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

text_unit merge fixes

Browse Source
This commit is contained in:
Blaise Tine
2021-10-19 00:16:22 -04:00
parent 0b23d8e935
commit 6edf38548f
7 changed files with 21 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
hw/rtl/VX_define.vh
View File

@@ -242,8 +242,11 @@
`define DBG_CACHE_REQ_MDATAW 0 `define DBG_CACHE_REQ_MDATAW 0
`endif `endif
// non-cacheable address bit // non-cacheable tag bits
`define NC_FLAG_BITS 1 `define NC_TAG_BIT 1
// texture tag bits
`define TEX_TAG_BIT 1
////////////////////////// Icache Configurable Knobs ////////////////////////// ////////////////////////// Icache Configurable Knobs //////////////////////////
@@ -285,15 +288,15 @@
// Core request tag bits // Core request tag bits
`define LSUQ_ADDR_BITS `LOG2UP(`LSUQ_SIZE) `define LSUQ_ADDR_BITS `LOG2UP(`LSUQ_SIZE)
`ifdef EXT_TEX_ENABLE `ifdef EXT_TEX_ENABLE
`define LSU_TAG_ID_BITS (`LSUQ_ADDR_BITS + `NC_FLAG_BITS + `SM_ENABLE) `define LSU_TAG_ID_BITS (`LSUQ_ADDR_BITS + `NC_TAG_BIT + `SM_ENABLE)
`define TEX_TAG_ID_BITS (2) `define TEX_TAG_ID_BITS (2)
`define LSU_TEX_TAG_ID_BITS `MAX(`LSU_TAG_ID_BITS, `TEX_TAG_ID_BITS) `define LSU_TEX_TAG_ID_BITS `MAX(`LSU_TAG_ID_BITS, `TEX_TAG_ID_BITS)
`define DCACHE_CORE_TAG_ID_BITS (`LSU_TEX_TAG_ID_BITS + `NC_FLAG_BITS) `define DCACHE_CORE_TAG_ID_BITS (`LSU_TEX_TAG_ID_BITS + `TEX_TAG_BIT)
`define LSU_DCACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + `LSU_TAG_ID_BITS) `define LSU_DCACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + `LSU_TAG_ID_BITS)
`define TEX_DCACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + `TEX_TAG_ID_BITS) `define TEX_DCACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + `TEX_TAG_ID_BITS)
`define LSU_TEX_DCACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + `LSU_TEX_TAG_ID_BITS) `define LSU_TEX_DCACHE_TAG_BITS (`DBG_CACHE_REQ_MDATAW + `LSU_TEX_TAG_ID_BITS)
`else `else
`define DCACHE_CORE_TAG_ID_BITS (`LSUQ_ADDR_BITS + `NC_FLAG_BITS + `SM_ENABLE) `define DCACHE_CORE_TAG_ID_BITS (`LSUQ_ADDR_BITS + `NC_TAG_BIT + `SM_ENABLE)
`endif `endif
`define DCACHE_CORE_TAG_WIDTH (`DBG_CACHE_REQ_MDATAW + `DCACHE_CORE_TAG_ID_BITS) `define DCACHE_CORE_TAG_WIDTH (`DBG_CACHE_REQ_MDATAW + `DCACHE_CORE_TAG_ID_BITS)
@@ -312,7 +315,7 @@
// Memory request tag bits // Memory request tag bits
`define _DMEM_ADDR_RATIO_W $clog2(`DCACHE_LINE_SIZE / `DCACHE_WORD_SIZE) `define _DMEM_ADDR_RATIO_W $clog2(`DCACHE_LINE_SIZE / `DCACHE_WORD_SIZE)
`define _DNC_MEM_TAG_WIDTH ($clog2(`DCACHE_NUM_REQS) + `_DMEM_ADDR_RATIO_W + `DCACHE_CORE_TAG_WIDTH) `define _DNC_MEM_TAG_WIDTH ($clog2(`DCACHE_NUM_REQS) + `_DMEM_ADDR_RATIO_W + `DCACHE_CORE_TAG_WIDTH)
`define DCACHE_MEM_TAG_WIDTH `MAX((`CLOG2(`DCACHE_NUM_BANKS) + `CLOG2(`DCACHE_MSHR_SIZE) + `NC_FLAG_BITS), `_DNC_MEM_TAG_WIDTH) `define DCACHE_MEM_TAG_WIDTH `MAX((`CLOG2(`DCACHE_NUM_BANKS) + `CLOG2(`DCACHE_MSHR_SIZE) + `NC_TAG_BIT), `_DNC_MEM_TAG_WIDTH)
// Merged D-cache/I-cache memory tag // Merged D-cache/I-cache memory tag
`define L1_MEM_TAG_WIDTH (`MAX(`ICACHE_MEM_TAG_WIDTH, `DCACHE_MEM_TAG_WIDTH) + `CLOG2(2)) `define L1_MEM_TAG_WIDTH (`MAX(`ICACHE_MEM_TAG_WIDTH, `DCACHE_MEM_TAG_WIDTH) + `CLOG2(2))
@@ -360,7 +363,7 @@
// Memory request tag bits // Memory request tag bits
`define _L2_MEM_ADDR_RATIO_W $clog2(`L2_CACHE_LINE_SIZE / `L2_WORD_SIZE) `define _L2_MEM_ADDR_RATIO_W $clog2(`L2_CACHE_LINE_SIZE / `L2_WORD_SIZE)
`define _L2_NC_MEM_TAG_WIDTH ($clog2(`L2_NUM_REQS) + `_L2_MEM_ADDR_RATIO_W + `L1_MEM_TAG_WIDTH) `define _L2_NC_MEM_TAG_WIDTH ($clog2(`L2_NUM_REQS) + `_L2_MEM_ADDR_RATIO_W + `L1_MEM_TAG_WIDTH)
`define _L2_MEM_TAG_WIDTH `MAX((`CLOG2(`L2_NUM_BANKS) + `CLOG2(`L2_MSHR_SIZE) + `NC_FLAG_BITS), `_L2_NC_MEM_TAG_WIDTH) `define _L2_MEM_TAG_WIDTH `MAX((`CLOG2(`L2_NUM_BANKS) + `CLOG2(`L2_MSHR_SIZE) + `NC_TAG_BIT), `_L2_NC_MEM_TAG_WIDTH)
`define L2_MEM_TAG_WIDTH ((`L2_ENABLE) ? `_L2_MEM_TAG_WIDTH : (`L1_MEM_TAG_WIDTH + `CLOG2(`L2_NUM_REQS))) `define L2_MEM_TAG_WIDTH ((`L2_ENABLE) ? `_L2_MEM_TAG_WIDTH : (`L1_MEM_TAG_WIDTH + `CLOG2(`L2_NUM_REQS)))
////////////////////////// L3cache Configurable Knobs ///////////////////////// ////////////////////////// L3cache Configurable Knobs /////////////////////////
@@ -392,7 +395,7 @@
// Memory request tag bits // Memory request tag bits
`define _L3_MEM_ADDR_RATIO_W $clog2(`L3_CACHE_LINE_SIZE / `L3_WORD_SIZE) `define _L3_MEM_ADDR_RATIO_W $clog2(`L3_CACHE_LINE_SIZE / `L3_WORD_SIZE)
`define _L3_NC_MEM_TAG_WIDTH ($clog2(`L3_NUM_REQS) + `_L3_MEM_ADDR_RATIO_W + `L2_MEM_TAG_WIDTH) `define _L3_NC_MEM_TAG_WIDTH ($clog2(`L3_NUM_REQS) + `_L3_MEM_ADDR_RATIO_W + `L2_MEM_TAG_WIDTH)
`define _L3_MEM_TAG_WIDTH `MAX((`CLOG2(`L3_NUM_BANKS) + `CLOG2(`L3_MSHR_SIZE) + `NC_FLAG_BITS), `_L3_NC_MEM_TAG_WIDTH) `define _L3_MEM_TAG_WIDTH `MAX((`CLOG2(`L3_NUM_BANKS) + `CLOG2(`L3_MSHR_SIZE) + `NC_TAG_BIT), `_L3_NC_MEM_TAG_WIDTH)
`define L3_MEM_TAG_WIDTH ((`L3_ENABLE) ? `_L3_MEM_TAG_WIDTH : (`L2_MEM_TAG_WIDTH + `CLOG2(`L3_NUM_REQS))) `define L3_MEM_TAG_WIDTH ((`L3_ENABLE) ? `_L3_MEM_TAG_WIDTH : (`L2_MEM_TAG_WIDTH + `CLOG2(`L3_NUM_REQS)))
/////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////

4
hw/rtl/VX_execute.sv
View File

@@ -102,7 +102,7 @@ module VX_execute #(
.LANES (`NUM_THREADS), .LANES (`NUM_THREADS),
.DATA_SIZE (4), .DATA_SIZE (4),
.TAG_IN_WIDTH (`LSU_TEX_DCACHE_TAG_BITS), .TAG_IN_WIDTH (`LSU_TEX_DCACHE_TAG_BITS),
.TAG_SEL_IDX (`NC_FLAG_BITS + `SM_ENABLE) .TAG_SEL_IDX (`NC_TAG_BIT + `SM_ENABLE)
) tex_lsu_arb ( ) tex_lsu_arb (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -142,7 +142,7 @@ module VX_execute #(
`endif `endif
`ifdef EXT_TEX_ENABLE `ifdef EXT_F_ENABLE
wire [`NUM_WARPS-1:0] csr_pending; wire [`NUM_WARPS-1:0] csr_pending;
wire [`NUM_WARPS-1:0] fpu_pending; wire [`NUM_WARPS-1:0] fpu_pending;
VX_fpu_to_csr_if fpu_to_csr_if(); VX_fpu_to_csr_if fpu_to_csr_if();

7
hw/rtl/VX_gpu_unit.sv
View File

@@ -12,10 +12,9 @@ module VX_gpu_unit #(
VX_gpu_req_if.slave gpu_req_if, VX_gpu_req_if.slave gpu_req_if,
`ifdef EXT_TEX_ENABLE `ifdef EXT_TEX_ENABLE
VX_tex_csr_if tex_csr_if, VX_dcache_req_if.master dcache_req_if,
VX_dcache_rsp_if.slave dcache_rsp_if,
VX_dcache_req_if dcache_req_if, VX_tex_csr_if.slave tex_csr_if,
VX_dcache_rsp_if dcache_rsp_if,
`endif `endif
// Outputs // Outputs

2
hw/rtl/VX_lsu_unit.sv
View File

@@ -24,7 +24,7 @@ module VX_lsu_unit #(
localparam REQ_ASHIFT = `CLOG2(`DCACHE_WORD_SIZE); localparam REQ_ASHIFT = `CLOG2(`DCACHE_WORD_SIZE);
localparam ADDR_TYPEW = `NC_FLAG_BITS + `SM_ENABLE; localparam ADDR_TYPEW = `NC_TAG_BIT + `SM_ENABLE;
`STATIC_ASSERT(0 == (`IO_BASE_ADDR % MEM_ASHIFT), ("invalid parameter")) `STATIC_ASSERT(0 == (`IO_BASE_ADDR % MEM_ASHIFT), ("invalid parameter"))
`STATIC_ASSERT(0 == (`SMEM_BASE_ADDR % MEM_ASHIFT), ("invalid parameter")) `STATIC_ASSERT(0 == (`SMEM_BASE_ADDR % MEM_ASHIFT), ("invalid parameter"))

1
hw/rtl/VX_mem_unit.sv
View File

@@ -206,6 +206,7 @@ module VX_mem_unit # (
.LANES (`NUM_THREADS), .LANES (`NUM_THREADS),
.DATA_SIZE (4), .DATA_SIZE (4),
.TAG_IN_WIDTH (`DCACHE_CORE_TAG_WIDTH), .TAG_IN_WIDTH (`DCACHE_CORE_TAG_WIDTH),
.TAG_SEL_IDX (0), // SM flag
.TYPE ("P"), .TYPE ("P"),
.BUFFERED_REQ (2), .BUFFERED_REQ (2),
.BUFFERED_RSP (1) .BUFFERED_RSP (1)

2
hw/rtl/cache/VX_shared_mem.sv vendored
View File

@@ -229,7 +229,7 @@ module VX_shared_mem #(
core_rsp_data_in = 'x; core_rsp_data_in = 'x;
bank_rsp_sel_n = bank_rsp_sel_r; bank_rsp_sel_n = bank_rsp_sel_r;
for (integer i = 0; i < NUM_BANKS; i++) begin for (integer i = 0; i < NUM_BANKS; i++) begin
if (per_bank_core_req_valid[i] if (core_req_read_mask[i]
&& (core_rsp_tag_in[CORE_TAG_ID_BITS-1:0] == per_bank_core_req_tag[i][CORE_TAG_ID_BITS-1:0])) begin && (core_rsp_tag_in[CORE_TAG_ID_BITS-1:0] == per_bank_core_req_tag[i][CORE_TAG_ID_BITS-1:0])) begin
core_rsp_valids_in[per_bank_core_req_tid[i]] = 1; core_rsp_valids_in[per_bank_core_req_tid[i]] = 1;
core_rsp_data_in[per_bank_core_req_tid[i]] = per_bank_core_rsp_data[i]; core_rsp_data_in[per_bank_core_req_tid[i]] = per_bank_core_rsp_data[i];

4
hw/rtl/tex_unit/VX_tex_mem.sv
View File

@@ -8,8 +8,8 @@ module VX_tex_mem #(
input wire reset, input wire reset,
// memory interface // memory interface
VX_dcache_req_if dcache_req_if, VX_dcache_req_if.master dcache_req_if,
VX_dcache_rsp_if dcache_rsp_if, VX_dcache_rsp_if.slave dcache_rsp_if,
// inputs // inputs
input wire req_valid, input wire req_valid,