Merge branch 'bug_fixes'

vortexgpgpu · Dec 5, 2024 · 18ae57c · 18ae57c
2 parents b48b605 + a760d90
commit 18ae57c
Show file tree

Hide file tree

Showing 38 changed files with 1,316 additions and 788 deletions.
diff --git a/ci/regression.sh.in b/ci/regression.sh.in
@@ -105,7 +105,7 @@ regression()
     ./ci/blackbox.sh --driver=simx --app=vecadd --rebuild=3
 
     # test for matmul
-    CONFIGS="-DTC_NUM=4 -DTC_SIZE=8" ./ci/blackbox.sh --cores=4 --app=matmul --driver=simx --threads=32 --warps=32 --args="-n128 -d1" 
+    CONFIGS="-DTC_NUM=4 -DTC_SIZE=8" ./ci/blackbox.sh --cores=4 --app=matmul --driver=simx --threads=32 --warps=32 --args="-n128 -d1"
 
     echo "regression tests done!"
 }
@@ -322,6 +322,10 @@ config2()
     CONFIGS="-DPLATFORM_MEMORY_INTERLEAVE=1" ./ci/blackbox.sh --driver=opae --app=mstress
     CONFIGS="-DPLATFORM_MEMORY_INTERLEAVE=0" ./ci/blackbox.sh --driver=opae --app=mstress
 
+    # test memory ports
+    CONFIGS="-DPLATFORM_MEMORY_BANKS=2" ./ci/blackbox.sh --driver=simx --app=demo
+    CONFIGS="-DPLATFORM_MEMORY_BANKS=2" ./ci/blackbox.sh --driver=simx --app=demo --threads=32
+
     echo "configuration-2 tests done!"
 }
 

diff --git a/hw/rtl/VX_config.vh b/hw/rtl/VX_config.vh
@@ -14,8 +14,6 @@
 `ifndef VX_CONFIG_VH
 `define VX_CONFIG_VH
 
-
-
 `ifndef MIN
 `define MIN(x, y)   (((x) < (y)) ? (x) : (y))
 `endif
@@ -170,8 +168,8 @@
 `define L3_LINE_SIZE `MEM_BLOCK_SIZE
 `endif
 
-`ifndef MEMORY_BANKS
-`define MEMORY_BANKS 2
+`ifndef PLATFORM_MEMORY_BANKS
+`define PLATFORM_MEMORY_BANKS 1
 `endif
 
 `ifdef XLEN_64
@@ -193,7 +191,7 @@
 `endif
 
 `ifdef VM_ENABLE
-`ifndef PAGE_TABLE_BASE_ADDR  
+`ifndef PAGE_TABLE_BASE_ADDR
 `define PAGE_TABLE_BASE_ADDR 64'h0F0000000
 `endif
 
@@ -218,7 +216,7 @@
 `endif
 
 `ifdef VM_ENABLE
-`ifndef PAGE_TABLE_BASE_ADDR  
+`ifndef PAGE_TABLE_BASE_ADDR
 `define PAGE_TABLE_BASE_ADDR 32'hF0000000
 `endif
 
@@ -303,13 +301,13 @@
         `ifndef VM_ADDR_MODE
         `define VM_ADDR_MODE SV32  //or BARE
         `endif
-        `ifndef PT_LEVEL 
+        `ifndef PT_LEVEL
         `define PT_LEVEL (2)
         `endif
         `ifndef PTE_SIZE
         `define PTE_SIZE (4)
         `endif
-        `ifndef NUM_PTE_ENTRY 
+        `ifndef NUM_PTE_ENTRY
         `define NUM_PTE_ENTRY (1024)
         `endif
         `ifndef PT_SIZE_LIMIT
@@ -319,13 +317,13 @@
         `ifndef VM_ADDR_MODE
         `define VM_ADDR_MODE SV39 //or BARE
         `endif
-        `ifndef PT_LEVEL 
+        `ifndef PT_LEVEL
         `define PT_LEVEL (3)
         `endif
         `ifndef PTE_SIZE
         `define PTE_SIZE (8)
         `endif
-        `ifndef NUM_PTE_ENTRY 
+        `ifndef NUM_PTE_ENTRY
         `define NUM_PTE_ENTRY (512)
         `endif
         `ifndef PT_SIZE_LIMIT
@@ -604,7 +602,7 @@
 
 // Number of Banks
 `ifndef DCACHE_NUM_BANKS
-`define DCACHE_NUM_BANKS `MIN(`NUM_LSU_LANES, 4)
+`define DCACHE_NUM_BANKS `MIN(DCACHE_NUM_REQS, 16)
 `endif
 
 // Core Response Queue Size
@@ -647,6 +645,15 @@
 `define DCACHE_REPL_POLICY 1
 `endif
 
+// Number of Memory Ports
+`ifndef L1_MEM_PORTS
+`ifdef L1_DISABLE
+`define L1_MEM_PORTS `MIN(DCACHE_NUM_REQS, `PLATFORM_MEMORY_BANKS)
+`else
+`define L1_MEM_PORTS `MIN(`DCACHE_NUM_BANKS, `PLATFORM_MEMORY_BANKS)
+`endif
+`endif
+
 // LMEM Configurable Knobs ////////////////////////////////////////////////////
 
 `ifndef LMEM_DISABLE
@@ -674,7 +681,7 @@
 
 // Number of Banks
 `ifndef L2_NUM_BANKS
-`define L2_NUM_BANKS `MIN(4, `NUM_SOCKETS)
+`define L2_NUM_BANKS `MIN(L2_NUM_REQS, 16)
 `endif
 
 // Core Response Queue Size
@@ -717,6 +724,15 @@
 `define L2_REPL_POLICY 1
 `endif
 
+// Number of Memory Ports
+`ifndef L2_MEM_PORTS
+`ifdef L2_ENABLE
+`define L2_MEM_PORTS `MIN(`L2_NUM_BANKS, `PLATFORM_MEMORY_BANKS)
+`else
+`define L2_MEM_PORTS `MIN(L2_NUM_REQS, `PLATFORM_MEMORY_BANKS)
+`endif
+`endif
+
 // L3cache Configurable Knobs /////////////////////////////////////////////////
 
 // Cache Size
@@ -726,7 +742,7 @@
 
 // Number of Banks
 `ifndef L3_NUM_BANKS
-`define L3_NUM_BANKS `MIN(8, `NUM_CLUSTERS)
+`define L3_NUM_BANKS `MIN(L3_NUM_REQS, 16)
 `endif
 
 // Core Response Queue Size
@@ -769,9 +785,13 @@
 `define L3_REPL_POLICY 1
 `endif
 
-// Number of Memory Ports from LLC
-`ifndef NUM_MEM_PORTS
-`define NUM_MEM_PORTS `MIN(`MEMORY_BANKS, `L3_NUM_BANKS)
+// Number of Memory Ports
+`ifndef L3_MEM_PORTS
+`ifdef L3_ENABLE
+`define L3_MEM_PORTS `MIN(`L3_NUM_BANKS, `PLATFORM_MEMORY_BANKS)
+`else
+`define L3_MEM_PORTS `MIN(L3_NUM_REQS, `PLATFORM_MEMORY_BANKS)
+`endif
 `endif
 
 // ISA Extensions /////////////////////////////////////////////////////////////

diff --git a/hw/rtl/VX_platform.vh b/hw/rtl/VX_platform.vh
@@ -163,6 +163,7 @@ endgenerate
 `define USE_BLOCK_BRAM  (* ramstyle = "block" *)
 `define USE_FAST_BRAM   (* ramstyle = "MLAB, no_rw_check" *)
 `define NO_RW_RAM_CHECK (* altera_attribute = "-name add_pass_through_logic_to_inferred_rams off" *)
+`define RW_RAM_CHECK    (* altera_attribute = "-name add_pass_through_logic_to_inferred_rams on" *)
 `define DISABLE_BRAM    (* ramstyle = "logic" *)
 `define PRESERVE_NET    (* preserve *)
 `define BLACKBOX_CELL   (* black_box *)
@@ -173,6 +174,7 @@ endgenerate
 `define USE_BLOCK_BRAM  (* ram_style = "block" *)
 `define USE_FAST_BRAM   (* ram_style = "distributed" *)
 `define NO_RW_RAM_CHECK (* rw_addr_collision = "no" *)
+`define RW_RAM_CHECK    (* rw_addr_collision = "yes" *)
 `define DISABLE_BRAM    (* ram_style = "registers" *)
 `define PRESERVE_NET    (* keep = "true" *)
 `define BLACKBOX_CELL   (* black_box *)
@@ -183,6 +185,7 @@ endgenerate
 `define USE_BLOCK_BRAM
 `define USE_FAST_BRAM
 `define NO_RW_RAM_CHECK
+`define RW_RAM_CHECK
 `define DISABLE_BRAM
 `define PRESERVE_NET
 `define BLACKBOX_CELL

diff --git a/hw/rtl/core/VX_mem_unit.sv b/hw/rtl/core/VX_mem_unit.sv
@@ -47,7 +47,7 @@ module VX_mem_unit import VX_gpu_pkg::*; #(
 
     for (genvar i = 0; i < `NUM_LSU_BLOCKS; ++i) begin : g_lmem_switches
         VX_lmem_switch #(
-            .REQ0_OUT_BUF (3),
+            .REQ0_OUT_BUF (1),
             .REQ1_OUT_BUF (0),
             .RSP_OUT_BUF  (1),
             .ARBITER      ("P")
@@ -78,7 +78,7 @@ module VX_mem_unit import VX_gpu_pkg::*; #(
             .TAG_SEL_BITS (LSU_TAG_WIDTH - `UUID_WIDTH),
             .ARBITER      ("P"),
             .REQ_OUT_BUF  (3),
-            .RSP_OUT_BUF  (0)
+            .RSP_OUT_BUF  (2)
         ) lmem_adapter (
             .clk        (clk),
             .reset      (reset),