diff --git a/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32.cc b/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32.cc index eb8f6f648a428b37358a9aff72e6c2843add8f55..5cca0d78ca1ba27f6fd6880007f3218c6af1ae8d 100644 --- a/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32.cc +++ b/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32.cc @@ -2631,7 +2631,7 @@ void conv_depthwise_3x3s1p1_bias_leakyRelu(float *dout, int cnt = cnt_col; asm volatile(INIT_S1 LEFT_COMPUTE_S1 LEFT_RESULT_S1_LEAKY_RELU MID_COMPUTE_S1 MID_RESULT_S1_LEAKY_RELU RIGHT_COMPUTE_S1 - RIGHT_RESULT_S1_LEAKY_RELU + RIGHT_RESULT_S1_LEAKY_RELU : [cnt] "+r"(cnt), [din_ptr0] "+r"(din_ptr0), [din_ptr1] "+r"(din_ptr1), @@ -2729,7 +2729,7 @@ void conv_depthwise_3x3s1p1_bias_leakyRelu(float *dout, unsigned int *vmask_ptr = vmask; asm volatile(INIT_S1 LEFT_COMPUTE_S1 LEFT_RESULT_S1_LEAKY_RELU MID_COMPUTE_S1 MID_RESULT_S1_LEAKY_RELU RIGHT_COMPUTE_S1 - RIGHT_RESULT_S1_LEAKY_RELU + RIGHT_RESULT_S1_LEAKY_RELU : [dout_ptr1] "+r"(doutr0), [dout_ptr2] "+r"(doutr1), [din0_ptr] "+r"(din_ptr0), diff --git a/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32_relu.cc b/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32_relu.cc index 8234502099277346d2b42bb829c739f70710dea2..e6c25fd3b585c636d408ff1985c2f1f634a385c4 100644 --- a/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32_relu.cc +++ b/lite/backends/arm/math/conv3x3s1p01_depthwise_fp32_relu.cc @@ -1376,34 +1376,34 @@ namespace math { [doutr1] "+r"(doutr1), [doutr2] "+r"(doutr2), [doutr3] "+r"(doutr3) - : [w0] "w"(wr0), - [w1] "w"(wr1), - [w2] "w"(wr2), - [bias_val] "r"(vbias), - [vmask] "r"(vmask), - [rmask] "r"(rmask), - [vzero] "w"(vzero) - : "cc", - "memory", - "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16", - "v17", - "v18", + : [w0] "w"(wr0), + [w1] "w"(wr1), + [w2] "w"(wr2), + [bias_val] "r"(vbias), + [vmask] "r"(vmask), + [rmask] "r"(rmask), + [vzero] "w"(vzero) + : "cc", + "memory", + "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16", + "v17", + "v18", "v19", "v20", "v21", @@ -1658,54 +1658,55 @@ void conv_depthwise_3x3s1p1_bias_relu(float *dout, } int cnt = cnt_col; - asm volatile(INIT_S1 LEFT_COMPUTE_S1 LEFT_RESULT_S1_RELU MID_COMPUTE_S1 - MID_RESULT_S1_RELU RIGHT_COMPUTE_S1 RIGHT_RESULT_S1_RELU - : [cnt] "+r"(cnt), - [din_ptr0] "+r"(din_ptr0), - [din_ptr1] "+r"(din_ptr1), - [din_ptr2] "+r"(din_ptr2), - [din_ptr3] "+r"(din_ptr3), - [din_ptr4] "+r"(din_ptr4), - [din_ptr5] "+r"(din_ptr5), - [doutr0] "+r"(doutr0), - [doutr1] "+r"(doutr1), - [doutr2] "+r"(doutr2), - [doutr3] "+r"(doutr3) - : [w0] "w"(wr0), - [w1] "w"(wr1), - [w2] "w"(wr2), - [bias_val] "r"(vbias), - [vmask] "r"(vmask), - [rmask] "r"(rmask), - [vzero] "w"(vzero) - : "cc", - "memory", - "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16", - "v17", - "v18", - "v19", - "v20", - "v21", - "v22", - "v23", - "v24", - "v25"); + asm volatile( + INIT_S1 LEFT_COMPUTE_S1 LEFT_RESULT_S1_RELU MID_COMPUTE_S1 + MID_RESULT_S1_RELU RIGHT_COMPUTE_S1 RIGHT_RESULT_S1_RELU + : [cnt] "+r"(cnt), + [din_ptr0] "+r"(din_ptr0), + [din_ptr1] "+r"(din_ptr1), + [din_ptr2] "+r"(din_ptr2), + [din_ptr3] "+r"(din_ptr3), + [din_ptr4] "+r"(din_ptr4), + [din_ptr5] "+r"(din_ptr5), + [doutr0] "+r"(doutr0), + [doutr1] "+r"(doutr1), + [doutr2] "+r"(doutr2)， + [doutr3] "+r"(doutr3) + : [w0] "w"(wr0), + [w1] "w"(wr1), + [w2] "w"(wr2), + [bias_val] "r"(vbias), + [vmask] "r"(vmask), + [rmask] "r"(rmask), + [vzero] "w"(vzero) + : "cc", + "memory", + "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16", + "v17", + "v18", + "v19", + "v20", + "v21", + "v22", + "v23", + "v24", + "v25"); dout_ptr = dout_ptr + 4 * w_out; } #else @@ -1755,36 +1756,37 @@ void conv_depthwise_3x3s1p1_bias_relu(float *dout, int cnt = cnt_col; unsigned int *rmask_ptr = rmask; unsigned int *vmask_ptr = vmask; - asm volatile(INIT_S1 LEFT_COMPUTE_S1 LEFT_RESULT_S1_RELU MID_COMPUTE_S1 - MID_RESULT_S1_RELU RIGHT_COMPUTE_S1 RIGHT_RESULT_S1_RELU - : [dout_ptr1] "+r"(doutr0), - [dout_ptr2] "+r"(doutr1), - [din0_ptr] "+r"(din_ptr0), - [din1_ptr] "+r"(din_ptr1), - [din2_ptr] "+r"(din_ptr2), - [din3_ptr] "+r"(din_ptr3), - [cnt] "+r"(cnt), - [rmask] "+r"(rmask_ptr), - [vmask] "+r"(vmask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias_val] "r"(bias_val), - [vzero] "w"(vzero) - : "cc", - "memory", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + asm volatile（ + INIT_S1 LEFT_COMPUTE_S1 LEFT_RESULT_S1_RELU MID_COMPUTE_S1 + MID_RESULT_S1_RELU RIGHT_COMPUTE_S1 RIGHT_RESULT_S1_RELU + : [dout_ptr1] "+r"(doutr0), + [dout_ptr2] "+r"(doutr1), + [din0_ptr] "+r"(din_ptr0), + [din1_ptr] "+r"(din_ptr1), + [din2_ptr] "+r"(din_ptr2), + [din3_ptr] "+r"(din_ptr3), + [cnt] "+r"(cnt), + [rmask] "+r"(rmask_ptr), + [vmask] "+r"(vmask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias_val] "r"(bias_val), + [vzero] "w"(vzero) + : "cc", + "memory", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); dout_ptr += 2 * w_out; } //! end of processing mid rows #endif @@ -1874,58 +1876,58 @@ void conv_depthwise_3x3s1p1_bias_s_no_relu(float *dout, [din1] "+r"(dr1), [din2] "+r"(dr2), [din3] "+r"(dr3) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [zero] "w"(vzero), - [mask] "w"(vmask_rp), - [bias] "w"(wbias), - [out1] "r"(out_buf1), - [out2] "r"(out_buf2) - : "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16", - "v17"); + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [zero] "w"(vzero), + [mask] "w"(vmask_rp), + [bias] "w"(wbias), + [out1] "r"(out_buf1), + [out2] "r"(out_buf2) + : "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16", + "v17"); #else asm volatile(COMPUTE_S_S1 RESULT_S_S1 : [din0] "+r"(dr0), - [din1] "+r"(dr1), - [din2] "+r"(dr2), - [din3] "+r"(dr3) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [vzero] "w"(vzero), - [mask] "w"(vmask_rp), - [bias] "w"(wbias), - [out1] "r"(out_buf1), - [out2] "r"(out_buf2) - : "cc", - "memory", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + [din1] "+r"(dr1), + [din2] "+r"(dr2), + [din3] "+r"(dr3) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [vzero] "w"(vzero), + [mask] "w"(vmask_rp), + [bias] "w"(wbias), + [out1] "r"(out_buf1), + [out2] "r"(out_buf2) + : "cc", + "memory", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); #endif for (int w = 0; w < w_out; ++w) { *doutr0++ = out_buf1[w]; @@ -2017,58 +2019,58 @@ void conv_depthwise_3x3s1p1_bias_s_relu(float *dout, [din1] "+r"(dr1), [din2] "+r"(dr2), [din3] "+r"(dr3) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [zero] "w"(vzero), - [mask] "w"(vmask_rp), - [bias] "w"(wbias), - [out1] "r"(out_buf1), - [out2] "r"(out_buf2) - : "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16", - "v17"); + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [zero] "w"(vzero), + [mask] "w"(vmask_rp), + [bias] "w"(wbias), + [out1] "r"(out_buf1), + [out2] "r"(out_buf2) + : "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16", + "v17"); #else asm volatile(COMPUTE_S_S1 RESULT_S_S1_RELU : [din0] "+r"(dr0), - [din1] "+r"(dr1), - [din2] "+r"(dr2), - [din3] "+r"(dr3) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [vzero] "w"(vzero), - [mask] "w"(vmask_rp), - [bias] "w"(wbias), - [out1] "r"(out_buf1), - [out2] "r"(out_buf2) - : "cc", - "memory", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + [din1] "+r"(dr1), + [din2] "+r"(dr2), + [din3] "+r"(dr3) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [vzero] "w"(vzero), + [mask] "w"(vmask_rp), + [bias] "w"(wbias), + [out1] "r"(out_buf1), + [out2] "r"(out_buf2) + : "cc", + "memory", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); #endif for (int w = 0; w < w_out; ++w) { *doutr0++ = out_buf1[w]; @@ -2237,7 +2239,8 @@ void conv_depthwise_3x3s1p0_bias_no_relu(float *dout, "ld1 {v11.4s}, [%[din_ptr5]] \n" /*vld1q_f32(din_ptr0)*/ MID_COMPUTE_S1 MID_RESULT_S1 "cmp %w[remain], #1 \n" - "blt 0f \n" RIGHT_COMPUTE_S1 RIGHT_RESULT_S1 "0: \n" + "blt 0f \n" RIGHT_COMPUTE_S1 RIGHT_RESULT_S1 + "0: \n" : [cnt] "+r"(cnt), [din_ptr0] "+r"(din_ptr0), [din_ptr1] "+r"(din_ptr1), @@ -2334,34 +2337,34 @@ void conv_depthwise_3x3s1p0_bias_no_relu(float *dout, "blt 0f \n" RIGHT_COMPUTE_S1 RIGHT_RESULT_S1 "0: \n" : [dout_ptr1] "+r"(doutr0), - [dout_ptr2] "+r"(doutr1), - [din0_ptr] "+r"(din_ptr0), - [din1_ptr] "+r"(din_ptr1), - [din2_ptr] "+r"(din_ptr2), - [din3_ptr] "+r"(din_ptr3), - [cnt] "+r"(cnt), - [rmask] "+r"(rmask_ptr), - [vmask] "+r"(vmask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias_val] "r"(bias_val), - [vzero] "w"(vzero), - [remain] "r"(remain) - : "cc", - "memory", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + [dout_ptr2] "+r"(doutr1), + [din0_ptr] "+r"(din_ptr0), + [din1_ptr] "+r"(din_ptr1), + [din2_ptr] "+r"(din_ptr2), + [din3_ptr] "+r"(din_ptr3), + [cnt] "+r"(cnt), + [rmask] "+r"(rmask_ptr), + [vmask] "+r"(vmask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias_val] "r"(bias_val), + [vzero] "w"(vzero), + [remain] "r"(remain) + : "cc", + "memory", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); dout_ptr += 2 * w_out; } //! end of processing mid rows #endif @@ -2618,34 +2621,34 @@ void conv_depthwise_3x3s1p0_bias_relu(float *dout, "blt 0f \n" RIGHT_COMPUTE_S1 RIGHT_RESULT_S1_RELU "0: \n" : [dout_ptr1] "+r"(doutr0), - [dout_ptr2] "+r"(doutr1), - [din0_ptr] "+r"(din_ptr0), - [din1_ptr] "+r"(din_ptr1), - [din2_ptr] "+r"(din_ptr2), - [din3_ptr] "+r"(din_ptr3), - [cnt] "+r"(cnt), - [rmask] "+r"(rmask_ptr), - [vmask] "+r"(vmask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias_val] "r"(bias_val), - [vzero] "w"(vzero), - [remain] "r"(remain) - : "cc", - "memory", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + [dout_ptr2] "+r"(doutr1), + [din0_ptr] "+r"(din_ptr0), + [din1_ptr] "+r"(din_ptr1), + [din2_ptr] "+r"(din_ptr2), + [din3_ptr] "+r"(din_ptr3), + [cnt] "+r"(cnt), + [rmask] "+r"(rmask_ptr), + [vmask] "+r"(vmask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias_val] "r"(bias_val), + [vzero] "w"(vzero), + [remain] "r"(remain) + : "cc", + "memory", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); dout_ptr += 2 * w_out; } //! end of processing mid rows #endif @@ -2746,35 +2749,35 @@ void conv_depthwise_3x3s1p0_bias_s_no_relu(float *dout, asm volatile(COMPUTE_S_S1_P0 RESULT_S_S1 : [din0] "+r"(dr0), [din1] "+r"(dr1), - [din2] "+r"(dr2), - [din3] "+r"(dr3) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [vbias] "w"(wbias), - [mask1] "w"(vmask_rp1), - [mask2] "w"(vmask_rp2), - [zero] "w"(vzero), - [out1] "r"(out_buf1), - [out2] "r"(out_buf2) - : "cc", - "memory", - "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15"); + [din2] "+r"(dr2), + [din3] "+r"(dr3) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [vbias] "w"(wbias), + [mask1] "w"(vmask_rp1), + [mask2] "w"(vmask_rp2), + [zero] "w"(vzero), + [out1] "r"(out_buf1), + [out2] "r"(out_buf2) + : "cc", + "memory", + "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15"); #else unsigned int *vmask_ptr = vmask; float bias_val = flag_bias ? bias[i] : 0.f; @@ -2905,35 +2908,35 @@ void conv_depthwise_3x3s1p0_bias_s_relu(float *dout, asm volatile(COMPUTE_S_S1_P0 RESULT_S_S1_RELU : [din0] "+r"(dr0), [din1] "+r"(dr1), - [din2] "+r"(dr2), - [din3] "+r"(dr3) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [vbias] "w"(wbias), - [mask1] "w"(vmask_rp1), - [mask2] "w"(vmask_rp2), - [zero] "w"(vzero), - [out1] "r"(out_buf1), - [out2] "r"(out_buf2) - : "cc", - "memory", - "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15"); + [din2] "+r"(dr2), + [din3] "+r"(dr3) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [vbias] "w"(wbias), + [mask1] "w"(vmask_rp1), + [mask2] "w"(vmask_rp2), + [zero] "w"(vzero), + [out1] "r"(out_buf1), + [out2] "r"(out_buf2) + : "cc", + "memory", + "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15"); #else unsigned int *vmask_ptr = vmask; float bias_val = flag_bias ? bias[i] : 0.f; diff --git a/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32.cc b/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32.cc index 4d63d5b8c9f7b2f117bf3feaf35ecf92490cf227..4936576de130c0c80b79991814f0aebb84d50fda 100644 --- a/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32.cc +++ b/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32.cc @@ -1747,53 +1747,52 @@ void conv_depthwise_3x3s2p1_bias_leakyRelu(float* dout, doutr1_ptr = write_ptr; } int cnt = cnt_col; - asm volatile( - INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2_LEAKY_RELU - MID_COMPUTE_S2 MID_RESULT_S2_LEAKY_RELU - RIGHT_COMPUTE_S2 RIGHT_RESULT_S2_LEAKY_RELU - : [inptr0] "+r"(din0_ptr), - [inptr1] "+r"(din1_ptr), - [inptr2] "+r"(din2_ptr), - [inptr3] "+r"(din3_ptr), - [inptr4] "+r"(din4_ptr), - [outptr0] "+r"(doutr0_ptr), - [outptr1] "+r"(doutr1_ptr), - [cnt] "+r"(cnt) - : [vzero] "w"(vzero), - [w0] "w"(wr0), - [w1] "w"(wr1), - [w2] "w"(wr2), - [remain] "r"(cnt_remain), - [scale_ptr] "r"(scale), - [mask1] "w"(vmask_rp1), - [mask2] "w"(vmask_rp2), - [wmask] "w"(wmask), - [vbias] "w"(wbias) - : "cc", - "memory", - "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16", - "v17", - "v18", - "v19", - "v20", - "v21", - "v22"); + asm volatile(INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2_LEAKY_RELU + MID_COMPUTE_S2 MID_RESULT_S2_LEAKY_RELU + RIGHT_COMPUTE_S2 RIGHT_RESULT_S2_LEAKY_RELU + : [inptr0] "+r"(din0_ptr), + [inptr1] "+r"(din1_ptr), + [inptr2] "+r"(din2_ptr), + [inptr3] "+r"(din3_ptr), + [inptr4] "+r"(din4_ptr), + [outptr0] "+r"(doutr0_ptr), + [outptr1] "+r"(doutr1_ptr), + [cnt] "+r"(cnt) + : [vzero] "w"(vzero), + [w0] "w"(wr0), + [w1] "w"(wr1), + [w2] "w"(wr2), + [remain] "r"(cnt_remain), + [scale_ptr] "r"(scale), + [mask1] "w"(vmask_rp1), + [mask2] "w"(vmask_rp2), + [wmask] "w"(wmask), + [vbias] "w"(wbias) + : "cc", + "memory", + "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16", + "v17", + "v18", + "v19", + "v20", + "v21", + "v22"); doutr0 = doutr0 + 2 * w_out; } #else @@ -1830,36 +1829,36 @@ void conv_depthwise_3x3s2p1_bias_leakyRelu(float* dout, } int cnt = cnt_col; unsigned int* mask_ptr = dmask; - asm volatile( - INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2_LEAKY_RELU MID_COMPUTE_S2 - MID_RESULT_S2_LEAKY_RELU RIGHT_COMPUTE_S2 RIGHT_RESULT_S2_LEAKY_RELU - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [outptr] "+r"(doutr0_ptr), - [cnt] "+r"(cnt), - [mask_ptr] "+r"(mask_ptr) - : [remain] "r"(cnt_remain), - [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [scale_ptr] "r"(scale), - [bias] "r"(bias_c) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + asm volatile(INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2_LEAKY_RELU + MID_COMPUTE_S2 MID_RESULT_S2_LEAKY_RELU + RIGHT_COMPUTE_S2 RIGHT_RESULT_S2_LEAKY_RELU + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [outptr] "+r"(doutr0_ptr), + [cnt] "+r"(cnt), + [mask_ptr] "+r"(mask_ptr) + : [remain] "r"(cnt_remain), + [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [scale_ptr] "r"(scale), + [bias] "r"(bias_c) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); doutr0 = doutr0 + w_out; } #endif @@ -2349,8 +2348,8 @@ void conv_depthwise_3x3s2p0_bias_relu6(float* dout, } int cnt = tile_w; unsigned int* mask_ptr = dmask; - asm volatile(INIT_S2 MID_COMPUTE_S2 MID_RESULT_S2_RELU6 - RIGHT_COMPUTE_S2 RIGHT_RESULT_S2_RELU6 + asm volatile(INIT_S2 MID_COMPUTE_S2 MID_RESULT_S2_RELU6 RIGHT_COMPUTE_S2 + RIGHT_RESULT_S2_RELU6 : [din0_ptr] "+r"(din0_ptr), [din1_ptr] "+r"(din1_ptr), [din2_ptr] "+r"(din2_ptr), diff --git a/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32_relu.cc b/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32_relu.cc index d06f73ef65ae62054b3568342cb7fda4a9b60547..3fe51cd93d7a9ae50905828949b2f640bcceac3a 100644 --- a/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32_relu.cc +++ b/lite/backends/arm/math/conv3x3s2p01_depthwise_fp32_relu.cc @@ -917,10 +917,10 @@ void conv_depthwise_3x3s2p1_bias_relu(float* dout, [cnt] "+r"(cnt), [mask_ptr] "+r"(mask_ptr) : [remain] "r"(cnt_remain), - [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c) + [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c) : "cc", "memory", "q3", @@ -931,7 +931,7 @@ void conv_depthwise_3x3s2p1_bias_relu(float* dout, "q8", "q9", "q10", - "q11", + "q11", "q12", "q13", "q14", @@ -1080,50 +1080,49 @@ void conv_depthwise_3x3s2p1_bias_no_relu(float* dout, doutr1_ptr = write_ptr; } int cnt = cnt_col; - asm volatile( - INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2 MID_COMPUTE_S2 - MID_RESULT_S2 RIGHT_COMPUTE_S2 RIGHT_RESULT_S2 - : [inptr0] "+r"(din0_ptr), - [inptr1] "+r"(din1_ptr), - [inptr2] "+r"(din2_ptr), - [inptr3] "+r"(din3_ptr), - [inptr4] "+r"(din4_ptr), - [outptr0] "+r"(doutr0_ptr), - [outptr1] "+r"(doutr1_ptr), - [cnt] "+r"(cnt) - : [vzero] "w"(vzero), - [w0] "w"(wr0), - [w1] "w"(wr1), - [w2] "w"(wr2), - [remain] "r"(cnt_remain), - [mask1] "w"(vmask_rp1), - [mask2] "w"(vmask_rp2), - [wmask] "w"(wmask), - [vbias] "w"(wbias) - : "cc", - "memory", - "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16", - "v17", - "v18", - "v19", - "v20", - "v21"); + asm volatile(INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2 MID_COMPUTE_S2 + MID_RESULT_S2 RIGHT_COMPUTE_S2 RIGHT_RESULT_S2 + : [inptr0] "+r"(din0_ptr), + [inptr1] "+r"(din1_ptr), + [inptr2] "+r"(din2_ptr), + [inptr3] "+r"(din3_ptr), + [inptr4] "+r"(din4_ptr), + [outptr0] "+r"(doutr0_ptr), + [outptr1] "+r"(doutr1_ptr), + [cnt] "+r"(cnt) + : [vzero] "w"(vzero), + [w0] "w"(wr0), + [w1] "w"(wr1), + [w2] "w"(wr2), + [remain] "r"(cnt_remain), + [mask1] "w"(vmask_rp1), + [mask2] "w"(vmask_rp2), + [wmask] "w"(wmask), + [vbias] "w"(wbias) + : "cc", + "memory", + "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16", + "v17", + "v18", + "v19", + "v20", + "v21"); doutr0 = doutr0 + 2 * w_out; } #else @@ -1160,35 +1159,34 @@ void conv_depthwise_3x3s2p1_bias_no_relu(float* dout, } int cnt = cnt_col; unsigned int* mask_ptr = dmask; - asm volatile( - INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2 MID_COMPUTE_S2 - MID_RESULT_S2 RIGHT_COMPUTE_S2 RIGHT_RESULT_S2 - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [outptr] "+r"(doutr0_ptr), - [cnt] "+r"(cnt), - [mask_ptr] "+r"(mask_ptr) - : [remain] "r"(cnt_remain), - [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + asm volatile(INIT_S2 LEFT_COMPUTE_S2 LEFT_RESULT_S2 MID_COMPUTE_S2 + MID_RESULT_S2 RIGHT_COMPUTE_S2 RIGHT_RESULT_S2 + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [outptr] "+r"(doutr0_ptr), + [cnt] "+r"(cnt), + [mask_ptr] "+r"(mask_ptr) + : [remain] "r"(cnt_remain), + [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); doutr0 = doutr0 + w_out; } #endif @@ -1266,54 +1264,54 @@ void conv_depthwise_3x3s2p1_bias_s_relu(float* dout, unsigned int* mask_ptr = dmask; #ifdef __aarch64__ - asm volatile(COMPUTE_S_S2 RESULT_S_S2_RELU - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [mask_ptr] "+r"(mask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "w"(vbias), - [out] "r"(out_buf) - : "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15"); + asm volatile(COMPUTE_S_S2 RESULT_S_S2_RELU + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [mask_ptr] "+r"(mask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "w"(vbias), + [out] "r"(out_buf) + : "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15"); #else - asm volatile(COMPUTE_S_S2 RESULT_S_S2_RELU - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [mask_ptr] "+r"(mask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c), - [out] "r"(out_buf) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + asm volatile(COMPUTE_S_S2 RESULT_S_S2_RELU + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [mask_ptr] "+r"(mask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c), + [out] "r"(out_buf) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); #endif for (int w = 0; w < w_out; ++w) { *dout_channel++ = out_buf[w]; @@ -1396,49 +1394,49 @@ void conv_depthwise_3x3s2p1_bias_s_no_relu(float* dout, [din1_ptr] "+r"(din1_ptr), [din2_ptr] "+r"(din2_ptr), [mask_ptr] "+r"(mask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "w"(vbias), - [out] "r"(out_buf) - : "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15"); + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "w"(vbias), + [out] "r"(out_buf) + : "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15"); #else asm volatile(COMPUTE_S_S2 RESULT_S_S2 : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [mask_ptr] "+r"(mask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c), - [out] "r"(out_buf) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [mask_ptr] "+r"(mask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c), + [out] "r"(out_buf) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); #endif for (int w = 0; w < w_out; ++w) { *dout_channel++ = out_buf[w]; @@ -1584,60 +1582,60 @@ void conv_depthwise_3x3s2p0_bias_relu(float* dout, doutr1_ptr = write_ptr; } int cnt = tile_w; - asm volatile( - INIT_S2 - "ld1 {v15.4s}, [%[inptr0]] \n" - "ld1 {v18.4s}, [%[inptr1]] \n" - "ld1 {v19.4s}, [%[inptr2]] \n" - "ld1 {v20.4s}, [%[inptr3]] \n" - "ld1 {v21.4s}, [%[inptr4]] \n" - "ext v10.16b, v0.16b, v15.16b, #4 \n" // v10 = {2,4,6,8} - MID_COMPUTE_S2 MID_RESULT_S2_RELU - "cmp %w[remain], #1 \n" - "blt 4f \n" RIGHT_COMPUTE_S2 - RIGHT_RESULT_S2_RELU - "4: \n" - : [inptr0] "+r"(din0_ptr), - [inptr1] "+r"(din1_ptr), - [inptr2] "+r"(din2_ptr), - [inptr3] "+r"(din3_ptr), - [inptr4] "+r"(din4_ptr), - [outptr0] "+r"(doutr0_ptr), - [outptr1] "+r"(doutr1_ptr), - [cnt] "+r"(cnt) - : [vzero] "w"(vzero), - [w0] "w"(wr0), - [w1] "w"(wr1), - [w2] "w"(wr2), - [remain] "r"(cnt_remain), - [mask1] "w"(vmask_rp1), - [mask2] "w"(vmask_rp2), - [wmask] "w"(wmask), - [vbias] "w"(wbias) - : "cc", - "memory", - "v0", - "v1", - "v2", - "v3", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16", - "v17", - "v18", - "v19", - "v20", - "v21"); + asm volatile( + INIT_S2 + "ld1 {v15.4s}, [%[inptr0]] \n" + "ld1 {v18.4s}, [%[inptr1]] \n" + "ld1 {v19.4s}, [%[inptr2]] \n" + "ld1 {v20.4s}, [%[inptr3]] \n" + "ld1 {v21.4s}, [%[inptr4]] \n" + "ext v10.16b, v0.16b, v15.16b, #4 \n" // v10 = {2,4,6,8} + MID_COMPUTE_S2 MID_RESULT_S2_RELU + "cmp %w[remain], #1 \n" + "blt 4f \n" RIGHT_COMPUTE_S2 + RIGHT_RESULT_S2_RELU + "4: \n" + : [inptr0] "+r"(din0_ptr), + [inptr1] "+r"(din1_ptr), + [inptr2] "+r"(din2_ptr), + [inptr3] "+r"(din3_ptr), + [inptr4] "+r"(din4_ptr), + [outptr0] "+r"(doutr0_ptr), + [outptr1] "+r"(doutr1_ptr), + [cnt] "+r"(cnt) + : [vzero] "w"(vzero), + [w0] "w"(wr0), + [w1] "w"(wr1), + [w2] "w"(wr2), + [remain] "r"(cnt_remain), + [mask1] "w"(vmask_rp1), + [mask2] "w"(vmask_rp2), + [wmask] "w"(wmask), + [vbias] "w"(wbias) + : "cc", + "memory", + "v0", + "v1", + "v2", + "v3", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16", + "v17", + "v18", + "v19", + "v20", + "v21"); doutr0 = doutr0 + 2 * w_out; } #else @@ -1665,34 +1663,34 @@ void conv_depthwise_3x3s2p0_bias_relu(float* dout, } int cnt = tile_w; unsigned int* mask_ptr = dmask; - asm volatile(INIT_S2 MID_COMPUTE_S2 MID_RESULT_S2_RELU - RIGHT_COMPUTE_S2 RIGHT_RESULT_S2_RELU - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [outptr] "+r"(doutr0_ptr), - [cnt] "+r"(cnt), - [mask_ptr] "+r"(mask_ptr) - : [remain] "r"(cnt_remain), - [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + asm volatile(INIT_S2 MID_COMPUTE_S2 MID_RESULT_S2_RELU + RIGHT_COMPUTE_S2 RIGHT_RESULT_S2_RELU + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [outptr] "+r"(doutr0_ptr), + [cnt] "+r"(cnt), + [mask_ptr] "+r"(mask_ptr) + : [remain] "r"(cnt_remain), + [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); doutr0 = doutr0 + w_out; } #endif @@ -1910,34 +1908,34 @@ void conv_depthwise_3x3s2p0_bias_no_relu(float* dout, } int cnt = tile_w; unsigned int* mask_ptr = dmask; - asm volatile(INIT_S2 MID_COMPUTE_S2 MID_RESULT_S2 RIGHT_COMPUTE_S2 - RIGHT_RESULT_S2 - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [outptr] "+r"(doutr0_ptr), - [cnt] "+r"(cnt), - [mask_ptr] "+r"(mask_ptr) - : [remain] "r"(cnt_remain), - [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + asm volatile(INIT_S2 MID_COMPUTE_S2 MID_RESULT_S2 + RIGHT_COMPUTE_S2 RIGHT_RESULT_S2 + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [outptr] "+r"(doutr0_ptr), + [cnt] "+r"(cnt), + [mask_ptr] "+r"(mask_ptr) + : [remain] "r"(cnt_remain), + [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); doutr0 = doutr0 + w_out; } #endif @@ -2020,57 +2018,57 @@ void conv_depthwise_3x3s2p0_bias_s_relu(float* dout, unsigned int* mask_ptr = dmask; #ifdef __aarch64__ - asm volatile(COMPUTE_S_S2_P0 RESULT_S_S2_P0_RELU - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [mask_ptr] "+r"(mask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "w"(vbias), - [out] "r"(out_buf) - : "cc", - "memory", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16"); + asm volatile(COMPUTE_S_S2_P0 RESULT_S_S2_P0_RELU + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [mask_ptr] "+r"(mask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "w"(vbias), + [out] "r"(out_buf) + : "cc", + "memory", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16"); #else - asm volatile(COMPUTE_S_S2_P0 RESULT_S_S2_P0_RELU - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c), - [out] "r"(out_buf), - [mask_ptr] "r"(dmask) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + asm volatile(COMPUTE_S_S2_P0 RESULT_S_S2_P0_RELU + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c), + [out] "r"(out_buf), + [mask_ptr] "r"(dmask) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); #endif for (int w = 0; w < w_out; ++w) { *dout_channel++ = out_buf[w]; @@ -2152,57 +2150,57 @@ void conv_depthwise_3x3s2p0_bias_s_no_relu(float* dout, unsigned int* mask_ptr = dmask; #ifdef __aarch64__ - asm volatile(COMPUTE_S_S2_P0 RESULT_S_S2_P0 - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr), - [mask_ptr] "+r"(mask_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "w"(vbias), - [out] "r"(out_buf) - : "cc", - "memory", - "v4", - "v5", - "v6", - "v7", - "v8", - "v9", - "v10", - "v11", - "v12", - "v13", - "v14", - "v15", - "v16"); + asm volatile(COMPUTE_S_S2_P0 RESULT_S_S2_P0 + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr), + [mask_ptr] "+r"(mask_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "w"(vbias), + [out] "r"(out_buf) + : "cc", + "memory", + "v4", + "v5", + "v6", + "v7", + "v8", + "v9", + "v10", + "v11", + "v12", + "v13", + "v14", + "v15", + "v16"); #else asm volatile(COMPUTE_S_S2_P0 RESULT_S_S2_P0 - : [din0_ptr] "+r"(din0_ptr), - [din1_ptr] "+r"(din1_ptr), - [din2_ptr] "+r"(din2_ptr) - : [wr0] "w"(wr0), - [wr1] "w"(wr1), - [wr2] "w"(wr2), - [bias] "r"(bias_c), - [out] "r"(out_buf), - [mask_ptr] "r"(dmask) - : "cc", - "memory", - "q3", - "q4", - "q5", - "q6", - "q7", - "q8", - "q9", - "q10", - "q11", - "q12", - "q13", - "q14", - "q15"); + : [din0_ptr] "+r"(din0_ptr), + [din1_ptr] "+r"(din1_ptr), + [din2_ptr] "+r"(din2_ptr) + : [wr0] "w"(wr0), + [wr1] "w"(wr1), + [wr2] "w"(wr2), + [bias] "r"(bias_c), + [out] "r"(out_buf), + [mask_ptr] "r"(dmask) + : "cc", + "memory", + "q3", + "q4", + "q5", + "q6", + "q7", + "q8", + "q9", + "q10", + "q11", + "q12", + "q13", + "q14", + "q15"); #endif for (int w = 0; w < w_out; ++w) { *dout_channel++ = out_buf[w];