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

Pass both A and D bundles to memfuzzer DPI

Browse Source
This commit is contained in:
Hansung Kim
2024-01-22 01:54:45 -08:00
parent e7340ba840
commit b2a83c788e
3 changed files with 70 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
src/main/resources/csrc/SimMemFuzzer.cc
View File

@@ -7,15 +7,23 @@
extern "C" void memfuzz_generate_rs(uint8_t *vec_a_ready, uint8_t *vec_a_valid, extern "C" void memfuzz_generate_rs(uint8_t *vec_a_ready, uint8_t *vec_a_valid,
long long *vec_a_address, long long *vec_a_address,
uint8_t *vec_d_ready, uint8_t *finished); uint8_t *vec_a_is_store, int *vec_a_size,
long long *vec_a_data, uint8_t *vec_d_ready,
uint8_t *vec_d_valid,
uint8_t *vec_d_is_store, int *vec_d_size,
uint8_t *finished);
extern "C" void memfuzz_init(int num_lanes) { extern "C" void memfuzz_init(int num_lanes) {
printf("from C: num_lanes=%d\n", num_lanes); printf("from C: num_lanes=%d\n", num_lanes);
} }
extern "C" void memfuzz_generate(uint8_t *vec_a_ready, uint8_t *vec_a_valid, extern "C" void memfuzz_generate(uint8_t *vec_a_ready, uint8_t *vec_a_valid,
long long *vec_a_address, uint8_t *vec_d_ready, long long *vec_a_address,
uint8_t *finished) { uint8_t *vec_a_is_store, int *vec_a_size,
memfuzz_generate_rs(vec_a_ready, vec_a_valid, vec_a_address, vec_d_ready, long long *vec_a_data, uint8_t *vec_d_ready,
finished); uint8_t *vec_d_valid, uint8_t *vec_d_is_store,
int *vec_d_size, uint8_t *finished) {
memfuzz_generate_rs(vec_a_ready, vec_a_valid, vec_a_address, vec_a_is_store,
vec_a_size, vec_a_data, vec_d_ready, vec_d_valid,
vec_d_is_store, vec_d_size, finished);
} }

66
src/main/resources/vsrc/SimMemFuzzer.v
View File

@@ -13,7 +13,15 @@ import "DPI-C" function void memfuzz_generate
input bit vec_a_ready[`MAX_NUM_LANES], input bit vec_a_ready[`MAX_NUM_LANES],
output bit vec_a_valid[`MAX_NUM_LANES], output bit vec_a_valid[`MAX_NUM_LANES],
output longint vec_a_address[`MAX_NUM_LANES], output longint vec_a_address[`MAX_NUM_LANES],
output bit vec_a_is_store[`MAX_NUM_LANES],
output int vec_a_size[`MAX_NUM_LANES],
output longint vec_a_data[`MAX_NUM_LANES],
output bit vec_d_ready[`MAX_NUM_LANES], output bit vec_d_ready[`MAX_NUM_LANES],
input bit vec_d_valid[`MAX_NUM_LANES],
input bit vec_d_is_store[`MAX_NUM_LANES],
input int vec_d_size[`MAX_NUM_LANES],
output bit finished output bit finished
); );
@@ -27,30 +35,47 @@ module SimMemFuzzer #(parameter NUM_LANES = 4) (
output [NUM_LANES-1:0] a_is_store, output [NUM_LANES-1:0] a_is_store,
output [`SIMMEM_LOGSIZE_WIDTH*NUM_LANES-1:0] a_size, output [`SIMMEM_LOGSIZE_WIDTH*NUM_LANES-1:0] a_size,
output [`SIMMEM_DATA_WIDTH*NUM_LANES-1:0] a_data, output [`SIMMEM_DATA_WIDTH*NUM_LANES-1:0] a_data,
output [NUM_LANES-1:0] d_ready, output [NUM_LANES-1:0] d_ready,
input [NUM_LANES-1:0] d_valid,
input [NUM_LANES-1:0] d_is_store,
input [`SIMMEM_LOGSIZE_WIDTH*NUM_LANES-1:0] d_size,
// TODO: d_mask
// TODO: d_data
output finished output finished
); );
// "in": verilog->C, "out": C->verilog
// need to be in ascending order to match with C indexing // need to be in ascending order to match with C indexing
// C array sizes are static, so need to use MAX_NUM_LANES // C array sizes are static, so need to use MAX_NUM_LANES
bit __out_a_ready [0:`MAX_NUM_LANES-1]; bit __out_a_ready [0:`MAX_NUM_LANES-1];
bit __in_a_valid [0:`MAX_NUM_LANES-1]; bit __in_a_valid [0:`MAX_NUM_LANES-1];
longint __in_a_address [0:`MAX_NUM_LANES-1]; longint __in_a_address [0:`MAX_NUM_LANES-1];
bit __in_a_is_store [0:`MAX_NUM_LANES-1]; bit __in_a_is_store [0:`MAX_NUM_LANES-1];
reg [`SIMMEM_LOGSIZE_WIDTH-1:0] __in_a_size [0:`MAX_NUM_LANES-1]; int __in_a_size [0:`MAX_NUM_LANES-1];
longint __in_a_data [0:`MAX_NUM_LANES-1]; longint __in_a_data [0:`MAX_NUM_LANES-1];
bit __in_d_ready [0:`MAX_NUM_LANES-1]; bit __in_d_ready [0:`MAX_NUM_LANES-1];
bit __in_finished; bit __out_d_valid [0:`MAX_NUM_LANES-1];
bit __out_d_is_store [0:`MAX_NUM_LANES-1];
int __out_d_size [0:`MAX_NUM_LANES-1];
bit __in_finished;
genvar g; genvar g;
generate generate
for (g = 0; g < NUM_LANES; g = g + 1) begin for (g = 0; g < NUM_LANES; g = g + 1) begin
assign __out_a_ready[g] = a_ready[g]; assign __out_a_ready[g] = a_ready[g];
assign a_valid[g] = __in_a_valid[g]; assign a_valid[g] = __in_a_valid[g];
assign a_address[`SIMMEM_DATA_WIDTH*g +: `SIMMEM_DATA_WIDTH] = __in_a_address[g]; assign a_address[`SIMMEM_DATA_WIDTH*g +: `SIMMEM_DATA_WIDTH]
= __in_a_address[g][`SIMMEM_DATA_WIDTH-1:0];
assign a_is_store[g] = __in_a_is_store[g]; assign a_is_store[g] = __in_a_is_store[g];
assign a_size[`SIMMEM_LOGSIZE_WIDTH*g +: `SIMMEM_LOGSIZE_WIDTH] = __in_a_size[g]; assign a_size[`SIMMEM_LOGSIZE_WIDTH*g +: `SIMMEM_LOGSIZE_WIDTH]
assign a_data[`SIMMEM_DATA_WIDTH*g +: `SIMMEM_DATA_WIDTH] = __in_a_data[g]; = __in_a_size[g][`SIMMEM_LOGSIZE_WIDTH-1:0];
assign a_data[`SIMMEM_DATA_WIDTH*g +: `SIMMEM_DATA_WIDTH]
= __in_a_data[g][`SIMMEM_DATA_WIDTH-1:0];
assign d_ready[g] = __in_d_ready[g];
assign __out_d_valid[g] = d_valid[g];
assign __out_d_is_store[g] = d_is_store[g];
assign __out_d_size[g] = d_size[`SIMMEM_LOGSIZE_WIDTH*g +: `SIMMEM_LOGSIZE_WIDTH];
end end
endgenerate endgenerate
assign finished = __in_finished; assign finished = __in_finished;
@@ -67,9 +92,8 @@ module SimMemFuzzer #(parameter NUM_LANES = 4) (
for (integer tid = 0; tid < NUM_LANES; tid = tid + 1) begin for (integer tid = 0; tid < NUM_LANES; tid = tid + 1) begin
__in_a_valid[tid] = 1'b0; __in_a_valid[tid] = 1'b0;
__in_a_address[tid] = `SIMMEM_DATA_WIDTH'b0; __in_a_address[tid] = `SIMMEM_DATA_WIDTH'b0;
__in_a_is_store[tid] = 1'b0; __in_a_is_store[tid] = 1'b0;
__in_a_size[tid] = `SIMMEM_LOGSIZE_WIDTH'b0; __in_a_size[tid] = 32'b0;
__in_a_data[tid] = `SIMMEM_DATA_WIDTH'b0; __in_a_data[tid] = `SIMMEM_DATA_WIDTH'b0;
__in_d_ready[tid] = 1'b0; __in_d_ready[tid] = 1'b0;
end end
@@ -79,15 +103,23 @@ module SimMemFuzzer #(parameter NUM_LANES = 4) (
__out_a_ready, __out_a_ready,
__in_a_valid, __in_a_valid,
__in_a_address, __in_a_address,
__in_a_is_store,
__in_a_size,
__in_a_data,
__in_d_ready, __in_d_ready,
__out_d_valid,
__out_d_is_store,
__out_d_size,
__in_finished __in_finished
); );
for (integer tid = 0; tid < NUM_LANES; tid = tid + 1) begin for (integer tid = 0; tid < NUM_LANES; tid = tid + 1) begin
$display("verilog: %04d valid[%d]=%d, address[%d]=%d", $display("verilog: %04d valid[%d]=%d, address[%d]=%d, __in_d_ready[%d]=%d",
$time, tid, __in_a_valid[tid], tid, __in_a_address[tid]); $time, tid, __in_a_valid[tid], tid, __in_a_address[tid], tid, __in_d_ready[tid]);
end end
if ($time >= 32'd200000) begin if ($time >= 64'd200000) begin
$finish; $finish;
end end
end end

16
src/main/scala/radiance/tile/VortexTile.scala
View File

@@ -178,12 +178,12 @@ class VortexTile private (
case Some(simtParam) => log2Ceil(simtParam.nSrcIds) case Some(simtParam) => log2Ceil(simtParam.nSrcIds)
case None => 4 case None => 4
} }
require( // require(
dmemSourceWidth >= 4, // dmemSourceWidth >= 4,
"Setting a small number of sourceIds may cause correctness bug inside " + // "Setting a small number of sourceIds may cause correctness bug inside " +
"Vortex core due to synchronization issues in vx_wspawn. " + // "Vortex core due to synchronization issues in vx_wspawn. " +
"We recommend setting nSrcIds to at least 16." // "We recommend setting nSrcIds to at least 16."
) // )
val smemSourceWidth = 4 // FIXME: hardcoded val smemSourceWidth = 4 // FIXME: hardcoded
@@ -295,9 +295,9 @@ class VortexTile private (
// Conditionally instantiate memory coalescer // Conditionally instantiate memory coalescer
val coalescerNode = p(CoalescerKey) match { val coalescerNode = p(CoalescerKey) match {
case Some(coalescerParam) => { case Some(coalParam) => {
val coal = LazyModule( val coal = LazyModule(
new CoalescingUnit(coalescerParam) new CoalescingUnit(coalParam)
) )
coal.cpuNode :=* dmemAggregateNode coal.cpuNode :=* dmemAggregateNode
coal.aggregateNode // N+1 lanes coal.aggregateNode // N+1 lanes