AMD FidelityFX Super Resolution v1.0.2 for mpv

FSR.glsl

// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.

// FidelityFX FSR v1.0.2 by AMD
// ported to mpv by agyild

// Changelog
// Made it compatible with pre-OpenGL 4.0 renderers
// Made it directly operate on LUMA plane, since the original shader was operating on LUMA by deriving it from RGB. This should cause a major increase in performance, especially on OpenGL 4.0+ renderers (4+2 texture lookups vs. 12+5)
// Removed transparency preservation mechanism since the alpha channel is a separate source plane than LUMA
// Added optional performance-saving lossy optimizations to EASU (Credit: atyuwen, https://atyuwen.github.io/posts/optimizing-fsr/)
// 
// Notes
// Per AMD's guidelines only upscales content up to 4x (e.g., 1080p -> 2160p, 720p -> 1440p etc.) and everything else in between,
// that means FSR will scale up to 4x at maximum, and any further scaling will be processed by mpv's scalers

//!HOOK LUMA
//!BIND HOOKED
//!SAVE EASUTEX
//!DESC FidelityFX Super Resolution v1.0.2 (EASU)
//!WHEN OUTPUT.w OUTPUT.h * LUMA.w LUMA.h * / 1.0 >
//!WIDTH OUTPUT.w OUTPUT.w LUMA.w 2 * < * LUMA.w 2 * OUTPUT.w LUMA.w 2 * > * + OUTPUT.w OUTPUT.w LUMA.w 2 * = * +
//!HEIGHT OUTPUT.h OUTPUT.h LUMA.h 2 * < * LUMA.h 2 * OUTPUT.h LUMA.h 2 * > * + OUTPUT.h OUTPUT.h LUMA.h 2 * = * +
//!COMPONENTS 1

// User variables - EASU
#define FSR_PQ 0 // Whether the source content has PQ gamma or not. Needs to be set to the same value for both passes. 0 or 1.
#define FSR_EASU_DERING 1 // If set to 0, disables deringing for a small increase in performance. 0 or 1.
#define FSR_EASU_SIMPLE_ANALYSIS 0 // If set to 1, uses a simpler single-pass direction and length analysis for an increase in performance. 0 or 1.
#define FSR_EASU_QUIT_EARLY 0 // If set to 1, uses bilinear filtering for non-edge pixels and skips EASU on those regions for an increase in performance. 0 or 1.

// Shader code

#ifndef FSR_EASU_DIR_THRESHOLD
	#if (FSR_EASU_QUIT_EARLY == 1)
		#define FSR_EASU_DIR_THRESHOLD 64.0
	#elif (FSR_EASU_QUIT_EARLY == 0)
		#define FSR_EASU_DIR_THRESHOLD 32768.0
	#endif
#endif

float APrxLoRcpF1(float a) {
	return uintBitsToFloat(uint(0x7ef07ebb) - floatBitsToUint(a));
}

float APrxLoRsqF1(float a) {
	return uintBitsToFloat(uint(0x5f347d74) - (floatBitsToUint(a) >> uint(1)));
}

float AMin3F1(float x, float y, float z) {
	return min(x, min(y, z));
}

float AMax3F1(float x, float y, float z) {
	return max(x, max(y, z));
}

#if (FSR_PQ == 1)

float ToGamma2(float a) { 
	return pow(a, 4.0);
}

#endif

 // Filtering for a given tap for the scalar.
 void FsrEasuTap(
	inout float aC,	// Accumulated color, with negative lobe.
	inout float aW, // Accumulated weight.
	vec2 off,       // Pixel offset from resolve position to tap.
	vec2 dir,       // Gradient direction.
	vec2 len,       // Length.
	float lob,      // Negative lobe strength.
	float clp,		// Clipping point.
	float c){		// Tap color.
	// Rotate offset by direction.
	vec2 v;
	v.x = (off.x * ( dir.x)) + (off.y * dir.y);
	v.y = (off.x * (-dir.y)) + (off.y * dir.x);
	// Anisotropy.
	v *= len;
	// Compute distance^2.
	float d2 = v.x * v.x + v.y * v.y;
	// Limit to the window as at corner, 2 taps can easily be outside.
	d2 = min(d2, clp);
	// Approximation of lancos2 without sin() or rcp(), or sqrt() to get x.
	//  (25/16 * (2/5 * x^2 - 1)^2 - (25/16 - 1)) * (1/4 * x^2 - 1)^2
	//  |_______________________________________|   |_______________|
	//                   base                             window
	// The general form of the 'base' is,
	//  (a*(b*x^2-1)^2-(a-1))
	// Where 'a=1/(2*b-b^2)' and 'b' moves around the negative lobe.
	float wB = float(2.0 / 5.0) * d2 + -1.0;
	float wA = lob * d2 + -1.0;
	wB *= wB;
	wA *= wA;
	wB = float(25.0 / 16.0) * wB + float(-(25.0 / 16.0 - 1.0));
	float w = wB * wA;
	// Do weighted average.
	aC += c * w;
	aW += w;
}

// Accumulate direction and length.
void FsrEasuSet(
	inout vec2 dir,
	inout float len,
	vec2 pp,
#if (FSR_EASU_SIMPLE_ANALYSIS == 1)
	float b, float c,
	float i, float j, float f, float e,
	float k, float l, float h, float g,
	float o, float n
#elif (FSR_EASU_SIMPLE_ANALYSIS == 0)
	bool biS, bool biT, bool biU, bool biV,
	float lA, float lB, float lC, float lD, float lE
#endif
	){
	// Compute bilinear weight, branches factor out as predicates are compiler time immediates.
	//  s t
	//  u v
#if (FSR_EASU_SIMPLE_ANALYSIS == 1)
	vec4 w = vec4(0.0);
	w.x = (1.0 - pp.x) * (1.0 - pp.y);
	w.y =        pp.x  * (1.0 - pp.y);
	w.z = (1.0 - pp.x) *        pp.y;
	w.w =        pp.x  *        pp.y;

	float lA = dot(w, vec4(b, c, f, g));
	float lB = dot(w, vec4(e, f, i, j));
	float lC = dot(w, vec4(f, g, j, k));
	float lD = dot(w, vec4(g, h, k, l));
	float lE = dot(w, vec4(j, k, n, o));
#elif (FSR_EASU_SIMPLE_ANALYSIS == 0)
	float w = 0.0;
	if (biS)
		w = (1.0 - pp.x) * (1.0 - pp.y);
	if (biT)
		w =        pp.x  * (1.0 - pp.y);
	if (biU)
		w = (1.0 - pp.x) *        pp.y;
	if (biV)
		w =        pp.x  *        pp.y;
#endif
	// Direction is the '+' diff.
	//    a
	//  b c d
	//    e
	// Then takes magnitude from abs average of both sides of 'c'.
	// Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms.
	float dc = lD - lC;
	float cb = lC - lB;
	float lenX = max(abs(dc), abs(cb));
	lenX = APrxLoRcpF1(lenX);
	float dirX = lD - lB;
	lenX = clamp(abs(dirX) * lenX, 0.0, 1.0);
	lenX *= lenX;
	// Repeat for the y axis.
	float ec = lE - lC;
	float ca = lC - lA;
	float lenY = max(abs(ec), abs(ca));
	lenY = APrxLoRcpF1(lenY);
	float dirY = lE - lA;
	lenY = clamp(abs(dirY) * lenY, 0.0, 1.0);
	lenY *= lenY;
#if (FSR_EASU_SIMPLE_ANALYSIS == 1)
	len = lenX + lenY;
	dir = vec2(dirX, dirY);
#elif (FSR_EASU_SIMPLE_ANALYSIS == 0)
	dir += vec2(dirX, dirY) * w;
	len += dot(vec2(w), vec2(lenX, lenY));
#endif
}

vec4 hook() {
	// Result
	vec4 pix = vec4(0.0, 0.0, 0.0, 1.0);

	//------------------------------------------------------------------------------------------------------------------------------
	//      +---+---+
	//      |   |   |
	//      +--(0)--+
	//      | b | c |
	//  +---F---+---+---+
	//  | e | f | g | h |
	//  +--(1)--+--(2)--+
	//  | i | j | k | l |
	//  +---+---+---+---+
	//      | n | o |
	//      +--(3)--+
	//      |   |   |
	//      +---+---+
	// Get position of 'F'.
	vec2 pp = HOOKED_pos * HOOKED_size - vec2(0.5);
	vec2 fp = floor(pp);
	pp -= fp;
	//------------------------------------------------------------------------------------------------------------------------------
	// 12-tap kernel.
	//    b c
	//  e f g h
	//  i j k l
	//    n o
	// Gather 4 ordering.
	//  a b
	//  r g
	// Allowing dead-code removal to remove the 'z's.
#if (defined(HOOKED_gather) && (__VERSION__ >= 400 || (GL_ES && __VERSION__ >= 310)))
	vec4 bczzL = HOOKED_gather(vec2((fp + vec2(1.0, -1.0)) * HOOKED_pt), 0);
	vec4 ijfeL = HOOKED_gather(vec2((fp + vec2(0.0,  1.0)) * HOOKED_pt), 0);
	vec4 klhgL = HOOKED_gather(vec2((fp + vec2(2.0,  1.0)) * HOOKED_pt), 0);
	vec4 zzonL = HOOKED_gather(vec2((fp + vec2(1.0,  3.0)) * HOOKED_pt), 0);
#else
	// pre-OpenGL 4.0 compatibility
	float b = HOOKED_tex(vec2((fp + vec2(0.5, -0.5)) * HOOKED_pt)).r;
	float c = HOOKED_tex(vec2((fp + vec2(1.5, -0.5)) * HOOKED_pt)).r;
	
	float e = HOOKED_tex(vec2((fp + vec2(-0.5, 0.5)) * HOOKED_pt)).r;
	float f = HOOKED_tex(vec2((fp + vec2( 0.5, 0.5)) * HOOKED_pt)).r;
	float g = HOOKED_tex(vec2((fp + vec2( 1.5, 0.5)) * HOOKED_pt)).r;
	float h = HOOKED_tex(vec2((fp + vec2( 2.5, 0.5)) * HOOKED_pt)).r;
	
	float i = HOOKED_tex(vec2((fp + vec2(-0.5, 1.5)) * HOOKED_pt)).r;
	float j = HOOKED_tex(vec2((fp + vec2( 0.5, 1.5)) * HOOKED_pt)).r;
	float k = HOOKED_tex(vec2((fp + vec2( 1.5, 1.5)) * HOOKED_pt)).r;
	float l = HOOKED_tex(vec2((fp + vec2( 2.5, 1.5)) * HOOKED_pt)).r;
	
	float n = HOOKED_tex(vec2((fp + vec2(0.5, 2.5) ) * HOOKED_pt)).r;
	float o = HOOKED_tex(vec2((fp + vec2(1.5, 2.5) ) * HOOKED_pt)).r;

	vec4 bczzL = vec4(b, c, 0.0, 0.0);
	vec4 ijfeL = vec4(i, j, f, e);
	vec4 klhgL = vec4(k, l, h, g);
	vec4 zzonL = vec4(0.0, 0.0, o, n);
#endif
	//------------------------------------------------------------------------------------------------------------------------------
	// Rename.
	float bL = bczzL.x;
	float cL = bczzL.y;
	float iL = ijfeL.x;
	float jL = ijfeL.y;
	float fL = ijfeL.z;
	float eL = ijfeL.w;
	float kL = klhgL.x;
	float lL = klhgL.y;
	float hL = klhgL.z;
	float gL = klhgL.w;
	float oL = zzonL.z;
	float nL = zzonL.w;

#if (FSR_PQ == 1)
	// Not the most performance-friendly solution, but should work until mpv adds proper gamma transformation functions for shaders
	bL = ToGamma2(bL);
	cL = ToGamma2(cL);
	iL = ToGamma2(iL);
	jL = ToGamma2(jL);
	fL = ToGamma2(fL);
	eL = ToGamma2(eL);
	kL = ToGamma2(kL);
	lL = ToGamma2(lL);
	hL = ToGamma2(hL);
	gL = ToGamma2(gL);
	oL = ToGamma2(oL);
	nL = ToGamma2(nL);
#endif

	// Accumulate for bilinear interpolation.
	vec2 dir = vec2(0.0);
	float len = 0.0;
#if (FSR_EASU_SIMPLE_ANALYSIS == 1)
	FsrEasuSet(dir, len, pp, bL, cL, iL, jL, fL, eL, kL, lL, hL, gL, oL, nL);
#elif (FSR_EASU_SIMPLE_ANALYSIS == 0)
	FsrEasuSet(dir, len, pp, true, false, false, false, bL, eL, fL, gL, jL);
	FsrEasuSet(dir, len, pp, false, true, false, false, cL, fL, gL, hL, kL);
	FsrEasuSet(dir, len, pp, false, false, true, false, fL, iL, jL, kL, nL);
	FsrEasuSet(dir, len, pp, false, false, false, true, gL, jL, kL, lL, oL);
#endif
	//------------------------------------------------------------------------------------------------------------------------------
	// Normalize with approximation, and cleanup close to zero.
	vec2 dir2 = dir * dir;
	float dirR = dir2.x + dir2.y;
	bool zro = dirR < float(1.0 / FSR_EASU_DIR_THRESHOLD);
	dirR = APrxLoRsqF1(dirR);
#if (FSR_EASU_QUIT_EARLY == 1)
	if (zro) {
		vec4 w = vec4(0.0);
		w.x = (1.0 - pp.x) * (1.0 - pp.y);
		w.y =        pp.x  * (1.0 - pp.y);
		w.z = (1.0 - pp.x) *        pp.y;
		w.w =        pp.x  *        pp.y;

		pix.r = clamp(dot(w, vec4(fL, gL, jL, kL)), 0.0, 1.0);
		return pix;
	}
#elif (FSR_EASU_QUIT_EARLY == 0)
	dirR = zro ? 1.0 : dirR;
	dir.x = zro ? 1.0 : dir.x;
#endif
	dir *= vec2(dirR);
	// Transform from {0 to 2} to {0 to 1} range, and shape with square.
	len = len * 0.5;
	len *= len;
	// Stretch kernel {1.0 vert|horz, to sqrt(2.0) on diagonal}.
	float stretch = (dir.x * dir.x + dir.y * dir.y) * APrxLoRcpF1(max(abs(dir.x), abs(dir.y)));
	// Anisotropic length after rotation,
	//  x := 1.0 lerp to 'stretch' on edges
	//  y := 1.0 lerp to 2x on edges
	vec2 len2 = vec2(1.0 + (stretch - 1.0) * len, 1.0 + -0.5 * len);
	// Based on the amount of 'edge',
	// the window shifts from +/-{sqrt(2.0) to slightly beyond 2.0}.
	float lob = 0.5 + float((1.0 / 4.0 - 0.04) - 0.5) * len;
	// Set distance^2 clipping point to the end of the adjustable window.
	float clp = APrxLoRcpF1(lob);
	//------------------------------------------------------------------------------------------------------------------------------
	// Accumulation
	//    b c
	//  e f g h
	//  i j k l
	//    n o
	float aC = 0.0;
	float aW = 0.0;
	FsrEasuTap(aC, aW, vec2( 0.0,-1.0) - pp, dir, len2, lob, clp, bL); // b
	FsrEasuTap(aC, aW, vec2( 1.0,-1.0) - pp, dir, len2, lob, clp, cL); // c
	FsrEasuTap(aC, aW, vec2(-1.0, 1.0) - pp, dir, len2, lob, clp, iL); // i
	FsrEasuTap(aC, aW, vec2( 0.0, 1.0) - pp, dir, len2, lob, clp, jL); // j
	FsrEasuTap(aC, aW, vec2( 0.0, 0.0) - pp, dir, len2, lob, clp, fL); // f
	FsrEasuTap(aC, aW, vec2(-1.0, 0.0) - pp, dir, len2, lob, clp, eL); // e
	FsrEasuTap(aC, aW, vec2( 1.0, 1.0) - pp, dir, len2, lob, clp, kL); // k
	FsrEasuTap(aC, aW, vec2( 2.0, 1.0) - pp, dir, len2, lob, clp, lL); // l
	FsrEasuTap(aC, aW, vec2( 2.0, 0.0) - pp, dir, len2, lob, clp, hL); // h
	FsrEasuTap(aC, aW, vec2( 1.0, 0.0) - pp, dir, len2, lob, clp, gL); // g
	FsrEasuTap(aC, aW, vec2( 1.0, 2.0) - pp, dir, len2, lob, clp, oL); // o
	FsrEasuTap(aC, aW, vec2( 0.0, 2.0) - pp, dir, len2, lob, clp, nL); // n
	//------------------------------------------------------------------------------------------------------------------------------
	// Normalize and dering.
	pix.r = aC / aW;
#if (FSR_EASU_DERING == 1)
	float min1 = min(AMin3F1(fL, gL, jL), kL);
	float max1 = max(AMax3F1(fL, gL, jL), kL);
	pix.r = clamp(pix.r, min1, max1);
#endif
	pix.r = clamp(pix.r, 0.0, 1.0);

	return pix;
}

//!HOOK LUMA
//!BIND EASUTEX
//!DESC FidelityFX Super Resolution v1.0.2 (RCAS)
//!WIDTH EASUTEX.w
//!HEIGHT EASUTEX.h
//!COMPONENTS 1

// User variables - RCAS
#define SHARPNESS 0.2 // Controls the amount of sharpening. The scale is {0.0 := maximum, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. 0.0 to 2.0.
#define FSR_RCAS_DENOISE 1 // If set to 1, lessens the sharpening on noisy areas. Can be disabled for better performance. 0 or 1.
#define FSR_PQ 0 // Whether the source content has PQ gamma or not. Needs to be set to the same value for both passes. 0 or 1.

// Shader code

#define FSR_RCAS_LIMIT (0.25 - (1.0 / 16.0)) // This is set at the limit of providing unnatural results for sharpening.

float APrxMedRcpF1(float a) {
	float b = uintBitsToFloat(uint(0x7ef19fff) - floatBitsToUint(a));
	return b * (-b * a + 2.0);
}

float AMax3F1(float x, float y, float z) {
	return max(x, max(y, z)); 
}

float AMin3F1(float x, float y, float z) {
	return min(x, min(y, z));
}

#if (FSR_PQ == 1)

float FromGamma2(float a) { 
	return sqrt(sqrt(a));
}

#endif

vec4 hook() {
	// Algorithm uses minimal 3x3 pixel neighborhood.
	//    b 
	//  d e f
	//    h
#if (defined(EASUTEX_gather) && (__VERSION__ >= 400 || (GL_ES && __VERSION__ >= 310)))
	vec3 bde = EASUTEX_gather(EASUTEX_pos + EASUTEX_pt * vec2(-0.5), 0).xyz;
	float b = bde.z;
	float d = bde.x;
	float e = bde.y;

	vec2 fh = EASUTEX_gather(EASUTEX_pos + EASUTEX_pt * vec2(0.5), 0).zx;
	float f = fh.x;
	float h = fh.y;
#else
	float b = EASUTEX_texOff(vec2( 0.0, -1.0)).r;
	float d = EASUTEX_texOff(vec2(-1.0,  0.0)).r;
	float e = EASUTEX_tex(EASUTEX_pos).r;
	float f = EASUTEX_texOff(vec2(1.0, 0.0)).r;
	float h = EASUTEX_texOff(vec2(0.0, 1.0)).r;
#endif

	// Min and max of ring.
	float mn1L = min(AMin3F1(b, d, f), h);
	float mx1L = max(AMax3F1(b, d, f), h);

	// Immediate constants for peak range.
	vec2 peakC = vec2(1.0, -1.0 * 4.0);

	// Limiters, these need to be high precision RCPs.
	float hitMinL = min(mn1L, e) / (4.0 * mx1L);
	float hitMaxL = (peakC.x - max(mx1L, e)) / (4.0 * mn1L + peakC.y);
	float lobeL = max(-hitMinL, hitMaxL);
	float lobe = max(float(-FSR_RCAS_LIMIT), min(lobeL, 0.0)) * exp2(-clamp(float(SHARPNESS), 0.0, 2.0));

	// Apply noise removal.
#if (FSR_RCAS_DENOISE == 1)
	// Noise detection.
	float nz = 0.25 * b + 0.25 * d + 0.25 * f + 0.25 * h - e;
	nz = clamp(abs(nz) * APrxMedRcpF1(AMax3F1(AMax3F1(b, d, e), f, h) - AMin3F1(AMin3F1(b, d, e), f, h)), 0.0, 1.0);
	nz = -0.5 * nz + 1.0;
	lobe *= nz;
#endif

	// Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes.
	float rcpL = APrxMedRcpF1(4.0 * lobe + 1.0);
	vec4 pix = vec4(0.0, 0.0, 0.0, 1.0);
	pix.r = float((lobe * b + lobe * d + lobe * h + lobe * f + e) * rcpL);
#if (FSR_PQ == 1)
	pix.r = FromGamma2(pix.r);
#endif

	return pix;
}

How do you use this?

https://mpv.io/manual/master/#options-glsl-shaders

Will there be a FSR2 even FSR3 version of this? The quality of 2 and 3 seems to be a lot improvement.
Is the quality of FSRCNNX still better than AMD FSR?

Will there be a FSR2 even FSR3 version of this? Is the quality of FSRCNNX better or AMD FSR better?

Check out this comparison page.

Will there be a FSR2 even FSR3 version of this? Is the quality of FSRCNNX better or AMD FSR better?

Check out this comparison page.

I have tested AMD FSR and FSRCNNX with a 240P video by myself, AMD FSR is still far worse than FSRCNNX, if somebody is able to implement FSR2 or even FSR3 in the future, the result may be different.

Will there be a FSR2 even FSR3 version of this? Is the quality of FSRCNNX better or AMD FSR better?

Check out this comparison page.

I have tested AMD FSR and FSRCNNX with a 240P video by myself, AMD FSR is still far worse than FSRCNNX, if somebody is able to implement FSR2 or even FSR3 in the future, the result may be different.

Video doesn't have motion vectors that are required for FSR 2+ so unlikely that would ever happen.

Will there be a FSR2 even FSR3 version of this? Is the quality of FSRCNNX better or AMD FSR better?

Check out this comparison page.

I have tested AMD FSR and FSRCNNX with a 240P video by myself, AMD FSR is still far worse than FSRCNNX, if somebody is able to implement FSR2 or even FSR3 in the future, the result may be different.

Video doesn't have motion vectors that are required for FSR 2+ so unlikely that would ever happen.

Do you mean that there is no improvement for video upscaling with FSR 2?
By the way, is there any benefit to run AMD FSR two times for a 404p video on my 1080p monitor?
Like this: 404p----(FSR)----808p----(FSR)----1616p----(SSimDownscaler)----1080p?

Will there be a FSR2 even FSR3 version of this? Is the quality of FSRCNNX better or AMD FSR better?

Check out this comparison page.

I have tested AMD FSR and FSRCNNX with a 240P video by myself, AMD FSR is still far worse than FSRCNNX, if somebody is able to implement FSR2 or even FSR3 in the future, the result may be different.

Video doesn't have motion vectors that are required for FSR 2+ so unlikely that would ever happen.

Do you mean that there is no improvement for video upscaling with FSR 2? By the way, is there any benefit to run AMD FSR two times for a 404p video on my 1080p monitor?

I'm saying it's not possible to implement it, video streams don't contain the information required by FSR 2.

By the way, is there any benefit to run AMD FSR two times for a 404p video on my 1080p monitor? Like this: 404p----(FSR)----808p----(FSR)----1616p----(SSimDownscaler)----1080p?

FSR scales directly to the target resolution so no benefit to run it twice.
Nvm, I was making assumptions, see below.

By the way, is there any benefit to run AMD FSR two times for a 404p video on my 1080p monitor? Like this: 404p----(FSR)----808p----(FSR)----1616p----(SSimDownscaler)----1080p?

FSR scales directly to the target resolution so no benefit to run it twice.

Wrong. It scales up to 4x (i.e., 2x per dimension), with anything more than that processed by the mpv's built-in scaler (i.e., --scale). Applying it more than once will produce an over-sharpened image with no new details.

Do you mean that there is no improvement for video upscaling with FSR 2?

Video upscaling can benefit from using data from the previous and even next frames (which is a unique benefit of pre-recorded video content), however implementing it requires low-level code changes in mpv since temporal upscaling cannot be fully implemented in pixel shaders which work entirely based on spatial information (i.e., the current frame). Also, there are the following challenges:

1. Temporal upscaling will result in ghosting artifacts because the whole scene information is used as input instead of individual object information within the scene. In a 3D-rendered environment such as a video game you have access to other information such as per-pixel motion direction and amount, and depth information along with when a scene significantly changes so you don't accidentally sample pixels from an entirely different scene in the next or previous frames.
2. Theoretically, one can use computer vision algorithms to guesstimate the aforementioned information, but the problem is such algorithms are highly domain-specific (i.e., each visual characteristic such as anime, 3d, real world, etc. requires different algorithm, and differentiating which scene is what automatically would be yet another computer vision algorithm on top of everything). This is a very performance demanding process and dumbed down real-time algorithms won't generate the necessary high resolution information.

So tl;dr is FSR2 is simply not feasible. The next best thing would be requesting the mpv developers to implement a compatibility layer for something like RTX Video Super Resolution which is basically a hardware accelerated and more powerful version of FSRCNN.

Sorry for being a nitwit but... TL;DR for newbies would be: if you have an NVidia GPU use the NVidia shader. If you have Intel/AMD/Whatever, use FSR.glsl + CAS-SCALED.glsl, correct?

And is your github the (un)official repository for mpv FSR? And what or who are the FSRCNNX shaders for? Please help me and other nitwits make sense of it all :)

Sorry for being a nitwit but... TL;DR for newbies would be: if you have an NVidia GPU use the NVidia shader. If you have Intel/AMD/Whatever, use FSR.glsl + CAS-SCALED.glsl, correct?

All shaders are vendor agnostic, use the one that is most aesthetically pleasing. They are just bunch of code that works the same everywhere with the NVIDIA/AMD label on them.

And is your github the (un)official repository for mpv FSR? And what or who are the FSRCNNX shaders for? Please help me and other nitwits make sense of it all :)

I would not call it official, but this gist is the only FSR 1 port for mpv that I am aware of. FSRCNNX is a completely separate shader maintained by igv, not me. Names are similar, the technology is different.

I have my 32" 4K monitor
I put two conditional profile in mpv.conf

[fsrcnnx]
glsl-shader=~~/shaders/FSRCNNX_x2_16-0-4-1.glsl
profile-cond=math.min(display_width / width, display_height / height) >= 3.0

[fsr]
glsl-shader=~~/shaders/AMD_FSR.glsl
profile-cond=math.min(display_width / width, display_height / height) < 3.0

No matter the video is 1920x1080 or 1280x720, only fsr profile was triggered

why can't I use fsr and cas at the same time.

why can't I use fsr and cas at the same time.

You don't need to. FSR already includes a sharpening pass called RCAS which is a variant of CAS specifically made to work with FSR.

nub question, is this FSR 3? if not any plans to add it? There is even talk about FSR adding ai at some point. will that help with mpv to watch anime?

I want to upscale 720p/1080p to 1440p resolution

nub question, is this FSR 3? if not any plans to add it? There is even talk about FSR adding ai at some point. will that help with mpv to watch anime?

I want to upscale 720p/1080p to 1440p resolution

It's not possible to implement.

Ty Zabooby.

1 port for mpv that I am aware of. FSRCNNX is a completely separate shader maintained by igv, not me. Names are similar, the technology is different.

confused by this. is FSRCNNX usable for RX 7800 XT AMD radeon gpu? On the FSRCNN site it says tensor. Isn't that a Nvidia only?

Also i think sharpening was being done by FCR already? so is this scale=ewa_lanczossharp still needed? or is that something different? Then what about "--profile=gpu-hq" is this entry needed or not at all when using FCR or anime4k shaders?

scale=ewa_lanczossharp

[fsrcnnx]
glsl-shader=~~/shaders/FSRCNNX_x2_16-0-4-1.glsl
profile-cond=math.min(display_width / width, display_height / height) >= 2.0

[fsr]
glsl-shader=~~/shaders/FSR.glsl
profile-cond=math.min(display_width / width, display_height / height) < 2.0

Yes FSRCNNX is usable, press i then 2 to check if mpv is reporting it is running or not.

scale= is used only when you don't have any external shaders running if I'm not mistaken.

profile=gpu-hq is outdated use profile=high-quality.

yeah i press i then 2. but for mpv.net which i am using, press t then 2 :}

i just wasnt sure if it was working or not. good to know, ty.

ah, that explains about gpu-hq. ty.

so basically i have to use profile=high-quality, and either fsr with fsrcnnx, OR

profile=high-quality with anime4k?

hope i got that right. Ty sensei

so basically i have to use profile=high-quality, and either fsr with fsrcnnx, OR

profile=high-quality with anime4k?

It is up to you to choose which shaders you want, I personally pick my shaders based on this comparison.

Is there a newer version available? Like FidelityFX Super Resolution 2.2 for mpv??

Re-read all the comments above.
tl;dr : IMPOSSIBLE.

Re-read all the comments above. tl;dr : IMPOSSIBLE.

Ah I see now. I thought it was about ffs3. But ffs2 was also mentioned as impossible. Donkers.

Why would they make ffs2 requirements that high you can't even upscale a video.. 😒?

Why would they make ffs2 requirements that high you can't even upscale a video.. 😒?

Requirements are not high, the underlying technologies are different. High-fidelity upscaling algorithms such as FSR2+ and DLSS are developed for use in video games, the fact that FSR1 and NIS can be used for video playback in the first place is simply a byproduct of their design or a hack.

Video playback does not include motion buffers because each individual scene is prebaked into the video stream. Video games have it because game engines create each scene via real-time rendering. Correspondingly, FSR2+ and DLSS won't be able to upscale pre-rendered in-game video content (e.g., cutscenes, TV screens, etc.) either.