1#include "ggml-impl.h"
 2#include "opt-step-adamw.cuh"
 3
 4#include <cstdint>
 5
 6static __global__ void opt_step_adamw_f32(
 7    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v,
 8    const float * __restrict__ pars, const int64_t k) {
 9
10    const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
11
12    if (i >= k) {
13        return;
14    }
15
16    const float alpha  = pars[0];
17    const float beta1  = pars[1];
18    const float beta2  = pars[2];
19    const float eps    = pars[3];
20    const float wd     = pars[4];
21    const float beta1h = pars[5];
22    const float beta2h = pars[6];
23
24    const float gi = g[i];
25    const float gmi = g_m[i]*beta1 +    gi*(1.0f - beta1);
26    const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
27
28    g_m[i] = gmi;
29    g_v[i] = gvi;
30
31    const float mh =       gmi*beta1h;
32    const float vh = sqrtf(gvi*beta2h) + eps;
33
34    x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
35}
36
37static void opt_step_adamw_f32_cuda(
38    float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
39
40    const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
41    const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
42    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, pars, k);
43}
44
45void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
46    const ggml_tensor * src0         = dst->src[0];
47    const ggml_tensor * src0_grad    = dst->src[1];
48    const ggml_tensor * src0_grad_m  = dst->src[2];
49    const ggml_tensor * src0_grad_v  = dst->src[3];
50    const ggml_tensor * adamw_params = dst->src[4];
51
52    GGML_ASSERT(src0->type         == GGML_TYPE_F32);
53    GGML_ASSERT(src0_grad->type    == GGML_TYPE_F32);
54    GGML_ASSERT(src0_grad_m->type  == GGML_TYPE_F32);
55    GGML_ASSERT(src0_grad_v->type  == GGML_TYPE_F32);
56    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
57    GGML_ASSERT(ggml_is_contiguous(src0));
58    GGML_ASSERT(ggml_is_contiguous(src0_grad));
59    GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
60    GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
61    GGML_ASSERT(ggml_is_contiguous(adamw_params));
62    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
63    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
64    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
65    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
66
67    float       * src0_d         = (float       *) src0->data;
68    const float * src0_grad_d    = (const float *) src0_grad->data;
69    float       * src0_grad_m_d  = (float       *) src0_grad_m->data;
70    float       * src0_grad_v_d  = (float       *) src0_grad_v->data;
71    const float * adamw_params_d = (const float *) adamw_params->data;
72
73    cudaStream_t stream = ctx.stream();
74
75    const int64_t ne = ggml_nelements(src0);
76
77    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
78}