1 #include "HalideRuntimeQurt.h"
2 #include "mini_qurt.h"
3 #include "printer.h"
4 #include "runtime_internal.h"
5
6 using namespace Halide::Runtime::Internal::Qurt;
7
8 extern "C" {
9
halide_qurt_hvx_lock(void * user_context,int size)10 WEAK int halide_qurt_hvx_lock(void *user_context, int size) {
11 qurt_hvx_mode_t mode;
12 switch (size) {
13 case 64:
14 mode = QURT_HVX_MODE_64B;
15 break;
16 case 128:
17 mode = QURT_HVX_MODE_128B;
18 break;
19 default:
20 error(user_context) << "HVX lock size must be 64 or 128.\n";
21 return -1;
22 }
23
24 debug(user_context) << "QuRT: qurt_hvx_lock(" << mode << ") ->\n";
25 int result = qurt_hvx_lock(mode);
26 debug(user_context) << " " << result << "\n";
27 if (result != QURT_EOK) {
28 error(user_context) << "qurt_hvx_lock failed\n";
29 return -1;
30 }
31 return 0;
32 }
33
halide_qurt_hvx_unlock(void * user_context)34 WEAK int halide_qurt_hvx_unlock(void *user_context) {
35 debug(user_context) << "QuRT: qurt_hvx_unlock ->\n";
36 int result = qurt_hvx_unlock();
37 debug(user_context) << " " << result << "\n";
38 if (result != QURT_EOK) {
39 error(user_context) << "qurt_hvx_unlock failed\n";
40 return -1;
41 }
42
43 return 0;
44 }
45
halide_qurt_hvx_unlock_as_destructor(void * user_context,void *)46 WEAK void halide_qurt_hvx_unlock_as_destructor(void *user_context, void * /*obj*/) {
47 halide_qurt_hvx_unlock(user_context);
48 }
49
50 // These need to inline, otherwise the extern call with the ptr
51 // parameter breaks a lot of optimizations.
_halide_prefetch_2d(const void * ptr,int width_bytes,int height,int stride_bytes)52 WEAK_INLINE int _halide_prefetch_2d(const void *ptr, int width_bytes, int height, int stride_bytes) {
53 // Notes:
54 // - Prefetches can be queued up to 3 deep (MAX_PREFETCH)
55 // - If 3 are already pending, the oldest request is dropped
56 // - USR:PFA status bit is set to indicate that prefetches are in progress
57 // - A l2fetch with any subfield set to zero cancels all pending prefetches
58 // - The l2fetch starting address must be in mapped memory but the range
59 // prefetched can go into unmapped memory without raising an exception
60 const int dir = 1;
61 uint64_t desc =
62 (static_cast<uint64_t>(dir) << 48) |
63 (static_cast<uint64_t>(stride_bytes) << 32) |
64 (static_cast<uint64_t>(width_bytes) << 16) |
65 (static_cast<uint64_t>(height) << 0);
66 __asm__ __volatile__("l2fetch(%0,%1)"
67 :
68 : "r"(ptr), "r"(desc));
69 return 0;
70 }
71
_halide_prefetch(const void * ptr,int size)72 WEAK_INLINE int _halide_prefetch(const void *ptr, int size) {
73 _halide_prefetch_2d(ptr, size, 1, 1);
74 return 0;
75 }
76
77 struct hexagon_buffer_t_arg {
78 uint64_t device;
79 uint8_t *host;
80 };
81
_halide_hexagon_buffer_get_host(const hexagon_buffer_t_arg * buf)82 WEAK_INLINE uint8_t *_halide_hexagon_buffer_get_host(const hexagon_buffer_t_arg *buf) {
83 return buf->host;
84 }
85
_halide_hexagon_buffer_get_device(const hexagon_buffer_t_arg * buf)86 WEAK_INLINE uint64_t _halide_hexagon_buffer_get_device(const hexagon_buffer_t_arg *buf) {
87 return buf->device;
88 }
89 }
90