simple_idct_arm.s

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5)
        mvn r11, r10             @ R11= NOT R10= 0x0000FFFF
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11)
        orr r8, r8, r9
        str r8, [r14, #0]

        add r8, r3, r5           @ R8=a2+b2
        add r9, r4, r7           @ R9=a3+b3
        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5)
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11)
        orr r8, r8, r9
        str r8, [r14, #4]

        sub r8, r4, r7           @ R8=a3-b3
        sub r9, r3, r5           @ R9=a2-b2
        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5)
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11)
        orr r8, r8, r9
        str r8, [r14, #8]

        sub r8, r2, r1           @ R8=a1-b1
        sub r9, r6, r0           @ R9=a0-b0
        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5)
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11)
        orr r8, r8, r9
        str r8, [r14, #12]

        bal __end_row_loop

__almost_empty_row:
        @@ the row was empty, except ROWr16[0], now, management of this special case
        @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
        @@                R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1],
        @@                R8=0xFFFF (temp), R9-R11 free
        mov r8, #0x10000         @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run).
        sub r8, r8, #1           @ R8 is now ready.
        and r5, r8, r6, lsl #3   @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF
        orr r5, r5, r5, lsl #16  @ R5=R5 | (R5<<16)
        str r5, [r14, #0]        @ R14[0]=ROWr32[0]=R5
        str r5, [r14, #4]        @ R14[4]=ROWr32[1]=R5
        str r5, [r14, #8]        @ R14[8]=ROWr32[2]=R5
        str r5, [r14, #12]       @ R14[12]=ROWr32[3]=R5

__end_row_loop:
        @@ at this point, R0-R11 (free)
        @@     R12=__const_ptr_, R14=&block[n]
        ldr r0, [sp, #0]         @ R0=block
        teq r0, r14              @ compare current &block[8*n] to block, when block is reached, the loop is finished.
        sub r14, r14, #16
        bne __row_loop



	@@ at this point, R0=block, R1-R11 (free)
	@@     R12=__const_ptr_, R14=&block[n]
	add r14, r0, #14        @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block.
__col_loop:

__b_evaluation2:
	@@ at this point, R0=block (temp),  R1-R11 (free)
	@@     R12=__const_ptr_, R14=&block[n]
	@@ proceed with b0-b3 first, followed by a0-a3
	@@ MUL16(b0, W1, col[8x1]);
	@@ MUL16(b1, W3, col[8x1]);
	@@ MUL16(b2, W5, col[8x1]);
	@@ MUL16(b3, W7, col[8x1]);
	@@ MAC16(b0, W3, col[8x3]);
	@@ MAC16(b1, -W7, col[8x3]);
	@@ MAC16(b2, -W1, col[8x3]);
	@@ MAC16(b3, -W5, col[8x3]);
	ldr r8, [r12, #offW1]    @ R8=W1
	ldrsh r7, [r14, #16]
	mul r0, r8, r7           @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
	ldr r9, [r12, #offW3]    @ R9=W3
	ldr r10, [r12, #offW5]   @ R10=W5
	mul r1, r9, r7           @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
	ldr r11, [r12, #offW7]   @ R11=W7
	mul r5, r10, r7          @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
	ldrsh r2, [r14, #48]
	mul r7, r11, r7          @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
	teq r2, #0               @ if 0, then avoid muls
	mlane r0, r9, r2, r0     @ R0+=W3*ROWr16[3]=b0 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
	rsbne r2, r2, #0         @ R2=-ROWr16[3]
	mlane r1, r11, r2, r1    @ R1-=W7*ROWr16[3]=b1 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
	mlane r5, r8, r2, r5     @ R5-=W1*ROWr16[3]=b2 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
	mlane r7, r10, r2, r7    @ R7-=W5*ROWr16[3]=b3 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)

	@@ at this point, R0=b0,  R1=b1, R2 (free), R3 (free), R4 (free),
	@@     R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
	@@     R12=__const_ptr_, R14=&block[n]
	@@ MAC16(b0, W5, col[5x8]);
	@@ MAC16(b2, W7, col[5x8]);
	@@ MAC16(b3, W3, col[5x8]);
	@@ MAC16(b1, -W1, col[5x8]);
	@@ MAC16(b0, W7, col[7x8]);
	@@ MAC16(b2, W3, col[7x8]);
	@@ MAC16(b3, -W1, col[7x8]);
	@@ MAC16(b1, -W5, col[7x8]);
	ldrsh r3, [r14, #80]     @ R3=COLr16[5x8]
	teq r3, #0               @ if 0 then avoid muls
	mlane r0, r10, r3, r0    @ R0+=W5*ROWr16[5x8]=b0
	mlane r5, r11, r3, r5    @ R5+=W7*ROWr16[5x8]=b2
	mlane r7, r9, r3, r7     @ R7+=W3*ROWr16[5x8]=b3
	rsbne r3, r3, #0         @ R3=-ROWr16[5x8]
	ldrsh r4, [r14, #112]    @ R4=COLr16[7x8]
	mlane r1, r8, r3, r1     @ R7-=W1*ROWr16[5x8]=b1
	@@ R3 is free now
	teq r4, #0               @ if 0 then avoid muls
	mlane r0, r11, r4, r0    @ R0+=W7*ROWr16[7x8]=b0
	mlane r5, r9, r4, r5     @ R5+=W3*ROWr16[7x8]=b2
	rsbne r4, r4, #0         @ R4=-ROWr16[7x8]
	mlane r7, r8, r4, r7     @ R7-=W1*ROWr16[7x8]=b3
	mlane r1, r10, r4, r1    @ R1-=W5*ROWr16[7x8]=b1
	@@ R4 is free now
__end_b_evaluation2:
	@@ at this point, R0=b0,  R1=b1, R2 (free), R3 (free), R4 (free),
	@@     R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
	@@     R12=__const_ptr_, R14=&block[n]

__a_evaluation2:
	@@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1));
	@@ a1 = a0 + W6 * row[2];
	@@ a2 = a0 - W6 * row[2];
	@@ a3 = a0 - W2 * row[2];
	@@ a0 = a0 + W2 * row[2];
	ldrsh r6, [r14, #0]
	ldr r9, [r12, #offW4]    @ R9=W4
	mul r6, r9, r6           @ R6=W4*ROWr16[0]
	ldr r10, [r12, #offW6]   @ R10=W6
	ldrsh r4, [r14, #32]      @ R4=ROWr16[2] (a3 not defined yet)
	add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0)
	mul r11, r10, r4         @ R11=W6*ROWr16[2]
	ldr r8, [r12, #offW2]    @ R8=W2
	add r2, r6, r11          @ R2=a0+W6*ROWr16[2] (a1)
	sub r3, r6, r11          @ R3=a0-W6*ROWr16[2] (a2)
	mul r11, r8, r4          @ R11=W2*ROWr16[2]
	sub r4, r6, r11          @ R4=a0-W2*ROWr16[2] (a3)
	add r6, r6, r11          @ R6=a0+W2*ROWr16[2] (a0)

	@@ at this point, R0=b0,  R1=b1, R2=a1, R3=a2, R4=a3,
	@@     R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
	@@     R12=__const_ptr_, R14=&block[n]
	@@ a0 += W4*row[4]
	@@ a1 -= W4*row[4]
	@@ a2 -= W4*row[4]
	@@ a3 += W4*row[4]
	ldrsh r11, [r14, #64]     @ R11=ROWr16[4]
	teq r11, #0              @ if null avoid muls
	mulne r11, r9, r11       @ R11=W4*ROWr16[4]
	@@ R9 is free now
	addne r6, r6, r11        @ R6+=W4*ROWr16[4] (a0)
	subne r2, r2, r11        @ R2-=W4*ROWr16[4] (a1)
	subne r3, r3, r11        @ R3-=W4*ROWr16[4] (a2)
	ldrsh r9, [r14, #96]     @ R9=ROWr16[6]
	addne r4, r4, r11        @ R4+=W4*ROWr16[4] (a3)
	@@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
	teq r9, #0               @ if null avoid muls
	mulne r11, r10, r9       @ R11=W6*ROWr16[6]
	addne r6, r6, r11        @ R6+=W6*ROWr16[6] (a0)
	mulne r10, r8, r9        @ R10=W2*ROWr16[6]
	@@ a0 += W6*row[6];
	@@ a3 -= W6*row[6];
	@@ a1 -= W2*row[6];
	@@ a2 += W2*row[6];
	subne r4, r4, r11        @ R4-=W6*ROWr16[6] (a3)
	subne r2, r2, r10        @ R2-=W2*ROWr16[6] (a1)
	addne r3, r3, r10        @ R3+=W2*ROWr16[6] (a2)
__end_a_evaluation2:
	@@ at this point, R0=b0,  R1=b1, R2=a1, R3=a2, R4=a3,
	@@     R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
	@@     R12=__const_ptr_, R14=&block[n]
	@@ col[0 ] = ((a0 + b0) >> COL_SHIFT);
	@@ col[8 ] = ((a1 + b1) >> COL_SHIFT);
	@@ col[16] = ((a2 + b2) >> COL_SHIFT);
	@@ col[24] = ((a3 + b3) >> COL_SHIFT);
	@@ col[32] = ((a3 - b3) >> COL_SHIFT);
	@@ col[40] = ((a2 - b2) >> COL_SHIFT);
	@@ col[48] = ((a1 - b1) >> COL_SHIFT);
	@@ col[56] = ((a0 - b0) >> COL_SHIFT);
	@@@@@ no optimisation here @@@@@
	add r8, r6, r0           @ R8=a0+b0
	add r9, r2, r1           @ R9=a1+b1
	mov r8, r8, asr #COL_SHIFT
	mov r9, r9, asr #COL_SHIFT
	strh r8, [r14, #0]
	strh r9, [r14, #16]
	add r8, r3, r5           @ R8=a2+b2
	add r9, r4, r7           @ R9=a3+b3
	mov r8, r8, asr #COL_SHIFT
	mov r9, r9, asr #COL_SHIFT
	strh r8, [r14, #32]
	strh r9, [r14, #48]
	sub r8, r4, r7           @ R8=a3-b3
	sub r9, r3, r5           @ R9=a2-b2
	mov r8, r8, asr #COL_SHIFT
	mov r9, r9, asr #COL_SHIFT
	strh r8, [r14, #64]
	strh r9, [r14, #80]
	sub r8, r2, r1           @ R8=a1-b1
	sub r9, r6, r0           @ R9=a0-b0
	mov r8, r8, asr #COL_SHIFT
	mov r9, r9, asr #COL_SHIFT
	strh r8, [r14, #96]
	strh r9, [r14, #112]

__end_col_loop:
	@@ at this point, R0-R11 (free)
	@@     R12=__const_ptr_, R14=&block[n]
	ldr r0, [sp, #0]         @ R0=block
	teq r0, r14              @ compare current &block[n] to block, when block is reached, the loop is finished.
	sub r14, r14, #2
	bne __col_loop




__end_simple_idct_ARM:
        @@ restore registers to previous status!
        add sp, sp, #8 @@ the local variables!
        ldmfd sp!, {r4-r11, r15} @@ update PC with LR content.



@@ kind of sub-function, here not to overload the common case.
__end_bef_a_evaluation:
	add r2, r6, r11          @ R2=a0+W6*ROWr16[2] (a1)
        mul r11, r8, r4          @ R11=W2*ROWr16[2]
        sub r4, r6, r11          @ R4=a0-W2*ROWr16[2] (a3)
        add r6, r6, r11          @ R6=a0+W2*ROWr16[2] (a0)
	bal __end_a_evaluation


__constant_ptr__:  @@ see #defines at the beginning of the source code for values.
	.align
        .word   W1
        .word   W2
        .word   W3
        .word   W4
        .word   W5
        .word   W6
        .word   W7
        .word   MASK_MSHW