void bumps_t::initialize (
const bump_specifier_t & b0, const bump_specifier_t & b1)
{
// Precompute the coefficients of four cubic polynomials in t, giving
// the two smoothstep regions of the each of the two bump functions.
v4f b0t = load4f (& b0.t0); // b0.t0 b0.t1 b0.t2 b0.t2
v4f b1t = load4f (& b1.t0); // b1.t0 b1.t1 b1.t2 b1.t2
v4f b0v = _mm_movelh_ps (load4f (& b0.v0), _mm_setzero_ps ()); // b0.v0 b0.v1
v4f b1v = _mm_movelh_ps (load4f (& b1.v0), _mm_setzero_ps ()); // b1.v0 b1.v1
v4f S = SHUFPS (b0t, b1t, (0, 2, 0, 2)); // b0.t0 b0.t2 b1.t0 b1.t2
v4f T = SHUFPS (b0t, b1t, (1, 3, 1, 3)); // b0.t1 b0.t3 b1.t1 b1.t3
v4f U = SHUFPS (b0v, b1v, (0, 2, 0, 2)); // b0.v0 0 b1.v0 0
v4f V1 = SHUFPS (b0v, b1v, (1, 0, 1, 0)); // b0.v1 b0.v0 b1.v1 b1.v0
v4f V2 = SHUFPS (b0v, b1v, (2, 1, 2, 1)); // 0 b0.v1 0 b1.v1
v4f V = V1 - V2;
v4f d = T - S;
v4f a = T + S;
v4f m = (V - U) / (d * d * d);
store4f (c [0], U + m * S * S * (a + d + d));
store4f (c [1], _mm_set1_ps (-6.0f) * m * S * T);
store4f (c [2], _mm_set1_ps (+3.0f) * m * a);
store4f (c [3], _mm_set1_ps (-2.0f) * m);
store4f (S0, S);
store4f (T0, T);
store4f (U0, U);
store4f (V0, V);
}
LinearFunc SamplerJitCache::CompileLinear(const SamplerID &id) {
_assert_msg_(G3D, id.linear, "Linear should be set on sampler id");
BeginWrite();
// We'll first write the nearest sampler, which we will CALL.
// This may differ slightly based on the "linear" flag.
const u8 *nearest = AlignCode16();
if (!Jit_ReadTextureFormat(id)) {
EndWrite();
SetCodePtr(const_cast<u8 *>(nearest));
return nullptr;
}
RET();
// Now the actual linear func, which is exposed externally.
const u8 *start = AlignCode16();
// NOTE: This doesn't use the general register mapping.
// POSIX: arg1=uptr, arg2=vptr, arg3=frac_u, arg4=frac_v, arg5=src, arg6=bufw, stack+8=level
// Win64: arg1=uptr, arg2=vptr, arg3=frac_u, arg4=frac_v, stack+40=src, stack+48=bufw, stack+56=level
//
// We map these to nearest CALLs, with order: u, v, src, bufw, level
// Let's start by saving a bunch of registers.
PUSH(R15);
PUSH(R14);
PUSH(R13);
PUSH(R12);
// Won't need frac_u/frac_v for a while.
PUSH(arg4Reg);
PUSH(arg3Reg);
// Extra space to restore alignment and save resultReg for lerp.
// TODO: Maybe use XMMs instead?
SUB(64, R(RSP), Imm8(24));
MOV(64, R(R12), R(arg1Reg));
MOV(64, R(R13), R(arg2Reg));
#ifdef _WIN32
// First arg now starts at 24 (extra space) + 48 (pushed stack) + 8 (ret address) + 32 (shadow space)
const int argOffset = 24 + 48 + 8 + 32;
MOV(64, R(R14), MDisp(RSP, argOffset));
MOV(32, R(R15), MDisp(RSP, argOffset + 8));
// level is at argOffset + 16.
#else
MOV(64, R(R14), R(arg5Reg));
MOV(32, R(R15), R(arg6Reg));
// level is at 24 + 48 + 8.
#endif
// Early exit on !srcPtr.
FixupBranch zeroSrc;
if (id.hasInvalidPtr) {
CMP(PTRBITS, R(R14), Imm8(0));
FixupBranch nonZeroSrc = J_CC(CC_NZ);
XOR(32, R(RAX), R(RAX));
zeroSrc = J(true);
SetJumpTarget(nonZeroSrc);
}
// At this point:
// R12=uptr, R13=vptr, stack+24=frac_u, stack+32=frac_v, R14=src, R15=bufw, stack+X=level
auto doNearestCall = [&](int off) {
MOV(32, R(uReg), MDisp(R12, off));
MOV(32, R(vReg), MDisp(R13, off));
MOV(64, R(srcReg), R(R14));
MOV(32, R(bufwReg), R(R15));
// Leave level, we just always load from RAM. Separate CLUTs is uncommon.
CALL(nearest);
MOV(32, MDisp(RSP, off), R(resultReg));
};
doNearestCall(0);
doNearestCall(4);
doNearestCall(8);
doNearestCall(12);
// Convert TL, TR, BL, BR to floats for easier blending.
if (!cpu_info.bSSE4_1) {
PXOR(XMM0, R(XMM0));
}
MOVD_xmm(fpScratchReg1, MDisp(RSP, 0));
MOVD_xmm(fpScratchReg2, MDisp(RSP, 4));
MOVD_xmm(fpScratchReg3, MDisp(RSP, 8));
MOVD_xmm(fpScratchReg4, MDisp(RSP, 12));
if (cpu_info.bSSE4_1) {
PMOVZXBD(fpScratchReg1, R(fpScratchReg1));
PMOVZXBD(fpScratchReg2, R(fpScratchReg2));
PMOVZXBD(fpScratchReg3, R(fpScratchReg3));
PMOVZXBD(fpScratchReg4, R(fpScratchReg4));
} else {
PUNPCKLBW(fpScratchReg1, R(XMM0));
PUNPCKLBW(fpScratchReg2, R(XMM0));
PUNPCKLBW(fpScratchReg3, R(XMM0));
PUNPCKLBW(fpScratchReg4, R(XMM0));
PUNPCKLWD(fpScratchReg1, R(XMM0));
PUNPCKLWD(fpScratchReg2, R(XMM0));
PUNPCKLWD(fpScratchReg3, R(XMM0));
PUNPCKLWD(fpScratchReg4, R(XMM0));
}
CVTDQ2PS(fpScratchReg1, R(fpScratchReg1));
CVTDQ2PS(fpScratchReg2, R(fpScratchReg2));
CVTDQ2PS(fpScratchReg3, R(fpScratchReg3));
CVTDQ2PS(fpScratchReg4, R(fpScratchReg4));
// Okay, now multiply the R sides by frac_u, and L by (256 - frac_u)...
MOVD_xmm(fpScratchReg5, MDisp(RSP, 24));
CVTDQ2PS(fpScratchReg5, R(fpScratchReg5));
SHUFPS(fpScratchReg5, R(fpScratchReg5), _MM_SHUFFLE(0, 0, 0, 0));
if (RipAccessible(by256)) {
MULPS(fpScratchReg5, M(by256)); // rip accessible
} else {
Crash(); // TODO
}
MOVAPS(XMM0, M(ones));
SUBPS(XMM0, R(fpScratchReg5));
MULPS(fpScratchReg1, R(XMM0));
MULPS(fpScratchReg2, R(fpScratchReg5));
MULPS(fpScratchReg3, R(XMM0));
MULPS(fpScratchReg4, R(fpScratchReg5));
// Now set top=fpScratchReg1, bottom=fpScratchReg3.
ADDPS(fpScratchReg1, R(fpScratchReg2));
ADDPS(fpScratchReg3, R(fpScratchReg4));
// Next, time for frac_v.
MOVD_xmm(fpScratchReg5, MDisp(RSP, 32));
CVTDQ2PS(fpScratchReg5, R(fpScratchReg5));
SHUFPS(fpScratchReg5, R(fpScratchReg5), _MM_SHUFFLE(0, 0, 0, 0));
MULPS(fpScratchReg5, M(by256));
MOVAPS(XMM0, M(ones));
SUBPS(XMM0, R(fpScratchReg5));
MULPS(fpScratchReg1, R(XMM0));
MULPS(fpScratchReg3, R(fpScratchReg5));
// Still at the 255 scale, now we're interpolated.
ADDPS(fpScratchReg1, R(fpScratchReg3));
// Time to convert back to a single 32 bit value.
CVTPS2DQ(fpScratchReg1, R(fpScratchReg1));
PACKSSDW(fpScratchReg1, R(fpScratchReg1));
PACKUSWB(fpScratchReg1, R(fpScratchReg1));
MOVD_xmm(R(resultReg), fpScratchReg1);
if (id.hasInvalidPtr) {
SetJumpTarget(zeroSrc);
}
ADD(64, R(RSP), Imm8(24));
POP(arg3Reg);
POP(arg4Reg);
POP(R12);
POP(R13);
POP(R14);
POP(R15);
RET();
EndWrite();
return (LinearFunc)start;
}
