2015-07-23 05:25:30 +02:00
|
|
|
// Copyright 2015 Citra Emulator Project
|
|
|
|
// Licensed under GPLv2 or any later version
|
|
|
|
// Refer to the license.txt file included.
|
|
|
|
|
2016-04-13 05:24:34 +02:00
|
|
|
#include <algorithm>
|
2016-04-30 17:34:51 +02:00
|
|
|
#include <cmath>
|
|
|
|
#include <cstdint>
|
|
|
|
#include <nihstro/shader_bytecode.h>
|
2016-12-12 10:23:08 +01:00
|
|
|
#include <smmintrin.h>
|
2016-09-21 09:13:46 +02:00
|
|
|
#include <xmmintrin.h>
|
2016-04-30 17:34:51 +02:00
|
|
|
#include "common/assert.h"
|
|
|
|
#include "common/logging/log.h"
|
|
|
|
#include "common/vector_math.h"
|
2015-08-12 06:19:20 +02:00
|
|
|
#include "common/x64/cpu_detect.h"
|
2016-12-12 10:23:08 +01:00
|
|
|
#include "common/x64/xbyak_abi.h"
|
|
|
|
#include "common/x64/xbyak_util.h"
|
2016-03-03 04:16:38 +01:00
|
|
|
#include "video_core/pica_state.h"
|
2016-04-30 17:34:51 +02:00
|
|
|
#include "video_core/pica_types.h"
|
2016-12-12 10:23:08 +01:00
|
|
|
#include "video_core/shader/shader.h"
|
2016-12-17 09:06:23 +01:00
|
|
|
#include "video_core/shader/shader_jit_x64_compiler.h"
|
2016-12-12 10:23:08 +01:00
|
|
|
|
|
|
|
using namespace Common::X64;
|
|
|
|
using namespace Xbyak::util;
|
|
|
|
using Xbyak::Label;
|
|
|
|
using Xbyak::Reg32;
|
|
|
|
using Xbyak::Reg64;
|
|
|
|
using Xbyak::Xmm;
|
2016-03-03 04:16:38 +01:00
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
namespace Pica {
|
|
|
|
|
|
|
|
namespace Shader {
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
typedef void (JitShader::*JitFunction)(Instruction instr);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
const JitFunction instr_table[64] = {
|
2018-05-18 22:28:58 +02:00
|
|
|
&JitShader::Compile_ADD, // add
|
|
|
|
&JitShader::Compile_DP3, // dp3
|
|
|
|
&JitShader::Compile_DP4, // dp4
|
|
|
|
&JitShader::Compile_DPH, // dph
|
|
|
|
nullptr, // unknown
|
|
|
|
&JitShader::Compile_EX2, // ex2
|
|
|
|
&JitShader::Compile_LG2, // lg2
|
|
|
|
nullptr, // unknown
|
|
|
|
&JitShader::Compile_MUL, // mul
|
|
|
|
&JitShader::Compile_SGE, // sge
|
|
|
|
&JitShader::Compile_SLT, // slt
|
|
|
|
&JitShader::Compile_FLR, // flr
|
|
|
|
&JitShader::Compile_MAX, // max
|
|
|
|
&JitShader::Compile_MIN, // min
|
|
|
|
&JitShader::Compile_RCP, // rcp
|
|
|
|
&JitShader::Compile_RSQ, // rsq
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
&JitShader::Compile_MOVA, // mova
|
|
|
|
&JitShader::Compile_MOV, // mov
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
&JitShader::Compile_DPH, // dphi
|
|
|
|
nullptr, // unknown
|
|
|
|
&JitShader::Compile_SGE, // sgei
|
|
|
|
&JitShader::Compile_SLT, // slti
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
nullptr, // unknown
|
|
|
|
&JitShader::Compile_NOP, // nop
|
|
|
|
&JitShader::Compile_END, // end
|
|
|
|
&JitShader::Compile_BREAKC, // breakc
|
|
|
|
&JitShader::Compile_CALL, // call
|
|
|
|
&JitShader::Compile_CALLC, // callc
|
|
|
|
&JitShader::Compile_CALLU, // callu
|
|
|
|
&JitShader::Compile_IF, // ifu
|
|
|
|
&JitShader::Compile_IF, // ifc
|
|
|
|
&JitShader::Compile_LOOP, // loop
|
|
|
|
&JitShader::Compile_EMIT, // emit
|
|
|
|
&JitShader::Compile_SETE, // sete
|
|
|
|
&JitShader::Compile_JMP, // jmpc
|
|
|
|
&JitShader::Compile_JMP, // jmpu
|
|
|
|
&JitShader::Compile_CMP, // cmp
|
|
|
|
&JitShader::Compile_CMP, // cmp
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // madi
|
|
|
|
&JitShader::Compile_MAD, // mad
|
|
|
|
&JitShader::Compile_MAD, // mad
|
|
|
|
&JitShader::Compile_MAD, // mad
|
|
|
|
&JitShader::Compile_MAD, // mad
|
|
|
|
&JitShader::Compile_MAD, // mad
|
|
|
|
&JitShader::Compile_MAD, // mad
|
|
|
|
&JitShader::Compile_MAD, // mad
|
|
|
|
&JitShader::Compile_MAD, // mad
|
2015-07-23 05:25:30 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
// The following is used to alias some commonly used registers. Generally, RAX-RDX and XMM0-XMM3 can
|
|
|
|
// be used as scratch registers within a compiler function. The other registers have designated
|
|
|
|
// purposes, as documented below:
|
|
|
|
|
|
|
|
/// Pointer to the uniform memory
|
2018-04-17 08:35:43 +02:00
|
|
|
static const Reg64 UNIFORMS = r9;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// The two 32-bit VS address offset registers set by the MOVA instruction
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Reg64 ADDROFFS_REG_0 = r10;
|
|
|
|
static const Reg64 ADDROFFS_REG_1 = r11;
|
2016-11-30 15:40:42 +01:00
|
|
|
/// VS loop count register (Multiplied by 16)
|
2016-12-13 06:44:15 +01:00
|
|
|
static const Reg32 LOOPCOUNT_REG = r12d;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Current VS loop iteration number (we could probably use LOOPCOUNT_REG, but this quicker)
|
2016-12-13 06:44:15 +01:00
|
|
|
static const Reg32 LOOPCOUNT = esi;
|
2016-11-30 15:40:42 +01:00
|
|
|
/// Number to increment LOOPCOUNT_REG by on each loop iteration (Multiplied by 16)
|
2016-12-13 06:44:15 +01:00
|
|
|
static const Reg32 LOOPINC = edi;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Result of the previous CMP instruction for the X-component comparison
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Reg64 COND0 = r13;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Result of the previous CMP instruction for the Y-component comparison
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Reg64 COND1 = r14;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Pointer to the UnitState instance for the current VS unit
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Reg64 STATE = r15;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// SIMD scratch register
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Xmm SCRATCH = xmm0;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Loaded with the first swizzled source register, otherwise can be used as a scratch register
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Xmm SRC1 = xmm1;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Loaded with the second swizzled source register, otherwise can be used as a scratch register
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Xmm SRC2 = xmm2;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Loaded with the third swizzled source register, otherwise can be used as a scratch register
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Xmm SRC3 = xmm3;
|
2015-08-24 06:46:10 +02:00
|
|
|
/// Additional scratch register
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Xmm SCRATCH2 = xmm4;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Constant vector of [1.0f, 1.0f, 1.0f, 1.0f], used to efficiently set a vector to one
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Xmm ONE = xmm14;
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Constant vector of [-0.f, -0.f, -0.f, -0.f], used to efficiently negate a vector with XOR
|
2016-12-12 10:23:08 +01:00
|
|
|
static const Xmm NEGBIT = xmm15;
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2015-08-26 09:12:14 +02:00
|
|
|
// State registers that must not be modified by external functions calls
|
|
|
|
// Scratch registers, e.g., SRC1 and SCRATCH, have to be saved on the side if needed
|
2016-12-12 10:23:08 +01:00
|
|
|
static const BitSet32 persistent_regs = BuildRegSet({
|
|
|
|
// Pointers to register blocks
|
2018-04-17 08:35:43 +02:00
|
|
|
UNIFORMS,
|
2018-03-09 18:54:43 +01:00
|
|
|
STATE,
|
2016-12-12 10:23:08 +01:00
|
|
|
// Cached registers
|
2018-03-09 18:54:43 +01:00
|
|
|
ADDROFFS_REG_0,
|
|
|
|
ADDROFFS_REG_1,
|
|
|
|
LOOPCOUNT_REG,
|
|
|
|
COND0,
|
|
|
|
COND1,
|
2016-12-12 10:23:08 +01:00
|
|
|
// Constants
|
2018-03-09 18:54:43 +01:00
|
|
|
ONE,
|
|
|
|
NEGBIT,
|
2017-01-31 09:38:31 +01:00
|
|
|
// Loop variables
|
2018-03-09 18:54:43 +01:00
|
|
|
LOOPCOUNT,
|
|
|
|
LOOPINC,
|
2016-12-12 10:23:08 +01:00
|
|
|
});
|
2015-08-26 09:12:14 +02:00
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
/// Raw constant for the source register selector that indicates no swizzling is performed
|
|
|
|
static const u8 NO_SRC_REG_SWIZZLE = 0x1b;
|
|
|
|
/// Raw constant for the destination register enable mask that indicates all components are enabled
|
|
|
|
static const u8 NO_DEST_REG_MASK = 0xf;
|
|
|
|
|
2016-04-02 06:02:03 +02:00
|
|
|
static void LogCritical(const char* msg) {
|
2018-06-29 13:18:07 +02:00
|
|
|
LOG_CRITICAL(HW_GPU, "{}", msg);
|
2016-04-02 06:02:03 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:35:36 +02:00
|
|
|
void JitShader::Compile_Assert(bool condition, const char* msg) {
|
2016-04-02 06:02:03 +02:00
|
|
|
if (!condition) {
|
2018-09-06 22:03:28 +02:00
|
|
|
mov(ABI_PARAM1, reinterpret_cast<std::size_t>(msg));
|
2016-12-12 10:23:08 +01:00
|
|
|
CallFarFunction(*this, LogCritical);
|
2016-04-02 06:02:03 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
/**
|
|
|
|
* Loads and swizzles a source register into the specified XMM register.
|
|
|
|
* @param instr VS instruction, used for determining how to load the source register
|
|
|
|
* @param src_num Number indicating which source register to load (1 = src1, 2 = src2, 3 = src3)
|
|
|
|
* @param src_reg SourceRegister object corresponding to the source register to load
|
|
|
|
* @param dest Destination XMM register to store the loaded, swizzled source register
|
|
|
|
*/
|
2016-09-18 02:38:01 +02:00
|
|
|
void JitShader::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg,
|
2016-12-12 10:23:08 +01:00
|
|
|
Xmm dest) {
|
|
|
|
Reg64 src_ptr;
|
2018-09-06 22:03:28 +02:00
|
|
|
std::size_t src_offset;
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
if (src_reg.GetRegisterType() == RegisterType::FloatUniform) {
|
2018-04-17 08:35:43 +02:00
|
|
|
src_ptr = UNIFORMS;
|
|
|
|
src_offset = Uniforms::GetFloatUniformOffset(src_reg.GetIndex());
|
2015-07-23 05:25:30 +02:00
|
|
|
} else {
|
2016-05-13 08:46:14 +02:00
|
|
|
src_ptr = STATE;
|
2016-12-16 08:57:10 +01:00
|
|
|
src_offset = UnitState::InputOffset(src_reg);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2015-09-07 07:49:57 +02:00
|
|
|
int src_offset_disp = (int)src_offset;
|
|
|
|
ASSERT_MSG(src_offset == src_offset_disp, "Source register offset too large for int type");
|
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
unsigned operand_desc_id;
|
2016-03-09 14:48:45 +01:00
|
|
|
|
2016-09-18 02:38:01 +02:00
|
|
|
const bool is_inverted =
|
|
|
|
(0 != (instr.opcode.Value().GetInfo().subtype & OpCode::Info::SrcInversed));
|
2016-03-09 14:48:45 +01:00
|
|
|
|
|
|
|
unsigned address_register_index;
|
|
|
|
unsigned offset_src;
|
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
|
|
|
|
instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
|
|
|
|
operand_desc_id = instr.mad.operand_desc_id;
|
2016-03-09 14:48:45 +01:00
|
|
|
offset_src = is_inverted ? 3 : 2;
|
|
|
|
address_register_index = instr.mad.address_register_index;
|
2015-07-23 05:25:30 +02:00
|
|
|
} else {
|
|
|
|
operand_desc_id = instr.common.operand_desc_id;
|
2016-03-09 14:48:45 +01:00
|
|
|
offset_src = is_inverted ? 2 : 1;
|
|
|
|
address_register_index = instr.common.address_register_index;
|
|
|
|
}
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-03-09 14:48:45 +01:00
|
|
|
if (src_num == offset_src && address_register_index != 0) {
|
|
|
|
switch (address_register_index) {
|
|
|
|
case 1: // address offset 1
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(dest, xword[src_ptr + ADDROFFS_REG_0 + src_offset_disp]);
|
2016-03-09 14:48:45 +01:00
|
|
|
break;
|
|
|
|
case 2: // address offset 2
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(dest, xword[src_ptr + ADDROFFS_REG_1 + src_offset_disp]);
|
2016-03-09 14:48:45 +01:00
|
|
|
break;
|
|
|
|
case 3: // address offset 3
|
2016-12-15 19:02:42 +01:00
|
|
|
movaps(dest, xword[src_ptr + LOOPCOUNT_REG.cvt64() + src_offset_disp]);
|
2016-03-09 14:48:45 +01:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
UNREACHABLE();
|
|
|
|
break;
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
2016-03-09 14:48:45 +01:00
|
|
|
} else {
|
|
|
|
// Load the source
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(dest, xword[src_ptr + src_offset_disp]);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-12-17 11:29:22 +01:00
|
|
|
SwizzlePattern swiz = {(*swizzle_data)[operand_desc_id]};
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Generate instructions for source register swizzling as needed
|
|
|
|
u8 sel = swiz.GetRawSelector(src_num);
|
|
|
|
if (sel != NO_SRC_REG_SWIZZLE) {
|
|
|
|
// Selector component order needs to be reversed for the SHUFPS instruction
|
|
|
|
sel = ((sel & 0xc0) >> 6) | ((sel & 3) << 6) | ((sel & 0xc) << 2) | ((sel & 0x30) >> 2);
|
|
|
|
|
|
|
|
// Shuffle inputs for swizzle
|
2016-12-12 10:23:08 +01:00
|
|
|
shufps(dest, dest, sel);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// If the source register should be negated, flip the negative bit using XOR
|
2016-09-18 02:38:01 +02:00
|
|
|
const bool negate[] = {swiz.negate_src1, swiz.negate_src2, swiz.negate_src3};
|
2015-07-23 05:25:30 +02:00
|
|
|
if (negate[src_num - 1]) {
|
2016-12-12 10:23:08 +01:00
|
|
|
xorps(dest, NEGBIT);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
void JitShader::Compile_DestEnable(Instruction instr, Xmm src) {
|
2015-07-23 05:25:30 +02:00
|
|
|
DestRegister dest;
|
|
|
|
unsigned operand_desc_id;
|
|
|
|
if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
|
|
|
|
instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
|
|
|
|
operand_desc_id = instr.mad.operand_desc_id;
|
|
|
|
dest = instr.mad.dest.Value();
|
|
|
|
} else {
|
|
|
|
operand_desc_id = instr.common.operand_desc_id;
|
|
|
|
dest = instr.common.dest.Value();
|
|
|
|
}
|
|
|
|
|
2016-12-17 11:29:22 +01:00
|
|
|
SwizzlePattern swiz = {(*swizzle_data)[operand_desc_id]};
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2018-09-06 22:03:28 +02:00
|
|
|
std::size_t dest_offset_disp = UnitState::OutputOffset(dest);
|
2015-09-07 07:49:57 +02:00
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
// If all components are enabled, write the result to the destination register
|
|
|
|
if (swiz.dest_mask == NO_DEST_REG_MASK) {
|
|
|
|
// Store dest back to memory
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(xword[STATE + dest_offset_disp], src);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
} else {
|
2016-09-18 02:38:01 +02:00
|
|
|
// Not all components are enabled, so mask the result when storing to the destination
|
|
|
|
// register...
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(SCRATCH, xword[STATE + dest_offset_disp]);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2015-08-12 23:42:13 +02:00
|
|
|
if (Common::GetCPUCaps().sse4_1) {
|
2016-09-18 02:38:01 +02:00
|
|
|
u8 mask = ((swiz.dest_mask & 1) << 3) | ((swiz.dest_mask & 8) >> 3) |
|
|
|
|
((swiz.dest_mask & 2) << 1) | ((swiz.dest_mask & 4) >> 1);
|
2016-12-12 10:23:08 +01:00
|
|
|
blendps(SCRATCH, src, mask);
|
2015-07-23 05:25:30 +02:00
|
|
|
} else {
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(SCRATCH2, src);
|
|
|
|
unpckhps(SCRATCH2, SCRATCH); // Unpack X/Y components of source and destination
|
|
|
|
unpcklps(SCRATCH, src); // Unpack Z/W components of source and destination
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-09-18 02:38:01 +02:00
|
|
|
// Compute selector to selectively copy source components to destination for SHUFPS
|
|
|
|
// instruction
|
2015-07-23 05:25:30 +02:00
|
|
|
u8 sel = ((swiz.DestComponentEnabled(0) ? 1 : 0) << 0) |
|
|
|
|
((swiz.DestComponentEnabled(1) ? 3 : 2) << 2) |
|
|
|
|
((swiz.DestComponentEnabled(2) ? 0 : 1) << 4) |
|
|
|
|
((swiz.DestComponentEnabled(3) ? 2 : 3) << 6);
|
2016-12-12 10:23:08 +01:00
|
|
|
shufps(SCRATCH, SCRATCH2, sel);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// Store dest back to memory
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(xword[STATE + dest_offset_disp], SCRATCH);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
void JitShader::Compile_SanitizedMul(Xmm src1, Xmm src2, Xmm scratch) {
|
2017-02-12 22:29:14 +01:00
|
|
|
// 0 * inf and inf * 0 in the PICA should return 0 instead of NaN. This can be implemented by
|
|
|
|
// checking for NaNs before and after the multiplication. If the multiplication result is NaN
|
|
|
|
// where neither source was, this NaN was generated by a 0 * inf multiplication, and so the
|
|
|
|
// result should be transformed to 0 to match PICA fp rules.
|
|
|
|
|
|
|
|
// Set scratch to mask of (src1 != NaN and src2 != NaN)
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(scratch, src1);
|
|
|
|
cmpordps(scratch, src2);
|
2015-08-24 06:48:15 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
mulps(src1, src2);
|
2015-08-24 06:48:15 +02:00
|
|
|
|
2017-02-12 22:29:14 +01:00
|
|
|
// Set src2 to mask of (result == NaN)
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(src2, src1);
|
|
|
|
cmpunordps(src2, src2);
|
2015-08-24 06:48:15 +02:00
|
|
|
|
2017-02-12 22:29:14 +01:00
|
|
|
// Clear components where scratch != src2 (i.e. if result is NaN where neither source was NaN)
|
2016-12-12 10:23:08 +01:00
|
|
|
xorps(scratch, src2);
|
|
|
|
andps(src1, scratch);
|
2015-08-24 06:48:15 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_EvaluateCondition(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
// Note: NXOR is used below to check for equality
|
|
|
|
switch (instr.flow_control.op) {
|
|
|
|
case Instruction::FlowControlType::Or:
|
2016-12-12 10:23:08 +01:00
|
|
|
mov(eax, COND0);
|
|
|
|
mov(ebx, COND1);
|
2017-06-17 21:20:22 +02:00
|
|
|
xor_(eax, (instr.flow_control.refx.Value() ^ 1));
|
|
|
|
xor_(ebx, (instr.flow_control.refy.Value() ^ 1));
|
|
|
|
or_(eax, ebx);
|
2015-07-23 05:25:30 +02:00
|
|
|
break;
|
|
|
|
|
|
|
|
case Instruction::FlowControlType::And:
|
2016-12-12 10:23:08 +01:00
|
|
|
mov(eax, COND0);
|
|
|
|
mov(ebx, COND1);
|
2017-06-17 21:20:22 +02:00
|
|
|
xor_(eax, (instr.flow_control.refx.Value() ^ 1));
|
|
|
|
xor_(ebx, (instr.flow_control.refy.Value() ^ 1));
|
|
|
|
and_(eax, ebx);
|
2015-07-23 05:25:30 +02:00
|
|
|
break;
|
|
|
|
|
|
|
|
case Instruction::FlowControlType::JustX:
|
2016-12-12 10:23:08 +01:00
|
|
|
mov(eax, COND0);
|
2017-06-17 21:20:22 +02:00
|
|
|
xor_(eax, (instr.flow_control.refx.Value() ^ 1));
|
2015-07-23 05:25:30 +02:00
|
|
|
break;
|
|
|
|
|
|
|
|
case Instruction::FlowControlType::JustY:
|
2016-12-12 10:23:08 +01:00
|
|
|
mov(eax, COND1);
|
2017-06-17 21:20:22 +02:00
|
|
|
xor_(eax, (instr.flow_control.refy.Value() ^ 1));
|
2015-07-23 05:25:30 +02:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_UniformCondition(Instruction instr) {
|
2018-09-06 22:03:28 +02:00
|
|
|
std::size_t offset = Uniforms::GetBoolUniformOffset(instr.flow_control.bool_uniform_id);
|
2018-04-17 08:35:43 +02:00
|
|
|
cmp(byte[UNIFORMS + offset], 0);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
BitSet32 JitShader::PersistentCallerSavedRegs() {
|
2015-08-26 09:12:14 +02:00
|
|
|
return persistent_regs & ABI_ALL_CALLER_SAVED;
|
2015-08-19 01:49:45 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_ADD(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
2016-12-12 10:23:08 +01:00
|
|
|
addps(SRC1, SRC2);
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_DP3(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
|
|
|
|
2015-08-24 06:48:15 +02:00
|
|
|
Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(SRC2, SRC1);
|
|
|
|
shufps(SRC2, SRC2, _MM_SHUFFLE(1, 1, 1, 1));
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(SRC3, SRC1);
|
|
|
|
shufps(SRC3, SRC3, _MM_SHUFFLE(2, 2, 2, 2));
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
|
|
|
|
addps(SRC1, SRC2);
|
|
|
|
addps(SRC1, SRC3);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_DP4(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
|
|
|
|
2015-08-24 06:48:15 +02:00
|
|
|
Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2017-12-12 16:37:00 +01:00
|
|
|
haddps(SRC1, SRC1);
|
|
|
|
haddps(SRC1, SRC1);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_DPH(Instruction instr) {
|
2015-08-21 12:49:21 +02:00
|
|
|
if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::DPHI) {
|
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
|
|
|
|
} else {
|
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (Common::GetCPUCaps().sse4_1) {
|
|
|
|
// Set 4th component to 1.0
|
2016-12-12 10:23:08 +01:00
|
|
|
blendps(SRC1, ONE, 0b1000);
|
2015-08-21 12:49:21 +02:00
|
|
|
} else {
|
2015-08-24 06:15:39 +02:00
|
|
|
// Set 4th component to 1.0
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(SCRATCH, SRC1);
|
|
|
|
unpckhps(SCRATCH, ONE); // XYZW, 1111 -> Z1__
|
|
|
|
unpcklpd(SRC1, SCRATCH); // XYZW, Z1__ -> XYZ1
|
2015-08-24 06:48:15 +02:00
|
|
|
}
|
2015-08-21 12:49:21 +02:00
|
|
|
|
2015-08-24 06:48:15 +02:00
|
|
|
Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
|
2015-08-21 12:49:21 +02:00
|
|
|
|
2017-12-12 16:37:00 +01:00
|
|
|
haddps(SRC1, SRC1);
|
|
|
|
haddps(SRC1, SRC1);
|
2015-08-21 12:49:21 +02:00
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_EX2(Instruction instr) {
|
2015-08-16 17:22:49 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
2017-11-25 04:18:55 +01:00
|
|
|
call(exp2_subroutine);
|
2015-08-16 17:22:49 +02:00
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_LG2(Instruction instr) {
|
2015-08-16 17:22:49 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
2017-11-25 04:18:55 +01:00
|
|
|
call(log2_subroutine);
|
2015-08-16 17:22:49 +02:00
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_MUL(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
2015-08-24 06:48:15 +02:00
|
|
|
Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_SGE(Instruction instr) {
|
2015-08-19 14:23:53 +02:00
|
|
|
if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SGEI) {
|
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
|
|
|
|
} else {
|
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
|
|
|
}
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
cmpleps(SRC2, SRC1);
|
|
|
|
andps(SRC2, ONE);
|
2015-08-19 14:23:53 +02:00
|
|
|
|
2015-08-31 08:09:07 +02:00
|
|
|
Compile_DestEnable(instr, SRC2);
|
2015-08-19 14:23:53 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_SLT(Instruction instr) {
|
2015-08-19 14:23:53 +02:00
|
|
|
if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SLTI) {
|
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
|
|
|
|
} else {
|
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
|
|
|
}
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
cmpltps(SRC1, SRC2);
|
|
|
|
andps(SRC1, ONE);
|
2015-08-19 14:23:53 +02:00
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_FLR(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
|
2015-08-12 23:42:13 +02:00
|
|
|
if (Common::GetCPUCaps().sse4_1) {
|
2016-12-12 10:23:08 +01:00
|
|
|
roundps(SRC1, SRC1, _MM_FROUND_FLOOR);
|
2015-07-23 05:25:30 +02:00
|
|
|
} else {
|
2016-12-12 10:23:08 +01:00
|
|
|
cvttps2dq(SRC1, SRC1);
|
|
|
|
cvtdq2ps(SRC1, SRC1);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_MAX(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
2015-08-24 06:46:58 +02:00
|
|
|
// SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
|
2016-12-12 10:23:08 +01:00
|
|
|
maxps(SRC1, SRC2);
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_MIN(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
2015-08-24 06:46:58 +02:00
|
|
|
// SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
|
2016-12-12 10:23:08 +01:00
|
|
|
minps(SRC1, SRC2);
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_MOVA(Instruction instr) {
|
2016-12-17 11:29:22 +01:00
|
|
|
SwizzlePattern swiz = {(*swizzle_data)[instr.common.operand_desc_id]};
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
if (!swiz.DestComponentEnabled(0) && !swiz.DestComponentEnabled(1)) {
|
|
|
|
return; // NoOp
|
|
|
|
}
|
|
|
|
|
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
|
2015-08-27 15:21:05 +02:00
|
|
|
// Convert floats to integers using truncation (only care about X and Y components)
|
2016-12-12 10:23:08 +01:00
|
|
|
cvttps2dq(SRC1, SRC1);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Get result
|
2016-12-12 10:23:08 +01:00
|
|
|
movq(rax, SRC1);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Handle destination enable
|
|
|
|
if (swiz.DestComponentEnabled(0) && swiz.DestComponentEnabled(1)) {
|
2015-07-31 18:40:09 +02:00
|
|
|
// Move and sign-extend low 32 bits
|
2016-12-12 10:23:08 +01:00
|
|
|
movsxd(ADDROFFS_REG_0, eax);
|
2015-07-31 18:40:09 +02:00
|
|
|
|
|
|
|
// Move and sign-extend high 32 bits
|
2016-12-12 10:23:08 +01:00
|
|
|
shr(rax, 32);
|
|
|
|
movsxd(ADDROFFS_REG_1, eax);
|
2015-07-31 18:40:09 +02:00
|
|
|
|
|
|
|
// Multiply by 16 to be used as an offset later
|
2016-12-12 10:23:08 +01:00
|
|
|
shl(ADDROFFS_REG_0, 4);
|
|
|
|
shl(ADDROFFS_REG_1, 4);
|
2015-07-23 05:25:30 +02:00
|
|
|
} else {
|
|
|
|
if (swiz.DestComponentEnabled(0)) {
|
2015-07-31 18:40:09 +02:00
|
|
|
// Move and sign-extend low 32 bits
|
2016-12-12 10:23:08 +01:00
|
|
|
movsxd(ADDROFFS_REG_0, eax);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2015-07-31 18:40:09 +02:00
|
|
|
// Multiply by 16 to be used as an offset later
|
2016-12-12 10:23:08 +01:00
|
|
|
shl(ADDROFFS_REG_0, 4);
|
2015-07-23 05:25:30 +02:00
|
|
|
} else if (swiz.DestComponentEnabled(1)) {
|
2015-07-31 18:40:09 +02:00
|
|
|
// Move and sign-extend high 32 bits
|
2016-12-12 10:23:08 +01:00
|
|
|
shr(rax, 32);
|
|
|
|
movsxd(ADDROFFS_REG_1, eax);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2015-07-31 18:40:09 +02:00
|
|
|
// Multiply by 16 to be used as an offset later
|
2016-12-12 10:23:08 +01:00
|
|
|
shl(ADDROFFS_REG_1, 4);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_MOV(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_RCP(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
|
2015-08-23 15:13:36 +02:00
|
|
|
// TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
|
2015-07-23 05:25:30 +02:00
|
|
|
// performs this operation more accurately. This should be checked on hardware.
|
2016-12-12 10:23:08 +01:00
|
|
|
rcpss(SRC1, SRC1);
|
|
|
|
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_RSQ(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
|
2015-08-23 15:13:36 +02:00
|
|
|
// TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
|
2015-07-23 05:25:30 +02:00
|
|
|
// performs this operation more accurately. This should be checked on hardware.
|
2016-12-12 10:23:08 +01:00
|
|
|
rsqrtss(SRC1, SRC1);
|
|
|
|
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-09-19 03:01:46 +02:00
|
|
|
void JitShader::Compile_NOP(Instruction instr) {}
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_END(Instruction instr) {
|
2018-05-28 13:28:03 +02:00
|
|
|
// Save conditional code
|
|
|
|
mov(byte[STATE + offsetof(UnitState, conditional_code[0])], COND0.cvt8());
|
|
|
|
mov(byte[STATE + offsetof(UnitState, conditional_code[1])], COND1.cvt8());
|
|
|
|
|
|
|
|
// Save address/loop registers
|
|
|
|
sar(ADDROFFS_REG_0, 4);
|
|
|
|
sar(ADDROFFS_REG_1, 4);
|
|
|
|
sar(LOOPCOUNT_REG, 4);
|
|
|
|
mov(dword[STATE + offsetof(UnitState, address_registers[0])], ADDROFFS_REG_0.cvt32());
|
|
|
|
mov(dword[STATE + offsetof(UnitState, address_registers[1])], ADDROFFS_REG_1.cvt32());
|
|
|
|
mov(dword[STATE + offsetof(UnitState, address_registers[2])], LOOPCOUNT_REG);
|
|
|
|
|
2017-02-03 13:22:43 +01:00
|
|
|
ABI_PopRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8, 16);
|
2016-12-12 10:23:08 +01:00
|
|
|
ret();
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2018-05-18 22:28:58 +02:00
|
|
|
void JitShader::Compile_BREAKC(Instruction instr) {
|
|
|
|
Compile_Assert(looping, "BREAKC must be inside a LOOP");
|
|
|
|
if (looping) {
|
|
|
|
Compile_EvaluateCondition(instr);
|
|
|
|
jnz(*loop_break_label);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_CALL(Instruction instr) {
|
2016-03-18 00:45:09 +01:00
|
|
|
// Push offset of the return
|
2016-12-12 10:23:08 +01:00
|
|
|
push(qword, (instr.flow_control.dest_offset + instr.flow_control.num_instructions));
|
2016-03-18 00:45:09 +01:00
|
|
|
|
2016-04-09 23:46:13 +02:00
|
|
|
// Call the subroutine
|
2016-12-12 10:23:08 +01:00
|
|
|
call(instruction_labels[instr.flow_control.dest_offset]);
|
2016-03-18 00:45:09 +01:00
|
|
|
|
2016-04-09 23:46:13 +02:00
|
|
|
// Skip over the return offset that's on the stack
|
2016-12-12 10:23:08 +01:00
|
|
|
add(rsp, 8);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_CALLC(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_EvaluateCondition(instr);
|
2016-12-12 10:23:08 +01:00
|
|
|
Label b;
|
|
|
|
jz(b);
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_CALL(instr);
|
2016-12-12 10:23:08 +01:00
|
|
|
L(b);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_CALLU(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_UniformCondition(instr);
|
2016-12-12 10:23:08 +01:00
|
|
|
Label b;
|
|
|
|
jz(b);
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_CALL(instr);
|
2016-12-12 10:23:08 +01:00
|
|
|
L(b);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_CMP(Instruction instr) {
|
2015-08-24 02:46:36 +02:00
|
|
|
using Op = Instruction::Common::CompareOpType::Op;
|
|
|
|
Op op_x = instr.common.compare_op.x;
|
|
|
|
Op op_y = instr.common.compare_op.y;
|
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
|
|
|
|
|
2015-08-24 02:46:36 +02:00
|
|
|
// SSE doesn't have greater-than (GT) or greater-equal (GE) comparison operators. You need to
|
|
|
|
// emulate them by swapping the lhs and rhs and using LT and LE. NLT and NLE can't be used here
|
|
|
|
// because they don't match when used with NaNs.
|
2016-09-18 02:38:01 +02:00
|
|
|
static const u8 cmp[] = {CMP_EQ, CMP_NEQ, CMP_LT, CMP_LE, CMP_LT, CMP_LE};
|
2015-08-24 02:46:36 +02:00
|
|
|
|
|
|
|
bool invert_op_x = (op_x == Op::GreaterThan || op_x == Op::GreaterEqual);
|
2016-12-12 10:23:08 +01:00
|
|
|
Xmm lhs_x = invert_op_x ? SRC2 : SRC1;
|
|
|
|
Xmm rhs_x = invert_op_x ? SRC1 : SRC2;
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2015-08-24 02:46:36 +02:00
|
|
|
if (op_x == op_y) {
|
2015-07-23 05:25:30 +02:00
|
|
|
// Compare X-component and Y-component together
|
2016-12-12 10:23:08 +01:00
|
|
|
cmpps(lhs_x, rhs_x, cmp[op_x]);
|
|
|
|
movq(COND0, lhs_x);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
mov(COND1, COND0);
|
2015-07-23 05:25:30 +02:00
|
|
|
} else {
|
2015-08-24 02:46:36 +02:00
|
|
|
bool invert_op_y = (op_y == Op::GreaterThan || op_y == Op::GreaterEqual);
|
2016-12-12 10:23:08 +01:00
|
|
|
Xmm lhs_y = invert_op_y ? SRC2 : SRC1;
|
|
|
|
Xmm rhs_y = invert_op_y ? SRC1 : SRC2;
|
2015-08-24 02:46:36 +02:00
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
// Compare X-component
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(SCRATCH, lhs_x);
|
|
|
|
cmpss(SCRATCH, rhs_x, cmp[op_x]);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Compare Y-component
|
2016-12-12 10:23:08 +01:00
|
|
|
cmpps(lhs_y, rhs_y, cmp[op_y]);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
movq(COND0, SCRATCH);
|
|
|
|
movq(COND1, lhs_y);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
shr(COND0.cvt32(), 31); // ignores upper 32 bits in source
|
|
|
|
shr(COND1, 63);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_MAD(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
Compile_SwizzleSrc(instr, 1, instr.mad.src1, SRC1);
|
|
|
|
|
|
|
|
if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.mad.src2i, SRC2);
|
|
|
|
Compile_SwizzleSrc(instr, 3, instr.mad.src3i, SRC3);
|
|
|
|
} else {
|
|
|
|
Compile_SwizzleSrc(instr, 2, instr.mad.src2, SRC2);
|
|
|
|
Compile_SwizzleSrc(instr, 3, instr.mad.src3, SRC3);
|
|
|
|
}
|
|
|
|
|
2015-08-24 06:48:15 +02:00
|
|
|
Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
|
2016-12-12 10:23:08 +01:00
|
|
|
addps(SRC1, SRC3);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
Compile_DestEnable(instr, SRC1);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_IF(Instruction instr) {
|
2016-09-18 02:38:01 +02:00
|
|
|
Compile_Assert(instr.flow_control.dest_offset >= program_counter,
|
|
|
|
"Backwards if-statements not supported");
|
2016-12-12 10:23:08 +01:00
|
|
|
Label l_else, l_endif;
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Evaluate the "IF" condition
|
|
|
|
if (instr.opcode.Value() == OpCode::Id::IFU) {
|
|
|
|
Compile_UniformCondition(instr);
|
|
|
|
} else if (instr.opcode.Value() == OpCode::Id::IFC) {
|
|
|
|
Compile_EvaluateCondition(instr);
|
|
|
|
}
|
2016-12-12 10:23:08 +01:00
|
|
|
jz(l_else, T_NEAR);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Compile the code that corresponds to the condition evaluating as true
|
2016-01-24 11:10:56 +01:00
|
|
|
Compile_Block(instr.flow_control.dest_offset);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// If there isn't an "ELSE" condition, we are done here
|
|
|
|
if (instr.flow_control.num_instructions == 0) {
|
2016-12-12 10:23:08 +01:00
|
|
|
L(l_else);
|
2015-07-23 05:25:30 +02:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
jmp(l_endif, T_NEAR);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
L(l_else);
|
2015-07-23 05:25:30 +02:00
|
|
|
// This code corresponds to the "ELSE" condition
|
|
|
|
// Comple the code that corresponds to the condition evaluating as false
|
2016-01-24 11:10:56 +01:00
|
|
|
Compile_Block(instr.flow_control.dest_offset + instr.flow_control.num_instructions);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
L(l_endif);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_LOOP(Instruction instr) {
|
2016-09-18 02:38:01 +02:00
|
|
|
Compile_Assert(instr.flow_control.dest_offset >= program_counter,
|
|
|
|
"Backwards loops not supported");
|
2016-04-13 05:35:36 +02:00
|
|
|
Compile_Assert(!looping, "Nested loops not supported");
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
looping = true;
|
|
|
|
|
2016-11-30 15:40:42 +01:00
|
|
|
// This decodes the fields from the integer uniform at index instr.flow_control.int_uniform_id.
|
|
|
|
// The Y (LOOPCOUNT_REG) and Z (LOOPINC) component are kept multiplied by 16 (Left shifted by
|
|
|
|
// 4 bits) to be used as an offset into the 16-byte vector registers later
|
2018-09-06 22:03:28 +02:00
|
|
|
std::size_t offset = Uniforms::GetIntUniformOffset(instr.flow_control.int_uniform_id);
|
2018-04-17 08:35:43 +02:00
|
|
|
mov(LOOPCOUNT, dword[UNIFORMS + offset]);
|
2016-12-13 06:44:15 +01:00
|
|
|
mov(LOOPCOUNT_REG, LOOPCOUNT);
|
|
|
|
shr(LOOPCOUNT_REG, 4);
|
2017-06-17 21:20:22 +02:00
|
|
|
and_(LOOPCOUNT_REG, 0xFF0); // Y-component is the start
|
2016-12-13 06:44:15 +01:00
|
|
|
mov(LOOPINC, LOOPCOUNT);
|
|
|
|
shr(LOOPINC, 12);
|
2017-06-17 21:20:22 +02:00
|
|
|
and_(LOOPINC, 0xFF0); // Z-component is the incrementer
|
2016-12-13 06:44:15 +01:00
|
|
|
movzx(LOOPCOUNT, LOOPCOUNT.cvt8()); // X-component is iteration count
|
|
|
|
add(LOOPCOUNT, 1); // Iteration count is X-component + 1
|
2016-12-12 10:23:08 +01:00
|
|
|
|
|
|
|
Label l_loop_start;
|
|
|
|
L(l_loop_start);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2018-05-18 22:28:58 +02:00
|
|
|
loop_break_label = Xbyak::Label();
|
2016-01-24 11:10:56 +01:00
|
|
|
Compile_Block(instr.flow_control.dest_offset + 1);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-13 06:44:15 +01:00
|
|
|
add(LOOPCOUNT_REG, LOOPINC); // Increment LOOPCOUNT_REG by Z-component
|
|
|
|
sub(LOOPCOUNT, 1); // Increment loop count by 1
|
|
|
|
jnz(l_loop_start); // Loop if not equal
|
2018-05-18 22:28:58 +02:00
|
|
|
L(*loop_break_label);
|
2018-10-05 12:37:55 +02:00
|
|
|
loop_break_label.reset();
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
looping = false;
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_JMP(Instruction instr) {
|
2015-07-23 05:25:30 +02:00
|
|
|
if (instr.opcode.Value() == OpCode::Id::JMPC)
|
|
|
|
Compile_EvaluateCondition(instr);
|
|
|
|
else if (instr.opcode.Value() == OpCode::Id::JMPU)
|
|
|
|
Compile_UniformCondition(instr);
|
|
|
|
else
|
|
|
|
UNREACHABLE();
|
|
|
|
|
2016-09-18 02:38:01 +02:00
|
|
|
bool inverted_condition =
|
|
|
|
(instr.opcode.Value() == OpCode::Id::JMPU) && (instr.flow_control.num_instructions & 1);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
Label& b = instruction_labels[instr.flow_control.dest_offset];
|
|
|
|
if (inverted_condition) {
|
|
|
|
jz(b, T_NEAR);
|
|
|
|
} else {
|
|
|
|
jnz(b, T_NEAR);
|
|
|
|
}
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2017-07-26 17:44:52 +02:00
|
|
|
static void Emit(GSEmitter* emitter, Math::Vec4<float24> (*output)[16]) {
|
|
|
|
emitter->Emit(*output);
|
|
|
|
}
|
|
|
|
|
|
|
|
void JitShader::Compile_EMIT(Instruction instr) {
|
|
|
|
Label have_emitter, end;
|
|
|
|
mov(rax, qword[STATE + offsetof(UnitState, emitter_ptr)]);
|
|
|
|
test(rax, rax);
|
|
|
|
jnz(have_emitter);
|
|
|
|
|
|
|
|
ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
|
2018-09-06 22:03:28 +02:00
|
|
|
mov(ABI_PARAM1, reinterpret_cast<std::size_t>("Execute EMIT on VS"));
|
2017-07-26 17:44:52 +02:00
|
|
|
CallFarFunction(*this, LogCritical);
|
|
|
|
ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
|
|
|
|
jmp(end);
|
|
|
|
|
|
|
|
L(have_emitter);
|
|
|
|
ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
|
|
|
|
mov(ABI_PARAM1, rax);
|
|
|
|
mov(ABI_PARAM2, STATE);
|
|
|
|
add(ABI_PARAM2, static_cast<Xbyak::uint32>(offsetof(UnitState, registers.output)));
|
|
|
|
CallFarFunction(*this, Emit);
|
|
|
|
ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
|
|
|
|
L(end);
|
|
|
|
}
|
|
|
|
|
|
|
|
void JitShader::Compile_SETE(Instruction instr) {
|
|
|
|
Label have_emitter, end;
|
|
|
|
mov(rax, qword[STATE + offsetof(UnitState, emitter_ptr)]);
|
|
|
|
test(rax, rax);
|
|
|
|
jnz(have_emitter);
|
|
|
|
|
|
|
|
ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
|
2018-09-06 22:03:28 +02:00
|
|
|
mov(ABI_PARAM1, reinterpret_cast<std::size_t>("Execute SETEMIT on VS"));
|
2017-07-26 17:44:52 +02:00
|
|
|
CallFarFunction(*this, LogCritical);
|
|
|
|
ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
|
|
|
|
jmp(end);
|
|
|
|
|
|
|
|
L(have_emitter);
|
|
|
|
mov(byte[rax + offsetof(GSEmitter, vertex_id)], instr.setemit.vertex_id);
|
|
|
|
mov(byte[rax + offsetof(GSEmitter, prim_emit)], instr.setemit.prim_emit);
|
|
|
|
mov(byte[rax + offsetof(GSEmitter, winding)], instr.setemit.winding);
|
|
|
|
L(end);
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_Block(unsigned end) {
|
2016-03-18 00:45:09 +01:00
|
|
|
while (program_counter < end) {
|
|
|
|
Compile_NextInstr();
|
|
|
|
}
|
|
|
|
}
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_Return() {
|
2016-03-18 00:45:09 +01:00
|
|
|
// Peek return offset on the stack and check if we're at that offset
|
2016-12-12 10:23:08 +01:00
|
|
|
mov(rax, qword[rsp + 8]);
|
|
|
|
cmp(eax, (program_counter));
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-03-18 00:45:09 +01:00
|
|
|
// If so, jump back to before CALL
|
2016-12-12 10:23:08 +01:00
|
|
|
Label b;
|
|
|
|
jnz(b);
|
|
|
|
ret();
|
|
|
|
L(b);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::Compile_NextInstr() {
|
2016-04-13 05:24:34 +02:00
|
|
|
if (std::binary_search(return_offsets.begin(), return_offsets.end(), program_counter)) {
|
2016-03-18 00:45:09 +01:00
|
|
|
Compile_Return();
|
|
|
|
}
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
L(instruction_labels[program_counter]);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-17 11:29:22 +01:00
|
|
|
Instruction instr = {(*program_code)[program_counter++]};
|
2016-03-18 00:51:43 +01:00
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
OpCode::Id opcode = instr.opcode.Value();
|
|
|
|
auto instr_func = instr_table[static_cast<unsigned>(opcode)];
|
|
|
|
|
|
|
|
if (instr_func) {
|
|
|
|
// JIT the instruction!
|
|
|
|
((*this).*instr_func)(instr);
|
|
|
|
} else {
|
|
|
|
// Unhandled instruction
|
2018-06-29 13:18:07 +02:00
|
|
|
LOG_CRITICAL(HW_GPU, "Unhandled instruction: 0x{:02x} (0x{:08x})",
|
2018-06-29 15:56:12 +02:00
|
|
|
static_cast<u32>(instr.opcode.Value().EffectiveOpCode()), instr.hex);
|
2015-07-23 05:25:30 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-04-13 05:34:03 +02:00
|
|
|
void JitShader::FindReturnOffsets() {
|
2016-03-18 00:45:09 +01:00
|
|
|
return_offsets.clear();
|
|
|
|
|
2018-09-06 22:03:28 +02:00
|
|
|
for (std::size_t offset = 0; offset < program_code->size(); ++offset) {
|
2016-12-17 11:29:22 +01:00
|
|
|
Instruction instr = {(*program_code)[offset]};
|
2016-03-18 00:45:09 +01:00
|
|
|
|
|
|
|
switch (instr.opcode.Value()) {
|
|
|
|
case OpCode::Id::CALL:
|
|
|
|
case OpCode::Id::CALLC:
|
|
|
|
case OpCode::Id::CALLU:
|
2016-09-18 02:38:01 +02:00
|
|
|
return_offsets.push_back(instr.flow_control.dest_offset +
|
|
|
|
instr.flow_control.num_instructions);
|
2016-03-18 00:45:09 +01:00
|
|
|
break;
|
2016-04-24 17:41:25 +02:00
|
|
|
default:
|
|
|
|
break;
|
2016-03-18 00:45:09 +01:00
|
|
|
}
|
|
|
|
}
|
2016-04-13 05:24:34 +02:00
|
|
|
|
|
|
|
// Sort for efficient binary search later
|
|
|
|
std::sort(return_offsets.begin(), return_offsets.end());
|
2016-03-18 00:45:09 +01:00
|
|
|
}
|
|
|
|
|
2017-05-08 23:37:38 +02:00
|
|
|
void JitShader::Compile(const std::array<u32, MAX_PROGRAM_CODE_LENGTH>* program_code_,
|
|
|
|
const std::array<u32, MAX_SWIZZLE_DATA_LENGTH>* swizzle_data_) {
|
2016-12-17 11:29:22 +01:00
|
|
|
program_code = program_code_;
|
|
|
|
swizzle_data = swizzle_data_;
|
|
|
|
|
2016-04-09 17:39:56 +02:00
|
|
|
// Reset flow control state
|
2016-12-12 10:23:08 +01:00
|
|
|
program = (CompiledShader*)getCurr();
|
2016-04-09 17:39:56 +02:00
|
|
|
program_counter = 0;
|
|
|
|
looping = false;
|
2016-12-12 10:23:08 +01:00
|
|
|
instruction_labels.fill(Xbyak::Label());
|
2016-04-09 17:39:56 +02:00
|
|
|
|
|
|
|
// Find all `CALL` instructions and identify return locations
|
|
|
|
FindReturnOffsets();
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2015-08-26 09:12:14 +02:00
|
|
|
// The stack pointer is 8 modulo 16 at the entry of a procedure
|
2017-02-03 13:22:43 +01:00
|
|
|
// We reserve 16 bytes and assign a dummy value to the first 8 bytes, to catch any potential
|
|
|
|
// return checks (see Compile_Return) that happen in shader main routine.
|
|
|
|
ABI_PushRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8, 16);
|
|
|
|
mov(qword[rsp + 8], 0xFFFFFFFFFFFFFFFFULL);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2018-04-17 08:35:43 +02:00
|
|
|
mov(UNIFORMS, ABI_PARAM1);
|
2016-12-12 10:23:08 +01:00
|
|
|
mov(STATE, ABI_PARAM2);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2018-05-28 13:28:03 +02:00
|
|
|
// Load address/loop registers
|
|
|
|
movsxd(ADDROFFS_REG_0, dword[STATE + offsetof(UnitState, address_registers[0])]);
|
|
|
|
movsxd(ADDROFFS_REG_1, dword[STATE + offsetof(UnitState, address_registers[1])]);
|
|
|
|
mov(LOOPCOUNT_REG, dword[STATE + offsetof(UnitState, address_registers[2])]);
|
|
|
|
shl(ADDROFFS_REG_0, 4);
|
|
|
|
shl(ADDROFFS_REG_1, 4);
|
|
|
|
shl(LOOPCOUNT_REG, 4);
|
|
|
|
|
|
|
|
// Load conditional code
|
|
|
|
mov(COND0, byte[STATE + offsetof(UnitState, conditional_code[0])]);
|
|
|
|
mov(COND1, byte[STATE + offsetof(UnitState, conditional_code[1])]);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Used to set a register to one
|
2016-09-18 02:38:01 +02:00
|
|
|
static const __m128 one = {1.f, 1.f, 1.f, 1.f};
|
2018-09-06 22:03:28 +02:00
|
|
|
mov(rax, reinterpret_cast<std::size_t>(&one));
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(ONE, xword[rax]);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
|
|
|
// Used to negate registers
|
2016-09-18 02:38:01 +02:00
|
|
|
static const __m128 neg = {-0.f, -0.f, -0.f, -0.f};
|
2018-09-06 22:03:28 +02:00
|
|
|
mov(rax, reinterpret_cast<std::size_t>(&neg));
|
2016-12-12 10:23:08 +01:00
|
|
|
movaps(NEGBIT, xword[rax]);
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-03-18 00:45:09 +01:00
|
|
|
// Jump to start of the shader program
|
2016-12-12 10:23:08 +01:00
|
|
|
jmp(ABI_PARAM3);
|
2016-03-18 00:45:09 +01:00
|
|
|
|
|
|
|
// Compile entire program
|
2016-12-17 11:29:22 +01:00
|
|
|
Compile_Block(static_cast<unsigned>(program_code->size()));
|
2015-08-12 06:00:44 +02:00
|
|
|
|
2016-04-13 05:29:25 +02:00
|
|
|
// Free memory that's no longer needed
|
2016-12-17 11:29:22 +01:00
|
|
|
program_code = nullptr;
|
|
|
|
swizzle_data = nullptr;
|
2016-04-13 05:29:25 +02:00
|
|
|
return_offsets.clear();
|
|
|
|
return_offsets.shrink_to_fit();
|
|
|
|
|
2016-12-12 10:23:08 +01:00
|
|
|
ready();
|
2015-07-23 05:25:30 +02:00
|
|
|
|
2016-12-17 11:29:22 +01:00
|
|
|
ASSERT_MSG(getSize() <= MAX_SHADER_SIZE, "Compiled a shader that exceeds the allocated size!");
|
2018-06-29 13:18:07 +02:00
|
|
|
LOG_DEBUG(HW_GPU, "Compiled shader size={}", getSize());
|
2015-08-12 06:00:44 +02:00
|
|
|
}
|
|
|
|
|
2017-11-25 04:18:55 +01:00
|
|
|
JitShader::JitShader() : Xbyak::CodeGenerator(MAX_SHADER_SIZE) {
|
|
|
|
CompilePrelude();
|
|
|
|
}
|
|
|
|
|
|
|
|
void JitShader::CompilePrelude() {
|
|
|
|
log2_subroutine = CompilePrelude_Log2();
|
|
|
|
exp2_subroutine = CompilePrelude_Exp2();
|
|
|
|
}
|
|
|
|
|
|
|
|
Xbyak::Label JitShader::CompilePrelude_Log2() {
|
|
|
|
Xbyak::Label subroutine;
|
|
|
|
|
|
|
|
// SSE does not have a log instruction, thus we must approximate.
|
|
|
|
// We perform this approximation first performaing a range reduction into the range [1.0, 2.0).
|
|
|
|
// A minimax polynomial which was fit for the function log2(x) / (x - 1) is then evaluated.
|
|
|
|
// We multiply the result by (x - 1) then restore the result into the appropriate range.
|
|
|
|
|
|
|
|
// Coefficients for the minimax polynomial.
|
|
|
|
// f(x) computes approximately log2(x) / (x - 1).
|
|
|
|
// f(x) = c4 + x * (c3 + x * (c2 + x * (c1 + x * c0)).
|
|
|
|
align(64);
|
|
|
|
const void* c0 = getCurr();
|
|
|
|
dd(0x3d74552f);
|
|
|
|
const void* c1 = getCurr();
|
|
|
|
dd(0xbeee7397);
|
|
|
|
const void* c2 = getCurr();
|
|
|
|
dd(0x3fbd96dd);
|
|
|
|
const void* c3 = getCurr();
|
|
|
|
dd(0xc02153f6);
|
|
|
|
const void* c4 = getCurr();
|
|
|
|
dd(0x4038d96c);
|
|
|
|
|
|
|
|
align(16);
|
|
|
|
const void* negative_infinity_vector = getCurr();
|
|
|
|
dd(0xff800000);
|
|
|
|
dd(0xff800000);
|
|
|
|
dd(0xff800000);
|
|
|
|
dd(0xff800000);
|
|
|
|
const void* default_qnan_vector = getCurr();
|
|
|
|
dd(0x7fc00000);
|
|
|
|
dd(0x7fc00000);
|
|
|
|
dd(0x7fc00000);
|
|
|
|
dd(0x7fc00000);
|
|
|
|
|
|
|
|
Xbyak::Label input_is_nan, input_is_zero, input_out_of_range;
|
|
|
|
|
|
|
|
align(16);
|
|
|
|
L(input_out_of_range);
|
|
|
|
je(input_is_zero);
|
|
|
|
movaps(SRC1, xword[rip + default_qnan_vector]);
|
|
|
|
ret();
|
|
|
|
L(input_is_zero);
|
|
|
|
movaps(SRC1, xword[rip + negative_infinity_vector]);
|
|
|
|
ret();
|
|
|
|
|
|
|
|
align(16);
|
|
|
|
L(subroutine);
|
|
|
|
|
|
|
|
// Here we handle edge cases: input in {NaN, 0, -Inf, Negative}.
|
|
|
|
xorps(SCRATCH, SCRATCH);
|
|
|
|
ucomiss(SCRATCH, SRC1);
|
|
|
|
jp(input_is_nan);
|
|
|
|
jae(input_out_of_range);
|
|
|
|
|
|
|
|
// Split input
|
|
|
|
movd(eax, SRC1);
|
|
|
|
mov(edx, eax);
|
|
|
|
and_(eax, 0x7f800000);
|
|
|
|
and_(edx, 0x007fffff);
|
|
|
|
movss(SCRATCH, xword[rip + c0]); // Preload c0.
|
|
|
|
or_(edx, 0x3f800000);
|
|
|
|
movd(SRC1, edx);
|
|
|
|
// SRC1 now contains the mantissa of the input.
|
|
|
|
mulss(SCRATCH, SRC1);
|
|
|
|
shr(eax, 23);
|
|
|
|
sub(eax, 0x7f);
|
|
|
|
cvtsi2ss(SCRATCH2, eax);
|
|
|
|
// SCRATCH2 now contains the exponent of the input.
|
|
|
|
|
|
|
|
// Complete computation of polynomial
|
|
|
|
addss(SCRATCH, xword[rip + c1]);
|
|
|
|
mulss(SCRATCH, SRC1);
|
|
|
|
addss(SCRATCH, xword[rip + c2]);
|
|
|
|
mulss(SCRATCH, SRC1);
|
|
|
|
addss(SCRATCH, xword[rip + c3]);
|
|
|
|
mulss(SCRATCH, SRC1);
|
|
|
|
subss(SRC1, ONE);
|
|
|
|
addss(SCRATCH, xword[rip + c4]);
|
|
|
|
mulss(SCRATCH, SRC1);
|
|
|
|
addss(SCRATCH2, SCRATCH);
|
|
|
|
|
|
|
|
// Duplicate result across vector
|
|
|
|
xorps(SRC1, SRC1); // break dependency chain
|
|
|
|
movss(SRC1, SCRATCH2);
|
|
|
|
L(input_is_nan);
|
|
|
|
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
|
|
|
|
|
|
|
|
ret();
|
|
|
|
|
|
|
|
return subroutine;
|
|
|
|
}
|
|
|
|
|
|
|
|
Xbyak::Label JitShader::CompilePrelude_Exp2() {
|
|
|
|
Xbyak::Label subroutine;
|
|
|
|
|
|
|
|
// SSE does not have a exp instruction, thus we must approximate.
|
|
|
|
// We perform this approximation first performaing a range reduction into the range [-0.5, 0.5).
|
|
|
|
// A minimax polynomial which was fit for the function exp2(x) is then evaluated.
|
|
|
|
// We then restore the result into the appropriate range.
|
|
|
|
|
|
|
|
align(64);
|
|
|
|
const void* input_max = getCurr();
|
|
|
|
dd(0x43010000);
|
|
|
|
const void* input_min = getCurr();
|
|
|
|
dd(0xc2fdffff);
|
|
|
|
const void* c0 = getCurr();
|
|
|
|
dd(0x3c5dbe69);
|
|
|
|
const void* half = getCurr();
|
|
|
|
dd(0x3f000000);
|
|
|
|
const void* c1 = getCurr();
|
|
|
|
dd(0x3d5509f9);
|
|
|
|
const void* c2 = getCurr();
|
|
|
|
dd(0x3e773cc5);
|
|
|
|
const void* c3 = getCurr();
|
|
|
|
dd(0x3f3168b3);
|
|
|
|
const void* c4 = getCurr();
|
|
|
|
dd(0x3f800016);
|
|
|
|
|
|
|
|
Xbyak::Label ret_label;
|
|
|
|
|
|
|
|
align(16);
|
|
|
|
L(subroutine);
|
|
|
|
|
|
|
|
// Handle edge cases
|
|
|
|
ucomiss(SRC1, SRC1);
|
|
|
|
jp(ret_label);
|
|
|
|
// Clamp to maximum range since we shift the value directly into the exponent.
|
|
|
|
minss(SRC1, xword[rip + input_max]);
|
|
|
|
maxss(SRC1, xword[rip + input_min]);
|
|
|
|
|
|
|
|
// Decompose input
|
|
|
|
movss(SCRATCH, SRC1);
|
|
|
|
movss(SCRATCH2, xword[rip + c0]); // Preload c0.
|
|
|
|
subss(SCRATCH, xword[rip + half]);
|
|
|
|
cvtss2si(eax, SCRATCH);
|
|
|
|
cvtsi2ss(SCRATCH, eax);
|
|
|
|
// SCRATCH now contains input rounded to the nearest integer.
|
|
|
|
add(eax, 0x7f);
|
|
|
|
subss(SRC1, SCRATCH);
|
|
|
|
// SRC1 contains input - round(input), which is in [-0.5, 0.5).
|
|
|
|
mulss(SCRATCH2, SRC1);
|
|
|
|
shl(eax, 23);
|
|
|
|
movd(SCRATCH, eax);
|
|
|
|
// SCRATCH contains 2^(round(input)).
|
|
|
|
|
|
|
|
// Complete computation of polynomial.
|
|
|
|
addss(SCRATCH2, xword[rip + c1]);
|
|
|
|
mulss(SCRATCH2, SRC1);
|
|
|
|
addss(SCRATCH2, xword[rip + c2]);
|
|
|
|
mulss(SCRATCH2, SRC1);
|
|
|
|
addss(SCRATCH2, xword[rip + c3]);
|
|
|
|
mulss(SRC1, SCRATCH2);
|
|
|
|
addss(SRC1, xword[rip + c4]);
|
|
|
|
mulss(SRC1, SCRATCH);
|
|
|
|
|
|
|
|
// Duplicate result across vector
|
|
|
|
L(ret_label);
|
|
|
|
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
|
|
|
|
|
|
|
|
ret();
|
|
|
|
|
|
|
|
return subroutine;
|
|
|
|
}
|
2015-08-12 06:00:44 +02:00
|
|
|
|
2015-07-23 05:25:30 +02:00
|
|
|
} // namespace Shader
|
|
|
|
|
|
|
|
} // namespace Pica
|