eden-emu/FFmpeg
20
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

mips/float_dsp: fix a bug in vector_fmul_window_mips

Browse source 
Loop was unrolled eight times although in heder there is assumption
that len is multiple of 4.
This is fixed, and assembly code is rewritten to be more optimal and
to simplify clobber list.

Signed-off-by: Nedeljko Babic <nedeljko.babic@imgtec.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This commit is contained in:
Nedeljko Babic 2015-03-05 12:23:59 +01:00 • committed by Michael Niedermayer
parent 45555a20c6
commit dfa9208074
1 changed files with 73 additions and 105 deletions
Download patch file Download diff file Expand all files Collapse all files

178
libavutil/mips/float_dsp_mips.c
View file

@ -144,114 +144,82 @@ static void vector_fmul_scalar_mips(float *dst, const float *src, float mul,
} }
static void vector_fmul_window_mips(float *dst, const float *src0, static void vector_fmul_window_mips(float *dst, const float *src0,
const float *src1, const float *win, int len) const float *src1, const float *win, int len)
{ {
int i, j; float * dst_j, *win_j, *src0_i, *src1_j, *dst_i, *win_i;
/* float temp, temp1, temp2, temp3;
* variables used in inline assembler float s0, s01, s1, s11;
*/ float wi, wi1, wi2, wi3;
float * dst_i, * dst_j, * dst_i2, * dst_j2; float wj, wj1, wj2, wj3;
float temp, temp1, temp2, temp3, temp4, temp5, temp6, temp7; const float * lp_end = win + len;
dst += len; win_i = (float *)win;
win += len; win_j = (float *)(win + 2 * len -1);
src0 += len; src1_j = (float *)(src1 + len - 1);
src0_i = (float *)src0;
dst_i = (float *)dst;
dst_j = (float *)(dst + 2 * len -1);
for (i = -len, j = len - 1; i < 0; i += 8, j -= 8) { /* loop unrolled 4 times */
__asm__ volatile (
dst_i = dst + i; "1:"
dst_j = dst + j; "lwc1 %[s1], 0(%[src1_j]) \n\t"
"lwc1 %[wi], 0(%[win_i]) \n\t"
dst_i2 = dst + i + 4; "lwc1 %[wj], 0(%[win_j]) \n\t"
dst_j2 = dst + j - 4; "lwc1 %[s11], -4(%[src1_j]) \n\t"
"lwc1 %[wi1], 4(%[win_i]) \n\t"
__asm__ volatile ( "lwc1 %[wj1], -4(%[win_j]) \n\t"
"mul.s %[temp], %[s1], %[wi] \n\t" "lwc1 %[s0], 0(%[src0_i]) \n\t"
"mul.s %[temp1], %[s1], %[wj] \n\t" "lwc1 %[s01], 4(%[src0_i]) \n\t"
"mul.s %[temp2], %[s11], %[wi1] \n\t" "mul.s %[temp], %[s1], %[wi] \n\t"
"mul.s %[temp3], %[s11], %[wj1] \n\t" "mul.s %[temp1], %[s1], %[wj] \n\t"
"mul.s %[temp2], %[s11], %[wi1] \n\t"
"msub.s %[temp], %[temp], %[s0], %[wj] \n\t" "mul.s %[temp3], %[s11], %[wj1] \n\t"
"madd.s %[temp1], %[temp1], %[s0], %[wi] \n\t" "lwc1 %[s1], -8(%[src1_j]) \n\t"
"msub.s %[temp2], %[temp2], %[s01], %[wj1] \n\t" "lwc1 %[wi2], 8(%[win_i]) \n\t"
"madd.s %[temp3], %[temp3], %[s01], %[wi1] \n\t" "lwc1 %[wj2], -8(%[win_j]) \n\t"
"lwc1 %[s11], -12(%[src1_j]) \n\t"
"swc1 %[temp], 0(%[dst_i]) \n\t" /* dst[i] = s0*wj - s1*wi; */ "msub.s %[temp], %[temp], %[s0], %[wj] \n\t"
"swc1 %[temp1], 0(%[dst_j]) \n\t" /* dst[j] = s0*wi + s1*wj; */ "madd.s %[temp1], %[temp1], %[s0], %[wi] \n\t"
"swc1 %[temp2], 4(%[dst_i]) \n\t" /* dst[i+1] = s01*wj1 - s11*wi1; */ "msub.s %[temp2], %[temp2], %[s01], %[wj1] \n\t"
"swc1 %[temp3], -4(%[dst_j]) \n\t" /* dst[j-1] = s01*wi1 + s11*wj1; */ "madd.s %[temp3], %[temp3], %[s01], %[wi1] \n\t"
"lwc1 %[wi3], 12(%[win_i]) \n\t"
"mul.s %[temp4], %[s12], %[wi2] \n\t" "lwc1 %[wj3], -12(%[win_j]) \n\t"
"mul.s %[temp5], %[s12], %[wj2] \n\t" "lwc1 %[s0], 8(%[src0_i]) \n\t"
"mul.s %[temp6], %[s13], %[wi3] \n\t" "lwc1 %[s01], 12(%[src0_i]) \n\t"
"mul.s %[temp7], %[s13], %[wj3] \n\t" "addiu %[src1_j],-16 \n\t"
"addiu %[win_i], 16 \n\t"
"msub.s %[temp4], %[temp4], %[s02], %[wj2] \n\t" "addiu %[win_j], -16 \n\t"
"madd.s %[temp5], %[temp5], %[s02], %[wi2] \n\t" "addiu %[src0_i], 16 \n\t"
"msub.s %[temp6], %[temp6], %[s03], %[wj3] \n\t" "swc1 %[temp], 0(%[dst_i]) \n\t" /* dst[i] = s0*wj - s1*wi; */
"madd.s %[temp7], %[temp7], %[s03], %[wi3] \n\t" "swc1 %[temp1], 0(%[dst_j]) \n\t" /* dst[j] = s0*wi + s1*wj; */
"swc1 %[temp2], 4(%[dst_i]) \n\t" /* dst[i+1] = s01*wj1 - s11*wi1; */
"swc1 %[temp4], 8(%[dst_i]) \n\t" /* dst[i+2] = s02*wj2 - s12*wi2; */ "swc1 %[temp3], -4(%[dst_j]) \n\t" /* dst[j-1] = s01*wi1 + s11*wj1; */
"swc1 %[temp5], -8(%[dst_j]) \n\t" /* dst[j-2] = s02*wi2 + s12*wj2; */ "mul.s %[temp], %[s1], %[wi2] \n\t"
"swc1 %[temp6], 12(%[dst_i]) \n\t" /* dst[i+2] = s03*wj3 - s13*wi3; */ "mul.s %[temp1], %[s1], %[wj2] \n\t"
"swc1 %[temp7], -12(%[dst_j]) \n\t" /* dst[j-3] = s03*wi3 + s13*wj3; */ "mul.s %[temp2], %[s11], %[wi3] \n\t"
: [temp]"=&f"(temp), [temp1]"=&f"(temp1), [temp2]"=&f"(temp2), "mul.s %[temp3], %[s11], %[wj3] \n\t"
[temp3]"=&f"(temp3), [temp4]"=&f"(temp4), [temp5]"=&f"(temp5), "msub.s %[temp], %[temp], %[s0], %[wj2] \n\t"
[temp6]"=&f"(temp6), [temp7]"=&f"(temp7) "madd.s %[temp1], %[temp1], %[s0], %[wi2] \n\t"
: [dst_j]"r"(dst_j), [dst_i]"r" (dst_i), "msub.s %[temp2], %[temp2], %[s01], %[wj3] \n\t"
[s0] "f"(src0[i]), [wj] "f"(win[j]), [s1] "f"(src1[j]), "madd.s %[temp3], %[temp3], %[s01], %[wi3] \n\t"
[wi] "f"(win[i]), [s01]"f"(src0[i + 1]),[wj1]"f"(win[j - 1]), "swc1 %[temp], 8(%[dst_i]) \n\t" /* dst[i+2] = s0*wj2 - s1*wi2; */
[s11]"f"(src1[j - 1]), [wi1]"f"(win[i + 1]), [s02]"f"(src0[i + 2]), "swc1 %[temp1], -8(%[dst_j]) \n\t" /* dst[j-2] = s0*wi2 + s1*wj2; */
[wj2]"f"(win[j - 2]), [s12]"f"(src1[j - 2]),[wi2]"f"(win[i + 2]), "swc1 %[temp2], 12(%[dst_i]) \n\t" /* dst[i+2] = s01*wj3 - s11*wi3; */
[s03]"f"(src0[i + 3]), [wj3]"f"(win[j - 3]), [s13]"f"(src1[j - 3]), "swc1 %[temp3], -12(%[dst_j]) \n\t" /* dst[j-3] = s01*wi3 + s11*wj3; */
[wi3]"f"(win[i + 3]) "addiu %[dst_i], 16 \n\t"
: "memory" "addiu %[dst_j], -16 \n\t"
); "bne %[win_i], %[lp_end], 1b \n\t"
: [temp]"=&f"(temp), [temp1]"=&f"(temp1), [temp2]"=&f"(temp2),
__asm__ volatile ( [temp3]"=&f"(temp3), [src0_i]"+r"(src0_i), [win_i]"+r"(win_i),
"mul.s %[temp], %[s1], %[wi] \n\t" [src1_j]"+r"(src1_j), [win_j]"+r"(win_j), [dst_i]"+r"(dst_i),
"mul.s %[temp1], %[s1], %[wj] \n\t" [dst_j]"+r"(dst_j), [s0] "=&f"(s0), [s01]"=&f"(s01), [s1] "=&f"(s1),
"mul.s %[temp2], %[s11], %[wi1] \n\t" [s11]"=&f"(s11), [wi] "=&f"(wi), [wj] "=&f"(wj), [wi2]"=&f"(wi2),
"mul.s %[temp3], %[s11], %[wj1] \n\t" [wj2]"=&f"(wj2), [wi3]"=&f"(wi3), [wj3]"=&f"(wj3), [wi1]"=&f"(wi1),
[wj1]"=&f"(wj1)
"msub.s %[temp], %[temp], %[s0], %[wj] \n\t" : [lp_end]"r"(lp_end)
"madd.s %[temp1], %[temp1], %[s0], %[wi] \n\t" : "memory"
"msub.s %[temp2], %[temp2], %[s01], %[wj1] \n\t" );
"madd.s %[temp3], %[temp3], %[s01], %[wi1] \n\t"
"swc1 %[temp], 0(%[dst_i2]) \n\t" /* dst[i] = s0*wj - s1*wi; */
"swc1 %[temp1], 0(%[dst_j2]) \n\t" /* dst[j] = s0*wi + s1*wj; */
"swc1 %[temp2], 4(%[dst_i2]) \n\t" /* dst[i+1] = s01*wj1 - s11*wi1; */
"swc1 %[temp3], -4(%[dst_j2]) \n\t" /* dst[j-1] = s01*wi1 + s11*wj1; */
"mul.s %[temp4], %[s12], %[wi2] \n\t"
"mul.s %[temp5], %[s12], %[wj2] \n\t"
"mul.s %[temp6], %[s13], %[wi3] \n\t"
"mul.s %[temp7], %[s13], %[wj3] \n\t"
"msub.s %[temp4], %[temp4], %[s02], %[wj2] \n\t"
"madd.s %[temp5], %[temp5], %[s02], %[wi2] \n\t"
"msub.s %[temp6], %[temp6], %[s03], %[wj3] \n\t"
"madd.s %[temp7], %[temp7], %[s03], %[wi3] \n\t"
"swc1 %[temp4], 8(%[dst_i2]) \n\t" /* dst[i+2] = s02*wj2 - s12*wi2; */
"swc1 %[temp5], -8(%[dst_j2]) \n\t" /* dst[j-2] = s02*wi2 + s12*wj2; */
"swc1 %[temp6], 12(%[dst_i2]) \n\t" /* dst[i+2] = s03*wj3 - s13*wi3; */
"swc1 %[temp7], -12(%[dst_j2]) \n\t" /* dst[j-3] = s03*wi3 + s13*wj3; */
: [temp]"=&f"(temp),
[temp1]"=&f"(temp1), [temp2]"=&f"(temp2), [temp3]"=&f"(temp3),
[temp4]"=&f"(temp4), [temp5]"=&f"(temp5), [temp6]"=&f"(temp6),
[temp7] "=&f" (temp7)
: [dst_j2]"r"(dst_j2), [dst_i2]"r"(dst_i2),
[s0] "f"(src0[i + 4]), [wj] "f"(win[j - 4]), [s1] "f"(src1[j - 4]),
[wi] "f"(win[i + 4]), [s01]"f"(src0[i + 5]),[wj1]"f"(win[j - 5]),
[s11]"f"(src1[j - 5]), [wi1]"f"(win[i + 5]), [s02]"f"(src0[i + 6]),
[wj2]"f"(win[j - 6]), [s12]"f"(src1[j - 6]),[wi2]"f"(win[i + 6]),
[s03]"f"(src0[i + 7]), [wj3]"f"(win[j - 7]), [s13]"f"(src1[j - 7]),
[wi3]"f"(win[i + 7])
: "memory"
);
}
} }
static void butterflies_float_mips(float *av_restrict v1, float *av_restrict v2, static void butterflies_float_mips(float *av_restrict v1, float *av_restrict v2,