tt-model push ornith-1.5-35b-a3b (container) (part 2)
Browse files
code/models/demos/blackhole/qwen36/tt/moe_decode/kernels/gu_compute.cpp
CHANGED
|
@@ -59,9 +59,9 @@ template <int ITERATIONS>
|
|
| 59 |
inline void round_sig7() {
|
| 60 |
for (int d = 0; d < ITERATIONS; d++) {
|
| 61 |
const sfpi::vFloat x = sfpi::dst_reg[0];
|
| 62 |
-
sfpi::vInt v = sfpi::
|
| 63 |
v = (v + 0x10000) & sfpi::vInt(0xFFFE0000);
|
| 64 |
-
sfpi::dst_reg[0] = sfpi::
|
| 65 |
sfpi::dst_reg++;
|
| 66 |
}
|
| 67 |
}
|
|
|
|
| 59 |
inline void round_sig7() {
|
| 60 |
for (int d = 0; d < ITERATIONS; d++) {
|
| 61 |
const sfpi::vFloat x = sfpi::dst_reg[0];
|
| 62 |
+
sfpi::vInt v = sfpi::as<sfpi::vInt>(x);
|
| 63 |
v = (v + 0x10000) & sfpi::vInt(0xFFFE0000);
|
| 64 |
+
sfpi::dst_reg[0] = sfpi::as<sfpi::vFloat>(v);
|
| 65 |
sfpi::dst_reg++;
|
| 66 |
}
|
| 67 |
}
|
code/models/demos/blackhole/qwen36/tt/moe_decode/kernels/gur_compute.cpp
CHANGED
|
@@ -49,9 +49,9 @@ template <int ITERATIONS>
|
|
| 49 |
inline void round_sig7() {
|
| 50 |
for (int d = 0; d < ITERATIONS; d++) {
|
| 51 |
const sfpi::vFloat x = sfpi::dst_reg[0];
|
| 52 |
-
sfpi::vInt v = sfpi::
|
| 53 |
v = (v + 0x10000) & sfpi::vInt(0xFFFE0000);
|
| 54 |
-
sfpi::dst_reg[0] = sfpi::
|
| 55 |
sfpi::dst_reg++;
|
| 56 |
}
|
| 57 |
}
|
|
|
|
| 49 |
inline void round_sig7() {
|
| 50 |
for (int d = 0; d < ITERATIONS; d++) {
|
| 51 |
const sfpi::vFloat x = sfpi::dst_reg[0];
|
| 52 |
+
sfpi::vInt v = sfpi::as<sfpi::vInt>(x);
|
| 53 |
v = (v + 0x10000) & sfpi::vInt(0xFFFE0000);
|
| 54 |
+
sfpi::dst_reg[0] = sfpi::as<sfpi::vFloat>(v);
|
| 55 |
sfpi::dst_reg++;
|
| 56 |
}
|
| 57 |
}
|