Merge pull request #313 from opencompl/sasha/stride-one

use stride of 1 in pooling kernels

Snakefile

@@ -653,6 +653,7 @@ rule kernel_generate_data_c:
     input:
         json="kernels/{kernel}/{shape}/params.json",
         h="kernels/{kernel}/{shape}/data.h",
+        gendata="kernels/{kernel}/gendata.py"
     output:
         "kernels/{kernel}/{shape}/data.c",
     wildcard_constraints:

kernels/pooling_nchw_max_d1_s2_3x3/baseline.c.template

@@ -3,10 +3,10 @@
 #include <stdint.h>
 
 void pooling_nchw_max_d1_s2_3x3(const double* restrict x, double* restrict y) {
-    for (int row = 0; row < H - 3 + 1; row += 2) {
-        for (int col = 0; col < W - 3 + 1; col += 2) {
-            int y_row = row / 2;
-            int y_col = col / 2;
+    for (int row = 0; row < H - 3 + 1; row += 1) {
+        for (int col = 0; col < W - 3 + 1; col += 1) {
+            int y_row = row / 1;
+            int y_col = col / 1;
             int y_index = (y_row * NEW_W) + y_col;
             // Load initial value in y
             double max_value = -10000.0;

kernels/pooling_nchw_max_d1_s2_3x3/data.h.template

@@ -2,8 +2,8 @@
 
 #define N 1
 #define C 1
-#define H {{(M - 1) * 2 + 3 + 1}}
-#define W {{(N - 1) * 2 + 3 + 1}}
+#define H {{(M - 1) * 1 + 3 + 1}}
+#define W {{(N - 1) * 1 + 3 + 1}}
 #define NEW_H {{M}}
 #define NEW_W {{N}}
 

kernels/pooling_nchw_max_d1_s2_3x3/gendata.py

@@ -16,7 +16,7 @@ def sum_pool_data(
 
     # Define the pooling parameters
     pool_size = (3, 3)
-    stride = 2
+    stride = 1
 
     new_h = M
     new_w = N
@@ -32,12 +32,15 @@ def sum_pool_data(
     y_in = np.random.uniform(rmin, rmax, (n, c, new_h, new_w))
     y_out = y_in.copy()
 
-    for row in range(0, H - pool_size[0] + 1, stride):
-        for col in range(0, W - pool_size[1] + 1, stride):
-            pooling_region = x[:, :, row : row + pool_size[0], col : col + pool_size[1]]
-            y_out[:, :, row // stride, col // stride] = np.max(
-                pooling_region, axis=(2, 3)
-            )
+    for row in range(new_h):
+        for col in range(new_w):
+            pooling_region = x[
+                :,
+                :,
+                row * stride : row * stride + pool_size[0],
+                col * stride : col * stride + pool_size[1],
+            ]
+            y_out[:, :, row, col] = np.max(pooling_region, axis=(2, 3))
 
     yield Define("N", n)
     yield Define("C", c)

kernels/pooling_nchw_max_d1_s2_3x3/linalg.mlir.template

@@ -1,7 +1,7 @@
-  func.func public @pooling_nchw_max_d1_s2_3x3(%X : tensor<1x1x{{(M - 1) * 2 + 3 + 1}}x{{(N - 1) * 2 + 3 + 1}}xf64> {"llvm.noalias"}, %Y : tensor<1x1x{{M}}x{{N}}xf64> {"llvm.noalias"}) -> tensor<1x1x{{M}}x{{N}}xf64> {
+  func.func public @pooling_nchw_max_d1_s2_3x3(%X : tensor<1x1x{{(M - 1) * 1 + 3 + 1}}x{{(N - 1) * 1 + 3 + 1}}xf64> {"llvm.noalias"}, %Y : tensor<1x1x{{M}}x{{N}}xf64> {"llvm.noalias"}) -> tensor<1x1x{{M}}x{{N}}xf64> {
     %min_val = arith.constant -10000.0 : f64
     %zeros = linalg.fill ins(%min_val : f64) outs(%Y : tensor<1x1x{{M}}x{{N}}xf64>) -> tensor<1x1x{{M}}x{{N}}xf64>
     %kernel = tensor.empty() : tensor<3x3xf64>
-    %res = linalg.pooling_nchw_max {"dilations" = dense<1> : vector<2xi64>, "strides" = dense<2> : vector<2xi64>} ins(%X, %kernel : tensor<1x1x{{(M - 1) * 2 + 3 + 1}}x{{(N - 1) * 2 + 3 + 1}}xf64>, tensor<3x3xf64>) outs(%zeros : tensor<1x1x{{M}}x{{N}}xf64>) -> tensor<1x1x{{M}}x{{N}}xf64>
+    %res = linalg.pooling_nchw_max {"dilations" = dense<1> : vector<2xi64>, "strides" = dense<1> : vector<2xi64>} ins(%X, %kernel : tensor<1x1x{{(M - 1) * 1 + 3 + 1}}x{{(N - 1) * 1 + 3 + 1}}xf64>, tensor<3x3xf64>) outs(%zeros : tensor<1x1x{{M}}x{{N}}xf64>) -> tensor<1x1x{{M}}x{{N}}xf64>
     func.return %res : tensor<1x1x{{M}}x{{N}}xf64>
   }

kernels/pooling_nchw_sum_d1_s2_3x3/baseline.c.template

@@ -3,12 +3,12 @@
 #include <stdint.h>
 
 void pooling_nchw_sum_d1_s2_3x3(const double* restrict x, double* restrict y) {
-    for (int row = 0; row < H - 3 + 1; row += 2) {
-        for (int col = 0; col < W - 3 + 1; col += 2) {
+    for (int row = 0; row < H - 3 + 1; row += 1) {
+        for (int col = 0; col < W - 3 + 1; col += 1) {
             for (int i = 0; i < N; i++) {
                 for (int j = 0; j < C; j++) {
-                    int y_row = row / 2;
-                    int y_col = col / 2;
+                    int y_row = row / 1;
+                    int y_col = col / 1;
                     int y_index = (i * (C * NEW_H * NEW_W)) + (j * (NEW_H * NEW_W)) +
                                   (y_row * NEW_W) + y_col;
                     double sum = 0.0;

kernels/pooling_nchw_sum_d1_s2_3x3/data.h.template

@@ -2,8 +2,8 @@
 
 #define N 1
 #define C 1
-#define H {{(M - 1) * 2 + 3 + 1}}
-#define W {{(N - 1) * 2 + 3 + 1}}
+#define H {{(M - 1) * 1 + 3 + 1}}
+#define W {{(N - 1) * 1 + 3 + 1}}
 #define NEW_H {{M}}
 #define NEW_W {{N}}
 

kernels/pooling_nchw_sum_d1_s2_3x3/gendata.py

@@ -16,7 +16,7 @@ def sum_pool_data(
 
     # Define the pooling parameters
     pool_size = (3, 3)
-    stride = 2
+    stride = 1
 
     new_h = M
     new_w = N
@@ -32,12 +32,15 @@ def sum_pool_data(
     y_in = np.random.uniform(rmin, rmax, (n, c, new_h, new_w))
     y_out = y_in.copy()
 
-    for row in range(0, H - pool_size[0] + 1, stride):
-        for col in range(0, W - pool_size[1] + 1, stride):
-            pooling_region = x[:, :, row : row + pool_size[0], col : col + pool_size[1]]
-            y_out[:, :, row // stride, col // stride] = np.sum(
-                pooling_region, axis=(2, 3)
-            )
+    for row in range(new_h):
+        for col in range(new_w):
+            pooling_region = x[
+                :,
+                :,
+                row * stride : row * stride + pool_size[0],
+                col * stride : col * stride + pool_size[1],
+            ]
+            y_out[:, :, row, col] = np.sum(pooling_region, axis=(2, 3))
 
     yield Define("N", n)
     yield Define("C", c)

kernels/pooling_nchw_sum_d1_s2_3x3/linalg.mlir.template

@@ -1,7 +1,7 @@
-  func.func public @pooling_nchw_sum_d1_s2_3x3(%X : tensor<1x1x{{(M - 1) * 2 + 3 + 1}}x{{(N - 1) * 2 + 3 + 1}}xf64> {"llvm.noalias"}, %Y : tensor<1x1x{{M}}x{{N}}xf64> {"llvm.noalias"}) -> tensor<1x1x{{M}}x{{N}}xf64> {
+  func.func public @pooling_nchw_sum_d1_s2_3x3(%X : tensor<1x1x{{(M - 1) * 1 + 3 + 1}}x{{(N - 1) * 1 + 3 + 1}}xf64> {"llvm.noalias"}, %Y : tensor<1x1x{{M}}x{{N}}xf64> {"llvm.noalias"}) -> tensor<1x1x{{M}}x{{N}}xf64> {
     %zero = arith.constant 0.0 : f64
     %zeros = linalg.fill ins(%zero : f64) outs(%Y : tensor<1x1x{{M}}x{{N}}xf64>) -> tensor<1x1x{{M}}x{{N}}xf64>
     %kernel = tensor.empty() : tensor<3x3xf64>
-    %res = linalg.pooling_nchw_sum {"dilations" = dense<1> : vector<2xi64>, "strides" = dense<2> : vector<2xi64>} ins(%X, %kernel : tensor<1x1x{{(M - 1) * 2 + 3 + 1}}x{{(N - 1) * 2 + 3 + 1}}xf64>, tensor<3x3xf64>) outs(%zeros : tensor<1x1x{{M}}x{{N}}xf64>) -> tensor<1x1x{{M}}x{{N}}xf64>
+    %res = linalg.pooling_nchw_sum {"dilations" = dense<1> : vector<2xi64>, "strides" = dense<1> : vector<2xi64>} ins(%X, %kernel : tensor<1x1x{{(M - 1) * 1 + 3 + 1}}x{{(N - 1) * 1 + 3 + 1}}xf64>, tensor<3x3xf64>) outs(%zeros : tensor<1x1x{{M}}x{{N}}xf64>) -> tensor<1x1x{{M}}x{{N}}xf64>
     func.return %res : tensor<1x1x{{M}}x{{N}}xf64>
   }

results/kernels.csv

@@ -25,9 +25,9 @@ matmul,4x16x8xf64,linalg_xdsl,708,1493,1490,2.811418685121107,0.0,512,578,1625,0
 matmul_transb,4x16x16xf32,baseline,3386,4184,4181,2.539660056657224,1.4921875,0,706,1793,0.20850561134081513,0.3935340022296544,1794,1528,1024,0.5298287064382753,0,64,1.0,1.0,1,0.0,1794,0.5561066336019839,1432,0,0,0.42291789722386297,0,799,0.0,0.9527466036621383,0.0
 matmul_transb,4x16x16xf32,snitch_stream,871,1660,1657,2.648367952522255,0.0,0,674,1785,0.7738231917336394,0.9519774011299436,708,0,0,0.8128587830080367,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.7793427230046949,94,0,0,0.1079219288174512,0,790,0.0,0.9207807118254879,0.0
 matmul_transb,4x16x16xf32,snrt,849,1612,1609,2.648367952522255,0.0,0,674,1785,0.7938751472320377,0.9519774011299436,708,0,0,0.833922261484099,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.8924731182795699,40,0,0,0.04711425206124853,0,764,0.0,0.8810365135453475,0.0
-pooling_nchw_max_d1_s2_3x3,4x4xf64,baseline,442,1213,1210,0.993103448275862,1.1192660550458715,0,145,144,0.32805429864253394,0.5370370370370371,270,122,109,0.6108597285067874,0,16,1.0,1.0,1,0.0,270,0.903010033444816,29,0,0,0.06561085972850679,0,772,0.0,0.6764705882352942,0.0
-pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg_xdsl,275,1023,1020,0.9943820224719101,0.0,0,178,177,0.6472727272727272,0.9888888888888889,180,0,0,0.6545454545454545,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5283018867924528,50,0,0,0.18181818181818182,0,749,0.0,0.8363636363636364,0.0
-pooling_nchw_sum_d1_s2_3x3,4x4xf64,baseline,582,1341,1338,2.9767441860465116,1.1018518518518519,0,129,384,0.22164948453608246,0.5098814229249012,253,119,108,0.43470790378006874,0,16,1.0,1.0,1,0.0,253,0.9730769230769231,7,0,0,0.012027491408934709,0,760,0.0,0.44673539518900346,0.0
+pooling_nchw_max_d1_s2_3x3,4x4xf64,baseline,584,1328,1325,0.995575221238938,1.1226415094339623,0,226,225,0.386986301369863,0.6330532212885154,357,119,106,0.6113013698630136,0,25,1.0,1.0,1,0.0,357,0.9153846153846154,33,0,0,0.05650684931506849,0,745,0.0,0.6678082191780821,0.0
+pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg_xdsl,275,1018,1015,0.9943820224719101,0.0,0,178,177,0.6472727272727272,0.9888888888888889,180,0,0,0.6545454545454545,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5283018867924528,50,0,0,0.18181818181818182,0,744,0.0,0.8363636363636364,0.0
+pooling_nchw_sum_d1_s2_3x3,4x4xf64,baseline,902,1647,1644,2.985074626865672,1.1904761904761905,0,201,600,0.22283813747228381,0.6072507552870091,331,125,105,0.3669623059866962,0,25,1.0,1.0,1,0.0,331,0.914364640883978,31,0,0,0.03436807095343681,0,746,0.0,0.401330376940133,0.0
 pooling_nchw_sum_d1_s2_3x3,4x4xf64,linalg_xdsl,271,1046,1043,2.6797752808988764,0.0,0,178,477,0.6568265682656826,0.9888888888888889,180,0,0,0.6642066420664207,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5384615384615384,48,0,0,0.17712177121771217,0,776,0.0,0.8413284132841328,0.0
 relu,4x4xf64,baseline,142,892,889,0.9444444444444444,1.0,0,18,17,0.1267605633802817,0.36,50,16,16,0.352112676056338,0,16,1.0,1.0,1,0.0,50,0.8771929824561403,7,0,0,0.04929577464788732,0,751,0.0,0.4014084507042253,0.0
 relu,4x4xf64,linalg_xdsl,72,817,814,0.9444444444444444,0.0,0,18,17,0.25,0.9,20,0,0,0.2777777777777778,0,0,3.333333333333333,3.3333333333333335,1,0.0,6,0.25,18,0,0,0.25,0,746,0.0,0.5277777777777778,0.0

results/kernels.fast.csv

@@ -25,9 +25,9 @@ matmul,4x16x8xf64,linalg_xdsl,708,1493,1490,2.811418685121107,0.0,512,578,1625,0
 matmul_transb,4x16x16xf32,baseline,3386,4184,4181,2.539660056657224,1.4921875,0,706,1793,0.20850561134081513,0.3935340022296544,1794,1528,1024,0.5298287064382753,0,64,1.0,1.0,1,0.0,1794,0.5561066336019839,1432,0,0,0.42291789722386297,0,799,0.0,0.9527466036621383,0.0
 matmul_transb,4x16x16xf32,snitch_stream,871,1660,1657,2.648367952522255,0.0,0,674,1785,0.7738231917336394,0.9519774011299436,708,0,0,0.8128587830080367,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.7793427230046949,94,0,0,0.1079219288174512,0,790,0.0,0.9207807118254879,0.0
 matmul_transb,4x16x16xf32,snrt,849,1612,1609,2.648367952522255,0.0,0,674,1785,0.7938751472320377,0.9519774011299436,708,0,0,0.833922261484099,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.8924731182795699,40,0,0,0.04711425206124853,0,764,0.0,0.8810365135453475,0.0
-pooling_nchw_max_d1_s2_3x3,4x4xf64,baseline,442,1213,1210,0.993103448275862,1.1192660550458715,0,145,144,0.32805429864253394,0.5370370370370371,270,122,109,0.6108597285067874,0,16,1.0,1.0,1,0.0,270,0.903010033444816,29,0,0,0.06561085972850679,0,772,0.0,0.6764705882352942,0.0
-pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg_xdsl,275,1023,1020,0.9943820224719101,0.0,0,178,177,0.6472727272727272,0.9888888888888889,180,0,0,0.6545454545454545,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5283018867924528,50,0,0,0.18181818181818182,0,749,0.0,0.8363636363636364,0.0
-pooling_nchw_sum_d1_s2_3x3,4x4xf64,baseline,582,1341,1338,2.9767441860465116,1.1018518518518519,0,129,384,0.22164948453608246,0.5098814229249012,253,119,108,0.43470790378006874,0,16,1.0,1.0,1,0.0,253,0.9730769230769231,7,0,0,0.012027491408934709,0,760,0.0,0.44673539518900346,0.0
+pooling_nchw_max_d1_s2_3x3,4x4xf64,baseline,584,1328,1325,0.995575221238938,1.1226415094339623,0,226,225,0.386986301369863,0.6330532212885154,357,119,106,0.6113013698630136,0,25,1.0,1.0,1,0.0,357,0.9153846153846154,33,0,0,0.05650684931506849,0,745,0.0,0.6678082191780821,0.0
+pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg_xdsl,275,1018,1015,0.9943820224719101,0.0,0,178,177,0.6472727272727272,0.9888888888888889,180,0,0,0.6545454545454545,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5283018867924528,50,0,0,0.18181818181818182,0,744,0.0,0.8363636363636364,0.0
+pooling_nchw_sum_d1_s2_3x3,4x4xf64,baseline,902,1647,1644,2.985074626865672,1.1904761904761905,0,201,600,0.22283813747228381,0.6072507552870091,331,125,105,0.3669623059866962,0,25,1.0,1.0,1,0.0,331,0.914364640883978,31,0,0,0.03436807095343681,0,746,0.0,0.401330376940133,0.0
 pooling_nchw_sum_d1_s2_3x3,4x4xf64,linalg_xdsl,271,1046,1043,2.6797752808988764,0.0,0,178,477,0.6568265682656826,0.9888888888888889,180,0,0,0.6642066420664207,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5384615384615384,48,0,0,0.17712177121771217,0,776,0.0,0.8413284132841328,0.0
 relu,4x4xf64,baseline,142,892,889,0.9444444444444444,1.0,0,18,17,0.1267605633802817,0.36,50,16,16,0.352112676056338,0,16,1.0,1.0,1,0.0,50,0.8771929824561403,7,0,0,0.04929577464788732,0,751,0.0,0.4014084507042253,0.0
 relu,4x4xf64,linalg_xdsl,72,817,814,0.9444444444444444,0.0,0,18,17,0.25,0.9,20,0,0,0.2777777777777778,0,0,3.333333333333333,3.3333333333333335,1,0.0,6,0.25,18,0,0,0.25,0,746,0.0,0.5277777777777778,0.0

results/pivoted.csv

@@ -5,8 +5,8 @@ dense 8x8xf64,3206,3530,,2741,2723
 fill 4x4xf64,50,,63,,
 matmul 4x16x8xf64,2495,,708,,
 matmul_transb 4x16x16xf32,3386,,,871,849
-pooling_nchw_max_d1_s2_3x3 4x4xf64,442,,275,,
-pooling_nchw_sum_d1_s2_3x3 4x4xf64,582,,271,,
+pooling_nchw_max_d1_s2_3x3 4x4xf64,584,,275,,
+pooling_nchw_sum_d1_s2_3x3 4x4xf64,902,,271,,
 relu 4x4xf64,142,,72,,
 relu 4x8xf32,297,210,,67,85
 saxpy 64xf32,634,634,,,140

results/pivoted.fast.csv

@@ -5,8 +5,8 @@ dense 8x8xf64,3206,3530,,2741,2723
 fill 4x4xf64,50,,63,,
 matmul 4x16x8xf64,2495,,708,,
 matmul_transb 4x16x16xf32,3386,,,871,849
-pooling_nchw_max_d1_s2_3x3 4x4xf64,442,,275,,
-pooling_nchw_sum_d1_s2_3x3 4x4xf64,582,,271,,
+pooling_nchw_max_d1_s2_3x3 4x4xf64,584,,275,,
+pooling_nchw_sum_d1_s2_3x3 4x4xf64,902,,271,,
 relu 4x4xf64,142,,72,,
 relu 4x8xf32,297,210,,67,85
 saxpy 64xf32,634,634,,,140

results/pivoted_fpu.csv

@@ -5,7 +5,7 @@ dense 8x8xf64,0.20,0.18,,0.26,0.26
 fill 4x4xf64,0.02,,0.29,,
 matmul 4x16x8xf64,0.21,,0.82,,
 matmul_transb 4x16x16xf32,0.21,,,0.77,0.79
-pooling_nchw_max_d1_s2_3x3 4x4xf64,0.33,,0.65,,
+pooling_nchw_max_d1_s2_3x3 4x4xf64,0.39,,0.65,,
 pooling_nchw_sum_d1_s2_3x3 4x4xf64,0.22,,0.66,,
 relu 4x4xf64,0.13,,0.25,,
 relu 4x8xf32,0.33,0.16,,0.28,0.22

results/pivoted_fpu.fast.csv

@@ -5,7 +5,7 @@ dense 8x8xf64,0.20,0.18,,0.26,0.26
 fill 4x4xf64,0.02,,0.29,,
 matmul 4x16x8xf64,0.21,,0.82,,
 matmul_transb 4x16x16xf32,0.21,,,0.77,0.79
-pooling_nchw_max_d1_s2_3x3 4x4xf64,0.33,,0.65,,
+pooling_nchw_max_d1_s2_3x3 4x4xf64,0.39,,0.65,,
 pooling_nchw_sum_d1_s2_3x3 4x4xf64,0.22,,0.66,,
 relu 4x4xf64,0.13,,0.25,,
 relu 4x8xf32,0.33,0.16,,0.28,0.22

results/pivoted_ipc.csv

@@ -5,8 +5,8 @@ dense 8x8xf64,0.51,0.55,,0.39,0.33
 fill 4x4xf64,0.46,,0.56,,
 matmul 4x16x8xf64,0.56,,0.93,,
 matmul_transb 4x16x16xf32,0.95,,,0.92,0.88
-pooling_nchw_max_d1_s2_3x3 4x4xf64,0.68,,0.84,,
-pooling_nchw_sum_d1_s2_3x3 4x4xf64,0.45,,0.84,,
+pooling_nchw_max_d1_s2_3x3 4x4xf64,0.67,,0.84,,
+pooling_nchw_sum_d1_s2_3x3 4x4xf64,0.40,,0.84,,
 relu 4x4xf64,0.40,,0.53,,
 relu 4x8xf32,0.57,0.51,,0.57,0.40
 saxpy 64xf32,0.93,0.93,,,0.65

results/pivoted_ipc.fast.csv

@@ -5,8 +5,8 @@ dense 8x8xf64,0.51,0.55,,0.39,0.33
 fill 4x4xf64,0.46,,0.56,,
 matmul 4x16x8xf64,0.56,,0.93,,
 matmul_transb 4x16x16xf32,0.95,,,0.92,0.88
-pooling_nchw_max_d1_s2_3x3 4x4xf64,0.68,,0.84,,
-pooling_nchw_sum_d1_s2_3x3 4x4xf64,0.45,,0.84,,
+pooling_nchw_max_d1_s2_3x3 4x4xf64,0.67,,0.84,,
+pooling_nchw_sum_d1_s2_3x3 4x4xf64,0.40,,0.84,,
 relu 4x4xf64,0.40,,0.53,,
 relu 4x8xf32,0.57,0.51,,0.57,0.40
 saxpy 64xf32,0.93,0.93,,,0.65