ \
+ while(num_blocks) \
+ { \
+ vec_8x16u zero_mask; \
+ \
+ draw_mask_bits = block->draw_mask_bits; \
+ dup_8x16b(draw_mask, draw_mask_bits); \
+ tst_8x16b(draw_mask, draw_mask, test_mask); \
+ \
+ pixels = block->texels; \
+ \
+ cmpeqz_8x16b(zero_mask, pixels); \
+ or_8x16b(zero_mask, draw_mask, zero_mask); \
+ \
+ shade_blocks_store_##target(zero_mask, pixels); \
+ \
+ num_blocks--; \
+ block++; \
+ } \
+} \
#ifndef NEON_BUILD
void shade_blocks_unshaded_untextured_direct(psx_gpu_struct *psx_gpu);
#ifndef NEON_BUILD
-
+
void shade_blocks_unshaded_untextured_indirect(psx_gpu_struct *psx_gpu)
{
}
triangle_area = triangle_signed_area_x2(a->x, a->y, b->x, b->y, c->x, c->y);
+#ifdef PROFILE
triangles++;
+#endif
if(triangle_area == 0)
{
+#ifdef PROFILE
trivial_rejects++;
+#endif
return;
}
if((y_bottom - y_top) >= 512)
{
+#ifdef PROFILE
trivial_rejects++;
+#endif
return;
}
if((c->x - a->x) >= 1024)
{
+#ifdef PROFILE
trivial_rejects++;
+#endif
return;
}
if(invalidate_texture_cache_region_viewport(psx_gpu, a->x, y_top, c->x,
y_bottom) == 0)
{
+#ifdef PROFILE
trivial_rejects++;
+#endif
return;
}
break;
}
+#ifdef PROFILE
spans += psx_gpu->num_spans;
+#endif
+
+ if(psx_gpu->interlace_mode & RENDER_INTERLACE_ENABLED)
+ {
+ u32 i;
+
+ if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+ {
+ for(i = 0; i < psx_gpu->num_spans; i++)
+ {
+ if((psx_gpu->span_edge_data[i].y & 1) == 0)
+ psx_gpu->span_edge_data[i].num_blocks = 0;
+ }
+ }
+ else
+ {
+ for(i = 0; i < psx_gpu->num_spans; i++)
+ {
+ if(psx_gpu->span_edge_data[i].y & 1)
+ psx_gpu->span_edge_data[i].num_blocks = 0;
+ }
+ }
+ }
u32 render_state = flags &
(RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND |
{
psx_gpu->render_state = render_state;
flush_render_block_buffer(psx_gpu);
+#ifdef PROFILE
state_changes++;
+#endif
}
psx_gpu->primitive_type = PRIMITIVE_TYPE_TRIANGLE;
u32 left_offset = u & 0x7;
u32 width_rounded = width + left_offset + 7;
- u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + (x - left_offset);
+ u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + (s32)(x - left_offset);
u32 right_width = width_rounded & 0x7;
u32 block_width = width_rounded / 8;
u32 fb_ptr_pitch = (1024 + 8) - (block_width * 8);
{
blocks_remaining = block_width - 1;
num_blocks += block_width;
+
+#ifdef PROFILE
sprite_blocks += block_width;
+#endif
if(num_blocks > MAX_BLOCKS)
{
s32 x_right = x + width - 1;
s32 y_bottom = y + height - 1;
+#ifdef PROFILE
+ sprites++;
+#endif
+
if(invalidate_texture_cache_region_viewport(psx_gpu, x, y, x_right,
y_bottom) == 0)
{
if((width <= 0) || (height <= 0))
return;
- sprites++;
-
+#ifdef PROFILE
span_pixels += width * height;
spans += height;
+#endif
u32 render_state = flags &
(RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND |
RENDER_FLAGS_TEXTURE_MAP);
render_state |=
(psx_gpu->render_state_base & ~RENDER_STATE_DITHER);
-
+
if((psx_gpu->render_state != render_state) ||
(psx_gpu->primitive_type != PRIMITIVE_TYPE_SPRITE))
{
psx_gpu->render_state = render_state;
flush_render_block_buffer(psx_gpu);
+#ifdef PROFILE
state_changes++;
+#endif
}
psx_gpu->primitive_type = PRIMITIVE_TYPE_SPRITE;
u32 control_mask;
+#ifdef PROFILE
lines++;
+#endif
if(vertex_a->x >= vertex_b->x)
{
void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
u32 width, u32 height)
{
+ if((width == 0) || (height == 0))
+ return;
+
+ flush_render_block_buffer(psx_gpu);
invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
-#ifndef NEON_BUILD
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);
+ u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
+ psx_gpu->mask_msb;
+ u32 color_32bpp = color_16bpp | (color_16bpp << 16);
- u16 *vram_ptr = psx_gpu->vram_ptr + x + (y * 1024);
- u32 draw_x, draw_y;
+ u32 *vram_ptr = (u32 *)(psx_gpu->vram_ptr + x + (y * 1024));
- for(draw_y = 0; draw_y < height; draw_y++)
+ u32 pitch = 512 - (width / 2);
+ u32 num_width;
+
+ if(psx_gpu->interlace_mode & RENDER_INTERLACE_ENABLED)
{
- for(draw_x = 0; draw_x < width; draw_x++)
+ pitch += 512;
+ height /= 2;
+
+ if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+ vram_ptr += 512;
+ }
+
+ while(height)
+ {
+ num_width = width;
+ while(num_width)
{
- vram_ptr[draw_x] = color_16bpp;
+ vram_ptr[0] = color_32bpp;
+ vram_ptr[1] = color_32bpp;
+ vram_ptr[2] = color_32bpp;
+ vram_ptr[3] = color_32bpp;
+ vram_ptr[4] = color_32bpp;
+ vram_ptr[5] = color_32bpp;
+ vram_ptr[6] = color_32bpp;
+ vram_ptr[7] = color_32bpp;
+
+ vram_ptr += 8;
+ num_width -= 16;
}
- vram_ptr += 1024;
+ vram_ptr += pitch;
+ height--;
}
-#else
- void render_block_fill_body(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
- u32 width, u32 height);
-
- render_block_fill_body(psx_gpu, color, x, y, width, height);
-#endif
}
void render_block_copy(psx_gpu_struct *psx_gpu, u16 *source, u32 x, u32 y,
{
u16 *vram_ptr = psx_gpu->vram_ptr + x + (y * 1024);
u32 draw_x, draw_y;
+ u32 mask_msb = psx_gpu->mask_msb;
- invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
+ if((width == 0) || (height == 0))
+ return;
- //printf("copy for %d, %d\n", width, height);
+ flush_render_block_buffer(psx_gpu);
+ invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
for(draw_y = 0; draw_y < height; draw_y++)
{
for(draw_x = 0; draw_x < width; draw_x++)
{
- vram_ptr[draw_x] = source[draw_x];
+ vram_ptr[draw_x] = source[draw_x] | mask_msb;
}
source += pitch;
psx_gpu->test_mask = test_mask;
- psx_gpu->pixel_count_mode = 0;
- psx_gpu->pixel_compare_mode = 0;
-
- psx_gpu->vram_pixel_counts_a = malloc(sizeof(u8) * 1024 * 512);
- psx_gpu->vram_pixel_counts_b = malloc(sizeof(u8) * 1024 * 512);
- memset(psx_gpu->vram_pixel_counts_a, 0, sizeof(u8) * 1024 * 512);
- memset(psx_gpu->vram_pixel_counts_b, 0, sizeof(u8) * 1024 * 512);
- psx_gpu->compare_vram = malloc(sizeof(u16) * 1024 * 512);
-
psx_gpu->dirty_textures_4bpp_mask = 0xFFFFFFFF;
psx_gpu->dirty_textures_8bpp_mask = 0xFFFFFFFF;
psx_gpu->dirty_textures_8bpp_alternate_mask = 0xFFFFFFFF;
psx_gpu->vram_ptr = vram;
+ psx_gpu->texture_page_base = psx_gpu->vram_ptr;
psx_gpu->texture_page_ptr = psx_gpu->vram_ptr;
psx_gpu->clut_ptr = psx_gpu->vram_ptr;
psx_gpu->mask_msb = 0;
+ psx_gpu->texture_window_x = 0;
+ psx_gpu->texture_window_y = 0;
+ psx_gpu->texture_mask_width = 0xFF;
+ psx_gpu->texture_mask_height = 0xFF;
+
+ psx_gpu->interlace_mode = 0;
+
memset(psx_gpu->vram_ptr, 0, sizeof(u16) * 1024 * 512);
initialize_reciprocal_table();
// d1: (2 3 6 7): y0
// d2: (4 5 6 7): x0 ^ y0
-
psx_gpu->dither_table[0] = dither_table_row(-4, 0, -3, 1);
psx_gpu->dither_table[1] = dither_table_row(2, -2, 3, -1);
psx_gpu->dither_table[2] = dither_table_row(-3, 1, -4, 0);
notaz.gp2x.de / pcsx_rearmed.git / blobdiff