Add exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5 (attention bounded by the used context)

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/00f1518df8d42f8c7c48cbb8d4080ed273fc370066401f5caf6316d3df3a9c72.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 10240u;
+if (i >= 10240u) { return; }
+out[i] = f32(b0[((i / 10240u) % 1u) * 10240u + ((i / 320u) % 32u) * 80u + ((i / 80u) % 4u) * 2560u + ((i / 1u) % 80u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/010ceb4c6d74413ae14f7a70e539bbcad465036da065b435f5f99436ba183f3a.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 1280u;
+if (i >= 1280u) { return; }
+let x = f32(b0[i]);
+out[i] = f32(cos(x));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/03780813af6ed9f38ba73c3512379e4a9db2730eb91497d8dde2d21a3984ae1b.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 640u;
+if (i >= 640u) { return; }
+out[i] = f32(b0[((i / 640u) % 1u) * 640u + ((i / 80u) % 8u) * 80u + ((i / 80u) % 1u) * 640u + ((i / 1u) % 80u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/04813c632839a0ae83168b6ec110ba2fe39c338097ed8689dd2d8541ea90e8f0.wgsl

@@ -0,0 +1,12 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 102400u;
+if (i >= 102400u) { return; }
+let coord = (i / 1u) % 102400u;
+if (coord >= 0u && coord < 52428u) { out[i] = f32(b0[(i / 102400u) * 52428u + (coord - 0u) * 1u + i % 1u]); }
+if (coord >= 52428u && coord < 102400u) { out[i] = f32(b1[(i / 102400u) * 49972u + (coord - 52428u) * 1u + i % 1u]); }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/0549459ddfcb213196091a0ea09245acf8c06f3d0d6272f4db4126192a956a6b.wgsl

@@ -0,0 +1,71 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<u32>;
+@group(0) @binding(2) var<storage, read> b2: array<f32>;
+@group(0) @binding(3) var<storage, read_write> out: array<f32>;
+fn dequant_1(index: u32) -> f32 {
+    let row = index / 2560u;
+    let col = index % 2560u;
+    let word = b1[row * 640u + col / 4u];
+    let code = i32((word >> ((col % 4u) * 8u)) & 255u) - 128;
+    return f32(code) * b2[row * 80u + col / 32u];
+}
+
+var<workgroup> tile_a: array<array<f32, 16>, 16>;
+var<workgroup> tile_b: array<array<f32, 64>, 16>;
+@compute @workgroup_size(16, 16)
+fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(local_invocation_id) local: vec3<u32>) {
+    let lane = local.y * 16u + local.x;
+    let batch = group.z;
+    let tile_n = group.x * 64u;
+    let tile_row = group.y * 16u;
+    var acc: array<array<f32, 4>, 1>;
+    for (var k0 = 0u; k0 < 2560u; k0 += 16u) {
+        for (var e = 0u; e < 1u; e++) {
+            let flat = lane + e * 256u;
+            let m_local = flat / 16u;
+            let row = tile_row + m_local;
+            let col = k0 + flat % 16u;
+            var value = 0.0;
+            if (row < 16u && col < 2560u) { value = f32(f32(b0[(batch * 16u + row) * 2560u + col])); }
+            tile_a[m_local][flat % 16u] = value;
+        }
+        if (lane < 256u) {
+            let n_local = lane / 4u;
+            let word_local = lane % 4u;
+            let n_index = tile_n + n_local;
+            let first = k0 + word_local * 4u;
+            var word = 0u;
+            var scale = 0.0;
+            if (n_index < 52428u && first < 2560u) {
+                word = b1[n_index * 640u + first / 4u];
+                scale = f32(b2[n_index * 80u + first / 32u]);
+            }
+            for (var e = 0u; e < 4u; e++) {
+                let code = i32((word >> (e * 8u)) & 255u) - 128;
+                var value = 0.0;
+                if (first + e < 2560u) { value = f32(code) * scale; }
+                tile_b[word_local * 4u + e][n_local] = value;
+            }
+        }
+        workgroupBarrier();
+        for (var kk = 0u; kk < 16u; kk++) {
+            var b_values: array<f32, 4>;
+            for (var c = 0u; c < 4u; c++) { b_values[c] = tile_b[kk][local.x * 4u + c]; }
+            for (var r = 0u; r < 1u; r++) {
+                let a_value = tile_a[local.y * 1u + r][kk];
+                for (var c = 0u; c < 4u; c++) { acc[r][c] += a_value * b_values[c]; }
+            }
+        }
+        workgroupBarrier();
+    }
+    for (var r = 0u; r < 1u; r++) {
+        let out_row = tile_row + local.y * 1u + r;
+        for (var c = 0u; c < 4u; c++) {
+            let out_col = tile_n + local.x * 4u + c;
+            if (out_row < 16u && out_col < 52428u) {
+                let i = (batch * 16u + out_row) * 52428u + out_col;
+                out[i] = f32(acc[r][c] + 0.0);
+            }
+        }
+    }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/056a7cde988532d104769fe6246161d0a150f08c4f910918110826ce621bb5a3.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 40960u;
+if (i >= 40960u) { return; }
+out[i] = f32(f32(b0[i]) * f32(b1[((i / 80u) % 64u) * 80u + ((i / 1u) % 80u) * 1u]));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/065de78c2ded88ff8cfca05f54ba48f64761693422c9a04ff2801fac2e85606a.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 5120u;
+if (i >= 5120u) { return; }
+let x = f32(b0[i]);
+out[i] = f32(sin(x));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/0b36e6f66583375c2c8a3426a96b1e0c913dfcbb16ab06838d9a186ec06d1289.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 5120u;
+if (i >= 5120u) { return; }
+out[i] = f32(b0[((i / 5120u) % 1u) * 10240u + ((i / 160u) % 32u) * 320u + ((i / 40u) % 4u) * 80u + (((i / 1u) % 40u) * 1u + 40u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/0cb55e74ef816571494ea173176bf81e797c73d0cb0b8a57a1c0a51469197a02.wgsl

@@ -0,0 +1,96 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read> b2: array<f32>;
+@group(0) @binding(3) var<storage, read> b3: array<f32>;
+@group(0) @binding(4) var<storage, read_write> out: array<f32>;
+
+var<workgroup> q_row: array<f32, 80>;
+var<workgroup> scores: array<f32, 1024>;
+var<workgroup> partial: array<f32, 64>;
+var<workgroup> limit_shared: u32;
+@compute @workgroup_size(64)
+fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(local_invocation_id) local: vec3<u32>) {
+    let lane = local.x;
+    let head = group.x / 64u;
+    let row = group.x % 64u;
+    let kv_head = head / 4u;
+    for (var d = lane; d < 80u; d += 64u) { q_row[d] = f32(b0[(head * 64u + row) * 80u + d]); }
+    // Bound the key loop by the last unmasked position so work follows the used context.
+    var last = 0.0;
+    for (var j = lane; j < 4096u; j += 64u) {
+        if (f32(b3[row * 4096u + j]) > -1000.0) { last = max(last, f32(j + 1u)); }
+    }
+    partial[lane] = last;
+    workgroupBarrier();
+    for (var stride = 32u; stride > 0u; stride /= 2u) {
+        if (lane < stride) { partial[lane] = max(partial[lane], partial[lane + stride]); }
+        workgroupBarrier();
+    }
+    if (lane == 0u) { limit_shared = u32(partial[0]); }
+    let limit = workgroupUniformLoad(&limit_shared);
+    var running_max = -3.4028234663852886e38;
+    var running_sum = 0.0;
+    var acc: array<f32, 2>;
+    for (var start = 0u; start < limit; start += 1024u) {
+        let count = min(1024u, limit - start);
+        var tile_max = -3.4028234663852886e38;
+        for (var t = lane; t < count; t += 64u) {
+            let j = start + t;
+            var score = -3.4028234663852886e38;
+            if (f32(b3[row * 4096u + j]) > -1000.0) {
+                var dot = 0.0;
+                for (var d = 0u; d < 80u; d++) { dot += q_row[d] * f32(b1[(((kv_head * 4096u + j) * 80u + d) / 80u) % 4096u * 640u + (((kv_head * 4096u + j) * 80u + d) / 327680u) % 8u * 80u + (((kv_head * 4096u + j) * 80u + d) / 1u) % 80u * 1u]); }
+                score = dot * 0.11180339887498948 + f32(b3[row * 4096u + j]);
+            }
+            scores[t] = score;
+            tile_max = max(tile_max, score);
+        }
+        partial[lane] = tile_max;
+        workgroupBarrier();
+        for (var stride = 32u; stride > 0u; stride /= 2u) {
+            if (lane < stride) { partial[lane] = max(partial[lane], partial[lane + stride]); }
+            workgroupBarrier();
+        }
+        let next_max = max(running_max, partial[0]);
+        workgroupBarrier();
+        var tile_sum = 0.0;
+        for (var t = lane; t < count; t += 64u) {
+            var p = 0.0;
+            if (scores[t] > -3.0e38) { p = exp(scores[t] - next_max); }
+            scores[t] = p;
+            tile_sum += p;
+        }
+        partial[lane] = tile_sum;
+        workgroupBarrier();
+        for (var stride = 32u; stride > 0u; stride /= 2u) {
+            if (lane < stride) { partial[lane] += partial[lane + stride]; }
+            workgroupBarrier();
+        }
+        var correction = 0.0;
+        if (running_max > -3.0e38) { correction = exp(running_max - next_max); }
+        running_sum = running_sum * correction + partial[0];
+        running_max = next_max;
+        for (var c = 0u; c < 2u; c++) {
+            let d = lane + c * 64u;
+            var total = acc[c] * correction;
+            if (d < 80u) {
+                for (var t = 0u; t < count; t++) {
+                    let p = scores[t];
+                    if (p != 0.0) {
+                        let j = start + t;
+                        total += p * f32(b2[(((kv_head * 4096u + j) * 80u + d) / 80u) % 4096u * 640u + (((kv_head * 4096u + j) * 80u + d) / 327680u) % 8u * 80u + (((kv_head * 4096u + j) * 80u + d) / 1u) % 80u * 1u]);
+                    }
+                }
+            }
+            acc[c] = total;
+        }
+        workgroupBarrier();
+    }
+    for (var c = 0u; c < 2u; c++) {
+        let d = lane + c * 64u;
+        if (d < 80u) {
+            let i = (head * 64u + row) * 80u + d;
+            out[i] = f32(acc[c] / running_sum);
+        }
+    }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/17a32140ebd06816dfb4d571da93b9fe24e622c069bb563882a65efdc9bd3045.wgsl

@@ -0,0 +1,12 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 640u;
+if (i >= 640u) { return; }
+let coord = (i / 1u) % 80u;
+if (coord >= 0u && coord < 40u) { out[i] = f32(b0[(i / 80u) * 40u + (coord - 0u) * 1u + i % 1u]); }
+if (coord >= 40u && coord < 80u) { out[i] = f32(b1[(i / 80u) * 40u + (coord - 40u) * 1u + i % 1u]); }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/17cf2e11eabbbfb1c28393ccc0cd4a6f4145566dd8d65f2a87be3e3481dd48e5.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 81920u;
+if (i >= 81920u) { return; }
+let x = f32(b0[i]);
+out[i] = f32(-x);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/189c65aa14859d6642ba29508ca2150726cc5d94b6cd1e72d8ddb2b14aaf603f.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 40960u;
+if (i >= 40960u) { return; }
+out[i] = f32(b0[((i / 40960u) % 1u) * 40960u + ((i / 5120u) % 8u) * 80u + ((i / 80u) % 64u) * 640u + ((i / 1u) % 80u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/18a66ec0875c456c9defd45d15e6ef0c148e64a1618e2825c7f96eadbbe8303a.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<i32>;
+@group(0) @binding(1) var<storage, read_write> out: array<i32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 64u;
+if (i >= 16u) { return; }
+out[i] = i32(b0[i]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1b3a44837ef6433e8faa7fb4dae08de71f68861ca914355ed5f8711a0b28719b.wgsl

@@ -0,0 +1,17 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 64u;
+if (i >= 40u) { return; }
+let batch = i / 40u;
+let row = (i / 1u) % 40u;
+let col = i % 1u;
+var acc = 0.0;
+for (var p = 0u; p < 1u; p++) {
+    acc += f32(b0[(0u) + row * 1u + p]) * f32(b1[(0u) + p * 1u + col]);
+}
+out[i] = f32(acc);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1b9e76d54a728dada6475c0d04d6c5cb5fec9a862cd2e01bd9ce77b16e0d96b6.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 1280u;
+if (i >= 1280u) { return; }
+out[i] = f32(f32(b0[i]) * 1.0);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1bf469f7b8f1f89afadd837b786ff5e2f9f210b8a163b844343cfed2865d790f.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 5120u;
+if (i >= 5120u) { return; }
+out[i] = f32(f32(b0[i]) * 1.0);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1e4e50d6a6d5c9380d37a3462475e30fa473e77585e8d36baff974e5c1152782.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 2560u;
+if (i >= 2560u) { return; }
+out[i] = f32(f32(b0[i]) + (f32(b1[i]) * 1.0));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1e9e57e0500e94c2e40db913222f7286a3886aa562c14f234e94361200a5fb2a.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 640u;
+if (i >= 640u) { return; }
+out[i] = f32(b0[((i / 640u) % 1u) * 640u + ((i / 40u) % 16u) * 1u + ((i / 1u) % 40u) * 16u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1ec850641f1c68248987e77bb1c9bb8f66bc8ded1c08941ea790ae4979bb7174.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 20480u;
+if (i >= 20480u) { return; }
+out[i] = f32(b0[((i / 20480u) % 1u) * 40960u + ((i / 640u) % 32u) * 1280u + ((i / 40u) % 16u) * 80u + (((i / 1u) % 40u) * 1u + 40u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1ece9b553926e5f15825e1e8cb6638808227e8e23c5063f4b6d35ef4ef05a83c.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 2560u;
+if (i >= 2560u) { return; }
+out[i] = f32(f32(b0[i]) * f32(b1[((i / 80u) % 4u) * 80u + ((i / 1u) % 80u) * 1u]));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/1ed82ae7c212a7174f1c1a80c3575f118523ff8c1de4ca54fb2d1fe542f4c031.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 320u;
+if (i >= 320u) { return; }
+let x = f32(b0[i]);
+out[i] = f32(cos(x));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/265f0d6e72ce92141e27fd00420e8565670ccc37e566aaf34ecfe9ec3d96725a.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 20480u;
+if (i >= 20480u) { return; }
+out[i] = f32(b0[((i / 20480u) % 1u) * 40960u + ((i / 2560u) % 8u) * 5120u + ((i / 40u) % 64u) * 80u + (((i / 1u) % 40u) * 1u + 40u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/27a921d3b5edfdca227a99e1c3328959881c09acb02b12cd09d07d598745b2d3.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<i32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 64u;
+if (i >= 1u) { return; }
+out[i] = f32(b0[i]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/307d2f0d7330af4c5fb41a9e52936470ccada38a6fe9bfa4530025d0188b8c3e.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 163840u;
+if (i >= 163840u) { return; }
+out[i] = f32(f32(b0[i]) + (f32(b1[i]) * 1.0));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/30fe0ee3a884aae9e3a4487b694d0bb86934bda87591466f4c4a8bda90e13c95.wgsl

@@ -0,0 +1,71 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<u32>;
+@group(0) @binding(2) var<storage, read> b2: array<f32>;
+@group(0) @binding(3) var<storage, read_write> out: array<f32>;
+fn dequant_1(index: u32) -> f32 {
+    let row = index / 2560u;
+    let col = index % 2560u;
+    let word = b1[row * 640u + col / 4u];
+    let code = i32((word >> ((col % 4u) * 8u)) & 255u) - 128;
+    return f32(code) * b2[row * 80u + col / 32u];
+}
+
+var<workgroup> tile_a: array<array<f32, 16>, 16>;
+var<workgroup> tile_b: array<array<f32, 64>, 16>;
+@compute @workgroup_size(16, 16)
+fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(local_invocation_id) local: vec3<u32>) {
+    let lane = local.y * 16u + local.x;
+    let batch = group.z;
+    let tile_n = group.x * 64u;
+    let tile_row = group.y * 16u;
+    var acc: array<array<f32, 4>, 1>;
+    for (var k0 = 0u; k0 < 2560u; k0 += 16u) {
+        for (var e = 0u; e < 1u; e++) {
+            let flat = lane + e * 256u;
+            let m_local = flat / 16u;
+            let row = tile_row + m_local;
+            let col = k0 + flat % 16u;
+            var value = 0.0;
+            if (row < 4u && col < 2560u) { value = f32(f32(b0[(batch * 4u + row) * 2560u + col])); }
+            tile_a[m_local][flat % 16u] = value;
+        }
+        if (lane < 256u) {
+            let n_local = lane / 4u;
+            let word_local = lane % 4u;
+            let n_index = tile_n + n_local;
+            let first = k0 + word_local * 4u;
+            var word = 0u;
+            var scale = 0.0;
+            if (n_index < 52428u && first < 2560u) {
+                word = b1[n_index * 640u + first / 4u];
+                scale = f32(b2[n_index * 80u + first / 32u]);
+            }
+            for (var e = 0u; e < 4u; e++) {
+                let code = i32((word >> (e * 8u)) & 255u) - 128;
+                var value = 0.0;
+                if (first + e < 2560u) { value = f32(code) * scale; }
+                tile_b[word_local * 4u + e][n_local] = value;
+            }
+        }
+        workgroupBarrier();
+        for (var kk = 0u; kk < 16u; kk++) {
+            var b_values: array<f32, 4>;
+            for (var c = 0u; c < 4u; c++) { b_values[c] = tile_b[kk][local.x * 4u + c]; }
+            for (var r = 0u; r < 1u; r++) {
+                let a_value = tile_a[local.y * 1u + r][kk];
+                for (var c = 0u; c < 4u; c++) { acc[r][c] += a_value * b_values[c]; }
+            }
+        }
+        workgroupBarrier();
+    }
+    for (var r = 0u; r < 1u; r++) {
+        let out_row = tile_row + local.y * 1u + r;
+        for (var c = 0u; c < 4u; c++) {
+            let out_col = tile_n + local.x * 4u + c;
+            if (out_row < 4u && out_col < 52428u) {
+                let i = (batch * 4u + out_row) * 52428u + out_col;
+                out[i] = f32(acc[r][c] + 0.0);
+            }
+        }
+    }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/31d236875acba515c98c35ba2b251fe95c92813111f4851a2c817589c5b872e8.wgsl

@@ -0,0 +1,24 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+var<workgroup> factor: f32;
+@compute @workgroup_size(64)
+fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(local_invocation_id) local: vec3<u32>) {
+    let row = group.x + group.y * 1u;
+    if (row >= 1u) { return; }
+    let lane = local.x;
+    if (lane == 0u) {
+        var total = 0.0;
+        for (var j = 0u; j < 2560u; j++) {
+            let v = f32(b0[row * 2560u + j]);
+            total += v * v;
+        }
+        factor = inverseSqrt(total / 2560.0 + 1e-05);
+    }
+    workgroupBarrier();
+    for (var p = lane; p < 2560u; p += 64u) {
+        let i = row * 2560u + p;
+        out[i] = f32(f32(b0[row * 2560u + p]) * factor * (f32(b1[p]) + 0.0));
+    }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/3203e08f2c99fa4848f553739b7115a7c677840344f17b1d4391d6130517f5a7.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<i32>;
+@group(0) @binding(1) var<storage, read_write> out: array<i32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 64u;
+if (i >= 4u) { return; }
+out[i] = i32(b0[i]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/33d5e3ca854722003d2032af6ad94488ac2caaefaec22cd9bed9df8eb51efc11.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 163840u;
+if (i >= 163840u) { return; }
+out[i] = f32(b0[((i / 163840u) % 1u) * 163840u + ((i / 2560u) % 64u) * 80u + ((i / 80u) % 32u) * 5120u + ((i / 1u) % 80u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/34789e50cf72ea181c86db87caa629346be4ae8221e39c85ebbd836bc22d0ca9.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 10240u;
+if (i >= 10240u) { return; }
+out[i] = f32(f32(b0[i]) * f32(b1[((i / 80u) % 16u) * 80u + ((i / 1u) % 80u) * 1u]));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/35c0d9ce0c1a0fc678ad4893e9b1508c759402b5649ccaca55d94993875f5c2c.wgsl

@@ -0,0 +1,11 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 1280u;
+if (i >= 1280u) { return; }
+let coord = (i / 1u) % 80u;
+if (coord >= 0u && coord < 40u) { out[i] = f32(b0[(i / 80u) * 40u + (coord - 0u) * 1u + i % 1u]); }
+if (coord >= 40u && coord < 80u) { out[i] = f32(b0[(i / 80u) * 40u + (coord - 40u) * 1u + i % 1u]); }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/36eb23461edc6900aa34d9d30acca79f0e3823a35bfbb413e72244886613cf1c.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 320u;
+if (i >= 320u) { return; }
+let x = f32(b0[i]);
+out[i] = f32(sin(x));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/397c272c70f921a18fa5b359f241c16191dabb59affdb25a5f21e0ac97b6ed39.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 10240u;
+if (i >= 10240u) { return; }
+out[i] = f32(b0[((i / 10240u) % 1u) * 10240u + ((i / 1280u) % 8u) * 80u + ((i / 80u) % 16u) * 640u + ((i / 1u) % 80u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/3c1557fa28d308c0aae510ad9480e3454a28c5f9f409906a5b3ff186abba9ef3.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 1280u;
+if (i >= 1280u) { return; }
+out[i] = f32(b0[((i / 1280u) % 1u) * 2560u + ((i / 40u) % 32u) * 80u + ((i / 40u) % 1u) * 80u + (((i / 1u) % 40u) * 1u + 40u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/3d1cd64fbe853586e7d03d72a2a44ee64b39cd9f0c4fd1ca598d09739f742e02.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 1280u;
+if (i >= 1280u) { return; }
+let x = f32(b0[i]);
+out[i] = f32(sin(x));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/3f52269ac3af898dffb89532a3d6d520ba633123c7903e3bcea2aa9ddb276f22.wgsl

@@ -0,0 +1,12 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 4194240u;
+if (i >= 6553600u) { return; }
+let coord = (i / 1u) % 102400u;
+if (coord >= 0u && coord < 52428u) { out[i] = f32(b0[(i / 102400u) * 52428u + (coord - 0u) * 1u + i % 1u]); }
+if (coord >= 52428u && coord < 102400u) { out[i] = f32(b1[(i / 102400u) * 49972u + (coord - 52428u) * 1u + i % 1u]); }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/4082a3d7d09bb1749393ef3ede46c932d3e55f0cf76a03de29293fec96bbb652.wgsl

@@ -0,0 +1,12 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 163840u;
+if (i >= 163840u) { return; }
+let coord = (i / 1u) % 80u;
+if (coord >= 0u && coord < 40u) { out[i] = f32(b0[(i / 80u) * 40u + (coord - 0u) * 1u + i % 1u]); }
+if (coord >= 40u && coord < 80u) { out[i] = f32(b1[(i / 80u) * 40u + (coord - 40u) * 1u + i % 1u]); }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/414d09288943439b0ef850b6e6395ffd2a84e8a26a0793ebef274572730e53d4.wgsl

@@ -0,0 +1,71 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<u32>;
+@group(0) @binding(2) var<storage, read> b2: array<f32>;
+@group(0) @binding(3) var<storage, read_write> out: array<f32>;
+fn dequant_1(index: u32) -> f32 {
+    let row = index / 2560u;
+    let col = index % 2560u;
+    let word = b1[row * 640u + col / 4u];
+    let code = i32((word >> ((col % 4u) * 8u)) & 255u) - 128;
+    return f32(code) * b2[row * 80u + col / 32u];
+}
+
+var<workgroup> tile_a: array<array<f32, 16>, 16>;
+var<workgroup> tile_b: array<array<f32, 64>, 16>;
+@compute @workgroup_size(16, 16)
+fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(local_invocation_id) local: vec3<u32>) {
+    let lane = local.y * 16u + local.x;
+    let batch = group.z;
+    let tile_n = group.x * 64u;
+    let tile_row = group.y * 16u;
+    var acc: array<array<f32, 4>, 1>;
+    for (var k0 = 0u; k0 < 2560u; k0 += 16u) {
+        for (var e = 0u; e < 1u; e++) {
+            let flat = lane + e * 256u;
+            let m_local = flat / 16u;
+            let row = tile_row + m_local;
+            let col = k0 + flat % 16u;
+            var value = 0.0;
+            if (row < 4u && col < 2560u) { value = f32(f32(b0[(batch * 4u + row) * 2560u + col])); }
+            tile_a[m_local][flat % 16u] = value;
+        }
+        if (lane < 256u) {
+            let n_local = lane / 4u;
+            let word_local = lane % 4u;
+            let n_index = tile_n + n_local;
+            let first = k0 + word_local * 4u;
+            var word = 0u;
+            var scale = 0.0;
+            if (n_index < 7168u && first < 2560u) {
+                word = b1[n_index * 640u + first / 4u];
+                scale = f32(b2[n_index * 80u + first / 32u]);
+            }
+            for (var e = 0u; e < 4u; e++) {
+                let code = i32((word >> (e * 8u)) & 255u) - 128;
+                var value = 0.0;
+                if (first + e < 2560u) { value = f32(code) * scale; }
+                tile_b[word_local * 4u + e][n_local] = value;
+            }
+        }
+        workgroupBarrier();
+        for (var kk = 0u; kk < 16u; kk++) {
+            var b_values: array<f32, 4>;
+            for (var c = 0u; c < 4u; c++) { b_values[c] = tile_b[kk][local.x * 4u + c]; }
+            for (var r = 0u; r < 1u; r++) {
+                let a_value = tile_a[local.y * 1u + r][kk];
+                for (var c = 0u; c < 4u; c++) { acc[r][c] += a_value * b_values[c]; }
+            }
+        }
+        workgroupBarrier();
+    }
+    for (var r = 0u; r < 1u; r++) {
+        let out_row = tile_row + local.y * 1u + r;
+        for (var c = 0u; c < 4u; c++) {
+            let out_col = tile_n + local.x * 4u + c;
+            if (out_row < 4u && out_col < 7168u) {
+                let i = (batch * 4u + out_row) * 7168u + out_col;
+                out[i] = f32(acc[r][c] + 0.0);
+            }
+        }
+    }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/42607a63976e415f493f17de2d5044b4a7a3f76d35b65ae2d14cadcc58e612ee.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 5120u;
+if (i >= 5120u) { return; }
+out[i] = f32(b0[((i / 5120u) % 1u) * 10240u + ((i / 640u) % 8u) * 1280u + ((i / 40u) % 16u) * 80u + (((i / 1u) % 40u) * 1u + 0u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/42c746147a03895998a4f97e31b3954f5e9556a50e17feb2e8da86c8b60b605b.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 320u;
+if (i >= 320u) { return; }
+out[i] = f32(b0[((i / 320u) % 1u) * 640u + ((i / 40u) % 8u) * 80u + ((i / 40u) % 1u) * 80u + (((i / 1u) % 40u) * 1u + 40u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/43e65eb3a65477fc67e79d0c196673b8b18f50fc79c4fe7e5f6bc137b5be8296.wgsl

@@ -0,0 +1,12 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 2560u;
+if (i >= 2560u) { return; }
+let coord = (i / 1u) % 80u;
+if (coord >= 0u && coord < 40u) { out[i] = f32(b0[(i / 80u) * 40u + (coord - 0u) * 1u + i % 1u]); }
+if (coord >= 40u && coord < 80u) { out[i] = f32(b1[(i / 80u) * 40u + (coord - 40u) * 1u + i % 1u]); }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/465b3e1dcae446e474286b793b2df123ff14fd88af419b1d1d38b154e09c0956.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 10240u;
+if (i >= 10240u) { return; }
+out[i] = f32(f32(b0[i]) + (f32(b1[i]) * 1.0));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/4e1484472c35832cda4bab8b4714b98868fff0ddc8e038954143e03f345cf905.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 5120u;
+if (i >= 5120u) { return; }
+out[i] = f32(b0[((i / 5120u) % 1u) * 10240u + ((i / 160u) % 32u) * 320u + ((i / 40u) % 4u) * 80u + (((i / 1u) % 40u) * 1u + 0u) * 1u]);
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/4ed96a253c0ffc0646197e209572477a26ded2619581c29f856a3855cd93c383.wgsl

@@ -0,0 +1,20 @@
+@group(0) @binding(0) var<storage, read> b0: array<i32>;
+@group(0) @binding(1) var<storage, read> b1: array<u32>;
+@group(0) @binding(2) var<storage, read> b2: array<f32>;
+@group(0) @binding(3) var<storage, read_write> out: array<f32>;
+fn dequant_1(index: u32) -> f32 {
+    let row = index / 2560u;
+    let col = index % 2560u;
+    let word = b1[row * 640u + col / 4u];
+    let code = i32((word >> ((col % 4u) * 8u)) & 255u) - 128;
+    return f32(code) * b2[row * 80u + col / 32u];
+}
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 2560u;
+if (i >= 2560u) { return; }
+let token = i32(b0[i / 2560u]);
+if (token < 52428 || token >= 102400) { out[i] = f32(0.0); return; }
+out[i] = f32(dequant_1(u32(token - 52428) * 2560u + i % 2560u));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/4ee0b4e25a053c3affa8817ece03ec3f9ddd112db15e3dc8ba998b801df5411d.wgsl

@@ -0,0 +1,10 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 10240u;
+if (i >= 10240u) { return; }
+out[i] = f32(f32(b0[i]) * f32(b1[((i / 80u) % 4u) * 80u + ((i / 1u) % 80u) * 1u]));
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/4f8d9c62e15a05268df079ee741a5f67f800dbf614a74c4a97ed75bdbf8bd37a.wgsl

@@ -0,0 +1,12 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read> b1: array<f32>;
+@group(0) @binding(2) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 10240u;
+if (i >= 10240u) { return; }
+let coord = (i / 1u) % 80u;
+if (coord >= 0u && coord < 40u) { out[i] = f32(b0[(i / 80u) * 40u + (coord - 0u) * 1u + i % 1u]); }
+if (coord >= 40u && coord < 80u) { out[i] = f32(b1[(i / 80u) * 40u + (coord - 40u) * 1u + i % 1u]); }
+}

exaone-deep-fp32-int8-g32-candidate-token-major-v2-v5/kernels/529fe85e34d79998156fe545c69dbc622f69ed375a016704f8e107fcf280c377.wgsl

@@ -0,0 +1,9 @@
+@group(0) @binding(0) var<storage, read> b0: array<f32>;
+@group(0) @binding(1) var<storage, read_write> out: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+let i = gid.x + gid.y * 5120u;
+if (i >= 5120u) { return; }
+out[i] = f32(b0[((i / 5120u) % 1u) * 10240u + ((i / 640u) % 8u) * 1280u + ((i / 40u) % 16u) * 80u + (((i / 1u) % 40u) * 1u + 40u) * 1u]);
+}