From 1f0f3234143c4fe82f2d0e8da500a5908f2421ac Mon Sep 17 00:00:00 2001 From: Nigel Tao Date: Mon, 12 Jun 2023 16:23:53 +1000 Subject: [PATCH] std/jpeg: memoize Huffman table selectors --- release/c/wuffs-unsupported-snapshot.c | 20 +++++++++++++------ std/jpeg/decode_jpeg.wuffs | 11 +++++++++- std/jpeg/decode_mcu_default.wuffs | 6 +++--- .../decode_mcu_progressive_ac_high_bits.wuffs | 2 +- .../decode_mcu_progressive_ac_low_bit.wuffs | 2 +- .../decode_mcu_progressive_dc_high_bits.wuffs | 4 ++-- 6 files changed, 31 insertions(+), 14 deletions(-) diff --git a/release/c/wuffs-unsupported-snapshot.c b/release/c/wuffs-unsupported-snapshot.c index 65f66bfa..238c02bb 100644 --- a/release/c/wuffs-unsupported-snapshot.c +++ b/release/c/wuffs-unsupported-snapshot.c @@ -8860,6 +8860,8 @@ struct wuffs_jpeg__decoder__struct { uint64_t f_mcu_blocks_offset[10]; uint32_t f_mcu_blocks_mx_mul[10]; uint32_t f_mcu_blocks_my_mul[10]; + uint8_t f_mcu_blocks_dc_hselector[10]; + uint8_t f_mcu_blocks_ac_hselector[10]; uint16_t f_mcu_previous_dc_values[4]; uint16_t f_restart_interval; uint16_t f_saved_restart_interval; @@ -40902,6 +40904,8 @@ wuffs_jpeg__decoder__calculate_single_component_scan_fields( self->private_impl.f_mcu_blocks_offset[0u] = self->private_impl.f_components_workbuf_offsets[v_csel]; self->private_impl.f_mcu_blocks_mx_mul[0u] = 8u; self->private_impl.f_mcu_blocks_my_mul[0u] = (8u * self->private_impl.f_components_workbuf_widths[v_csel]); + self->private_impl.f_mcu_blocks_dc_hselector[0u] = (0u | self->private_impl.f_scan_comps_td[0u]); + self->private_impl.f_mcu_blocks_ac_hselector[0u] = (4u | self->private_impl.f_scan_comps_ta[0u]); self->private_impl.f_scan_width_in_mcus = wuffs_jpeg__decoder__quantize_dimension(self, self->private_impl.f_width, self->private_impl.f_components_h[v_csel], self->private_impl.f_max_incl_components_h); self->private_impl.f_scan_height_in_mcus = wuffs_jpeg__decoder__quantize_dimension(self, self->private_impl.f_height, self->private_impl.f_components_v[v_csel], self->private_impl.f_max_incl_components_v); return wuffs_base__make_empty_struct(); @@ -40920,6 +40924,7 @@ wuffs_jpeg__decoder__calculate_multiple_component_scan_fields( uint32_t v_b = 0; uint32_t v_bx_offset = 0; uint32_t v_by_offset = 0; + uint8_t v_ssel = 0; uint8_t v_csel = 0; v_total_hv = 0u; @@ -40955,10 +40960,13 @@ wuffs_jpeg__decoder__calculate_multiple_component_scan_fields( self->private_impl.f_mcu_num_blocks = v_total_hv; v_b = 0u; while (v_b < self->private_impl.f_mcu_num_blocks) { - v_csel = self->private_impl.f_scan_comps_cselector[self->private_impl.f_mcu_blocks_sselector[v_b]]; + v_ssel = self->private_impl.f_mcu_blocks_sselector[v_b]; + v_csel = self->private_impl.f_scan_comps_cselector[v_ssel]; self->private_impl.f_mcu_blocks_offset[v_b] = (self->private_impl.f_components_workbuf_offsets[v_csel] + (8u * ((uint64_t)(self->private_impl.f_scan_comps_bx_offset[v_b]))) + (8u * ((uint64_t)(self->private_impl.f_scan_comps_by_offset[v_b])) * ((uint64_t)(self->private_impl.f_components_workbuf_widths[v_csel])))); self->private_impl.f_mcu_blocks_mx_mul[v_b] = (8u * ((uint32_t)(self->private_impl.f_components_h[v_csel]))); self->private_impl.f_mcu_blocks_my_mul[v_b] = (8u * ((uint32_t)(self->private_impl.f_components_v[v_csel])) * self->private_impl.f_components_workbuf_widths[v_csel]); + self->private_impl.f_mcu_blocks_dc_hselector[v_b] = (0u | self->private_impl.f_scan_comps_td[v_ssel]); + self->private_impl.f_mcu_blocks_ac_hselector[v_b] = (4u | self->private_impl.f_scan_comps_ta[v_ssel]); v_b += 1u; } self->private_impl.f_scan_width_in_mcus = self->private_impl.f_width_in_mcus; @@ -41624,7 +41632,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default( v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u)); iop_v_r += ((63u - (v_n_bits & 63u)) >> 3u); v_n_bits |= 56u; - v_dc_h = (0u | self->private_impl.f_scan_comps_td[self->private_impl.f_mcu_blocks_sselector[self->private_impl.f_mcu_current_block]]); + v_dc_h = self->private_impl.f_mcu_blocks_dc_hselector[self->private_impl.f_mcu_current_block]; v_dc_ht_fast = ((uint32_t)(self->private_impl.f_huff_tables_fast[v_dc_h][(v_bits >> 56u)])); v_dc_bl = (v_dc_ht_fast >> 8u); if (v_n_bits >= v_dc_bl) { @@ -41674,7 +41682,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default( v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u)); iop_v_r += ((63u - (v_n_bits & 63u)) >> 3u); v_n_bits |= 56u; - v_ac_h = (4u | self->private_impl.f_scan_comps_ta[self->private_impl.f_mcu_blocks_sselector[self->private_impl.f_mcu_current_block]]); + v_ac_h = self->private_impl.f_mcu_blocks_ac_hselector[self->private_impl.f_mcu_current_block]; v_ac_ht_fast = ((uint32_t)(self->private_impl.f_huff_tables_fast[v_ac_h][(v_bits >> 56u)])); v_ac_bl = (v_ac_ht_fast >> 8u); if (v_n_bits >= v_ac_bl) { @@ -41833,7 +41841,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits( v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u)); iop_v_r += ((63u - (v_n_bits & 63u)) >> 3u); v_n_bits |= 56u; - v_ac_h = (4u | self->private_impl.f_scan_comps_ta[self->private_impl.f_mcu_blocks_sselector[0u]]); + v_ac_h = self->private_impl.f_mcu_blocks_ac_hselector[0u]; v_ac_ht_fast = ((uint32_t)(self->private_impl.f_huff_tables_fast[v_ac_h][(v_bits >> 56u)])); v_ac_bl = (v_ac_ht_fast >> 8u); if (v_n_bits >= v_ac_bl) { @@ -41984,7 +41992,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_low_bit( v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u)); iop_v_r += ((63u - (v_n_bits & 63u)) >> 3u); v_n_bits |= 56u; - v_ac_h = (4u | self->private_impl.f_scan_comps_ta[self->private_impl.f_mcu_blocks_sselector[0u]]); + v_ac_h = self->private_impl.f_mcu_blocks_ac_hselector[0u]; v_ac_ht_fast = ((uint32_t)(self->private_impl.f_huff_tables_fast[v_ac_h][(v_bits >> 56u)])); v_ac_bl = (v_ac_ht_fast >> 8u); if (v_n_bits >= v_ac_bl) { @@ -42228,7 +42236,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits( v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u)); iop_v_r += ((63u - (v_n_bits & 63u)) >> 3u); v_n_bits |= 56u; - v_dc_h = (0u | self->private_impl.f_scan_comps_td[self->private_impl.f_mcu_blocks_sselector[self->private_impl.f_mcu_current_block]]); + v_dc_h = self->private_impl.f_mcu_blocks_dc_hselector[self->private_impl.f_mcu_current_block]; v_dc_ht_fast = ((uint32_t)(self->private_impl.f_huff_tables_fast[v_dc_h][(v_bits >> 56u)])); v_dc_bl = (v_dc_ht_fast >> 8u); if (v_n_bits >= v_dc_bl) { diff --git a/std/jpeg/decode_jpeg.wuffs b/std/jpeg/decode_jpeg.wuffs index 0884ba3b..dbb4ce15 100644 --- a/std/jpeg/decode_jpeg.wuffs +++ b/std/jpeg/decode_jpeg.wuffs @@ -242,6 +242,9 @@ pub struct decoder? implements base.image_decoder( mcu_blocks_mx_mul : array[10] base.u32[..= 0x00_0020], // 8 * 4. mcu_blocks_my_mul : array[10] base.u32[..= 0x20_0100], // 8 * 4 * 0x1_0008. + mcu_blocks_dc_hselector : array[10] base.u8[..= 3], + mcu_blocks_ac_hselector : array[10] base.u8[..= 7], + mcu_previous_dc_values : array[4] base.u16, restart_interval : base.u16, @@ -1391,6 +1394,8 @@ pri func decoder.calculate_single_component_scan_fields!() { this.mcu_blocks_offset[0] = this.components_workbuf_offsets[csel] this.mcu_blocks_mx_mul[0] = 8 this.mcu_blocks_my_mul[0] = 8 * this.components_workbuf_widths[csel] + this.mcu_blocks_dc_hselector[0] = 0 | this.scan_comps_td[0] + this.mcu_blocks_ac_hselector[0] = 4 | this.scan_comps_ta[0] this.scan_width_in_mcus = this.quantize_dimension( width: this.width, h: this.components_h[csel], max_incl_h: this.max_incl_components_h) @@ -1407,6 +1412,7 @@ pri func decoder.calculate_multiple_component_scan_fields!() base.bool { var b : base.u32 var bx_offset : base.u32 var by_offset : base.u32 + var ssel : base.u8[..= 3] var csel : base.u8[..= 3] total_hv = 0 @@ -1450,12 +1456,15 @@ pri func decoder.calculate_multiple_component_scan_fields!() base.bool { b = 0 while b < this.mcu_num_blocks { assert b < 10 via "a < b: a < c; c <= b"(c: this.mcu_num_blocks) - csel = this.scan_comps_cselector[this.mcu_blocks_sselector[b]] + ssel = this.mcu_blocks_sselector[b] + csel = this.scan_comps_cselector[ssel] this.mcu_blocks_offset[b] = this.components_workbuf_offsets[csel] + (8 * (this.scan_comps_bx_offset[b] as base.u64)) + (8 * (this.scan_comps_by_offset[b] as base.u64) * (this.components_workbuf_widths[csel] as base.u64)) this.mcu_blocks_mx_mul[b] = 8 * (this.components_h[csel] as base.u32) this.mcu_blocks_my_mul[b] = 8 * (this.components_v[csel] as base.u32) * this.components_workbuf_widths[csel] + this.mcu_blocks_dc_hselector[b] = 0 | this.scan_comps_td[ssel] + this.mcu_blocks_ac_hselector[b] = 4 | this.scan_comps_ta[ssel] b += 1 } endwhile diff --git a/std/jpeg/decode_mcu_default.wuffs b/std/jpeg/decode_mcu_default.wuffs index e69d12ca..bceaae89 100644 --- a/std/jpeg/decode_mcu_default.wuffs +++ b/std/jpeg/decode_mcu_default.wuffs @@ -28,7 +28,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m // etc_bl and etc_blm1 are the Huffman code's "bit length" and "bit length // minus 1" for the DC (first) and AC (later) components. - var dc_h : base.u8[..= 7] + var dc_h : base.u8[..= 3] var dc_symbol : base.u32[..= 0x0F] var dc_ht_fast : base.u32 var dc_bl : base.u32 @@ -81,7 +81,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m n_bits |= 56 // Read the Huffman-encoded dc_symbol, up to 16 bits long. - dc_h = 0 | this.scan_comps_td[this.mcu_blocks_sselector[this.mcu_current_block]] + dc_h = this.mcu_blocks_dc_hselector[this.mcu_current_block] dc_ht_fast = this.huff_tables_fast[dc_h][bits >> 56] as base.u32 dc_bl = dc_ht_fast >> 8 if n_bits >= dc_bl { @@ -144,7 +144,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m n_bits |= 56 // Read the Huffman-encoded ac_symbol, up to 16 bits long. - ac_h = 4 | this.scan_comps_ta[this.mcu_blocks_sselector[this.mcu_current_block]] + ac_h = this.mcu_blocks_ac_hselector[this.mcu_current_block] ac_ht_fast = this.huff_tables_fast[ac_h][bits >> 56] as base.u32 ac_bl = ac_ht_fast >> 8 if n_bits >= ac_bl { diff --git a/std/jpeg/decode_mcu_progressive_ac_high_bits.wuffs b/std/jpeg/decode_mcu_progressive_ac_high_bits.wuffs index a5f1dde6..8abc51c3 100644 --- a/std/jpeg/decode_mcu_progressive_ac_high_bits.wuffs +++ b/std/jpeg/decode_mcu_progressive_ac_high_bits.wuffs @@ -61,7 +61,7 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx n_bits |= 56 // Read the Huffman-encoded ac_symbol, up to 16 bits long. - ac_h = 4 | this.scan_comps_ta[this.mcu_blocks_sselector[0]] + ac_h = this.mcu_blocks_ac_hselector[0] ac_ht_fast = this.huff_tables_fast[ac_h][bits >> 56] as base.u32 ac_bl = ac_ht_fast >> 8 if n_bits >= ac_bl { diff --git a/std/jpeg/decode_mcu_progressive_ac_low_bit.wuffs b/std/jpeg/decode_mcu_progressive_ac_low_bit.wuffs index 265a4922..d4622d58 100644 --- a/std/jpeg/decode_mcu_progressive_ac_low_bit.wuffs +++ b/std/jpeg/decode_mcu_progressive_ac_low_bit.wuffs @@ -77,7 +77,7 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx: n_bits |= 56 // Read the Huffman-encoded ac_symbol, up to 16 bits long. - ac_h = 4 | this.scan_comps_ta[this.mcu_blocks_sselector[0]] + ac_h = this.mcu_blocks_ac_hselector[0] ac_ht_fast = this.huff_tables_fast[ac_h][bits >> 56] as base.u32 ac_bl = ac_ht_fast >> 8 if n_bits >= ac_bl { diff --git a/std/jpeg/decode_mcu_progressive_dc_high_bits.wuffs b/std/jpeg/decode_mcu_progressive_dc_high_bits.wuffs index 6f16ba4f..a744916e 100644 --- a/std/jpeg/decode_mcu_progressive_dc_high_bits.wuffs +++ b/std/jpeg/decode_mcu_progressive_dc_high_bits.wuffs @@ -23,7 +23,7 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx var r : base.io_reader var pos : base.u32 - var dc_h : base.u8[..= 7] + var dc_h : base.u8[..= 3] var dc_symbol : base.u32[..= 0x0F] var dc_ht_fast : base.u32 var dc_bl : base.u32 @@ -56,7 +56,7 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx n_bits |= 56 // Read the Huffman-encoded dc_symbol, up to 16 bits long. - dc_h = 0 | this.scan_comps_td[this.mcu_blocks_sselector[this.mcu_current_block]] + dc_h = this.mcu_blocks_dc_hselector[this.mcu_current_block] dc_ht_fast = this.huff_tables_fast[dc_h][bits >> 56] as base.u32 dc_bl = dc_ht_fast >> 8 if n_bits >= dc_bl {