Bug Summary

File:nnc/cmd/scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c
Warning:line 132, column 28
Array access (from variable 'amp2') results in a null pointer dereference

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-unknown-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name ccv_nnc_scaled_dot_product_attention_cpu_ref.c -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model pic -pic-level 2 -pic-is-pie -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -target-feature +sse2 -tune-cpu generic -debugger-tuning=gdb -fdebug-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc/cmd -fcoverage-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc/cmd -resource-dir /usr/local/lib/clang/19 -I ../../ -I .. -I /usr/local/cuda/include -D HAVE_CBLAS -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_PTHREAD -D HAVE_LIBLINEAR -D HAVE_TESSERACT -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_AVUTIL -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_GSL -D HAVE_CUDA -D HAVE_CUDNN -D HAVE_NCCL -D USE_SYSTEM_CUB -I /usr/local/include -internal-isystem /usr/local/lib/clang/19/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/12/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O3 -ferror-limit 19 -fgnuc-version=4.2.1 -fskip-odr-check-in-gmf -vectorize-loops -vectorize-slp -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/liu/actions-runner/_work/ccv/ccv/_analyze/2025-11-03-220905-61206-1 -x c scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c
1#include "ccv.h"
2#include "ccv_internal.h"
3#include "nnc/ccv_nnc.h"
4#include "nnc/ccv_nnc_easy.h"
5#include "nnc/ccv_nnc_internal.h"
6#ifdef USE_OPENMP
7#include <omp.h>
8#endif
9#ifdef USE_DISPATCH
10#include <dispatch/dispatch.h>
11#endif
12
13// Shared methods.
14#include "../_ccv_nnc_cpu_ref.h"
15
16static int _ccv_nnc_scaled_dot_product_attention_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
17{
18 assert(input_size >= 3)((void) sizeof ((input_size >= 3) ? 1 : 0), __extension__ (
{ if (input_size >= 3) ; else __assert_fail ("input_size >= 3"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 18, __extension__ __PRETTY_FUNCTION__); }));
1
Assuming 'input_size' is >= 3
→
2
←
Taking true branch
→
19 assert(output_size >= 1)((void) sizeof ((output_size >= 1) ? 1 : 0), __extension__
({ if (output_size >= 1) ; else __assert_fail ("output_size >= 1"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 19, __extension__ __PRETTY_FUNCTION__); }));
3
←
Assuming 'output_size' is >= 1
→
4
←
Taking true branch
→
20 ccv_nnc_tensor_view_t* const q = (ccv_nnc_tensor_view_t*)inputs[0];
21 ccv_nnc_tensor_view_t* const k = (ccv_nnc_tensor_view_t*)inputs[1];
22 ccv_nnc_tensor_view_t* const v = (ccv_nnc_tensor_view_t*)inputs[2];
23 ccv_nnc_tensor_view_t* const attn_mask = input_size > 3 ? (ccv_nnc_tensor_view_t*)inputs[3] : 0;
5
←
Assuming 'input_size' is > 3
→
6
←
'?' condition is true
→
24 ccv_nnc_tensor_view_t* const w = input_size > 4 ? (ccv_nnc_tensor_view_t*)inputs[4] : 0;
7
←
Assuming 'input_size' is > 4
→
8
←
'?' condition is true
→
25 ccv_nnc_tensor_view_t* const bias = input_size > 5 ? (ccv_nnc_tensor_view_t*)inputs[5] : 0;
9
←
Assuming 'input_size' is > 5
→
10
←
'?' condition is true
→
26 if (bias) // bias always requires a weight matrix.
11
←
Assuming 'bias' is null
→
27 { assert(w)((void) sizeof ((w) ? 1 : 0), __extension__ ({ if (w) ; else __assert_fail
("w", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 27, __extension__ __PRETTY_FUNCTION__); })); }
28 ccv_nnc_tensor_view_t* const c = (w) ? (ccv_nnc_tensor_view_t*)outputs[2] : (ccv_nnc_tensor_view_t*)outputs[0];
12
←
Taking false branch
→
13
←
Assuming 'w' is null
→
14
←
'?' condition is false
→
29 const int q_nd = ccv_nnc_tensor_nd(q->info.dim);
30 assert(q_nd == 3 || q_nd == 4)((void) sizeof ((q_nd == 3 || q_nd == 4) ? 1 : 0), __extension__
({ if (q_nd == 3 || q_nd == 4) ; else __assert_fail ("q_nd == 3 || q_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 30, __extension__ __PRETTY_FUNCTION__); }));
15
←
Assuming 'q_nd' is not equal to 3
→
16
←
Assuming 'q_nd' is equal to 4
→
17
←
Taking true branch
→
31 const int k_nd = ccv_nnc_tensor_nd(k->info.dim);
32 assert(k_nd == 3 || k_nd == 4)((void) sizeof ((k_nd == 3 || k_nd == 4) ? 1 : 0), __extension__
({ if (k_nd == 3 || k_nd == 4) ; else __assert_fail ("k_nd == 3 || k_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 32, __extension__ __PRETTY_FUNCTION__); }));
18
←
Assuming 'k_nd' is not equal to 3
→
19
←
Assuming 'k_nd' is equal to 4
→
20
←
Taking true branch
→
33 const int v_nd = ccv_nnc_tensor_nd(v->info.dim);
34 assert(v_nd == 3 || v_nd == 4)((void) sizeof ((v_nd == 3 || v_nd == 4) ? 1 : 0), __extension__
({ if (v_nd == 3 || v_nd == 4) ; else __assert_fail ("v_nd == 3 || v_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 34, __extension__ __PRETTY_FUNCTION__); }));
21
←
Assuming 'v_nd' is not equal to 3
→
22
←
Assuming 'v_nd' is equal to 4
→
23
←
Taking true branch
→
35 const int c_nd = ccv_nnc_tensor_nd(c->info.dim);
36 assert(c_nd == 3 || c_nd == 4)((void) sizeof ((c_nd == 3 || c_nd == 4) ? 1 : 0), __extension__
({ if (c_nd == 3 || c_nd == 4) ; else __assert_fail ("c_nd == 3 || c_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 36, __extension__ __PRETTY_FUNCTION__); }));
24
←
Assuming 'c_nd' is not equal to 3
→
25
←
Assuming 'c_nd' is equal to 4
→
26
←
Taking true branch
→
37 assert(q_nd == k_nd && k_nd == v_nd && v_nd == c_nd)((void) sizeof ((q_nd == k_nd && k_nd == v_nd &&
v_nd == c_nd) ? 1 : 0), __extension__ ({ if (q_nd == k_nd &&
k_nd == v_nd && v_nd == c_nd) ; else __assert_fail (
"q_nd == k_nd && k_nd == v_nd && v_nd == c_nd"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 37, __extension__ __PRETTY_FUNCTION__); }));
27
←
Taking true branch
→
38 // Assuming this is float 32.
39 int qdim[CCV_NNC_MAX_DIM_ALLOC(12)];
40 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
41 int vdim[CCV_NNC_MAX_DIM_ALLOC(12)];
42 int cdim[CCV_NNC_MAX_DIM_ALLOC(12)];
43 int amdim[CCV_NNC_MAX_DIM_ALLOC(12)];
44 ccv_nnc_tensor_view_get_dim(q, qdim);
45 ccv_nnc_tensor_view_get_dim(k, kdim);
46 ccv_nnc_tensor_view_get_dim(v, vdim);
47 ccv_nnc_tensor_view_get_dim(c, cdim);
48 if (q_nd
27.1
'q_nd' is not equal to 3
 == 3)
28
←
Taking false branch
→
49 {
50 qdim[0] = qdim[1], qdim[1] = qdim[2], qdim[2] = 1;
51 kdim[0] = kdim[1], kdim[1] = kdim[2], kdim[2] = 1;
52 vdim[0] = vdim[1], vdim[1] = vdim[2], vdim[2] = 1;
53 cdim[0] = cdim[1], cdim[1] = cdim[2], cdim[2] = 1;
54 }
55 assert(qdim[0] == kdim[0] && kdim[0] == vdim[0] && vdim[0] == cdim[0])((void) sizeof ((qdim[0] == kdim[0] && kdim[0] == vdim
[0] && vdim[0] == cdim[0]) ? 1 : 0), __extension__ ({
if (qdim[0] == kdim[0] && kdim[0] == vdim[0] &&
vdim[0] == cdim[0]) ; else __assert_fail ("qdim[0] == kdim[0] && kdim[0] == vdim[0] && vdim[0] == cdim[0]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 55, __extension__ __PRETTY_FUNCTION__); }));
29
←
Assuming the condition is true
→
30
←
Assuming the condition is true
→
31
←
Assuming the condition is true
→
32
←
Taking true branch
→
56 assert(qdim[2] == cdim[2])((void) sizeof ((qdim[2] == cdim[2]) ? 1 : 0), __extension__ (
{ if (qdim[2] == cdim[2]) ; else __assert_fail ("qdim[2] == cdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 56, __extension__ __PRETTY_FUNCTION__); }));
33
←
Assuming the condition is true
→
34
←
Taking true branch
→
57 assert(kdim[2] == vdim[2])((void) sizeof ((kdim[2] == vdim[2]) ? 1 : 0), __extension__ (
{ if (kdim[2] == vdim[2]) ; else __assert_fail ("kdim[2] == vdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 57, __extension__ __PRETTY_FUNCTION__); }));
35
←
Assuming the condition is true
→
36
←
Taking true branch
→
58 assert(qdim[2] % kdim[2] == 0)((void) sizeof ((qdim[2] % kdim[2] == 0) ? 1 : 0), __extension__
({ if (qdim[2] % kdim[2] == 0) ; else __assert_fail ("qdim[2] % kdim[2] == 0"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 58, __extension__ __PRETTY_FUNCTION__); }));
37
←
Assuming the condition is true
→
38
←
Taking true branch
→
59 assert(qdim[2] >= kdim[2])((void) sizeof ((qdim[2] >= kdim[2]) ? 1 : 0), __extension__
({ if (qdim[2] >= kdim[2]) ; else __assert_fail ("qdim[2] >= kdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 59, __extension__ __PRETTY_FUNCTION__); }));
39
←
Assuming the condition is true
→
40
←
Taking true branch
→
60 assert(qdim[3] == kdim[3])((void) sizeof ((qdim[3] == kdim[3]) ? 1 : 0), __extension__ (
{ if (qdim[3] == kdim[3]) ; else __assert_fail ("qdim[3] == kdim[3]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 60, __extension__ __PRETTY_FUNCTION__); }));
41
←
Assuming the condition is true
→
42
←
Taking true branch
→
61 assert(kdim[1] == vdim[1])((void) sizeof ((kdim[1] == vdim[1]) ? 1 : 0), __extension__ (
{ if (kdim[1] == vdim[1]) ; else __assert_fail ("kdim[1] == vdim[1]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 61, __extension__ __PRETTY_FUNCTION__); }));
43
←
Assuming the condition is true
→
44
←
Taking true branch
→
62 assert(cdim[1] == qdim[1])((void) sizeof ((cdim[1] == qdim[1]) ? 1 : 0), __extension__ (
{ if (cdim[1] == qdim[1]) ; else __assert_fail ("cdim[1] == qdim[1]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 62, __extension__ __PRETTY_FUNCTION__); }));
45
←
Assuming the condition is true
→
46
←
Taking true branch
→
63 assert(cdim[3] == vdim[3])((void) sizeof ((cdim[3] == vdim[3]) ? 1 : 0), __extension__ (
{ if (cdim[3] == vdim[3]) ; else __assert_fail ("cdim[3] == vdim[3]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 63, __extension__ __PRETTY_FUNCTION__); }));
47
←
Assuming the condition is true
→
48
←
Taking true branch
→
64 assert(CCV_NNC_MAX_DIM == 2)((void) sizeof (((2) == 2) ? 1 : 0), __extension__ ({ if ((2)
== 2) ; else __assert_fail ("CCV_NNC_MAX_DIM == 2", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 64, __extension__ __PRETTY_FUNCTION__); })); // Need to change this logic for CCV_NNC_MAX_DIM == other number.
49
←
Taking true branch
→
65 int qstride[CCV_NNC_MAX_DIM_ALLOC(12)];
66 int kstride[CCV_NNC_MAX_DIM_ALLOC(12)];
67 int vstride[CCV_NNC_MAX_DIM_ALLOC(12)];
68 int cstride[CCV_NNC_MAX_DIM_ALLOC(12)];
69 int amstride[CCV_NNC_MAX_DIM_ALLOC(12)];
70 ccv_nnc_tensor_view_get_stride(q, qstride);
71 ccv_nnc_tensor_view_get_stride(k, kstride);
72 ccv_nnc_tensor_view_get_stride(v, vstride);
73 ccv_nnc_tensor_view_get_stride(c, cstride);
74 if (q_nd
49.1
'q_nd' is not equal to 3
 == 3)
50
←
Taking false branch
→
75 {
76 qstride[0] = qstride[1], qstride[1] = qstride[2], qstride[2] = qstride[3];
77 kstride[0] = kstride[1], kstride[1] = kstride[2], kstride[2] = kstride[3];
78 vstride[0] = vstride[1], vstride[1] = vstride[2], vstride[2] = vstride[3];
79 cstride[0] = cstride[1], cstride[1] = cstride[2], cstride[2] = cstride[3];
80 }
81 if (attn_mask)
51
←
Assuming 'attn_mask' is non-null
→
52
←
Taking true branch
→
82 {
83 ccv_nnc_tensor_view_get_dim(attn_mask, amdim);
84 ccv_nnc_tensor_view_get_stride(attn_mask, amstride);
85 assert(amdim[0] == qdim[0] || amdim[0] == 1)((void) sizeof ((amdim[0] == qdim[0] || amdim[0] == 1) ? 1 : 0
), __extension__ ({ if (amdim[0] == qdim[0] || amdim[0] == 1)
; else __assert_fail ("amdim[0] == qdim[0] || amdim[0] == 1"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 85, __extension__ __PRETTY_FUNCTION__); }));
53
←
Assuming the condition is true
→
86 assert(amdim[1] == qdim[2] || amdim[1] == 1)((void) sizeof ((amdim[1] == qdim[2] || amdim[1] == 1) ? 1 : 0
), __extension__ ({ if (amdim[1] == qdim[2] || amdim[1] == 1)
; else __assert_fail ("amdim[1] == qdim[2] || amdim[1] == 1"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 86, __extension__ __PRETTY_FUNCTION__); }));
54
←
Assuming the condition is true
→
87 assert(amdim[2] == qdim[1])((void) sizeof ((amdim[2] == qdim[1]) ? 1 : 0), __extension__
({ if (amdim[2] == qdim[1]) ; else __assert_fail ("amdim[2] == qdim[1]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 87, __extension__ __PRETTY_FUNCTION__); }));
55
←
Assuming the condition is true
→
56
←
Taking true branch
→
88 assert(amdim[3] == kdim[1])((void) sizeof ((amdim[3] == kdim[1]) ? 1 : 0), __extension__
({ if (amdim[3] == kdim[1]) ; else __assert_fail ("amdim[3] == kdim[1]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 88, __extension__ __PRETTY_FUNCTION__); }));
57
←
Assuming the condition is true
→
58
←
Taking true branch
→
89 }
90 int i[CCV_NNC_MAX_DIM(2) + 2];
91 float* qk = ccv_nnc_stream_context_get_workspace(stream_context, sizeof(float) * qdim[1] * kdim[1], CCV_TENSOR_CPU_MEMORY);
92 const float* const qp = q->data.f32;
93 const float* const kp = k->data.f32;
94 const float* const vp = v->data.f32;
95 const float* const amp = attn_mask
58.1
'attn_mask' is non-null
 ? attn_mask->data.f32 : 0;
59
←
'?' condition is true
→
96 float* const cp = c->data.f32;
97 const float scale = cmd.info.scaled_dot_product_attention.scale;
98 const int is_causal = cmd.info.scaled_dot_product_attention.is_causal;
99 const int h_h_k_ratio = qdim[2] / kdim[2];
100 assert(kdim[2] == vdim[2])((void) sizeof ((kdim[2] == vdim[2]) ? 1 : 0), __extension__ (
{ if (kdim[2] == vdim[2]) ; else __assert_fail ("kdim[2] == vdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 100, __extension__ __PRETTY_FUNCTION__); }));
60
←
Taking true branch
→
101 assert(qdim[2] >= kdim[2])((void) sizeof ((qdim[2] >= kdim[2]) ? 1 : 0), __extension__
({ if (qdim[2] >= kdim[2]) ; else __assert_fail ("qdim[2] >= kdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 101, __extension__ __PRETTY_FUNCTION__); }));
61
←
Taking true branch
→
102 assert(qdim[2] % kdim[2] == 0)((void) sizeof ((qdim[2] % kdim[2] == 0) ? 1 : 0), __extension__
({ if (qdim[2] % kdim[2] == 0) ; else __assert_fail ("qdim[2] % kdim[2] == 0"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 102, __extension__ __PRETTY_FUNCTION__); }));
62
←
Taking true branch
→
103 for (i[0] = 0; i[0] < qdim[0]; i[0]++)
63
←
Assuming the condition is true
→
64
←
Loop condition is true. Entering loop body
→
104 {
105 const float* const qp0 = qp + i[0] * qstride[0];
106 const float* const kp0 = kp + i[0] * kstride[0];
107 const float* const vp0 = vp + i[0] * vstride[0];
108 const float* const amp0 = amp && amdim[0] > 1 ? amp + i[0] * amstride[0] : amp;
65
←
Assuming 'amp' is null
→
109 float* const cp0 = cp + i[0] * cstride[0];
110 for (i[1] = 0; i[1] < qdim[2]; i[1]++)
66
←
Assuming the condition is true
→
67
←
Loop condition is true. Entering loop body
→
111 {
112 const float* const qp1 = qp0 + i[1] * qstride[2];
113 const float* const kp1 = kp0 + (i[1] / h_h_k_ratio) * kstride[2];
114 const float* const vp1 = vp0 + (i[1] / h_h_k_ratio) * vstride[2];
115 const float* const amp1 = amp
67.1
'amp' is null
 && amdim[1] > 1 ? amp0 + i[1] * amstride[1] : amp0;
116 float* const cp1 = cp0 + i[1] * cstride[2];
117 // Compute Q @ K^T
118 parallel_for(x, qdim[1]){ int x; for ((x) = 0; (x) < (qdim[1]); (x)++) { {
68
←
Assuming the condition is true
→
69
←
Loop condition is true. Entering loop body
→
119 int y, k;
120 const float* const qp2 = qp1 + x * qstride[1];
121 float* const cp2 = cp1 + x * cstride[1];
122 float* const qk0 = qk + x * kdim[1];
123 const float* const amp2 = amp1
69.1
'amp1' is null
 ? amp1 + x * amstride[2] : 0;
70
←
'?' condition is false
→
71
←
'amp2' initialized to a null pointer value
→
124 if (attn_mask
71.1
'attn_mask' is non-null
)
72
←
Taking true branch
→
125 {
126 for (y = 0; y < kdim[1]; y++)
73
←
Assuming the condition is true
→
74
←
Loop condition is true. Entering loop body
→
127 {
128 const float* const kp2 = kp1 + y * kstride[1];
129 float v = 0;
130 for (k = 0; k < qdim[3]; k++)
75
←
Assuming the condition is false
→
76
←
Loop condition is false. Execution continues on line 132
→
131 v += qp2[k * qstride[3]] * kp2[k * kstride[3]];
132 qk0[y] = scale * v + amp2[y * amstride[3]];
77
←
Array access (from variable 'amp2') results in a null pointer dereference
133 }
134 } else {
135 for (y = 0; y < kdim[1]; y++)
136 {
137 const float* const kp2 = kp1 + y * kstride[1];
138 float v = 0;
139 for (k = 0; k < qdim[3]; k++)
140 v += qp2[k * qstride[3]] * kp2[k * kstride[3]];
141 qk0[y] = scale * v;
142 }
143 }
144 // Compute softmax on qk.
145 if (is_causal)
146 {
147 const int x_end = ccv_max(x - qdim[1] + kdim[1] + 1, 0)({ typeof (x - qdim[1] + kdim[1] + 1) _a = (x - qdim[1] + kdim
[1] + 1); typeof (0) _b = (0); (_a > _b) ? _a : _b; });
148 for (y = x_end; y < kdim[1]; y++)
149 qk0[y] = 0;
150 double maxval = qk0[0];
151 for (y = 1; y < x_end; y++)
152 if (qk0[y] > maxval)
153 maxval = qk0[y];
154 double sumval = 0;
155 for (y = 0; y < x_end; y++)
156 sumval += (qk0[y] = expf(qk0[y] - maxval));
157 sumval = 1.0 / sumval;
158 for (y = 0; y < x_end; y++)
159 qk0[y] *= sumval;
160 } else {
161 double maxval = qk0[0];
162 for (y = 1; y < kdim[1]; y++)
163 if (qk0[y] > maxval)
164 maxval = qk0[y];
165 double sumval = 0;
166 for (y = 0; y < kdim[1]; y++)
167 sumval += (qk0[y] = expf(qk0[y] - maxval));
168 sumval = 1.0 / sumval;
169 for (y = 0; y < kdim[1]; y++)
170 qk0[y] *= sumval;
171 }
172 for (k = 0; k < vdim[3]; k++)
173 cp2[k * cstride[3]] = 0;
174 for (y = 0; y < kdim[1]; y++)
175 {
176 const float* const vp2 = vp1 + y * vstride[1];
177 const float v = qk0[y];
178 for (k = 0; k < vdim[3]; k++)
179 cp2[k * cstride[3]] += v * vp2[k * vstride[3]];
180 }
181 } parallel_endfor} }
182 }
183 }
184 if (w)
185 {
186 const int num_heads = cdim[2];
187 ccv_nnc_tensor_view_t* const d = (ccv_nnc_tensor_view_t*)outputs[0];
188 const int w_nd = ccv_nnc_tensor_nd(w->info.dim);
189 assert(w_nd == 2)((void) sizeof ((w_nd == 2) ? 1 : 0), __extension__ ({ if (w_nd
== 2) ; else __assert_fail ("w_nd == 2", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 189, __extension__ __PRETTY_FUNCTION__); }));
190 assert(CCV_IS_TENSOR_CONTIGUOUS(w))((void) sizeof (((!((*(int*)(w)) & CCV_TENSOR_VIEW) || ((
(ccv_nnc_tensor_view_t*)w)->contiguous == 1))) ? 1 : 0), __extension__
({ if ((!((*(int*)(w)) & CCV_TENSOR_VIEW) || (((ccv_nnc_tensor_view_t
*)w)->contiguous == 1))) ; else __assert_fail ("CCV_IS_TENSOR_CONTIGUOUS(w)"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 190, __extension__ __PRETTY_FUNCTION__); }));
191 const int d_nd = ccv_nnc_tensor_nd(d->info.dim);
192 assert(d_nd == 3)((void) sizeof ((d_nd == 3) ? 1 : 0), __extension__ ({ if (d_nd
== 3) ; else __assert_fail ("d_nd == 3", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 192, __extension__ __PRETTY_FUNCTION__); }));
193 int ddim[CCV_NNC_MAX_DIM_ALLOC(12)];
194 int dstride[CCV_NNC_MAX_DIM_ALLOC(12)];
195 ccv_nnc_tensor_view_get_dim(d, ddim);
196 ccv_nnc_tensor_view_get_stride(d, dstride);
197 assert(ddim[2] == cdim[1])((void) sizeof ((ddim[2] == cdim[1]) ? 1 : 0), __extension__ (
{ if (ddim[2] == cdim[1]) ; else __assert_fail ("ddim[2] == cdim[1]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 197, __extension__ __PRETTY_FUNCTION__); }));
198 assert(ddim[3] == num_heads * cdim[3])((void) sizeof ((ddim[3] == num_heads * cdim[3]) ? 1 : 0), __extension__
({ if (ddim[3] == num_heads * cdim[3]) ; else __assert_fail (
"ddim[3] == num_heads * cdim[3]", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 198, __extension__ __PRETTY_FUNCTION__); }));
199 assert(w->info.dim[1] == ddim[3])((void) sizeof ((w->info.dim[1] == ddim[3]) ? 1 : 0), __extension__
({ if (w->info.dim[1] == ddim[3]) ; else __assert_fail ("w->info.dim[1] == ddim[3]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 199, __extension__ __PRETTY_FUNCTION__); }));
200 assert(w->info.dim[0] == ddim[3])((void) sizeof ((w->info.dim[0] == ddim[3]) ? 1 : 0), __extension__
({ if (w->info.dim[0] == ddim[3]) ; else __assert_fail ("w->info.dim[0] == ddim[3]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 200, __extension__ __PRETTY_FUNCTION__); }));
201 float* const dp = d->data.f32;
202 const float* const wp = w->data.f32;
203 const float* const cp = c->data.f32;
204 if (bias)
205 {
206 assert(ccv_nnc_tensor_count(bias->info) == ddim[3])((void) sizeof ((ccv_nnc_tensor_count(bias->info) == ddim[
3]) ? 1 : 0), __extension__ ({ if (ccv_nnc_tensor_count(bias->
info) == ddim[3]) ; else __assert_fail ("ccv_nnc_tensor_count(bias->info) == ddim[3]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 206, __extension__ __PRETTY_FUNCTION__); }));
207 assert(CCV_IS_TENSOR_CONTIGUOUS(bias))((void) sizeof (((!((*(int*)(bias)) & CCV_TENSOR_VIEW) ||
(((ccv_nnc_tensor_view_t*)bias)->contiguous == 1))) ? 1 :
0), __extension__ ({ if ((!((*(int*)(bias)) & CCV_TENSOR_VIEW
) || (((ccv_nnc_tensor_view_t*)bias)->contiguous == 1))) ;
else __assert_fail ("CCV_IS_TENSOR_CONTIGUOUS(bias)", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 207, __extension__ __PRETTY_FUNCTION__); }));
208 const float* const biasp = bias->data.f32;
209 for (i[0] = 0; i[0] < ddim[1]; i[0]++)
210 {
211 const float* const cp0 = cp + i[0] * cstride[0];
212 float* const dp0 = dp + i[0] * dstride[1];
213 parallel_for(y, ddim[2]){ int y; for ((y) = 0; (y) < (ddim[2]); (y)++) { {
214 int x, j, k;
215 const float* const cp1 = cp0 + y * cstride[1];
216 float* const dp1 = dp0 + y * dstride[2];
217 for (x = 0; x < ddim[3]; x++)
218 {
219 const float* const wp0 = wp + x * ddim[3];
220 float v = biasp[x];
221 for (j = 0; j < num_heads; j++)
222 {
223 const float* const cp2 = cp1 + j * cstride[2];
224 for (k = 0; k < cdim[3]; k++)
225 v += wp0[j * cdim[3] + k] * cp2[k * cstride[3]];
226 }
227 dp1[x * dstride[3]] = v;
228 }
229 } parallel_endfor} }
230 }
231 } else {
232 for (i[0] = 0; i[0] < ddim[1]; i[0]++)
233 {
234 const float* const cp0 = cp + i[0] * cstride[0];
235 float* const dp0 = dp + i[0] * dstride[1];
236 parallel_for(y, ddim[2]){ int y; for ((y) = 0; (y) < (ddim[2]); (y)++) { {
237 int x, j, k;
238 const float* const cp1 = cp0 + y * cstride[1];
239 float* const dp1 = dp0 + y * dstride[2];
240 for (x = 0; x < ddim[3]; x++)
241 {
242 const float* const wp0 = wp + x * ddim[3];
243 float v = 0;
244 for (j = 0; j < num_heads; j++)
245 {
246 const float* const cp2 = cp1 + j * cstride[2];
247 for (k = 0; k < cdim[3]; k++)
248 v += wp0[j * cdim[3] + k] * cp2[k * cstride[3]];
249 }
250 dp1[x * dstride[3]] = v;
251 }
252 } parallel_endfor} }
253 }
254 }
255 }
256 return CCV_NNC_EXEC_SUCCESS;
257}
258
259static int _ccv_nnc_scaled_dot_product_attention_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
260{
261 // Assuming no saved_softmax, we need to recompute from q, k, v.
262 // We cannot do this with masks (yet).
263 assert(input_size >= 6)((void) sizeof ((input_size >= 6) ? 1 : 0), __extension__ (
{ if (input_size >= 6) ; else __assert_fail ("input_size >= 6"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 263, __extension__ __PRETTY_FUNCTION__); }));
264 ccv_nnc_tensor_view_t* const g = (ccv_nnc_tensor_view_t*)inputs[0];
265 ccv_nnc_tensor_view_t* const q = (ccv_nnc_tensor_view_t*)inputs[3];
266 ccv_nnc_tensor_view_t* const k = (ccv_nnc_tensor_view_t*)inputs[4];
267 ccv_nnc_tensor_view_t* const v = (ccv_nnc_tensor_view_t*)inputs[5];
268 ccv_nnc_tensor_view_t* const dq = (ccv_nnc_tensor_view_t*)outputs[0];
269 ccv_nnc_tensor_view_t* const dk = (ccv_nnc_tensor_view_t*)outputs[1];
270 ccv_nnc_tensor_view_t* const dv = (ccv_nnc_tensor_view_t*)outputs[2];
271 const int q_nd = ccv_nnc_tensor_nd(q->info.dim);
272 assert(q_nd == 3 || q_nd == 4)((void) sizeof ((q_nd == 3 || q_nd == 4) ? 1 : 0), __extension__
({ if (q_nd == 3 || q_nd == 4) ; else __assert_fail ("q_nd == 3 || q_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 272, __extension__ __PRETTY_FUNCTION__); }));
273 const int k_nd = ccv_nnc_tensor_nd(k->info.dim);
274 assert(k_nd == 3 || k_nd == 4)((void) sizeof ((k_nd == 3 || k_nd == 4) ? 1 : 0), __extension__
({ if (k_nd == 3 || k_nd == 4) ; else __assert_fail ("k_nd == 3 || k_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 274, __extension__ __PRETTY_FUNCTION__); }));
275 const int v_nd = ccv_nnc_tensor_nd(v->info.dim);
276 assert(v_nd == 3 || v_nd == 4)((void) sizeof ((v_nd == 3 || v_nd == 4) ? 1 : 0), __extension__
({ if (v_nd == 3 || v_nd == 4) ; else __assert_fail ("v_nd == 3 || v_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 276, __extension__ __PRETTY_FUNCTION__); }));
277 const int g_nd = ccv_nnc_tensor_nd(g->info.dim);
278 assert(g_nd == 3 || g_nd == 4)((void) sizeof ((g_nd == 3 || g_nd == 4) ? 1 : 0), __extension__
({ if (g_nd == 3 || g_nd == 4) ; else __assert_fail ("g_nd == 3 || g_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 278, __extension__ __PRETTY_FUNCTION__); }));
279 const int dq_nd = ccv_nnc_tensor_nd(dq->info.dim);
280 assert(dq_nd == 3 || dq_nd == 4)((void) sizeof ((dq_nd == 3 || dq_nd == 4) ? 1 : 0), __extension__
({ if (dq_nd == 3 || dq_nd == 4) ; else __assert_fail ("dq_nd == 3 || dq_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 280, __extension__ __PRETTY_FUNCTION__); }));
281 assert(dq_nd == q_nd)((void) sizeof ((dq_nd == q_nd) ? 1 : 0), __extension__ ({ if
(dq_nd == q_nd) ; else __assert_fail ("dq_nd == q_nd", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 281, __extension__ __PRETTY_FUNCTION__); }));
282 const int dk_nd = ccv_nnc_tensor_nd(dk->info.dim);
283 assert(dk_nd == 3 || dk_nd == 4)((void) sizeof ((dk_nd == 3 || dk_nd == 4) ? 1 : 0), __extension__
({ if (dk_nd == 3 || dk_nd == 4) ; else __assert_fail ("dk_nd == 3 || dk_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 283, __extension__ __PRETTY_FUNCTION__); }));
284 assert(dk_nd == k_nd)((void) sizeof ((dk_nd == k_nd) ? 1 : 0), __extension__ ({ if
(dk_nd == k_nd) ; else __assert_fail ("dk_nd == k_nd", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 284, __extension__ __PRETTY_FUNCTION__); }));
285 const int dv_nd = ccv_nnc_tensor_nd(dv->info.dim);
286 assert(dv_nd == 3 || dv_nd == 4)((void) sizeof ((dv_nd == 3 || dv_nd == 4) ? 1 : 0), __extension__
({ if (dv_nd == 3 || dv_nd == 4) ; else __assert_fail ("dv_nd == 3 || dv_nd == 4"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 286, __extension__ __PRETTY_FUNCTION__); }));
287 assert(dv_nd == v_nd)((void) sizeof ((dv_nd == v_nd) ? 1 : 0), __extension__ ({ if
(dv_nd == v_nd) ; else __assert_fail ("dv_nd == v_nd", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 287, __extension__ __PRETTY_FUNCTION__); }));
288 assert(q_nd == k_nd && k_nd == v_nd && v_nd == g_nd)((void) sizeof ((q_nd == k_nd && k_nd == v_nd &&
v_nd == g_nd) ? 1 : 0), __extension__ ({ if (q_nd == k_nd &&
k_nd == v_nd && v_nd == g_nd) ; else __assert_fail (
"q_nd == k_nd && k_nd == v_nd && v_nd == g_nd"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 288, __extension__ __PRETTY_FUNCTION__); }));
289 // Assuming this is float 32.
290 int qdim[CCV_NNC_MAX_DIM_ALLOC(12)];
291 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
292 int vdim[CCV_NNC_MAX_DIM_ALLOC(12)];
293 int gdim[CCV_NNC_MAX_DIM_ALLOC(12)];
294 int dqdim[CCV_NNC_MAX_DIM_ALLOC(12)];
295 int dkdim[CCV_NNC_MAX_DIM_ALLOC(12)];
296 int dvdim[CCV_NNC_MAX_DIM_ALLOC(12)];
297 ccv_nnc_tensor_view_get_dim(q, qdim);
298 ccv_nnc_tensor_view_get_dim(k, kdim);
299 ccv_nnc_tensor_view_get_dim(v, vdim);
300 ccv_nnc_tensor_view_get_dim(g, gdim);
301 ccv_nnc_tensor_view_get_dim(dq, dqdim);
302 ccv_nnc_tensor_view_get_dim(dk, dkdim);
303 ccv_nnc_tensor_view_get_dim(dv, dvdim);
304 if (q_nd == 3)
305 {
306 qdim[0] = qdim[1], qdim[1] = qdim[2], qdim[2] = 1;
307 kdim[0] = kdim[1], kdim[1] = kdim[2], kdim[2] = 1;
308 vdim[0] = vdim[1], vdim[1] = vdim[2], vdim[2] = 1;
309 gdim[0] = gdim[1], gdim[1] = gdim[2], gdim[2] = 1;
310 dqdim[0] = dqdim[1], dqdim[1] = dqdim[2], dqdim[2] = 1;
311 dkdim[0] = dkdim[1], dkdim[1] = dkdim[2], dkdim[2] = 1;
312 dvdim[0] = dvdim[1], dvdim[1] = dvdim[2], dvdim[2] = 1;
313 }
314 assert(qdim[0] == kdim[0] && kdim[0] == vdim[0] && vdim[0] == gdim[0])((void) sizeof ((qdim[0] == kdim[0] && kdim[0] == vdim
[0] && vdim[0] == gdim[0]) ? 1 : 0), __extension__ ({
if (qdim[0] == kdim[0] && kdim[0] == vdim[0] &&
vdim[0] == gdim[0]) ; else __assert_fail ("qdim[0] == kdim[0] && kdim[0] == vdim[0] && vdim[0] == gdim[0]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 314, __extension__ __PRETTY_FUNCTION__); }));
315 assert(qdim[2] == gdim[2])((void) sizeof ((qdim[2] == gdim[2]) ? 1 : 0), __extension__ (
{ if (qdim[2] == gdim[2]) ; else __assert_fail ("qdim[2] == gdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 315, __extension__ __PRETTY_FUNCTION__); }));
316 assert(kdim[2] == vdim[2])((void) sizeof ((kdim[2] == vdim[2]) ? 1 : 0), __extension__ (
{ if (kdim[2] == vdim[2]) ; else __assert_fail ("kdim[2] == vdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 316, __extension__ __PRETTY_FUNCTION__); }));
317 assert(qdim[2] % kdim[2] == 0)((void) sizeof ((qdim[2] % kdim[2] == 0) ? 1 : 0), __extension__
({ if (qdim[2] % kdim[2] == 0) ; else __assert_fail ("qdim[2] % kdim[2] == 0"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 317, __extension__ __PRETTY_FUNCTION__); }));
318 assert(qdim[2] >= kdim[2])((void) sizeof ((qdim[2] >= kdim[2]) ? 1 : 0), __extension__
({ if (qdim[2] >= kdim[2]) ; else __assert_fail ("qdim[2] >= kdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 318, __extension__ __PRETTY_FUNCTION__); }));
319 assert(qdim[3] == kdim[3])((void) sizeof ((qdim[3] == kdim[3]) ? 1 : 0), __extension__ (
{ if (qdim[3] == kdim[3]) ; else __assert_fail ("qdim[3] == kdim[3]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 319, __extension__ __PRETTY_FUNCTION__); }));
320 assert(kdim[1] == vdim[1])((void) sizeof ((kdim[1] == vdim[1]) ? 1 : 0), __extension__ (
{ if (kdim[1] == vdim[1]) ; else __assert_fail ("kdim[1] == vdim[1]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 320, __extension__ __PRETTY_FUNCTION__); }));
321 assert(gdim[1] == qdim[1])((void) sizeof ((gdim[1] == qdim[1]) ? 1 : 0), __extension__ (
{ if (gdim[1] == qdim[1]) ; else __assert_fail ("gdim[1] == qdim[1]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 321, __extension__ __PRETTY_FUNCTION__); }));
322 assert(gdim[3] == vdim[3])((void) sizeof ((gdim[3] == vdim[3]) ? 1 : 0), __extension__ (
{ if (gdim[3] == vdim[3]) ; else __assert_fail ("gdim[3] == vdim[3]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 322, __extension__ __PRETTY_FUNCTION__); }));
323 assert(CCV_NNC_MAX_DIM == 2)((void) sizeof (((2) == 2) ? 1 : 0), __extension__ ({ if ((2)
== 2) ; else __assert_fail ("CCV_NNC_MAX_DIM == 2", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 323, __extension__ __PRETTY_FUNCTION__); })); // Need to change this logic for CCV_NNC_MAX_DIM == other number.
324 int qstride[CCV_NNC_MAX_DIM_ALLOC(12)];
325 int kstride[CCV_NNC_MAX_DIM_ALLOC(12)];
326 int vstride[CCV_NNC_MAX_DIM_ALLOC(12)];
327 int gstride[CCV_NNC_MAX_DIM_ALLOC(12)];
328 int dqstride[CCV_NNC_MAX_DIM_ALLOC(12)];
329 int dkstride[CCV_NNC_MAX_DIM_ALLOC(12)];
330 int dvstride[CCV_NNC_MAX_DIM_ALLOC(12)];
331 ccv_nnc_tensor_view_get_stride(q, qstride);
332 ccv_nnc_tensor_view_get_stride(k, kstride);
333 ccv_nnc_tensor_view_get_stride(v, vstride);
334 ccv_nnc_tensor_view_get_stride(g, gstride);
335 ccv_nnc_tensor_view_get_stride(dq, dqstride);
336 ccv_nnc_tensor_view_get_stride(dk, dkstride);
337 ccv_nnc_tensor_view_get_stride(dv, dvstride);
338 if (q_nd == 3)
339 {
340 qstride[0] = qstride[1], qstride[1] = qstride[2], qstride[2] = qstride[3];
341 kstride[0] = kstride[1], kstride[1] = kstride[2], kstride[2] = kstride[3];
342 vstride[0] = vstride[1], vstride[1] = vstride[2], vstride[2] = vstride[3];
343 gstride[0] = gstride[1], gstride[1] = gstride[2], gstride[2] = gstride[3];
344 dqstride[0] = dqstride[1], dqstride[1] = dqstride[2], dqstride[2] = dqstride[3];
345 dkstride[0] = dkstride[1], dkstride[1] = dkstride[2], dkstride[2] = dkstride[3];
346 dvstride[0] = dvstride[1], dvstride[1] = dvstride[2], dvstride[2] = dvstride[3];
347 }
348 int i[CCV_NNC_MAX_DIM(2) + 2];
349 float* qk = ccv_nnc_stream_context_get_workspace(stream_context, sizeof(float) * 2 * kdim[1], CCV_TENSOR_CPU_MEMORY);
350 const float* const qp = q->data.f32;
351 const float* const kp = k->data.f32;
352 const float* const vp = v->data.f32;
353 const float* const gp = g->data.f32;
354 float* const dqp = dq->data.f32;
355 float* const dkp = dk->data.f32;
356 float* const dvp = dv->data.f32;
357 const float scale = cmd.info.scaled_dot_product_attention.scale;
358 const int is_causal = cmd.info.scaled_dot_product_attention.is_causal;
359 const int h_h_k_ratio = qdim[2] / kdim[2];
360 for (i[0] = 0; i[0] < qdim[0]; i[0]++)
361 {
362 const float* const qp0 = qp + i[0] * qstride[0];
363 const float* const kp0 = kp + i[0] * kstride[0];
364 const float* const vp0 = vp + i[0] * vstride[0];
365 const float* const gp0 = gp + i[0] * gstride[0];
366 float* const dqp0 = dqp + i[0] * dqstride[0];
367 float* const dkp0 = dkp + i[0] * dkstride[0];
368 float* const dvp0 = dvp + i[0] * dvstride[0];
369 for (i[1] = 0; i[1] < qdim[2]; i[1]++)
370 {
371 const float* const qp1 = qp0 + i[1] * qstride[2];
372 const float* const kp1 = kp0 + (i[1] / h_h_k_ratio) * kstride[2];
373 const float* const vp1 = vp0 + (i[1] / h_h_k_ratio) * vstride[2];
374 const float* const gp1 = gp0 + i[1] * gstride[2];
375 float* const dqp1 = dqp0 + i[1] * dqstride[2];
376 float* const dkp1 = dkp0 + (i[1] / h_h_k_ratio) * dkstride[2];
377 float* const dvp1 = dvp0 + (i[1] / h_h_k_ratio) * dvstride[2];
378 // Compute Q @ K^T
379 int x, y, k;
380 for (x = 0; x < qdim[1]; x++)
381 {
382 float* const dqp2 = dqp1 + x * dqstride[1];
383 for (k = 0; k < qdim[3]; k++)
384 dqp2[k * dqstride[3]] = 0;
385 }
386 // Only zero out when it is at 0-index.
387 if (i[1] % h_h_k_ratio == 0)
388 for (y = 0; y < kdim[1]; y++)
389 {
390 float* const dkp2 = dkp1 + y * dkstride[1];
391 for (k = 0; k < qdim[3]; k++)
392 dkp2[k * dkstride[3]] = 0;
393 }
394 // Only zero out when it is at 0-index.
395 if (i[1] % h_h_k_ratio == 0)
396 for (y = 0; y < kdim[1]; y++)
397 {
398 float* const dvp2 = dvp1 + y * dvstride[1];
399 for (k = 0; k < vdim[3]; k++)
400 dvp2[k * dvstride[3]] = 0;
401 }
402 for (x = 0; x < qdim[1]; x++)
403 {
404 const float* const qp2 = qp1 + x * qstride[1];
405 const float* const gp2 = gp1 + x * gstride[1];
406 float* const qk0 = qk;
407 float* const qks0 = qk + kdim[1];
408 for (y = 0; y < kdim[1]; y++)
409 {
410 const float* const kp2 = kp1 + y * kstride[1];
411 float v = 0;
412 for (k = 0; k < qdim[3]; k++)
413 v += qp2[k * qstride[3]] * kp2[k * kstride[3]];
414 qk0[y] = scale * v;
415 }
416 // Compute softmax on qk.
417 if (is_causal)
418 {
419 const int x_end = ccv_max(x - qdim[1] + kdim[1] + 1, 0)({ typeof (x - qdim[1] + kdim[1] + 1) _a = (x - qdim[1] + kdim
[1] + 1); typeof (0) _b = (0); (_a > _b) ? _a : _b; });
420 for (y = x_end; y < kdim[1]; y++)
421 qk0[y] = 0;
422 double maxval = qk0[0];
423 for (y = 1; y < x_end; y++)
424 if (qk0[y] > maxval)
425 maxval = qk0[y];
426 double sumval = 0;
427 for (y = 0; y < x_end; y++)
428 sumval += (qk0[y] = expf(qk0[y] - maxval));
429 sumval = 1.0 / sumval;
430 for (y = 0; y < x_end; y++)
431 qk0[y] *= sumval;
432 } else {
433 double maxval = qk0[0];
434 for (y = 1; y < kdim[1]; y++)
435 if (qk0[y] > maxval)
436 maxval = qk0[y];
437 double sumval = 0;
438 for (y = 0; y < kdim[1]; y++)
439 sumval += (qk0[y] = expf(qk0[y] - maxval));
440 sumval = 1.0 / sumval;
441 for (y = 0; y < kdim[1]; y++)
442 qk0[y] *= sumval;
443 }
444 for (y = 0; y < kdim[1]; y++)
445 {
446 float* const dvp2 = dvp1 + y * dvstride[1];
447 const float v = qk0[y];
448 for (k = 0; k < vdim[3]; k++)
449 dvp2[k * dvstride[3]] += v * gp2[k * gstride[3]];
450 }
451 double sumval = 0;
452 for (y = 0; y < kdim[1]; y++)
453 {
454 const float* const vp2 = vp1 + y * vstride[1];
455 float v = 0;
456 for (k = 0; k < vdim[3]; k++)
457 v += gp2[k * gstride[3]] * vp2[k * vstride[3]];
458 qks0[y] = v;
459 sumval += v * qk0[y];
460 }
461 for (y = 0; y < kdim[1]; y++)
462 qk0[y] = (qks0[y] - sumval) * qk0[y];
463 float* const dqp2 = dqp1 + x * dqstride[1];
464 for (y = 0; y < kdim[1]; y++)
465 {
466 const float* const kp2 = kp1 + y * kstride[1];
467 float* const dkp2 = dkp1 + y * dkstride[1];
468 const float v = scale * qk0[y];
469 for (k = 0; k < qdim[3]; k++)
470 {
471 dqp2[k * dqstride[3]] += v * kp2[k * kstride[3]];
472 dkp2[k * dkstride[3]] += v * qp2[k * qstride[3]];
473 }
474 }
475 }
476 }
477 }
478 return CCV_NNC_EXEC_SUCCESS;
479}
480
481REGISTER_COMMAND_BACKEND(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD, CCV_NNC_BACKEND_CPU_REF)void _register_command_CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD_backend_CCV_NNC_BACKEND_CPU_REF(ccv_nnc_cmd_backend_registry_t* const registry)
482{
483 registry->tensor_formats = CCV_TENSOR_FORMAT_NHWC;
484 registry->tensor_datatypes = CCV_32F;
485 registry->tensor_memory = CCV_TENSOR_CPU_MEMORY;
486 registry->algorithms = 1;
487 registry->exec = _ccv_nnc_scaled_dot_product_attention_forw;
488}
489
490REGISTER_COMMAND_BACKEND(CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD, CCV_NNC_BACKEND_CPU_REF)void _register_command_CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_BACKWARD_backend_CCV_NNC_BACKEND_CPU_REF(ccv_nnc_cmd_backend_registry_t* const registry)
491{
492 registry->tensor_formats = CCV_TENSOR_FORMAT_NHWC;
493 registry->tensor_datatypes = CCV_32F;
494 registry->tensor_memory = CCV_TENSOR_CPU_MEMORY;
495 registry->algorithms = 1;
496 registry->exec = _ccv_nnc_scaled_dot_product_attention_back;
497}