Quantcast
Channel: MSP low-power microcontrollers
Viewing all articles
Browse latest Browse all 62309

Forum Post: MSP430F5529: Hardware Multiplier MPY32 for 8x8 signed matrix multiplication

$
0
0
Part Number: MSP430F5529 Hi, I want to use the combination of DMA and MPY32 of MSP430F5529 for matrix multiplication. I have achieved the task and the code gives correct outputs for smaller dimensions of input matrices like (1x3)*(3x2). But for my application, I want to do the dot product of (1x128) and (128x5) matrix. In this case, the outputs are not as expected and I get only one element out of the 5 which is erroneous. I'm using the MPY32 in MACS mode for 8x8 signed multiplication. Please find the code below. It would be of great help if someone can find the bug in the code. Thanks in advance! int8_t a[128] = {3,8,-1,7,9,12,3,10,13,11,4,10,2,12,10,2,-3,7,11,13,4,7,8,7,9,6,9,7,2,7,6,-2,9,0,13,3,5,11,-6,12,5,8,3,11,11,10,1,9,11,1,9,4,7,1,7,-2,7,11,14,11,3,9,5,8,9,9,-3,7,8,12,9,11,10,9,9,13,10,7,3,-5,11,12,11,7,7,4,2,9,2,6,12,9,7,9,9,7,8,4,9,6,7,6,7,11,9,9,3,12,0,11,1,9,8,10,8,2,9,9,1,5,3,13,5,8,-1,9,12,10}; int8_t beta[640]={-9,-4,-8,4,5,8,8,5,-9,-9,-9,-8,7,5,7,1,-4,2,1,4,8,8,8,-6,-8,8,-2,-2,8,3,-6,-9,-5,2,2,-3,-3,-1,7,3,7,7,-9,4,1,2,-1,-2,-3,-8,-8,-6,3,0,-9,2,3,0,-5,9,8,-6,-9,3,4,9,-8,9,-8,-9,2,-2,-8,-9,2,4,6,0,9,-2,-6,7,2,-9,-4,-6,-7,-3,1,-8,-1,1,5,-1,3,4,5,9,4,-1,-1,-8,-9,5,6,5,-3,7,9,-9,6,3,-6,-6,-2,6,-4,2,9,-7,5,9,0,-3,-2,-1,-5,6,8,6,-5,3,9,6,-4,8,-3,7,-5,1,4,-6,0,3,-3,-7,3,1,-5,-9,7,-6,4,8,-7,-1,-8,-9,2,8,6,4,-3,5,9,-9,-7,8,-9,7,2,7,-8,-9,-2,4,0,7,9,7,-2,8,5,9,6,0,-6,-3,-4,0,5,-6,8,9,6,-3,-9,1,-6,-2,9,1,6,6,2,-7,-3,-5,9,2,-5,-4,-4,6,-4,-2,-6,-8,8,1,-7,0,-2,7,7,-9,0,0,0,-8,8,1,-6,1,-1,-4,-6,-8,-6,2,-6,8,0,5,9,1,6,4,4,2,4,5,-9,8,-7,-7,-5,8,6,4,3,0,-4,-7,4,4,1,-5,7,6,-4,-5,-7,9,-9,-2,7,9,1,8,-1,8,0,-5,9,1,-4,7,-8,3,9,-2,-7,-1,9,6,-2,-8,-8,2,9,-7,-5,9,0,3,-4,-3,-8,-2,3,3,7,0,-4,-2,-1,9,-7,8,-3,0,1,7,0,-2,0,-8,4,-6,8,7,7,-5,-1,-3,-3,-2,5,9,-4,-4,6,9,5,4,-9,1,-4,0,-7,0,4,-3,0,2,-5,9,5,2,5,5,1,-6,0,-8,-2,7,-1,-9,-5,8,-6,4,4,-5,-1,4,7,5,0,0,3,6,6,8,-1,2,6,3,0,-5,5,8,-5,6,0,-4,2,2,5,-6,5,2,0,2,-9,-7,7,-8,2,-5,-6,7,-4,-8,-8,-2,-9,-2,5,-1,-2,4,1,3,3,8,2,0,3,-4,4,-6,2,4,-3,-7,3,5,-2,5,-8,3,-6,0,2,3,5,6,9,-8,4,-4,-9,2,-8,-3,1,-6,-9,-4,-3,-1,-6,-1,3,-7,-6,-7,-2,2,-5,-1,-4,9,-5,-3,-3,4,8,9,3,-1,4,-1,4,2,5,2,-8,-9,-5,7,4,3,-9,-8,-8,-4,0,0,6,1,4,-9,8,-5,4,-4,6,-7,-5,-3,0,-8,4,-5,-5,-9,3,2,-1,-1,0,-1,0,9,-2,-5,6,-9,4,0,-7,2,6,-8,6,8,-8,-3,1,3,-9,-5,1,-5,8,-1,5,-6,-4,8,0,-1,-2,-6,-9,-2,0,7,-2,3,9,5,-8,-5,5,-6,0,6,7,-9,-1,4,-1,2,7,1,-5,9,9,5,-9,5,-6,4,2,1,-6,-8,-3,-8,-2,5,-4,-4,-2,-4,8,-7,-3,7,0,-8,4,9,-2,9,2,7,-3,0,2,-1,-8,-7,-9,-1,-1,-6,-5,1,-7,-9,-9,-5,7,1,7,1,8,-6,1,-5,4,-4,-7,-2}; uint16_t hid_dim = 128; uint8_t outputDim = 5; int16_t g[5]; int main(void) { WDTCTL = WDTPW | WDTHOLD; // Stop watchdog timer MACS_8X8(a,beta,g,hid_dim,outputDim); //operand 1, operand 2, result,no.of.rows in operand 2, no.of.elements in result } void MACS_8X8(int8_t *pHidden, int8_t *pWeight, int16_t *pOutput, uint16_t hiddenDimenson, uint8_t outputDimenson) { initDMAsforMPY8X8(hiddenDimenson, pHidden, pWeight, pOutput); int i; for (i=0; i<outputDimenson-1; i++) { pOutput += 1; // results is 16-bit, uint16_t type ptr needs to increase by 1 pWeight += hiddenDimenson; __data16_write_addr((unsigned short) &DMA2DA,(unsigned long) pOutput); __data16_write_addr((unsigned short) &DMA0SA, (unsigned long) pWeight); // restart DMA RESHI=0; RESLO=0; MACS_L=0x00; DMA1CTL &= ~DMAIFG; DMA0CTL |= DMAEN; // enable DMA channel 0 DMA1CTL |= DMAEN; // enable DMA channel 1 DMA2CTL |= DMAEN; // enable DMA channel 2 OP2 = 0x00; // start MACS to get first DMA trigger } } void initDMAsforMPY8X8(uint16_t inputVectDimenson, int8_t *pInput, int8_t *pWeight, int16_t *pOutput) { // DMA channel 0: from pWeight (increment) to &MAC, single transfer per trigger, byte transfer // triggered by DMA_TRIGGERSOURCE_29 (MPY ready signal), totaly transfer inputVectDimenson times DMACTL0 = DMA0TSEL_29; // MPY ready signal triggered DMA0CTL = DMADT_0+DMASRCINCR_3+DMADSTBYTE+DMASRCBYTE; // Single tranfer (DMADT_0), inc src (DMASRCINCR_3), Interrupt disenabled // unchange dst (DMADSTINCR_0), byte to byte, rising edge trigger (default) DMA0SZ = inputVectDimenson; // transfer size __data16_write_addr((unsigned short) &DMA0SA, (unsigned long) pWeight); // Source address __data16_write_addr((unsigned short) &DMA0DA,(unsigned long) &MACS); // Destination address // DMA channel 1: from pInput (increment) to &OP2, single transfer per trigger // triggered by DMA_TRIGGERSOURCE_29 (MPY ready signal), totaly transfer inputVectDimenson times DMACTL0 |= DMA1TSEL_29; // MPY ready signal triggered for channel 1 DMA1CTL = DMADT_0+DMASRCINCR_3+DMADSTBYTE+DMASRCBYTE; // Single tranfer (DMADT_0), inc src (DMASRCINCR_3), Interrupt disenabled // unchange dst (DMADSTINCR_0), byte to byte, rising edge trigger (default) DMA1SZ = inputVectDimenson; // transfer size __data16_write_addr((unsigned short) &DMA1SA, (unsigned long) pInput); // Source address __data16_write_addr((unsigned short) &DMA1DA,(unsigned long) &OP2); // Destination address // DMA channel 2: from &RESLO (increment) to pOutput (increment), 2 transfers (DMA_TRANSFER_BLOCK, with size of 2) // triggered by DMA_TRIGGERSOURCE_30 (DMA1IFG signal) DMACTL1 |= DMA2TSEL_30; // MPY ready signal triggered for channel 2 DMA2CTL = DMADT_1+DMASRCINCR_3+DMADSTINCR_3; // Block tranfer (DMADT_1), inc src (DMASRCINCR_3), Interrupt disenabled // inc dst (DMADSTINCR_0), word to word, rising edge trigger (default) DMA2SZ = 1; // transfer size __data16_write_addr((unsigned short) &DMA2SA, (unsigned long) &RESLO); // Source address __data16_write_addr((unsigned short) &DMA2DA,(unsigned long) pOutput); // Destination address // initiate MPY by setting output and OP1 registers to be 0x0000 RESLO=0; RESHI=0; MACS_L=0x00; DMA2CTL |= DMAIE; DMA0CTL |= DMAEN; // enable DMA channel 0 DMA1CTL |= DMAEN; // enable DMA channel 1 DMA2CTL |= DMAEN; // enable DMA channel 2 OP2 = 0x00; // start MACS to get first DMA trigger }

Viewing all articles
Browse latest Browse all 62309

Trending Articles



<script src="https://jsc.adskeeper.com/r/s/rssing.com.1596347.js" async> </script>