// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. #include #include "lite/backends/arm/math/conv_block_utils.h" #include "lite/backends/arm/math/conv_depthwise.h" #include "lite/backends/arm/math/conv_impl.h" #include "lite/core/context.h" #include "lite/operators/op_params.h" #ifdef ARM_WITH_OMP #include #endif namespace paddle { namespace lite { namespace arm { namespace math { #define ROUNDUP(a, b) ((((a) + (b)-1) / (b)) * (b)) template void conv_depthwise_3x3s2_int8(Dtype* dout, const int8_t* din, const int8_t* weights, const float* scale, const float* bias, bool flag_bias, bool flag_relu, int num, int chin, int hin, int win, int hout, int wout, int padw, int padh, ARMContext* ctx) { const int threads = ctx->threads(); int llc_size = ctx->llc_size() / 4; const int hout_c_block = 8; const int hout_r_kernel = 1; const int wout_block = 4; const int wout_round = ROUNDUP(wout, wout_block); const int win_round = wout_round * 2 /*stride*/ + 1; //! get h block //! llc_size = threads * win_round * hin_r_block * sizeof(int8_t) + wout_round //! * hout_c_block * hout_r_block * threads * sizeof(int32_t) //! win_round = wout_round + 2 //! hin_r_block = hout_r_block + 2 int hout_r_block = (llc_size - 2 * win_round * threads) / (2 * win_round * threads + hout_c_block * wout_round * threads * 4); hout_r_block = hout_r_block > hout ? hout : hout_r_block; hout_r_block = ((hout_r_block + hout_r_kernel - 1) / hout_r_kernel) * hout_r_kernel; hout_r_block = hout_r_block < hout_r_kernel ? hout_r_kernel : hout_r_block; const int hin_r_block = hout_r_block * 2 /*stride*/ + 1; auto tmp_work_space = ctx->workspace_data(); int8_t ptr_zero[win_round]; // NOLINT memset(ptr_zero, 0, sizeof(int8_t) * win_round); Dtype ptr_write[wout_round]; // NOLINT int in_len = win_round * hout_c_block; int pre_in_size = hin_r_block * in_len; pre_in_size = ROUNDUP(pre_in_size, 4); int pre_out_size = hout_c_block * hout_r_block * wout_round; int8_t* tmp_din = tmp_work_space; int size_in_channel = win * hin; int size_out_channel = wout * hout; int w_stride = 9; // kernel_w * kernel_h; int ws = -padw; int we = ws + win_round; int w_loop = wout_round / 4; int chout = chin; int out_row_stride = hout_c_block * wout_round; for (int n = 0; n < num; ++n) { const int8_t* din_batch = din + n * chin * size_in_channel; int8_t* dout_batch = reinterpret_cast(dout) + n * chout * size_out_channel * sizeof(Dtype); for (int h = 0; h < hout; h += hout_r_block) { int h_kernel = hout_r_block; if (h + hout_r_block > hout) { h_kernel = hout - h; } int hs = h * 2 /*stride*/ - padh; int he = hs + h_kernel * 2 /*stride*/ + 1; #pragma omp parallel for num_threads(threads) for (int c = 0; c < chout; c += hout_c_block) { #ifdef ARM_WITH_OMP int8_t* pre_din = tmp_din + omp_get_thread_num() * (pre_in_size + pre_out_size * 4); int32_t* pre_out = reinterpret_cast(pre_din + pre_in_size); #else int32_t* pre_out = reinterpret_cast(tmp_din + pre_in_size); auto pre_din = tmp_din; #endif prepack_input_nxw_c8_int8(din_batch, pre_din, c, c + hout_c_block, hs, he, ws, we, chin, win, hin); const int8_t* block_inr0 = pre_din; const int8_t* block_inr1 = block_inr0 + in_len; const int8_t* block_inr2 = block_inr1 + in_len; const int8_t* weight_c = weights + c * w_stride; float bias_local[8] = {0, 0, 0, 0, 0, 0, 0, 0}; if (flag_bias) { bias_local[0] = bias[c]; bias_local[1] = bias[c + 1]; bias_local[2] = bias[c + 2]; bias_local[3] = bias[c + 3]; bias_local[4] = bias[c + 4]; bias_local[5] = bias[c + 5]; bias_local[6] = bias[c + 6]; bias_local[7] = bias[c + 7]; } #ifdef __aarch64__ int8x8_t vw0 = vld1_s8(weight_c); int8x8_t vw1 = vld1_s8(weight_c + 8); int8x8_t vw2 = vld1_s8(weight_c + 16); int8x8_t vw3 = vld1_s8(weight_c + 24); int8x8_t vw4 = vld1_s8(weight_c + 32); int8x8_t vw5 = vld1_s8(weight_c + 40); int8x8_t vw6 = vld1_s8(weight_c + 48); int8x8_t vw7 = vld1_s8(weight_c + 56); int8x8_t vw8 = vld1_s8(weight_c + 64); #endif for (int hk = 0; hk < h_kernel; hk += hout_r_kernel) { int cnt = w_loop; const int8_t* inr0 = block_inr0; const int8_t* inr1 = block_inr1; const int8_t* inr2 = block_inr2; int32_t* ptr_out0 = pre_out + hk * out_row_stride; #ifdef __aarch64__ asm volatile( "ld1 {v0.8b, v1.8b, v2.8b, v3.8b}, [%[r0]], #32\n" "ld1 {v4.8b, v5.8b, v6.8b, v7.8b}, [%[r0]], #32\n" "1:\n" /* inr0 -> outr0 */ "smull v20.8h, v0.8b, %[w0].8b\n" /* int16, out0 */ "smull v21.8h, v2.8b, %[w0].8b\n" /* int16, out1 */ "smull v22.8h, v4.8b, %[w0].8b\n" /* int16, out2 */ "smull v23.8h, v6.8b, %[w0].8b\n" /* int16, out3 */ "smlal v20.8h, v1.8b, %[w1].8b\n" /* int16, out0 */ "smlal v21.8h, v3.8b, %[w1].8b\n" /* int16, out1 */ "smlal v22.8h, v5.8b, %[w1].8b\n" /* int16, out2 */ "smlal v23.8h, v7.8b, %[w1].8b\n" /* int16, out3 */ "ldr d8, [%[r0]]\n" /* load r0, 8 */ "ldp d0, d1, [%[r1]], #16\n" /* load r1, 0,1 */ "sxtl v24.4s, v20.4h\n" "sxtl2 v25.4s, v20.8h\n" "smull v20.8h, v2.8b, %[w2].8b\n" /* int16, out0 */ "ldp d2, d3, [%[r1]], #16\n" /* load r1, 2,3 */ "sxtl v26.4s, v21.4h\n" "sxtl2 v27.4s, v21.8h\n" "smull v21.8h, v4.8b, %[w2].8b\n" /* int16, out1 */ "ldp d4, d5, [%[r1]], #16\n" /* load r1, 4,5 */ "sxtl v28.4s, v22.4h\n" "sxtl2 v29.4s, v22.8h\n" "smull v22.8h, v6.8b, %[w2].8b\n" /* int16, out2 */ "ldp d6, d7, [%[r1]], #16\n" /* load r1, 6,7 */ "sxtl v30.4s, v23.4h\n" "sxtl2 v31.4s, v23.8h\n" "smull v23.8h, v8.8b, %[w2].8b\n" /* int16, out3 */ "smlal v20.8h, v0.8b, %[w3].8b\n" /* int16, out0 */ "smlal v21.8h, v2.8b, %[w3].8b\n" /* int16, out1 */ "smlal v22.8h, v4.8b, %[w3].8b\n" /* int16, out2 */ "smlal v23.8h, v6.8b, %[w3].8b\n" /* int16, out3 */ "saddw v24.4s, v24.4s, v20.4h\n" "saddw2 v25.4s, v25.4s, v20.8h\n" "saddw v26.4s, v26.4s, v21.4h\n" "saddw2 v27.4s, v27.4s, v21.8h\n" "ldr d8, [%[r1]]\n" /* load r1, 8 */ "saddw v28.4s, v28.4s, v22.4h\n" "saddw2 v29.4s, v29.4s, v22.8h\n" "saddw v30.4s, v30.4s, v23.4h\n" "saddw2 v31.4s, v31.4s, v23.8h\n" "smull v20.8h, v1.8b, %[w4].8b\n" /* int16, out0 */ "smull v21.8h, v3.8b, %[w4].8b\n" /* int16, out1 */ "smull v22.8h, v5.8b, %[w4].8b\n" /* int16, out1 */ "smull v23.8h, v7.8b, %[w4].8b\n" /* int16, out1 */ "ldp d0, d1, [%[r2]], #16\n" /* load r2, 0,1 */ "smlal v20.8h, v2.8b, %[w5].8b\n" /* int16, out0 */ "smlal v21.8h, v4.8b, %[w5].8b\n" /* int16, out1 */ "ldp d2, d3, [%[r2]], #16\n" /* load r2, 2,3 */ "smlal v22.8h, v6.8b, %[w5].8b\n" /* int16, out2 */ "smlal v23.8h, v8.8b, %[w5].8b\n" /* int16, out3 */ "ldp d4, d5, [%[r2]], #16\n" /* load r2, 4,5 */ "saddw v24.4s, v24.4s, v20.4h\n" "saddw2 v25.4s, v25.4s, v20.8h\n" "saddw v26.4s, v26.4s, v21.4h\n" "saddw2 v27.4s, v27.4s, v21.8h\n" "ldp d6, d7, [%[r2]], #16\n" /* load r2, 6,7 */ "saddw v28.4s, v28.4s, v22.4h\n" "saddw2 v29.4s, v29.4s, v22.8h\n" "saddw v30.4s, v30.4s, v23.4h\n" "saddw2 v31.4s, v31.4s, v23.8h\n" "smull v20.8h, v0.8b, %[w6].8b\n" /* int16, out0 */ "smull v21.8h, v2.8b, %[w6].8b\n" /* int16, out1 */ "smull v22.8h, v4.8b, %[w6].8b\n" /* int16, out1 */ "smull v23.8h, v6.8b, %[w6].8b\n" /* int16, out1 */ "smlal v20.8h, v1.8b, %[w7].8b\n" /* int16, out0 */ "smlal v21.8h, v3.8b, %[w7].8b\n" /* int16, out1 */ "smlal v22.8h, v5.8b, %[w7].8b\n" /* int16, out1 */ "smlal v23.8h, v7.8b, %[w7].8b\n" /* int16, out1 */ "ldp d0, d1, [%[r0]], #16\n" /* load r0, 0,1 */ "saddw v24.4s, v24.4s, v20.4h\n" "saddw2 v25.4s, v25.4s, v20.8h\n" "saddw v26.4s, v26.4s, v21.4h\n" "saddw2 v27.4s, v27.4s, v21.8h\n" "ldr d8, [%[r2]]\n" /* load r2 */ "saddw v28.4s, v28.4s, v22.4h\n" "saddw2 v29.4s, v29.4s, v22.8h\n" "saddw v30.4s, v30.4s, v23.4h\n" "saddw2 v31.4s, v31.4s, v23.8h\n" "smull v20.8h, v2.8b, %[w8].8b\n" /* int16, out0 */ "smull v21.8h, v4.8b, %[w8].8b\n" /* int16, out1 */ "ldp d2, d3, [%[r0]], #16\n" /* load r0, 2,3 */ "smull v22.8h, v6.8b, %[w8].8b\n" /* int16, out1 */ "smull v23.8h, v8.8b, %[w8].8b\n" /* int16, out1 */ "ldp d4, d5, [%[r0]], #16\n" /* load r0, 5 */ "saddw v24.4s, v24.4s, v20.4h\n" "saddw2 v25.4s, v25.4s, v20.8h\n" "saddw v26.4s, v26.4s, v21.4h\n" "saddw2 v27.4s, v27.4s, v21.8h\n" "ldp d6, d7, [%[r0]], #16\n" /* load r0, 6 */ "stp q24, q25, [%[ptr_out0]], #32\n" "saddw v28.4s, v28.4s, v22.4h\n" "saddw2 v29.4s, v29.4s, v22.8h\n" "stp q26, q27, [%[ptr_out0]], #32\n" "saddw v30.4s, v30.4s, v23.4h\n" "saddw2 v31.4s, v31.4s, v23.8h\n" "subs %w[cnt], %w[cnt], #1\n" "stp q28, q29, [%[ptr_out0]], #32\n" "stp q30, q31, [%[ptr_out0]], #32\n" "bne 1b\n" : [cnt] "+r"(cnt), [r0] "+r"(inr0), [r1] "+r"(inr1), [r2] "+r"(inr2), [ptr_out0] "+r"(ptr_out0) : [w0] "w"(vw0), [w1] "w"(vw1), [w2] "w"(vw2), [w3] "w"(vw3), [w4] "w"(vw4), [w5] "w"(vw5), [w6] "w"(vw6), [w7] "w"(vw7), [w8] "w"(vw8) : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" ); #else auto wptr = weight_c; asm volatile( "vld1.32 {d0-d3}, [%[r0]]!\n" /* load r0, 0-3 */ "vld1.32 {d6-d7}, [%[wptr]]!\n" /* load w0-w1 */ "vld1.32 {d4-d5}, [%[r0]]!\n" /* load r0, 4-5 */ "1:\n" /* inr0 -> outr0 */ "vmull.s8 q4, d1, d7\n" /* int16, out0 */ "vld1.32 {d1}, [%[r0]]!\n" /* load r0, 6 */ "vmull.s8 q5, d3, d7\n" /* int16, out1 */ "vld1.32 {d3}, [%[r0]]!\n" /* load r0, 7 */ "vmull.s8 q6, d5, d7\n" /* int16, out2 */ "vld1.32 {d5}, [%[r0]]\n" /* load r0, 8 */ "vmull.s8 q7, d1, d6\n" /* int16, out0 */ "vmlal.s8 q4, d0, d6\n" /* int16, out3 */ "vmlal.s8 q5, d2, d6\n" /* int16, out1 */ "vmlal.s8 q6, d4, d6\n" /* int16, out2 */ "vmlal.s8 q7, d3, d7\n" /* int16, out3 */ "vmovl.s16 q8, d8\n" "vmovl.s16 q9, d9\n" "vld1.32 {d6-d7}, [%[wptr]]!\n" /* load w2-w3 */ "vmovl.s16 q10, d10\n" "vmovl.s16 q11, d11\n" "vmovl.s16 q12, d12\n" "vmovl.s16 q13, d13\n" "vmovl.s16 q14, d14\n" "vmovl.s16 q15, d15\n" "vmull.s8 q4, d2, d6\n" /* int16, out0 */ "vmull.s8 q6, d1, d6\n" /* int16, out2 */ "vld1.32 {d0-d3}, [%[r1]]!\n" /* load r1, 0-3 */ "vmull.s8 q5, d4, d6\n" /* int16, out1 */ "vmull.s8 q7, d5, d6\n" /* int16, out3 */ "vld1.32 {d4-d5}, [%[r1]]!\n" /* load r1, 4,5 */ /* inr1 -> outr0 */ "vmlal.s8 q4, d0, d7\n" /* int16, out0 */ "vld1.32 {d0}, [%[r1]]!\n" /* load r1, 6 */ "vmlal.s8 q5, d2, d7\n" /* int16, out1 */ "vmlal.s8 q6, d4, d7\n" /* int16, out2 */ "vaddw.s16 q8, q8, d8\n" "vaddw.s16 q9, q9, d9\n" "vmlal.s8 q7, d0, d7\n" /* int16, out3 */ "vaddw.s16 q10, q10, d10\n" "vaddw.s16 q11, q11, d11\n" "vld1.32 {d6-d7}, [%[wptr]]!\n" /* load w4-w5 */ "vaddw.s16 q12, q12, d12\n" "vaddw.s16 q13, q13, d13\n" "vaddw.s16 q14, q14, d14\n" "vaddw.s16 q15, q15, d15\n" "vmull.s8 q4, d1, d6\n" /* int16, out0 */ "vld1.32 {d1}, [%[r1]]!\n" /* load r1, 7 */ "vmull.s8 q5, d3, d6\n" /* int16, out1 */ "vld1.32 {d3}, [%[r1]]\n" /* load r1, 8 */ "vmull.s8 q6, d5, d6\n" /* int16, out2 */ "vmull.s8 q7, d1, d6\n" /* int16, out3 */ "vmlal.s8 q4, d2, d7\n" /* int16, out0 */ "vmlal.s8 q5, d4, d7\n" /* int16, out2 */ "vmlal.s8 q6, d0, d7\n" /* int16, out1 */ "vmlal.s8 q7, d3, d7\n" /* int16, out3 */ "vld1.32 {d0-d3}, [%[r2]]!\n" /* load r2, 0-3 */ "vaddw.s16 q8, q8, d8\n" "vaddw.s16 q9, q9, d9\n" "vaddw.s16 q10, q10, d10\n" "vaddw.s16 q11, q11, d11\n" "vld1.32 {d6-d7}, [%[wptr]]!\n" /* load w6-w7 */ "vaddw.s16 q12, q12, d12\n" "vaddw.s16 q13, q13, d13\n" "vaddw.s16 q14, q14, d14\n" "vaddw.s16 q15, q15, d15\n" "vld1.32 {d4-d5}, [%[r2]]!\n" /* load r2, 4-5 */ /* inr2 -> outr0 */ "vmull.s8 q4, d1, d7\n" /* int16, out0 */ "vld1.32 {d1}, [%[r2]]!\n" /* load r2, 6 */ "vmull.s8 q5, d3, d7\n" /* int16, out1 */ "vld1.32 {d3}, [%[r2]]!\n" /* load r2, 7 */ "vmull.s8 q6, d5, d7\n" /* int16, out2 */ "vld1.32 {d5}, [%[r2]]\n" /* load r2, 8 */ "vmull.s8 q7, d1, d6\n" /* int16, out3 */ "vmlal.s8 q4, d0, d6\n" /* int16, out0 */ "vmlal.s8 q5, d2, d6\n" /* int16, out1 */ "vmlal.s8 q6, d4, d6\n" /* int16, out2 */ "vmlal.s8 q7, d3, d7\n" /* int16, out3 */ "vld1.32 {d6}, [%[wptr]]!\n" /* load w8 */ "vaddw.s16 q8, q8, d8\n" "vaddw.s16 q9, q9, d9\n" "vaddw.s16 q10, q10, d10\n" "vaddw.s16 q11, q11, d11\n" "vaddw.s16 q12, q12, d12\n" "vaddw.s16 q13, q13, d13\n" "vaddw.s16 q14, q14, d14\n" "vaddw.s16 q15, q15, d15\n" "sub %[wptr], %[wptr], #72\n" "vmull.s8 q4, d2, d6\n" /* int16, out0 */ "vmull.s8 q5, d4, d6\n" /* int16, out1 */ "vmull.s8 q6, d1, d6\n" /* int16, out2 */ "vmull.s8 q7, d5, d6\n" /* int16, out3 */ "vld1.32 {d0-d3}, [%[r0]]!\n" /* load r0, 0-3 */ "vaddw.s16 q8, q8, d8\n" "vaddw.s16 q9, q9, d9\n" "vld1.32 {d4-d5}, [%[r0]]!\n" /* load r0, 4-5 */ "vaddw.s16 q10, q10, d10\n" "vaddw.s16 q11, q11, d11\n" "vst1.32 {d16-d19}, [%[ptr_out0]]!\n" "vld1.32 {d6-d7}, [%[wptr]]!\n" /* load w0-w1 */ "vaddw.s16 q12, q12, d12\n" "vaddw.s16 q13, q13, d13\n" "vst1.32 {d20-d23}, [%[ptr_out0]]!\n" "vaddw.s16 q14, q14, d14\n" "vaddw.s16 q15, q15, d15\n" "subs %[cnt], #1\n" "vst1.32 {d24-d27}, [%[ptr_out0]]!\n" "vst1.32 {d28-d31}, [%[ptr_out0]]!\n" "bne 1b\n" : [cnt] "+r"(cnt), [r0] "+r"(inr0), [r1] "+r"(inr1), [r2] "+r"(inr2), [ptr_out0] "+r"(ptr_out0), [wptr] "+r"(wptr) : : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"); #endif block_inr0 = block_inr2; block_inr1 = block_inr0 + in_len; block_inr2 = block_inr1 + in_len; } write_int32_nchwc8_to_nchw(pre_out, reinterpret_cast(dout_batch), c, c + hout_c_block, h, h + h_kernel, 0, wout_round, chout, hout, wout, flag_relu, bias_local, flag_bias, ptr_write, scale + c); } } } } template void conv_depthwise_3x3s2_int8(int8_t* dout, const int8_t* din, const int8_t* weights, const float* scale, const float* bias, bool flag_bias, bool flag_relu, int num, int chin, int hin, int win, int hout, int wout, int padw, int padh, ARMContext* ctx); template void conv_depthwise_3x3s2_int8(float* dout, const int8_t* din, const int8_t* weights, const float* scale, const float* bias, bool flag_bias, bool flag_relu, int num, int chin, int hin, int win, int hout, int wout, int padw, int padh, ARMContext* ctx); } // namespace math } // namespace arm } // namespace lite } // namespace paddle