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

OPAE hw snooping fixes

Browse Source
This commit is contained in:
Blaise Tine
2020-04-04 05:07:45 -07:00
parent 1f63139ce5
commit 07ec0ef344
9 changed files with 96 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
driver/hw/sources.txt
View File

@@ -68,6 +68,7 @@ vortex_afu.json
../../rtl/VX_cache/VX_cache_miss_resrv.v
../../rtl/VX_cache/VX_fill_invalidator.v
../../rtl/VX_cache/VX_tag_data_structure.v
../../rtl/VX_cache/VX_prefetcher.v
../../rtl/cache/VX_generic_pe.v
../../rtl/cache/cache_set.v
../../rtl/cache/VX_d_cache.v

2
driver/hw/vortex_afu.json
View File

@@ -14,7 +14,7 @@
"cmd-type-read": 1,
"cmd-type-write": 2,
"cmd-type-run": 3,
"cmd-type-snoop": 4,
"cmd-type-clflush": 4,
"afu-top-interface":
{

81
driver/hw/vortex_afu.sv
View File

@@ -34,7 +34,9 @@ module vortex_afu #(
);
localparam AVS_RD_QUEUE_SIZE = 16;
localparam VX_SNOOPING_DELAY = 300;
localparam VX_SNOOP_DELAY = 300;
localparam VX_SNOOP_LEVELS = 2;
localparam AFU_ID_L = 16'h0002; // AFU ID Lower
localparam AFU_ID_H = 16'h0004; // AFU ID Higher
@@ -42,7 +44,7 @@ localparam AFU_ID_H = 16'h0004; // AFU ID Higher
localparam CMD_TYPE_READ = `AFU_IMAGE_CMD_TYPE_READ;
localparam CMD_TYPE_WRITE = `AFU_IMAGE_CMD_TYPE_WRITE;
localparam CMD_TYPE_RUN = `AFU_IMAGE_CMD_TYPE_RUN;
localparam CMD_TYPE_SNOOP = `AFU_IMAGE_CMD_TYPE_SNOOP;
localparam CMD_TYPE_CLFLUSH = `AFU_IMAGE_CMD_TYPE_CLFLUSH;
localparam MMIO_CSR_CMD = `AFU_IMAGE_MMIO_CSR_CMD;
localparam MMIO_CSR_STATUS = `AFU_IMAGE_MMIO_CSR_STATUS;
@@ -52,13 +54,12 @@ localparam MMIO_CSR_DATA_SIZE = `AFU_IMAGE_MMIO_CSR_DATA_SIZE;
logic [127:0] afu_id = `AFU_ACCEL_UUID;
typedef enum logic[2:0] {
typedef enum logic[3:0] {
STATE_IDLE,
STATE_READ,
STATE_WRITE,
STATE_RUN,
STATE_SNOOP1,
STATE_SNOOP2
STATE_CLFLUSH
} state_t;
state_t state;
@@ -192,7 +193,8 @@ logic [31:0] cci_write_ctr;
logic [31:0] avs_read_ctr;
logic [31:0] avs_write_ctr;
logic [31:0] vx_snoop_ctr;
logic [31:0] vx_snoop_delay;
logic [9:0] vx_snoop_delay;
logic [1:0] vx_snoop_level;
logic vx_reset;
always_ff @(posedge clk)
@@ -210,21 +212,21 @@ begin
STATE_IDLE: begin
case (csr_cmd)
CMD_TYPE_READ: begin
$display("%t: CMD READ: ia=%h da=%h sz=%0d", $time, csr_io_addr, csr_mem_addr, csr_data_size);
$display("%t: STATE READ: ia=%h da=%h sz=%0d", $time, csr_io_addr, csr_mem_addr, csr_data_size);
state <= STATE_READ;
end
CMD_TYPE_WRITE: begin
$display("%t: CMD WRITE: ia=%h da=%h sz=%0d", $time, csr_io_addr, csr_mem_addr, csr_data_size);
$display("%t: STATE WRITE: ia=%h da=%h sz=%0d", $time, csr_io_addr, csr_mem_addr, csr_data_size);
state <= STATE_WRITE;
end
CMD_TYPE_RUN: begin
$display("%t: CMD START", $time);
$display("%t: STATE START", $time);
vx_reset <= 1;
state <= STATE_RUN;
end
CMD_TYPE_SNOOP: begin
$display("%t: CMD SNOOP: da=%h sz=%0d", $time, csr_mem_addr, csr_data_size);
state <= STATE_SNOOP1;
CMD_TYPE_CLFLUSH: begin
$display("%t: STATE CFLUSH: da=%h sz=%0d", $time, csr_mem_addr, csr_data_size);
state <= STATE_CLFLUSH;
end
endcase
end
@@ -251,16 +253,8 @@ begin
end
end
STATE_SNOOP1: begin
if (vx_snoop_delay >= VX_SNOOPING_DELAY)
begin
// TODO: Allow both RUN and SNOOP states to use the AVS bus
state <= STATE_SNOOP2;
end
end
STATE_SNOOP2: begin
if (vx_snoop_delay >= VX_SNOOPING_DELAY)
STATE_CLFLUSH: begin
if (vx_snoop_level >= VX_SNOOP_LEVELS)
begin
state <= STATE_IDLE;
end
@@ -322,7 +316,7 @@ begin
end
end
STATE_RUN: begin
STATE_RUN, STATE_CLFLUSH: begin
if (vx_dram_req_read
&& !vx_dram_req_delay)
begin
@@ -350,15 +344,20 @@ begin
end
// Vortex DRAM requests stalling
assign vx_dram_req_delay = !((STATE_RUN == state)
&& !avs_waitrequest
&& !avs_raq_full
&& !avs_rdq_full);
// Vortex DRAM fill response
logic vortex_enabled;
always_comb
begin
vx_dram_fill_rsp = (STATE_RUN == state) && !avs_rdq_empty && vx_dram_fill_accept;
vortex_enabled = (STATE_RUN == state) || (STATE_CLFLUSH == state);
vx_dram_req_delay = !vortex_enabled || avs_waitrequest || avs_raq_full || avs_rdq_full;
end
// Vortex DRAM fill response
always_comb
begin
vx_dram_fill_rsp = vortex_enabled && !avs_rdq_empty && vx_dram_fill_accept;
vx_dram_fill_rsp_addr = (avs_raq_dout << 6);
{>>{vx_dram_fill_rsp_data}} = avs_rdq_dout;
end
@@ -522,35 +521,39 @@ begin
vx_snp_req <= 0;
vx_snoop_ctr <= 0;
vx_snoop_delay <= 0;
vx_snoop_level <= 0;
end
else begin
if (STATE_IDLE == state)
begin
vx_snoop_ctr <= 0;
vx_snoop_ctr <= 0;
vx_snoop_delay <= 0;
vx_snoop_level <= 0;
end
vx_snp_req <= 0;
if ((STATE_SNOOP1 == state
|| STATE_SNOOP2 == state)
if ((STATE_CLFLUSH == state)
&& vx_snoop_ctr < csr_data_size
&& vx_snp_req_delay)
&& vx_snoop_level < VX_SNOOP_LEVELS
&& !vx_snp_req_delay)
begin
vx_snp_req <= 1;
vx_snoop_ctr <= vx_snoop_ctr + 1;
vx_snp_req_addr <= (csr_mem_addr + vx_snoop_ctr) << 6;
vx_snp_req <= 1;
vx_snoop_ctr <= vx_snoop_ctr + 1;
end
if ((vx_snoop_ctr >= csr_data_size)
&& (vx_snoop_delay < VX_SNOOPING_DELAY))
if ((vx_snoop_ctr == csr_data_size)
&& (vx_snoop_delay < VX_SNOOP_DELAY))
begin
vx_snoop_delay <= vx_snoop_delay + 1;
end
if (vx_snoop_delay >= VX_SNOOPING_DELAY)
if (vx_snoop_delay == VX_SNOOP_DELAY)
begin
vx_snoop_ctr <= 0;
vx_snoop_ctr <= 0;
vx_snoop_delay <= 0;
vx_snoop_level <= vx_snoop_level + 1;
end
end
end

5
driver/hw/wave.do
View File

@@ -27,6 +27,7 @@ add wave -noupdate -label avs_raq_full /ase_top/ase_top_generic/platform_shim_cc
add wave -noupdate -label avs_rdq_full /ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/avs_rdq_full
add wave -noupdate -label avs_raq_empty /ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/avs_raq_empty
add wave -noupdate -label avs_rdq_empty /ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/avs_rdq_empty
add wave -noupdate -label vortex_enabled /ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/vortex_enabled
add wave -noupdate -label vx_reset /ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/vx_soc/reset
add wave -noupdate -label vx_dram_req_read /ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/vx_dram_req_read
add wave -noupdate -label vx_dram_req_write /ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/vx_dram_req_write
@@ -49,7 +50,7 @@ add wave -noupdate -label warp_stalled {/ase_top/ase_top_generic/platform_shim_c
add wave -noupdate -label warp_lock {/ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/vx_soc/genblk1/Vortex_Cluster/genblk1[0]/vortex_core/vx_front_end/vx_fetch/warp_scheduler/warp_lock}
add wave -noupdate -label use_active {/ase_top/ase_top_generic/platform_shim_ccip_std_afu/ccip_std_afu/vortex_afu_inst/vx_soc/genblk1/Vortex_Cluster/genblk1[0]/vortex_core/vx_front_end/vx_fetch/warp_scheduler/use_active}
TreeUpdate [SetDefaultTree]
WaveRestoreCursors {{Cursor 2} {620643200 ps} 0}
WaveRestoreCursors {{Cursor 2} {293228800 ps} 0}
quietly wave cursor active 1
configure wave -namecolwidth 195
configure wave -valuecolwidth 100
@@ -65,4 +66,4 @@ configure wave -griddelta 40
configure wave -timeline 0
configure wave -timelineunits ps
update
WaveRestoreZoom {620460856 ps} {620825544 ps}
WaveRestoreZoom {293046456 ps} {293411144 ps}