From a4bba9c594c4022c9f85192bb8fd3593ad1a8d3c Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 2 Apr 2019 15:27:52 +0100 Subject: COMPMID-1995: Fix 32-bit NEDepthwiseConvolution errors. -Updates padding handling in assembly depthwise kernels. -Fixes 32-bit runs issues for depthwise convolution. Change-Id: I3fe6369397c1d13f5629dd34c068ce4af53c95cd Signed-off-by: Georgios Pinitas Reviewed-on: https://review.mlplatform.org/c/939 Reviewed-by: Giuseppe Rossini Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- .../depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp | 1515 ++++++++++++++++++++ 1 file changed, 1515 insertions(+) (limited to 'src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp') diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp index 4ac6276123..010dd81bce 100644 --- a/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp +++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp @@ -429,6 +429,491 @@ void Conv::execute_tile( ); } +template <> +template <> +void Conv::execute_tile( + int n_channels, + const void *weight_bias_ptr, + const float *inptrs[Base::inner_tile_rows][Base::inner_tile_cols], + float *outptrs[Base::output_tile_rows][Base::output_tile_cols] +) +{ + __asm __volatile( + "mov x23, xzr\n" + "mov x24, xzr\n" + "and x25, %[n_channels], #3\n" + "lsr x26, %[n_channels], #2\n" + "cbz x26, 4f\n" + "1:\n" + "ldr q13, [%[wbptr]]\n" + "ldr x19, [%[inptrs], 0]\n" + "mov v10.16b, v13.16b\n" + "ldr q12, [%[wbptr], #16]\n" + "mov v8.16b, v13.16b\n" + "ldr q6, [%[wbptr], #32]\n" + "mov v9.16b, v13.16b\n" + "ldr q5, [%[wbptr], #48]\n" + "mov v7.16b, v13.16b\n" + "ldr q11, [%[wbptr], #64]\n" + "ldr q4, [%[wbptr], #80]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr q3, [%[wbptr], #96]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr q2, [%[wbptr], #112]\n" + "ldr x27, [%[inptrs], 120]\n" + "ldr q1, [%[wbptr], #128]\n" + "subs x26, x26, #1\n" + "ldr q0, [%[wbptr], #144]\n" + "ldr q14, [x19, x23]\n" + "fmla v10.4s, v14.4s, v12.4s\n" + "ldr q18, [x20, x23]\n" + "ldr q14, [x21, x23]\n" + "ldr x19, [%[inptrs], 8]\n" + "ldr q16, [x27, x23]\n" + "ldr x20, [%[inptrs], 48]\n" + "ldr q19, [x19, x23]\n" + "ldr x21, [%[inptrs], 88]\n" + "fmla v10.4s, v18.4s, v11.4s\n" + "ldr q15, [x20, x23]\n" + "ldr q18, [x21, x23]\n" + "ldr x19, [%[inptrs], 16]\n" + "ldr q13, [x19, x23]\n" + "fmla v10.4s, v19.4s, v6.4s\n" + "fmla v10.4s, v14.4s, v2.4s\n" + "beq 3f\n" + "2:\n" + "fmla v8.4s, v14.4s, v12.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v10.4s, v15.4s, v4.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v9.4s, v13.4s, v12.4s\n" + "ldr q14, [x20, x23]\n" + "ldr q17, [x19, x23]\n" + "ldr x22, [%[inptrs], 160]\n" + "fmla v8.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 128]\n" + "fmla v10.4s, v13.4s, v5.4s\n" + "ldr q15, [x22, x23]\n" + "fmla v9.4s, v14.4s, v11.4s\n" + "ldr q19, [x27, x23]\n" + "ldr x21, [%[inptrs], 96]\n" + "ldr x20, [%[inptrs], 64]\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v8.4s, v18.4s, v6.4s\n" + "ldr x22, [%[inptrs], 168]\n" + "fmla v10.4s, v18.4s, v1.4s\n" + "ldr q13, [x21, x23]\n" + "fmla v9.4s, v17.4s, v6.4s\n" + "ldr q18, [x20, x23]\n" + "fmla v7.4s, v13.4s, v12.4s\n" + "ldr q17, [x19, x23]\n" + "fmla v8.4s, v15.4s, v2.4s\n" + "ldr q15, [x22, x23]\n" + "fmla v10.4s, v14.4s, v3.4s\n" + "ldr x27, [%[inptrs], 136]\n" + "fmla v9.4s, v13.4s, v2.4s\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr q16, [x27, x23]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v8.4s, v19.4s, v4.4s\n" + "ldr q19, [x21, x23]\n" + "fmla v10.4s, v13.4s, v0.4s\n" + "ldr q12, [x20, x23]\n" + "fmla v9.4s, v18.4s, v4.4s\n" + "ldr x22, [%[inptrs], 176]\n" + "fmla v7.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 144]\n" + "fmla v8.4s, v13.4s, v5.4s\n" + "ldr q11, [x22, x23]\n" + "ldr q13, [x27, x23]\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v9.4s, v17.4s, v5.4s\n" + "ldr x22, [%[inptrs], 184]\n" + "fmla v7.4s, v19.4s, v6.4s\n" + "ldr q14, [x21, x23]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr q17, [x22, x23]\n" + "ldr x27, [%[inptrs], 152]\n" + "ldr x22, [%[inptrs], 192]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v9.4s, v19.4s, v1.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str q10, [x21, x24]\n" + "fmla v7.4s, v11.4s, v2.4s\n" + "fmla v8.4s, v16.4s, v3.4s\n" + "ldr q16, [x27, x23]\n" + "ldr q15, [x22, x23]\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v9.4s, v12.4s, v3.4s\n" + "add %[wbptr], %[wbptr], #160\n" + "fmla v7.4s, v13.4s, v4.4s\n" + "ldr q13, [%[wbptr]]\n" + "fmla v8.4s, v11.4s, v0.4s\n" + "ldr q12, [%[wbptr], #16]\n" + "mov v10.16b, v13.16b\n" + "ldr q6, [%[wbptr], #32]\n" + "fmla v9.4s, v14.4s, v0.4s\n" + "ldr q11, [%[wbptr], #64]\n" + "fmla v7.4s, v14.4s, v5.4s\n" + "ldr q4, [%[wbptr], #80]\n" + "str q8, [x28, x24]\n" + "add x23, x23, #16\n" + "mov v8.16b, v13.16b\n" + "ldr q2, [%[wbptr], #112]\n" + "str q9, [x21, x24]\n" + "ldr x28, [%[outptrs], 24]\n" + "fmla v7.4s, v17.4s, v1.4s\n" + "ldr q5, [%[wbptr], #48]\n" + "mov v9.16b, v13.16b\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "ldr x19, [%[inptrs], 0]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr x27, [%[inptrs], 120]\n" + "subs x26, x26, #1\n" + "fmla v7.4s, v16.4s, v3.4s\n" + "ldr q1, [%[wbptr], #128]\n" + "ldr q14, [x19, x23]\n" + "fmla v10.4s, v14.4s, v12.4s\n" + "ldr q18, [x20, x23]\n" + "ldr q14, [x21, x23]\n" + "ldr x19, [%[inptrs], 8]\n" + "fmla v7.4s, v15.4s, v0.4s\n" + "ldr q3, [%[wbptr], #96]\n" + "ldr q19, [x19, x23]\n" + "ldr x20, [%[inptrs], 48]\n" + "fmla v10.4s, v18.4s, v11.4s\n" + "ldr q16, [x27, x23]\n" + "ldr q15, [x20, x23]\n" + "ldr x19, [%[inptrs], 16]\n" + "str q7, [x28, x24]\n" + "ldr x21, [%[inptrs], 88]\n" + "mov v7.16b, v13.16b\n" + "ldr q0, [%[wbptr], #144]\n" + "fmla v10.4s, v19.4s, v6.4s\n" + "ldr q13, [x19, x23]\n" + "ldr q18, [x21, x23]\n" + "add x24, x24, #16\n" + "fmla v10.4s, v14.4s, v2.4s\n" + "bne 2b\n" + "3:\n" + "fmla v8.4s, v14.4s, v12.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v10.4s, v15.4s, v4.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v9.4s, v13.4s, v12.4s\n" + "ldr q14, [x20, x23]\n" + "ldr q17, [x19, x23]\n" + "ldr x22, [%[inptrs], 160]\n" + "fmla v8.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 128]\n" + "fmla v10.4s, v13.4s, v5.4s\n" + "ldr q15, [x22, x23]\n" + "fmla v9.4s, v14.4s, v11.4s\n" + "ldr q19, [x27, x23]\n" + "ldr x21, [%[inptrs], 96]\n" + "ldr x20, [%[inptrs], 64]\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v8.4s, v18.4s, v6.4s\n" + "ldr x22, [%[inptrs], 168]\n" + "fmla v10.4s, v18.4s, v1.4s\n" + "ldr q13, [x21, x23]\n" + "fmla v9.4s, v17.4s, v6.4s\n" + "ldr q18, [x20, x23]\n" + "fmla v7.4s, v13.4s, v12.4s\n" + "ldr q17, [x19, x23]\n" + "fmla v8.4s, v15.4s, v2.4s\n" + "ldr q15, [x22, x23]\n" + "fmla v10.4s, v14.4s, v3.4s\n" + "ldr x27, [%[inptrs], 136]\n" + "fmla v9.4s, v13.4s, v2.4s\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr q16, [x27, x23]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v8.4s, v19.4s, v4.4s\n" + "ldr q19, [x21, x23]\n" + "fmla v10.4s, v13.4s, v0.4s\n" + "ldr q12, [x20, x23]\n" + "fmla v9.4s, v18.4s, v4.4s\n" + "ldr x22, [%[inptrs], 176]\n" + "fmla v7.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 144]\n" + "fmla v8.4s, v13.4s, v5.4s\n" + "ldr q11, [x22, x23]\n" + "ldr q13, [x27, x23]\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v9.4s, v17.4s, v5.4s\n" + "ldr x22, [%[inptrs], 184]\n" + "fmla v7.4s, v19.4s, v6.4s\n" + "ldr q14, [x21, x23]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr q17, [x22, x23]\n" + "ldr x27, [%[inptrs], 152]\n" + "ldr x22, [%[inptrs], 192]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v9.4s, v19.4s, v1.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str q10, [x21, x24]\n" + "fmla v7.4s, v11.4s, v2.4s\n" + "fmla v8.4s, v16.4s, v3.4s\n" + "ldr q16, [x27, x23]\n" + "ldr q15, [x22, x23]\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v9.4s, v12.4s, v3.4s\n" + "add %[wbptr], %[wbptr], #160\n" + "fmla v7.4s, v13.4s, v4.4s\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "fmla v8.4s, v11.4s, v0.4s\n" + "add x23, x23, #16\n" + "fmla v9.4s, v14.4s, v0.4s\n" + "fmla v7.4s, v14.4s, v5.4s\n" + "str q8, [x28, x24]\n" + "ldr x28, [%[outptrs], 24]\n" + "str q9, [x21, x24]\n" + "fmla v7.4s, v17.4s, v1.4s\n" + "fmla v7.4s, v16.4s, v3.4s\n" + "fmla v7.4s, v15.4s, v0.4s\n" + "str q7, [x28, x24]\n" + "add x24, x24, #16\n" + "4:\n" + "cbz x25, 7f\n" + "ldr s13, [%[wbptr]]\n" + "mov v10.16b, v13.16b\n" + "ldr s12, [%[wbptr], #4]\n" + "mov v8.16b, v13.16b\n" + "ldr s6, [%[wbptr], #8]\n" + "mov v9.16b, v13.16b\n" + "ldr s5, [%[wbptr], #12]\n" + "mov v7.16b, v13.16b\n" + "ldr s11, [%[wbptr], #16]\n" + "ldr s4, [%[wbptr], #20]\n" + "ldr x19, [%[inptrs], 0]\n" + "ldr s3, [%[wbptr], #24]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr s2, [%[wbptr], #28]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr s1, [%[wbptr], #32]\n" + "ldr x27, [%[inptrs], 120]\n" + "ldr s0, [%[wbptr], #36]\n" + "subs x25, x25, #1\n" + "ldr s14, [x19, x23]\n" + "ldr s18, [x20, x23]\n" + "fmla v10.4s, v14.4s, v12.4s\n" + "ldr s14, [x21, x23]\n" + "ldr s16, [x27, x23]\n" + "ldr x19, [%[inptrs], 8]\n" + "ldr x20, [%[inptrs], 48]\n" + "ldr x21, [%[inptrs], 88]\n" + "ldr s19, [x19, x23]\n" + "fmla v10.4s, v18.4s, v11.4s\n" + "ldr s15, [x20, x23]\n" + "ldr s18, [x21, x23]\n" + "ldr x19, [%[inptrs], 16]\n" + "ldr s13, [x19, x23]\n" + "fmla v10.4s, v19.4s, v6.4s\n" + "fmla v10.4s, v14.4s, v2.4s\n" + "beq 6f\n" + "5:\n" + "fmla v8.4s, v14.4s, v12.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v10.4s, v15.4s, v4.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v9.4s, v13.4s, v12.4s\n" + "ldr s14, [x20, x23]\n" + "ldr s17, [x19, x23]\n" + "ldr x22, [%[inptrs], 160]\n" + "fmla v8.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 128]\n" + "fmla v10.4s, v13.4s, v5.4s\n" + "ldr s15, [x22, x23]\n" + "fmla v9.4s, v14.4s, v11.4s\n" + "ldr s19, [x27, x23]\n" + "ldr x21, [%[inptrs], 96]\n" + "ldr x20, [%[inptrs], 64]\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v8.4s, v18.4s, v6.4s\n" + "ldr x22, [%[inptrs], 168]\n" + "fmla v10.4s, v18.4s, v1.4s\n" + "ldr s13, [x21, x23]\n" + "fmla v9.4s, v17.4s, v6.4s\n" + "ldr s18, [x20, x23]\n" + "fmla v7.4s, v13.4s, v12.4s\n" + "ldr s17, [x19, x23]\n" + "fmla v8.4s, v15.4s, v2.4s\n" + "ldr s15, [x22, x23]\n" + "fmla v10.4s, v14.4s, v3.4s\n" + "ldr x27, [%[inptrs], 136]\n" + "fmla v9.4s, v13.4s, v2.4s\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr s16, [x27, x23]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v8.4s, v19.4s, v4.4s\n" + "ldr s19, [x21, x23]\n" + "fmla v10.4s, v13.4s, v0.4s\n" + "ldr s12, [x20, x23]\n" + "fmla v9.4s, v18.4s, v4.4s\n" + "ldr x22, [%[inptrs], 176]\n" + "fmla v7.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 144]\n" + "fmla v8.4s, v13.4s, v5.4s\n" + "ldr s11, [x22, x23]\n" + "ldr s13, [x27, x23]\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v9.4s, v17.4s, v5.4s\n" + "ldr x22, [%[inptrs], 184]\n" + "fmla v7.4s, v19.4s, v6.4s\n" + "ldr s14, [x21, x23]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr s17, [x22, x23]\n" + "ldr x27, [%[inptrs], 152]\n" + "ldr x22, [%[inptrs], 192]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v9.4s, v19.4s, v1.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str s10, [x21, x24]\n" + "fmla v7.4s, v11.4s, v2.4s\n" + "fmla v8.4s, v16.4s, v3.4s\n" + "ldr s16, [x27, x23]\n" + "ldr s15, [x22, x23]\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v9.4s, v12.4s, v3.4s\n" + "add %[wbptr], %[wbptr], #40\n" + "fmla v7.4s, v13.4s, v4.4s\n" + "ldr s13, [%[wbptr]]\n" + "fmla v8.4s, v11.4s, v0.4s\n" + "ldr s12, [%[wbptr], #4]\n" + "mov v10.16b, v13.16b\n" + "ldr s6, [%[wbptr], #8]\n" + "fmla v9.4s, v14.4s, v0.4s\n" + "ldr s11, [%[wbptr], #16]\n" + "fmla v7.4s, v14.4s, v5.4s\n" + "ldr s4, [%[wbptr], #20]\n" + "str s8, [x28, x24]\n" + "add x23, x23, #4\n" + "mov v8.16b, v13.16b\n" + "ldr s2, [%[wbptr], #28]\n" + "str s9, [x21, x24]\n" + "ldr x28, [%[outptrs], 24]\n" + "fmla v7.4s, v17.4s, v1.4s\n" + "ldr s5, [%[wbptr], #12]\n" + "mov v9.16b, v13.16b\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "ldr x19, [%[inptrs], 0]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr x27, [%[inptrs], 120]\n" + "subs x25, x25, #1\n" + "fmla v7.4s, v16.4s, v3.4s\n" + "ldr s1, [%[wbptr], #32]\n" + "ldr s14, [x19, x23]\n" + "fmla v10.4s, v14.4s, v12.4s\n" + "ldr s18, [x20, x23]\n" + "ldr s14, [x21, x23]\n" + "ldr x19, [%[inptrs], 8]\n" + "fmla v7.4s, v15.4s, v0.4s\n" + "ldr s3, [%[wbptr], #24]\n" + "ldr s19, [x19, x23]\n" + "ldr x20, [%[inptrs], 48]\n" + "fmla v10.4s, v18.4s, v11.4s\n" + "ldr s16, [x27, x23]\n" + "ldr s15, [x20, x23]\n" + "ldr x19, [%[inptrs], 16]\n" + "str s7, [x28, x24]\n" + "ldr x21, [%[inptrs], 88]\n" + "mov v7.16b, v13.16b\n" + "ldr s0, [%[wbptr], #36]\n" + "fmla v10.4s, v19.4s, v6.4s\n" + "ldr s13, [x19, x23]\n" + "ldr s18, [x21, x23]\n" + "add x24, x24, #4\n" + "fmla v10.4s, v14.4s, v2.4s\n" + "bne 5b\n" + "6:\n" + "fmla v8.4s, v14.4s, v12.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v10.4s, v15.4s, v4.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v9.4s, v13.4s, v12.4s\n" + "ldr s14, [x20, x23]\n" + "ldr s17, [x19, x23]\n" + "ldr x22, [%[inptrs], 160]\n" + "fmla v8.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 128]\n" + "fmla v10.4s, v13.4s, v5.4s\n" + "ldr s15, [x22, x23]\n" + "fmla v9.4s, v14.4s, v11.4s\n" + "ldr s19, [x27, x23]\n" + "ldr x21, [%[inptrs], 96]\n" + "ldr x20, [%[inptrs], 64]\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v8.4s, v18.4s, v6.4s\n" + "ldr x22, [%[inptrs], 168]\n" + "fmla v10.4s, v18.4s, v1.4s\n" + "ldr s13, [x21, x23]\n" + "fmla v9.4s, v17.4s, v6.4s\n" + "ldr s18, [x20, x23]\n" + "fmla v7.4s, v13.4s, v12.4s\n" + "ldr s17, [x19, x23]\n" + "fmla v8.4s, v15.4s, v2.4s\n" + "ldr s15, [x22, x23]\n" + "fmla v10.4s, v14.4s, v3.4s\n" + "ldr x27, [%[inptrs], 136]\n" + "fmla v9.4s, v13.4s, v2.4s\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr s16, [x27, x23]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v8.4s, v19.4s, v4.4s\n" + "ldr s19, [x21, x23]\n" + "fmla v10.4s, v13.4s, v0.4s\n" + "ldr s12, [x20, x23]\n" + "fmla v9.4s, v18.4s, v4.4s\n" + "ldr x22, [%[inptrs], 176]\n" + "fmla v7.4s, v16.4s, v11.4s\n" + "ldr x27, [%[inptrs], 144]\n" + "fmla v8.4s, v13.4s, v5.4s\n" + "ldr s11, [x22, x23]\n" + "ldr s13, [x27, x23]\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v9.4s, v17.4s, v5.4s\n" + "ldr x22, [%[inptrs], 184]\n" + "fmla v7.4s, v19.4s, v6.4s\n" + "ldr s14, [x21, x23]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr s17, [x22, x23]\n" + "ldr x27, [%[inptrs], 152]\n" + "ldr x22, [%[inptrs], 192]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v9.4s, v19.4s, v1.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str s10, [x21, x24]\n" + "fmla v7.4s, v11.4s, v2.4s\n" + "fmla v8.4s, v16.4s, v3.4s\n" + "ldr s16, [x27, x23]\n" + "ldr s15, [x22, x23]\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v9.4s, v12.4s, v3.4s\n" + "add %[wbptr], %[wbptr], #40\n" + "fmla v7.4s, v13.4s, v4.4s\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "fmla v8.4s, v11.4s, v0.4s\n" + "add x23, x23, #4\n" + "fmla v9.4s, v14.4s, v0.4s\n" + "fmla v7.4s, v14.4s, v5.4s\n" + "str s8, [x28, x24]\n" + "ldr x28, [%[outptrs], 24]\n" + "str s9, [x21, x24]\n" + "fmla v7.4s, v17.4s, v1.4s\n" + "fmla v7.4s, v16.4s, v3.4s\n" + "fmla v7.4s, v15.4s, v0.4s\n" + "str s7, [x28, x24]\n" + "add x24, x24, #4\n" + "7:\n" + : [wbptr] "+r" (weight_bias_ptr) + : [inptrs] "r" (inptrs), [n_channels] "r" ((long) n_channels), [outptrs] "r" (outptrs) + : "cc", "v0", "v1", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "memory" + ); +} + template <> template <> void Conv::execute_tile( @@ -848,6 +1333,511 @@ void Conv::execute_tile( ); } +template <> +template <> +void Conv::execute_tile( + int n_channels, + const void *weight_bias_ptr, + const float *inptrs[Base::inner_tile_rows][Base::inner_tile_cols], + float *outptrs[Base::output_tile_rows][Base::output_tile_cols] +) +{ + __asm __volatile( + "mov x22, xzr\n" + "mov x26, xzr\n" + "and x23, %[n_channels], #3\n" + "lsr x24, %[n_channels], #2\n" + "cbz x24, 4f\n" + "1:\n" + "ldr q14, [%[wbptr]]\n" + "ldr x19, [%[inptrs], 0]\n" + "mov v3.16b, v14.16b\n" + "ldr q13, [%[wbptr], #16]\n" + "mov v1.16b, v14.16b\n" + "ldr q11, [%[wbptr], #32]\n" + "mov v2.16b, v14.16b\n" + "ldr q4, [%[wbptr], #48]\n" + "mov v0.16b, v14.16b\n" + "ldr q12, [%[wbptr], #64]\n" + "ldr q9, [%[wbptr], #80]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr q8, [%[wbptr], #96]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr q7, [%[wbptr], #112]\n" + "ldr x25, [%[inptrs], 120]\n" + "ldr q6, [%[wbptr], #128]\n" + "subs x24, x24, #1\n" + "ldr q5, [%[wbptr], #144]\n" + "ldr q15, [x19, x22]\n" + "fmla v3.4s, v15.4s, v13.4s\n" + "ldr q17, [x20, x22]\n" + "ldr q16, [x21, x22]\n" + "ldr x19, [%[inptrs], 8]\n" + "ldr q15, [x25, x22]\n" + "ldr x20, [%[inptrs], 48]\n" + "ldr q10, [x19, x22]\n" + "ldr x21, [%[inptrs], 88]\n" + "fmla v3.4s, v17.4s, v12.4s\n" + "ldr q17, [x20, x22]\n" + "ldr q14, [x21, x22]\n" + "ldr x19, [%[inptrs], 16]\n" + "ldr q18, [x19, x22]\n" + "fmla v3.4s, v10.4s, v11.4s\n" + "fmla v3.4s, v16.4s, v7.4s\n" + "beq 3f\n" + "2:\n" + "fmla v1.4s, v16.4s, v13.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v3.4s, v17.4s, v9.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v2.4s, v18.4s, v13.4s\n" + "ldr q16, [x20, x22]\n" + "movi v10.16b, #0\n" + "ldr q17, [x19, x22]\n" + "fmla v1.4s, v15.4s, v12.4s\n" + "ldr x27, [%[inptrs], 160]\n" + "fmla v3.4s, v18.4s, v4.4s\n" + "ldr x25, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v12.4s\n" + "ldr q18, [x27, x22]\n" + "ldr q15, [x25, x22]\n" + "ldr x21, [%[inptrs], 96]\n" + "fmla v1.4s, v14.4s, v11.4s\n" + "ldr x20, [%[inptrs], 64]\n" + "fmla v3.4s, v14.4s, v6.4s\n" + "ldr q14, [x21, x22]\n" + "fmla v2.4s, v17.4s, v11.4s\n" + "ldr q17, [x20, x22]\n" + "fmla v0.4s, v14.4s, v13.4s\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v1.4s, v18.4s, v7.4s\n" + "ldr x27, [%[inptrs], 168]\n" + "fmla v3.4s, v16.4s, v8.4s\n" + "ldr q18, [x19, x22]\n" + "fmla v2.4s, v14.4s, v7.4s\n" + "ldr q13, [x27, x22]\n" + "ldr x25, [%[inptrs], 136]\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v1.4s, v15.4s, v9.4s\n" + "ldr x27, [%[inptrs], 176]\n" + "fmla v3.4s, v14.4s, v5.4s\n" + "ldr q16, [x25, x22]\n" + "fmla v2.4s, v17.4s, v9.4s\n" + "ldr q17, [x21, x22]\n" + "fmla v0.4s, v16.4s, v12.4s\n" + "ldr q12, [x20, x22]\n" + "fmla v1.4s, v14.4s, v4.4s\n" + "ldr q15, [x27, x22]\n" + "fmax v3.4s, v3.4s, v10.4s\n" + "ldr x25, [%[inptrs], 144]\n" + "fmla v2.4s, v18.4s, v4.4s\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v0.4s, v17.4s, v11.4s\n" + "ldr q14, [x25, x22]\n" + "fmla v1.4s, v13.4s, v6.4s\n" + "ldr q11, [x21, x22]\n" + "ldr x27, [%[inptrs], 184]\n" + "ldr x25, [%[inptrs], 152]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str q3, [x21, x26]\n" + "fmla v0.4s, v15.4s, v7.4s\n" + "fmla v1.4s, v16.4s, v8.4s\n" + "ldr q18, [x27, x22]\n" + "ldr q17, [x25, x22]\n" + "ldr x27, [%[inptrs], 192]\n" + "fmla v2.4s, v12.4s, v8.4s\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v0.4s, v14.4s, v9.4s\n" + "ldr q16, [x27, x22]\n" + "fmla v1.4s, v15.4s, v5.4s\n" + "add %[wbptr], %[wbptr], #160\n" + "ldr q14, [%[wbptr]]\n" + "add x22, x22, #16\n" + "fmla v2.4s, v11.4s, v5.4s\n" + "ldr q13, [%[wbptr], #16]\n" + "fmla v0.4s, v11.4s, v4.4s\n" + "ldr q11, [%[wbptr], #32]\n" + "fmax v1.4s, v1.4s, v10.4s\n" + "ldr q12, [%[wbptr], #64]\n" + "mov v3.16b, v14.16b\n" + "ldr q9, [%[wbptr], #80]\n" + "fmax v2.4s, v2.4s, v10.4s\n" + "ldr q7, [%[wbptr], #112]\n" + "str q1, [x28, x26]\n" + "fmla v0.4s, v18.4s, v6.4s\n" + "mov v1.16b, v14.16b\n" + "ldr q4, [%[wbptr], #48]\n" + "str q2, [x21, x26]\n" + "ldr x28, [%[outptrs], 24]\n" + "mov v2.16b, v14.16b\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "fmla v0.4s, v17.4s, v8.4s\n" + "ldr q6, [%[wbptr], #128]\n" + "ldr x19, [%[inptrs], 0]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr x25, [%[inptrs], 120]\n" + "subs x24, x24, #1\n" + "ldr q15, [x19, x22]\n" + "fmla v0.4s, v16.4s, v5.4s\n" + "ldr q8, [%[wbptr], #96]\n" + "fmla v3.4s, v15.4s, v13.4s\n" + "ldr q17, [x20, x22]\n" + "ldr q16, [x21, x22]\n" + "ldr x19, [%[inptrs], 8]\n" + "ldr q15, [x25, x22]\n" + "ldr x20, [%[inptrs], 48]\n" + "fmax v0.4s, v0.4s, v10.4s\n" + "ldr q5, [%[wbptr], #144]\n" + "fmla v3.4s, v17.4s, v12.4s\n" + "ldr q10, [x19, x22]\n" + "ldr q17, [x20, x22]\n" + "ldr x19, [%[inptrs], 16]\n" + "str q0, [x28, x26]\n" + "ldr x21, [%[inptrs], 88]\n" + "mov v0.16b, v14.16b\n" + "ldr q18, [x19, x22]\n" + "fmla v3.4s, v10.4s, v11.4s\n" + "ldr q14, [x21, x22]\n" + "add x26, x26, #16\n" + "fmla v3.4s, v16.4s, v7.4s\n" + "bne 2b\n" + "3:\n" + "fmla v1.4s, v16.4s, v13.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v3.4s, v17.4s, v9.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v2.4s, v18.4s, v13.4s\n" + "ldr q16, [x20, x22]\n" + "movi v10.16b, #0\n" + "ldr q17, [x19, x22]\n" + "fmla v1.4s, v15.4s, v12.4s\n" + "ldr x27, [%[inptrs], 160]\n" + "fmla v3.4s, v18.4s, v4.4s\n" + "ldr x25, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v12.4s\n" + "ldr q18, [x27, x22]\n" + "ldr q15, [x25, x22]\n" + "ldr x21, [%[inptrs], 96]\n" + "fmla v1.4s, v14.4s, v11.4s\n" + "ldr x20, [%[inptrs], 64]\n" + "fmla v3.4s, v14.4s, v6.4s\n" + "ldr q14, [x21, x22]\n" + "fmla v2.4s, v17.4s, v11.4s\n" + "ldr q17, [x20, x22]\n" + "fmla v0.4s, v14.4s, v13.4s\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v1.4s, v18.4s, v7.4s\n" + "ldr x27, [%[inptrs], 168]\n" + "fmla v3.4s, v16.4s, v8.4s\n" + "ldr q18, [x19, x22]\n" + "fmla v2.4s, v14.4s, v7.4s\n" + "ldr q13, [x27, x22]\n" + "ldr x25, [%[inptrs], 136]\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v1.4s, v15.4s, v9.4s\n" + "ldr x27, [%[inptrs], 176]\n" + "fmla v3.4s, v14.4s, v5.4s\n" + "ldr q16, [x25, x22]\n" + "fmla v2.4s, v17.4s, v9.4s\n" + "ldr q17, [x21, x22]\n" + "fmla v0.4s, v16.4s, v12.4s\n" + "ldr q12, [x20, x22]\n" + "fmla v1.4s, v14.4s, v4.4s\n" + "ldr q15, [x27, x22]\n" + "fmax v3.4s, v3.4s, v10.4s\n" + "ldr x25, [%[inptrs], 144]\n" + "fmla v2.4s, v18.4s, v4.4s\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v0.4s, v17.4s, v11.4s\n" + "ldr q14, [x25, x22]\n" + "fmla v1.4s, v13.4s, v6.4s\n" + "ldr q11, [x21, x22]\n" + "ldr x27, [%[inptrs], 184]\n" + "ldr x25, [%[inptrs], 152]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str q3, [x21, x26]\n" + "fmla v0.4s, v15.4s, v7.4s\n" + "fmla v1.4s, v16.4s, v8.4s\n" + "ldr q18, [x27, x22]\n" + "ldr q17, [x25, x22]\n" + "ldr x27, [%[inptrs], 192]\n" + "fmla v2.4s, v12.4s, v8.4s\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v0.4s, v14.4s, v9.4s\n" + "ldr q16, [x27, x22]\n" + "fmla v1.4s, v15.4s, v5.4s\n" + "add %[wbptr], %[wbptr], #160\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "add x22, x22, #16\n" + "fmla v2.4s, v11.4s, v5.4s\n" + "fmla v0.4s, v11.4s, v4.4s\n" + "fmax v1.4s, v1.4s, v10.4s\n" + "fmax v2.4s, v2.4s, v10.4s\n" + "str q1, [x28, x26]\n" + "fmla v0.4s, v18.4s, v6.4s\n" + "ldr x28, [%[outptrs], 24]\n" + "str q2, [x21, x26]\n" + "fmla v0.4s, v17.4s, v8.4s\n" + "fmla v0.4s, v16.4s, v5.4s\n" + "fmax v0.4s, v0.4s, v10.4s\n" + "str q0, [x28, x26]\n" + "add x26, x26, #16\n" + "4:\n" + "cbz x23, 7f\n" + "ldr s14, [%[wbptr]]\n" + "mov v3.16b, v14.16b\n" + "ldr s13, [%[wbptr], #4]\n" + "mov v1.16b, v14.16b\n" + "ldr s11, [%[wbptr], #8]\n" + "mov v2.16b, v14.16b\n" + "ldr s4, [%[wbptr], #12]\n" + "mov v0.16b, v14.16b\n" + "ldr s12, [%[wbptr], #16]\n" + "ldr s9, [%[wbptr], #20]\n" + "ldr x19, [%[inptrs], 0]\n" + "ldr s8, [%[wbptr], #24]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr s7, [%[wbptr], #28]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr s6, [%[wbptr], #32]\n" + "ldr x25, [%[inptrs], 120]\n" + "ldr s5, [%[wbptr], #36]\n" + "subs x23, x23, #1\n" + "ldr s15, [x19, x22]\n" + "ldr s17, [x20, x22]\n" + "fmla v3.4s, v15.4s, v13.4s\n" + "ldr s16, [x21, x22]\n" + "ldr s15, [x25, x22]\n" + "ldr x19, [%[inptrs], 8]\n" + "ldr x20, [%[inptrs], 48]\n" + "ldr x21, [%[inptrs], 88]\n" + "ldr s10, [x19, x22]\n" + "fmla v3.4s, v17.4s, v12.4s\n" + "ldr s17, [x20, x22]\n" + "ldr s14, [x21, x22]\n" + "ldr x19, [%[inptrs], 16]\n" + "ldr s18, [x19, x22]\n" + "fmla v3.4s, v10.4s, v11.4s\n" + "fmla v3.4s, v16.4s, v7.4s\n" + "beq 6f\n" + "5:\n" + "fmla v1.4s, v16.4s, v13.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v3.4s, v17.4s, v9.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v2.4s, v18.4s, v13.4s\n" + "ldr s16, [x20, x22]\n" + "movi v10.16b, #0\n" + "ldr s17, [x19, x22]\n" + "fmla v1.4s, v15.4s, v12.4s\n" + "ldr x27, [%[inptrs], 160]\n" + "fmla v3.4s, v18.4s, v4.4s\n" + "ldr x25, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v12.4s\n" + "ldr s18, [x27, x22]\n" + "ldr s15, [x25, x22]\n" + "ldr x21, [%[inptrs], 96]\n" + "fmla v1.4s, v14.4s, v11.4s\n" + "ldr x20, [%[inptrs], 64]\n" + "fmla v3.4s, v14.4s, v6.4s\n" + "ldr s14, [x21, x22]\n" + "fmla v2.4s, v17.4s, v11.4s\n" + "ldr s17, [x20, x22]\n" + "fmla v0.4s, v14.4s, v13.4s\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v1.4s, v18.4s, v7.4s\n" + "ldr x27, [%[inptrs], 168]\n" + "fmla v3.4s, v16.4s, v8.4s\n" + "ldr s18, [x19, x22]\n" + "fmla v2.4s, v14.4s, v7.4s\n" + "ldr s13, [x27, x22]\n" + "ldr x25, [%[inptrs], 136]\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v1.4s, v15.4s, v9.4s\n" + "ldr x27, [%[inptrs], 176]\n" + "fmla v3.4s, v14.4s, v5.4s\n" + "ldr s16, [x25, x22]\n" + "fmla v2.4s, v17.4s, v9.4s\n" + "ldr s17, [x21, x22]\n" + "fmla v0.4s, v16.4s, v12.4s\n" + "ldr s12, [x20, x22]\n" + "fmla v1.4s, v14.4s, v4.4s\n" + "ldr s15, [x27, x22]\n" + "fmax v3.4s, v3.4s, v10.4s\n" + "ldr x25, [%[inptrs], 144]\n" + "fmla v2.4s, v18.4s, v4.4s\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v0.4s, v17.4s, v11.4s\n" + "ldr s14, [x25, x22]\n" + "fmla v1.4s, v13.4s, v6.4s\n" + "ldr s11, [x21, x22]\n" + "ldr x27, [%[inptrs], 184]\n" + "ldr x25, [%[inptrs], 152]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str s3, [x21, x26]\n" + "fmla v0.4s, v15.4s, v7.4s\n" + "fmla v1.4s, v16.4s, v8.4s\n" + "ldr s18, [x27, x22]\n" + "ldr s17, [x25, x22]\n" + "ldr x27, [%[inptrs], 192]\n" + "fmla v2.4s, v12.4s, v8.4s\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v0.4s, v14.4s, v9.4s\n" + "ldr s16, [x27, x22]\n" + "fmla v1.4s, v15.4s, v5.4s\n" + "add %[wbptr], %[wbptr], #40\n" + "ldr s14, [%[wbptr]]\n" + "add x22, x22, #4\n" + "fmla v2.4s, v11.4s, v5.4s\n" + "ldr s13, [%[wbptr], #4]\n" + "fmla v0.4s, v11.4s, v4.4s\n" + "ldr s11, [%[wbptr], #8]\n" + "fmax v1.4s, v1.4s, v10.4s\n" + "ldr s12, [%[wbptr], #16]\n" + "mov v3.16b, v14.16b\n" + "ldr s9, [%[wbptr], #20]\n" + "fmax v2.4s, v2.4s, v10.4s\n" + "ldr s7, [%[wbptr], #28]\n" + "str s1, [x28, x26]\n" + "fmla v0.4s, v18.4s, v6.4s\n" + "mov v1.16b, v14.16b\n" + "ldr s4, [%[wbptr], #12]\n" + "str s2, [x21, x26]\n" + "ldr x28, [%[outptrs], 24]\n" + "mov v2.16b, v14.16b\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "fmla v0.4s, v17.4s, v8.4s\n" + "ldr s6, [%[wbptr], #32]\n" + "ldr x19, [%[inptrs], 0]\n" + "ldr x20, [%[inptrs], 40]\n" + "ldr x21, [%[inptrs], 80]\n" + "ldr x25, [%[inptrs], 120]\n" + "subs x23, x23, #1\n" + "ldr s15, [x19, x22]\n" + "fmla v0.4s, v16.4s, v5.4s\n" + "ldr s8, [%[wbptr], #24]\n" + "fmla v3.4s, v15.4s, v13.4s\n" + "ldr s17, [x20, x22]\n" + "ldr s16, [x21, x22]\n" + "ldr x19, [%[inptrs], 8]\n" + "ldr s15, [x25, x22]\n" + "ldr x20, [%[inptrs], 48]\n" + "fmax v0.4s, v0.4s, v10.4s\n" + "ldr s5, [%[wbptr], #36]\n" + "fmla v3.4s, v17.4s, v12.4s\n" + "ldr s10, [x19, x22]\n" + "ldr s17, [x20, x22]\n" + "ldr x19, [%[inptrs], 16]\n" + "str s0, [x28, x26]\n" + "ldr x21, [%[inptrs], 88]\n" + "mov v0.16b, v14.16b\n" + "ldr s18, [x19, x22]\n" + "fmla v3.4s, v10.4s, v11.4s\n" + "ldr s14, [x21, x22]\n" + "add x26, x26, #4\n" + "fmla v3.4s, v16.4s, v7.4s\n" + "bne 5b\n" + "6:\n" + "fmla v1.4s, v16.4s, v13.4s\n" + "ldr x20, [%[inptrs], 56]\n" + "fmla v3.4s, v17.4s, v9.4s\n" + "ldr x19, [%[inptrs], 24]\n" + "fmla v2.4s, v18.4s, v13.4s\n" + "ldr s16, [x20, x22]\n" + "movi v10.16b, #0\n" + "ldr s17, [x19, x22]\n" + "fmla v1.4s, v15.4s, v12.4s\n" + "ldr x27, [%[inptrs], 160]\n" + "fmla v3.4s, v18.4s, v4.4s\n" + "ldr x25, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v12.4s\n" + "ldr s18, [x27, x22]\n" + "ldr s15, [x25, x22]\n" + "ldr x21, [%[inptrs], 96]\n" + "fmla v1.4s, v14.4s, v11.4s\n" + "ldr x20, [%[inptrs], 64]\n" + "fmla v3.4s, v14.4s, v6.4s\n" + "ldr s14, [x21, x22]\n" + "fmla v2.4s, v17.4s, v11.4s\n" + "ldr s17, [x20, x22]\n" + "fmla v0.4s, v14.4s, v13.4s\n" + "ldr x19, [%[inptrs], 32]\n" + "fmla v1.4s, v18.4s, v7.4s\n" + "ldr x27, [%[inptrs], 168]\n" + "fmla v3.4s, v16.4s, v8.4s\n" + "ldr s18, [x19, x22]\n" + "fmla v2.4s, v14.4s, v7.4s\n" + "ldr s13, [x27, x22]\n" + "ldr x25, [%[inptrs], 136]\n" + "ldr x21, [%[inptrs], 104]\n" + "ldr x20, [%[inptrs], 72]\n" + "fmla v1.4s, v15.4s, v9.4s\n" + "ldr x27, [%[inptrs], 176]\n" + "fmla v3.4s, v14.4s, v5.4s\n" + "ldr s16, [x25, x22]\n" + "fmla v2.4s, v17.4s, v9.4s\n" + "ldr s17, [x21, x22]\n" + "fmla v0.4s, v16.4s, v12.4s\n" + "ldr s12, [x20, x22]\n" + "fmla v1.4s, v14.4s, v4.4s\n" + "ldr s15, [x27, x22]\n" + "fmax v3.4s, v3.4s, v10.4s\n" + "ldr x25, [%[inptrs], 144]\n" + "fmla v2.4s, v18.4s, v4.4s\n" + "ldr x21, [%[inptrs], 112]\n" + "fmla v0.4s, v17.4s, v11.4s\n" + "ldr s14, [x25, x22]\n" + "fmla v1.4s, v13.4s, v6.4s\n" + "ldr s11, [x21, x22]\n" + "ldr x27, [%[inptrs], 184]\n" + "ldr x25, [%[inptrs], 152]\n" + "ldr x21, [%[outptrs], 0]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr x28, [%[outptrs], 16]\n" + "str s3, [x21, x26]\n" + "fmla v0.4s, v15.4s, v7.4s\n" + "fmla v1.4s, v16.4s, v8.4s\n" + "ldr s18, [x27, x22]\n" + "ldr s17, [x25, x22]\n" + "ldr x27, [%[inptrs], 192]\n" + "fmla v2.4s, v12.4s, v8.4s\n" + "ldr x21, [%[outptrs], 8]\n" + "fmla v0.4s, v14.4s, v9.4s\n" + "ldr s16, [x27, x22]\n" + "fmla v1.4s, v15.4s, v5.4s\n" + "add %[wbptr], %[wbptr], #40\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "add x22, x22, #4\n" + "fmla v2.4s, v11.4s, v5.4s\n" + "fmla v0.4s, v11.4s, v4.4s\n" + "fmax v1.4s, v1.4s, v10.4s\n" + "fmax v2.4s, v2.4s, v10.4s\n" + "str s1, [x28, x26]\n" + "fmla v0.4s, v18.4s, v6.4s\n" + "ldr x28, [%[outptrs], 24]\n" + "str s2, [x21, x26]\n" + "fmla v0.4s, v17.4s, v8.4s\n" + "fmla v0.4s, v16.4s, v5.4s\n" + "fmax v0.4s, v0.4s, v10.4s\n" + "str s0, [x28, x26]\n" + "add x26, x26, #4\n" + "7:\n" + : [wbptr] "+r" (weight_bias_ptr) + : [inptrs] "r" (inptrs), [n_channels] "r" ((long) n_channels), [outptrs] "r" (outptrs) + : "cc", "v0", "v1", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "memory" + ); +} + template <> template <> void Conv::execute_tile( @@ -1287,6 +2277,531 @@ void Conv::execute_tile( ); } +template <> +template <> +void Conv::execute_tile( + int n_channels, + const void *weight_bias_ptr, + const float *inptrs[Base::inner_tile_rows][Base::inner_tile_cols], + float *outptrs[Base::output_tile_rows][Base::output_tile_cols] +) +{ + __asm __volatile( + "mov x27, xzr\n" + "mov x28, xzr\n" + "and x26, %[n_channels], #3\n" + "lsr x25, %[n_channels], #2\n" + "cbz x25, 4f\n" + "1:\n" + "ldr q15, [%[wbptr]]\n" + "ldr x21, [%[inptrs], 0]\n" + "mov v8.16b, v15.16b\n" + "ldr q14, [%[wbptr], #16]\n" + "mov v3.16b, v15.16b\n" + "ldr q10, [%[wbptr], #32]\n" + "mov v2.16b, v15.16b\n" + "ldr q7, [%[wbptr], #48]\n" + "mov v4.16b, v15.16b\n" + "ldr q13, [%[wbptr], #64]\n" + "ldr q5, [%[wbptr], #80]\n" + "ldr x22, [%[inptrs], 40]\n" + "ldr q0, [%[wbptr], #96]\n" + "ldr x20, [%[inptrs], 80]\n" + "ldr q9, [%[wbptr], #112]\n" + "ldr x23, [%[inptrs], 120]\n" + "ldr q6, [%[wbptr], #128]\n" + "subs x25, x25, #1\n" + "ldr q1, [%[wbptr], #144]\n" + "ldr q17, [x21, x27]\n" + "fmla v8.4s, v17.4s, v14.4s\n" + "ldr q18, [x22, x27]\n" + "ldr q16, [x20, x27]\n" + "ldr x21, [%[inptrs], 8]\n" + "ldr q17, [x23, x27]\n" + "ldr x22, [%[inptrs], 48]\n" + "ldr q11, [x21, x27]\n" + "ldr x20, [%[inptrs], 88]\n" + "fmla v8.4s, v18.4s, v13.4s\n" + "ldr q19, [x22, x27]\n" + "ldr q15, [x20, x27]\n" + "ldr x21, [%[inptrs], 16]\n" + "ldr q12, [x21, x27]\n" + "fmla v8.4s, v11.4s, v10.4s\n" + "fmla v8.4s, v16.4s, v9.4s\n" + "beq 3f\n" + "2:\n" + "fmla v3.4s, v16.4s, v14.4s\n" + "ldr x22, [%[inptrs], 56]\n" + "fmla v8.4s, v19.4s, v5.4s\n" + "ldr x21, [%[inptrs], 24]\n" + "fmla v2.4s, v12.4s, v14.4s\n" + "ldr q16, [x22, x27]\n" + "movi v11.16b, #0\n" + "ldr q18, [x21, x27]\n" + "fmla v3.4s, v17.4s, v13.4s\n" + "ldr x20, [%[inptrs], 160]\n" + "fmla v8.4s, v12.4s, v7.4s\n" + "ldr x23, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v13.4s\n" + "ldr q19, [x20, x27]\n" + "fmov v12.4s, #6.0\n" + "ldr q17, [x23, x27]\n" + "fmla v3.4s, v15.4s, v10.4s\n" + "ldr x20, [%[inptrs], 96]\n" + "fmla v8.4s, v15.4s, v6.4s\n" + "ldr x22, [%[inptrs], 64]\n" + "fmla v2.4s, v18.4s, v10.4s\n" + "ldr q15, [x20, x27]\n" + "fmla v4.4s, v15.4s, v14.4s\n" + "ldr q18, [x22, x27]\n" + "fmla v3.4s, v19.4s, v9.4s\n" + "ldr x21, [%[inptrs], 32]\n" + "fmla v8.4s, v16.4s, v0.4s\n" + "ldr x20, [%[inptrs], 168]\n" + "fmla v2.4s, v15.4s, v9.4s\n" + "ldr q19, [x21, x27]\n" + "ldr q16, [x20, x27]\n" + "ldr x23, [%[inptrs], 136]\n" + "fmla v3.4s, v17.4s, v5.4s\n" + "ldr x20, [%[inptrs], 104]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr q14, [x23, x27]\n" + "fmla v2.4s, v18.4s, v5.4s\n" + "ldr q17, [x20, x27]\n" + "fmla v4.4s, v14.4s, v13.4s\n" + "ldr x22, [%[inptrs], 72]\n" + "fmla v3.4s, v15.4s, v7.4s\n" + "ldr x20, [%[inptrs], 176]\n" + "fmax v8.4s, v8.4s, v11.4s\n" + "ldr q18, [x22, x27]\n" + "fmla v2.4s, v19.4s, v7.4s\n" + "ldr q13, [x20, x27]\n" + "fmla v4.4s, v17.4s, v10.4s\n" + "ldr x23, [%[inptrs], 144]\n" + "fmla v3.4s, v16.4s, v6.4s\n" + "ldr x20, [%[inptrs], 112]\n" + "fmin v8.4s, v8.4s, v12.4s\n" + "ldr q10, [x23, x27]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr q15, [x20, x27]\n" + "fmla v4.4s, v13.4s, v9.4s\n" + "ldr x20, [%[inptrs], 184]\n" + "fmla v3.4s, v14.4s, v0.4s\n" + "ldr x23, [%[inptrs], 152]\n" + "ldr q9, [x20, x27]\n" + "ldr x22, [%[outptrs], 0]\n" + "fmla v2.4s, v18.4s, v0.4s\n" + "ldr q19, [x23, x27]\n" + "str q8, [x22, x28]\n" + "fmla v4.4s, v10.4s, v5.4s\n" + "fmla v3.4s, v13.4s, v1.4s\n" + "ldr x20, [%[inptrs], 192]\n" + "ldr x22, [%[outptrs], 8]\n" + "ldr x24, [%[outptrs], 16]\n" + "add %[wbptr], %[wbptr], #160\n" + "fmla v2.4s, v15.4s, v1.4s\n" + "ldr q16, [x20, x27]\n" + "fmla v4.4s, v15.4s, v7.4s\n" + "ldr q15, [%[wbptr]]\n" + "fmax v3.4s, v3.4s, v11.4s\n" + "ldr q14, [%[wbptr], #16]\n" + "mov v8.16b, v15.16b\n" + "ldr q10, [%[wbptr], #32]\n" + "fmax v2.4s, v2.4s, v11.4s\n" + "ldr q13, [%[wbptr], #64]\n" + "fmla v4.4s, v9.4s, v6.4s\n" + "ldr q7, [%[wbptr], #48]\n" + "fmin v3.4s, v3.4s, v12.4s\n" + "ldr q5, [%[wbptr], #80]\n" + "fmin v2.4s, v2.4s, v12.4s\n" + "ldr q9, [%[wbptr], #112]\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "add x27, x27, #16\n" + "str q3, [x24, x28]\n" + "fmla v4.4s, v19.4s, v0.4s\n" + "str q2, [x22, x28]\n" + "mov v3.16b, v15.16b\n" + "mov v2.16b, v15.16b\n" + "ldr q6, [%[wbptr], #128]\n" + "ldr x24, [%[outptrs], 24]\n" + "ldr x21, [%[inptrs], 0]\n" + "ldr x22, [%[inptrs], 40]\n" + "fmla v4.4s, v16.4s, v1.4s\n" + "ldr q0, [%[wbptr], #96]\n" + "ldr q17, [x21, x27]\n" + "ldr x20, [%[inptrs], 80]\n" + "fmla v8.4s, v17.4s, v14.4s\n" + "ldr q18, [x22, x27]\n" + "ldr q16, [x20, x27]\n" + "ldr x21, [%[inptrs], 8]\n" + "fmax v4.4s, v4.4s, v11.4s\n" + "ldr q1, [%[wbptr], #144]\n" + "ldr q11, [x21, x27]\n" + "ldr x22, [%[inptrs], 48]\n" + "fmla v8.4s, v18.4s, v13.4s\n" + "ldr x21, [%[inptrs], 16]\n" + "fmin v4.4s, v4.4s, v12.4s\n" + "ldr q19, [x22, x27]\n" + "ldr q12, [x21, x27]\n" + "ldr x23, [%[inptrs], 120]\n" + "ldr x20, [%[inptrs], 88]\n" + "subs x25, x25, #1\n" + "str q4, [x24, x28]\n" + "mov v4.16b, v15.16b\n" + "ldr q17, [x23, x27]\n" + "fmla v8.4s, v11.4s, v10.4s\n" + "ldr q15, [x20, x27]\n" + "add x28, x28, #16\n" + "fmla v8.4s, v16.4s, v9.4s\n" + "bne 2b\n" + "3:\n" + "fmla v3.4s, v16.4s, v14.4s\n" + "ldr x22, [%[inptrs], 56]\n" + "fmla v8.4s, v19.4s, v5.4s\n" + "ldr x21, [%[inptrs], 24]\n" + "fmla v2.4s, v12.4s, v14.4s\n" + "ldr q16, [x22, x27]\n" + "movi v11.16b, #0\n" + "ldr q18, [x21, x27]\n" + "fmla v3.4s, v17.4s, v13.4s\n" + "ldr x20, [%[inptrs], 160]\n" + "fmla v8.4s, v12.4s, v7.4s\n" + "ldr x23, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v13.4s\n" + "ldr q19, [x20, x27]\n" + "fmov v12.4s, #6.0\n" + "ldr q17, [x23, x27]\n" + "fmla v3.4s, v15.4s, v10.4s\n" + "ldr x20, [%[inptrs], 96]\n" + "fmla v8.4s, v15.4s, v6.4s\n" + "ldr x22, [%[inptrs], 64]\n" + "fmla v2.4s, v18.4s, v10.4s\n" + "ldr q15, [x20, x27]\n" + "fmla v4.4s, v15.4s, v14.4s\n" + "ldr q18, [x22, x27]\n" + "fmla v3.4s, v19.4s, v9.4s\n" + "ldr x21, [%[inptrs], 32]\n" + "fmla v8.4s, v16.4s, v0.4s\n" + "ldr x20, [%[inptrs], 168]\n" + "fmla v2.4s, v15.4s, v9.4s\n" + "ldr q19, [x21, x27]\n" + "ldr q16, [x20, x27]\n" + "ldr x23, [%[inptrs], 136]\n" + "fmla v3.4s, v17.4s, v5.4s\n" + "ldr x20, [%[inptrs], 104]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr q14, [x23, x27]\n" + "fmla v2.4s, v18.4s, v5.4s\n" + "ldr q17, [x20, x27]\n" + "fmla v4.4s, v14.4s, v13.4s\n" + "ldr x22, [%[inptrs], 72]\n" + "fmla v3.4s, v15.4s, v7.4s\n" + "ldr x20, [%[inptrs], 176]\n" + "fmax v8.4s, v8.4s, v11.4s\n" + "ldr q18, [x22, x27]\n" + "fmla v2.4s, v19.4s, v7.4s\n" + "ldr q13, [x20, x27]\n" + "fmla v4.4s, v17.4s, v10.4s\n" + "ldr x23, [%[inptrs], 144]\n" + "fmla v3.4s, v16.4s, v6.4s\n" + "ldr x20, [%[inptrs], 112]\n" + "fmin v8.4s, v8.4s, v12.4s\n" + "ldr q10, [x23, x27]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr q15, [x20, x27]\n" + "fmla v4.4s, v13.4s, v9.4s\n" + "ldr x20, [%[inptrs], 184]\n" + "fmla v3.4s, v14.4s, v0.4s\n" + "ldr x23, [%[inptrs], 152]\n" + "ldr q9, [x20, x27]\n" + "ldr x22, [%[outptrs], 0]\n" + "fmla v2.4s, v18.4s, v0.4s\n" + "ldr q19, [x23, x27]\n" + "str q8, [x22, x28]\n" + "fmla v4.4s, v10.4s, v5.4s\n" + "fmla v3.4s, v13.4s, v1.4s\n" + "ldr x20, [%[inptrs], 192]\n" + "ldr x22, [%[outptrs], 8]\n" + "ldr x24, [%[outptrs], 16]\n" + "add %[wbptr], %[wbptr], #160\n" + "fmla v2.4s, v15.4s, v1.4s\n" + "ldr q16, [x20, x27]\n" + "fmla v4.4s, v15.4s, v7.4s\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "fmax v3.4s, v3.4s, v11.4s\n" + "add x27, x27, #16\n" + "fmax v2.4s, v2.4s, v11.4s\n" + "fmla v4.4s, v9.4s, v6.4s\n" + "fmin v3.4s, v3.4s, v12.4s\n" + "fmin v2.4s, v2.4s, v12.4s\n" + "str q3, [x24, x28]\n" + "fmla v4.4s, v19.4s, v0.4s\n" + "str q2, [x22, x28]\n" + "ldr x24, [%[outptrs], 24]\n" + "fmla v4.4s, v16.4s, v1.4s\n" + "fmax v4.4s, v4.4s, v11.4s\n" + "fmin v4.4s, v4.4s, v12.4s\n" + "str q4, [x24, x28]\n" + "add x28, x28, #16\n" + "4:\n" + "cbz x26, 7f\n" + "ldr s15, [%[wbptr]]\n" + "mov v8.16b, v15.16b\n" + "ldr s14, [%[wbptr], #4]\n" + "mov v3.16b, v15.16b\n" + "ldr s10, [%[wbptr], #8]\n" + "mov v2.16b, v15.16b\n" + "ldr s7, [%[wbptr], #12]\n" + "mov v4.16b, v15.16b\n" + "ldr s13, [%[wbptr], #16]\n" + "ldr s5, [%[wbptr], #20]\n" + "ldr x21, [%[inptrs], 0]\n" + "ldr s0, [%[wbptr], #24]\n" + "ldr x22, [%[inptrs], 40]\n" + "ldr s9, [%[wbptr], #28]\n" + "ldr x20, [%[inptrs], 80]\n" + "ldr s6, [%[wbptr], #32]\n" + "ldr x23, [%[inptrs], 120]\n" + "ldr s1, [%[wbptr], #36]\n" + "subs x26, x26, #1\n" + "ldr s17, [x21, x27]\n" + "ldr s18, [x22, x27]\n" + "fmla v8.4s, v17.4s, v14.4s\n" + "ldr s16, [x20, x27]\n" + "ldr s17, [x23, x27]\n" + "ldr x21, [%[inptrs], 8]\n" + "ldr x22, [%[inptrs], 48]\n" + "ldr x20, [%[inptrs], 88]\n" + "ldr s11, [x21, x27]\n" + "fmla v8.4s, v18.4s, v13.4s\n" + "ldr s19, [x22, x27]\n" + "ldr s15, [x20, x27]\n" + "ldr x21, [%[inptrs], 16]\n" + "ldr s12, [x21, x27]\n" + "fmla v8.4s, v11.4s, v10.4s\n" + "fmla v8.4s, v16.4s, v9.4s\n" + "beq 6f\n" + "5:\n" + "fmla v3.4s, v16.4s, v14.4s\n" + "ldr x22, [%[inptrs], 56]\n" + "fmla v8.4s, v19.4s, v5.4s\n" + "ldr x21, [%[inptrs], 24]\n" + "fmla v2.4s, v12.4s, v14.4s\n" + "ldr s16, [x22, x27]\n" + "movi v11.16b, #0\n" + "ldr s18, [x21, x27]\n" + "fmla v3.4s, v17.4s, v13.4s\n" + "ldr x20, [%[inptrs], 160]\n" + "fmla v8.4s, v12.4s, v7.4s\n" + "ldr x23, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v13.4s\n" + "ldr s19, [x20, x27]\n" + "fmov v12.4s, #6.0\n" + "ldr s17, [x23, x27]\n" + "fmla v3.4s, v15.4s, v10.4s\n" + "ldr x20, [%[inptrs], 96]\n" + "fmla v8.4s, v15.4s, v6.4s\n" + "ldr x22, [%[inptrs], 64]\n" + "fmla v2.4s, v18.4s, v10.4s\n" + "ldr s15, [x20, x27]\n" + "fmla v4.4s, v15.4s, v14.4s\n" + "ldr s18, [x22, x27]\n" + "fmla v3.4s, v19.4s, v9.4s\n" + "ldr x21, [%[inptrs], 32]\n" + "fmla v8.4s, v16.4s, v0.4s\n" + "ldr x20, [%[inptrs], 168]\n" + "fmla v2.4s, v15.4s, v9.4s\n" + "ldr s19, [x21, x27]\n" + "ldr s16, [x20, x27]\n" + "ldr x23, [%[inptrs], 136]\n" + "fmla v3.4s, v17.4s, v5.4s\n" + "ldr x20, [%[inptrs], 104]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr s14, [x23, x27]\n" + "fmla v2.4s, v18.4s, v5.4s\n" + "ldr s17, [x20, x27]\n" + "fmla v4.4s, v14.4s, v13.4s\n" + "ldr x22, [%[inptrs], 72]\n" + "fmla v3.4s, v15.4s, v7.4s\n" + "ldr x20, [%[inptrs], 176]\n" + "fmax v8.4s, v8.4s, v11.4s\n" + "ldr s18, [x22, x27]\n" + "fmla v2.4s, v19.4s, v7.4s\n" + "ldr s13, [x20, x27]\n" + "fmla v4.4s, v17.4s, v10.4s\n" + "ldr x23, [%[inptrs], 144]\n" + "fmla v3.4s, v16.4s, v6.4s\n" + "ldr x20, [%[inptrs], 112]\n" + "fmin v8.4s, v8.4s, v12.4s\n" + "ldr s10, [x23, x27]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr s15, [x20, x27]\n" + "fmla v4.4s, v13.4s, v9.4s\n" + "ldr x20, [%[inptrs], 184]\n" + "fmla v3.4s, v14.4s, v0.4s\n" + "ldr x23, [%[inptrs], 152]\n" + "ldr s9, [x20, x27]\n" + "ldr x22, [%[outptrs], 0]\n" + "fmla v2.4s, v18.4s, v0.4s\n" + "ldr s19, [x23, x27]\n" + "str s8, [x22, x28]\n" + "fmla v4.4s, v10.4s, v5.4s\n" + "fmla v3.4s, v13.4s, v1.4s\n" + "ldr x20, [%[inptrs], 192]\n" + "ldr x22, [%[outptrs], 8]\n" + "ldr x24, [%[outptrs], 16]\n" + "add %[wbptr], %[wbptr], #40\n" + "fmla v2.4s, v15.4s, v1.4s\n" + "ldr s16, [x20, x27]\n" + "fmla v4.4s, v15.4s, v7.4s\n" + "ldr s15, [%[wbptr]]\n" + "fmax v3.4s, v3.4s, v11.4s\n" + "ldr s14, [%[wbptr], #4]\n" + "mov v8.16b, v15.16b\n" + "ldr s10, [%[wbptr], #8]\n" + "fmax v2.4s, v2.4s, v11.4s\n" + "ldr s13, [%[wbptr], #16]\n" + "fmla v4.4s, v9.4s, v6.4s\n" + "ldr s7, [%[wbptr], #12]\n" + "fmin v3.4s, v3.4s, v12.4s\n" + "ldr s5, [%[wbptr], #20]\n" + "fmin v2.4s, v2.4s, v12.4s\n" + "ldr s9, [%[wbptr], #28]\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "add x27, x27, #4\n" + "str s3, [x24, x28]\n" + "fmla v4.4s, v19.4s, v0.4s\n" + "str s2, [x22, x28]\n" + "mov v3.16b, v15.16b\n" + "mov v2.16b, v15.16b\n" + "ldr s6, [%[wbptr], #32]\n" + "ldr x24, [%[outptrs], 24]\n" + "ldr x21, [%[inptrs], 0]\n" + "ldr x22, [%[inptrs], 40]\n" + "fmla v4.4s, v16.4s, v1.4s\n" + "ldr s0, [%[wbptr], #24]\n" + "ldr s17, [x21, x27]\n" + "ldr x20, [%[inptrs], 80]\n" + "fmla v8.4s, v17.4s, v14.4s\n" + "ldr s18, [x22, x27]\n" + "ldr s16, [x20, x27]\n" + "ldr x21, [%[inptrs], 8]\n" + "fmax v4.4s, v4.4s, v11.4s\n" + "ldr s1, [%[wbptr], #36]\n" + "ldr s11, [x21, x27]\n" + "ldr x22, [%[inptrs], 48]\n" + "fmla v8.4s, v18.4s, v13.4s\n" + "ldr x21, [%[inptrs], 16]\n" + "fmin v4.4s, v4.4s, v12.4s\n" + "ldr s19, [x22, x27]\n" + "ldr s12, [x21, x27]\n" + "ldr x23, [%[inptrs], 120]\n" + "ldr x20, [%[inptrs], 88]\n" + "subs x26, x26, #1\n" + "str s4, [x24, x28]\n" + "mov v4.16b, v15.16b\n" + "ldr s17, [x23, x27]\n" + "fmla v8.4s, v11.4s, v10.4s\n" + "ldr s15, [x20, x27]\n" + "add x28, x28, #4\n" + "fmla v8.4s, v16.4s, v9.4s\n" + "bne 5b\n" + "6:\n" + "fmla v3.4s, v16.4s, v14.4s\n" + "ldr x22, [%[inptrs], 56]\n" + "fmla v8.4s, v19.4s, v5.4s\n" + "ldr x21, [%[inptrs], 24]\n" + "fmla v2.4s, v12.4s, v14.4s\n" + "ldr s16, [x22, x27]\n" + "movi v11.16b, #0\n" + "ldr s18, [x21, x27]\n" + "fmla v3.4s, v17.4s, v13.4s\n" + "ldr x20, [%[inptrs], 160]\n" + "fmla v8.4s, v12.4s, v7.4s\n" + "ldr x23, [%[inptrs], 128]\n" + "fmla v2.4s, v16.4s, v13.4s\n" + "ldr s19, [x20, x27]\n" + "fmov v12.4s, #6.0\n" + "ldr s17, [x23, x27]\n" + "fmla v3.4s, v15.4s, v10.4s\n" + "ldr x20, [%[inptrs], 96]\n" + "fmla v8.4s, v15.4s, v6.4s\n" + "ldr x22, [%[inptrs], 64]\n" + "fmla v2.4s, v18.4s, v10.4s\n" + "ldr s15, [x20, x27]\n" + "fmla v4.4s, v15.4s, v14.4s\n" + "ldr s18, [x22, x27]\n" + "fmla v3.4s, v19.4s, v9.4s\n" + "ldr x21, [%[inptrs], 32]\n" + "fmla v8.4s, v16.4s, v0.4s\n" + "ldr x20, [%[inptrs], 168]\n" + "fmla v2.4s, v15.4s, v9.4s\n" + "ldr s19, [x21, x27]\n" + "ldr s16, [x20, x27]\n" + "ldr x23, [%[inptrs], 136]\n" + "fmla v3.4s, v17.4s, v5.4s\n" + "ldr x20, [%[inptrs], 104]\n" + "fmla v8.4s, v15.4s, v1.4s\n" + "ldr s14, [x23, x27]\n" + "fmla v2.4s, v18.4s, v5.4s\n" + "ldr s17, [x20, x27]\n" + "fmla v4.4s, v14.4s, v13.4s\n" + "ldr x22, [%[inptrs], 72]\n" + "fmla v3.4s, v15.4s, v7.4s\n" + "ldr x20, [%[inptrs], 176]\n" + "fmax v8.4s, v8.4s, v11.4s\n" + "ldr s18, [x22, x27]\n" + "fmla v2.4s, v19.4s, v7.4s\n" + "ldr s13, [x20, x27]\n" + "fmla v4.4s, v17.4s, v10.4s\n" + "ldr x23, [%[inptrs], 144]\n" + "fmla v3.4s, v16.4s, v6.4s\n" + "ldr x20, [%[inptrs], 112]\n" + "fmin v8.4s, v8.4s, v12.4s\n" + "ldr s10, [x23, x27]\n" + "fmla v2.4s, v17.4s, v6.4s\n" + "ldr s15, [x20, x27]\n" + "fmla v4.4s, v13.4s, v9.4s\n" + "ldr x20, [%[inptrs], 184]\n" + "fmla v3.4s, v14.4s, v0.4s\n" + "ldr x23, [%[inptrs], 152]\n" + "ldr s9, [x20, x27]\n" + "ldr x22, [%[outptrs], 0]\n" + "fmla v2.4s, v18.4s, v0.4s\n" + "ldr s19, [x23, x27]\n" + "str s8, [x22, x28]\n" + "fmla v4.4s, v10.4s, v5.4s\n" + "fmla v3.4s, v13.4s, v1.4s\n" + "ldr x20, [%[inptrs], 192]\n" + "ldr x22, [%[outptrs], 8]\n" + "ldr x24, [%[outptrs], 16]\n" + "add %[wbptr], %[wbptr], #40\n" + "fmla v2.4s, v15.4s, v1.4s\n" + "ldr s16, [x20, x27]\n" + "fmla v4.4s, v15.4s, v7.4s\n" + "prfm pldl1keep, [%[wbptr], #64]\n" + "fmax v3.4s, v3.4s, v11.4s\n" + "add x27, x27, #4\n" + "fmax v2.4s, v2.4s, v11.4s\n" + "fmla v4.4s, v9.4s, v6.4s\n" + "fmin v3.4s, v3.4s, v12.4s\n" + "fmin v2.4s, v2.4s, v12.4s\n" + "str s3, [x24, x28]\n" + "fmla v4.4s, v19.4s, v0.4s\n" + "str s2, [x22, x28]\n" + "ldr x24, [%[outptrs], 24]\n" + "fmla v4.4s, v16.4s, v1.4s\n" + "fmax v4.4s, v4.4s, v11.4s\n" + "fmin v4.4s, v4.4s, v12.4s\n" + "str s4, [x24, x28]\n" + "add x28, x28, #4\n" + "7:\n" + : [wbptr] "+r" (weight_bias_ptr) + : [inptrs] "r" (inptrs), [outptrs] "r" (outptrs), [n_channels] "r" ((long) n_channels) + : "cc", "v0", "v1", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "memory" + ); +} + #endif // __aarch64__ template class DepthwiseConvolution<2, 2, 3, 3, 2, 2, float, float, float>; -- cgit v1.2.1