opae rtl fixes
This commit is contained in:
@@ -34,6 +34,7 @@ RTL_INCLUDE = -I../../hw/rtl -I../../hw/rtl/libs -I../../hw/rtl/interfaces -I../
|
|||||||
VL_FLAGS += --language 1800-2009 --assert -Wall -Wpedantic $(MULTICORE)
|
VL_FLAGS += --language 1800-2009 --assert -Wall -Wpedantic $(MULTICORE)
|
||||||
VL_FLAGS += -Wno-DECLFILENAME
|
VL_FLAGS += -Wno-DECLFILENAME
|
||||||
VL_FLAGS += --x-initial unique
|
VL_FLAGS += --x-initial unique
|
||||||
|
VL_FLAGS += --x-assign unique
|
||||||
|
|
||||||
# Enable Verilator multithreaded simulation
|
# Enable Verilator multithreaded simulation
|
||||||
#THREADS ?= $(shell python3 -c 'import multiprocessing as mp; print(max(1, mp.cpu_count() // 2))')
|
#THREADS ?= $(shell python3 -c 'import multiprocessing as mp; print(max(1, mp.cpu_count() // 2))')
|
||||||
|
|||||||
@@ -1,8 +1,8 @@
|
|||||||
#ifndef _COMMON_H_
|
#ifndef _COMMON_H_
|
||||||
#define _COMMON_H_
|
#define _COMMON_H_
|
||||||
|
|
||||||
#define DEV_MEM_SRC_ADDR 0x10000000
|
#define DEV_MEM_SRC_ADDR 0x10000040
|
||||||
#define DEV_MEM_DST_ADDR 0x20000000
|
#define DEV_MEM_DST_ADDR 0x20000080
|
||||||
#define NUM_BLOCKS 64
|
#define NUM_BLOCKS 64
|
||||||
|
|
||||||
#endif
|
#endif
|
||||||
BIN
driver/tests/basic/kernel.bin
Executable file → Normal file
BIN
driver/tests/basic/kernel.bin
Executable file → Normal file
Binary file not shown.
@@ -91,7 +91,7 @@ int run_test(vx_device_h device,
|
|||||||
int ref = i + i;
|
int ref = i + i;
|
||||||
int cur = buf_ptr[i];
|
int cur = buf_ptr[i];
|
||||||
if (cur != ref) {
|
if (cur != ref) {
|
||||||
std::cout << "error at 0x" << std::hex << (buf_ptr + i)
|
std::cout << "error at value " << i
|
||||||
<< ": actual 0x" << cur << ", expected 0x" << ref << std::endl;
|
<< ": actual 0x" << cur << ", expected 0x" << ref << std::endl;
|
||||||
++errors;
|
++errors;
|
||||||
}
|
}
|
||||||
@@ -150,23 +150,39 @@ int main(int argc, char *argv[]) {
|
|||||||
RT_CHECK(vx_alloc_dev_mem(device, buf_size, &value));
|
RT_CHECK(vx_alloc_dev_mem(device, buf_size, &value));
|
||||||
kernel_arg.dst_ptr = value;
|
kernel_arg.dst_ptr = value;
|
||||||
|
|
||||||
|
std::cout << "dev_src0=" << std::hex << kernel_arg.src0_ptr << std::endl;
|
||||||
|
std::cout << "dev_src1=" << std::hex << kernel_arg.src1_ptr << std::endl;
|
||||||
|
std::cout << "dev_dst=" << std::hex << kernel_arg.dst_ptr << std::endl;
|
||||||
|
|
||||||
// allocate shared memory
|
// allocate shared memory
|
||||||
std::cout << "allocate shared memory" << std::endl;
|
std::cout << "allocate shared memory" << std::endl;
|
||||||
uint32_t alloc_size = std::max<uint32_t>(buf_size, sizeof(kernel_arg_t));
|
uint32_t alloc_size = std::max<uint32_t>(buf_size, sizeof(kernel_arg_t));
|
||||||
RT_CHECK(vx_alloc_shared_mem(device, alloc_size, &buffer));
|
RT_CHECK(vx_alloc_shared_mem(device, alloc_size, &buffer));
|
||||||
|
|
||||||
// populate source buffer values
|
// populate source buffer0 values
|
||||||
std::cout << "populate source buffer values" << std::endl;
|
std::cout << "populate source buffer0 values" << std::endl;
|
||||||
{
|
{
|
||||||
auto buf_ptr = (int*)vx_host_ptr(buffer);
|
auto buf_ptr = (int*)vx_host_ptr(buffer);
|
||||||
for (uint32_t i = 0; i < num_points; ++i) {
|
for (uint32_t i = 0; i < num_points; ++i) {
|
||||||
buf_ptr[i] = i;
|
buf_ptr[i] = i-1;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
// upload source buffers
|
// upload source buffer0
|
||||||
std::cout << "upload source buffers" << std::endl;
|
std::cout << "upload source buffer0" << std::endl;
|
||||||
RT_CHECK(vx_copy_to_dev(buffer, kernel_arg.src0_ptr, buf_size, 0));
|
RT_CHECK(vx_copy_to_dev(buffer, kernel_arg.src0_ptr, buf_size, 0));
|
||||||
|
|
||||||
|
// populate source buffer1 values
|
||||||
|
std::cout << "populate source buffer1 values" << std::endl;
|
||||||
|
{
|
||||||
|
auto buf_ptr = (int*)vx_host_ptr(buffer);
|
||||||
|
for (uint32_t i = 0; i < num_points; ++i) {
|
||||||
|
buf_ptr[i] = i+1;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
// upload source buffer1
|
||||||
|
std::cout << "upload source buffer1" << std::endl;
|
||||||
RT_CHECK(vx_copy_to_dev(buffer, kernel_arg.src1_ptr, buf_size, 0));
|
RT_CHECK(vx_copy_to_dev(buffer, kernel_arg.src1_ptr, buf_size, 0));
|
||||||
|
|
||||||
// upload kernel argument
|
// upload kernel argument
|
||||||
|
|||||||
@@ -4,8 +4,8 @@ CF += -std=c++11 -fms-extensions
|
|||||||
|
|
||||||
VF += --language 1800-2009 --assert -Wall -Wpedantic
|
VF += --language 1800-2009 --assert -Wall -Wpedantic
|
||||||
VF += -Wno-DECLFILENAME
|
VF += -Wno-DECLFILENAME
|
||||||
VF += --x-initial unique
|
VF += --x-initial unique
|
||||||
|
VF += --x-assign unique
|
||||||
VF += -exe $(SRCS) $(INCLUDE)
|
VF += -exe $(SRCS) $(INCLUDE)
|
||||||
|
|
||||||
#MULTICORE += -DNUM_CLUSTERS=2 -DNUM_CORES=4
|
#MULTICORE += -DNUM_CLUSTERS=2 -DNUM_CORES=4
|
||||||
|
|||||||
@@ -1,6 +1,16 @@
|
|||||||
|
`ifndef NOPAE
|
||||||
`include "platform_if.vh"
|
`include "platform_if.vh"
|
||||||
import local_mem_cfg_pkg::*;
|
import local_mem_cfg_pkg::*;
|
||||||
`include "afu_json_info.vh"
|
`include "afu_json_info.vh"
|
||||||
|
`include "VX_define.vh"
|
||||||
|
`else
|
||||||
|
`include "vortex_afu.vh"
|
||||||
|
/* verilator lint_off IMPORTSTAR */
|
||||||
|
import ccip_if_pkg::*;
|
||||||
|
import local_mem_cfg_pkg::*;
|
||||||
|
/* verilator lint_on IMPORTSTAR */
|
||||||
|
`endif
|
||||||
|
|
||||||
`include "VX_define.vh"
|
`include "VX_define.vh"
|
||||||
|
|
||||||
`define VX_TO_DRAM_ADDR(x) x[`VX_DRAM_ADDR_WIDTH-1:(`VX_DRAM_ADDR_WIDTH-DRAM_ADDR_WIDTH)]
|
`define VX_TO_DRAM_ADDR(x) x[`VX_DRAM_ADDR_WIDTH-1:(`VX_DRAM_ADDR_WIDTH-DRAM_ADDR_WIDTH)]
|
||||||
@@ -93,55 +103,68 @@ logic [`VX_SNP_TAG_WIDTH-1:0] vx_snp_req_tag;
|
|||||||
logic vx_snp_req_ready;
|
logic vx_snp_req_ready;
|
||||||
|
|
||||||
logic vx_snp_rsp_valid;
|
logic vx_snp_rsp_valid;
|
||||||
|
`DEBUG_BEGIN
|
||||||
logic [`VX_SNP_TAG_WIDTH-1:0] vx_snp_rsp_tag;
|
logic [`VX_SNP_TAG_WIDTH-1:0] vx_snp_rsp_tag;
|
||||||
|
`DEBUG_END
|
||||||
logic vx_snp_rsp_ready;
|
logic vx_snp_rsp_ready;
|
||||||
|
|
||||||
|
logic vx_reset;
|
||||||
logic vx_busy;
|
logic vx_busy;
|
||||||
|
|
||||||
// AVS Queues /////////////////////////////////////////////////////////////////
|
// AVS Queues /////////////////////////////////////////////////////////////////
|
||||||
|
|
||||||
logic avs_rtq_push;
|
logic avs_rtq_push;
|
||||||
logic avs_rtq_pop;
|
logic avs_rtq_pop;
|
||||||
|
`DEBUG_BEGIN
|
||||||
logic avs_rtq_empty;
|
logic avs_rtq_empty;
|
||||||
logic avs_rtq_full;
|
logic avs_rtq_full;
|
||||||
|
`DEBUG_BEGIN
|
||||||
|
|
||||||
logic avs_rdq_push;
|
logic avs_rdq_push;
|
||||||
logic avs_rdq_pop;
|
logic avs_rdq_pop;
|
||||||
t_local_mem_data avs_rdq_dout;
|
t_local_mem_data avs_rdq_dout;
|
||||||
logic avs_rdq_empty;
|
logic avs_rdq_empty;
|
||||||
|
`DEBUG_BEGIN
|
||||||
logic avs_rdq_full;
|
logic avs_rdq_full;
|
||||||
|
`DEBUG_END
|
||||||
|
|
||||||
// CSR variables //////////////////////////////////////////////////////////////
|
// CSR variables //////////////////////////////////////////////////////////////
|
||||||
|
|
||||||
logic [2:0] csr_cmd;
|
logic [2:0] csr_cmd;
|
||||||
t_ccip_clAddr csr_io_addr;
|
t_ccip_clAddr csr_io_addr;
|
||||||
t_local_mem_addr csr_mem_addr;
|
logic[DRAM_ADDR_WIDTH-1:0] csr_mem_addr;
|
||||||
t_ccip_clAddr csr_data_size;
|
logic[DRAM_ADDR_WIDTH-1:0] csr_data_size;
|
||||||
|
|
||||||
// MMIO controller ////////////////////////////////////////////////////////////
|
// MMIO controller ////////////////////////////////////////////////////////////
|
||||||
|
|
||||||
t_ccip_c0_ReqMmioHdr mmioHdr = t_ccip_c0_ReqMmioHdr'(cp2af_sRxPort.c0.hdr);
|
`IGNORE_WARNINGS_BEGIN
|
||||||
|
t_ccip_c0_ReqMmioHdr mmio_hdr;
|
||||||
|
`IGNORE_WARNINGS_END
|
||||||
|
assign mmio_hdr = t_ccip_c0_ReqMmioHdr'(cp2af_sRxPort.c0.hdr);
|
||||||
|
|
||||||
|
t_if_ccip_c2_Tx mmio_tx;
|
||||||
|
assign af2cp_sTxPort.c2 = mmio_tx;
|
||||||
|
|
||||||
always_ff @(posedge clk)
|
always_ff @(posedge clk)
|
||||||
begin
|
begin
|
||||||
if (SoftReset) begin
|
if (SoftReset) begin
|
||||||
af2cp_sTxPort.c2.hdr <= 0;
|
mmio_tx.hdr <= 0;
|
||||||
af2cp_sTxPort.c2.data <= 0;
|
mmio_tx.data <= 0;
|
||||||
af2cp_sTxPort.c2.mmioRdValid <= 0;
|
mmio_tx.mmioRdValid <= 0;
|
||||||
csr_cmd <= 0;
|
csr_cmd <= 0;
|
||||||
csr_io_addr <= 0;
|
csr_io_addr <= 0;
|
||||||
csr_mem_addr <= 0;
|
csr_mem_addr <= 0;
|
||||||
csr_data_size <= 0;
|
csr_data_size <= 0;
|
||||||
end
|
end
|
||||||
else begin
|
else begin
|
||||||
|
|
||||||
csr_cmd <= 0;
|
csr_cmd <= 0;
|
||||||
af2cp_sTxPort.c2.mmioRdValid <= 0;
|
mmio_tx.mmioRdValid <= 0;
|
||||||
|
|
||||||
// serve MMIO write request
|
// serve MMIO write request
|
||||||
if (cp2af_sRxPort.c0.mmioWrValid)
|
if (cp2af_sRxPort.c0.mmioWrValid)
|
||||||
begin
|
begin
|
||||||
case (mmioHdr.address)
|
case (mmio_hdr.address)
|
||||||
MMIO_CSR_IO_ADDR: begin
|
MMIO_CSR_IO_ADDR: begin
|
||||||
csr_io_addr <= t_ccip_clAddr'(cp2af_sRxPort.c0.data);
|
csr_io_addr <= t_ccip_clAddr'(cp2af_sRxPort.c0.data);
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
@@ -168,7 +191,7 @@ begin
|
|||||||
end
|
end
|
||||||
default: begin
|
default: begin
|
||||||
// user-defined CSRs
|
// user-defined CSRs
|
||||||
//if (mmioHdr.addres >= MMIO_CSR_USER) begin
|
//if (mmio_hdr.addres >= MMIO_CSR_USER) begin
|
||||||
// write Vortex CRS
|
// write Vortex CRS
|
||||||
//end
|
//end
|
||||||
end
|
end
|
||||||
@@ -177,10 +200,10 @@ begin
|
|||||||
|
|
||||||
// serve MMIO read requests
|
// serve MMIO read requests
|
||||||
if (cp2af_sRxPort.c0.mmioRdValid) begin
|
if (cp2af_sRxPort.c0.mmioRdValid) begin
|
||||||
af2cp_sTxPort.c2.hdr.tid <= mmioHdr.tid; // copy TID
|
mmio_tx.hdr.tid <= mmio_hdr.tid; // copy TID
|
||||||
case (mmioHdr.address)
|
case (mmio_hdr.address)
|
||||||
// AFU header
|
// AFU header
|
||||||
16'h0000: af2cp_sTxPort.c2.data <= {
|
16'h0000: mmio_tx.data <= {
|
||||||
4'b0001, // Feature type = AFU
|
4'b0001, // Feature type = AFU
|
||||||
8'b0, // reserved
|
8'b0, // reserved
|
||||||
4'b0, // afu minor revision = 0
|
4'b0, // afu minor revision = 0
|
||||||
@@ -190,37 +213,31 @@ begin
|
|||||||
4'b0, // afu major revision = 0
|
4'b0, // afu major revision = 0
|
||||||
12'b0 // feature ID = 0
|
12'b0 // feature ID = 0
|
||||||
};
|
};
|
||||||
AFU_ID_L: af2cp_sTxPort.c2.data <= afu_id[63:0]; // afu id low
|
AFU_ID_L: mmio_tx.data <= afu_id[63:0]; // afu id low
|
||||||
AFU_ID_H: af2cp_sTxPort.c2.data <= afu_id[127:64]; // afu id hi
|
AFU_ID_H: mmio_tx.data <= afu_id[127:64]; // afu id hi
|
||||||
16'h0006: af2cp_sTxPort.c2.data <= 64'h0; // next AFU
|
16'h0006: mmio_tx.data <= 64'h0; // next AFU
|
||||||
16'h0008: af2cp_sTxPort.c2.data <= 64'h0; // reserved
|
16'h0008: mmio_tx.data <= 64'h0; // reserved
|
||||||
MMIO_CSR_STATUS: begin
|
MMIO_CSR_STATUS: begin
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
if (state != af2cp_sTxPort.c2.data) begin
|
if (state != mmio_tx.data) begin
|
||||||
$display("%t: STATUS: state=%0d", $time, state);
|
$display("%t: STATUS: state=%0d", $time, state);
|
||||||
end
|
end
|
||||||
`endif
|
`endif
|
||||||
af2cp_sTxPort.c2.data <= state;
|
mmio_tx.data <= {60'b0, state};
|
||||||
end
|
end
|
||||||
default: af2cp_sTxPort.c2.data <= 64'h0;
|
default: mmio_tx.data <= 64'h0;
|
||||||
endcase
|
endcase
|
||||||
af2cp_sTxPort.c2.mmioRdValid <= 1; // post response
|
mmio_tx.mmioRdValid <= 1; // post response
|
||||||
end
|
end
|
||||||
end
|
end
|
||||||
end
|
end
|
||||||
|
|
||||||
// COMMAND FSM ////////////////////////////////////////////////////////////////
|
// COMMAND FSM ////////////////////////////////////////////////////////////////
|
||||||
|
|
||||||
t_ccip_clAddr cci_wr_req_ctr;
|
|
||||||
logic [DRAM_ADDR_WIDTH-1:0] avs_rd_req_ctr;
|
|
||||||
logic [DRAM_ADDR_WIDTH-1:0] avs_wr_req_ctr;
|
|
||||||
logic vx_reset;
|
|
||||||
|
|
||||||
logic cmd_read_done;
|
logic cmd_read_done;
|
||||||
logic cmd_write_done;
|
logic cmd_write_done;
|
||||||
logic cmd_clflush_done;
|
logic cmd_clflush_done;
|
||||||
|
logic cmd_run_done;
|
||||||
logic cmd_run_done = !vx_busy;
|
|
||||||
|
|
||||||
always_ff @(posedge clk)
|
always_ff @(posedge clk)
|
||||||
begin
|
begin
|
||||||
@@ -260,6 +277,9 @@ begin
|
|||||||
`endif
|
`endif
|
||||||
state <= STATE_CLFLUSH;
|
state <= STATE_CLFLUSH;
|
||||||
end
|
end
|
||||||
|
default: begin
|
||||||
|
state <= state;
|
||||||
|
end
|
||||||
endcase
|
endcase
|
||||||
end
|
end
|
||||||
|
|
||||||
@@ -291,6 +311,10 @@ begin
|
|||||||
end
|
end
|
||||||
end
|
end
|
||||||
|
|
||||||
|
default: begin
|
||||||
|
state <= state;
|
||||||
|
end
|
||||||
|
|
||||||
endcase
|
endcase
|
||||||
end
|
end
|
||||||
end
|
end
|
||||||
@@ -304,7 +328,9 @@ t_cci_rdq_data cci_rdq_dout;
|
|||||||
logic cci_dram_rd_req_fire;
|
logic cci_dram_rd_req_fire;
|
||||||
logic cci_dram_wr_req_fire;
|
logic cci_dram_wr_req_fire;
|
||||||
logic vx_dram_rd_req_fire;
|
logic vx_dram_rd_req_fire;
|
||||||
|
`DEBUG_BEGIN
|
||||||
logic vx_dram_wr_req_fire;
|
logic vx_dram_wr_req_fire;
|
||||||
|
`DEBUG_END
|
||||||
logic vx_dram_rd_rsp_fire;
|
logic vx_dram_rd_rsp_fire;
|
||||||
|
|
||||||
t_local_mem_byte_mask vx_dram_req_byteen_;
|
t_local_mem_byte_mask vx_dram_req_byteen_;
|
||||||
@@ -315,15 +341,17 @@ logic [DRAM_ADDR_WIDTH-1:0] cci_dram_rd_req_addr, cci_dram_wr_req_addr;
|
|||||||
logic cci_dram_rd_req_enable, cci_dram_wr_req_enable;
|
logic cci_dram_rd_req_enable, cci_dram_wr_req_enable;
|
||||||
logic vx_dram_req_enable, vx_dram_rd_req_enable, vx_dram_wr_req_enable;
|
logic vx_dram_req_enable, vx_dram_rd_req_enable, vx_dram_wr_req_enable;
|
||||||
|
|
||||||
|
logic [DRAM_ADDR_WIDTH-1:0] cci_dram_rd_req_ctr, cci_dram_wr_req_ctr;
|
||||||
|
|
||||||
assign vortex_enabled = (STATE_RUN == state) || (STATE_CLFLUSH == state);
|
assign vortex_enabled = (STATE_RUN == state) || (STATE_CLFLUSH == state);
|
||||||
|
|
||||||
assign cci_dram_rd_req_enable = (state == STATE_READ)
|
assign cci_dram_rd_req_enable = (state == STATE_READ)
|
||||||
&& (avs_pending_reads < AVS_RD_QUEUE_SIZE)
|
&& (avs_pending_reads < AVS_RD_QUEUE_SIZE)
|
||||||
&& (avs_rd_req_ctr != 0);
|
&& (cci_dram_rd_req_ctr != 0);
|
||||||
|
|
||||||
assign cci_dram_wr_req_enable = (state == STATE_WRITE)
|
assign cci_dram_wr_req_enable = (state == STATE_WRITE)
|
||||||
&& !cci_rdq_empty
|
&& !cci_rdq_empty
|
||||||
&& (avs_wr_req_ctr != 0);
|
&& (cci_dram_wr_req_ctr < csr_data_size);
|
||||||
|
|
||||||
assign vx_dram_req_enable = vortex_enabled && (avs_pending_reads < AVS_RD_QUEUE_SIZE);
|
assign vx_dram_req_enable = vortex_enabled && (avs_pending_reads < AVS_RD_QUEUE_SIZE);
|
||||||
assign vx_dram_rd_req_enable = vx_dram_req_enable && vx_dram_req_valid && ~vx_dram_req_rw;
|
assign vx_dram_rd_req_enable = vx_dram_req_enable && vx_dram_req_valid && ~vx_dram_req_rw;
|
||||||
@@ -338,24 +366,22 @@ assign vx_dram_wr_req_fire = vx_dram_wr_req_enable && ~avs_waitrequest;
|
|||||||
assign vx_dram_rd_rsp_fire = vx_dram_rsp_valid && vx_dram_rsp_ready;
|
assign vx_dram_rd_rsp_fire = vx_dram_rsp_valid && vx_dram_rsp_ready;
|
||||||
|
|
||||||
assign avs_pending_reads_next = avs_pending_reads
|
assign avs_pending_reads_next = avs_pending_reads
|
||||||
+ ((cci_dram_rd_req_fire || vx_dram_rd_req_fire) && ~avs_rdq_pop) ? 1 :
|
+ (((cci_dram_rd_req_fire || vx_dram_rd_req_fire) && ~avs_rdq_pop) ? 1 :
|
||||||
(~(cci_dram_rd_req_fire || vx_dram_rd_req_fire) && avs_rdq_pop) ? -1 : 0;
|
(~(cci_dram_rd_req_fire || vx_dram_rd_req_fire) && avs_rdq_pop) ? -1 : 0);
|
||||||
|
|
||||||
assign cmd_write_done = (0 == avs_wr_req_ctr);
|
|
||||||
|
|
||||||
if (`VX_DRAM_LINE_WIDTH != DRAM_LINE_WIDTH) begin
|
if (`VX_DRAM_LINE_WIDTH != DRAM_LINE_WIDTH) begin
|
||||||
assign vx_dram_req_offset = {{VX_DRAM_LINE_LW{1'b0}}, vx_dram_req_addr[(DRAM_LINE_LW-VX_DRAM_LINE_LW)-1:0]} << VX_DRAM_LINE_LW;
|
assign vx_dram_req_offset = ((DRAM_LINE_LW)'(vx_dram_req_addr[(DRAM_LINE_LW-VX_DRAM_LINE_LW)-1:0])) << VX_DRAM_LINE_LW;
|
||||||
assign vx_dram_req_byteen_ = vx_dram_req_byteen << ({(VX_DRAM_LINE_LW - 3)'(0), vx_dram_req_addr[(DRAM_LINE_LW-VX_DRAM_LINE_LW)-1:0]} << (VX_DRAM_LINE_LW - 3));
|
assign vx_dram_req_byteen_ = 64'(vx_dram_req_byteen) << (6'(vx_dram_req_addr[(DRAM_LINE_LW-VX_DRAM_LINE_LW)-1:0]) << (VX_DRAM_LINE_LW - 3));
|
||||||
end else begin
|
end else begin
|
||||||
assign vx_dram_req_offset = 0;
|
assign vx_dram_req_offset = 0;
|
||||||
assign vx_dram_req_byteen_ = 64'hffffffffffffffff;
|
assign vx_dram_req_byteen_ = vx_dram_req_byteen;
|
||||||
end
|
end
|
||||||
|
|
||||||
always_comb
|
always_comb
|
||||||
begin
|
begin
|
||||||
case (state)
|
case (state)
|
||||||
CMD_TYPE_READ: avs_address = cci_dram_rd_req_addr;
|
CMD_TYPE_READ: avs_address = cci_dram_rd_req_addr;
|
||||||
CMD_TYPE_WRITE: avs_address = cci_dram_wr_req_addr;
|
CMD_TYPE_WRITE: avs_address = cci_dram_wr_req_addr + ((DRAM_ADDR_WIDTH)'(t_cci_rdq_tag'(cci_rdq_dout)));
|
||||||
default: avs_address = `VX_TO_DRAM_ADDR(vx_dram_req_addr);
|
default: avs_address = `VX_TO_DRAM_ADDR(vx_dram_req_addr);
|
||||||
endcase
|
endcase
|
||||||
|
|
||||||
@@ -367,51 +393,53 @@ begin
|
|||||||
|
|
||||||
case (state)
|
case (state)
|
||||||
CMD_TYPE_WRITE: avs_writedata = cci_rdq_dout[$bits(t_ccip_clData) + $bits(t_cci_rdq_tag)-1:$bits(t_cci_rdq_tag)];
|
CMD_TYPE_WRITE: avs_writedata = cci_rdq_dout[$bits(t_ccip_clData) + $bits(t_cci_rdq_tag)-1:$bits(t_cci_rdq_tag)];
|
||||||
default: avs_writedata = vx_dram_req_data << vx_dram_req_offset;
|
default: avs_writedata = (DRAM_LINE_WIDTH)'(vx_dram_req_data) << vx_dram_req_offset;
|
||||||
endcase
|
endcase
|
||||||
end
|
end
|
||||||
|
|
||||||
assign avs_read = cci_dram_rd_req_enable || vx_dram_rd_req_enable;
|
assign avs_read = cci_dram_rd_req_enable || vx_dram_rd_req_enable;
|
||||||
assign avs_write = cci_dram_wr_req_enable || vx_dram_wr_req_enable;
|
assign avs_write = cci_dram_wr_req_enable || vx_dram_wr_req_enable;
|
||||||
|
|
||||||
|
assign cmd_write_done = (cci_dram_wr_req_ctr >= csr_data_size);
|
||||||
|
|
||||||
always_ff @(posedge clk)
|
always_ff @(posedge clk)
|
||||||
begin
|
begin
|
||||||
if (SoftReset)
|
if (SoftReset)
|
||||||
begin
|
begin
|
||||||
mem_bank_select <= 0;
|
mem_bank_select <= 0;
|
||||||
avs_burstcount <= 1;
|
avs_burstcount <= 1;
|
||||||
avs_rd_req_ctr <= 0;
|
|
||||||
avs_wr_req_ctr <= 0;
|
|
||||||
avs_pending_reads <= 0;
|
|
||||||
cci_dram_rd_req_addr <= 0;
|
cci_dram_rd_req_addr <= 0;
|
||||||
cci_dram_wr_req_addr <= 0;
|
cci_dram_wr_req_addr <= 0;
|
||||||
|
cci_dram_rd_req_ctr <= 0;
|
||||||
|
cci_dram_wr_req_ctr <= 0;
|
||||||
|
avs_pending_reads <= 0;
|
||||||
end
|
end
|
||||||
else begin
|
else begin
|
||||||
|
|
||||||
if (state == STATE_IDLE) begin
|
if (state == STATE_IDLE) begin
|
||||||
if (CMD_TYPE_READ == csr_cmd) begin
|
if (CMD_TYPE_READ == csr_cmd) begin
|
||||||
cci_dram_rd_req_addr <= csr_mem_addr;
|
cci_dram_rd_req_addr <= csr_mem_addr;
|
||||||
avs_rd_req_ctr <= csr_data_size;
|
cci_dram_rd_req_ctr <= csr_data_size;
|
||||||
end
|
end
|
||||||
else if (CMD_TYPE_WRITE == csr_cmd) begin
|
else if (CMD_TYPE_WRITE == csr_cmd) begin
|
||||||
cci_dram_wr_req_addr <= csr_mem_addr;
|
cci_dram_wr_req_addr <= csr_mem_addr;
|
||||||
avs_wr_req_ctr <= csr_data_size;
|
cci_dram_wr_req_ctr <= 0;
|
||||||
end
|
end
|
||||||
end
|
end
|
||||||
|
|
||||||
if (cci_dram_rd_req_fire) begin
|
if (cci_dram_rd_req_fire) begin
|
||||||
cci_dram_rd_req_addr <= cci_dram_rd_req_addr + 1;
|
cci_dram_rd_req_addr <= cci_dram_rd_req_addr + 1;
|
||||||
avs_rd_req_ctr <= avs_rd_req_ctr - 1;
|
cci_dram_rd_req_ctr <= cci_dram_rd_req_ctr - 1;
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
$display("%t: AVS Rd Req: addr=%0h, rem=%0d, pending=%0d", $time, `DRAM_TO_BYTE_ADDR(avs_address), (avs_rd_req_ctr - 1), avs_pending_reads_next);
|
$display("%t: AVS Rd Req: addr=%0h, rem=%0d, pending=%0d", $time, `DRAM_TO_BYTE_ADDR(avs_address), (cci_dram_rd_req_ctr - 1), avs_pending_reads_next);
|
||||||
`endif
|
`endif
|
||||||
end
|
end
|
||||||
|
|
||||||
if (cci_dram_wr_req_fire) begin
|
if (cci_dram_wr_req_fire) begin
|
||||||
cci_dram_wr_req_addr <= ((cci_dram_wr_req_addr + 1) & ~(CCI_RD_WINDOW_SIZE-1)) | t_cci_rdq_tag'(cci_rdq_dout);
|
cci_dram_wr_req_addr <= cci_dram_wr_req_addr + ((t_cci_rdq_tag'(cci_dram_wr_req_ctr) == (DRAM_ADDR_WIDTH)'(CCI_RD_WINDOW_SIZE-1)) ? (DRAM_ADDR_WIDTH)'(CCI_RD_WINDOW_SIZE) : 0);
|
||||||
avs_wr_req_ctr <= avs_wr_req_ctr - 1;
|
cci_dram_wr_req_ctr <= cci_dram_wr_req_ctr + 1;
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
$display("%t: AVS Wr Req: addr=%0h, data=%0h, rem=%0d", $time, `DRAM_TO_BYTE_ADDR(avs_address), avs_writedata, (avs_wr_req_ctr - 1));
|
$display("%t: AVS Wr Req: addr=%0h, data=%0h, rem=%0d", $time, `DRAM_TO_BYTE_ADDR(avs_address), avs_writedata, (cci_dram_wr_req_ctr + 1));
|
||||||
`endif
|
`endif
|
||||||
end
|
end
|
||||||
|
|
||||||
@@ -441,7 +469,7 @@ assign vx_dram_req_ready = vx_dram_req_enable && !avs_waitrequest;
|
|||||||
|
|
||||||
assign vx_dram_rsp_valid = vortex_enabled && !avs_rdq_empty;
|
assign vx_dram_rsp_valid = vortex_enabled && !avs_rdq_empty;
|
||||||
if (`VX_DRAM_LINE_WIDTH != DRAM_LINE_WIDTH) begin
|
if (`VX_DRAM_LINE_WIDTH != DRAM_LINE_WIDTH) begin
|
||||||
assign vx_dram_rsp_data = (avs_rdq_dout >> vx_dram_rsp_offset);
|
assign vx_dram_rsp_data = (`VX_DRAM_LINE_WIDTH)'(avs_rdq_dout >> vx_dram_rsp_offset);
|
||||||
end else begin
|
end else begin
|
||||||
assign vx_dram_rsp_data = avs_rdq_dout;
|
assign vx_dram_rsp_data = avs_rdq_dout;
|
||||||
end
|
end
|
||||||
@@ -462,7 +490,8 @@ VX_generic_queue #(
|
|||||||
.pop (avs_rtq_pop),
|
.pop (avs_rtq_pop),
|
||||||
.data_out ({vx_dram_rsp_tag, vx_dram_rsp_offset}),
|
.data_out ({vx_dram_rsp_tag, vx_dram_rsp_offset}),
|
||||||
.empty (avs_rtq_empty),
|
.empty (avs_rtq_empty),
|
||||||
.full (avs_rtq_full)
|
.full (avs_rtq_full),
|
||||||
|
`UNUSED_PIN (size)
|
||||||
);
|
);
|
||||||
|
|
||||||
// AVS data read response queue ///////////////////////////////////////////////
|
// AVS data read response queue ///////////////////////////////////////////////
|
||||||
@@ -483,25 +512,27 @@ VX_generic_queue #(
|
|||||||
.pop (avs_rdq_pop),
|
.pop (avs_rdq_pop),
|
||||||
.data_out (avs_rdq_dout),
|
.data_out (avs_rdq_dout),
|
||||||
.empty (avs_rdq_empty),
|
.empty (avs_rdq_empty),
|
||||||
.full (avs_rdq_full)
|
.full (avs_rdq_full),
|
||||||
|
`UNUSED_PIN (size)
|
||||||
);
|
);
|
||||||
|
|
||||||
// CCI-P Read Request ///////////////////////////////////////////////////////////
|
// CCI-P Read Request ///////////////////////////////////////////////////////////
|
||||||
|
|
||||||
logic [$clog2(CCI_RD_QUEUE_SIZE+1)-1:0] cci_pending_reads, cci_pending_reads_next;
|
logic [$clog2(CCI_RD_QUEUE_SIZE+1)-1:0] cci_pending_reads, cci_pending_reads_next;
|
||||||
t_ccip_clAddr cci_rd_req_addr, cci_rd_req_ctr, cci_rd_req_ctr_next;
|
logic [DRAM_ADDR_WIDTH-1:0] cci_rd_req_ctr, cci_rd_req_ctr_next;
|
||||||
|
t_ccip_clAddr cci_rd_req_addr;
|
||||||
t_cci_rdq_tag cci_rd_rsp_ctr;
|
t_cci_rdq_tag cci_rd_rsp_ctr;
|
||||||
|
|
||||||
logic cci_rd_req_fire, cci_rd_rsp_fire;
|
logic cci_rd_req_fire, cci_rd_rsp_fire;
|
||||||
logic cci_rd_req_enable, cci_rd_req_wait;
|
logic cci_rd_req_enable, cci_rd_req_wait;
|
||||||
|
|
||||||
logic cci_rdq_full, cci_rdq_push, cci_rdq_pop;
|
logic cci_rdq_push, cci_rdq_pop;
|
||||||
t_cci_rdq_data cci_rdq_din;
|
t_cci_rdq_data cci_rdq_din;
|
||||||
|
|
||||||
always_comb begin
|
always_comb begin
|
||||||
af2cp_sTxPort.c0.hdr = t_ccip_c0_ReqMemHdr'(0);
|
af2cp_sTxPort.c0.hdr = t_ccip_c0_ReqMemHdr'(0);
|
||||||
af2cp_sTxPort.c0.hdr.address = cci_rd_req_addr;
|
af2cp_sTxPort.c0.hdr.address = cci_rd_req_addr;
|
||||||
af2cp_sTxPort.c0.hdr.mdata = t_cci_rdq_tag'(cci_rd_req_ctr);
|
af2cp_sTxPort.c0.hdr.mdata = t_ccip_mdata'(t_cci_rdq_tag'(cci_rd_req_ctr));
|
||||||
end
|
end
|
||||||
|
|
||||||
assign cci_rd_req_fire = af2cp_sTxPort.c0.valid && !cp2af_sRxPort.c0TxAlmFull;
|
assign cci_rd_req_fire = af2cp_sTxPort.c0.valid && !cp2af_sRxPort.c0TxAlmFull;
|
||||||
@@ -514,8 +545,8 @@ assign cci_rdq_push = cci_rd_rsp_fire;
|
|||||||
assign cci_rdq_din = {cp2af_sRxPort.c0.data, t_cci_rdq_tag'(cp2af_sRxPort.c0.hdr.mdata)};
|
assign cci_rdq_din = {cp2af_sRxPort.c0.data, t_cci_rdq_tag'(cp2af_sRxPort.c0.hdr.mdata)};
|
||||||
|
|
||||||
assign cci_pending_reads_next = cci_pending_reads
|
assign cci_pending_reads_next = cci_pending_reads
|
||||||
+ (cci_rd_req_fire && ~cci_rdq_pop) ? 1 :
|
+ ((cci_rd_req_fire && ~cci_rdq_pop) ? 1 :
|
||||||
(~cci_rd_req_fire && cci_rdq_pop) ? -1 : 0;
|
(~cci_rd_req_fire && cci_rdq_pop) ? -1 : 0);
|
||||||
|
|
||||||
assign af2cp_sTxPort.c0.valid = cci_rd_req_enable && ~cci_rd_req_wait;
|
assign af2cp_sTxPort.c0.valid = cci_rd_req_enable && ~cci_rd_req_wait;
|
||||||
|
|
||||||
@@ -549,7 +580,7 @@ begin
|
|||||||
if (cci_rd_req_fire) begin
|
if (cci_rd_req_fire) begin
|
||||||
cci_rd_req_addr <= cci_rd_req_addr + 1;
|
cci_rd_req_addr <= cci_rd_req_addr + 1;
|
||||||
cci_rd_req_ctr <= cci_rd_req_ctr_next;
|
cci_rd_req_ctr <= cci_rd_req_ctr_next;
|
||||||
if (t_cci_rdq_tag'(cci_rd_req_ctr) == (CCI_RD_WINDOW_SIZE-1)) begin
|
if (t_cci_rdq_tag'(cci_rd_req_ctr) == t_cci_rdq_tag'(CCI_RD_WINDOW_SIZE-1)) begin
|
||||||
cci_rd_req_wait <= 1; // end current request batch
|
cci_rd_req_wait <= 1; // end current request batch
|
||||||
end
|
end
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
@@ -559,7 +590,7 @@ begin
|
|||||||
|
|
||||||
if (cci_rd_rsp_fire) begin
|
if (cci_rd_rsp_fire) begin
|
||||||
cci_rd_rsp_ctr <= cci_rd_rsp_ctr + 1;
|
cci_rd_rsp_ctr <= cci_rd_rsp_ctr + 1;
|
||||||
if (cci_rd_rsp_ctr == (CCI_RD_WINDOW_SIZE-1)) begin
|
if (cci_rd_rsp_ctr == t_cci_rdq_tag'(CCI_RD_WINDOW_SIZE-1)) begin
|
||||||
cci_rd_req_wait <= 0; // restart new request batch
|
cci_rd_req_wait <= 0; // restart new request batch
|
||||||
end
|
end
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
@@ -589,12 +620,14 @@ VX_generic_queue #(
|
|||||||
.pop (cci_rdq_pop),
|
.pop (cci_rdq_pop),
|
||||||
.data_out (cci_rdq_dout),
|
.data_out (cci_rdq_dout),
|
||||||
.empty (cci_rdq_empty),
|
.empty (cci_rdq_empty),
|
||||||
.full (cci_rdq_full)
|
`UNUSED_PIN (full),
|
||||||
|
`UNUSED_PIN (size)
|
||||||
);
|
);
|
||||||
|
|
||||||
// CCI-P Write Request //////////////////////////////////////////////////////////
|
// CCI-P Write Request //////////////////////////////////////////////////////////
|
||||||
|
|
||||||
logic [$clog2(CCI_RW_QUEUE_SIZE+1)-1:0] cci_pending_writes, cci_pending_writes_next;
|
logic [$clog2(CCI_RW_QUEUE_SIZE+1)-1:0] cci_pending_writes, cci_pending_writes_next;
|
||||||
|
logic [DRAM_ADDR_WIDTH-1:0] cci_wr_req_ctr;
|
||||||
t_ccip_clAddr cci_wr_req_addr;
|
t_ccip_clAddr cci_wr_req_addr;
|
||||||
logic cci_wr_req_enable, cci_wr_rsp_fire;
|
logic cci_wr_req_enable, cci_wr_rsp_fire;
|
||||||
|
|
||||||
@@ -609,8 +642,8 @@ assign cci_wr_req_fire = af2cp_sTxPort.c1.valid && !cp2af_sRxPort.c1TxAlmFull;
|
|||||||
assign cci_wr_rsp_fire = (STATE_READ == state) && cp2af_sRxPort.c1.rspValid;
|
assign cci_wr_rsp_fire = (STATE_READ == state) && cp2af_sRxPort.c1.rspValid;
|
||||||
|
|
||||||
assign cci_pending_writes_next = cci_pending_writes
|
assign cci_pending_writes_next = cci_pending_writes
|
||||||
+ (cci_wr_req_fire && ~cci_wr_rsp_fire) ? 1 :
|
+ ((cci_wr_req_fire && ~cci_wr_rsp_fire) ? 1 :
|
||||||
(~cci_wr_req_fire && cci_wr_rsp_fire) ? -1 : 0;
|
(~cci_wr_req_fire && cci_wr_rsp_fire) ? -1 : 0);
|
||||||
|
|
||||||
assign cmd_read_done = (0 == cci_wr_req_ctr) && (0 == cci_pending_writes);
|
assign cmd_read_done = (0 == cci_wr_req_ctr) && (0 == cci_pending_writes);
|
||||||
|
|
||||||
@@ -660,7 +693,8 @@ end
|
|||||||
|
|
||||||
logic [`VX_DRAM_ADDR_WIDTH-1:0] snp_req_size;
|
logic [`VX_DRAM_ADDR_WIDTH-1:0] snp_req_size;
|
||||||
logic [`VX_DRAM_ADDR_WIDTH-1:0] snp_req_baseaddr;
|
logic [`VX_DRAM_ADDR_WIDTH-1:0] snp_req_baseaddr;
|
||||||
logic [`VX_DRAM_ADDR_WIDTH-1:0] snp_req_ctr, snp_rsp_ctr;
|
logic [`VX_DRAM_ADDR_WIDTH-1:0] snp_req_ctr, snp_req_ctr_next;
|
||||||
|
logic [`VX_DRAM_ADDR_WIDTH-1:0] snp_rsp_ctr, snp_rsp_ctr_next;
|
||||||
|
|
||||||
logic vx_snp_req_fire, vx_snp_rsp_fire;
|
logic vx_snp_req_fire, vx_snp_rsp_fire;
|
||||||
|
|
||||||
@@ -674,6 +708,10 @@ end
|
|||||||
|
|
||||||
assign vx_snp_req_fire = vx_snp_req_valid && vx_snp_req_ready;
|
assign vx_snp_req_fire = vx_snp_req_valid && vx_snp_req_ready;
|
||||||
assign vx_snp_rsp_fire = vx_snp_rsp_valid && vx_snp_rsp_ready;
|
assign vx_snp_rsp_fire = vx_snp_rsp_valid && vx_snp_rsp_ready;
|
||||||
|
|
||||||
|
assign snp_req_ctr_next = vx_snp_req_fire ? (snp_req_ctr + 1) : snp_req_ctr;
|
||||||
|
assign snp_rsp_ctr_next = vx_snp_rsp_fire ? (snp_rsp_ctr - 1) : snp_rsp_ctr;
|
||||||
|
|
||||||
assign cmd_clflush_done = (0 == snp_rsp_ctr);
|
assign cmd_clflush_done = (0 == snp_rsp_ctr);
|
||||||
|
|
||||||
always_ff @(posedge clk)
|
always_ff @(posedge clk)
|
||||||
@@ -691,38 +729,40 @@ begin
|
|||||||
if ((STATE_IDLE == state)
|
if ((STATE_IDLE == state)
|
||||||
&& (CMD_TYPE_CLFLUSH == csr_cmd)) begin
|
&& (CMD_TYPE_CLFLUSH == csr_cmd)) begin
|
||||||
vx_snp_req_addr <= snp_req_baseaddr;
|
vx_snp_req_addr <= snp_req_baseaddr;
|
||||||
snp_req_ctr <= snp_req_size;
|
vx_snp_req_tag <= 0;
|
||||||
|
snp_req_ctr <= 0;
|
||||||
snp_rsp_ctr <= snp_req_size;
|
snp_rsp_ctr <= snp_req_size;
|
||||||
vx_snp_req_valid <= (snp_req_size != 0);
|
vx_snp_req_valid <= (snp_req_size != 0);
|
||||||
vx_snp_rsp_ready <= (snp_req_size != 0);
|
vx_snp_rsp_ready <= (snp_req_size != 0);
|
||||||
end
|
end
|
||||||
|
|
||||||
if ((STATE_CLFLUSH == state)
|
if ((STATE_CLFLUSH == state)
|
||||||
&& (0 == snp_rsp_ctr)) begin
|
&& (snp_req_ctr_next >= snp_req_size)) begin
|
||||||
vx_snp_rsp_ready <= 0;
|
vx_snp_req_valid <= 0;
|
||||||
end
|
end
|
||||||
|
|
||||||
if ((STATE_CLFLUSH == state)
|
if ((STATE_CLFLUSH == state)
|
||||||
&& (0 == snp_req_ctr)) begin
|
&& (0 == snp_rsp_ctr_next)) begin
|
||||||
vx_snp_req_valid <= 0;
|
vx_snp_rsp_ready <= 0;
|
||||||
end
|
end
|
||||||
|
|
||||||
if (vx_snp_req_fire)
|
if (vx_snp_req_fire)
|
||||||
begin
|
begin
|
||||||
|
assert(snp_req_ctr < snp_req_size);
|
||||||
vx_snp_req_addr <= vx_snp_req_addr + 1;
|
vx_snp_req_addr <= vx_snp_req_addr + 1;
|
||||||
vx_snp_req_tag <= snp_req_ctr[`VX_SNP_TAG_WIDTH-1:0];
|
vx_snp_req_tag <= (`VX_SNP_TAG_WIDTH)'(snp_req_ctr_next);
|
||||||
snp_req_ctr <= snp_req_ctr - 1;
|
snp_req_ctr <= snp_req_ctr_next;
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
$display("%t: AFU Snp Req: addr=%0h, tag=%0d, rem=%0d", $time, `DRAM_TO_BYTE_ADDR(vx_snp_req_addr), vx_snp_req_tag, (snp_req_ctr - 1));
|
$display("%t: AFU Snp Req: addr=%0h, tag=%0d, rem=%0d", $time, `DRAM_TO_BYTE_ADDR(vx_snp_req_addr), (`VX_SNP_TAG_WIDTH)'(snp_req_ctr_next), (snp_req_size - snp_req_ctr_next));
|
||||||
`endif
|
`endif
|
||||||
end
|
end
|
||||||
|
|
||||||
if ((STATE_CLFLUSH == state)
|
if ((STATE_CLFLUSH == state)
|
||||||
&& vx_snp_rsp_fire) begin
|
&& vx_snp_rsp_fire) begin
|
||||||
assert(snp_rsp_ctr != 0);
|
assert(snp_rsp_ctr != 0);
|
||||||
snp_rsp_ctr <= snp_rsp_ctr - 1;
|
snp_rsp_ctr <= snp_rsp_ctr_next;
|
||||||
`ifdef DBG_PRINT_OPAE
|
`ifdef DBG_PRINT_OPAE
|
||||||
$display("%t: AFU Snp Rsp: tag=%0d, rem=%0d", $time, vx_snp_rsp_tag, (snp_rsp_ctr - 1));
|
$display("%t: AFU Snp Rsp: tag=%0d, rem=%0d", $time, vx_snp_rsp_tag, snp_rsp_ctr_next);
|
||||||
`endif
|
`endif
|
||||||
end
|
end
|
||||||
end
|
end
|
||||||
@@ -730,6 +770,8 @@ end
|
|||||||
|
|
||||||
// Vortex binding /////////////////////////////////////////////////////////////
|
// Vortex binding /////////////////////////////////////////////////////////////
|
||||||
|
|
||||||
|
assign cmd_run_done = !vx_busy;
|
||||||
|
|
||||||
Vortex_Socket #() vx_socket (
|
Vortex_Socket #() vx_socket (
|
||||||
.clk (clk),
|
.clk (clk),
|
||||||
.reset (vx_reset),
|
.reset (vx_reset),
|
||||||
@@ -761,23 +803,23 @@ Vortex_Socket #() vx_socket (
|
|||||||
.snp_rsp_ready (vx_snp_rsp_ready),
|
.snp_rsp_ready (vx_snp_rsp_ready),
|
||||||
|
|
||||||
// I/O request
|
// I/O request
|
||||||
.io_req_valid (),
|
`UNUSED_PIN (io_req_valid),
|
||||||
.io_req_rw (),
|
`UNUSED_PIN (io_req_rw),
|
||||||
.io_req_byteen (),
|
`UNUSED_PIN (io_req_byteen),
|
||||||
.io_req_addr (),
|
`UNUSED_PIN (io_req_addr),
|
||||||
.io_req_data (),
|
`UNUSED_PIN (io_req_data),
|
||||||
.io_req_tag (),
|
`UNUSED_PIN (io_req_tag),
|
||||||
.io_req_ready (1),
|
.io_req_ready (1),
|
||||||
|
|
||||||
// I/O response
|
// I/O response
|
||||||
.io_rsp_valid (0),
|
.io_rsp_valid (0),
|
||||||
.io_rsp_data (0),
|
.io_rsp_data (0),
|
||||||
.io_rsp_tag (0),
|
.io_rsp_tag (0),
|
||||||
.io_rsp_ready (),
|
`UNUSED_PIN (io_rsp_ready),
|
||||||
|
|
||||||
// status
|
// status
|
||||||
.busy (vx_busy),
|
.busy (vx_busy),
|
||||||
.ebreak ()
|
`UNUSED_PIN (ebreak)
|
||||||
);
|
);
|
||||||
|
|
||||||
endmodule
|
endmodule
|
||||||
@@ -48,6 +48,7 @@
|
|||||||
`define CLOG2(x) $clog2(x)
|
`define CLOG2(x) $clog2(x)
|
||||||
`define FLOG2(x) ($clog2(x) - (((1 << $clog2(x)) > (x)) ? 1 : 0))
|
`define FLOG2(x) ($clog2(x) - (((1 << $clog2(x)) > (x)) ? 1 : 0))
|
||||||
`define LOG2UP(x) (((x) > 1) ? $clog2(x) : 1)
|
`define LOG2UP(x) (((x) > 1) ? $clog2(x) : 1)
|
||||||
|
`define ISPOW2(x) (((x) != 0) && (0 == ((x) & ((x) - 1))))
|
||||||
|
|
||||||
`define MIN(x, y) ((x < y) ? (x) : (y))
|
`define MIN(x, y) ((x < y) ? (x) : (y))
|
||||||
`define MAX(x, y) ((x > y) ? (x) : (y))
|
`define MAX(x, y) ((x > y) ? (x) : (y))
|
||||||
|
|||||||
@@ -70,7 +70,7 @@ module VX_lsu_unit #(
|
|||||||
|
|
||||||
for (i = 0; i < `NUM_THREADS; ++i) begin
|
for (i = 0; i < `NUM_THREADS; ++i) begin
|
||||||
assign mem_req_addr[i] = use_address[i][31:2];
|
assign mem_req_addr[i] = use_address[i][31:2];
|
||||||
assign mem_req_offset[i] = {3'b0, use_address[i][1:0]} << 3;
|
assign mem_req_offset[i] = (5'(use_address[i][1:0])) << 3;
|
||||||
assign mem_req_byteen[i] = (wmask << use_address[i][1:0]);
|
assign mem_req_byteen[i] = (wmask << use_address[i][1:0]);
|
||||||
assign mem_req_data[i] = (use_store_data[i] << mem_req_offset[i]);
|
assign mem_req_data[i] = (use_store_data[i] << mem_req_offset[i]);
|
||||||
end
|
end
|
||||||
|
|||||||
@@ -123,7 +123,7 @@ module Vortex #(
|
|||||||
assign io_req_tag = io_core_req_if.core_req_tag[0];
|
assign io_req_tag = io_core_req_if.core_req_tag[0];
|
||||||
assign io_core_req_if.core_req_ready = io_req_ready;
|
assign io_core_req_if.core_req_ready = io_req_ready;
|
||||||
|
|
||||||
assign io_core_rsp_if.core_rsp_valid = {{`NUM_THREADS-1{1'b0}}, io_rsp_valid};
|
assign io_core_rsp_if.core_rsp_valid = {{(`NUM_THREADS-1){1'b0}}, io_rsp_valid};
|
||||||
assign io_core_rsp_if.core_rsp_data[0] = io_rsp_data;
|
assign io_core_rsp_if.core_rsp_data[0] = io_rsp_data;
|
||||||
assign io_core_rsp_if.core_rsp_tag = io_rsp_tag;
|
assign io_core_rsp_if.core_rsp_tag = io_rsp_tag;
|
||||||
assign io_rsp_ready = io_core_rsp_if.core_rsp_ready;
|
assign io_rsp_ready = io_core_rsp_if.core_rsp_ready;
|
||||||
|
|||||||
16
hw/rtl/cache/VX_bank.v
vendored
16
hw/rtl/cache/VX_bank.v
vendored
@@ -718,12 +718,12 @@ module VX_bank #(
|
|||||||
`ifdef DBG_PRINT_CACHE_BANK
|
`ifdef DBG_PRINT_CACHE_BANK
|
||||||
if (NUM_BANKS == 1) begin
|
if (NUM_BANKS == 1) begin
|
||||||
always_ff @(posedge clk) begin
|
always_ff @(posedge clk) begin
|
||||||
/*if (core_req_valid && core_req_ready) begin
|
if (core_req_valid && core_req_ready) begin
|
||||||
$display("%t: bank%01d%01d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(core_req_addr), core_req_tag);
|
$display("%t: bank%01d%01d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(core_req_addr), core_req_tag);
|
||||||
end
|
end
|
||||||
if (core_rsp_valid && core_rsp_ready) begin
|
if (core_rsp_valid && core_rsp_ready) begin
|
||||||
$display("%t: bank%01d%01d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
$display("%t: bank%01d%01d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
||||||
end*/
|
end
|
||||||
if (dram_fill_req_valid && dram_fill_req_ready) begin
|
if (dram_fill_req_valid && dram_fill_req_ready) begin
|
||||||
$display("%t: bank%01d%01d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_req_addr));
|
$display("%t: bank%01d%01d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_req_addr));
|
||||||
end
|
end
|
||||||
@@ -733,21 +733,21 @@ module VX_bank #(
|
|||||||
if (dram_fill_rsp_valid && dram_fill_rsp_ready) begin
|
if (dram_fill_rsp_valid && dram_fill_rsp_ready) begin
|
||||||
$display("%t: bank%01d%01d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_rsp_addr), dram_fill_rsp_data);
|
$display("%t: bank%01d%01d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_rsp_addr), dram_fill_rsp_data);
|
||||||
end
|
end
|
||||||
/*if (snp_req_valid && snp_req_ready) begin
|
if (snp_req_valid && snp_req_ready) begin
|
||||||
$display("%t: bank%01d%01d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(snp_req_addr), snp_req_tag);
|
$display("%t: bank%01d%01d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(snp_req_addr), snp_req_tag);
|
||||||
end
|
end
|
||||||
if (snp_rsp_valid && snp_rsp_ready) begin
|
if (snp_rsp_valid && snp_rsp_ready) begin
|
||||||
$display("%t: bank%01d%01d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
$display("%t: bank%01d%01d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
||||||
end*/
|
end
|
||||||
end
|
end
|
||||||
end else begin
|
end else begin
|
||||||
always_ff @(posedge clk) begin
|
always_ff @(posedge clk) begin
|
||||||
/*if ((|core_req_valid) && core_req_ready) begin
|
if ((|core_req_valid) && core_req_ready) begin
|
||||||
$display("%t: bank%01d%01d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(core_req_addr, BANK_ID), core_req_tag);
|
$display("%t: bank%01d%01d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(core_req_addr, BANK_ID), core_req_tag);
|
||||||
end
|
end
|
||||||
if (core_rsp_valid && core_rsp_ready) begin
|
if (core_rsp_valid && core_rsp_ready) begin
|
||||||
$display("%t: bank%01d%01d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
$display("%t: bank%01d%01d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
||||||
end*/
|
end
|
||||||
if (dram_fill_req_valid && dram_fill_req_ready) begin
|
if (dram_fill_req_valid && dram_fill_req_ready) begin
|
||||||
$display("%t: bank%01d%01d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_req_addr, BANK_ID));
|
$display("%t: bank%01d%01d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_req_addr, BANK_ID));
|
||||||
end
|
end
|
||||||
@@ -757,12 +757,12 @@ module VX_bank #(
|
|||||||
if (dram_fill_rsp_valid && dram_fill_rsp_ready) begin
|
if (dram_fill_rsp_valid && dram_fill_rsp_ready) begin
|
||||||
$display("%t: bank%01d%01d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_rsp_addr, BANK_ID), dram_fill_rsp_data);
|
$display("%t: bank%01d%01d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_rsp_addr, BANK_ID), dram_fill_rsp_data);
|
||||||
end
|
end
|
||||||
/*if (snp_req_valid && snp_req_ready) begin
|
if (snp_req_valid && snp_req_ready) begin
|
||||||
$display("%t: bank%01d%01d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(snp_req_addr, BANK_ID), snp_req_tag);
|
$display("%t: bank%01d%01d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(snp_req_addr, BANK_ID), snp_req_tag);
|
||||||
end
|
end
|
||||||
if (snp_rsp_valid && snp_rsp_ready) begin
|
if (snp_rsp_valid && snp_rsp_ready) begin
|
||||||
$display("%t: bank%01d%01d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
$display("%t: bank%01d%01d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
||||||
end*/
|
end
|
||||||
end
|
end
|
||||||
end
|
end
|
||||||
`endif
|
`endif
|
||||||
|
|||||||
3
hw/rtl/cache/VX_cache_core_rsp_merge.v
vendored
3
hw/rtl/cache/VX_cache_core_rsp_merge.v
vendored
@@ -46,6 +46,7 @@ module VX_cache_core_rsp_merge #(
|
|||||||
assign core_rsp_tag = per_bank_core_rsp_tag[main_bank_index];
|
assign core_rsp_tag = per_bank_core_rsp_tag[main_bank_index];
|
||||||
always @(*) begin
|
always @(*) begin
|
||||||
core_rsp_valid = 0;
|
core_rsp_valid = 0;
|
||||||
|
core_rsp_data = 0;
|
||||||
for (i = 0; i < NUM_BANKS; i++) begin
|
for (i = 0; i < NUM_BANKS; i++) begin
|
||||||
if (per_bank_core_rsp_valid[i]
|
if (per_bank_core_rsp_valid[i]
|
||||||
&& (per_bank_core_rsp_tag[i][CORE_TAG_ID_BITS-1:0] == per_bank_core_rsp_tag[main_bank_index][CORE_TAG_ID_BITS-1:0])) begin
|
&& (per_bank_core_rsp_tag[i][CORE_TAG_ID_BITS-1:0] == per_bank_core_rsp_tag[main_bank_index][CORE_TAG_ID_BITS-1:0])) begin
|
||||||
@@ -60,6 +61,8 @@ module VX_cache_core_rsp_merge #(
|
|||||||
end else begin
|
end else begin
|
||||||
always @(*) begin
|
always @(*) begin
|
||||||
core_rsp_valid = 0;
|
core_rsp_valid = 0;
|
||||||
|
core_rsp_data = 0;
|
||||||
|
core_rsp_tag = 0;
|
||||||
for (i = 0; i < NUM_BANKS; i++) begin
|
for (i = 0; i < NUM_BANKS; i++) begin
|
||||||
if (per_bank_core_rsp_valid[i]
|
if (per_bank_core_rsp_valid[i]
|
||||||
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
|
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
|
||||||
|
|||||||
2
hw/rtl/cache/VX_snp_forwarder.v
vendored
2
hw/rtl/cache/VX_snp_forwarder.v
vendored
@@ -97,7 +97,7 @@ module VX_snp_forwarder #(
|
|||||||
always @(posedge clk) begin
|
always @(posedge clk) begin
|
||||||
if (reset) begin
|
if (reset) begin
|
||||||
fwdin_sel <= 0;
|
fwdin_sel <= 0;
|
||||||
end else begin
|
end else if (NUM_REQUESTS > 1) begin
|
||||||
fwdin_sel <= fwdin_sel + 1;
|
fwdin_sel <= fwdin_sel + 1;
|
||||||
end
|
end
|
||||||
end
|
end
|
||||||
|
|||||||
2
hw/rtl/cache/VX_tag_data_access.v
vendored
2
hw/rtl/cache/VX_tag_data_access.v
vendored
@@ -136,7 +136,7 @@ module VX_tag_data_access #(
|
|||||||
end
|
end
|
||||||
|
|
||||||
assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || ~DRAM_ENABLE; // If shared memory, always valid
|
assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || ~DRAM_ENABLE; // If shared memory, always valid
|
||||||
assign use_read_dirty_st1e = read_dirty_st1c[STAGE_1_CYCLES-1] && DRAM_ENABLE; // Dirty only applies in Dcache
|
assign use_read_dirty_st1e = read_dirty_st1c[STAGE_1_CYCLES-1] && DRAM_ENABLE && WRITE_ENABLE; // Dirty only applies in Dcache
|
||||||
assign use_read_tag_st1e = DRAM_ENABLE ? read_tag_st1c[STAGE_1_CYCLES-1] : writetag_st1e; // Tag is always the same in SM
|
assign use_read_tag_st1e = DRAM_ENABLE ? read_tag_st1c[STAGE_1_CYCLES-1] : writetag_st1e; // Tag is always the same in SM
|
||||||
assign use_read_dirtyb_st1e= read_dirtyb_st1c[STAGE_1_CYCLES-1];
|
assign use_read_dirtyb_st1e= read_dirtyb_st1c[STAGE_1_CYCLES-1];
|
||||||
assign use_read_data_st1e = read_data_st1c[STAGE_1_CYCLES-1];
|
assign use_read_data_st1e = read_data_st1c[STAGE_1_CYCLES-1];
|
||||||
|
|||||||
3
hw/rtl/cache/VX_tag_data_structure.v
vendored
3
hw/rtl/cache/VX_tag_data_structure.v
vendored
@@ -33,7 +33,7 @@ module VX_tag_data_structure #(
|
|||||||
reg [`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0][7:0] data [`BANK_LINE_COUNT-1:0];
|
reg [`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0][7:0] data [`BANK_LINE_COUNT-1:0];
|
||||||
reg [`TAG_SELECT_BITS-1:0] tag [`BANK_LINE_COUNT-1:0];
|
reg [`TAG_SELECT_BITS-1:0] tag [`BANK_LINE_COUNT-1:0];
|
||||||
reg [`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0] dirtyb[`BANK_LINE_COUNT-1:0];
|
reg [`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0] dirtyb[`BANK_LINE_COUNT-1:0];
|
||||||
reg dirty[`BANK_LINE_COUNT-1:0];
|
reg [`BANK_LINE_COUNT-1:0] dirty;
|
||||||
reg [`BANK_LINE_COUNT-1:0] valid;
|
reg [`BANK_LINE_COUNT-1:0] valid;
|
||||||
|
|
||||||
assign read_valid = valid [read_addr];
|
assign read_valid = valid [read_addr];
|
||||||
@@ -49,6 +49,7 @@ module VX_tag_data_structure #(
|
|||||||
if (reset) begin
|
if (reset) begin
|
||||||
for (i = 0; i < `BANK_LINE_COUNT; i++) begin
|
for (i = 0; i < `BANK_LINE_COUNT; i++) begin
|
||||||
valid[i] <= 0;
|
valid[i] <= 0;
|
||||||
|
dirty[i] <= 0;
|
||||||
end
|
end
|
||||||
end else if (!stall_bank_pipe) begin
|
end else if (!stall_bank_pipe) begin
|
||||||
if (do_write) begin
|
if (do_write) begin
|
||||||
|
|||||||
@@ -15,6 +15,8 @@ module VX_generic_queue #(
|
|||||||
output wire full,
|
output wire full,
|
||||||
output wire [`LOG2UP(SIZE+1)-1:0] size
|
output wire [`LOG2UP(SIZE+1)-1:0] size
|
||||||
);
|
);
|
||||||
|
`STATIC_ASSERT(0 == SIZE || `ISPOW2(SIZE), "must be 0 or power of 2!");
|
||||||
|
|
||||||
if (SIZE == 0) begin
|
if (SIZE == 0) begin
|
||||||
|
|
||||||
assign empty = 1;
|
assign empty = 1;
|
||||||
@@ -88,6 +90,7 @@ module VX_generic_queue #(
|
|||||||
if (writing) begin
|
if (writing) begin
|
||||||
data[wr_ptr_a] <= data_in;
|
data[wr_ptr_a] <= data_in;
|
||||||
wr_ptr_r <= wr_ptr_r + 1;
|
wr_ptr_r <= wr_ptr_r + 1;
|
||||||
|
|
||||||
if (!reading) begin
|
if (!reading) begin
|
||||||
size_r <= size_r + 1;
|
size_r <= size_r + 1;
|
||||||
end
|
end
|
||||||
@@ -120,16 +123,17 @@ module VX_generic_queue #(
|
|||||||
|
|
||||||
always @(posedge clk) begin
|
always @(posedge clk) begin
|
||||||
if (reset) begin
|
if (reset) begin
|
||||||
size_r <= 0;
|
|
||||||
empty_r <= 1;
|
|
||||||
full_r <= 0;
|
|
||||||
wr_ptr_r <= 0;
|
wr_ptr_r <= 0;
|
||||||
rd_ptr_r <= 0;
|
rd_ptr_r <= 0;
|
||||||
rd_ptr_next_r <= 1;
|
rd_ptr_next_r <= 1;
|
||||||
|
empty_r <= 1;
|
||||||
|
full_r <= 0;
|
||||||
|
size_r <= 0;
|
||||||
end else begin
|
end else begin
|
||||||
if (writing) begin
|
if (writing) begin
|
||||||
data[wr_ptr_r] <= data_in;
|
data[wr_ptr_r] <= data_in;
|
||||||
wr_ptr_r <= wr_ptr_r + 1;
|
wr_ptr_r <= wr_ptr_r + 1;
|
||||||
|
|
||||||
if (!reading) begin
|
if (!reading) begin
|
||||||
empty_r <= 0;
|
empty_r <= 0;
|
||||||
if (size_r == SIZE-1) begin
|
if (size_r == SIZE-1) begin
|
||||||
@@ -140,15 +144,17 @@ module VX_generic_queue #(
|
|||||||
end
|
end
|
||||||
|
|
||||||
if (reading) begin
|
if (reading) begin
|
||||||
rd_ptr_r <= rd_ptr_next_r;
|
rd_ptr_r <= rd_ptr_next_r;
|
||||||
if (SIZE == 2) begin
|
|
||||||
rd_ptr_next_r <= ~rd_ptr_next_r;
|
if (SIZE > 2) begin
|
||||||
end else if (SIZE > 2) begin
|
|
||||||
rd_ptr_next_r <= rd_ptr_r + 2;
|
rd_ptr_next_r <= rd_ptr_r + 2;
|
||||||
|
end else begin // (SIZE == 2);
|
||||||
|
rd_ptr_next_r <= ~rd_ptr_next_r;
|
||||||
end
|
end
|
||||||
|
|
||||||
if (!writing) begin
|
if (!writing) begin
|
||||||
if (size_r == 1) begin
|
if (size_r == 1) begin
|
||||||
|
assert(rd_ptr_next_r == wr_ptr_r);
|
||||||
empty_r <= 1;
|
empty_r <= 1;
|
||||||
end;
|
end;
|
||||||
full_r <= 0;
|
full_r <= 0;
|
||||||
@@ -156,7 +162,9 @@ module VX_generic_queue #(
|
|||||||
end
|
end
|
||||||
end
|
end
|
||||||
|
|
||||||
bypass_r <= writing && (empty_r || (1 == size_r) && reading);
|
bypass_r <= writing
|
||||||
|
&& (empty_r || ((1 == size_r) && reading)); // empty or about to go empty
|
||||||
|
|
||||||
curr_r <= data_in;
|
curr_r <= data_in;
|
||||||
head_r <= data[reading ? rd_ptr_next_r : rd_ptr_r];
|
head_r <= data[reading ? rd_ptr_next_r : rd_ptr_r];
|
||||||
end
|
end
|
||||||
|
|||||||
@@ -10,9 +10,8 @@ double sc_time_stamp() {
|
|||||||
}
|
}
|
||||||
|
|
||||||
Simulator::Simulator() {
|
Simulator::Simulator() {
|
||||||
// force random values for unitialized signals
|
// force random values for unitialized signals
|
||||||
const char* args[] = {"", "+verilator+rand+reset+2", "+verilator+seed+50"};
|
Verilated::randReset(2);
|
||||||
Verilated::commandArgs(3, args);
|
|
||||||
|
|
||||||
ram_ = nullptr;
|
ram_ = nullptr;
|
||||||
vortex_ = new VVortex_Socket();
|
vortex_ = new VVortex_Socket();
|
||||||
|
|||||||
Reference in New Issue
Block a user