wordpress 如何提交表單關鍵詞整站排名優(yōu)化
Apache TVM 是一個端到端的深度學習編譯框架，適用于 CPU、GPU 和各種機器學習加速芯片。更多 TVM 中文文檔可訪問 → https://tvm.hyper.ai/
作者：Masahiro Masuda，Truman Tian
本文介紹如何將 cuDNN 或 cuBLAS 等外部庫與 Relay 一起使用。
Relay 內(nèi)部用 TVM 來生成 target-specific 的代碼。例如，TVM 使用 CUDA 后端為用戶提供的網(wǎng)絡中的所有層生成 CUDA 內(nèi)核。有時也可將各個供應商開發(fā)的外部庫合并到 Relay 中，TVM 有一種機制可以透明地調(diào)用這些庫——對于 Relay 用戶，只需要設置一個適當?shù)?target 字符串。
使用 Relay 的外部庫前，用你要用的庫構建 TVM。例如，要用 cuDNN，需啟用 cmake/config.cmake 中的 USE_CUDNN 選項，必要時要指定 cuDNN 頭文件和庫目錄。
首先導入 Relay 和 TVM。
import tvm
from tvm import te
import numpy as np
from tvm.contrib import graph_executor as runtime
from tvm import relay
from tvm.relay import testing
import tvm.testing
創(chuàng)建一個簡單網(wǎng)絡?
下面創(chuàng)建一個簡單網(wǎng)絡進行演示，它由 convolution，batch normalization 和 ReLU activation 組成。
out_channels = 16
batch_size = 1data = relay.var("data", relay.TensorType((batch_size, 3, 224, 224), "float32"))
weight = relay.var("weight")
bn_gamma = relay.var("bn_gamma")
bn_beta = relay.var("bn_beta")
bn_mmean = relay.var("bn_mean")
bn_mvar = relay.var("bn_var")simple_net = relay.nn.conv2d(data=data, weight=weight, kernel_size=(3, 3), channels=out_channels, padding=(1, 1)
)
simple_net = relay.nn.batch_norm(simple_net, bn_gamma, bn_beta, bn_mmean, bn_mvar)[0]
simple_net = relay.nn.relu(simple_net)
simple_net = relay.Function(relay.analysis.free_vars(simple_net), simple_net)data_shape = (batch_size, 3, 224, 224)
net, params = testing.create_workload(simple_net)
使用 CUDA 后端構建和運行?
正常使用 CUDA 后端構建和運行這個網(wǎng)絡。設置日志記錄級別為 DEBUG，Relay 計算圖編譯的結(jié)果將作為偽代碼轉(zhuǎn)儲。
import logginglogging.basicConfig(level=logging.DEBUG) # to dump TVM IR after fusiontarget = "cuda"
lib = relay.build_module.build(net, target, params=params)dev = tvm.device(target, 0)
data = np.random.uniform(-1, 1, size=data_shape).astype("float32")
module = runtime.GraphModule(lib["default"](dev))
module.set_input("data", data)
module.run()
out_shape = (batch_size, out_channels, 224, 224)
out = module.get_output(0, tvm.nd.empty(out_shape))
out_cuda = out.numpy()
輸出結(jié)果：
/workspace/python/tvm/driver/build_module.py:268: UserWarning: target_host parameter is going to be deprecated. Please pass in tvm.target.Target(target, host=target_host) instead."target_host parameter is going to be deprecated. "
生成的偽代碼應如下。注意 bias add，batch normalization 和 ReLU activation 是如何融合到卷積核中的。 TVM 從這個表示中生成一個單一的融合內(nèi)核。
produce tensor {// attr [iter_var(blockIdx.z, , blockIdx.z)] thread_extent = 1// attr [compute] storage_scope = "local"allocate compute[float32 * 32]// attr [pad_temp.shared] storage_scope = "shared"allocate pad_temp.shared[float32 * 180]// attr [placeholder.shared] storage_scope = "shared"allocate placeholder.shared[float32 * 144]// attr [iter_var(blockIdx.y, , blockIdx.y)] thread_extent = 28// attr [iter_var(blockIdx.x, , blockIdx.x)] thread_extent = 14// attr [iter_var(threadIdx.z, , threadIdx.z)] thread_extent = 4// attr [iter_var(threadIdx.y, , threadIdx.y)] thread_extent = 1// attr [iter_var(threadIdx.x, , threadIdx.x)] thread_extent = 16produce compute {compute[0] = 0.000000fcompute[1] = 0.000000fcompute[2] = 0.000000fcompute[3] = 0.000000fcompute[4] = 0.000000fcompute[5] = 0.000000fcompute[6] = 0.000000fcompute[7] = 0.000000fcompute[8] = 0.000000fcompute[9] = 0.000000fcompute[10] = 0.000000fcompute[11] = 0.000000fcompute[12] = 0.000000fcompute[13] = 0.000000fcompute[14] = 0.000000fcompute[15] = 0.000000fcompute[16] = 0.000000fcompute[17] = 0.000000fcompute[18] = 0.000000fcompute[19] = 0.000000fcompute[20] = 0.000000fcompute[21] = 0.000000fcompute[22] = 0.000000fcompute[23] = 0.000000fcompute[24] = 0.000000fcompute[25] = 0.000000fcompute[26] = 0.000000fcompute[27] = 0.000000fcompute[28] = 0.000000fcompute[29] = 0.000000fcompute[30] = 0.000000fcompute[31] = 0.000000ffor (rc.outer, 0, 3) {produce pad_temp.shared {// attr [iter_var(threadIdx.z, , threadIdx.z)] thread_extent = 4// attr [iter_var(threadIdx.y, , threadIdx.y)] thread_extent = 1// attr [iter_var(threadIdx.x, , threadIdx.x)] thread_extent = 16if (likely(((threadIdx.z*15) < (60 - threadIdx.x)))) {if (likely((threadIdx.x < 15))) {pad_temp.shared[(((((threadIdx.z*15) + threadIdx.x)/60)*180) + ((((((threadIdx.z*15) + threadIdx.x)/6) % 10)*18) + ((((threadIdx.z*3) + threadIdx.x)*3) % 18)))] = tvm_if_then_else((((((1 - ((((threadIdx.z*15) + threadIdx.x)/6) % 10)) <= (blockIdx.y*8)) && ((blockIdx.y*8) < (225 - ((((threadIdx.z*15) + threadIdx.x)/6) % 10)))) && ((1 - ((((threadIdx.z*3) + threadIdx.x)*3) % 18)) <= (blockIdx.x*16))) && ((blockIdx.x*16) < (225 - ((((threadIdx.z*3) + threadIdx.x)*3) % 18)))), placeholder[((((((((blockIdx.y*112) + blockIdx.x) + (rc.outer*3136)) + ((((threadIdx.z*15) + threadIdx.x)/60)*9408))*16) + ((((threadIdx.z*3) + threadIdx.x)*3) % 18)) + (((((threadIdx.z*15) + threadIdx.x)/6) % 10)*224)) + -225)], 0.000000f)pad_temp.shared[(((((((threadIdx.z*15) + threadIdx.x)*3) + 1)/180)*180) + ((((((((threadIdx.z*15) + threadIdx.x)*3) + 1)/18) % 10)*18) + (((((threadIdx.z*3) + threadIdx.x)*3) + 1) % 18)))] = tvm_if_then_else((((((1 - ((((((threadIdx.z*15) + threadIdx.x)*3) + 1)/18) % 10)) <= (blockIdx.y*8)) && ((blockIdx.y*8) < (225 - ((((((threadIdx.z*15) + threadIdx.x)*3) + 1)/18) % 10)))) && ((1 - (((((threadIdx.z*3) + threadIdx.x)*3) + 1) % 18)) <= (blockIdx.x*16))) && ((blockIdx.x*16) < (225 - (((((threadIdx.z*3) + threadIdx.x)*3) + 1) % 18)))), placeholder[((((((((blockIdx.y*112) + blockIdx.x) + (rc.outer*3136)) + ((((((threadIdx.z*15) + threadIdx.x)*3) + 1)/180)*9408))*16) + (((((threadIdx.z*3) + threadIdx.x)*3) + 1) % 18)) + (((((((threadIdx.z*15) + threadIdx.x)*3) + 1)/18) % 10)*224)) + -225)], 0.000000f)pad_temp.shared[(((((((threadIdx.z*15) + threadIdx.x)*3) + 2)/180)*180) + ((((((((threadIdx.z*15) + threadIdx.x)*3) + 2)/18) % 10)*18) + (((((threadIdx.z*3) + threadIdx.x)*3) + 2) % 18)))] = tvm_if_then_else((((((1 - ((((((threadIdx.z*15) + threadIdx.x)*3) + 2)/18) % 10)) <= (blockIdx.y*8)) && ((blockIdx.y*8) < (225 - ((((((threadIdx.z*15) + threadIdx.x)*3) + 2)/18) % 10)))) && ((1 - (((((threadIdx.z*3) + threadIdx.x)*3) + 2) % 18)) <= (blockIdx.x*16))) && ((blockIdx.x*16) < (225 - (((((threadIdx.z*3) + threadIdx.x)*3) + 2) % 18)))), placeholder[((((((((blockIdx.y*112) + blockIdx.x) + (rc.outer*3136)) + ((((((threadIdx.z*15) + threadIdx.x)*3) + 2)/180)*9408))*16) + (((((threadIdx.z*3) + threadIdx.x)*3) + 2) % 18)) + (((((((threadIdx.z*15) + threadIdx.x)*3) + 2)/18) % 10)*224)) + -225)], 0.000000f)}}}produce placeholder.shared {// attr [iter_var(threadIdx.z, , threadIdx.z)] thread_extent = 4// attr [iter_var(threadIdx.y, , threadIdx.y)] thread_extent = 1// attr [iter_var(threadIdx.x, , threadIdx.x)] thread_extent = 16if (likely(((threadIdx.z*4) < (16 - (threadIdx.x/3))))) {if (likely(((threadIdx.z*12) < (48 - threadIdx.x)))) {if (likely((threadIdx.x < 12))) {placeholder.shared[(((((threadIdx.z*4) + (threadIdx.x/3))*3) + (threadIdx.x % 3))*3)] = placeholder[(((((rc.outer + (threadIdx.z*12)) + ((threadIdx.x/3)*3))*3) + (threadIdx.x % 3))*3)]placeholder.shared[((((((threadIdx.z*4) + (threadIdx.x/3))*3) + (threadIdx.x % 3))*3) + 1)] = placeholder[((((((rc.outer + (threadIdx.z*12)) + ((threadIdx.x/3)*3))*3) + (threadIdx.x % 3))*3) + 1)]placeholder.shared[((((((threadIdx.z*4) + (threadIdx.x/3))*3) + (threadIdx.x % 3))*3) + 2)] = placeholder[((((((rc.outer + (threadIdx.z*12)) + ((threadIdx.x/3)*3))*3) + (threadIdx.x % 3))*3) + 2)]}}}}compute[0] = (compute[0] + (pad_temp.shared[threadIdx.x]*placeholder.shared[(threadIdx.z*36)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[(threadIdx.z*36)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[(threadIdx.z*36)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[(threadIdx.z*36)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[(threadIdx.z*36)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[(threadIdx.z*36)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[(threadIdx.z*36)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[(threadIdx.z*36)]))compute[8] = (compute[8] + (pad_temp.shared[threadIdx.x]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 9)]))compute[16] = (compute[16] + (pad_temp.shared[threadIdx.x]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 18)]))compute[24] = (compute[24] + (pad_temp.shared[threadIdx.x]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 27)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 1)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 1)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 1)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 10)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 1)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 19)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 1)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 28)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 2)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 2)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 2)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 11)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 2)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 20)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 2)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 29)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 3)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 12)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 21)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 18)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 30)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 4)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 13)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 22)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 19)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 31)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 5)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 14)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 23)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 20)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 32)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 162)]*placeholder.shared[((threadIdx.z*36) + 6)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 162)]*placeholder.shared[((threadIdx.z*36) + 15)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 162)]*placeholder.shared[((threadIdx.z*36) + 24)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 36)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 54)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 72)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 90)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 108)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 126)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 144)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 162)]*placeholder.shared[((threadIdx.z*36) + 33)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 163)]*placeholder.shared[((threadIdx.z*36) + 7)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 163)]*placeholder.shared[((threadIdx.z*36) + 16)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 163)]*placeholder.shared[((threadIdx.z*36) + 25)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 37)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 55)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 73)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 91)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 109)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 127)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 145)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 163)]*placeholder.shared[((threadIdx.z*36) + 34)]))compute[0] = (compute[0] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[1] = (compute[1] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[2] = (compute[2] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[3] = (compute[3] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[4] = (compute[4] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[5] = (compute[5] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[6] = (compute[6] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[7] = (compute[7] + (pad_temp.shared[(threadIdx.x + 164)]*placeholder.shared[((threadIdx.z*36) + 8)]))compute[8] = (compute[8] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[9] = (compute[9] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[10] = (compute[10] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[11] = (compute[11] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[12] = (compute[12] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[13] = (compute[13] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[14] = (compute[14] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[15] = (compute[15] + (pad_temp.shared[(threadIdx.x + 164)]*placeholder.shared[((threadIdx.z*36) + 17)]))compute[16] = (compute[16] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[17] = (compute[17] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[18] = (compute[18] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[19] = (compute[19] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[20] = (compute[20] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[21] = (compute[21] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[22] = (compute[22] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[23] = (compute[23] + (pad_temp.shared[(threadIdx.x + 164)]*placeholder.shared[((threadIdx.z*36) + 26)]))compute[24] = (compute[24] + (pad_temp.shared[(threadIdx.x + 38)]*placeholder.shared[((threadIdx.z*36) + 35)]))compute[25] = (compute[25] + (pad_temp.shared[(threadIdx.x + 56)]*placeholder.shared[((threadIdx.z*36) + 35)]))compute[26] = (compute[26] + (pad_temp.shared[(threadIdx.x + 74)]*placeholder.shared[((threadIdx.z*36) + 35)]))compute[27] = (compute[27] + (pad_temp.shared[(threadIdx.x + 92)]*placeholder.shared[((threadIdx.z*36) + 35)]))compute[28] = (compute[28] + (pad_temp.shared[(threadIdx.x + 110)]*placeholder.shared[((threadIdx.z*36) + 35)]))compute[29] = (compute[29] + (pad_temp.shared[(threadIdx.x + 128)]*placeholder.shared[((threadIdx.z*36) + 35)]))compute[30] = (compute[30] + (pad_temp.shared[(threadIdx.x + 146)]*placeholder.shared[((threadIdx.z*36) + 35)]))compute[31] = (compute[31] + (pad_temp.shared[(threadIdx.x + 164)]*placeholder.shared[((threadIdx.z*36) + 35)]))}}tensor[(((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x)] = max(((compute[0]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 224)] = max(((compute[1]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 448)] = max(((compute[2]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 672)] = max(((compute[3]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 896)] = max(((compute[4]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 1120)] = max(((compute[5]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 1344)] = max(((compute[6]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 1568)] = max(((compute[7]*placeholder[(threadIdx.z*4)]) + placeholder[(threadIdx.z*4)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 50176)] = max(((compute[8]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 50400)] = max(((compute[9]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 50624)] = max(((compute[10]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 50848)] = max(((compute[11]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 51072)] = max(((compute[12]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 51296)] = max(((compute[13]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 51520)] = max(((compute[14]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 51744)] = max(((compute[15]*placeholder[((threadIdx.z*4) + 1)]) + placeholder[((threadIdx.z*4) + 1)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 100352)] = max(((compute[16]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 100576)] = max(((compute[17]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 100800)] = max(((compute[18]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 101024)] = max(((compute[19]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 101248)] = max(((compute[20]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 101472)] = max(((compute[21]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 101696)] = max(((compute[22]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 101920)] = max(((compute[23]*placeholder[((threadIdx.z*4) + 2)]) + placeholder[((threadIdx.z*4) + 2)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 150528)] = max(((compute[24]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 150752)] = max(((compute[25]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 150976)] = max(((compute[26]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 151200)] = max(((compute[27]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 151424)] = max(((compute[28]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 151648)] = max(((compute[29]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 151872)] = max(((compute[30]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)tensor[((((((blockIdx.y*112) + blockIdx.x) + (threadIdx.z*12544))*16) + threadIdx.x) + 152096)] = max(((compute[31]*placeholder[((threadIdx.z*4) + 3)]) + placeholder[((threadIdx.z*4) + 3)]), 0.000000f)
}
將 cuDNN 用于卷積層?
將選項 “-libs=cudnn” 附加到 target 字符串，從而用 cuDNN 將卷積核替換為 cuDNN。
net, params = testing.create_workload(simple_net)
target = "cuda -libs=cudnn" # use cudnn for convolution
lib = relay.build_module.build(net, target, params=params)dev = tvm.device(target, 0)
data = np.random.uniform(-1, 1, size=data_shape).astype("float32")
module = runtime.GraphModule(lib["default"](dev))
module.set_input("data", data)
module.run()
out_shape = (batch_size, out_channels, 224, 224)
out = module.get_output(0, tvm.nd.empty(out_shape))
out_cudnn = out.numpy()
輸出結(jié)果：
/workspace/python/tvm/driver/build_module.py:268: UserWarning: target_host parameter is going to be deprecated. Please pass in tvm.target.Target(target, host=target_host) instead."target_host parameter is going to be deprecated. "
注意，若用 cuDNN，Relay 無法將卷積與其后面的層融合。因為層融合發(fā)生在 TVM internal representation（IR）級別。 Relay 將外部庫視為黑盒，因此無法將它們與 TVM IR 融合。
下面的偽代碼顯示了 cuDNN 卷積 + bias add + batch norm + ReLU 變成了兩個計算階段，一個用于 cuDNN 調(diào)用，另一個用于其余操作。
// attr [y] storage_scope = "global"
allocate y[float32 * 802816]
produce y {// attr [0] extern_scope = 0tvm_call_packed("tvm.contrib.cudnn.conv2d.forward", 1, 0, 1, 1, 1, 1, 1, 1, 1, tvm_stack_make_array(placeholder, tvm_stack_make_shape(1, 3, 224, 224), 0, 4, 0.000000f, 0), tvm_stack_make_array(placeholder, tvm_stack_make_shape(16, 3, 3, 3), 0, 4, 0.000000f, 0), tvm_stack_make_array(y, tvm_stack_make_shape(1, 16, 224, 224), 0, 4, 0.000000f, 0))
}
produce tensor {// attr [iter_var(blockIdx.x, , blockIdx.x)] thread_extent = 256// attr [iter_var(threadIdx.x, , threadIdx.x)] thread_extent = 512for (ax0.ax1.fused.ax2.fused.ax3.fused.outer, 0, 7) {if (likely(((blockIdx.x*512) < ((802816 - (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072)) - threadIdx.x)))) {tensor[(((((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/802816)*802816) + (((((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/224) % 224)*224) + ((((blockIdx.x*64) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*32)) % 224))) + ((((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/50176) % 16)*50176))] = max(((y[(((((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/802816)*802816) + (((((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/224) % 224)*224) + ((((blockIdx.x*64) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*32)) % 224))) + ((((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/50176) % 16)*50176))]*placeholder[(((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/50176) % 16)]) + placeholder[(((((blockIdx.x*512) + threadIdx.x) + (ax0.ax1.fused.ax2.fused.ax3.fused.outer*131072))/50176) % 16)]), 0.000000f)}}
}
驗證結(jié)果?
檢查兩次運行的結(jié)果是否匹配。
tvm.testing.assert_allclose(out_cuda, out_cudnn, rtol=1e-5)
結(jié)論?
本教程介紹了 cuDNN 與 Relay 的使用，此外還支持 cuBLAS。若啟用了 cuBLAS，它將在全連接層（relay.dense）內(nèi)使用。若要用 cuBLAS，請將 target 字符串設置為 “cuda -libs=cublas”。也可以將 cuDNN 和 cuBLAS 與 “cuda -libs=cudnn,cublas” 一起使用。
對于 ROCm 后端，支持 MIOpen 和 rocBLAS。將 target 設置為 “rocm -libs=miopen,rocblas” 以啟用它們。
使用外部庫的注意事項：
首先，使用外部庫可能會限制 TVM 和 Relay 的使用。例如，MIOpen 目前只支持 NCHW 布局和 fp32 數(shù)據(jù)類型，因此不能在 TVM 中使用其他布局或數(shù)據(jù)類型。
其次，外部庫限制了計算圖編譯期間算子融合的可能性，如上所示。TVM 和 Relay 旨在通過聯(lián)合算子級別和計算圖級別優(yōu)化，在各種硬件上實現(xiàn)最佳性能。為了實現(xiàn)這個目標，應該繼續(xù)為 TVM 和 Relay 開發(fā)更好的優(yōu)化，同時在必要時使用外部庫回退到現(xiàn)有實現(xiàn)。
下載 Python 源代碼：using_external_lib.py
下載 Jupyter Notebook：using_external_lib.ipynb