Skip to content

Commit 977d7b9

Browse files
netrunnerevenetrunnereve
committed
i might be too tired
1 parent beb4f8e commit 977d7b9

File tree

2 files changed

+47
-47
lines changed

2 files changed

+47
-47
lines changed

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

Lines changed: 36 additions & 36 deletions
Original file line numberDiff line numberDiff line change
@@ -44,45 +44,45 @@ 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]);
85-
}
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]);
85+
}
8686
}
8787

8888
// sum up partial sums and write back result

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp‎

Lines changed: 11 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -51,20 +51,20 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
5151
B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8];
5252
B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
5353
B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
54-
54+
5555
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
5656
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
5757
f16vec2 d = data_a[ib0 + i].d;
5858
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
5959
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
60-
60+
6161
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ];
6262
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
6363
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
6464
uvec4 scale0 = uvec4(unpack8(scale0_u32));
6565
uvec4 scale4 = uvec4(unpack8(scale4_u32));
6666
uvec4 scale8 = uvec4(unpack8(scale8_u32));
67-
67+
6868
const uint32_t sc0 = ( scale0.x & 0x3f);
6969
const uint32_t sc1 = ( scale0.y & 0x3f);
7070
const uint32_t sc2 = ( scale4.x & 0x3f);
@@ -73,32 +73,32 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
7373
const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2));
7474
const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
7575
const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
76-
76+
7777
uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
7878
uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
79-
79+
8080
uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
8181
uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
8282
uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
8383
uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
84-
84+
8585
uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
86-
86+
8787
uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
8888
uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
8989
uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
9090
uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
91-
91+
9292
qs0_16_u32_lo4 += qs0_16_lo4_offset16;
9393
qs0_16_u32_hi4 += qs0_16_hi4_offset16;
9494
qs64_80_u32_lo4 += qs64_80_lo4_offset16;
9595
qs64_80_u32_hi4 += qs64_80_hi4_offset16;
96-
96+
9797
uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
9898
uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
9999
uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
100100
uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
101-
101+
102102
const uint32_t q4_0 = qs0_16_lo4.x;
103103
const uint32_t q4_1 = qs0_16_lo4.y;
104104
const uint32_t q4_2 = qs0_16_lo4.z;
@@ -115,7 +115,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
115115
const uint32_t q4_13 = qs64_80_hi4.y;
116116
const uint32_t q4_14 = qs64_80_hi4.z;
117117
const uint32_t q4_15 = qs64_80_hi4.w;
118-
118+
119119
const FLOAT_TYPE sx =
120120
fma(FLOAT_TYPE(by10.x), q4_0,
121121
fma(FLOAT_TYPE(by10.y), q4_1,

0 commit comments

Comments
 (0)