diff options
Diffstat (limited to 'simd')
-rw-r--r-- | simd/arm/arm/jsimd_neon.S | 607 | ||||
-rw-r--r-- | simd/arm/arm64/jsimd_neon.S | 642 | ||||
-rw-r--r-- | simd/arm/common/jidctint-neon.c | 747 |
3 files changed, 747 insertions, 1249 deletions
diff --git a/simd/arm/arm/jsimd_neon.S b/simd/arm/arm/jsimd_neon.S index 9d567233..d19e619e 100644 --- a/simd/arm/arm/jsimd_neon.S +++ b/simd/arm/arm/jsimd_neon.S @@ -65,613 +65,6 @@ _\fname: /*****************************************************************************/ /* - * Perform dequantization and inverse DCT on one block of coefficients. - * - * GLOBAL(void) - * jsimd_idct_islow_neon(void *dct_table, JCOEFPTR coef_block, - * JSAMPARRAY output_buf, JDIMENSION output_col) - */ - -#define FIX_0_298631336 (2446) -#define FIX_0_390180644 (3196) -#define FIX_0_541196100 (4433) -#define FIX_0_765366865 (6270) -#define FIX_0_899976223 (7373) -#define FIX_1_175875602 (9633) -#define FIX_1_501321110 (12299) -#define FIX_1_847759065 (15137) -#define FIX_1_961570560 (16069) -#define FIX_2_053119869 (16819) -#define FIX_2_562915447 (20995) -#define FIX_3_072711026 (25172) - -#define FIX_1_175875602_MINUS_1_961570560 (FIX_1_175875602 - FIX_1_961570560) -#define FIX_1_175875602_MINUS_0_390180644 (FIX_1_175875602 - FIX_0_390180644) -#define FIX_0_541196100_MINUS_1_847759065 (FIX_0_541196100 - FIX_1_847759065) -#define FIX_3_072711026_MINUS_2_562915447 (FIX_3_072711026 - FIX_2_562915447) -#define FIX_0_298631336_MINUS_0_899976223 (FIX_0_298631336 - FIX_0_899976223) -#define FIX_1_501321110_MINUS_0_899976223 (FIX_1_501321110 - FIX_0_899976223) -#define FIX_2_053119869_MINUS_2_562915447 (FIX_2_053119869 - FIX_2_562915447) -#define FIX_0_541196100_PLUS_0_765366865 (FIX_0_541196100 + FIX_0_765366865) - -/* - * Reference SIMD-friendly 1-D ISLOW iDCT C implementation. - * Uses some ideas from the comments in 'simd/jiss2int-64.asm' - */ -#define REF_1D_IDCT(xrow0, xrow1, xrow2, xrow3, xrow4, xrow5, xrow6, xrow7) { \ - DCTELEM row0, row1, row2, row3, row4, row5, row6, row7; \ - JLONG q1, q2, q3, q4, q5, q6, q7; \ - JLONG tmp11_plus_tmp2, tmp11_minus_tmp2; \ - \ - /* 1-D iDCT input data */ \ - row0 = xrow0; \ - row1 = xrow1; \ - row2 = xrow2; \ - row3 = xrow3; \ - row4 = xrow4; \ - row5 = xrow5; \ - row6 = xrow6; \ - row7 = xrow7; \ - \ - q5 = row7 + row3; \ - q4 = row5 + row1; \ - q6 = MULTIPLY(q5, FIX_1_175875602_MINUS_1_961570560) + \ - MULTIPLY(q4, FIX_1_175875602); \ - q7 = MULTIPLY(q5, FIX_1_175875602) + \ - MULTIPLY(q4, FIX_1_175875602_MINUS_0_390180644); \ - q2 = MULTIPLY(row2, FIX_0_541196100) + \ - MULTIPLY(row6, FIX_0_541196100_MINUS_1_847759065); \ - q4 = q6; \ - q3 = ((JLONG)row0 - (JLONG)row4) << 13; \ - q6 += MULTIPLY(row5, -FIX_2_562915447) + \ - MULTIPLY(row3, FIX_3_072711026_MINUS_2_562915447); \ - /* now we can use q1 (reloadable constants have been used up) */ \ - q1 = q3 + q2; \ - q4 += MULTIPLY(row7, FIX_0_298631336_MINUS_0_899976223) + \ - MULTIPLY(row1, -FIX_0_899976223); \ - q5 = q7; \ - q1 = q1 + q6; \ - q7 += MULTIPLY(row7, -FIX_0_899976223) + \ - MULTIPLY(row1, FIX_1_501321110_MINUS_0_899976223); \ - \ - /* (tmp11 + tmp2) has been calculated (out_row1 before descale) */ \ - tmp11_plus_tmp2 = q1; \ - row1 = 0; \ - \ - q1 = q1 - q6; \ - q5 += MULTIPLY(row5, FIX_2_053119869_MINUS_2_562915447) + \ - MULTIPLY(row3, -FIX_2_562915447); \ - q1 = q1 - q6; \ - q6 = MULTIPLY(row2, FIX_0_541196100_PLUS_0_765366865) + \ - MULTIPLY(row6, FIX_0_541196100); \ - q3 = q3 - q2; \ - \ - /* (tmp11 - tmp2) has been calculated (out_row6 before descale) */ \ - tmp11_minus_tmp2 = q1; \ - \ - q1 = ((JLONG)row0 + (JLONG)row4) << 13; \ - q2 = q1 + q6; \ - q1 = q1 - q6; \ - \ - /* pick up the results */ \ - tmp0 = q4; \ - tmp1 = q5; \ - tmp2 = (tmp11_plus_tmp2 - tmp11_minus_tmp2) / 2; \ - tmp3 = q7; \ - tmp10 = q2; \ - tmp11 = (tmp11_plus_tmp2 + tmp11_minus_tmp2) / 2; \ - tmp12 = q3; \ - tmp13 = q1; \ -} - -#define XFIX_0_899976223 d0[0] -#define XFIX_0_541196100 d0[1] -#define XFIX_2_562915447 d0[2] -#define XFIX_0_298631336_MINUS_0_899976223 d0[3] -#define XFIX_1_501321110_MINUS_0_899976223 d1[0] -#define XFIX_2_053119869_MINUS_2_562915447 d1[1] -#define XFIX_0_541196100_PLUS_0_765366865 d1[2] -#define XFIX_1_175875602 d1[3] -#define XFIX_1_175875602_MINUS_0_390180644 d2[0] -#define XFIX_0_541196100_MINUS_1_847759065 d2[1] -#define XFIX_3_072711026_MINUS_2_562915447 d2[2] -#define XFIX_1_175875602_MINUS_1_961570560 d2[3] - -.balign 16 -jsimd_idct_islow_neon_consts: - .short FIX_0_899976223 /* d0[0] */ - .short FIX_0_541196100 /* d0[1] */ - .short FIX_2_562915447 /* d0[2] */ - .short FIX_0_298631336_MINUS_0_899976223 /* d0[3] */ - .short FIX_1_501321110_MINUS_0_899976223 /* d1[0] */ - .short FIX_2_053119869_MINUS_2_562915447 /* d1[1] */ - .short FIX_0_541196100_PLUS_0_765366865 /* d1[2] */ - .short FIX_1_175875602 /* d1[3] */ - /* reloadable constants */ - .short FIX_1_175875602_MINUS_0_390180644 /* d2[0] */ - .short FIX_0_541196100_MINUS_1_847759065 /* d2[1] */ - .short FIX_3_072711026_MINUS_2_562915447 /* d2[2] */ - .short FIX_1_175875602_MINUS_1_961570560 /* d2[3] */ - -asm_function jsimd_idct_islow_neon - - DCT_TABLE .req r0 - COEF_BLOCK .req r1 - OUTPUT_BUF .req r2 - OUTPUT_COL .req r3 - TMP1 .req r0 - TMP2 .req r1 - TMP3 .req r2 - TMP4 .req ip - - ROW0L .req d16 - ROW0R .req d17 - ROW1L .req d18 - ROW1R .req d19 - ROW2L .req d20 - ROW2R .req d21 - ROW3L .req d22 - ROW3R .req d23 - ROW4L .req d24 - ROW4R .req d25 - ROW5L .req d26 - ROW5R .req d27 - ROW6L .req d28 - ROW6R .req d29 - ROW7L .req d30 - ROW7R .req d31 - - /* Load and dequantize coefficients into NEON registers - * with the following allocation: - * 0 1 2 3 | 4 5 6 7 - * ---------+-------- - * 0 | d16 | d17 ( q8 ) - * 1 | d18 | d19 ( q9 ) - * 2 | d20 | d21 ( q10 ) - * 3 | d22 | d23 ( q11 ) - * 4 | d24 | d25 ( q12 ) - * 5 | d26 | d27 ( q13 ) - * 6 | d28 | d29 ( q14 ) - * 7 | d30 | d31 ( q15 ) - */ - adr ip, jsimd_idct_islow_neon_consts - vld1.16 {d16, d17, d18, d19}, [COEF_BLOCK, :128]! - vld1.16 {d0, d1, d2, d3}, [DCT_TABLE, :128]! - vld1.16 {d20, d21, d22, d23}, [COEF_BLOCK, :128]! - vmul.s16 q8, q8, q0 - vld1.16 {d4, d5, d6, d7}, [DCT_TABLE, :128]! - vmul.s16 q9, q9, q1 - vld1.16 {d24, d25, d26, d27}, [COEF_BLOCK, :128]! - vmul.s16 q10, q10, q2 - vld1.16 {d0, d1, d2, d3}, [DCT_TABLE, :128]! - vmul.s16 q11, q11, q3 - vld1.16 {d28, d29, d30, d31}, [COEF_BLOCK, :128] - vmul.s16 q12, q12, q0 - vld1.16 {d4, d5, d6, d7}, [DCT_TABLE, :128]! - vmul.s16 q14, q14, q2 - vmul.s16 q13, q13, q1 - vld1.16 {d0, d1, d2, d3}, [ip, :128] /* load constants */ - add ip, ip, #16 - vmul.s16 q15, q15, q3 - vpush {d8-d15} /* save NEON registers */ - /* 1-D IDCT, pass 1, left 4x8 half */ - vadd.s16 d4, ROW7L, ROW3L - vadd.s16 d5, ROW5L, ROW1L - vmull.s16 q6, d4, XFIX_1_175875602_MINUS_1_961570560 - vmlal.s16 q6, d5, XFIX_1_175875602 - vmull.s16 q7, d4, XFIX_1_175875602 - /* Check for the zero coefficients in the right 4x8 half */ - push {r4, r5} - vmlal.s16 q7, d5, XFIX_1_175875602_MINUS_0_390180644 - vsubl.s16 q3, ROW0L, ROW4L - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 1 * 8))] - vmull.s16 q2, ROW2L, XFIX_0_541196100 - vmlal.s16 q2, ROW6L, XFIX_0_541196100_MINUS_1_847759065 - orr r0, r4, r5 - vmov q4, q6 - vmlsl.s16 q6, ROW5L, XFIX_2_562915447 - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 2 * 8))] - vmlal.s16 q6, ROW3L, XFIX_3_072711026_MINUS_2_562915447 - vshl.s32 q3, q3, #13 - orr r0, r0, r4 - vmlsl.s16 q4, ROW1L, XFIX_0_899976223 - orr r0, r0, r5 - vadd.s32 q1, q3, q2 - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 3 * 8))] - vmov q5, q7 - vadd.s32 q1, q1, q6 - orr r0, r0, r4 - vmlsl.s16 q7, ROW7L, XFIX_0_899976223 - orr r0, r0, r5 - vmlal.s16 q7, ROW1L, XFIX_1_501321110_MINUS_0_899976223 - vrshrn.s32 ROW1L, q1, #11 - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 4 * 8))] - vsub.s32 q1, q1, q6 - vmlal.s16 q5, ROW5L, XFIX_2_053119869_MINUS_2_562915447 - orr r0, r0, r4 - vmlsl.s16 q5, ROW3L, XFIX_2_562915447 - orr r0, r0, r5 - vsub.s32 q1, q1, q6 - vmull.s16 q6, ROW2L, XFIX_0_541196100_PLUS_0_765366865 - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 5 * 8))] - vmlal.s16 q6, ROW6L, XFIX_0_541196100 - vsub.s32 q3, q3, q2 - orr r0, r0, r4 - vrshrn.s32 ROW6L, q1, #11 - orr r0, r0, r5 - vadd.s32 q1, q3, q5 - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 6 * 8))] - vsub.s32 q3, q3, q5 - vaddl.s16 q5, ROW0L, ROW4L - orr r0, r0, r4 - vrshrn.s32 ROW2L, q1, #11 - orr r0, r0, r5 - vrshrn.s32 ROW5L, q3, #11 - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 7 * 8))] - vshl.s32 q5, q5, #13 - vmlal.s16 q4, ROW7L, XFIX_0_298631336_MINUS_0_899976223 - orr r0, r0, r4 - vadd.s32 q2, q5, q6 - orrs r0, r0, r5 - vsub.s32 q1, q5, q6 - vadd.s32 q6, q2, q7 - ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 0 * 8))] - vsub.s32 q2, q2, q7 - vadd.s32 q5, q1, q4 - orr r0, r4, r5 - vsub.s32 q3, q1, q4 - pop {r4, r5} - vrshrn.s32 ROW7L, q2, #11 - vrshrn.s32 ROW3L, q5, #11 - vrshrn.s32 ROW0L, q6, #11 - vrshrn.s32 ROW4L, q3, #11 - - beq 3f /* Go to do some special handling for the sparse - right 4x8 half */ - - /* 1-D IDCT, pass 1, right 4x8 half */ - vld1.s16 {d2}, [ip, :64] /* reload constants */ - vadd.s16 d10, ROW7R, ROW3R - vadd.s16 d8, ROW5R, ROW1R - /* Transpose left 4x8 half */ - vtrn.16 ROW6L, ROW7L - vmull.s16 q6, d10, XFIX_1_175875602_MINUS_1_961570560 - vmlal.s16 q6, d8, XFIX_1_175875602 - vtrn.16 ROW2L, ROW3L - vmull.s16 q7, d10, XFIX_1_175875602 - vmlal.s16 q7, d8, XFIX_1_175875602_MINUS_0_390180644 - vtrn.16 ROW0L, ROW1L - vsubl.s16 q3, ROW0R, ROW4R - vmull.s16 q2, ROW2R, XFIX_0_541196100 - vmlal.s16 q2, ROW6R, XFIX_0_541196100_MINUS_1_847759065 - vtrn.16 ROW4L, ROW5L - vmov q4, q6 - vmlsl.s16 q6, ROW5R, XFIX_2_562915447 - vmlal.s16 q6, ROW3R, XFIX_3_072711026_MINUS_2_562915447 - vtrn.32 ROW1L, ROW3L - vshl.s32 q3, q3, #13 - vmlsl.s16 q4, ROW1R, XFIX_0_899976223 - vtrn.32 ROW4L, ROW6L - vadd.s32 q1, q3, q2 - vmov q5, q7 - vadd.s32 q1, q1, q6 - vtrn.32 ROW0L, ROW2L - vmlsl.s16 q7, ROW7R, XFIX_0_899976223 - vmlal.s16 q7, ROW1R, XFIX_1_501321110_MINUS_0_899976223 - vrshrn.s32 ROW1R, q1, #11 - vtrn.32 ROW5L, ROW7L - vsub.s32 q1, q1, q6 - vmlal.s16 q5, ROW5R, XFIX_2_053119869_MINUS_2_562915447 - vmlsl.s16 q5, ROW3R, XFIX_2_562915447 - vsub.s32 q1, q1, q6 - vmull.s16 q6, ROW2R, XFIX_0_541196100_PLUS_0_765366865 - vmlal.s16 q6, ROW6R, XFIX_0_541196100 - vsub.s32 q3, q3, q2 - vrshrn.s32 ROW6R, q1, #11 - vadd.s32 q1, q3, q5 - vsub.s32 q3, q3, q5 - vaddl.s16 q5, ROW0R, ROW4R - vrshrn.s32 ROW2R, q1, #11 - vrshrn.s32 ROW5R, q3, #11 - vshl.s32 q5, q5, #13 - vmlal.s16 q4, ROW7R, XFIX_0_298631336_MINUS_0_899976223 - vadd.s32 q2, q5, q6 - vsub.s32 q1, q5, q6 - vadd.s32 q6, q2, q7 - vsub.s32 q2, q2, q7 - vadd.s32 q5, q1, q4 - vsub.s32 q3, q1, q4 - vrshrn.s32 ROW7R, q2, #11 - vrshrn.s32 ROW3R, q5, #11 - vrshrn.s32 ROW0R, q6, #11 - vrshrn.s32 ROW4R, q3, #11 - /* Transpose right 4x8 half */ - vtrn.16 ROW6R, ROW7R - vtrn.16 ROW2R, ROW3R - vtrn.16 ROW0R, ROW1R - vtrn.16 ROW4R, ROW5R - vtrn.32 ROW1R, ROW3R - vtrn.32 ROW4R, ROW6R - vtrn.32 ROW0R, ROW2R - vtrn.32 ROW5R, ROW7R - -1: /* 1-D IDCT, pass 2 (normal variant), left 4x8 half */ - vld1.s16 {d2}, [ip, :64] /* reload constants */ - vmull.s16 q6, ROW1R, XFIX_1_175875602 /* ROW5L <-> ROW1R */ - vmlal.s16 q6, ROW1L, XFIX_1_175875602 - vmlal.s16 q6, ROW3R, XFIX_1_175875602_MINUS_1_961570560 /* ROW7L <-> ROW3R */ - vmlal.s16 q6, ROW3L, XFIX_1_175875602_MINUS_1_961570560 - vmull.s16 q7, ROW3R, XFIX_1_175875602 /* ROW7L <-> ROW3R */ - vmlal.s16 q7, ROW3L, XFIX_1_175875602 - vmlal.s16 q7, ROW1R, XFIX_1_175875602_MINUS_0_390180644 /* ROW5L <-> ROW1R */ - vmlal.s16 q7, ROW1L, XFIX_1_175875602_MINUS_0_390180644 - vsubl.s16 q3, ROW0L, ROW0R /* ROW4L <-> ROW0R */ - vmull.s16 q2, ROW2L, XFIX_0_541196100 - vmlal.s16 q2, ROW2R, XFIX_0_541196100_MINUS_1_847759065 /* ROW6L <-> ROW2R */ - vmov q4, q6 - vmlsl.s16 q6, ROW1R, XFIX_2_562915447 /* ROW5L <-> ROW1R */ - vmlal.s16 q6, ROW3L, XFIX_3_072711026_MINUS_2_562915447 - vshl.s32 q3, q3, #13 - vmlsl.s16 q4, ROW1L, XFIX_0_899976223 - vadd.s32 q1, q3, q2 - vmov q5, q7 - vadd.s32 q1, q1, q6 - vmlsl.s16 q7, ROW3R, XFIX_0_899976223 /* ROW7L <-> ROW3R */ - vmlal.s16 q7, ROW1L, XFIX_1_501321110_MINUS_0_899976223 - vshrn.s32 ROW1L, q1, #16 - vsub.s32 q1, q1, q6 - vmlal.s16 q5, ROW1R, XFIX_2_053119869_MINUS_2_562915447 /* ROW5L <-> ROW1R */ - vmlsl.s16 q5, ROW3L, XFIX_2_562915447 - vsub.s32 q1, q1, q6 - vmull.s16 q6, ROW2L, XFIX_0_541196100_PLUS_0_765366865 - vmlal.s16 q6, ROW2R, XFIX_0_541196100 /* ROW6L <-> ROW2R */ - vsub.s32 q3, q3, q2 - vshrn.s32 ROW2R, q1, #16 /* ROW6L <-> ROW2R */ - vadd.s32 q1, q3, q5 - vsub.s32 q3, q3, q5 - vaddl.s16 q5, ROW0L, ROW0R /* ROW4L <-> ROW0R */ - vshrn.s32 ROW2L, q1, #16 - vshrn.s32 ROW1R, q3, #16 /* ROW5L <-> ROW1R */ - vshl.s32 q5, q5, #13 - vmlal.s16 q4, ROW3R, XFIX_0_298631336_MINUS_0_899976223 /* ROW7L <-> ROW3R */ - vadd.s32 q2, q5, q6 - vsub.s32 q1, q5, q6 - vadd.s32 q6, q2, q7 - vsub.s32 q2, q2, q7 - vadd.s32 q5, q1, q4 - vsub.s32 q3, q1, q4 - vshrn.s32 ROW3R, q2, #16 /* ROW7L <-> ROW3R */ - vshrn.s32 ROW3L, q5, #16 - vshrn.s32 ROW0L, q6, #16 - vshrn.s32 ROW0R, q3, #16 /* ROW4L <-> ROW0R */ - /* 1-D IDCT, pass 2, right 4x8 half */ - vld1.s16 {d2}, [ip, :64] /* reload constants */ - vmull.s16 q6, ROW5R, XFIX_1_175875602 - vmlal.s16 q6, ROW5L, XFIX_1_175875602 /* ROW5L <-> ROW1R */ - vmlal.s16 q6, ROW7R, XFIX_1_175875602_MINUS_1_961570560 - vmlal.s16 q6, ROW7L, XFIX_1_175875602_MINUS_1_961570560 /* ROW7L <-> ROW3R */ - vmull.s16 q7, ROW7R, XFIX_1_175875602 - vmlal.s16 q7, ROW7L, XFIX_1_175875602 /* ROW7L <-> ROW3R */ - vmlal.s16 q7, ROW5R, XFIX_1_175875602_MINUS_0_390180644 - vmlal.s16 q7, ROW5L, XFIX_1_175875602_MINUS_0_390180644 /* ROW5L <-> ROW1R */ - vsubl.s16 q3, ROW4L, ROW4R /* ROW4L <-> ROW0R */ - vmull.s16 q2, ROW6L, XFIX_0_541196100 /* ROW6L <-> ROW2R */ - vmlal.s16 q2, ROW6R, XFIX_0_541196100_MINUS_1_847759065 - vmov q4, q6 - vmlsl.s16 q6, ROW5R, XFIX_2_562915447 - vmlal.s16 q6, ROW7L, XFIX_3_072711026_MINUS_2_562915447 /* ROW7L <-> ROW3R */ - vshl.s32 q3, q3, #13 - vmlsl.s16 q4, ROW5L, XFIX_0_899976223 /* ROW5L <-> ROW1R */ - vadd.s32 q1, q3, q2 - vmov q5, q7 - vadd.s32 q1, q1, q6 - vmlsl.s16 q7, ROW7R, XFIX_0_899976223 - vmlal.s16 q7, ROW5L, XFIX_1_501321110_MINUS_0_899976223 /* ROW5L <-> ROW1R */ - vshrn.s32 ROW5L, q1, #16 /* ROW5L <-> ROW1R */ - vsub.s32 q1, q1, q6 - vmlal.s16 q5, ROW5R, XFIX_2_053119869_MINUS_2_562915447 - vmlsl.s16 q5, ROW7L, XFIX_2_562915447 /* ROW7L <-> ROW3R */ - vsub.s32 q1, q1, q6 - vmull.s16 q6, ROW6L, XFIX_0_541196100_PLUS_0_765366865 /* ROW6L <-> ROW2R */ - vmlal.s16 q6, ROW6R, XFIX_0_541196100 - vsub.s32 q3, q3, q2 - vshrn.s32 ROW6R, q1, #16 - vadd.s32 q1, q3, q5 - vsub.s32 q3, q3, q5 - vaddl.s16 q5, ROW4L, ROW4R /* ROW4L <-> ROW0R */ - vshrn.s32 ROW6L, q1, #16 /* ROW6L <-> ROW2R */ - vshrn.s32 ROW5R, q3, #16 - vshl.s32 q5, q5, #13 - vmlal.s16 q4, ROW7R, XFIX_0_298631336_MINUS_0_899976223 - vadd.s32 q2, q5, q6 - vsub.s32 q1, q5, q6 - vadd.s32 q6, q2, q7 - vsub.s32 q2, q2, q7 - vadd.s32 q5, q1, q4 - vsub.s32 q3, q1, q4 - vshrn.s32 ROW7R, q2, #16 - vshrn.s32 ROW7L, q5, #16 /* ROW7L <-> ROW3R */ - vshrn.s32 ROW4L, q6, #16 /* ROW4L <-> ROW0R */ - vshrn.s32 ROW4R, q3, #16 - -2: /* Descale to 8-bit and range limit */ - vqrshrn.s16 d16, q8, #2 - vqrshrn.s16 d17, q9, #2 - vqrshrn.s16 d18, q10, #2 - vqrshrn.s16 d19, q11, #2 - vpop {d8-d15} /* restore NEON registers */ - vqrshrn.s16 d20, q12, #2 - /* Transpose the final 8-bit samples and do signed->unsigned conversion */ - vtrn.16 q8, q9 - vqrshrn.s16 d21, q13, #2 - vqrshrn.s16 d22, q14, #2 - vmov.u8 q0, #(CENTERJSAMPLE) - vqrshrn.s16 d23, q15, #2 - vtrn.8 d16, d17 - vtrn.8 d18, d19 - vadd.u8 q8, q8, q0 - vadd.u8 q9, q9, q0 - vtrn.16 q10, q11 - /* Store results to the output buffer */ - ldmia OUTPUT_BUF!, {TMP1, TMP2} - add TMP1, TMP1, OUTPUT_COL - add TMP2, TMP2, OUTPUT_COL - vst1.8 {d16}, [TMP1] - vtrn.8 d20, d21 - vst1.8 {d17}, [TMP2] - ldmia OUTPUT_BUF!, {TMP1, TMP2} - add TMP1, TMP1, OUTPUT_COL - add TMP2, TMP2, OUTPUT_COL - vst1.8 {d18}, [TMP1] - vadd.u8 q10, q10, q0 - vst1.8 {d19}, [TMP2] - ldmia OUTPUT_BUF, {TMP1, TMP2, TMP3, TMP4} - add TMP1, TMP1, OUTPUT_COL - add TMP2, TMP2, OUTPUT_COL - add TMP3, TMP3, OUTPUT_COL - add TMP4, TMP4, OUTPUT_COL - vtrn.8 d22, d23 - vst1.8 {d20}, [TMP1] - vadd.u8 q11, q11, q0 - vst1.8 {d21}, [TMP2] - vst1.8 {d22}, [TMP3] - vst1.8 {d23}, [TMP4] - bx lr - -3: /* Left 4x8 half is done, right 4x8 half contains mostly zeros */ - - /* Transpose left 4x8 half */ - vtrn.16 ROW6L, ROW7L - vtrn.16 ROW2L, ROW3L - vtrn.16 ROW0L, ROW1L - vtrn.16 ROW4L, ROW5L - vshl.s16 ROW0R, ROW0R, #2 /* PASS1_BITS */ - vtrn.32 ROW1L, ROW3L - vtrn.32 ROW4L, ROW6L - vtrn.32 ROW0L, ROW2L - vtrn.32 ROW5L, ROW7L - - cmp r0, #0 - beq 4f /* Right 4x8 half has all zeros, go to 'sparse' second - pass */ - - /* Only row 0 is non-zero for the right 4x8 half */ - vdup.s16 ROW1R, ROW0R[1] - vdup.s16 ROW2R, ROW0R[2] - vdup.s16 ROW3R, ROW0R[3] - vdup.s16 ROW4R, ROW0R[0] - vdup.s16 ROW5R, ROW0R[1] - vdup.s16 ROW6R, ROW0R[2] - vdup.s16 ROW7R, ROW0R[3] - vdup.s16 ROW0R, ROW0R[0] - b 1b /* Go to 'normal' second pass */ - -4: /* 1-D IDCT, pass 2 (sparse variant with zero rows 4-7), left 4x8 half */ - vld1.s16 {d2}, [ip, :64] /* reload constants */ - vmull.s16 q6, ROW1L, XFIX_1_175875602 - vmlal.s16 q6, ROW3L, XFIX_1_175875602_MINUS_1_961570560 - vmull.s16 q7, ROW3L, XFIX_1_175875602 - vmlal.s16 q7, ROW1L, XFIX_1_175875602_MINUS_0_390180644 - vmull.s16 q2, ROW2L, XFIX_0_541196100 - vshll.s16 q3, ROW0L, #13 - vmov q4, q6 - vmlal.s16 q6, ROW3L, XFIX_3_072711026_MINUS_2_562915447 - vmlsl.s16 q4, ROW1L, XFIX_0_899976223 - vadd.s32 q1, q3, q2 - vmov q5, q7 - vmlal.s16 q7, ROW1L, XFIX_1_501321110_MINUS_0_899976223 - vadd.s32 q1, q1, q6 - vadd.s32 q6, q6, q6 - vmlsl.s16 q5, ROW3L, XFIX_2_562915447 - vshrn.s32 ROW1L, q1, #16 - vsub.s32 q1, q1, q6 - vmull.s16 q6, ROW2L, XFIX_0_541196100_PLUS_0_765366865 - vsub.s32 q3, q3, q2 - vshrn.s32 ROW2R, q1, #16 /* ROW6L <-> ROW2R */ - vadd.s32 q1, q3, q5 - vsub.s32 q3, q3, q5 - vshll.s16 q5, ROW0L, #13 - vshrn.s32 ROW2L, q1, #16 - vshrn.s32 ROW1R, q3, #16 /* ROW5L <-> ROW1R */ - vadd.s32 q2, q5, q6 - vsub.s32 q1, q5, q6 - vadd.s32 q6, q2, q7 - vsub.s32 q2, q2, q7 - vadd.s32 q5, q1, q4 - vsub.s32 q3, q1, q4 - vshrn.s32 ROW3R, q2, #16 /* ROW7L <-> ROW3R */ - vshrn.s32 ROW3L, q5, #16 - vshrn.s32 ROW0L, q6, #16 - vshrn.s32 ROW0R, q3, #16 /* ROW4L <-> ROW0R */ - /* 1-D IDCT, pass 2 (sparse variant with zero rows 4-7), right 4x8 half */ - vld1.s16 {d2}, [ip, :64] /* reload constants */ - vmull.s16 q6, ROW5L, XFIX_1_175875602 - vmlal.s16 q6, ROW7L, XFIX_1_175875602_MINUS_1_961570560 - vmull.s16 q7, ROW7L, XFIX_1_175875602 - vmlal.s16 q7, ROW5L, XFIX_1_175875602_MINUS_0_390180644 - vmull.s16 q2, ROW6L, XFIX_0_541196100 - vshll.s16 q3, ROW4L, #13 - vmov q4, q6 - vmlal.s16 q6, ROW7L, XFIX_3_072711026_MINUS_2_562915447 - vmlsl.s16 q4, ROW5L, XFIX_0_899976223 - vadd.s32 q1, q3, q2 - vmov q5, q7 - vmlal.s16 q7, ROW5L, XFIX_1_501321110_MINUS_0_899976223 - vadd.s32 q1, q1, q6 - vadd.s32 q6, q6, q6 - vmlsl.s16 q5, ROW7L, XFIX_2_562915447 - vshrn.s32 ROW5L, q1, #16 /* ROW5L <-> ROW1R */ - vsub.s32 q1, q1, q6 - vmull.s16 q6, ROW6L, XFIX_0_541196100_PLUS_0_765366865 - vsub.s32 q3, q3, q2 - vshrn.s32 ROW6R, q1, #16 - vadd.s32 q1, q3, q5 - vsub.s32 q3, q3, q5 - vshll.s16 q5, ROW4L, #13 - vshrn.s32 ROW6L, q1, #16 /* ROW6L <-> ROW2R */ - vshrn.s32 ROW5R, q3, #16 - vadd.s32 q2, q5, q6 - vsub.s32 q1, q5, q6 - vadd.s32 q6, q2, q7 - vsub.s32 q2, q2, q7 - vadd.s32 q5, q1, q4 - vsub.s32 q3, q1, q4 - vshrn.s32 ROW7R, q2, #16 - vshrn.s32 ROW7L, q5, #16 /* ROW7L <-> ROW3R */ - vshrn.s32 ROW4L, q6, #16 /* ROW4L <-> ROW0R */ - vshrn.s32 ROW4R, q3, #16 - b 2b /* Go to epilogue */ - - .unreq DCT_TABLE - .unreq COEF_BLOCK - .unreq OUTPUT_BUF - .unreq OUTPUT_COL - .unreq TMP1 - .unreq TMP2 - .unreq TMP3 - .unreq TMP4 - - .unreq ROW0L - .unreq ROW0R - .unreq ROW1L - .unreq ROW1R - .unreq ROW2L - .unreq ROW2R - .unreq ROW3L - .unreq ROW3R - .unreq ROW4L - .unreq ROW4R - .unreq ROW5L - .unreq ROW5R - .unreq ROW6L - .unreq ROW6R - .unreq ROW7L - .unreq ROW7R - - -/*****************************************************************************/ - -/* * jsimd_idct_ifast_neon * * This function contains a fast, not so accurate integer implementation of diff --git a/simd/arm/arm64/jsimd_neon.S b/simd/arm/arm64/jsimd_neon.S index 5077a168..4d387416 100644 --- a/simd/arm/arm64/jsimd_neon.S +++ b/simd/arm/arm64/jsimd_neon.S @@ -90,648 +90,6 @@ _\fname: /*****************************************************************************/ /* - * Perform dequantization and inverse DCT on one block of coefficients. - * - * GLOBAL(void) - * jsimd_idct_islow_neon(void *dct_table, JCOEFPTR coef_block, - * JSAMPARRAY output_buf, JDIMENSION output_col) - */ - -#define CONST_BITS 13 -#define PASS1_BITS 2 - -#define F_0_298 2446 /* FIX(0.298631336) */ -#define F_0_390 3196 /* FIX(0.390180644) */ -#define F_0_541 4433 /* FIX(0.541196100) */ -#define F_0_765 6270 /* FIX(0.765366865) */ -#define F_0_899 7373 /* FIX(0.899976223) */ -#define F_1_175 9633 /* FIX(1.175875602) */ -#define F_1_501 12299 /* FIX(1.501321110) */ -#define F_1_847 15137 /* FIX(1.847759065) */ -#define F_1_961 16069 /* FIX(1.961570560) */ -#define F_2_053 16819 /* FIX(2.053119869) */ -#define F_2_562 20995 /* FIX(2.562915447) */ -#define F_3_072 25172 /* FIX(3.072711026) */ - -.balign 16 -Ljsimd_idct_islow_neon_consts: - .short F_0_298 - .short -F_0_390 - .short F_0_541 - .short F_0_765 - .short - F_0_899 - .short F_1_175 - .short F_1_501 - .short - F_1_847 - .short - F_1_961 - .short F_2_053 - .short - F_2_562 - .short F_3_072 - .short 0 /* padding */ - .short 0 - .short 0 - .short 0 - -#undef F_0_298 -#undef F_0_390 -#undef F_0_541 -#undef F_0_765 -#undef F_0_899 -#undef F_1_175 -#undef F_1_501 -#undef F_1_847 -#undef F_1_961 -#undef F_2_053 -#undef F_2_562 -#undef F_3_072 - -#define XFIX_P_0_298 v0.h[0] -#define XFIX_N_0_390 v0.h[1] -#define XFIX_P_0_541 v0.h[2] -#define XFIX_P_0_765 v0.h[3] -#define XFIX_N_0_899 v0.h[4] -#define XFIX_P_1_175 v0.h[5] -#define XFIX_P_1_501 v0.h[6] -#define XFIX_N_1_847 v0.h[7] -#define XFIX_N_1_961 v1.h[0] -#define XFIX_P_2_053 v1.h[1] -#define XFIX_N_2_562 v1.h[2] -#define XFIX_P_3_072 v1.h[3] - -asm_function jsimd_idct_islow_neon - DCT_TABLE .req x0 - COEF_BLOCK .req x1 - OUTPUT_BUF .req x2 - OUTPUT_COL .req x3 - TMP1 .req x0 - TMP2 .req x1 - TMP3 .req x9 - TMP4 .req x10 - TMP5 .req x11 - TMP6 .req x12 - TMP7 .req x13 - TMP8 .req x14 - - /* OUTPUT_COL is a JDIMENSION (unsigned int) argument, so the ABI doesn't - guarantee that the upper (unused) 32 bits of x3 are valid. This - instruction ensures that those bits are set to zero. */ - uxtw x3, w3 - - sub sp, sp, #64 - adr x15, Ljsimd_idct_islow_neon_consts - mov x10, sp - st1 {v8.8b, v9.8b, v10.8b, v11.8b}, [x10], #32 - st1 {v12.8b, v13.8b, v14.8b, v15.8b}, [x10], #32 - ld1 {v0.8h, v1.8h}, [x15] - ld1 {v2.8h, v3.8h, v4.8h, v5.8h}, [COEF_BLOCK], #64 - ld1 {v18.8h, v19.8h, v20.8h, v21.8h}, [DCT_TABLE], #64 - ld1 {v6.8h, v7.8h, v8.8h, v9.8h}, [COEF_BLOCK], #64 - ld1 {v22.8h, v23.8h, v24.8h, v25.8h}, [DCT_TABLE], #64 - - cmeq v16.8h, v3.8h, #0 - cmeq v26.8h, v4.8h, #0 - cmeq v27.8h, v5.8h, #0 - cmeq v28.8h, v6.8h, #0 - cmeq v29.8h, v7.8h, #0 - cmeq v30.8h, v8.8h, #0 - cmeq v31.8h, v9.8h, #0 - - and v10.16b, v16.16b, v26.16b - and v11.16b, v27.16b, v28.16b - and v12.16b, v29.16b, v30.16b - and v13.16b, v31.16b, v10.16b - and v14.16b, v11.16b, v12.16b - mul v2.8h, v2.8h, v18.8h - and v15.16b, v13.16b, v14.16b - shl v10.8h, v2.8h, #(PASS1_BITS) - sqxtn v16.8b, v15.8h - mov TMP1, v16.d[0] - mvn TMP2, TMP1 - - cbnz TMP2, 2f - /* case all AC coeffs are zeros */ - dup v2.2d, v10.d[0] - dup v6.2d, v10.d[1] - mov v3.16b, v2.16b - mov v7.16b, v6.16b - mov v4.16b, v2.16b - mov v8.16b, v6.16b - mov v5.16b, v2.16b - mov v9.16b, v6.16b -1: - /* for this transpose, we should organise data like this: - * 00, 01, 02, 03, 40, 41, 42, 43 - * 10, 11, 12, 13, 50, 51, 52, 53 - * 20, 21, 22, 23, 60, 61, 62, 63 - * 30, 31, 32, 33, 70, 71, 72, 73 - * 04, 05, 06, 07, 44, 45, 46, 47 - * 14, 15, 16, 17, 54, 55, 56, 57 - * 24, 25, 26, 27, 64, 65, 66, 67 - * 34, 35, 36, 37, 74, 75, 76, 77 - */ - trn1 v28.8h, v2.8h, v3.8h - trn1 v29.8h, v4.8h, v5.8h - trn1 v30.8h, v6.8h, v7.8h - trn1 v31.8h, v8.8h, v9.8h - trn2 v16.8h, v2.8h, v3.8h - trn2 v17.8h, v4.8h, v5.8h - trn2 v18.8h, v6.8h, v7.8h - trn2 v19.8h, v8.8h, v9.8h - trn1 v2.4s, v28.4s, v29.4s - trn1 v6.4s, v30.4s, v31.4s - trn1 v3.4s, v16.4s, v17.4s - trn1 v7.4s, v18.4s, v19.4s - trn2 v4.4s, v28.4s, v29.4s - trn2 v8.4s, v30.4s, v31.4s - trn2 v5.4s, v16.4s, v17.4s - trn2 v9.4s, v18.4s, v19.4s - /* Even part: reverse the even part of the forward DCT. */ - add v18.8h, v4.8h, v8.8h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */ - add v22.8h, v2.8h, v6.8h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - smull2 v19.4s, v18.8h, XFIX_P_0_541 /* z1h z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */ - sub v26.8h, v2.8h, v6.8h /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - smull v18.4s, v18.4h, XFIX_P_0_541 /* z1l z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */ - sshll2 v23.4s, v22.8h, #(CONST_BITS) /* tmp0h tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */ - mov v21.16b, v19.16b /* tmp3 = z1 */ - mov v20.16b, v18.16b /* tmp3 = z1 */ - smlal2 v19.4s, v8.8h, XFIX_N_1_847 /* tmp2h tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065); */ - smlal v18.4s, v8.4h, XFIX_N_1_847 /* tmp2l tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065); */ - sshll2 v27.4s, v26.8h, #(CONST_BITS) /* tmp1h tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */ - smlal2 v21.4s, v4.8h, XFIX_P_0_765 /* tmp3h tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */ - smlal v20.4s, v4.4h, XFIX_P_0_765 /* tmp3l tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */ - sshll v22.4s, v22.4h, #(CONST_BITS) /* tmp0l tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */ - sshll v26.4s, v26.4h, #(CONST_BITS) /* tmp1l tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */ - add v2.4s, v22.4s, v20.4s /* tmp10l tmp10 = tmp0 + tmp3; */ - sub v6.4s, v22.4s, v20.4s /* tmp13l tmp13 = tmp0 - tmp3; */ - add v8.4s, v26.4s, v18.4s /* tmp11l tmp11 = tmp1 + tmp2; */ - sub v4.4s, v26.4s, v18.4s /* tmp12l tmp12 = tmp1 - tmp2; */ - add v28.4s, v23.4s, v21.4s /* tmp10h tmp10 = tmp0 + tmp3; */ - sub v31.4s, v23.4s, v21.4s /* tmp13h tmp13 = tmp0 - tmp3; */ - add v29.4s, v27.4s, v19.4s /* tmp11h tmp11 = tmp1 + tmp2; */ - sub v30.4s, v27.4s, v19.4s /* tmp12h tmp12 = tmp1 - tmp2; */ - - /* Odd part per figure 8; the matrix is unitary and hence its - * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively. - */ - - add v22.8h, v9.8h, v5.8h /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v24.8h, v7.8h, v3.8h /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v18.8h, v9.8h, v3.8h /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v20.8h, v7.8h, v5.8h /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v26.8h, v22.8h, v24.8h /* z5 = z3 + z4 */ - - smull2 v11.4s, v9.8h, XFIX_P_0_298 /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */ - smull2 v13.4s, v7.8h, XFIX_P_2_053 /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */ - smull2 v15.4s, v5.8h, XFIX_P_3_072 /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */ - smull2 v17.4s, v3.8h, XFIX_P_1_501 /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */ - smull2 v27.4s, v26.8h, XFIX_P_1_175 /* z5h z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */ - smull2 v23.4s, v22.8h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560) */ - smull2 v25.4s, v24.8h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644) */ - smull2 v19.4s, v18.8h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223) */ - smull2 v21.4s, v20.8h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447) */ - - smull v10.4s, v9.4h, XFIX_P_0_298 /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */ - smull v12.4s, v7.4h, XFIX_P_2_053 /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */ - smull v14.4s, v5.4h, XFIX_P_3_072 /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */ - smull v16.4s, v3.4h, XFIX_P_1_501 /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */ - smull v26.4s, v26.4h, XFIX_P_1_175 /* z5l z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */ - smull v22.4s, v22.4h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560) */ - smull v24.4s, v24.4h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644) */ - smull v18.4s, v18.4h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223) */ - smull v20.4s, v20.4h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447) */ - - add v23.4s, v23.4s, v27.4s /* z3 += z5 */ - add v22.4s, v22.4s, v26.4s /* z3 += z5 */ - add v25.4s, v25.4s, v27.4s /* z4 += z5 */ - add v24.4s, v24.4s, v26.4s /* z4 += z5 */ - - add v11.4s, v11.4s, v19.4s /* tmp0 += z1 */ - add v10.4s, v10.4s, v18.4s /* tmp0 += z1 */ - add v13.4s, v13.4s, v21.4s /* tmp1 += z2 */ - add v12.4s, v12.4s, v20.4s /* tmp1 += z2 */ - add v15.4s, v15.4s, v21.4s /* tmp2 += z2 */ - add v14.4s, v14.4s, v20.4s /* tmp2 += z2 */ - add v17.4s, v17.4s, v19.4s /* tmp3 += z1 */ - add v16.4s, v16.4s, v18.4s /* tmp3 += z1 */ - - add v11.4s, v11.4s, v23.4s /* tmp0 += z3 */ - add v10.4s, v10.4s, v22.4s /* tmp0 += z3 */ - add v13.4s, v13.4s, v25.4s /* tmp1 += z4 */ - add v12.4s, v12.4s, v24.4s /* tmp1 += z4 */ - add v17.4s, v17.4s, v25.4s /* tmp3 += z4 */ - add v16.4s, v16.4s, v24.4s /* tmp3 += z4 */ - add v15.4s, v15.4s, v23.4s /* tmp2 += z3 */ - add v14.4s, v14.4s, v22.4s /* tmp2 += z3 */ - - /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */ - - add v18.4s, v2.4s, v16.4s /* tmp10 + tmp3 */ - add v19.4s, v28.4s, v17.4s /* tmp10 + tmp3 */ - sub v20.4s, v2.4s, v16.4s /* tmp10 - tmp3 */ - sub v21.4s, v28.4s, v17.4s /* tmp10 - tmp3 */ - add v22.4s, v8.4s, v14.4s /* tmp11 + tmp2 */ - add v23.4s, v29.4s, v15.4s /* tmp11 + tmp2 */ - sub v24.4s, v8.4s, v14.4s /* tmp11 - tmp2 */ - sub v25.4s, v29.4s, v15.4s /* tmp11 - tmp2 */ - add v26.4s, v4.4s, v12.4s /* tmp12 + tmp1 */ - add v27.4s, v30.4s, v13.4s /* tmp12 + tmp1 */ - sub v28.4s, v4.4s, v12.4s /* tmp12 - tmp1 */ - sub v29.4s, v30.4s, v13.4s /* tmp12 - tmp1 */ - add v14.4s, v6.4s, v10.4s /* tmp13 + tmp0 */ - add v15.4s, v31.4s, v11.4s /* tmp13 + tmp0 */ - sub v16.4s, v6.4s, v10.4s /* tmp13 - tmp0 */ - sub v17.4s, v31.4s, v11.4s /* tmp13 - tmp0 */ - - shrn v2.4h, v18.4s, #16 /* wsptr[DCTSIZE*0] = (int)DESCALE(tmp10 + tmp3, CONST_BITS+PASS1_BITS+3) */ - shrn v9.4h, v20.4s, #16 /* wsptr[DCTSIZE*7] = (int)DESCALE(tmp10 - tmp3, CONST_BITS+PASS1_BITS+3) */ - shrn v3.4h, v22.4s, #16 /* wsptr[DCTSIZE*1] = (int)DESCALE(tmp11 + tmp2, CONST_BITS+PASS1_BITS+3) */ - shrn v8.4h, v24.4s, #16 /* wsptr[DCTSIZE*6] = (int)DESCALE(tmp11 - tmp2, CONST_BITS+PASS1_BITS+3) */ - shrn v4.4h, v26.4s, #16 /* wsptr[DCTSIZE*2] = (int)DESCALE(tmp12 + tmp1, CONST_BITS+PASS1_BITS+3) */ - shrn v7.4h, v28.4s, #16 /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS+PASS1_BITS+3) */ - shrn v5.4h, v14.4s, #16 /* wsptr[DCTSIZE*3] = (int)DESCALE(tmp13 + tmp0, CONST_BITS+PASS1_BITS+3) */ - shrn v6.4h, v16.4s, #16 /* wsptr[DCTSIZE*4] = (int)DESCALE(tmp13 - tmp0, CONST_BITS+PASS1_BITS+3) */ - shrn2 v2.8h, v19.4s, #16 /* wsptr[DCTSIZE*0] = (int)DESCALE(tmp10 + tmp3, CONST_BITS+PASS1_BITS+3) */ - shrn2 v9.8h, v21.4s, #16 /* wsptr[DCTSIZE*7] = (int)DESCALE(tmp10 - tmp3, CONST_BITS+PASS1_BITS+3) */ - shrn2 v3.8h, v23.4s, #16 /* wsptr[DCTSIZE*1] = (int)DESCALE(tmp11 + tmp2, CONST_BITS+PASS1_BITS+3) */ - shrn2 v8.8h, v25.4s, #16 /* wsptr[DCTSIZE*6] = (int)DESCALE(tmp11 - tmp2, CONST_BITS+PASS1_BITS+3) */ - shrn2 v4.8h, v27.4s, #16 /* wsptr[DCTSIZE*2] = (int)DESCALE(tmp12 + tmp1, CONST_BITS+PASS1_BITS+3) */ - shrn2 v7.8h, v29.4s, #16 /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS+PASS1_BITS+3) */ - shrn2 v5.8h, v15.4s, #16 /* wsptr[DCTSIZE*3] = (int)DESCALE(tmp13 + tmp0, CONST_BITS+PASS1_BITS+3) */ - shrn2 v6.8h, v17.4s, #16 /* wsptr[DCTSIZE*4] = (int)DESCALE(tmp13 - tmp0, CONST_BITS+PASS1_BITS+3) */ - movi v0.16b, #(CENTERJSAMPLE) - /* Prepare pointers (dual-issue with NEON instructions) */ - ldp TMP1, TMP2, [OUTPUT_BUF], 16 - sqrshrn v28.8b, v2.8h, #(CONST_BITS+PASS1_BITS+3-16) - ldp TMP3, TMP4, [OUTPUT_BUF], 16 - sqrshrn v29.8b, v3.8h, #(CONST_BITS+PASS1_BITS+3-16) - add TMP1, TMP1, OUTPUT_COL - sqrshrn v30.8b, v4.8h, #(CONST_BITS+PASS1_BITS+3-16) - add TMP2, TMP2, OUTPUT_COL - sqrshrn v31.8b, v5.8h, #(CONST_BITS+PASS1_BITS+3-16) - add TMP3, TMP3, OUTPUT_COL - sqrshrn2 v28.16b, v6.8h, #(CONST_BITS+PASS1_BITS+3-16) - add TMP4, TMP4, OUTPUT_COL - sqrshrn2 v29.16b, v7.8h, #(CONST_BITS+PASS1_BITS+3-16) - ldp TMP5, TMP6, [OUTPUT_BUF], 16 - sqrshrn2 v30.16b, v8.8h, #(CONST_BITS+PASS1_BITS+3-16) - ldp TMP7, TMP8, [OUTPUT_BUF], 16 - sqrshrn2 v31.16b, v9.8h, #(CONST_BITS+PASS1_BITS+3-16) - add TMP5, TMP5, OUTPUT_COL - add v16.16b, v28.16b, v0.16b - add TMP6, TMP6, OUTPUT_COL - add v18.16b, v29.16b, v0.16b - add TMP7, TMP7, OUTPUT_COL - add v20.16b, v30.16b, v0.16b - add TMP8, TMP8, OUTPUT_COL - add v22.16b, v31.16b, v0.16b - - /* Transpose the final 8-bit samples */ - trn1 v28.16b, v16.16b, v18.16b - trn1 v30.16b, v20.16b, v22.16b - trn2 v29.16b, v16.16b, v18.16b - trn2 v31.16b, v20.16b, v22.16b - - trn1 v16.8h, v28.8h, v30.8h - trn2 v18.8h, v28.8h, v30.8h - trn1 v20.8h, v29.8h, v31.8h - trn2 v22.8h, v29.8h, v31.8h - - uzp1 v28.4s, v16.4s, v18.4s - uzp2 v30.4s, v16.4s, v18.4s - uzp1 v29.4s, v20.4s, v22.4s - uzp2 v31.4s, v20.4s, v22.4s - - /* Store results to the output buffer */ - st1 {v28.d}[0], [TMP1] - st1 {v29.d}[0], [TMP2] - st1 {v28.d}[1], [TMP3] - st1 {v29.d}[1], [TMP4] - st1 {v30.d}[0], [TMP5] - st1 {v31.d}[0], [TMP6] - st1 {v30.d}[1], [TMP7] - st1 {v31.d}[1], [TMP8] - ld1 {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], #32 - ld1 {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], #32 - blr x30 - -.balign 16 -2: - mul v3.8h, v3.8h, v19.8h - mul v4.8h, v4.8h, v20.8h - mul v5.8h, v5.8h, v21.8h - add TMP4, xzr, TMP2, LSL #32 - mul v6.8h, v6.8h, v22.8h - mul v7.8h, v7.8h, v23.8h - adds TMP3, xzr, TMP2, LSR #32 - mul v8.8h, v8.8h, v24.8h - mul v9.8h, v9.8h, v25.8h - b.ne 3f - /* Right AC coef is zero */ - dup v15.2d, v10.d[1] - /* Even part: reverse the even part of the forward DCT. */ - add v18.4h, v4.4h, v8.4h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */ - add v22.4h, v2.4h, v6.4h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - sub v26.4h, v2.4h, v6.4h /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - smull v18.4s, v18.4h, XFIX_P_0_541 /* z1l z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */ - sshll v22.4s, v22.4h, #(CONST_BITS) /* tmp0l tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */ - mov v20.16b, v18.16b /* tmp3 = z1 */ - sshll v26.4s, v26.4h, #(CONST_BITS) /* tmp1l tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */ - smlal v18.4s, v8.4h, XFIX_N_1_847 /* tmp2l tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065); */ - smlal v20.4s, v4.4h, XFIX_P_0_765 /* tmp3l tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */ - add v2.4s, v22.4s, v20.4s /* tmp10l tmp10 = tmp0 + tmp3; */ - sub v6.4s, v22.4s, v20.4s /* tmp13l tmp13 = tmp0 - tmp3; */ - add v8.4s, v26.4s, v18.4s /* tmp11l tmp11 = tmp1 + tmp2; */ - sub v4.4s, v26.4s, v18.4s /* tmp12l tmp12 = tmp1 - tmp2; */ - - /* Odd part per figure 8; the matrix is unitary and hence its - * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively. - */ - - add v22.4h, v9.4h, v5.4h /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v24.4h, v7.4h, v3.4h /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v18.4h, v9.4h, v3.4h /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v20.4h, v7.4h, v5.4h /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v26.4h, v22.4h, v24.4h /* z5 = z3 + z4 */ - - smull v10.4s, v9.4h, XFIX_P_0_298 /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */ - smull v12.4s, v7.4h, XFIX_P_2_053 /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */ - smull v14.4s, v5.4h, XFIX_P_3_072 /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */ - smull v16.4s, v3.4h, XFIX_P_1_501 /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */ - smull v26.4s, v26.4h, XFIX_P_1_175 /* z5l z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */ - smull v22.4s, v22.4h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560) */ - smull v24.4s, v24.4h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644) */ - smull v18.4s, v18.4h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223) */ - smull v20.4s, v20.4h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447) */ - - add v22.4s, v22.4s, v26.4s /* z3 += z5 */ - add v24.4s, v24.4s, v26.4s /* z4 += z5 */ - - add v10.4s, v10.4s, v18.4s /* tmp0 += z1 */ - add v12.4s, v12.4s, v20.4s /* tmp1 += z2 */ - add v14.4s, v14.4s, v20.4s /* tmp2 += z2 */ - add v16.4s, v16.4s, v18.4s /* tmp3 += z1 */ - - add v10.4s, v10.4s, v22.4s /* tmp0 += z3 */ - add v12.4s, v12.4s, v24.4s /* tmp1 += z4 */ - add v16.4s, v16.4s, v24.4s /* tmp3 += z4 */ - add v14.4s, v14.4s, v22.4s /* tmp2 += z3 */ - - /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */ - - add v18.4s, v2.4s, v16.4s /* tmp10 + tmp3 */ - sub v20.4s, v2.4s, v16.4s /* tmp10 - tmp3 */ - add v22.4s, v8.4s, v14.4s /* tmp11 + tmp2 */ - sub v24.4s, v8.4s, v14.4s /* tmp11 - tmp2 */ - add v26.4s, v4.4s, v12.4s /* tmp12 + tmp1 */ - sub v28.4s, v4.4s, v12.4s /* tmp12 - tmp1 */ - add v14.4s, v6.4s, v10.4s /* tmp13 + tmp0 */ - sub v16.4s, v6.4s, v10.4s /* tmp13 - tmp0 */ - - rshrn v2.4h, v18.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*0] = (int)DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */ - rshrn v3.4h, v22.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*1] = (int)DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */ - rshrn v4.4h, v26.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*2] = (int)DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */ - rshrn v5.4h, v14.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*3] = (int)DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */ - rshrn2 v2.8h, v16.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*4] = (int)DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */ - rshrn2 v3.8h, v28.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */ - rshrn2 v4.8h, v24.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*6] = (int)DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */ - rshrn2 v5.8h, v20.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*7] = (int)DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */ - mov v6.16b, v15.16b - mov v7.16b, v15.16b - mov v8.16b, v15.16b - mov v9.16b, v15.16b - b 1b - -.balign 16 -3: - cbnz TMP4, 4f - /* Left AC coef is zero */ - dup v14.2d, v10.d[0] - /* Even part: reverse the even part of the forward DCT. */ - add v18.8h, v4.8h, v8.8h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */ - add v22.8h, v2.8h, v6.8h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - smull2 v19.4s, v18.8h, XFIX_P_0_541 /* z1h z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */ - sub v26.8h, v2.8h, v6.8h /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - sshll2 v23.4s, v22.8h, #(CONST_BITS) /* tmp0h tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */ - mov v21.16b, v19.16b /* tmp3 = z1 */ - smlal2 v19.4s, v8.8h, XFIX_N_1_847 /* tmp2h tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065); */ - sshll2 v27.4s, v26.8h, #(CONST_BITS) /* tmp1h tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */ - smlal2 v21.4s, v4.8h, XFIX_P_0_765 /* tmp3h tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */ - add v28.4s, v23.4s, v21.4s /* tmp10h tmp10 = tmp0 + tmp3; */ - sub v31.4s, v23.4s, v21.4s /* tmp13h tmp13 = tmp0 - tmp3; */ - add v29.4s, v27.4s, v19.4s /* tmp11h tmp11 = tmp1 + tmp2; */ - sub v30.4s, v27.4s, v19.4s /* tmp12h tmp12 = tmp1 - tmp2; */ - - /* Odd part per figure 8; the matrix is unitary and hence its - * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively. - */ - - add v22.8h, v9.8h, v5.8h /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v24.8h, v7.8h, v3.8h /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v18.8h, v9.8h, v3.8h /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v20.8h, v7.8h, v5.8h /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v26.8h, v22.8h, v24.8h /* z5 = z3 + z4 */ - - smull2 v11.4s, v9.8h, XFIX_P_0_298 /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */ - smull2 v13.4s, v7.8h, XFIX_P_2_053 /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */ - smull2 v15.4s, v5.8h, XFIX_P_3_072 /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */ - smull2 v17.4s, v3.8h, XFIX_P_1_501 /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */ - smull2 v27.4s, v26.8h, XFIX_P_1_175 /* z5h z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */ - smull2 v23.4s, v22.8h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560) */ - smull2 v25.4s, v24.8h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644) */ - smull2 v19.4s, v18.8h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223) */ - smull2 v21.4s, v20.8h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447) */ - - add v23.4s, v23.4s, v27.4s /* z3 += z5 */ - add v22.4s, v22.4s, v26.4s /* z3 += z5 */ - add v25.4s, v25.4s, v27.4s /* z4 += z5 */ - add v24.4s, v24.4s, v26.4s /* z4 += z5 */ - - add v11.4s, v11.4s, v19.4s /* tmp0 += z1 */ - add v13.4s, v13.4s, v21.4s /* tmp1 += z2 */ - add v15.4s, v15.4s, v21.4s /* tmp2 += z2 */ - add v17.4s, v17.4s, v19.4s /* tmp3 += z1 */ - - add v11.4s, v11.4s, v23.4s /* tmp0 += z3 */ - add v13.4s, v13.4s, v25.4s /* tmp1 += z4 */ - add v17.4s, v17.4s, v25.4s /* tmp3 += z4 */ - add v15.4s, v15.4s, v23.4s /* tmp2 += z3 */ - - /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */ - - add v19.4s, v28.4s, v17.4s /* tmp10 + tmp3 */ - sub v21.4s, v28.4s, v17.4s /* tmp10 - tmp3 */ - add v23.4s, v29.4s, v15.4s /* tmp11 + tmp2 */ - sub v25.4s, v29.4s, v15.4s /* tmp11 - tmp2 */ - add v27.4s, v30.4s, v13.4s /* tmp12 + tmp1 */ - sub v29.4s, v30.4s, v13.4s /* tmp12 - tmp1 */ - add v15.4s, v31.4s, v11.4s /* tmp13 + tmp0 */ - sub v17.4s, v31.4s, v11.4s /* tmp13 - tmp0 */ - - mov v2.16b, v14.16b - mov v3.16b, v14.16b - mov v4.16b, v14.16b - mov v5.16b, v14.16b - rshrn v6.4h, v19.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*0] = (int)DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */ - rshrn v7.4h, v23.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*1] = (int)DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */ - rshrn v8.4h, v27.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*2] = (int)DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */ - rshrn v9.4h, v15.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*3] = (int)DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */ - rshrn2 v6.8h, v17.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*4] = (int)DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */ - rshrn2 v7.8h, v29.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */ - rshrn2 v8.8h, v25.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*6] = (int)DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */ - rshrn2 v9.8h, v21.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*7] = (int)DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */ - b 1b - -.balign 16 -4: - /* "No" AC coef is zero */ - /* Even part: reverse the even part of the forward DCT. */ - add v18.8h, v4.8h, v8.8h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */ - add v22.8h, v2.8h, v6.8h /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - smull2 v19.4s, v18.8h, XFIX_P_0_541 /* z1h z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */ - sub v26.8h, v2.8h, v6.8h /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */ - smull v18.4s, v18.4h, XFIX_P_0_541 /* z1l z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */ - sshll2 v23.4s, v22.8h, #(CONST_BITS) /* tmp0h tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */ - mov v21.16b, v19.16b /* tmp3 = z1 */ - mov v20.16b, v18.16b /* tmp3 = z1 */ - smlal2 v19.4s, v8.8h, XFIX_N_1_847 /* tmp2h tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065); */ - smlal v18.4s, v8.4h, XFIX_N_1_847 /* tmp2l tmp2 = z1 + MULTIPLY(z3, -FIX_1_847759065); */ - sshll2 v27.4s, v26.8h, #(CONST_BITS) /* tmp1h tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */ - smlal2 v21.4s, v4.8h, XFIX_P_0_765 /* tmp3h tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */ - smlal v20.4s, v4.4h, XFIX_P_0_765 /* tmp3l tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */ - sshll v22.4s, v22.4h, #(CONST_BITS) /* tmp0l tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */ - sshll v26.4s, v26.4h, #(CONST_BITS) /* tmp1l tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */ - add v2.4s, v22.4s, v20.4s /* tmp10l tmp10 = tmp0 + tmp3; */ - sub v6.4s, v22.4s, v20.4s /* tmp13l tmp13 = tmp0 - tmp3; */ - add v8.4s, v26.4s, v18.4s /* tmp11l tmp11 = tmp1 + tmp2; */ - sub v4.4s, v26.4s, v18.4s /* tmp12l tmp12 = tmp1 - tmp2; */ - add v28.4s, v23.4s, v21.4s /* tmp10h tmp10 = tmp0 + tmp3; */ - sub v31.4s, v23.4s, v21.4s /* tmp13h tmp13 = tmp0 - tmp3; */ - add v29.4s, v27.4s, v19.4s /* tmp11h tmp11 = tmp1 + tmp2; */ - sub v30.4s, v27.4s, v19.4s /* tmp12h tmp12 = tmp1 - tmp2; */ - - /* Odd part per figure 8; the matrix is unitary and hence its - * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively. - */ - - add v22.8h, v9.8h, v5.8h /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v24.8h, v7.8h, v3.8h /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v18.8h, v9.8h, v3.8h /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */ - add v20.8h, v7.8h, v5.8h /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */ - add v26.8h, v22.8h, v24.8h /* z5 = z3 + z4 */ - - smull2 v11.4s, v9.8h, XFIX_P_0_298 /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */ - smull2 v13.4s, v7.8h, XFIX_P_2_053 /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */ - smull2 v15.4s, v5.8h, XFIX_P_3_072 /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */ - smull2 v17.4s, v3.8h, XFIX_P_1_501 /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */ - smull2 v27.4s, v26.8h, XFIX_P_1_175 /* z5h z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */ - smull2 v23.4s, v22.8h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560) */ - smull2 v25.4s, v24.8h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644) */ - smull2 v19.4s, v18.8h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223) */ - smull2 v21.4s, v20.8h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447) */ - - smull v10.4s, v9.4h, XFIX_P_0_298 /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */ - smull v12.4s, v7.4h, XFIX_P_2_053 /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */ - smull v14.4s, v5.4h, XFIX_P_3_072 /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */ - smull v16.4s, v3.4h, XFIX_P_1_501 /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */ - smull v26.4s, v26.4h, XFIX_P_1_175 /* z5l z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */ - smull v22.4s, v22.4h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560) */ - smull v24.4s, v24.4h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644) */ - smull v18.4s, v18.4h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223) */ - smull v20.4s, v20.4h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447) */ - - add v23.4s, v23.4s, v27.4s /* z3 += z5 */ - add v22.4s, v22.4s, v26.4s /* z3 += z5 */ - add v25.4s, v25.4s, v27.4s /* z4 += z5 */ - add v24.4s, v24.4s, v26.4s /* z4 += z5 */ - - add v11.4s, v11.4s, v19.4s /* tmp0 += z1 */ - add v10.4s, v10.4s, v18.4s /* tmp0 += z1 */ - add v13.4s, v13.4s, v21.4s /* tmp1 += z2 */ - add v12.4s, v12.4s, v20.4s /* tmp1 += z2 */ - add v15.4s, v15.4s, v21.4s /* tmp2 += z2 */ - add v14.4s, v14.4s, v20.4s /* tmp2 += z2 */ - add v17.4s, v17.4s, v19.4s /* tmp3 += z1 */ - add v16.4s, v16.4s, v18.4s /* tmp3 += z1 */ - - add v11.4s, v11.4s, v23.4s /* tmp0 += z3 */ - add v10.4s, v10.4s, v22.4s /* tmp0 += z3 */ - add v13.4s, v13.4s, v25.4s /* tmp1 += z4 */ - add v12.4s, v12.4s, v24.4s /* tmp1 += z4 */ - add v17.4s, v17.4s, v25.4s /* tmp3 += z4 */ - add v16.4s, v16.4s, v24.4s /* tmp3 += z4 */ - add v15.4s, v15.4s, v23.4s /* tmp2 += z3 */ - add v14.4s, v14.4s, v22.4s /* tmp2 += z3 */ - - /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */ - - add v18.4s, v2.4s, v16.4s /* tmp10 + tmp3 */ - add v19.4s, v28.4s, v17.4s /* tmp10 + tmp3 */ - sub v20.4s, v2.4s, v16.4s /* tmp10 - tmp3 */ - sub v21.4s, v28.4s, v17.4s /* tmp10 - tmp3 */ - add v22.4s, v8.4s, v14.4s /* tmp11 + tmp2 */ - add v23.4s, v29.4s, v15.4s /* tmp11 + tmp2 */ - sub v24.4s, v8.4s, v14.4s /* tmp11 - tmp2 */ - sub v25.4s, v29.4s, v15.4s /* tmp11 - tmp2 */ - add v26.4s, v4.4s, v12.4s /* tmp12 + tmp1 */ - add v27.4s, v30.4s, v13.4s /* tmp12 + tmp1 */ - sub v28.4s, v4.4s, v12.4s /* tmp12 - tmp1 */ - sub v29.4s, v30.4s, v13.4s /* tmp12 - tmp1 */ - add v14.4s, v6.4s, v10.4s /* tmp13 + tmp0 */ - add v15.4s, v31.4s, v11.4s /* tmp13 + tmp0 */ - sub v16.4s, v6.4s, v10.4s /* tmp13 - tmp0 */ - sub v17.4s, v31.4s, v11.4s /* tmp13 - tmp0 */ - - rshrn v2.4h, v18.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*0] = (int)DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */ - rshrn v3.4h, v22.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*1] = (int)DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */ - rshrn v4.4h, v26.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*2] = (int)DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */ - rshrn v5.4h, v14.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*3] = (int)DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */ - rshrn v6.4h, v19.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*0] = (int)DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */ - rshrn v7.4h, v23.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*1] = (int)DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */ - rshrn v8.4h, v27.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*2] = (int)DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */ - rshrn v9.4h, v15.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*3] = (int)DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */ - rshrn2 v2.8h, v16.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*4] = (int)DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */ - rshrn2 v3.8h, v28.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */ - rshrn2 v4.8h, v24.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*6] = (int)DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */ - rshrn2 v5.8h, v20.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*7] = (int)DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */ - rshrn2 v6.8h, v17.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*4] = (int)DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */ - rshrn2 v7.8h, v29.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */ - rshrn2 v8.8h, v25.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*6] = (int)DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */ - rshrn2 v9.8h, v21.4s, #(CONST_BITS-PASS1_BITS) /* wsptr[DCTSIZE*7] = (int)DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */ - b 1b - - .unreq DCT_TABLE - .unreq COEF_BLOCK - .unreq OUTPUT_BUF - .unreq OUTPUT_COL - .unreq TMP1 - .unreq TMP2 - .unreq TMP3 - .unreq TMP4 - .unreq TMP5 - .unreq TMP6 - .unreq TMP7 - .unreq TMP8 - -#undef CENTERJSAMPLE -#undef CONST_BITS -#undef PASS1_BITS -#undef XFIX_P_0_298 -#undef XFIX_N_0_390 -#undef XFIX_P_0_541 -#undef XFIX_P_0_765 -#undef XFIX_N_0_899 -#undef XFIX_P_1_175 -#undef XFIX_P_1_501 -#undef XFIX_N_1_847 -#undef XFIX_N_1_961 -#undef XFIX_P_2_053 -#undef XFIX_N_2_562 -#undef XFIX_P_3_072 - - -/*****************************************************************************/ - -/* * jsimd_idct_ifast_neon * * This function contains a fast, not so accurate integer implementation of diff --git a/simd/arm/common/jidctint-neon.c b/simd/arm/common/jidctint-neon.c new file mode 100644 index 00000000..7fb683b4 --- /dev/null +++ b/simd/arm/common/jidctint-neon.c @@ -0,0 +1,747 @@ +/* + * jidctint-neon.c - slow IDCT (Arm NEON) + * + * Copyright 2019 The Chromium Authors. All Rights Reserved. + * + * This software is provided 'as-is', without any express or implied + * warranty. In no event will the authors be held liable for any damages + * arising from the use of this software. + * + * Permission is granted to anyone to use this software for any purpose, + * including commercial applications, and to alter it and redistribute it + * freely, subject to the following restrictions: + * + * 1. The origin of this software must not be misrepresented; you must not + * claim that you wrote the original software. If you use this software + * in a product, an acknowledgment in the product documentation would be + * appreciated but is not required. + * 2. Altered source versions must be plainly marked as such, and must not be + * misrepresented as being the original software. + * 3. This notice may not be removed or altered from any source distribution. + */ + +#define JPEG_INTERNALS +#include "../../../jinclude.h" +#include "../../../jpeglib.h" +#include "../../../jsimd.h" +#include "../../../jdct.h" +#include "../../../jsimddct.h" +#include "../../jsimd.h" + +#include <arm_neon.h> + +#define CONST_BITS 13 +#define PASS1_BITS 2 + +#define DESCALE_P1 (CONST_BITS - PASS1_BITS) +#define DESCALE_P2 (CONST_BITS + PASS1_BITS + 3) + +/* The computation of the inverse DCT requires the use of constants known at + * compile-time. Scaled integer constants are used to avoid floating-point + * arithmetic: + * 0.298631336 = 2446 * 2^-13 + * 0.390180644 = 3196 * 2^-13 + * 0.541196100 = 4433 * 2^-13 + * 0.765366865 = 6270 * 2^-13 + * 0.899976223 = 7373 * 2^-13 + * 1.175875602 = 9633 * 2^-13 + * 1.501321110 = 12299 * 2^-13 + * 1.847759065 = 15137 * 2^-13 + * 1.961570560 = 16069 * 2^-13 + * 2.053119869 = 16819 * 2^-13 + * 2.562915447 = 20995 * 2^-13 + * 3.072711026 = 25172 * 2^-13 + */ + +#define F_0_298 2446 +#define F_0_390 3196 +#define F_0_541 4433 +#define F_0_765 6270 +#define F_0_899 7373 +#define F_1_175 9633 +#define F_1_501 12299 +#define F_1_847 15137 +#define F_1_961 16069 +#define F_2_053 16819 +#define F_2_562 20995 +#define F_3_072 25172 + +#define F_1_175_MINUS_1_961 (F_1_175 - F_1_961) +#define F_1_175_MINUS_0_390 (F_1_175 - F_0_390) +#define F_0_541_MINUS_1_847 (F_0_541 - F_1_847) +#define F_3_072_MINUS_2_562 (F_3_072 - F_2_562) +#define F_0_298_MINUS_0_899 (F_0_298 - F_0_899) +#define F_1_501_MINUS_0_899 (F_1_501 - F_0_899) +#define F_2_053_MINUS_2_562 (F_2_053 - F_2_562) +#define F_0_541_PLUS_0_765 (F_0_541 + F_0_765) + +__attribute__ ((aligned(8))) static int16_t jsimd_idct_islow_neon_consts[] = { + F_0_899, F_0_541, + F_2_562, F_0_298_MINUS_0_899, + F_1_501_MINUS_0_899, F_2_053_MINUS_2_562, + F_0_541_PLUS_0_765, F_1_175, + F_1_175_MINUS_0_390, F_0_541_MINUS_1_847, + F_3_072_MINUS_2_562, F_1_175_MINUS_1_961, + 0, 0, 0, 0 + }; + +/* Forward declaration of regular and sparse IDCT helper functions. */ + +static inline void jsimd_idct_islow_pass1_regular(int16x4_t row0, + int16x4_t row1, + int16x4_t row2, + int16x4_t row3, + int16x4_t row4, + int16x4_t row5, + int16x4_t row6, + int16x4_t row7, + int16x4_t quant_row0, + int16x4_t quant_row1, + int16x4_t quant_row2, + int16x4_t quant_row3, + int16x4_t quant_row4, + int16x4_t quant_row5, + int16x4_t quant_row6, + int16x4_t quant_row7, + int16_t *workspace_1, + int16_t *workspace_2); + +static inline void jsimd_idct_islow_pass1_sparse(int16x4_t row0, + int16x4_t row1, + int16x4_t row2, + int16x4_t row3, + int16x4_t quant_row0, + int16x4_t quant_row1, + int16x4_t quant_row2, + int16x4_t quant_row3, + int16_t *workspace_1, + int16_t *workspace_2); + +static inline void jsimd_idct_islow_pass2_regular(int16_t *workspace, + JSAMPARRAY output_buf, + JDIMENSION output_col, + unsigned buf_offset); + +static inline void jsimd_idct_islow_pass2_sparse(int16_t *workspace, + JSAMPARRAY output_buf, + JDIMENSION output_col, + unsigned buf_offset); + + +/* Performs dequantization and inverse DCT on one block of coefficients. For + * reference, the C implementation 'jpeg_idct_slow' can be found jidctint.c. + * + * Optimization techniques used for data access: + * + * In each pass, the inverse DCT is computed on the left and right 4x8 halves + * of the DCT block. This avoids spilling due to register pressure and the + * increased granularity allows an optimized calculation depending on the + * values of the DCT coefficients. Between passes, intermediate data is stored + * in 4x8 workspace buffers. + * + * Transposing the 8x8 DCT block after each pass can be achieved by transposing + * each of the four 4x4 quadrants, and swapping quadrants 1 and 2 (in the + * diagram below.) Swapping quadrants is cheap as the second pass can just load + * from the other workspace buffer. + * + * +-------+-------+ +-------+-------+ + * | | | | | | + * | 0 | 1 | | 0 | 2 | + * | | | transpose | | | + * +-------+-------+ ------> +-------+-------+ + * | | | | | | + * | 2 | 3 | | 1 | 3 | + * | | | | | | + * +-------+-------+ +-------+-------+ + * + * Optimization techniques used to accelerate the inverse DCT calculation: + * + * In a DCT coefficient block, the coefficients are increasingly likely to be 0 + * moving diagonally from top left to bottom right. If whole rows of + * coefficients are 0, the inverse DCT calculation can be simplified. In this + * NEON implementation, on the first pass of the inverse DCT, we test for three + * special cases before defaulting to a full 'regular' inverse DCT: + * + * i) AC and DC coefficients are all zero. (Only tested for the right 4x8 + * half of the DCT coefficient block.) In this case the inverse DCT result + * is all zero. We do no work here, signalling that the 'sparse' case is + * required in the second pass. + * ii) AC coefficients (all but the top row) are zero. In this case, the value + * of the inverse DCT of the AC coefficients is just the DC coefficients. + * iii) Coefficients of rows 4, 5, 6 and 7 are all zero. In this case we opt to + * execute a 'sparse' simplified inverse DCT. + * + * In the second pass, only a single special case is tested: whether the the AC + * and DC coefficients were all zero in the right 4x8 block in the first pass + * (case 'i'). If this is the case, a 'sparse' variant of the second pass + * inverse DCT is executed for both the left and right halves of the DCT block. + * (The transposition after the first pass would have made the bottom half of + * the block all zero.) + */ + +void jsimd_idct_islow_neon(void *dct_table, + JCOEFPTR coef_block, + JSAMPARRAY output_buf, + JDIMENSION output_col) +{ + ISLOW_MULT_TYPE *quantptr = dct_table; + + int16_t workspace_l[8 * DCTSIZE / 2]; + int16_t workspace_r[8 * DCTSIZE / 2]; + + /* Compute IDCT first pass on left 4x8 coefficient block. */ + /* Load DCT coefficients in left 4x8 block. */ + int16x4_t row0 = vld1_s16(coef_block + 0 * DCTSIZE); + int16x4_t row1 = vld1_s16(coef_block + 1 * DCTSIZE); + int16x4_t row2 = vld1_s16(coef_block + 2 * DCTSIZE); + int16x4_t row3 = vld1_s16(coef_block + 3 * DCTSIZE); + int16x4_t row4 = vld1_s16(coef_block + 4 * DCTSIZE); + int16x4_t row5 = vld1_s16(coef_block + 5 * DCTSIZE); + int16x4_t row6 = vld1_s16(coef_block + 6 * DCTSIZE); + int16x4_t row7 = vld1_s16(coef_block + 7 * DCTSIZE); + + /* Load quantization table for left 4x8 block. */ + int16x4_t quant_row0 = vld1_s16(quantptr + 0 * DCTSIZE); + int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE); + int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE); + int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE); + int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE); + int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE); + int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE); + int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE); + + /* Construct bitmap to test if DCT coefficients in left 4x8 block are 0. */ + int16x4_t bitmap = vorr_s16(row7, row6); + bitmap = vorr_s16(bitmap, row5); + bitmap = vorr_s16(bitmap, row4); + int64_t bitmap_rows_4567 = vreinterpret_s64_s16(bitmap); + + if (bitmap_rows_4567 == 0) { + bitmap = vorr_s16(bitmap, row3); + bitmap = vorr_s16(bitmap, row2); + bitmap = vorr_s16(bitmap, row1); + int64_t left_ac_bitmap = vreinterpret_s64_s16(bitmap); + + if (left_ac_bitmap == 0) { + int16x4_t dcval = vshl_n_s16(vmul_s16(row0, quant_row0), PASS1_BITS); + int16x4x4_t quadrant = { dcval, dcval, dcval, dcval }; + /* Store 4x4 blocks to workspace, transposing in the process. */ + vst4_s16(workspace_l, quadrant); + vst4_s16(workspace_r, quadrant); + } else { + jsimd_idct_islow_pass1_sparse(row0, row1, row2, row3, quant_row0, + quant_row1, quant_row2, quant_row3, + workspace_l, workspace_r); + } + } else { + jsimd_idct_islow_pass1_regular(row0, row1, row2, row3, row4, row5, + row6, row7, quant_row0, quant_row1, + quant_row2, quant_row3, quant_row4, + quant_row5, quant_row6, quant_row7, + workspace_l, workspace_r); + } + + /* Compute IDCT first pass on right 4x8 coefficient block.*/ + /* Load DCT coefficients for right 4x8 block. */ + row0 = vld1_s16(coef_block + 0 * DCTSIZE + 4); + row1 = vld1_s16(coef_block + 1 * DCTSIZE + 4); + row2 = vld1_s16(coef_block + 2 * DCTSIZE + 4); + row3 = vld1_s16(coef_block + 3 * DCTSIZE + 4); + row4 = vld1_s16(coef_block + 4 * DCTSIZE + 4); + row5 = vld1_s16(coef_block + 5 * DCTSIZE + 4); + row6 = vld1_s16(coef_block + 6 * DCTSIZE + 4); + row7 = vld1_s16(coef_block + 7 * DCTSIZE + 4); + + /* Load quantization table for right 4x8 block. */ + quant_row0 = vld1_s16(quantptr + 0 * DCTSIZE + 4); + quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4); + quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4); + quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4); + quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE + 4); + quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE + 4); + quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4); + quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4); + + /* Construct bitmap to test if DCT coefficients in right 4x8 block are 0. */ + bitmap = vorr_s16(row7, row6); + bitmap = vorr_s16(bitmap, row5); + bitmap = vorr_s16(bitmap, row4); + bitmap_rows_4567 = vreinterpret_s64_s16(bitmap); + bitmap = vorr_s16(bitmap, row3); + bitmap = vorr_s16(bitmap, row2); + bitmap = vorr_s16(bitmap, row1); + int64_t right_ac_bitmap = vreinterpret_s64_s16(bitmap); + + /* Initialise to non-zero value: defaults to regular second pass. */ + int64_t right_ac_dc_bitmap = 1; + + if (right_ac_bitmap == 0) { + bitmap = vorr_s16(bitmap, row0); + right_ac_dc_bitmap = vreinterpret_s64_s16(bitmap); + + if (right_ac_dc_bitmap != 0) { + int16x4_t dcval = vshl_n_s16(vmul_s16(row0, quant_row0), PASS1_BITS); + int16x4x4_t quadrant = { dcval, dcval, dcval, dcval }; + /* Store 4x4 blocks to workspace, transposing in the process. */ + vst4_s16(workspace_l + 4 * DCTSIZE / 2, quadrant); + vst4_s16(workspace_r + 4 * DCTSIZE / 2, quadrant); + } + } else { + if (bitmap_rows_4567 == 0) { + jsimd_idct_islow_pass1_sparse(row0, row1, row2, row3, quant_row0, + quant_row1, quant_row2, quant_row3, + workspace_l + 4 * DCTSIZE / 2, + workspace_r + 4 * DCTSIZE / 2); + } else { + jsimd_idct_islow_pass1_regular(row0, row1, row2, row3, row4, row5, + row6, row7, quant_row0, quant_row1, + quant_row2, quant_row3, quant_row4, + quant_row5, quant_row6, quant_row7, + workspace_l + 4 * DCTSIZE / 2, + workspace_r + 4 * DCTSIZE / 2); + } + } + + /* Second pass: compute IDCT on rows in workspace. */ + /* If all coefficients in right 4x8 block are 0, use 'sparse' second pass. */ + if (right_ac_dc_bitmap == 0) { + jsimd_idct_islow_pass2_sparse(workspace_l, output_buf, output_col, 0); + jsimd_idct_islow_pass2_sparse(workspace_r, output_buf, output_col, 4); + } else { + jsimd_idct_islow_pass2_regular(workspace_l, output_buf, output_col, 0); + jsimd_idct_islow_pass2_regular(workspace_r, output_buf, output_col, 4); + } +} + + +/* Performs dequantization and the first pass of the slow-but-accurate inverse + * DCT on a 4x8 block of coefficients. (To process the full 8x8 DCT block this + * function - or some other optimized variant - needs to be called on both the + * right and left 4x8 blocks.) + * + * This 'regular' version assumes that no optimization can be made to the IDCT + * calculation since no useful set of AC coefficients are all 0. + * + * The original C implementation of the slow IDCT 'jpeg_idct_slow' can be found + * in jidctint.c. Algorithmic changes made here are documented inline. + */ + +static inline void jsimd_idct_islow_pass1_regular(int16x4_t row0, + int16x4_t row1, + int16x4_t row2, + int16x4_t row3, + int16x4_t row4, + int16x4_t row5, + int16x4_t row6, + int16x4_t row7, + int16x4_t quant_row0, + int16x4_t quant_row1, + int16x4_t quant_row2, + int16x4_t quant_row3, + int16x4_t quant_row4, + int16x4_t quant_row5, + int16x4_t quant_row6, + int16x4_t quant_row7, + int16_t *workspace_1, + int16_t *workspace_2) +{ + /* Load constants for IDCT calculation. */ + const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts); + + /* Even part. */ + int16x4_t z2_s16 = vmul_s16(row2, quant_row2); + int16x4_t z3_s16 = vmul_s16(row6, quant_row6); + + int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1); + int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2); + tmp2 = vmlal_lane_s16(tmp2, z3_s16, consts.val[2], 1); + tmp3 = vmlal_lane_s16(tmp3, z3_s16, consts.val[0], 1); + + z2_s16 = vmul_s16(row0, quant_row0); + z3_s16 = vmul_s16(row4, quant_row4); + + int32x4_t tmp0 = vshll_n_s16(vadd_s16(z2_s16, z3_s16), CONST_BITS); + int32x4_t tmp1 = vshll_n_s16(vsub_s16(z2_s16, z3_s16), CONST_BITS); + + int32x4_t tmp10 = vaddq_s32(tmp0, tmp3); + int32x4_t tmp13 = vsubq_s32(tmp0, tmp3); + int32x4_t tmp11 = vaddq_s32(tmp1, tmp2); + int32x4_t tmp12 = vsubq_s32(tmp1, tmp2); + + /* Odd part. */ + int16x4_t tmp0_s16 = vmul_s16(row7, quant_row7); + int16x4_t tmp1_s16 = vmul_s16(row5, quant_row5); + int16x4_t tmp2_s16 = vmul_s16(row3, quant_row3); + int16x4_t tmp3_s16 = vmul_s16(row1, quant_row1); + + z3_s16 = vadd_s16(tmp0_s16, tmp2_s16); + int16x4_t z4_s16 = vadd_s16(tmp1_s16, tmp3_s16); + + /* Implementation as per 'jpeg_idct_islow' in jidctint.c: + * z5 = (z3 + z4) * 1.175875602; + * z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; + * z3 += z5; z4 += z5; + * + * This implementation: + * z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; + * z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); + */ + + int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3); + int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3); + z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3); + z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0); + + /* Implementation as per 'jpeg_idct_islow' in jidctint.c: + * z1 = tmp0 + tmp3; z2 = tmp1 + tmp2; + * tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869; + * tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110; + * z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; + * tmp0 += z1 + z3; tmp1 += z2 + z4; + * tmp2 += z2 + z3; tmp3 += z1 + z4; + * + * This implementation: + * tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223; + * tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447; + * tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447); + * tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223); + * tmp0 += z3; tmp1 += z4; + * tmp2 += z3; tmp3 += z4; + */ + + tmp0 = vmull_lane_s16(tmp0_s16, consts.val[0], 3); + tmp1 = vmull_lane_s16(tmp1_s16, consts.val[1], 1); + tmp2 = vmull_lane_s16(tmp2_s16, consts.val[2], 2); + tmp3 = vmull_lane_s16(tmp3_s16, consts.val[1], 0); + + tmp0 = vmlsl_lane_s16(tmp0, tmp3_s16, consts.val[0], 0); + tmp1 = vmlsl_lane_s16(tmp1, tmp2_s16, consts.val[0], 2); + tmp2 = vmlsl_lane_s16(tmp2, tmp1_s16, consts.val[0], 2); + tmp3 = vmlsl_lane_s16(tmp3, tmp0_s16, consts.val[0], 0); + + tmp0 = vaddq_s32(tmp0, z3); + tmp1 = vaddq_s32(tmp1, z4); + tmp2 = vaddq_s32(tmp2, z3); + tmp3 = vaddq_s32(tmp3, z4); + + /* Final output stage: descale and narrow to 16-bit. */ + int16x4x4_t rows_0123 = { vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1), + vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1), + vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1), + vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1) + }; + int16x4x4_t rows_4567 = { vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1), + vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1), + vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1), + vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1) + }; + + /* Store 4x4 blocks to the intermediate workspace ready for second pass. */ + /* (VST4 transposes the blocks - we need to operate on rows in next pass.) */ + vst4_s16(workspace_1, rows_0123); + vst4_s16(workspace_2, rows_4567); +} + + +/* Performs dequantization and the first pass of the slow-but-accurate inverse + * DCT on a 4x8 block of coefficients. + * + * This 'sparse' version assumes that the AC coefficients in rows 4, 5, 6 and 7 + * are all 0. This simplifies the IDCT calculation, accelerating overall + * performance. + */ + +static inline void jsimd_idct_islow_pass1_sparse(int16x4_t row0, + int16x4_t row1, + int16x4_t row2, + int16x4_t row3, + int16x4_t quant_row0, + int16x4_t quant_row1, + int16x4_t quant_row2, + int16x4_t quant_row3, + int16_t *workspace_1, + int16_t *workspace_2) +{ + /* Load constants for IDCT computation. */ + const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts); + + /* Even part. */ + int16x4_t z2_s16 = vmul_s16(row2, quant_row2); + /* z3 is all 0. */ + + int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1); + int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2); + + z2_s16 = vmul_s16(row0, quant_row0); + int32x4_t tmp0 = vshll_n_s16(z2_s16, CONST_BITS); + int32x4_t tmp1 = vshll_n_s16(z2_s16, CONST_BITS); + + int32x4_t tmp10 = vaddq_s32(tmp0, tmp3); + int32x4_t tmp13 = vsubq_s32(tmp0, tmp3); + int32x4_t tmp11 = vaddq_s32(tmp1, tmp2); + int32x4_t tmp12 = vsubq_s32(tmp1, tmp2); + + /* Odd part. */ + /* tmp0 and tmp1 are both all 0. */ + int16x4_t tmp2_s16 = vmul_s16(row3, quant_row3); + int16x4_t tmp3_s16 = vmul_s16(row1, quant_row1); + + int16x4_t z3_s16 = tmp2_s16; + int16x4_t z4_s16 = tmp3_s16; + + int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3); + int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3); + z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3); + z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0); + + tmp0 = vmlsl_lane_s16(z3, tmp3_s16, consts.val[0], 0); + tmp1 = vmlsl_lane_s16(z4, tmp2_s16, consts.val[0], 2); + tmp2 = vmlal_lane_s16(z3, tmp2_s16, consts.val[2], 2); + tmp3 = vmlal_lane_s16(z4, tmp3_s16, consts.val[1], 0); + + /* Final output stage: descale and narrow to 16-bit. */ + int16x4x4_t rows_0123 = { vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1), + vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1), + vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1), + vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1) + }; + int16x4x4_t rows_4567 = { vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1), + vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1), + vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1), + vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1) + }; + + /* Store 4x4 blocks to the intermediate workspace ready for second pass. */ + /* (VST4 transposes the blocks - we need to operate on rows in next pass.) */ + vst4_s16(workspace_1, rows_0123); + vst4_s16(workspace_2, rows_4567); +} + + +/* Performs the second pass of the slow-but-accurate inverse DCT on a 4x8 block + * of coefficients. (To process the full 8x8 DCT block this function - or some + * other optimized variant - needs to be called on both the right and left 4x8 + * blocks.) + * + * This 'regular' version assumes that no optimization can be made to the IDCT + * calculation since no useful set of coefficient values are all 0 after the + * first pass. + * + * Again, the original C implementation of the slow IDCT 'jpeg_idct_slow' can + * be found in jidctint.c. Algorithmic changes made here are documented inline. + */ + +static inline void jsimd_idct_islow_pass2_regular(int16_t *workspace, + JSAMPARRAY output_buf, + JDIMENSION output_col, + unsigned buf_offset) +{ + /* Load constants for IDCT computation. */ + const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts); + + /* Even part. */ + int16x4_t z2_s16 = vld1_s16(workspace + 2 * DCTSIZE / 2); + int16x4_t z3_s16 = vld1_s16(workspace + 6 * DCTSIZE / 2); + + int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1); + int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2); + tmp2 = vmlal_lane_s16(tmp2, z3_s16, consts.val[2], 1); + tmp3 = vmlal_lane_s16(tmp3, z3_s16, consts.val[0], 1); + + z2_s16 = vld1_s16(workspace + 0 * DCTSIZE / 2); + z3_s16 = vld1_s16(workspace + 4 * DCTSIZE / 2); + + int32x4_t tmp0 = vshll_n_s16(vadd_s16(z2_s16, z3_s16), CONST_BITS); + int32x4_t tmp1 = vshll_n_s16(vsub_s16(z2_s16, z3_s16), CONST_BITS); + + int32x4_t tmp10 = vaddq_s32(tmp0, tmp3); + int32x4_t tmp13 = vsubq_s32(tmp0, tmp3); + int32x4_t tmp11 = vaddq_s32(tmp1, tmp2); + int32x4_t tmp12 = vsubq_s32(tmp1, tmp2); + + /* Odd part. */ + int16x4_t tmp0_s16 = vld1_s16(workspace + 7 * DCTSIZE / 2); + int16x4_t tmp1_s16 = vld1_s16(workspace + 5 * DCTSIZE / 2); + int16x4_t tmp2_s16 = vld1_s16(workspace + 3 * DCTSIZE / 2); + int16x4_t tmp3_s16 = vld1_s16(workspace + 1 * DCTSIZE / 2); + + z3_s16 = vadd_s16(tmp0_s16, tmp2_s16); + int16x4_t z4_s16 = vadd_s16(tmp1_s16, tmp3_s16); + + /* Implementation as per 'jpeg_idct_islow' in jidctint.c: + * z5 = (z3 + z4) * 1.175875602; + * z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; + * z3 += z5; z4 += z5; + * + * This implementation: + * z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; + * z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); + */ + + int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3); + int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3); + z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3); + z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0); + + /* Implementation as per 'jpeg_idct_islow' in jidctint.c: + * z1 = tmp0 + tmp3; z2 = tmp1 + tmp2; + * tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869; + * tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110; + * z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; + * tmp0 += z1 + z3; tmp1 += z2 + z4; + * tmp2 += z2 + z3; tmp3 += z1 + z4; + * + * This implementation: + * tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223; + * tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447; + * tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447); + * tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223); + * tmp0 += z3; tmp1 += z4; + * tmp2 += z3; tmp3 += z4; + */ + + tmp0 = vmull_lane_s16(tmp0_s16, consts.val[0], 3); + tmp1 = vmull_lane_s16(tmp1_s16, consts.val[1], 1); + tmp2 = vmull_lane_s16(tmp2_s16, consts.val[2], 2); + tmp3 = vmull_lane_s16(tmp3_s16, consts.val[1], 0); + + tmp0 = vmlsl_lane_s16(tmp0, tmp3_s16, consts.val[0], 0); + tmp1 = vmlsl_lane_s16(tmp1, tmp2_s16, consts.val[0], 2); + tmp2 = vmlsl_lane_s16(tmp2, tmp1_s16, consts.val[0], 2); + tmp3 = vmlsl_lane_s16(tmp3, tmp0_s16, consts.val[0], 0); + + tmp0 = vaddq_s32(tmp0, z3); + tmp1 = vaddq_s32(tmp1, z4); + tmp2 = vaddq_s32(tmp2, z3); + tmp3 = vaddq_s32(tmp3, z4); + + /* Final output stage: descale and narrow to 16-bit. */ + int16x8_t cols_02_s16 = vcombine_s16(vaddhn_s32(tmp10, tmp3), + vaddhn_s32(tmp12, tmp1)); + int16x8_t cols_13_s16 = vcombine_s16(vaddhn_s32(tmp11, tmp2), + vaddhn_s32(tmp13, tmp0)); + int16x8_t cols_46_s16 = vcombine_s16(vsubhn_s32(tmp13, tmp0), + vsubhn_s32(tmp11, tmp2)); + int16x8_t cols_57_s16 = vcombine_s16(vsubhn_s32(tmp12, tmp1), + vsubhn_s32(tmp10, tmp3)); + /* Descale and narrow to 8-bit. */ + int8x8_t cols_02_s8 = vqrshrn_n_s16(cols_02_s16, DESCALE_P2 - 16); + int8x8_t cols_13_s8 = vqrshrn_n_s16(cols_13_s16, DESCALE_P2 - 16); + int8x8_t cols_46_s8 = vqrshrn_n_s16(cols_46_s16, DESCALE_P2 - 16); + int8x8_t cols_57_s8 = vqrshrn_n_s16(cols_57_s16, DESCALE_P2 - 16); + /* Clamp to range [0-255]. */ + uint8x8_t cols_02_u8 = vadd_u8(cols_02_s8, vdup_n_u8(CENTERJSAMPLE)); + uint8x8_t cols_13_u8 = vadd_u8(cols_13_s8, vdup_n_u8(CENTERJSAMPLE)); + uint8x8_t cols_46_u8 = vadd_u8(cols_46_s8, vdup_n_u8(CENTERJSAMPLE)); + uint8x8_t cols_57_u8 = vadd_u8(cols_57_s8, vdup_n_u8(CENTERJSAMPLE)); + + /* Transpose 4x8 block and store to memory. */ + /* Zipping adjacent columns together allows us to store 16-bit elements. */ + uint8x8x2_t cols_01_23 = vzip_u8(cols_02_u8, cols_13_u8); + uint8x8x2_t cols_45_67 = vzip_u8(cols_46_u8, cols_57_u8); + uint16x4x4_t cols_01_23_45_67 = { vreinterpret_u16_u8(cols_01_23.val[0]), + vreinterpret_u16_u8(cols_01_23.val[1]), + vreinterpret_u16_u8(cols_45_67.val[0]), + vreinterpret_u16_u8(cols_45_67.val[1]) + }; + /* VST4 of 16-bit elements completes the transpose. */ + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 0] + output_col), + cols_01_23_45_67, 0); + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 1] + output_col), + cols_01_23_45_67, 1); + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 2] + output_col), + cols_01_23_45_67, 2); + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 3] + output_col), + cols_01_23_45_67, 3); +} + + +/* Performs the second pass of the slow-but-accurate inverse DCT on a 4x8 block + * of coefficients. + * + * This 'sparse' version assumes that the coefficient values (after the first + * pass) in rows 4, 5, 6 and 7 are all 0. This simplifies the IDCT calculation, + * accelerating overall performance. + */ + +static inline void jsimd_idct_islow_pass2_sparse(int16_t *workspace, + JSAMPARRAY output_buf, + JDIMENSION output_col, + unsigned buf_offset) +{ + /* Load constants for IDCT computation. */ + const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts); + + /* Even part. */ + int16x4_t z2_s16 = vld1_s16(workspace + 2 * DCTSIZE / 2); + /* z3 is all 0. */ + + int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1); + int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2); + + z2_s16 = vld1_s16(workspace + 0 * DCTSIZE / 2); + int32x4_t tmp0 = vshll_n_s16(z2_s16, CONST_BITS); + int32x4_t tmp1 = vshll_n_s16(z2_s16, CONST_BITS); + + int32x4_t tmp10 = vaddq_s32(tmp0, tmp3); + int32x4_t tmp13 = vsubq_s32(tmp0, tmp3); + int32x4_t tmp11 = vaddq_s32(tmp1, tmp2); + int32x4_t tmp12 = vsubq_s32(tmp1, tmp2); + + /* Odd part. */ + /* tmp0 and tmp1 are both all 0. */ + int16x4_t tmp2_s16 = vld1_s16(workspace + 3 * DCTSIZE / 2); + int16x4_t tmp3_s16 = vld1_s16(workspace + 1 * DCTSIZE / 2); + + int16x4_t z3_s16 = tmp2_s16; + int16x4_t z4_s16 = tmp3_s16; + + int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3); + z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3); + int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3); + z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0); + + tmp0 = vmlsl_lane_s16(z3, tmp3_s16, consts.val[0], 0); + tmp1 = vmlsl_lane_s16(z4, tmp2_s16, consts.val[0], 2); + tmp2 = vmlal_lane_s16(z3, tmp2_s16, consts.val[2], 2); + tmp3 = vmlal_lane_s16(z4, tmp3_s16, consts.val[1], 0); + + /* Final output stage: descale and narrow to 16-bit. */ + int16x8_t cols_02_s16 = vcombine_s16(vaddhn_s32(tmp10, tmp3), + vaddhn_s32(tmp12, tmp1)); + int16x8_t cols_13_s16 = vcombine_s16(vaddhn_s32(tmp11, tmp2), + vaddhn_s32(tmp13, tmp0)); + int16x8_t cols_46_s16 = vcombine_s16(vsubhn_s32(tmp13, tmp0), + vsubhn_s32(tmp11, tmp2)); + int16x8_t cols_57_s16 = vcombine_s16(vsubhn_s32(tmp12, tmp1), + vsubhn_s32(tmp10, tmp3)); + /* Descale and narrow to 8-bit. */ + int8x8_t cols_02_s8 = vqrshrn_n_s16(cols_02_s16, DESCALE_P2 - 16); + int8x8_t cols_13_s8 = vqrshrn_n_s16(cols_13_s16, DESCALE_P2 - 16); + int8x8_t cols_46_s8 = vqrshrn_n_s16(cols_46_s16, DESCALE_P2 - 16); + int8x8_t cols_57_s8 = vqrshrn_n_s16(cols_57_s16, DESCALE_P2 - 16); + /* Clamp to range [0-255]. */ + uint8x8_t cols_02_u8 = vadd_u8(cols_02_s8, vdup_n_u8(CENTERJSAMPLE)); + uint8x8_t cols_13_u8 = vadd_u8(cols_13_s8, vdup_n_u8(CENTERJSAMPLE)); + uint8x8_t cols_46_u8 = vadd_u8(cols_46_s8, vdup_n_u8(CENTERJSAMPLE)); + uint8x8_t cols_57_u8 = vadd_u8(cols_57_s8, vdup_n_u8(CENTERJSAMPLE)); + + /* Transpose 4x8 block and store to memory. */ + /* Zipping adjacent columns together allow us to store 16-bit elements. */ + uint8x8x2_t cols_01_23 = vzip_u8(cols_02_u8, cols_13_u8); + uint8x8x2_t cols_45_67 = vzip_u8(cols_46_u8, cols_57_u8); + uint16x4x4_t cols_01_23_45_67 = { vreinterpret_u16_u8(cols_01_23.val[0]), + vreinterpret_u16_u8(cols_01_23.val[1]), + vreinterpret_u16_u8(cols_45_67.val[0]), + vreinterpret_u16_u8(cols_45_67.val[1]) + }; + /* VST4 of 16-bit elements completes the transpose. */ + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 0] + output_col), + cols_01_23_45_67, 0); + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 1] + output_col), + cols_01_23_45_67, 1); + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 2] + output_col), + cols_01_23_45_67, 2); + vst4_lane_u16((uint16_t *)(output_buf[buf_offset + 3] + output_col), + cols_01_23_45_67, 3); +} |