自動(dòng)調(diào)度GPU的卷積層
這是有關(guān)如何對(duì)GPU使用自動(dòng)調(diào)度程序的文檔。
與依靠手動(dòng)模板定義搜索空間的基于模板的autotvm不同，自動(dòng)調(diào)度程序不需要任何模板。用戶只需要編寫計(jì)算聲明，而無需任何調(diào)度命令或模板。自動(dòng)調(diào)度程序可以自動(dòng)生成較大的搜索空間，并在該空間中找到良好的調(diào)度。
本文以卷積層為例。
注意，本文無法在Windows或最新版本的macOS上運(yùn)行。要使其運(yùn)行，需要將本文的內(nèi)容包裝在一個(gè)if name == “main”:塊中。
import os

import numpy as np
import tvm
from tvm import te, auto_scheduler, topi
from tvm.topi.testing import conv2d_nchw_python
定義計(jì)算
首先，定義卷積層的計(jì)算。該函數(shù)應(yīng)返回輸入/輸出張量的列表。通過這些張量，自動(dòng)調(diào)度器可以獲取整個(gè)計(jì)算圖。
@auto_scheduler.register_workload
def conv2d_layer(N, H, W, CO, CI, KH, KW, stride, padding):
data = te.placeholder((N, CI, H, W), name=“data”)
kernel = te.placeholder((CO, CI, KH, KW), name=“kernel”)
bias = te.placeholder((1, CO, 1, 1), name=“bias”)
conv = topi.nn.conv2d_nchw(data, kernel, stride, padding, dilation=1, out_dtype=“float32”)
out = topi.nn.relu(conv + bias)
return [data, kernel, bias, out]
創(chuàng)建搜索任務(wù)
然后，為resnet中的最后一個(gè)卷積層創(chuàng)建搜索任務(wù)。
target = tvm.target.Target(“cuda”)

Use the last layer in ResNet-50

N, H, W, CO, CI, KH, KW, strides, padding = 1, 7, 7, 512, 512, 3, 3, (1, 1), (1, 1)
task = auto_scheduler.SearchTask(
func=conv2d_layer, args=(N, H, W, CO, CI, KH, KW, strides, padding), target=target
)

Inspect the computational graph

print(“Computational DAG:”)
print(task.compute_dag)
輸出：
Computational DAG:
data = PLACEHOLDER [1, 512, 7, 7]
pad_temp(i0, i1, i2, i3) = tir.if_then_else(((((i2 >= 1) && (i2 < 8)) && (i3 >= 1)) && (i3 < 8)), data[i0, i1, (i2 - 1), (i3 - 1)], 0f)
kernel = PLACEHOLDER [512, 512, 3, 3]
compute(nn, ff, yy, xx) += (pad_temp[nn, rc, (yy + ry), (xx + rx)]*kernel[ff, rc, ry, rx])
bias = PLACEHOLDER [1, 512, 1, 1]
T_add(ax0, ax1, ax2, ax3) = (compute[ax0, ax1, ax2, ax3] + bias[ax0, ax1, 0, 0])
compute(i0, i1, i2, i3) = max(T_add[i0, i1, i2, i3], 0f)
接下來，為自動(dòng)調(diào)度程序設(shè)置參數(shù)。這些參數(shù)主要指定在搜索過程中如何進(jìn)行測(cè)量。
? measure_ctx啟動(dòng)不同的測(cè)量過程以提供隔離。它可以保護(hù)主進(jìn)程免受測(cè)量期間GPU崩潰的影響，并避免其他運(yùn)行時(shí)runtime沖突。
? min_repeat_ms定義每次測(cè)量中一次“重復(fù)”的最小持續(xù)時(shí)間。這樣可以預(yù)熱GPU，對(duì)于獲得準(zhǔn)確的測(cè)量結(jié)果是必不可少的。通常，建議閾值> = 300 ms。
? num_measure_trials是在搜索過程中可以使用的測(cè)量試驗(yàn)的數(shù)量。為了快速演示，在本文中僅進(jìn)行10次試用。實(shí)際上，1000是搜索收斂的一個(gè)很好的值。可以根據(jù)自己的時(shí)間預(yù)算進(jìn)行更多試驗(yàn)。
? 此外，還用于RecordToFile將測(cè)量記錄轉(zhuǎn)儲(chǔ)到文件conv2d.json中。測(cè)量記錄可用于最好地查詢歷史記錄，恢復(fù)搜索以及以后進(jìn)行更多分析。
? 有關(guān)更多參數(shù)， 請(qǐng)參見auto_scheduler.TuningOptions，auto_scheduler.LocalRPCMeasureContext。
log_file = “conv2d.json”
measure_ctx = auto_scheduler.LocalRPCMeasureContext(min_repeat_ms=300)
tune_option = auto_scheduler.TuningOptions(
num_measure_trials=10, # change this to 1000 to achieve the best performance
runner=measure_ctx.runner,
measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
verbose=2,
)
輸出：
Get devices for measurement successfully!
運(yùn)行搜索
現(xiàn)在準(zhǔn)備好所有輸入。開始搜索，讓自動(dòng)調(diào)度程序發(fā)揮作用。經(jīng)過一些測(cè)量試驗(yàn)后，可以從日志文件中加載最佳調(diào)度并應(yīng)用它。

Run auto-tuning (search)

task.tune(tune_option)

Apply the best schedule

sch, args = task.apply_best(log_file)

Kill the measurement process

del measure_ctx
輸出：
可以降低調(diào)度，以便在自動(dòng)調(diào)度后查看IR。自動(dòng)調(diào)度程序可以正確執(zhí)行優(yōu)化，包括多層平鋪，協(xié)作提取，展開和算子融合。
print(“Lowered TIR:”)
print(tvm.lower(sch, args, simple_mode=True))
輸出：
Lowered TIR:
primfn(data_1: handle, kernel_1: handle, bias_1: handle, compute_1: handle) -> ()
attr = {“global_symbol”: “main”, “tir.noalias”: True}
buffers = {compute: Buffer(compute_2: Pointer(float32), float32, [1, 512, 7, 7], []),
bias: Buffer(bias_2: Pointer(float32), float32, [1, 512, 1, 1], []),
kernel: Buffer(kernel_2: Pointer(float32), float32, [512, 512, 3, 3], []),
data: Buffer(data_2: Pointer(float32), float32, [1, 512, 7, 7], [])}
buffer_map = {data_1: data, kernel_1: kernel, bias_1: bias, compute_1: compute} {
attr [IterVar(blockIdx.x: int32, (nullptr), “ThreadIndex”, “blockIdx.x”)] “thread_extent” = 28;
attr [compute_3: Pointer(float32)] “storage_scope” = “l(fā)ocal”;
allocate(compute_3, float32, [14]);
attr [pad_temp.shared: Pointer(float32)] “storage_scope” = “shared”;
allocate(pad_temp.shared, float32, [72]);
attr [kernel.shared: Pointer(float32)] “storage_scope” = “shared”;
allocate(kernel.shared, float32, [3072]);
attr [IterVar(threadIdx.x: int32, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64 {
compute_3[0] = 0f32
compute_3[1] = 0f32
compute_3[2] = 0f32
compute_3[3] = 0f32
compute_3[4] = 0f32
compute_3[5] = 0f32
compute_3[6] = 0f32
compute_3[7] = 0f32
compute_3[8] = 0f32
compute_3[9] = 0f32
compute_3[10] = 0f32
compute_3[11] = 0f32
compute_3[12] = 0f32
compute_3[13] = 0f32
for (rc.outer.outer: int32, 0, 64) {
for (ry.outer.outer: int32, 0, 3) {
attr [IterVar(threadIdx.x_1: int32, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64 {
if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
pad_temp.shared[(threadIdx.x_14)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod((threadIdx.x_14), 9))) && (floormod((threadIdx.x_14), 9) < 8)), (float32)data_2[((((((rc.outer.outer392) + (floordiv((threadIdx.x_14), 9)49)) + (ry.outer.outer7)) + (floormod(blockIdx.x, 7)7)) + floormod((threadIdx.x_14), 9)) - 8)], 0f32, dtype=float32)
}
if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
pad_temp.shared[((threadIdx.x_14) + 1)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_14) + 1), 9))) && (floormod(((threadIdx.x_14) + 1), 9) < 8)), (float32)data_2[((((((rc.outer.outer392) + (floordiv(((threadIdx.x_14) + 1), 9)49)) + (ry.outer.outer7)) + (floormod(blockIdx.x, 7)7)) + floormod(((threadIdx.x_14) + 1), 9)) - 8)], 0f32, dtype=float32)
}
if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
pad_temp.shared[((threadIdx.x_14) + 2)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_14) + 2), 9))) && (floormod(((threadIdx.x_14) + 2), 9) < 8)), (float32)data_2[((((((rc.outer.outer392) + (floordiv(((threadIdx.x_14) + 2), 9)49)) + (ry.outer.outer7)) + (floormod(blockIdx.x, 7)7)) + floormod(((threadIdx.x_14) + 2), 9)) - 8)], 0f32, dtype=float32)
}
if @tir.likely((threadIdx.x_1 < 18), dtype=bool) {
pad_temp.shared[((threadIdx.x_14) + 3)] = @tir.if_then_else(((((1 <= (ry.outer.outer + floormod(blockIdx.x, 7))) && ((ry.outer.outer + floormod(blockIdx.x, 7)) < 8)) && (1 <= floormod(((threadIdx.x_14) + 3), 9))) && (floormod(((threadIdx.x_14) + 3), 9) < 8)), (float32)data_2[((((((rc.outer.outer392) + (floordiv(((threadIdx.x_14) + 3), 9)49)) + (ry.outer.outer7)) + (floormod(blockIdx.x, 7)7)) + floormod(((threadIdx.x_14) + 3), 9)) - 8)], 0f32, dtype=float32)
}
}
attr [IterVar(threadIdx.x_2: int32, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[threadIdx.x_2] = (float32*)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 64)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 64), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 128)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 128), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 192)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 36864)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 256)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 256), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 320)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 320), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 384)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 73728)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 448)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 448), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 512)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 512), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 576)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 110592)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 640)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 640), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 704)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 704), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 768)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 147456)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 832)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 832), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 896)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 896), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 960)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 184320)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1024)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1024), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1088)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1088), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1152)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 221184)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1216)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1216), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1280)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1280), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1344)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 258048)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1408)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1408), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1472)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1472), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1536)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 294912)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1600)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1600), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1664)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1664), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1728)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 331776)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1792)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1792), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1856)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1856), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1920)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 368640)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 1984)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 1984), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2048)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2048), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2112)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 405504)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2176)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2176), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2240)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2240), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2304)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 442368)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2368)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2368), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2432)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2432), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2496)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 479232)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2560)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2560), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2624)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2624), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2688)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 516096)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2752)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2752), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2816)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2816), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2880)] = (float32)kernel_2[(((((((floordiv(blockIdx.x, 7)589824) + (floordiv(threadIdx.x_2, 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod(threadIdx.x_2, 24), 3)9)) + (ry.outer.outer3)) + floormod(threadIdx.x_2, 3)) + 552960)]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 2944)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 2944), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 16), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 1), 3))]
attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 64;
kernel.shared[(threadIdx.x_2 + 3008)] = (float32)kernel_2[((((((floordiv(blockIdx.x, 7)589824) + (floordiv((threadIdx.x_2 + 3008), 24)4608)) + (rc.outer.outer72)) + (floordiv(floormod((threadIdx.x_2 + 8), 24), 3)9)) + (ry.outer.outer3)) + floormod((threadIdx.x_2 + 2), 3))]
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[0](float32)kernel.shared[(threadIdx.x48)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[9](float32)kernel.shared[((threadIdx.x48) + 3)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[1](float32)kernel.shared[(threadIdx.x48)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[10](float32)kernel.shared[((threadIdx.x48) + 3)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[2](float32)kernel.shared[(threadIdx.x48)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[11](float32)kernel.shared[((threadIdx.x48) + 3)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[3](float32)kernel.shared[(threadIdx.x48)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[12](float32)kernel.shared[((threadIdx.x48) + 3)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[4](float32)kernel.shared[(threadIdx.x48)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[13](float32)kernel.shared[((threadIdx.x48) + 3)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[5](float32)kernel.shared[(threadIdx.x48)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[14](float32)kernel.shared[((threadIdx.x48) + 3)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[6](float32)kernel.shared[(threadIdx.x48)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[15](float32)kernel.shared[((threadIdx.x48) + 3)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[0](float32)kernel.shared[((threadIdx.x48) + 24)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[9](float32)kernel.shared[((threadIdx.x48) + 27)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[1](float32)kernel.shared[((threadIdx.x48) + 24)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[10](float32)kernel.shared[((threadIdx.x48) + 27)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[2](float32)kernel.shared[((threadIdx.x48) + 24)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[11](float32)kernel.shared[((threadIdx.x48) + 27)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[3](float32)kernel.shared[((threadIdx.x48) + 24)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[12](float32)kernel.shared[((threadIdx.x48) + 27)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[4](float32)kernel.shared[((threadIdx.x48) + 24)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[13](float32)kernel.shared[((threadIdx.x48) + 27)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[5](float32)kernel.shared[((threadIdx.x48) + 24)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[14](float32)kernel.shared[((threadIdx.x48) + 27)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[6](float32)kernel.shared[((threadIdx.x48) + 24)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[15](float32)kernel.shared[((threadIdx.x48) + 27)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[1](float32)kernel.shared[((threadIdx.x48) + 1)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[10](float32)kernel.shared[((threadIdx.x48) + 4)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[2](float32)kernel.shared[((threadIdx.x48) + 1)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[11](float32)kernel.shared[((threadIdx.x48) + 4)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[3](float32)kernel.shared[((threadIdx.x48) + 1)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[12](float32)kernel.shared[((threadIdx.x48) + 4)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[4](float32)kernel.shared[((threadIdx.x48) + 1)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[13](float32)kernel.shared[((threadIdx.x48) + 4)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[5](float32)kernel.shared[((threadIdx.x48) + 1)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[14](float32)kernel.shared[((threadIdx.x48) + 4)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[6](float32)kernel.shared[((threadIdx.x48) + 1)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[15](float32)kernel.shared[((threadIdx.x48) + 4)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[7](float32)kernel.shared[((threadIdx.x48) + 1)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[16](float32)kernel.shared[((threadIdx.x48) + 4)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[1](float32)kernel.shared[((threadIdx.x48) + 25)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[10](float32)kernel.shared[((threadIdx.x48) + 28)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[2](float32)kernel.shared[((threadIdx.x48) + 25)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[11](float32)kernel.shared[((threadIdx.x48) + 28)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[3](float32)kernel.shared[((threadIdx.x48) + 25)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[12](float32)kernel.shared[((threadIdx.x48) + 28)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[4](float32)kernel.shared[((threadIdx.x48) + 25)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[13](float32)kernel.shared[((threadIdx.x48) + 28)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[5](float32)kernel.shared[((threadIdx.x48) + 25)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[14](float32)kernel.shared[((threadIdx.x48) + 28)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[6](float32)kernel.shared[((threadIdx.x48) + 25)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[15](float32)kernel.shared[((threadIdx.x48) + 28)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[7](float32)kernel.shared[((threadIdx.x48) + 25)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[16](float32)kernel.shared[((threadIdx.x48) + 28)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[2](float32)kernel.shared[((threadIdx.x48) + 2)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[11](float32)kernel.shared[((threadIdx.x48) + 5)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[3](float32)kernel.shared[((threadIdx.x48) + 2)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[12](float32)kernel.shared[((threadIdx.x48) + 5)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[4](float32)kernel.shared[((threadIdx.x48) + 2)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[13](float32)kernel.shared[((threadIdx.x48) + 5)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[5](float32)kernel.shared[((threadIdx.x48) + 2)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[14](float32)kernel.shared[((threadIdx.x48) + 5)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[6](float32)kernel.shared[((threadIdx.x48) + 2)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[15](float32)kernel.shared[((threadIdx.x48) + 5)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[7](float32)kernel.shared[((threadIdx.x48) + 2)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[16](float32)kernel.shared[((threadIdx.x48) + 5)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[8](float32)kernel.shared[((threadIdx.x48) + 2)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[17](float32)kernel.shared[((threadIdx.x48) + 5)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[2](float32)kernel.shared[((threadIdx.x48) + 26)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[11](float32)kernel.shared[((threadIdx.x48) + 29)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[3](float32)kernel.shared[((threadIdx.x48) + 26)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[12](float32)kernel.shared[((threadIdx.x48) + 29)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[4](float32)kernel.shared[((threadIdx.x48) + 26)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[13](float32)kernel.shared[((threadIdx.x48) + 29)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[5](float32)kernel.shared[((threadIdx.x48) + 26)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[14](float32)kernel.shared[((threadIdx.x48) + 29)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[6](float32)kernel.shared[((threadIdx.x48) + 26)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[15](float32)kernel.shared[((threadIdx.x48) + 29)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[7](float32)kernel.shared[((threadIdx.x48) + 26)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[16](float32)kernel.shared[((threadIdx.x48) + 29)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[8](float32)kernel.shared[((threadIdx.x48) + 26)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[17](float32)kernel.shared[((threadIdx.x48) + 29)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[18](float32)kernel.shared[((threadIdx.x48) + 6)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[27](float32)kernel.shared[((threadIdx.x48) + 9)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[19](float32)kernel.shared[((threadIdx.x48) + 6)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[28](float32)kernel.shared[((threadIdx.x48) + 9)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[20](float32)kernel.shared[((threadIdx.x48) + 6)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[29](float32)kernel.shared[((threadIdx.x48) + 9)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[21](float32)kernel.shared[((threadIdx.x48) + 6)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[30](float32)kernel.shared[((threadIdx.x48) + 9)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[22](float32)kernel.shared[((threadIdx.x48) + 6)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[31](float32)kernel.shared[((threadIdx.x48) + 9)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[23](float32)kernel.shared[((threadIdx.x48) + 6)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[32](float32)kernel.shared[((threadIdx.x48) + 9)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[24](float32)kernel.shared[((threadIdx.x48) + 6)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[33](float32)kernel.shared[((threadIdx.x48) + 9)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[18](float32)kernel.shared[((threadIdx.x48) + 30)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[27](float32)kernel.shared[((threadIdx.x48) + 33)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[19](float32)kernel.shared[((threadIdx.x48) + 30)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[28](float32)kernel.shared[((threadIdx.x48) + 33)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[20](float32)kernel.shared[((threadIdx.x48) + 30)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[29](float32)kernel.shared[((threadIdx.x48) + 33)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[21](float32)kernel.shared[((threadIdx.x48) + 30)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[30](float32)kernel.shared[((threadIdx.x48) + 33)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[22](float32)kernel.shared[((threadIdx.x48) + 30)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[31](float32)kernel.shared[((threadIdx.x48) + 33)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[23](float32)kernel.shared[((threadIdx.x48) + 30)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[32](float32)kernel.shared[((threadIdx.x48) + 33)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[24](float32)kernel.shared[((threadIdx.x48) + 30)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[33](float32)kernel.shared[((threadIdx.x48) + 33)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[19](float32)kernel.shared[((threadIdx.x48) + 7)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[28](float32)kernel.shared[((threadIdx.x48) + 10)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[20](float32)kernel.shared[((threadIdx.x48) + 7)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[29](float32)kernel.shared[((threadIdx.x48) + 10)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[21](float32)kernel.shared[((threadIdx.x48) + 7)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[30](float32)kernel.shared[((threadIdx.x48) + 10)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[22](float32)kernel.shared[((threadIdx.x48) + 7)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[31](float32)kernel.shared[((threadIdx.x48) + 10)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[23](float32)kernel.shared[((threadIdx.x48) + 7)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[32](float32)kernel.shared[((threadIdx.x48) + 10)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[24](float32)kernel.shared[((threadIdx.x48) + 7)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[33](float32)kernel.shared[((threadIdx.x48) + 10)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[25](float32)kernel.shared[((threadIdx.x48) + 7)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[34](float32)kernel.shared[((threadIdx.x48) + 10)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[19](float32)kernel.shared[((threadIdx.x48) + 31)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[28](float32)kernel.shared[((threadIdx.x48) + 34)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[20](float32)kernel.shared[((threadIdx.x48) + 31)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[29](float32)kernel.shared[((threadIdx.x48) + 34)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[21](float32)kernel.shared[((threadIdx.x48) + 31)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[30](float32)kernel.shared[((threadIdx.x48) + 34)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[22](float32)kernel.shared[((threadIdx.x48) + 31)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[31](float32)kernel.shared[((threadIdx.x48) + 34)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[23](float32)kernel.shared[((threadIdx.x48) + 31)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[32](float32)kernel.shared[((threadIdx.x48) + 34)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[24](float32)kernel.shared[((threadIdx.x48) + 31)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[33](float32)kernel.shared[((threadIdx.x48) + 34)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[25](float32)kernel.shared[((threadIdx.x48) + 31)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[34](float32)kernel.shared[((threadIdx.x48) + 34)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[20](float32)kernel.shared[((threadIdx.x48) + 8)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[29](float32)kernel.shared[((threadIdx.x48) + 11)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[21](float32)kernel.shared[((threadIdx.x48) + 8)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[30](float32)kernel.shared[((threadIdx.x48) + 11)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[22](float32)kernel.shared[((threadIdx.x48) + 8)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[31](float32)kernel.shared[((threadIdx.x48) + 11)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[23](float32)kernel.shared[((threadIdx.x48) + 8)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[32](float32)kernel.shared[((threadIdx.x48) + 11)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[24](float32)kernel.shared[((threadIdx.x48) + 8)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[33](float32)kernel.shared[((threadIdx.x48) + 11)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[25](float32)kernel.shared[((threadIdx.x48) + 8)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[34](float32)kernel.shared[((threadIdx.x48) + 11)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[26](float32)kernel.shared[((threadIdx.x48) + 8)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[35](float32)kernel.shared[((threadIdx.x48) + 11)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[20](float32)kernel.shared[((threadIdx.x48) + 32)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[29](float32)kernel.shared[((threadIdx.x48) + 35)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[21](float32)kernel.shared[((threadIdx.x48) + 32)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[30](float32)kernel.shared[((threadIdx.x48) + 35)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[22](float32)kernel.shared[((threadIdx.x48) + 32)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[31](float32)kernel.shared[((threadIdx.x48) + 35)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[23](float32)kernel.shared[((threadIdx.x48) + 32)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[32](float32)kernel.shared[((threadIdx.x48) + 35)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[24](float32)kernel.shared[((threadIdx.x48) + 32)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[33](float32)kernel.shared[((threadIdx.x48) + 35)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[25](float32)kernel.shared[((threadIdx.x48) + 32)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[34](float32)kernel.shared[((threadIdx.x48) + 35)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[26](float32)kernel.shared[((threadIdx.x48) + 32)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[35](float32)kernel.shared[((threadIdx.x48) + 35)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[36](float32)kernel.shared[((threadIdx.x48) + 12)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[45](float32)kernel.shared[((threadIdx.x48) + 15)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[37](float32)kernel.shared[((threadIdx.x48) + 12)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[46](float32)kernel.shared[((threadIdx.x48) + 15)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[38](float32)kernel.shared[((threadIdx.x48) + 12)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[47](float32)kernel.shared[((threadIdx.x48) + 15)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[39](float32)kernel.shared[((threadIdx.x48) + 12)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[48](float32)kernel.shared[((threadIdx.x48) + 15)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[40](float32)kernel.shared[((threadIdx.x48) + 12)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[49](float32)kernel.shared[((threadIdx.x48) + 15)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[41](float32)kernel.shared[((threadIdx.x48) + 12)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[50](float32)kernel.shared[((threadIdx.x48) + 15)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[42](float32)kernel.shared[((threadIdx.x48) + 12)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[51](float32)kernel.shared[((threadIdx.x48) + 15)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[36](float32)kernel.shared[((threadIdx.x48) + 36)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[45](float32)kernel.shared[((threadIdx.x48) + 39)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[37](float32)kernel.shared[((threadIdx.x48) + 36)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[46](float32)kernel.shared[((threadIdx.x48) + 39)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[38](float32)kernel.shared[((threadIdx.x48) + 36)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[47](float32)kernel.shared[((threadIdx.x48) + 39)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[39](float32)kernel.shared[((threadIdx.x48) + 36)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[48](float32)kernel.shared[((threadIdx.x48) + 39)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[40](float32)kernel.shared[((threadIdx.x48) + 36)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[49](float32)kernel.shared[((threadIdx.x48) + 39)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[41](float32)kernel.shared[((threadIdx.x48) + 36)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[50](float32)kernel.shared[((threadIdx.x48) + 39)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[42](float32)kernel.shared[((threadIdx.x48) + 36)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[51](float32)kernel.shared[((threadIdx.x48) + 39)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[37](float32)kernel.shared[((threadIdx.x48) + 13)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[46](float32)kernel.shared[((threadIdx.x48) + 16)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[38](float32)kernel.shared[((threadIdx.x48) + 13)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[47](float32)kernel.shared[((threadIdx.x48) + 16)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[39](float32)kernel.shared[((threadIdx.x48) + 13)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[48](float32)kernel.shared[((threadIdx.x48) + 16)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[40](float32)kernel.shared[((threadIdx.x48) + 13)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[49](float32)kernel.shared[((threadIdx.x48) + 16)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[41](float32)kernel.shared[((threadIdx.x48) + 13)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[50](float32)kernel.shared[((threadIdx.x48) + 16)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[42](float32)kernel.shared[((threadIdx.x48) + 13)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[51](float32)kernel.shared[((threadIdx.x48) + 16)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[43](float32)kernel.shared[((threadIdx.x48) + 13)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[52](float32)kernel.shared[((threadIdx.x48) + 16)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[37](float32)kernel.shared[((threadIdx.x48) + 37)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[46](float32)kernel.shared[((threadIdx.x48) + 40)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[38](float32)kernel.shared[((threadIdx.x48) + 37)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[47](float32)kernel.shared[((threadIdx.x48) + 40)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[39](float32)kernel.shared[((threadIdx.x48) + 37)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[48](float32)kernel.shared[((threadIdx.x48) + 40)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[40](float32)kernel.shared[((threadIdx.x48) + 37)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[49](float32)kernel.shared[((threadIdx.x48) + 40)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[41](float32)kernel.shared[((threadIdx.x48) + 37)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[50](float32)kernel.shared[((threadIdx.x48) + 40)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[42](float32)kernel.shared[((threadIdx.x48) + 37)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[51](float32)kernel.shared[((threadIdx.x48) + 40)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[43](float32)kernel.shared[((threadIdx.x48) + 37)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[52](float32)kernel.shared[((threadIdx.x48) + 40)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[38](float32)kernel.shared[((threadIdx.x48) + 14)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[47](float32)kernel.shared[((threadIdx.x48) + 17)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[39](float32)kernel.shared[((threadIdx.x48) + 14)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[48](float32)kernel.shared[((threadIdx.x48) + 17)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[40](float32)kernel.shared[((threadIdx.x48) + 14)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[49](float32)kernel.shared[((threadIdx.x48) + 17)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[41](float32)kernel.shared[((threadIdx.x48) + 14)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[50](float32)kernel.shared[((threadIdx.x48) + 17)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[42](float32)kernel.shared[((threadIdx.x48) + 14)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[51](float32)kernel.shared[((threadIdx.x48) + 17)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[43](float32)kernel.shared[((threadIdx.x48) + 14)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[52](float32)kernel.shared[((threadIdx.x48) + 17)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[44](float32)kernel.shared[((threadIdx.x48) + 14)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[53](float32)kernel.shared[((threadIdx.x48) + 17)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[38](float32)kernel.shared[((threadIdx.x48) + 38)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[47](float32)kernel.shared[((threadIdx.x48) + 41)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[39](float32)kernel.shared[((threadIdx.x48) + 38)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[48](float32)kernel.shared[((threadIdx.x48) + 41)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[40](float32)kernel.shared[((threadIdx.x48) + 38)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[49](float32)kernel.shared[((threadIdx.x48) + 41)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[41](float32)kernel.shared[((threadIdx.x48) + 38)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[50](float32)kernel.shared[((threadIdx.x48) + 41)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[42](float32)kernel.shared[((threadIdx.x48) + 38)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[51](float32)kernel.shared[((threadIdx.x48) + 41)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[43](float32)kernel.shared[((threadIdx.x48) + 38)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[52](float32)kernel.shared[((threadIdx.x48) + 41)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[44](float32)kernel.shared[((threadIdx.x48) + 38)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[53](float32)kernel.shared[((threadIdx.x48) + 41)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[54](float32)kernel.shared[((threadIdx.x48) + 18)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[63](float32)kernel.shared[((threadIdx.x48) + 21)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[55](float32)kernel.shared[((threadIdx.x48) + 18)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[64](float32)kernel.shared[((threadIdx.x48) + 21)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[56](float32)kernel.shared[((threadIdx.x48) + 18)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[65](float32)kernel.shared[((threadIdx.x48) + 21)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[57](float32)kernel.shared[((threadIdx.x48) + 18)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[66](float32)kernel.shared[((threadIdx.x48) + 21)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[58](float32)kernel.shared[((threadIdx.x48) + 18)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[67](float32)kernel.shared[((threadIdx.x48) + 21)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[59](float32)kernel.shared[((threadIdx.x48) + 18)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[68](float32)kernel.shared[((threadIdx.x48) + 21)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[60](float32)kernel.shared[((threadIdx.x48) + 18)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[69](float32)kernel.shared[((threadIdx.x48) + 21)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[54](float32)kernel.shared[((threadIdx.x48) + 42)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[63](float32)kernel.shared[((threadIdx.x48) + 45)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[55](float32)kernel.shared[((threadIdx.x48) + 42)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[64](float32)kernel.shared[((threadIdx.x48) + 45)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[56](float32)kernel.shared[((threadIdx.x48) + 42)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[65](float32)kernel.shared[((threadIdx.x48) + 45)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[57](float32)kernel.shared[((threadIdx.x48) + 42)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[66](float32)kernel.shared[((threadIdx.x48) + 45)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[58](float32)kernel.shared[((threadIdx.x48) + 42)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[67](float32)kernel.shared[((threadIdx.x48) + 45)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[59](float32)kernel.shared[((threadIdx.x48) + 42)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[68](float32)kernel.shared[((threadIdx.x48) + 45)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[60](float32)kernel.shared[((threadIdx.x48) + 42)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[69](float32)kernel.shared[((threadIdx.x48) + 45)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[55](float32)kernel.shared[((threadIdx.x48) + 19)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[64](float32)kernel.shared[((threadIdx.x48) + 22)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[56](float32)kernel.shared[((threadIdx.x48) + 19)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[65](float32)kernel.shared[((threadIdx.x48) + 22)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[57](float32)kernel.shared[((threadIdx.x48) + 19)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[66](float32)kernel.shared[((threadIdx.x48) + 22)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[58](float32)kernel.shared[((threadIdx.x48) + 19)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[67](float32)kernel.shared[((threadIdx.x48) + 22)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[59](float32)kernel.shared[((threadIdx.x48) + 19)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[68](float32)kernel.shared[((threadIdx.x48) + 22)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[60](float32)kernel.shared[((threadIdx.x48) + 19)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[69](float32)kernel.shared[((threadIdx.x48) + 22)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[61](float32)kernel.shared[((threadIdx.x48) + 19)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[70](float32)kernel.shared[((threadIdx.x48) + 22)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[55](float32)kernel.shared[((threadIdx.x48) + 43)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[64](float32)kernel.shared[((threadIdx.x48) + 46)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[56](float32)kernel.shared[((threadIdx.x48) + 43)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[65](float32)kernel.shared[((threadIdx.x48) + 46)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[57](float32)kernel.shared[((threadIdx.x48) + 43)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[66](float32)kernel.shared[((threadIdx.x48) + 46)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[58](float32)kernel.shared[((threadIdx.x48) + 43)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[67](float32)kernel.shared[((threadIdx.x48) + 46)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[59](float32)kernel.shared[((threadIdx.x48) + 43)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[68](float32)kernel.shared[((threadIdx.x48) + 46)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[60](float32)kernel.shared[((threadIdx.x48) + 43)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[69](float32)kernel.shared[((threadIdx.x48) + 46)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[61](float32)kernel.shared[((threadIdx.x48) + 43)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[70](float32)kernel.shared[((threadIdx.x48) + 46)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[56](float32)kernel.shared[((threadIdx.x48) + 20)]))
compute_3[0] = ((float32)compute_3[0] + ((float32*)pad_temp.shared[65](float32)kernel.shared[((threadIdx.x48) + 23)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[57](float32)kernel.shared[((threadIdx.x48) + 20)]))
compute_3[1] = ((float32)compute_3[1] + ((float32*)pad_temp.shared[66](float32)kernel.shared[((threadIdx.x48) + 23)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[58](float32)kernel.shared[((threadIdx.x48) + 20)]))
compute_3[2] = ((float32)compute_3[2] + ((float32*)pad_temp.shared[67](float32)kernel.shared[((threadIdx.x48) + 23)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[59](float32)kernel.shared[((threadIdx.x48) + 20)]))
compute_3[3] = ((float32)compute_3[3] + ((float32*)pad_temp.shared[68](float32)kernel.shared[((threadIdx.x48) + 23)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[60](float32)kernel.shared[((threadIdx.x48) + 20)]))
compute_3[4] = ((float32)compute_3[4] + ((float32*)pad_temp.shared[69](float32)kernel.shared[((threadIdx.x48) + 23)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[61](float32)kernel.shared[((threadIdx.x48) + 20)]))
compute_3[5] = ((float32)compute_3[5] + ((float32*)pad_temp.shared[70](float32)kernel.shared[((threadIdx.x48) + 23)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[62](float32)kernel.shared[((threadIdx.x48) + 20)]))
compute_3[6] = ((float32)compute_3[6] + ((float32*)pad_temp.shared[71](float32)kernel.shared[((threadIdx.x48) + 23)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[56](float32)kernel.shared[((threadIdx.x48) + 44)]))
compute_3[7] = ((float32)compute_3[7] + ((float32*)pad_temp.shared[65](float32)kernel.shared[((threadIdx.x48) + 47)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[57](float32)kernel.shared[((threadIdx.x48) + 44)]))
compute_3[8] = ((float32)compute_3[8] + ((float32*)pad_temp.shared[66](float32)kernel.shared[((threadIdx.x48) + 47)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[58](float32)kernel.shared[((threadIdx.x48) + 44)]))
compute_3[9] = ((float32)compute_3[9] + ((float32*)pad_temp.shared[67](float32)kernel.shared[((threadIdx.x48) + 47)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[59](float32)kernel.shared[((threadIdx.x48) + 44)]))
compute_3[10] = ((float32)compute_3[10] + ((float32*)pad_temp.shared[68](float32)kernel.shared[((threadIdx.x48) + 47)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[60](float32)kernel.shared[((threadIdx.x48) + 44)]))
compute_3[11] = ((float32)compute_3[11] + ((float32*)pad_temp.shared[69](float32)kernel.shared[((threadIdx.x48) + 47)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[61](float32)kernel.shared[((threadIdx.x48) + 44)]))
compute_3[12] = ((float32)compute_3[12] + ((float32*)pad_temp.shared[70](float32)kernel.shared[((threadIdx.x48) + 47)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[62](float32)kernel.shared[((threadIdx.x48) + 44)]))
compute_3[13] = ((float32)compute_3[13] + ((float32*)pad_temp.shared[71](float32)kernel.shared[((threadIdx.x48) + 47)]))
}
}
for (i1.inner: int32, 0, 2) {
for (i3.inner: int32, 0, 7) {
compute_2[(((((floordiv(blockIdx.x, 7)6272) + (threadIdx.x98)) + (i1.inner49)) + (floormod(blockIdx.x, 7)7)) + i3.inner)] = max(((float32)compute_3[((i1.inner7) + i3.inner)] + (float32)bias_2[(((floordiv(blockIdx.x, 7)128) + (threadIdx.x2)) + i1.inner)]), 0f32)
}
}
}
}
檢查正確性并評(píng)估性能
構(gòu)建二進(jìn)制文件并檢查其正確性和性能。
func = tvm.build(sch, args, target)

Check correctness

data_np = np.random.uniform(size=(N, CI, H, W)).astype(np.float32)
weight_np = np.random.uniform(size=(CO, CI, KH, KW)).astype(np.float32)
bias_np = np.random.uniform(size=(1, CO, 1, 1)).astype(np.float32)
conv_np = conv2d_nchw_python(data_np, weight_np, strides, padding)
out_np = np.maximum(conv_np + bias_np, 0.0)

ctx = tvm.gpu()
data_tvm = tvm.nd.array(data_np, ctx=ctx)
weight_tvm = tvm.nd.array(weight_np, ctx=ctx)
bias_tvm = tvm.nd.array(bias_np, ctx=ctx)
out_tvm = tvm.nd.empty(out_np.shape, ctx=ctx)
func(data_tvm, weight_tvm, bias_tvm, out_tvm)

Check results

np.testing.assert_allclose(out_np, out_tvm.asnumpy(), rtol=1e-3)

Evaluate execution time

evaluator = func.time_evaluator(func.entry_name, ctx, min_repeat_ms=500)
print(
“Execution time of this operator: %.3f ms”
% (np.median(evaluator(data_tvm, weight_tvm, bias_tvm, out_tvm).results) * 1000)
)
出：
Execution time of this operator: 0.417 ms
使用記錄文件
搜索期間，所有測(cè)量記錄都將轉(zhuǎn)儲(chǔ)到記錄文件“ conv2d.json”中。測(cè)量記錄可用于重新應(yīng)用搜索結(jié)果，繼續(xù)搜索以及執(zhí)行其他分析。
這是一個(gè)示例，其中從文件加載最佳調(diào)度，并打印等效的python，調(diào)度API和CUDA源代碼。可用于調(diào)試和學(xué)習(xí)自動(dòng)調(diào)度程序的行為。
print(“Equivalent python schedule:”)
print(task.print_best(log_file, print_mode=“schedule”))

print(“CUDA source code:”)
print(task.print_best(log_file, print_mode=“cuda”))
輸出：
Equivalent python schedule:
pad_temp_i0, pad_temp_i1, pad_temp_i2, pad_temp_i3 = tuple(pad_temp.op.axis) + tuple(pad_temp.op.reduce_axis)
compute_nn, compute_ff, compute_yy, compute_xx, compute_rc, compute_ry, compute_rx = tuple(compute.op.axis) + tuple(compute.op.reduce_axis)
T_add_ax0, T_add_ax1, T_add_ax2, T_add_ax3 = tuple(T_add.op.axis) + tuple(T_add.op.reduce_axis)
compute_i0, compute_i1, compute_i2, compute_i3 = tuple(compute.op.axis) + tuple(compute.op.reduce_axis)
s[T_add].compute_inline()
compute_nn_o_i, compute_nn_i = s[compute].split(compute_nn, factor=1)
compute_nn_o_o_i, compute_nn_o_i = s[compute].split(compute_nn_o_i, factor=1)
compute_nn_o_o_o_i, compute_nn_o_o_i = s[compute].split(compute_nn_o_o_i, factor=1)
compute_nn_o_o_o_o, compute_nn_o_o_o_i = s[compute].split(compute_nn_o_o_o_i, factor=1)
compute_ff_o_i, compute_ff_i = s[compute].split(compute_ff, factor=1)
compute_ff_o_o_i, compute_ff_o_i = s[compute].split(compute_ff_o_i, factor=2)
compute_ff_o_o_o_i, compute_ff_o_o_i = s[compute].split(compute_ff_o_o_i, factor=64)
compute_ff_o_o_o_o, compute_ff_o_o_o_i = s[compute].split(compute_ff_o_o_o_i, factor=1)
compute_yy_o_i, compute_yy_i = s[compute].split(compute_yy, factor=1)
compute_yy_o_o_i, compute_yy_o_i = s[compute].split(compute_yy_o_i, factor=1)
compute_yy_o_o_o_i, compute_yy_o_o_i = s[compute].split(compute_yy_o_o_i, factor=1)
compute_yy_o_o_o_o, compute_yy_o_o_o_i = s[compute].split(compute_yy_o_o_o_i, factor=1)
compute_xx_o_i, compute_xx_i = s[compute].split(compute_xx, factor=1)
compute_xx_o_o_i, compute_xx_o_i = s[compute].split(compute_xx_o_i, factor=7)
compute_xx_o_o_o_i, compute_xx_o_o_i = s[compute].split(compute_xx_o_o_i, factor=1)
compute_xx_o_o_o_o, compute_xx_o_o_o_i = s[compute].split(compute_xx_o_o_o_i, factor=1)
compute_rc_o_i, compute_rc_i = s[compute].split(compute_rc, factor=2)
compute_rc_o_o, compute_rc_o_i = s[compute].split(compute_rc_o_i, factor=4)
compute_ry_o_i, compute_ry_i = s[compute].split(compute_ry, factor=1)
compute_ry_o_o, compute_ry_o_i = s[compute].split(compute_ry_o_i, factor=1)
compute_rx_o_i, compute_rx_i = s[compute].split(compute_rx, factor=1)
compute_rx_o_o, compute_rx_o_i = s[compute].split(compute_rx_o_i, factor=3)
s[compute].reorder(compute_nn_o_o_o_o, compute_ff_o_o_o_o, compute_yy_o_o_o_o, compute_xx_o_o_o_o, compute_nn_o_o_o_i, compute_ff_o_o_o_i, compute_yy_o_o_o_i, compute_xx_o_o_o_i, compute_nn_o_o_i, compute_ff_o_o_i, compute_yy_o_o_i, compute_xx_o_o_i, compute_rc_o_o, compute_ry_o_o, compute_rx_o_o, compute_rc_o_i, compute_ry_o_i, compute_rx_o_i, compute_nn_o_i, compute_ff_o_i, compute_yy_o_i, compute_xx_o_i, compute_rc_i, compute_ry_i, compute_rx_i, compute_nn_i, compute_ff_i, compute_yy_i, compute_xx_i)
compute_i0_o_i, compute_i0_i = s[compute].split(compute_i0, factor=1)
compute_i0_o_o_i, compute_i0_o_i = s[compute].split(compute_i0_o_i, factor=1)
compute_i0_o_o_o, compute_i0_o_o_i = s[compute].split(compute_i0_o_o_i, factor=1)
compute_i1_o_i, compute_i1_i = s[compute].split(compute_i1, factor=2)
compute_i1_o_o_i, compute_i1_o_i = s[compute].split(compute_i1_o_i, factor=64)
compute_i1_o_o_o, compute_i1_o_o_i = s[compute].split(compute_i1_o_o_i, factor=1)
compute_i2_o_i, compute_i2_i = s[compute].split(compute_i2, factor=1)
compute_i2_o_o_i, compute_i2_o_i = s[compute].split(compute_i2_o_i, factor=1)
compute_i2_o_o_o, compute_i2_o_o_i = s[compute].split(compute_i2_o_o_i, factor=1)
compute_i3_o_i, compute_i3_i = s[compute].split(compute_i3, factor=7)
compute_i3_o_o_i, compute_i3_o_i = s[compute].split(compute_i3_o_i, factor=1)
compute_i3_o_o_o, compute_i3_o_o_i = s[compute].split(compute_i3_o_o_i, factor=1)
s[compute].reorder(compute_i0_o_o_o, compute_i1_o_o_o, compute_i2_o_o_o, compute_i3_o_o_o, compute_i0_o_o_i, compute_i1_o_o_i, compute_i2_o_o_i, compute_i3_o_o_i, compute_i0_o_i, compute_i1_o_i, compute_i2_o_i, compute_i3_o_i, compute_i0_i, compute_i1_i, compute_i2_i, compute_i3_i)
s[compute].compute_at(s[compute], compute_i3_o_i)
kernel_shared = s.cache_read(kernel, “shared”, [compute])
kernel_shared_ax0, kernel_shared_ax1, kernel_shared_ax2, kernel_shared_ax3 = tuple(kernel_shared.op.axis)
s[kernel_shared].compute_at(s[compute], compute_rx_o_o)
pad_temp_shared = s.cache_read(pad_temp, “shared”, [compute])
pad_temp_shared_ax0, pad_temp_shared_ax1, pad_temp_shared_ax2, pad_temp_shared_ax3 = tuple(pad_temp_shared.op.axis)
s[pad_temp_shared].compute_at(s[compute], compute_rx_o_o)
s[pad_temp].compute_inline()
compute_i0_o_o_o_i1_o_o_o_fused_i2_o_o_o_fused_i3_o_o_o_fused = s[compute].fuse(compute_i0_o_o_o, compute_i1_o_o_o, compute_i2_o_o_o, compute_i3_o_o_o)
s[compute].bind(compute_i0_o_o_o_i1_o_o_o_fused_i2_o_o_o_fused_i3_o_o_o_fused, te.thread_axis(“blockIdx.x”))
compute_i0_o_o_i_i1_o_o_i_fused_i2_o_o_i_fused_i3_o_o_i_fused = s[compute].fuse(compute_i0_o_o_i, compute_i1_o_o_i, compute_i2_o_o_i, compute_i3_o_o_i)
s[compute].bind(compute_i0_o_o_i_i1_o_o_i_fused_i2_o_o_i_fused_i3_o_o_i_fused, te.thread_axis(“vthread”))
compute_i0_o_i_i1_o_i_fused_i2_o_i_fused_i3_o_i_fused = s[compute].fuse(compute_i0_o_i, compute_i1_o_i, compute_i2_o_i, compute_i3_o_i)
s[compute].bind(compute_i0_o_i_i1_o_i_fused_i2_o_i_fused_i3_o_i_fused, te.thread_axis(“threadIdx.x”))
kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused = s[kernel_shared].fuse(kernel_shared_ax0, kernel_shared_ax1, kernel_shared_ax2, kernel_shared_ax3)
kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=1)
s[kernel_shared].vectorize(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i)
kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[kernel_shared].split(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=64)
s[kernel_shared].bind(kernel_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i, te.thread_axis(“threadIdx.x”))
pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused = s[pad_temp_shared].fuse(pad_temp_shared_ax0, pad_temp_shared_ax1, pad_temp_shared_ax2, pad_temp_shared_ax3)
pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused, factor=4)
s[pad_temp_shared].vectorize(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_i)
pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_o, pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i = s[pad_temp_shared].split(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o, factor=64)
s[pad_temp_shared].bind(pad_temp_shared_ax0_ax1_fused_ax2_fused_ax3_fused_o_i, te.thread_axis(“threadIdx.x”))
s[compute].pragma(compute_nn_o_o_o_o, “auto_unroll_max_step”, 512)
s[compute].pragma(compute_nn_o_o_o_o, “unroll_explicit”, True)

CUDA source code:
extern “C” global void default_function_kernel0(float* restrict data, float* restrict kernel, float* restrict compute, float* restrict bias) {
float compute1[14];
shared float pad_temp_shared[72];
shared float kernel_shared[3072];
compute1[(0)] = 0.000000e+00f;
compute1[(1)] = 0.000000e+00f;
compute1[(2)] = 0.000000e+00f;
compute1[(3)] = 0.000000e+00f;
compute1[(4)] = 0.000000e+00f;
compute1[(5)] = 0.000000e+00f;
compute1[(6)] = 0.000000e+00f;
compute1[(7)] = 0.000000e+00f;
compute1[(8)] = 0.000000e+00f;
compute1[(9)] = 0.000000e+00f;
compute1[(10)] = 0.000000e+00f;
compute1[(11)] = 0.000000e+00f;
compute1[(12)] = 0.000000e+00f;
compute1[(13)] = 0.000000e+00f;
for (int rc_outer_outer = 0; rc_outer_outer < 64; ++rc_outer_outer) {
for (int ry_outer_outer = 0; ry_outer_outer < 3; ++ry_outer_outer) {
__syncthreads();
if (((int)threadIdx.x) < 18) {
pad_temp_shared[((((int)threadIdx.x) * 4))] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= ((((int)threadIdx.x) * 4) % 9))) && (((((int)threadIdx.x) * 4) % 9) < 8)) ? data[(((((((rc_outer_outer * 392) + (((((int)threadIdx.x) * 4) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + ((((int)threadIdx.x) * 4) % 9)) - 8))] : 0.000000e+00f);
}
if (((int)threadIdx.x) < 18) {
pad_temp_shared[(((((int)threadIdx.x) * 4) + 1))] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 1) % 9))) && ((((((int)threadIdx.x) * 4) + 1) % 9) < 8)) ? data[(((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 1) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 1) % 9)) - 8))] : 0.000000e+00f);
}
if (((int)threadIdx.x) < 18) {
pad_temp_shared[(((((int)threadIdx.x) * 4) + 2))] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 2) % 9))) && ((((((int)threadIdx.x) * 4) + 2) % 9) < 8)) ? data[(((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 2) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 2) % 9)) - 8))] : 0.000000e+00f);
}
if (((int)threadIdx.x) < 18) {
pad_temp_shared[(((((int)threadIdx.x) * 4) + 3))] = (((((1 <= (ry_outer_outer + (((int)blockIdx.x) % 7))) && ((ry_outer_outer + (((int)blockIdx.x) % 7)) < 8)) && (1 <= (((((int)threadIdx.x) * 4) + 3) % 9))) && ((((((int)threadIdx.x) * 4) + 3) % 9) < 8)) ? data[(((((((rc_outer_outer * 392) + ((((((int)threadIdx.x) * 4) + 3) / 9) * 49)) + (ry_outer_outer * 7)) + ((((int)blockIdx.x) % 7) * 7)) + (((((int)threadIdx.x) * 4) + 3) % 9)) - 8))] : 0.000000e+00f);
}
kernel_shared[(((int)threadIdx.x))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)))];
kernel_shared[((((int)threadIdx.x) + 64))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 64) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 128))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 128) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 192))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 36864))];
kernel_shared[((((int)threadIdx.x) + 256))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 256) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 320))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 320) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 384))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 73728))];
kernel_shared[((((int)threadIdx.x) + 448))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 448) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 512))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 512) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 576))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 110592))];
kernel_shared[((((int)threadIdx.x) + 640))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 640) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 704))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 704) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 768))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 147456))];
kernel_shared[((((int)threadIdx.x) + 832))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 832) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 896))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 896) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 960))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 184320))];
kernel_shared[((((int)threadIdx.x) + 1024))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1024) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1088))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1088) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1152))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 221184))];
kernel_shared[((((int)threadIdx.x) + 1216))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1216) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1280))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1280) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1344))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 258048))];
kernel_shared[((((int)threadIdx.x) + 1408))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1408) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1472))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1472) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1536))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 294912))];
kernel_shared[((((int)threadIdx.x) + 1600))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1600) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1664))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1664) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1728))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 331776))];
kernel_shared[((((int)threadIdx.x) + 1792))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1792) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1856))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1856) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 1920))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 368640))];
kernel_shared[((((int)threadIdx.x) + 1984))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 1984) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2048))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2048) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2112))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 405504))];
kernel_shared[((((int)threadIdx.x) + 2176))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2176) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2240))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2240) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2304))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 442368))];
kernel_shared[((((int)threadIdx.x) + 2368))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2368) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2432))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2432) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2496))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 479232))];
kernel_shared[((((int)threadIdx.x) + 2560))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2560) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2624))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2624) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2688))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 516096))];
kernel_shared[((((int)threadIdx.x) + 2752))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2752) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2816))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2816) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
kernel_shared[((((int)threadIdx.x) + 2880))] = kernel[(((((((((((int)blockIdx.x) / 7) * 589824) + ((((int)threadIdx.x) / 24) * 4608)) + (rc_outer_outer * 72)) + (((((int)threadIdx.x) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + (((int)threadIdx.x) % 3)) + 552960))];
kernel_shared[((((int)threadIdx.x) + 2944))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 2944) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 16) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 1) % 3)))];
kernel_shared[((((int)threadIdx.x) + 3008))] = kernel[((((((((((int)blockIdx.x) / 7) * 589824) + (((((int)threadIdx.x) + 3008) / 24) * 4608)) + (rc_outer_outer * 72)) + ((((((int)threadIdx.x) + 8) % 24) / 3) * 9)) + (ry_outer_outer * 3)) + ((((int)threadIdx.x) + 2) % 3)))];
__syncthreads();
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(0)] * kernel_shared[((((int)threadIdx.x) * 48))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(9)] * kernel_shared[(((((int)threadIdx.x) * 48) + 3))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(1)] * kernel_shared[((((int)threadIdx.x) * 48))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(10)] * kernel_shared[(((((int)threadIdx.x) * 48) + 3))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(2)] * kernel_shared[((((int)threadIdx.x) * 48))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(11)] * kernel_shared[(((((int)threadIdx.x) * 48) + 3))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(3)] * kernel_shared[((((int)threadIdx.x) * 48))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(12)] * kernel_shared[(((((int)threadIdx.x) * 48) + 3))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(4)] * kernel_shared[((((int)threadIdx.x) * 48))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(13)] * kernel_shared[(((((int)threadIdx.x) * 48) + 3))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(5)] * kernel_shared[((((int)threadIdx.x) * 48))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(14)] * kernel_shared[(((((int)threadIdx.x) * 48) + 3))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(6)] * kernel_shared[((((int)threadIdx.x) * 48))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(15)] * kernel_shared[(((((int)threadIdx.x) * 48) + 3))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(0)] * kernel_shared[(((((int)threadIdx.x) * 48) + 24))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(9)] * kernel_shared[(((((int)threadIdx.x) * 48) + 27))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(1)] * kernel_shared[(((((int)threadIdx.x) * 48) + 24))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(10)] * kernel_shared[(((((int)threadIdx.x) * 48) + 27))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(2)] * kernel_shared[(((((int)threadIdx.x) * 48) + 24))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(11)] * kernel_shared[(((((int)threadIdx.x) * 48) + 27))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(3)] * kernel_shared[(((((int)threadIdx.x) * 48) + 24))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(12)] * kernel_shared[(((((int)threadIdx.x) * 48) + 27))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(4)] * kernel_shared[(((((int)threadIdx.x) * 48) + 24))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(13)] * kernel_shared[(((((int)threadIdx.x) * 48) + 27))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(5)] * kernel_shared[(((((int)threadIdx.x) * 48) + 24))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(14)] * kernel_shared[(((((int)threadIdx.x) * 48) + 27))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(6)] * kernel_shared[(((((int)threadIdx.x) * 48) + 24))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(15)] * kernel_shared[(((((int)threadIdx.x) * 48) + 27))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(1)] * kernel_shared[(((((int)threadIdx.x) * 48) + 1))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(10)] * kernel_shared[(((((int)threadIdx.x) * 48) + 4))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(2)] * kernel_shared[(((((int)threadIdx.x) * 48) + 1))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(11)] * kernel_shared[(((((int)threadIdx.x) * 48) + 4))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(3)] * kernel_shared[(((((int)threadIdx.x) * 48) + 1))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(12)] * kernel_shared[(((((int)threadIdx.x) * 48) + 4))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(4)] * kernel_shared[(((((int)threadIdx.x) * 48) + 1))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(13)] * kernel_shared[(((((int)threadIdx.x) * 48) + 4))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(5)] * kernel_shared[(((((int)threadIdx.x) * 48) + 1))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(14)] * kernel_shared[(((((int)threadIdx.x) * 48) + 4))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(6)] * kernel_shared[(((((int)threadIdx.x) * 48) + 1))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(15)] * kernel_shared[(((((int)threadIdx.x) * 48) + 4))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(7)] * kernel_shared[(((((int)threadIdx.x) * 48) + 1))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(16)] * kernel_shared[(((((int)threadIdx.x) * 48) + 4))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(1)] * kernel_shared[(((((int)threadIdx.x) * 48) + 25))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(10)] * kernel_shared[(((((int)threadIdx.x) * 48) + 28))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(2)] * kernel_shared[(((((int)threadIdx.x) * 48) + 25))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(11)] * kernel_shared[(((((int)threadIdx.x) * 48) + 28))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(3)] * kernel_shared[(((((int)threadIdx.x) * 48) + 25))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(12)] * kernel_shared[(((((int)threadIdx.x) * 48) + 28))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(4)] * kernel_shared[(((((int)threadIdx.x) * 48) + 25))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(13)] * kernel_shared[(((((int)threadIdx.x) * 48) + 28))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(5)] * kernel_shared[(((((int)threadIdx.x) * 48) + 25))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(14)] * kernel_shared[(((((int)threadIdx.x) * 48) + 28))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(6)] * kernel_shared[(((((int)threadIdx.x) * 48) + 25))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(15)] * kernel_shared[(((((int)threadIdx.x) * 48) + 28))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(7)] * kernel_shared[(((((int)threadIdx.x) * 48) + 25))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(16)] * kernel_shared[(((((int)threadIdx.x) * 48) + 28))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(2)] * kernel_shared[(((((int)threadIdx.x) * 48) + 2))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(11)] * kernel_shared[(((((int)threadIdx.x) * 48) + 5))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(3)] * kernel_shared[(((((int)threadIdx.x) * 48) + 2))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(12)] * kernel_shared[(((((int)threadIdx.x) * 48) + 5))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(4)] * kernel_shared[(((((int)threadIdx.x) * 48) + 2))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(13)] * kernel_shared[(((((int)threadIdx.x) * 48) + 5))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(5)] * kernel_shared[(((((int)threadIdx.x) * 48) + 2))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(14)] * kernel_shared[(((((int)threadIdx.x) * 48) + 5))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(6)] * kernel_shared[(((((int)threadIdx.x) * 48) + 2))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(15)] * kernel_shared[(((((int)threadIdx.x) * 48) + 5))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(7)] * kernel_shared[(((((int)threadIdx.x) * 48) + 2))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(16)] * kernel_shared[(((((int)threadIdx.x) * 48) + 5))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(8)] * kernel_shared[(((((int)threadIdx.x) * 48) + 2))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(17)] * kernel_shared[(((((int)threadIdx.x) * 48) + 5))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(2)] * kernel_shared[(((((int)threadIdx.x) * 48) + 26))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(11)] * kernel_shared[(((((int)threadIdx.x) * 48) + 29))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(3)] * kernel_shared[(((((int)threadIdx.x) * 48) + 26))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(12)] * kernel_shared[(((((int)threadIdx.x) * 48) + 29))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(4)] * kernel_shared[(((((int)threadIdx.x) * 48) + 26))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(13)] * kernel_shared[(((((int)threadIdx.x) * 48) + 29))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(5)] * kernel_shared[(((((int)threadIdx.x) * 48) + 26))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(14)] * kernel_shared[(((((int)threadIdx.x) * 48) + 29))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(6)] * kernel_shared[(((((int)threadIdx.x) * 48) + 26))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(15)] * kernel_shared[(((((int)threadIdx.x) * 48) + 29))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(7)] * kernel_shared[(((((int)threadIdx.x) * 48) + 26))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(16)] * kernel_shared[(((((int)threadIdx.x) * 48) + 29))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(8)] * kernel_shared[(((((int)threadIdx.x) * 48) + 26))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(17)] * kernel_shared[(((((int)threadIdx.x) * 48) + 29))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(18)] * kernel_shared[(((((int)threadIdx.x) * 48) + 6))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(27)] * kernel_shared[(((((int)threadIdx.x) * 48) + 9))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(19)] * kernel_shared[(((((int)threadIdx.x) * 48) + 6))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(28)] * kernel_shared[(((((int)threadIdx.x) * 48) + 9))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(20)] * kernel_shared[(((((int)threadIdx.x) * 48) + 6))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(29)] * kernel_shared[(((((int)threadIdx.x) * 48) + 9))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(21)] * kernel_shared[(((((int)threadIdx.x) * 48) + 6))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(30)] * kernel_shared[(((((int)threadIdx.x) * 48) + 9))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(22)] * kernel_shared[(((((int)threadIdx.x) * 48) + 6))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(31)] * kernel_shared[(((((int)threadIdx.x) * 48) + 9))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(23)] * kernel_shared[(((((int)threadIdx.x) * 48) + 6))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(32)] * kernel_shared[(((((int)threadIdx.x) * 48) + 9))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(24)] * kernel_shared[(((((int)threadIdx.x) * 48) + 6))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(33)] * kernel_shared[(((((int)threadIdx.x) * 48) + 9))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(18)] * kernel_shared[(((((int)threadIdx.x) * 48) + 30))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(27)] * kernel_shared[(((((int)threadIdx.x) * 48) + 33))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(19)] * kernel_shared[(((((int)threadIdx.x) * 48) + 30))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(28)] * kernel_shared[(((((int)threadIdx.x) * 48) + 33))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(20)] * kernel_shared[(((((int)threadIdx.x) * 48) + 30))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(29)] * kernel_shared[(((((int)threadIdx.x) * 48) + 33))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(21)] * kernel_shared[(((((int)threadIdx.x) * 48) + 30))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(30)] * kernel_shared[(((((int)threadIdx.x) * 48) + 33))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(22)] * kernel_shared[(((((int)threadIdx.x) * 48) + 30))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(31)] * kernel_shared[(((((int)threadIdx.x) * 48) + 33))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(23)] * kernel_shared[(((((int)threadIdx.x) * 48) + 30))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(32)] * kernel_shared[(((((int)threadIdx.x) * 48) + 33))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(24)] * kernel_shared[(((((int)threadIdx.x) * 48) + 30))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(33)] * kernel_shared[(((((int)threadIdx.x) * 48) + 33))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(19)] * kernel_shared[(((((int)threadIdx.x) * 48) + 7))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(28)] * kernel_shared[(((((int)threadIdx.x) * 48) + 10))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(20)] * kernel_shared[(((((int)threadIdx.x) * 48) + 7))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(29)] * kernel_shared[(((((int)threadIdx.x) * 48) + 10))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(21)] * kernel_shared[(((((int)threadIdx.x) * 48) + 7))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(30)] * kernel_shared[(((((int)threadIdx.x) * 48) + 10))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(22)] * kernel_shared[(((((int)threadIdx.x) * 48) + 7))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(31)] * kernel_shared[(((((int)threadIdx.x) * 48) + 10))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(23)] * kernel_shared[(((((int)threadIdx.x) * 48) + 7))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(32)] * kernel_shared[(((((int)threadIdx.x) * 48) + 10))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(24)] * kernel_shared[(((((int)threadIdx.x) * 48) + 7))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(33)] * kernel_shared[(((((int)threadIdx.x) * 48) + 10))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(25)] * kernel_shared[(((((int)threadIdx.x) * 48) + 7))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(34)] * kernel_shared[(((((int)threadIdx.x) * 48) + 10))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(19)] * kernel_shared[(((((int)threadIdx.x) * 48) + 31))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(28)] * kernel_shared[(((((int)threadIdx.x) * 48) + 34))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(20)] * kernel_shared[(((((int)threadIdx.x) * 48) + 31))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(29)] * kernel_shared[(((((int)threadIdx.x) * 48) + 34))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(21)] * kernel_shared[(((((int)threadIdx.x) * 48) + 31))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(30)] * kernel_shared[(((((int)threadIdx.x) * 48) + 34))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(22)] * kernel_shared[(((((int)threadIdx.x) * 48) + 31))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(31)] * kernel_shared[(((((int)threadIdx.x) * 48) + 34))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(23)] * kernel_shared[(((((int)threadIdx.x) * 48) + 31))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(32)] * kernel_shared[(((((int)threadIdx.x) * 48) + 34))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(24)] * kernel_shared[(((((int)threadIdx.x) * 48) + 31))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(33)] * kernel_shared[(((((int)threadIdx.x) * 48) + 34))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(25)] * kernel_shared[(((((int)threadIdx.x) * 48) + 31))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(34)] * kernel_shared[(((((int)threadIdx.x) * 48) + 34))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(20)] * kernel_shared[(((((int)threadIdx.x) * 48) + 8))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(29)] * kernel_shared[(((((int)threadIdx.x) * 48) + 11))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(21)] * kernel_shared[(((((int)threadIdx.x) * 48) + 8))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(30)] * kernel_shared[(((((int)threadIdx.x) * 48) + 11))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(22)] * kernel_shared[(((((int)threadIdx.x) * 48) + 8))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(31)] * kernel_shared[(((((int)threadIdx.x) * 48) + 11))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(23)] * kernel_shared[(((((int)threadIdx.x) * 48) + 8))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(32)] * kernel_shared[(((((int)threadIdx.x) * 48) + 11))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(24)] * kernel_shared[(((((int)threadIdx.x) * 48) + 8))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(33)] * kernel_shared[(((((int)threadIdx.x) * 48) + 11))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(25)] * kernel_shared[(((((int)threadIdx.x) * 48) + 8))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(34)] * kernel_shared[(((((int)threadIdx.x) * 48) + 11))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(26)] * kernel_shared[(((((int)threadIdx.x) * 48) + 8))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(35)] * kernel_shared[(((((int)threadIdx.x) * 48) + 11))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(20)] * kernel_shared[(((((int)threadIdx.x) * 48) + 32))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(29)] * kernel_shared[(((((int)threadIdx.x) * 48) + 35))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(21)] * kernel_shared[(((((int)threadIdx.x) * 48) + 32))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(30)] * kernel_shared[(((((int)threadIdx.x) * 48) + 35))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(22)] * kernel_shared[(((((int)threadIdx.x) * 48) + 32))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(31)] * kernel_shared[(((((int)threadIdx.x) * 48) + 35))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(23)] * kernel_shared[(((((int)threadIdx.x) * 48) + 32))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(32)] * kernel_shared[(((((int)threadIdx.x) * 48) + 35))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(24)] * kernel_shared[(((((int)threadIdx.x) * 48) + 32))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(33)] * kernel_shared[(((((int)threadIdx.x) * 48) + 35))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(25)] * kernel_shared[(((((int)threadIdx.x) * 48) + 32))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(34)] * kernel_shared[(((((int)threadIdx.x) * 48) + 35))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(26)] * kernel_shared[(((((int)threadIdx.x) * 48) + 32))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(35)] * kernel_shared[(((((int)threadIdx.x) * 48) + 35))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(36)] * kernel_shared[(((((int)threadIdx.x) * 48) + 12))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(45)] * kernel_shared[(((((int)threadIdx.x) * 48) + 15))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(37)] * kernel_shared[(((((int)threadIdx.x) * 48) + 12))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(46)] * kernel_shared[(((((int)threadIdx.x) * 48) + 15))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(38)] * kernel_shared[(((((int)threadIdx.x) * 48) + 12))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(47)] * kernel_shared[(((((int)threadIdx.x) * 48) + 15))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(39)] * kernel_shared[(((((int)threadIdx.x) * 48) + 12))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(48)] * kernel_shared[(((((int)threadIdx.x) * 48) + 15))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(40)] * kernel_shared[(((((int)threadIdx.x) * 48) + 12))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(49)] * kernel_shared[(((((int)threadIdx.x) * 48) + 15))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(41)] * kernel_shared[(((((int)threadIdx.x) * 48) + 12))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(50)] * kernel_shared[(((((int)threadIdx.x) * 48) + 15))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(42)] * kernel_shared[(((((int)threadIdx.x) * 48) + 12))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(51)] * kernel_shared[(((((int)threadIdx.x) * 48) + 15))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(36)] * kernel_shared[(((((int)threadIdx.x) * 48) + 36))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(45)] * kernel_shared[(((((int)threadIdx.x) * 48) + 39))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(37)] * kernel_shared[(((((int)threadIdx.x) * 48) + 36))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(46)] * kernel_shared[(((((int)threadIdx.x) * 48) + 39))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(38)] * kernel_shared[(((((int)threadIdx.x) * 48) + 36))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(47)] * kernel_shared[(((((int)threadIdx.x) * 48) + 39))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(39)] * kernel_shared[(((((int)threadIdx.x) * 48) + 36))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(48)] * kernel_shared[(((((int)threadIdx.x) * 48) + 39))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(40)] * kernel_shared[(((((int)threadIdx.x) * 48) + 36))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(49)] * kernel_shared[(((((int)threadIdx.x) * 48) + 39))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(41)] * kernel_shared[(((((int)threadIdx.x) * 48) + 36))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(50)] * kernel_shared[(((((int)threadIdx.x) * 48) + 39))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(42)] * kernel_shared[(((((int)threadIdx.x) * 48) + 36))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(51)] * kernel_shared[(((((int)threadIdx.x) * 48) + 39))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(37)] * kernel_shared[(((((int)threadIdx.x) * 48) + 13))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(46)] * kernel_shared[(((((int)threadIdx.x) * 48) + 16))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(38)] * kernel_shared[(((((int)threadIdx.x) * 48) + 13))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(47)] * kernel_shared[(((((int)threadIdx.x) * 48) + 16))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(39)] * kernel_shared[(((((int)threadIdx.x) * 48) + 13))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(48)] * kernel_shared[(((((int)threadIdx.x) * 48) + 16))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(40)] * kernel_shared[(((((int)threadIdx.x) * 48) + 13))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(49)] * kernel_shared[(((((int)threadIdx.x) * 48) + 16))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(41)] * kernel_shared[(((((int)threadIdx.x) * 48) + 13))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(50)] * kernel_shared[(((((int)threadIdx.x) * 48) + 16))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(42)] * kernel_shared[(((((int)threadIdx.x) * 48) + 13))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(51)] * kernel_shared[(((((int)threadIdx.x) * 48) + 16))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(43)] * kernel_shared[(((((int)threadIdx.x) * 48) + 13))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(52)] * kernel_shared[(((((int)threadIdx.x) * 48) + 16))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(37)] * kernel_shared[(((((int)threadIdx.x) * 48) + 37))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(46)] * kernel_shared[(((((int)threadIdx.x) * 48) + 40))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(38)] * kernel_shared[(((((int)threadIdx.x) * 48) + 37))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(47)] * kernel_shared[(((((int)threadIdx.x) * 48) + 40))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(39)] * kernel_shared[(((((int)threadIdx.x) * 48) + 37))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(48)] * kernel_shared[(((((int)threadIdx.x) * 48) + 40))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(40)] * kernel_shared[(((((int)threadIdx.x) * 48) + 37))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(49)] * kernel_shared[(((((int)threadIdx.x) * 48) + 40))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(41)] * kernel_shared[(((((int)threadIdx.x) * 48) + 37))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(50)] * kernel_shared[(((((int)threadIdx.x) * 48) + 40))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(42)] * kernel_shared[(((((int)threadIdx.x) * 48) + 37))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(51)] * kernel_shared[(((((int)threadIdx.x) * 48) + 40))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(43)] * kernel_shared[(((((int)threadIdx.x) * 48) + 37))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(52)] * kernel_shared[(((((int)threadIdx.x) * 48) + 40))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(38)] * kernel_shared[(((((int)threadIdx.x) * 48) + 14))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(47)] * kernel_shared[(((((int)threadIdx.x) * 48) + 17))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(39)] * kernel_shared[(((((int)threadIdx.x) * 48) + 14))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(48)] * kernel_shared[(((((int)threadIdx.x) * 48) + 17))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(40)] * kernel_shared[(((((int)threadIdx.x) * 48) + 14))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(49)] * kernel_shared[(((((int)threadIdx.x) * 48) + 17))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(41)] * kernel_shared[(((((int)threadIdx.x) * 48) + 14))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(50)] * kernel_shared[(((((int)threadIdx.x) * 48) + 17))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(42)] * kernel_shared[(((((int)threadIdx.x) * 48) + 14))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(51)] * kernel_shared[(((((int)threadIdx.x) * 48) + 17))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(43)] * kernel_shared[(((((int)threadIdx.x) * 48) + 14))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(52)] * kernel_shared[(((((int)threadIdx.x) * 48) + 17))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(44)] * kernel_shared[(((((int)threadIdx.x) * 48) + 14))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(53)] * kernel_shared[(((((int)threadIdx.x) * 48) + 17))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(38)] * kernel_shared[(((((int)threadIdx.x) * 48) + 38))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(47)] * kernel_shared[(((((int)threadIdx.x) * 48) + 41))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(39)] * kernel_shared[(((((int)threadIdx.x) * 48) + 38))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(48)] * kernel_shared[(((((int)threadIdx.x) * 48) + 41))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(40)] * kernel_shared[(((((int)threadIdx.x) * 48) + 38))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(49)] * kernel_shared[(((((int)threadIdx.x) * 48) + 41))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(41)] * kernel_shared[(((((int)threadIdx.x) * 48) + 38))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(50)] * kernel_shared[(((((int)threadIdx.x) * 48) + 41))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(42)] * kernel_shared[(((((int)threadIdx.x) * 48) + 38))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(51)] * kernel_shared[(((((int)threadIdx.x) * 48) + 41))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(43)] * kernel_shared[(((((int)threadIdx.x) * 48) + 38))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(52)] * kernel_shared[(((((int)threadIdx.x) * 48) + 41))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(44)] * kernel_shared[(((((int)threadIdx.x) * 48) + 38))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(53)] * kernel_shared[(((((int)threadIdx.x) * 48) + 41))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(54)] * kernel_shared[(((((int)threadIdx.x) * 48) + 18))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(63)] * kernel_shared[(((((int)threadIdx.x) * 48) + 21))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(55)] * kernel_shared[(((((int)threadIdx.x) * 48) + 18))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(64)] * kernel_shared[(((((int)threadIdx.x) * 48) + 21))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(56)] * kernel_shared[(((((int)threadIdx.x) * 48) + 18))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(65)] * kernel_shared[(((((int)threadIdx.x) * 48) + 21))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(57)] * kernel_shared[(((((int)threadIdx.x) * 48) + 18))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(66)] * kernel_shared[(((((int)threadIdx.x) * 48) + 21))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(58)] * kernel_shared[(((((int)threadIdx.x) * 48) + 18))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(67)] * kernel_shared[(((((int)threadIdx.x) * 48) + 21))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(59)] * kernel_shared[(((((int)threadIdx.x) * 48) + 18))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(68)] * kernel_shared[(((((int)threadIdx.x) * 48) + 21))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(60)] * kernel_shared[(((((int)threadIdx.x) * 48) + 18))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(69)] * kernel_shared[(((((int)threadIdx.x) * 48) + 21))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(54)] * kernel_shared[(((((int)threadIdx.x) * 48) + 42))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(63)] * kernel_shared[(((((int)threadIdx.x) * 48) + 45))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(55)] * kernel_shared[(((((int)threadIdx.x) * 48) + 42))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(64)] * kernel_shared[(((((int)threadIdx.x) * 48) + 45))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(56)] * kernel_shared[(((((int)threadIdx.x) * 48) + 42))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(65)] * kernel_shared[(((((int)threadIdx.x) * 48) + 45))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(57)] * kernel_shared[(((((int)threadIdx.x) * 48) + 42))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(66)] * kernel_shared[(((((int)threadIdx.x) * 48) + 45))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(58)] * kernel_shared[(((((int)threadIdx.x) * 48) + 42))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(67)] * kernel_shared[(((((int)threadIdx.x) * 48) + 45))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(59)] * kernel_shared[(((((int)threadIdx.x) * 48) + 42))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(68)] * kernel_shared[(((((int)threadIdx.x) * 48) + 45))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(60)] * kernel_shared[(((((int)threadIdx.x) * 48) + 42))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(69)] * kernel_shared[(((((int)threadIdx.x) * 48) + 45))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(55)] * kernel_shared[(((((int)threadIdx.x) * 48) + 19))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(64)] * kernel_shared[(((((int)threadIdx.x) * 48) + 22))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(56)] * kernel_shared[(((((int)threadIdx.x) * 48) + 19))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(65)] * kernel_shared[(((((int)threadIdx.x) * 48) + 22))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(57)] * kernel_shared[(((((int)threadIdx.x) * 48) + 19))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(66)] * kernel_shared[(((((int)threadIdx.x) * 48) + 22))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(58)] * kernel_shared[(((((int)threadIdx.x) * 48) + 19))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(67)] * kernel_shared[(((((int)threadIdx.x) * 48) + 22))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(59)] * kernel_shared[(((((int)threadIdx.x) * 48) + 19))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(68)] * kernel_shared[(((((int)threadIdx.x) * 48) + 22))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(60)] * kernel_shared[(((((int)threadIdx.x) * 48) + 19))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(69)] * kernel_shared[(((((int)threadIdx.x) * 48) + 22))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(61)] * kernel_shared[(((((int)threadIdx.x) * 48) + 19))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(70)] * kernel_shared[(((((int)threadIdx.x) * 48) + 22))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(55)] * kernel_shared[(((((int)threadIdx.x) * 48) + 43))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(64)] * kernel_shared[(((((int)threadIdx.x) * 48) + 46))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(56)] * kernel_shared[(((((int)threadIdx.x) * 48) + 43))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(65)] * kernel_shared[(((((int)threadIdx.x) * 48) + 46))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(57)] * kernel_shared[(((((int)threadIdx.x) * 48) + 43))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(66)] * kernel_shared[(((((int)threadIdx.x) * 48) + 46))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(58)] * kernel_shared[(((((int)threadIdx.x) * 48) + 43))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(67)] * kernel_shared[(((((int)threadIdx.x) * 48) + 46))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(59)] * kernel_shared[(((((int)threadIdx.x) * 48) + 43))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(68)] * kernel_shared[(((((int)threadIdx.x) * 48) + 46))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(60)] * kernel_shared[(((((int)threadIdx.x) * 48) + 43))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(69)] * kernel_shared[(((((int)threadIdx.x) * 48) + 46))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(61)] * kernel_shared[(((((int)threadIdx.x) * 48) + 43))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(70)] * kernel_shared[(((((int)threadIdx.x) * 48) + 46))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(56)] * kernel_shared[(((((int)threadIdx.x) * 48) + 20))]));
compute1[(0)] = (compute1[(0)] + (pad_temp_shared[(65)] * kernel_shared[(((((int)threadIdx.x) * 48) + 23))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(57)] * kernel_shared[(((((int)threadIdx.x) * 48) + 20))]));
compute1[(1)] = (compute1[(1)] + (pad_temp_shared[(66)] * kernel_shared[(((((int)threadIdx.x) * 48) + 23))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(58)] * kernel_shared[(((((int)threadIdx.x) * 48) + 20))]));
compute1[(2)] = (compute1[(2)] + (pad_temp_shared[(67)] * kernel_shared[(((((int)threadIdx.x) * 48) + 23))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(59)] * kernel_shared[(((((int)threadIdx.x) * 48) + 20))]));
compute1[(3)] = (compute1[(3)] + (pad_temp_shared[(68)] * kernel_shared[(((((int)threadIdx.x) * 48) + 23))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(60)] * kernel_shared[(((((int)threadIdx.x) * 48) + 20))]));
compute1[(4)] = (compute1[(4)] + (pad_temp_shared[(69)] * kernel_shared[(((((int)threadIdx.x) * 48) + 23))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(61)] * kernel_shared[(((((int)threadIdx.x) * 48) + 20))]));
compute1[(5)] = (compute1[(5)] + (pad_temp_shared[(70)] * kernel_shared[(((((int)threadIdx.x) * 48) + 23))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(62)] * kernel_shared[(((((int)threadIdx.x) * 48) + 20))]));
compute1[(6)] = (compute1[(6)] + (pad_temp_shared[(71)] * kernel_shared[(((((int)threadIdx.x) * 48) + 23))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(56)] * kernel_shared[(((((int)threadIdx.x) * 48) + 44))]));
compute1[(7)] = (compute1[(7)] + (pad_temp_shared[(65)] * kernel_shared[(((((int)threadIdx.x) * 48) + 47))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(57)] * kernel_shared[(((((int)threadIdx.x) * 48) + 44))]));
compute1[(8)] = (compute1[(8)] + (pad_temp_shared[(66)] * kernel_shared[(((((int)threadIdx.x) * 48) + 47))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(58)] * kernel_shared[(((((int)threadIdx.x) * 48) + 44))]));
compute1[(9)] = (compute1[(9)] + (pad_temp_shared[(67)] * kernel_shared[(((((int)threadIdx.x) * 48) + 47))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(59)] * kernel_shared[(((((int)threadIdx.x) * 48) + 44))]));
compute1[(10)] = (compute1[(10)] + (pad_temp_shared[(68)] * kernel_shared[(((((int)threadIdx.x) * 48) + 47))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(60)] * kernel_shared[(((((int)threadIdx.x) * 48) + 44))]));
compute1[(11)] = (compute1[(11)] + (pad_temp_shared[(69)] * kernel_shared[(((((int)threadIdx.x) * 48) + 47))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(61)] * kernel_shared[(((((int)threadIdx.x) * 48) + 44))]));
compute1[(12)] = (compute1[(12)] + (pad_temp_shared[(70)] * kernel_shared[(((((int)threadIdx.x) * 48) + 47))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(62)] * kernel_shared[(((((int)threadIdx.x) * 48) + 44))]));
compute1[(13)] = (compute1[(13)] + (pad_temp_shared[(71)] * kernel_shared[(((((int)threadIdx.x) * 48) + 47))]));
}
}
for (int i1_inner = 0; i1_inner < 2; ++i1_inner) {
for (int i3_inner = 0; i3_inner < 7; ++i3_inner) {
compute[(((((((((int)blockIdx.x) / 7) * 6272) + (((int)threadIdx.x) * 98)) + (i1_inner * 49)) + ((((int)blockIdx.x) % 7) * 7)) + i3_inner))] = max((compute1[(((i1_inner * 7) + i3_inner))] + bias[(((((((int)blockIdx.x) / 7) * 128) + (((int)threadIdx.x) * 2)) + i1_inner))]), 0.000000e+00f);
}
}
}
一個(gè)更復(fù)雜的示例是繼續(xù)搜索。在這種情況下，需要自己創(chuàng)建搜索策略和成本模型，并使用日志文件恢復(fù)搜索策略和成本模型的狀態(tài)。在下面的示例中，恢復(fù)狀態(tài)并進(jìn)行5次以上的試用。
cost_model = auto_scheduler.XGBModel()
cost_model.update_from_file(log_file)
search_policy = auto_scheduler.SketchPolicy(
task, cost_model, init_search_callbacks=[auto_scheduler.PreloadMeasuredStates(log_file)]
)
measure_ctx = auto_scheduler.LocalRPCMeasureContext(min_repeat_ms=300)
tune_option = auto_scheduler.TuningOptions(
num_measure_trials=5,
runner=measure_ctx.runner,
measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
)
task.tune(tune_option, search_policy=search_policy)

Kill the measurement process

del measure_ctx
輸出：
Get devices for measurement successfully!
腳本的總運(yùn)行時(shí)間：（ 1分鐘34.433秒）
https://tvm.apache.org/docs/tutorials/auto_scheduler/tune_conv2d_layer_cuda.html#sphx-glr-tutorials-auto-scheduler-tune-conv2d-layer-cuda-py

總結(jié)

以上是生活随笔為你收集整理的自动调度GPU的卷积层的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。