From 2fc42b6a82a841de5539f4c8ba8d456f4e060d7f Mon Sep 17 00:00:00 2001 From: Zhiyuan Li Date: Sun, 3 Nov 2024 01:12:52 +1100 Subject: [PATCH] wkv on sycl --- docs/backend/SYCL.md | 2 +- ggml/src/ggml-sycl.cpp | 4 + ggml/src/ggml-sycl/backend.hpp | 1 + ggml/src/ggml-sycl/wkv6.cpp | 140 +++++++++++++++++++++++++++++++++ ggml/src/ggml-sycl/wkv6.hpp | 10 +++ 5 files changed, 156 insertions(+), 1 deletion(-) create mode 100644 ggml/src/ggml-sycl/wkv6.cpp create mode 100644 ggml/src/ggml-sycl/wkv6.hpp diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md index ea34182e4..bc8c0f886 100644 --- a/docs/backend/SYCL.md +++ b/docs/backend/SYCL.md @@ -377,7 +377,7 @@ found 2 SYCL devices: |Chosen Device ID|Setting| |-|-| -|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action| +|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:0"` or no action| |1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`| |0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`| diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index a62c67f4f..53b8cec1a 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -4803,6 +4803,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens case GGML_OP_TIMESTEP_EMBEDDING: func = ggml_sycl_op_timestep_embedding; break; + case GGML_OP_RWKV_WKV6: + func = ggml_sycl_op_rwkv_wkv6; + break; default: return false; } @@ -5251,6 +5254,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_PAD: case GGML_OP_LEAKY_RELU: case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_RWKV_WKV6: return true; default: return false; diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index d21b5f8dd..17811725c 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -26,5 +26,6 @@ #include "softmax.hpp" #include "tsembd.hpp" #include "im2col.hpp" +#include "wkv6.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/wkv6.cpp b/ggml/src/ggml-sycl/wkv6.cpp new file mode 100644 index 000000000..93a2edd56 --- /dev/null +++ b/ggml/src/ggml-sycl/wkv6.cpp @@ -0,0 +1,140 @@ +// Copyright (C) 2024 Zhiyuan Li + + +#include +#include "wkv6.hpp" + +constexpr int WKV_BLOCK_SIZE = 64; // Matching CUDA_WKV_BLOCK_SIZE + +// Helper function for the main kernel +static void rwkv_wkv_f32_kernel( + const int B, const int T, const int C, const int H, + const float* k, const float* v, const float* r, + const float* tf, const float* td, const float* s, + float* dst, const sycl::nd_item<3>& item_ct1, float* shared_mem) { + + const int tid = item_ct1.get_local_id(2); + const int bid = item_ct1.get_group(2); + + const int head_size = WKV_BLOCK_SIZE; + const int batch_i = bid / H; + const int head_i = bid % H; + const int state_size = C * head_size; + const int n_seq_tokens = T / B; + + // Set up shared memory pointers + float* _k = shared_mem; + float* _r = _k + head_size; + float* _tf = _r + head_size; + float* _td = _tf + head_size; + + // Local state array + float state[WKV_BLOCK_SIZE]; + + // Load initial state + #pragma unroll + for (int i = 0; i < head_size; i++) { + state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid]; + } + + // Sync threads before shared memory operations + item_ct1.barrier(sycl::access::fence_space::local_space); + + // Load time-mixing parameters + _tf[tid] = tf[head_i * head_size + tid]; + item_ct1.barrier(sycl::access::fence_space::local_space); + + // Main sequence processing loop + for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; + t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; + t += C) { + + item_ct1.barrier(sycl::access::fence_space::local_space); + + // Load current timestep data to shared memory + _k[tid] = k[t]; + _r[tid] = r[t]; + _td[tid] = td[t]; + + item_ct1.barrier(sycl::access::fence_space::local_space); + + const float _v = v[t]; + float y = 0; + + // Process in chunks of 4 for better vectorization + #pragma unroll + for (int j = 0; j < head_size; j += 4) { + // Load data in vec4 chunks + sycl::float4 k4(_k[j], _k[j+1], _k[j+2], _k[j+3]); + sycl::float4 r4(_r[j], _r[j+1], _r[j+2], _r[j+3]); + sycl::float4 tf4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]); + sycl::float4 td4(_td[j], _td[j+1], _td[j+2], _td[j+3]); + sycl::float4 s4(state[j], state[j+1], state[j+2], state[j+3]); + + // Compute key-value product + sycl::float4 kv4 = k4 * _v; + + // Accumulate weighted sum + y += sycl::dot(r4, tf4 * kv4 + s4); + + // Update state + s4 = s4 * td4 + kv4; + + // Store updated state + state[j] = s4.x(); + state[j+1] = s4.y(); + state[j+2] = s4.z(); + state[j+3] = s4.w(); + } + + dst[t] = y; + } + + // Save final state + #pragma unroll + for (int i = 0; i < head_size; i++) { + dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i]; + } +} + +void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst) { + + const float* k_d = (const float*)dst->src[0]->data; + const float* v_d = (const float*)dst->src[1]->data; + const float* r_d = (const float*)dst->src[2]->data; + const float* tf_d = (const float*)dst->src[3]->data; + const float* td_d = (const float*)dst->src[4]->data; + const float* s_d = (const float*)dst->src[5]->data; + float* dst_d = (float*)dst->data; + + const int64_t B = dst->src[5]->ne[1]; + const int64_t T = dst->src[0]->ne[3]; + const int64_t C = dst->ne[0]; + const int64_t H = dst->src[0]->ne[2]; + + GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32); + GGML_ASSERT(C % H == 0); + GGML_ASSERT(C / H == WKV_BLOCK_SIZE); + + dpct::queue_ptr stream = ctx.stream(); + + // Calculate execution configuration + const size_t shared_mem_size = WKV_BLOCK_SIZE * 4 * sizeof(float); // For k, r, tf, td + sycl::range<3> block_dims(1, 1, C / H); + sycl::range<3> grid_dims(1, 1, B * H); + + // Submit kernel + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); + + cgh.parallel_for( + sycl::nd_range<3>(grid_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rwkv_wkv_f32_kernel( + B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, + item_ct1, shared_mem_acc.get_pointer() + ); + }); + }); +} \ No newline at end of file diff --git a/ggml/src/ggml-sycl/wkv6.hpp b/ggml/src/ggml-sycl/wkv6.hpp new file mode 100644 index 000000000..d51bf858f --- /dev/null +++ b/ggml/src/ggml-sycl/wkv6.hpp @@ -0,0 +1,10 @@ +#ifndef GGML_SYCL_WKV6_HPP +#define GGML_SYCL_WKV6_HPP + +#include "common.hpp" + +void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor * dst); + + +#endif // GGML_SYCL_WKV6_HPP \ No newline at end of file