minor updates

This commit is contained in:
Blaise Tine
2023-12-30 00:52:44 -08:00
parent 7425446b15
commit 031d24e695
65 changed files with 5705 additions and 18904 deletions

View File

@@ -82,64 +82,9 @@ module VX_cluster import VX_gpu_pkg::*; #(
`endif
VX_mem_bus_if #(
.DATA_SIZE (L2_WORD_SIZE),
.TAG_WIDTH (L2_TAG_WIDTH)
) l2_mem_bus_if[L2_NUM_REQS]();
VX_mem_bus_if #(
.DATA_SIZE (ICACHE_LINE_SIZE),
.TAG_WIDTH (ICACHE_MEM_TAG_WIDTH)
) per_socket_icache_mem_bus_if[`NUM_SOCKETS]();
VX_mem_bus_if #(
.DATA_SIZE (DCACHE_LINE_SIZE),
.TAG_WIDTH (DCACHE_MEM_TAG_WIDTH)
) per_socket_dcache_mem_bus_if[`NUM_SOCKETS]();
VX_mem_bus_if #(
.DATA_SIZE (ICACHE_LINE_SIZE),
.TAG_WIDTH (ICACHE_MEM_ARB_TAG_WIDTH)
) icache_mem_bus_if[1]();
VX_mem_bus_if #(
.DATA_SIZE (DCACHE_LINE_SIZE),
.TAG_WIDTH (DCACHE_MEM_ARB_TAG_WIDTH)
) dcache_mem_bus_if[1]();
`RESET_RELAY (l1_mem_arb_reset, reset);
VX_mem_arb #(
.NUM_INPUTS (`NUM_SOCKETS),
.DATA_SIZE (ICACHE_LINE_SIZE),
.TAG_WIDTH (ICACHE_MEM_TAG_WIDTH),
.TAG_SEL_IDX (1), // Skip 0 for NC flag
.ARBITER ("R"),
.OUT_REG_REQ (2),
.OUT_REG_RSP (2)
) icache_mem_arb (
.clk (clk),
.reset (l1_mem_arb_reset),
.bus_in_if (per_socket_icache_mem_bus_if),
.bus_out_if (icache_mem_bus_if)
);
VX_mem_arb #(
.NUM_INPUTS (`NUM_SOCKETS),
.DATA_SIZE (DCACHE_LINE_SIZE),
.TAG_WIDTH (DCACHE_MEM_TAG_WIDTH),
.TAG_SEL_IDX (1), // Skip 0 for NC flag
.ARBITER ("R"),
.OUT_REG_REQ (2),
.OUT_REG_RSP (2)
) dcache_mem_arb (
.clk (clk),
.reset (l1_mem_arb_reset),
.bus_in_if (per_socket_dcache_mem_bus_if),
.bus_out_if (dcache_mem_bus_if)
);
`ASSIGN_VX_MEM_BUS_IF_X (l2_mem_bus_if[ICACHE_MEM_ARB_IDX], icache_mem_bus_if[0], L1_MEM_TAG_WIDTH, ICACHE_MEM_ARB_TAG_WIDTH);
`ASSIGN_VX_MEM_BUS_IF_X (l2_mem_bus_if[DCACHE_MEM_ARB_IDX], dcache_mem_bus_if[0], L1_MEM_TAG_WIDTH, DCACHE_MEM_ARB_TAG_WIDTH);
.DATA_SIZE (`L1_LINE_SIZE),
.TAG_WIDTH (L1_MEM_ARB_TAG_WIDTH)
) per_socket_mem_bus_if[`NUM_SOCKETS]();
`RESET_RELAY (l2_reset, reset);
@@ -155,7 +100,7 @@ module VX_cluster import VX_gpu_pkg::*; #(
.MSHR_SIZE (`L2_MSHR_SIZE),
.MRSQ_SIZE (`L2_MRSQ_SIZE),
.MREQ_SIZE (`L2_MREQ_SIZE),
.TAG_WIDTH (L1_MEM_TAG_WIDTH),
.TAG_WIDTH (L2_TAG_WIDTH),
.WRITE_ENABLE (1),
.UUID_WIDTH (`UUID_WIDTH),
.CORE_OUT_REG (2),
@@ -168,7 +113,7 @@ module VX_cluster import VX_gpu_pkg::*; #(
`ifdef PERF_ENABLE
.cache_perf (mem_perf_tmp_if.l2cache),
`endif
.core_bus_if (l2_mem_bus_if),
.core_bus_if (per_socket_mem_bus_if),
.mem_bus_if (mem_bus_if)
);
@@ -209,8 +154,7 @@ module VX_cluster import VX_gpu_pkg::*; #(
.dcr_bus_if (socket_dcr_bus_if),
.icache_mem_bus_if (per_socket_icache_mem_bus_if[i]),
.dcache_mem_bus_if (per_socket_dcache_mem_bus_if[i]),
.mem_bus_if (per_socket_mem_bus_if[i]),
`ifdef GBAR_ENABLE
.gbar_bus_if (per_socket_gbar_bus_if[i]),

View File

@@ -292,8 +292,8 @@
// Floating-Point Units ///////////////////////////////////////////////////////
// Size of FPU Request Queue
`ifndef FPU_REQ_QUEUE_SIZE
`define FPU_REQ_QUEUE_SIZE (2 * (`NUM_THREADS / `NUM_FPU_LANES))
`ifndef FPUQ_SIZE
`define FPUQ_SIZE (2 * (`NUM_THREADS / `NUM_FPU_LANES))
`endif
// FNCP Latency

View File

@@ -141,10 +141,9 @@ package VX_gpu_pkg;
/////////////////////////////// L1 Parameters /////////////////////////////
localparam ICACHE_MEM_ARB_TAG_WIDTH = (ICACHE_MEM_TAG_WIDTH + `CLOG2(`NUM_SOCKETS));
localparam DCACHE_MEM_ARB_TAG_WIDTH = (DCACHE_MEM_TAG_WIDTH + `CLOG2(`NUM_SOCKETS));
localparam L1_MEM_TAG_WIDTH = `MAX(ICACHE_MEM_ARB_TAG_WIDTH, DCACHE_MEM_ARB_TAG_WIDTH);
localparam L1_MEM_TAG_WIDTH = `MAX(ICACHE_MEM_TAG_WIDTH, DCACHE_MEM_TAG_WIDTH);
localparam L1_MEM_ARB_TAG_WIDTH = (L1_MEM_TAG_WIDTH + `CLOG2(2));
/////////////////////////////// L2 Parameters /////////////////////////////
localparam ICACHE_MEM_ARB_IDX = 0;
@@ -154,10 +153,10 @@ package VX_gpu_pkg;
localparam L2_WORD_SIZE = `L1_LINE_SIZE;
// Input request size
localparam L2_NUM_REQS = 2;
localparam L2_NUM_REQS = `NUM_SOCKETS;
// Core request tag bits
localparam L2_TAG_WIDTH = L1_MEM_TAG_WIDTH;
localparam L2_TAG_WIDTH = L1_MEM_ARB_TAG_WIDTH;
// Memory request data bits
localparam L2_MEM_DATA_WIDTH = (`L2_LINE_SIZE * 8);

View File

@@ -30,8 +30,7 @@ module VX_socket import VX_gpu_pkg::*; #(
VX_dcr_bus_if.slave dcr_bus_if,
// Memory
VX_mem_bus_if.master icache_mem_bus_if,
VX_mem_bus_if.master dcache_mem_bus_if,
VX_mem_bus_if.master mem_bus_if,
`ifdef GBAR_ENABLE
// Barrier
@@ -79,6 +78,11 @@ module VX_socket import VX_gpu_pkg::*; #(
.TAG_WIDTH (ICACHE_TAG_WIDTH)
) per_core_icache_bus_if[`SOCKET_SIZE]();
VX_mem_bus_if #(
.DATA_SIZE (ICACHE_LINE_SIZE),
.TAG_WIDTH (ICACHE_MEM_TAG_WIDTH)
) icache_mem_bus_if();
`RESET_RELAY (icache_reset, reset);
VX_cache_cluster #(
@@ -117,6 +121,11 @@ module VX_socket import VX_gpu_pkg::*; #(
.DATA_SIZE (DCACHE_WORD_SIZE),
.TAG_WIDTH (DCACHE_NOSM_TAG_WIDTH)
) per_core_dcache_bus_if[`SOCKET_SIZE * DCACHE_NUM_REQS]();
VX_mem_bus_if #(
.DATA_SIZE (DCACHE_LINE_SIZE),
.TAG_WIDTH (DCACHE_MEM_TAG_WIDTH)
) dcache_mem_bus_if();
`RESET_RELAY (dcache_reset, reset);
@@ -151,6 +160,40 @@ module VX_socket import VX_gpu_pkg::*; #(
.mem_bus_if (dcache_mem_bus_if)
);
///////////////////////////////////////////////////////////////////////////
VX_mem_bus_if #(
.DATA_SIZE (`L1_LINE_SIZE),
.TAG_WIDTH (L1_MEM_TAG_WIDTH)
) l1_mem_bus_if[2]();
VX_mem_bus_if #(
.DATA_SIZE (`L1_LINE_SIZE),
.TAG_WIDTH (L1_MEM_ARB_TAG_WIDTH)
) l1_mem_arb_bus_if[1]();
`ASSIGN_VX_MEM_BUS_IF_X (l1_mem_bus_if[0], icache_mem_bus_if, L1_MEM_TAG_WIDTH, ICACHE_MEM_TAG_WIDTH);
`ASSIGN_VX_MEM_BUS_IF_X (l1_mem_bus_if[1], dcache_mem_bus_if, L1_MEM_TAG_WIDTH, DCACHE_MEM_TAG_WIDTH);
`RESET_RELAY (mem_arb_reset, reset);
VX_mem_arb #(
.NUM_INPUTS (2),
.DATA_SIZE (`L1_LINE_SIZE),
.TAG_WIDTH (L1_MEM_TAG_WIDTH),
.TAG_SEL_IDX (1), // Skip 0 for NC flag
.ARBITER ("R"),
.OUT_REG_REQ (2),
.OUT_REG_RSP (2)
) mem_arb (
.clk (clk),
.reset (mem_arb_reset),
.bus_in_if (l1_mem_bus_if),
.bus_out_if (l1_mem_arb_bus_if)
);
`ASSIGN_VX_MEM_BUS_IF (mem_bus_if, l1_mem_arb_bus_if[0]);
///////////////////////////////////////////////////////////////////////////
wire [`SOCKET_SIZE-1:0] per_core_sim_ebreak;

View File

@@ -30,7 +30,7 @@ module VX_fpu_unit import VX_fpu_pkg::*; #(
localparam NUM_LANES = `NUM_FPU_LANES;
localparam PID_BITS = `CLOG2(`NUM_THREADS / NUM_LANES);
localparam PID_WIDTH = `UP(PID_BITS);
localparam TAG_WIDTH = `LOG2UP(`FPU_REQ_QUEUE_SIZE);
localparam TAG_WIDTH = `LOG2UP(`FPUQ_SIZE);
localparam PARTIAL_BW = (BLOCK_SIZE != `ISSUE_WIDTH) || (NUM_LANES != `NUM_THREADS);
VX_execute_if #(
@@ -87,7 +87,7 @@ module VX_fpu_unit import VX_fpu_pkg::*; #(
VX_index_buffer #(
.DATAW (`UUID_WIDTH + `NW_WIDTH + NUM_LANES + `XLEN + `NR_BITS + PID_WIDTH + 1 + 1),
.SIZE (`FPU_REQ_QUEUE_SIZE)
.SIZE (`FPUQ_SIZE)
) tag_store (
.clk (clk),
.reset (reset),