shader_jit: Add optimizations up to x86-64-v4 (#6668)

This commit is contained in:
Wunk 2023-07-11 09:21:37 -07:00 committed by GitHub
parent 6da4853360
commit a94af8ea62
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -338,15 +338,39 @@ void JitShader::Compile_SanitizedMul(Xmm src1, Xmm src2, Xmm scratch) {
// 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.
if (host_caps.has(Cpu::tAVX512F | Cpu::tAVX512VL | Cpu::tAVX512DQ)) {
vmulps(scratch, src1, src2);
// Mask of any NaN values found in the result
const Xbyak::Opmask zero_mask = k1;
vcmpunordps(zero_mask, scratch, scratch);
// Mask of any non-NaN inputs producing NaN results
vcmpordps(zero_mask | zero_mask, src1, src2);
knotb(zero_mask, zero_mask);
vmovaps(src1 | zero_mask | T_z, scratch);
return;
}
// Set scratch to mask of (src1 != NaN and src2 != NaN)
movaps(scratch, src1);
cmpordps(scratch, src2);
if (host_caps.has(Cpu::tAVX)) {
vcmpordps(scratch, src1, src2);
} else {
movaps(scratch, src1);
cmpordps(scratch, src2);
}
mulps(src1, src2);
// Set src2 to mask of (result == NaN)
movaps(src2, src1);
cmpunordps(src2, src2);
if (host_caps.has(Cpu::tAVX)) {
vcmpunordps(src2, src2, src1);
} else {
movaps(src2, src1);
cmpunordps(src2, src2);
}
// Clear components where scratch != src2 (i.e. if result is NaN where neither source was NaN)
xorps(scratch, src2);
@ -406,13 +430,20 @@ void JitShader::Compile_DP3(Instruction instr) {
Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
movaps(SRC2, SRC1);
shufps(SRC2, SRC2, _MM_SHUFFLE(1, 1, 1, 1));
if (host_caps.has(Cpu::tAVX)) {
vshufps(SRC3, SRC1, SRC1, _MM_SHUFFLE(2, 2, 2, 2));
vshufps(SRC2, SRC1, SRC1, _MM_SHUFFLE(1, 1, 1, 1));
vshufps(SRC1, SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
} else {
movaps(SRC2, SRC1);
shufps(SRC2, SRC2, _MM_SHUFFLE(1, 1, 1, 1));
movaps(SRC3, SRC1);
shufps(SRC3, SRC3, _MM_SHUFFLE(2, 2, 2, 2));
movaps(SRC3, SRC1);
shufps(SRC3, SRC3, _MM_SHUFFLE(2, 2, 2, 2));
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
}
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
addps(SRC1, SRC2);
addps(SRC1, SRC3);
@ -589,9 +620,15 @@ void JitShader::Compile_MOV(Instruction instr) {
void JitShader::Compile_RCP(Instruction instr) {
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
// TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
// performs this operation more accurately. This should be checked on hardware.
rcpss(SRC1, SRC1);
if (host_caps.has(Cpu::tAVX512F | Cpu::tAVX512VL)) {
// Accurate to 14 bits of precisions rather than 12 bits of rcpss
vrcp14ss(SRC1, SRC1, SRC1);
} else {
// TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
// performs this operation more accurately. This should be checked on hardware.
rcpss(SRC1, SRC1);
}
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
Compile_DestEnable(instr, SRC1);
@ -600,9 +637,15 @@ void JitShader::Compile_RCP(Instruction instr) {
void JitShader::Compile_RSQ(Instruction instr) {
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
// TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
// performs this operation more accurately. This should be checked on hardware.
rsqrtss(SRC1, SRC1);
if (host_caps.has(Cpu::tAVX512F | Cpu::tAVX512VL)) {
// Accurate to 14 bits of precisions rather than 12 bits of rsqrtss
vrsqrt14ss(SRC1, SRC1, SRC1);
} else {
// TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
// performs this operation more accurately. This should be checked on hardware.
rsqrtss(SRC1, SRC1);
}
shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
Compile_DestEnable(instr, SRC1);
@ -1050,32 +1093,47 @@ Xbyak::Label JitShader::CompilePrelude_Log2() {
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.
// Split input: SRC1=MANT[1,2) SCRATCH2=Exponent
if (host_caps.has(Cpu::tAVX512F | Cpu::tAVX512VL)) {
vgetexpss(SCRATCH2, SRC1, SRC1);
vgetmantss(SRC1, SRC1, SRC1, 0x0'0);
} else {
movd(eax, SRC1);
mov(edx, eax);
and_(eax, 0x7f800000);
and_(edx, 0x007fffff);
or_(edx, 0x3f800000);
movd(SRC1, edx);
// SRC1 now contains the mantissa of the input.
shr(eax, 23);
sub(eax, 0x7f);
cvtsi2ss(SCRATCH2, eax);
// SCRATCH2 now contains the exponent of the input.
}
movss(SCRATCH, xword[rip + c0]);
// 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);
if (host_caps.has(Cpu::tFMA)) {
vfmadd213ss(SCRATCH, SRC1, xword[rip + c1]);
vfmadd213ss(SCRATCH, SRC1, xword[rip + c2]);
vfmadd213ss(SCRATCH, SRC1, xword[rip + c3]);
vfmadd213ss(SCRATCH, SRC1, xword[rip + c4]);
subss(SRC1, ONE);
vfmadd231ss(SCRATCH2, SCRATCH, SRC1);
} else {
mulss(SCRATCH, SRC1);
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
@ -1122,33 +1180,69 @@ Xbyak::Label JitShader::CompilePrelude_Exp2() {
// 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)).
// Decompose input:
// SCRATCH=2^round(input)
// SRC1=input-round(input) [-0.5, 0.5)
if (host_caps.has(Cpu::tAVX512F | Cpu::tAVX512VL)) {
// input - 0.5
vsubss(SCRATCH, SRC1, xword[rip + half]);
// trunc(input - 0.5)
vrndscaless(SCRATCH2, SCRATCH, SCRATCH, _MM_FROUND_TRUNC);
// SCRATCH = 1 * 2^(trunc(input - 0.5))
vscalefss(SCRATCH, ONE, SCRATCH2);
// SRC1 = input-trunc(input - 0.5)
vsubss(SRC1, SRC1, SCRATCH2);
} else {
// 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]);
if (host_caps.has(Cpu::tAVX)) {
vsubss(SCRATCH, SRC1, xword[rip + half]);
} else {
movss(SCRATCH, SRC1);
subss(SCRATCH, xword[rip + half]);
}
if (host_caps.has(Cpu::tSSE41)) {
roundss(SCRATCH, SCRATCH, _MM_FROUND_TRUNC);
cvtss2si(eax, SCRATCH);
} else {
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).
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]);
movss(SCRATCH2, xword[rip + c0]);
if (host_caps.has(Cpu::tFMA)) {
vfmadd213ss(SCRATCH2, SRC1, xword[rip + c1]);
vfmadd213ss(SCRATCH2, SRC1, xword[rip + c2]);
vfmadd213ss(SCRATCH2, SRC1, xword[rip + c3]);
vfmadd213ss(SRC1, SCRATCH2, xword[rip + c4]);
} else {
mulss(SCRATCH2, SRC1);
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