2 ; jchuff-sse2.asm - Huffman entropy encoding (SSE2)
4 ; Copyright (C) 2009-2011, 2014-2017, D. R. Commander.
5 ; Copyright (C) 2015, Matthieu Darbois.
7 ; Based on the x86 SIMD extension for IJG JPEG library
8 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
11 ; This file should be assembled with NASM (Netwide Assembler),
12 ; can *not* be assembled with Microsoft's MASM or any compatible
13 ; assembler (including Borland's Turbo Assembler).
14 ; NASM is available from http://nasm.sourceforge.net/ or
15 ; http://sourceforge.net/project/showfiles.php?group_id=6208
17 ; This file contains an SSE2 implementation for Huffman coding of one block.
18 ; The following code is based directly on jchuff.c; see jchuff.c for more
23 %include "jsimdext.inc"
25 ; --------------------------------------------------------------------------
29 global EXTN(jconst_huff_encode_one_block)
31 EXTN(jconst_huff_encode_one_block):
33 %include "jpeg_nbits_table.inc"
37 ; --------------------------------------------------------------------------
41 ; These macros perform the same task as the emit_bits() function in the
42 ; original libjpeg code. In addition to reducing overhead by explicitly
43 ; inlining the code, additional performance is achieved by taking into
44 ; account the size of the bit buffer and waiting until it is almost full
45 ; before emptying it. This mostly benefits 64-bit platforms, since 6
46 ; bytes can be stored in a 64-bit bit buffer before it has to be emptied.
49 sub put_bits, 8 ; put_bits -= 8;
52 shr edx, cl ; c = (JOCTET)GETJOCTET(put_buffer >> put_bits);
53 mov byte [eax], dl ; *buffer++ = c;
55 cmp dl, 0xFF ; need to stuff a zero byte?
57 mov byte [eax], 0 ; *buffer++ = 0;
63 add put_bits, ecx ; put_bits += size;
64 shl put_buffer, cl ; put_buffer = (put_buffer << size);
69 cmp put_bits, 16 ; if (put_bits > 31) {
71 mov eax, POINTER [esp+buffer]
74 mov POINTER [esp+buffer], eax
83 %macro kloop_prepare 37 ;(ko, jno0, ..., jno31, xmm0, xmm1, xmm2, xmm3)
84 pxor xmm4, xmm4 ; __m128i neg = _mm_setzero_si128();
85 pxor xmm5, xmm5 ; __m128i neg = _mm_setzero_si128();
86 pxor xmm6, xmm6 ; __m128i neg = _mm_setzero_si128();
87 pxor xmm7, xmm7 ; __m128i neg = _mm_setzero_si128();
88 pinsrw %34, word [esi + %2 * SIZEOF_WORD], 0 ; xmm_shadow[0] = block[jno0];
89 pinsrw %35, word [esi + %10 * SIZEOF_WORD], 0 ; xmm_shadow[8] = block[jno8];
90 pinsrw %36, word [esi + %18 * SIZEOF_WORD], 0 ; xmm_shadow[16] = block[jno16];
91 pinsrw %37, word [esi + %26 * SIZEOF_WORD], 0 ; xmm_shadow[24] = block[jno24];
92 pinsrw %34, word [esi + %3 * SIZEOF_WORD], 1 ; xmm_shadow[1] = block[jno1];
93 pinsrw %35, word [esi + %11 * SIZEOF_WORD], 1 ; xmm_shadow[9] = block[jno9];
94 pinsrw %36, word [esi + %19 * SIZEOF_WORD], 1 ; xmm_shadow[17] = block[jno17];
95 pinsrw %37, word [esi + %27 * SIZEOF_WORD], 1 ; xmm_shadow[25] = block[jno25];
96 pinsrw %34, word [esi + %4 * SIZEOF_WORD], 2 ; xmm_shadow[2] = block[jno2];
97 pinsrw %35, word [esi + %12 * SIZEOF_WORD], 2 ; xmm_shadow[10] = block[jno10];
98 pinsrw %36, word [esi + %20 * SIZEOF_WORD], 2 ; xmm_shadow[18] = block[jno18];
99 pinsrw %37, word [esi + %28 * SIZEOF_WORD], 2 ; xmm_shadow[26] = block[jno26];
100 pinsrw %34, word [esi + %5 * SIZEOF_WORD], 3 ; xmm_shadow[3] = block[jno3];
101 pinsrw %35, word [esi + %13 * SIZEOF_WORD], 3 ; xmm_shadow[11] = block[jno11];
102 pinsrw %36, word [esi + %21 * SIZEOF_WORD], 3 ; xmm_shadow[19] = block[jno19];
103 pinsrw %37, word [esi + %29 * SIZEOF_WORD], 3 ; xmm_shadow[27] = block[jno27];
104 pinsrw %34, word [esi + %6 * SIZEOF_WORD], 4 ; xmm_shadow[4] = block[jno4];
105 pinsrw %35, word [esi + %14 * SIZEOF_WORD], 4 ; xmm_shadow[12] = block[jno12];
106 pinsrw %36, word [esi + %22 * SIZEOF_WORD], 4 ; xmm_shadow[20] = block[jno20];
107 pinsrw %37, word [esi + %30 * SIZEOF_WORD], 4 ; xmm_shadow[28] = block[jno28];
108 pinsrw %34, word [esi + %7 * SIZEOF_WORD], 5 ; xmm_shadow[5] = block[jno5];
109 pinsrw %35, word [esi + %15 * SIZEOF_WORD], 5 ; xmm_shadow[13] = block[jno13];
110 pinsrw %36, word [esi + %23 * SIZEOF_WORD], 5 ; xmm_shadow[21] = block[jno21];
111 pinsrw %37, word [esi + %31 * SIZEOF_WORD], 5 ; xmm_shadow[29] = block[jno29];
112 pinsrw %34, word [esi + %8 * SIZEOF_WORD], 6 ; xmm_shadow[6] = block[jno6];
113 pinsrw %35, word [esi + %16 * SIZEOF_WORD], 6 ; xmm_shadow[14] = block[jno14];
114 pinsrw %36, word [esi + %24 * SIZEOF_WORD], 6 ; xmm_shadow[22] = block[jno22];
115 pinsrw %37, word [esi + %32 * SIZEOF_WORD], 6 ; xmm_shadow[30] = block[jno30];
116 pinsrw %34, word [esi + %9 * SIZEOF_WORD], 7 ; xmm_shadow[7] = block[jno7];
117 pinsrw %35, word [esi + %17 * SIZEOF_WORD], 7 ; xmm_shadow[15] = block[jno15];
118 pinsrw %36, word [esi + %25 * SIZEOF_WORD], 7 ; xmm_shadow[23] = block[jno23];
120 pinsrw %37, word [esi + %33 * SIZEOF_WORD], 7 ; xmm_shadow[31] = block[jno31];
122 pinsrw %37, ecx, 7 ; xmm_shadow[31] = block[jno31];
124 pcmpgtw xmm4, %34 ; neg = _mm_cmpgt_epi16(neg, x1);
125 pcmpgtw xmm5, %35 ; neg = _mm_cmpgt_epi16(neg, x1);
126 pcmpgtw xmm6, %36 ; neg = _mm_cmpgt_epi16(neg, x1);
127 pcmpgtw xmm7, %37 ; neg = _mm_cmpgt_epi16(neg, x1);
128 paddw %34, xmm4 ; x1 = _mm_add_epi16(x1, neg);
129 paddw %35, xmm5 ; x1 = _mm_add_epi16(x1, neg);
130 paddw %36, xmm6 ; x1 = _mm_add_epi16(x1, neg);
131 paddw %37, xmm7 ; x1 = _mm_add_epi16(x1, neg);
132 pxor %34, xmm4 ; x1 = _mm_xor_si128(x1, neg);
133 pxor %35, xmm5 ; x1 = _mm_xor_si128(x1, neg);
134 pxor %36, xmm6 ; x1 = _mm_xor_si128(x1, neg);
135 pxor %37, xmm7 ; x1 = _mm_xor_si128(x1, neg);
136 pxor xmm4, %34 ; neg = _mm_xor_si128(neg, x1);
137 pxor xmm5, %35 ; neg = _mm_xor_si128(neg, x1);
138 pxor xmm6, %36 ; neg = _mm_xor_si128(neg, x1);
139 pxor xmm7, %37 ; neg = _mm_xor_si128(neg, x1);
140 movdqa XMMWORD [esp + t1 + %1 * SIZEOF_WORD], %34 ; _mm_storeu_si128((__m128i *)(t1 + ko), x1);
141 movdqa XMMWORD [esp + t1 + (%1 + 8) * SIZEOF_WORD], %35 ; _mm_storeu_si128((__m128i *)(t1 + ko + 8), x1);
142 movdqa XMMWORD [esp + t1 + (%1 + 16) * SIZEOF_WORD], %36 ; _mm_storeu_si128((__m128i *)(t1 + ko + 16), x1);
143 movdqa XMMWORD [esp + t1 + (%1 + 24) * SIZEOF_WORD], %37 ; _mm_storeu_si128((__m128i *)(t1 + ko + 24), x1);
144 movdqa XMMWORD [esp + t2 + %1 * SIZEOF_WORD], xmm4 ; _mm_storeu_si128((__m128i *)(t2 + ko), neg);
145 movdqa XMMWORD [esp + t2 + (%1 + 8) * SIZEOF_WORD], xmm5 ; _mm_storeu_si128((__m128i *)(t2 + ko + 8), neg);
146 movdqa XMMWORD [esp + t2 + (%1 + 16) * SIZEOF_WORD], xmm6 ; _mm_storeu_si128((__m128i *)(t2 + ko + 16), neg);
147 movdqa XMMWORD [esp + t2 + (%1 + 24) * SIZEOF_WORD], xmm7 ; _mm_storeu_si128((__m128i *)(t2 + ko + 24), neg);
151 ; Encode a single block's worth of coefficients.
154 ; jsimd_huff_encode_one_block_sse2 (working_state *state, JOCTET *buffer,
155 ; JCOEFPTR block, int last_dc_val,
156 ; c_derived_tbl *dctbl, c_derived_tbl *actbl)
159 ; eax + 8 = working_state *state
160 ; eax + 12 = JOCTET *buffer
161 ; eax + 16 = JCOEFPTR block
162 ; eax + 20 = int last_dc_val
163 ; eax + 24 = c_derived_tbl *dctbl
164 ; eax + 28 = c_derived_tbl *actbl
166 %define pad 6*SIZEOF_DWORD ; Align to 16 bytes
168 %define t2 t1+(DCTSIZE2*SIZEOF_WORD)
169 %define block t2+(DCTSIZE2*SIZEOF_WORD)
170 %define actbl block+SIZEOF_DWORD
171 %define buffer actbl+SIZEOF_DWORD
172 %define temp buffer+SIZEOF_DWORD
173 %define temp2 temp+SIZEOF_DWORD
174 %define temp3 temp2+SIZEOF_DWORD
175 %define temp4 temp3+SIZEOF_DWORD
176 %define temp5 temp4+SIZEOF_DWORD
177 %define gotptr temp5+SIZEOF_DWORD ; void *gotptr
178 %define put_buffer ebx
182 global EXTN(jsimd_huff_encode_one_block_sse2)
184 EXTN(jsimd_huff_encode_one_block_sse2):
186 mov eax,esp ; eax = original ebp
188 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
190 mov ebp,esp ; ebp = aligned ebp
191 sub esp, temp5+9*SIZEOF_DWORD-pad
194 ; push edx ; need not be preserved
199 mov esi, POINTER [eax+8] ; (working_state *state)
200 mov put_buffer, DWORD [esi+8] ; put_buffer = state->cur.put_buffer;
201 mov put_bits, DWORD [esi+12] ; put_bits = state->cur.put_bits;
202 push esi ; esi is now scratch
204 get_GOT edx ; get GOT address
205 movpic POINTER [esp+gotptr], edx ; save GOT address
207 mov ecx, POINTER [eax+28]
208 mov edx, POINTER [eax+16]
209 mov esi, POINTER [eax+12]
210 mov POINTER [esp+actbl], ecx
211 mov POINTER [esp+block], edx
212 mov POINTER [esp+buffer], esi
214 ; Encode the DC coefficient difference per section F.1.2.1
215 mov esi, POINTER [esp+block] ; block
216 movsx ecx, word [esi] ; temp = temp2 = block[0] - last_dc_val;
217 sub ecx, DWORD [eax+20]
220 ; This is a well-known technique for obtaining the absolute value
221 ; without a branch. It is derived from an assembly language technique
222 ; presented in "How to Optimize for the Pentium Processors",
223 ; Copyright (c) 1996, 1997 by Agner Fog.
225 sar edx, 31 ; temp3 = temp >> (CHAR_BIT * sizeof(int) - 1);
226 xor ecx, edx ; temp ^= temp3;
227 sub ecx, edx ; temp -= temp3;
229 ; For a negative input, want temp2 = bitwise complement of abs(input)
230 ; This code assumes we are on a two's complement machine
231 add esi, edx ; temp2 += temp3;
232 mov DWORD [esp+temp], esi ; backup temp2 in temp
234 ; Find the number of bits needed for the magnitude of the coefficient
235 movpic ebp, POINTER [esp+gotptr] ; load GOT address (ebp)
236 movzx edx, byte [GOTOFF(ebp, jpeg_nbits_table + ecx)] ; nbits = JPEG_NBITS(temp);
237 mov DWORD [esp+temp2], edx ; backup nbits in temp2
239 ; Emit the Huffman-coded symbol for the number of bits
240 mov ebp, POINTER [eax+24] ; After this point, arguments are not accessible anymore
241 mov eax, INT [ebp + edx * 4] ; code = dctbl->ehufco[nbits];
242 movzx ecx, byte [ebp + edx + 1024] ; size = dctbl->ehufsi[nbits];
243 EMIT_BITS eax ; EMIT_BITS(code, size)
245 mov ecx, DWORD [esp+temp2] ; restore nbits
247 ; Mask off any extra bits in code
251 and eax, DWORD [esp+temp] ; temp2 &= (((JLONG) 1)<<nbits) - 1;
253 ; Emit that number of bits of the value, if positive,
254 ; or the complement of its magnitude, if negative.
255 EMIT_BITS eax ; EMIT_BITS(temp2, nbits)
259 mov esi, POINTER [esp+block]
260 kloop_prepare 0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, \
261 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, \
262 27, 20, 13, 6, 7, 14, 21, 28, 35, \
263 xmm0, xmm1, xmm2, xmm3
264 kloop_prepare 32, 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, \
265 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, \
266 53, 60, 61, 54, 47, 55, 62, 63, 63, \
267 xmm0, xmm1, xmm2, xmm3
270 movdqa xmm0, XMMWORD [esp + t1 + 0 * SIZEOF_WORD] ; __m128i tmp0 = _mm_loadu_si128((__m128i *)(t1 + 0));
271 movdqa xmm1, XMMWORD [esp + t1 + 8 * SIZEOF_WORD] ; __m128i tmp1 = _mm_loadu_si128((__m128i *)(t1 + 8));
272 movdqa xmm2, XMMWORD [esp + t1 + 16 * SIZEOF_WORD] ; __m128i tmp2 = _mm_loadu_si128((__m128i *)(t1 + 16));
273 movdqa xmm3, XMMWORD [esp + t1 + 24 * SIZEOF_WORD] ; __m128i tmp3 = _mm_loadu_si128((__m128i *)(t1 + 24));
274 pcmpeqw xmm0, xmm7 ; tmp0 = _mm_cmpeq_epi16(tmp0, zero);
275 pcmpeqw xmm1, xmm7 ; tmp1 = _mm_cmpeq_epi16(tmp1, zero);
276 pcmpeqw xmm2, xmm7 ; tmp2 = _mm_cmpeq_epi16(tmp2, zero);
277 pcmpeqw xmm3, xmm7 ; tmp3 = _mm_cmpeq_epi16(tmp3, zero);
278 packsswb xmm0, xmm1 ; tmp0 = _mm_packs_epi16(tmp0, tmp1);
279 packsswb xmm2, xmm3 ; tmp2 = _mm_packs_epi16(tmp2, tmp3);
280 pmovmskb edx, xmm0 ; index = ((uint64_t)_mm_movemask_epi8(tmp0)) << 0;
281 pmovmskb ecx, xmm2 ; index = ((uint64_t)_mm_movemask_epi8(tmp2)) << 16;
284 not edx ; index = ~index;
287 mov ebp, POINTER [esp+actbl] ; ebp = actbl
290 bsf ecx, edx ; r = __builtin_ctzl(index);
292 lea esi, [esi+ecx*2] ; k += r;
293 shr edx, cl ; index >>= r;
294 mov DWORD [esp+temp3], edx
296 cmp ecx, 16 ; while (r > 15) {
298 sub ecx, 16 ; r -= 16;
299 mov DWORD [esp+temp], ecx
300 mov eax, INT [ebp + 240 * 4] ; code_0xf0 = actbl->ehufco[0xf0];
301 movzx ecx, byte [ebp + 1024 + 240] ; size_0xf0 = actbl->ehufsi[0xf0];
302 EMIT_BITS eax ; EMIT_BITS(code_0xf0, size_0xf0)
303 mov ecx, DWORD [esp+temp]
306 movsx eax, word [esi] ; temp = t1[k];
307 movpic edx, POINTER [esp+gotptr] ; load GOT address (edx)
308 movzx eax, byte [GOTOFF(edx, jpeg_nbits_table + eax)] ; nbits = JPEG_NBITS(temp);
309 mov DWORD [esp+temp2], eax
310 ; Emit Huffman symbol for run length / number of bits
311 shl ecx, 4 ; temp3 = (r << 4) + nbits;
313 mov eax, INT [ebp + ecx * 4] ; code = actbl->ehufco[temp3];
314 movzx ecx, byte [ebp + ecx + 1024] ; size = actbl->ehufsi[temp3];
317 movsx edx, word [esi+DCTSIZE2*2] ; temp2 = t2[k];
318 ; Mask off any extra bits in code
319 mov ecx, DWORD [esp+temp2]
323 and eax, edx ; temp2 &= (((JLONG) 1)<<nbits) - 1;
324 EMIT_BITS eax ; PUT_BITS(temp2, nbits)
325 mov edx, DWORD [esp+temp3]
327 shr edx, 1 ; index >>= 1;
331 movdqa xmm0, XMMWORD [esp + t1 + 32 * SIZEOF_WORD] ; __m128i tmp0 = _mm_loadu_si128((__m128i *)(t1 + 0));
332 movdqa xmm1, XMMWORD [esp + t1 + 40 * SIZEOF_WORD] ; __m128i tmp1 = _mm_loadu_si128((__m128i *)(t1 + 8));
333 movdqa xmm2, XMMWORD [esp + t1 + 48 * SIZEOF_WORD] ; __m128i tmp2 = _mm_loadu_si128((__m128i *)(t1 + 16));
334 movdqa xmm3, XMMWORD [esp + t1 + 56 * SIZEOF_WORD] ; __m128i tmp3 = _mm_loadu_si128((__m128i *)(t1 + 24));
335 pcmpeqw xmm0, xmm7 ; tmp0 = _mm_cmpeq_epi16(tmp0, zero);
336 pcmpeqw xmm1, xmm7 ; tmp1 = _mm_cmpeq_epi16(tmp1, zero);
337 pcmpeqw xmm2, xmm7 ; tmp2 = _mm_cmpeq_epi16(tmp2, zero);
338 pcmpeqw xmm3, xmm7 ; tmp3 = _mm_cmpeq_epi16(tmp3, zero);
339 packsswb xmm0, xmm1 ; tmp0 = _mm_packs_epi16(tmp0, tmp1);
340 packsswb xmm2, xmm3 ; tmp2 = _mm_packs_epi16(tmp2, tmp3);
341 pmovmskb edx, xmm0 ; index = ((uint64_t)_mm_movemask_epi8(tmp0)) << 0;
342 pmovmskb ecx, xmm2 ; index = ((uint64_t)_mm_movemask_epi8(tmp2)) << 16;
345 not edx ; index = ~index;
347 lea eax, [esp + t1 + (DCTSIZE2/2) * 2]
350 bsf ecx, edx ; r = __builtin_ctzl(index);
352 shr edx, cl ; index >>= r;
354 lea esi, [esi+ecx*2] ; k += r;
355 mov DWORD [esp+temp3], edx
358 bsf ecx, edx ; r = __builtin_ctzl(index);
360 lea esi, [esi+ecx*2] ; k += r;
361 shr edx, cl ; index >>= r;
362 mov DWORD [esp+temp3], edx
364 cmp ecx, 16 ; while (r > 15) {
366 sub ecx, 16 ; r -= 16;
367 mov DWORD [esp+temp], ecx
368 mov eax, INT [ebp + 240 * 4] ; code_0xf0 = actbl->ehufco[0xf0];
369 movzx ecx, byte [ebp + 1024 + 240] ; size_0xf0 = actbl->ehufsi[0xf0];
370 EMIT_BITS eax ; EMIT_BITS(code_0xf0, size_0xf0)
371 mov ecx, DWORD [esp+temp]
374 movsx eax, word [esi] ; temp = t1[k];
375 bsr eax, eax ; nbits = 32 - __builtin_clz(temp);
377 mov DWORD [esp+temp2], eax
378 ; Emit Huffman symbol for run length / number of bits
379 shl ecx, 4 ; temp3 = (r << 4) + nbits;
381 mov eax, INT [ebp + ecx * 4] ; code = actbl->ehufco[temp3];
382 movzx ecx, byte [ebp + ecx + 1024] ; size = actbl->ehufsi[temp3];
385 movsx edx, word [esi+DCTSIZE2*2] ; temp2 = t2[k];
386 ; Mask off any extra bits in code
387 mov ecx, DWORD [esp+temp2]
391 and eax, edx ; temp2 &= (((JLONG) 1)<<nbits) - 1;
392 EMIT_BITS eax ; PUT_BITS(temp2, nbits)
393 mov edx, DWORD [esp+temp3]
395 shr edx, 1 ; index >>= 1;
399 ; If the last coef(s) were zero, emit an end-of-block code
400 lea edx, [esp + t1 + (DCTSIZE2-1) * 2] ; r = DCTSIZE2-1-k;
401 cmp edx, esi ; if (r > 0) {
403 mov eax, INT [ebp] ; code = actbl->ehufco[0];
404 movzx ecx, byte [ebp + 1024] ; size = actbl->ehufsi[0];
407 mov eax, [esp+buffer]
409 ; Save put_buffer & put_bits
410 mov DWORD [esi+8], put_buffer ; state->cur.put_buffer = put_buffer;
411 mov DWORD [esi+12], put_bits ; state->cur.put_bits = put_bits;
416 ; pop edx ; need not be preserved
419 mov esp,ebp ; esp <- aligned ebp
420 pop esp ; esp <- original ebp
424 ; For some reason, the OS X linker does not honor the request to align the
425 ; segment unless we do this.