scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu

Bug Summary

File:	nnc/cmd/scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c
Warning:	line 258, column 28 Array access (from variable 'amp2') results in a null pointer dereference
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
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/2026-05-27-144330-3062811-1 -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	const int is_varlen = cmd.info.scaled_dot_product_attention.is_varlen;
21	ccv_nnc_tensor_view_t* const q = (ccv_nnc_tensor_view_t*)inputs[0];
22	ccv_nnc_tensor_view_t* const k = (ccv_nnc_tensor_view_t*)inputs[1];
23	ccv_nnc_tensor_view_t* const v = (ccv_nnc_tensor_view_t*)inputs[2];
24	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→
25	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→
26	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→
27	ccv_nnc_tensor_view_t* const q_seq_offsets = is_varlen && input_size > 6 ? (ccv_nnc_tensor_view_t*)inputs[6] : 0;
11
←
Assuming 'is_varlen' is 0→
28	ccv_nnc_tensor_view_t* const kv_seq_offsets = is_varlen11.1
'is_varlen' is 0
 && input_size > 7 ? (ccv_nnc_tensor_view_t*)inputs[7] : 0;
29	const int attention_sinks = cmd.info.scaled_dot_product_attention.attention_sinks;
30	ccv_nnc_tensor_view_t* const sinks = attention_sinks && input_size > 8 ? (ccv_nnc_tensor_view_t*)inputs[8] : 0;
12
←
Assuming 'attention_sinks' is not equal to 0→
13
←
Assuming 'input_size' is > 8→
14
←
'?' condition is true→
31	if (bias) // bias always requires a weight matrix.
15
←
Assuming 'bias' is null→
32		{ 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"
, 32, __extension__ __PRETTY_FUNCTION__); })); }
33	if (is_varlen15.1
'is_varlen' is 0
 && (attn_mask || w || bias || !q_seq_offsets || !kv_seq_offsets))
34		return CCV_NNC_EXEC_INVALID;
35	if (attention_sinks15.2
'attention_sinks' is not equal to 0
 && !sinks)
16
←
Assuming 'sinks' is non-null→
36		return CCV_NNC_EXEC_INVALID;
37	ccv_nnc_tensor_view_t* const c = (w) ? (ccv_nnc_tensor_view_t*)outputs[2] : (ccv_nnc_tensor_view_t*)outputs[0];
17
←
Taking false branch→
18
←
Assuming 'w' is non-null→
19
←
'?' condition is true→
38	const int q_nd = ccv_nnc_tensor_nd(q->info.dim);
39	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"
, 39, __extension__ __PRETTY_FUNCTION__); }));
20
←
Assuming 'q_nd' is not equal to 3→
21
←
Assuming 'q_nd' is equal to 4→
22
←
Taking true branch→
40	const int k_nd = ccv_nnc_tensor_nd(k->info.dim);
41	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"
, 41, __extension__ __PRETTY_FUNCTION__); }));
23
←
Assuming 'k_nd' is not equal to 3→
24
←
Assuming 'k_nd' is equal to 4→
25
←
Taking true branch→
42	const int v_nd = ccv_nnc_tensor_nd(v->info.dim);
43	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"
, 43, __extension__ __PRETTY_FUNCTION__); }));
26
←
Assuming 'v_nd' is not equal to 3→
27
←
Assuming 'v_nd' is equal to 4→
28
←
Taking true branch→
44	const int c_nd = ccv_nnc_tensor_nd(c->info.dim);
45	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"
, 45, __extension__ __PRETTY_FUNCTION__); }));
29
←
Assuming 'c_nd' is not equal to 3→
30
←
Assuming 'c_nd' is equal to 4→
31
←
Taking true branch→
46	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"
, 46, __extension__ __PRETTY_FUNCTION__); }));
32
←
Taking true branch→
47	if (is_varlen32.1
'is_varlen' is 0
 && q_nd != 4)
48		return CCV_NNC_EXEC_INVALID;
49	// Assuming this is float 32.
50	int qdim[CCV_NNC_MAX_DIM_ALLOC(12)];
51	int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
52	int vdim[CCV_NNC_MAX_DIM_ALLOC(12)];
53	int cdim[CCV_NNC_MAX_DIM_ALLOC(12)];
54	int amdim[CCV_NNC_MAX_DIM_ALLOC(12)];
55	ccv_nnc_tensor_view_get_dim(q, qdim);
56	ccv_nnc_tensor_view_get_dim(k, kdim);
57	ccv_nnc_tensor_view_get_dim(v, vdim);
58	ccv_nnc_tensor_view_get_dim(c, cdim);
59	if (is_varlen32.2
'is_varlen' is 0
)
33
←
Taking false branch→
60	{
61		assert(q_seq_offsets->info.datatype == CCV_32S)((void) sizeof ((q_seq_offsets->info.datatype == CCV_32S) ?
 1 : 0), __extension__ ({ if (q_seq_offsets->info.datatype
 == CCV_32S) ; else __assert_fail ("q_seq_offsets->info.datatype == CCV_32S"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 61, __extension__ __PRETTY_FUNCTION__); }));
62		assert(kv_seq_offsets->info.datatype == CCV_32S)((void) sizeof ((kv_seq_offsets->info.datatype == CCV_32S)
 ? 1 : 0), __extension__ ({ if (kv_seq_offsets->info.datatype
 == CCV_32S) ; else __assert_fail ("kv_seq_offsets->info.datatype == CCV_32S"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 62, __extension__ __PRETTY_FUNCTION__); }));
63		assert(CCV_IS_TENSOR_CONTIGUOUS(q_seq_offsets))((void) sizeof (((!((*(int*)(q_seq_offsets)) & CCV_TENSOR_VIEW
) || (((ccv_nnc_tensor_view_t*)q_seq_offsets)->contiguous ==
 1))) ? 1 : 0), __extension__ ({ if ((!((*(int*)(q_seq_offsets
)) & CCV_TENSOR_VIEW) || (((ccv_nnc_tensor_view_t*)q_seq_offsets
)->contiguous == 1))) ; else __assert_fail ("CCV_IS_TENSOR_CONTIGUOUS(q_seq_offsets)"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 63, __extension__ __PRETTY_FUNCTION__); }));
64		assert(CCV_IS_TENSOR_CONTIGUOUS(kv_seq_offsets))((void) sizeof (((!((*(int*)(kv_seq_offsets)) & CCV_TENSOR_VIEW
) || (((ccv_nnc_tensor_view_t*)kv_seq_offsets)->contiguous
 == 1))) ? 1 : 0), __extension__ ({ if ((!((*(int*)(kv_seq_offsets
)) & CCV_TENSOR_VIEW) || (((ccv_nnc_tensor_view_t*)kv_seq_offsets
)->contiguous == 1))) ; else __assert_fail ("CCV_IS_TENSOR_CONTIGUOUS(kv_seq_offsets)"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 64, __extension__ __PRETTY_FUNCTION__); }));
65	}
66	if (q_nd33.1
'q_nd' is not equal to 3
 == 3)
34
←
Taking false branch→
67	{
68		qdim[0] = qdim[1], qdim[1] = qdim[2], qdim[2] = 1;
69		kdim[0] = kdim[1], kdim[1] = kdim[2], kdim[2] = 1;
70		vdim[0] = vdim[1], vdim[1] = vdim[2], vdim[2] = 1;
71		cdim[0] = cdim[1], cdim[1] = cdim[2], cdim[2] = 1;
72	}
73	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"
, 73, __extension__ __PRETTY_FUNCTION__); }));
35
←
Assuming the condition is true→
36
←
Assuming the condition is true→
37
←
Assuming the condition is true→
38
←
Taking true branch→
74	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"
, 74, __extension__ __PRETTY_FUNCTION__); }));
39
←
Assuming the condition is true→
40
←
Taking true branch→
75	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"
, 75, __extension__ __PRETTY_FUNCTION__); }));
41
←
Assuming the condition is true→
42
←
Taking true branch→
76	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"
, 76, __extension__ __PRETTY_FUNCTION__); }));
43
←
Assuming the condition is true→
44
←
Taking true branch→
77	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"
, 77, __extension__ __PRETTY_FUNCTION__); }));
45
←
Assuming the condition is true→
46
←
Taking true branch→
78	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"
, 78, __extension__ __PRETTY_FUNCTION__); }));
47
←
Assuming the condition is true→
48
←
Taking true branch→
79	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"
, 79, __extension__ __PRETTY_FUNCTION__); }));
49
←
Assuming the condition is true→
50
←
Taking true branch→
80	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"
, 80, __extension__ __PRETTY_FUNCTION__); }));
51
←
Assuming the condition is true→
52
←
Taking true branch→
81	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"
, 81, __extension__ __PRETTY_FUNCTION__); }));
53
←
Assuming the condition is true→
54
←
Taking true branch→
82	const float* const sinkp = sinks54.1
'sinks' is non-null
 ? sinks->data.f32 : 0;
55
←
'?' condition is true→
83	int sink_head_stride = 0;
84	if (attention_sinks55.1
'attention_sinks' is not equal to 0
)
56
←
Taking true branch→
85	{
86		assert(CCV_IS_TENSOR_CONTIGUOUS(sinks))((void) sizeof (((!((*(int*)(sinks)) & CCV_TENSOR_VIEW) ||
 (((ccv_nnc_tensor_view_t*)sinks)->contiguous == 1))) ? 1 :
 0), __extension__ ({ if ((!((*(int*)(sinks)) & CCV_TENSOR_VIEW
) || (((ccv_nnc_tensor_view_t*)sinks)->contiguous == 1))) ;
 else __assert_fail ("CCV_IS_TENSOR_CONTIGUOUS(sinks)", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 86, __extension__ __PRETTY_FUNCTION__); }));
57
←
Assuming the condition is true→
87		const int sink_count = ccv_nnc_tensor_count(sinks->info);
88		assert(sink_count == 1 || sink_count == qdim[2])((void) sizeof ((sink_count == 1 || sink_count == qdim[2]) ? 1
 : 0), __extension__ ({ if (sink_count == 1 || sink_count == qdim
[2]) ; else __assert_fail ("sink_count == 1 || sink_count == qdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 88, __extension__ __PRETTY_FUNCTION__); }));
58
←
Assuming 'sink_count' is equal to 1→
89		sink_head_stride = (sink_count58.1
'sink_count' is equal to 1
 == 1) ? 0 : 1;
59
←
'?' condition is true→
90	}
91	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"
, 91, __extension__ __PRETTY_FUNCTION__); })); // Need to change this logic for CCV_NNC_MAX_DIM == other number.
60
←
Taking true branch→
92	int qstride[CCV_NNC_MAX_DIM_ALLOC(12)];
93	int kstride[CCV_NNC_MAX_DIM_ALLOC(12)];
94	int vstride[CCV_NNC_MAX_DIM_ALLOC(12)];
95	int cstride[CCV_NNC_MAX_DIM_ALLOC(12)];
96	int amstride[CCV_NNC_MAX_DIM_ALLOC(12)];
97	ccv_nnc_tensor_view_get_stride(q, qstride);
98	ccv_nnc_tensor_view_get_stride(k, kstride);
99	ccv_nnc_tensor_view_get_stride(v, vstride);
100	ccv_nnc_tensor_view_get_stride(c, cstride);
101	if (q_nd60.1
'q_nd' is not equal to 3
 == 3)
61
←
Taking false branch→
102	{
103		qstride[0] = qstride[1], qstride[1] = qstride[2], qstride[2] = qstride[3];
104		kstride[0] = kstride[1], kstride[1] = kstride[2], kstride[2] = kstride[3];
105		vstride[0] = vstride[1], vstride[1] = vstride[2], vstride[2] = vstride[3];
106		cstride[0] = cstride[1], cstride[1] = cstride[2], cstride[2] = cstride[3];
107	}
108	if (attn_mask)
62
←
Assuming 'attn_mask' is non-null→
63
←
Taking true branch→
109	{
110		ccv_nnc_tensor_view_get_dim(attn_mask, amdim);
111		ccv_nnc_tensor_view_get_stride(attn_mask, amstride);
112		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"
, 112, __extension__ __PRETTY_FUNCTION__); }));
64
←
Assuming the condition is true→
113		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"
, 113, __extension__ __PRETTY_FUNCTION__); }));
65
←
Assuming the condition is true→
114		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"
, 114, __extension__ __PRETTY_FUNCTION__); }));
66
←
Assuming the condition is true→
67
←
Taking true branch→
115		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"
, 115, __extension__ __PRETTY_FUNCTION__); }));
68
←
Assuming the condition is true→
69
←
Taking true branch→
116	}
117	int i[CCV_NNC_MAX_DIM(2) + 2];
118	float* qk = ccv_nnc_stream_context_get_workspace(stream_context, sizeof(float) * qdim[1] * kdim[1], CCV_TENSOR_CPU_MEMORY);
119	const float* const qp = q->data.f32;
120	const float* const kp = k->data.f32;
121	const float* const vp = v->data.f32;
122	const float* const amp = attn_mask69.1
'attn_mask' is non-null
 ? attn_mask->data.f32 : 0;
70
←
'?' condition is true→
123	float* const cp = c->data.f32;
124	const float scale = cmd.info.scaled_dot_product_attention.scale;
125	const int is_causal = cmd.info.scaled_dot_product_attention.is_causal;
126	const int h_h_k_ratio = qdim[2] / kdim[2];
127	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"
, 127, __extension__ __PRETTY_FUNCTION__); }));
71
←
Taking true branch→
128	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"
, 128, __extension__ __PRETTY_FUNCTION__); }));
72
←
Taking true branch→
129	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"
, 129, __extension__ __PRETTY_FUNCTION__); }));
73
←
Taking true branch→
130	if (is_varlen73.1
'is_varlen' is 0
)
74
←
Taking false branch→
131	{
132		const int batch_size = ccv_nnc_tensor_count(q_seq_offsets->info) - 1;
133		assert(batch_size > 0)((void) sizeof ((batch_size > 0) ? 1 : 0), __extension__ (
{ if (batch_size > 0) ; else __assert_fail ("batch_size > 0"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 133, __extension__ __PRETTY_FUNCTION__); }));
134		assert(ccv_nnc_tensor_count(kv_seq_offsets->info) == batch_size + 1)((void) sizeof ((ccv_nnc_tensor_count(kv_seq_offsets->info
) == batch_size + 1) ? 1 : 0), __extension__ ({ if (ccv_nnc_tensor_count
(kv_seq_offsets->info) == batch_size + 1) ; else __assert_fail
 ("ccv_nnc_tensor_count(kv_seq_offsets->info) == batch_size + 1"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 134, __extension__ __PRETTY_FUNCTION__); }));
135		assert(qdim[0] == 1)((void) sizeof ((qdim[0] == 1) ? 1 : 0), __extension__ ({ if (
qdim[0] == 1) ; else __assert_fail ("qdim[0] == 1", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 135, __extension__ __PRETTY_FUNCTION__); }));
136		assert(kdim[0] == 1)((void) sizeof ((kdim[0] == 1) ? 1 : 0), __extension__ ({ if (
kdim[0] == 1) ; else __assert_fail ("kdim[0] == 1", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 136, __extension__ __PRETTY_FUNCTION__); }));
137		assert(vdim[0] == 1)((void) sizeof ((vdim[0] == 1) ? 1 : 0), __extension__ ({ if (
vdim[0] == 1) ; else __assert_fail ("vdim[0] == 1", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 137, __extension__ __PRETTY_FUNCTION__); }));
138		assert(cdim[0] == 1)((void) sizeof ((cdim[0] == 1) ? 1 : 0), __extension__ ({ if (
cdim[0] == 1) ; else __assert_fail ("cdim[0] == 1", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 138, __extension__ __PRETTY_FUNCTION__); }));
139		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"
, 139, __extension__ __PRETTY_FUNCTION__); }));
140		assert(cdim[2] == qdim[2])((void) sizeof ((cdim[2] == qdim[2]) ? 1 : 0), __extension__ (
{ if (cdim[2] == qdim[2]) ; else __assert_fail ("cdim[2] == qdim[2]"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 140, __extension__ __PRETTY_FUNCTION__); }));
141		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"
, 141, __extension__ __PRETTY_FUNCTION__); }));
142		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"
, 142, __extension__ __PRETTY_FUNCTION__); }));
143		const int* const q_offset = q_seq_offsets->data.i32;
144		const int* const kv_offset = kv_seq_offsets->data.i32;
145		assert(q_offset[0] == 0)((void) sizeof ((q_offset[0] == 0) ? 1 : 0), __extension__ ({
 if (q_offset[0] == 0) ; else __assert_fail ("q_offset[0] == 0"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 145, __extension__ __PRETTY_FUNCTION__); }));
146		assert(kv_offset[0] == 0)((void) sizeof ((kv_offset[0] == 0) ? 1 : 0), __extension__ (
{ if (kv_offset[0] == 0) ; else __assert_fail ("kv_offset[0] == 0"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 146, __extension__ __PRETTY_FUNCTION__); }));
147		assert(q_offset[batch_size] == qdim[1])((void) sizeof ((q_offset[batch_size] == qdim[1]) ? 1 : 0), __extension__
 ({ if (q_offset[batch_size] == qdim[1]) ; else __assert_fail
 ("q_offset[batch_size] == qdim[1]", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 147, __extension__ __PRETTY_FUNCTION__); }));
148		assert(kv_offset[batch_size] == kdim[1])((void) sizeof ((kv_offset[batch_size] == kdim[1]) ? 1 : 0), __extension__
 ({ if (kv_offset[batch_size] == kdim[1]) ; else __assert_fail
 ("kv_offset[batch_size] == kdim[1]", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 148, __extension__ __PRETTY_FUNCTION__); }));
149		for (i[0] = 0; i[0] < batch_size; i[0]++)
150		{
151			const int q_start = q_offset[i[0]];
152			const int q_end = q_offset[i[0] + 1];
153			const int k_start = kv_offset[i[0]];
154			const int k_end = kv_offset[i[0] + 1];
155			assert(q_start <= q_end)((void) sizeof ((q_start <= q_end) ? 1 : 0), __extension__
 ({ if (q_start <= q_end) ; else __assert_fail ("q_start <= q_end"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 155, __extension__ __PRETTY_FUNCTION__); }));
156			assert(k_start <= k_end)((void) sizeof ((k_start <= k_end) ? 1 : 0), __extension__
 ({ if (k_start <= k_end) ; else __assert_fail ("k_start <= k_end"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 156, __extension__ __PRETTY_FUNCTION__); }));
157			const int R = q_end - q_start;
158			const int K = k_end - k_start;
159			assert(R > 0)((void) sizeof ((R > 0) ? 1 : 0), __extension__ ({ if (R >
 0) ; else __assert_fail ("R > 0", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 159, __extension__ __PRETTY_FUNCTION__); }));
160			assert(K > 0)((void) sizeof ((K > 0) ? 1 : 0), __extension__ ({ if (K >
 0) ; else __assert_fail ("K > 0", "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 160, __extension__ __PRETTY_FUNCTION__); }));
161			assert(R <= cmd.info.scaled_dot_product_attention.max_seqlen_q)((void) sizeof ((R <= cmd.info.scaled_dot_product_attention
.max_seqlen_q) ? 1 : 0), __extension__ ({ if (R <= cmd.info
.scaled_dot_product_attention.max_seqlen_q) ; else __assert_fail
 ("R <= cmd.info.scaled_dot_product_attention.max_seqlen_q"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 161, __extension__ __PRETTY_FUNCTION__); }));
162			assert(K <= cmd.info.scaled_dot_product_attention.max_seqlen_kv)((void) sizeof ((K <= cmd.info.scaled_dot_product_attention
.max_seqlen_kv) ? 1 : 0), __extension__ ({ if (K <= cmd.info
.scaled_dot_product_attention.max_seqlen_kv) ; else __assert_fail
 ("K <= cmd.info.scaled_dot_product_attention.max_seqlen_kv"
, "scaled_dot_product_attention/ccv_nnc_scaled_dot_product_attention_cpu_ref.c"
, 162, __extension__ __PRETTY_FUNCTION__); }));
163			const float* const qp0 = qp + q_start * qstride[1];
164			const float* const kp0 = kp + k_start * kstride[1];
165			const float* const vp0 = vp + k_start * vstride[1];
166			float* const cp0 = cp + q_start * cstride[1];
167			for (i[1] = 0; i[1] < qdim[2]; i[1]++)
168			{
169				const float* const qp1 = qp0 + i[1] * qstride[2];
170				const float* const kp1 = kp0 + (i[1] / h_h_k_ratio) * kstride[2];
171				const float* const vp1 = vp0 + (i[1] / h_h_k_ratio) * vstride[2];
172				float* const cp1 = cp0 + i[1] * cstride[2];
173				const double sink = attention_sinks ? (double)sinkp[i[1] * sink_head_stride] : 0;
174				parallel_for(x, R){ int x; for ((x) = 0; (x) < (R); (x)++) { {
175					int y, k;
176					const float* const qp2 = qp1 + x * qstride[1];
177					float* const cp2 = cp1 + x * cstride[1];
178					float* const qk0 = qk + x * K;
179					for (y = 0; y < K; y++)
180					{
181						const float* const kp2 = kp1 + y * kstride[1];
182						float v = 0;
183						for (k = 0; k < qdim[3]; k++)
184							v += qp2[k * qstride[3]] * kp2[k * kstride[3]];
185						qk0[y] = scale * v;
186					}
187					if (is_causal)
188					{
189						const int x_end = ccv_max(x - R + K + 1, 0)({ typeof (x - R + K + 1) _a = (x - R + K + 1); typeof (0) _b
 = (0); (_a > _b) ? _a : _b; });
190						for (y = x_end; y < K; y++)
191							qk0[y] = 0;
192						double maxval = attention_sinks ? sink : qk0[0];
193						for (y = attention_sinks ? 0 : 1; y < x_end; y++)
194							if (qk0[y] > maxval)
195								maxval = qk0[y];
196						double sumval = attention_sinks ? expf(sink - maxval) : 0;
197						for (y = 0; y < x_end; y++)
198							sumval += (qk0[y] = expf(qk0[y] - maxval));
199						sumval = 1.0 / sumval;
200						for (y = 0; y < x_end; y++)
201							qk0[y] *= sumval;
202					} else {
203						double maxval = attention_sinks ? sink : qk0[0];
204						for (y = attention_sinks ? 0 : 1; y < K; y++)
205							if (qk0[y] > maxval)
206								maxval = qk0[y];
207						double sumval = attention_sinks ? expf(sink - maxval) : 0;
208						for (y = 0; y < K; y++)
209							sumval += (qk0[y] = expf(qk0[y] - maxval));
210						sumval = 1.0 / sumval;
211						for (y = 0; y < K; y++)
212							qk0[y] *= sumval;
213					}
214					for (k = 0; k < vdim[3]; k++)
215						cp2[k * cstride[3]] = 0;
216					for (y = 0; y < K; y++)
217					{
218						const float* const vp2 = vp1 + y * vstride[1];
219						const float v = qk0[y];
220						for (k = 0; k < vdim[3]; k++)
221							cp2[k * cstride[3]] += v * vp2[k * vstride[3]];
222					}
223				} parallel_endfor} }
224			}
225		}
226		return CCV_NNC_EXEC_SUCCESS;
227	}
228	for (i[0] = 0; i[0] < qdim[0]; i[0]++)
75
←
Assuming the condition is true→
76
←
Loop condition is true.  Entering loop body→
229	{
230		const float* const qp0 = qp + i[0] * qstride[0];
231		const float* const kp0 = kp + i[0] * kstride[0];
232		const float* const vp0 = vp + i[0] * vstride[0];
233		const float* const amp0 = amp && amdim[0] > 1 ? amp + i[0] * amstride[0] : amp;
77
←
Assuming 'amp' is null→
234		float* const cp0 = cp + i[0] * cstride[0];
235		for (i[1] = 0; i[1] < qdim[2]; i[1]++)
78
←
Assuming the condition is true→
79
←
Loop condition is true.  Entering loop body→
236		{
237			const float* const qp1 = qp0 + i[1] * qstride[2];
238			const float* const kp1 = kp0 + (i[1] / h_h_k_ratio) * kstride[2];
239			const float* const vp1 = vp0 + (i[1] / h_h_k_ratio) * vstride[2];
240			const float* const amp1 = amp79.1
'amp' is null
 && amdim[1] > 1 ? amp0 + i[1] * amstride[1] : amp0;
241			float* const cp1 = cp0 + i[1] * cstride[2];
242			const double sink = attention_sinks79.2
'attention_sinks' is not equal to 0
 ? (double)sinkp[i[1] * sink_head_stride] : 0;
80
←
'?' condition is true→
243			// Compute Q @ K^T
244			parallel_for(x, qdim[1]){ int x; for ((x) = 0; (x) < (qdim[1]); (x)++) { {
81
←
Assuming the condition is true→
82
←
Loop condition is true.  Entering loop body→
245				int y, k;
246				const float* const qp2 = qp1 + x * qstride[1];
247				float* const cp2 = cp1 + x * cstride[1];
248				float* const qk0 = qk + x * kdim[1];
249				const float* const amp2 = amp182.1
'amp1' is null
 ? amp1 + x * amstride[2] : 0;
83
←
'?' condition is false→
84
←
'amp2' initialized to a null pointer value→
250				if (attn_mask84.1
'attn_mask' is non-null
)
85
←
Taking true branch→
251				{
252					for (y = 0; y < kdim[1]; y++)
86
←
Assuming the condition is true→
87
←
Loop condition is true.  Entering loop body→
253					{
254						const float* const kp2 = kp1 + y * kstride[1];
255						float v = 0;
256						for (k = 0; k < qdim[3]; k++)
88
←
Assuming the condition is false→
89
←
Loop condition is false. Execution continues on line 258→
257							v += qp2[k * qstride[3]] * kp2[k * kstride[3]];
258						qk0[y] = scale * v + amp2[y * amstride[3]];
90
←
Array access (from variable 'amp2') results in a null pointer dereference
259					}
260				} else {
261					for (y = 0; y < kdim[1]; y++)
262					{
263						const float* const kp2 = kp1 + y * kstride[1];
264						float v = 0;
265						for (k = 0; k < qdim[3]; k++)
266							v += qp2[k * qstride[3]] * kp2[k * kstride[3]];
267						qk0[y] = scale * v;
268					}
269				}
270				// Compute softmax on qk.
271				if (is_causal)
272				{
273					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; });
274					for (y = x_end; y < kdim[1]; y++)
275						qk0[y] = 0;
276					double maxval = attention_sinks ? sink : qk0[0];
277					for (y = attention_sinks ? 0 : 1; y < x_end; y++)
278						if (qk0[y] > maxval)
279							maxval = qk0[y];
280					double sumval = attention_sinks ? expf(sink - maxval) : 0;
281					for (y = 0; y < x_end; y++)
282						sumval += (qk0[y] = expf(qk0[y] - maxval));
283					sumval = 1.0 / sumval;
284					for (y = 0; y < x_end; y++)
285						qk0[y] *= sumval;
286				} else {
287					double maxval = attention_sinks ? sink : qk0[0];
288					for (y = attention_sinks ? 0 : 1; y < kdim[1]; y++)
289						if (qk0[y] > maxval)
290							maxval = qk0[y];
291					double sumval = attention_sinks ? expf(sink - maxval) : 0;
292					for (y = 0; y < kdim[1]; y++)
293						sumval += (qk0[y] = expf(qk0[y] - maxval));
294					sumval = 1.0 / sumval;
295					for (y = 0; y < kdim[1]; y++)
296						qk0[y] *= sumval;
297				}
298				for (k = 0; k < vdim[3]; k++)
299					cp2[k * cstride[3]] = 0;
300				for (y = 0; y < kdim[1]; y++)
301				{
302					const float* const vp2 = vp1 + y * vstride[1];
303					const float v = qk0[y];
304					for (k = 0; k < vdim[3]; k++)
305						cp2[k * cstride[3]] += v * vp2[k * vstride[3]];
306				}
307			} parallel_endfor} }
308		}
309	}
310	if (w)
311	{
312		const int num_heads = cdim[2];
313		ccv_nnc_tensor_view_t* const d = (ccv_nnc_tensor_view_t*)outputs[0];
314		const int w_nd = ccv_nnc_tensor_nd(w->info.dim);
315		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"
, 315, __extension__ __PRETTY_FUNCTION__); }));
316		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"
, 316, __extension__ __PRETTY_FUNCTION__); }));
317		const int d_nd = ccv_nnc_tensor_nd(d->info.dim);
318		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"
, 318, __extension__ __PRETTY_FUNCTION__); }));
319		int ddim[CCV_NNC_MAX_DIM_ALLOC(12)];
320		int dstride[CCV_NNC_MAX_DIM_ALLOC(12)];
321		ccv_nnc_tensor_view_get_dim(d, ddim);
322		ccv_nnc_tensor_view_get_stride(d, dstride);
323		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"
, 323, __extension__ __PRETTY_FUNCTION__); }));
324		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"
, 324, __extension__ __PRETTY_FUNCTION__); }));
325		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"
, 325, __extension__ __PRETTY_FUNCTION__); }));
326		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"
, 326, __extension__ __PRETTY_FUNCTION__); }));
327		float* const dp = d->data.f32;
328		const float* const wp = w->data.f32;
329		const float* const cp = c->data.f32;
330		if (bias)
331		{
332			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"
, 332, __extension__ __PRETTY_FUNCTION__); }));
333			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"
, 333, __extension__ __PRETTY_FUNCTION__); }));
334			const float* const biasp = bias->data.f32;
335			for (i[0] = 0; i[0] < ddim[1]; i[0]++)
336			{
337				const float* const cp0 = cp + i[0] * cstride[0];
338				float* const dp0 = dp + i[0] * dstride[1];
339				parallel_for(y, ddim[2]){ int y; for ((y) = 0; (y) < (ddim[2]); (y)++) { {
340					int x, j, k;
341					const float* const cp1 = cp0 + y * cstride[1];
342					float* const dp1 = dp0 + y * dstride[2];
343					for (x = 0; x < ddim[3]; x++)
344					{
345						const float* const wp0 = wp + x * ddim[3];
346						float v = biasp[x];
347						for (j = 0; j < num_heads; j++)
348						{
349							const float* const cp2 = cp1 + j * cstride[2];
350							for (k = 0; k < cdim[3]; k++)
351								v += wp0[j * cdim[3] + k] * cp2[k * cstride[3]];
352						}
353						dp1[x * dstride[3]] = v;
354					}
355				} parallel_endfor} }
356			}
357		} else {
358			for (i[0] = 0; i[0] < ddim[1]; i[0]++)
359			{
360				const float* const cp0 = cp + i[0] * cstride[0];
361				float* const dp0 = dp + i[0] * dstride[1];
362				parallel_for(y, ddim[2]){ int y; for ((y) = 0; (y) < (ddim[2]); (y)++) { {
363					int x, j, k;
364					const float* const cp1 = cp0 + y * cstride[1];
365					float* const dp1 = dp0 + y * dstride[2];
366					for (x = 0; x < ddim[3]; x++)
367					{
368						const float* const wp0 = wp + x * ddim[3];
369						float v = 0;
370						for (j = 0; j < num_heads; j++)
371						{
372							const float* const cp2 = cp1 + j * cstride[2];
373							for (k = 0; k < cdim[3]; k++)
374								v += wp0[j * cdim[3] + k] * cp2[k * cstride[3]];
375						}
376						dp1[x * dstride[3]] = v;
377					}
378				} parallel_endfor} }
379			}
380		}
381	}
382	return CCV_NNC_EXEC_SUCCESS;
383}
384 
385static 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)
386{
387	// Assuming no saved_softmax, we need to recompute from q, k, v.
388	// We cannot do this with masks (yet).
389	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"
, 389, __extension__ __PRETTY_FUNCTION__); }));
390	if (cmd.info.scaled_dot_product_attention.is_varlen)
391		return CCV_NNC_EXEC_INVALID;
392	if (cmd.info.scaled_dot_product_attention.attention_sinks)
393		return CCV_NNC_EXEC_INVALID;
394	ccv_nnc_tensor_view_t* const g = (ccv_nnc_tensor_view_t*)inputs[0];
395	ccv_nnc_tensor_view_t* const q = (ccv_nnc_tensor_view_t*)inputs[3];
396	ccv_nnc_tensor_view_t* const k = (ccv_nnc_tensor_view_t*)inputs[4];
397	ccv_nnc_tensor_view_t* const v = (ccv_nnc_tensor_view_t*)inputs[5];
398	ccv_nnc_tensor_view_t* const dq = (ccv_nnc_tensor_view_t*)outputs[0];
399	ccv_nnc_tensor_view_t* const dk = (ccv_nnc_tensor_view_t*)outputs[1];
400	ccv_nnc_tensor_view_t* const dv = (ccv_nnc_tensor_view_t*)outputs[2];
401	const int q_nd = ccv_nnc_tensor_nd(q->info.dim);
402	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"
, 402, __extension__ __PRETTY_FUNCTION__); }));
403	const int k_nd = ccv_nnc_tensor_nd(k->info.dim);
404	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"
, 404, __extension__ __PRETTY_FUNCTION__); }));
405	const int v_nd = ccv_nnc_tensor_nd(v->info.dim);
406	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"
, 406, __extension__ __PRETTY_FUNCTION__); }));
407	const int g_nd = ccv_nnc_tensor_nd(g->info.dim);
408	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"
, 408, __extension__ __PRETTY_FUNCTION__); }));
409	const int dq_nd = ccv_nnc_tensor_nd(dq->info.dim);
410	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"
, 410, __extension__ __PRETTY_FUNCTION__); }));
411	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"
, 411, __extension__ __PRETTY_FUNCTION__); }));
412	const int dk_nd = ccv_nnc_tensor_nd(dk->info.dim);
413	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"
, 413, __extension__ __PRETTY_FUNCTION__); }));
414	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"
, 414, __extension__ __PRETTY_FUNCTION__); }));
415	const int dv_nd = ccv_nnc_tensor_nd(dv->info.dim);
416	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"
, 416, __extension__ __PRETTY_FUNCTION__); }));
417	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"
, 417, __extension__ __PRETTY_FUNCTION__); }));
418	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"
, 418, __extension__ __PRETTY_FUNCTION__); }));
419	// Assuming this is float 32.
420	int qdim[CCV_NNC_MAX_DIM_ALLOC(12)];
421	int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
422	int vdim[CCV_NNC_MAX_DIM_ALLOC(12)];
423	int gdim[CCV_NNC_MAX_DIM_ALLOC(12)];
424	int dqdim[CCV_NNC_MAX_DIM_ALLOC(12)];
425	int dkdim[CCV_NNC_MAX_DIM_ALLOC(12)];
426	int dvdim[CCV_NNC_MAX_DIM_ALLOC(12)];
427	ccv_nnc_tensor_view_get_dim(q, qdim);
428	ccv_nnc_tensor_view_get_dim(k, kdim);
429	ccv_nnc_tensor_view_get_dim(v, vdim);
430	ccv_nnc_tensor_view_get_dim(g, gdim);
431	ccv_nnc_tensor_view_get_dim(dq, dqdim);
432	ccv_nnc_tensor_view_get_dim(dk, dkdim);
433	ccv_nnc_tensor_view_get_dim(dv, dvdim);
434	if (q_nd == 3)
435	{
436		qdim[0] = qdim[1], qdim[1] = qdim[2], qdim[2] = 1;
437		kdim[0] = kdim[1], kdim[1] = kdim[2], kdim[2] = 1;
438		vdim[0] = vdim[1], vdim[1] = vdim[2], vdim[2] = 1;
439		gdim[0] = gdim[1], gdim[1] = gdim[2], gdim[2] = 1;
440		dqdim[0] = dqdim[1], dqdim[1] = dqdim[2], dqdim[2] = 1;
441		dkdim[0] = dkdim[1], dkdim[1] = dkdim[2], dkdim[2] = 1;
442		dvdim[0] = dvdim[1], dvdim[1] = dvdim[2], dvdim[2] = 1;
443	}
444	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"
, 444, __extension__ __PRETTY_FUNCTION__); }));
445	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"
, 445, __extension__ __PRETTY_FUNCTION__); }));
446	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"
, 446, __extension__ __PRETTY_FUNCTION__); }));
447	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"
, 447, __extension__ __PRETTY_FUNCTION__); }));
448	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"
, 448, __extension__ __PRETTY_FUNCTION__); }));
449	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"
, 449, __extension__ __PRETTY_FUNCTION__); }));
450	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"
, 450, __extension__ __PRETTY_FUNCTION__); }));
451	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"
, 451, __extension__ __PRETTY_FUNCTION__); }));
452	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"
, 452, __extension__ __PRETTY_FUNCTION__); }));
453	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"
, 453, __extension__ __PRETTY_FUNCTION__); })); // Need to change this logic for CCV_NNC_MAX_DIM == other number.
454	int qstride[CCV_NNC_MAX_DIM_ALLOC(12)];
455	int kstride[CCV_NNC_MAX_DIM_ALLOC(12)];
456	int vstride[CCV_NNC_MAX_DIM_ALLOC(12)];
457	int gstride[CCV_NNC_MAX_DIM_ALLOC(12)];
458	int dqstride[CCV_NNC_MAX_DIM_ALLOC(12)];
459	int dkstride[CCV_NNC_MAX_DIM_ALLOC(12)];
460	int dvstride[CCV_NNC_MAX_DIM_ALLOC(12)];
461	ccv_nnc_tensor_view_get_stride(q, qstride);
462	ccv_nnc_tensor_view_get_stride(k, kstride);
463	ccv_nnc_tensor_view_get_stride(v, vstride);
464	ccv_nnc_tensor_view_get_stride(g, gstride);
465	ccv_nnc_tensor_view_get_stride(dq, dqstride);
466	ccv_nnc_tensor_view_get_stride(dk, dkstride);
467	ccv_nnc_tensor_view_get_stride(dv, dvstride);
468	if (q_nd == 3)
469	{
470		qstride[0] = qstride[1], qstride[1] = qstride[2], qstride[2] = qstride[3];
471		kstride[0] = kstride[1], kstride[1] = kstride[2], kstride[2] = kstride[3];
472		vstride[0] = vstride[1], vstride[1] = vstride[2], vstride[2] = vstride[3];
473		gstride[0] = gstride[1], gstride[1] = gstride[2], gstride[2] = gstride[3];
474		dqstride[0] = dqstride[1], dqstride[1] = dqstride[2], dqstride[2] = dqstride[3];
475		dkstride[0] = dkstride[1], dkstride[1] = dkstride[2], dkstride[2] = dkstride[3];
476		dvstride[0] = dvstride[1], dvstride[1] = dvstride[2], dvstride[2] = dvstride[3];
477	}
478	int i[CCV_NNC_MAX_DIM(2) + 2];
479	float* qk = ccv_nnc_stream_context_get_workspace(stream_context, sizeof(float) * 2 * kdim[1], CCV_TENSOR_CPU_MEMORY);
480	const float* const qp = q->data.f32;
481	const float* const kp = k->data.f32;
482	const float* const vp = v->data.f32;
483	const float* const gp = g->data.f32;
484	float* const dqp = dq->data.f32;
485	float* const dkp = dk->data.f32;
486	float* const dvp = dv->data.f32;
487	const float scale = cmd.info.scaled_dot_product_attention.scale;
488	const int is_causal = cmd.info.scaled_dot_product_attention.is_causal;
489	const int h_h_k_ratio = qdim[2] / kdim[2];
490	for (i[0] = 0; i[0] < qdim[0]; i[0]++)
491	{
492		const float* const qp0 = qp + i[0] * qstride[0];
493		const float* const kp0 = kp + i[0] * kstride[0];
494		const float* const vp0 = vp + i[0] * vstride[0];
495		const float* const gp0 = gp + i[0] * gstride[0];
496		float* const dqp0 = dqp + i[0] * dqstride[0];
497		float* const dkp0 = dkp + i[0] * dkstride[0];
498		float* const dvp0 = dvp + i[0] * dvstride[0];
499		for (i[1] = 0; i[1] < qdim[2]; i[1]++)
500		{
501			const float* const qp1 = qp0 + i[1] * qstride[2];
502			const float* const kp1 = kp0 + (i[1] / h_h_k_ratio) * kstride[2];
503			const float* const vp1 = vp0 + (i[1] / h_h_k_ratio) * vstride[2];
504			const float* const gp1 = gp0 + i[1] * gstride[2];
505			float* const dqp1 = dqp0 + i[1] * dqstride[2];
506			float* const dkp1 = dkp0 + (i[1] / h_h_k_ratio) * dkstride[2];
507			float* const dvp1 = dvp0 + (i[1] / h_h_k_ratio) * dvstride[2];
508			// Compute Q @ K^T
509			int x, y, k;
510			for (x = 0; x < qdim[1]; x++)
511			{
512				float* const dqp2 = dqp1 + x * dqstride[1];
513				for (k = 0; k < qdim[3]; k++)
514					dqp2[k * dqstride[3]] = 0;
515			}
516			// Only zero out when it is at 0-index.
517			if (i[1] % h_h_k_ratio == 0)
518				for (y = 0; y < kdim[1]; y++)
519				{
520					float* const dkp2 = dkp1 + y * dkstride[1];
521					for (k = 0; k < qdim[3]; k++)
522						dkp2[k * dkstride[3]] = 0;
523				}
524			// Only zero out when it is at 0-index.
525			if (i[1] % h_h_k_ratio == 0)
526				for (y = 0; y < kdim[1]; y++)
527				{
528					float* const dvp2 = dvp1 + y * dvstride[1];
529					for (k = 0; k < vdim[3]; k++)
530						dvp2[k * dvstride[3]] = 0;
531				}
532			for (x = 0; x < qdim[1]; x++)
533			{
534				const float* const qp2 = qp1 + x * qstride[1];
535				const float* const gp2 = gp1 + x * gstride[1];
536				float* const qk0 = qk;
537				float* const qks0 = qk + kdim[1];
538				for (y = 0; y < kdim[1]; y++)
539				{
540					const float* const kp2 = kp1 + y * kstride[1];
541					float v = 0;
542					for (k = 0; k < qdim[3]; k++)
543						v += qp2[k * qstride[3]] * kp2[k * kstride[3]];
544					qk0[y] = scale * v;
545				}
546				// Compute softmax on qk.
547				if (is_causal)
548				{
549					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; });
550					for (y = x_end; y < kdim[1]; y++)
551						qk0[y] = 0;
552					double maxval = qk0[0];
553					for (y = 1; y < x_end; y++)
554						if (qk0[y] > maxval)
555							maxval = qk0[y];
556					double sumval = 0;
557					for (y = 0; y < x_end; y++)
558						sumval += (qk0[y] = expf(qk0[y] - maxval));
559					sumval = 1.0 / sumval;
560					for (y = 0; y < x_end; y++)
561						qk0[y] *= sumval;
562				} else {
563					double maxval = qk0[0];
564					for (y = 1; y < kdim[1]; y++)
565						if (qk0[y] > maxval)
566							maxval = qk0[y];
567					double sumval = 0;
568					for (y = 0; y < kdim[1]; y++)
569						sumval += (qk0[y] = expf(qk0[y] - maxval));
570					sumval = 1.0 / sumval;
571					for (y = 0; y < kdim[1]; y++)
572						qk0[y] *= sumval;
573				}
574				for (y = 0; y < kdim[1]; y++)
575				{
576					float* const dvp2 = dvp1 + y * dvstride[1];
577					const float v = qk0[y];
578					for (k = 0; k < vdim[3]; k++)
579						dvp2[k * dvstride[3]] += v * gp2[k * gstride[3]];
580				}
581				double sumval = 0;
582				for (y = 0; y < kdim[1]; y++)
583				{
584					const float* const vp2 = vp1 + y * vstride[1];
585					float v = 0;
586					for (k = 0; k < vdim[3]; k++)
587						v += gp2[k * gstride[3]] * vp2[k * vstride[3]];
588					qks0[y] = v;
589					sumval += v * qk0[y];
590				}
591				for (y = 0; y < kdim[1]; y++)
592					qk0[y] = (qks0[y] - sumval) * qk0[y];
593				float* const dqp2 = dqp1 + x * dqstride[1];
594				for (y = 0; y < kdim[1]; y++)
595				{
596					const float* const kp2 = kp1 + y * kstride[1];
597					float* const dkp2 = dkp1 + y * dkstride[1];
598					const float v = scale * qk0[y];
599					for (k = 0; k < qdim[3]; k++)
600					{
601						dqp2[k * dqstride[3]] += v * kp2[k * kstride[3]];
602						dkp2[k * dkstride[3]] += v * qp2[k * qstride[3]];
603					}
604				}
605			}
606		}
607	}
608	return CCV_NNC_EXEC_SUCCESS;
609}
610 
611REGISTER_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)
612{
613	registry->tensor_formats = CCV_TENSOR_FORMAT_NHWC;
614	registry->tensor_datatypes = CCV_32F | CCV_32S;
615	registry->tensor_memory = CCV_TENSOR_CPU_MEMORY;
616	registry->algorithms = 1;
617	registry->exec = _ccv_nnc_scaled_dot_product_attention_forw;
618}
619 
620REGISTER_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)
621{
622	registry->tensor_formats = CCV_TENSOR_FORMAT_NHWC;
623	registry->tensor_datatypes = CCV_32F;
624	registry->tensor_memory = CCV_TENSOR_CPU_MEMORY;
625	registry->algorithms = 1;
626	registry->exec = _ccv_nnc_scaled_dot_product_attention_back;
627}