ShuriZma
/
xenia-canary
mirror of https://github.com/xenia-canary/xenia-canary.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[D3D12] Restore W and apply half-pixel offset in VS

This commit is contained in:
Triang3l 2018-07-31 17:13:44 +03:00
parent fb1051b610
commit bb045cae70
3 changed files with 47 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
src/xenia/gpu/d3d12/d3d12_command_processor.cc
View File

@ -793,6 +793,9 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(Endian index_endian) {
// shaders (for rectangle list drawing, for instance) to the 2560x2560
// viewport that is used to emulate unnormalized coordinates.
// Z scale/offset is to convert from OpenGL NDC to Direct3D NDC if needed.
// Also apply half-pixel offset to reproduce Direct3D 9 rasterization rules.
// TODO(Triang3l): Check if pixel coordinates need to offset depending on a
// different register.
bool gl_clip_space_def =
!(pa_cl_clip_cntl & (1 << 19)) && (pa_cl_vte_cntl & (1 << 4));
float ndc_scale_x = (pa_cl_vte_cntl & (1 << 0)) ? 1.0f / 1280.0f : 1.0f;
@ -801,51 +804,39 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(Endian index_endian) {
float ndc_offset_x = (pa_cl_vte_cntl & (1 << 1)) ? -1.0f : 0.0f;
float ndc_offset_y = (pa_cl_vte_cntl & (1 << 3)) ? -1.0f : 0.0f;
float ndc_offset_z = gl_clip_space_def ? 0.5f : 0.0f;
float pixel_half_pixel_offset = 0.0f;
if (pa_su_vtx_cntl & (1 << 0)) {
if (pa_cl_vte_cntl & (1 << 0)) {
float viewport_scale_x = regs[XE_GPU_REG_PA_CL_VPORT_XSCALE].f32;
if (viewport_scale_x != 0.0f) {
ndc_offset_x -= 0.5f / viewport_scale_x;
}
} else {
ndc_offset_x -= 1.0f / 2560.0f;
}
if (pa_cl_vte_cntl & (1 << 2)) {
float viewport_scale_y = regs[XE_GPU_REG_PA_CL_VPORT_YSCALE].f32;
if (viewport_scale_y != 0.0f) {
ndc_offset_y -= 0.5f / viewport_scale_y;
}
} else {
ndc_offset_y -= 1.0f / 2560.0f;
}
pixel_half_pixel_offset = -0.5f;
}
dirty |= system_constants_.ndc_scale[0] != ndc_scale_x;
dirty |= system_constants_.ndc_scale[1] != ndc_scale_y;
dirty |= system_constants_.ndc_scale[2] != ndc_scale_z;
dirty |= system_constants_.ndc_offset[0] != ndc_offset_x;
dirty |= system_constants_.ndc_offset[1] != ndc_offset_y;
dirty |= system_constants_.ndc_offset[2] != ndc_offset_z;
dirty |= system_constants_.pixel_half_pixel_offset != pixel_half_pixel_offset;
system_constants_.ndc_scale[0] = ndc_scale_x;
system_constants_.ndc_scale[1] = ndc_scale_y;
system_constants_.ndc_scale[2] = ndc_scale_z;
system_constants_.ndc_offset[0] = ndc_offset_x;
system_constants_.ndc_offset[1] = ndc_offset_y;
system_constants_.ndc_offset[2] = ndc_offset_z;
// Half-pixel offset for vertex and pixel coordinates.
// TODO(Triang3l): Check if pixel coordinates need to offset depending on a
// different register.
float vertex_half_pixel_offset[2], pixel_half_pixel_offset;
if (pa_su_vtx_cntl & (1 << 0)) {
if (pa_cl_vte_cntl & (1 << 0)) {
float viewport_scale_x = regs[XE_GPU_REG_PA_CL_VPORT_XSCALE].f32;
vertex_half_pixel_offset[0] =
viewport_scale_x != 0.0f ? -0.5f / viewport_scale_x : 0.0f;
} else {
vertex_half_pixel_offset[0] = -1.0f / 2560.0f;
}
if (pa_cl_vte_cntl & (1 << 2)) {
float viewport_scale_y = regs[XE_GPU_REG_PA_CL_VPORT_YSCALE].f32;
vertex_half_pixel_offset[1] =
viewport_scale_y != 0.0f ? -0.5f / viewport_scale_y : 0.0f;
} else {
vertex_half_pixel_offset[1] = -1.0f / 2560.0f;
}
pixel_half_pixel_offset = -0.5f;
} else {
vertex_half_pixel_offset[0] = 0.0f;
vertex_half_pixel_offset[1] = 0.0f;
pixel_half_pixel_offset = 0.0f;
}
dirty |= system_constants_.vertex_half_pixel_offset[0] !=
vertex_half_pixel_offset[0];
dirty |= system_constants_.vertex_half_pixel_offset[1] !=
vertex_half_pixel_offset[1];
dirty |= system_constants_.pixel_half_pixel_offset != pixel_half_pixel_offset;
system_constants_.vertex_half_pixel_offset[0] = vertex_half_pixel_offset[0];
system_constants_.vertex_half_pixel_offset[1] = vertex_half_pixel_offset[1];
system_constants_.pixel_half_pixel_offset = pixel_half_pixel_offset;
// Pixel position register.

23
src/xenia/gpu/hlsl_shader_translator.cc
View File

@ -167,9 +167,14 @@ std::vector<uint8_t> HlslShaderTranslator::CompleteTranslation() {
// Only up to 14 constant buffers can be used on binding tiers 1 and 2.
source.Append(
"cbuffer xe_system_constants : register(b0) {\n"
" float2 xe_viewport_inv_scale;\n"
" float3 xe_mul_rcp_w;\n"
" uint xe_vertex_index_endian;\n"
" float3 xe_ndc_scale;\n"
" uint xe_textures_are_3d;\n"
" float3 xe_ndc_offset;\n"
" float xe_pixel_half_pixel_offset;\n"
" float2 xe_ssaa_inv_scale;\n"
" uint xe_pixel_pos_reg;\n"
"};\n"
"\n"
"cbuffer xe_loop_bool_constants : register(b1) {\n"
@ -313,6 +318,22 @@ std::vector<uint8_t> HlslShaderTranslator::CompleteTranslation() {
" break;\n"
" }\n"
" } while (xe_pc != 0xFFFFu);\n");
if (is_vertex_shader()) {
// Restore the original W if the shader has already taken its reciprocal,
// and restore the original XYZ if the shader has divided them by W. Also
// normalize the coordinates to the viewport if the shader has returned
// unnormalized ones (for rectangle lists, for instance) and apply the half-
// pixel offset.
source.Append(
" [flatten] if (xe_mul_rcp_w.z == 0.0) {\n"
" xe_output.position.w = rcp(xe_output.position.w);\n"
" }\n"
" xe_output.position.xyz *=\n"
" lerp((1.0).xxx, xe_output.position.www, xe_mul_rcp_w.xxy);\n"
" xe_output.position.xyz =\n"
" xe_output.position.xyz * xe_ndc_scale +\n"
" xe_ndc_offset * xe_output.position.www;\n");
}
// TODO(Triang3l): Window offset, half pixel offset, alpha test, gamma.
source.Append(
" return xe_output;\n"

4
src/xenia/gpu/hlsl_shader_translator.h
View File

@ -35,10 +35,8 @@ class HlslShaderTranslator : public ShaderTranslator {
float ndc_offset[3];
float pixel_half_pixel_offset;
// vec4 3
float vertex_half_pixel_offset[2];
uint32_t pixel_pos_reg;
// vec4 4
float ssaa_inv_scale[2];
uint32_t pixel_pos_reg;
};
enum class SRVType : uint32_t {