vkml 0.0.2

High-level Vulkan-based machine learning library
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195

mod f32_cpu;
pub mod push_constants;

use crate::VKMLError;
use crate::gpu::vk_gpu::Gpu;
use crate::instruction::reducemean::f32_cpu::f32_cpu;
use crate::instruction::reducemean::push_constants::ReduceMeanPushConstants;
use crate::instruction::{GPUOperation, Instruction};
use crate::utils::as_bytes;
use crate::{ComputeManager, tensor::TensorDesc, tensor_graph::TensorId};
use onnx_extractor::DataType;
use std::fmt::{Debug, Formatter, Result as FmtResult};
use vulkanalia::vk;

pub struct ReduceMeanInstruction {
    pub src: TensorId,
    pub axes: Option<Vec<i64>>,
    pub keepdims: i64,
    pub noop_with_empty_axes: i64,
    pub dst: TensorId,
}

impl Debug for ReduceMeanInstruction {
    fn fmt(&self, f: &mut Formatter<'_>) -> FmtResult {
        write!(
            f,
            "ReduceMean(src={}, axes={:?}, keepdims={}, dst={})",
            self.src, self.axes, self.keepdims, self.dst
        )
    }
}

impl Instruction for ReduceMeanInstruction {
    fn get_input_tensor_ids(&self) -> Vec<TensorId> {
        vec![self.src]
    }

    fn get_output_tensor_ids(&self) -> Vec<TensorId> {
        vec![self.dst]
    }

    fn remap_tensor_ids(&mut self, new_inputs: &[TensorId], new_outputs: &[TensorId]) {
        if !new_inputs.is_empty() {
            self.src = new_inputs[0];
        }
        if !new_outputs.is_empty() {
            self.dst = new_outputs[0];
        }
    }

    fn record_into_command_buffer(
        &self,
        gpu: &Gpu,
        command_buffer: vk::CommandBuffer,
        cm: &ComputeManager,
    ) -> Result<(), VKMLError> {
        // GPU implementation: two-pass reduction (sum then scale)
        let src_t = cm.tensor_read(self.src);
        let src_mem = src_t.get_gpu_memory_or_panic();
        let dst_t = cm.tensor_read(self.dst);
        let dst_mem = dst_t.get_gpu_memory_or_panic();

        // Determine axes to reduce
        let rank = src_t.desc().ndim() as i64;
        let axes_vec: Vec<i64> = if let Some(a) = &self.axes {
            a.clone()
        } else if self.noop_with_empty_axes != 0 {
            Vec::new()
        } else {
            (0..rank).collect()
        };

        // If noop and empty axes, nothing to do on GPU (copy handled elsewhere)
        if axes_vec.is_empty() && self.noop_with_empty_axes != 0 {
            return Ok(());
        }

        // compute total elements and reduction_size and output elements
        let _total_elements: u64 = src_t.desc().dims().iter().map(|d| *d as u64).product();
        let mut reduction_size: u64 = 1;
        for &a in &axes_vec {
            reduction_size *= src_t.desc().dims()[a as usize] as u64;
        }

        // compute output dims and elements
        let mut out_dims: Vec<i64> = Vec::new();
        for (i, &d) in src_t.desc().dims().iter().enumerate() {
            if axes_vec.contains(&(i as i64)) {
                if self.keepdims != 0 {
                    out_dims.push(1);
                }
            } else {
                out_dims.push(d);
            }
        }
        if out_dims.is_empty() {
            out_dims.push(1);
        }
        let out_elements: u64 = out_dims.iter().map(|d| *d as u64).product();

        let mean_pc = ReduceMeanPushConstants {
            total: out_elements as u32,
            reduction_size: reduction_size as u32,
        };
        let mean_pc_bytes = as_bytes(&mean_pc);

        let src_dtype = src_t.desc().data_type();
        let dst_dtype = dst_t.desc().data_type();

        // Select GPUOperation based on DataType trio (src,dst)
        if src_dtype != dst_dtype {
            return Err(VKMLError::Instruction(format!(
                "GPU ReduceMean unimplemented for DataType src:{:?}, dst:{:?}",
                src_dtype, dst_dtype
            )));
        }

        let gpu_op = GPUOperation::ReduceMean;

        // Choose a local size for dispatch (1D op)
        let local_size = gpu.optimal_workgroup_size_1d(out_elements);

        gpu.bind_slang_compute_pipeline(command_buffer, gpu_op, dst_dtype, local_size);
        gpu.bind_storage_buffers(command_buffer, &[src_mem, dst_mem]);
        gpu.bind_push_constants(command_buffer, gpu_op, mean_pc_bytes);
        gpu.dispatch(command_buffer, local_size, [out_elements, 1, 1]);

        Ok(())
    }

    fn execute_cpu(&self, cm: &ComputeManager) {
        // Basic CPU implementation: compute mean over axes
        let src_t = cm.tensor_read(self.src);
        let src_desc = src_t.desc();
        let src_dims = src_desc.dims().to_vec();
        let rank = src_dims.len() as i64;

        // Determine axes to reduce
        let axes_vec: Vec<i64> = if let Some(a) = &self.axes {
            a.clone()
        } else {
            // No axes => reduce over all axes unless noop_with_empty_axes == 1
            if self.noop_with_empty_axes != 0 {
                Vec::new()
            } else {
                (0..rank).collect()
            }
        };

        // If noop_with_empty_axes==1 and axes empty => copy input to output
        if axes_vec.is_empty() && self.noop_with_empty_axes != 0 {
            let src_bytes = src_t.get_cpu_memory_slice_or_panic();
            let dst_t = cm.tensor_write(self.dst);
            *dst_t.desc_mut() = src_desc.clone();
            dst_t
                .get_cpu_memory_mut_slice_or_panic()
                .copy_from_slice(src_bytes);
            return;
        }

        // Compute output shape
        let keep = self.keepdims != 0;
        let mut out_dims: Vec<i64> = Vec::new();
        for (i, &d) in src_dims.iter().enumerate() {
            if axes_vec.contains(&(i as i64)) {
                if keep {
                    out_dims.push(1);
                }
            } else {
                out_dims.push(d);
            }
        }
        if !keep && out_dims.is_empty() {
            out_dims.push(1); // scalar -> 1-element tensor representation
        }

        // Update dst descriptor
        let dst_t = cm.tensor_write(self.dst);
        *dst_t.desc_mut() = TensorDesc::new(out_dims.clone(), src_desc.data_type());

        let src_bytes = src_t.get_cpu_memory_slice_or_panic();
        let out_bytes = dst_t.get_cpu_memory_mut_slice_or_panic();

        // For simplicity, only implement numeric float32 path
        match src_desc.data_type() {
            DataType::Float => {
                f32_cpu(src_bytes, &src_dims, &axes_vec, keep, out_bytes);
            }
            _ => unimplemented!(
                "ReduceMean CPU: DataType {:?} not implemented",
                src_desc.data_type()
            ),
        }
    }
}