Skip to content

Commit 4a57d36

Browse files
jeffbolznvjeffbolznv
authored
vulkan: optimize Q2_K and Q3_K mul_mat_vec (#10459)
1 parent c9b00a7 commit 4a57d36

File tree

2 files changed

+87
-33
lines changed

2 files changed

+87
-33
lines changed

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

+52-22
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,5 @@
11
#version 450
2+
#extension GL_EXT_shader_explicit_arithmetic_types : require
23

34
#include "mul_mat_vec_base.comp"
45

@@ -32,38 +33,67 @@ void main() {
3233
const uint s_offset = 8*v_im;
3334
const uint y_offset = 128*v_im + l0;
3435

35-
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
36+
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
3637

3738
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
3839
const uint y_idx = i * QUANT_K + y_offset;
3940

40-
const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
41-
const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
41+
f16vec2 d = data_a[ib0 + i].d;
42+
const FLOAT_TYPE dall = d.x;
43+
const FLOAT_TYPE dmin = d.y;
44+
45+
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
46+
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
47+
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
48+
B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
49+
B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
50+
B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
51+
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
52+
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
53+
54+
uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
55+
uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
56+
57+
uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
58+
uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
59+
uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
60+
uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
61+
62+
uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
63+
uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
64+
uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
65+
uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
66+
67+
uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
68+
uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
69+
uvec2 qs0 = uvec2(unpack8(qs0_u16));
70+
uvec2 qs16 = uvec2(unpack8(qs16_u16));
4271

4372
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
4473
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
45-
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
46-
sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3),
47-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3),
48-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3),
49-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3),
50-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3),
51-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3),
52-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3),
53-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1))))))));
54-
sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF),
55-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF),
56-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF),
57-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF),
58-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF),
59-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF),
60-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF),
61-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2))))))));
74+
[[unroll]] for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
75+
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
76+
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
77+
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
78+
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
79+
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
80+
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
81+
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
82+
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
83+
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
84+
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
85+
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
86+
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
87+
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
88+
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
89+
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
90+
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
6291
}
63-
const uint tmp_idx = 16 * ix + tid;
64-
tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx]));
92+
temp = fma(dall, sum1, fma(-dmin, sum2, temp));
6593
}
6694

95+
tmp[gl_LocalInvocationID.x] = temp;
96+
6797
// sum up partial sums and write back result
6898
barrier();
6999
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {

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

+35-11
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,5 @@
11
#version 450
2+
#extension GL_EXT_shader_explicit_arithmetic_types : require
23

34
#include "mul_mat_vec_base.comp"
45

@@ -33,7 +34,7 @@ void main() {
3334
const uint q_offset = 32*v_im + l0;
3435
const uint y_offset = 128*v_im + l0;
3536

36-
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
37+
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
3738

3839
const uint s_shift = 4 * v_im;
3940

@@ -42,21 +43,44 @@ void main() {
4243

4344
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
4445

46+
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
47+
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
48+
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
49+
B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
50+
B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
51+
B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
52+
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
53+
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
54+
55+
uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
56+
uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
57+
uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
58+
uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
59+
uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
60+
uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
61+
u8vec2 s0 = unpack8(s0_16);
62+
u8vec2 s2 = unpack8(s2_16);
63+
u8vec2 s4 = unpack8(s4_16);
64+
u8vec2 s6 = unpack8(s6_16);
65+
u8vec2 s8 = unpack8(s8_16);
66+
u8vec2 s10 = unpack8(s10_16);
67+
4568
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
4669
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
47-
sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (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)),
48-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (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)),
49-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (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)),
50-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (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)),
51-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (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)),
52-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (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)),
53-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (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)),
54-
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (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))))))));
70+
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)),
71+
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)),
72+
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)),
73+
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)),
74+
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)),
75+
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)),
76+
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)),
77+
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))))))));
5578
}
56-
const uint tmp_idx = 16 * ix + tid;
57-
tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]);
79+
temp = fma(d, sum, temp);
5880
}
5981

82+
tmp[gl_LocalInvocationID.x] = temp;
83+
6084
// sum up partial sums and write back result
6185
barrier();
6286
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {

0 commit comments

Comments
 (0)