projekte
/
bfloat16nn
26
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Matrice finished, but timeout?!

Browse Source
master
kaqu 1 year ago
parent 45fa4f4f12
commit e99f510d9e
2 changed files with 79 additions and 58 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 94
      libmodules/bfloat16nncore.py
  2. 43
      software/source/bfloat16nnlib.c

94
libmodules/bfloat16nncore.py
Unescape View File

@ -169,16 +169,15 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
self.submodules += Loader_fsm
self.Loader_Delay = Signal(32, reset_less=True)
self.Loader_Active = Signal()
self.Loader_Active = Signal()
Loader_fsm.act("Loader_IDLE",
If(self.LU_CacheValid & ~self.Loader_Active, # Enter if not active already
NextValue(self.Loader_Active, True), # Loader up & running
NextValue(self.Loader_Delay, 0), # Reset read delay timer
NextValue(LoadUnit.b9Offset1.storage, LUCacheSize - 1), # Adjust offset to read sentinel
NextValue(LoadUnit.b9Offset2.storage, LUCacheSize >> 1), # Adjust offset to start of 2nd array
NextValue(LoadUnit.b9Offset2.storage, LUCacheSize >> 1), # Adjust offset to start of 2nd array
NextValue(self.b16Result1.storage, 0), # Indicate # delays
NextValue(self.b16Result2.storage, 0), # Indicate # delays
NextValue(self.b16Status.storage[0], True), # Current status
NextValue(self.b16Result2.storage, 0), # Indicate # delays
NextValue(self.b16Value1_1.storage, 0), # Nothing loaded so far ...
NextValue(self.b16Value1_2.storage, 0),
NextValue(self.b16Value2_1.storage, 0),
@ -191,7 +190,7 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
)
)
Loader_fsm.act("Loader_LOAD1",
NextValue(self.b16Status.storage[1], True), # Current status added
NextValue(self.b16Status.storage[0], True), # Current status added
If(LoadUnit.b32Data1.storage == self.b32Sentinel.storage, # Valid last entry?
NextValue(LoadUnit.b9Offset1.storage, 0), # 1st value offset preparation
NextState("Loader_LOAD2")
@ -199,9 +198,11 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
NextState("Loader_IDLE") # Abort!
)
)
#-----> LOOP ENTRY ! (2nd loop onward: fs3 already prepared!)
Loader_fsm.act("Loader_LOAD2",
NextValue(self.b16Status.storage[2], True), # Current status added
If(self.Loader_Delay > RAMWaitTime,
NextValue(self.b16Status.storage[1], True), # Current status added
If(self.Loader_Delay > RAMWaitTime, # Required only for 1st entry ...
# FPU#1
NextValue(self.b16Value1_1.storage, LoadUnit.b32Data1.storage & 0xFFFF), # Pick 1st date
NextValue(self.b16Value1_2.storage, LoadUnit.b32Data1.storage >> 16), # Pick 2nd date
@ -215,56 +216,59 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
NextValue(fpu2.fs2, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data2.storage[16:32])),
NextValue(LoadUnit.b9Offset2.storage, LoadUnit.b9Offset2.storage + 1), # Move on to next entry
#NextValue(self.Loader_Delay, 0), # Reset delay
#NextState("Loader_LOAD3")
NextState("Loader_EXEC1")
).Else( # MEM wait cycles
NextValue(self.Loader_Delay, self.Loader_Delay + 1), # Increment
)
)
"""
Loader_fsm.act("Loader_LOAD3",
NextValue(self.b16Status.storage[3], True), # Current status added
If(self.Loader_Delay > RAMWaitTime,
NextValue(self.b16Value3.storage, LoadUnit.b32Data.storage & 0xFFFF), # Pick 3rd date
NextValue(self.b16Value4.storage, LoadUnit.b32Data.storage >> 16), # Pick 4th date
NextValue(fpu1.fs3, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data.storage[0:16])),
NextValue(fpu2.fs1, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data.storage[16:32])),
NextValue(LoadUnit.b9Offset.storage, 2), # 3rd value offset preparation
NextValue(self.Loader_Delay, 0), # Reset delay
NextState("Loader_LOAD4")
).Else( # MEM wait cycles
NextValue(self.Loader_Delay, self.Loader_Delay + 1), # Increment
)
)
Loader_fsm.act("Loader_LOAD4",
NextValue(self.b16Status.storage[4], True), # Current status added
If(self.Loader_Delay > RAMWaitTime,
NextValue(self.b16Value5.storage, LoadUnit.b32Data.storage & 0xFFFF), # Pick 5th date
NextValue(self.b16Value6.storage, LoadUnit.b32Data.storage >> 16), # Pick 6th date
NextValue(fpu2.fs2, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data.storage[0:16])),
NextValue(fpu2.fs3, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data.storage[16:32])),
NextState("Loader_EXEC1")
).Else( # MEM wait cycles
NextValue(self.Loader_Delay, self.Loader_Delay + 1), # Increment
)
)
"""
)
Loader_fsm.act("Loader_EXEC1",
NextValue(self.b16Status.storage[5], True), # Current status added
NextValue(fpu1.fadd, True), # 1st ADD requested
NextValue(fpu2.fadd, True),
NextValue(self.b16Status.storage[2], True), # Current status added
If(LoadUnit.b9Offset1.storage < self.b9ArrayWordLen.storage, #(LUCacheSize >> 1),
If(LoadUnit.b9Offset1.storage == 1, # As pointer already moved ahead 1!
NextValue(fpu1.fmul, True), # 1st ADD requested
NextValue(fpu2.fmul, True),
).Else(
NextValue(fpu1.fmadd, True), # 2nd ... last MUL/ADD requested
NextValue(fpu2.fmadd, True),
)
),
NextValue(fpu1.fready, False), # Engage trigger FPU#1
NextValue(fpu2.fready, False), # Engage trigger FPU#2
NextState("Loader_EXEC2")
)
Loader_fsm.act("Loader_EXEC2",
NextValue(self.b16Status.storage[6], True), # Current status added
NextValue(self.b16Status.storage[3], True), # Current status added
If(fpu1.fready & fpu2.fready,
NextValue(fpu1.fadd, False), # Clear command request FPU#1
NextValue(fpu2.fadd, False), # Clear command request FPU#2
If(LoadUnit.b9Offset1.storage == 1, # As pointer already moved ahead 1!
NextValue(fpu1.fmul, False), # Clear command request FPU#1
NextValue(fpu2.fmul, False), # Clear command request FPU#2
).Else(
NextValue(fpu1.fmadd, False), # Clear command request FPU#1
NextValue(fpu2.fmadd, False), # Clear command request FPU#2
),
NextValue(fpu1.fs3, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, fpu1.fresult[16:32])), # Sum will be used for fmadd.s
NextValue(fpu2.fs3, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, fpu2.fresult[16:32])), # Sum will be used for fmadd.s
If(LoadUnit.b9Offset1.storage < self.b9ArrayWordLen.storage, #(LUCacheSize >> 1), # Words 0 .. 255
NextState("Loader_LOAD2")
).Else( # Finally prepare ADD both result sums (on FPU#1 only!)
NextValue(fpu1.fs1, fpu1.fresult),
NextValue(fpu1.fs2, fpu2.fresult),
NextState("Loader_EXEC3")
)
)
)
Loader_fsm.act("Loader_EXEC3",
NextValue(self.b16Status.storage[4], True), # Current status added
NextValue(fpu1.fadd, True), # Final ADD requested
NextValue(fpu1.fready, False), # Engage trigger FPU#1 (only!)
NextState("Loader_EXEC4")
)
Loader_fsm.act("Loader_EXEC4",
NextValue(self.b16Status.storage[5], True), # Current status added
If(fpu1.fready,
NextValue(fpu1.fadd, False), # Clear command request FPU#1
NextValue(self.b16Result1.storage, fpu1.fresult[16:32]), # Pick result (little endian, high word!)
NextValue(self.b16Result2.storage, fpu2.fresult[16:32]), # Pick result (little endian, high word!)
NextValue(self.b16Result2.storage, fpu2.fresult[16:32]), # Useless (control only ...)
NextValue(self.b16Status.storage[15], True), # Indicate readyness ...
NextValue(self.bReady, True), # Indicate readyness (LED on!) (TODO: Remove!)
NextState("Loader_IDLE")

43
software/source/bfloat16nnlib.c
Unescape View File

@ -47,15 +47,18 @@ extern int key_eval(void);
#define DRAMDATABASE 0x40190000
#define DRAMDATASIZE 512
static int fpgaload(uint32_t *mempt, int16_t len)
static int fpgaload(uint32_t *mempt, int16_t len, int16_t calclen)
{
uint32_t *sentinel = (uint32_t *)(DRAMDATABASE + (DRAMDATASIZE - 1) * sizeof(int32_t));
if((len < 4) | (len > 512)) return -1; // Verify length of transfer was understood!
flush_l2_cache(); // Strictly nec. for longer transfers
bfloat16nn_b32DRAMLoadAddress_write((uint32_t)mempt); // Indicate memory to load from
bfloat16nn_b32Sentinel_write(*sentinel);
bfloat16nn_b9ArrayWordLen_write(calclen); // Indicate array length for calc.
bfloat16nn_bEnable_write(1); // Finally: Engage!
for(int i=0;i<10;i++) { // Max. 100ms delay
for(int i=0;i<100;i++) { // Max. 100ms delay
if(bfloat16nn_b16Status_read() & 0x8000) {
bfloat16nn_bEnable_write(0); // Disable transfer
return 1; // Ok, ready!
@ -132,6 +135,8 @@ int key_eval(void)
float fp1_1, fp1_2, fpResult1;
float fp2_1, fp2_2, fpResult2;
#define MAXCALCLEN 4
switch(kbhit()) {
case 'r': // Reload
printf("\e[35;1m*** Reload ***\e[0m\n");
@ -140,14 +145,24 @@ int key_eval(void)
// FPU#1
ui16ptr1 = (uint16_t *)(DRAMDATABASE + 0 * sizeof(uint32_t)); // Absolute: bytes!
for(i=1;i<=DRAMDATASIZE/2;i++)
*ui16ptr1++ = f2ui16(1.0 * (float)i );
for(i=1;i<=MAXCALCLEN;i++)
*ui16ptr1++ = f2ui16(1.0 * (float)i );
// FPU#2
ui16ptr2 = (uint16_t *)(DRAMDATABASE + (DRAMDATASIZE/2) * sizeof(uint32_t)); // Absolute: bytes!
for(i=1;i<=DRAMDATASIZE/2;i++)
for(i=1;i<=MAXCALCLEN;i++)
*ui16ptr2++ = f2ui16(1.0 * (float)i );
if(fpgaload((uint32_t *)DRAMDATABASE, 512)) { // 512 * 32-bit= 2048 bytes = 2kB
// TODO: Control procedure w/ regular code (matrice inner product)
float sum = 0.0;
for(i=1;i<=MAXCALCLEN;i+=2) {
sum += (1.0 * (float)i) * (1.0 * (float)(i+1));
}
sum *= 2.0; // Sim. 2 FPUs!
printf("Elements/FPU: %d", MAXCALCLEN);
printf1(" S/W SUM (2xFPU): %8.3f\n", sum);
if(fpgaload((uint32_t *)DRAMDATABASE, 512, MAXCALCLEN)) { // 512*32-bit=2048 bytes = 2kB, 256 Words/FPU to calc.
fp1_1 = fp1_1_read();
fp1_2 = fp1_2_read();
fp2_1 = fp2_1_read();
@ -157,7 +172,7 @@ int key_eval(void)
printf("S=%04Xh:\n", (uint32_t)bfloat16nn_b16Status_read());
printf1("V1_1=%6.3f ", fp1_1);
printf1("V1_2=%6.3f ", fp1_2);
printf1("RESULT1=%8.4f\n", fpResult1);
printf1("TOTAL SUM=%8.4f\n", fpResult1);
printf1("V2_1=%6.3f ", fp2_1);
printf1("V2_2=%6.3f ", fp2_2);
printf1("RESULT2=%8.4f\n", fpResult2);
@ -171,11 +186,13 @@ int key_eval(void)
printf("%d: %d\n", DRAMDATASIZE - 1, dram2fpga_b32Data2_read());
*/
}
else
printf("CURRENT: Timeout!");
else {
printf("CURRENT: Timeout! %04Xh ", (uint32_t)bfloat16nn_b16Status_read());
printf("Offset 1: %d ", (uint32_t)dram2fpga_b9Offset1_read());
printf("Offset 2: %d\n", (uint32_t)dram2fpga_b9Offset2_read());
}
*sentinel = 0; // Invalidate data!
if(fpgaload((uint32_t *)DRAMDATABASE, 512)) {
if(fpgaload((uint32_t *)DRAMDATABASE, 512, MAXCALCLEN)) {
fp1_1 = fp1_1_read();
fp1_2 = fp1_2_read();
fp2_1 = fp2_1_read();
@ -191,8 +208,8 @@ int key_eval(void)
printf1("V2_2=%6.3f ", fp2_2);
printf1("RESULT2=%8.4f\n", fpResult2);
}
else
printf("INVALIDATED: Timeout!");
else
printf("INVALIDATED: Timeout! %04Xh:\n", (uint32_t)bfloat16nn_b16Status_read());
break;
case 's':
fp1_1 = fp1_1_read();

Write Preview
Loading…
Cancel
Save