From a4ff60233fce134e8f779ef001872dfd9a8f9923 Mon Sep 17 00:00:00 2001 From: skal Date: Sat, 21 Mar 2026 08:38:29 +0100 Subject: feat(cnn_v3): Phase 3 complete — WGSL U-Net inference shaders MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 5 compute shaders + cnn_v3/common snippet: enc0: Conv(20→4,3×3) + FiLM + ReLU full-res enc1: AvgPool + Conv(4→8,3×3) + FiLM + ReLU half-res bottleneck: AvgPool + Conv(8→8,1×1) + ReLU quarter-res dec1: NearestUp + cat(enc1) + Conv(16→4) + FiLM half-res dec0: NearestUp + cat(enc0) + Conv(8→4) + FiLM + Sigmoid full-res Parity rules: zero-pad conv, AvgPool down, NearestUp, FiLM after conv+bias, skip=concat, OIHW weights+bias layout. Matches PyTorch train_cnn_v3.py forward() exactly. Registered in workspaces/main/assets.txt + src/effects/shaders.cc. Weight layout + Params struct documented in cnn_v3/docs/HOWTO.md §7. Next: Phase 4 — C++ CNNv3Effect + FiLM uniform upload. Co-Authored-By: Claude Sonnet 4.6 --- cnn_v3/shaders/cnn_v3_bottleneck.wgsl | 73 +++++++++++++++++++++++++++++ cnn_v3/shaders/cnn_v3_common.wgsl | 23 +++++++++ cnn_v3/shaders/cnn_v3_dec0.wgsl | 73 +++++++++++++++++++++++++++++ cnn_v3/shaders/cnn_v3_dec1.wgsl | 72 ++++++++++++++++++++++++++++ cnn_v3/shaders/cnn_v3_enc0.wgsl | 75 +++++++++++++++++++++++++++++ cnn_v3/shaders/cnn_v3_enc1.wgsl | 88 +++++++++++++++++++++++++++++++++++ 6 files changed, 404 insertions(+) create mode 100644 cnn_v3/shaders/cnn_v3_bottleneck.wgsl create mode 100644 cnn_v3/shaders/cnn_v3_common.wgsl create mode 100644 cnn_v3/shaders/cnn_v3_dec0.wgsl create mode 100644 cnn_v3/shaders/cnn_v3_dec1.wgsl create mode 100644 cnn_v3/shaders/cnn_v3_enc0.wgsl create mode 100644 cnn_v3/shaders/cnn_v3_enc1.wgsl (limited to 'cnn_v3/shaders') diff --git a/cnn_v3/shaders/cnn_v3_bottleneck.wgsl b/cnn_v3/shaders/cnn_v3_bottleneck.wgsl new file mode 100644 index 0000000..909fd41 --- /dev/null +++ b/cnn_v3/shaders/cnn_v3_bottleneck.wgsl @@ -0,0 +1,73 @@ +// CNN v3 — Bottleneck +// AvgPool2x2(enc1) + Conv(8->8, 1x1) + ReLU (no FiLM) +// +// Input: enc1_tex (rgba32uint, 8xf16) half-res +// Output: bottleneck_out (rgba32uint, 8xf16) quarter-res (dispatch at quarter-res dims) +// +// Weight layout (f16, OIHW + bias): +// [0 .. 8*8*1) conv: w[out][in] (1x1 kernel) +// [64 .. +8) bias: b[out] + +#include "cnn_v3/common" + +const BN_IN: u32 = 8u; +const BN_OUT: u32 = 8u; + +struct Params { + weight_offset: u32, + _pad0: u32, + _pad1: u32, + _pad2: u32, +} + +@group(0) @binding(0) var enc1_tex: texture_2d; +@group(0) @binding(1) var weights: array; +@group(0) @binding(2) var params: Params; +@group(0) @binding(3) var bottleneck_out: texture_storage_2d; + +// Avg-pool 2x2 from enc1_tex at quarter-res coord qcoord. +// Returns zeros for OOB quarter-res coords (zero-padding for the 1x1 conv). +fn load_enc1_avg(qcoord: vec2i, half_dims: vec2i) -> array { + let quart_dims = half_dims / 2; + if (qcoord.x < 0 || qcoord.y < 0 || qcoord.x >= quart_dims.x || qcoord.y >= quart_dims.y) { + return array(0., 0., 0., 0., 0., 0., 0., 0.); + } + let base = qcoord * 2; + var s: array; + for (var dy: i32 = 0; dy < 2; dy++) { + for (var dx: i32 = 0; dx < 2; dx++) { + let hc = clamp(base + vec2i(dx, dy), vec2i(0), half_dims - vec2i(1)); + let f = unpack_8ch(enc1_tex, hc); + for (var i: u32 = 0u; i < BN_IN; i++) { s[i] += f[i]; } + } + } + for (var i: u32 = 0u; i < BN_IN; i++) { s[i] *= 0.25; } + return s; +} + +@compute @workgroup_size(8, 8) +fn bottleneck_main(@builtin(global_invocation_id) id: vec3u) { + let half_dims = vec2i(textureDimensions(enc1_tex)); + let quart_dims = half_dims / 2; + let coord = vec2i(id.xy); + if (coord.x >= quart_dims.x || coord.y >= quart_dims.y) { return; } + + let wo = params.weight_offset; + let feat = load_enc1_avg(coord, half_dims); + var out: array; + + for (var o: u32 = 0u; o < BN_OUT; o++) { + var sum = get_w(wo, BN_OUT * BN_IN + o); // bias (1x1 kernel: no spatial idx) + for (var i: u32 = 0u; i < BN_IN; i++) { + sum += get_w(wo, o * BN_IN + i) * feat[i]; + } + out[o] = max(0.0, sum); + } + + textureStore(bottleneck_out, coord, vec4u( + pack2x16float(vec2f(out[0], out[1])), + pack2x16float(vec2f(out[2], out[3])), + pack2x16float(vec2f(out[4], out[5])), + pack2x16float(vec2f(out[6], out[7])) + )); +} diff --git a/cnn_v3/shaders/cnn_v3_common.wgsl b/cnn_v3/shaders/cnn_v3_common.wgsl new file mode 100644 index 0000000..54b0f3d --- /dev/null +++ b/cnn_v3/shaders/cnn_v3_common.wgsl @@ -0,0 +1,23 @@ +// CNN v3 shared helpers — included by all inference compute shaders. +// Requires the host shader to declare: +// @group(?) @binding(?) var weights: array; + +// Read one f16 value from the packed-f16 weights buffer. +// `base` — weight_offset from Params (f16 index of the layer start) +// `idx` — local f16 index within the layer (conv weight or bias) +fn get_w(base: u32, idx: u32) -> f32 { + let i = base + idx; + let v = unpack2x16float(weights[i >> 1u]); + return select(v.y, v.x, (i & 1u) == 0u); +} + +// Unpack 8 f16 channels from an rgba32uint texel (pack2x16float layout: +// u32[0]=ch0|ch1, u32[1]=ch2|ch3, u32[2]=ch4|ch5, u32[3]=ch6|ch7) +fn unpack_8ch(tex: texture_2d, coord: vec2i) -> array { + let t = textureLoad(tex, coord, 0); + let v0 = unpack2x16float(t.x); + let v1 = unpack2x16float(t.y); + let v2 = unpack2x16float(t.z); + let v3 = unpack2x16float(t.w); + return array(v0.x, v0.y, v1.x, v1.y, v2.x, v2.y, v3.x, v3.y); +} diff --git a/cnn_v3/shaders/cnn_v3_dec0.wgsl b/cnn_v3/shaders/cnn_v3_dec0.wgsl new file mode 100644 index 0000000..7a4e7c9 --- /dev/null +++ b/cnn_v3/shaders/cnn_v3_dec0.wgsl @@ -0,0 +1,73 @@ +// CNN v3 — Decoder level 0 + output +// NearestUp2x(dec1) + cat(enc0_skip) -> Conv(8->4, 3x3, zero-pad) + FiLM + ReLU + Sigmoid +// +// Inputs: dec1_tex (rgba16float, 4ch) half-res +// enc0_tex (rgba16float, 4ch) full-res (skip connection) +// Output: output_tex (rgba16float, 4ch) full-res (dispatch at full-res dims) +// +// Weight layout (f16, OIHW + bias): +// [0 .. 8*4*9) conv: w[out][in][ky][kx] (in=8: 4 dec1 + 4 enc0 skip) +// [288 .. +4) bias: b[out] +// +// Parity note: sigmoid applied directly to dec0 output (matches train_cnn_v3.py forward()). + +#include "cnn_v3/common" + +const DEC0_IN: u32 = 8u; +const DEC0_OUT: u32 = 4u; + +struct Params { + weight_offset: u32, + _pad: vec3u, + gamma: vec4f, + beta: vec4f, +} + +@group(0) @binding(0) var dec1_tex: texture_2d; +@group(0) @binding(1) var enc0_tex: texture_2d; +@group(0) @binding(2) var weights: array; +@group(0) @binding(3) var params: Params; +@group(0) @binding(4) var output_tex: texture_storage_2d; + +// Load 8 concatenated channels at full-res coord: +// ch 0-3: dec1 nearest-up (dec1_tex[coord/2]) +// ch 4-7: enc0 skip (enc0_tex[coord]) +// Returns zeros for OOB coord (zero-padding for the conv). +fn load_dec0_concat(coord: vec2i, full_dims: vec2i) -> array { + if (coord.x < 0 || coord.y < 0 || coord.x >= full_dims.x || coord.y >= full_dims.y) { + return array(0., 0., 0., 0., 0., 0., 0., 0.); + } + let half_dims = vec2i(textureDimensions(dec1_tex)); + let hc = clamp(coord / 2, vec2i(0), half_dims - vec2i(1)); + let d = textureLoad(dec1_tex, hc, 0); + let e = textureLoad(enc0_tex, coord, 0); + return array(d.x, d.y, d.z, d.w, e.x, e.y, e.z, e.w); +} + +@compute @workgroup_size(8, 8) +fn dec0_main(@builtin(global_invocation_id) id: vec3u) { + let full_dims = vec2i(textureDimensions(enc0_tex)); + let coord = vec2i(id.xy); + if (coord.x >= full_dims.x || coord.y >= full_dims.y) { return; } + + let wo = params.weight_offset; + var out: array; + + for (var o: u32 = 0u; o < DEC0_OUT; o++) { + var sum = get_w(wo, DEC0_OUT * DEC0_IN * 9u + o); // bias + for (var ky: i32 = -1; ky <= 1; ky++) { + for (var kx: i32 = -1; kx <= 1; kx++) { + let feat = load_dec0_concat(coord + vec2i(kx, ky), full_dims); + let ki = u32(ky + 1) * 3u + u32(kx + 1); + for (var i: u32 = 0u; i < DEC0_IN; i++) { + sum += get_w(wo, o * DEC0_IN * 9u + i * 9u + ki) * feat[i]; + } + } + } + // FiLM + ReLU + Sigmoid (matches training forward()) + let v = max(0.0, params.gamma[o] * sum + params.beta[o]); + out[o] = 1.0 / (1.0 + exp(-v)); + } + + textureStore(output_tex, coord, vec4f(out[0], out[1], out[2], out[3])); +} diff --git a/cnn_v3/shaders/cnn_v3_dec1.wgsl b/cnn_v3/shaders/cnn_v3_dec1.wgsl new file mode 100644 index 0000000..28ae3dc --- /dev/null +++ b/cnn_v3/shaders/cnn_v3_dec1.wgsl @@ -0,0 +1,72 @@ +// CNN v3 — Decoder level 1 +// NearestUp2x(bottleneck) + cat(enc1_skip) -> Conv(16->4, 3x3, zero-pad) + FiLM + ReLU +// +// Inputs: bottleneck_tex (rgba32uint, 8xf16) quarter-res +// enc1_tex (rgba32uint, 8xf16) half-res (skip connection) +// Output: dec1_out (rgba16float, 4ch) half-res (dispatch at half-res dims) +// +// Weight layout (f16, OIHW + bias): +// [0 .. 16*4*9) conv: w[out][in][ky][kx] (in=16: 8 bottleneck + 8 enc1 skip) +// [576 .. +4) bias: b[out] + +#include "cnn_v3/common" + +const DEC1_IN: u32 = 16u; +const DEC1_OUT: u32 = 4u; + +struct Params { + weight_offset: u32, + _pad: vec3u, + gamma: vec4f, + beta: vec4f, +} + +@group(0) @binding(0) var bottleneck_tex: texture_2d; +@group(0) @binding(1) var enc1_tex: texture_2d; +@group(0) @binding(2) var weights: array; +@group(0) @binding(3) var params: Params; +@group(0) @binding(4) var dec1_out: texture_storage_2d; + +// Load 16 concatenated channels at half-res coord hcoord: +// ch 0-7: bottleneck nearest-up (bottleneck_tex[hcoord/2]) +// ch 8-15: enc1 skip (enc1_tex[hcoord]) +// Returns zeros for OOB hcoord (zero-padding for the conv). +fn load_dec1_concat(hcoord: vec2i, half_dims: vec2i) -> array { + if (hcoord.x < 0 || hcoord.y < 0 || hcoord.x >= half_dims.x || hcoord.y >= half_dims.y) { + return array(0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.); + } + let quart_dims = half_dims / 2; + let qc = clamp(hcoord / 2, vec2i(0), quart_dims - vec2i(1)); + let b = unpack_8ch(bottleneck_tex, qc); + let s = unpack_8ch(enc1_tex, hcoord); + return array( + b[0], b[1], b[2], b[3], b[4], b[5], b[6], b[7], + s[0], s[1], s[2], s[3], s[4], s[5], s[6], s[7] + ); +} + +@compute @workgroup_size(8, 8) +fn dec1_main(@builtin(global_invocation_id) id: vec3u) { + let half_dims = vec2i(textureDimensions(enc1_tex)); + let coord = vec2i(id.xy); + if (coord.x >= half_dims.x || coord.y >= half_dims.y) { return; } + + let wo = params.weight_offset; + var out: array; + + for (var o: u32 = 0u; o < DEC1_OUT; o++) { + var sum = get_w(wo, DEC1_OUT * DEC1_IN * 9u + o); // bias + for (var ky: i32 = -1; ky <= 1; ky++) { + for (var kx: i32 = -1; kx <= 1; kx++) { + let feat = load_dec1_concat(coord + vec2i(kx, ky), half_dims); + let ki = u32(ky + 1) * 3u + u32(kx + 1); + for (var i: u32 = 0u; i < DEC1_IN; i++) { + sum += get_w(wo, o * DEC1_IN * 9u + i * 9u + ki) * feat[i]; + } + } + } + out[o] = max(0.0, params.gamma[o] * sum + params.beta[o]); + } + + textureStore(dec1_out, coord, vec4f(out[0], out[1], out[2], out[3])); +} diff --git a/cnn_v3/shaders/cnn_v3_enc0.wgsl b/cnn_v3/shaders/cnn_v3_enc0.wgsl new file mode 100644 index 0000000..f52a167 --- /dev/null +++ b/cnn_v3/shaders/cnn_v3_enc0.wgsl @@ -0,0 +1,75 @@ +// CNN v3 — Encoder level 0 +// Conv(20->4, 3x3, zero-pad) + FiLM + ReLU +// +// Input: feat_tex0 (rgba32uint, 8xf16), feat_tex1 (rgba32uint, 12xu8) full-res +// Output: enc0_out (rgba16float, 4ch) full-res +// +// Weight layout (f16, OIHW + bias): +// [0 .. 20*4*9) conv: w[out][in][ky][kx] +// [720 .. +4) bias: b[out] + +#include "cnn_v3/common" + +const ENC0_IN: u32 = 20u; +const ENC0_OUT: u32 = 4u; + +struct Params { + weight_offset: u32, + _pad: vec3u, + gamma: vec4f, + beta: vec4f, +} + +@group(0) @binding(0) var feat_tex0: texture_2d; +@group(0) @binding(1) var feat_tex1: texture_2d; +@group(0) @binding(2) var weights: array; +@group(0) @binding(3) var params: Params; +@group(0) @binding(4) var enc0_out: texture_storage_2d; + +// Unpack all 20 feature channels at coord. Returns zeros for OOB (zero-padding). +fn load_feat(coord: vec2i, dims: vec2i) -> array { + if (coord.x < 0 || coord.y < 0 || coord.x >= dims.x || coord.y >= dims.y) { + return array(0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.); + } + let t0 = textureLoad(feat_tex0, coord, 0); + let t1 = textureLoad(feat_tex1, coord, 0); + let a = unpack2x16float(t0.x); + let b = unpack2x16float(t0.y); + let c = unpack2x16float(t0.z); + let d = unpack2x16float(t0.w); + let e = unpack4x8unorm(t1.x); + let f = unpack4x8unorm(t1.y); + let g = unpack4x8unorm(t1.z); + return array( + a.x, a.y, b.x, b.y, c.x, c.y, d.x, d.y, + e.x, e.y, e.z, e.w, + f.x, f.y, f.z, f.w, + g.x, g.y, g.z, g.w + ); +} + +@compute @workgroup_size(8, 8) +fn enc0_main(@builtin(global_invocation_id) id: vec3u) { + let coord = vec2i(id.xy); + let dims = vec2i(textureDimensions(feat_tex0)); + if (coord.x >= dims.x || coord.y >= dims.y) { return; } + + let wo = params.weight_offset; + var out: array; + + for (var o: u32 = 0u; o < ENC0_OUT; o++) { + var sum = get_w(wo, ENC0_OUT * ENC0_IN * 9u + o); // bias + for (var ky: i32 = -1; ky <= 1; ky++) { + for (var kx: i32 = -1; kx <= 1; kx++) { + let feat = load_feat(coord + vec2i(kx, ky), dims); + let ki = u32(ky + 1) * 3u + u32(kx + 1); + for (var i: u32 = 0u; i < ENC0_IN; i++) { + sum += get_w(wo, o * ENC0_IN * 9u + i * 9u + ki) * feat[i]; + } + } + } + out[o] = max(0.0, params.gamma[o] * sum + params.beta[o]); + } + + textureStore(enc0_out, coord, vec4f(out[0], out[1], out[2], out[3])); +} diff --git a/cnn_v3/shaders/cnn_v3_enc1.wgsl b/cnn_v3/shaders/cnn_v3_enc1.wgsl new file mode 100644 index 0000000..23e485d --- /dev/null +++ b/cnn_v3/shaders/cnn_v3_enc1.wgsl @@ -0,0 +1,88 @@ +// CNN v3 — Encoder level 1 +// AvgPool2x2(enc0) + Conv(4->8, 3x3, zero-pad) + FiLM + ReLU +// +// Input: enc0_tex (rgba16float, 4ch) full-res +// Output: enc1_out (rgba32uint, 8xf16) half-res (dispatch at half-res dims) +// +// Weight layout (f16, OIHW + bias): +// [0 .. 4*8*9) conv: w[out][in][ky][kx] +// [288 .. +8) bias: b[out] + +#include "cnn_v3/common" + +const ENC1_IN: u32 = 4u; +const ENC1_OUT: u32 = 8u; + +struct Params { + weight_offset: u32, + _pad: vec3u, + gamma_lo: vec4f, // FiLM gamma ch 0-3 + gamma_hi: vec4f, // FiLM gamma ch 4-7 + beta_lo: vec4f, // FiLM beta ch 0-3 + beta_hi: vec4f, // FiLM beta ch 4-7 +} + +@group(0) @binding(0) var enc0_tex: texture_2d; +@group(0) @binding(1) var weights: array; +@group(0) @binding(2) var params: Params; +@group(0) @binding(3) var enc1_out: texture_storage_2d; + +fn film_gamma(o: u32) -> f32 { + if (o < 4u) { return params.gamma_lo[o]; } + return params.gamma_hi[o - 4u]; +} +fn film_beta(o: u32) -> f32 { + if (o < 4u) { return params.beta_lo[o]; } + return params.beta_hi[o - 4u]; +} + +// Avg-pool 2x2 from enc0_tex at half-res coord hcoord. +// Returns zeros for OOB half-res coords (zero-padding for the conv). +fn load_enc0_avg(hcoord: vec2i, full_dims: vec2i) -> array { + let half_dims = full_dims / 2; + if (hcoord.x < 0 || hcoord.y < 0 || hcoord.x >= half_dims.x || hcoord.y >= half_dims.y) { + return array(0., 0., 0., 0.); + } + let base = hcoord * 2; + var s = vec4f(0.); + for (var dy: i32 = 0; dy < 2; dy++) { + for (var dx: i32 = 0; dx < 2; dx++) { + let fc = clamp(base + vec2i(dx, dy), vec2i(0), full_dims - vec2i(1)); + s += textureLoad(enc0_tex, fc, 0); + } + } + let avg = s * 0.25; + return array(avg.x, avg.y, avg.z, avg.w); +} + +@compute @workgroup_size(8, 8) +fn enc1_main(@builtin(global_invocation_id) id: vec3u) { + let full_dims = vec2i(textureDimensions(enc0_tex)); + let half_dims = full_dims / 2; + let coord = vec2i(id.xy); + if (coord.x >= half_dims.x || coord.y >= half_dims.y) { return; } + + let wo = params.weight_offset; + var out: array; + + for (var o: u32 = 0u; o < ENC1_OUT; o++) { + var sum = get_w(wo, ENC1_OUT * ENC1_IN * 9u + o); // bias + for (var ky: i32 = -1; ky <= 1; ky++) { + for (var kx: i32 = -1; kx <= 1; kx++) { + let feat = load_enc0_avg(coord + vec2i(kx, ky), full_dims); + let ki = u32(ky + 1) * 3u + u32(kx + 1); + for (var i: u32 = 0u; i < ENC1_IN; i++) { + sum += get_w(wo, o * ENC1_IN * 9u + i * 9u + ki) * feat[i]; + } + } + } + out[o] = max(0.0, film_gamma(o) * sum + film_beta(o)); + } + + textureStore(enc1_out, coord, vec4u( + pack2x16float(vec2f(out[0], out[1])), + pack2x16float(vec2f(out[2], out[3])), + pack2x16float(vec2f(out[4], out[5])), + pack2x16float(vec2f(out[6], out[7])) + )); +} -- cgit v1.2.3