Stash for debugging

candle-core/src/metal_backend.rs

@@ -795,15 +795,16 @@ impl BackendStorage for MetalStorage {
         rhs_l: &Layout,
     ) -> Result<Self> {
         // Create descriptors
-
-        let (type_id, size) = match self.dtype {
+        let (type_id, size, name) = match self.dtype {
             DType::F32 => (
                 metal::mps::MPS_FLOATBIT_ENCODING | 32,
                 core::mem::size_of::<f32>() as NSUInteger,
+                "sgemm",
             ),
             DType::F16 => (
                 metal::mps::MPS_FLOATBIT_ENCODING | 16,
                 core::mem::size_of::<f16>() as NSUInteger,
+                "hgemm",
             ),
             dtype => todo!("Dtype for matmul {dtype:?} is not supported"),
         };
@@ -837,60 +838,37 @@ impl BackendStorage for MetalStorage {
                 mnk: (m, n, k),
             })?
         };
-        let b = b as NSUInteger;
-        let m = m as NSUInteger;
-        let n = n as NSUInteger;
-        let k = k as NSUInteger;
 
-        let left_matrix = self.matrix(
-            (b, m, k),
-            transpose_left,
-            size,
-            lhs_l.start_offset() as NSUInteger * size,
-            type_id,
-        )?;
-        let right_matrix = rhs.matrix(
-            (b, k, n),
-            transpose_right,
-            size,
-            rhs_l.start_offset() as NSUInteger * size,
-            type_id,
-        )?;
-        let (result_matrix, out_buffer) =
-            self.device
-                .new_matrix((b, m, n), size, type_id, self.dtype)?;
+        let result_buffer = self.device.new_buffer(b * m * n, self.dtype);
 
         let command_buffer = self.device.command_buffer();
 
-        let alpha = 1.0f64;
-        let beta = 0.0f64;
-        // Create kernel
-        let matrix_multiplication = MatrixMultiplication::init(
-            &self.device,
+        command_buffer.set_label("mfa gemm");
+
+        candle_metal_kernels::call_mfa_gemm(
+            &self.device.device,
+            &command_buffer,
+            &self.device.kernels,
+            name,
+            &self.buffer,
+            lhs_l.shape().dims(),
+            &rhs.buffer,
+            rhs_l.shape().dims(),
+            &result_buffer,
+            (b, m, n, k),
             transpose_left,
             transpose_right,
-            m,
-            n,
-            k,
-            alpha,
-            beta,
         )
-        .ok_or_else(|| {
-            MetalError::from("Failed to create matrix multiplication kernel".to_string())
-        })?;
+        .map_err(MetalError::from)?;
 
-        // Encode kernel to command buffer
-        matrix_multiplication.encode_to_command_buffer(
-            &command_buffer,
-            &left_matrix,
-            &right_matrix,
-            &result_matrix,
-        );
-        command_buffer.set_label("matmul");
         drop(command_buffer);
         self.device.commit();
 
-        Ok(Self::new(out_buffer, self.device.clone(), self.dtype()))
+        Ok(Self::new(
+            self.buffer.clone(),
+            self.device.clone(),
+            self.dtype(),
+        ))
     }
 
     fn copy_strided_src(&self, dst: &mut Self, dst_offset: usize, src_l: &Layout) -> Result<()> {

candle-metal-kernels/Cargo.toml

@@ -11,6 +11,7 @@ license = "MIT OR Apache-2.0"
 
 [dependencies]
 metal = { version = "0.27.1", features = ["mps"], package="candle-metal" }
+metal-flash-attention = { path = "../../../metal-flash-attention" }
 once_cell = "1.18.0"
 thiserror = "1"
 tracing = "0.1.37"

candle-metal-kernels/src/lib.rs

@@ -1,9 +1,12 @@
 use metal::{
     Buffer, CommandBufferRef, CompileOptions, ComputeCommandEncoderRef, ComputePipelineState,
-    Device, Function, Library, MTLSize,
+    Device, Function, FunctionConstantValues, Library, MTLDataType, MTLResourceUsage, MTLSize,
+    NSUInteger,
 };
-use std::collections::HashMap;
+use std::collections::{BTreeMap, HashMap};
 use std::ffi::c_void;
+use std::hash::Hash;
+use std::io::{stdout, Write};
 use std::sync::RwLock;
 
 const AFFINE: &str = include_str!("affine.metal");
@@ -13,6 +16,7 @@ const BINARY: &str = include_str!("binary.metal");
 const TERNARY: &str = include_str!("ternary.metal");
 const CAST: &str = include_str!("cast.metal");
 const REDUCE: &str = include_str!("reduce.metal");
+const MFA_LIB: &[u8] = include_bytes!("mfa.metallib");
 
 fn linear_split(pipeline: &ComputePipelineState, length: usize) -> (MTLSize, MTLSize) {
     let size = length as u64;
@@ -105,6 +109,7 @@ pub enum Source {
     Ternary,
     Cast,
     Reduce,
+    MetalFlashAttention,
 }
 
 macro_rules! ops{
@@ -179,7 +184,7 @@ impl<T> From<std::sync::PoisonError<T>> for MetalKernelError {
     }
 }
 
-type KernelMap<T> = HashMap<&'static str, T>;
+type KernelMap<T> = HashMap<KernelKey, T>;
 type Libraries = HashMap<Source, Library>;
 type Pipelines = KernelMap<ComputePipelineState>;
 
@@ -189,6 +194,22 @@ pub struct Kernels {
     pipelines: RwLock<Pipelines>,
 }
 
+enum LibraryDefinition {
+    Source(&'static str),
+    Data(&'static [u8]),
+}
+
+impl From<&'static str> for LibraryDefinition {
+    fn from(s: &'static str) -> Self {
+        Self::Source(s)
+    }
+}
+impl From<&'static [u8]> for LibraryDefinition {
+    fn from(s: &'static [u8]) -> Self {
+        Self::Data(s)
+    }
+}
+
 impl Kernels {
     pub fn new() -> Self {
         let libraries = RwLock::new(Libraries::new());
@@ -199,15 +220,16 @@ impl Kernels {
         }
     }
 
-    fn get_library_source(&self, source: Source) -> &'static str {
+    fn get_library_source(&self, source: Source) -> LibraryDefinition {
         match source {
-            Source::Affine => AFFINE,
-            Source::Unary => UNARY,
-            Source::Binary => BINARY,
-            Source::Ternary => TERNARY,
-            Source::Indexing => INDEXING,
-            Source::Cast => CAST,
-            Source::Reduce => REDUCE,
+            Source::Affine => AFFINE.into(),
+            Source::Unary => UNARY.into(),
+            Source::Binary => BINARY.into(),
+            Source::Ternary => TERNARY.into(),
+            Source::Indexing => INDEXING.into(),
+            Source::Cast => CAST.into(),
+            Source::Reduce => REDUCE.into(),
+            Source::MetalFlashAttention => MFA_LIB.into(),
         }
     }
 
@@ -220,10 +242,15 @@ impl Kernels {
         if let Some(lib) = libraries.get(&source) {
             Ok(lib.clone())
         } else {
-            let source_content = self.get_library_source(source);
-            let lib = device
-                .new_library_with_source(source_content, &CompileOptions::new())
-                .map_err(|e| MetalKernelError::LoadLibraryError(e.to_string()))?;
+            let lib = match self.get_library_source(source) {
+                LibraryDefinition::Source(source_content) => device
+                    .new_library_with_source(source_content, &CompileOptions::new())
+                    .map_err(|e| MetalKernelError::LoadLibraryError(e.to_string()))?,
+                LibraryDefinition::Data(data) => device
+                    .new_library_with_data(data)
+                    .map_err(|e| MetalKernelError::LoadLibraryError(e.to_string()))?,
+            };
+
             libraries.insert(source, lib.clone());
             Ok(lib)
         }
@@ -233,43 +260,190 @@ impl Kernels {
         &self,
         device: &Device,
         source: Source,
-        name: &'static str,
+        key: KernelKey,
     ) -> Result<Function, MetalKernelError> {
         let func = self
             .load_library(device, source)?
-            .get_function(name, None)
+            .get_function(
+                key.name,
+                key.constants.map(|c| c.create_function_constant_values()),
+            )
             .map_err(|e| MetalKernelError::LoadFunctionError(e.to_string()))?;
         Ok(func)
-        // let mut funcs = self.funcs.write()?;
-        // if let Some(func) = funcs.get(name) {
-        //     Ok(func.clone())
-        // } else {
-        //     funcs.insert(name, func.clone());
-        //     Ok(func)
-        // }
     }
 
-    pub fn load_pipeline(
+    pub fn load_pipeline<T: Into<KernelKey>>(
         &self,
         device: &Device,
         source: Source,
-        name: &'static str,
+        key: T,
     ) -> Result<ComputePipelineState, MetalKernelError> {
+        let key: KernelKey = key.into();
         let mut pipelines = self.pipelines.write()?;
-        if let Some(pipeline) = pipelines.get(name) {
+        if let Some(pipeline) = pipelines.get(&key) {
             Ok(pipeline.clone())
         } else {
-            let func = self.load_function(device, source, name)?;
+            let func = self.load_function(device, source, key.clone())?;
             let pipeline = device
                 .new_compute_pipeline_state_with_function(&func)
                 .map_err(|e| MetalKernelError::FailedToCreatePipeline(e.to_string()))?;
-            pipelines.insert(name, pipeline.clone());
+            pipelines.insert(key, pipeline.clone());
 
             Ok(pipeline)
         }
     }
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+struct KernelKey {
+    name: &'static str,
+    constants: Option<ConstantValues>,
+}
+
+impl KernelKey {
+    fn new(name: &'static str) -> Self {
+        Self {
+            name,
+            constants: None,
+        }
+    }
+
+    fn with_constants(mut self, constants: ConstantValues) -> Self {
+        self.constants = Some(constants);
+        self
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
+enum ConstantValueId {
+    Index(NSUInteger),
+    Name(&'static str),
+}
+
+trait MetalDType {
+    const MTL_DATA_TYPE: MTLDataType;
+}
+macro_rules! metal_dtype {
+    ($ty:ty, $mtl_data_type:ident) => {
+        impl MetalDType for $ty {
+            const MTL_DATA_TYPE: MTLDataType = MTLDataType::$mtl_data_type;
+        }
+    };
+}
+metal_dtype!(f32, Float);
+metal_dtype!(u32, UInt);
+metal_dtype!(u16, UShort);
+metal_dtype!(bool, Bool);
+
+#[derive(Debug, Clone, PartialEq)]
+enum ConstantValueType {
+    Float(f32),
+    Uint(u32),
+    UShort(u16),
+    Bool(bool),
+}
+
+impl Hash for ConstantValueType {
+    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
+        use ConstantValueType::*;
+        match self {
+            Float(v) => v.to_bits().hash(state),
+            Uint(v) => v.hash(state),
+            UShort(v) => v.hash(state),
+            Bool(v) => v.hash(state),
+        }
+    }
+}
+
+impl Eq for ConstantValueType {}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct ConstantValues(BTreeMap<ConstantValueId, ConstantValueType>);
+
+macro_rules! add_indexed_constant {
+    ($fcv:expr, $value:expr, $ty:ty, $idx:expr) => {
+        $fcv.set_constant_value_at_index(
+            $value as *const $ty as *const c_void,
+            <$ty>::MTL_DATA_TYPE,
+            $idx,
+        )
+    };
+}
+macro_rules! add_named_constant {
+    ($fcv:expr, $value:expr, $ty:ty, $name:expr) => {
+        $fcv.set_constant_value_with_name(
+            $value as *const $ty as *const c_void,
+            <$ty>::MTL_DATA_TYPE,
+            $name,
+        )
+    };
+}
+
+impl Hash for ConstantValues {
+    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
+        for (id, value) in &self.0 {
+            id.hash(state);
+            value.hash(state);
+        }
+    }
+}
+
+impl ConstantValues {
+    fn new() -> Self {
+        Self(BTreeMap::new())
+    }
+
+    fn set(mut self, id: impl Into<ConstantValueId>, value: impl Into<ConstantValueType>) -> Self {
+        self.0.insert(id.into(), value.into());
+        self
+    }
+
+    fn create_function_constant_values(&self) -> FunctionConstantValues {
+        use ConstantValueId::*;
+        use ConstantValueType::*;
+        let mut function_values = FunctionConstantValues::new();
+
+        for (id, value) in &self.0 {
+            match (&id, &value) {
+                (Index(index), Float(value)) => {
+                    add_indexed_constant!(function_values, value, f32, *index);
+                }
+                (Index(index), Uint(value)) => {
+                    add_indexed_constant!(function_values, value, u32, *index);
+                }
+                (Index(index), UShort(value)) => {
+                    add_indexed_constant!(function_values, value, u16, *index);
+                }
+                (Index(index), Bool(value)) => {
+                    add_indexed_constant!(function_values, value, bool, *index);
+                }
+                (Name(name), Float(value)) => {
+                    add_named_constant!(function_values, value, f32, name);
+                }
+                (Name(name), Uint(value)) => {
+                    add_named_constant!(function_values, value, u32, name);
+                }
+                (Name(name), UShort(value)) => {
+                    add_named_constant!(function_values, value, u16, name);
+                }
+                (Name(name), Bool(value)) => {
+                    add_named_constant!(function_values, value, bool, name);
+                }
+            }
+        }
+        function_values
+    }
+}
+
+impl From<&'static str> for KernelKey {
+    fn from(name: &'static str) -> Self {
+        Self {
+            name,
+            constants: None,
+        }
+    }
+}
+
 #[allow(clippy::too_many_arguments)]
 pub fn call_unary_contiguous(
     device: &Device,
@@ -706,5 +880,230 @@ pub fn call_index_select(
     Ok(())
 }
 
+impl From<NSUInteger> for ConstantValueId {
+    fn from(idx: NSUInteger) -> Self {
+        Self::Index(idx)
+    }
+}
+
+impl From<usize> for ConstantValueId {
+    fn from(idx: usize) -> Self {
+        ConstantValueId::from(idx as NSUInteger)
+    }
+}
+
+impl From<i32> for ConstantValueId {
+    fn from(idx: i32) -> Self {
+        ConstantValueId::from(idx as NSUInteger)
+    }
+}
+
+impl From<&'static str> for ConstantValueId {
+    fn from(name: &'static str) -> Self {
+        Self::Name(name)
+    }
+}
+
+macro_rules! to_constant_value {
+    ($ty:ty, $constant_value_type:ident) => {
+        to_constant_value!($ty, $ty, $constant_value_type);
+    };
+    ($ty:ty, $via:ty, $constant_value_type:ident) => {
+        impl From<$ty> for ConstantValueType {
+            fn from(v: $ty) -> Self {
+                Self::$constant_value_type(v as $via)
+            }
+        }
+    };
+}
+to_constant_value!(f32, Float);
+to_constant_value!(u32, Uint);
+to_constant_value!(usize, u32, Uint);
+to_constant_value!(u16, UShort);
+to_constant_value!(bool, Bool);
+
+struct MFAGemmConfig {
+    m: usize,
+    k: usize,
+    n: usize,
+    transpose_left: bool,
+    transpose_right: bool,
+    batched: bool,
+    m_simd: u16,
+    n_simd: u16,
+    k_simd: u16,
+    m_splits: u16,
+    n_splits: u16,
+    m_group: u16,
+    n_group: u16,
+}
+
+impl From<MFAGemmConfig> for ConstantValues {
+    fn from(conf: MFAGemmConfig) -> Self {
+        ConstantValues::new()
+            .set(0, conf.m)
+            .set(1, conf.k)
+            .set(2, conf.n)
+            .set(10, conf.transpose_left)
+            .set(11, conf.transpose_right)
+            .set(12, false)
+            .set(20, 1.0)
+            .set(21, 0.0)
+            .set(100, conf.batched)
+            .set(101, false)
+            .set(50001, false)
+            .set(200, conf.m_simd)
+            .set(201, conf.n_simd)
+            .set(202, conf.k_simd)
+            .set(210, conf.m_splits)
+            .set(211, conf.n_splits)
+            // garbage
+            .set(102, false)
+            .set(103, false)
+            .set(113, false)
+            .set(50000, false)
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+pub fn call_mfa_gemm(
+    device: &Device,
+    command_buffer: &CommandBufferRef,
+    kernels: &Kernels,
+    name: &'static str,
+    lhs: &Buffer,
+    lhs_dims: &[usize],
+    rhs: &Buffer,
+    rhs_dims: &[usize],
+    output: &Buffer,
+    (b, m, n, k): (usize, usize, usize, usize),
+    transpose_left: bool,
+    transpose_right: bool,
+) -> Result<(), MetalKernelError> {
+    let batched = b > 1;
+
+    let mut c_elements = m * n;
+    if batched {
+        c_elements *= 2;
+    }
+
+    let is_half = name == "hgemm";
+    let is_float = name == "sgemm";
+
+    let mut m_group = 32;
+    let mut n_group = 32;
+    let mut k_simd = 32;
+    if c_elements > 10 ^ 6 {
+        m_group = 48;
+        n_group = 48;
+    }
+    // If K_simd is perfectly equal to matrix K, the compiler can elide a large
+    // amount of logic in the kernel.
+    if k >= 33 && k <= 40 {
+        k_simd = 40;
+    } else if is_half && k >= 73 && k >= 80 {
+        k_simd = 80;
+    } else if c_elements > 10 ^ 6 {
+        if k <= 16 {
+            k_simd = 16;
+        } else if k <= 24 {
+            k_simd = 24;
+        } else if k <= 32 {
+            k_simd = 32;
+        } else if k <= 48 {
+            k_simd = 24;
+        } else if k <= 64 {
+            k_simd = 32;
+        } else if is_float {
+            k_simd = 24;
+        }
+    }
+
+    let m_splits = 2;
+    let n_splits = 2;
+    let m_simd = m_group / m_splits;
+    let n_simd = n_group / n_splits;
+
+    let config = MFAGemmConfig {
+        m,
+        k,
+        n,
+        transpose_left,
+        transpose_right,
+        batched,
+        m_simd,
+        n_simd,
+        k_simd,
+        m_splits,
+        n_splits,
+        m_group,
+        n_group,
+    };
+
+    let pipeline = kernels.load_pipeline(
+        device,
+        Source::MetalFlashAttention,
+        KernelKey::new(name).with_constants(config.into()),
+    )?;
+    let block_type_size = if is_half { 2 } else { 4 };
+    let a_block_bytes = m_group * k_simd * block_type_size;
+    let b_block_bytes = k_simd * n_group * block_type_size;
+    let c_block_bytes = m_group * n_group * block_type_size;
+    let mut thread_group_memory_length = a_block_bytes + b_block_bytes;
+
+    if m % 8 > 0 && n % 8 > 0 {
+        thread_group_memory_length = core::cmp::max(thread_group_memory_length, c_block_bytes);
+    }
+
+    let encoder = command_buffer.new_compute_command_encoder();
+    encoder.set_compute_pipeline_state(&pipeline);
+    encoder.set_threadgroup_memory_length(0, thread_group_memory_length as NSUInteger);
+    encoder.use_resources(&[&lhs, &rhs], MTLResourceUsage::Read);
+    encoder.use_resource(&output, MTLResourceUsage::Write);
+    encoder.set_buffers(0, &[Some(lhs), Some(rhs), Some(output)], &[0; 3]);
+
+    let ceil_divide = |a, b| (a + b - 1) / b;
+
+    let mut grid_z = 1;
+
+    if batched {
+        grid_z = lhs_dims[..lhs_dims.len() - 2].iter().product();
+        let byte_stride = |shape: &[usize]| -> u64 {
+            let rank = shape.len();
+            let mut output = core::mem::size_of::<f32>() * shape[rank - 2] * shape[rank - 1];
+            if shape[..shape.len() - 2].iter().product::<usize>() == 1 {
+                output = 0;
+            }
+            output as u64
+        };
+        let byte_stride_a = byte_stride(lhs_dims);
+        let byte_stride_b = byte_stride(rhs_dims);
+        let byte_stride_c = byte_stride(&[m, n]);
+
+        type BatchConfig = (u64, u64, u64, u64);
+        let mut batching_conf: Vec<BatchConfig> = vec![];
+        for i in 0..grid_z {
+            batching_conf.push((
+                i as u64 * byte_stride_a,
+                i as u64 * byte_stride_b,
+                i as u64 * byte_stride_c,
+                0u64,
+            ));
+        }
+        set_param(encoder, 10, batching_conf.as_slice());
+    }
+
+    let grid_size = MTLSize::new(
+        ceil_divide(n as NSUInteger, n_group as NSUInteger),
+        ceil_divide(m as NSUInteger, m_group as NSUInteger),
+        grid_z as NSUInteger,
+    );
+
+    let group_size = MTLSize::new((32 * m_splits * n_splits) as NSUInteger, 1, 1);
+    encoder.dispatch_thread_groups(grid_size, group_size);
+    encoder.end_encoding();
+    Ok(())
+}
+
 #[cfg(test)]
 mod tests;