add scf_xdsl rule and collect cycles

kernels/Makefile

@@ -31,6 +31,8 @@ CONV_8_TESTS += $(CONV_8)/baseline.csv
 CONV_8_TESTS += $(CONV_8)/linalg.csv
 CONV_8_TESTS += $(CONV_8)/snitch_stream.csv
 CONV_8_TESTS += $(CONV_8)/snrt.csv
+# spill to j register...
+# CONV_8_TESTS += $(CONV_8)/scf_xdsl.csv
 
 $(CONV_8)/tests.csv: $(CONV_8_TESTS)
 	./generate_tests_csv.sh $@ $^
@@ -47,6 +49,8 @@ DENSE_8_TESTS += $(DENSE_8)/fused.csv
 DENSE_8_TESTS += $(DENSE_8)/linalg.csv
 DENSE_8_TESTS += $(DENSE_8)/snrt.csv
 DENSE_8_TESTS += $(DENSE_8)/snitch_stream.csv
+# spill to j register...
+# DENSE_8_TESTS += $(DENSE_8)/scf_xdsl.csv
 
 $(DENSE_8)/tests.csv: $(DENSE_8_TESTS)
 	./generate_tests_csv.sh $@ $^
@@ -65,6 +69,7 @@ DSUM_8_16_TESTS += $(DSUM_8_16)/snrt.csv
 DSUM_8_16_TESTS += $(DSUM_8_16)/ssr1d.csv
 DSUM_8_16_TESTS += $(DSUM_8_16)/ssr2d.csv
 DSUM_8_16_TESTS += $(DSUM_8_16)/snitch_stream.csv
+DSUM_8_16_TESTS += $(DSUM_8_16)/scf_xdsl.csv
 
 $(DSUM_8_16)/tests.csv: $(DSUM_8_16_TESTS)
 	./generate_tests_csv.sh $@ $^
@@ -80,6 +85,7 @@ MATMUL_8_TESTS += $(MATMUL_8)/baseline.csv
 MATMUL_8_TESTS += $(MATMUL_8)/linalg.csv
 MATMUL_8_TESTS += $(MATMUL_8)/snitch_stream.csv
 MATMUL_8_TESTS += $(MATMUL_8)/snrt.csv
+MATMUL_8_TESTS += $(MATMUL_8)/scf_xdsl.csv
 
 $(MATMUL_8)/tests.csv: $(MATMUL_8_TESTS)
 	./generate_tests_csv.sh $@ $^
@@ -95,6 +101,7 @@ MAX_POOL_16_TESTS += $(MAX_POOL_16)/baseline.csv
 MAX_POOL_16_TESTS += $(MAX_POOL_16)/linalg.csv
 MAX_POOL_16_TESTS += $(MAX_POOL_16)/snitch_stream.csv
 MAX_POOL_16_TESTS += $(MAX_POOL_16)/snrt.csv
+MAX_POOL_16_TESTS += $(MAX_POOL_16)/scf_xdsl.csv
 
 $(MAX_POOL_16)/tests.csv: $(MAX_POOL_16_TESTS)
 	./generate_tests_csv.sh $@ $^
@@ -110,6 +117,7 @@ SUM_POOL_16_TESTS += $(SUM_POOL_16)/baseline.csv
 SUM_POOL_16_TESTS += $(SUM_POOL_16)/linalg.csv
 SUM_POOL_16_TESTS += $(SUM_POOL_16)/snitch_stream.csv
 SUM_POOL_16_TESTS += $(SUM_POOL_16)/snrt.csv
+SUM_POOL_16_TESTS += $(SUM_POOL_16)/scf_xdsl.csv
 
 $(SUM_POOL_16)/tests.csv: $(SUM_POOL_16_TESTS)
 	./generate_tests_csv.sh $@ $^
@@ -127,6 +135,7 @@ RELU_16_TESTS += $(RELU_16)/ssr_frep.csv
 RELU_16_TESTS += $(RELU_16)/snrt.csv
 RELU_16_TESTS += $(RELU_16)/linalg.csv
 RELU_16_TESTS += $(RELU_16)/snitch_stream.csv
+RELU_16_TESTS += $(RELU_16)/scf_xdsl.csv
 
 $(RELU_16)/tests.csv: $(RELU_16_TESTS)
 	./generate_tests_csv.sh $@ $^
@@ -141,6 +150,7 @@ FILL_16_TESTS += $(FILL_16)/baseline.csv
 FILL_16_TESTS += $(FILL_16)/linalg.csv
 FILL_16_TESTS += $(FILL_16)/snitch_stream.csv
 FILL_16_TESTS += $(FILL_16)/snrt.csv
+FILL_16_TESTS += $(FILL_16)/scf_xdsl.csv
 
 
 $(FILL_16)/tests.csv: $(FILL_16_TESTS)

kernels/dsum/8x16xf32/Makefile

@@ -11,5 +11,6 @@ TESTS += snitch_stream.x
 TESTS += linalg.x
 TESTS += linalg_xdsl.x
 TESTS += scf.x
+TESTS += scf_xdsl.x
 
 include ../../Makefile.kernels

kernels/dsum/8x16xf32/scf_xdsl.csv

@@ -0,0 +1 @@
+3564

kernels/dsum/8x16xf32/scf_xdsl.xdsl.mlir

@@ -0,0 +1,17 @@
+module {
+  func.func public @dsum(%arg0: memref<8x16xf64>, %arg1: memref<8x16xf64>, %arg2: memref<8x16xf64>) -> memref<8x16xf64> {
+    %c0 = arith.constant 0 : index
+    %c8 = arith.constant 8 : index
+    %c1 = arith.constant 1 : index
+    %c16 = arith.constant 16 : index
+    scf.for %arg3 = %c0 to %c8 step %c1 {
+      scf.for %arg4 = %c0 to %c16 step %c1 {
+        %0 = memref.load %arg0[%arg3, %arg4] : memref<8x16xf64>
+        %1 = memref.load %arg1[%arg3, %arg4] : memref<8x16xf64>
+        %2 = arith.addf %0, %1 : f64
+        memref.store %2, %arg2[%arg3, %arg4] : memref<8x16xf64>
+      }
+    }
+    return %arg2 : memref<8x16xf64>
+  }
+}

kernels/dsum/8x16xf32/tests.csv

@@ -5,3 +5,4 @@ snrt,187
 ssr1d,253
 ssr2d,273
 snitch_stream,206
+scf_xdsl,3564

kernels/dsum/params.csv

@@ -5,3 +5,4 @@
 8x16xf32,ssr1d,253
 8x16xf32,ssr2d,273
 8x16xf32,snitch_stream,206
+8x16xf32,scf_xdsl,3564

kernels/fill/16x16xf64/scf_xdsl.csv

@@ -0,0 +1 @@
+2676

kernels/fill/16x16xf64/scf_xdsl.xdsl.mlir

@@ -0,0 +1,13 @@
+module {
+  func.func public @fill(%arg0: f64, %arg1: memref<16x16xf64>) -> memref<16x16xf64> {
+    %c0 = arith.constant 0 : index
+    %c16 = arith.constant 16 : index
+    %c1 = arith.constant 1 : index
+    scf.for %arg2 = %c0 to %c16 step %c1 {
+      scf.for %arg3 = %c0 to %c16 step %c1 {
+        memref.store %arg0, %arg1[%arg2, %arg3] : memref<16x16xf64>
+      }
+    }
+    return %arg1 : memref<16x16xf64>
+  }
+}

kernels/fill/16x16xf64/tests.csv

@@ -2,3 +2,4 @@ baseline,370
 linalg,347
 snitch_stream,296
 snrt,299
+scf_xdsl,2676

kernels/fill/params.csv

@@ -2,3 +2,4 @@
 16x16xf64,linalg,347
 16x16xf64,snitch_stream,296
 16x16xf64,snrt,299
+16x16xf64,scf_xdsl,2676

kernels/kernels.csv

@@ -14,26 +14,32 @@ dsum,8x16xf32,snrt,187
 dsum,8x16xf32,ssr1d,253
 dsum,8x16xf32,ssr2d,273
 dsum,8x16xf32,snitch_stream,206
+dsum,8x16xf32,scf_xdsl,3564
 fill,16x16xf64,baseline,370
 fill,16x16xf64,linalg,347
 fill,16x16xf64,snitch_stream,296
 fill,16x16xf64,snrt,299
+fill,16x16xf64,scf_xdsl,2676
 matmul,8x8xf64,baseline,4230
 matmul,8x8xf64,linalg,6220
 matmul,8x8xf64,snitch_stream,2339
 matmul,8x8xf64,snrt,2321
+matmul,8x8xf64,scf_xdsl,21363
 pooling_nchw_max_d1_s2_3x3,1x16x16x1xf64,baseline,1406
 pooling_nchw_max_d1_s2_3x3,1x16x16x1xf64,linalg,2330
 pooling_nchw_max_d1_s2_3x3,1x16x16x1xf64,snitch_stream,1116
 pooling_nchw_max_d1_s2_3x3,1x16x16x1xf64,snrt,1124
+pooling_nchw_max_d1_s2_3x3,1x16x16x1xf64,scf_xdsl,24729
 pooling_nchw_sum_d1_s2_3x3,1x16x16x1xf64,baseline,1982
 pooling_nchw_sum_d1_s2_3x3,1x16x16x1xf64,linalg,3212
 pooling_nchw_sum_d1_s2_3x3,1x16x16x1xf64,snitch_stream,2012
 pooling_nchw_sum_d1_s2_3x3,1x16x16x1xf64,snrt,2006
+pooling_nchw_sum_d1_s2_3x3,1x16x16x1xf64,scf_xdsl,24729
 relu,16x16xf64,baseline,1339
 relu,16x16xf64,ssr,846
 relu,16x16xf64,ssr_frep,327
 relu,16x16xf64,snrt,334
 relu,16x16xf64,linalg,1335
 relu,16x16xf64,snitch_stream,322
+relu,16x16xf64,scf_xdsl,5020
 softmax,16xf64,baseline,32611

kernels/matmul/8x8xf64/scf_xdsl.csv

@@ -0,0 +1 @@
+21363

kernels/matmul/8x8xf64/scf_xdsl.xdsl.mlir

@@ -0,0 +1,20 @@
+module {
+  func.func public @matmul(%arg0: memref<8x8xf64>, %arg1: memref<8x8xf64>, %arg2: memref<8x8xf64>) -> memref<8x8xf64> {
+    %c0 = arith.constant 0 : index
+    %c8 = arith.constant 8 : index
+    %c1 = arith.constant 1 : index
+    scf.for %arg3 = %c0 to %c8 step %c1 {
+      scf.for %arg4 = %c0 to %c8 step %c1 {
+        scf.for %arg5 = %c0 to %c8 step %c1 {
+          %0 = memref.load %arg0[%arg3, %arg5] : memref<8x8xf64>
+          %1 = memref.load %arg1[%arg5, %arg4] : memref<8x8xf64>
+          %2 = memref.load %arg2[%arg3, %arg4] : memref<8x8xf64>
+          %3 = arith.mulf %0, %1 : f64
+          %4 = arith.addf %2, %3 : f64
+          memref.store %4, %arg2[%arg3, %arg4] : memref<8x8xf64>
+        }
+      }
+    }
+    return %arg2 : memref<8x8xf64>
+  }
+}

kernels/matmul/8x8xf64/tests.csv

@@ -2,3 +2,4 @@ baseline,4230
 linalg,6220
 snitch_stream,2339
 snrt,2321
+scf_xdsl,21363

kernels/matmul/params.csv

@@ -2,3 +2,4 @@
 8x8xf64,linalg,6220
 8x8xf64,snitch_stream,2339
 8x8xf64,snrt,2321
+8x8xf64,scf_xdsl,21363

kernels/pivoted.csv

@@ -1,10 +1,10 @@
-combined,baseline,linalg,linalg_xdsl,snitch_stream,snrt,min_llvm_mlir,speedup
-conv2d_d1_s1_3x3 1x8x8x1xf64,1850,3946,,1500,1498,1850,1.23x
-dense 8x8xf64,3240,7069,,2721,2711,3240,1.19x
-dsum 8x16xf32,1202,1077,194,206,187,1077,5.23x
-fill 16x16xf64,370,347,,296,299,347,1.17x
-matmul 8x8xf64,4230,6220,,2339,2321,4230,1.81x
-pooling_nchw_max_d1_s2_3x3 1x16x16x1xf64,1406,2330,,1116,1124,1406,1.26x
-pooling_nchw_sum_d1_s2_3x3 1x16x16x1xf64,1982,3212,,2012,2006,1982,0.99x
-relu 16x16xf64,1339,1335,,322,334,1335,4.15x
-softmax 16xf64,32611,,,,,32611,?x
+combined,baseline,linalg,linalg_xdsl,scf_xdsl,snitch_stream,snrt,min_llvm_mlir,speedup
+conv2d_d1_s1_3x3 1x8x8x1xf64,1850,3946,,,1500,1498,1850,1.23x
+dense 8x8xf64,3240,7069,,,2721,2711,3240,1.19x
+dsum 8x16xf32,1202,1077,194,3564,206,187,1077,5.23x
+fill 16x16xf64,370,347,,2676,296,299,347,1.17x
+matmul 8x8xf64,4230,6220,,21363,2339,2321,4230,1.81x
+pooling_nchw_max_d1_s2_3x3 1x16x16x1xf64,1406,2330,,24729,1116,1124,1406,1.26x
+pooling_nchw_sum_d1_s2_3x3 1x16x16x1xf64,1982,3212,,24729,2012,2006,1982,0.99x
+relu 16x16xf64,1339,1335,,5020,322,334,1335,4.15x
+softmax 16xf64,32611,,,,,,32611,?x

kernels/pooling_nchw_max_d1_s2_3x3/1x16x16x1xf64/scf_xdsl.csv

@@ -0,0 +1 @@
+24729

kernels/pooling_nchw_max_d1_s2_3x3/1x16x16x1xf64/scf_xdsl.xdsl.mlir

@@ -0,0 +1,26 @@
+module {
+  func.func public @pooling_nchw_max_d1_s2_3x3(%arg0: memref<1x1x16x16xf64>, %arg1: memref<1x1x7x7xf64>) -> memref<1x1x7x7xf64> {
+    %c2 = arith.constant 2 : index
+    %c0 = arith.constant 0 : index
+    %c1 = arith.constant 1 : index
+    %c7 = arith.constant 7 : index
+    %c3 = arith.constant 3 : index
+    scf.for %arg2 = %c0 to %c7 step %c1 {
+      scf.for %arg3 = %c0 to %c7 step %c1 {
+        scf.for %arg4 = %c0 to %c3 step %c1 {
+          scf.for %arg5 = %c0 to %c3 step %c1 {
+            %0 = arith.muli %arg2, %c2 : index
+            %1 = arith.addi %0, %arg4 : index
+            %2 = arith.muli %arg3, %c2 : index
+            %3 = arith.addi %2, %arg5 : index
+            %4 = memref.load %arg0[%c0, %c0, %1, %3] : memref<1x1x16x16xf64>
+            %5 = memref.load %arg1[%c0, %c0, %arg2, %arg3] : memref<1x1x7x7xf64>
+            %6 = arith.maxf %5, %4 : f64
+            memref.store %6, %arg1[%c0, %c0, %arg2, %arg3] : memref<1x1x7x7xf64>
+          }
+        }
+      }
+    }
+    return %arg1 : memref<1x1x7x7xf64>
+  }
+}

kernels/pooling_nchw_max_d1_s2_3x3/1x16x16x1xf64/tests.csv

@@ -2,3 +2,4 @@ baseline,1406
 linalg,2330
 snitch_stream,1116
 snrt,1124
+scf_xdsl,24729

kernels/pooling_nchw_max_d1_s2_3x3/params.csv

@@ -2,3 +2,4 @@
 1x16x16x1xf64,linalg,2330
 1x16x16x1xf64,snitch_stream,1116
 1x16x16x1xf64,snrt,1124
+1x16x16x1xf64,scf_xdsl,24729

kernels/pooling_nchw_sum_d1_s2_3x3/1x16x16x1xf64/scf_xdsl.csv

@@ -0,0 +1 @@
+24729

kernels/pooling_nchw_sum_d1_s2_3x3/1x16x16x1xf64/scf_xdsl.xdsl.mlir

@@ -0,0 +1,26 @@
+module {
+  func.func public @pooling_nchw_sum_d1_s2_3x3(%arg0: memref<1x1x16x16xf64>, %arg1: memref<1x1x7x7xf64>) -> memref<1x1x7x7xf64> {
+    %c2 = arith.constant 2 : index
+    %c0 = arith.constant 0 : index
+    %c1 = arith.constant 1 : index
+    %c7 = arith.constant 7 : index
+    %c3 = arith.constant 3 : index
+    scf.for %arg2 = %c0 to %c7 step %c1 {
+      scf.for %arg3 = %c0 to %c7 step %c1 {
+        scf.for %arg4 = %c0 to %c3 step %c1 {
+          scf.for %arg5 = %c0 to %c3 step %c1 {
+            %0 = arith.muli %arg2, %c2 : index
+            %1 = arith.addi %0, %arg4 : index
+            %2 = arith.muli %arg3, %c2 : index
+            %3 = arith.addi %2, %arg5 : index
+            %4 = memref.load %arg0[%c0, %c0, %1, %3] : memref<1x1x16x16xf64>
+            %5 = memref.load %arg1[%c0, %c0, %arg2, %arg3] : memref<1x1x7x7xf64>
+            %6 = arith.addf %5, %4 : f64
+            memref.store %6, %arg1[%c0, %c0, %arg2, %arg3] : memref<1x1x7x7xf64>
+          }
+        }
+      }
+    }
+    return %arg1 : memref<1x1x7x7xf64>
+  }
+}

kernels/pooling_nchw_sum_d1_s2_3x3/1x16x16x1xf64/tests.csv

@@ -2,3 +2,4 @@ baseline,1982
 linalg,3212
 snitch_stream,2012
 snrt,2006
+scf_xdsl,24729

kernels/pooling_nchw_sum_d1_s2_3x3/params.csv

@@ -2,3 +2,4 @@
 1x16x16x1xf64,linalg,3212
 1x16x16x1xf64,snitch_stream,2012
 1x16x16x1xf64,snrt,2006
+1x16x16x1xf64,scf_xdsl,24729

kernels/relu/16x16xf64/scf_xdsl.csv

@@ -0,0 +1 @@
+5020

kernels/relu/16x16xf64/scf_xdsl.xdsl.mlir

@@ -0,0 +1,16 @@
+module {
+  func.func public @relu(%arg0: memref<16x16xf64>, %arg1: memref<16x16xf64>) -> memref<16x16xf64> {
+    %c0 = arith.constant 0 : index
+    %c16 = arith.constant 16 : index
+    %c1 = arith.constant 1 : index
+    %cst = arith.constant 0.000000e+00 : f64
+    scf.for %arg2 = %c0 to %c16 step %c1 {
+      scf.for %arg3 = %c0 to %c16 step %c1 {
+        %0 = memref.load %arg0[%arg2, %arg3] : memref<16x16xf64>
+        %1 = arith.maxf %0, %cst : f64
+        memref.store %1, %arg1[%arg2, %arg3] : memref<16x16xf64>
+      }
+    }
+    return %arg1 : memref<16x16xf64>
+  }
+}

kernels/relu/16x16xf64/tests.csv

@@ -4,3 +4,4 @@ ssr_frep,327
 snrt,334
 linalg,1335
 snitch_stream,322
+scf_xdsl,5020

kernels/relu/params.csv

@@ -4,3 +4,4 @@
 16x16xf64,snrt,334
 16x16xf64,linalg,1335
 16x16xf64,snitch_stream,322
+16x16xf64,scf_xdsl,5020

scripts/pivot.py

@@ -7,7 +7,7 @@
 
 pivoted = df.pivot(index="combined", columns="impl")["cycles"]
 
-PIVOTED_COLS = set(("linalg", "baseline", "snitch_stream", "snrt", "linalg_xdsl"))
+PIVOTED_COLS = set(("linalg", "baseline", "snitch_stream", "snrt", "linalg_xdsl", "scf_xdsl"))
 
 for col in pivoted:
     if col not in PIVOTED_COLS:

snitch/Makefile.rules

@@ -130,6 +130,10 @@ XDSLOPTFLAGS += -t riscv-asm
 %/linalg_xdsl.xdsl.mlir: %/linalg.mlir
 	$(MLIROPT) $(MLIROPTFLAGS_XDSL) --mlir-print-local-scope -o $@ $<
 
+.PRECIOUS: %/scf_xdsl.xdsl.mlir
+%/scf_xdsl.xdsl.mlir: %/linalg.mlir
+	$(MLIROPT) $(MLIROPTFLAGS_SCF_XDSL) --mlir-print-local-scope -o $@ $<
+
 %.S: %.xdsl.mlir $(XDSL_COMMIT_FILE)
 	$(XDSLOPT) $(XDSLOPTFLAGS) $< -o $@
 
@@ -151,10 +155,12 @@ MLIROPTFLAGS_XDSL += -empty-tensor-to-alloc-tensor
 MLIROPTFLAGS_XDSL += --one-shot-bufferize='bufferize-function-boundaries function-boundary-type-conversion=identity-layout-map'
 MLIROPTFLAGS_XDSL += --canonicalize
 
-MLIROPTFLAGS = $(MLIROPTFLAGS_XDSL)
-MLIROPTFLAGS += --convert-linalg-to-loops
-MLIROPTFLAGS += --lower-affine
-MLIROPTFLAGS += --canonicalize
+MLIROPTFLAGS_SCF_XDSL = $(MLIROPTFLAGS_XDSL)
+MLIROPTFLAGS_SCF_XDSL += --convert-linalg-to-loops
+MLIROPTFLAGS_SCF_XDSL += --lower-affine
+MLIROPTFLAGS_SCF_XDSL += --canonicalize
+
+MLIROPTFLAGS = $(MLIROPTFLAGS_SCF_XDSL)
 MLIROPTFLAGS += --convert-scf-to-cf
 MLIROPTFLAGS += --canonicalize
 MLIROPTFLAGS += --cse
@@ -204,6 +210,9 @@ MLIROPTFLAGS += --reconcile-unrealized-casts
 %/fused.x: %/fused.o %/main.o %/data.o
 	cd $(dir $<); $(LD) $(LDFLAGS) $(notdir $^) -o $(notdir $@)
 
+%/scf_xdsl.x: %/scf_xdsl.o %/main.o %/data.o
+	cd $(dir $<); $(LD) $(LDFLAGS) $(notdir $^) -o $(notdir $@)
+
 # Trace rules
 
 LOG_DIR = $<.logs

xdsl_commit.txt

@@ -1 +1 @@
-d40ad7beed1fc9e9bb8a2cdbba43545645927f7d
+8bddd0aada21c3591a4ab0d325b44116515bf19e