Merge pull request #801 from xmos/convf32

Convf32
xmos · Aug 18, 2023 · f32acdb · f32acdb
2 parents 1e40e6f + 48957ad
commit f32acdb
Show file tree

Hide file tree

Showing 5 changed files with 73 additions and 5 deletions.
diff --git a/third_party/lib_tflite_micro b/third_party/lib_tflite_micro
diff --git a/xformer/IR/XCoreOps.td b/xformer/IR/XCoreOps.td
@@ -132,6 +132,47 @@ def XC_MulOp : XC_Op<"mul", [Pure]> {
   let results = (outs TensorOf<[QI8]> : $output);
 }
 
+def XC_Beta_ConvF32Op : XC_Op<"beta_convf32", [Pure]> {
+  let summary = "Beta ConvF32 op";
+
+  let description = [{Beta ConvF32 op.}];
+
+  let arguments = (ins
+    TensorOf<[F32]>:$input,
+    TensorOf<[F32]>:$kernels,
+    TensorOf<[F32]>:$bias
+  );
+
+  let results = (outs TensorOf<[F32]> : $output);
+}
+
+def XC_Beta_TransposeConvF32Op : XC_Op<"beta_transposeconvf32", [Pure]> {
+  let summary = "Beta Transpose ConvF32 op";
+
+  let description = [{Beta Transpose ConvF32 op.}];
+
+  let arguments = (ins
+    TensorOf<[F32]>:$input,
+    TensorOf<[F32]>:$kernels,
+    TensorOf<[F32]>:$bias
+  );
+
+  let results = (outs TensorOf<[F32]> : $output);
+}
+
+def XC_Beta_FcF32Op : XC_Op<"beta_fcf32", [Pure]> {
+  let summary = "Beta Fc F32 op";
+
+  let description = [{Beta Fc ConvF32 op.}];
+
+  let arguments = (ins
+    TensorOf<[F32]>:$input,
+    TensorOf<[F32]>:$kernels
+  );
+
+  let results = (outs TensorOf<[F32]> : $output);
+}
+
 def XC_Conv2DV2Op : XC_Op<"conv2d_v2", [Pure]> {
   let summary = "Conv2D V2 op";
 

diff --git a/xformer/Transforms/ConvPatterns.td b/xformer/Transforms/ConvPatterns.td
@@ -25,8 +25,8 @@ def IsConstOp
 
 // TFL_Conv2D() -> XC_FakeConv2D()
 def :
-Pat<(TFL_Conv2DOp: $output TensorOf<[QI8]>:$input, TensorOf<[QI8]>:$f, AnyTypeOf<[TensorOf<[I32,QI32]>, NoneType]>:$b, $dh, $dw, $faf, $p, $sh, $sw),
-          (XC_FakeConv2DOp $input, $f, $b, $dh, $dw, $faf, $p, (CreateNoneValue $input), $sh, $sw),[
+Pat<(TFL_Conv2DOp: $output TensorOf<[QI8]>:$input, TensorOf<[QI8]>:$f, AnyTypeOf<[TensorOf<[I32,QI32]>, NoneType]>:$b, $dh, $dw, $faf, $wf, $sh, $sw),
+          (XC_FakeConv2DOp $input, $f, $b, $dh, $dw, $faf, $wf, (CreateNoneValue $input), $sh, $sw),[
               (HasMultipleOfNBytesPerPixel<4> $input),
               (HasMultipleOfNBytesPerPixel<4> $output),
               (HasEqualChannels $input, $f),
@@ -35,9 +35,24 @@ Pat<(TFL_Conv2DOp: $output TensorOf<[QI8]>:$input, TensorOf<[QI8]>:$f, AnyTypeOf
 
 // TFL_DepthwiseConv2D() -> XC_FakeDepthwiseConv2D()
 def :
-Pat<(TFL_DepthwiseConv2DOp: $output TensorOf<[QI8]>:$input, TensorOf<[QI8]>:$f, TensorOf<[I32,QI32]>:$b, $dh, $dw, $faf, $p, $sh, $sw, $dm),
-          (XC_FakeDepthwiseConv2DOp $input, $f, $b, $dh, $dw, $faf, $p, (CreateNoneValue $input), $sh, $sw, $dm),[
+Pat<(TFL_DepthwiseConv2DOp: $output TensorOf<[QI8]>:$input, TensorOf<[QI8]>:$f, TensorOf<[I32,QI32]>:$b, $dh, $dw, $faf, $wf, $sh, $sw, $dm),
+          (XC_FakeDepthwiseConv2DOp $input, $f, $b, $dh, $dw, $faf, $wf, (CreateNoneValue $input), $sh, $sw, $dm),[
               (HasMultipleOfNBytesPerPixel<4> $input),
               (HasMultipleOfNBytesPerPixel<4> $output),
               (IsConstOp $f),
               ]>;
+
+// F32 TFL_Conv2D() -> XC_Beta_ConvF32()
+def :
+Pat<(TFL_Conv2DOp: $output TensorOf<[F32]>:$input, TensorOf<[F32]>:$f, TensorOf<[F32]>:$b, $dh, $dw, $faf, $wf, $sh, $sw),
+          (XC_Beta_ConvF32Op $input, $f, $b)>;
+
+// F32 TFL_TransposeConv2D() -> XC_Beta_TransposeConvF32()
+def :
+Pat<(TFL_TransposeConvOp: $output $outshape, TensorOf<[F32]>:$f, TensorOf<[F32]>:$input, TensorOf<[F32]>:$b, $wf, $sh, $sw, $faf),
+          (XC_Beta_TransposeConvF32Op $input, $f, $b)>;
+
+// F32 TFL_FullyConnected() -> XC_Beta_FcF32()
+def :
+Pat<(TFL_FullyConnectedOp: $output TensorOf<[F32]>:$input, TensorOf<[F32]>:$f, $b, $faf, $wf, $knd, $aqi),
+          (XC_Beta_FcF32Op $input, $f)>;
diff --git a/xformer/Transforms/TranslateToCustomOp.cpp b/xformer/Transforms/TranslateToCustomOp.cpp
@@ -14,6 +14,11 @@ namespace mlir {
 namespace xcore {
 
 std::vector<uint8_t> Bsign8Op::buildCustomOptions() { return {}; }
+std::vector<uint8_t> Beta_ConvF32Op::buildCustomOptions() { return {}; }
+std::vector<uint8_t> Beta_TransposeConvF32Op::buildCustomOptions() {
+  return {};
+}
+std::vector<uint8_t> Beta_FcF32Op::buildCustomOptions() { return {}; }
 
 std::vector<uint8_t> LookupOp::buildCustomOptions() {
   flexbuffers::Builder fbb;
@@ -171,6 +176,9 @@ void TranslateToCustomOp::runOnOperation() {
   patterns.insert<RewriteToCustomOp<PadOp>>(ctx);
   patterns.insert<RewriteToCustomOp<Pad3To4Op>>(ctx);
   patterns.insert<RewriteToCustomOp<StridedSliceOp>>(ctx);
+  patterns.insert<RewriteToCustomOp<Beta_ConvF32Op>>(ctx);
+  patterns.insert<RewriteToCustomOp<Beta_TransposeConvF32Op>>(ctx);
+  patterns.insert<RewriteToCustomOp<Beta_FcF32Op>>(ctx);
 
   (void)applyPatternsAndFoldGreedily(func, std::move(patterns));
 }

diff --git a/xformer/lib_tflite_micro.BUILD b/xformer/lib_tflite_micro.BUILD
@@ -36,5 +36,9 @@ filegroup(
         "lib_tflite_micro/src/tflite-xcore-kernels/xcore_3_to_4.cc",
         "lib_tflite_micro/src/tflite-xcore-kernels/xcore_strided_slice.cc",
         "lib_tflite_micro/src/tflite-xcore-kernels/xcore_mul.cc",
+        "lib_tflite_micro/src/tflite-xcore-kernels/xcore_beta_convf32.cc",
+        "lib_tflite_micro/src/tflite-xcore-kernels/xcore_beta_transposeconvf32.cc",
+        "lib_tflite_micro/src/tflite-xcore-kernels/xcore_beta_fcf32.cc",
+        "lib_tflite_micro/src/tflite-xcore-kernels/conv2d_float.c",
     ],
 )
+2 −2		Jenkinsfile
+4 −0		cmakefiles/xtflm.cmake
+3 −3		lib_tflite_micro/src/flash_server.xc
+302 −0		lib_tflite_micro/src/tflite-xcore-kernels/conv2d_float.c
+155 −0		lib_tflite_micro/src/tflite-xcore-kernels/conv2d_float.h
+150 −0		lib_tflite_micro/src/tflite-xcore-kernels/xcore_beta_convf32.cc
+132 −0		lib_tflite_micro/src/tflite-xcore-kernels/xcore_beta_fcf32.cc
+150 −0		lib_tflite_micro/src/tflite-xcore-kernels/xcore_beta_transposeconvf32.cc
+8 −4		lib_tflite_micro/src/tflite-xcore-kernels/xcore_conv2d_v2.cc
+5 −0		lib_tflite_micro/src/tflite-xcore-kernels/xcore_ops.cc
+8 −0		lib_tflite_micro/src/tflite-xcore-kernels/xcore_ops.h
+31 −10		tflite_micro_compiler/src/Compiler.cc
+3 −3		tflite_micro_compiler/src/Compiler.h