| 2198 | } |
| 2199 | |
| 2200 | static webgpu_encoded_op ggml_webgpu_concat(webgpu_context & ctx, |
| 2201 | ggml_tensor * src0, |
| 2202 | ggml_tensor * src1, |
| 2203 | ggml_tensor * dst) { |
| 2204 | uint32_t ne = (uint32_t) ggml_nelements(dst); |
| 2205 | uint32_t dim = (uint32_t) dst->op_params[0]; |
| 2206 | |
| 2207 | std::vector<uint32_t> params = { |
| 2208 | ne, |
| 2209 | (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)), |
| 2210 | (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)), |
| 2211 | (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)), |
| 2212 | (uint32_t) (src0->nb[0] / ggml_type_size(src0->type)), |
| 2213 | (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)), |
| 2214 | (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)), |
| 2215 | (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)), |
| 2216 | (uint32_t) (src1->nb[0] / ggml_type_size(src1->type)), |
| 2217 | (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)), |
| 2218 | (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)), |
| 2219 | (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)), |
| 2220 | (uint32_t) dst->ne[0], |
| 2221 | (uint32_t) dst->ne[1], |
| 2222 | (uint32_t) dst->ne[2], |
| 2223 | (uint32_t) dst->ne[3], |
| 2224 | dim, |
| 2225 | (uint32_t) src0->ne[dim] |
| 2226 | }; |
| 2227 | |
| 2228 | std::vector<wgpu::BindGroupEntry> entries = { |
| 2229 | ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0), |
| 2230 | ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, src1), |
| 2231 | ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst), |
| 2232 | }; |
| 2233 | |
| 2234 | ggml_webgpu_shader_lib_context shader_lib_ctx = {}; |
| 2235 | shader_lib_ctx.src0 = src0; |
| 2236 | shader_lib_ctx.src1 = src1; |
| 2237 | shader_lib_ctx.dst = dst; |
| 2238 | shader_lib_ctx.max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup; |
| 2239 | |
| 2240 | webgpu_pipeline pipeline = ctx->shader_lib->get_concat_pipeline(shader_lib_ctx); |
| 2241 | auto * decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get()); |
| 2242 | uint32_t wg_x = CEIL_DIV(ne, decisions->wg_size); |
| 2243 | return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); |
| 2244 | } |
| 2245 | |
| 2246 | static webgpu_encoded_op ggml_webgpu_repeat(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * dst) { |
| 2247 | uint32_t ne = (uint32_t) ggml_nelements(dst); |
no test coverage detected