github-mirrors/vortex
1
0
Fork 0
mirror of https://github.com/vortexgpgpu/vortex.git synced 2025-04-24 05:47:35 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

minor update

Browse source
This commit is contained in:
Blaise Tine 2024-10-23 12:27:44 -07:00
parent 1fa4603fa2
commit 1c384c096d
3 changed files with 95 additions and 81 deletions
Download patch file Download diff file Expand all files Collapse all files

2
hw/rtl/core/VX_mem_unit.sv
View file

@ -127,7 +127,7 @@ module VX_mem_unit import VX_gpu_pkg::*; #(
.TAG_WIDTH (DCACHE_TAG_WIDTH)
) dcache_coalesced_if[`NUM_LSU_BLOCKS]();
if (LSU_WORD_SIZE != DCACHE_WORD_SIZE) begin : g_enabled
if ((`NUM_LSU_LANES > 1) && (LSU_WORD_SIZE != DCACHE_WORD_SIZE)) begin : g_enabled
for (genvar i = 0; i < `NUM_LSU_BLOCKS; ++i) begin : g_coalescers
VX_mem_coalescer #(

1
hw/rtl/libs/VX_mem_coalescer.sv
View file

@ -74,6 +74,7 @@ module VX_mem_coalescer #(
output wire out_rsp_ready
);
`UNUSED_SPARAM (INSTANCE_ID)
`STATIC_ASSERT ((NUM_REQS > 1), ("invalid parameter"))
`STATIC_ASSERT (`IS_DIVISBLE(NUM_REQS * DATA_IN_WIDTH, DATA_OUT_WIDTH), ("invalid parameter"))
`STATIC_ASSERT ((NUM_REQS * DATA_IN_WIDTH >= DATA_OUT_WIDTH), ("invalid parameter"))
`RUNTIME_ASSERT ((~in_req_valid || in_req_mask != 0), ("%t: invalid request mask", $time))

173
hw/rtl/libs/VX_mem_scheduler.sv
View file

@ -32,7 +32,7 @@ module VX_mem_scheduler #(
parameter WORD_WIDTH = WORD_SIZE * 8,
parameter LINE_WIDTH = LINE_SIZE * 8,
parameter COALESCE_ENABLE = (LINE_SIZE != WORD_SIZE),
parameter COALESCE_ENABLE = (CORE_REQS > 1) && (LINE_SIZE != WORD_SIZE),
parameter PER_LINE_REQS = LINE_SIZE / WORD_SIZE,
parameter MERGED_REQS = CORE_REQS / PER_LINE_REQS,
parameter MEM_BATCHES = `CDIV(MERGED_REQS, MEM_CHANNELS),
@ -94,6 +94,7 @@ module VX_mem_scheduler #(
localparam CORE_BATCHES = COALESCE_ENABLE ? 1 : MEM_BATCHES;
localparam CORE_BATCH_BITS = `CLOG2(CORE_BATCHES);
`STATIC_ASSERT ((MEM_CHANNELS <= CORE_REQS), ("invalid parameter"))
`STATIC_ASSERT (`IS_DIVISBLE(CORE_REQS * WORD_SIZE, LINE_SIZE), ("invalid parameter"))
`STATIC_ASSERT ((TAG_WIDTH >= UUID_WIDTH), ("invalid parameter"))
`RUNTIME_ASSERT((~core_req_valid || core_req_mask != 0), ("%t: invalid request mask", $time))
@ -411,99 +412,113 @@ module VX_mem_scheduler #(
// Handle memory responses ////////////////////////////////////////////////
reg [CORE_QUEUE_SIZE-1:0][CORE_REQS-1:0] rsp_rem_mask;
wire [CORE_REQS-1:0] rsp_rem_mask_n, curr_mask;
wire [BATCH_SEL_WIDTH-1:0] rsp_batch_idx;
if (CORE_BATCHES > 1) begin : g_rsp_batch_idx
assign rsp_batch_idx = mem_rsp_tag_s[CORE_BATCH_BITS-1:0];
end else begin : g_rsp_batch_idx_0
assign rsp_batch_idx = '0;
end
for (genvar r = 0; r < CORE_REQS; ++r) begin : g_curr_mask
localparam i = r / CORE_CHANNELS;
localparam j = r % CORE_CHANNELS;
assign curr_mask[r] = (BATCH_SEL_WIDTH'(i) == rsp_batch_idx) && mem_rsp_mask_s[j];
end
assign rsp_rem_mask_n = rsp_rem_mask[ibuf_raddr] & ~curr_mask;
wire rsp_complete = ~(| rsp_rem_mask_n);
wire mem_rsp_fire_s = mem_rsp_valid_s && mem_rsp_ready_s;
always @(posedge clk) begin
if (ibuf_push) begin
rsp_rem_mask[ibuf_waddr] <= core_req_mask;
end
if (mem_rsp_fire_s) begin
rsp_rem_mask[ibuf_raddr] <= rsp_rem_mask_n;
end
end
if (RSP_PARTIAL != 0 || CORE_REQS == 1) begin : g_rsp_partial
reg [CORE_QUEUE_SIZE-1:0] rsp_sop_r;
always @(posedge clk) begin
if (ibuf_push) begin
rsp_sop_r[ibuf_waddr] <= 1;
end
if (mem_rsp_fire_s) begin
rsp_sop_r[ibuf_raddr] <= 0;
end
end
if (CORE_REQS == 1) begin : g_rsp_1
assign crsp_valid = mem_rsp_valid_s;
assign crsp_mask = curr_mask;
assign crsp_sop = rsp_sop_r[ibuf_raddr];
for (genvar r = 0; r < CORE_REQS; ++r) begin : g_crsp_data
localparam j = r % CORE_CHANNELS;
assign crsp_data[r] = mem_rsp_data_s[j];
end
assign crsp_mask = mem_rsp_mask_s;
assign crsp_sop = 1'b1;
assign crsp_eop = 1'b1;
assign crsp_data = mem_rsp_data_s;
assign mem_rsp_ready_s = crsp_ready;
end else begin : g_rsp_full
end else begin : g_rsp_N
wire [CORE_CHANNELS-1:0][CORE_BATCHES-1:0][WORD_WIDTH-1:0] rsp_store_n;
reg [CORE_REQS-1:0] rsp_orig_mask [CORE_QUEUE_SIZE-1:0];
reg [CORE_QUEUE_SIZE-1:0][CORE_REQS-1:0] rsp_rem_mask;
wire [CORE_REQS-1:0] rsp_rem_mask_n, curr_mask;
wire [BATCH_SEL_WIDTH-1:0] rsp_batch_idx;
for (genvar i = 0; i < CORE_CHANNELS; ++i) begin : g_rsp_store
for (genvar j = 0; j < CORE_BATCHES; ++j) begin : g_j
reg [WORD_WIDTH-1:0] rsp_store [0:CORE_QUEUE_SIZE-1];
wire rsp_wren = mem_rsp_fire_s
&& (BATCH_SEL_WIDTH'(j) == rsp_batch_idx)
&& ((CORE_CHANNELS == 1) || mem_rsp_mask_s[i]);
always @(posedge clk) begin
if (rsp_wren) begin
rsp_store[ibuf_raddr] <= mem_rsp_data_s[i];
end
end
assign rsp_store_n[i][j] = rsp_wren ? mem_rsp_data_s[i] : rsp_store[ibuf_raddr];
end
if (CORE_BATCHES > 1) begin : g_rsp_batch_idx
assign rsp_batch_idx = mem_rsp_tag_s[CORE_BATCH_BITS-1:0];
end else begin : g_rsp_batch_idx_0
assign rsp_batch_idx = '0;
end
for (genvar r = 0; r < CORE_REQS; ++r) begin : g_curr_mask
localparam i = r / CORE_CHANNELS;
localparam j = r % CORE_CHANNELS;
assign curr_mask[r] = (BATCH_SEL_WIDTH'(i) == rsp_batch_idx) && mem_rsp_mask_s[j];
end
assign rsp_rem_mask_n = rsp_rem_mask[ibuf_raddr] & ~curr_mask;
wire mem_rsp_fire_s = mem_rsp_valid_s && mem_rsp_ready_s;
always @(posedge clk) begin
if (ibuf_push) begin
rsp_orig_mask[ibuf_waddr] <= core_req_mask;
rsp_rem_mask[ibuf_waddr] <= core_req_mask;
end
if (mem_rsp_fire_s) begin
rsp_rem_mask[ibuf_raddr] <= rsp_rem_mask_n;
end
end
assign crsp_valid = mem_rsp_valid_s && rsp_complete;
assign crsp_mask = rsp_orig_mask[ibuf_raddr];
assign crsp_sop = 1'b1;
wire rsp_complete = ~(| rsp_rem_mask_n) || (CORE_REQS == 1);
for (genvar r = 0; r < CORE_REQS; ++r) begin : g_crsp_data
localparam i = r / CORE_CHANNELS;
localparam j = r % CORE_CHANNELS;
assign crsp_data[r] = rsp_store_n[j][i];
if (RSP_PARTIAL != 0) begin : g_rsp_partial
reg [CORE_QUEUE_SIZE-1:0] rsp_sop_r;
always @(posedge clk) begin
if (ibuf_push) begin
rsp_sop_r[ibuf_waddr] <= 1;
end
if (mem_rsp_fire_s) begin
rsp_sop_r[ibuf_raddr] <= 0;
end
end
assign crsp_valid = mem_rsp_valid_s;
assign crsp_mask = curr_mask;
assign crsp_sop = rsp_sop_r[ibuf_raddr];
for (genvar r = 0; r < CORE_REQS; ++r) begin : g_crsp_data
localparam j = r % CORE_CHANNELS;
assign crsp_data[r] = mem_rsp_data_s[j];
end
assign mem_rsp_ready_s = crsp_ready;
end else begin : g_rsp_full
wire [CORE_CHANNELS-1:0][CORE_BATCHES-1:0][WORD_WIDTH-1:0] rsp_store_n;
reg [CORE_REQS-1:0] rsp_orig_mask [CORE_QUEUE_SIZE-1:0];
for (genvar i = 0; i < CORE_CHANNELS; ++i) begin : g_rsp_store
for (genvar j = 0; j < CORE_BATCHES; ++j) begin : g_j
reg [WORD_WIDTH-1:0] rsp_store [0:CORE_QUEUE_SIZE-1];
wire rsp_wren = mem_rsp_fire_s
&& (BATCH_SEL_WIDTH'(j) == rsp_batch_idx)
&& ((CORE_CHANNELS == 1) || mem_rsp_mask_s[i]);
always @(posedge clk) begin
if (rsp_wren) begin
rsp_store[ibuf_raddr] <= mem_rsp_data_s[i];
end
end
assign rsp_store_n[i][j] = rsp_wren ? mem_rsp_data_s[i] : rsp_store[ibuf_raddr];
end
end
always @(posedge clk) begin
if (ibuf_push) begin
rsp_orig_mask[ibuf_waddr] <= core_req_mask;
end
end
assign crsp_valid = mem_rsp_valid_s && rsp_complete;
assign crsp_mask = rsp_orig_mask[ibuf_raddr];
assign crsp_sop = 1'b1;
for (genvar r = 0; r < CORE_REQS; ++r) begin : g_crsp_data
localparam i = r / CORE_CHANNELS;
localparam j = r % CORE_CHANNELS;
assign crsp_data[r] = rsp_store_n[j][i];
end
assign mem_rsp_ready_s = crsp_ready || ~rsp_complete;
end
assign mem_rsp_ready_s = crsp_ready || ~rsp_complete;
assign crsp_eop = rsp_complete;
end
if (UUID_WIDTH != 0) begin : g_crsp_tag
@ -512,8 +527,6 @@ module VX_mem_scheduler #(
assign crsp_tag = ibuf_dout;
end
assign crsp_eop = rsp_complete;
// Send response to caller
VX_elastic_buffer #(
@ -525,7 +538,7 @@ module VX_mem_scheduler #(
.reset (reset),
.valid_in (crsp_valid),
.ready_in (crsp_ready),
.data_in ({crsp_mask, crsp_sop, crsp_eop, crsp_data, crsp_tag}),
.data_in ({crsp_mask, crsp_sop, crsp_eop, crsp_data, crsp_tag}),
.data_out ({core_rsp_mask, core_rsp_sop, core_rsp_eop, core_rsp_data, core_rsp_tag}),
.valid_out (core_rsp_valid),
.ready_out (core_rsp_ready)