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
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
impl CudaExecutor {
/// CORRECTNESS-011: RoPE NEOX Indirect (CUDA Graph compatible)
///
/// Same as rope_neox_into but reads position from device memory.
/// CORRECTNESS-013: When CORRECTNESS_MODE=1, uses PreciseRopeNeoxIndirect kernel
/// with polynomial sin/cos approximation for CPU-matching precision.
pub fn rope_neox_indirect_into(
&mut self,
input: &GpuBuffer<f32>,
output: &GpuBuffer<f32>,
position_buf: &GpuBuffer<u32>,
num_heads: u32,
head_dim: u32,
theta: f32,
) -> Result<(), GpuError> {
// CORRECTNESS-013: Check if precise mode is requested
static PRECISE_MODE: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
let use_precise = *PRECISE_MODE.get_or_init(|| {
let mode = std::env::var("CORRECTNESS_MODE")
.map(|v| v == "1")
.unwrap_or(false);
if mode {
eprintln!(
"[CORRECTNESS-013] RoPE NEOX using PreciseRopeIndirectKernel (polynomial trig)"
);
}
mode
});
// Choose kernel type based on mode
let (kernel_type, cache_key) = if use_precise {
(
KernelType::PreciseRopeNeoxIndirect {
num_heads,
head_dim,
theta,
},
format!("rope_precise_indirect_{}_{}", num_heads, head_dim),
)
} else {
(
KernelType::RopeNeoxIndirect {
num_heads,
head_dim,
theta,
},
format!("rope_neox_indirect_{}_{}", num_heads, head_dim),
)
};
let kernel_name = self.kernels.kernel_name(&kernel_type);
if !self.modules.contains_key(&cache_key) {
let ptx = self.kernels.generate_ptx(&kernel_type);
let module = self.compile_ptx(&ptx)?;
self.modules.insert(cache_key.clone(), module);
}
let module = self
.modules
.get_mut(&cache_key)
.expect("module just inserted");
// Grid: num_heads blocks, each with half_dim threads
let config = LaunchConfig::grid_2d(num_heads, 1, head_dim / 2, 1);
let mut ptr_input = input.as_ptr();
let mut ptr_output = output.as_ptr();
let mut ptr_position = position_buf.as_ptr();
// SAFETY: Memory safety ensured by bounds checking and alignment
unsafe {
self.stream.launch_kernel(
module,
kernel_name,
&config,
&mut [
std::ptr::from_mut(&mut ptr_input) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut ptr_position) as *mut std::ffi::c_void,
],
)?;
}
Ok(())
}
// =========================================================================
// PAR-023: Host Convenience Methods for Activation Kernels
// =========================================================================
/// PAR-023: SiLU activation with host memory (convenience)
///
/// Uploads input, runs kernel, syncs, downloads result.
pub fn silu_host(&mut self, input: &[f32], output: &mut [f32]) -> Result<(), GpuError> {
let n = input.len() as u32;
let input_gpu = GpuBuffer::from_host(&self.context, input)?;
let output_gpu = self.silu_gpu(&input_gpu, n)?;
self.stream.synchronize()?;
output_gpu.copy_to_host(output)?;
Ok(())
}
/// PAR-023: GELU activation with host memory (convenience)
pub fn gelu_host(&mut self, input: &[f32], output: &mut [f32]) -> Result<(), GpuError> {
let n = input.len() as u32;
let input_gpu = GpuBuffer::from_host(&self.context, input)?;
let output_gpu = self.gelu_async(&input_gpu, n)?;
self.stream.synchronize()?;
output_gpu.copy_to_host(output)?;
Ok(())
}
/// PAR-023: Element-wise multiply with host memory (convenience)
pub fn elementwise_mul_host(
&mut self,
a: &[f32],
b: &[f32],
output: &mut [f32],
) -> Result<(), GpuError> {
let n = a.len() as u32;
let a_gpu = GpuBuffer::from_host(&self.context, a)?;
let b_gpu = GpuBuffer::from_host(&self.context, b)?;
let output_gpu = self.elementwise_mul_gpu(&a_gpu, &b_gpu, n)?;
self.stream.synchronize()?;
output_gpu.copy_to_host(output)?;
Ok(())
}
/// PAR-023: Fused SwiGLU with host memory (convenience)
pub fn fused_swiglu_host(
&mut self,
gate: &[f32],
up: &[f32],
output: &mut [f32],
) -> Result<(), GpuError> {
let n = gate.len() as u32;
let gate_gpu = GpuBuffer::from_host(&self.context, gate)?;
let up_gpu = GpuBuffer::from_host(&self.context, up)?;
let output_gpu = self.fused_swiglu_gpu(&gate_gpu, &up_gpu, n)?;
self.stream.synchronize()?;
output_gpu.copy_to_host(output)?;
Ok(())
}
/// PAR-014: Add two GPU buffers element-wise (residual connection)
///
/// Computes: output[i] += input[i] for all i
/// Uses simple element-wise kernel for residual connections.
pub fn add_residual_gpu(
&mut self,
output: &GpuBuffer<f32>,
input: &GpuBuffer<f32>,
n: u32,
) -> Result<(), GpuError> {
// Use BiasActivation kernel with no activation - it adds "bias" to output
// We repurpose this by treating input as "bias" to add to output
let kernel_type = KernelType::BiasActivation {
n,
bias_size: n, // Same size as output
activation: 0, // No activation
};
let kernel_name = self.kernels.kernel_name(&kernel_type);
let cache_key = format!("residual_{}", n);
if !self.modules.contains_key(&cache_key) {
let ptx = self.kernels.generate_ptx(&kernel_type);
let module = self.compile_ptx(&ptx)?;
self.modules.insert(cache_key.clone(), module);
}
let module = self
.modules
.get_mut(&cache_key)
.expect("module just inserted");
let threads_per_block = 256u32;
let blocks = (n + threads_per_block - 1) / threads_per_block;
let config = LaunchConfig::grid_2d(blocks, 1, threads_per_block, 1);
let mut ptr_output = output.as_ptr();
let mut ptr_input = input.as_ptr();
let mut n_val = n;
// SAFETY: Memory safety ensured by bounds checking and alignment
unsafe {
self.stream.launch_kernel(
module,
kernel_name,
&config,
&mut [
std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut ptr_input) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut n_val) as *mut std::ffi::c_void,
],
)?;
}
// No sync - caller can batch operations
Ok(())
}
/// PAR-014: Q4K GEMV operating on GPU buffers (no CPU round-trip)
///
/// Input and output are GPU-resident buffers. Only weight name lookup uses CPU.
/// Part of persistent GPU tensor optimization for M4 milestone.
pub fn q4k_gemv_gpu(
&mut self,
weight_name: &str,
input: &GpuBuffer<f32>,
output: &GpuBuffer<f32>,
n: u32,
k: u32,
) -> Result<(), GpuError> {
let weight_ptr = self.get_quantized_weight_ptr(weight_name)?;
let kernel_type = KernelType::Q4KGemv { k, n };
let kernel_name = self.kernels.kernel_name(&kernel_type);
let cache_key = format!("q4k_gemv_{}_{}", k, n);
if !self.modules.contains_key(&cache_key) {
let ptx = self.kernels.generate_ptx(&kernel_type);
let module = self.compile_ptx(&ptx)?;
self.modules.insert(cache_key.clone(), module);
}
let module = self
.modules
.get_mut(&cache_key)
.expect("module just inserted");
let config = LaunchConfig::grid_2d(n, 1, 32, 1);
let mut ptr_output = output.as_ptr();
let mut ptr_weights = weight_ptr;
let mut ptr_input = input.as_ptr();
let mut k_val = k;
let mut n_val = n;
// SAFETY: Memory safety ensured by bounds checking and alignment
unsafe {
self.stream.launch_kernel(
module,
kernel_name,
&config,
&mut [
std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut ptr_weights) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut ptr_input) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut k_val) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut n_val) as *mut std::ffi::c_void,
],
)?;
}
// No sync - caller can batch operations
Ok(())
}
/// PAR-094: Tensor Core Q4K GEMM for batched speculative decode
///
/// Enables M>1 batched forward pass with fused dequant+GEMM using tensor cores.
/// Target: 8x speedup over GEMV for M≥16 speculative tokens.
///
/// # Arguments
/// * `weight_name` - Name of cached Q4K weight
/// * `input` - Input activations [M, K] in FP16
/// * `output` - Output buffer [M, N] in FP16
/// * `m` - Batch size (number of tokens)
/// * `k` - Input dimension (must be multiple of 256)
/// * `n` - Output dimension
///
/// # Errors
/// Returns error if weight not cached or kernel launch fails
#[allow(clippy::too_many_arguments)]
pub fn tensor_core_q4k_gemm(
&mut self,
weight_name: &str,
input: &GpuBuffer<f32>,
output: &GpuBuffer<f32>,
m: u32,
k: u32,
n: u32,
) -> Result<(), GpuError> {
let weight_ptr = self.get_quantized_weight_ptr(weight_name)?;
let kernel_type = KernelType::TensorCoreQ4KGemm { m, k, n };
let kernel_name = self.kernels.kernel_name(&kernel_type);
let cache_key = format!("tc_q4k_gemm_{}_{}_{}", m, k, n);
if !self.modules.contains_key(&cache_key) {
let ptx = self.kernels.generate_ptx(&kernel_type);
let module = self.compile_ptx(&ptx)?;
self.modules.insert(cache_key.clone(), module);
}
let module = self
.modules
.get_mut(&cache_key)
.expect("module just inserted");
// Grid: ceil(N/16) x ceil(M/16) blocks, 32 threads per block (1 warp for WMMA)
let grid_x = (n + 15) / 16;
let grid_y = (m + 15) / 16;
let config = LaunchConfig::grid_2d(grid_x, grid_y, 32, 1);
let mut ptr_input = input.as_ptr();
let mut ptr_weights = weight_ptr;
let mut ptr_output = output.as_ptr();
// SAFETY: Memory safety ensured by bounds checking and alignment
unsafe {
self.stream.launch_kernel(
module,
kernel_name,
&config,
&mut [
std::ptr::from_mut(&mut ptr_input) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut ptr_weights) as *mut std::ffi::c_void,
std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
],
)?;
}
// No sync - caller can batch operations
Ok(())
}
/// PAR-095: Tensor Core Q4K GEMM with CPU input/output
///
/// Batched forward pass for speculative decode verification.
/// Input and output are CPU slices; computation uses GPU-resident Q4K weights.
///
/// # Arguments
/// * `weight_name` - Name of cached Q4K weight
/// * `input` - Input activations [M, K] in FP32 (converted to FP16 on GPU)
/// * `output` - Output buffer [M, N] in FP32
/// * `m` - Batch size (number of tokens)
/// * `k` - Input dimension (must be multiple of 256)
/// * `n` - Output dimension
///
/// # Errors
/// Returns error if weight not cached or kernel launch fails
#[allow(clippy::too_many_arguments)]
pub fn tensor_core_q4k_gemm_cached(
&mut self,
weight_name: &str,
input: &[f32],
output: &mut [f32],
m: u32,
k: u32,
n: u32,
) -> Result<(), GpuError> {
// Validate dimensions
let expected_input = (m as usize) * (k as usize);
let expected_output = (m as usize) * (n as usize);
if input.len() != expected_input {
return Err(GpuError::InvalidLaunchConfig(format!(
"PAR-095: Input size {} != expected M*K = {}*{} = {}",
input.len(),
m,
k,
expected_input
)));
}
if output.len() != expected_output {
return Err(GpuError::InvalidLaunchConfig(format!(
"PAR-095: Output size {} != expected M*N = {}*{} = {}",
output.len(),
m,
n,
expected_output
)));
}
// Get cached weight buffer (ALB-098: checks pool first, then individual cache)
let _weight_ptr = self.get_quantized_weight_ptr(weight_name)?;
// Upload input to GPU
let input_buf = GpuBuffer::from_host(&self.context, input)?;
let output_buf = GpuBuffer::new(&self.context, expected_output)?;
// Execute kernel
self.tensor_core_q4k_gemm(weight_name, &input_buf, &output_buf, m, k, n)?;
// Sync and download output
self.stream.synchronize()?;
output_buf.copy_to_host(output)?;
Ok(())
}
}