numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
Documentation 
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
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319

//! CUDA kernel launchers for GEMM epilogue operations.

use cudarc::driver::PushKernelArg;
use cudarc::driver::safe::{CudaContext, CudaStream};
use std::sync::Arc;

use super::super::loader::{
    get_kernel_function, get_or_load_module, kernel_name, matmul_batched_launch_config,
    matmul_launch_config,
};
use crate::algorithm::TileConfig;
use crate::dtype::DType;
use crate::error::{Error, Result};
use crate::ops::GemmActivation;

const GEMM_EPILOGUE_MODULE: &str = "gemm_epilogue";

fn activation_to_u32(activation: GemmActivation) -> u32 {
    match activation {
        GemmActivation::None => 0,
        GemmActivation::ReLU => 1,
        GemmActivation::GELU => 2,
        GemmActivation::SiLU => 3,
        GemmActivation::Sigmoid => 4,
        GemmActivation::Tanh => 5,
    }
}

fn default_tile_config(dtype: DType) -> TileConfig {
    match dtype {
        DType::F64 => TileConfig {
            block_m: 32,
            block_n: 32,
            block_k: 8,
            thread_m: 4,
            thread_n: 4,
        },
        _ => TileConfig {
            block_m: 64,
            block_n: 64,
            block_k: 8,
            thread_m: 8,
            thread_n: 8,
        },
    }
}

/// Launch fused GEMM + bias + activation kernel.
///
/// # Safety
/// All pointers must be valid device memory.
#[allow(clippy::too_many_arguments)]
pub unsafe fn launch_gemm_bias_act_kernel(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    bias_ptr: u64,
    c_ptr: u64,
    m: usize,
    n: usize,
    k: usize,
    activation: GemmActivation,
) -> Result<()> {
    let tile_cfg = default_tile_config(dtype);
    let module = get_or_load_module(context, device_index, GEMM_EPILOGUE_MODULE)?;
    let func_name = kernel_name("gemm_bias_act", dtype);
    let func = get_kernel_function(&module, &func_name)?;

    let elem_size = dtype.size_in_bytes();
    let shared_elem_size = match dtype {
        DType::F16 | DType::BF16 => 4,
        _ => elem_size,
    };

    let cfg = matmul_launch_config(m, n, &tile_cfg, shared_elem_size);
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;
    let act_u32 = activation_to_u32(activation);
    let block_m = tile_cfg.block_m as u32;
    let block_n = tile_cfg.block_n as u32;
    let block_k = tile_cfg.block_k as u32;
    let thread_m = tile_cfg.thread_m as u32;
    let thread_n = tile_cfg.thread_n as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&bias_ptr);
        builder.arg(&c_ptr);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.arg(&act_u32);
        builder.arg(&block_m);
        builder.arg(&block_n);
        builder.arg(&block_k);
        builder.arg(&thread_m);
        builder.arg(&thread_n);

        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!("CUDA gemm_bias_act kernel launch failed: {:?}", e))
        })?;
    }

    Ok(())
}

/// Launch batched fused GEMM + bias + activation kernel.
///
/// # Safety
/// All pointers must be valid device memory.
#[allow(clippy::too_many_arguments)]
pub unsafe fn launch_gemm_bias_act_batched_kernel(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    bias_ptr: u64,
    c_ptr: u64,
    batch: usize,
    m: usize,
    n: usize,
    k: usize,
    activation: GemmActivation,
) -> Result<()> {
    let tile_cfg = default_tile_config(dtype);
    let module = get_or_load_module(context, device_index, GEMM_EPILOGUE_MODULE)?;
    let func_name = kernel_name("gemm_bias_act_batched", dtype);
    let func = get_kernel_function(&module, &func_name)?;

    let elem_size = dtype.size_in_bytes();
    let shared_elem_size = match dtype {
        DType::F16 | DType::BF16 => 4,
        _ => elem_size,
    };

    let cfg = matmul_batched_launch_config(batch, m, n, &tile_cfg, shared_elem_size);
    let batch_u32 = batch as u32;
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;
    let act_u32 = activation_to_u32(activation);
    let block_m = tile_cfg.block_m as u32;
    let block_n = tile_cfg.block_n as u32;
    let block_k = tile_cfg.block_k as u32;
    let thread_m = tile_cfg.thread_m as u32;
    let thread_n = tile_cfg.thread_n as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&bias_ptr);
        builder.arg(&c_ptr);
        builder.arg(&batch_u32);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.arg(&act_u32);
        builder.arg(&block_m);
        builder.arg(&block_n);
        builder.arg(&block_k);
        builder.arg(&thread_m);
        builder.arg(&thread_n);

        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!(
                "CUDA gemm_bias_act_batched kernel launch failed: {:?}",
                e
            ))
        })?;
    }

    Ok(())
}

/// Launch fused GEMM + bias + residual kernel.
///
/// # Safety
/// All pointers must be valid device memory.
#[allow(clippy::too_many_arguments)]
pub unsafe fn launch_gemm_bias_residual_kernel(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    bias_ptr: u64,
    residual_ptr: u64,
    c_ptr: u64,
    m: usize,
    n: usize,
    k: usize,
) -> Result<()> {
    let tile_cfg = default_tile_config(dtype);
    let module = get_or_load_module(context, device_index, GEMM_EPILOGUE_MODULE)?;
    let func_name = kernel_name("gemm_bias_residual", dtype);
    let func = get_kernel_function(&module, &func_name)?;

    let elem_size = dtype.size_in_bytes();
    let shared_elem_size = match dtype {
        DType::F16 | DType::BF16 => 4,
        _ => elem_size,
    };

    let cfg = matmul_launch_config(m, n, &tile_cfg, shared_elem_size);
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;
    let block_m = tile_cfg.block_m as u32;
    let block_n = tile_cfg.block_n as u32;
    let block_k = tile_cfg.block_k as u32;
    let thread_m = tile_cfg.thread_m as u32;
    let thread_n = tile_cfg.thread_n as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&bias_ptr);
        builder.arg(&residual_ptr);
        builder.arg(&c_ptr);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.arg(&block_m);
        builder.arg(&block_n);
        builder.arg(&block_k);
        builder.arg(&thread_m);
        builder.arg(&thread_n);

        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!(
                "CUDA gemm_bias_residual kernel launch failed: {:?}",
                e
            ))
        })?;
    }

    Ok(())
}

/// Launch batched fused GEMM + bias + residual kernel.
///
/// # Safety
/// All pointers must be valid device memory.
#[allow(clippy::too_many_arguments)]
pub unsafe fn launch_gemm_bias_residual_batched_kernel(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    bias_ptr: u64,
    residual_ptr: u64,
    c_ptr: u64,
    batch: usize,
    m: usize,
    n: usize,
    k: usize,
) -> Result<()> {
    let tile_cfg = default_tile_config(dtype);
    let module = get_or_load_module(context, device_index, GEMM_EPILOGUE_MODULE)?;
    let func_name = kernel_name("gemm_bias_residual_batched", dtype);
    let func = get_kernel_function(&module, &func_name)?;

    let elem_size = dtype.size_in_bytes();
    let shared_elem_size = match dtype {
        DType::F16 | DType::BF16 => 4,
        _ => elem_size,
    };

    let cfg = matmul_batched_launch_config(batch, m, n, &tile_cfg, shared_elem_size);
    let batch_u32 = batch as u32;
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;
    let block_m = tile_cfg.block_m as u32;
    let block_n = tile_cfg.block_n as u32;
    let block_k = tile_cfg.block_k as u32;
    let thread_m = tile_cfg.thread_m as u32;
    let thread_n = tile_cfg.thread_n as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&bias_ptr);
        builder.arg(&residual_ptr);
        builder.arg(&c_ptr);
        builder.arg(&batch_u32);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.arg(&block_m);
        builder.arg(&block_n);
        builder.arg(&block_k);
        builder.arg(&thread_m);
        builder.arg(&thread_n);

        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!(
                "CUDA gemm_bias_residual_batched kernel launch failed: {:?}",
                e
            ))
        })?;
    }

    Ok(())
}