diff --git a/llvm/projects/hpvm-tensor-rt/dnn_sources/src/unit_tests.cc b/llvm/projects/hpvm-tensor-rt/dnn_sources/src/unit_tests.cc
index ae9fdc1731f17a7f3dbc55df809c6a6f956e7ed4..1ccc5bd1b4e48a0019a78c6550b7ea9d1f679b91 100644
--- a/llvm/projects/hpvm-tensor-rt/dnn_sources/src/unit_tests.cc
+++ b/llvm/projects/hpvm-tensor-rt/dnn_sources/src/unit_tests.cc
@@ -369,10 +369,32 @@ void testTensorHalfGroupConv(){
}
-void testTensorPool(){
+void testTensorPooling(){
+
void* x = create4DTensor(CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 2, 1, 4, 4);
- fillTensorWithOnes(x);
- void* output = tensorPooling(x, 0, 2, 2, 0, 0, 1, 1);
+ fillTensorWithOnes(x);
+
+ float* data_arr = (float*) ((Tensor*) x)->host_data;
+ for(int i = 0; i < ((Tensor*) x)->num_elems; i += 4){
+ data_arr[i] = i;
+ }
+
+ void* output = tensorPooling(x, 0, 2, 2, 0, 0, 2, 2);
+ printTensorValues(output);
+}
+
+
+void testTensorHalfPooling(){
+
+ void* x = create4DTensor(CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 2, 1, 4, 4);
+ fillTensorWithOnes(x);
+
+ float* data_arr = (float*) ((Tensor*) x)->host_data;
+ for(int i = 0; i < ((Tensor*) x)->num_elems; i += 4){
+ data_arr[i] = i;
+ }
+
+ void* output = tensorPooling(x, 0, 2, 2, 0, 0, 2, 2);
printTensorValues(output);
}
@@ -701,8 +723,142 @@ void testSampleFilter(){
+
+void testPerforationCalls(void* input, void* filter,
+ int pad_h, int pad_w,
+ int stride_h, int stride_w,
+ int row, int col){
+
+
+ float interpolation_rate = 1.0;
+ for (int offset = 0; offset < 2; offset++){
+
+ printf("\n\n\n\**Test -- pad_h = %d pad_w = %d stride_h = %d stride_w = %d row = %d col = %d offset= %d \n\n",
+ pad_h, pad_w, stride_h, stride_w, row, col, offset);
+
+
+ void* res_exact = tensorConvolution(input, filter, pad_h, pad_w,
+ stride_h, stride_w,
+ 1, 1);
+
+ printf ("tensorConvolution Result :");
+ printTensorValues(res_exact);
+
+
+ void* res_exact2 = tensorConvApprox(input, filter, pad_h, pad_w,
+ stride_h, stride_w,
+ 1, 1, 1, 1, 1, 1);
+
+ printf ("\nBaseline Result :");
+ printTensorValues(res_exact2);
+
+
+ void* res_exact3 = tensorConvApproxHalf2(input, filter, pad_h, pad_w,
+ stride_h, stride_w,
+ 1, 1, 1, 1, 1, 1);
+ convertToFP32((struct Tensor*) res_exact3);
+
+ printf ("\nFP16_Baseline Result :");
+ printTensorValues(res_exact3);
+
+
+ void* res_sim = tensorConvPerfCuda(input, filter,
+ pad_h, pad_w,
+ stride_h, stride_w,
+ 1, 1,
+ row, col,
+ offset);
+
+ printf ("\nConvPerfCuda Result :");
+ printTensorValues(res_sim);
+
+
+ void* res = tensorConvApprox(input, filter,
+ pad_h, pad_w,
+ stride_h, stride_w,
+ 1, 1,
+ row, col,
+ 1, offset);
+
+
+ printf ("\nConvApprox Result :");
+ printTensorValues(res);
+
+
+ void* res_half = tensorConvApproxHalf2(input, filter,
+ pad_h, pad_w,
+ stride_h, stride_w,
+ 1, 1,
+ row, col,
+ 1, offset);
+
+ convertToFP32((struct Tensor*) res_half);
+
+ printf ("\nConvApproxHalf2 Result :");
+ printTensorValues(res_half);
+
+ }
+
+
+ printf ("\n\n\n\ --- End of Test \n\n\n");
+}
+
+
+
+
+
+/**** Tests Perforation for a set of different inputs */
void testPerforation(){
+
+ printf("***** Tests Sample for a sample 3 * 3 Filter ***** \n\n");
+ Tensor* input = (Tensor*) create4DTensor(CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 1, 3, 4, 4);
+ fillTensorWithVal(input, 1);
+
+ Tensor* filter = (Tensor*) create4DTensor(CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 1, 3, 3, 3);
+ fillTensorWithVal(filter, 1);
+
+
+ /*
+ float* host_ptr = (float*) ((struct Tensor*) filter)->host_data;
+ host_ptr[0] = 2;
+ host_ptr[2] = 2;
+ host_ptr[4] = 2;
+ host_ptr[6] = 2;
+ host_ptr[8] = 2;
+ host_ptr[10] = 2;
+ host_ptr[12] = 2;
+ host_ptr[14] = 2;
+ host_ptr[16] = 2;
+ host_ptr[18] = 2;
+ host_ptr[20] = 2;
+ host_ptr[22] = 2;
+ host_ptr[24] = 2;
+ host_ptr[26] = 2;
+ */
+
+
+ testPerforationCalls(input, filter, 0, 0, 1, 1, 1, 2);
+
+ testPerforationCalls(input, filter, 0, 0, 1, 1, 2, 1);
+
+
+ testPerforationCalls(input, filter, 1, 1, 1, 1, 1, 3);
+
+ testPerforationCalls(input, filter, 1, 1, 1, 1, 3, 1);
+
+
+ testPerforationCalls(input, filter, 1, 1, 2, 2, 1, 4);
+
+ testPerforationCalls(input, filter, 1, 1, 2, 2, 4, 1);
+
+}
+
+
+
+
+/*void testPerforation(){
+
printf("***** Testing Perforation ***** \n\n");
Tensor* input = (Tensor*) create4DTensor(CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 1, 1, 6, 6);
fillTensorWithVal(input, 3);
@@ -718,8 +874,6 @@ void testPerforation(){
host_ptr[3] = 0;
host_ptr[4] = 0;
host_ptr[5] = 0;
- //host_ptr[6] = 0;
- //host_ptr[7] = 0;
//printTensorValues(input);
@@ -746,7 +900,7 @@ void testPerforation(){
//printTensorValues(res4);
}
-
+*/
@@ -859,22 +1013,13 @@ void testSampling(){
void testSamplingCalls(void* input, void* filter,
int pad_h, int pad_w,
- int stride_h, int stride_w, int skip_every){
-
-
- //for (int i = 0; i < 2; i++){
-
- // float interpolation_rate;
- // if (i == 0)
- // interpolation_rate = 1.0;
- // else
- // interpolation_rate = 0.5;
+ int stride_h, int stride_w,
+ int skip_every){
float interpolation_rate = 1.0;
for (int offset = 0; offset < 2; offset++){
-
-
+
printf("\n\n\n\**Test -- pad_h = %d pad_w = %d stride_h = %d stride_w = %d skip_every = %d offset= %d interpolation_rate = %f \n\n",
pad_h, pad_w, stride_h, stride_w, skip_every, offset, interpolation_rate);
@@ -1252,9 +1397,13 @@ int main(){
testTensorBatchNorm();
testTensorHalfBatchNorm();
- ///testTensorPooling();
- ///testTensorHalfPooling();
+ testTensorPooling();
+ testTensorHalfPooling();
+
+ testSampling_3_3();
+ testSampling_1_1();
+ testPerforation();
//testTensorError();
//testQuantization();
@@ -1280,10 +1429,6 @@ int main(){
/********* SAMPLING TESTS ****
- testSampling_3_3();
-
-
- testSampling_1_1();
*************/