vulkan: support softmax/FA batch and broadcast (ggml-org#14449)

Author: jeffbolznv

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -633,6 +633,7 @@ struct vk_flash_attn_push_constants {
     uint32_t nev2;
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nem2;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -643,7 +644,6 @@ struct vk_flash_attn_push_constants {
     uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
-    uint32_t nb31;
 
     float scale;
     float max_bias;
@@ -658,6 +658,7 @@ struct vk_flash_attn_push_constants {
     uint32_t split_kv;
     uint32_t k_num;
 };
+static_assert(sizeof(vk_flash_attn_push_constants) <= 128, "sizeof(vk_flash_attn_push_constants) must be <= 128");
 
 struct vk_op_push_constants {
     uint32_t KX;
@@ -756,6 +757,14 @@ struct vk_op_rope_push_constants {
 struct vk_op_soft_max_push_constants {
     uint32_t KX;
     uint32_t KY;
+    uint32_t ne00;
+    uint32_t ne01;
+    uint32_t ne02;
+    uint32_t ne12;
+    uint32_t ne13;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
     float scale;
     float max_bias;
     float m0;
@@ -6040,7 +6049,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
     const uint32_t nem1 = mask ? mask->ne[1] : 0;
-    const uint32_t nbm1 = mask ? mask->nb[1] : 0;
+    const uint32_t nem2 = mask ? mask->ne[2] : 0;
 
     const uint32_t D = neq0;
     uint32_t N = neq1;
@@ -6203,7 +6212,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     // Try to use split_k when KV is large enough to be worth the overhead
     if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) {
         // Try to run two workgroups per SM.
-        split_k = ctx->device->shader_core_count * 2 / workgroups_y;
+        split_k = ctx->device->shader_core_count * 2 / (workgroups_y * workgroups_z);
         if (split_k > 1) {
             // Try to evenly split KV into split_k chunks, but it needs to be a multiple
             // of "align", so recompute split_k based on that.
@@ -6213,9 +6222,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
 
-    // Reserve space for split_k temporaries. For each split, we need to store the O matrix (D x ne1)
-    // and the per-row m and L values (ne1 rows).
-    const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k : 0;
+    // Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
+    // and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
+    const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
     if (split_k_size > ctx->device->max_memory_allocation_size) {
         GGML_ABORT("Requested preallocation size is too large");
     }
@@ -6307,11 +6316,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                               (uint32_t)neq2, (uint32_t)neq3,
                                               (uint32_t)nek2, (uint32_t)nek3,
                                               (uint32_t)nev2, (uint32_t)nev3,
-                                              nem1,
+                                              nem1, nem2,
                                               q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
                                               k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
                                               v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
-                                              nbm1,
                                               scale, max_bias, logit_softcap,
                                               mask != nullptr, n_head_log2, m0, m1,
                                               gqa_ratio, split_kv, split_k };
@@ -6334,13 +6342,13 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(subctx);
-        const std::array<uint32_t, 3> pc2 = { D, (uint32_t)ne1, split_k };
+        const std::array<uint32_t, 4> pc2 = { D, (uint32_t)ne1, (uint32_t)ne3, split_k };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {
                                         vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
-                                    pc2, { (uint32_t)ne1, 1, 1 });
+                                    pc2, { (uint32_t)ne1, 1, (uint32_t)ne3 });
     } else {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
@@ -7666,7 +7674,13 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     const uint32_t nrows_x = (uint32_t)ggml_nrows(src0);
     const uint32_t nrows_y = (uint32_t)src0->ne[1];
 
-    const uint32_t n_head_kv   = nrows_x/nrows_y;
+    const uint32_t ne12 = src1 ? (uint32_t)(src1->ne[2]) : 0u;
+    const uint32_t ne13 = src1 ? (uint32_t)(src1->ne[3]) : 0u;
+    const uint32_t nb11 = src1 ? (uint32_t)(src1->nb[1] / src1->nb[0]) : 0u;
+    const uint32_t nb12 = src1 ? (uint32_t)(src1->nb[2] / src1->nb[0]) : 0u;
+    const uint32_t nb13 = src1 ? (uint32_t)(src1->nb[3] / src1->nb[0]) : 0u;
+
+    const uint32_t n_head_kv   = src0->ne[2];
     const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -7675,6 +7689,9 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX, {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
+        ne12, ne13,
+        nb11, nb12, nb13,
         scale, max_bias,
         m0, m1,
         n_head_log2,
@@ -10248,11 +10265,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
                     return false;
                 }
-                // TODO: support broadcast
-                // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-                if (op->src[0]->ne[3] != 1) {
-                    return false;
-                }
                 // It's straightforward to support different K/V dequant, but would
                 // significantly increase the number of pipelines
                 if (op->src[1]->type != op->src[2]->type) {
@@ -10413,13 +10425,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_DIAG_MASK_INF:
             return true;
         case GGML_OP_SOFT_MAX:
-            // TODO: support batching
-            if (op->src[0]->ne[3] != 1) {
-                return false;
-            }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
         case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM:

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -11,8 +11,7 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t HSK_per_thread = HSK / D_split;
-const uint32_t HSV_per_thread = HSV / D_split;
+const uint32_t D_per_thread = D / D_split;
 
 const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -30,7 +29,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * HSV + c;
+    uint32_t offset = (iq2 + r) * D + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -39,7 +38,7 @@ shared FLOAT_TYPE tmpsh[WorkGroupSize];
 shared vec4 tmpshv4[WorkGroupSize];
 
 shared float masksh[Bc][Br];
-shared vec4 Qf[Br][HSK / 4];
+shared vec4 Qf[Br][D / 4];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -54,18 +53,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (HSK / 4);
-        uint32_t r = (idx + tid) / (HSK / 4);
-        if (r < Br && d < HSK / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (D / 4);
+        uint32_t r = (idx + tid) / (D / 4);
+        if (r < Br && d < D / 4 &&
             i * Br + r < N) {
             Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
         }
     }
     barrier();
 
-    vec4 Of[Br][HSV_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+    vec4 Of[Br][D_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] = vec4(0.0);
         }
@@ -101,8 +100,8 @@ void main() {
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
     uint32_t m_offset = 0;
-    if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+    if (p.nem2 != 1) {
+        m_offset = (iq3 % p.nem2) * p.nem1 * KV;
     }
 
     [[dont_unroll]]
@@ -117,7 +116,7 @@ void main() {
 
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -149,7 +148,7 @@ void main() {
             }
         }
 
-        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+        if (p.mask != 0) {
 
             [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
                 uint32_t c = (idx + tid) % Bc;
@@ -196,14 +195,14 @@ void main() {
             Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
         }
 
-        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                 Of[r][d] = eMf[r] * Of[r][d];
             }
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -260,7 +259,7 @@ void main() {
         Lf[r] = tmpsh[d_tid];
         barrier();
 
-        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
 
             Of[r][d] = eMf * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -282,19 +281,19 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
+        uint32_t o_offset = D * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
                 }
             }
         }
 
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = D * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -310,18 +309,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] *= Lfrcp[r];
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*D;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -331,9 +330,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (i * Br + r < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp

@@ -4,10 +4,10 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
 layout (constant_id = 1) const uint32_t Br = 1;
 layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t HSK = 32;
-layout (constant_id = 4) const uint32_t HSV = 32;
-layout (constant_id = 5) const uint32_t Clamp = 0;
-layout (constant_id = 6) const uint32_t D_split = 16;
+layout (constant_id = 3) const uint32_t D = 32;
+layout (constant_id = 4) const uint32_t Clamp = 0;
+layout (constant_id = 5) const uint32_t D_split = 16;
+
 
 layout (push_constant) uniform parameter {
     uint32_t N;
@@ -25,7 +25,6 @@ layout (push_constant) uniform parameter {
     uint32_t nev3;
     uint32_t nem1;
     uint32_t nem2;
-    uint32_t nem3;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -41,7 +40,8 @@ layout (push_constant) uniform parameter {
     float max_bias;
     float logit_softcap;
 
-    uint32_t mask_n_head_log2;
+    uint32_t mask;
+    uint32_t n_head_log2;
     float m0;
     float m1;
 
@@ -50,9 +50,6 @@ layout (push_constant) uniform parameter {
     uint32_t k_num;
 } p;
 
-#define MASK_ENABLE_BIT (1<<16)
-#define N_LOG2_MASK 0xFFFF
-
 layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
 
 #if defined(A_TYPE_PACKED16)
@@ -103,10 +100,8 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
 {
     const uint32_t h = iq2 + (r % p.gqa_ratio);
 
-    uint32_t n_head_log2 = p.mask_n_head_log2 & N_LOG2_MASK;
-
-    const ACC_TYPE base = ACC_TYPE(h < n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1);
+    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
+    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
 
     return ACC_TYPE(pow(base, ACC_TYPE(exph)));
 }