question about arm cortex-a9 neon optimization(4x4 matrix mul)

=======================================

for matrix 4 by 4 multiplication, neon programming is slower than natural code with

auto-vectorization option. (Xilinx Zynq 702 EVM board - cortex a9 with gcc complier option

-mfloat-abi=softfp -mfpu=neon-fp16 -ftree-vectorize)

=========================================

could you tell me what's wrong ?

=========================================

here are 3 versions.

1 - natural code

2 - neon code (intrinsic)

3 - neon inline asm code

=========================================

operation results of three ways are all same (perfect matched).

=========================================

I have a problem in speed results

  => natural is faster than neon intrinsic code.

  => intrinsic is faster than inline asm code.

=========================================

natural code needs 119 lines in asm file.

neon code needs 46 lines in asm file.

inline asm code needs 23 lines in asm file.

=========================================

for two 1280x720 source arrays, it's results(time tick) are as follow (in zynq).

natural - 7160144 tick (100.0%)

intrinsic - 7847183 tick (9.8% ↑)

inline asm - 8720648 tick (21.8%↑)

=========================================

here are c codes for each version.

SRC_T and RES_T mean float.

FP32X4 and FP32X2 mean float32x4_t and float32x2_t each.

========================================================

1 - natural code

static int nat_matrix_mul( const SRC_T * RESTRICT src0_4x4, const SRC_T * RESTRICT src1_4x4,

RES_T * RESTRICT res4x4)

{

res4x4[ 0] = (src0_4x4[ 0] * src1_4x4[ 0]) + (src0_4x4[ 1] * src1_4x4[ 4]) + (src0_4x4[ 2] *

src1_4x4[ 8]) + (src0_4x4[ 3] * src1_4x4[12]);

res4x4[ 1] = (src0_4x4[ 0] * src1_4x4[ 1]) + (src0_4x4[ 1] * src1_4x4[ 5]) + (src0_4x4[ 2] *

src1_4x4[ 9]) + (src0_4x4[ 3] * src1_4x4[13]);

res4x4[ 2] = (src0_4x4[ 0] * src1_4x4[ 2]) + (src0_4x4[ 1] * src1_4x4[ 6]) + (src0_4x4[ 2] *

src1_4x4[10]) + (src0_4x4[ 3] * src1_4x4[14]);

res4x4[ 3] = (src0_4x4[ 0] * src1_4x4[ 3]) + (src0_4x4[ 1] * src1_4x4[ 7]) + (src0_4x4[ 2] *

src1_4x4[11]) + (src0_4x4[ 3] * src1_4x4[15]);

res4x4[ 4] = (src0_4x4[ 4] * src1_4x4[ 0]) + (src0_4x4[ 5] * src1_4x4[ 4]) + (src0_4x4[ 6] *

src1_4x4[ 8]) + (src0_4x4[ 7] * src1_4x4[12]);

res4x4[ 5] = (src0_4x4[ 4] * src1_4x4[ 1]) + (src0_4x4[ 5] * src1_4x4[ 5]) + (src0_4x4[ 6] *

src1_4x4[ 9]) + (src0_4x4[ 7] * src1_4x4[13]);

res4x4[ 6] = (src0_4x4[ 4] * src1_4x4[ 2]) + (src0_4x4[ 5] * src1_4x4[ 6]) + (src0_4x4[ 6] *

src1_4x4[10]) + (src0_4x4[ 7] * src1_4x4[14]);

res4x4[ 7] = (src0_4x4[ 4] * src1_4x4[ 3]) + (src0_4x4[ 5] * src1_4x4[ 7]) + (src0_4x4[ 6] *

src1_4x4[11]) + (src0_4x4[ 7] * src1_4x4[15]);

res4x4[ 8] = (src0_4x4[ 8] * src1_4x4[ 0]) + (src0_4x4[ 9] * src1_4x4[ 4]) + (src0_4x4[10] *

src1_4x4[ 8]) + (src0_4x4[11] * src1_4x4[12]);

res4x4[ 9] = (src0_4x4[ 8] * src1_4x4[ 1]) + (src0_4x4[ 9] * src1_4x4[ 5]) + (src0_4x4[10] *

src1_4x4[ 9]) + (src0_4x4[11] * src1_4x4[13]);

res4x4[10] = (src0_4x4[ 8] * src1_4x4[ 2]) + (src0_4x4[ 9] * src1_4x4[ 6]) + (src0_4x4[10] *

src1_4x4[10]) + (src0_4x4[11] * src1_4x4[14]);

res4x4[11] = (src0_4x4[ 8] * src1_4x4[ 3]) + (src0_4x4[ 9] * src1_4x4[ 7]) + (src0_4x4[10] *

src1_4x4[11]) + (src0_4x4[11] * src1_4x4[15]);

res4x4[12] = (src0_4x4[12] * src1_4x4[ 0]) + (src0_4x4[13] * src1_4x4[ 4]) + (src0_4x4[14] *

src1_4x4[ 8]) + (src0_4x4[15] * src1_4x4[12]);

res4x4[13] = (src0_4x4[12] * src1_4x4[ 1]) + (src0_4x4[13] * src1_4x4[ 5]) + (src0_4x4[14] *

src1_4x4[ 9]) + (src0_4x4[15] * src1_4x4[13]);

res4x4[14] = (src0_4x4[12] * src1_4x4[ 2]) + (src0_4x4[13] * src1_4x4[ 6]) + (src0_4x4[14] *

src1_4x4[10]) + (src0_4x4[15] * src1_4x4[14]);

res4x4[15] = (src0_4x4[12] * src1_4x4[ 3]) + (src0_4x4[13] * src1_4x4[ 7]) + (src0_4x4[14] *

src1_4x4[11]) + (src0_4x4[15] * src1_4x4[15]);

return 0;

}

========================================================

2 - neon code (intrinsic)

int opt_matrix_mul( const SRC_T * RESTRICT src0_4x4, const SRC_T * RESTRICT src1_4x4,

RES_T * RESTRICT res4x4)

{

FP32X4 F32X4_IN_0_3210, F32X4_IN_0_7654, F32X4_IN_0_BA98, F32X4_IN_0_FEDC;

FP32X2 F32X2_IN_0_10, F32X2_IN_0_32, F32X2_IN_0_54, F32X2_IN_0_76, F32X2_IN_0_98,

F32X2_IN_0_BA, F32X2_IN_0_DC, F32X2_IN_0_FE;

FP32X4 F32X4_IN_1_3210, F32X4_IN_1_7654, F32X4_IN_1_BA98, F32X4_IN_1_FEDC;

FP32X4 F32X4_OUT_0_3210, F32X4_OUT_0_7654, F32X4_OUT_0_BA98, F32X4_OUT_0_FEDC;

F32X4_IN_1_3210 = vld1q_f32(&src1_4x4[ 0]);

F32X4_IN_1_7654 = vld1q_f32(&src1_4x4[ 4]);

F32X4_IN_1_BA98 = vld1q_f32(&src1_4x4[ 8]);

F32X4_IN_1_FEDC = vld1q_f32(&src1_4x4[12]);

F32X4_IN_0_3210 = vld1q_f32(&src0_4x4[ 0]);

F32X4_IN_0_7654 = vld1q_f32(&src0_4x4[ 4]);

F32X4_IN_0_BA98 = vld1q_f32(&src0_4x4[ 8]);

F32X4_IN_0_FEDC = vld1q_f32(&src0_4x4[12]);

F32X2_IN_0_10 = vget_low_f32 (F32X4_IN_0_3210);

F32X2_IN_0_32 = vget_high_f32(F32X4_IN_0_3210);

F32X2_IN_0_54 = vget_low_f32 (F32X4_IN_0_7654);

F32X2_IN_0_76 = vget_high_f32(F32X4_IN_0_7654);

F32X2_IN_0_98 = vget_low_f32 (F32X4_IN_0_BA98);

F32X2_IN_0_BA = vget_high_f32(F32X4_IN_0_BA98);

F32X2_IN_0_DC = vget_low_f32 (F32X4_IN_0_FEDC);

F32X2_IN_0_FE = vget_high_f32(F32X4_IN_0_FEDC);

F32X4_OUT_0_3210 = vmulq_lane_f32( F32X4_IN_1_3210, F32X2_IN_0_10, 0);

F32X4_OUT_0_3210 = vmlaq_lane_f32(F32X4_OUT_0_3210, F32X4_IN_1_7654, F32X2_IN_0_10, 1);

F32X4_OUT_0_3210 = vmlaq_lane_f32(F32X4_OUT_0_3210, F32X4_IN_1_BA98, F32X2_IN_0_32, 0);

F32X4_OUT_0_3210 = vmlaq_lane_f32(F32X4_OUT_0_3210, F32X4_IN_1_FEDC, F32X2_IN_0_32, 1);

F32X4_OUT_0_7654 = vmulq_lane_f32( F32X4_IN_1_3210, F32X2_IN_0_54, 0);

F32X4_OUT_0_7654 = vmlaq_lane_f32(F32X4_OUT_0_7654, F32X4_IN_1_7654, F32X2_IN_0_54, 1);

F32X4_OUT_0_7654 = vmlaq_lane_f32(F32X4_OUT_0_7654, F32X4_IN_1_BA98, F32X2_IN_0_76, 0);

F32X4_OUT_0_7654 = vmlaq_lane_f32(F32X4_OUT_0_7654, F32X4_IN_1_FEDC, F32X2_IN_0_76, 1);

F32X4_OUT_0_BA98 = vmulq_lane_f32( F32X4_IN_1_3210, F32X2_IN_0_98, 0);

F32X4_OUT_0_BA98 = vmlaq_lane_f32(F32X4_OUT_0_BA98, F32X4_IN_1_7654, F32X2_IN_0_98, 1);

F32X4_OUT_0_BA98 = vmlaq_lane_f32(F32X4_OUT_0_BA98, F32X4_IN_1_BA98, F32X2_IN_0_BA, 0);

F32X4_OUT_0_BA98 = vmlaq_lane_f32(F32X4_OUT_0_BA98, F32X4_IN_1_FEDC, F32X2_IN_0_BA, 1);

F32X4_OUT_0_FEDC = vmulq_lane_f32( F32X4_IN_1_3210, F32X2_IN_0_DC, 0);

F32X4_OUT_0_FEDC = vmlaq_lane_f32(F32X4_OUT_0_FEDC, F32X4_IN_1_7654, F32X2_IN_0_DC, 1);

F32X4_OUT_0_FEDC = vmlaq_lane_f32(F32X4_OUT_0_FEDC, F32X4_IN_1_BA98, F32X2_IN_0_FE, 0);

F32X4_OUT_0_FEDC = vmlaq_lane_f32(F32X4_OUT_0_FEDC, F32X4_IN_1_FEDC, F32X2_IN_0_FE, 1);

vst1q_f32(&res4x4[ 0], F32X4_OUT_0_3210);

vst1q_f32(&res4x4[ 4], F32X4_OUT_0_7654);

vst1q_f32(&res4x4[ 8], F32X4_OUT_0_BA98);

vst1q_f32(&res4x4[12], F32X4_OUT_0_FEDC);

return 0;

}

========================================================

3 - neon inline asm code

int opt_matrix_mul( const SRC_T * RESTRICT src0_4x4, const SRC_T * RESTRICT src1_4x4,

RES_T * RESTRICT res4x4)

{

/* load src0_4x4 */

asm ("vld1.32 { q8-q9 }, [r1]! @,");

asm ("vld1.32 { q10-q11 }, [r1]! @,");

/* load src1_4x4 */

asm ("vld1.32 { q0-q1 }, [r0]! @,");

asm ("vld1.32 { q2-q3 }, [r0]! @,");

/* mul res0-3 */

asm ("vmul.f32 q12, q8, d0[0] @,");

asm ("vmul.f32 q13, q8, d2[0] @,");

asm ("vmul.f32 q14, q8, d4[0] @,");

asm ("vmul.f32 q15, q8, d6[0] @,");

asm ("vmla.f32 q12, q9, d0[1] @,");

asm ("vmla.f32 q13, q9, d2[1] @,");

asm ("vmla.f32 q14, q9, d4[1] @,");

asm ("vmla.f32 q15, q9, d6[1] @,");

asm ("vmla.f32 q12, q10, d1[0] @,");

asm ("vmla.f32 q13, q10, d3[0] @,");

asm ("vmla.f32 q14, q10, d5[0] @,");

asm ("vmla.f32 q15, q10, d7[0] @,");

asm ("vmla.f32 q12, q11, d1[1] @,");

asm ("vmla.f32 q13, q11, d3[1] @,");

asm ("vmla.f32 q14, q11, d5[1] @,");

asm ("vmla.f32 q15, q11, d7[1] @,");

/* save res4x4 */

asm ("vst1.32 { q12-q13 }, [r2]! @,");

asm ("vst1.32 { q14-q15 }, [r2]! @,");

/* return 0 */

asm ("mov r0, #0 @,");

}

=========================================

here are asm codes for each version

========================================================

1 - natural code

add r3, r1, ip @ D.14387, p_src_c1, ivtmp.11

add r2, r0, ip @ D.14387, p_src_c0, ivtmp.11

flds s28, [r3, #16] @ D.14385, MEM[base: _575, offset: 16B]

flds s10, [r2, #4] @ D.14385, MEM[base: _576, offset: 4B]

fmuls s7, s10, s28 @ D.14385, D.14385, D.14385

flds s23, [r3] @ D.14385, MEM[base: _575, offset: 0B]

flds s31, [r2] @ D.14385, MEM[base: _576, offset: 0B]

flds s27, [r3, #20] @ D.14385, MEM[base: _575, offset: 20B]

flds s25, [r3, #24] @ D.14385, MEM[base: _575, offset: 24B]

flds s24, [r3, #28] @ D.14385, MEM[base: _575, offset: 28B]

flds s14, [r2, #20] @ D.14385, MEM[base: _576, offset: 20B]

flds s26, [r2, #36] @ D.14385, MEM[base: _576, offset: 36B]

flds s22, [r3, #4] @ D.14385, MEM[base: _575, offset: 4B]

flds s20, [r3, #8] @ D.14385, MEM[base: _575, offset: 8B]

flds s19, [r3, #12] @ D.14385, MEM[base: _575, offset: 12B]

flds s29, [r2, #8] @ D.14385, MEM[base: _576, offset: 8B]

flds s18, [r3, #32] @ D.14385, MEM[base: _575, offset: 32B]

flds s30, [r2, #16] @ D.14385, MEM[base: _576, offset: 16B]

flds s21, [r2, #32] @ D.14385, MEM[base: _576, offset: 32B]

flds s17, [r3, #36] @ D.14385, MEM[base: _575, offset: 36B]

flds s0, [r3, #40] @ D.14385, MEM[base: _575, offset: 40B]

flds s1, [r3, #44] @ D.14385, MEM[base: _575, offset: 44B]

flds s2, [r3, #48] @ D.14385, MEM[base: _575, offset: 48B]

flds s16, [r2, #40] @ D.14385, MEM[base: _576, offset: 40B]

flds s3, [r3, #52] @ D.14385, MEM[base: _575, offset: 52B]

flds s5, [r3, #56] @ D.14385, MEM[base: _575, offset: 56B]

flds s6, [r3, #60] @ D.14385, MEM[base: _575, offset: 60B]

flds s4, [r2, #44] @ D.14385, MEM[base: _576, offset: 44B]

fmacs s7, s31, s23 @ D.14385, D.14385, D.14385

add r3, r5, ip @ D.14386, p_res, ivtmp.11

add r4, r4, #16 @ idx, idx,

add ip, ip, #64 @ ivtmp.11, ivtmp.11,

cmp r4, r6 @ idx, D.14384

fmuls s8, s10, s27 @ D.14385, D.14385, D.14385

fmuls s9, s10, s25 @ D.14385, D.14385, D.14385

fmuls s11, s28, s14 @ D.14385, D.14385, D.14385

fmuls s12, s27, s14 @ D.14385, D.14385, D.14385

fmuls s13, s25, s14 @ D.14385, D.14385, D.14385

fmuls s15, s28, s26 @ D.14385, D.14385, D.14385

fmuls s10, s10, s24 @ D.14385, D.14385, D.14385

fmuls s14, s24, s14 @ D.14385, D.14385, D.14385

fmacs s7, s29, s18 @ D.14385, D.14385, D.14385

fmacs s8, s31, s22 @ D.14385, D.14385, D.14385

fmacs s9, s31, s20 @ D.14385, D.14385, D.14385

fmacs s10, s31, s19 @ D.14385, D.14385, D.14385

fmacs s11, s23, s30 @ D.14385, D.14385, D.14385

fmacs s12, s22, s30 @ D.14385, D.14385, D.14385

fmacs s13, s20, s30 @ D.14385, D.14385, D.14385

fmacs s14, s19, s30 @ D.14385, D.14385, D.14385

fmacs s15, s23, s21 @ D.14385, D.14385, D.14385

fmacs s8, s29, s17 @ D.14385, D.14385, D.14385

fmacs s9, s29, s0 @ D.14385, D.14385, D.14385

fmacs s10, s29, s1 @ D.14385, D.14385, D.14385

flds s29, [r2, #24] @, MEM[base: _576, offset: 24B]

fmacs s15, s18, s16 @ D.14385, D.14385, D.14385

fmacs s11, s18, s29 @ D.14385, D.14385,

fmacs s12, s17, s29 @ D.14385, D.14385,

fmacs s13, s0, s29 @ D.14385, D.14385,

fmacs s14, s1, s29 @ D.14385, D.14385,

flds s29, [r2, #12] @, MEM[base: _576, offset: 12B]

fmacs s7, s29, s2 @ D.14385,, D.14385

fmacs s10, s29, s6 @ D.14385,, D.14385

fsts s7, [r3] @ D.14385, MEM[base: _568, offset: 0B]

flds s7, [r2, #28] @, MEM[base: _576, offset: 28B]

fmacs s15, s2, s4 @ D.14385, D.14385, D.14385

fsts s10, [r3, #12] @ D.14385, MEM[base: _568, offset: 12B]

fmacs s11, s2, s7 @ D.14385, D.14385,

fmacs s12, s3, s7 @ D.14385, D.14385,

fsts s11, [r3, #16] @ D.14385, MEM[base: _568, offset: 16B]

fmacs s14, s6, s7 @ D.14385, D.14385,

fmacs s8, s29, s3 @ D.14385,, D.14385

fmacs s9, s29, s5 @ D.14385,, D.14385

fsts s8, [r3, #4] @ D.14385, MEM[base: _568, offset: 4B]

fsts s9, [r3, #8] @ D.14385, MEM[base: _568, offset: 8B]

fsts s12, [r3, #20] @ D.14385, MEM[base: _568, offset: 20B]

flds s10, [r2, #52] @ D.14385, MEM[base: _576, offset: 52B]

fsts s14, [r3, #28] @ D.14385, MEM[base: _568, offset: 28B]

fsts s15, [r3, #32] @ D.14385, MEM[base: _568, offset: 32B]

fmuls s14, s27, s26 @ D.14385, D.14385, D.14385

flds s11, [r2, #48] @ D.14385, MEM[base: _576, offset: 48B]

flds s12, [r2, #56] @ D.14385, MEM[base: _576, offset: 56B]

fmuls s15, s25, s26 @ D.14385, D.14385, D.14385

fmuls s28, s28, s10 @ D.14385, D.14385, D.14385

fmuls s26, s24, s26 @ D.14385, D.14385, D.14385

fmuls s27, s27, s10 @ D.14385, D.14385, D.14385

fmuls s25, s25, s10 @ D.14385, D.14385, D.14385

fmuls s24, s24, s10 @ D.14385, D.14385, D.14385

fmacs s28, s23, s11 @ D.14385, D.14385, D.14385

fmacs s14, s22, s21 @ D.14385, D.14385, D.14385

fmacs s27, s22, s11 @ D.14385, D.14385, D.14385

fmacs s15, s20, s21 @ D.14385, D.14385, D.14385

fmacs s26, s19, s21 @ D.14385, D.14385, D.14385

fmacs s25, s20, s11 @ D.14385, D.14385, D.14385

fmacs s24, s19, s11 @ D.14385, D.14385, D.14385

fmacs s13, s5, s7 @ D.14385, D.14385,

fmacs s28, s18, s12 @ D.14385, D.14385, D.14385

fsts s13, [r3, #24] @ D.14385, MEM[base: _568, offset: 24B]

flds s13, [r2, #60] @ D.14385, MEM[base: _576, offset: 60B]

fmacs s14, s17, s16 @ D.14385, D.14385, D.14385

fmacs s27, s17, s12 @ D.14385, D.14385, D.14385

fmacs s15, s0, s16 @ D.14385, D.14385, D.14385

fmacs s26, s1, s16 @ D.14385, D.14385, D.14385

fmacs s25, s0, s12 @ D.14385, D.14385, D.14385

fmacs s24, s1, s12 @ D.14385, D.14385, D.14385

fmacs s28, s2, s13 @ D.14385, D.14385, D.14385

fmacs s14, s3, s4 @ D.14385, D.14385, D.14385

fsts s28, [r3, #48] @ D.14385, MEM[base: _568, offset: 48B]

fmacs s27, s3, s13 @ D.14385, D.14385, D.14385

fsts s14, [r3, #36] @ D.14385, MEM[base: _568, offset: 36B]

fmacs s15, s5, s4 @ D.14385, D.14385, D.14385

fsts s27, [r3, #52] @ D.14385, MEM[base: _568, offset: 52B]

fmacs s26, s6, s4 @ D.14385, D.14385, D.14385

fsts s15, [r3, #40] @ D.14385, MEM[base: _568, offset: 40B]

fmacs s25, s5, s13 @ D.14385, D.14385, D.14385

fsts s26, [r3, #44] @ D.14385, MEM[base: _568, offset: 44B]

fmacs s24, s6, s13 @ D.14385, D.14385, D.14385

fsts s25, [r3, #56] @ D.14385, MEM[base: _568, offset: 56B]

fsts s24, [r3, #60] @ D.14385, MEM[base: _568, offset: 60B]

========================================================

2 - neon code (intrinsic)

add r4, r0, #32 @ tmp159, src0_4x4,

vld1.32 {d16-d17}, [r3]! @ D.14261, MEM[(const __builtin_neon_sf[4] *)src0_4x4_10(D)]

add r0, r0, #48 @ tmp160, src0_4x4,

vmov d1, d16 @ v2sf @ D.14262, D.14261

add r5, r2, #32 @ tmp162, res4x4,

vld1.32 {d18-d19}, [r3] @ D.14261, MEM[(const __builtin_neon_sf[4] *)_12]

vmov d5, d17 @ v2sf @ D.14262, D.14261

vmov d0, d18 @ v2sf @ D.14262, D.14261

add r3, r1, #32 @ tmp156, src1_4x4,

vld1.32 {d20-d21}, [r4] @ D.14261, MEM[(const __builtin_neon_sf[4] *)_14]

vmov d4, d19 @ v2sf @ D.14262, D.14261

vmov d2, d20 @ v2sf @ D.14262, D.14261

add r1, r1, #48 @ tmp157, src1_4x4,

vld1.32 {d18-d19}, [r0] @ D.14261, MEM[(const __builtin_neon_sf[4] *)_16]

vmov d6, d21 @ v2sf @ D.14262, D.14261

vmov d3, d18 @ v2sf @ D.14262, D.14261

add r4, r2, #48 @ tmp163, res4x4,

vld1.32 {d16-d17}, [ip]! @ D.14261, MEM[(const __builtin_neon_sf[4] *)src1_4x4_2(D)]

vmov d7, d19 @ v2sf @ D.14262, D.14261

mov r0, #0 @,

vmul.f32 q10, q8, d0[0] @ D.14261, D.14261, D.14262,

vld1.32 {d28-d29}, [ip] @ D.14261, MEM[(const __builtin_neon_sf[4] *)_4]

vmul.f32 q11, q8, d1[0] @ D.14261, D.14261, D.14262,

vld1.32 {d26-d27}, [r3] @ D.14261, MEM[(const __builtin_neon_sf[4] *)_6]

vmul.f32 q9, q8, d2[0] @ D.14261, D.14261, D.14262,

vld1.32 {d24-d25}, [r1] @ D.14261, MEM[(const __builtin_neon_sf[4] *)_8]

vmul.f32 q8, q8, d3[0] @ D.14261, D.14261, D.14262,

vmla.f32 q10, q14, d0[1] @ D.14261, D.14261, D.14262,

vmla.f32 q11, q14, d1[1] @ D.14261, D.14261, D.14262,

vmla.f32 q9, q14, d2[1] @ D.14261, D.14261, D.14262,

vmla.f32 q8, q14, d3[1] @ D.14261, D.14261, D.14262,

vmla.f32 q10, q13, d4[0] @ D.14261, D.14261, D.14262,

vmla.f32 q11, q13, d5[0] @ D.14261, D.14261, D.14262,

vmla.f32 q9, q13, d6[0] @ D.14261, D.14261, D.14262,

vmla.f32 q8, q13, d7[0] @ D.14261, D.14261, D.14262,

vmla.f32 q10, q12, d4[1] @ D.14261, D.14261, D.14262,

vmla.f32 q11, q12, d5[1] @ D.14261, D.14261, D.14262,

vmla.f32 q9, q12, d6[1] @ D.14261, D.14261, D.14262,

vmla.f32 q8, q12, d7[1] @ D.14261, D.14261, D.14262,

vst1.32 {d22-d23}, [r2]! @ D.14261, MEM[(__builtin_neon_sf[4] *)res4x4_42(D)]

vst1.32 {d20-d21}, [r2] @ D.14261, MEM[(__builtin_neon_sf[4] *)_44]

vst1.32 {d18-d19}, [r5] @ D.14261, MEM[(__builtin_neon_sf[4] *)_46]

vst1.32 {d16-d17}, [r4] @ D.14261, MEM[(__builtin_neon_sf[4] *)_48]

========================================================

3 - neon inline asm code

@ 68 "../src/mat.c" 1

vld1.32 { q8-q9 }, [r1]! @,

@ 0 "" 2

@ 69 "../src/mat.c" 1

vld1.32 { q10-q11 }, [r1]! @,

@ 0 "" 2

@ 72 "../src/mat.c" 1

vld1.32 { q0-q1 }, [r0]! @,

@ 0 "" 2

@ 73 "../src/mat.c" 1

vld1.32 { q2-q3 }, [r0]! @,

@ 0 "" 2

@ 76 "../src/mat.c" 1

vmul.f32 q12, q8, d0[0] @,

@ 0 "" 2

@ 77 "../src/mat.c" 1

vmul.f32 q13, q8, d2[0] @,

@ 0 "" 2

@ 78 "../src/mat.c" 1

vmul.f32 q14, q8, d4[0] @,

@ 0 "" 2

@ 79 "../src/mat.c" 1

vmul.f32 q15, q8, d6[0] @,

@ 0 "" 2

@ 81 "../src/mat.c" 1

vmla.f32 q12, q9, d0[1] @,

@ 0 "" 2

@ 82 "../src/mat.c" 1

vmla.f32 q13, q9, d2[1] @,

@ 0 "" 2

@ 83 "../src/mat.c" 1

vmla.f32 q14, q9, d4[1] @,

@ 0 "" 2

@ 84 "../src/mat.c" 1

vmla.f32 q15, q9, d6[1] @,

@ 0 "" 2

@ 86 "../src/mat.c" 1

vmla.f32 q12, q10, d1[0] @,

@ 0 "" 2

@ 87 "../src/mat.c" 1

vmla.f32 q13, q10, d3[0] @,

@ 0 "" 2

@ 88 "../src/mat.c" 1

vmla.f32 q14, q10, d5[0] @,

@ 0 "" 2

@ 89 "../src/mat.c" 1

vmla.f32 q15, q10, d7[0] @,

@ 0 "" 2

@ 91 "../src/mat.c" 1

vmla.f32 q12, q11, d1[1] @,

@ 0 "" 2

@ 92 "../src/mat.c" 1

vmla.f32 q13, q11, d3[1] @,

@ 0 "" 2

@ 93 "../src/mat.c" 1

vmla.f32 q14, q11, d5[1] @,

@ 0 "" 2

@ 94 "../src/mat.c" 1

vmla.f32 q15, q11, d7[1] @,

@ 0 "" 2

@ 97 "../src/mat.c" 1

vst1.32 { q12-q13 }, [r2]! @,

@ 0 "" 2

@ 98 "../src/mat.c" 1

vst1.32 { q14-q15 }, [r2]! @,

@ 0 "" 2

@ 100 "../src/mat.c" 1

mov r0, #0 @,