mirrors/FreeRDP
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fixed profiler API

Browse Source 
* Duplicate the name internally to avoid issues with stack
* Made API opaque and removed direct dereferencing of struct.
This commit is contained in:
Armin Novak 2018-11-15 17:52:43 +01:00
parent fe1a79759e
commit d2e1248b09
5 changed files with 80 additions and 96 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
include/freerdp/utils/profiler.h
View File

@ -23,18 +23,13 @@
#include <freerdp/api.h>
#include <freerdp/utils/stopwatch.h>
struct _PROFILER
{
char* name;
STOPWATCH* stopwatch;
};
typedef struct _PROFILER PROFILER;
#ifdef __cplusplus
extern "C" {
#endif
FREERDP_API PROFILER* profiler_create(char* name);
FREERDP_API PROFILER* profiler_create(const char* name);
FREERDP_API void profiler_free(PROFILER* profiler);
FREERDP_API void profiler_enter(PROFILER* profiler);
@ -49,7 +44,7 @@ FREERDP_API void profiler_print_footer(void);
#endif
#ifdef WITH_PROFILER
#define IF_PROFILER(then) then
#define PROFILER_RENAME(prof, name) do { profiler_free(prof); prof = profiler_create(name); } while(0)
#define PROFILER_DEFINE(prof) PROFILER* prof;
#define PROFILER_CREATE(prof,name) prof = profiler_create(name);
#define PROFILER_FREE(prof) profiler_free(prof);
@ -59,7 +54,7 @@ FREERDP_API void profiler_print_footer(void);
#define PROFILER_PRINT(prof) profiler_print(prof);
#define PROFILER_PRINT_FOOTER profiler_print_footer();
#else
#define IF_PROFILER(then) do { } while (0)
#define PROFILER_RENAME(prof, name) do { } while (0)
#define PROFILER_DEFINE(prof)
#define PROFILER_CREATE(prof,name) do { } while (0);

2
libfreerdp/codec/nsc_sse2.c
View File

@ -401,6 +401,6 @@ void nsc_init_sse2(NSC_CONTEXT* context)
if (!IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
return;
IF_PROFILER(context->priv->prof_nsc_encode->name = "nsc_encode_sse2");
PROFILER_RENAME(context->priv->prof_nsc_encode, "nsc_encode_sse2");
context->encode = nsc_encode_sse2;
}

41
libfreerdp/codec/rfx_neon.c
View File

@ -82,7 +82,6 @@ rfx_dwt_2d_decode_block_horiz_NEON(INT16 * l, INT16 * h, INT16 * dst, int subban
{
// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
int16x8_t l_n = vld1q_s16(l_ptr);
int16x8_t h_n = vld1q_s16(h_ptr);
int16x8_t h_n_m = vld1q_s16(h_ptr - 1);
@ -95,14 +94,12 @@ rfx_dwt_2d_decode_block_horiz_NEON(INT16 * l, INT16 * h, INT16 * dst, int subban
int16x8_t tmp_n = vaddq_s16(h_n, h_n_m);
tmp_n = vaddq_s16(tmp_n, vdupq_n_s16(1));
tmp_n = vshrq_n_s16(tmp_n, 1);
int16x8_t dst_n = vsubq_s16(l_n, tmp_n);
vst1q_s16(l_ptr, dst_n);
l_ptr += 8;
h_ptr += 8;
}
l_ptr -= subband_width;
h_ptr -= subband_width;
@ -110,14 +107,12 @@ rfx_dwt_2d_decode_block_horiz_NEON(INT16 * l, INT16 * h, INT16 * dst, int subban
for (n = 0; n < subband_width; n += 8)
{
// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
int16x8_t h_n = vld1q_s16(h_ptr);
h_n = vshlq_n_s16(h_n, 1);
int16x8x2_t dst_n;
dst_n.val[0] = vld1q_s16(l_ptr);
int16x8_t dst_n_p = vld1q_s16(l_ptr + 1);
if (n == subband_width - 8)
{
int16_t last = vgetq_lane_s16(dst_n_p, 6);
@ -126,11 +121,8 @@ rfx_dwt_2d_decode_block_horiz_NEON(INT16 * l, INT16 * h, INT16 * dst, int subban
dst_n.val[1] = vaddq_s16(dst_n_p, dst_n.val[0]);
dst_n.val[1] = vshrq_n_s16(dst_n.val[1], 1);
dst_n.val[1] = vaddq_s16(dst_n.val[1], h_n);
vst2q_s16(dst_ptr, dst_n);
l_ptr += 8;
h_ptr += 8;
dst_ptr += 16;
@ -145,7 +137,6 @@ rfx_dwt_2d_decode_block_vert_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband
INT16* l_ptr = l;
INT16* h_ptr = h;
INT16* dst_ptr = dst;
int total_width = subband_width + subband_width;
/* Even coefficients */
@ -154,11 +145,10 @@ rfx_dwt_2d_decode_block_vert_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband
for (x = 0; x < total_width; x += 8)
{
// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
int16x8_t l_n = vld1q_s16(l_ptr);
int16x8_t h_n = vld1q_s16(h_ptr);
int16x8_t tmp_n = vaddq_s16(h_n, vdupq_n_s16(1));;
if (n == 0)
tmp_n = vaddq_s16(tmp_n, h_n);
else
@ -166,15 +156,15 @@ rfx_dwt_2d_decode_block_vert_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband
int16x8_t h_n_m = vld1q_s16((h_ptr - total_width));
tmp_n = vaddq_s16(tmp_n, h_n_m);
}
tmp_n = vshrq_n_s16(tmp_n, 1);
tmp_n = vshrq_n_s16(tmp_n, 1);
int16x8_t dst_n = vsubq_s16(l_n, tmp_n);
vst1q_s16(dst_ptr, dst_n);
l_ptr += 8;
h_ptr += 8;
dst_ptr += 8;
}
dst_ptr += total_width;
}
@ -189,10 +179,9 @@ rfx_dwt_2d_decode_block_vert_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband
// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
int16x8_t h_n = vld1q_s16(h_ptr);
int16x8_t dst_n_m = vld1q_s16(dst_ptr - total_width);
h_n = vshlq_n_s16(h_n, 1);
int16x8_t tmp_n = dst_n_m;
if (n == subband_width - 1)
tmp_n = vaddq_s16(tmp_n, dst_n_m);
else
@ -200,14 +189,14 @@ rfx_dwt_2d_decode_block_vert_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband
int16x8_t dst_n_p = vld1q_s16((dst_ptr + total_width));
tmp_n = vaddq_s16(tmp_n, dst_n_p);
}
tmp_n = vshrq_n_s16(tmp_n, 1);
tmp_n = vshrq_n_s16(tmp_n, 1);
int16x8_t dst_n = vaddq_s16(tmp_n, h_n);
vst1q_s16(dst_ptr, dst_n);
h_ptr += 8;
dst_ptr += 8;
}
dst_ptr += total_width;
}
}
@ -217,24 +206,18 @@ rfx_dwt_2d_decode_block_NEON(INT16 * buffer, INT16 * idwt, int subband_width)
{
INT16* hl, * lh, * hh, * ll;
INT16* l_dst, * h_dst;
/* Inverse DWT in horizontal direction, results in 2 sub-bands in L, H order in tmp buffer idwt. */
/* The 4 sub-bands are stored in HL(0), LH(1), HH(2), LL(3) order. */
/* The lower part L uses LL(3) and HL(0). */
/* The higher part H uses LH(1) and HH(2). */
ll = buffer + subband_width * subband_width * 3;
hl = buffer;
l_dst = idwt;
rfx_dwt_2d_decode_block_horiz_NEON(ll, hl, l_dst, subband_width);
lh = buffer + subband_width * subband_width;
hh = buffer + subband_width * subband_width * 2;
h_dst = idwt + subband_width * subband_width * 2;
rfx_dwt_2d_decode_block_horiz_NEON(lh, hh, h_dst, subband_width);
/* Inverse DWT in vertical direction, results are stored in original buffer. */
rfx_dwt_2d_decode_block_vert_NEON(l_dst, h_dst, buffer, subband_width);
}
@ -251,11 +234,9 @@ void rfx_init_neon(RFX_CONTEXT * context)
if (IsProcessorFeaturePresent(PF_ARM_NEON_INSTRUCTIONS_AVAILABLE))
{
DEBUG_RFX("Using NEON optimizations");
IF_PROFILER(context->priv->prof_rfx_ycbcr_to_rgb->name = "rfx_decode_YCbCr_to_RGB_NEON");
IF_PROFILER(context->priv->prof_rfx_quantization_decode->name = "rfx_quantization_decode_NEON");
IF_PROFILER(context->priv->prof_rfx_dwt_2d_decode->name = "rfx_dwt_2d_decode_NEON");
PROFILER_RENAME(context->priv->prof_rfx_ycbcr_to_rgb, "rfx_decode_YCbCr_to_RGB_NEON");
PROFILER_RENAME(context->priv->prof_rfx_quantization_decode, "rfx_quantization_decode_NEON");
PROFILER_RENAME(context->priv->prof_rfx_dwt_2d_decode, "rfx_dwt_2d_decode_NEON");
context->quantization_decode = rfx_quantization_decode_NEON;
context->dwt_2d_decode = rfx_dwt_2d_decode_NEON;
}

8
libfreerdp/codec/rfx_sse2.c
View File

@ -455,10 +455,10 @@ void rfx_init_sse2(RFX_CONTEXT* context)
if (!IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
return;
IF_PROFILER(context->priv->prof_rfx_quantization_decode->name = "rfx_quantization_decode_sse2");
IF_PROFILER(context->priv->prof_rfx_quantization_encode->name = "rfx_quantization_encode_sse2");
IF_PROFILER(context->priv->prof_rfx_dwt_2d_decode->name = "rfx_dwt_2d_decode_sse2");
IF_PROFILER(context->priv->prof_rfx_dwt_2d_encode->name = "rfx_dwt_2d_encode_sse2");
PROFILER_RENAME(context->priv->prof_rfx_quantization_decode, "rfx_quantization_decode_sse2");
PROFILER_RENAME(context->priv->prof_rfx_quantization_encode, "rfx_quantization_encode_sse2");
PROFILER_RENAME(context->priv->prof_rfx_dwt_2d_decode, "rfx_dwt_2d_decode_sse2");
PROFILER_RENAME(context->priv->prof_rfx_dwt_2d_encode, "rfx_dwt_2d_encode_sse2");
context->quantization_decode = rfx_quantization_decode_sse2;
context->quantization_encode = rfx_quantization_encode_sse2;
context->dwt_2d_decode = rfx_dwt_2d_decode_sse2;

26
libfreerdp/utils/profiler.c
View File

@ -29,30 +29,38 @@
#define TAG FREERDP_TAG("utils")
PROFILER* profiler_create(char* name)
struct _PROFILER
{
PROFILER* profiler;
profiler = (PROFILER*) malloc(sizeof(PROFILER));
char* name;
STOPWATCH* stopwatch;
};
PROFILER* profiler_create(const char* name)
{
PROFILER* profiler = (PROFILER*) calloc(1, sizeof(PROFILER));
if (!profiler)
return NULL;
profiler->name = name;
profiler->name = _strdup(name);
profiler->stopwatch = stopwatch_create();
if (!profiler->stopwatch)
{
free(profiler);
return NULL;
}
if (!profiler->name || !profiler->stopwatch)
goto fail;
return profiler;
fail:
profiler_free(profiler);
return NULL;
}
void profiler_free(PROFILER* profiler)
{
if (profiler)
{
free(profiler->name);
stopwatch_free(profiler->stopwatch);
}
free(profiler);
}