#version 440
#extension GL_ARB_shader_clock : enable

#define PARALLEL_SLICES 1

#define ENABLE_TIMING 0

layout(local_size_x = 64*PARALLEL_SLICES) in;
layout(r8ui) uniform restrict readonly uimage2D cum2sym_tex;
layout(rg16ui) uniform restrict readonly uimage2D dsyms_tex;
layout(r8) uniform restrict writeonly image2D out_tex;
layout(r16i) uniform restrict writeonly iimage2D coeff_tex;
uniform int num_blocks;

const uint prob_bits = 12;
const uint prob_scale = 1 << prob_bits;
const uint NUM_SYMS = 256;
const uint ESCAPE_LIMIT = NUM_SYMS - 1;
const uint BLOCKS_PER_STREAM = 320;

// These need to be folded into quant_matrix.
const float dc_scalefac = 8.0;
const float quant_scalefac = 4.0;

const float quant_matrix[64] = {
         8, 16, 19, 22, 26, 27, 29, 34,
        16, 16, 22, 24, 27, 29, 34, 37,
        19, 22, 26, 27, 29, 34, 34, 38,
        22, 22, 26, 27, 29, 34, 37, 40,
        22, 26, 27, 29, 32, 35, 40, 48,
        26, 27, 29, 32, 35, 40, 48, 58,
        26, 27, 29, 34, 38, 46, 56, 69,
        27, 29, 35, 38, 46, 56, 69, 83
};
const uint ff_zigzag_direct[64] = {
    0,   1,  8, 16,  9,  2,  3, 10,
    17, 24, 32, 25, 18, 11,  4,  5,
    12, 19, 26, 33, 40, 48, 41, 34,
    27, 20, 13,  6,  7, 14, 21, 28,
    35, 42, 49, 56, 57, 50, 43, 36,
    29, 22, 15, 23, 30, 37, 44, 51,
    58, 59, 52, 45, 38, 31, 39, 46,
    53, 60, 61, 54, 47, 55, 62, 63
};
const uint stream_mapping[64] = {
	0, 0, 1, 1, 2, 2, 3, 3,
	0, 0, 1, 2, 2, 2, 3, 3,
	1, 1, 2, 2, 2, 3, 3, 3,
	1, 1, 2, 2, 2, 3, 3, 3,
	1, 2, 2, 2, 2, 3, 3, 3,
	2, 2, 2, 2, 3, 3, 3, 3,
	2, 2, 3, 3, 3, 3, 3, 3,
	3, 3, 3, 3, 3, 3, 3, 3,
};

layout(std430, binding = 9) buffer layoutName
{
	uint data_SSBO[];
};
layout(std430, binding = 10) buffer layoutName2
{
	uvec2 timing[10 * 64];
};

struct CoeffStream {
	uint src_offset, src_len;
};
layout(std430, binding = 0) buffer whatever3
{
	CoeffStream streams[];
};
uniform uint sign_bias_per_model[16];

const uint RANS_BYTE_L = (1u << 23);  // lower bound of our normalization interval

uint get_rans_byte(uint offset)
{
	// We assume little endian.
	return bitfieldExtract(data_SSBO[offset >> 2], 8 * int(offset & 3u), 8);
}

uint RansDecInit(inout uint offset)
{
	uint x;

	x  = get_rans_byte(offset);
	x |= get_rans_byte(offset + 1) << 8;
	x |= get_rans_byte(offset + 2) << 16;
	x |= get_rans_byte(offset + 3) << 24;
	offset += 4;

	return x;
}

uint RansDecGet(uint r, uint scale_bits)
{
	return r & ((1u << scale_bits) - 1);
}

void RansDecAdvance(inout uint rans, inout uint offset, const uint start, const uint freq, uint prob_bits)
{
	const uint mask = (1u << prob_bits) - 1;
	rans = freq * (rans >> prob_bits) + (rans & mask) - start;
	
	// renormalize
	while (rans < RANS_BYTE_L) {
		rans = (rans << 8) | get_rans_byte(offset++);
	}
}

uint cum2sym(uint bits, uint table)
{
	return imageLoad(cum2sym_tex, ivec2(bits, table)).x;
}

uvec2 get_dsym(uint k, uint table)
{
	return imageLoad(dsyms_tex, ivec2(k, table)).xy;
}

void idct_1d(inout float y0, inout float y1, inout float y2, inout float y3, inout float y4, inout float y5, inout float y6, inout float y7)
{
	const float a1 = 0.7071067811865474;   // sqrt(2)
	const float a2 = 0.5411961001461971;   // cos(3/8 pi) * sqrt(2)
	const float a4 = 1.3065629648763766;   // cos(pi/8) * sqrt(2)
	// static const float a5 = 0.5 * (a4 - a2);
	const float a5 = 0.3826834323650897;

	// phase 2 (phase 1 is just moving around)
	const float p2_4 = y5 - y3;
	const float p2_5 = y1 + y7;
	const float p2_6 = y1 - y7;
	const float p2_7 = y5 + y3;

	// phase 3
	const float p3_2 = y2 - y6;
	const float p3_3 = y2 + y6;
	const float p3_5 = p2_5 - p2_7;
	const float p3_7 = p2_5 + p2_7;

	// phase 4
	const float p4_2 = a1 * p3_2;
	const float p4_4 = p2_4 * a2 + (p2_4 + p2_6) * a5;  // Inverted.
	const float p4_5 = a1 * p3_5;
	const float p4_6 = p2_6 * a4 - (p2_4 + p2_6) * a5;

	// phase 5
	const float p5_0 = y0 + y4;
	const float p5_1 = y0 - y4;
	const float p5_3 = p4_2 + p3_3;

	// phase 6
	const float p6_0 = p5_0 + p5_3;
	const float p6_1 = p5_1 + p4_2;
	const float p6_2 = p5_1 - p4_2;
	const float p6_3 = p5_0 - p5_3;
	const float p6_5 = p4_5 + p4_4;
	const float p6_6 = p4_5 + p4_6;
	const float p6_7 = p4_6 + p3_7;

	// phase 7
	y0 = p6_0 + p6_7;
	y1 = p6_1 + p6_6;
	y2 = p6_2 + p6_5;
	y3 = p6_3 + p4_4;
	y4 = p6_3 - p4_4;
	y5 = p6_2 - p6_5;
	y6 = p6_1 - p6_6;
	y7 = p6_0 - p6_7;
}

shared float temp[64 * 8 * PARALLEL_SLICES];

void pick_timer(inout uvec2 start, inout uvec2 t)
{
#if ENABLE_TIMING
	uvec2 now = clock2x32ARB();

	uvec2 delta = now - start;
	if (now.x < start.x) {
		--delta.y;
	}

	uvec2 new_t = t + delta;
	if (new_t.x < t.x) {
		++new_t.y;
	}
	t = new_t;

	start = clock2x32ARB();
#endif
}

void main()
{
	uvec2 local_timing[10];
#if ENABLE_TIMING
	for (int timer_idx = 0; timer_idx < 10; ++timer_idx) {
		local_timing[timer_idx] = uvec2(0, 0);
	}
	uvec2 start = clock2x32ARB();
#else
	uvec2 start = uvec2(0, 0);
	local_timing[0] = start;
#endif

	const uint blocks_per_row = (imageSize(out_tex).x + 7) / 8;

	const uint local_x = gl_LocalInvocationID.x % 8;
	const uint local_y = (gl_LocalInvocationID.x / 8) % 8;
	const uint local_z = gl_LocalInvocationID.x / 64;

	const uint slice_num = local_z;
	const uint thread_num = local_y * 8 + local_x;

	const uint block_row = gl_WorkGroupID.y * PARALLEL_SLICES + slice_num;
	//const uint coeff_num = ff_zigzag_direct[thread_num];
	const uint coeff_num = thread_num;
	const uint stream_num = coeff_num * num_blocks + block_row;
	const uint model_num = stream_mapping[coeff_num];
	const uint sign_bias = sign_bias_per_model[model_num];

	// Initialize rANS decoder.
	uint offset = streams[stream_num].src_offset;
	uint rans = RansDecInit(offset);

	float q = (coeff_num == 0) ? 1.0 : (quant_matrix[coeff_num] * quant_scalefac / 128.0 / sqrt(2.0));  // FIXME: fold
	q *= (1.0 / 255.0);
	//int w = (coeff_num == 0) ? 32 : int(quant_matrix[coeff_num]);
	int last_k = 128;

	pick_timer(start, local_timing[0]);

	for (uint block_idx = BLOCKS_PER_STREAM / 8; block_idx --> 0; ) {
		pick_timer(start, local_timing[1]);

		// rANS decode one coefficient across eight blocks (so 64x8 coefficients).
		for (uint subblock_idx = 8; subblock_idx --> 0; ) {
			// Read a symbol.
			uint bottom_bits = RansDecGet(rans, prob_bits + 1);
			bool sign = false;
			if (bottom_bits >= sign_bias) {
				bottom_bits -= sign_bias;
				rans -= sign_bias;
				sign = true;
			}
			int k = int(cum2sym(bottom_bits, model_num));  // Can go out-of-bounds; that will return zero.
			uvec2 sym = get_dsym(k, model_num);
			RansDecAdvance(rans, offset, sym.x, sym.y, prob_bits + 1);

			if (k == ESCAPE_LIMIT) {
				k = int(RansDecGet(rans, prob_bits));
				RansDecAdvance(rans, offset, k, 1, prob_bits);
			}
			if (sign) {
				k = -k;
			}

			if (coeff_num == 0) {
				k += last_k;
				last_k = k;
			}

#if 0
			uint y = block_row * 16 + block_y * 8 + local_y;
			uint x = block_x * 64 + subblock_idx * 8 + local_x;
			imageStore(coeff_tex, ivec2(x, y), ivec4(k, 0,0,0));
#endif

			temp[slice_num * 64 * 8 + subblock_idx * 64 + coeff_num] = k * q;
			//temp[subblock_idx * 64 + 8 * y + x] = (2 * k * w * 4) / 32;  // 100% matching unquant
		}

		pick_timer(start, local_timing[2]);

		memoryBarrierShared();
		barrier();

		pick_timer(start, local_timing[3]);

		// Horizontal DCT one row (so 64 rows).
		idct_1d(temp[slice_num * 64 * 8 + thread_num * 8 + 0],
		        temp[slice_num * 64 * 8 + thread_num * 8 + 1],
		        temp[slice_num * 64 * 8 + thread_num * 8 + 2],
		        temp[slice_num * 64 * 8 + thread_num * 8 + 3],
		        temp[slice_num * 64 * 8 + thread_num * 8 + 4],
		        temp[slice_num * 64 * 8 + thread_num * 8 + 5],
		        temp[slice_num * 64 * 8 + thread_num * 8 + 6],
		        temp[slice_num * 64 * 8 + thread_num * 8 + 7]);

		pick_timer(start, local_timing[4]);

		memoryBarrierShared();
		barrier();

		pick_timer(start, local_timing[5]);

		// Vertical DCT one row (so 64 columns).
		uint row_offset = local_z * 64 * 8 + local_y * 64 + local_x;
		idct_1d(temp[row_offset + 0 * 8],
		        temp[row_offset + 1 * 8],
		        temp[row_offset + 2 * 8],
		        temp[row_offset + 3 * 8],
		        temp[row_offset + 4 * 8],
		        temp[row_offset + 5 * 8],
		        temp[row_offset + 6 * 8],
		        temp[row_offset + 7 * 8]);

		pick_timer(start, local_timing[6]);

		uint global_block_idx = (block_row * 40 + block_idx) * 8 + local_y;
		uint block_x = global_block_idx % blocks_per_row;
		uint block_y = global_block_idx / blocks_per_row;

		uint y = block_y * 8;
		uint x = block_x * 8 + local_x;
		for (uint yl = 0; yl < 8; ++yl) {
			imageStore(out_tex, ivec2(x, yl + y), vec4(temp[row_offset + yl * 8], 0.0, 0.0, 1.0));
		}

		pick_timer(start, local_timing[7]);

		memoryBarrierShared();  // is this needed?
		barrier();

		pick_timer(start, local_timing[8]);
		pick_timer(start, local_timing[9]);  // should be nearly nothing
	}

#if ENABLE_TIMING
	for (int timer_idx = 0; timer_idx < 10; ++timer_idx) {
		uint global_idx = thread_num * 10 + timer_idx;

		uint old_val = atomicAdd(timing[global_idx].x, local_timing[timer_idx].x);
		if (old_val + local_timing[timer_idx].x < old_val) {
			++local_timing[timer_idx].y;
		}
		atomicAdd(timing[global_idx].y, local_timing[timer_idx].y);
	}
#endif
}