mirror of
https://github.com/hrydgard/ppsspp.git
synced 2026-11-18 11:54:30 +01:00
interp: Use software dot for better accuracy.
This gets some vdot/similar tests to give more accurate results compared to hardware. Also added flushing of zero and NaNs.
This commit is contained in:
1 parent
730fd0616e
commit
b8e3a3070a
3 files changed
+149
-36
No files matched your search
+63
-36
@@ -463,20 +463,21 @@ namespace MIPSInt
|
||||
|
||||
for (int a = 0; a < n; a++) {
|
||||
for (int b = 0; b < n; b++) {
|
||||
float sum = 0.0f;
|
||||
union { float f; uint32_t u; } sum = { 0.0f };
|
||||
if (a == n - 1 && b == n - 1) {
|
||||
// S and T prefixes work on the final (or maybe first, in reverse?) dot.
|
||||
ApplySwizzleS(&s[b * 4], V_Quad);
|
||||
ApplySwizzleT(&t[a * 4], V_Quad);
|
||||
for (int c = 0; c < 4; c++) {
|
||||
sum += s[b * 4 + c] * t[a * 4 + c];
|
||||
}
|
||||
} else {
|
||||
for (int c = 0; c < n; c++) {
|
||||
sum += s[b * 4 + c] * t[a * 4 + c];
|
||||
}
|
||||
}
|
||||
d[a * 4 + b] = sum;
|
||||
|
||||
sum.f = vfpu_dot(&s[b * 4], &t[a * 4]);
|
||||
if (my_isnan(sum.f)) {
|
||||
sum.u = 0x7f800001;
|
||||
} else if ((sum.u & 0x7F800000) == 0) {
|
||||
sum.u &= 0xFF800000;
|
||||
}
|
||||
|
||||
d[a * 4 + b] = sum.f;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1116,7 +1117,7 @@ namespace MIPSInt
|
||||
|
||||
void Int_VDot(MIPSOpcode op) {
|
||||
float s[4]{}, t[4]{};
|
||||
float d;
|
||||
union { float f; uint32_t u; } d;
|
||||
int vd = _VD;
|
||||
int vs = _VS;
|
||||
int vt = _VT;
|
||||
@@ -1125,12 +1126,16 @@ namespace MIPSInt
|
||||
ApplySwizzleS(s, V_Quad);
|
||||
ReadVector(t, sz, vt);
|
||||
ApplySwizzleT(t, V_Quad);
|
||||
d = 0.0f;
|
||||
for (int i = 0; i < 4; i++) {
|
||||
d += s[i] * t[i];
|
||||
|
||||
d.f = vfpu_dot(s, t);
|
||||
if (my_isnan(d.f)) {
|
||||
d.u = 0x7f800001;
|
||||
} else if ((d.u & 0x7F800000) == 0) {
|
||||
d.u &= 0xFF800000;
|
||||
}
|
||||
ApplyPrefixD(&d, V_Single);
|
||||
WriteVector(&d, V_Single, vd);
|
||||
|
||||
ApplyPrefixD(&d.f, V_Single);
|
||||
WriteVector(&d.f, V_Single, vd);
|
||||
PC += 4;
|
||||
EatPrefixes();
|
||||
}
|
||||
@@ -1558,7 +1563,8 @@ namespace MIPSInt
|
||||
}
|
||||
|
||||
void Int_Vtfm(MIPSOpcode op) {
|
||||
float s[16]{}, t[4]{}, d[4];
|
||||
float s[16]{}, t[4]{};
|
||||
FloatBits d;
|
||||
int vd = _VD;
|
||||
int vs = _VS;
|
||||
int vt = _VT;
|
||||
@@ -1572,13 +1578,24 @@ namespace MIPSInt
|
||||
ReadMatrix(s, msz, vs);
|
||||
ReadVector(t, sz, vt);
|
||||
|
||||
for (int i = 0; i < ins; i++) {
|
||||
d[i] = s[i * 4] * t[0];
|
||||
for (int k = 1; k < tn; k++) {
|
||||
d[i] += s[i * 4 + k] * t[k];
|
||||
float t2[4];
|
||||
for (int i = 0; i < 4; i++) {
|
||||
if (i < tn) {
|
||||
t2[i] = t[i];
|
||||
} else if (i == ins) {
|
||||
t2[i] = 1.0f;
|
||||
} else {
|
||||
t2[i] = 0.0f;
|
||||
}
|
||||
if (ins >= n) {
|
||||
d[i] += s[i * 4 + ins];
|
||||
}
|
||||
|
||||
for (int i = 0; i < ins; i++) {
|
||||
d.f[i] = vfpu_dot(&s[i * 4], t2);
|
||||
|
||||
if (my_isnan(d.f[i])) {
|
||||
d.u[i] = 0x7f800001;
|
||||
} else if ((d.u[i] & 0x7F800000) == 0) {
|
||||
d.u[i] &= 0xFF800000;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1599,17 +1616,20 @@ namespace MIPSInt
|
||||
ApplyPrefixST(t, VFPURewritePrefix(VFPU_CTRL_TPREFIX, tprefixRemove, tprefixAdd), V_Quad);
|
||||
|
||||
// Really this is the operation all rows probably use (with constant wiring.)
|
||||
d[ins] = s[ins * 4] * t[0];
|
||||
for (int k = 1; k < 4; k++) {
|
||||
d[ins] += s[ins * 4 + k] * t[k];
|
||||
d.f[ins] = vfpu_dot(&s[ins * 4], t);
|
||||
|
||||
if (my_isnan(d.f[ins])) {
|
||||
d.u[ins] = 0x7f800001;
|
||||
} else if ((d.u[ins] & 0x7F800000) == 0) {
|
||||
d.u[ins] &= 0xFF800000;
|
||||
}
|
||||
|
||||
// D prefix applies to the last element only.
|
||||
u32 lastmask = (currentMIPS->vfpuCtrl[VFPU_CTRL_DPREFIX] & (1 << 8)) << ins;
|
||||
u32 lastsat = (currentMIPS->vfpuCtrl[VFPU_CTRL_DPREFIX] & 3) << (ins + ins);
|
||||
currentMIPS->vfpuCtrl[VFPU_CTRL_DPREFIX] = lastmask | lastsat;
|
||||
ApplyPrefixD(d, sz);
|
||||
WriteVector(d, sz, vd);
|
||||
ApplyPrefixD(d.f, sz);
|
||||
WriteVector(d.f, sz, vd);
|
||||
PC += 4;
|
||||
EatPrefixes();
|
||||
}
|
||||
@@ -1956,7 +1976,8 @@ namespace MIPSInt
|
||||
}
|
||||
|
||||
void Int_VecDo3(MIPSOpcode op) {
|
||||
float s[4], t[4], d[4];
|
||||
float s[4], t[4];
|
||||
FloatBits d;
|
||||
int vd = _VD;
|
||||
int vs = _VS;
|
||||
int vt = _VT;
|
||||
@@ -2000,10 +2021,16 @@ namespace MIPSInt
|
||||
|
||||
for (int i = 0; i < n; i++) {
|
||||
switch (optype) {
|
||||
case 0: d[i] = s[i] + t[i]; break; //vadd
|
||||
case 1: d[i] = s[i] - t[i]; break; //vsub
|
||||
case 7: d[i] = s[i] / t[i]; break; //vdiv
|
||||
case 8: d[i] = s[i] * t[i]; break; //vmul
|
||||
case 0: d.f[i] = s[i] + t[i]; break; //vadd
|
||||
case 1: d.f[i] = s[i] - t[i]; break; //vsub
|
||||
case 7: d.f[i] = s[i] / t[i]; break; //vdiv
|
||||
case 8: d.f[i] = s[i] * t[i]; break; //vmul
|
||||
}
|
||||
|
||||
if (my_isnan(d.f[i])) {
|
||||
d.u[i] = 0x7f800001;
|
||||
} else if ((d.u[i] & 0x7F800000) == 0) {
|
||||
d.u[i] &= 0xFF800000;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -2012,12 +2039,12 @@ namespace MIPSInt
|
||||
u32 lastmask = (currentMIPS->vfpuCtrl[VFPU_CTRL_DPREFIX] & (1 << 8)) << (n - 1);
|
||||
u32 lastsat = (currentMIPS->vfpuCtrl[VFPU_CTRL_DPREFIX] & 3) << (n + n - 2);
|
||||
currentMIPS->vfpuCtrl[VFPU_CTRL_DPREFIX] = lastmask | lastsat;
|
||||
ApplyPrefixD(d, sz);
|
||||
ApplyPrefixD(d.f, sz);
|
||||
} else {
|
||||
RetainInvalidSwizzleST(d, sz);
|
||||
ApplyPrefixD(d, sz);
|
||||
RetainInvalidSwizzleST(d.f, sz);
|
||||
ApplyPrefixD(d.f, sz);
|
||||
}
|
||||
WriteVector(d, sz, vd);
|
||||
WriteVector(d.f, sz, vd);
|
||||
PC += 4;
|
||||
EatPrefixes();
|
||||
}
|
||||
|
||||
@@ -607,3 +607,87 @@ float Float16ToFloat32(unsigned short l)
|
||||
}
|
||||
return f;
|
||||
}
|
||||
|
||||
static uint32_t get_uexp(uint32_t x) {
|
||||
return (x >> 23) & 0xFF;
|
||||
}
|
||||
|
||||
static int32_t get_mant(uint32_t x) {
|
||||
// Note: this returns the hidden 1.
|
||||
return (x & 0x007FFFFF) | 0x00800000;
|
||||
}
|
||||
|
||||
static int32_t get_sign(uint32_t x) {
|
||||
return x & 0x80000000;
|
||||
}
|
||||
|
||||
float vfpu_dot(float a[4], float b[4]) {
|
||||
static const int EXTRA_BITS = 2;
|
||||
union float2int {
|
||||
uint32_t i;
|
||||
float f;
|
||||
} intermed[4];
|
||||
|
||||
for (int i = 0; i < 4; i++) {
|
||||
intermed[i].f = a[i] * b[i];
|
||||
}
|
||||
|
||||
uint32_t exps[4];
|
||||
int32_t mants[4];
|
||||
uint32_t max_exp = 0;
|
||||
|
||||
for (int i = 0; i < 4; i++) {
|
||||
exps[i] = get_uexp(intermed[i].i);
|
||||
// Preserve extra bits of precision in the mantissa during the add.
|
||||
mants[i] = get_mant(intermed[i].i) << EXTRA_BITS;
|
||||
if (exps[i] > max_exp) {
|
||||
max_exp = exps[i];
|
||||
}
|
||||
}
|
||||
|
||||
int32_t mant_sum = 0;
|
||||
for (int i = 0; i < 4; i++) {
|
||||
int exp = max_exp - exps[i];
|
||||
if (exp >= 32) {
|
||||
mants[i] = 0;
|
||||
} else {
|
||||
mants[i] >>= max_exp - exps[i];
|
||||
}
|
||||
if (get_sign(intermed[i].i)) {
|
||||
mants[i] = -mants[i];
|
||||
}
|
||||
mant_sum += mants[i];
|
||||
}
|
||||
|
||||
uint32_t sign_sum = 0;
|
||||
if (mant_sum < 0) {
|
||||
sign_sum = 0x80000000;
|
||||
mant_sum = -mant_sum;
|
||||
}
|
||||
// Chop off the extra bits.
|
||||
mant_sum >>= EXTRA_BITS;
|
||||
|
||||
if (mant_sum == 0 || max_exp == 0) {
|
||||
return 0.0f;
|
||||
}
|
||||
|
||||
while (mant_sum < 0x00800000) {
|
||||
mant_sum <<= 1;
|
||||
max_exp -= 1;
|
||||
}
|
||||
while (mant_sum >= 0x01000000) {
|
||||
mant_sum >>= 1;
|
||||
max_exp += 1;
|
||||
}
|
||||
_dbg_assert_(JIT, (mant_sum & 0x00800000) != 0);
|
||||
|
||||
if (max_exp >= 255) {
|
||||
max_exp = 255;
|
||||
mant_sum = 0;
|
||||
} else if (max_exp == 0) {
|
||||
return 0.0f;
|
||||
}
|
||||
|
||||
intermed[0].i = sign_sum | (max_exp << 23) | (mant_sum & 0x007FFFFF);
|
||||
return intermed[0].f;
|
||||
}
|
||||
@@ -97,6 +97,8 @@ inline float vfpu_clamp(float v, float min, float max) {
|
||||
return v >= max ? max : (v <= min ? min : v);
|
||||
}
|
||||
|
||||
float vfpu_dot(float a[4], float b[4]);
|
||||
|
||||
#define VFPU_FLOAT16_EXP_MAX 0x1f
|
||||
#define VFPU_SH_FLOAT16_SIGN 15
|
||||
#define VFPU_MASK_FLOAT16_SIGN 0x1
|
||||
|
||||
Reference in new issue
Block a user