reset networks optimization

This commit is contained in:
Blaise Tine
2020-11-16 01:12:02 -08:00
parent 1bc4b8e7a8
commit a1fcdd467a
14 changed files with 259 additions and 219 deletions

View File

@@ -20,19 +20,19 @@ install:
- export PATH=$VERILATOR_ROOT/bin:$PATH
script:
- make -j
- make -j > /dev/null 2>&1
- ci/test_runtime.sh
- ci/test_driver.sh
- ci/test_riscv_isa.sh
- ci/test_opencl.sh
- ci/blackbox.sh -run_debug
- ci/blackbox.sh -run_scope
- ci/blackbox.sh -run_1c
- ci/blackbox.sh -run_2c
- ci/blackbox.sh -run_4c
- ci/blackbox.sh -run_4c_l2
- ci/blackbox.sh -run_8c_2l2
- ci/blackbox.sh -run_16c_4l2_l3
- ci/blackbox.sh -run_debug
- ci/blackbox.sh -run_scope
after_success:
# Gather code coverage

View File

@@ -4,7 +4,7 @@ run_1c()
{
# test single core
make -C driver/opae/vlsim clean
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=1" make -C driver/opae/vlsim
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=1" make -C driver/opae/vlsim > /dev/null 2>&1
make -C driver/tests/dogfood run-vlsim
make -C benchmarks/opencl/sgemm run-vlsim
}
@@ -13,7 +13,7 @@ run_2c()
{
# test 2 cores
make -C driver/opae/vlsim clean
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0" make -C driver/opae/vlsim
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0" make -C driver/opae/vlsim > /dev/null 2>&1
make -C driver/tests/dogfood run-vlsim
make -C benchmarks/opencl/sgemm run-vlsim
}
@@ -22,7 +22,7 @@ run_4c()
{
# test 4 cores
make -C driver/opae/vlsim clean
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=0" make -C driver/opae/vlsim
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=0" make -C driver/opae/vlsim > /dev/null 2>&1
make -C driver/tests/dogfood run-vlsim
make -C benchmarks/opencl/sgemm run-vlsim
}
@@ -31,7 +31,7 @@ run_4c_l2()
{
# test 4 cores with L2
make -C driver/opae/vlsim clean
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=1" make -C driver/opae/vlsim
CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=1" make -C driver/opae/vlsim > /dev/null 2>&1
make -C driver/tests/dogfood run-vlsim
make -C benchmarks/opencl/sgemm run-vlsim
}
@@ -40,7 +40,7 @@ run_8c_2l2()
{
# test 8 cores with 2xL2
make -C driver/opae/vlsim clean
CONFIGS="-DNUM_CLUSTERS=2 -DNUM_CORES=4 -DL2_ENABLE=1" make -C driver/opae/vlsim
CONFIGS="-DNUM_CLUSTERS=2 -DNUM_CORES=4 -DL2_ENABLE=1" make -C driver/opae/vlsim > /dev/null 2>&1
make -C driver/tests/dogfood run-vlsim
make -C benchmarks/opencl/sgemm run-vlsim
}
@@ -49,7 +49,7 @@ run_16c_4l2_l3()
{
# test 16 cores with L2 and L3
make -C driver/opae/vlsim clean
CONFIGS="-DNUM_CLUSTERS=4 -DNUM_CORES=4 -DL2_ENABLE=1 -DL3_ENABLE=1" make -C driver/opae/vlsim
CONFIGS="-DNUM_CLUSTERS=4 -DNUM_CORES=4 -DL2_ENABLE=1 -DL3_ENABLE=1" make -C driver/opae/vlsim > /dev/null 2>&1
make -C driver/tests/dogfood run-vlsim
make -C benchmarks/opencl/sgemm run-vlsim
}
@@ -58,7 +58,7 @@ run_debug()
{
# test debug build
make -C driver/opae/vlsim clean
DEBUG=1 CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=1" make -C driver/opae/vlsim
DEBUG=1 CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=1" make -C driver/opae/vlsim > /dev/null 2>&1
make -C driver/tests/demo run-vlsim
}
@@ -66,7 +66,7 @@ run_scope()
{
# test build with scope analyzer
make -C driver/opae clean
SCOPE=1 CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=1" make -C driver/opae
SCOPE=1 CONFIGS="-DNUM_CLUSTERS=1 -DNUM_CORES=1" make -C driver/opae > /dev/null 2>&1
make -C driver/tests/demo run-vlsim
}

View File

@@ -115,12 +115,12 @@ wire vx_dram_rsp_ready;
reg vx_snp_req_valid;
reg [`VX_DRAM_ADDR_WIDTH-1:0] vx_snp_req_addr;
wire vx_snp_req_invalidate = 0;
reg [`VX_SNP_TAG_WIDTH-1:0] vx_snp_req_tag;
wire [`VX_SNP_TAG_WIDTH-1:0] vx_snp_req_tag;
wire vx_snp_req_ready;
reg vx_snp_rsp_valid;
wire vx_snp_rsp_valid;
`DEBUG_BEGIN
reg [`VX_SNP_TAG_WIDTH-1:0] vx_snp_rsp_tag;
wire [`VX_SNP_TAG_WIDTH-1:0] vx_snp_rsp_tag;
`DEBUG_END
reg vx_snp_rsp_ready;
@@ -135,9 +135,10 @@ wire vx_csr_io_rsp_valid;
wire [31:0] vx_csr_io_rsp_data;
wire vx_csr_io_rsp_ready;
wire vx_busy;
reg vx_reset;
reg vx_enabled;
wire vx_busy;
// CMD variables //////////////////////////////////////////////////////////////
@@ -208,127 +209,125 @@ always @(posedge clk) begin
`ifndef VERILATOR
$asserton; // enable assertions
`endif
mmio_tx.hdr <= 0;
mmio_tx.mmioRdValid <= 0;
mmio_tx.hdr <= 0;
`ifdef SCOPE
scope_start <= 0;
`endif
end else begin
mmio_tx.mmioRdValid <= cp2af_sRxPort.c0.mmioRdValid;
mmio_tx.hdr.tid <= mmio_hdr.tid;
`ifdef SCOPE
scope_start <= cp2af_sRxPort.c0.mmioWrValid;
`endif
end
else begin
mmio_tx.mmioRdValid <= 0;
// serve MMIO write request
if (cp2af_sRxPort.c0.mmioWrValid)
begin
`ifdef SCOPE
scope_start <= 1;
`endif
case (mmio_hdr.address)
MMIO_IO_ADDR: begin
cmd_io_addr <= t_ccip_clAddr'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_IO_ADDR: addr=%0h, data=0x%0h", $time, mmio_hdr.address, t_ccip_clAddr'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_MEM_ADDR: begin
cmd_mem_addr <= t_local_mem_addr'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_MEM_ADDR: addr=%0h, data=0x%0h", $time, mmio_hdr.address, t_local_mem_addr'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_DATA_SIZE: begin
cmd_data_size <= $bits(cmd_data_size)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_DATA_SIZE: addr=%0h, data=%0d", $time, mmio_hdr.address, $bits(cmd_data_size)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_CMD_TYPE: begin
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CMD_TYPE: addr=%0h, data=%0d", $time, mmio_hdr.address, $bits(cmd_type)'(cp2af_sRxPort.c0.data));
`endif
end
`ifdef SCOPE
MMIO_SCOPE_WRITE: begin
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_SCOPE_WRITE: addr=%0h, data=%0h", $time, mmio_hdr.address, 64'(cp2af_sRxPort.c0.data));
`endif
end
`endif
MMIO_CSR_CORE: begin
cmd_csr_core <= $bits(cmd_csr_core)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_CORE: addr=%0h, %0h", $time, mmio_hdr.address, $bits(cmd_csr_core)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_CSR_ADDR: begin
cmd_csr_addr <= $bits(cmd_csr_addr)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_ADDR: addr=%0h, %0h", $time, mmio_hdr.address, $bits(cmd_csr_addr)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_CSR_DATA: begin
cmd_csr_wdata <= $bits(cmd_csr_wdata)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_DATA: addr=%0h, %0h", $time, mmio_hdr.address, $bits(cmd_csr_wdata)'(cp2af_sRxPort.c0.data));
`endif
end
default: begin
`ifdef DBG_PRINT_OPAE
$display("%t: Unknown MMIO Wr: addr=%0h, data=%0h", $time, mmio_hdr.address, $bits(cmd_csr_wdata)'(cp2af_sRxPort.c0.data));
`endif
end
endcase
end
// serve MMIO read requests
if (cp2af_sRxPort.c0.mmioRdValid) begin
mmio_tx.hdr.tid <= mmio_hdr.tid; // copy TID
case (mmio_hdr.address)
// AFU header
16'h0000: mmio_tx.data <= {
4'b0001, // Feature type = AFU
8'b0, // reserved
4'b0, // afu minor revision = 0
7'b0, // reserved
1'b1, // end of DFH list = 1
24'b0, // next DFH offset = 0
4'b0, // afu major revision = 0
12'b0 // feature ID = 0
};
AFU_ID_L: mmio_tx.data <= afu_id[63:0]; // afu id low
AFU_ID_H: mmio_tx.data <= afu_id[127:64]; // afu id hi
16'h0006: mmio_tx.data <= 64'h0; // next AFU
16'h0008: mmio_tx.data <= 64'h0; // reserved
MMIO_STATUS: begin
mmio_tx.data <= 64'(state);
// serve MMIO write request
if (cp2af_sRxPort.c0.mmioWrValid) begin
case (mmio_hdr.address)
MMIO_IO_ADDR: begin
cmd_io_addr <= t_ccip_clAddr'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_IO_ADDR: addr=%0h, data=0x%0h", $time, mmio_hdr.address, t_ccip_clAddr'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_MEM_ADDR: begin
cmd_mem_addr <= t_local_mem_addr'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_MEM_ADDR: addr=%0h, data=0x%0h", $time, mmio_hdr.address, t_local_mem_addr'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_DATA_SIZE: begin
cmd_data_size <= $bits(cmd_data_size)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_DATA_SIZE: addr=%0h, data=%0d", $time, mmio_hdr.address, $bits(cmd_data_size)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_CMD_TYPE: begin
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CMD_TYPE: addr=%0h, data=%0d", $time, mmio_hdr.address, $bits(cmd_type)'(cp2af_sRxPort.c0.data));
`endif
end
`ifdef SCOPE
MMIO_SCOPE_WRITE: begin
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_SCOPE_WRITE: addr=%0h, data=%0h", $time, mmio_hdr.address, 64'(cp2af_sRxPort.c0.data));
`endif
end
`endif
MMIO_CSR_CORE: begin
cmd_csr_core <= $bits(cmd_csr_core)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_CORE: addr=%0h, %0h", $time, mmio_hdr.address, $bits(cmd_csr_core)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_CSR_ADDR: begin
cmd_csr_addr <= $bits(cmd_csr_addr)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_ADDR: addr=%0h, %0h", $time, mmio_hdr.address, $bits(cmd_csr_addr)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_CSR_DATA: begin
cmd_csr_wdata <= $bits(cmd_csr_wdata)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_DATA: addr=%0h, %0h", $time, mmio_hdr.address, $bits(cmd_csr_wdata)'(cp2af_sRxPort.c0.data));
`endif
end
default: begin
`ifdef DBG_PRINT_OPAE
if (state != STATE_WIDTH'(mmio_tx.data)) begin
$display("%t: MMIO_STATUS: addr=%0h, state=%0d", $time, mmio_hdr.address, state);
end
`endif
end
MMIO_CSR_READ: begin
mmio_tx.data <= 64'(cmd_csr_rdata);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_READ: addr=%0h, data=%0h", $time, mmio_hdr.address, cmd_csr_rdata);
`endif
end
`ifdef SCOPE
MMIO_SCOPE_READ: begin
mmio_tx.data <= cmd_scope_rdata;
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_SCOPE_READ: addr=%0h, data=%0h", $time, mmio_hdr.address, cmd_scope_rdata);
$display("%t: Unknown MMIO Wr: addr=%0h, data=%0h", $time, mmio_hdr.address, $bits(cmd_csr_wdata)'(cp2af_sRxPort.c0.data));
`endif
end
endcase
end
// serve MMIO read requests
if (cp2af_sRxPort.c0.mmioRdValid) begin
case (mmio_hdr.address)
// AFU header
16'h0000: mmio_tx.data <= {
4'b0001, // Feature type = AFU
8'b0, // reserved
4'b0, // afu minor revision = 0
7'b0, // reserved
1'b1, // end of DFH list = 1
24'b0, // next DFH offset = 0
4'b0, // afu major revision = 0
12'b0 // feature ID = 0
};
AFU_ID_L: mmio_tx.data <= afu_id[63:0]; // afu id low
AFU_ID_H: mmio_tx.data <= afu_id[127:64]; // afu id hi
16'h0006: mmio_tx.data <= 64'h0; // next AFU
16'h0008: mmio_tx.data <= 64'h0; // reserved
MMIO_STATUS: begin
mmio_tx.data <= 64'(state);
`ifdef DBG_PRINT_OPAE
if (state != STATE_WIDTH'(mmio_tx.data)) begin
$display("%t: MMIO_STATUS: addr=%0h, state=%0d", $time, mmio_hdr.address, state);
end
`endif
default: begin
mmio_tx.data <= 64'h0;
`ifdef DBG_PRINT_OPAE
$display("%t: Unknown MMIO Rd: addr=%0h", $time, mmio_hdr.address);
`endif
end
endcase
mmio_tx.mmioRdValid <= 1; // post response
end
end
MMIO_CSR_READ: begin
mmio_tx.data <= 64'(cmd_csr_rdata);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CSR_READ: addr=%0h, data=%0h", $time, mmio_hdr.address, cmd_csr_rdata);
`endif
end
`ifdef SCOPE
MMIO_SCOPE_READ: begin
mmio_tx.data <= cmd_scope_rdata;
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_SCOPE_READ: addr=%0h, data=%0h", $time, mmio_hdr.address, cmd_scope_rdata);
`endif
end
`endif
default: begin
mmio_tx.data <= 64'h0;
`ifdef DBG_PRINT_OPAE
$display("%t: Unknown MMIO Rd: addr=%0h", $time, mmio_hdr.address);
`endif
end
endcase
end
end
@@ -687,6 +686,7 @@ always @(posedge clk) begin
cci_rd_req_enable <= 0;
cci_rd_req_wait <= 0;
cci_dram_wr_req_ctr <= 0;
cci_dram_wr_req_addr_unqual <= 0;
end
else begin
if ((STATE_IDLE == state)
@@ -819,6 +819,7 @@ begin
cci_wr_req_ctr <= 0;
cci_pending_writes <= 0;
cci_dram_rd_req_ctr <= 0;
cci_dram_rd_req_addr_unqual <= 0;
end
else begin
if ((STATE_IDLE == state)
@@ -871,6 +872,8 @@ end else begin
assign snp_req_size = cmd_data_size;
end
assign vx_snp_req_tag = (`VX_SNP_TAG_WIDTH)'(snp_req_ctr);
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;
@@ -886,17 +889,14 @@ always @(posedge clk) begin
vx_snp_rsp_ready <= 0;
snp_req_ctr <= 0;
snp_rsp_ctr <= 0;
end
else begin
end else begin
if ((STATE_IDLE == state)
&& (CMD_CLFLUSH == cmd_type)) begin
vx_snp_req_addr <= snp_req_baseaddr;
vx_snp_req_tag <= 0;
snp_req_ctr <= 0;
snp_rsp_ctr <= snp_req_size;
vx_snp_req_valid <= (snp_req_size != 0);
vx_snp_req_addr <= snp_req_baseaddr;
vx_snp_rsp_ready <= (snp_req_size != 0);
snp_req_ctr <= 0;
snp_rsp_ctr <= snp_req_size;
end
if ((STATE_CLFLUSH == state)
@@ -909,11 +909,9 @@ always @(posedge clk) begin
vx_snp_rsp_ready <= 0;
end
if (vx_snp_req_fire)
begin
if (vx_snp_req_fire) begin
assert(snp_req_ctr < snp_req_size);
vx_snp_req_addr <= vx_snp_req_addr + `VX_DRAM_ADDR_WIDTH'(1);
vx_snp_req_tag <= (`VX_SNP_TAG_WIDTH)'(snp_req_ctr_next);
snp_req_ctr <= snp_req_ctr_next;
`ifdef DBG_PRINT_OPAE
$display("%t: AFU Snp Req: addr=%0h, tag=%0h, rem=%0d", $time, `DRAM_TO_BYTE_ADDR(vx_snp_req_addr), (`VX_SNP_TAG_WIDTH)'(vx_snp_req_tag), (snp_req_size - snp_req_ctr_next));
@@ -949,19 +947,19 @@ assign cmd_csr_done = (STATE_CSR_WRITE == state) ? vx_csr_io_req_ready : vx_csr_
always @(posedge clk) begin
if (reset) begin
csr_io_req_sent <= 0;
end
else begin
end else begin
if (vx_csr_io_req_valid && vx_csr_io_req_ready) begin
csr_io_req_sent <= 1;
end
if (cmd_csr_done) begin
csr_io_req_sent <= 0;
end
if ((STATE_CSR_READ == state)
&& vx_csr_io_rsp_ready
&& vx_csr_io_rsp_valid) begin
cmd_csr_rdata <= vx_csr_io_rsp_data;
end
end
if ((STATE_CSR_READ == state)
&& vx_csr_io_rsp_ready
&& vx_csr_io_rsp_valid) begin
cmd_csr_rdata <= vx_csr_io_rsp_data;
end
end

View File

@@ -40,7 +40,6 @@ module VX_gpr_bypass #(
delayed_push <= push;
assert(!use_buffer2 || use_buffer);
if (pop) begin
buffer <= buffer2;
use_buffer <= use_buffer2;
use_buffer2 <= 0;
end
@@ -48,18 +47,29 @@ module VX_gpr_bypass #(
if (use_buffer) begin
assert(!use_buffer2); // full!
use_buffer <= 1;
if (pop) begin
buffer <= data_in;
end else begin
buffer2 <= data_in;
if (!pop) begin
use_buffer2 <= 1;
end
end else if (!pop) begin
buffer <= data_in;
use_buffer <= 1;
end
end
end
if (pop) begin
buffer <= buffer2;
end
if (delayed_push) begin
if (use_buffer) begin
if (pop) begin
buffer <= data_in;
end else begin
buffer2 <= data_in;
end
end else if (!pop) begin
buffer <= data_in;
end
end
end
assign data_out = use_buffer ? buffer : data_in;

View File

@@ -38,15 +38,16 @@ module VX_gpr_stage #(
always @(posedge clk) begin
if (reset) begin
rsp_valid <= 0;
rsp_valid <= 0;
end else begin
rsp_valid <= gpr_req_if.valid;
rsp_wid <= gpr_req_if.wid;
rsp_pc <= gpr_req_if.PC;
rs1_is_zero <= (0 == gpr_req_if.rs1);
rs2_is_zero <= (0 == gpr_req_if.rs2);
rsp_valid <= gpr_req_if.valid;
end
end
rsp_wid <= gpr_req_if.wid;
rsp_pc <= gpr_req_if.PC;
rs1_is_zero <= (0 == gpr_req_if.rs1);
rs2_is_zero <= (0 == gpr_req_if.rs2);
end
`ifdef EXT_F_ENABLE
@@ -62,16 +63,19 @@ module VX_gpr_stage #(
end else begin
if (rs3_delay) begin
read_rs3 <= 1;
save_rs3 <= 1;
end else if (read_fire) begin
read_rs3 <= 0;
end
if (save_rs3) begin
rs3_data <= rs1_data;
save_rs3 <= 0;
end
end
assert(!read_rs3 || rsp_wid == gpr_req_if.wid);
end
end
if (rs3_delay) begin
save_rs3 <= 1;
end
if (save_rs3) begin
rs3_data <= rs1_data;
save_rs3 <= 0;
end
end
assign raddr1 = {gpr_req_if.wid, (rs3_delay ? gpr_req_if.rs3 : gpr_req_if.rs1)};

View File

@@ -60,23 +60,20 @@ module VX_ibuffer #(
if (reset) begin
size_r[i] <= 0;
end else begin
if (writing) begin
if (is_slot0) begin
q_data_out[i] <= q_data_in;
end
if (!reading) begin
size_r[i] <= size_r[i] + SIZEW'(1);
end
if (writing && !reading) begin
size_r[i] <= size_r[i] + SIZEW'(1);
end
if (reading) begin
if (size_r[i] != 1) begin
q_data_out[i] <= q_data_prev[i];
end
if (!writing) begin
size_r[i] <= size_r[i] - SIZEW'(1);
end
if (reading && !writing) begin
size_r[i] <= size_r[i] - SIZEW'(1);
end
end
end
if (writing && is_slot0) begin
q_data_out[i] <= q_data_in;
end
if (reading && (size_r[i] != 1)) begin
q_data_out[i] <= q_data_prev[i];
end
end
assign q_full[i] = (size_r[i] == SIZE);

View File

@@ -126,14 +126,15 @@ module VX_bank_core_req_arb #(
end else begin
pop_mask[sel_idx] <= 1;
end
end
if ((0 == q_valids_cnt_r) || pop) begin
sel_tid <= sel_idx;
sel_byteen <= q_byteen[sel_idx];
sel_addr <= q_addr[sel_idx];
sel_writedata <= q_writedata[sel_idx];
end
end
if ((0 == q_valids_cnt_r) || pop) begin
sel_tid <= sel_idx;
sel_byteen <= q_byteen[sel_idx];
sel_addr <= q_addr[sel_idx];
sel_writedata <= q_writedata[sel_idx];
end
end
if (CORE_TAG_ID_BITS != 0) begin

View File

@@ -123,8 +123,7 @@ module VX_cache_miss_resrv #(
head_ptr <= 0;
tail_ptr <= 0;
size <= 0;
end else begin
end else begin
if (update_ready_st0) begin
ready_table <= ready_table | valid_address_match;
end
@@ -140,7 +139,6 @@ module VX_cache_miss_resrv #(
end else begin
valid_table[tail_ptr] <= 1;
ready_table[tail_ptr] <= enqueue_ready_st3;
addr_table[tail_ptr] <= enqueue_addr_st3;
tail_ptr <= tail_ptr + $bits(tail_ptr)'(1);
size <= size + $bits(size)'(1);
end
@@ -159,6 +157,12 @@ module VX_cache_miss_resrv #(
end
end
always @(posedge clk) begin
if (enqueue_st3 && !enqueue_msrq_st3) begin
addr_table[tail_ptr] <= enqueue_addr_st3;
end
end
VX_dp_ram #(
.DATAW(`MRVQ_METADATA_WIDTH),
.SIZE(MRVQ_SIZE),

View File

@@ -32,10 +32,13 @@ module VX_bypass_buffer #(
end
if (valid_in && ~ready_out) begin
assert(!buffer_valid);
buffer <= data_in;
buffer_valid <= 1;
end
end
if (valid_in && ~ready_out) begin
buffer <= data_in;
end
end
assign ready_in = ready_out || !buffer_valid;

View File

@@ -52,10 +52,6 @@ module VX_cam_buffer #(
full_r <= 1'b0;
write_addr_r <= ADDRW'(1'b0);
end else begin
if (acquire_slot) begin
assert(1 == free_slots[write_addr]) else $display("%t: inused slot at port %d", $time, write_addr);
entries[write_addr] <= write_data;
end
for (integer i = 0; i < CPORTS; i++) begin
if (release_slot[i]) begin
assert(0 == free_slots[release_addr[i]]) else $display("%t: freed slot at port %d", $time, release_addr[i]);
@@ -65,6 +61,11 @@ module VX_cam_buffer #(
write_addr_r <= free_index;
full_r <= ~free_valid;
end
if (acquire_slot) begin
assert(1 == free_slots[write_addr]) else $display("%t: inused slot at port %d", $time, write_addr);
entries[write_addr] <= write_data;
end
end
for (genvar i = 0; i < RPORTS; i++) begin

View File

@@ -42,7 +42,6 @@ module VX_index_queue #(
valid <= 0;
end else begin
if (enqueue) begin
entries[wr_a] <= write_data;
valid[wr_a] <= 1;
wr_ptr <= wr_ptr + 1;
end
@@ -53,6 +52,10 @@ module VX_index_queue #(
valid[read_addr] <= 0;
end
end
if (enqueue) begin
entries[wr_a] <= write_data;
end
end
assign write_addr = wr_a;

View File

@@ -77,6 +77,7 @@ module VX_scope #(
read_delta <= 0;
data_valid <= 0;
timestamp <= 0;
start_time <= 0;
end else begin
timestamp <= timestamp + 1;
@@ -177,6 +178,20 @@ module VX_scope #(
end
end
end
if (recording) begin
if (UPDW_ENABLE) begin
if (delta_flush
|| changed
|| (trigger_id != prev_trigger_id)) begin
delta_store[waddr] <= delta;
data_store[waddr] <= data_in;
end
end else begin
delta_store[waddr] <= 0;
data_store[waddr] <= data_in;
end
end
end
always @(*) begin

View File

@@ -58,32 +58,32 @@ module VX_serial_div #(
if (reset) begin
cntr <= 0;
is_busy <= 0;
end
else begin
end else begin
if (push) begin
for (integer i = 0; i < LANES; ++i) begin
working[i] <= {{WIDTHD{1'b0}}, numer_qual[i], 1'b0};
denom_r[i] <= denom_qual[i];
inv_quot[i] <= (denom[i] != 0) && signed_mode && (numer[i][31] ^ denom[i][31]);
inv_rem[i] <= signed_mode && numer[i][31];
end
tag_r <= tag_in;
cntr <= WIDTHN;
cntr <= WIDTHN;
is_busy <= 1;
end
else begin
if (!done) begin
for (integer i = 0; i < LANES; ++i) begin
working[i] <= sub_result[i][WIDTHD] ? {working[i][WIDTHN+MIN_ND-1:0], 1'b0} :
{sub_result[i][WIDTHD-1:0], working[i][WIDTHN-1:0], 1'b1};
end
cntr <= cntr - CNTRW'(1);
end
end else if (!done) begin
cntr <= cntr - CNTRW'(1);
end
if (pop) begin
is_busy <= 0;
end
end
if (push) begin
for (integer i = 0; i < LANES; ++i) begin
working[i] <= {{WIDTHD{1'b0}}, numer_qual[i], 1'b0};
denom_r[i] <= denom_qual[i];
inv_quot[i] <= (denom[i] != 0) && signed_mode && (numer[i][31] ^ denom[i][31]);
inv_rem[i] <= signed_mode && numer[i][31];
end
tag_r <= tag_in;
end else if (!done) begin
for (integer i = 0; i < LANES; ++i) begin
working[i] <= sub_result[i][WIDTHD] ? {working[i][WIDTHN+MIN_ND-1:0], 1'b0} :
{sub_result[i][WIDTHD-1:0], working[i][WIDTHN-1:0], 1'b1};
end
end
end
for (genvar i = 0; i < LANES; ++i) begin

View File

@@ -27,18 +27,22 @@ module VX_skid_buffer #(
if (ready_out) begin
use_buffer <= 0;
end
if (push) begin
buffer <= data_in;
if (valid_out_r && !ready_out) begin
assert(!use_buffer);
use_buffer <= 1;
end
if (push && valid_out_r && !ready_out) begin
assert(!use_buffer);
use_buffer <= 1;
end
if (!valid_out_r || ready_out) begin
valid_out_r <= valid_in || use_buffer;
data_out_r <= use_buffer ? buffer : data_in;
end
end
if (push) begin
buffer <= data_in;
end
if (!valid_out_r || ready_out) begin
data_out_r <= use_buffer ? buffer : data_in;
end
end
assign ready_in = !use_buffer;