riscv
/
riscv-isa-sim
mirror of https://github.com/riscv-software-src/riscv-isa-sim.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

Merge pull request #1972 from riscv-software-src/fix-vlen-32 

Fix mask element accesses under VLEN=32
pull/1974/head
Andrew Waterman 1 year ago
committed by GitHub
parent
d85cd1081a 7988172561
commit
a3f126b3bc
No known key found for this signature in database GPG Key ID: B5690EEEBB952194
9 changed files with 49 additions and 89 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 6
      riscv/insns/vcompress_vm.h
  2. 12
      riscv/insns/vcpop_m.h
  3. 3
      riscv/insns/vfirst_m.h
  4. 8
      riscv/insns/viota_m.h
  5. 17
      riscv/insns/vmsbf_m.h
  6. 18
      riscv/insns/vmsif_m.h
  7. 16
      riscv/insns/vmsof_m.h
  8. 47
      riscv/v_ext_macros.h
  9. 11
      riscv/vector_unit.h

6
riscv/insns/vcompress_vm.h
View File

@ -9,11 +9,7 @@ require_noover(insn.rd(), P.VU.vflmul, insn.rs1(), 1);
reg_t pos = 0;
VI_GENERAL_LOOP_BASE
const int midx = i / 64;
const int mpos = i % 64;
bool do_mask = (P.VU.elt<uint64_t>(rs1_num, midx) >> mpos) & 0x1;
if (do_mask) {
if (P.VU.mask_elt(rs1_num, i)) {
switch (sew) {
case e8:
P.VU.elt<uint8_t>(rd_num, pos, true) = P.VU.elt<uint8_t>(rs2_num, i);

12
riscv/insns/vcpop_m.h
View File

@ -6,15 +6,7 @@ reg_t rs2_num = insn.rs2();
require(P.VU.vstart->read() == 0);
reg_t popcount = 0;
for (reg_t i=P.VU.vstart->read(); i<vl; ++i) {
const int midx = i / 32;
const int mpos = i % 32;
bool vs2_lsb = ((P.VU.elt<uint32_t>(rs2_num, midx ) >> mpos) & 0x1) == 1;
if (insn.v_vm() == 1) {
popcount += vs2_lsb;
} else {
bool do_mask = (P.VU.elt<uint32_t>(0, midx) >> mpos) & 0x1;
popcount += (vs2_lsb && do_mask);
}
bool vs2_bit = P.VU.mask_elt(rs2_num, i);
popcount += vs2_bit && (insn.v_vm() || P.VU.mask_elt(0, i));
}
WRITE_RD(popcount);

3
riscv/insns/vfirst_m.h
View File

@ -8,8 +8,7 @@ reg_t pos = -1;
for (reg_t i=P.VU.vstart->read(); i < vl; ++i) {
VI_LOOP_ELEMENT_SKIP()
bool vs2_lsb = ((P.VU.elt<uint64_t>(rs2_num, midx ) >> mpos) & 0x1) == 1;
if (vs2_lsb) {
if (P.VU.mask_elt(rs2_num, i)) {
pos = i;
break;
}

8
riscv/insns/viota_m.h
View File

@ -12,15 +12,11 @@ require_noover(rd_num, P.VU.vflmul, rs2_num, 1);
int cnt = 0;
for (reg_t i = 0; i < vl; ++i) {
const int midx = i / 64;
const int mpos = i % 64;
bool vs2_lsb = ((P.VU.elt<uint64_t>(rs2_num, midx) >> mpos) & 0x1) == 1;
bool do_mask = (P.VU.elt<uint64_t>(0, midx) >> mpos) & 0x1;
bool do_mask = P.VU.mask_elt(0, i);
bool has_one = false;
if (insn.v_vm() == 1 || (insn.v_vm() == 0 && do_mask)) {
if (vs2_lsb) {
if (P.VU.mask_elt(rs2_num, i)) {
has_one = true;
}
}

17
riscv/insns/vmsbf_m.h
View File

@ -11,22 +11,17 @@ reg_t rs2_num = insn.rs2();
bool has_one = false;
for (reg_t i = P.VU.vstart->read(); i < vl; ++i) {
const int midx = i / 64;
const int mpos = i % 64;
const uint64_t mmask = UINT64_C(1) << mpos; \
bool vs2_lsb = ((P.VU.elt<uint64_t>(rs2_num, midx) >> mpos) & 0x1) == 1;
bool do_mask = (P.VU.elt<uint64_t>(0, midx) >> mpos) & 0x1;
bool vs2_lsb = P.VU.mask_elt(rs2_num, i);
bool do_mask = P.VU.mask_elt(0, i);
if (insn.v_vm() == 1 || (insn.v_vm() == 0 && do_mask)) {
auto &vd = P.VU.elt<uint64_t>(rd_num, midx, true);
uint64_t res = 0;
bool res = false;
if (!has_one && !vs2_lsb) {
res = 1;
res = true;
} else if (!has_one && vs2_lsb) {
has_one = true;
}
vd = (vd & ~mmask) | ((res << mpos) & mmask);
P.VU.set_mask_elt(rd_num, i, res);
}
}

18
riscv/insns/vmsif_m.h
View File

@ -11,22 +11,18 @@ reg_t rs2_num = insn.rs2();
bool has_one = false;
for (reg_t i = P.VU.vstart->read(); i < vl; ++i) {
const int midx = i / 64;
const int mpos = i % 64;
const uint64_t mmask = UINT64_C(1) << mpos; \
bool vs2_lsb = ((P.VU.elt<uint64_t>(rs2_num, midx ) >> mpos) & 0x1) == 1;
bool do_mask = (P.VU.elt<uint64_t>(0, midx) >> mpos) & 0x1;
bool vs2_lsb = P.VU.mask_elt(rs2_num, i);
bool do_mask = P.VU.mask_elt(0, i);
if (insn.v_vm() == 1 || (insn.v_vm() == 0 && do_mask)) {
auto &vd = P.VU.elt<uint64_t>(rd_num, midx, true);
uint64_t res = 0;
bool res = false;
if (!has_one && !vs2_lsb) {
res = 1;
res = true;
} else if (!has_one && vs2_lsb) {
has_one = true;
res = 1;
res = true;
}
vd = (vd & ~mmask) | ((res << mpos) & mmask);
P.VU.set_mask_elt(rd_num, i, res);
}
}

16
riscv/insns/vmsof_m.h
View File

@ -11,20 +11,16 @@ reg_t rs2_num = insn.rs2();
bool has_one = false;
for (reg_t i = P.VU.vstart->read() ; i < vl; ++i) {
const int midx = i / 64;
const int mpos = i % 64;
const uint64_t mmask = UINT64_C(1) << mpos; \
bool vs2_lsb = ((P.VU.elt<uint64_t>(rs2_num, midx ) >> mpos) & 0x1) == 1;
bool do_mask = (P.VU.elt<uint64_t>(0, midx) >> mpos) & 0x1;
bool vs2_lsb = P.VU.mask_elt(rs2_num, i);
bool do_mask = P.VU.mask_elt(0, i);
if (insn.v_vm() == 1 || (insn.v_vm() == 0 && do_mask)) {
uint64_t &vd = P.VU.elt<uint64_t>(rd_num, midx, true);
uint64_t res = 0;
bool res = false;
if (!has_one && vs2_lsb) {
has_one = true;
res = 1;
res = true;
}
vd = (vd & ~mmask) | ((res << mpos) & mmask);
P.VU.set_mask_elt(rd_num, i, res);
}
}

47
riscv/v_ext_macros.h
View File

@ -8,16 +8,10 @@
//
// vector: masking skip helper
//
#define VI_MASK_VARS \
const int midx = i / 64; \
const int mpos = i % 64;
#define VI_LOOP_ELEMENT_SKIP(BODY) \
VI_MASK_VARS \
if (insn.v_vm() == 0) { \
BODY; \
bool skip = ((P.VU.elt<uint64_t>(0, midx) >> mpos) & 0x1) == 0; \
if (skip) { \
if (!P.VU.mask_elt(0, i)) { \
continue; \
} \
}
@ -231,24 +225,18 @@ static inline bool is_overlapped_widen(const int astart, int asize,
#define VI_LOOP_CARRY_BASE \
VI_GENERAL_LOOP_BASE \
VI_MASK_VARS \
auto v0 = P.VU.elt<uint64_t>(0, midx); \
const uint64_t mmask = UINT64_C(1) << mpos; \
const uint128_t op_mask = (UINT64_MAX >> (64 - sew)); \
uint64_t carry = insn.v_vm() == 0 ? (v0 >> mpos) & 0x1 : 0; \
uint128_t res = 0; \
auto &vd = P.VU.elt<uint64_t>(rd_num, midx, true);
uint64_t carry = insn.v_vm() == 0 ? P.VU.mask_elt(0, i) : 0; \
bool res = false;
#define VI_LOOP_CARRY_END \
vd = (vd & ~mmask) | (((res) << mpos) & mmask); \
P.VU.set_mask_elt(insn.rd(), i, res); \
} \
P.VU.vstart->write(0);
#define VI_LOOP_WITH_CARRY_BASE \
VI_GENERAL_LOOP_BASE \
VI_MASK_VARS \
auto &v0 = P.VU.elt<uint64_t>(0, midx); \
const uint128_t op_mask = (UINT64_MAX >> (64 - sew)); \
uint64_t carry = (v0 >> mpos) & 0x1;
uint64_t carry = P.VU.mask_elt(0, i);
#define VI_LOOP_CMP_BASE \
require(P.VU.vsew >= e8 && P.VU.vsew <= e64); \
@ -260,12 +248,10 @@ static inline bool is_overlapped_widen(const int astart, int asize,
reg_t rs2_num = insn.rs2(); \
for (reg_t i = P.VU.vstart->read(); i < vl; ++i) { \
VI_LOOP_ELEMENT_SKIP(); \
uint64_t mmask = UINT64_C(1) << mpos; \
uint64_t &vdi = P.VU.elt<uint64_t>(insn.rd(), midx, true); \
uint64_t res = 0;
bool res = false;
#define VI_LOOP_CMP_END \
vdi = (vdi & ~mmask) | (((res) << mpos) & mmask); \
P.VU.set_mask_elt(insn.rd(), i, res); \
} \
P.VU.vstart->write(0);
@ -274,13 +260,9 @@ static inline bool is_overlapped_widen(const int astart, int asize,
require_vector(true); \
reg_t vl = P.VU.vl->read(); \
for (reg_t i = P.VU.vstart->read(); i < vl; ++i) { \
int midx = i / 64; \
int mpos = i % 64; \
uint64_t mmask = UINT64_C(1) << mpos; \
uint64_t vs2 = P.VU.elt<uint64_t>(insn.rs2(), midx); \
uint64_t vs1 = P.VU.elt<uint64_t>(insn.rs1(), midx); \
uint64_t &res = P.VU.elt<uint64_t>(insn.rd(), midx, true); \
res = (res & ~mmask) | ((op) & (1ULL << mpos)); \
bool vs2 = P.VU.mask_elt(insn.rs2(), i); \
bool vs1 = P.VU.mask_elt(insn.rs1(), i); \
P.VU.set_mask_elt(insn.rd(), i, (op)); \
} \
P.VU.vstart->write(0);
@ -523,8 +505,7 @@ static inline bool is_overlapped_widen(const int astart, int asize,
// merge and copy loop
#define VI_MERGE_VARS \
VI_MASK_VARS \
bool UNUSED use_first = (P.VU.elt<uint64_t>(0, midx) >> mpos) & 0x1;
bool UNUSED use_first = P.VU.mask_elt(0, i);
#define VI_MERGE_LOOP_BASE \
VI_GENERAL_LOOP_BASE \
@ -1482,9 +1463,7 @@ VI_VX_ULOOP({ \
VI_VFP_COMMON \
for (reg_t i = P.VU.vstart->read(); i < vl; ++i) { \
VI_LOOP_ELEMENT_SKIP(); \
uint64_t mmask = UINT64_C(1) << mpos; \
uint64_t &vd = P.VU.elt<uint64_t>(rd_num, midx, true); \
uint64_t res = 0;
bool res = false;
#define VI_VFP_LOOP_REDUCTION_BASE(width) \
float##width##_t vd_0 = P.VU.elt<float##width##_t>(rd_num, 0); \
@ -1562,7 +1541,7 @@ VI_VX_ULOOP({ \
case e16: \
case e32: \
case e64: { \
vd = (vd & ~mmask) | (((res) << mpos) & mmask); \
P.VU.set_mask_elt(insn.rd(), i, res); \
break; \
} \
default: \

11
riscv/vector_unit.h
View File

@ -108,6 +108,17 @@ public:
template<typename EG> EG&
elt_group(reg_t vReg, reg_t n, bool is_write = false);
bool mask_elt(reg_t vReg, reg_t n)
{
return (elt<uint8_t>(vReg, n / 8) >> (n % 8)) & 1;
}
void set_mask_elt(reg_t vReg, reg_t n, bool value)
{
auto& e = elt<uint8_t>(vReg, n / 8, true);
e = (e & ~(1U << (n % 8))) | (value << (n % 8));
}
public:
void reset();

Write Preview
Loading…
Cancel
Save