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

Fixed Stall Pipeline Logic

Browse source
This commit is contained in:
felsabbagh3 2020-03-09 22:08:46 -07:00
parent 469334f23e
commit dea271eb6b
5 changed files with 88 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

18
rtl/VX_cache/VX_bank.v
View file

@ -37,6 +37,8 @@ module VX_bank
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Forward SNP Queue
parameter FFSQ_SIZE = 8,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
@ -95,7 +97,10 @@ module VX_bank
// Snp Request
input wire snp_req,
input wire[31:0] snp_req_addr
input wire[31:0] snp_req_addr,
output wire snp_fwd,
output wire[31:0] snp_fwd_addr
);
@ -511,7 +516,7 @@ module VX_bank
// Enqueue to miss reserv if it's a valid miss
assign miss_add = valid_st2 && miss_st2 && !mrvq_full && !((cwbq_push && cwbq_full) || (dwbq_push && dwbq_full) || (dram_fill_req && dram_fill_req_queue_full));
assign miss_add = valid_st2 && miss_st2 && !mrvq_full && !(((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full));
assign miss_add_pc = pc_st2;
assign miss_add_addr = addr_st2;
assign miss_add_data = writeword_st2;
@ -519,7 +524,7 @@ module VX_bank
// Enqueue to CWB Queue
wire cwbq_push = (valid_st2 && !miss_st2) && !cwbq_full && !((FUNC_ID == `LLFUNC_ID) && (miss_add_wb == 0));
wire cwbq_push = (valid_st2 && !miss_st2) && !cwbq_full && !((FUNC_ID == `LLFUNC_ID) && (miss_add_wb == 0)) && !( (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full));
wire [`WORD_SIZE_RNG] cwbq_data = readword_st2;
wire [`vx_clog2(NUMBER_REQUESTS)-1:0] cwbq_tid = miss_add_tid;
wire [4:0] cwbq_rd = miss_add_rd;
@ -544,7 +549,7 @@ module VX_bank
);
// Enqueue to DWB Queue
wire dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && !dwbq_full && !(!fill_saw_dirty_st2 && mrvq_full);
wire dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && !dwbq_full && !(((valid_st2 && !miss_st2) && cwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full));
wire[31:0] dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_ADDR_END:0]} & `BASE_ADDR_MASK;
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dwbq_req_data = readdata_st2;
wire dwbq_empty;
@ -603,9 +608,12 @@ module VX_bank
.full (dwbq_full)
);
wire snp_fwd_push;
wire snp_fwd_pop;
assign stall_bank_pipe = (cwbq_push && cwbq_full) || (dwbq_push && dwbq_full) || (miss_add && mrvq_full) || (dram_fill_req && dram_fill_req_queue_full);
assign stall_bank_pipe = ((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full);
endmodule

3
rtl/VX_cache/VX_cache.v
View file

@ -37,6 +37,8 @@ module VX_cache
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Forward SNP Queue
parameter FFSQ_SIZE = 8,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
@ -343,6 +345,7 @@ module VX_cache
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FFSQ_SIZE (FFSQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)

71
rtl/VX_define.v
View file

@ -125,7 +125,11 @@
-199
`define NUMBER_CORES 2
`define NUMBER_CORES_PER_CLUSTERS (2)
`define NUMBER_CLUSTERS (1)
`define NUMBER_CORES (`NUMBER_CORES_PER_CLUSTERS*`NUMBER_CLUSTERS)
// `define SINGLE_CORE_BENCH 0
`define GLOBAL_BLOCK_SIZE_BYTES 16
// ========================================= Dcache Configurable Knobs =========================================
@ -169,6 +173,8 @@
`define DDFQQ_SIZE `DREQQ_SIZE
// Lower Level Cache Hit Queue Size
`define DLLVQ_SIZE 0
// Fill Forward SNP Queue
`define DFFSQ_SIZE 8
// Fill Invalidator Size {Fill invalidator must be active}
`define DFILL_INVALIDAOR_SIZE 16
@ -220,6 +226,8 @@
`define IDFQQ_SIZE `IREQQ_SIZE
// Lower Level Cache Hit Queue Size
`define ILLVQ_SIZE 0
// Fill Forward SNP Queue
`define IFFSQ_SIZE 8
// Fill Invalidator Size {Fill invalidator must be active}
`define IFILL_INVALIDAOR_SIZE 16
@ -270,6 +278,8 @@
`define SDFQQ_SIZE 0
// Lower Level Cache Hit Queue Size
`define SLLVQ_SIZE 0
// Fill Forward SNP Queue
`define SFFSQ_SIZE 0
// Fill Invalidator Size {Fill invalidator must be active}
`define SFILL_INVALIDAOR_SIZE 16
@ -293,7 +303,7 @@
// Size of a word in bytes
`define LLWORD_SIZE_BYTES (`LLBANK_LINE_SIZE_BYTES)
// Number of Word requests per cycle {1, 2, 4, 8, ...}
`define LLNUMBER_REQUESTS (2*`NUMBER_CORES)
`define LLNUMBER_REQUESTS (2*`NUMBER_CORES_PER_CLUSTERS)
// Number of cycles to complete stage 1 (read from memory)
`define LLSTAGE_1_CYCLES 2
// Function ID
@ -305,7 +315,7 @@
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
`define LLREQQ_SIZE (`NT*`NW*`NUMBER_CORES)
`define LLREQQ_SIZE (`NT*`NW*`NUMBER_CORES_PER_CLUSTERS)
// Miss Reserv Queue Knob
`define LLMRVQ_SIZE `LLREQQ_SIZE
// Dram Fill Rsp Queue Size
@ -322,6 +332,8 @@
`define LLDFQQ_SIZE `LLREQQ_SIZE
// Lower Level Cache Hit Queue Size
`define LLLLVQ_SIZE 0
// Fill Forward SNP Queue
`define LLFFSQ_SIZE 8
// Fill Invalidator Size {Fill invalidator must be active}
`define LLFILL_INVALIDAOR_SIZE 16
@ -332,4 +344,57 @@
// ========================================= L2cache Configurable Knobs =========================================
// ========================================= L3cache Configurable Knobs =========================================
// General Cache Knobs
// Size of cache in bytes
`define L3CACHE_SIZE_BYTES 1024
// Size of line inside a bank in bytes
`define L3BANK_LINE_SIZE_BYTES `GLOBAL_BLOCK_SIZE_BYTES
// Number of banks {1, 2, 4, 8,...}
`define L3NUMBER_BANKS 8
// Size of a word in bytes
`define L3WORD_SIZE_BYTES (`LLBANK_LINE_SIZE_BYTES)
// Number of Word requests per cycle {1, 2, 4, 8, ...}
`define L3NUMBER_REQUESTS (2*`NUMBER_CLUSTERS)
// Number of cycles to complete stage 1 (read from memory)
`define L3STAGE_1_CYCLES 2
// Function ID
`define L3FUNC_ID 3
// Bank Number of words in a line
`define L3BANK_LINE_SIZE_WORDS (`LLBANK_LINE_SIZE_BYTES / `LLWORD_SIZE_BYTES)
`define L3BANK_LINE_SIZE_RNG `LLBANK_LINE_SIZE_WORDS-1:0
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
`define L3REQQ_SIZE (`NT*`NW*`NUMBER_CLUSTERS)
// Miss Reserv Queue Knob
`define L3MRVQ_SIZE `LLREQQ_SIZE
// Dram Fill Rsp Queue Size
`define L3DFPQ_SIZE 2
// Snoop Req Queue
`define L3SNRQ_SIZE 8
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
`define L3CWBQ_SIZE `LLREQQ_SIZE
// Dram Writeback Queue Size
`define L3DWBQ_SIZE 4
// Dram Fill Req Queue Size
`define L3DFQQ_SIZE `LLREQQ_SIZE
// Lower Level Cache Hit Queue Size
`define L3LLVQ_SIZE 0
// Fill Forward SNP Queue
`define L3FFSQ_SIZE 8
// Fill Invalidator Size {Fill invalidator must be active}
`define L3FILL_INVALIDAOR_SIZE 16
// Dram knobs
`define L3SIMULATED_DRAM_LATENCY_CYCLES 10
// ========================================= L3cache Configurable Knobs =========================================
`endif

3
rtl/VX_dmem_controller.v
View file

@ -92,6 +92,7 @@ module VX_dmem_controller (
.DWBQ_SIZE (`SDWBQ_SIZE),
.DFQQ_SIZE (`SDFQQ_SIZE),
.LLVQ_SIZE (`SLLVQ_SIZE),
.FFSQ_SIZE (`SFFSQ_SIZE),
.FILL_INVALIDAOR_SIZE (`SFILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(`SSIMULATED_DRAM_LATENCY_CYCLES)
)
@ -167,6 +168,7 @@ module VX_dmem_controller (
.DWBQ_SIZE (`DDWBQ_SIZE),
.DFQQ_SIZE (`DDFQQ_SIZE),
.LLVQ_SIZE (`DLLVQ_SIZE),
.FFSQ_SIZE (`DFFSQ_SIZE),
.FILL_INVALIDAOR_SIZE (`DFILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(`DSIMULATED_DRAM_LATENCY_CYCLES)
)
@ -244,6 +246,7 @@ module VX_dmem_controller (
.DWBQ_SIZE (`IDWBQ_SIZE),
.DFQQ_SIZE (`IDFQQ_SIZE),
.LLVQ_SIZE (`ILLVQ_SIZE),
.FFSQ_SIZE (`IFFSQ_SIZE),
.FILL_INVALIDAOR_SIZE (`IFILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(`ISIMULATED_DRAM_LATENCY_CYCLES)
)

1
rtl/Vortex_SOC.v
View file

@ -216,6 +216,7 @@ module Vortex_SOC (
.DWBQ_SIZE (`LLDWBQ_SIZE),
.DFQQ_SIZE (`LLDFQQ_SIZE),
.LLVQ_SIZE (`LLLLVQ_SIZE),
.FFSQ_SIZE (`LLFFSQ_SIZE),
.FILL_INVALIDAOR_SIZE (`LLFILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(`LLSIMULATED_DRAM_LATENCY_CYCLES)
)