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

[D3D12] Format DXBC and HLSL shader translator code

This commit is contained in:
Triang3l 2018-09-09 14:12:40 +03:00
parent 01bbba32cb
commit 3b7d6bfdcb
2 changed files with 55 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

61
src/xenia/gpu/dxbc_shader_translator.cc
View File

@ -220,9 +220,9 @@ void DxbcShaderTranslator::StartVertexShader_LoadVertexIndex() {
shader_code_.push_back(1); shader_code_.push_back(1);
rdef_constants_used_ |= 1ull rdef_constants_used_ |= 1ull
<< uint32_t(RdefConstantIndex::kSysVertexIndexEndian); << uint32_t(RdefConstantIndex::kSysVertexIndexEndian);
shader_code_.push_back(EncodeVectorReplicatedOperand( shader_code_.push_back(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER, kSysConst_VertexIndexEndian_Comp, EncodeVectorReplicatedOperand(D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
3)); kSysConst_VertexIndexEndian_Comp, 3));
shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants)); shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants));
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants)); shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_VertexIndexEndian_Vec); shader_code_.push_back(kSysConst_VertexIndexEndian_Vec);
@ -338,9 +338,9 @@ void DxbcShaderTranslator::StartVertexShader_LoadVertexIndex() {
shader_code_.push_back(reg); shader_code_.push_back(reg);
rdef_constants_used_ |= 1ull rdef_constants_used_ |= 1ull
<< uint32_t(RdefConstantIndex::kSysVertexBaseIndex); << uint32_t(RdefConstantIndex::kSysVertexBaseIndex);
shader_code_.push_back(EncodeVectorSelectOperand( shader_code_.push_back(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER, kSysConst_VertexBaseIndex_Comp, EncodeVectorSelectOperand(D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
3)); kSysConst_VertexBaseIndex_Comp, 3));
shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants)); shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants));
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants)); shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_VertexBaseIndex_Vec); shader_code_.push_back(kSysConst_VertexBaseIndex_Vec);
@ -506,7 +506,8 @@ void DxbcShaderTranslator::CompleteVertexShader() {
shader_code_.push_back(EncodeVectorSwizzledOperand( shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER, D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_NDCScale_Comp | ((kSysConst_NDCScale_Comp + 1) << 2) | kSysConst_NDCScale_Comp | ((kSysConst_NDCScale_Comp + 1) << 2) |
((kSysConst_NDCScale_Comp + 2) << 4), 3)); ((kSysConst_NDCScale_Comp + 2) << 4),
3));
shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants)); shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants));
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants)); shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_NDCScale_Vec); shader_code_.push_back(kSysConst_NDCScale_Vec);
@ -524,7 +525,8 @@ void DxbcShaderTranslator::CompleteVertexShader() {
shader_code_.push_back(EncodeVectorSwizzledOperand( shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER, D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER,
kSysConst_NDCOffset_Comp | ((kSysConst_NDCOffset_Comp + 1) << 2) | kSysConst_NDCOffset_Comp | ((kSysConst_NDCOffset_Comp + 1) << 2) |
((kSysConst_NDCOffset_Comp + 2) << 4), 3)); ((kSysConst_NDCOffset_Comp + 2) << 4),
3));
shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants)); shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants));
shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants)); shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
shader_code_.push_back(kSysConst_NDCOffset_Vec); shader_code_.push_back(kSysConst_NDCOffset_Vec);
@ -554,8 +556,7 @@ void DxbcShaderTranslator::CompletePixelShader() {
// TODO(Triang3l): Alpha testing. // TODO(Triang3l): Alpha testing.
// Apply color exponent bias (the constant contains 2.0^bias). // Apply color exponent bias (the constant contains 2.0^bias).
rdef_constants_used_ |= 1ull rdef_constants_used_ |= 1ull << uint32_t(RdefConstantIndex::kSysColorExpBias);
<< uint32_t(RdefConstantIndex::kSysColorExpBias);
for (uint32_t i = 0; i < 4; ++i) { for (uint32_t i = 0; i < 4; ++i) {
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MUL) | shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MUL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9)); ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(9));
@ -1651,8 +1652,8 @@ void DxbcShaderTranslator::StoreResult(const InstructionResult& result,
shader_code_.push_back( shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1)); EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
shader_code_.push_back(i); shader_code_.push_back(i);
shader_code_.push_back(EncodeVectorReplicatedOperand( shader_code_.push_back(
D3D10_SB_OPERAND_TYPE_TEMP, i & 3, 1)); EncodeVectorReplicatedOperand(D3D10_SB_OPERAND_TYPE_TEMP, i & 3, 1));
shader_code_.push_back(gpr_movc_mask_register); shader_code_.push_back(gpr_movc_mask_register);
shader_code_.push_back(EncodeVectorSwizzledOperand( shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1)); D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
@ -2075,8 +2076,8 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
extract_offsets[2] = 16; extract_offsets[2] = 16;
extract_offsets[3] = 24; extract_offsets[3] = 24;
normalize_scales[0] = normalize_scales[1] = normalize_scales[2] = normalize_scales[0] = normalize_scales[1] = normalize_scales[2] =
normalize_scales[3] = instr.attributes.is_signed ? (1.0f / 127.0f) normalize_scales[3] =
: (1.0f / 255.0f); instr.attributes.is_signed ? (1.0f / 127.0f) : (1.0f / 255.0f);
break; break;
case VertexFormat::k_2_10_10_10: case VertexFormat::k_2_10_10_10:
extract_widths[0] = extract_widths[1] = extract_widths[2] = 10; extract_widths[0] = extract_widths[1] = extract_widths[2] = 10;
@ -2151,8 +2152,8 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
ENCODE_D3D10_SB_OPCODE_TYPE(extract_signed ? D3D11_SB_OPCODE_IBFE ENCODE_D3D10_SB_OPCODE_TYPE(extract_signed ? D3D11_SB_OPCODE_IBFE
: D3D11_SB_OPCODE_UBFE) | : D3D11_SB_OPCODE_UBFE) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(15)); ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(15));
shader_code_.push_back(EncodeVectorMaskedOperand( shader_code_.push_back(EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP,
D3D10_SB_OPERAND_TYPE_TEMP, result_write_mask, 1)); result_write_mask, 1));
shader_code_.push_back(system_temp_pv_); shader_code_.push_back(system_temp_pv_);
shader_code_.push_back(EncodeVectorSwizzledOperand( shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_IMMEDIATE32, kSwizzleXYZW, 0)); D3D10_SB_OPERAND_TYPE_IMMEDIATE32, kSwizzleXYZW, 0));
@ -2183,8 +2184,8 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
shader_code_.push_back( shader_code_.push_back(
ENCODE_D3D10_SB_OPCODE_TYPE(D3D11_SB_OPCODE_F16TOF32) | ENCODE_D3D10_SB_OPCODE_TYPE(D3D11_SB_OPCODE_F16TOF32) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5)); ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));
shader_code_.push_back(EncodeVectorMaskedOperand( shader_code_.push_back(EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP,
D3D10_SB_OPERAND_TYPE_TEMP, result_write_mask, 1)); result_write_mask, 1));
shader_code_.push_back(system_temp_pv_); shader_code_.push_back(system_temp_pv_);
shader_code_.push_back(EncodeVectorSwizzledOperand( shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1)); D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
@ -2195,12 +2196,12 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
// If no normalize_scales, it's a float value already. Otherwise, convert to // If no normalize_scales, it's a float value already. Otherwise, convert to
// float and normalize if needed. // float and normalize if needed.
shader_code_.push_back( shader_code_.push_back(
ENCODE_D3D10_SB_OPCODE_TYPE( ENCODE_D3D10_SB_OPCODE_TYPE(instr.attributes.is_signed
instr.attributes.is_signed ? D3D10_SB_OPCODE_ITOF ? D3D10_SB_OPCODE_ITOF
: D3D10_SB_OPCODE_UTOF) | : D3D10_SB_OPCODE_UTOF) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5)); ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));
shader_code_.push_back(EncodeVectorMaskedOperand( shader_code_.push_back(EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP,
D3D10_SB_OPERAND_TYPE_TEMP, result_write_mask, 1)); result_write_mask, 1));
shader_code_.push_back(system_temp_pv_); shader_code_.push_back(system_temp_pv_);
shader_code_.push_back(EncodeVectorSwizzledOperand( shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1)); D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
@ -2271,8 +2272,8 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
if (instr.attributes.exp_adjust != 0) { if (instr.attributes.exp_adjust != 0) {
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MUL) | shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MUL) |
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(10)); ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(10));
shader_code_.push_back(EncodeVectorMaskedOperand( shader_code_.push_back(EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP,
D3D10_SB_OPERAND_TYPE_TEMP, result_write_mask, 1)); result_write_mask, 1));
shader_code_.push_back(system_temp_pv_); shader_code_.push_back(system_temp_pv_);
shader_code_.push_back(EncodeVectorSwizzledOperand( shader_code_.push_back(EncodeVectorSwizzledOperand(
D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1)); D3D10_SB_OPERAND_TYPE_TEMP, kSwizzleXYZW, 1));
@ -3022,9 +3023,9 @@ void DxbcShaderTranslator::ProcessScalarAluInstruction(
++stat_.instruction_count; ++stat_.instruction_count;
++stat_.float_instruction_count; ++stat_.float_instruction_count;
// src0.y <= 0.0 // src0.y <= 0.0
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_GE) | shader_code_.push_back(
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH( ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_GE) |
5 + operand_lengths[0])); ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5 + operand_lengths[0]));
shader_code_.push_back( shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0100, 1)); EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0100, 1));
shader_code_.push_back(minus_max_mask); shader_code_.push_back(minus_max_mask);
@ -3074,9 +3075,9 @@ void DxbcShaderTranslator::ProcessScalarAluInstruction(
++stat_.uint_instruction_count; ++stat_.uint_instruction_count;
// Calculate the product for the regular path of the instruction. // Calculate the product for the regular path of the instruction.
// ps = src0.x * ps // ps = src0.x * ps
shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MUL) | shader_code_.push_back(
ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH( ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MUL) |
5 + operand_lengths[0])); ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5 + operand_lengths[0]));
shader_code_.push_back( shader_code_.push_back(
EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0001, 1)); EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b0001, 1));
shader_code_.push_back(system_temp_ps_pc_p0_a0_); shader_code_.push_back(system_temp_ps_pc_p0_a0_);

15
src/xenia/gpu/hlsl_shader_translator.cc
View File

@ -979,7 +979,8 @@ void HlslShaderTranslator::ProcessVertexFetchInstruction(
} }
if (!instr.attributes.is_integer) { if (!instr.attributes.is_integer) {
if (instr.attributes.is_signed) { if (instr.attributes.is_signed) {
EmitSourceDepth("xe_pv = max(xe_pv * float4((1.0 / 511.0).xxx, 1.0), " EmitSourceDepth(
"xe_pv = max(xe_pv * float4((1.0 / 511.0).xxx, 1.0), "
"(-1.0).xxxx);\n"); "(-1.0).xxxx);\n");
} else { } else {
EmitSourceDepth("xe_pv *= float4((1.0 / 1023.0).xxx, 1.0 / 3.0);\n"); EmitSourceDepth("xe_pv *= float4((1.0 / 1023.0).xxx, 1.0 / 3.0);\n");
@ -1529,25 +1530,29 @@ void HlslShaderTranslator::ProcessVectorAluInstruction(
case AluVectorOpcode::kSetpEqPush: case AluVectorOpcode::kSetpEqPush:
cf_exec_pred_ = false; cf_exec_pred_ = false;
EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w == 0.0;\n"); EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w == 0.0;\n");
EmitSourceDepth("xe_pv = (xe_src0.x == 0.0 && xe_src1.x == 0.0 ? " EmitSourceDepth(
"xe_pv = (xe_src0.x == 0.0 && xe_src1.x == 0.0 ? "
"0.0 : xe_src0.x + 1.0).xxxx;\n"); "0.0 : xe_src0.x + 1.0).xxxx;\n");
break; break;
case AluVectorOpcode::kSetpNePush: case AluVectorOpcode::kSetpNePush:
cf_exec_pred_ = false; cf_exec_pred_ = false;
EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w != 0.0;\n"); EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w != 0.0;\n");
EmitSourceDepth("xe_pv = (xe_src0.x == 0.0 && xe_src1.x != 0.0 ? " EmitSourceDepth(
"xe_pv = (xe_src0.x == 0.0 && xe_src1.x != 0.0 ? "
"0.0 : xe_src0.x + 1.0).xxxx;\n"); "0.0 : xe_src0.x + 1.0).xxxx;\n");
break; break;
case AluVectorOpcode::kSetpGtPush: case AluVectorOpcode::kSetpGtPush:
cf_exec_pred_ = false; cf_exec_pred_ = false;
EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w > 0.0;\n"); EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w > 0.0;\n");
EmitSourceDepth("xe_pv = (xe_src0.x == 0.0 && xe_src1.x > 0.0 ? " EmitSourceDepth(
"xe_pv = (xe_src0.x == 0.0 && xe_src1.x > 0.0 ? "
"0.0 : xe_src0.x + 1.0).xxxx;\n"); "0.0 : xe_src0.x + 1.0).xxxx;\n");
break; break;
case AluVectorOpcode::kSetpGePush: case AluVectorOpcode::kSetpGePush:
cf_exec_pred_ = false; cf_exec_pred_ = false;
EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w >= 0.0;\n"); EmitSourceDepth("xe_p0 = xe_src0.w == 0.0 && xe_src1.w >= 0.0;\n");
EmitSourceDepth("xe_pv = (xe_src0.x == 0.0 && xe_src1.x >= 0.0 ? " EmitSourceDepth(
"xe_pv = (xe_src0.x == 0.0 && xe_src1.x >= 0.0 ? "
"0.0 : xe_src0.x + 1.0).xxxx;\n"); "0.0 : xe_src0.x + 1.0).xxxx;\n");
break; break;
case AluVectorOpcode::kKillEq: case AluVectorOpcode::kKillEq: