← Back to C-Kernel-Engine Docs Doxygen Source Documentation
 
Loading...
Searching...
No Matches
audio_conv1d_checked.c File Reference
#include "ckernel_audio_conv_checked.h"
#include "checked_float_buffer.h"
#include "audio_conv1d_checked_internal.h"
#include <limits.h>
#include <math.h>
#include <stdint.h>
#include <string.h>

Go to the source code of this file.

Macros

#define CK_AUDIO_CONV1D_COMPUTE_BODY
 
#define CK_AUDIO_CONV_X86_TARGET_FMA   0
 

Functions

int audio_conv1d_checked_avx2_fma_available (void)
 
int audio_conv1d_checked_channel_major_f32 (const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t output_frames)
 
int audio_conv1d_checked_channel_major_f32_avx2_fma (const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t output_frames)
 
int audio_conv1d_checked_channel_major_f32_hwfma (const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t output_frames)
 
int audio_conv1d_checked_hardware_fma_available (void)
 
int audio_conv1d_checked_workspace (size_t input_channels, size_t output_channels, size_t input_frames, size_t output_frames, size_t *scratch_elements)
 
static void audio_conv1d_compute_scalar (const ck_audio_conv1d_job *job, void *unused)
 
int audio_conv1d_dilated_checked_channel_major_f32 (const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t dilation, size_t output_frames)
 
int audio_conv1d_dilated_checked_channel_major_f32_avx2_fma (const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t dilation, size_t output_frames)
 
int audio_conv1d_dilated_checked_channel_major_f32_hwfma (const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t dilation, size_t output_frames)
 
int ck_audio_conv1d_checked_execute (const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t dilation, size_t output_frames, ck_audio_conv1d_compute_fn compute, void *compute_context)
 

Macro Definition Documentation

◆ CK_AUDIO_CONV1D_COMPUTE_BODY

#define CK_AUDIO_CONV1D_COMPUTE_BODY
Value:
for (size_t oc = job->channel_begin; oc < job->channel_end; ++oc) { \
for (size_t frame = 0; frame < job->output_frames; ++frame) { \
float sum = job->bias[oc]; \
for (size_t ic = 0; ic < job->input_channels; ++ic) { \
const float *row = job->weight + \
(oc * job->input_channels + ic) * job->kernel_size; \
for (size_t tap = 0; tap < job->kernel_size; ++tap) { \
int64_t source = (int64_t)frame * job->stride + \
(int64_t)tap * job->dilation - job->padding; \
if (source >= 0 && (size_t)source < job->input_frames) \
sum = fmaf(job->input[ic * job->input_stride + \
(size_t)source], row[tap], sum); \
} \
} \
job->staged[oc * job->output_frames + frame] = sum; \
} \
}

Definition at line 24 of file audio_conv1d_checked.c.

25 { \
26 for (size_t frame = 0; frame < job->output_frames; ++frame) { \
27 float sum = job->bias[oc]; \
28 for (size_t ic = 0; ic < job->input_channels; ++ic) { \
29 const float *row = job->weight + \
30 (oc * job->input_channels + ic) * job->kernel_size; \
31 for (size_t tap = 0; tap < job->kernel_size; ++tap) { \
32 int64_t source = (int64_t)frame * job->stride + \
33 (int64_t)tap * job->dilation - job->padding; \
34 if (source >= 0 && (size_t)source < job->input_frames) \
35 sum = fmaf(job->input[ic * job->input_stride + \
36 (size_t)source], row[tap], sum); \
37 } \
38 } \
39 job->staged[oc * job->output_frames + frame] = sum; \
40 } \
41 }

◆ CK_AUDIO_CONV_X86_TARGET_FMA

#define CK_AUDIO_CONV_X86_TARGET_FMA   0

Definition at line 19 of file audio_conv1d_checked.c.

Function Documentation

◆ audio_conv1d_checked_avx2_fma_available()

int audio_conv1d_checked_avx2_fma_available ( void  )

Definition at line 118 of file audio_conv1d_checked.c.

118 {
119#if CK_AUDIO_CONV_X86_TARGET_FMA
120 __builtin_cpu_init();
121 return __builtin_cpu_supports("avx2") && __builtin_cpu_supports("fma");
122#else
123 return 0;
124#endif
125}

Referenced by audio_conv1d_checked_channel_major_f32_avx2_fma(), audio_conv1d_dilated_checked_channel_major_f32_avx2_fma(), and audio_conv1d_parallel_checked_impl().

◆ audio_conv1d_checked_channel_major_f32()

int audio_conv1d_checked_channel_major_f32 ( const float *  input,
size_t  input_elements,
size_t  input_stride,
const float *  weight,
size_t  weight_elements,
const float *  bias,
size_t  bias_elements,
float *  output,
size_t  output_elements,
size_t  output_stride,
float *  scratch,
size_t  scratch_elements,
size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  kernel_size,
size_t  stride,
size_t  padding,
size_t  output_frames 
)

Definition at line 219 of file audio_conv1d_checked.c.

226 {
227 return ck_audio_conv1d_checked_execute(input, input_elements, input_stride,
228 weight, weight_elements, bias, bias_elements, output, output_elements,
229 output_stride, scratch, scratch_elements, input_channels,
230 output_channels, input_frames, kernel_size, stride, padding, 1,
231 output_frames, audio_conv1d_compute_scalar, NULL);
232}
int ck_audio_conv1d_checked_execute(const float *input, size_t input_elements, size_t input_stride, const float *weight, size_t weight_elements, const float *bias, size_t bias_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t input_channels, size_t output_channels, size_t input_frames, size_t kernel_size, size_t stride, size_t padding, size_t dilation, size_t output_frames, ck_audio_conv1d_compute_fn compute, void *compute_context)
static void audio_conv1d_compute_scalar(const ck_audio_conv1d_job *job, void *unused)

References audio_conv1d_compute_scalar(), and ck_audio_conv1d_checked_execute().

◆ audio_conv1d_checked_channel_major_f32_avx2_fma()

int audio_conv1d_checked_channel_major_f32_avx2_fma ( const float *  input,
size_t  input_elements,
size_t  input_stride,
const float *  weight,
size_t  weight_elements,
const float *  bias,
size_t  bias_elements,
float *  output,
size_t  output_elements,
size_t  output_stride,
float *  scratch,
size_t  scratch_elements,
size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  kernel_size,
size_t  stride,
size_t  padding,
size_t  output_frames 
)

Definition at line 305 of file audio_conv1d_checked.c.

312 {
313#if CK_AUDIO_CONV_X86_TARGET_FMA
316 return ck_audio_conv1d_checked_execute(input, input_elements, input_stride,
317 weight, weight_elements, bias, bias_elements, output, output_elements,
318 output_stride, scratch, scratch_elements, input_channels,
319 output_channels, input_frames, kernel_size, stride, padding, 1,
320 output_frames, audio_conv1d_compute_avx2, NULL);
321#else
322 (void)input; (void)input_elements; (void)input_stride;
323 (void)weight; (void)weight_elements; (void)bias; (void)bias_elements;
324 (void)output; (void)output_elements; (void)output_stride;
325 (void)scratch; (void)scratch_elements; (void)input_channels;
326 (void)output_channels; (void)input_frames; (void)kernel_size;
327 (void)stride; (void)padding; (void)output_frames;
329#endif
330}
int audio_conv1d_checked_avx2_fma_available(void)
void audio_conv1d_compute_avx2(const ck_audio_conv1d_job *job, void *unused)
@ CK_AUDIO_CONV_CHECKED_UNSUPPORTED

References audio_conv1d_checked_avx2_fma_available(), audio_conv1d_compute_avx2(), ck_audio_conv1d_checked_execute(), and CK_AUDIO_CONV_CHECKED_UNSUPPORTED.

◆ audio_conv1d_checked_channel_major_f32_hwfma()

int audio_conv1d_checked_channel_major_f32_hwfma ( const float *  input,
size_t  input_elements,
size_t  input_stride,
const float *  weight,
size_t  weight_elements,
const float *  bias,
size_t  bias_elements,
float *  output,
size_t  output_elements,
size_t  output_stride,
float *  scratch,
size_t  scratch_elements,
size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  kernel_size,
size_t  stride,
size_t  padding,
size_t  output_frames 
)

Definition at line 250 of file audio_conv1d_checked.c.

257 {
258#if CK_AUDIO_CONV_X86_TARGET_FMA
261 return ck_audio_conv1d_checked_execute(input, input_elements, input_stride,
262 weight, weight_elements, bias, bias_elements, output, output_elements,
263 output_stride, scratch, scratch_elements, input_channels,
264 output_channels, input_frames, kernel_size, stride, padding, 1,
265 output_frames, audio_conv1d_compute_hwfma, NULL);
266#else
267 (void)input; (void)input_elements; (void)input_stride;
268 (void)weight; (void)weight_elements; (void)bias; (void)bias_elements;
269 (void)output; (void)output_elements; (void)output_stride;
270 (void)scratch; (void)scratch_elements; (void)input_channels;
271 (void)output_channels; (void)input_frames; (void)kernel_size;
272 (void)stride; (void)padding; (void)output_frames;
274#endif
275}
int audio_conv1d_checked_hardware_fma_available(void)

References audio_conv1d_checked_hardware_fma_available(), ck_audio_conv1d_checked_execute(), and CK_AUDIO_CONV_CHECKED_UNSUPPORTED.

◆ audio_conv1d_checked_hardware_fma_available()

int audio_conv1d_checked_hardware_fma_available ( void  )

Definition at line 109 of file audio_conv1d_checked.c.

109 {
110#if CK_AUDIO_CONV_X86_TARGET_FMA
111 __builtin_cpu_init();
112 return __builtin_cpu_supports("fma") != 0;
113#else
114 return 0;
115#endif
116}

Referenced by audio_conv1d_checked_channel_major_f32_hwfma(), and audio_conv1d_dilated_checked_channel_major_f32_hwfma().

◆ audio_conv1d_checked_workspace()

int audio_conv1d_checked_workspace ( size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  output_frames,
size_t *  scratch_elements 
)

Definition at line 127 of file audio_conv1d_checked.c.

128 {
129 size_t staged;
130 if (!scratch_elements || !input_channels || !output_channels ||
131 !input_frames || !output_frames ||
132 input_channels > SIZE_MAX / input_frames ||
133 output_channels > SIZE_MAX / output_frames)
135 staged = output_channels * output_frames;
136 if (staged > SIZE_MAX / sizeof(float))
138 *scratch_elements = staged;
140}
@ CK_AUDIO_CONV_CHECKED_OVERFLOW
@ CK_AUDIO_CONV_CHECKED_OK

References CK_AUDIO_CONV_CHECKED_OK, and CK_AUDIO_CONV_CHECKED_OVERFLOW.

Referenced by ck_audio_conv1d_checked_execute().

◆ audio_conv1d_compute_scalar()

static void audio_conv1d_compute_scalar ( const ck_audio_conv1d_job *  job,
void *  unused 
)
static

◆ audio_conv1d_dilated_checked_channel_major_f32()

int audio_conv1d_dilated_checked_channel_major_f32 ( const float *  input,
size_t  input_elements,
size_t  input_stride,
const float *  weight,
size_t  weight_elements,
const float *  bias,
size_t  bias_elements,
float *  output,
size_t  output_elements,
size_t  output_stride,
float *  scratch,
size_t  scratch_elements,
size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  kernel_size,
size_t  stride,
size_t  padding,
size_t  dilation,
size_t  output_frames 
)

Definition at line 234 of file audio_conv1d_checked.c.

242 {
243 return ck_audio_conv1d_checked_execute(input, input_elements, input_stride,
244 weight, weight_elements, bias, bias_elements, output, output_elements,
245 output_stride, scratch, scratch_elements, input_channels,
246 output_channels, input_frames, kernel_size, stride, padding, dilation,
247 output_frames, audio_conv1d_compute_scalar, NULL);
248}

References audio_conv1d_compute_scalar(), and ck_audio_conv1d_checked_execute().

◆ audio_conv1d_dilated_checked_channel_major_f32_avx2_fma()

int audio_conv1d_dilated_checked_channel_major_f32_avx2_fma ( const float *  input,
size_t  input_elements,
size_t  input_stride,
const float *  weight,
size_t  weight_elements,
const float *  bias,
size_t  bias_elements,
float *  output,
size_t  output_elements,
size_t  output_stride,
float *  scratch,
size_t  scratch_elements,
size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  kernel_size,
size_t  stride,
size_t  padding,
size_t  dilation,
size_t  output_frames 
)

Definition at line 332 of file audio_conv1d_checked.c.

340 {
341#if CK_AUDIO_CONV_X86_TARGET_FMA
344 return ck_audio_conv1d_checked_execute(input, input_elements, input_stride,
345 weight, weight_elements, bias, bias_elements, output, output_elements,
346 output_stride, scratch, scratch_elements, input_channels,
347 output_channels, input_frames, kernel_size, stride, padding, dilation,
348 output_frames, audio_conv1d_compute_avx2, NULL);
349#else
350 (void)input; (void)input_elements; (void)input_stride;
351 (void)weight; (void)weight_elements; (void)bias; (void)bias_elements;
352 (void)output; (void)output_elements; (void)output_stride;
353 (void)scratch; (void)scratch_elements; (void)input_channels;
354 (void)output_channels; (void)input_frames; (void)kernel_size;
355 (void)stride; (void)padding; (void)dilation; (void)output_frames;
357#endif
358}

References audio_conv1d_checked_avx2_fma_available(), audio_conv1d_compute_avx2(), ck_audio_conv1d_checked_execute(), and CK_AUDIO_CONV_CHECKED_UNSUPPORTED.

◆ audio_conv1d_dilated_checked_channel_major_f32_hwfma()

int audio_conv1d_dilated_checked_channel_major_f32_hwfma ( const float *  input,
size_t  input_elements,
size_t  input_stride,
const float *  weight,
size_t  weight_elements,
const float *  bias,
size_t  bias_elements,
float *  output,
size_t  output_elements,
size_t  output_stride,
float *  scratch,
size_t  scratch_elements,
size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  kernel_size,
size_t  stride,
size_t  padding,
size_t  dilation,
size_t  output_frames 
)

Definition at line 277 of file audio_conv1d_checked.c.

285 {
286#if CK_AUDIO_CONV_X86_TARGET_FMA
289 return ck_audio_conv1d_checked_execute(input, input_elements, input_stride,
290 weight, weight_elements, bias, bias_elements, output, output_elements,
291 output_stride, scratch, scratch_elements, input_channels,
292 output_channels, input_frames, kernel_size, stride, padding, dilation,
293 output_frames, audio_conv1d_compute_hwfma, NULL);
294#else
295 (void)input; (void)input_elements; (void)input_stride;
296 (void)weight; (void)weight_elements; (void)bias; (void)bias_elements;
297 (void)output; (void)output_elements; (void)output_stride;
298 (void)scratch; (void)scratch_elements; (void)input_channels;
299 (void)output_channels; (void)input_frames; (void)kernel_size;
300 (void)stride; (void)padding; (void)dilation; (void)output_frames;
302#endif
303}

References audio_conv1d_checked_hardware_fma_available(), ck_audio_conv1d_checked_execute(), and CK_AUDIO_CONV_CHECKED_UNSUPPORTED.

◆ ck_audio_conv1d_checked_execute()

int ck_audio_conv1d_checked_execute ( const float *  input,
size_t  input_elements,
size_t  input_stride,
const float *  weight,
size_t  weight_elements,
const float *  bias,
size_t  bias_elements,
float *  output,
size_t  output_elements,
size_t  output_stride,
float *  scratch,
size_t  scratch_elements,
size_t  input_channels,
size_t  output_channels,
size_t  input_frames,
size_t  kernel_size,
size_t  stride,
size_t  padding,
size_t  dilation,
size_t  output_frames,
ck_audio_conv1d_compute_fn  compute,
void *  compute_context 
)

Definition at line 142 of file audio_conv1d_checked.c.

151 {
152 size_t input_span, output_span, weight_count, scratch_need;
153 if (!input_channels || !output_channels || !input_frames ||
154 !kernel_size || !stride || !dilation || !output_frames ||
155 input_channels > INT_MAX || output_channels > INT_MAX ||
156 input_frames > INT_MAX || kernel_size > INT_MAX ||
157 stride > INT_MAX || padding > INT_MAX || dilation > INT_MAX ||
158 output_frames > INT_MAX)
160 const int64_t effective_kernel = 1 +
161 ((int64_t)kernel_size - 1) * (int64_t)dilation;
162 const int64_t padded = (int64_t)input_frames +
163 2 * (int64_t)padding - effective_kernel;
164 if (padded < 0 || padded / stride + 1 != output_frames)
166 if (!ck_checked_span((size_t)input_channels, (size_t)input_frames,
167 input_stride, &input_span) ||
168 !ck_checked_span((size_t)output_channels, (size_t)output_frames,
169 output_stride, &output_span) ||
170 audio_conv1d_checked_workspace((size_t)input_channels,
171 (size_t)output_channels, (size_t)input_frames,
172 (size_t)output_frames, &scratch_need))
174 if ((size_t)output_channels > SIZE_MAX / (size_t)input_channels ||
175 (size_t)output_channels * (size_t)input_channels > SIZE_MAX / (size_t)kernel_size)
177 weight_count = (size_t)output_channels * (size_t)input_channels * (size_t)kernel_size;
178 if (weight_count > SIZE_MAX / sizeof(float)) return CK_AUDIO_CONV_CHECKED_OVERFLOW;
179 if (input_elements < input_span || weight_elements < weight_count ||
180 bias_elements < (size_t)output_channels || output_elements < output_span ||
181 scratch_elements < scratch_need)
183 if (!ck_checked_region(input, input_span) ||
184 !ck_checked_region(weight, weight_count) ||
185 !ck_checked_region(bias, (size_t)output_channels) ||
186 !ck_checked_region(output, output_span) ||
187 !ck_checked_region(scratch, scratch_need))
189 if (ck_checked_overlap(output, output_span, input, input_span) ||
190 ck_checked_overlap(output, output_span, weight, weight_count) ||
191 ck_checked_overlap(output, output_span, bias, (size_t)output_channels) ||
192 ck_checked_overlap(output, output_span, scratch, scratch_need) ||
193 ck_checked_overlap(scratch, scratch_need, input, input_span) ||
194 ck_checked_overlap(scratch, scratch_need, weight, weight_count) ||
195 ck_checked_overlap(scratch, scratch_need, bias, (size_t)output_channels))
197 if (!ck_checked_finite(weight, weight_count) ||
198 !ck_checked_finite(bias, (size_t)output_channels))
200 for (size_t channel = 0; channel < input_channels; ++channel)
201 if (!ck_checked_finite(input + (size_t)channel * input_stride,
202 (size_t)input_frames))
204 /* Validation and staged commit remain shared by both providers. */
205 const ck_audio_conv1d_job job = {input, weight, bias, scratch,
206 input_stride, input_channels, output_channels, input_frames,
207 kernel_size, stride, padding, dilation, output_frames, 0,
208 output_channels};
209 compute(&job, compute_context);
210 if (!ck_checked_finite(scratch, (size_t)output_channels * (size_t)output_frames))
212 for (size_t channel = 0; channel < output_channels; ++channel)
213 memcpy(output + (size_t)channel * output_stride,
214 scratch + (size_t)channel * (size_t)output_frames,
215 (size_t)output_frames * sizeof(float));
217}
int audio_conv1d_checked_workspace(size_t input_channels, size_t output_channels, size_t input_frames, size_t output_frames, size_t *scratch_elements)
static int ck_checked_finite(const float *ptr, size_t n)
static int ck_checked_span(size_t rows, size_t width, size_t stride, size_t *span)
static int ck_checked_region(const float *ptr, size_t n)
static int ck_checked_overlap(const float *a, size_t na, const float *b, size_t nb)
@ CK_AUDIO_CONV_CHECKED_CAPACITY
@ CK_AUDIO_CONV_CHECKED_NONFINITE
@ CK_AUDIO_CONV_CHECKED_INVALID

References audio_conv1d_checked_workspace(), CK_AUDIO_CONV_CHECKED_CAPACITY, CK_AUDIO_CONV_CHECKED_INVALID, CK_AUDIO_CONV_CHECKED_NONFINITE, CK_AUDIO_CONV_CHECKED_OK, CK_AUDIO_CONV_CHECKED_OVERFLOW, ck_checked_finite(), ck_checked_overlap(), ck_checked_region(), and ck_checked_span().

Referenced by audio_conv1d_checked_channel_major_f32(), audio_conv1d_checked_channel_major_f32_avx2_fma(), audio_conv1d_checked_channel_major_f32_hwfma(), audio_conv1d_dilated_checked_channel_major_f32(), audio_conv1d_dilated_checked_channel_major_f32_avx2_fma(), audio_conv1d_dilated_checked_channel_major_f32_hwfma(), and audio_conv1d_parallel_checked_impl().