1#include "ggml-impl.h"
 2#include "opt-step-sgd.cuh"
 3
 4#include <cstdint>
 5
 6static __global__ void opt_step_sgd_f32(
 7    float * __restrict__ x, const float * __restrict__ g,
 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    x[i] = x[i] * (1.0f - pars[0] * pars[1]) - pars[0] * g[i];
16}
17
18static void opt_step_sgd_f32_cuda(
19    float * x, const float * g, const float * __restrict__ pars, const int64_t k, cudaStream_t stream) {
20
21    const dim3 block_dims(CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
22    const dim3 block_nums((k + CUDA_OPT_STEP_SGD_BLOCK_SIZE - 1) / CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
23    opt_step_sgd_f32<<<block_nums, block_dims, 0, stream>>>(x, g, pars, k);
24}
25
26void ggml_cuda_opt_step_sgd(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
27    const ggml_tensor * src0      = dst->src[0];
28    const ggml_tensor * src0_grad = dst->src[1];
29    const ggml_tensor * params    = dst->src[2];
30
31    GGML_ASSERT(src0->type      == GGML_TYPE_F32);
32    GGML_ASSERT(src0_grad->type == GGML_TYPE_F32);
33    GGML_ASSERT(params->type    == GGML_TYPE_F32);
34    GGML_ASSERT(ggml_is_contiguous(src0));
35    GGML_ASSERT(ggml_is_contiguous(src0_grad));
36    GGML_ASSERT(ggml_is_contiguous(params));
37    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
38    GGML_ASSERT(ggml_nelements(params) == 2);
39
40    float       * src0_d      = (float       *) src0->data;
41    const float * src0_grad_d = (const float *) src0_grad->data;
42    const float * params_d    = (const float *) params->data;
43
44    cudaStream_t stream = ctx.stream();
45
46    const int64_t ne = ggml_nelements(src0);
47
48    opt_step_sgd_f32_cuda(src0_d, src0_grad_d, params_d, ne, stream);
49}