@@ -990,6 +990,7 @@ static void conv2d_implicit_cuda_f16(ggml_backend_cuda_context & ctx, const floa
990990
991991static void conv2d_implicit_cuda_f32 (ggml_backend_cuda_context & ctx, const float * X_D, const float * K_D, float * Y_D, int cc, const param_t P, cudaStream_t st) {
992992 conv2d_implicit_cuda<float , 1 >(X_D, K_D, Y_D, P, st);
993+ GGML_UNUSED (ctx);
993994}
994995
995996void ggml_cuda_op_conv2d_implicit (ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -1033,14 +1034,13 @@ void ggml_cuda_op_conv2d_implicit(ggml_backend_cuda_context & ctx, ggml_tensor *
10331034 const uint OC = kernel->ne [3 ]; // ouptut_chanles
10341035 const uint B = input->ne [3 ]; // n_batches
10351036
1036- param_t params = { B, IC, IH, IW, OC, KH, KW, ST_Y, ST_X, PD_Y, PD_X, DL_Y, DL_X, OH, OW };
1037- params.SC_fastdiv = init_fastdiv_values (KW*IC);
1038- params.OW_fastdiv = init_fastdiv_values (OW);
1039- params.OHOW_fastdiv = init_fastdiv_values (OW*OH);
1040- params.C_fastdiv = init_fastdiv_values (IC);
1041- params.RS_fastdiv = init_fastdiv_values (KW*KH);
1042- params.S_fastdiv = init_fastdiv_values (KW);
1043- // params.layout = LT;
1037+ param_t params = { B, IC, IH, IW, OC, KH, KW, ST_Y, ST_X, PD_Y, PD_X, DL_Y, DL_X, OH, OW,
1038+ init_fastdiv_values (KW*IC),
1039+ init_fastdiv_values (OW),
1040+ init_fastdiv_values (IC),
1041+ init_fastdiv_values (KW*KH),
1042+ init_fastdiv_values (KW),
1043+ init_fastdiv_values (OW*OH)};
10441044
10451045 if (kernel->type == GGML_TYPE_F16) {
10461046 conv2d_implicit_cuda_f16 (ctx, X_D, (half *) K_D, Y_D, cc, params, st);
0 commit comments