PRIM_STATIC pstatus_t sse2_RGBToRGB_16s8u_P3AC4R(
const INT16 *pSrc[3], /* 16-bit R,G, and B arrays */
INT32 srcStep, /* bytes between rows in source data */
BYTE *pDst, /* 32-bit interleaved ARGB (ABGR?) data */
INT32 dstStep, /* bytes between rows in dest data */
const prim_size_t *roi) /* region of interest */
{
const UINT16 *r = (const UINT16 *) (pSrc[0]);
const UINT16 *g = (const UINT16 *) (pSrc[1]);
const UINT16 *b = (const UINT16 *) (pSrc[2]);
BYTE *out;
int srcbump, dstbump, y;
/* Ensure 16-byte alignment on all pointers,
* that width is a multiple of 8,
* and that the next row will also remain aligned.
* Since this is usually used for 64x64 aligned arrays,
* these checks should presumably pass.
*/
if ((((ULONG_PTR) (pSrc[0]) & 0x0f) != 0)
|| (((ULONG_PTR) (pSrc[1]) & 0x0f) != 0)
|| (((ULONG_PTR) (pSrc[2]) & 0x0f) != 0)
|| (((ULONG_PTR) pDst & 0x0f) != 0)
|| (roi->width & 0x0f)
|| (srcStep & 0x0f)
|| (dstStep & 0x0f))
{
return general_RGBToRGB_16s8u_P3AC4R(pSrc, srcStep, pDst, dstStep, roi);
}
out = (BYTE *) pDst;
srcbump = (srcStep - (roi->width * sizeof(UINT16))) / sizeof(UINT16);
dstbump = (dstStep - (roi->width * sizeof(UINT32)));
for (y=0; y<roi->height; ++y)
{
int width = roi->width;
do {
__m128i R0, R1, R2, R3, R4;
/* The comments below pretend these are 8-byte registers
* rather than 16-byte, for readability.
*/
R0 = LOAD128(b); b += 8; /* R0 = 00B300B200B100B0 */
R1 = LOAD128(b); b += 8; /* R1 = 00B700B600B500B4 */
PACKUSWB(R0,R1); /* R0 = B7B6B5B4B3B2B1B0 */
R1 = LOAD128(g); g += 8; /* R1 = 00G300G200G100G0 */
R2 = LOAD128(g); g += 8; /* R2 = 00G700G600G500G4 */
PACKUSWB(R1,R2); /* R1 = G7G6G5G4G3G2G1G0 */
R2 = R1; /* R2 = G7G6G5G4G3G2G1G0 */
PUNPCKLBW(R2,R0); /* R2 = G3B3G2B2G1B1G0B0 */
PUNPCKHBW(R1,R0); /* R1 = G7B7G6B7G5B5G4B4 */
R0 = LOAD128(r); r += 8; /* R0 = 00R300R200R100R0 */
R3 = LOAD128(r); r += 8; /* R3 = 00R700R600R500R4 */
PACKUSWB(R0,R3); /* R0 = R7R6R5R4R3R2R1R0 */
R3 = XMM_ALL_ONES; /* R3 = FFFFFFFFFFFFFFFF */
R4 = R3; /* R4 = FFFFFFFFFFFFFFFF */
PUNPCKLBW(R4,R0); /* R4 = FFR3FFR2FFR1FFR0 */
PUNPCKHBW(R3,R0); /* R3 = FFR7FFR6FFR5FFR4 */
R0 = R4; /* R0 = R4 */
PUNPCKLWD(R0,R2); /* R0 = FFR1G1B1FFR0G0B0 */
PUNPCKHWD(R4,R2); /* R4 = FFR3G3B3FFR2G2B2 */
R2 = R3; /* R2 = R3 */
PUNPCKLWD(R2,R1); /* R2 = FFR5G5B5FFR4G4B4 */
PUNPCKHWD(R3,R1); /* R3 = FFR7G7B7FFR6G6B6 */
STORE128(out, R0); out += 16; /* FFR1G1B1FFR0G0B0 */
STORE128(out, R4); out += 16; /* FFR3G3B3FFR2G2B2 */
STORE128(out, R2); out += 16; /* FFR5G5B5FFR4G4B4 */
STORE128(out, R3); out += 16; /* FFR7G7B7FFR6G6B6 */
} while (width -= 16);
/* Jump to next row. */
r += srcbump;
g += srcbump;
b += srcbump;
out += dstbump;
}
return PRIMITIVES_SUCCESS;
}
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;
}
