projekte
/
bfloat16nn
28
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

2 transfers, working (but wrong solution!)

Browse Source
master
kaqu 2 years ago
parent 8ae7085df7
commit c0189b15c7
4 changed files with 196 additions and 131 deletions
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Show Stats Download Patch File Download Diff File
  1. 4
      bfloat16nn.py
  2. 107
      libmodules/bfloat16nncore.py
  3. 46
      libmodules/dramtransfer.py
  4. 170
      software/source/bfloat16nnlib.c

4
bfloat16nn.py
Unescape View File

@ -183,7 +183,7 @@ class BaseSoC(SoCCore):
self.add_csr("systime")
# DRAM access section
MAXWORDS = 512 #512 # Transfer length 32 x 32-bit, FIFO depth (511 L1 cache currently possible = 9-bit!)
MAXWORDS = 512 # Max. FIFO!
# Load unit memory access
self.submodules.dma_reader = dma_reader = LiteDRAMDMAReader(self.sdram.crossbar.get_port(), fifo_depth=MAXWORDS, fifo_buffered=True)
dma_reader.add_csr()
@ -209,7 +209,7 @@ class BaseSoC(SoCCore):
self.add_csr("fpga2dram")
"""
# Integrate bfloat16NN processor
RAMWAITTIME=1 # Minimum wait!
RAMWAITTIME = 1 # Minimum wait!
self.submodules.bfloat16nn = bfloat16nn = bfloat16NeuralNetworkCore(
RAMWaitTime=RAMWAITTIME,
LUCacheSize=MAXWORDS,

107
libmodules/bfloat16nncore.py
Unescape View File

@ -42,11 +42,11 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
:b32DRAMAddress: New DRAM address from where to load into local memory
:b32Sentinel: Write control word to last address (same as [b32DRAMAddress+511] value)
:b32Sentinel: Write control word to last address (same as [b32DRAMAddress+LEN-1] value)
:bEnable: To enable running (after data preparation)
:b9ArrayWordLen: Number of words used for calculation of scalar (inner) product
:b10ArrayWordLen: Number of words used for calculation of scalar (inner) product
Outputs:
########
@ -78,8 +78,8 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
description="""
Enable free run
""")
self.b9ArrayWordLen = CSRStorage(9, reset_less=False,
fields=[CSRField("ArrayWordLen", size=9, description="*Field*: 9-Bit value")],
self.b10ArrayWordLen = CSRStorage(10, reset_less=False,
fields=[CSRField("ArrayWordLen", size=10, description="*Field*: 10-Bit value")],
description="""
Word length of array used for calculation
""")
@ -95,28 +95,6 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
description="""
FPU states: Low FPU#1, High FPU#2
""")
""" TODO: Remove!
self.b16Value1_1 = CSRStorage(16, reset_less=False,
fields=[CSRField("Value", size=16, description="*Field*: 16-Bit value")],
description="
FPU#1 Float register 1
")
self.b16Value1_2 = CSRStorage(16, reset_less=False,
fields=[CSRField("Value", size=16, description="*Field*: 16-Bit value")],
description="
#FPU#1 Float register 2
")
self.b16Value2_1 = CSRStorage(16, reset_less=False,
fields=[CSRField("Value", size=16, description="*Field*: 16-Bit value")],
description="
#FPU#2 Float register 1
")
self.b16Value2_2 = CSRStorage(16, reset_less=False,
fields=[CSRField("Value", size=16, description="*Field*: 16-Bit value")],
description=" " "
#FPU#2 Float register 2
" " ")
"""
self.b16Result1 = CSRStorage(16, reset_less=False,
fields=[CSRField("Result1", size=16, description="*Field*: 16-Bit value")],
description="""
@ -130,13 +108,13 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
self.bReady = Signal() # To be wired to data pin ... ;)
# Local vars.
# - none yet -
self.b10CurrentOffest = Signal(10, reset_less=True)
#---------------- Load unit (LU) -------------------------------------------------------------
LU_fsm = FSM(reset_state="LU_IDLE") # FSM starts idling ...
self.submodules += LU_fsm
self.LU_CacheOffset = Signal(9, reset_less=True) # 0..511 log2_int(LUCacheSize, False)) # Cache reading offset (0..(Size-1))=>Bits)
self.LU_CacheOffset = Signal(10, reset_less=True) # 0..1023 log2_int(LUCacheSize, False)) # Cache reading offset (0..(Size-1))=>Bits)
self.LU_CacheValid = Signal() # Indicate loaded LU cache
self.LU_CacheDelay = Signal(11, reset_less=True) # Evaluate load length in cycles (2048 max.)
LU_fsm.act("LU_IDLE", # If cache not valid fill it!
@ -181,14 +159,11 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
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(self.b10CurrentOffest, 0), # Actual offset (=DRAM local offset)
NextValue(LoadUnit.b10Offset1.storage, LUCacheSize - 1), # Adjust offset to read sentinel
NextValue(LoadUnit.b10Offset2.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.b16Value1_1.storage, 0), # TODO: Remove! Nothing loaded so far ...
#NextValue(self.b16Value1_2.storage, 0),
#NextValue(self.b16Value2_1.storage, 0),
#NextValue(self.b16Value2_2.storage, 0),
NextValue(self.bReady, False), # LED off!
NextState("Loader_LOAD1")
).Elif(~self.bEnable.storage, # Externally aborted?
@ -197,9 +172,9 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
)
)
Loader_fsm.act("Loader_LOAD1",
NextValue(self.b16Status.storage[0], 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
NextValue(LoadUnit.b10Offset1.storage, 0), # 1st value offset preparation
NextState("Loader_LOAD2")
).Elif(~self.bEnable.storage, # Enable withdrawn?
NextState("Loader_IDLE") # Abort!
@ -210,18 +185,16 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
Loader_fsm.act("Loader_LOAD2",
NextValue(self.b16Status.storage[1], True), # Current status added
If(self.Loader_Delay > RAMWaitTime, # Required only for 1st entry ...
NextValue(self.b10CurrentOffest, self.b10CurrentOffest + 1), # Increment (total) offset
# FPU#1
#NextValue(self.b16Value1_1.storage, LoadUnit.b32Data1.storage & 0xFFFF), # TODO: Remove! Pick 1st date
#NextValue(self.b16Value1_2.storage, LoadUnit.b32Data1.storage >> 16), # TODO: Remove! Pick 2nd date
NextValue(fpu1.fs1, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data1.storage[0:16])),
NextValue(fpu1.fs2, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data1.storage[16:32])),
NextValue(LoadUnit.b9Offset1.storage, LoadUnit.b9Offset1.storage + 1), # Move on to next entry
NextValue(LoadUnit.b10Offset1.storage, LoadUnit.b10Offset1.storage + 1), # Move on to next entry
# FPU#2
#NextValue(self.b16Value2_1.storage, LoadUnit.b32Data2.storage & 0xFFFF), # TODO: Remove! Pick 1st date
#NextValue(self.b16Value2_2.storage, LoadUnit.b32Data2.storage >> 16), # TODO: Remove! Pick 2nd date
NextValue(fpu2.fs1, Cat(0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0, LoadUnit.b32Data2.storage[0:16])),
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(LoadUnit.b10Offset2.storage, LoadUnit.b10Offset2.storage + 1), # Move on to next entry
NextState("Loader_EXEC1")
).Else( # MEM wait cycles
@ -230,7 +203,7 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
)
Loader_fsm.act("Loader_EXEC1",
NextValue(self.b16Status.storage[2], True), # Current status added
If(LoadUnit.b9Offset1.storage == 1, # As pointer already moved ahead 1!
If(LoadUnit.b10Offset1.storage == 1, # As pointer already moved ahead 1!
NextValue(fpu1.fmul, True), # 1st ADD requested
NextValue(fpu2.fmul, True),
).Else(
@ -246,7 +219,7 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
NextValue(self.b16Status.storage[8], fpu1.fready), # TODO: Remove!
NextValue(self.b16Status.storage[9], fpu2.fready), # TODO: Remove!
If(fpu1.fready & fpu2.fready,
If(LoadUnit.b9Offset1.storage == 1, # As pointer already moved ahead 1! (Actually: Entry #0)
If(LoadUnit.b10Offset1.storage == 1, # As pointer already moved ahead 1! (Actually: Entry #0)
NextValue(fpu1.fmul, False), # Clear command request FPU#1
NextValue(fpu2.fmul, False), # Clear command request FPU#2
).Else( # Entries 1 .. (maxlen-1)
@ -254,24 +227,58 @@ class bfloat16NeuralNetworkCore(Module, AutoCSR, AutoDoc, ModuleDoc):
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, # Words 0 .. 255
NextState("Loader_LOAD2")
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(self.b10CurrentOffest < self.b10ArrayWordLen.storage, # Words 0 .. LEN-1
If(LoadUnit.b10Offset1.storage < LUCacheSize, # Words 0 .. Cachelen
NextState("Loader_LOAD2") # Next value @offset
).Else( # Cache empty ...
NextValue(self.b32DRAMLoadAddress.storage, self.b32DRAMLoadAddress.storage + LUCacheSize), # Prepare DRAM address
NextState("Loader_XLOAD0") # Fill cache again
)
).Else( # Finally prepare ADD both result sums (on FPU#1 only!)
NextValue(fpu1.fs1, fpu1.fresult),
NextValue(fpu1.fs2, fpu2.fresult),
NextState("Loader_EXEC3")
NextState("Loader_EXEC3") # -> Final ADD logic & finishing cleanup
)
)
)
# Extended (2nd) cache load
Loader_fsm.act("Loader_XLOAD0",
NextValue(self.b16Status.storage[4], True), # Current status added
NextValue(self.LU_CacheValid, 0), # Engage refill (address safely adjusted by now ...)
NextState("Loader_XLOAD1")
)
Loader_fsm.act("Loader_XLOAD1", # Extended load ...
NextValue(self.b16Status.storage[5], True), # Current status added
If(self.LU_CacheValid, # Wait until filled ...
#NextValue(self.Loader_Delay, 0), # Reset read delay timer
NextValue(LoadUnit.b10Offset1.storage, LUCacheSize - 1), # Adjust offset to read sentinel
NextValue(LoadUnit.b10Offset2.storage, LUCacheSize >> 1), # Adjust offset to start of 2nd array
NextState("Loader_XLOAD2")
).Elif(~self.bEnable.storage, # Externally aborted?
NextState("Loader_IDLE") # Abort!
)
)
Loader_fsm.act("Loader_XLOAD2",
NextValue(self.b16Status.storage[6], True), # Current status added
If(LoadUnit.b32Data1.storage == (self.b32Sentinel.storage + 1), # Valid last entry? +1!!!
NextValue(LoadUnit.b10Offset1.storage, 0), # 1st value offset preparation
NextState("Loader_LOAD2") # Continue w/ loop
).Elif(~self.bEnable.storage, # Enable withdrawn?
NextState("Loader_IDLE") # Abort!
)
)
# Final ADD of results
Loader_fsm.act("Loader_EXEC3",
NextValue(self.b16Status.storage[4], True), # Current status added
NextValue(self.b16Status.storage[7], 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
NextValue(self.b16Status.storage[8], 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!)

46
libmodules/dramtransfer.py
Unescape View File

@ -36,7 +36,7 @@ class DRAM2FPGA(Module, AutoCSR, AutoDoc, ModuleDoc):
#. Once ``bValid`` becomes true (1), FPGA local memory is loaded, deactivate ``bEnable``
#. To retrieve, load ``b9Offset`` with offset (from base adress) to read from (0 .. 511),
#. To retrieve, load ``b10Offset`` with offset (from base adress) to read from (0 .. 1023),
``b32Data`` will contain the 32-bit value (from local FPGA memory @offset)
Inputs:
@ -46,18 +46,18 @@ class DRAM2FPGA(Module, AutoCSR, AutoDoc, ModuleDoc):
:bEnable: To enable running (after initialization)
:b9Offset1: Offset #1 (0..511) into local FPGA memory to read from
:b10Offset1: Offset #1 (0..1023) into local FPGA memory to read from
:b9Offset2: Offset #2 (0..511) into local FPGA memory to read from
:b10Offset2: Offset #2 (0..1023) into local FPGA memory to read from
Output:
#######
:bValid: Indicate validity of local FPGA memory, i.e. 'loaded'
:b32Data1: Local FPGA memory at b9Offset1
:b32Data1: Local FPGA memory at b10Offset1
:b32Data2: Local FPGA memory at b9Offset2
:b32Data2: Local FPGA memory at b10Offset2
"""
def __init__(self, maxwords=8, dma_reader=None, sync_fifo=None):
@ -76,13 +76,13 @@ class DRAM2FPGA(Module, AutoCSR, AutoDoc, ModuleDoc):
description="""
Enable/disabling DRAM access
""")
self.b9Offset1 = CSRStorage(9, reset_less=True,
fields=[CSRField("Offset1", size=9, description="*Field*: 9-Bit value (0..511)")],
self.b10Offset1 = CSRStorage(10, reset_less=True,
fields=[CSRField("Offset1", size=10, description="*Field*: 10-Bit value (0..1023)")],
description="""
Offset added to base address, port #1
""")
self.b9Offset2 = CSRStorage(9, reset_less=True,
fields=[CSRField("Offset2", size=9, description="*Field*: 9-Bit value (0..511)")],
self.b10Offset2 = CSRStorage(10, reset_less=True,
fields=[CSRField("Offset2", size=10, description="*Field*: 10-Bit value (0..1023)")],
description="""
Offset added to base address, port #2
""")
@ -190,32 +190,32 @@ class DRAM2FPGA(Module, AutoCSR, AutoDoc, ModuleDoc):
rdport2 = storage.get_port()
self.specials += rdport2
self.comb += [ # Read from (FPGA local) memory
self.b2Address1.eq(self.b9Offset1.storage[0:2]), # Filter bits 0..1 (range 0-3)
If(self.b9Offset1.storage < maxwords,
#rdport.adr.eq(self.b9Offset1.storage), # w/ translation!
self.b2Address1.eq(self.b10Offset1.storage[0:2]), # Filter bits 0..1 (range 0-3)
If(self.b10Offset1.storage < maxwords,
#rdport.adr.eq(self.b10Offset1.storage), # w/ translation!
If(self.b2Address1 == 0,
rdport1.adr.eq(self.b9Offset1.storage | 3) # 0->3
rdport1.adr.eq(self.b10Offset1.storage | 3) # 0->3
).Elif(self.b2Address1 == 1,
rdport1.adr.eq((self.b9Offset1.storage & 0x1FC) | 2) # 1->2
rdport1.adr.eq((self.b10Offset1.storage & 0x1FC) | 2) # 1->2
).Elif(self.b2Address1 == 2,
rdport1.adr.eq((self.b9Offset1.storage & 0x1FC) | 1) # 2->1
rdport1.adr.eq((self.b10Offset1.storage & 0x1FC) | 1) # 2->1
).Elif(self.b2Address1 == 3,
rdport1.adr.eq(self.b9Offset1.storage & 0x1FC) # 3->0
rdport1.adr.eq(self.b10Offset1.storage & 0x1FC) # 3->0
),
self.bData1.eq(rdport1.dat_r) # Assign to external var. ...
),
self.b2Address2.eq(self.b9Offset2.storage[0:2]), # Filter bits 0..1 (range 0-3)
If(self.b9Offset2.storage < maxwords,
#rdport.adr.eq(self.b9Offset2.storage), # w/ translation!
self.b2Address2.eq(self.b10Offset2.storage[0:2]), # Filter bits 0..1 (range 0-3)
If(self.b10Offset2.storage < maxwords,
#rdport.adr.eq(self.b10Offset2.storage), # w/ translation!
If(self.b2Address2 == 0,
rdport2.adr.eq(self.b9Offset2.storage | 3) # 0->3
rdport2.adr.eq(self.b10Offset2.storage | 3) # 0->3
).Elif(self.b2Address2 == 1,
rdport2.adr.eq((self.b9Offset2.storage & 0x1FC) | 2) # 1->2
rdport2.adr.eq((self.b10Offset2.storage & 0x1FC) | 2) # 1->2
).Elif(self.b2Address2 == 2,
rdport2.adr.eq((self.b9Offset2.storage & 0x1FC) | 1) # 2->1
rdport2.adr.eq((self.b10Offset2.storage & 0x1FC) | 1) # 2->1
).Elif(self.b2Address2 == 3,
rdport2.adr.eq(self.b9Offset2.storage & 0x1FC) # 3->0
rdport2.adr.eq(self.b10Offset2.storage & 0x1FC) # 3->0
),
self.bData2.eq(rdport2.dat_r) # Assign to external var. ...
),

170
software/source/bfloat16nnlib.c
Unescape View File

@ -45,33 +45,46 @@ extern char kbhit(void);
extern int key_eval(void);
#define DRAMDATABASE 0x40190000
#define DRAMDATASIZE 512
#define DRAMDATASIZE 1024 // 512 OK, 800 FAIL => 2 Load cycles (2*512)!
static uint32_t fpgastate, fpustates;
static int fpgaload(uint32_t *mempt, int16_t len, int16_t calclen)
{
uint32_t *sentinel = (uint32_t *)(DRAMDATABASE + (DRAMDATASIZE - 1) * sizeof(int32_t));
uint32_t *sentinel1 = (uint32_t *)(DRAMDATABASE + (DRAMDATASIZE/2 - 1) * sizeof(int32_t));
uint32_t *sentinel2 = (uint32_t *)(DRAMDATABASE + (DRAMDATASIZE - 1) * sizeof(int32_t));
if((len < 4) | (len > 512)) return -1; // Verify length of transfer was understood!
if((calclen < 2) | (calclen > len/2)) return -2; // Reasonable calc amount?
if((len < 4) | (len > DRAMDATASIZE)) {
printf("*** fpgaload: len out of range!");
return -1; // Verify length of transfer was understood!
}
if((calclen < 2) | (calclen > len/2)) {
printf("*** fpgaload: calclen out of range!");
return -2; // Reasonable calc amount?
}
bfloat16nn_bEnable_write(0); // Disable transfer (if still active for some reason ...)
*sentinel1 = 0x41434142; // Just some marker pattern ;)
*sentinel2 = 0x41434142 + 1; // Just some marker pattern ;)
bfloat16nn_b32Sentinel_write(*sentinel1);
flush_l2_cache(); // Strictly nec. for longer transfers
bfloat16nn_b9ArrayWordLen_write(calclen); // Indicate array length for calc.
bfloat16nn_b10ArrayWordLen_write(calclen); // Indicate array length for calc.
bfloat16nn_b32DRAMLoadAddress_write((uint32_t)mempt); // Indicate memory to load from
//*sentinel = 0xAAFF01A3;
bfloat16nn_b32Sentinel_write(*sentinel);
bfloat16nn_bEnable_write(1); // Finally: Engage!
for(int i=0;i<10;i++) { // Max. 100ms delay
for(int i=0;i<2000;i++) { // Max. 100ms delay
if(bfloat16nn_b16Status_read() & 0x8000) {
bfloat16nn_bEnable_write(0); // Disable transfer
fpgastate = 0;
fpustates = 0;
return 1; // Ok, ready!
}
else
busy_wait(10); // Just wait some time ...
busy_wait(1); // Just wait some time ...
}
fpgastate = (uint32_t)bfloat16nn_b16Status_read();
fpustates = (uint32_t)bfloat16nn_b16FPUStates_read();
bfloat16nn_bEnable_write(0); // Disable transfer
return 0; // Timeout
}
/*
static float fp1_1_read(void)
{
uint32_t v __attribute__((aligned(16))) = 0;
@ -100,6 +113,7 @@ static float fp2_2_read(void)
float *fpt = (float *)&v;
return *fpt;
}
*/
static float fpResult1_read(void)
{
uint32_t v __attribute__((aligned(16))) = 0;
@ -134,17 +148,44 @@ int key_eval(void)
uint32_t *ui32ptr;
uint16_t *ui16ptr1, *ui16ptr2;
int i;
float fp1_1, fp1_2, fpResult1;
float fp2_1, fp2_2, fpResult2;
//float fp1_1, fp1_2;
float fpResult1;
//float fp2_1, fp2_2;
float fpResult2;
uint32_t starttime;
uint32_t deltatime;
#define MAXCALCLEN 16
#define MAXCALCLEN (284) //784 //16 OK
switch(kbhit()) {
case 'r': // Reload
printf("\e[35;1m*** Reload ***\e[0m\n");
printf("Elements/FPU: %d\n", MAXCALCLEN);
for(i=0, ui32ptr = (uint32_t *)DRAMDATABASE;i<DRAMDATASIZE;i++) // Setup test data
*ui32ptr++ = 0; // Clear all memory ...
for(i=0, ui32ptr = (uint32_t *)DRAMDATABASE;i<DRAMDATASIZE;i++) // Setup test data
*ui32ptr++ = i+1; // Clear all memory ...
/*
// TODO: Control procedure w/ regular code (matrice inner product)
float *floatptr = (float *)DRAMDATABASE;
float *floatptr2 = (float *)(DRAMDATABASE + (MAXCALCLEN/2) * sizeof(float)); // Absolute: bytes!
for(i=1;i<=MAXCALCLEN/2;i++) {
*floatptr++ = (1.0 * (float)i);
*floatptr2++ = (1.0 * (float)i);
}
floatptr = (float *)DRAMDATABASE;
floatptr2 = (float *)(DRAMDATABASE + (MAXCALCLEN/2) * sizeof(float)); // Absolute: bytes!
starttime = systime(0);
float sum = 0.0;
for(i=1;i<=MAXCALCLEN/2;i++) {
sum += ((*floatptr++) * (*floatptr2++)); // 1*1+2*2+3*3+4*4 = 1+4+9+16 = 5+9+16 = 14+16=30
}
deltatime = systime(0)-starttime;
printf("S/W Delta t: %dms ", deltatime);
printf1("\t\t\tS/W SUM=%8.4f\n", sum);
// FPU#1
ui16ptr1 = (uint16_t *)(DRAMDATABASE + 0 * sizeof(uint32_t)); // Absolute: bytes!
for(i=1;i<=MAXCALCLEN;i++)
@ -154,77 +195,94 @@ int key_eval(void)
ui16ptr2 = (uint16_t *)(DRAMDATABASE + (DRAMDATASIZE/2) * sizeof(uint32_t)); // Absolute: bytes!
for(i=1;i<=MAXCALCLEN;i++)
*ui16ptr2++ = f2ui16(1.0 * (float)i );
// 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();
*/
starttime = systime(0);
if(fpgaload((uint32_t *)DRAMDATABASE, DRAMDATASIZE, MAXCALCLEN/2)) { // 800*32-bit=3200 bytes, 400 Words/FPU to calc.
deltatime = systime(0)-starttime;
printf("H/W Delta t: %dms ", deltatime);
/*fp1_1 = fp1_1_read();
fp1_2 = fp1_2_read();
fp2_1 = fp2_1_read();
fp2_2 = fp2_2_read();
fp2_2 = fp2_2_read();*/
fpResult1 = fpResult1_read();
fpResult2 = fpResult2_read();
printf("S=%04Xh: FS=%04Xh\n", (uint32_t)bfloat16nn_b16Status_read(), (uint32_t)bfloat16nn_b16FPUStates_read());
printf1("V1_1=%6.3f ", fp1_1);
printf1("V1_2=%6.3f ", fp1_2);
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);
printf("(S=%04Xh: FS=%04Xh)", fpgastate, fpustates);
/*printf1("V1_1=%6.3f ", fp1_1);
printf1("V1_2=%6.3f ", fp1_2);*/
printf1("\tS/W SUM=%8.4f\n", fpResult1);
/*printf1("V2_1=%6.3f ", fp2_1);
printf1("V2_2=%6.3f ", fp2_2);*/
//printf1("(RESULT2=%8.4f)", fpResult2);
/*
for(i=DRAMDATASIZE/2;i<DRAMDATASIZE/2+3;i++) {
dram2fpga_b9Offset2_write(i);
dram2fpga_b10Offset2_write(i);
printf("%d: %d\n", i, dram2fpga_b32Data2_read());
}
dram2fpga_b9Offset2_write(DRAMDATASIZE - 1);
dram2fpga_b10Offset2_write(DRAMDATASIZE - 1);
printf("%d: %d\n", DRAMDATASIZE - 1, dram2fpga_b32Data2_read());
*/
}
else {
printf("CURRENT: TIMEOUT! S=%04Xh: FS=%04Xh\n", (uint32_t)bfloat16nn_b16Status_read(), (uint32_t)bfloat16nn_b16FPUStates_read());
printf("Offset 1: %d ", (uint32_t)dram2fpga_b9Offset1_read());
printf("Offset 2: %d\n", (uint32_t)dram2fpga_b9Offset2_read());
printf("CURRENT TIMEOUT: S=%04Xh: FS=%04Xh ", fpgastate, fpustates);
printf("Offset 1: %d (%d) ", (uint32_t)dram2fpga_b10Offset1_read(), dram2fpga_b32Data1_read());
printf("Offset 2: %d (%d)", (uint32_t)dram2fpga_b10Offset2_read(), dram2fpga_b32Data2_read());
printf("Sentinels: %08Xh %08Xh\n", bfloat16nn_b32Sentinel_read(), dram2fpga_b32Data1_read());
for(i=0;i<10;i++) {
dram2fpga_b10Offset1_write(i);
dram2fpga_b10Offset2_write(i);
printf("%d: %d=%d\n", i, dram2fpga_b32Data1_read(), dram2fpga_b32Data2_read());
}
for(i=DRAMDATASIZE/2 - 5;i<(DRAMDATASIZE/2 + 5);i++) {
dram2fpga_b10Offset1_write(i);
dram2fpga_b10Offset2_write(i);
printf("%d: %d=%d\n", i, dram2fpga_b32Data1_read(), dram2fpga_b32Data2_read());
}
for(i=512 - 5;i<(512 + 5);i++) {
dram2fpga_b10Offset1_write(i);
dram2fpga_b10Offset2_write(i);
printf("%d: %d=%d\n", i, dram2fpga_b32Data1_read(), dram2fpga_b32Data2_read());
}
for(i=DRAMDATASIZE-10;i<DRAMDATASIZE-1;i++) {
dram2fpga_b10Offset1_write(i);
dram2fpga_b10Offset2_write(i);
printf("%d: %d=%d\n", i, dram2fpga_b32Data1_read(), dram2fpga_b32Data2_read());
}
dram2fpga_b10Offset1_write(DRAMDATASIZE - 1);
dram2fpga_b10Offset2_write(DRAMDATASIZE - 1);
printf("%d:*%d=%d*\n", DRAMDATASIZE - 1, dram2fpga_b32Data1_read(), dram2fpga_b32Data2_read());
}
*sentinel = 0; // Invalidate data!
if(fpgaload((uint32_t *)DRAMDATABASE, 512, MAXCALCLEN)) {
fp1_1 = fp1_1_read();
if(fpgaload((uint32_t *)DRAMDATABASE, DRAMDATASIZE, MAXCALCLEN/2)) {
/*fp1_1 = fp1_1_read();
fp1_2 = fp1_2_read();
fp2_1 = fp2_1_read();
fp2_2 = fp2_2_read();
fp2_2 = fp2_2_read();*/
fpResult1 = fpResult1_read();
fpResult2 = fpResult2_read();
printf("INVALIDATED: S=%04Xh: FS=%04Xh\n", (uint32_t)bfloat16nn_b16Status_read(), (uint32_t)bfloat16nn_b16FPUStates_read());
printf1("V1_1=%6.3f ", fp1_1);
printf1("V1_2=%6.3f ", fp1_2);
printf1("RESULT1=%8.4f\n", fpResult1);
printf1("V2_1=%6.3f ", fp2_1);
printf1("V2_2=%6.3f ", fp2_2);
printf1("RESULT2=%8.4f\n", fpResult2);
printf("INVALIDATED: S=%04Xh: FS=%04Xh\n", fpgastate, fpustates);
/*printf1("V1_1=%6.3f ", fp1_1);
printf1("V1_2=%6.3f ", fp1_2);*/
//printf1("RESULT1=%8.4f\n", fpResult1);
/*printf1("V2_1=%6.3f ", fp2_1);
printf1("V2_2=%6.3f ", fp2_2);*/
//printf1("RESULT2=%8.4f\n", fpResult2);
}
else
printf("INVALIDATED: S=%04Xh: FS=%04Xh\n", (uint32_t)bfloat16nn_b16Status_read(), (uint32_t)bfloat16nn_b16FPUStates_read());
printf("INVALIDATED TIMEOUT: S=%04Xh: FS=%04Xh\n", fpgastate, fpustates);
break;
case 's':
fp1_1 = fp1_1_read();
/*fp1_1 = fp1_1_read();
fp1_2 = fp1_2_read();
fp2_1 = fp2_1_read();
fp2_2 = fp2_2_read();
fp2_2 = fp2_2_read();*/
fpResult1 = fpResult1_read();
fpResult2 = fpResult2_read();
printf("REQUESTED: S=%04Xh: FS=%04Xh\n", (uint32_t)bfloat16nn_b16Status_read(), (uint32_t)bfloat16nn_b16FPUStates_read());
printf1("V1_1=%6.3f ", fp1_1);
printf1("V1_2=%6.3f ", fp1_2);
/*printf1("V1_1=%6.3f ", fp1_1);
printf1("V1_2=%6.3f ", fp1_2);*/
printf1("RESULT1=%8.4f\n", fpResult1);
printf1("V2_1=%6.3f ", fp2_1);
printf1("V2_2=%6.3f ", fp2_2);
/*printf1("V2_1=%6.3f ", fp2_1);
printf1("V2_2=%6.3f ", fp2_2);*/
printf1("RESULT2=%8.4f\n", fpResult2);
break;
case 'x': return 1; // Abort indication

Write Preview
Loading…
Cancel
Save