Skip to content

Commit 745d401

Browse files
netrunnerevenetrunnereve
committed
oops
1 parent 977d7b9 commit 745d401

File tree

1 file changed

+35
-35
lines changed

1 file changed

+35
-35
lines changed

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp‎

Lines changed: 35 additions & 35 deletions
Original file line numberDiff line numberDiff line change
@@ -44,44 +44,44 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
4444
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
4545
const uint y_idx = i * QUANT_K + y_offset;
4646

47-
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
48-
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
49-
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
50-
B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
51-
B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
52-
B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
53-
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
54-
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
47+
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
48+
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
49+
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
50+
B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
51+
B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
52+
B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
53+
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
54+
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
5555

5656
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
5757
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
58-
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
59-
60-
uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
61-
uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
62-
uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
63-
uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
64-
uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
65-
uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
66-
u8vec2 s0 = unpack8(s0_16);
67-
u8vec2 s2 = unpack8(s2_16);
68-
u8vec2 s4 = unpack8(s4_16);
69-
u8vec2 s6 = unpack8(s6_16);
70-
u8vec2 s8 = unpack8(s8_16);
71-
u8vec2 s10 = unpack8(s10_16);
72-
73-
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
74-
[[unroll]] for (int l = 0; l < 2; ++l) {
75-
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
76-
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
77-
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
78-
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
79-
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
80-
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
81-
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
82-
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
83-
}
84-
temp[n] = fma(d, sum, temp[n]);
58+
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
59+
60+
uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
61+
uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
62+
uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
63+
uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
64+
uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
65+
uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
66+
u8vec2 s0 = unpack8(s0_16);
67+
u8vec2 s2 = unpack8(s2_16);
68+
u8vec2 s4 = unpack8(s4_16);
69+
u8vec2 s6 = unpack8(s6_16);
70+
u8vec2 s8 = unpack8(s8_16);
71+
u8vec2 s10 = unpack8(s10_16);
72+
73+
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
74+
[[unroll]] for (int l = 0; l < 2; ++l) {
75+
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
76+
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
77+
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
78+
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
79+
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
80+
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
81+
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
82+
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
83+
}
84+
temp[n] = fma(d, sum, temp[n]);
8585
}
8686
}
8787

0 commit comments

Comments
 (0)