wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

adding fast DPI implemntation of imul and idiv

Browse Source
This commit is contained in:
Blaise Tine
2021-06-22 09:02:41 -07:00
parent 65a3704479
commit c331da5ff7
5 changed files with 136 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
hw/rtl/VX_muldiv.v
View File

@@ -1,5 +1,9 @@
`include "VX_define.vh"
`ifndef SYNTHESIS
`include "util_dpi.vh"
`endif
module VX_muldiv (
input wire clk,
input wire reset,
@@ -43,13 +47,42 @@ module VX_muldiv (
wire mul_valid_out;
wire mul_valid_in = valid_in && !is_div_op;
wire mul_ready_in = ~stall_out || ~mul_valid_out;
wire is_mulh_in = (alu_op != `MUL_MUL);
wire is_mulh_in = (alu_op != `MUL_MUL);
wire is_signed_mul_a = (alu_op != `MUL_MULHU);
wire is_signed_mul_b = (alu_op != `MUL_MULHU && alu_op != `MUL_MULHSU);
`ifdef IMUL_DPI
wire [`NUM_THREADS-1:0][31:0] mul_result_tmp;
for (genvar i = 0; i < `NUM_THREADS; i++) begin
wire [31:0] mul_resultl, mul_resulth;
always @(*) begin
dpi_imul (alu_in1[i], alu_in2[i], is_signed_mul_a, is_signed_mul_b, mul_resultl, mul_resulth);
end
assign mul_result_tmp[i] = is_mulh_in ? mul_resulth : mul_resultl;
end
VX_shift_register #(
.DATAW (1 + `NW_BITS + `NUM_THREADS + 32 + `NR_BITS + 1 + (`NUM_THREADS * 32)),
.DEPTH (`LATENCY_IMUL),
.RESETW (1)
) mul_shift_reg (
.clk(clk),
.reset (reset),
.enable (mul_ready_in),
.data_in ({mul_valid_in, wid_in, tmask_in, PC_in, rd_in, wb_in, mul_result_tmp}),
.data_out ({mul_valid_out, mul_wid_out, mul_tmask_out, mul_PC_out, mul_rd_out, mul_wb_out, mul_result})
);
`else
wire is_mulh_out;
for (genvar i = 0; i < `NUM_THREADS; i++) begin
wire [32:0] mul_in1 = {(alu_op != `MUL_MULHU) & alu_in1[i][31], alu_in1[i]};
wire [32:0] mul_in2 = {(alu_op != `MUL_MULHU && alu_op != `MUL_MULHSU) & alu_in2[i][31], alu_in2[i]};
wire [32:0] mul_in1 = {is_signed_mul_a & alu_in1[i][31], alu_in1[i]};
wire [32:0] mul_in2 = {is_signed_mul_b & alu_in2[i][31], alu_in2[i]};
`IGNORE_WARNINGS_BEGIN
wire [65:0] mul_result_tmp;
`IGNORE_WARNINGS_END
@@ -83,9 +116,11 @@ module VX_muldiv (
.data_out ({mul_valid_out, mul_wid_out, mul_tmask_out, mul_PC_out, mul_rd_out, mul_wb_out, is_mulh_out})
);
`endif
///////////////////////////////////////////////////////////////////////////
wire [`NUM_THREADS-1:0][31:0] div_result_tmp, rem_result_tmp;
wire [`NUM_THREADS-1:0][31:0] div_result;
wire [`NW_BITS-1:0] div_wid_out;
wire [`NUM_THREADS-1:0] div_tmask_out;
wire [31:0] div_PC_out;
@@ -98,6 +133,36 @@ module VX_muldiv (
wire div_ready_out = ~stall_out && ~mul_valid_out; // arbitration prioritizes MUL
wire div_ready_in;
wire div_valid_out;
`ifdef IDIV_DPI
wire [`NUM_THREADS-1:0][31:0] div_result_tmp;
for (genvar i = 0; i < `NUM_THREADS; i++) begin
wire [31:0] div_quotient, div_remainder;
always @(*) begin
dpi_idiv (alu_in1[i], alu_in2[i], is_signed_div, div_quotient, div_remainder);
end
assign div_result_tmp[i] = is_rem_op_in ? div_remainder : div_quotient;
end
VX_shift_register #(
.DATAW (1 + `NW_BITS + `NUM_THREADS + 32 + `NR_BITS + 1 + (`NUM_THREADS * 32)),
.DEPTH (`LATENCY_IMUL),
.RESETW (1)
) div_shift_reg (
.clk(clk),
.reset (reset),
.enable (div_ready_in),
.data_in ({div_valid_in, wid_in, tmask_in, PC_in, rd_in, wb_in, div_result_tmp}),
.data_out ({div_valid_out, div_wid_out, div_tmask_out, div_PC_out, div_rd_out, div_wb_out, div_result})
);
assign div_ready_in = div_ready_out || ~div_valid_out;
`else
wire [`NUM_THREADS-1:0][31:0] div_result_tmp, rem_result_tmp;
wire is_rem_op_out;
VX_serial_div #(
@@ -123,7 +188,9 @@ module VX_muldiv (
.tag_out ({div_wid_out, div_tmask_out, div_PC_out, div_rd_out, div_wb_out, is_rem_op_out})
);
wire [`NUM_THREADS-1:0][31:0] div_result = is_rem_op_out ? rem_result_tmp : div_result_tmp;
assign div_result = is_rem_op_out ? rem_result_tmp : div_result_tmp;
`endif
///////////////////////////////////////////////////////////////////////////