91 lines
3.1 KiB
Verilog
91 lines
3.1 KiB
Verilog
`include "VX_define.vh"
|
|
|
|
module VX_gpgpu_inst (
|
|
// Input
|
|
VX_gpu_inst_req_inter VX_gpu_inst_req,
|
|
|
|
// Output
|
|
VX_warp_ctl_inter VX_warp_ctl
|
|
);
|
|
|
|
|
|
wire[`NUM_THREADS-1:0] curr_valids = VX_gpu_inst_req.valid;
|
|
wire is_split = (VX_gpu_inst_req.is_split);
|
|
|
|
wire[`NUM_THREADS-1:0] tmc_new_mask;
|
|
wire all_threads = `NUM_THREADS < VX_gpu_inst_req.a_reg_data[0];
|
|
genvar curr_t;
|
|
generate
|
|
for (curr_t = 0; curr_t < `NUM_THREADS; curr_t=curr_t+1) begin : tmc_new_mask_init
|
|
assign tmc_new_mask[curr_t] = all_threads ? 1 : curr_t < VX_gpu_inst_req.a_reg_data[0];
|
|
end
|
|
endgenerate
|
|
|
|
wire valid_inst = (|curr_valids);
|
|
|
|
assign VX_warp_ctl.warp_num = VX_gpu_inst_req.warp_num;
|
|
assign VX_warp_ctl.change_mask = (VX_gpu_inst_req.is_tmc) && valid_inst;
|
|
assign VX_warp_ctl.thread_mask = VX_gpu_inst_req.is_tmc ? tmc_new_mask : 0;
|
|
|
|
// assign VX_warp_ctl.ebreak = (VX_gpu_inst_req.a_reg_data[0] == 0) && valid_inst;
|
|
assign VX_warp_ctl.ebreak = VX_warp_ctl.change_mask && (VX_warp_ctl.thread_mask == 0);
|
|
|
|
|
|
wire wspawn = VX_gpu_inst_req.is_wspawn;
|
|
wire[31:0] wspawn_pc = VX_gpu_inst_req.rd2;
|
|
wire all_active = `NUM_WARPS < VX_gpu_inst_req.a_reg_data[0];
|
|
wire[`NUM_WARPS-1:0] wspawn_new_active;
|
|
genvar curr_w;
|
|
generate
|
|
for (curr_w = 0; curr_w < `NUM_WARPS; curr_w=curr_w+1) begin : wspawn_new_active_init
|
|
assign wspawn_new_active[curr_w] = all_active ? 1 : curr_w < VX_gpu_inst_req.a_reg_data[0];
|
|
end
|
|
endgenerate
|
|
|
|
|
|
assign VX_warp_ctl.is_barrier = VX_gpu_inst_req.is_barrier && valid_inst;
|
|
assign VX_warp_ctl.barrier_id = VX_gpu_inst_req.a_reg_data[0];
|
|
|
|
wire[31:0] num_warps_m1 = VX_gpu_inst_req.rd2 - 1;
|
|
assign VX_warp_ctl.num_warps = num_warps_m1[$clog2(`NUM_WARPS):0];
|
|
|
|
assign VX_warp_ctl.wspawn = wspawn;
|
|
assign VX_warp_ctl.wspawn_pc = wspawn_pc;
|
|
assign VX_warp_ctl.wspawn_new_active = wspawn_new_active;
|
|
|
|
wire[`NUM_THREADS-1:0] split_new_use_mask;
|
|
wire[`NUM_THREADS-1:0] split_new_later_mask;
|
|
|
|
// VX_gpu_inst_req.pc
|
|
genvar curr_s_t;
|
|
generate
|
|
for (curr_s_t = 0; curr_s_t < `NUM_THREADS; curr_s_t=curr_s_t+1) begin : masks_init
|
|
wire curr_bool = (VX_gpu_inst_req.a_reg_data[curr_s_t] == 32'b1);
|
|
|
|
assign split_new_use_mask[curr_s_t] = curr_valids[curr_s_t] & (curr_bool);
|
|
assign split_new_later_mask[curr_s_t] = curr_valids[curr_s_t] & (!curr_bool);
|
|
end
|
|
endgenerate
|
|
|
|
wire[$clog2(`NUM_THREADS):0] num_valids;
|
|
|
|
VX_countones #(.N(`NUM_THREADS)) valids_counter (
|
|
.valids(curr_valids),
|
|
.count (num_valids)
|
|
);
|
|
|
|
// wire[`NW_BITS-1:0] num_valids = $countones(curr_valids);
|
|
|
|
|
|
assign VX_warp_ctl.is_split = is_split && (num_valids > 1);
|
|
assign VX_warp_ctl.dont_split = VX_warp_ctl.is_split && ((split_new_use_mask == 0) || (split_new_use_mask == {`NUM_THREADS{1'b1}}));
|
|
assign VX_warp_ctl.split_new_mask = split_new_use_mask;
|
|
assign VX_warp_ctl.split_later_mask = split_new_later_mask;
|
|
assign VX_warp_ctl.split_save_pc = VX_gpu_inst_req.pc_next;
|
|
assign VX_warp_ctl.split_warp_num = VX_gpu_inst_req.warp_num;
|
|
|
|
// VX_gpu_inst_req.is_wspawn
|
|
// VX_gpu_inst_req.is_split
|
|
// VX_gpu_inst_req.is_barrier
|
|
|
|
endmodule |