Add non-optimized MatMulInteger operator

With this it is possible to run a quantized GPT-2 model ... very slowly.

rten-convert/rten_convert/schema_generated.py

@@ -115,6 +115,7 @@ class OperatorType(object):
     DequantizeLinear = 105
     QuantizeLinear = 106
     DynamicQuantizeLinear = 107
+    MatMulInteger = 108
 
 
 class RNNDirection(object):

src/model_builder.rs

@@ -77,6 +77,7 @@ pub enum OpType<'a> {
     Log,
     LogSoftmax(LogSoftmax),
     MatMul,
+    MatMulInteger,
     Max,
     MaxPool(MaxPool),
     Mean,
@@ -614,6 +615,7 @@ impl<'mb, 'a> GraphBuilder<'mb, 'a> {
                 }
             ),
             OpType::MatMul => op!(MatMul),
+            OpType::MatMulInteger => op!(MatMulInteger),
             OpType::Max => op!(Max),
             OpType::MaxPool(args) => op_with_attrs!(MaxPool, MaxPoolAttrs, {
                 let pad_args = pad_args_from_padding(args.padding);

src/op_registry.rs

@@ -133,6 +133,7 @@ impl OpRegistry {
         register_op!(LogSoftmax);
         register_op!(LSTM);
         register_op!(MatMul);
+        register_op!(MatMulInteger);
         register_op!(Max);
         register_op!(MaxPool);
         register_op!(Mean);
@@ -610,6 +611,7 @@ impl_read_op!(LSTM, attrs_as_lstmattrs, |attrs: sg::LSTMAttrs| {
     })
 });
 impl_read_op!(MatMul);
+impl_read_op!(MatMulInteger);
 impl_read_op!(Max);
 impl_read_op!(
     MaxPool,

src/ops/matmul.rs

@@ -5,6 +5,7 @@ use rten_tensor::{Tensor, TensorView};
 
 use crate::check_dims;
 use crate::gemm::{GemmExecutor, GemmInT, GemmInputA, GemmInputB, GemmOutT};
+use crate::iter_util::range_chunks;
 use crate::ops::binary_elementwise::broadcast_shapes;
 use crate::ops::layout::expand_to;
 use crate::ops::{InputList, IntoOpResult, OpError, Operator, OutputList};
@@ -266,6 +267,117 @@ impl Operator for MatMul {
     }
 }
 
+pub fn matmul_integer(
+    pool: &TensorPool,
+    a: TensorView<u8>,
+    b: TensorView<i8>,
+    a_zero_point: Option<TensorView<u8>>,
+    b_zero_point: Option<TensorView<i8>>,
+) -> Result<Tensor<i32>, OpError> {
+    if a.ndim() < 2 || b.ndim() < 2 {
+        return Err(OpError::InvalidValue("Inputs must have >= 2 dimensions"));
+    }
+
+    let a_rows = a.size(a.ndim() - 2);
+    let a_cols = a.size(a.ndim() - 1);
+
+    let b_rows = b.size(b.ndim() - 2);
+    let b_cols = b.size(b.ndim() - 1);
+
+    if a_cols != b_rows {
+        return Err(OpError::IncompatibleInputShapes(
+            "Columns of first matrix does not match rows of second matrix",
+        ));
+    }
+
+    let a_prefix = &a.shape()[..a.ndim() - 2];
+    let b_prefix = &b.shape()[..b.ndim() - 2];
+
+    let out_prefix = broadcast_shapes(a_prefix, b_prefix)
+        .ok_or(OpError::IncompatibleInputShapes("Cannot broadcast shapes"))?;
+    let out_shape = &[out_prefix.as_slice(), &[a_rows, b_cols]].concat();
+
+    let mut output = Tensor::<i32>::uninit_in(pool, out_shape);
+    if output.is_empty() {
+        // nb. We don't need to alloc from the pool here, since the buffer
+        // is already empty.
+        return Ok(Tensor::zeros(out_shape));
+    }
+
+    let a_broadcast_shape = [out_prefix.as_slice(), &[a_rows, a_cols]].concat();
+    let b_broadcast_shape = [out_prefix.as_slice(), &[b_rows, b_cols]].concat();
+
+    let a_broadcast = a.broadcast(a_broadcast_shape.as_slice());
+    let b_broadcast = b.broadcast(b_broadcast_shape.as_slice());
+
+    let out_row_stride = output.stride(output.ndim() - 2);
+    let out_batches = output
+        .data_mut()
+        .unwrap()
+        .chunks_mut(out_row_stride * a_rows);
+
+    let a_zero = a_zero_point.and_then(|zp| zp.item()).copied().unwrap_or(0) as i32;
+    let b_zero = b_zero_point.and_then(|zp| zp.item()).copied().unwrap_or(0) as i32;
+
+    a_broadcast
+        .inner_iter::<2>()
+        .zip(b_broadcast.inner_iter::<2>())
+        .zip(out_batches)
+        .par_bridge()
+        .for_each(|((a_mat, b_mat), out_mat)| {
+            let [m, k] = a_mat.shape();
+            let [_k, n] = b_mat.shape();
+
+            // Do some extremely rudimentary cache blocking.
+            for col_block in range_chunks(0..n, 32) {
+                for depth_block in range_chunks(0..k, 32) {
+                    for row_block in range_chunks(0..m, 32) {
+                        for j in col_block.clone() {
+                            for i in row_block.clone() {
+                                let mut out = 0i32;
+                                for k in depth_block.clone() {
+                                    let a = unsafe { *a_mat.get_unchecked([i, k]) } as i32 - a_zero;
+                                    let b = unsafe { *b_mat.get_unchecked([k, j]) } as i32 - b_zero;
+                                    out += a * b;
+                                }
+                                unsafe {
+                                    let el = out_mat.get_unchecked_mut((i * out_row_stride) + j);
+                                    if depth_block.start == 0 {
+                                        el.write(out);
+                                    } else {
+                                        el.write(el.assume_init() + out);
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        });
+
+    // Safety: Loop above initialized all output elements.
+    let output = unsafe { output.assume_init() };
+
+    Ok(output)
+}
+
+#[derive(Debug)]
+pub struct MatMulInteger {}
+
+impl Operator for MatMulInteger {
+    fn name(&self) -> &str {
+        "MatMulInteger"
+    }
+
+    fn run(&self, pool: &TensorPool, inputs: InputList) -> Result<OutputList, OpError> {
+        let a = inputs.require_as(0)?;
+        let b = inputs.require_as(1)?;
+        let a_zero_point = inputs.get_as(2)?;
+        let b_zero_point = inputs.get_as(3)?;
+        matmul_integer(pool, a, b, a_zero_point, b_zero_point).into_op_result()
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use std::error::Error;
@@ -277,10 +389,11 @@ mod tests {
     use rten_tensor::{Tensor, TensorView, TensorViewMut};
 
     use crate::gemm::gemm;
+    use crate::ops::binary_elementwise::broadcast_shapes;
     use crate::ops::tests::new_pool;
     use crate::tensor_pool::AutoReturn;
 
-    use super::{gemm_op, matmul, matmul_impl, MatmulStrategy, OpError};
+    use super::{gemm_op, matmul, matmul_impl, matmul_integer, MatmulStrategy, OpError};
 
     fn gemm_tensors(c: &mut Tensor, a: &Tensor, b: &Tensor, alpha: f32, beta: f32) {
         c.make_contiguous();
@@ -323,6 +436,53 @@ mod tests {
             });
     }
 
+    fn reference_matmul_integer(
+        a: TensorView<u8>,
+        b: TensorView<i8>,
+        a_zero_point: Option<TensorView<u8>>,
+        b_zero_point: Option<TensorView<i8>>,
+    ) -> Tensor<i32> {
+        let a_batch_dims = a.ndim() - 2;
+        let b_batch_dims = b.ndim() - 2;
+
+        let a_prefix = &a.shape()[..a.ndim() - 2];
+        let b_prefix = &b.shape()[..b.ndim() - 2];
+        let out_prefix = broadcast_shapes(a_prefix, b_prefix).unwrap();
+        let mut out_shape = out_prefix.to_vec();
+        out_shape.push(a.size(a.ndim() - 2));
+        out_shape.push(b.size(b.ndim() - 1));
+        let mut out = Tensor::<i32>::zeros(&out_shape);
+
+        let a_bcast = [out_prefix.as_slice(), &a.shape()[a_batch_dims..]].concat();
+        let b_bcast = [out_prefix.as_slice(), &b.shape()[b_batch_dims..]].concat();
+
+        let a_zero_point = a_zero_point.and_then(|zp| zp.item()).copied().unwrap_or(0) as i32;
+        let b_zero_point = b_zero_point.and_then(|zp| zp.item()).copied().unwrap_or(0) as i32;
+
+        a.broadcast(a_bcast.as_slice())
+            .inner_iter::<2>()
+            .zip(b.broadcast(b_bcast.as_slice()).inner_iter::<2>())
+            .zip(out.inner_iter_mut::<2>())
+            .for_each(|((a, b), mut c)| {
+                let [n_rows, n_cols] = c.shape();
+                let depth = a.size(1);
+
+                for i in 0..n_rows {
+                    for j in 0..n_cols {
+                        let mut y = 0;
+                        for k in 0..depth {
+                            let a_el = (a[[i, k]] as i32) - a_zero_point;
+                            let b_el = (b[[k, j]] as i32) - b_zero_point;
+                            y += a_el * b_el;
+                        }
+                        c[[i, j]] = y;
+                    }
+                }
+            });
+
+        out
+    }
+
     #[test]
     fn test_gemm_op() -> Result<(), Box<dyn Error>> {
         let pool = new_pool();
@@ -573,6 +733,60 @@ mod tests {
         }
     }
 
+    #[test]
+    fn test_matmul_integer() -> Result<(), Box<dyn Error>> {
+        struct Case {
+            a: Tensor<u8>,
+            b: Tensor<i8>,
+            a_zero_point: Option<Tensor<u8>>,
+            b_zero_point: Option<Tensor<i8>>,
+        }
+
+        let cases = [
+            // No zero point
+            Case {
+                a: Tensor::from([[1, 2], [3, 4]]),
+                b: Tensor::from([[5, 6], [7, 8]]),
+                a_zero_point: None,
+                b_zero_point: None,
+            },
+            // Scalar zero points
+            Case {
+                a: Tensor::from([[1, 2], [3, 4]]),
+                b: Tensor::from([[5, 6], [7, 8]]),
+                a_zero_point: Some(Tensor::from(127)),
+                b_zero_point: Some(Tensor::from(-50)),
+            },
+        ];
+
+        let pool = new_pool();
+
+        for Case {
+            a,
+            b,
+            a_zero_point,
+            b_zero_point,
+        } in cases
+        {
+            let result = matmul_integer(
+                &pool,
+                a.view(),
+                b.view(),
+                a_zero_point.as_ref().map(|zp| zp.view()),
+                b_zero_point.as_ref().map(|zp| zp.view()),
+            )?;
+            let expected = reference_matmul_integer(
+                a.view(),
+                b.view(),
+                a_zero_point.as_ref().map(|zp| zp.view()),
+                b_zero_point.as_ref().map(|zp| zp.view()),
+            );
+            assert_eq!(result, expected);
+        }
+
+        Ok(())
+    }
+
     #[test]
     #[ignore]
     fn bench_matmul() {

src/ops/mod.rs

@@ -82,7 +82,7 @@ pub use layout::{
     expand, flatten, reshape, squeeze, squeeze_in_place, Expand, Flatten, Reshape, Shape, Size,
     Squeeze, Transpose, Unsqueeze,
 };
-pub use matmul::{gemm_op, matmul, Gemm, MatMul};
+pub use matmul::{gemm_op, matmul, Gemm, MatMul, MatMulInteger};
 pub use non_max_suppression::{non_max_suppression, BoxOrder, NonMaxSuppression};
 pub use norm::{
     batch_norm, batch_norm_in_place, instance_normalization, layer_normalization, log_softmax,

src/schema.fbs

@@ -121,6 +121,7 @@ enum OperatorType: ubyte {
   DequantizeLinear,
   QuantizeLinear,
   DynamicQuantizeLinear,
+  MatMulInteger,
 }
 
 enum RNNDirection: ubyte {

src/schema_generated.rs

@@ -18,13 +18,13 @@ pub const ENUM_MIN_OPERATOR_TYPE: u8 = 0;
     since = "2.0.0",
     note = "Use associated constants instead. This will no longer be generated in 2021."
 )]
-pub const ENUM_MAX_OPERATOR_TYPE: u8 = 107;
+pub const ENUM_MAX_OPERATOR_TYPE: u8 = 108;
 #[deprecated(
     since = "2.0.0",
     note = "Use associated constants instead. This will no longer be generated in 2021."
 )]
 #[allow(non_camel_case_types)]
-pub const ENUM_VALUES_OPERATOR_TYPE: [OperatorType; 108] = [
+pub const ENUM_VALUES_OPERATOR_TYPE: [OperatorType; 109] = [
     OperatorType::Add,
     OperatorType::ArgMin,
     OperatorType::ArgMax,
@@ -133,6 +133,7 @@ pub const ENUM_VALUES_OPERATOR_TYPE: [OperatorType; 108] = [
     OperatorType::DequantizeLinear,
     OperatorType::QuantizeLinear,
     OperatorType::DynamicQuantizeLinear,
+    OperatorType::MatMulInteger,
 ];
 
 #[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Default)]
@@ -248,9 +249,10 @@ impl OperatorType {
     pub const DequantizeLinear: Self = Self(105);
     pub const QuantizeLinear: Self = Self(106);
     pub const DynamicQuantizeLinear: Self = Self(107);
+    pub const MatMulInteger: Self = Self(108);
 
     pub const ENUM_MIN: u8 = 0;
-    pub const ENUM_MAX: u8 = 107;
+    pub const ENUM_MAX: u8 = 108;
     pub const ENUM_VALUES: &'static [Self] = &[
         Self::Add,
         Self::ArgMin,
@@ -360,6 +362,7 @@ impl OperatorType {
         Self::DequantizeLinear,
         Self::QuantizeLinear,
         Self::DynamicQuantizeLinear,
+        Self::MatMulInteger,
     ];
     /// Returns the variant's name or "" if unknown.
     pub fn variant_name(self) -> Option<&'static str> {
@@ -472,6 +475,7 @@ impl OperatorType {
             Self::DequantizeLinear => Some("DequantizeLinear"),
             Self::QuantizeLinear => Some("QuantizeLinear"),
             Self::DynamicQuantizeLinear => Some("DynamicQuantizeLinear"),
+            Self::MatMulInteger => Some("MatMulInteger"),
             _ => None,
         }
     }