@ int32_t dot_offset(const uint8_t* bytes, const int32_t* inputs, int n) @ @ The sum of bytes[i] x inputs[i], 16 at a time (n is a multiple of 16), the @ inner loop of every matrix product. ARM code in the fast internal RAM: @ - the bytes (weights, int8 + 128) come from cartridge ROM four words per @ LDM, a sequential burst, much cheaper than four separate loads @ - each byte is unpacked with one AND (a shifted operand), which is why they @ are stored offset to 0..255 rather than signed @ - the byte is MLA's second operand: the ARM7 multiplies faster when that @ operand is small @ Registers: r0 bytes, r1 inputs, r2 end of bytes, r3 sum, r4-r7 sixteen @ bytes, r8-r11 four inputs, r12 the mask 255, lr one unpacked byte. .section .iwram.text, "ax" .arm .align 2 .global dot_offset dot_offset: push {r4-r11, lr} mov r3, #0 mov r12, #255 add r2, r0, r2 1: ldmia r0!, {r4-r7} .irp word, r4, r5, r6, r7 ldmia r1!, {r8-r11} and lr, r12, \word mla r3, r8, lr, r3 and lr, r12, \word, lsr #8 mla r3, r9, lr, r3 and lr, r12, \word, lsr #16 mla r3, r10, lr, r3 mov lr, \word, lsr #24 mla r3, r11, lr, r3 .endr cmp r0, r2 blo 1b mov r0, r3 pop {r4-r11, lr} bx lr