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

[restrict] remove from in/out buffers

Browse Source
This commit is contained in:
Armin Novak 2023-10-11 13:56:23 +02:00 committed by akallabeth
parent 1761edf2f8
commit 4778203340
6 changed files with 46 additions and 59 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
libfreerdp/codec/progressive.c
View File

@ -762,14 +762,13 @@ static INLINE size_t progressive_rfx_get_band_h_count(size_t level)
return (64 + (1 << (level - 1))) >> level; return (64 + (1 << (level - 1))) >> level;
} }
static INLINE void progressive_rfx_dwt_2d_decode_block(INT16* WINPR_RESTRICT buffer, static INLINE void progressive_rfx_dwt_2d_decode_block(INT16* buffer, INT16* temp, size_t level)
INT16* WINPR_RESTRICT temp, size_t level)
{ {
size_t nDstStepX; size_t nDstStepX;
size_t nDstStepY; size_t nDstStepY;
INT16 *WINPR_RESTRICT HL, *WINPR_RESTRICT LH; INT16 *HL, *LH;
INT16 *WINPR_RESTRICT HH, *WINPR_RESTRICT LL; INT16 *HH, *LL;
INT16 *WINPR_RESTRICT L, *WINPR_RESTRICT H, *WINPR_RESTRICT LLx; INT16 *L, *H, *LLx;
const size_t nBandL = progressive_rfx_get_band_l_count(level); const size_t nBandL = progressive_rfx_get_band_l_count(level);
const size_t nBandH = progressive_rfx_get_band_h_count(level); const size_t nBandH = progressive_rfx_get_band_h_count(level);
@ -801,7 +800,7 @@ static INLINE void progressive_rfx_dwt_2d_decode_block(INT16* WINPR_RESTRICT buf
nBandL + nBandH); nBandL + nBandH);
} }
void rfx_dwt_2d_extrapolate_decode(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT temp) void rfx_dwt_2d_extrapolate_decode(INT16* buffer, INT16* temp)
{ {
WINPR_ASSERT(buffer); WINPR_ASSERT(buffer);
WINPR_ASSERT(temp); WINPR_ASSERT(temp);

37
libfreerdp/codec/rfx_dwt.c
View File

@ -25,17 +25,13 @@
#include "rfx_dwt.h" #include "rfx_dwt.h"
static void rfx_dwt_2d_decode_block(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT idwt, static void rfx_dwt_2d_decode_block(INT16* buffer, INT16* idwt, size_t subband_width)
int subband_width)
{ {
INT16 *WINPR_RESTRICT dst, *WINPR_RESTRICT l, *WINPR_RESTRICT h; INT16 *dst, *l, *h;
INT16 *WINPR_RESTRICT l_dst, *WINPR_RESTRICT h_dst; INT16 *l_dst, *h_dst;
INT16 *WINPR_RESTRICT hl, *WINPR_RESTRICT lh, *WINPR_RESTRICT hh, *WINPR_RESTRICT ll; INT16 *hl, *lh, *hh, *ll;
int total_width;
int x, y;
int n;
total_width = subband_width << 1; const size_t total_width = subband_width << 1;
/* Inverse DWT in horizontal direction, results in 2 sub-bands in L, H order in tmp buffer idwt. /* Inverse DWT in horizontal direction, results in 2 sub-bands in L, H order in tmp buffer idwt.
*/ */
@ -51,26 +47,28 @@ static void rfx_dwt_2d_decode_block(INT16* WINPR_RESTRICT buffer, INT16* WINPR_R
hh = buffer + subband_width * subband_width * 2; hh = buffer + subband_width * subband_width * 2;
h_dst = idwt + subband_width * subband_width * 2; h_dst = idwt + subband_width * subband_width * 2;
for (y = 0; y < subband_width; y++) for (size_t y = 0; y < subband_width; y++)
{ {
/* Even coefficients */ /* Even coefficients */
l_dst[0] = ll[0] - ((hl[0] + hl[0] + 1) >> 1); l_dst[0] = ll[0] - ((hl[0] + hl[0] + 1) >> 1);
h_dst[0] = lh[0] - ((hh[0] + hh[0] + 1) >> 1); h_dst[0] = lh[0] - ((hh[0] + hh[0] + 1) >> 1);
for (n = 1; n < subband_width; n++) for (size_t n = 1; n < subband_width; n++)
{ {
x = n << 1; const size_t x = n << 1;
l_dst[x] = ll[n] - ((hl[n - 1] + hl[n] + 1) >> 1); l_dst[x] = ll[n] - ((hl[n - 1] + hl[n] + 1) >> 1);
h_dst[x] = lh[n] - ((hh[n - 1] + hh[n] + 1) >> 1); h_dst[x] = lh[n] - ((hh[n - 1] + hh[n] + 1) >> 1);
} }
/* Odd coefficients */ /* Odd coefficients */
size_t n = 0;
for (n = 0; n < subband_width - 1; n++) for (n = 0; n < subband_width - 1; n++)
{ {
x = n << 1; const size_t x = n << 1;
l_dst[x + 1] = (hl[n] << 1) + ((l_dst[x] + l_dst[x + 2]) >> 1); l_dst[x + 1] = (hl[n] << 1) + ((l_dst[x] + l_dst[x + 2]) >> 1);
h_dst[x + 1] = (hh[n] << 1) + ((h_dst[x] + h_dst[x + 2]) >> 1); h_dst[x + 1] = (hh[n] << 1) + ((h_dst[x] + h_dst[x + 2]) >> 1);
} }
x = n << 1;
const size_t x = n << 1;
l_dst[x + 1] = (hl[n] << 1) + (l_dst[x]); l_dst[x + 1] = (hl[n] << 1) + (l_dst[x]);
h_dst[x + 1] = (hh[n] << 1) + (h_dst[x]); h_dst[x + 1] = (hh[n] << 1) + (h_dst[x]);
@ -84,7 +82,7 @@ static void rfx_dwt_2d_decode_block(INT16* WINPR_RESTRICT buffer, INT16* WINPR_R
} }
/* Inverse DWT in vertical direction, results are stored in original buffer. */ /* Inverse DWT in vertical direction, results are stored in original buffer. */
for (x = 0; x < total_width; x++) for (size_t x = 0; x < total_width; x++)
{ {
l = idwt + x; l = idwt + x;
h = idwt + x + subband_width * total_width; h = idwt + x + subband_width * total_width;
@ -92,7 +90,7 @@ static void rfx_dwt_2d_decode_block(INT16* WINPR_RESTRICT buffer, INT16* WINPR_R
*dst = *l - ((*h * 2 + 1) >> 1); *dst = *l - ((*h * 2 + 1) >> 1);
for (n = 1; n < subband_width; n++) for (size_t n = 1; n < subband_width; n++)
{ {
l += total_width; l += total_width;
h += total_width; h += total_width;
@ -110,7 +108,7 @@ static void rfx_dwt_2d_decode_block(INT16* WINPR_RESTRICT buffer, INT16* WINPR_R
} }
} }
void rfx_dwt_2d_decode(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_buffer) void rfx_dwt_2d_decode(INT16* buffer, INT16* dwt_buffer)
{ {
WINPR_ASSERT(buffer); WINPR_ASSERT(buffer);
WINPR_ASSERT(dwt_buffer); WINPR_ASSERT(dwt_buffer);
@ -120,8 +118,7 @@ void rfx_dwt_2d_decode(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_b
rfx_dwt_2d_decode_block(&buffer[0], dwt_buffer, 32); rfx_dwt_2d_decode_block(&buffer[0], dwt_buffer, 32);
} }
static void rfx_dwt_2d_encode_block(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt, static void rfx_dwt_2d_encode_block(INT16* buffer, INT16* dwt, UINT32 subband_width)
UINT32 subband_width)
{ {
INT16 *src, *l, *h; INT16 *src, *l, *h;
INT16 *l_src, *h_src; INT16 *l_src, *h_src;
@ -198,7 +195,7 @@ static void rfx_dwt_2d_encode_block(INT16* WINPR_RESTRICT buffer, INT16* WINPR_R
} }
} }
void rfx_dwt_2d_encode(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_buffer) void rfx_dwt_2d_encode(INT16* buffer, INT16* dwt_buffer)
{ {
WINPR_ASSERT(buffer); WINPR_ASSERT(buffer);
WINPR_ASSERT(dwt_buffer); WINPR_ASSERT(dwt_buffer);

9
libfreerdp/codec/rfx_dwt.h
View File

@ -24,11 +24,8 @@
#include <freerdp/codec/rfx.h> #include <freerdp/codec/rfx.h>
#include <freerdp/api.h> #include <freerdp/api.h>
FREERDP_LOCAL void rfx_dwt_2d_decode(INT16* WINPR_RESTRICT buffer, FREERDP_LOCAL void rfx_dwt_2d_decode(INT16* buffer, INT16* dwt_buffer);
INT16* WINPR_RESTRICT dwt_buffer); FREERDP_LOCAL void rfx_dwt_2d_encode(INT16* buffer, INT16* dwt_buffer);
FREERDP_LOCAL void rfx_dwt_2d_encode(INT16* WINPR_RESTRICT buffer, FREERDP_LOCAL void rfx_dwt_2d_extrapolate_decode(INT16* buffer, INT16* dwt_buffer);
INT16* WINPR_RESTRICT dwt_buffer);
FREERDP_LOCAL void rfx_dwt_2d_extrapolate_decode(INT16* WINPR_RESTRICT buffer,
INT16* WINPR_RESTRICT dwt_buffer);
#endif /* FREERDP_LIB_CODEC_RFX_DWT_H */ #endif /* FREERDP_LIB_CODEC_RFX_DWT_H */

32
libfreerdp/codec/rfx_neon.c
View File

@ -67,17 +67,16 @@ static void rfx_quantization_decode_NEON(INT16* buffer, const UINT32* WINPR_REST
static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_decode_block_horiz_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRICT h, rfx_dwt_2d_decode_block_horiz_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRICT h,
INT16* WINPR_RESTRICT dst, int subband_width) INT16* WINPR_RESTRICT dst, size_t subband_width)
{ {
int y, n;
INT16* l_ptr = l; INT16* l_ptr = l;
INT16* h_ptr = h; INT16* h_ptr = h;
INT16* dst_ptr = dst; INT16* dst_ptr = dst;
for (y = 0; y < subband_width; y++) for (size_t y = 0; y < subband_width; y++)
{ {
/* Even coefficients */ /* Even coefficients */
for (n = 0; n < subband_width; n += 8) for (size_t n = 0; n < subband_width; n += 8)
{ {
// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1); // dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
int16x8_t l_n = vld1q_s16(l_ptr); int16x8_t l_n = vld1q_s16(l_ptr);
@ -103,7 +102,7 @@ rfx_dwt_2d_decode_block_horiz_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRIC
h_ptr -= subband_width; h_ptr -= subband_width;
/* Odd coefficients */ /* Odd coefficients */
for (n = 0; n < subband_width; n += 8) for (size_t n = 0; n < subband_width; n += 8)
{ {
// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1); // dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
int16x8_t h_n = vld1q_s16(h_ptr); int16x8_t h_n = vld1q_s16(h_ptr);
@ -131,18 +130,17 @@ rfx_dwt_2d_decode_block_horiz_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRIC
static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_decode_block_vert_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRICT h, rfx_dwt_2d_decode_block_vert_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRICT h,
INT16* WINPR_RESTRICT dst, int subband_width) INT16* WINPR_RESTRICT dst, size_t subband_width)
{ {
int x, n;
INT16* l_ptr = l; INT16* l_ptr = l;
INT16* h_ptr = h; INT16* h_ptr = h;
INT16* dst_ptr = dst; INT16* dst_ptr = dst;
int total_width = subband_width + subband_width; const size_t total_width = subband_width + subband_width;
/* Even coefficients */ /* Even coefficients */
for (n = 0; n < subband_width; n++) for (size_t n = 0; n < subband_width; n++)
{ {
for (x = 0; x < total_width; x += 8) for (size_t x = 0; x < total_width; x += 8)
{ {
// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1); // dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
int16x8_t l_n = vld1q_s16(l_ptr); int16x8_t l_n = vld1q_s16(l_ptr);
@ -172,9 +170,9 @@ rfx_dwt_2d_decode_block_vert_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRICT
dst_ptr = dst + total_width; dst_ptr = dst + total_width;
/* Odd coefficients */ /* Odd coefficients */
for (n = 0; n < subband_width; n++) for (size_t n = 0; n < subband_width; n++)
{ {
for (x = 0; x < total_width; x += 8) for (size_t x = 0; x < total_width; x += 8)
{ {
// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1); // dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
int16x8_t h_n = vld1q_s16(h_ptr); int16x8_t h_n = vld1q_s16(h_ptr);
@ -203,7 +201,7 @@ rfx_dwt_2d_decode_block_vert_NEON(INT16* WINPR_RESTRICT l, INT16* WINPR_RESTRICT
static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
rfx_dwt_2d_decode_block_NEON(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT idwt, rfx_dwt_2d_decode_block_NEON(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT idwt,
int subband_width) size_t subband_width)
{ {
INT16 *hl, *lh, *hh, *ll; INT16 *hl, *lh, *hh, *ll;
INT16 *l_dst, *h_dst; INT16 *l_dst, *h_dst;
@ -224,7 +222,7 @@ rfx_dwt_2d_decode_block_NEON(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT
rfx_dwt_2d_decode_block_vert_NEON(l_dst, h_dst, buffer, subband_width); rfx_dwt_2d_decode_block_vert_NEON(l_dst, h_dst, buffer, subband_width);
} }
static void rfx_dwt_2d_decode_NEON(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_buffer) static void rfx_dwt_2d_decode_NEON(INT16* buffer, INT16* dwt_buffer)
{ {
rfx_dwt_2d_decode_block_NEON(buffer + 3840, dwt_buffer, 8); rfx_dwt_2d_decode_block_NEON(buffer + 3840, dwt_buffer, 8);
rfx_dwt_2d_decode_block_NEON(buffer + 3072, dwt_buffer, 16); rfx_dwt_2d_decode_block_NEON(buffer + 3072, dwt_buffer, 16);
@ -468,8 +466,7 @@ static INLINE size_t prfx_get_band_h_count(size_t level)
return (64 + (1 << (level - 1))) >> level; return (64 + (1 << (level - 1))) >> level;
} }
static INLINE void rfx_dwt_2d_decode_extrapolate_block_neon(INT16* WINPR_RESTRICT buffer, static INLINE void rfx_dwt_2d_decode_extrapolate_block_neon(INT16* buffer, INT16* temp,
INT16* WINPR_RESTRICT temp,
size_t level) size_t level)
{ {
size_t nDstStepX; size_t nDstStepX;
@ -511,8 +508,7 @@ static INLINE void rfx_dwt_2d_decode_extrapolate_block_neon(INT16* WINPR_RESTRIC
nBandL + nBandH); nBandL + nBandH);
} }
static void rfx_dwt_2d_extrapolate_decode_neon(INT16* WINPR_RESTRICT buffer, static void rfx_dwt_2d_extrapolate_decode_neon(INT16* buffer, INT16* temp)
INT16* WINPR_RESTRICT temp)
{ {
WINPR_ASSERT(buffer); WINPR_ASSERT(buffer);
WINPR_ASSERT(temp); WINPR_ASSERT(temp);

4
libfreerdp/codec/rfx_sse2.c
View File

@ -319,7 +319,7 @@ rfx_dwt_2d_decode_block_sse2(INT16* buffer, INT16* idwt, int subband_width)
rfx_dwt_2d_decode_block_vert_sse2(l_dst, h_dst, buffer, subband_width); rfx_dwt_2d_decode_block_vert_sse2(l_dst, h_dst, buffer, subband_width);
} }
static void rfx_dwt_2d_decode_sse2(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_buffer) static void rfx_dwt_2d_decode_sse2(INT16* buffer, INT16* dwt_buffer)
{ {
WINPR_ASSERT(buffer); WINPR_ASSERT(buffer);
WINPR_ASSERT(dwt_buffer); WINPR_ASSERT(dwt_buffer);
@ -455,7 +455,7 @@ rfx_dwt_2d_encode_block_sse2(INT16* buffer, INT16* dwt, int subband_width)
rfx_dwt_2d_encode_block_horiz_sse2(h_src, lh, hh, subband_width); rfx_dwt_2d_encode_block_horiz_sse2(h_src, lh, hh, subband_width);
} }
static void rfx_dwt_2d_encode_sse2(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_buffer) static void rfx_dwt_2d_encode_sse2(INT16* buffer, INT16* dwt_buffer)
{ {
WINPR_ASSERT(buffer); WINPR_ASSERT(buffer);
WINPR_ASSERT(dwt_buffer); WINPR_ASSERT(dwt_buffer);

12
libfreerdp/codec/rfx_types.h
View File

@ -165,13 +165,11 @@ struct S_RFX_CONTEXT
struct S_RFX_MESSAGE currentMessage; struct S_RFX_MESSAGE currentMessage;
/* routines */ /* routines */
void (*quantization_decode)(INT16* WINPR_RESTRICT buffer, void (*quantization_decode)(INT16* buffer, const UINT32* WINPR_RESTRICT quantization_values);
const UINT32* WINPR_RESTRICT quantization_values); void (*quantization_encode)(INT16* buffer, const UINT32* WINPR_RESTRICT quantization_values);
void (*quantization_encode)(INT16* WINPR_RESTRICT buffer, void (*dwt_2d_decode)(INT16* buffer, INT16* dwt_buffer);
const UINT32* WINPR_RESTRICT quantization_values); void (*dwt_2d_extrapolate_decode)(INT16* src, INT16* temp);
void (*dwt_2d_decode)(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_buffer); void (*dwt_2d_encode)(INT16* buffer, INT16* dwt_buffer);
void (*dwt_2d_extrapolate_decode)(INT16* WINPR_RESTRICT src, INT16* WINPR_RESTRICT temp);
void (*dwt_2d_encode)(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RESTRICT dwt_buffer);
int (*rlgr_decode)(RLGR_MODE mode, const BYTE* WINPR_RESTRICT data, UINT32 data_size, int (*rlgr_decode)(RLGR_MODE mode, const BYTE* WINPR_RESTRICT data, UINT32 data_size,
INT16* WINPR_RESTRICT buffer, UINT32 buffer_size); INT16* WINPR_RESTRICT buffer, UINT32 buffer_size);
int (*rlgr_encode)(RLGR_MODE mode, const INT16* WINPR_RESTRICT data, UINT32 data_size, int (*rlgr_encode)(RLGR_MODE mode, const INT16* WINPR_RESTRICT data, UINT32 data_size,