Files
vortex/hw/rtl/fpu/VX_tensor_dpu.sv

342 lines
11 KiB
Systemverilog

`ifdef EXT_T_ENABLE
`include "VX_fpu_define.vh"
// Module that contains the threadgroups with DPUs + operand buffer.
module VX_tensor_threadgroups #(
parameter ISW,
parameter OCTET,
parameter FP16,
// @perf: has big impact on throughput. A rule of thumb is to set it to
// the pipeline length of FEDPs in order to make sure there are enough
// entries to fully saturate the pipeline, but this is still rough
parameter OPERAND_BUFFER_DEPTH = `LATENCY_HMMA
) (
input clk,
input reset,
input valid_in,
output ready_in,
// [rows][cols][dtype]
// (m,n,k) = (4,4,2)
input [3:0][1:0][31:0] A_tile,
input [1:0][3:0][31:0] B_tile,
input [3:0][3:0][31:0] C_tile,
input [`NW_WIDTH-1:0] wid,
output valid_out,
input ready_out,
output [3:0][3:0][31:0] D_tile,
output [`NW_WIDTH-1:0] D_wid
);
// logic [3:0][3:0][31:0] result_hmma;
// always @(*) begin
// dpi_hmma(valid_in, A_tile, B_tile, C_tile, result_hmma);
// end
// // fixed-latency queue
// VX_shift_register #(
// .DATAW (1 + $bits(wid)/* + $bits(D_tile)*/),
// .DEPTH (`LATENCY_HMMA + 1),
// .RESETW (1)
// ) shift_reg (
// .clk (clk),
// .reset (reset),
// .enable (ready_out),
// .data_in ({valid_in && ready_in, wid /*, result_hmma*/}),
// .data_out ({valid_out, D_wid/*, D_tile */})
// );
// ready as soon as valid_out
// assign ready_in = valid_out;
// fully pipelined; ready_in is coupled to ready_out by immediately
// stalling
// assign ready_in = ready_out;
wire [3:0][1:0][31:0] A_tile_buf;
wire [1:0][3:0][31:0] B_tile_buf;
wire [3:0][3:0][31:0] C_tile_buf;
wire wid_empty;
wire wid_full;
wire empty;
wire full;
// sync operand buffer and wid buffer
assign ready_in = !full && !wid_full;
wire [1:0] threadgroup_valids_out;
wire [1:0] threadgroup_readys_in;
wire threadgroup_valid_in = !empty;
wire threadgroup_fire_in = threadgroup_valid_in && &(threadgroup_readys_in);
wire enq = valid_in && ready_in;
wire deq = threadgroup_fire_in;
// Operand buffer for the dot product units.
//
// This exists to decouple the execution of the dot-product unit from
// the operand arrival. Operands from the upstream execute_if can arrive
// intermittently depending on the frontend's behavior, whereas downstream
// writeback happens at a regular cadence. Therefore to achieve full
// throughput of the dpu, we need to decouple the operand arrival from the
// direct input to the dpu.
VX_fifo_queue #(
.DATAW ($bits(A_tile) + $bits(B_tile) + $bits(C_tile)),
.DEPTH (OPERAND_BUFFER_DEPTH)
) operand_buffer (
.clk (clk),
.reset (reset),
.push (enq),
.pop (deq),
.data_in ({A_tile, B_tile, C_tile}),
.data_out ({A_tile_buf, B_tile_buf, C_tile_buf}),
.empty (empty),
`UNUSED_PIN(alm_empty),
.full (full),
`UNUSED_PIN(alm_full),
`UNUSED_PIN(size)
);
// Split A_tile and C_tile by rows (0-1, 2-3) and parallelize in two
// threadgroup DPUs; B_tile is shared across the two threadgroups. See
// Figure 13 in paper
VX_tensor_threadgroup #(
.FP16(FP16)
) threadgroup_0 (
.clk (clk),
.reset (reset),
.valid_in (threadgroup_valid_in),
.ready_in (threadgroup_readys_in[0]),
.stall (!ready_out),
.A_frag (A_tile_buf[1:0]),
.B_frag (B_tile_buf),
.C_frag (C_tile_buf[1:0]),
.valid_out (threadgroup_valids_out[0]),
.D_frag (D_tile[1:0])
);
VX_tensor_threadgroup #(
.FP16(FP16)
) threadgroup_1 (
.clk (clk),
.reset (reset),
.valid_in (threadgroup_valid_in),
.ready_in (threadgroup_readys_in[1]),
.stall (!ready_out),
.A_frag (A_tile_buf[3:2]),
.B_frag (B_tile_buf),
.C_frag (C_tile_buf[3:2]),
.valid_out (threadgroup_valids_out[1]),
.D_frag (D_tile[3:2])
);
`RUNTIME_ASSERT(&(threadgroup_valids_out) == |(threadgroup_valids_out),
("threadgroups went out of sync!"))
`RUNTIME_ASSERT(&(threadgroup_readys_in) == |(threadgroup_readys_in),
("threadgroups went out of sync!"))
wire wid_enq = valid_in && ready_in;
wire wid_deq = valid_out && ready_out;
assign valid_out = &(threadgroup_valids_out);
// need to pass along warp id's to do multithreading
VX_fifo_queue #(
.DATAW ($bits(wid)),
// @perf: seems to require deeper depth than the FEDP issue queues to
// not cause stalls.
.DEPTH (2 * OPERAND_BUFFER_DEPTH)
) wid_queue (
.clk (clk),
.reset (reset),
.push (wid_enq),
.pop (wid_deq),
.data_in (wid),
.data_out (D_wid),
.empty (wid_empty),
`UNUSED_PIN(alm_empty),
.full (wid_full),
`UNUSED_PIN(alm_full),
`UNUSED_PIN(size)
);
`RUNTIME_ASSERT(reset || !(deq && empty),
("dequeueing from empty warp id queue!"))
endmodule
// does (m,n,k) = (2,4,2) matmul compute over 2 cycles.
// see Figure 10(b) of the paper.
module VX_tensor_threadgroup #(
parameter FP16
) (
input clk,
input reset,
input valid_in,
output ready_in,
input stall,
// all *_frag are row-major
// (m,n,k) = (2,4,2)
input [1:0][1:0][31:0] A_frag,
input [1:0][3:0][31:0] B_frag,
input [1:0][3:0][31:0] C_frag,
output valid_out,
output [1:0][3:0][31:0] D_frag
);
wire fedp_valid_in;
wire fedp_ready_in;
wire fedp_fire_in = fedp_valid_in && fedp_ready_in;
wire [3:0] fedp_valids;
wire fedp_valid_out = &(fedp_valids);
wire fedp_ready_out = !stall;
wire fedp_fire_out = fedp_valid_out && fedp_ready_out;
assign fedp_valid_in = valid_in;
// coupled ready; backpressure immediately reaches input from output
assign fedp_ready_in = fedp_ready_out;
// The dot product units take 2 cycles to finish computing A_frag * B_frag
// + C_frag. substep_in and substep_out keeps track of which cycle they're at
// & when they have to pop from input queue and push to result queue.
// Note that substep is different from the "step" defined in the HMMA
// instruction set; it is similar in meaning to the substeps in
// VX_tensor_octet.
//
// substep_in == 0: FEDP uses first half from operand buffer
// substep_in == 1: FEDP uses last half and pops from operand buffer
wire substep_in;
// substep_out == 0: FEDP produces first half of D_frag
// substep_out == 1: FEDP produces last half of D_frag and asserts valid_out
wire substep_out;
assign ready_in = fedp_fire_in && (substep_in == 1'b1);
wire [3:0][31:0] D_reg;
logic [3:0][31:0] D_reg_n;
// staging buffer that latches the D half-tile
VX_tensor_reg #(
.DATAW($bits(D_reg))
) staging_d (
.clk(clk),
.reset(reset),
.d(D_reg_n),
.en(1'b1),
.q(D_reg)
);
// latch the first-half result of D_frag
wire [3:0][31:0] D_half;
always @(*) begin
D_reg_n = D_reg;
if (fedp_fire_out) begin
if (substep_out == 1'b0) begin
D_reg_n = D_half;
end
end
end
// flip substep_in/substep_out on FEDP in/out fire, respectively
VX_tensor_reg #(
.DATAW(1)
) staging_substep_in (
.clk(clk),
.reset(reset),
.d(~substep_in),
.en(fedp_fire_in),
.q(substep_in)
);
VX_tensor_reg #(
.DATAW(1)
) staging_substep_out (
.clk(clk),
.reset(reset),
.d(~substep_out),
.en(fedp_fire_out),
.q(substep_out)
);
// TODO: Instead of latching half-result and constructing a full D tile,
// we should be able to send these half fragments down to commit stage
// immediately, saving flop space
assign D_frag[0][0] = D_reg[0];
assign D_frag[0][2] = D_reg[1];
assign D_frag[1][0] = D_reg[2];
assign D_frag[1][2] = D_reg[3];
assign D_frag[0][1] = D_half[0];
assign D_frag[0][3] = D_half[1];
assign D_frag[1][1] = D_half[2];
assign D_frag[1][3] = D_half[3];
wire [1:0][3:0][15:0] A_frag_fp16 = A_frag;
wire [3:0][3:0][15:0] B_frag_fp16 = B_frag;
// 4 FEDPs per threadgroup
for (genvar i = 0; i < 4; ++i) begin
// Determine which elements in the D matrix the dot-product units get
// mapped to.
//
// At substep == 0, the 0th and 2nd columns of D begins compute;
// At substep == 1, the 1st and 3rd columns of D begins compute.
// There are two row elements for each column, rounding out to
// 4 elements computed by 4 FEDPs at every cycle
// (see Figure 10(b)).
// i : 0, 1, 2, 3
// d_row : 0, 0, 1, 1
// d_col : 0, 2, 0, 2
// d_col_sel: 1, 3, 1, 3
//
// substep 0:
// [ 0 x 2 x ] (0~3 is 'i', the dpu id)
// [ 1 x 3 x ]
// substep 1:
// [ x 0 x 2 ]
// [ x 1 x 3 ]
localparam int d_row = i / 2;
localparam int d_col = (i % 2) * 2;
wire [31:0] d_col_sel = (substep_in == 1'b0) ? d_col : (d_col + 1);
// Dot product (FEDP) unit generated from Chisel
if (FP16 != 0) begin
TensorDotProductUnit fedp (
.clock (clk),
.reset (reset),
.io_in_valid (fedp_fire_in),
.io_in_bits_a_0 (A_frag[d_row][0][15: 0]),
.io_in_bits_a_1 (A_frag[d_row][0][31:16]),
.io_in_bits_a_2 (A_frag[d_row][1][15: 0]),
.io_in_bits_a_3 (A_frag[d_row][1][31:16]),
.io_in_bits_b_0 (B_frag[0][d_col_sel][15: 0]),
.io_in_bits_b_1 (B_frag[0][d_col_sel][31:16]),
.io_in_bits_b_2 (B_frag[1][d_col_sel][15: 0]),
.io_in_bits_b_3 (B_frag[1][d_col_sel][31:16]),
.io_in_bits_c (C_frag[d_row][d_col_sel]),
.io_stall (stall),
.io_out_valid (fedp_valids[i]),
.io_out_bits_data (D_half[i])
);
end else begin
TensorDotProductUnit fedp (
.clock (clk),
.reset (reset),
.io_in_valid (fedp_fire_in),
.io_in_bits_a_0 (A_frag[d_row][0]),
.io_in_bits_a_1 (A_frag[d_row][1]),
.io_in_bits_b_0 (B_frag[0][d_col_sel]),
.io_in_bits_b_1 (B_frag[1][d_col_sel]),
.io_in_bits_c (C_frag[d_row][d_col_sel]),
.io_stall (stall),
.io_out_valid (fedp_valids[i]),
.io_out_bits_data (D_half[i])
);
end
end
assign valid_out = fedp_valid_out && (substep_out == 1'b1);
endmodule
`endif