Skip to content

Commit 5109563

Browse files
zhiyuan1izhiyuan1i
committed
wkv on sycl
1 parent 7febf6e commit 5109563

File tree

5 files changed

+156
-1
lines changed

5 files changed

+156
-1
lines changed

‎docs/backend/SYCL.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -377,7 +377,7 @@ found 2 SYCL devices:
377377

378378
|Chosen Device ID|Setting|
379379
|-|-|
380-
|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
380+
|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:0"` or no action|
381381
|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
382382
|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
383383

‎ggml/src/ggml-sycl.cpp

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -4803,6 +4803,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
48034803
case GGML_OP_TIMESTEP_EMBEDDING:
48044804
func = ggml_sycl_op_timestep_embedding;
48054805
break;
4806+
case GGML_OP_RWKV_WKV6:
4807+
func = ggml_sycl_op_rwkv_wkv6;
4808+
break;
48064809
default:
48074810
return false;
48084811
}
@@ -5251,6 +5254,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
52515254
case GGML_OP_PAD:
52525255
case GGML_OP_LEAKY_RELU:
52535256
case GGML_OP_TIMESTEP_EMBEDDING:
5257+
case GGML_OP_RWKV_WKV6:
52545258
return true;
52555259
default:
52565260
return false;

‎ggml/src/ggml-sycl/backend.hpp

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -26,5 +26,6 @@
2626
#include "softmax.hpp"
2727
#include "tsembd.hpp"
2828
#include "im2col.hpp"
29+
#include "wkv6.hpp"
2930

3031
#endif // GGML_SYCL_BACKEND_HPP

‎ggml/src/ggml-sycl/wkv6.cpp

Lines changed: 140 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,140 @@
1+
// Copyright (C) 2024 Zhiyuan Li
2+
3+
4+
#include <sycl/sycl.hpp>
5+
#include "wkv6.hpp"
6+
7+
constexpr int WKV_BLOCK_SIZE = 64; // Matching CUDA_WKV_BLOCK_SIZE
8+
9+
// Helper function for the main kernel
10+
static void rwkv_wkv_f32_kernel(
11+
const int B, const int T, const int C, const int H,
12+
const float* k, const float* v, const float* r,
13+
const float* tf, const float* td, const float* s,
14+
float* dst, const sycl::nd_item<3>& item_ct1, float* shared_mem) {
15+
16+
const int tid = item_ct1.get_local_id(2);
17+
const int bid = item_ct1.get_group(2);
18+
19+
const int head_size = WKV_BLOCK_SIZE;
20+
const int batch_i = bid / H;
21+
const int head_i = bid % H;
22+
const int state_size = C * head_size;
23+
const int n_seq_tokens = T / B;
24+
25+
// Set up shared memory pointers
26+
float* _k = shared_mem;
27+
float* _r = _k + head_size;
28+
float* _tf = _r + head_size;
29+
float* _td = _tf + head_size;
30+
31+
// Local state array
32+
float state[WKV_BLOCK_SIZE];
33+
34+
// Load initial state
35+
#pragma unroll
36+
for (int i = 0; i < head_size; i++) {
37+
state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
38+
}
39+
40+
// Sync threads before shared memory operations
41+
item_ct1.barrier(sycl::access::fence_space::local_space);
42+
43+
// Load time-mixing parameters
44+
_tf[tid] = tf[head_i * head_size + tid];
45+
item_ct1.barrier(sycl::access::fence_space::local_space);
46+
47+
// Main sequence processing loop
48+
for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid;
49+
t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid;
50+
t += C) {
51+
52+
item_ct1.barrier(sycl::access::fence_space::local_space);
53+
54+
// Load current timestep data to shared memory
55+
_k[tid] = k[t];
56+
_r[tid] = r[t];
57+
_td[tid] = td[t];
58+
59+
item_ct1.barrier(sycl::access::fence_space::local_space);
60+
61+
const float _v = v[t];
62+
float y = 0;
63+
64+
// Process in chunks of 4 for better vectorization
65+
#pragma unroll
66+
for (int j = 0; j < head_size; j += 4) {
67+
// Load data in vec4 chunks
68+
sycl::float4 k4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
69+
sycl::float4 r4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
70+
sycl::float4 tf4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
71+
sycl::float4 td4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
72+
sycl::float4 s4(state[j], state[j+1], state[j+2], state[j+3]);
73+
74+
// Compute key-value product
75+
sycl::float4 kv4 = k4 * _v;
76+
77+
// Accumulate weighted sum
78+
y += sycl::dot(r4, tf4 * kv4 + s4);
79+
80+
// Update state
81+
s4 = s4 * td4 + kv4;
82+
83+
// Store updated state
84+
state[j] = s4.x();
85+
state[j+1] = s4.y();
86+
state[j+2] = s4.z();
87+
state[j+3] = s4.w();
88+
}
89+
90+
dst[t] = y;
91+
}
92+
93+
// Save final state
94+
#pragma unroll
95+
for (int i = 0; i < head_size; i++) {
96+
dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
97+
}
98+
}
99+
100+
void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
101+
const ggml_tensor* src1, ggml_tensor* dst) {
102+
103+
const float* k_d = (const float*)dst->src[0]->data;
104+
const float* v_d = (const float*)dst->src[1]->data;
105+
const float* r_d = (const float*)dst->src[2]->data;
106+
const float* tf_d = (const float*)dst->src[3]->data;
107+
const float* td_d = (const float*)dst->src[4]->data;
108+
const float* s_d = (const float*)dst->src[5]->data;
109+
float* dst_d = (float*)dst->data;
110+
111+
const int64_t B = dst->src[5]->ne[1];
112+
const int64_t T = dst->src[0]->ne[3];
113+
const int64_t C = dst->ne[0];
114+
const int64_t H = dst->src[0]->ne[2];
115+
116+
GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
117+
GGML_ASSERT(C % H == 0);
118+
GGML_ASSERT(C / H == WKV_BLOCK_SIZE);
119+
120+
dpct::queue_ptr stream = ctx.stream();
121+
122+
// Calculate execution configuration
123+
const size_t shared_mem_size = WKV_BLOCK_SIZE * 4 * sizeof(float); // For k, r, tf, td
124+
sycl::range<3> block_dims(1, 1, C / H);
125+
sycl::range<3> grid_dims(1, 1, B * H);
126+
127+
// Submit kernel
128+
stream->submit([&](sycl::handler& cgh) {
129+
sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
130+
131+
cgh.parallel_for(
132+
sycl::nd_range<3>(grid_dims * block_dims, block_dims),
133+
[=](sycl::nd_item<3> item_ct1) {
134+
rwkv_wkv_f32_kernel(
135+
B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
136+
item_ct1, shared_mem_acc.get_pointer()
137+
);
138+
});
139+
});
140+
}

‎ggml/src/ggml-sycl/wkv6.hpp

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,10 @@
1+
#ifndef GGML_SYCL_WKV6_HPP
2+
#define GGML_SYCL_WKV6_HPP
3+
4+
#include "common.hpp"
5+
6+
void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
7+
const ggml_tensor *src1, ggml_tensor * dst);
8+
9+
10+
#endif // GGML_SYCL_WKV6_HPP

0 commit comments

Comments
 (0)