fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[D3D12] Experimental 2x resolution scale

Browse Source
This commit is contained in:
Triang3l
2018-12-06 10:19:07 +03:00
parent 132af3e266
commit 9427667a27
173 changed files with 16680 additions and 4985 deletions
Download Patch File Download Diff File Expand all files Collapse all files

297
src/xenia/gpu/dxbc_shader_translator.cc
View File

@@ -1037,8 +1037,8 @@ void DxbcShaderTranslator::StartPixelShader() {
++stat_.instruction_count;
++stat_.float_instruction_count;
}
// Floor VPOS so with AA, the resulting pixel corner/center is written, not
// the sample position (games likely don't expect it to be fractional).
// Floor VPOS so with SSAA, the resulting pixel corner/center is written, not
// the sample position.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ROUND_NI) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));
shader_code_.push_back(
@@ -1071,6 +1071,44 @@ void DxbcShaderTranslator::StartPixelShader() {
shader_code_.push_back(kSysConst_PixelHalfPixelOffset_Vec);
++stat_.instruction_count;
++stat_.float_instruction_count;
// Undo 2x resolution scale in VPOS.
if (edram_rov_used_) {
// Get inverse of the width/height scale.
system_constants_used_ |= 1ull << kSysConst_EDRAMResolutionScaleLog2_Index;
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_IMAD) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(11));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0100, 1));
shader_code_.push_back(param_gen_value_temp);
shader_code_.push_back(
EncodeVectorSelectOperand(D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_EDRAMResolutionScaleLog2_Comp, 3));
shader_code_.push_back(cbuffer_index_system_constants_);
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_EDRAMResolutionScaleLog2_Vec);
shader_code_.push_back(
EncodeScalarOperand(D3D10_SB_OPERAND_TYPE_IMMEDIATE32, 0));
shader_code_.push_back(uint32_t(-(1 << 23)));
shader_code_.push_back(
EncodeScalarOperand(D3D10_SB_OPERAND_TYPE_IMMEDIATE32, 0));
shader_code_.push_back(0x3F800000);
++stat_.instruction_count;
++stat_.int_instruction_count;
// Convert to guest pixels.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MUL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0011, 1));
shader_code_.push_back(param_gen_value_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_INPUT, kSwizzleXYZW, 1));
shader_code_.push_back(uint32_t(InOutRegister::kPSInPosition));
shader_code_.push_back(
EncodeVectorReplicatedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 2, 1));
shader_code_.push_back(param_gen_value_temp);
++stat_.instruction_count;
++stat_.float_instruction_count;
}
// Write point sprite coordinates to ZW.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));
@@ -2431,6 +2469,97 @@ void DxbcShaderTranslator::CompletePixelShader_WriteToROV_DepthStencil(
++stat_.instruction_count;
++stat_.int_instruction_count;
// Apply pixel width and height scale.
system_constants_used_ |= 1ull << kSysConst_EDRAMResolutionScaleLog2_Index;
for (uint32_t i = 0; i < 2; ++i) {
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ISHL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(edram_dword_offset_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(edram_dword_offset_temp);
shader_code_.push_back(EncodeVectorReplicatedOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_EDRAMResolutionScaleLog2_Comp, 3));
shader_code_.push_back(cbuffer_index_system_constants_);
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_EDRAMResolutionScaleLog2_Vec);
++stat_.instruction_count;
++stat_.int_instruction_count;
}
// Choose the pixel for 2x scaling.
uint32_t resolution_scale_pixel_temp = PushSystemTemp();
// 1) Convert pixel position to integer.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_FTOU) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0011, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_INPUT, kSwizzleXYZW, 1));
shader_code_.push_back(uint32_t(InOutRegister::kPSInPosition));
++stat_.instruction_count;
++stat_.conversion_instruction_count;
// 2) For 2x, get the current pixel in the quad. For 1x, write 0 for it.
system_constants_used_ |= 1ull << kSysConst_EDRAMResolutionScaleLog2_Index;
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_AND) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0011, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(
EncodeVectorSwizzledOperand(D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(EncodeVectorReplicatedOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_EDRAMResolutionScaleLog2_Comp, 3));
shader_code_.push_back(cbuffer_index_system_constants_);
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_EDRAMResolutionScaleLog2_Vec);
++stat_.instruction_count;
++stat_.uint_instruction_count;
// 3) Calculate dword offset of the pixel in the quad.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_UMAD) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0001, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(
EncodeVectorSelectOperand(D3D10_SB_OPERAND_TYPE_TEMP, 1, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(
EncodeScalarOperand(D3D10_SB_OPERAND_TYPE_IMMEDIATE32, 0));
shader_code_.push_back(2);
shader_code_.push_back(
EncodeVectorSelectOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
++stat_.instruction_count;
++stat_.uint_instruction_count;
// 4) Add the quad pixel offset.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_IADD) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(edram_dword_offset_temp);
shader_code_.push_back(
EncodeVectorSwizzledOperand(D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(edram_dword_offset_temp);
shader_code_.push_back(
EncodeVectorReplicatedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
++stat_.instruction_count;
++stat_.int_instruction_count;
// Release resolution_scale_pixel_temp.
PopSystemTemp();
// Load the previous depth/stencil values.
uint32_t depth_values_temp = PushSystemTemp();
for (uint32_t i = 0; i < 4; ++i) {
@@ -5132,6 +5261,26 @@ void DxbcShaderTranslator::CompletePixelShader_WriteToROV() {
++stat_.instruction_count;
++stat_.conversion_instruction_count;
// Get guest pixel position as if increased resolution is disabled - addresses
// within the quad with 2x resolution will be calculated later.
system_constants_used_ |= 1ull << kSysConst_EDRAMResolutionScaleLog2_Index;
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_USHR) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0011, 1));
shader_code_.push_back(edram_coord_pixel_temp);
shader_code_.push_back(
EncodeVectorSwizzledOperand(D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(edram_coord_pixel_temp);
shader_code_.push_back(EncodeVectorReplicatedOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_EDRAMResolutionScaleLog2_Comp, 3));
shader_code_.push_back(cbuffer_index_system_constants_);
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_EDRAMResolutionScaleLog2_Vec);
++stat_.instruction_count;
++stat_.uint_instruction_count;
// Convert the position from pixels to samples.
system_constants_used_ |= 1ull << kSysConst_SampleCountLog2_Index;
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ISHL) |
@@ -5590,6 +5739,112 @@ void DxbcShaderTranslator::CompletePixelShader_WriteToROV() {
// Release rt_write_masks_temp.
PopSystemTemp();
// Apply pixel width and height scale.
system_constants_used_ |= 1ull << kSysConst_EDRAMResolutionScaleLog2_Index;
for (uint32_t i = 0; i < 2; ++i) {
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ISHL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(edram_coord_pixel_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(edram_coord_pixel_temp);
shader_code_.push_back(EncodeVectorReplicatedOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_EDRAMResolutionScaleLog2_Comp, 3));
shader_code_.push_back(cbuffer_index_system_constants_);
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_EDRAMResolutionScaleLog2_Vec);
++stat_.instruction_count;
++stat_.int_instruction_count;
}
// Choose the pixel for 2x scaling.
uint32_t resolution_scale_pixel_temp = PushSystemTemp();
// 1) Convert pixel position to integer.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_FTOU) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0011, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_INPUT, kSwizzleXYZW, 1));
shader_code_.push_back(uint32_t(InOutRegister::kPSInPosition));
++stat_.instruction_count;
++stat_.conversion_instruction_count;
// 2) For 2x, get the current pixel in the quad. For 1x, write 0 for it.
system_constants_used_ |= 1ull << kSysConst_EDRAMResolutionScaleLog2_Index;
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_AND) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0011, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(EncodeVectorReplicatedOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_EDRAMResolutionScaleLog2_Comp, 3));
shader_code_.push_back(cbuffer_index_system_constants_);
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_EDRAMResolutionScaleLog2_Vec);
++stat_.instruction_count;
++stat_.uint_instruction_count;
// 3) Calculate dword offset of the pixel in the quad.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_UMAD) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0001, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(
EncodeVectorSelectOperand(D3D10_SB_OPERAND_TYPE_TEMP, 1, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(
EncodeScalarOperand(D3D10_SB_OPERAND_TYPE_IMMEDIATE32, 0));
shader_code_.push_back(2);
shader_code_.push_back(
EncodeVectorSelectOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
++stat_.instruction_count;
++stat_.uint_instruction_count;
// 4) Multiply the quad pixel offset by dword count per pixel for each RT.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ISHL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(
EncodeVectorReplicatedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(rt_64bpp_temp);
++stat_.instruction_count;
++stat_.int_instruction_count;
// 5) Add the quad pixel offsets.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_IADD) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(edram_coord_pixel_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(edram_coord_pixel_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(resolution_scale_pixel_temp);
++stat_.instruction_count;
++stat_.int_instruction_count;
// Release resolution_scale_pixel_temp.
PopSystemTemp();
// Get what render targets need gamma conversion.
uint32_t rt_gamma_temp = PushSystemTemp();
system_constants_used_ |= 1ull << kSysConst_Flags_Index;
@@ -5757,27 +6012,54 @@ void DxbcShaderTranslator::CompletePixelShader_WriteToROV() {
++stat_.instruction_count;
++stat_.uint_instruction_count;
// Get per-sample EDRAM addresses offsets. First, multiply the relative
// sample offset by sample size.
// Get per-sample EDRAM addresses offsets.
uint32_t edram_coord_sample_temp = PushSystemTemp();
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ISHL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(10));
// 1) Choose the strides according to the resolution scale (1x or 2x2x).
system_constants_used_ |= 1ull
<< kSysConst_EDRAMResolutionScaleLog2_Index;
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOVC) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(17));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(edram_coord_sample_temp);
shader_code_.push_back(EncodeVectorReplicatedOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_EDRAMResolutionScaleLog2_Comp, 3));
shader_code_.push_back(cbuffer_index_system_constants_);
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_EDRAMResolutionScaleLog2_Vec);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_IMMEDIATE32, kSwizzleXYZW, 0));
shader_code_.push_back(0);
shader_code_.push_back(320);
shader_code_.push_back(4);
shader_code_.push_back(324);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_IMMEDIATE32, kSwizzleXYZW, 0));
shader_code_.push_back(0);
shader_code_.push_back(80);
shader_code_.push_back(1);
shader_code_.push_back(81);
++stat_.instruction_count;
++stat_.movc_instruction_count;
// 2) Multiply the relative sample offset by sample size.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_ISHL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(edram_coord_sample_temp);
shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
shader_code_.push_back(edram_coord_sample_temp);
shader_code_.push_back(
EncodeVectorReplicatedOperand(D3D10_SB_OPERAND_TYPE_TEMP, i, 1));
shader_code_.push_back(rt_64bpp_temp);
++stat_.instruction_count;
++stat_.int_instruction_count;
// Add the EDRAM base for the render target to the sample EDRAM addresses.
// 3) Add the first sample EDRAM addresses to the sample offsets.
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_IADD) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
shader_code_.push_back(
@@ -12978,6 +13260,7 @@ const DxbcShaderTranslator::SystemConstantRdef DxbcShaderTranslator::
{"xe_edram_poly_offset_front", RdefTypeIndex::kFloat2, 136, 8},
// vec4 9
{"xe_edram_poly_offset_back", RdefTypeIndex::kFloat2, 144, 8},
{"xe_edram_resolution_scale_log2", RdefTypeIndex::kUint, 152, 4},
// vec4 10
{"xe_edram_stencil_reference", RdefTypeIndex::kUint, 160, 4},
{"xe_edram_stencil_read_mask", RdefTypeIndex::kUint, 164, 4},