vkml 0.0.3

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
196
197
198
199
200
201
202
203
204
205

mod f32_f32_f32_cpu;
mod push_constants;

use crate::ComputeManager;
use crate::VKMLError;
use crate::instruction::mul::push_constants::MulPushConstants;
use crate::utils::as_bytes;
use crate::{
    gpu::vk_gpu::Gpu,
    instruction::{
        Instruction, gpu_operations::GPUOperation, mul::f32_f32_f32_cpu::f32_f32_f32_cpu,
    },
    tensor::TensorDesc,
    tensor_graph::TensorId,
};
use onnx_extractor::DataType;
use std::fmt::{Debug, Formatter, Result as FmtResult};
use vulkanalia::vk;

pub struct MulInstruction {
    pub src1: TensorId,
    pub src2: TensorId,
    pub dst: TensorId,
}

impl Debug for MulInstruction {
    fn fmt(&self, f: &mut Formatter<'_>) -> FmtResult {
        write!(
            f,
            "Mul(src1={}, src2={}, dst={})",
            self.src1, self.src2, self.dst
        )
    }
}

impl Instruction for MulInstruction {
    fn get_input_tensor_ids(&self) -> Vec<TensorId> {
        vec![self.src1, self.src2]
    }

    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.len() >= 2 {
            self.src1 = new_inputs[0];
            self.src2 = new_inputs[1];
        }

        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> {
        let src1_tensor = cm.tensor_read(self.src1);
        let src1_mem = src1_tensor.get_gpu_memory_or_panic();
        let src2_tensor = cm.tensor_read(self.src2);
        let src2_mem = src2_tensor.get_gpu_memory_or_panic();
        let dst_tensor = cm.tensor_read(self.dst);
        let dst_mem = dst_tensor.get_gpu_memory_or_panic();

        // Get tensor descriptions for calculating broadcast shapes and strides
        let src1_desc = src1_tensor.desc();
        let src2_desc = src2_tensor.desc();
        let dst_desc = dst_tensor.desc();

        let src1_dims = src1_desc.dims();
        let src2_dims = src2_desc.dims();
        let dst_dims = dst_desc.dims();

        // Prepare push constant data using shared PushConstants

        let rank = dst_dims.len() as u32;
        assert!(
            rank <= 8,
            "Mul: tensor rank {} exceeds maximum supported rank of 8",
            rank
        );

        let mut dims_arr = [0u32; 8];
        for (i, &d) in dst_dims.iter().enumerate().take(8) {
            dims_arr[i] = d as u32;
        }

        // Calculate broadcast shape and strides (similar to execute_cpu)
        let broadcast_dims =
            TensorDesc::broadcast_shape(src1_dims, src2_dims).unwrap_or_else(|| {
                panic!(
                    "GPU Mul: Can't broadcast {:?} vs {:?}",
                    src1_dims, src2_dims
                )
            });

        assert_eq!(
            broadcast_dims.as_slice(),
            dst_dims,
            "GPU Mul: Broadcast shape {:?} != dst shape {:?}",
            broadcast_dims,
            dst_dims
        );

        let strides_a_usize = TensorDesc::broadcast_strides(src1_dims, dst_dims);
        let strides_b_usize = TensorDesc::broadcast_strides(src2_dims, dst_dims);

        let mut strides_a_arr = [0u32; 8];
        for (i, &s) in strides_a_usize.iter().enumerate().take(8) {
            strides_a_arr[i] = s as u32;
        }

        let mut strides_b_arr = [0u32; 8];
        for (i, &s) in strides_b_usize.iter().enumerate().take(8) {
            strides_b_arr[i] = s as u32;
        }

        let total_elements: u64 = dst_dims.iter().map(|d| *d as u64).product();

        let push_const_values = MulPushConstants {
            rank,
            pad: 0, // Padding value, 0 is fine
            total: total_elements as u32,
            dims: dims_arr,
            strides_a: strides_a_arr,
            strides_b: strides_b_arr,
        };

        let push_constant_bytes = as_bytes(&push_const_values);

        let src1_dtype = src1_desc.data_type();
        let src2_dtype = src2_desc.data_type();
        let dst_dtype = dst_desc.data_type();

        if src1_dtype != src2_dtype || src1_dtype != dst_dtype {
            return Err(VKMLError::Instruction(format!(
                "GPU Mul unimplemented for mixed DataType src1:{:?}, src2:{:?}, dst:{:?}",
                src1_dtype, src2_dtype, dst_dtype
            )));
        }

        let op_name = GPUOperation::Multiply;

        // Choose 1D local workgroup size and bind pipeline/descriptors
        let local_size = gpu.optimal_workgroup_size_1d(total_elements);

        gpu.bind_slang_compute_pipeline(command_buffer, op_name, dst_dtype, local_size);
        gpu.bind_storage_buffers(command_buffer, &[src1_mem, src2_mem, dst_mem]);

        // Push constants to the shader
        gpu.bind_push_constants(command_buffer, op_name, push_constant_bytes);

        let num_elements: u64 = dst_dims.iter().map(|d| *d as u64).product();

        gpu.dispatch(command_buffer, local_size, [num_elements, 1, 1]);

        Ok(())
    }

    fn execute_cpu(&self, cm: &ComputeManager) {
        assert!(
            self.src1 != self.dst && self.src2 != self.dst,
            "Cannot use Mul for in-place operation. Use MulInplace instead."
        );

        let src1_tensor = cm.tensor_read(self.src1);
        let src2_tensor = cm.tensor_read(self.src2);
        let dst_tensor = cm.tensor_write(self.dst);

        let a = src1_tensor.desc().dims();
        let b = src2_tensor.desc().dims();
        let c = dst_tensor.desc().dims().to_vec();

        let bc = TensorDesc::broadcast_shape(a, b)
            .unwrap_or_else(|| panic!("Can't broadcast {:?} vs {:?}", a, b));
        assert_eq!(bc, c, "Broadcast {:?} != dst {:?}", bc, c);

        let sa = TensorDesc::broadcast_strides(a, &c);
        let sb = TensorDesc::broadcast_strides(b, &c);

        let src1_dtype = src1_tensor.desc().data_type();
        let src2_dtype = src2_tensor.desc().data_type();
        let dst_dtype = dst_tensor.desc().data_type();

        let src1_bytes = src1_tensor.get_cpu_memory_slice_or_panic();
        let src2_bytes = src2_tensor.get_cpu_memory_slice_or_panic();
        let dst_ptr = dst_tensor.get_cpu_memory_mut_slice_or_panic();

        match (src1_dtype, src2_dtype, dst_dtype) {
            (DataType::Float, DataType::Float, DataType::Float) => {
                f32_f32_f32_cpu(sa, sb, c, src1_bytes, src2_bytes, dst_ptr)
            }
            _ => unimplemented!(
                "mul.rs unimplemented cpu instruction for DataType src1:{:?}, src2:{:?}, dst:{:?}",
                src1_dtype,
                src2_dtype,
                dst_dtype
            ),
        }
    }
}