whispercpp 1.3.6 → 1.3.7

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl

@@ -0,0 +1,303 @@
+#ifdef U32_DEQUANT_HELPERS
+#define SRC0_TYPE u32
+
+fn byte_of(v: u32, b: u32) -> u32 {
+    return (v >> (b * 8u)) & 0xFFu;
+}
+
+fn sbyte_of(v: u32, b: u32) -> i32 {
+    let raw = i32((v >> (b * 8u)) & 0xFFu);
+    return select(raw, raw - 256, raw >= 128);
+}
+#endif
+
+#define SRC0_TYPE SRC0_INNER_TYPE
+#define SRC1_TYPE SRC1_INNER_TYPE
+
+#ifdef LEGACY_QUANTS
+#define BLOCK_SIZE 32
+#define THREADS_PER_BLOCK 4
+#elif K_QUANTS
+#define BLOCK_SIZE 256
+#define THREADS_PER_BLOCK 16
+#endif
+
+#define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
+#define Q8_BLOCK_SIZE 32
+
+#ifdef MUL_ACC_Q4_0
+#define BLOCK_SIZE_BYTES 18
+#define B_DS_TYPE vec2<f32>
+fn repack_a(block_byte_base: u32, inner_id: u32) -> vec2<u32> {
+    let qs_packed = load_u32_at_src0(block_byte_base + 2u + 4u * inner_id);
+
+    return vec2<u32>(
+        qs_packed & 0x0F0F0F0Fu,
+        (qs_packed >> 4u) & 0x0F0F0F0Fu
+    );
+}
+fn repack_b_qs(block:u32, inner_id: u32) -> vec2<u32> {
+    return vec2<u32>(
+            src1q[block].qs[inner_id],
+            src1q[block].qs[inner_id + 4u],
+        );
+}
+fn repack_b_dm(block: u32) -> B_DS_TYPE {
+    return B_DS_TYPE(
+        f32(src1q[block].d),
+        f32(src1q[block].s)
+    );
+}
+fn get_dm(block_byte_base: u32) -> f32 {
+    return f32(load_f16_at_src0(block_byte_base));
+}
+fn mul_q8_1(row_sum: i32, da: f32, b_ds: B_DS_TYPE) -> f32 {
+    return f32(row_sum) * (da * b_ds.x) - 8.0 * da * b_ds.y / THREADS_PER_BLOCK;
+}
+#endif
+
+#ifdef MUL_ACC_Q4_1
+#define BLOCK_SIZE_BYTES 20
+#define B_DS_TYPE vec2<f32>
+fn repack_a(block_byte_base: u32, inner_id: u32) -> vec2<u32> {
+    let qs_packed = load_u32_at_src0(block_byte_base + 4u + 4u * inner_id);
+
+    return vec2<u32>(
+        qs_packed & 0x0F0F0F0Fu,
+        (qs_packed >> 4u) & 0x0F0F0F0Fu
+    );
+}
+fn repack_b_qs(block:u32, inner_id: u32) -> vec2<u32> {
+    return vec2<u32>(
+            src1q[block].qs[inner_id],
+            src1q[block].qs[inner_id + 4u],
+        );
+}
+fn repack_b_dm(block: u32) -> B_DS_TYPE {
+    return B_DS_TYPE(
+        f32(src1q[block].d),
+        f32(src1q[block].s)
+    );
+}
+fn get_dm(block_byte_base: u32) -> vec2<f32> {
+    return vec2<f32>(
+        f32(load_f16_at_src0(block_byte_base)),
+        f32(load_f16_at_src0(block_byte_base + 2u))
+    );
+}
+fn mul_q8_1(row_sum: i32, dma: vec2<f32>, b_ds: B_DS_TYPE) -> f32 {
+    return f32(row_sum) * (dma.x * b_ds.x) + dma.y * b_ds.y / THREADS_PER_BLOCK;
+}
+#endif
+
+#ifdef MUL_ACC_Q8_0
+#define BLOCK_SIZE_BYTES 34
+#define B_DS_TYPE f32
+fn repack_a(block_byte_base: u32, inner_id: u32) -> vec2<u32> {
+    return vec2<u32>(
+        load_u32_at_src0(block_byte_base + 2u + 4u * (inner_id * 2u)),
+        load_u32_at_src0(block_byte_base + 2u + 4u * (inner_id * 2u + 1))
+    );
+}
+fn repack_b_qs(block:u32, inner_id: u32) -> vec2<u32> {
+    return vec2<u32>(
+            src1q[block].qs[inner_id * 2u],
+            src1q[block].qs[inner_id * 2u + 1],
+        );
+}
+fn repack_b_dm(block: u32) -> B_DS_TYPE {
+    return B_DS_TYPE(src1q[block].d);
+}
+fn get_dm(block_byte_base: u32) -> f32 {
+    return f32(load_f16_at_src0(block_byte_base));
+}
+fn mul_q8_1(row_sum: i32, da: f32, b_ds: B_DS_TYPE) -> f32 {
+    return f32(row_sum) * (da * b_ds);
+}
+#endif
+
+#ifdef LEGACY_QUANTS
+fn mmvq_dot_product(a_byte_base: u32, b_inner_id: u32, b_repacked: vec2<u32>, b_ds: B_DS_TYPE) -> f32 {
+    var row_sum = 0;
+    let a_repacked = repack_a(a_byte_base, b_inner_id);
+
+    row_sum += dot4I8Packed(a_repacked[0], b_repacked[0]);
+    row_sum += dot4I8Packed(a_repacked[1], b_repacked[1]);
+
+    return mul_q8_1(row_sum, get_dm(a_byte_base), b_ds);
+}
+
+fn accumulate_vec_q_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1q_idx_base: u32) -> array<f32, OUTPUTS_PER_WG> {
+    var acc: array<f32, OUTPUTS_PER_WG>;
+
+    let num_blocks = params.k / BLOCK_SIZE;
+
+    for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
+        let b_inner_id = thread_id % THREADS_PER_BLOCK;
+        let b_block_idx = src1q_idx_base + block;
+
+        let b_repacked = repack_b_qs(b_block_idx, b_inner_id);
+        let b_ds = repack_b_dm(b_block_idx);
+
+        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
+            let output_row = row_base + row;
+            if (output_row < params.m) {
+                let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
+                acc[row] += mmvq_dot_product(block_byte_base, b_inner_id, b_repacked, b_ds);
+            }
+        }
+    }
+
+    return acc;
+}
+#endif
+
+#ifdef MUL_ACC_Q2_K
+#define BLOCK_SIZE_BYTES 84
+#define B_DS_TYPE f32
+fn repack_a(block_byte_base: u32, tid: u32) -> vec4<u32> {
+    let ih2 = tid / 8u;
+    let phase = tid % 2u;
+    let iq4_idx = 2u * ih2 + phase;
+    let qs_byte_base = block_byte_base + 16u + 16u * iq4_idx;
+    let qs_shift = tid & 6u;
+    return vec4<u32>(
+        (load_u32_at_src0_aligned(qs_byte_base) >> qs_shift) & 0x03030303u,
+        (load_u32_at_src0_aligned(qs_byte_base + 4u) >> qs_shift) & 0x03030303u,
+        (load_u32_at_src0_aligned(qs_byte_base + 8u) >> qs_shift) & 0x03030303u,
+        (load_u32_at_src0_aligned(qs_byte_base + 12u) >> qs_shift) & 0x03030303u,
+    );
+}
+fn repack_b_qs(q8_block_idx: u32, tid: u32) -> vec4<u32> {
+    let phase = tid % 2u;
+    return vec4<u32>(
+        src1q[q8_block_idx].qs[4u * phase],
+        src1q[q8_block_idx].qs[4u * phase + 1u],
+        src1q[q8_block_idx].qs[4u * phase + 2u],
+        src1q[q8_block_idx].qs[4u * phase + 3u],
+    );
+}
+fn repack_b_dm(q8_block_idx: u32) -> B_DS_TYPE {
+    return B_DS_TYPE(src1q[q8_block_idx].d);
+}
+fn get_dm(block_byte_base: u32) -> vec2<f32> {
+    return vec2<f32>(
+        f32(load_f16_at_src0(block_byte_base + 80u)),
+        f32(load_f16_at_src0(block_byte_base + 82u)),
+    );
+}
+fn get_scale_min(block_byte_base: u32, tid: u32) -> vec2<f32> {
+    let scale_byte = block_byte_base + tid;
+    let scale = byte_of(load_u32_at_src0_aligned(scale_byte), scale_byte & 3u);
+    return vec2<f32>(f32(scale & 0xFu), f32(scale >> 4u));
+}
+fn mmvq_dot_product(a_byte_base: u32, tid: u32, b_repacked: vec4<u32>, b_ds: B_DS_TYPE) -> f32 {
+    let a_repacked = repack_a(a_byte_base, tid);
+    let dm = get_dm(a_byte_base);
+    let scale_min = get_scale_min(a_byte_base, tid);
+
+    let scale_q = i32(scale_min.x);
+    let scale_m_i8x4 = u32(scale_min.y) * 0x01010101u;
+
+    let row_sum_d = (dot4I8Packed(b_repacked[0], a_repacked[0]) + dot4I8Packed(b_repacked[1], a_repacked[1])
+                   + dot4I8Packed(b_repacked[2], a_repacked[2]) + dot4I8Packed(b_repacked[3], a_repacked[3])) * scale_q;
+    let row_sum_m = dot4I8Packed(b_repacked[0], scale_m_i8x4) + dot4I8Packed(b_repacked[1], scale_m_i8x4)
+                  + dot4I8Packed(b_repacked[2], scale_m_i8x4) + dot4I8Packed(b_repacked[3], scale_m_i8x4);
+
+    return b_ds * (dm.x * f32(row_sum_d) - dm.y * f32(row_sum_m));
+}
+#endif
+
+#ifdef MUL_ACC_Q4_K
+#define BLOCK_SIZE_BYTES 144
+#define B_DS_TYPE vec2<f32>
+fn repack_a(block_byte_base: u32, tid: u32) -> vec4<u32> {
+    let iq4 = tid / 4u;
+    let phase = tid % 2u;
+    let nibble = (tid >> 1u) % 2u;
+    let q_qs_byte_base = block_byte_base + 16u + 32u * iq4 + 16u * phase;
+    let qs_shift = 4u * nibble;
+    return vec4<u32>(
+        (load_u32_at_src0_aligned(q_qs_byte_base) >> qs_shift) & 0x0F0F0F0Fu,
+        (load_u32_at_src0_aligned(q_qs_byte_base + 4u) >> qs_shift) & 0x0F0F0F0Fu,
+        (load_u32_at_src0_aligned(q_qs_byte_base + 8u) >> qs_shift) & 0x0F0F0F0Fu,
+        (load_u32_at_src0_aligned(q_qs_byte_base + 12u) >> qs_shift) & 0x0F0F0F0Fu,
+    );
+}
+fn repack_b_qs(q8_block_idx: u32, tid: u32) -> vec4<u32> {
+    let phase = tid % 2u;
+    return vec4<u32>(
+        src1q[q8_block_idx].qs[4u * phase],
+        src1q[q8_block_idx].qs[4u * phase + 1u],
+        src1q[q8_block_idx].qs[4u * phase + 2u],
+        src1q[q8_block_idx].qs[4u * phase + 3u],
+    );
+}
+fn repack_b_dm(q8_block_idx: u32) -> B_DS_TYPE {
+    return B_DS_TYPE(
+        f32(src1q[q8_block_idx].d),
+        f32(src1q[q8_block_idx].s),
+    );
+}
+fn get_dm(block_byte_base: u32) -> vec2<f32> {
+    return vec2<f32>(
+        f32(load_f16_at_src0(block_byte_base + 0u)),
+        f32(load_f16_at_src0(block_byte_base + 2u)),
+    );
+}
+fn get_scale_min(block_byte_base: u32, tid: u32) -> vec2<f32> {
+    let sc_m_idx = tid / 2u;
+    let scales_byte_base = block_byte_base + 4u;
+    let scales0_3  = load_u32_at_src0_aligned(scales_byte_base);
+    let scales4_7  = load_u32_at_src0_aligned(scales_byte_base + 4u);
+    let scales8_11 = load_u32_at_src0_aligned(scales_byte_base + 8u);
+
+    let byte_idx = sc_m_idx & 3u;
+    let is_high = sc_m_idx >= 4u;
+
+    let sc_low  = byte_of(scales0_3, byte_idx) & 0x3Fu;
+    let sc_high = (byte_of(scales8_11, byte_idx) & 0x0Fu) | ((byte_of(scales0_3, byte_idx) & 0xC0u) >> 2u);
+    let scale = f32(select(sc_low, sc_high, is_high));
+
+    let mn_low  = byte_of(scales4_7, byte_idx) & 0x3Fu;
+    let mn_high = (byte_of(scales8_11, byte_idx) >> 4u) | ((byte_of(scales4_7, byte_idx) & 0xC0u) >> 2u);
+    let min_val = f32(select(mn_low, mn_high, is_high));
+
+    return vec2<f32>(scale, min_val);
+}
+fn mmvq_dot_product(a_byte_base: u32, tid: u32, b_repacked: vec4<u32>, b_ds: B_DS_TYPE) -> f32 {
+    let a_repacked = repack_a(a_byte_base, tid);
+    let dm = get_dm(a_byte_base);
+    let scale_min = get_scale_min(a_byte_base, tid);
+
+    let row_sum = dot4I8Packed(a_repacked[0], b_repacked[0]) + dot4I8Packed(a_repacked[1], b_repacked[1])
+                + dot4I8Packed(a_repacked[2], b_repacked[2]) + dot4I8Packed(a_repacked[3], b_repacked[3]);
+
+    // Each thread covers half of the Q8_1 block, so add only b_ds.y/2.
+    return b_ds.x * dm.x * scale_min.x * f32(row_sum) - dm.y * scale_min.y * (b_ds.y / (Q8_BLOCK_SIZE / ELEMS_PER_THREAD));
+}
+#endif
+
+#ifdef K_QUANTS
+fn accumulate_vec_q_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1q_idx_base: u32) -> array<f32, OUTPUTS_PER_WG> {
+    var acc: array<f32, OUTPUTS_PER_WG>;
+
+    let tid = thread_id % THREADS_PER_BLOCK;
+
+    for (var block = thread_id / THREADS_PER_BLOCK; block < params.k / BLOCK_SIZE; block += WG_SIZE / THREADS_PER_BLOCK) {
+        let src1q_idx = src1q_idx_base + (block * BLOCK_SIZE + ELEMS_PER_THREAD * tid) / Q8_BLOCK_SIZE;
+        let b_repacked = repack_b_qs(src1q_idx, tid);
+        let b_ds = repack_b_dm(src1q_idx);
+
+        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
+            let output_row = row_base + row;
+            if (output_row < params.m) {
+                let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
+                acc[row] += mmvq_dot_product(block_byte_base, tid, b_repacked, b_ds);
+            }
+        }
+    }
+
+    return acc;
+}
+#endif

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl

@@ -0,0 +1,21 @@
+#ifdef U32_DEQUANT_HELPERS
+fn dequant_q4_0_packed_to_shmem(q_packed: u32, d: f16, dst_idx: u32) {
+    let scale = QUANT_OUT_TYPE(d);
+    for (var k = 0u; k < 4u; k++) {
+        let q_byte = get_byte(q_packed, k);
+        let q_hi = (QUANT_OUT_TYPE((q_byte >> 4) & 0xFu) - QUANT_OUT_TYPE(8.0)) * scale;
+        let q_lo = (QUANT_OUT_TYPE(q_byte & 0xFu) - QUANT_OUT_TYPE(8.0)) * scale;
+        QUANT_SHMEM[dst_idx + k] = q_lo;
+        QUANT_SHMEM[dst_idx + k + 16u] = q_hi;
+    }
+}
+
+fn dequant_q8_0_packed_to_shmem(q_packed: u32, d: f16, dst_idx: u32) {
+    let scale = QUANT_OUT_TYPE(d);
+    for (var k = 0u; k < 4u; k++) {
+        let q_byte = get_byte_i32(q_packed, k);
+        let q_val = QUANT_OUT_TYPE(q_byte) * scale;
+        QUANT_SHMEM[dst_idx + k] = q_val;
+    }
+}
+#endif

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl

@@ -0,0 +1,173 @@
+#ifdef USE_SUBGROUP_REDUCTION
+enable subgroups;
+#endif
+enable f16;
+
+requires packed_4x8_integer_dot_product;
+
+#include "common_decls.tmpl"
+
+struct Params {
+    offset_src1: u32,
+    stride_12: u32,
+    stride_13: u32,
+    ne0: u32,
+    ne2: u32,
+    ne3: u32,
+};
+
+#define SRC1_TYPE vec4<SRC1_INNER_TYPE>
+
+@group(0) @binding(0) var<storage, read_write> src1: array<SRC1_TYPE>;
+@group(0) @binding(1) var<storage, read_write> src1q: array<q8_1>;
+
+@group(0) @binding(2) var<uniform> params: Params;
+
+#ifdef USE_SUBGROUP_REDUCTION
+fn cluster_max_8(v: f32) -> f32 {
+    var r = v;
+    r = max(r, subgroupShuffleXor(r, 1u));
+    r = max(r, subgroupShuffleXor(r, 2u));
+    r = max(r, subgroupShuffleXor(r, 4u));
+    return r;
+}
+
+#if defined(MUL_ACC_Q4_0) || defined(MUL_ACC_Q4_1) || defined(MUL_ACC_Q4_K)
+fn cluster_add_i4x8(v: i32) -> i32 {
+    var r= v;
+    r += subgroupShuffleXor(r, 1u);
+    r += subgroupShuffleXor(r, 2u);
+    r += subgroupShuffleXor(r, 4u);
+    return r;
+}
+#endif
+#endif
+
+#ifdef USE_WORKGROUP_REDUCTION
+#define CLUSTER_SIZE 8
+
+var<workgroup> partial_amaxs: array<array<f32, CLUSTER_SIZE>, WG_SIZE / CLUSTER_SIZE>;
+var<workgroup> partial_sums:  array<array<i32, CLUSTER_SIZE>, WG_SIZE / CLUSTER_SIZE>;
+#endif
+
+@compute @workgroup_size(WG_SIZE)
+fn main(
+    @builtin(local_invocation_id) local_id: vec3<u32>,
+    @builtin(workgroup_id) wg_id: vec3<u32>,
+    @builtin(num_workgroups) num_wg: vec3<u32>
+) {
+    let thread_id = local_id.x;
+    let num_vec4 = params.ne0 / 4u;
+
+    let wg_per_vec = (num_vec4 + (WG_SIZE - 1u)) / WG_SIZE;
+    let total_batches = wg_per_vec * params.ne2 * params.ne3;
+
+    let wg_linear = wg_id.y * num_wg.x + wg_id.x;
+    if (wg_linear >= total_batches) {
+        return;
+    }
+
+    let src13_idx = wg_linear / (params.ne2 * wg_per_vec);
+    let src12_idx = (wg_linear - src13_idx * (params.ne2 * wg_per_vec)) / wg_per_vec;
+    let src11_wg_idx = wg_linear % wg_per_vec;
+    let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
+    let src1_idx_vec4_base = src1_idx_base / 4u;
+
+    let blocks_per_row = params.ne0 / 32u;
+    let blocks_per_wg = (WG_SIZE * 4u) / 32u;
+    let src1q_idx_base = (src13_idx * params.ne2 + src12_idx) * blocks_per_row;
+    let src1q_idx = src1q_idx_base + src11_wg_idx * blocks_per_wg + thread_id / 8u;
+    let qs_idx = thread_id % 8u;
+
+    // reduction
+    var q4 = vec4<f32>(0.0);
+    var q4_quants = 0u;
+    var thread_amax = 0.0;
+
+    let src11_vec4_idx = src11_wg_idx * WG_SIZE + thread_id;
+    let is_valid = src11_vec4_idx < num_vec4;
+
+#ifdef USE_SUBGROUP_REDUCTION
+
+    var d = 0.0;
+
+    if (is_valid) {
+        q4 = src1[src1_idx_vec4_base + src11_vec4_idx];
+        let abs_q4 = abs(q4);
+        thread_amax = max(max(abs_q4[0u], abs_q4[1u]), max(abs_q4[2], abs_q4[3]));
+    }
+
+    d = cluster_max_8(thread_amax) / 127.0;
+
+    if (is_valid) {
+        let id = select(0.0, 1.0 / d, d > 0.0);
+        q4_quants = pack4xI8(vec4<i32>(round(q4 * id)));
+        if (qs_idx == 0u) {
+            src1q[src1q_idx].d = f16(d);
+        }
+        src1q[src1q_idx].qs[qs_idx] = q4_quants;
+    }
+
+#if defined(MUL_ACC_Q4_0) || defined(MUL_ACC_Q4_1) || defined(MUL_ACC_Q4_K)
+    let q4_quants_sum = dot4I8Packed(q4_quants, 0x01010101u);
+    let s = f16(d * f32(cluster_add_i4x8(q4_quants_sum)));
+
+    if (is_valid) {
+        if (qs_idx == 0u) {
+            src1q[src1q_idx].s = s;
+        }
+    }
+#endif
+#endif
+
+#ifdef USE_WORKGROUP_REDUCTION
+
+    var d = 0.0;
+    let cluster_id = thread_id / 8u;
+
+    if (is_valid) {
+        q4 = src1[src1_idx_vec4_base + src11_vec4_idx];
+        let abs_q4 = abs(q4);
+        thread_amax = max(max(abs_q4[0], abs_q4[1]), max(abs_q4[2], abs_q4[3]));
+        partial_amaxs[cluster_id][qs_idx] = thread_amax;
+    }
+
+    workgroupBarrier();
+
+    if (is_valid) {
+        let amax = max(
+                    max(
+                        max(partial_amaxs[cluster_id][0], partial_amaxs[cluster_id][1]), max(partial_amaxs[cluster_id][2], partial_amaxs[cluster_id][3])),
+                    max(
+                        max(partial_amaxs[cluster_id][4], partial_amaxs[cluster_id][5]), max(partial_amaxs[cluster_id][6], partial_amaxs[cluster_id][7]))
+                );
+
+        d = amax / 127.0;
+        let id = select(0.0f, 1.0f / d, d > 0.0f);
+
+        q4_quants = pack4xI8(vec4<i32>(round(q4 * id)));
+        src1q[src1q_idx].qs[qs_idx] = q4_quants;
+
+        if (qs_idx == 0u) {
+            src1q[src1q_idx].d = f16(d);
+        }
+    }
+
+#if defined(MUL_ACC_Q4_0) || defined(MUL_ACC_Q4_1) || defined(MUL_ACC_Q4_K)
+
+    partial_sums[cluster_id][qs_idx] = dot4I8Packed(q4_quants, 0x01010101u);
+
+    workgroupBarrier();
+
+    if (is_valid) {
+        if (qs_idx == 0u) {
+            let s = d * f32(partial_sums[cluster_id][0] + partial_sums[cluster_id][1] + partial_sums[cluster_id][2] + partial_sums[cluster_id][3]
+                                    + partial_sums[cluster_id][4] + partial_sums[cluster_id][5] + partial_sums[cluster_id][6] + partial_sums[cluster_id][7]);
+            src1q[src1q_idx].s = f16(s);
+        }
+    }
+
+#endif
+#endif
+
+}

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl

@@ -0,0 +1,152 @@
+#ifdef OVERLAP
+
+@group(0) @binding(0)
+var<storage, read_write> rn_src: array<f32>;
+
+@group(0) @binding(1)
+var<storage, read_write> mul_src: array<f32>;
+
+@group(0) @binding(2)
+var<uniform> params: Params;
+
+fn update(rn_src_offset: u32, dst_offset: u32, scale: f32, mul_src_offset: u32) {
+    mul_src[dst_offset] = scale * rn_src[rn_src_offset] * mul_src[mul_src_offset];
+}
+
+#elif INPLACE
+
+@group(0) @binding(0)
+var<storage, read_write> rn_src: array<f32>;
+
+@group(0) @binding(1)
+var<storage, read_write> mul_src: array<f32>;
+
+@group(0) @binding(2)
+var<uniform> params: Params;
+
+fn update(rn_src_offset: u32, dst_offset: u32, scale: f32, mul_src_offset: u32) {
+    rn_src[dst_offset] = scale * rn_src[rn_src_offset] * mul_src[mul_src_offset];
+}
+
+#elif SRC_OVERLAP
+
+@group(0) @binding(0)
+var<storage, read_write> merged_src: array<f32>;
+
+@group(0) @binding(1)
+var<storage, read_write> dst: array<f32>;
+
+@group(0) @binding(2)
+var<uniform> params: Params;
+
+fn update(rn_src_offset: u32, dst_offset: u32, scale: f32, mul_src_offset: u32) {
+    dst[dst_offset] = scale * merged_src[rn_src_offset] * merged_src[mul_src_offset];
+}
+
+#else
+
+@group(0) @binding(0)
+var<storage, read_write> rn_src: array<f32>;
+
+@group(0) @binding(1)
+var<storage, read_write> mul_src: array<f32>;
+
+@group(0) @binding(2)
+var<storage, read_write> dst: array<f32>;
+
+@group(0) @binding(3)
+var<uniform> params: Params;
+
+fn update(rn_src_offset: u32, dst_offset: u32, scale: f32, mul_src_offset: u32) {
+    dst[dst_offset] = scale * rn_src[rn_src_offset] * mul_src[mul_src_offset];
+}
+
+#endif
+
+struct Params {
+    offset_rn_src: u32,
+    offset_mul_src: u32,
+    offset_dst: u32,
+
+    stride_rn_src1: u32,
+    stride_rn_src2: u32,
+    stride_rn_src3: u32,
+
+    stride_mul_src1: u32,
+    stride_mul_src2: u32,
+    stride_mul_src3: u32,
+
+    stride_dst1: u32,
+    stride_dst2: u32,
+    stride_dst3: u32,
+
+    mul_src_ne0: u32,
+    mul_src_ne1: u32,
+    mul_src_ne2: u32,
+    mul_src_ne3: u32,
+
+    ne0: u32,
+    ne1: u32,
+    ne2: u32,
+    ne3: u32,
+
+    eps: f32
+};
+
+var<workgroup> scratch: array<f32, WG_SIZE>;
+
+@compute @workgroup_size(WG_SIZE)
+fn main(@builtin(workgroup_id) wid: vec3<u32>,
+        @builtin(local_invocation_id) lid: vec3<u32>) {
+
+    // one thread per row
+    var i = wid.x;
+    let i3 = i / (params.ne2 * params.ne1);
+    i = i % (params.ne2 * params.ne1);
+    let i2 = i / params.ne1;
+    let i1 = i % params.ne1;
+    let i_rn_src_row = params.offset_rn_src + i3 * params.stride_rn_src3 + i2 * params.stride_rn_src2 + i1 * params.stride_rn_src1;
+    let i_mul_src_row = params.offset_mul_src + (i3 % params.mul_src_ne3) * params.stride_mul_src3 + (i2 % params.mul_src_ne2) * params.stride_mul_src2 + (i1 % params.mul_src_ne1) * params.stride_mul_src1;
+    let i_dst_row = params.offset_dst + i3 * params.stride_dst3 + i2 * params.stride_dst2 + i1 * params.stride_dst1;
+
+    let elems = (params.ne0 + WG_SIZE - 1) / WG_SIZE;
+
+    var sum = 0.0f;
+    var col = lid.x;
+    for (var j: u32 = 0; j < elems; j++) {
+        if (col >= params.ne0) {
+            break;
+        }
+#ifdef SRC_OVERLAP
+        sum += pow(merged_src[i_rn_src_row + col], 2.0);
+#else
+        sum += pow(rn_src[i_rn_src_row + col], 2.0);
+#endif
+        col += WG_SIZE;
+    }
+
+    scratch[lid.x] = sum;
+
+    workgroupBarrier();
+
+    var offset: u32 = WG_SIZE / 2;
+    while (offset > 0) {
+        if (lid.x < offset) {
+            scratch[lid.x] += scratch[lid.x + offset];
+        }
+        offset = offset / 2;
+        workgroupBarrier();
+    }
+    sum = scratch[0];
+
+    let scale = 1.0/sqrt(sum/f32(params.ne0) + params.eps);
+
+    col = lid.x;
+    for (var j: u32 = 0; j < elems; j++) {
+        if (col >= params.ne0) {
+            break;
+        }
+        update(i_rn_src_row + col, i_dst_row + col, scale, i_mul_src_row + col % params.mul_src_ne0);
+        col += WG_SIZE;
+    }
+}