Function encode_bc4

external/basisu/transcoder/basisu_transcoder.cpp:14592–14710

	const uint8_t* s_uastc_to_bc1_weights[6] = { nullptr, s_uastc1_to_bc1, s_uastc2_to_bc1, s_uastc3_to_bc1, s_uastc4_to_bc1, s_uastc5_to_bc1 };
				
	void encode_bc4(void* pDst, const uint8_t* pPixels, uint32_t stride)
	{
		uint32_t min0_v, max0_v, min1_v, max1_v,min2_v, max2_v, min3_v, max3_v;

		{
			min0_v = max0_v = pPixels[0 * stride];
			min1_v = max1_v = pPixels[1 * stride];
			min2_v = max2_v = pPixels[2 * stride];
			min3_v = max3_v = pPixels[3 * stride];
		}

		{
			uint32_t v0 = pPixels[4 * stride]; min0_v = basisu::minimum(min0_v, v0); max0_v = basisu::maximum(max0_v, v0);
			uint32_t v1 = pPixels[5 * stride]; min1_v = basisu::minimum(min1_v, v1); max1_v = basisu::maximum(max1_v, v1);
			uint32_t v2 = pPixels[6 * stride]; min2_v = basisu::minimum(min2_v, v2); max2_v = basisu::maximum(max2_v, v2);
			uint32_t v3 = pPixels[7 * stride]; min3_v = basisu::minimum(min3_v, v3); max3_v = basisu::maximum(max3_v, v3);
		}

		{
			uint32_t v0 = pPixels[8 * stride]; min0_v = basisu::minimum(min0_v, v0); max0_v = basisu::maximum(max0_v, v0);
			uint32_t v1 = pPixels[9 * stride]; min1_v = basisu::minimum(min1_v, v1); max1_v = basisu::maximum(max1_v, v1);
			uint32_t v2 = pPixels[10 * stride]; min2_v = basisu::minimum(min2_v, v2); max2_v = basisu::maximum(max2_v, v2);
			uint32_t v3 = pPixels[11 * stride]; min3_v = basisu::minimum(min3_v, v3); max3_v = basisu::maximum(max3_v, v3);
		}

		{
			uint32_t v0 = pPixels[12 * stride]; min0_v = basisu::minimum(min0_v, v0); max0_v = basisu::maximum(max0_v, v0);
			uint32_t v1 = pPixels[13 * stride]; min1_v = basisu::minimum(min1_v, v1); max1_v = basisu::maximum(max1_v, v1);
			uint32_t v2 = pPixels[14 * stride]; min2_v = basisu::minimum(min2_v, v2); max2_v = basisu::maximum(max2_v, v2);
			uint32_t v3 = pPixels[15 * stride]; min3_v = basisu::minimum(min3_v, v3); max3_v = basisu::maximum(max3_v, v3);
		}

		const uint32_t min_v = basisu::minimum(min0_v, min1_v, min2_v, min3_v);
		const uint32_t max_v = basisu::maximum(max0_v, max1_v, max2_v, max3_v);

		uint8_t* pDst_bytes = static_cast<uint8_t*>(pDst);
		pDst_bytes[0] = (uint8_t)max_v;
		pDst_bytes[1] = (uint8_t)min_v;

		if (max_v == min_v)
		{
			memset(pDst_bytes + 2, 0, 6);
			return;
		}

		const uint32_t delta = max_v - min_v;

		// min_v is now 0. Compute thresholds between values by scaling max_v. It's x14 because we're adding two x7 scale factors.
		const int t0 = delta * 13;
		const int t1 = delta * 11;
		const int t2 = delta * 9;
		const int t3 = delta * 7;
		const int t4 = delta * 5;
		const int t5 = delta * 3;
		const int t6 = delta * 1;

		// BC4 floors in its divisions, which we compensate for with the 4 bias.
		// This function is optimal for all possible inputs (i.e. it outputs the same results as checking all 8 values and choosing the closest one).