Skip to content

Commit a0b2015

Browse files
remyoudomphengremyoudompheng
committed
vulkan: initial support for IQ2_XXS
1 parent c03579d commit a0b2015

16 files changed

+276
-63
lines changed

‎ggml/src/ggml-vulkan/ggml-vulkan.cpp

Lines changed: 27 additions & 22 deletions
Large diffs are not rendered by default.

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

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -12,8 +12,8 @@ layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
1212
#endif
1313

1414
void main() {
15-
#if defined(DATA_A_IQ4_NL)
16-
init_iq4nl_shmem();
15+
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
16+
init_iq_shmem();
1717
if (gl_LocalInvocationIndex.x != 0) {
1818
return;
1919
}

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

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -217,8 +217,8 @@ void quantize(uint dst_idx, uint src_idx)
217217
#endif
218218

219219
void main() {
220-
#if defined(DATA_A_IQ4_NL)
221-
init_iq4nl_shmem();
220+
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
221+
init_iq_shmem();
222222
if (gl_LocalInvocationIndex.x != 0) {
223223
return;
224224
}

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

Lines changed: 48 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -88,6 +88,53 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
8888
}
8989
#endif
9090

91+
#if defined(DATA_A_IQ2_XXS)
92+
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
93+
const uint ib32 = iqs / 32;
94+
const uint ib8 = (iqs / 8) % 4;
95+
const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8];
96+
// Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
97+
const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2],
98+
data_a_packed16[a_offset + ib].qs[4 * ib32 + 3]));
99+
const float db = 0.25 * (0.5 + (signs >> 28));
100+
const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
101+
// Add parity bit
102+
const uint sign8 = sign7 | (bitCount(sign7) << 7);
103+
const uint sign = sign8 >> (iqs % 8);
104+
const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4)));
105+
bool sign0 = (sign & 1) != 0;
106+
bool sign1 = (sign & 2) != 0;
107+
return db * vec2(
108+
grid.x * (sign0 ? -1.0 : 1.0),
109+
grid.y * (sign1 ? -1.0 : 1.0)
110+
);
111+
}
112+
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
113+
const uint ib32 = iqs / 32;
114+
const uint ib8 = (iqs / 8) % 4;
115+
const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8];
116+
// Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
117+
const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2],
118+
data_a_packed16[a_offset + ib].qs[4 * ib32 + 3]));
119+
const float db = 0.25 * (0.5 + (signs >> 28));
120+
const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
121+
// Add parity bit
122+
const uint sign8 = sign7 | (bitCount(sign7) << 7);
123+
const uint sign = sign8 >> (iqs % 8);
124+
const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4)));
125+
bool sign0 = (sign & 1) != 0;
126+
bool sign1 = (sign & 2) != 0;
127+
bool sign2 = (sign & 4) != 0;
128+
bool sign3 = (sign & 8) != 0;
129+
return db * vec4(
130+
grid.x * (sign0 ? -1.0 : 1.0),
131+
grid.y * (sign1 ? -1.0 : 1.0),
132+
grid.z * (sign2 ? -1.0 : 1.0),
133+
grid.w * (sign3 ? -1.0 : 1.0)
134+
);
135+
}
136+
#endif
137+
91138
#if defined(DATA_A_IQ3_XXS)
92139
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
93140
const uint ib4 = iqs / 4;
@@ -190,7 +237,7 @@ vec2 get_dm(uint ib, uint a_offset) {
190237
}
191238
#endif
192239

193-
#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
240+
#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
194241
vec2 get_dm(uint ib, uint a_offset) {
195242
return vec2(float(data_a[a_offset + ib].d), 0);
196243
}

‎ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp

Lines changed: 48 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,48 @@
1+
#version 450
2+
3+
#include "dequant_head.comp"
4+
5+
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
6+
7+
layout (binding = 0) readonly buffer A {block_iq2_xxs data_a[];};
8+
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
9+
10+
void main() {
11+
// Each thread handles 1 scale block (32 values)
12+
// Each block is described by 4 lattice indices, 4x7 sign bits and 4 scale bits
13+
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
14+
15+
init_iq_shmem();
16+
17+
if (ib >= p.nel / 256) {
18+
return;
19+
}
20+
21+
const uint is = gl_LocalInvocationID.x % 8;
22+
const uint b_idx = 256 * ib + 32 * is;
23+
24+
const float d = float(data_a[ib].d);
25+
uint signscale = pack32(u8vec4(
26+
data_a[ib].qs[8*is + 4],
27+
data_a[ib].qs[8*is + 5],
28+
data_a[ib].qs[8*is + 6],
29+
data_a[ib].qs[8*is + 7]
30+
));
31+
const float db = d * (0.5 + (signscale >> 28)) * 0.25;
32+
33+
[[unroll]] for (uint l = 0; l < 4; ++l) {
34+
const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
35+
const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit
36+
const uvec2 grid = iq2xxs_grid[data_a[ib].qs[8 * is + l]];
37+
const u8vec4 grid0 = unpack8(grid.x);
38+
const u8vec4 grid1 = unpack8(grid.y);
39+
data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));
40+
data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0));
41+
data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0));
42+
data_b[b_idx + 8 * l + 3] = D_TYPE(db * grid0.w * ((sign8 & 8) != 0 ? -1.0 : 1.0));
43+
data_b[b_idx + 8 * l + 4] = D_TYPE(db * grid1.x * ((sign8 & 16) != 0 ? -1.0 : 1.0));
44+
data_b[b_idx + 8 * l + 5] = D_TYPE(db * grid1.y * ((sign8 & 32) != 0 ? -1.0 : 1.0));
45+
data_b[b_idx + 8 * l + 6] = D_TYPE(db * grid1.z * ((sign8 & 64) != 0 ? -1.0 : 1.0));
46+
data_b[b_idx + 8 * l + 7] = D_TYPE(db * grid1.w * ((sign8 & 128) != 0 ? -1.0 : 1.0));
47+
}
48+
}

‎ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,7 @@ void main() {
1212
// Each block contains 4 scale bytes (8 scales) for 256 output values.
1313
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
1414

15-
init_iq3s_shmem();
15+
init_iq_shmem();
1616

1717
if (ib >= p.nel / 256) {
1818
return;

‎ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,7 @@ void main() {
1212
// 8 threads handle 1 superblock
1313
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
1414

15-
init_iq3xxs_shmem();
15+
init_iq_shmem();
1616

1717
if (ib >= p.nel / 256) {
1818
return;

‎ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
1010
void main() {
1111
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
1212

13-
init_iq4nl_shmem();
13+
init_iq_shmem();
1414

1515
const uint tid = gl_LocalInvocationID.x % 64;
1616
const uint il = tid/32;

‎ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

Lines changed: 2 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -104,10 +104,8 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
104104
#endif
105105

106106
void main() {
107-
#if defined(DATA_A_IQ3_S)
108-
init_iq3s_shmem();
109-
#elif defined(DATA_A_IQ4_NL)
110-
init_iq4nl_shmem();
107+
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
108+
init_iq_shmem();
111109
#endif
112110

113111
const uint32_t N = p.N;

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

Lines changed: 2 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -12,12 +12,8 @@ void main() {
1212
const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
1313
const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
1414

15-
#if defined(DATA_A_IQ3_XXS)
16-
init_iq3xxs_shmem();
17-
#elif defined(DATA_A_IQ3_S)
18-
init_iq3s_shmem();
19-
#elif defined(DATA_A_IQ4_NL)
20-
init_iq4nl_shmem();
15+
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
16+
init_iq_shmem();
2117
#endif
2218

2319
if (i00 >= p.ne00) {

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

Lines changed: 2 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -133,12 +133,8 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
133133
void main() {
134134
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
135135

136-
#if defined(DATA_A_IQ3_XXS)
137-
init_iq3xxs_shmem();
138-
#elif defined(DATA_A_IQ3_S)
139-
init_iq3s_shmem();
140-
#elif defined(DATA_A_IQ4_NL)
141-
init_iq4nl_shmem();
136+
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
137+
init_iq_shmem();
142138
#endif
143139

144140
// do NUM_ROWS at a time, unless there aren't enough remaining rows

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

Lines changed: 27 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -95,12 +95,8 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
9595
#endif
9696

9797
void main() {
98-
#if defined(DATA_A_IQ3_XXS)
99-
init_iq3xxs_shmem();
100-
#elif defined(DATA_A_IQ3_S)
101-
init_iq3s_shmem();
102-
#elif defined(DATA_A_IQ4_NL)
103-
init_iq4nl_shmem();
98+
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
99+
init_iq_shmem();
104100
#endif
105101

106102
#ifdef MUL_MAT_ID
@@ -443,6 +439,31 @@ void main() {
443439

444440
buf_a[buf_idx ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi ] >> qhshift) & 3) << 4)) - 32));
445441
buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
442+
#elif defined(DATA_A_IQ2_XXS)
443+
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
444+
const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
445+
446+
const uint ib = idx / 128; // 2 values per idx
447+
const uint ib32 = (idx % 128) / 16; // 0..7
448+
const uint ib8 = (idx / 4) % 4;
449+
450+
const float d = float(data_a[ib].d);
451+
const uint qs = data_a[ib].qs[8 * ib32 + ib8];
452+
const uint signs = pack32(u8vec4(
453+
data_a[ib].qs[8*ib32 + 4],
454+
data_a[ib].qs[8*ib32 + 5],
455+
data_a[ib].qs[8*ib32 + 6],
456+
data_a[ib].qs[8*ib32 + 7]
457+
));
458+
const float db = d * 0.25 * (0.5 + (signs >> 28));
459+
const uint32_t sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
460+
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
461+
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
462+
const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1));
463+
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
464+
465+
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
466+
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
446467
#elif defined(DATA_A_IQ3_XXS)
447468
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
448469
const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

Lines changed: 2 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -106,12 +106,8 @@ D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem
106106
#endif
107107

108108
void main() {
109-
#if defined(DATA_A_IQ3_XXS)
110-
init_iq3xxs_shmem();
111-
#elif defined(DATA_A_IQ3_S)
112-
init_iq3s_shmem();
113-
#elif defined(DATA_A_IQ4_NL)
114-
init_iq4nl_shmem();
109+
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
110+
init_iq_shmem();
115111
#endif
116112

117113
#ifdef MUL_MAT_ID

0 commit comments

Comments
 (0)