ggml-webgpu: enable FLASH_ATTN_EXT on browser without subgroup matrix (ggml-org#22199)

* ggml-webgpu: add tile flash attention fallback

* ggml-webgpu: add new fields and discard usage of mnk for tile version

* ggml-webgpu: modify the vec path to discard the mnk parameter

* ggml-webgpu: enable flash attention vec and tile version for broswer

* ggml-webgpu: stagging KV for flash attention tile version

* formatting

* turn on subgroup uniformity check

* remove Q_TILE as it is always 1 for vec path

* make row_max and exp_sum to local register

* make different bindings with same underlying buffer to have the same usage flags

* move path selection into the shader library and have the host consume a single flash-attn decision object.

* turn off skip_validation and address buffer overlapping when nwg==1

* formatting

* merge binding when kv overlap

Author: ArberSephirotheca

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -138,26 +138,55 @@ struct Params {
 };
 
 @group(0) @binding(0) var<storage, read_write> Q: array<f32>;
+#ifdef KV_OVERLAP
+@group(0) @binding(1) var<storage, read_write> K: array<KV_TYPE>;
+#define V K
+#else
 @group(0) @binding(1) var<storage, read_write> K: array<KV_TYPE>;
 @group(0) @binding(2) var<storage, read_write> V: array<KV_TYPE>;
+#endif
 
 #if defined(MASK) && defined(SINKS)
+#ifdef KV_OVERLAP
+@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
+@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
+#define DST_BINDING 4
+#define PARAMS_BINDING 5
+#else
 @group(0) @binding(3) var<storage, read_write> mask: array<f16>;
 @group(0) @binding(4) var<storage, read_write> sinks: array<f32>;
 #define DST_BINDING 5
 #define PARAMS_BINDING 6
+#endif
 #elif defined(MASK)
+#ifdef KV_OVERLAP
+@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
+#define DST_BINDING 3
+#define PARAMS_BINDING 4
+#else
 @group(0) @binding(3) var<storage, read_write> mask: array<f16>;
 #define DST_BINDING 4
 #define PARAMS_BINDING 5
+#endif
 #elif defined(SINKS)
+#ifdef KV_OVERLAP
+@group(0) @binding(2) var<storage, read_write> sinks: array<f32>;
+#define DST_BINDING 3
+#define PARAMS_BINDING 4
+#else
 @group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
 #define DST_BINDING 4
 #define PARAMS_BINDING 5
+#endif
+#else
+#ifdef KV_OVERLAP
+#define DST_BINDING 2
+#define PARAMS_BINDING 3
 #else
 #define DST_BINDING 3
 #define PARAMS_BINDING 4
 #endif
+#endif
 
 @group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<vec4<f32>>;
 @group(0) @binding(PARAMS_BINDING) var<uniform> params: Params;

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -0,0 +1,330 @@
+enable f16;
+enable subgroups;
+
+#define HEAD_DIM_QK 64
+#define HEAD_DIM_V 64
+#define KV_STAGE_STRIDE 64
+#define Q_TILE 4
+#define KV_TILE 64
+#define WG_SIZE 128
+
+struct Params {
+    offset_q: u32,
+    offset_k: u32,
+    offset_v: u32,
+    offset_mask: u32,
+    offset_sinks: u32,
+    offset_dst: u32,
+
+    n_heads: u32,
+    seq_len_q: u32,
+    seq_len_kv: u32,
+
+    stride_q1: u32,
+    stride_q2: u32,
+    stride_q3: u32,
+    stride_k1: u32,
+    stride_k2: u32,
+    stride_k3: u32,
+    stride_v1: u32,
+    stride_v2: u32,
+    stride_v3: u32,
+    stride_mask3: u32,
+
+    q_per_kv: u32,
+
+    scale: f32,
+    max_bias: f32,
+    logit_softcap: f32,
+    n_head_log2: f32,
+    m0: f32,
+    m1: f32,
+};
+
+@group(0) @binding(0) var<storage, read_write> Q: array<f32>;
+#ifdef KV_OVERLAP
+@group(0) @binding(1) var<storage, read_write> K: array<vec4<f16>>;
+#define V K
+#else
+@group(0) @binding(1) var<storage, read_write> K: array<vec4<f16>>;
+@group(0) @binding(2) var<storage, read_write> V: array<vec4<f16>>;
+#endif
+
+#if defined(MASK) && defined(SINKS)
+#ifdef KV_OVERLAP
+@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
+@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
+#define DST_BINDING 4
+#define PARAMS_BINDING 5
+#else
+@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
+@group(0) @binding(4) var<storage, read_write> sinks: array<f32>;
+#define DST_BINDING 5
+#define PARAMS_BINDING 6
+#endif
+#elif defined(MASK)
+#ifdef KV_OVERLAP
+@group(0) @binding(2) var<storage, read_write> mask: array<f16>;
+#define DST_BINDING 3
+#define PARAMS_BINDING 4
+#else
+@group(0) @binding(3) var<storage, read_write> mask: array<f16>;
+#define DST_BINDING 4
+#define PARAMS_BINDING 5
+#endif
+#elif defined(SINKS)
+#ifdef KV_OVERLAP
+@group(0) @binding(2) var<storage, read_write> sinks: array<f32>;
+#define DST_BINDING 3
+#define PARAMS_BINDING 4
+#else
+@group(0) @binding(3) var<storage, read_write> sinks: array<f32>;
+#define DST_BINDING 4
+#define PARAMS_BINDING 5
+#endif
+#else
+#ifdef KV_OVERLAP
+#define DST_BINDING 2
+#define PARAMS_BINDING 3
+#else
+#define DST_BINDING 3
+#define PARAMS_BINDING 4
+#endif
+#endif
+
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<vec4<f32>>;
+@group(0) @binding(PARAMS_BINDING) var<uniform> params: Params;
+
+const FLOAT_MIN: f32 = -1.0e9;
+const Q_CHUNKS: u32 = HEAD_DIM_QK / 4u;
+const V_CHUNKS: u32 = HEAD_DIM_V / 4u;
+const SCORE_REGS_PER_LANE: u32 = (KV_TILE + MAX_SUBGROUP_SIZE - 1u) / MAX_SUBGROUP_SIZE;
+const OUT_REGS_PER_LANE: u32 = (V_CHUNKS + MAX_SUBGROUP_SIZE - 1u) / MAX_SUBGROUP_SIZE;
+
+var<workgroup> q_shmem: array<f16, Q_TILE * HEAD_DIM_QK>;
+var<workgroup> kv_shmem: array<f16, KV_TILE * KV_STAGE_STRIDE>;
+var<workgroup> p_shmem: array<f32, Q_TILE * KV_TILE>;
+
+@compute @workgroup_size(WG_SIZE)
+fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
+        @builtin(local_invocation_id) local_id: vec3<u32>,
+        @builtin(subgroup_id) subgroup_id: u32,
+        @builtin(subgroup_size) subgroup_size: u32,
+        @builtin(num_subgroups) num_subgroups: u32,
+        @builtin(subgroup_invocation_id) sg_inv_id: u32) {
+    if (subgroup_size == 0u || num_subgroups < Q_TILE) {
+        return;
+    }
+
+    let wg_per_head = (params.seq_len_q + Q_TILE - 1u) / Q_TILE;
+    let wg_per_batch = wg_per_head * params.n_heads;
+
+    let dst2_stride = HEAD_DIM_V * params.n_heads;
+    let dst3_stride = dst2_stride * params.seq_len_q;
+
+    let batch_idx = wg_id.x / wg_per_batch;
+    let q_batch_offset = params.offset_q + batch_idx * params.stride_q3;
+    let k_batch_offset = params.offset_k + batch_idx * params.stride_k3;
+    let v_batch_offset = params.offset_v + batch_idx * params.stride_v3;
+    let dst_batch_offset = params.offset_dst + batch_idx * dst3_stride;
+    let wg_in_batch = wg_id.x % wg_per_batch;
+
+    let head_idx = wg_in_batch / wg_per_head;
+    let q_head_offset = q_batch_offset + head_idx * params.stride_q2;
+    let k_head_idx = head_idx / params.q_per_kv;
+    let v_head_offset = v_batch_offset + k_head_idx * params.stride_v2;
+    let k_head_offset = k_batch_offset + k_head_idx * params.stride_k2;
+
+    let wg_in_head = wg_in_batch % wg_per_head;
+    let q_row_start = wg_in_head * Q_TILE;
+    let global_q_row = q_row_start + subgroup_id;
+    let row_active = subgroup_id < Q_TILE && global_q_row < params.seq_len_q;
+
+#ifdef MASK
+    let mask_global_offset = params.offset_mask + batch_idx * params.stride_mask3 + q_row_start * params.seq_len_kv;
+#endif
+
+    let dst_global_offset = dst_batch_offset + q_row_start * dst2_stride + head_idx * HEAD_DIM_V;
+
+    let head = f32(head_idx);
+    let slope = select(1.0,
+                       select(pow(params.m1, 2.0 * (head - params.n_head_log2) + 1.0),
+                              pow(params.m0, head + 1.0),
+                              head < params.n_head_log2),
+                       params.max_bias > 0.0);
+
+    for (var elem_idx = local_id.x; elem_idx < Q_TILE * HEAD_DIM_QK; elem_idx += WG_SIZE) {
+        let q_tile_row = elem_idx / HEAD_DIM_QK;
+        let q_col = elem_idx % HEAD_DIM_QK;
+        let head_q_row = q_row_start + q_tile_row;
+        let global_q_row_offset = q_head_offset + head_q_row * params.stride_q1;
+        q_shmem[elem_idx] = f16(select(
+            0.0,
+            Q[global_q_row_offset + q_col] * params.scale,
+            head_q_row < params.seq_len_q));
+    }
+
+    workgroupBarrier();
+
+    var row_max = FLOAT_MIN;
+    var exp_sum = 0.0;
+    var out_regs: array<vec4<f32>, OUT_REGS_PER_LANE>;
+    for (var reg_idx = 0u; reg_idx < OUT_REGS_PER_LANE; reg_idx += 1u) {
+        out_regs[reg_idx] = vec4<f32>(0.0);
+    }
+
+    let q_base = subgroup_id * HEAD_DIM_QK;
+    let subgroup_p_offset = subgroup_id * KV_TILE;
+
+    for (var kv_tile = 0u; kv_tile < params.seq_len_kv; kv_tile += KV_TILE) {
+        let kv_count = min(KV_TILE, params.seq_len_kv - kv_tile);
+        let score_slots = min(SCORE_REGS_PER_LANE, (kv_count + subgroup_size - 1u) / subgroup_size);
+        let out_slots = min(OUT_REGS_PER_LANE, (V_CHUNKS + subgroup_size - 1u) / subgroup_size);
+        var local_scores: array<f32, SCORE_REGS_PER_LANE>;
+        for (var slot = 0u; slot < SCORE_REGS_PER_LANE; slot += 1u) {
+            local_scores[slot] = FLOAT_MIN;
+        }
+
+        for (var vec_idx_local = local_id.x; vec_idx_local < kv_count * Q_CHUNKS; vec_idx_local += WG_SIZE) {
+            let kv_local = vec_idx_local / Q_CHUNKS;
+            let chunk = vec_idx_local % Q_CHUNKS;
+            let global_k_row = kv_tile + kv_local;
+            let k_vec_index = (k_head_offset + global_k_row * params.stride_k1 + chunk * 4u) >> 2u;
+            let k4 = K[k_vec_index];
+            let kv_off = kv_local * KV_STAGE_STRIDE + chunk * 4u;
+            kv_shmem[kv_off + 0u] = k4.x;
+            kv_shmem[kv_off + 1u] = k4.y;
+            kv_shmem[kv_off + 2u] = k4.z;
+            kv_shmem[kv_off + 3u] = k4.w;
+        }
+
+        workgroupBarrier();
+
+        var local_max = FLOAT_MIN;
+        if (row_active) {
+            for (var slot = 0u; slot < score_slots; slot += 1u) {
+                let kv_local = sg_inv_id + slot * subgroup_size;
+                if (kv_local >= kv_count) {
+                    continue;
+                }
+
+                let global_k_row = kv_tile + kv_local;
+                var dot_val = 0.0;
+                for (var chunk = 0u; chunk < Q_CHUNKS; chunk += 1u) {
+                    let q_off = q_base + chunk * 4u;
+                    let qv = vec4<f32>(
+                        f32(q_shmem[q_off + 0u]),
+                        f32(q_shmem[q_off + 1u]),
+                        f32(q_shmem[q_off + 2u]),
+                        f32(q_shmem[q_off + 3u]));
+                    let kv_off = kv_local * KV_STAGE_STRIDE + chunk * 4u;
+                    let kv = vec4<f32>(
+                        f32(kv_shmem[kv_off + 0u]),
+                        f32(kv_shmem[kv_off + 1u]),
+                        f32(kv_shmem[kv_off + 2u]),
+                        f32(kv_shmem[kv_off + 3u]));
+                    dot_val += dot(qv, kv);
+                }
+#ifdef LOGIT_SOFTCAP
+                dot_val = params.logit_softcap * tanh(dot_val);
+#endif
+#ifdef MASK
+                let mask_idx = mask_global_offset + subgroup_id * params.seq_len_kv + global_k_row;
+                dot_val += slope * f32(mask[mask_idx]);
+#endif
+                local_scores[slot] = dot_val;
+                local_max = max(local_max, dot_val);
+            }
+        }
+
+        let tile_max = subgroupMax(local_max);
+        let new_max = max(row_max, tile_max);
+        let cur_exp = exp(row_max - new_max);
+        exp_sum *= cur_exp;
+        for (var reg_idx = 0u; reg_idx < OUT_REGS_PER_LANE; reg_idx += 1u) {
+            out_regs[reg_idx] *= cur_exp;
+        }
+
+        var local_sum = 0.0;
+        for (var slot = 0u; slot < score_slots; slot += 1u) {
+            let kv_local = sg_inv_id + slot * subgroup_size;
+            if (row_active && kv_local < kv_count) {
+                let p = exp(local_scores[slot] - new_max);
+                p_shmem[subgroup_p_offset + kv_local] = p;
+                local_sum += p;
+            }
+        }
+
+        workgroupBarrier();
+
+        for (var vec_idx_local = local_id.x; vec_idx_local < kv_count * V_CHUNKS; vec_idx_local += WG_SIZE) {
+            let kv_local = vec_idx_local / V_CHUNKS;
+            let chunk = vec_idx_local % V_CHUNKS;
+            let global_v_row = kv_tile + kv_local;
+            let v_vec_index = (v_head_offset + global_v_row * params.stride_v1 + chunk * 4u) >> 2u;
+            let v4 = V[v_vec_index];
+            let kv_off = kv_local * KV_STAGE_STRIDE + chunk * 4u;
+            kv_shmem[kv_off + 0u] = v4.x;
+            kv_shmem[kv_off + 1u] = v4.y;
+            kv_shmem[kv_off + 2u] = v4.z;
+            kv_shmem[kv_off + 3u] = v4.w;
+        }
+
+        workgroupBarrier();
+
+        let tile_sum = subgroupAdd(local_sum);
+        exp_sum += tile_sum;
+        row_max = new_max;
+
+        if (row_active) {
+            for (var reg_idx = 0u; reg_idx < out_slots; reg_idx += 1u) {
+                let chunk = sg_inv_id + reg_idx * subgroup_size;
+                if (chunk >= V_CHUNKS) {
+                    continue;
+                }
+
+                var acc = out_regs[reg_idx];
+                for (var kv_local = 0u; kv_local < kv_count; kv_local += 1u) {
+                    let p = p_shmem[subgroup_p_offset + kv_local];
+                    let kv_off = kv_local * KV_STAGE_STRIDE + chunk * 4u;
+                    let v4 = vec4<f32>(
+                        f32(kv_shmem[kv_off + 0u]),
+                        f32(kv_shmem[kv_off + 1u]),
+                        f32(kv_shmem[kv_off + 2u]),
+                        f32(kv_shmem[kv_off + 3u]));
+                    acc += p * v4;
+                }
+                out_regs[reg_idx] = acc;
+            }
+        }
+
+        workgroupBarrier();
+    }
+
+#ifdef SINKS
+    if (row_active) {
+        let sink_score = sinks[params.offset_sinks + head_idx];
+        let sink_max = max(row_max, sink_score);
+        let sink_scale = exp(row_max - sink_max);
+        for (var reg_idx = 0u; reg_idx < OUT_REGS_PER_LANE; reg_idx += 1u) {
+            out_regs[reg_idx] *= sink_scale;
+        }
+        exp_sum = exp_sum * sink_scale + exp(sink_score - sink_max);
+        row_max = sink_max;
+    }
+#endif
+
+    if (row_active) {
+        let inv_exp_sum = select(0.0, 1.0 / exp_sum, exp_sum != 0.0);
+        let row_base = dst_global_offset + subgroup_id * dst2_stride;
+        let out_slots = min(OUT_REGS_PER_LANE, (V_CHUNKS + subgroup_size - 1u) / subgroup_size);
+        for (var reg_idx = 0u; reg_idx < out_slots; reg_idx += 1u) {
+            let chunk = sg_inv_id + reg_idx * subgroup_size;
+            if (chunk >= V_CHUNKS) {
+                continue;
+            }
+            let dst_vec_index = (row_base + chunk * 4u) >> 2u;
+            dst[dst_vec_index] = out_regs[reg_idx] * inv_exp_sum;
+        }
+    }
+}

ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl

@@ -15,7 +15,7 @@ struct Params {
     nblk1: u32,
 };
 
-@group(0) @binding(0) var<storage, read> mask: array<f16>;
+@group(0) @binding(0) var<storage, read_write> mask: array<f16>;
 @group(0) @binding(1) var<storage, read_write> blk: array<u32>;
 @group(0) @binding(2) var<uniform> params: Params;