Skip to content

Commit 6def73b

Browse files
0cc4m0cc4m
committed
vulkan: do 8 calculations per invocation instead of 32 in mul_mat_vecq, similar to mul_mat_vec
1 parent c19ec8f commit 6def73b

File tree

6 files changed

+67
-134
lines changed

6 files changed

+67
-134
lines changed

‎ggml/src/ggml-vulkan/ggml-vulkan.cpp‎

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -366,7 +366,7 @@ struct vk_device_struct {
366366
bool uma;
367367
bool prefer_host_memory;
368368
bool float_controls_rte_fp16;
369-
bool subgroup_add;
369+
bool subgroup_arithmetic;
370370
bool subgroup_shuffle;
371371
bool subgroup_clustered;
372372

@@ -2797,7 +2797,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
27972797

27982798
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
27992799
if (device->integer_dot_product) {
2800-
if (device->subgroup_clustered && device->subgroup_require_full_support && device->vendor_id != VK_VENDOR_ID_INTEL) {
2800+
if (device->subgroup_arithmetic && device->subgroup_require_full_support && device->vendor_id != VK_VENDOR_ID_INTEL) {
28012801
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_q8_1_f32_subgroup_len, mul_mat_vec_q4_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
28022802
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_q8_1_f32_subgroup_len, mul_mat_vec_q4_1_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
28032803
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_q8_1_f32_subgroup_len, mul_mat_vec_q5_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
@@ -2908,7 +2908,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
29082908
}
29092909

29102910
for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
2911-
if (device->subgroup_add && device->subgroup_require_full_support) {
2911+
if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
29122912
ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true);
29132913
} else {
29142914
ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
@@ -3381,8 +3381,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
33813381
}
33823382
device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;
33833383

3384-
device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
3385-
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
3384+
device->subgroup_arithmetic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
3385+
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
33863386
device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
33873387
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
33883388
device->subgroup_clustered = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&

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

Lines changed: 17 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -97,7 +97,23 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
9797
layout (constant_id = 1) const uint NUM_ROWS = 1;
9898
layout (constant_id = 2) const uint NUM_COLS = 1;
9999

100-
#if !defined(MMQ) || !defined(USE_SUBGROUPS)
100+
#ifdef USE_SUBGROUPS
101+
void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
102+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
103+
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
104+
temp[j][n] = subgroupAdd(temp[j][n]);
105+
}
106+
}
107+
108+
if (tid == 0) {
109+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
110+
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
111+
data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
112+
}
113+
}
114+
}
115+
}
116+
#else
101117
shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
102118

103119
void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {

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

Lines changed: 33 additions & 118 deletions
Original file line numberDiff line numberDiff line change
@@ -5,95 +5,47 @@
55

66
#ifdef USE_SUBGROUPS
77
#extension GL_KHR_shader_subgroup_basic : require
8-
#extension GL_KHR_shader_subgroup_clustered : require
8+
#extension GL_KHR_shader_subgroup_arithmetic : require
99

1010
#define INVOCATION_ID gl_SubgroupInvocationID.x
1111
#else
1212
#define INVOCATION_ID gl_LocalInvocationID.x
1313
#endif
1414

1515
#define MMQ
16-
#define B_TYPE block_q8_1_x4_packed128
16+
#define B_TYPE block_q8_1_x4
1717

1818
#include "mul_mat_vec_base.comp"
1919

2020
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
2121

22-
#define K_PER_ITER 32
23-
24-
const uint GROUP_SIZE = 8;
25-
const uint GROUPS_PER_WARP = (BLOCK_SIZE / GROUP_SIZE);
22+
#define K_PER_ITER 8
2623

2724
#include "mul_mmq_funcs.comp"
2825

29-
uint a_offset, b_offset, d_offset, y_offset;
26+
uint a_offset, b_offset, d_offset;
3027

31-
#ifdef USE_SUBGROUPS
32-
void reduce_result_grouped(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid_in_group) {
33-
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
34-
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
35-
temp[j][n] = subgroupClusteredAdd(temp[j][n], GROUP_SIZE);
36-
}
37-
}
38-
39-
if (tid_in_group == 0) {
40-
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
41-
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
42-
data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
43-
}
44-
}
45-
}
46-
}
47-
#else
48-
void reduce_result_grouped(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid_in_group) {
49-
const uint tid = INVOCATION_ID;
50-
// sum up partial sums and write back result
51-
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
52-
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
53-
tmpsh[j][n][tid] = temp[j][n];
54-
}
55-
}
56-
barrier();
57-
if (tid_in_group < 4) {
58-
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
59-
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
60-
tmpsh[j][n][tid] += tmpsh[j][n][tid + 4];
61-
}
62-
}
63-
}
64-
barrier();
65-
if (tid_in_group < 2) {
66-
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
67-
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
68-
tmpsh[j][n][tid] += tmpsh[j][n][tid + 2];
69-
}
70-
}
71-
}
72-
barrier();
73-
if (tid_in_group == 0) {
74-
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
75-
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
76-
data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][tid] + tmpsh[j][n][tid + 1]);
77-
}
78-
}
79-
}
80-
}
81-
#endif
82-
83-
ivec4 cache_b_qs[2];
28+
int32_t cache_b_qs[2];
8429
vec2 cache_b_ds;
8530

86-
void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid_in_group, const uint i) {
31+
void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i) {
8732
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
88-
const uint col = i*GROUP_SIZE + K_PER_ITER*tid_in_group;
33+
const uint col = i*BLOCK_SIZE + tid*K_PER_ITER;
8934

9035
// Preload data_b block
9136
const uint b_block_idx = (j*p.batch_stride_b + col) / QUANT_K_Q8_1 + b_offset;
37+
const uint b_qs_idx = tid % 4;
9238
const uint b_block_idx_outer = b_block_idx / 4;
9339
const uint b_block_idx_inner = b_block_idx % 4;
9440
cache_b_ds = vec2(data_b[b_block_idx_outer].ds[b_block_idx_inner]);
95-
cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 2];
96-
cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 2 + 1];
41+
42+
#if QUANT_R == 2
43+
cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx];
44+
cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx + 4];
45+
#else
46+
cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2];
47+
cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2 + 1];
48+
#endif
9749

9850
uint ibi = first_row*p.ncols;
9951
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -102,71 +54,36 @@ void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const
10254

10355
int32_t q_sum = 0;
10456
#if QUANT_R == 2
105-
i32vec2 data_a_qs = repack(a_block_idx, 0);
106-
q_sum += dotPacked4x8EXT(data_a_qs.x,
107-
cache_b_qs[0].x);
108-
q_sum += dotPacked4x8EXT(data_a_qs.y,
109-
cache_b_qs[1].x);
110-
data_a_qs = repack(a_block_idx, 1);
111-
q_sum += dotPacked4x8EXT(data_a_qs.x,
112-
cache_b_qs[0].y);
113-
q_sum += dotPacked4x8EXT(data_a_qs.y,
114-
cache_b_qs[1].y);
115-
data_a_qs = repack(a_block_idx, 2);
57+
const i32vec2 data_a_qs = repack(a_block_idx, b_qs_idx);
11658
q_sum += dotPacked4x8EXT(data_a_qs.x,
117-
cache_b_qs[0].z);
59+
cache_b_qs[0]);
11860
q_sum += dotPacked4x8EXT(data_a_qs.y,
119-
cache_b_qs[1].z);
120-
data_a_qs = repack(a_block_idx, 3);
121-
q_sum += dotPacked4x8EXT(data_a_qs.x,
122-
cache_b_qs[0].w);
123-
q_sum += dotPacked4x8EXT(data_a_qs.y,
124-
cache_b_qs[1].w);
61+
cache_b_qs[1]);
12562
#else
126-
int32_t data_a_qs = repack(a_block_idx, 0);
127-
q_sum += dotPacked4x8EXT(data_a_qs,
128-
cache_b_qs[0].x);
129-
data_a_qs = repack(a_block_idx, 1);
130-
q_sum += dotPacked4x8EXT(data_a_qs,
131-
cache_b_qs[0].y);
132-
data_a_qs = repack(a_block_idx, 2);
133-
q_sum += dotPacked4x8EXT(data_a_qs,
134-
cache_b_qs[0].z);
135-
data_a_qs = repack(a_block_idx, 3);
63+
int32_t data_a_qs = repack(a_block_idx, b_qs_idx * 2);
13664
q_sum += dotPacked4x8EXT(data_a_qs,
137-
cache_b_qs[0].w);
138-
data_a_qs = repack(a_block_idx, 4);
65+
cache_b_qs[0]);
66+
data_a_qs = repack(a_block_idx, b_qs_idx * 2 + 1);
13967
q_sum += dotPacked4x8EXT(data_a_qs,
140-
cache_b_qs[1].x);
141-
data_a_qs = repack(a_block_idx, 5);
142-
q_sum += dotPacked4x8EXT(data_a_qs,
143-
cache_b_qs[1].y);
144-
data_a_qs = repack(a_block_idx, 6);
145-
q_sum += dotPacked4x8EXT(data_a_qs,
146-
cache_b_qs[1].z);
147-
data_a_qs = repack(a_block_idx, 7);
148-
q_sum += dotPacked4x8EXT(data_a_qs,
149-
cache_b_qs[1].w);
68+
cache_b_qs[1]);
15069
#endif
15170

15271
#if QUANT_AUXF == 1
153-
temp[j][n] += mul_q8_1(q_sum, get_d(a_block_idx), cache_b_ds);
72+
temp[j][n] += mul_q8_1(q_sum, get_d(a_block_idx), cache_b_ds, 4);
15473
#else
155-
temp[j][n] += mul_q8_1(q_sum, get_dm(a_block_idx), cache_b_ds);
74+
temp[j][n] += mul_q8_1(q_sum, get_dm(a_block_idx), cache_b_ds, 4);
15675
#endif
15776
}
15877
}
15978
}
16079

16180
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
162-
const uint tid_in_group = INVOCATION_ID % GROUP_SIZE;
81+
const uint tid = INVOCATION_ID;
16382

16483
get_offsets(a_offset, b_offset, d_offset);
16584
a_offset /= QUANT_K;
16685
b_offset /= QUANT_K_Q8_1;
16786

168-
y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
169-
17087
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
17188

17289
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
@@ -175,8 +92,8 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
17592
}
17693
}
17794

178-
uint num_iters = p.ncols / (K_PER_ITER * GROUP_SIZE);
179-
if (num_iters * K_PER_ITER * GROUP_SIZE + K_PER_ITER*tid_in_group < p.ncols) {
95+
uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
96+
if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
18097
num_iters++;
18198
}
18299
int unroll_count = 4;
@@ -186,7 +103,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
186103
while (i < unrolled_iters) {
187104
// Manually partially unroll the loop
188105
[[unroll]] for (uint k = 0; k < unroll_count; ++k) {
189-
iter(temp, first_row, num_rows, tid_in_group, i*K_PER_ITER);
106+
iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
190107
i++;
191108
}
192109
}
@@ -205,22 +122,20 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
205122
while (i < unrolled_iters) {
206123
// Manually partially unroll the loop
207124
[[unroll]] for (uint k = 0; k < unroll_count; ++k) {
208-
iter(temp, first_row, num_rows, tid_in_group, i*K_PER_ITER);
125+
iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
209126
i++;
210127
}
211128
}
212129
while (i < num_iters) {
213-
iter(temp, first_row, num_rows, tid_in_group, i*K_PER_ITER);
130+
iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
214131
i++;
215132
}
216133

217-
reduce_result_grouped(temp, d_offset, first_row, num_rows, tid_in_group);
134+
reduce_result(temp, d_offset, first_row, num_rows, tid);
218135
}
219136

220137
void main() {
221-
const uint group_id = INVOCATION_ID / GROUP_SIZE;
222-
// 8 threads work together on a NUM_ROWS * NUM_COLS block/slice
223-
const uint first_row = NUM_ROWS * (GROUPS_PER_WARP * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z) + group_id);
138+
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
224139

225140
// do NUM_ROWS at a time, unless there aren't enough remaining rows
226141
if (first_row + NUM_ROWS <= p.stride_d) {

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

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -356,7 +356,7 @@ void main() {
356356
cache_b_qs[cc * (BK / 4) + idx_k]);
357357
}
358358

359-
sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc]);
359+
sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc], 1);
360360
}
361361
}
362362
}

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

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -16,8 +16,8 @@ i32vec2 repack(uint ib, uint iqs) {
1616
(vui >> 4) & 0x0F0F0F0F);
1717
}
1818

19-
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
20-
return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y));
19+
ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
20+
return ACC_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
2121
}
2222
#endif
2323

@@ -29,8 +29,8 @@ i32vec2 repack(uint ib, uint iqs) {
2929
(vui >> 4) & 0x0F0F0F0F);
3030
}
3131

32-
ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
33-
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
32+
ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
33+
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
3434
}
3535
#endif
3636

@@ -50,8 +50,8 @@ i32vec2 repack(uint ib, uint iqs) {
5050
return i32vec2(v0, v1);
5151
}
5252

53-
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
54-
return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y));
53+
ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
54+
return ACC_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
5555
}
5656
#endif
5757

@@ -69,8 +69,8 @@ i32vec2 repack(uint ib, uint iqs) {
6969
return i32vec2(v0, v1);
7070
}
7171

72-
ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
73-
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
72+
ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
73+
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
7474
}
7575
#endif
7676

@@ -81,7 +81,7 @@ int32_t repack(uint ib, uint iqs) {
8181
data_a[ib].qs[iqs * 2 + 1]));
8282
}
8383

84-
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
84+
ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
8585
return ACC_TYPE(float(q_sum) * da * dsb.x);
8686
}
8787
#endif

‎ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -86,7 +86,9 @@ void quantize() {
8686
data_b[ibx4_outer].qs[ibx4_inner * 8 + iqs] = pack32(i8vec4(round(vals)));
8787
#endif
8888

89+
#ifndef USE_SUBGROUPS
8990
barrier();
91+
#endif
9092

9193
// Calculate the sum for each block
9294
const float thread_sum = vals.x + vals.y + vals.z + vals.w;

0 commit comments

Comments
 (0)