DBT-RISE/DBT-RISE-TGC
12
0
Fork 0
Code Issues 1 Pull Requests 1 Projects Releases Wiki Activity

changes to ternary functions for Multiply-Add Instructions

Browse Source
This commit is contained in:
Eyck-Alexander Jentzsch 2025-02-17 15:48:30 +01:00
parent 9ba9d2432c
commit ac1322d66b
2 changed files with 51 additions and 52 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
gen_input/templates/interp/CORENAME.cpp.gtl
View File

@ -385,7 +385,6 @@ if(vector != null) {%>
throw new std::runtime_error("Unsupported sew bit value"); throw new std::runtime_error("Unsupported sew bit value");
} }
} }
<%}%> <%}%>
uint64_t fetch_count{0}; uint64_t fetch_count{0};
uint64_t tval{0}; uint64_t tval{0};

102
src/vm/vector_functions.hpp
View File

@ -94,40 +94,40 @@ template <> struct twice<uint64_t> { using type = __uint128_t; };
template <class T> using twice_t = typename twice<T>::type; // for convenience template <class T> using twice_t = typename twice<T>::type; // for convenience
template <typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_t = dest_elem_t> template <typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_t = dest_elem_t>
std::function<dest_elem_t(src2_elem_t, src1_elem_t)> get_funct(unsigned funct6, unsigned funct3) { std::function<dest_elem_t(dest_elem_t, src2_elem_t, src1_elem_t)> get_funct(unsigned funct6, unsigned funct3) {
if(funct3 == OPIVV || funct3 == OPIVX || funct3 == OPIVI) if(funct3 == OPIVV || funct3 == OPIVX || funct3 == OPIVI)
switch(funct6) { switch(funct6) {
case 0b000000: // VADD case 0b000000: // VADD
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; };
case 0b000010: // VSUB case 0b000010: // VSUB
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 - vs1; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 - vs1; };
case 0b000011: // VRSUB case 0b000011: // VRSUB
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 - vs2; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs1 - vs2; };
case 0b000100: // VMINU case 0b000100: // VMINU
return [](src2_elem_t vs2, src1_elem_t vs1) { return std::min(vs2, static_cast<src2_elem_t>(vs1)); }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return std::min(vs2, static_cast<src2_elem_t>(vs1)); };
case 0b000101: // VMIN case 0b000101: // VMIN
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return std::min(static_cast<std::make_signed_t<src2_elem_t>>(vs2), static_cast<std::make_signed_t<src2_elem_t>>(vs1)); return std::min(static_cast<std::make_signed_t<src2_elem_t>>(vs2), static_cast<std::make_signed_t<src2_elem_t>>(vs1));
}; };
case 0b000110: // VMAXU case 0b000110: // VMAXU
return [](src2_elem_t vs2, src1_elem_t vs1) { return std::max(vs2, static_cast<src2_elem_t>(vs1)); }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return std::max(vs2, static_cast<src2_elem_t>(vs1)); };
case 0b000111: // VMAX case 0b000111: // VMAX
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return std::max(static_cast<std::make_signed_t<src2_elem_t>>(vs2), static_cast<std::make_signed_t<src2_elem_t>>(vs1)); return std::max(static_cast<std::make_signed_t<src2_elem_t>>(vs2), static_cast<std::make_signed_t<src2_elem_t>>(vs1));
}; };
case 0b001001: // VAND case 0b001001: // VAND
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 & vs2; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs1 & vs2; };
case 0b001010: // VOR case 0b001010: // VOR
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 | vs2; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs1 | vs2; };
case 0b001011: // VXOR case 0b001011: // VXOR
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 ^ vs2; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs1 ^ vs2; };
// case 0b001100: // VRGATHER // case 0b001100: // VRGATHER
// case 0b001110: // VRGATHEREI16 // case 0b001110: // VRGATHEREI16
// case 0b001111: // VLSLIDEDOWN // case 0b001111: // VLSLIDEDOWN
case 0b010000: // VADC case 0b010000: // VADC
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; };
case 0b010010: // VSBC case 0b010010: // VSBC
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) - return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) -
static_cast<std::make_signed_t<src1_elem_t>>(vs1)); static_cast<std::make_signed_t<src1_elem_t>>(vs1));
}; };
@ -137,19 +137,19 @@ std::function<dest_elem_t(src2_elem_t, src1_elem_t)> get_funct(unsigned funct6,
// case 0b100010: // VSSUBU // case 0b100010: // VSSUBU
// case 0b100011: // VSSUB // case 0b100011: // VSSUB
case 0b100101: // VSLL case 0b100101: // VSLL
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 << (vs1 & shift_mask<src2_elem_t>()); }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 << (vs1 & shift_mask<src2_elem_t>()); };
// case 0b100111: // VSMUL // case 0b100111: // VSMUL
// case 0b100111: // VMV<NR>R // case 0b100111: // VMV<NR>R
case 0b101000: // VSRL case 0b101000: // VSRL
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 >> (vs1 & shift_mask<src2_elem_t>()); }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 >> (vs1 & shift_mask<src2_elem_t>()); };
case 0b101001: // VSRA case 0b101001: // VSRA
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<src2_elem_t>>(vs2) >> (vs1 & shift_mask<src2_elem_t>()); return static_cast<std::make_signed_t<src2_elem_t>>(vs2) >> (vs1 & shift_mask<src2_elem_t>());
}; };
case 0b101100: // VNSRL case 0b101100: // VNSRL
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 >> (vs1 & shift_mask<src2_elem_t>()); }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 >> (vs1 & shift_mask<src2_elem_t>()); };
case 0b101101: // VNSRA case 0b101101: // VNSRA
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<src2_elem_t>>(vs2) >> (vs1 & shift_mask<src2_elem_t>()); return static_cast<std::make_signed_t<src2_elem_t>>(vs2) >> (vs1 & shift_mask<src2_elem_t>());
}; };
// case 0b101110: // VNCLIPU // case 0b101110: // VNCLIPU
@ -185,14 +185,14 @@ std::function<dest_elem_t(src2_elem_t, src1_elem_t)> get_funct(unsigned funct6,
// case 0b011110: // VMNOR // case 0b011110: // VMNOR
// case 0b011111: // VMXNOR // case 0b011111: // VMXNOR
case 0b100000: // VDIVU case 0b100000: // VDIVU
return [](src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t {
if(vs1 == 0) if(vs1 == 0)
return -1; return -1;
else else
return vs2 / vs1; return vs2 / vs1;
}; };
case 0b100001: // VDIV case 0b100001: // VDIV
return [](src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t {
if(vs1 == 0) if(vs1 == 0)
return -1; return -1;
else if(vs2 == std::numeric_limits<std::make_signed_t<src2_elem_t>>::min() && else if(vs2 == std::numeric_limits<std::make_signed_t<src2_elem_t>>::min() &&
@ -202,14 +202,14 @@ std::function<dest_elem_t(src2_elem_t, src1_elem_t)> get_funct(unsigned funct6,
return static_cast<std::make_signed_t<src2_elem_t>>(vs2) / static_cast<std::make_signed_t<src1_elem_t>>(vs1); return static_cast<std::make_signed_t<src2_elem_t>>(vs2) / static_cast<std::make_signed_t<src1_elem_t>>(vs1);
}; };
case 0b100010: // VREMU case 0b100010: // VREMU
return [](src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t {
if(vs1 == 0) if(vs1 == 0)
return vs2; return vs2;
else else
return vs2 % vs1; return vs2 % vs1;
}; };
case 0b100011: // VREM case 0b100011: // VREM
return [](src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) -> dest_elem_t {
if(vs1 == 0) if(vs1 == 0)
return vs2; return vs2;
else if(vs2 == std::numeric_limits<std::make_signed_t<src2_elem_t>>::min() && else if(vs2 == std::numeric_limits<std::make_signed_t<src2_elem_t>>::min() &&
@ -219,68 +219,72 @@ std::function<dest_elem_t(src2_elem_t, src1_elem_t)> get_funct(unsigned funct6,
return static_cast<std::make_signed_t<src2_elem_t>>(vs2) % static_cast<std::make_signed_t<src1_elem_t>>(vs1); return static_cast<std::make_signed_t<src2_elem_t>>(vs2) % static_cast<std::make_signed_t<src1_elem_t>>(vs1);
}; };
case 0b100100: // VMULHU case 0b100100: // VMULHU
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return (static_cast<twice_t<src2_elem_t>>(vs2) * static_cast<twice_t<src2_elem_t>>(vs1)) >> sizeof(dest_elem_t) * 8; return (static_cast<twice_t<src2_elem_t>>(vs2) * static_cast<twice_t<src2_elem_t>>(vs1)) >> sizeof(dest_elem_t) * 8;
}; };
case 0b100101: // VMUL case 0b100101: // VMUL
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<src2_elem_t>>(vs2) * static_cast<std::make_signed_t<src1_elem_t>>(vs1); return static_cast<std::make_signed_t<src2_elem_t>>(vs2) * static_cast<std::make_signed_t<src1_elem_t>>(vs1);
}; };
case 0b100110: // VMULHSU case 0b100110: // VMULHSU
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) * return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) *
static_cast<twice_t<src2_elem_t>>(vs1)) >> static_cast<twice_t<src2_elem_t>>(vs1)) >>
sizeof(dest_elem_t) * 8; sizeof(dest_elem_t) * 8;
}; };
case 0b100111: // VMULH case 0b100111: // VMULH
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) * return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) *
static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src1_elem_t>>(vs1))) >> static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src1_elem_t>>(vs1))) >>
sizeof(dest_elem_t) * 8; sizeof(dest_elem_t) * 8;
}; };
// case 0b101001: // VMADD case 0b101001: // VMADD
// case 0b101011: // VNMSUB return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs1 * vd + vs2; };
// case 0b101101: // VMACC case 0b101011: // VNMSUB
// case 0b101111: // VNMSAC return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return -1 * (vs1 * vd) + vs2; };
case 0b101101: // VMACC
return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs1 * vs2 + vd; };
case 0b101111: // VNMSAC
return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return -1 * (vs1 * vs2) + vd; };
case 0b110000: // VWADDU case 0b110000: // VWADDU
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; };
case 0b110001: // VWADD case 0b110001: // VWADD
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) + return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) +
static_cast<std::make_signed_t<src1_elem_t>>(vs1)); static_cast<std::make_signed_t<src1_elem_t>>(vs1));
}; };
case 0b110010: // VWSUBU case 0b110010: // VWSUBU
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 - vs1; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 - vs1; };
case 0b110011: // VWSUB case 0b110011: // VWSUB
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) - return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) -
static_cast<std::make_signed_t<src1_elem_t>>(vs1)); static_cast<std::make_signed_t<src1_elem_t>>(vs1));
}; };
case 0b110100: // VWADDU.W case 0b110100: // VWADDU.W
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; };
case 0b110101: // VWADD.W case 0b110101: // VWADD.W
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) + return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) +
static_cast<std::make_signed_t<src1_elem_t>>(vs1)); static_cast<std::make_signed_t<src1_elem_t>>(vs1));
}; };
case 0b110110: // VWSUBU.W case 0b110110: // VWSUBU.W
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 - vs1; }; return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs2 - vs1; };
case 0b110111: // VWSUB.W case 0b110111: // VWSUB.W
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) - return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) -
static_cast<std::make_signed_t<src1_elem_t>>(vs1)); static_cast<std::make_signed_t<src1_elem_t>>(vs1));
}; };
case 0b111000: // VWMULU case 0b111000: // VWMULU
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return (static_cast<twice_t<src2_elem_t>>(vs2) * static_cast<twice_t<src2_elem_t>>(vs1)); return (static_cast<twice_t<src2_elem_t>>(vs2) * static_cast<twice_t<src2_elem_t>>(vs1));
}; };
case 0b111010: // VWMULSU case 0b111010: // VWMULSU
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) * return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) *
static_cast<twice_t<src2_elem_t>>(vs1)); static_cast<twice_t<src2_elem_t>>(vs1));
}; };
case 0b111011: // VWMUL case 0b111011: // VWMUL
return [](src2_elem_t vs2, src1_elem_t vs1) { return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) * return (static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) *
static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src1_elem_t>>(vs1))); static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src1_elem_t>>(vs1)));
}; };
@ -310,19 +314,18 @@ void vector_vector_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl,
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) { for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
bool mask_active = vm ? 1 : mask_reg[idx]; bool mask_active = vm ? 1 : mask_reg[idx];
if(mask_active) { if(mask_active) {
auto res = fn(vs2_view[idx], vs1_view[idx]); vd_view[idx] = fn(vd_view[idx], vs2_view[idx], vs1_view[idx]);
vd_view[idx] = res;
} else { } else {
vd_view[idx] = vtype.vma() ? vd_view[idx] : vd_view[idx]; vd_view[idx] = vtype.vma() ? vd_view[idx] : vd_view[idx];
} }
} }
} else if(carry == carry_t::SUB_CARRY) { } else if(carry == carry_t::SUB_CARRY) {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) { for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
vd_view[idx] = fn(vs2_view[idx], vs1_view[idx]) - mask_reg[idx]; vd_view[idx] = fn(vd_view[idx], vs2_view[idx], vs1_view[idx]) - mask_reg[idx];
} }
} else { } else {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) { for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
vd_view[idx] = fn(vs2_view[idx], vs1_view[idx]) + mask_reg[idx]; vd_view[idx] = fn(vd_view[idx], vs2_view[idx], vs1_view[idx]) + mask_reg[idx];
} }
} }
// elements w/ index larger than elem_count are in the tail (fractional LMUL) // elements w/ index larger than elem_count are in the tail (fractional LMUL)
@ -347,21 +350,18 @@ void vector_imm_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, ui
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) { for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
bool mask_active = vm ? 1 : mask_reg[idx]; bool mask_active = vm ? 1 : mask_reg[idx];
if(mask_active) { if(mask_active) {
vd_view[idx] = fn(vs2_view[idx], imm); vd_view[idx] = fn(vd_view[idx], vs2_view[idx], imm);
} else { } else {
vd_view[idx] = vtype.vma() ? vd_view[idx] : vd_view[idx]; vd_view[idx] = vtype.vma() ? vd_view[idx] : vd_view[idx];
} }
} }
} else if(carry == carry_t::SUB_CARRY) { } else if(carry == carry_t::SUB_CARRY) {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) { for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
auto val1 = fn(vs2_view[idx], imm); vd_view[idx] = fn(vd_view[idx], vs2_view[idx], imm) - mask_reg[idx];
auto val2 = static_cast<std::make_signed_t<dest_elem_t>>(mask_reg[idx]);
auto diff = val1 - val2;
vd_view[idx] = fn(vs2_view[idx], imm) - mask_reg[idx];
} }
} else { } else {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) { for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
vd_view[idx] = fn(vs2_view[idx], imm) + mask_reg[idx]; vd_view[idx] = fn(vd_view[idx], vs2_view[idx], imm) + mask_reg[idx];
} }
} }
// elements w/ index larger than elem_count are in the tail (fractional LMUL) // elements w/ index larger than elem_count are in the tail (fractional LMUL)