aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorStephan Walter <stephan@walter.name>2023-04-13 14:59:50 +0000
committerGitHub <noreply@github.com>2023-04-13 17:59:50 +0300
commit6232f2d7fd7a22d5eeb62182b2f21fcf01359754 (patch)
tree431ada6d84ca8db32acee8a54ec0a888899489e2
parent6c248707f51c8a50f7792e7f7787ec481881db88 (diff)
ggml : optimize non-SIMD Q4_0 vector dot product (#703)
-rw-r--r--ggml.c12
1 files changed, 7 insertions, 5 deletions
diff --git a/ggml.c b/ggml.c
index 281fd8e..eb47d82 100644
--- a/ggml.c
+++ b/ggml.c
@@ -2160,18 +2160,20 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
const uint8_t * restrict p0 = x[i].qs;
const uint8_t * restrict p1 = y[i].qs;
+ int sumi = 0;
for (int j = 0; j < QK/2; j++) {
const uint8_t v0 = p0[j];
const uint8_t v1 = p1[j];
- const float f0 = d0*((int8_t) (v0 & 0xf) - 8);
- const float f1 = d0*((int8_t) (v0 >> 4) - 8);
+ const int8_t i0 = (int8_t) (v0 & 0xf) - 8;
+ const int8_t i1 = (int8_t) (v0 >> 4) - 8;
- const float f2 = d1*((int8_t) (v1 & 0xf) - 8);
- const float f3 = d1*((int8_t) (v1 >> 4) - 8);
+ const int8_t i2 = (int8_t) (v1 & 0xf) - 8;
+ const int8_t i3 = (int8_t) (v1 >> 4) - 8;
- sumf += f0*f2 + f1*f3;
+ sumi += i0*i2 + i1*i3;
}
+ sumf += d0 * d1 * sumi;
}
#endif