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

non-cacheable memory address critical paths optimizations

Browse Source
This commit is contained in:
Blaise Tine
2021-06-10 12:47:18 -07:00
parent 41069ba188
commit adf033b0aa
11 changed files with 258 additions and 194 deletions
Download Patch File Download Diff File Expand all files Collapse all files

191
hw/rtl/cache/VX_nc_bypass.v vendored
View File

@@ -97,8 +97,7 @@ module VX_nc_bypass #(
reg [NUM_REQS-1:0] core_req_ready_in_r;
wire [NUM_REQS-1:0] core_req_valid_in_nc;
wire [CORE_REQ_TIDW-1:0] core_req_nc_tid;
for (genvar i = 0; i < NUM_REQS; ++i) begin
assign core_req_valid_in_nc[i] = core_req_valid_in[i] && core_req_tag_in[i][NC_TAG_BIT];
end
@@ -107,14 +106,45 @@ module VX_nc_bypass #(
for (integer i = 0; i < NUM_REQS; ++i) begin
if (core_req_valid_in_nc[i]) begin
core_req_valid_out_r[i] = 0;
core_req_ready_in_r[i] = mem_req_ready_out && (core_req_nc_tid == CORE_REQ_TIDW'(i));
end else begin
core_req_valid_out_r[i] = core_req_valid_in[i];
core_req_ready_in_r[i] = core_req_ready_out[i];
end
end
end
wire [`UP(CORE_REQ_TIDW)-1:0] core_req_nc_tid;
wire core_req_nc_valid;
VX_priority_encoder #(
.N (NUM_REQS)
) core_req_sel (
.data_in (core_req_valid_in_nc),
.index (core_req_nc_tid),
`UNUSED_PIN (onehot),
.valid_out (core_req_nc_valid)
);
if (NUM_REQS > 1) begin
always @(*) begin
for (integer i = 0; i < NUM_REQS; ++i) begin
if (core_req_valid_in_nc[i]) begin
core_req_ready_in_r[i] = mem_req_ready_out && (core_req_nc_tid == CORE_REQ_TIDW'(i));
end else begin
core_req_ready_in_r[i] = core_req_ready_out[i];
end
end
end
end else begin
`UNUSED_VAR (core_req_nc_tid)
always @(*) begin
if (core_req_valid_in_nc) begin
core_req_ready_in_r = mem_req_ready_out;
end else begin
core_req_ready_in_r = core_req_ready_out;
end
end
end
assign core_req_valid_out = core_req_valid_out_r;
assign core_req_rw_out = core_req_rw_in;
assign core_req_addr_out = core_req_addr_in;
@@ -131,57 +161,92 @@ module VX_nc_bypass #(
reg [MEM_ADDR_WIDTH-1:0] mem_req_addr_out_r;
reg [MEM_DATA_WIDTH-1:0] mem_req_data_out_r;
reg [MEM_TAG_WIDTH-1:0] mem_req_tag_out_r;
reg mem_req_ready_in_r;
wire core_req_nc_valid;
VX_priority_encoder #(
.N (NUM_REQS)
) core_req_sel (
.data_in (core_req_valid_in_nc),
.index (core_req_nc_tid),
`UNUSED_PIN (onehot),
.valid_out (core_req_nc_valid)
);
reg mem_req_ready_in_r;
always @(*) begin
if (core_req_nc_valid) begin
mem_req_valid_out_r = 1;
mem_req_rw_out_r = core_req_rw_in[core_req_nc_tid];
mem_req_addr_out_r = core_req_addr_in[core_req_nc_tid][D +: MEM_ADDR_WIDTH];
for (integer i = 0; i < P; ++i) begin
mem_req_data_out_r[i * CORE_DATA_WIDTH +: CORE_DATA_WIDTH] = core_req_data_in[core_req_nc_tid];
end
mem_req_ready_in_r = 0;
mem_req_valid_out_r = 1;
mem_req_ready_in_r = 0;
end else begin
mem_req_valid_out_r = mem_req_valid_in;
mem_req_rw_out_r = mem_req_rw_in;
mem_req_addr_out_r = mem_req_addr_in;
mem_req_data_out_r = mem_req_data_in;
mem_req_ready_in_r = mem_req_ready_out;
mem_req_valid_out_r = mem_req_valid_in;
mem_req_ready_in_r = mem_req_ready_out;
end
end
if (D != 0) begin
wire [D-1:0] req_addr_idx = core_req_addr_in[core_req_nc_tid][D-1:0];
if (NUM_REQS > 1) begin
always @(*) begin
if (core_req_nc_valid) begin
mem_req_byteen_out_r = 0;
mem_req_byteen_out_r[req_addr_idx * CORE_DATA_SIZE +: CORE_DATA_SIZE] = core_req_byteen_in[core_req_nc_tid];
mem_req_tag_out_r = MEM_TAG_WIDTH'({core_req_nc_tid, req_addr_idx, core_req_tag_in[core_req_nc_tid]});
mem_req_rw_out_r = core_req_rw_in[core_req_nc_tid];
mem_req_addr_out_r = core_req_addr_in[core_req_nc_tid][D +: MEM_ADDR_WIDTH];
for (integer i = 0; i < P; ++i) begin
mem_req_data_out_r[i * CORE_DATA_WIDTH +: CORE_DATA_WIDTH] = core_req_data_in[core_req_nc_tid];
end
end else begin
mem_req_byteen_out_r = mem_req_byteen_in;
mem_req_tag_out_r = mem_req_tag_in;
mem_req_rw_out_r = mem_req_rw_in;
mem_req_addr_out_r = mem_req_addr_in;
mem_req_data_out_r = mem_req_data_in;
end
end
end else begin
if (D != 0) begin
wire [D-1:0] req_addr_idx = core_req_addr_in[core_req_nc_tid][D-1:0];
always @(*) begin
if (core_req_nc_valid) begin
mem_req_byteen_out_r = 0;
mem_req_byteen_out_r[req_addr_idx * CORE_DATA_SIZE +: CORE_DATA_SIZE] = core_req_byteen_in[core_req_nc_tid];
mem_req_tag_out_r = MEM_TAG_WIDTH'({core_req_nc_tid, req_addr_idx, core_req_tag_in[core_req_nc_tid]});
end else begin
mem_req_byteen_out_r = mem_req_byteen_in;
mem_req_tag_out_r = mem_req_tag_in;
end
end
end else begin
always @(*) begin
if (core_req_nc_valid) begin
mem_req_byteen_out_r = core_req_byteen_in[core_req_nc_tid];
mem_req_tag_out_r = MEM_TAG_WIDTH'({core_req_nc_tid, core_req_tag_in[core_req_nc_tid]});
end else begin
mem_req_byteen_out_r = mem_req_byteen_in;
mem_req_tag_out_r = mem_req_tag_in;
end
end
end
end else begin
always @(*) begin
if (core_req_nc_valid) begin
mem_req_byteen_out_r = core_req_byteen_in[core_req_nc_tid];
mem_req_tag_out_r = MEM_TAG_WIDTH'({core_req_nc_tid, core_req_tag_in[core_req_nc_tid]});
mem_req_rw_out_r = core_req_rw_in;
mem_req_addr_out_r = core_req_addr_in[0][D +: MEM_ADDR_WIDTH];
for (integer i = 0; i < P; ++i) begin
mem_req_data_out_r[i * CORE_DATA_WIDTH +: CORE_DATA_WIDTH] = core_req_data_in;
end
end else begin
mem_req_byteen_out_r = mem_req_byteen_in;
mem_req_tag_out_r = mem_req_tag_in;
mem_req_rw_out_r = mem_req_rw_in;
mem_req_addr_out_r = mem_req_addr_in;
mem_req_data_out_r = mem_req_data_in;
end
end
if (D != 0) begin
wire [D-1:0] req_addr_idx = core_req_addr_in[0][D-1:0];
always @(*) begin
if (core_req_nc_valid) begin
mem_req_byteen_out_r = 0;
mem_req_byteen_out_r[req_addr_idx * CORE_DATA_SIZE +: CORE_DATA_SIZE] = core_req_byteen_in;
mem_req_tag_out_r = MEM_TAG_WIDTH'({req_addr_idx, core_req_tag_in});
end else begin
mem_req_byteen_out_r = mem_req_byteen_in;
mem_req_tag_out_r = mem_req_tag_in;
end
end
end else begin
always @(*) begin
if (core_req_nc_valid) begin
mem_req_byteen_out_r = core_req_byteen_in;
mem_req_tag_out_r = MEM_TAG_WIDTH'(core_req_tag_in);
end else begin
mem_req_byteen_out_r = mem_req_byteen_in;
mem_req_tag_out_r = mem_req_tag_in;
end
end
end
end
@@ -201,26 +266,41 @@ module VX_nc_bypass #(
reg [NUM_RSP_TAGS-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag_out_r;
reg [NUM_RSP_TAGS-1:0] core_rsp_ready_in_r;
wire [CORE_REQ_TIDW-1:0] rsp_tid = mem_rsp_tag_in[(CORE_TAG_WIDTH + D) +: CORE_REQ_TIDW];
wire is_mem_rsp_nc = mem_rsp_valid_in && mem_rsp_tag_in[NC_TAG_BIT];
if (NUM_REQS > 1) begin
always @(*) begin
if (is_mem_rsp_nc) begin
core_rsp_valid_out_r = 0;
core_rsp_valid_out_r[rsp_tid] = 1;
for (integer i = 0; i < NUM_RSP_TAGS; ++i) begin
core_rsp_tag_out_r[i] = mem_rsp_tag_in[CORE_TAG_WIDTH-1:0];
end
core_rsp_ready_in_r = 0;
end else begin
core_rsp_valid_out_r = core_rsp_valid_in;
core_rsp_tag_out_r = core_rsp_tag_in;
core_rsp_ready_in_r = core_rsp_ready_out;
wire [CORE_REQ_TIDW-1:0] rsp_tid = mem_rsp_tag_in[(CORE_TAG_WIDTH + D) +: CORE_REQ_TIDW];
if (NUM_RSP_TAGS > 1) begin
always @(*) begin
for (integer i = 0; i < NUM_REQS; ++i) begin
if (is_mem_rsp_nc && (rsp_tid == CORE_REQ_TIDW'(i))) begin
core_rsp_valid_out_r[i] = 1;
core_rsp_tag_out_r[i] = mem_rsp_tag_in[CORE_TAG_WIDTH-1:0];
core_rsp_ready_in_r[i] = 0;
end else begin
core_rsp_valid_out_r[i] = core_rsp_valid_in[i];
core_rsp_tag_out_r[i] = core_rsp_tag_in[i];
core_rsp_ready_in_r[i] = core_rsp_ready_out[i];
end
end
end
end else begin
always @(*) begin
if (is_mem_rsp_nc) begin
core_rsp_valid_out_r = 0;
core_rsp_valid_out_r[rsp_tid] = 1;
for (integer i = 0; i < NUM_RSP_TAGS; ++i) begin
core_rsp_tag_out_r[i] = mem_rsp_tag_in[CORE_TAG_WIDTH-1:0];
end
core_rsp_ready_in_r = 0;
end else begin
core_rsp_valid_out_r = core_rsp_valid_in;
core_rsp_tag_out_r = core_rsp_tag_in;
core_rsp_ready_in_r = core_rsp_ready_out;
end
end
end
end else begin
end else begin
always @(*) begin
if (is_mem_rsp_nc) begin
core_rsp_valid_out_r = 1;
@@ -276,6 +356,7 @@ module VX_nc_bypass #(
end
if (NUM_RSP_TAGS > 1) begin
wire [CORE_REQ_TIDW-1:0] rsp_tid = mem_rsp_tag_in[(CORE_TAG_WIDTH + D) +: CORE_REQ_TIDW];
always @(*) begin
if (is_mem_rsp_nc) begin
mem_rsp_ready_in_r = core_rsp_ready_out[rsp_tid];