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

pipeline refactoring

Browse Source
This commit is contained in:
Blaise Tine
2020-07-19 05:03:47 -04:00
parent 9cf8bf6149
commit 25f66e6490
71 changed files with 2242 additions and 2379 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
hw/rtl/cache/VX_bank.v vendored
View File

@@ -106,8 +106,8 @@ module VX_bank #(
`ifdef DBG_CORE_REQ_INFO
/* verilator lint_off UNUSED */
wire[31:0] debug_use_pc_st0;
wire[1:0] debug_wb_st0;
wire[4:0] debug_rd_st0;
wire[`WB_BITS-1:0] debug_wb_st0;
wire[`NR_BITS-1:0] debug_rd_st0;
wire[`NW_BITS-1:0] debug_warp_num_st0;
wire debug_rw_st0;
wire[WORD_SIZE-1:0] debug_byteen_st0;
@@ -115,8 +115,8 @@ module VX_bank #(
wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st0;
wire[31:0] debug_use_pc_st1e;
wire[1:0] debug_wb_st1e;
wire[4:0] debug_rd_st1e;
wire[`WB_BITS-1:0] debug_wb_st1e;
wire[`NR_BITS-1:0] debug_rd_st1e;
wire[`NW_BITS-1:0] debug_warp_num_st1e;
wire debug_rw_st1e;
wire[WORD_SIZE-1:0] debug_byteen_st1e;
@@ -124,8 +124,8 @@ module VX_bank #(
wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st1e;
wire[31:0] debug_use_pc_st2;
wire[1:0] debug_wb_st2;
wire[4:0] debug_rd_st2;
wire[`WB_BITS-1:0] debug_wb_st2;
wire[`NR_BITS-1:0] debug_rd_st2;
wire[`NW_BITS-1:0] debug_warp_num_st2;
wire debug_rw_st2;
wire[WORD_SIZE-1:0] debug_byteen_st2;
@@ -370,7 +370,7 @@ module VX_bank #(
.clk (clk),
.reset (reset),
.stall (stall_bank_pipe),
.flush (1'b0),
.flush (0),
.in ({qual_is_mrvq_st0, qual_is_snp_st0, qual_snp_invalidate_st0, qual_going_to_write_st0, qual_valid_st0, qual_addr_st0, qual_wsel_st0, qual_writeword_st0, qual_inst_meta_st0, qual_is_fill_st0, qual_writedata_st0}),
.out ({is_mrvq_st1[0] , is_snp_st1[0], snp_invalidate_st1[0], going_to_write_st1[0], valid_st1[0], addr_st1[0], wsel_st1[0], writeword_st1[0], inst_meta_st1[0], is_fill_st1[0], writedata_st1[0]})
);
@@ -383,7 +383,7 @@ module VX_bank #(
.clk (clk),
.reset (reset),
.stall (stall_bank_pipe),
.flush (1'b0),
.flush (0),
.in ({is_mrvq_st1[i-1], is_snp_st1[i-1], snp_invalidate_st1[i-1], going_to_write_st1[i-1], valid_st1[i-1], addr_st1[i-1], wsel_st1[i-1], writeword_st1[i-1], inst_meta_st1[i-1], is_fill_st1[i-1], writedata_st1[i-1]}),
.out ({is_mrvq_st1[i] , is_snp_st1[i], snp_invalidate_st1[i], going_to_write_st1[i], valid_st1[i], addr_st1[i], wsel_st1[i], writeword_st1[i], inst_meta_st1[i], is_fill_st1[i], writedata_st1[i]})
);
@@ -512,7 +512,7 @@ module VX_bank #(
.clk (clk),
.reset (reset),
.stall (stall_bank_pipe),
.flush (1'b0),
.flush (0),
.in ({mrvq_recover_ready_state_st1e, is_mrvq_st1e_st2, mrvq_init_ready_state_st1e , snp_to_mrvq_st1e, is_snp_st1e, snp_invalidate_st1e, fill_saw_dirty_st1e, is_fill_st1[STAGE_1_CYCLES-1] , qual_valid_st1e_2, addr_st1e, wsel_st1[STAGE_1_CYCLES-1], writeword_st1[STAGE_1_CYCLES-1], readword_st1e, readdata_st1e, readtag_st1e, miss_st1e, dirty_st1e, dirtyb_st1e, inst_meta_st1[STAGE_1_CYCLES-1]}),
.out ({mrvq_recover_ready_state_st2 , is_mrvq_st2 , mrvq_init_ready_state_unqual_st2, snp_to_mrvq_st2 , is_snp_st2 , snp_invalidate_st2, fill_saw_dirty_st2 , is_fill_st2 , valid_st2 , addr_st2 , wsel_st2, writeword_st2 , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , dirtyb_st2, inst_meta_st2 })
);
@@ -765,4 +765,4 @@ module VX_bank #(
`SCOPE_ASSIGN(scope_bank_addr_st1, `LINE_TO_BYTE_ADDR(addr_st1e, BANK_ID));
`SCOPE_ASSIGN(scope_bank_addr_st2, `LINE_TO_BYTE_ADDR(addr_st2, BANK_ID));
endmodule : VX_bank
endmodule

4
hw/rtl/cache/VX_cache.v vendored
View File

@@ -130,8 +130,8 @@ module VX_cache #(
`ifdef DBG_CORE_REQ_INFO
/* verilator lint_off UNUSED */
wire[31:0] debug_core_req_use_pc;
wire[1:0] debug_core_req_wb;
wire[4:0] debug_core_req_rd;
wire[`WB_BITS-1:0] debug_core_req_wb;
wire[`NR_BITS-1:0] debug_core_req_rd;
wire[`NW_BITS-1:0] debug_core_req_warp_num;
wire[`LOG2UP(CREQ_SIZE)-1:0] debug_core_req_idx;
/* verilator lint_on UNUSED */

58
hw/rtl/cache/VX_cache_core_rsp_merge.v vendored
View File

@@ -23,9 +23,9 @@ module VX_cache_core_rsp_merge #(
output wire [NUM_BANKS-1:0] per_bank_core_rsp_ready,
// Core Writeback
output reg [NUM_REQUESTS-1:0] core_rsp_valid,
output reg [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_rsp_data,
output reg [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
output wire [NUM_REQUESTS-1:0] core_rsp_valid,
output wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_rsp_data,
output wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
input wire core_rsp_ready
);
@@ -41,47 +41,63 @@ module VX_cache_core_rsp_merge #(
`UNUSED_PIN (grant_onehot)
);
reg [NUM_BANKS-1:0] per_bank_core_rsp_pop_unqual;
reg [NUM_REQUESTS-1:0] core_rsp_valid_unqual;
reg [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_rsp_data_unqual;
reg [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag_unqual;
reg [NUM_BANKS-1:0] core_rsp_bank_select;
assign per_bank_core_rsp_ready = per_bank_core_rsp_pop_unqual & {NUM_BANKS{core_rsp_ready}};
wire stall = ~core_rsp_ready;
integer i;
if (CORE_TAG_ID_BITS != 0) begin
assign core_rsp_tag = per_bank_core_rsp_tag[main_bank_index];
always @(*) begin
core_rsp_valid = 0;
core_rsp_data = 0;
core_rsp_valid_unqual = 0;
core_rsp_data_unqual = 0;
core_rsp_tag_unqual = per_bank_core_rsp_tag[main_bank_index];
for (i = 0; i < NUM_BANKS; i++) begin
if (per_bank_core_rsp_valid[i]
&& (per_bank_core_rsp_tag[i][CORE_TAG_ID_BITS-1:0] == per_bank_core_rsp_tag[main_bank_index][CORE_TAG_ID_BITS-1:0])) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
per_bank_core_rsp_pop_unqual[i] = 1;
core_rsp_valid_unqual[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data_unqual[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_bank_select[i] = 1;
end else begin
per_bank_core_rsp_pop_unqual[i] = 0;
core_rsp_bank_select[i] = 0;
end
end
end
end else begin
always @(*) begin
core_rsp_valid = 0;
core_rsp_data = 0;
core_rsp_tag = 0;
core_rsp_valid_unqual = 0;
core_rsp_data_unqual = 0;
core_rsp_tag_unqual = 0;
for (i = 0; i < NUM_BANKS; i++) begin
if (per_bank_core_rsp_valid[i]
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
&& !core_rsp_valid_unqual[per_bank_core_rsp_tid[i]]
&& ((main_bank_index == `BANK_BITS'(i))
|| (per_bank_core_rsp_tid[i] != per_bank_core_rsp_tid[main_bank_index]))) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_tag[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_tag[i];
per_bank_core_rsp_pop_unqual[i] = 1;
core_rsp_valid_unqual[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data_unqual[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_tag_unqual[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_tag[i];
core_rsp_bank_select[i] = 1;
end else begin
per_bank_core_rsp_pop_unqual[i] = 0;
core_rsp_bank_select[i] = 0;
end
end
end
end
VX_generic_register #(
.N(NUM_REQUESTS + (NUM_REQUESTS *`WORD_WIDTH) + (`CORE_REQ_TAG_COUNT * CORE_TAG_WIDTH))
) core_wb_reg (
.clk (clk),
.reset (reset),
.stall (stall),
.flush (0),
.in ({core_rsp_valid_unqual, core_rsp_data_unqual, core_rsp_tag_unqual}),
.out ({core_rsp_valid, core_rsp_data, core_rsp_tag})
);
assign per_bank_core_rsp_ready = core_rsp_bank_select & {NUM_BANKS{~stall}};
endmodule

2
hw/rtl/cache/VX_snp_forwarder.v vendored
View File

@@ -59,7 +59,7 @@ module VX_snp_forwarder #(
assign sfq_push = snp_req_valid && !sfq_full && fwdout_ready;
assign sfq_pop = snp_rsp_valid;
VX_indexable_queue #(
VX_index_queue #(
.DATAW (`LOG2UP(SNRQ_SIZE) + 1 +`DRAM_ADDR_WIDTH+SNP_REQ_TAG_WIDTH),
.SIZE (SNRQ_SIZE)
) snp_fwd_queue (

4
hw/rtl/cache/VX_tag_data_access.v vendored
View File

@@ -116,7 +116,7 @@ module VX_tag_data_access #(
.clk (clk),
.reset (reset),
.stall (stall),
.flush (1'b0),
.flush (0),
.in ({qual_read_valid_st1, qual_read_dirty_st1, qual_read_dirtyb_st1, qual_read_tag_st1, qual_read_data_st1}),
.out ({read_valid_st1c[0], read_dirty_st1c[0], read_dirtyb_st1c[0], read_tag_st1c[0], read_data_st1c[0]})
);
@@ -129,7 +129,7 @@ module VX_tag_data_access #(
.clk (clk),
.reset (reset),
.stall (stall),
.flush (1'b0),
.flush (0),
.in ({read_valid_st1c[i-1], read_dirty_st1c[i-1], read_dirtyb_st1c[i-1], read_tag_st1c[i-1], read_data_st1c[i-1]}),
.out ({read_valid_st1c[i], read_dirty_st1c[i], read_dirtyb_st1c[i], read_tag_st1c[i], read_data_st1c[i]})
);