media: vicodec: add the FWHT software codec

Add a software codec based on the Fast Walsh Hadamard Transform.

The original FWHT codec was developed by Tom aan de Wiel, and it was
turned into 'proper' kernel code by Hans Verkuil, with a lot of
performance and memory improvements.

Signed-off-by: Tom aan de Wiel <tom.aandewiel@gmail.com>
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Signed-off-by: Mauro Carvalho Chehab <mchehab+samsung@kernel.org>

drivers/media/platform/vicodec/vicodec-codec.c

@@ -0,0 +1,797 @@
+// SPDX-License-Identifier: GPL-2.0+
+/*
+ * Copyright 2016 Tom aan de Wiel
+ * Copyright 2018 Cisco Systems, Inc. and/or its affiliates. All rights reserved.
+ *
+ * 8x8 Fast Walsh Hadamard Transform in sequency order based on the paper:
+ *
+ * A Recursive Algorithm for Sequency-Ordered Fast Walsh Transforms,
+ * R.D. Brown, 1977
+ */
+
+#include <linux/string.h>
+#include "vicodec-codec.h"
+
+#define ALL_ZEROS 15
+#define DEADZONE_WIDTH 20
+
+static const uint8_t zigzag[64] = {
+	0,
+	1,  8,
+	2,  9, 16,
+	3, 10, 17, 24,
+	4, 11, 18, 25, 32,
+	5, 12, 19, 26, 33, 40,
+	6, 13, 20, 27, 34, 41, 48,
+	7, 14, 21, 28, 35, 42, 49, 56,
+	15, 22, 29, 36, 43, 50, 57,
+	23, 30, 37, 44, 51, 58,
+	31, 38, 45, 52, 59,
+	39, 46, 53, 60,
+	47, 54, 61,
+	55, 62,
+	63,
+};
+
+
+static int rlc(const s16 *in, __be16 *output, int blocktype)
+{
+	s16 block[8 * 8];
+	s16 *wp = block;
+	int i = 0;
+	int x, y;
+	int ret = 0;
+
+	/* read in block from framebuffer */
+	int lastzero_run = 0;
+	int to_encode;
+
+	for (y = 0; y < 8; y++) {
+		for (x = 0; x < 8; x++) {
+			*wp = in[x + y * 8];
+			wp++;
+		}
+	}
+
+	/* keep track of amount of trailing zeros */
+	for (i = 63; i >= 0 && !block[zigzag[i]]; i--)
+		lastzero_run++;
+
+	*output++ = (blocktype == PBLOCK ? htons(PFRAME_BIT) : 0);
+	ret++;
+
+	to_encode = 8 * 8 - (lastzero_run > 14 ? lastzero_run : 0);
+
+	i = 0;
+	while (i < to_encode) {
+		int cnt = 0;
+		int tmp;
+
+		/* count leading zeros */
+		while ((tmp = block[zigzag[i]]) == 0 && cnt < 14) {
+			cnt++;
+			i++;
+			if (i == to_encode) {
+				cnt--;
+				break;
+			}
+		}
+		/* 4 bits for run, 12 for coefficient (quantization by 4) */
+		*output++ = htons((cnt | tmp << 4));
+		i++;
+		ret++;
+	}
+	if (lastzero_run > 14) {
+		*output = htons(ALL_ZEROS | 0);
+		ret++;
+	}
+
+	return ret;
+}
+
+/*
+ * This function will worst-case increase rlc_in by 65*2 bytes:
+ * one s16 value for the header and 8 * 8 coefficients of type s16.
+ */
+static s16 derlc(const __be16 **rlc_in, s16 *dwht_out)
+{
+	/* header */
+	const __be16 *input = *rlc_in;
+	s16 ret = ntohs(*input++);
+	int dec_count = 0;
+	s16 block[8 * 8 + 16];
+	s16 *wp = block;
+	int i;
+
+	/*
+	 * Now de-compress, it expands one byte to up to 15 bytes
+	 * (or fills the remainder of the 64 bytes with zeroes if it
+	 * is the last byte to expand).
+	 *
+	 * So block has to be 8 * 8 + 16 bytes, the '+ 16' is to
+	 * allow for overflow if the incoming data was malformed.
+	 */
+	while (dec_count < 8 * 8) {
+		s16 in = ntohs(*input++);
+		int length = in & 0xf;
+		int coeff = in >> 4;
+
+		/* fill remainder with zeros */
+		if (length == 15) {
+			for (i = 0; i < 64 - dec_count; i++)
+				*wp++ = 0;
+			break;
+		}
+
+		for (i = 0; i < length; i++)
+			*wp++ = 0;
+		*wp++ = coeff;
+		dec_count += length + 1;
+	}
+
+	wp = block;
+
+	for (i = 0; i < 64; i++) {
+		int pos = zigzag[i];
+		int y = pos / 8;
+		int x = pos % 8;
+
+		dwht_out[x + y * 8] = *wp++;
+	}
+	*rlc_in = input;
+	return ret;
+}
+
+static const int quant_table[] = {
+	2, 2, 2, 2, 2, 2,  2,  2,
+	2, 2, 2, 2, 2, 2,  2,  2,
+	2, 2, 2, 2, 2, 2,  2,  3,
+	2, 2, 2, 2, 2, 2,  3,  6,
+	2, 2, 2, 2, 2, 3,  6,  6,
+	2, 2, 2, 2, 3, 6,  6,  6,
+	2, 2, 2, 3, 6, 6,  6,  6,
+	2, 2, 3, 6, 6, 6,  6,  8,
+};
+
+static const int quant_table_p[] = {
+	3, 3, 3, 3, 3, 3,  3,  3,
+	3, 3, 3, 3, 3, 3,  3,  3,
+	3, 3, 3, 3, 3, 3,  3,  3,
+	3, 3, 3, 3, 3, 3,  3,  6,
+	3, 3, 3, 3, 3, 3,  6,  6,
+	3, 3, 3, 3, 3, 6,  6,  9,
+	3, 3, 3, 3, 6, 6,  9,  9,
+	3, 3, 3, 6, 6, 9,  9,  10,
+};
+
+static void quantize_intra(s16 *coeff, s16 *de_coeff)
+{
+	const int *quant = quant_table;
+	int i, j;
+
+	for (j = 0; j < 8; j++) {
+		for (i = 0; i < 8; i++, quant++, coeff++, de_coeff++) {
+			*coeff >>= *quant;
+			if (*coeff >= -DEADZONE_WIDTH &&
+			    *coeff <= DEADZONE_WIDTH)
+				*coeff = *de_coeff = 0;
+			else
+				*de_coeff = *coeff << *quant;
+		}
+	}
+}
+
+static void dequantize_intra(s16 *coeff)
+{
+	const int *quant = quant_table;
+	int i, j;
+
+	for (j = 0; j < 8; j++)
+		for (i = 0; i < 8; i++, quant++, coeff++)
+			*coeff <<= *quant;
+}
+
+static void quantize_inter(s16 *coeff, s16 *de_coeff)
+{
+	const int *quant = quant_table_p;
+	int i, j;
+
+	for (j = 0; j < 8; j++) {
+		for (i = 0; i < 8; i++, quant++, coeff++, de_coeff++) {
+			*coeff >>= *quant;
+			if (*coeff >= -DEADZONE_WIDTH &&
+			    *coeff <= DEADZONE_WIDTH)
+				*coeff = *de_coeff = 0;
+			else
+				*de_coeff = *coeff << *quant;
+		}
+	}
+}
+
+static void dequantize_inter(s16 *coeff)
+{
+	const int *quant = quant_table_p;
+	int i, j;
+
+	for (j = 0; j < 8; j++)
+		for (i = 0; i < 8; i++, quant++, coeff++)
+			*coeff <<= *quant;
+}
+
+static void fwht(const u8 *block, s16 *output_block, unsigned int stride,
+		 unsigned int input_step, bool intra)
+{
+	/* we'll need more than 8 bits for the transformed coefficients */
+	s32 workspace1[8], workspace2[8];
+	const u8 *tmp = block;
+	s16 *out = output_block;
+	int add = intra ? 256 : 0;
+	unsigned int i;
+
+	/* stage 1 */
+	stride *= input_step;
+
+	for (i = 0; i < 8; i++, tmp += stride, out += 8) {
+		if (input_step == 1) {
+			workspace1[0]  = tmp[0] + tmp[1] - add;
+			workspace1[1]  = tmp[0] - tmp[1];
+
+			workspace1[2]  = tmp[2] + tmp[3] - add;
+			workspace1[3]  = tmp[2] - tmp[3];
+
+			workspace1[4]  = tmp[4] + tmp[5] - add;
+			workspace1[5]  = tmp[4] - tmp[5];
+
+			workspace1[6]  = tmp[6] + tmp[7] - add;
+			workspace1[7]  = tmp[6] - tmp[7];
+		} else {
+			workspace1[0]  = tmp[0] + tmp[2] - add;
+			workspace1[1]  = tmp[0] - tmp[2];
+
+			workspace1[2]  = tmp[4] + tmp[6] - add;
+			workspace1[3]  = tmp[4] - tmp[6];
+
+			workspace1[4]  = tmp[8] + tmp[10] - add;
+			workspace1[5]  = tmp[8] - tmp[10];
+
+			workspace1[6]  = tmp[12] + tmp[14] - add;
+			workspace1[7]  = tmp[12] - tmp[14];
+		}
+
+		/* stage 2 */
+		workspace2[0] = workspace1[0] + workspace1[2];
+		workspace2[1] = workspace1[0] - workspace1[2];
+		workspace2[2] = workspace1[1] - workspace1[3];
+		workspace2[3] = workspace1[1] + workspace1[3];
+
+		workspace2[4] = workspace1[4] + workspace1[6];
+		workspace2[5] = workspace1[4] - workspace1[6];
+		workspace2[6] = workspace1[5] - workspace1[7];
+		workspace2[7] = workspace1[5] + workspace1[7];
+
+		/* stage 3 */
+		out[0] = workspace2[0] + workspace2[4];
+		out[1] = workspace2[0] - workspace2[4];
+		out[2] = workspace2[1] - workspace2[5];
+		out[3] = workspace2[1] + workspace2[5];
+		out[4] = workspace2[2] + workspace2[6];
+		out[5] = workspace2[2] - workspace2[6];
+		out[6] = workspace2[3] - workspace2[7];
+		out[7] = workspace2[3] + workspace2[7];
+	}
+
+	out = output_block;
+
+	for (i = 0; i < 8; i++, out++) {
+		/* stage 1 */
+		workspace1[0]  = out[0] + out[1 * 8];
+		workspace1[1]  = out[0] - out[1 * 8];
+
+		workspace1[2]  = out[2 * 8] + out[3 * 8];
+		workspace1[3]  = out[2 * 8] - out[3 * 8];
+
+		workspace1[4]  = out[4 * 8] + out[5 * 8];
+		workspace1[5]  = out[4 * 8] - out[5 * 8];
+
+		workspace1[6]  = out[6 * 8] + out[7 * 8];
+		workspace1[7]  = out[6 * 8] - out[7 * 8];
+
+		/* stage 2 */
+		workspace2[0] = workspace1[0] + workspace1[2];
+		workspace2[1] = workspace1[0] - workspace1[2];
+		workspace2[2] = workspace1[1] - workspace1[3];
+		workspace2[3] = workspace1[1] + workspace1[3];
+
+		workspace2[4] = workspace1[4] + workspace1[6];
+		workspace2[5] = workspace1[4] - workspace1[6];
+		workspace2[6] = workspace1[5] - workspace1[7];
+		workspace2[7] = workspace1[5] + workspace1[7];
+		/* stage 3 */
+		out[0 * 8] = workspace2[0] + workspace2[4];
+		out[1 * 8] = workspace2[0] - workspace2[4];
+		out[2 * 8] = workspace2[1] - workspace2[5];
+		out[3 * 8] = workspace2[1] + workspace2[5];
+		out[4 * 8] = workspace2[2] + workspace2[6];
+		out[5 * 8] = workspace2[2] - workspace2[6];
+		out[6 * 8] = workspace2[3] - workspace2[7];
+		out[7 * 8] = workspace2[3] + workspace2[7];
+	}
+}
+
+/*
+ * Not the nicest way of doing it, but P-blocks get twice the range of
+ * that of the I-blocks. Therefore we need a type bigger than 8 bits.
+ * Furthermore values can be negative... This is just a version that
+ * works with 16 signed data
+ */
+static void fwht16(const s16 *block, s16 *output_block, int stride, int intra)
+{
+	/* we'll need more than 8 bits for the transformed coefficients */
+	s32 workspace1[8], workspace2[8];
+	const s16 *tmp = block;
+	s16 *out = output_block;
+	int i;
+
+	for (i = 0; i < 8; i++, tmp += stride, out += 8) {
+		/* stage 1 */
+		workspace1[0]  = tmp[0] + tmp[1];
+		workspace1[1]  = tmp[0] - tmp[1];
+
+		workspace1[2]  = tmp[2] + tmp[3];
+		workspace1[3]  = tmp[2] - tmp[3];
+
+		workspace1[4]  = tmp[4] + tmp[5];
+		workspace1[5]  = tmp[4] - tmp[5];
+
+		workspace1[6]  = tmp[6] + tmp[7];
+		workspace1[7]  = tmp[6] - tmp[7];
+
+		/* stage 2 */
+		workspace2[0] = workspace1[0] + workspace1[2];
+		workspace2[1] = workspace1[0] - workspace1[2];
+		workspace2[2] = workspace1[1] - workspace1[3];
+		workspace2[3] = workspace1[1] + workspace1[3];
+
+		workspace2[4] = workspace1[4] + workspace1[6];
+		workspace2[5] = workspace1[4] - workspace1[6];
+		workspace2[6] = workspace1[5] - workspace1[7];
+		workspace2[7] = workspace1[5] + workspace1[7];
+
+		/* stage 3 */
+		out[0] = workspace2[0] + workspace2[4];
+		out[1] = workspace2[0] - workspace2[4];
+		out[2] = workspace2[1] - workspace2[5];
+		out[3] = workspace2[1] + workspace2[5];
+		out[4] = workspace2[2] + workspace2[6];
+		out[5] = workspace2[2] - workspace2[6];
+		out[6] = workspace2[3] - workspace2[7];
+		out[7] = workspace2[3] + workspace2[7];
+	}
+
+	out = output_block;
+
+	for (i = 0; i < 8; i++, out++) {
+		/* stage 1 */
+		workspace1[0]  = out[0] + out[1*8];
+		workspace1[1]  = out[0] - out[1*8];
+
+		workspace1[2]  = out[2*8] + out[3*8];
+		workspace1[3]  = out[2*8] - out[3*8];
+
+		workspace1[4]  = out[4*8] + out[5*8];
+		workspace1[5]  = out[4*8] - out[5*8];
+
+		workspace1[6]  = out[6*8] + out[7*8];
+		workspace1[7]  = out[6*8] - out[7*8];
+
+		/* stage 2 */
+		workspace2[0] = workspace1[0] + workspace1[2];
+		workspace2[1] = workspace1[0] - workspace1[2];
+		workspace2[2] = workspace1[1] - workspace1[3];
+		workspace2[3] = workspace1[1] + workspace1[3];
+
+		workspace2[4] = workspace1[4] + workspace1[6];
+		workspace2[5] = workspace1[4] - workspace1[6];
+		workspace2[6] = workspace1[5] - workspace1[7];
+		workspace2[7] = workspace1[5] + workspace1[7];
+
+		/* stage 3 */
+		out[0*8] = workspace2[0] + workspace2[4];
+		out[1*8] = workspace2[0] - workspace2[4];
+		out[2*8] = workspace2[1] - workspace2[5];
+		out[3*8] = workspace2[1] + workspace2[5];
+		out[4*8] = workspace2[2] + workspace2[6];
+		out[5*8] = workspace2[2] - workspace2[6];
+		out[6*8] = workspace2[3] - workspace2[7];
+		out[7*8] = workspace2[3] + workspace2[7];
+	}
+}
+
+static void ifwht(const s16 *block, s16 *output_block, int intra)
+{
+	/*
+	 * we'll need more than 8 bits for the transformed coefficients
+	 * use native unit of cpu
+	 */
+	int workspace1[8], workspace2[8];
+	int inter = intra ? 0 : 1;
+	const s16 *tmp = block;
+	s16 *out = output_block;
+	int i;
+
+	for (i = 0; i < 8; i++, tmp += 8, out += 8) {
+		/* stage 1 */
+		workspace1[0]  = tmp[0] + tmp[1];
+		workspace1[1]  = tmp[0] - tmp[1];
+
+		workspace1[2]  = tmp[2] + tmp[3];
+		workspace1[3]  = tmp[2] - tmp[3];
+
+		workspace1[4]  = tmp[4] + tmp[5];
+		workspace1[5]  = tmp[4] - tmp[5];
+
+		workspace1[6]  = tmp[6] + tmp[7];
+		workspace1[7]  = tmp[6] - tmp[7];
+
+		/* stage 2 */
+		workspace2[0] = workspace1[0] + workspace1[2];
+		workspace2[1] = workspace1[0] - workspace1[2];
+		workspace2[2] = workspace1[1] - workspace1[3];
+		workspace2[3] = workspace1[1] + workspace1[3];
+
+		workspace2[4] = workspace1[4] + workspace1[6];
+		workspace2[5] = workspace1[4] - workspace1[6];
+		workspace2[6] = workspace1[5] - workspace1[7];
+		workspace2[7] = workspace1[5] + workspace1[7];
+
+		/* stage 3 */
+		out[0] = workspace2[0] + workspace2[4];
+		out[1] = workspace2[0] - workspace2[4];
+		out[2] = workspace2[1] - workspace2[5];
+		out[3] = workspace2[1] + workspace2[5];
+		out[4] = workspace2[2] + workspace2[6];
+		out[5] = workspace2[2] - workspace2[6];
+		out[6] = workspace2[3] - workspace2[7];
+		out[7] = workspace2[3] + workspace2[7];
+	}
+
+	out = output_block;
+
+	for (i = 0; i < 8; i++, out++) {
+		/* stage 1 */
+		workspace1[0]  = out[0] + out[1 * 8];
+		workspace1[1]  = out[0] - out[1 * 8];
+
+		workspace1[2]  = out[2 * 8] + out[3 * 8];
+		workspace1[3]  = out[2 * 8] - out[3 * 8];
+
+		workspace1[4]  = out[4 * 8] + out[5 * 8];
+		workspace1[5]  = out[4 * 8] - out[5 * 8];
+
+		workspace1[6]  = out[6 * 8] + out[7 * 8];
+		workspace1[7]  = out[6 * 8] - out[7 * 8];
+
+		/* stage 2 */
+		workspace2[0] = workspace1[0] + workspace1[2];
+		workspace2[1] = workspace1[0] - workspace1[2];
+		workspace2[2] = workspace1[1] - workspace1[3];
+		workspace2[3] = workspace1[1] + workspace1[3];
+
+		workspace2[4] = workspace1[4] + workspace1[6];
+		workspace2[5] = workspace1[4] - workspace1[6];
+		workspace2[6] = workspace1[5] - workspace1[7];
+		workspace2[7] = workspace1[5] + workspace1[7];
+
+		/* stage 3 */
+		if (inter) {
+			int d;
+
+			out[0 * 8] = workspace2[0] + workspace2[4];
+			out[1 * 8] = workspace2[0] - workspace2[4];
+			out[2 * 8] = workspace2[1] - workspace2[5];
+			out[3 * 8] = workspace2[1] + workspace2[5];
+			out[4 * 8] = workspace2[2] + workspace2[6];
+			out[5 * 8] = workspace2[2] - workspace2[6];
+			out[6 * 8] = workspace2[3] - workspace2[7];
+			out[7 * 8] = workspace2[3] + workspace2[7];
+
+			for (d = 0; d < 8; d++)
+				out[8 * d] >>= 6;
+		} else {
+			int d;
+
+			out[0 * 8] = workspace2[0] + workspace2[4];
+			out[1 * 8] = workspace2[0] - workspace2[4];
+			out[2 * 8] = workspace2[1] - workspace2[5];
+			out[3 * 8] = workspace2[1] + workspace2[5];
+			out[4 * 8] = workspace2[2] + workspace2[6];
+			out[5 * 8] = workspace2[2] - workspace2[6];
+			out[6 * 8] = workspace2[3] - workspace2[7];
+			out[7 * 8] = workspace2[3] + workspace2[7];
+
+			for (d = 0; d < 8; d++) {
+				out[8 * d] >>= 6;
+				out[8 * d] += 128;
+			}
+		}
+	}
+}
+
+static void fill_encoder_block(const u8 *input, s16 *dst,
+			       unsigned int stride, unsigned int input_step)
+{
+	int i, j;
+
+	for (i = 0; i < 8; i++) {
+		for (j = 0; j < 8; j++, input += input_step)
+			*dst++ = *input;
+		input += (stride - 8) * input_step;
+	}
+}
+
+static int var_intra(const s16 *input)
+{
+	int32_t mean = 0;
+	int32_t ret = 0;
+	const s16 *tmp = input;
+	int i;
+
+	for (i = 0; i < 8 * 8; i++, tmp++)
+		mean += *tmp;
+	mean /= 64;
+	tmp = input;
+	for (i = 0; i < 8 * 8; i++, tmp++)
+		ret += (*tmp - mean) < 0 ? -(*tmp - mean) : (*tmp - mean);
+	return ret;
+}
+
+static int var_inter(const s16 *old, const s16 *new)
+{
+	int32_t ret = 0;
+	int i;
+
+	for (i = 0; i < 8 * 8; i++, old++, new++)
+		ret += (*old - *new) < 0 ? -(*old - *new) : (*old - *new);
+	return ret;
+}
+
+static int decide_blocktype(const u8 *current, const u8 *reference,
+			    s16 *deltablock, unsigned int stride,
+			    unsigned int input_step)
+{
+	s16 tmp[64];
+	s16 old[64];
+	s16 *work = tmp;
+	unsigned int k, l;
+	int vari;
+	int vard;
+
+	fill_encoder_block(current, tmp, stride, input_step);
+	fill_encoder_block(reference, old, 8, 1);
+	vari = var_intra(tmp);
+
+	for (k = 0; k < 8; k++) {
+		for (l = 0; l < 8; l++) {
+			*deltablock = *work - *reference;
+			deltablock++;
+			work++;
+			reference++;
+		}
+	}
+	deltablock -= 64;
+	vard = var_inter(old, tmp);
+	return vari <= vard ? IBLOCK : PBLOCK;
+}
+
+static void fill_decoder_block(u8 *dst, const s16 *input, int stride)
+{
+	int i, j;
+
+	for (i = 0; i < 8; i++) {
+		for (j = 0; j < 8; j++)
+			*dst++ = *input++;
+		dst += stride - 8;
+	}
+}
+
+static void add_deltas(s16 *deltas, const u8 *ref, int stride)
+{
+	int k, l;
+
+	for (k = 0; k < 8; k++) {
+		for (l = 0; l < 8; l++) {
+			*deltas += *ref++;
+			/*
+			 * Due to quantizing, it might possible that the
+			 * decoded coefficients are slightly out of range
+			 */
+			if (*deltas < 0)
+				*deltas = 0;
+			else if (*deltas > 255)
+				*deltas = 255;
+			deltas++;
+		}
+		ref += stride - 8;
+	}
+}
+
+static u32 encode_plane(u8 *input, u8 *refp, __be16 **rlco, __be16 *rlco_max,
+			struct cframe *cf, u32 height, u32 width,
+			unsigned int input_step,
+			bool is_intra, bool next_is_intra)
+{
+	u8 *input_start = input;
+	__be16 *rlco_start = *rlco;
+	s16 deltablock[64];
+	__be16 pframe_bit = htons(PFRAME_BIT);
+	u32 encoding = 0;
+	unsigned int last_size = 0;
+	unsigned int i, j;
+
+	for (j = 0; j < height / 8; j++) {
+		for (i = 0; i < width / 8; i++) {
+			/* intra code, first frame is always intra coded. */
+			int blocktype = IBLOCK;
+			unsigned int size;
+
+			if (!is_intra)
+				blocktype = decide_blocktype(input, refp,
+					deltablock, width, input_step);
+			if (is_intra || blocktype == IBLOCK) {
+				fwht(input, cf->coeffs, width, input_step, 1);
+				quantize_intra(cf->coeffs, cf->de_coeffs);
+				blocktype = IBLOCK;
+			} else {
+				/* inter code */
+				encoding |= FRAME_PCODED;
+				fwht16(deltablock, cf->coeffs, 8, 0);
+				quantize_inter(cf->coeffs, cf->de_coeffs);
+			}
+			if (!next_is_intra) {
+				ifwht(cf->de_coeffs, cf->de_fwht, blocktype);
+
+				if (blocktype == PBLOCK)
+					add_deltas(cf->de_fwht, refp, 8);
+				fill_decoder_block(refp, cf->de_fwht, 8);
+			}
+
+			input += 8 * input_step;
+			refp += 8 * 8;
+
+			if (encoding & FRAME_UNENCODED)
+				continue;
+
+			size = rlc(cf->coeffs, *rlco, blocktype);
+			if (last_size == size &&
+			    !memcmp(*rlco + 1, *rlco - size + 1, 2 * size - 2)) {
+				__be16 *last_rlco = *rlco - size;
+				s16 hdr = ntohs(*last_rlco);
+
+				if (!((*last_rlco ^ **rlco) & pframe_bit) &&
+				    (hdr & DUPS_MASK) < DUPS_MASK)
+					*last_rlco = htons(hdr + 2);
+				else
+					*rlco += size;
+			} else {
+				*rlco += size;
+			}
+			if (*rlco >= rlco_max)
+				encoding |= FRAME_UNENCODED;
+			last_size = size;
+		}
+		input += width * 7 * input_step;
+	}
+	if (encoding & FRAME_UNENCODED) {
+		u8 *out = (u8 *)rlco_start;
+
+		input = input_start;
+		/*
+		 * The compressed stream should never contain the magic
+		 * header, so when we copy the YUV data we replace 0xff
+		 * by 0xfe. Since YUV is limited range such values
+		 * shouldn't appear anyway.
+		 */
+		for (i = 0; i < height * width; i++, input += input_step)
+			*out++ = (*input == 0xff) ? 0xfe : *input;
+		*rlco = (__be16 *)out;
+	}
+	return encoding;
+}
+
+u32 encode_frame(struct raw_frame *frm, struct raw_frame *ref_frm,
+		 struct cframe *cf, bool is_intra, bool next_is_intra)
+{
+	unsigned int size = frm->height * frm->width;
+	__be16 *rlco = cf->rlc_data;
+	__be16 *rlco_max;
+	u32 encoding;
+
+	rlco_max = rlco + size / 2 - 256;
+	encoding = encode_plane(frm->luma, ref_frm->luma, &rlco, rlco_max, cf,
+				  frm->height, frm->width,
+				  1, is_intra, next_is_intra);
+	if (encoding & FRAME_UNENCODED)
+		encoding |= LUMA_UNENCODED;
+	encoding &= ~FRAME_UNENCODED;
+	rlco_max = rlco + size / 8 - 256;
+	encoding |= encode_plane(frm->cb, ref_frm->cb, &rlco, rlco_max, cf,
+				   frm->height / 2, frm->width / 2,
+				   frm->chroma_step, is_intra, next_is_intra);
+	if (encoding & FRAME_UNENCODED)
+		encoding |= CB_UNENCODED;
+	encoding &= ~FRAME_UNENCODED;
+	rlco_max = rlco + size / 8 - 256;
+	encoding |= encode_plane(frm->cr, ref_frm->cr, &rlco, rlco_max, cf,
+				   frm->height / 2, frm->width / 2,
+				   frm->chroma_step, is_intra, next_is_intra);
+	if (encoding & FRAME_UNENCODED)
+		encoding |= CR_UNENCODED;
+	encoding &= ~FRAME_UNENCODED;
+	cf->size = (rlco - cf->rlc_data) * sizeof(*rlco);
+	return encoding;
+}
+
+static void decode_plane(struct cframe *cf, const __be16 **rlco, u8 *ref,
+			 u32 height, u32 width, bool uncompressed)
+{
+	unsigned int copies = 0;
+	s16 copy[8 * 8];
+	s16 stat;
+	unsigned int i, j;
+
+	if (uncompressed) {
+		memcpy(ref, *rlco, width * height);
+		*rlco += width * height / 2;
+		return;
+	}
+
+	/*
+	 * When decoding each macroblock the rlco pointer will be increased
+	 * by 65 * 2 bytes worst-case.
+	 * To avoid overflow the buffer has to be 65/64th of the actual raw
+	 * image size, just in case someone feeds it malicious data.
+	 */
+	for (j = 0; j < height / 8; j++) {
+		for (i = 0; i < width / 8; i++) {
+			u8 *refp = ref + j * 8 * width + i * 8;
+
+			if (copies) {
+				memcpy(cf->de_fwht, copy, sizeof(copy));
+				if (stat & PFRAME_BIT)
+					add_deltas(cf->de_fwht, refp, width);
+				fill_decoder_block(refp, cf->de_fwht, width);
+				copies--;
+				continue;
+			}
+
+			stat = derlc(rlco, cf->coeffs);
+
+			if (stat & PFRAME_BIT)
+				dequantize_inter(cf->coeffs);
+			else
+				dequantize_intra(cf->coeffs);
+
+			ifwht(cf->coeffs, cf->de_fwht,
+			      (stat & PFRAME_BIT) ? 0 : 1);
+
+			copies = (stat & DUPS_MASK) >> 1;
+			if (copies)
+				memcpy(copy, cf->de_fwht, sizeof(copy));
+			if (stat & PFRAME_BIT)
+				add_deltas(cf->de_fwht, refp, width);
+			fill_decoder_block(refp, cf->de_fwht, width);
+		}
+	}
+}
+
+void decode_frame(struct cframe *cf, struct raw_frame *ref, u32 hdr_flags)
+{
+	const __be16 *rlco = cf->rlc_data;
+
+	decode_plane(cf, &rlco, ref->luma, cf->height, cf->width,
+		     hdr_flags & VICODEC_FL_LUMA_IS_UNCOMPRESSED);
+	decode_plane(cf, &rlco, ref->cb, cf->height / 2, cf->width / 2,
+		     hdr_flags & VICODEC_FL_CB_IS_UNCOMPRESSED);
+	decode_plane(cf, &rlco, ref->cr, cf->height / 2, cf->width / 2,
+		     hdr_flags & VICODEC_FL_CR_IS_UNCOMPRESSED);
+}

drivers/media/platform/vicodec/vicodec-codec.h

@@ -0,0 +1,129 @@
+/* SPDX-License-Identifier: GPL-2.0+ */
+/*
+ * Copyright 2016 Tom aan de Wiel
+ * Copyright 2018 Cisco Systems, Inc. and/or its affiliates. All rights reserved.
+ */
+
+#ifndef VICODEC_RLC_H
+#define VICODEC_RLC_H
+
+#include <linux/types.h>
+#include <linux/bitops.h>
+#include <asm/byteorder.h>
+
+/*
+ * The compressed format consists of a cframe_hdr struct followed by the
+ * compressed frame data. The header contains the size of that data.
+ * Each Y, Cb and Cr plane is compressed separately. If the compressed
+ * size of each plane becomes larger than the uncompressed size, then
+ * that plane is stored uncompressed and the corresponding bit is set
+ * in the flags field of the header.
+ *
+ * Each compressed plane consists of macroblocks and each macroblock
+ * is run-length-encoded. Each macroblock starts with a 16 bit value.
+ * Bit 15 indicates if this is a P-coded macroblock (1) or not (0).
+ * P-coded macroblocks contain a delta against the previous frame.
+ *
+ * Bits 1-12 contain a number. If non-zero, then this same macroblock
+ * repeats that number of times. This results in a high degree of
+ * compression for generated images like colorbars.
+ *
+ * Following this macroblock header the MB coefficients are run-length
+ * encoded: the top 12 bits contain the coefficient, the bottom 4 bits
+ * tell how many times this coefficient occurs. The value 0xf indicates
+ * that the remainder of the macroblock should be filled with zeroes.
+ *
+ * All 16 and 32 bit values are stored in big-endian (network) order.
+ *
+ * Each cframe_hdr starts with an 8 byte magic header that is
+ * guaranteed not to occur in the compressed frame data. This header
+ * can be used to sync to the next frame.
+ *
+ * This codec uses the Fast Walsh Hadamard Transform. Tom aan de Wiel
+ * developed this as part of a university project, specifically for use
+ * with this driver. His project report can be found here:
+ *
+ * https://hverkuil.home.xs4all.nl/fwht.pdf
+ */
+
+/*
+ * Note: bit 0 of the header must always be 0. Otherwise it cannot
+ * be guaranteed that the magic 8 byte sequence (see below) can
+ * never occur in the rlc output.
+ */
+#define PFRAME_BIT (1 << 15)
+#define DUPS_MASK 0x1ffe
+
+/*
+ * This is a sequence of 8 bytes with the low 4 bits set to 0xf.
+ *
+ * This sequence cannot occur in the encoded data
+ */
+#define VICODEC_MAGIC1 0x4f4f4f4f
+#define VICODEC_MAGIC2 0xffffffff
+
+#define VICODEC_VERSION 1
+
+#define VICODEC_MAX_WIDTH 3840
+#define VICODEC_MAX_HEIGHT 2160
+#define VICODEC_MIN_WIDTH 640
+#define VICODEC_MIN_HEIGHT 480
+
+#define PBLOCK 0
+#define IBLOCK 1
+
+/* Set if this is an interlaced format */
+#define VICODEC_FL_IS_INTERLACED	BIT(0)
+/* Set if this is a bottom-first (NTSC) interlaced format */
+#define VICODEC_FL_IS_BOTTOM_FIRST	BIT(1)
+/* Set if each 'frame' contains just one field */
+#define VICODEC_FL_IS_ALTERNATE		BIT(2)
+/*
+ * If VICODEC_FL_IS_ALTERNATE was set, then this is set if this
+ * 'frame' is the bottom field, else it is the top field.
+ */
+#define VICODEC_FL_IS_BOTTOM_FIELD	BIT(3)
+/* Set if this frame is uncompressed */
+#define VICODEC_FL_LUMA_IS_UNCOMPRESSED	BIT(4)
+#define VICODEC_FL_CB_IS_UNCOMPRESSED	BIT(5)
+#define VICODEC_FL_CR_IS_UNCOMPRESSED	BIT(6)
+
+struct cframe_hdr {
+	u32 magic1;
+	u32 magic2;
+	__be32 version;
+	__be32 width, height;
+	__be32 flags;
+	__be32 colorspace;
+	__be32 xfer_func;
+	__be32 ycbcr_enc;
+	__be32 quantization;
+	__be32 size;
+};
+
+struct cframe {
+	unsigned int width, height;
+	__be16 *rlc_data;
+	s16 coeffs[8 * 8];
+	s16 de_coeffs[8 * 8];
+	s16 de_fwht[8 * 8];
+	u32 size;
+};
+
+struct raw_frame {
+	unsigned int width, height;
+	unsigned int chroma_step;
+	u8 *luma, *cb, *cr;
+};
+
+#define FRAME_PCODED	BIT(0)
+#define FRAME_UNENCODED	BIT(1)
+#define LUMA_UNENCODED	BIT(2)
+#define CB_UNENCODED	BIT(3)
+#define CR_UNENCODED	BIT(4)
+
+u32 encode_frame(struct raw_frame *frm, struct raw_frame *ref_frm,
+		 struct cframe *cf, bool is_intra, bool next_is_intra);
+void decode_frame(struct cframe *cf, struct raw_frame *ref, u32 hdr_flags);
+
+#endif