Skip to content

Commit b4a2825

Browse files
netrunnerevemglambda
netrunnereve
authored and
mglambda
committed
vulkan: scale caching for k quants + misc fixes (ggml-org#11081)
* q6_k scale caching * 16 bit unpack * q4_k test (slow) * revert it * q3_k * q2_k * little stuff * try precalculating products of a and q2_k scales * Revert "try precalculating products of a and q2_k scales" This reverts commit 65110b8. * unpack should be u16, add vim swap to gitignore (about time) * better q4_k scales * q5_k * better q6_k with separate paths for all threads and partial threads in use, plus some more optimizations * q2_k better dequant * q3_k optimizations * q3_k use hmask simd from cpu avx version * make the caches happy * q3_k separate out calculation * q2_k separate out * little stuff * use calc_superblock everywhere * q2_k optimize scale calculation * more barriers
1 parent 9a52053 commit b4a2825

File tree

6 files changed

+455
-385
lines changed

6 files changed

+455
-385
lines changed

‎.gitignore

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,7 @@
1818
*.metallib
1919
*.o
2020
*.so
21+
*.swp
2122
*.tmp
2223

2324
# IDE / OS

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

Lines changed: 84 additions & 70 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,80 @@
55

66
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
77

8+
shared FLOAT_TYPE sccache1[BLOCK_SIZE/16][16];
9+
shared FLOAT_TYPE sccache2[BLOCK_SIZE/16][16];
10+
11+
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
12+
13+
void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
14+
const uint y_idx = i * QUANT_K + y_offset;
15+
16+
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
17+
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
18+
19+
barrier();
20+
if (!all_threads) { // when we don't have enough blocks to use all threads
21+
if (i < num_blocks_per_row) {
22+
const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]);
23+
sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF);
24+
sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
25+
}
26+
barrier();
27+
28+
if (i >= num_blocks_per_row)
29+
continue;
30+
} else {
31+
const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]);
32+
sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF);
33+
sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
34+
barrier();
35+
}
36+
37+
const uint32_t qs_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);
38+
const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303));
39+
const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303));
40+
const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
41+
const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
42+
43+
vec2 d = vec2(data_a[ib0 + i].d);
44+
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
45+
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
46+
47+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
48+
vec2 b0 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0]);
49+
vec2 b16 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8]);
50+
vec2 b32 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
51+
vec2 b48 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
52+
vec2 b64 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
53+
vec2 b80 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
54+
vec2 b96 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
55+
vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
56+
57+
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
58+
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
59+
[[unroll]] for (int l = 0; l < 2; ++l) {
60+
sum1 = fma(FLOAT_TYPE(b0[l]), sccache1[ix][ 8*v_im] * qs_u32_0[l ],
61+
fma(FLOAT_TYPE(b16[l]), sccache1[ix][1 + 8*v_im] * qs_u32_0[l+2],
62+
fma(FLOAT_TYPE(b32[l]), sccache1[ix][2 + 8*v_im] * qs_u32_2[l ],
63+
fma(FLOAT_TYPE(b48[l]), sccache1[ix][3 + 8*v_im] * qs_u32_2[l+2],
64+
fma(FLOAT_TYPE(b64[l]), sccache1[ix][4 + 8*v_im] * qs_u32_4[l ],
65+
fma(FLOAT_TYPE(b80[l]), sccache1[ix][5 + 8*v_im] * qs_u32_4[l+2],
66+
fma(FLOAT_TYPE(b96[l]), sccache1[ix][6 + 8*v_im] * qs_u32_6[l ],
67+
fma(FLOAT_TYPE(b112[l]), sccache1[ix][7 + 8*v_im] * qs_u32_6[l+2], sum1))))))));
68+
sum2 = fma(FLOAT_TYPE(b0[l]), sccache2[ix][ 8*v_im],
69+
fma(FLOAT_TYPE(b16[l]), sccache2[ix][1 + 8*v_im],
70+
fma(FLOAT_TYPE(b32[l]), sccache2[ix][2 + 8*v_im],
71+
fma(FLOAT_TYPE(b48[l]), sccache2[ix][3 + 8*v_im],
72+
fma(FLOAT_TYPE(b64[l]), sccache2[ix][4 + 8*v_im],
73+
fma(FLOAT_TYPE(b80[l]), sccache2[ix][5 + 8*v_im],
74+
fma(FLOAT_TYPE(b96[l]), sccache2[ix][6 + 8*v_im],
75+
fma(FLOAT_TYPE(b112[l]), sccache2[ix][7 + 8*v_im], sum2))))))));
76+
}
77+
temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
78+
}
79+
}
80+
}
81+
882
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
983
uint a_offset, b_offset, d_offset;
1084
get_offsets(a_offset, b_offset, d_offset);
@@ -14,88 +88,28 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
1488
// 16 threads are used to process each block
1589
const uint it_size = gl_WorkGroupSize.x/16;
1690
const uint tid = gl_LocalInvocationID.x;
17-
const uint itid = tid%16; // 0...16
18-
const uint ix = tid/16;
19-
20-
const uint step = 8;
91+
const uint itid = tid%16; // 0...15
92+
const uint ix = tid/16;
2193

22-
const uint v_im = itid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
23-
const uint v_in = itid - step*v_im; // 0...15 or 0...7
94+
const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128...
95+
const uint v_in = itid - 8*v_im; // 0...7
2496

2597
const uint l0 = 2*v_in; // 0...15
2698
const uint q_offset = 32*v_im + l0;
27-
const uint s_offset = 8*v_im;
2899
const uint y_offset = 128*v_im + l0;
29100

30-
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
31-
32101
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
33102
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
34103
temp[j][i] = FLOAT_TYPE(0);
35104
}
36105
}
37106

38-
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
39-
const uint y_idx = i * QUANT_K + y_offset;
40-
41-
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
42-
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
43-
vec2 d = vec2(data_a[ib0 + i].d);
44-
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
45-
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
46-
47-
uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
48-
uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
49-
50-
uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
51-
uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
52-
uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
53-
uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
54-
55-
uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
56-
uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
57-
uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
58-
uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
59-
60-
uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
61-
uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
62-
uvec2 qs0 = uvec2(unpack8(qs0_u16));
63-
uvec2 qs16 = uvec2(unpack8(qs16_u16));
64-
65-
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
66-
vec2 b0 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0]);
67-
vec2 b16 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8]);
68-
vec2 b32 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
69-
vec2 b48 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
70-
vec2 b64 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
71-
vec2 b80 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
72-
vec2 b96 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
73-
vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
74-
75-
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
76-
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
77-
[[unroll]] for (int l = 0; l < 2; ++l) {
78-
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
79-
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
80-
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
81-
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
82-
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
83-
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
84-
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
85-
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
86-
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
87-
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
88-
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
89-
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
90-
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
91-
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
92-
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
93-
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
94-
}
95-
temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
96-
}
97-
}
98-
}
107+
const uint nbr_par_th = num_blocks_per_row%it_size;
108+
const uint nbr_all_th = num_blocks_per_row - nbr_par_th;
109+
uint i0 = 0;
110+
[[unroll]] for (; i0 < nbr_all_th; i0 += it_size)
111+
calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true);
112+
calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false);
99113

100114
reduce_result(temp, d_offset, first_row, num_rows, tid);
101115
}

0 commit comments

Comments
 (0)