vulkan : support ggml_mean (llama/15393)

* vulkan : support ggml_mean

* vulkan : support sum, sum_rows and mean with non-contiguous tensors

* vulkan : fix subbuffer size not accounting for misalign offset

* tests : add backend-op tests for non-contiguous sum_rows

* cuda : require contiguous src for SUM_ROWS, MEAN support
* sycl : require contiguous src for SUM, SUM_ROWS, ARGSORT support

* require ggml_contiguous_rows in supports_op and expect nb00=1 in the shader

diff --git a/src/ggml-cuda/ggml-cuda.cu b/src/ggml-cuda/ggml-cuda.cu
index d29a0b573f19300af868f388caf5db085e205663..aa45ab39ed89efd093e49428daaa4bb5a87e46e5 100644 (file)
--- a/src/ggml-cuda/ggml-cuda.cu
+++ b/src/ggml-cuda/ggml-cuda.cu
@@ -3485,11 +3485,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_UPSCALE:

diff --git a/src/ggml-sycl/ggml-sycl.cpp b/src/ggml-sycl/ggml-sycl.cpp
index a0a650e92e442c169a63820281d71b1a3283265a..12dd5dd2e6287a045bee3b68b4749fa65c8ddecc 100644 (file)
--- a/src/ggml-sycl/ggml-sycl.cpp
+++ b/src/ggml-sycl/ggml-sycl.cpp
@@ -4391,10 +4391,11 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return true;
         case GGML_OP_UPSCALE:
             return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
-        case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_POOL_2D:
         case GGML_OP_ACC:
         case GGML_OP_PAD:
         case GGML_OP_LEAKY_RELU:

diff --git a/src/ggml-vulkan/ggml-vulkan.cpp b/src/ggml-vulkan/ggml-vulkan.cpp
index 2c5678f4884cf21a1c927ad1a62c3f2f1b4d11cb..007556cf4f403e2f4a042a014b83008344b57a42 100644 (file)
--- a/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1015,6 +1015,39 @@ struct vk_op_upscale_push_constants {
     float sf0; float sf1; float sf2; float sf3;
 };
 
+struct vk_op_sum_rows_push_constants
+{
+    uint32_t n_cols;
+    uint32_t ne01, ne02;
+    uint32_t nb01, nb02, nb03;
+    uint32_t nb11, nb12, nb13;
+    float weight;
+    uint32_t misalign_offsets;
+    uint32_t ne0_12mp, ne0_12L;
+    uint32_t ne0_1mp, ne0_1L;
+};
+
+vk_op_sum_rows_push_constants vk_op_sum_rows_push_constants_init(const ggml_tensor * src, const ggml_tensor * dst, int64_t n_cols) {
+    uint32_t type_size = (uint32_t)ggml_type_size(src->type);
+    vk_op_sum_rows_push_constants p = {};
+    p.n_cols = (uint32_t)n_cols;
+    p.ne01 = (uint32_t)src->ne[1];
+    p.ne02 = (uint32_t)src->ne[2];
+    p.nb01 = (uint32_t)src->nb[1] / type_size;
+    p.nb02 = (uint32_t)src->nb[2] / type_size;
+    p.nb03 = (uint32_t)src->nb[3] / type_size;
+    p.nb11 = (uint32_t)dst->nb[1] / type_size;
+    p.nb12 = (uint32_t)dst->nb[2] / type_size;
+    p.nb13 = (uint32_t)dst->nb[3] / type_size;
+    p.weight = 1.0f;
+    return p;
+}
+
+template <> void init_pushconst_fastdiv(vk_op_sum_rows_push_constants &p) {
+    init_fastdiv_values(p.ne01*p.ne02, p.ne0_12mp, p.ne0_12L);
+    init_fastdiv_values(p.ne01,        p.ne0_1mp,  p.ne0_1L);
+}
+
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -3128,7 +3161,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
 
@@ -7249,6 +7282,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_SUM:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_sum_rows_f32;
         }
@@ -7387,6 +7421,9 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_IM2COL:
     case GGML_OP_SET_ROWS:
+    case GGML_OP_SUM:
+    case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
         return true;
     default:
         return false;
@@ -7421,6 +7458,16 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     GGML_UNUSED(src2);
 }
 
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_sum_rows_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.misalign_offsets = (a_offset << 16) | d_offset;
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
+
 template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
@@ -7571,10 +7618,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
 
     if (op_supports_incontiguous) {
-        x_sz = ggml_nbytes(src0);
-        y_sz = use_src1 ? ggml_nbytes(src1) : 0;
-        z_sz = use_src2 ? ggml_nbytes(src2) : 0;
-        d_sz = ggml_nbytes(dst);
+        x_sz = ggml_nbytes(src0) + get_misalign_bytes(ctx, src0);
+        y_sz = use_src1 ? ggml_nbytes(src1) + get_misalign_bytes(ctx, src1) : 0;
+        z_sz = use_src2 ? ggml_nbytes(src2) + get_misalign_bytes(ctx, src2) : 0;
+        d_sz = ggml_nbytes(dst) + get_misalign_bytes(ctx, dst);
 
         if (x_buf_offset + x_sz >= d_X->size) {
             x_sz = VK_WHOLE_SIZE;
@@ -7602,6 +7649,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_SOFT_MAX:
     case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
         {
             const uint32_t nr = ggml_nrows(src0);
@@ -8588,11 +8636,19 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
 }
 
 static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, ggml_nelements(src0));
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, p, dryrun);
 }
 
 static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun);
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p, dryrun);
+}
+
+static void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
+    p.weight = 1.0f / (float)src0->ne[0];
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_MEAN, p, dryrun);
 }
 
 static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -9815,6 +9871,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_ARGSORT:
     case GGML_OP_SUM:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
@@ -9884,6 +9941,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_ARGMAX:
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
@@ -10087,6 +10145,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_SUM_ROWS:
         ggml_vk_sum_rows(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_MEAN:
+        ggml_vk_mean(ctx, compute_ctx, src0, node, dryrun);
+
         break;
     case GGML_OP_ARGMAX:
         ggml_vk_argmax(ctx, compute_ctx, src0, node, dryrun);
@@ -10246,6 +10308,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_ARGSORT:
     case GGML_OP_SUM:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
@@ -11483,8 +11546,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
+            return true;
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
+            return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_ARGMAX:
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
@@ -12043,6 +12109,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         tensor_clone = ggml_sum(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_SUM_ROWS) {
         tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]);
+    } else if (tensor->op == GGML_OP_MEAN) {
+        tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_ARGMAX) {
         tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_COUNT_EQUAL) {

diff --git a/src/ggml-vulkan/vulkan-shaders/sum_rows.comp b/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
index 961e5ffa1f56f42b6b35a855ecc582d47ac7d298..759204afaf9aaa373cf6192c339efa722e199ec7 100644 (file)
--- a/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
+++ b/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
@@ -1,9 +1,9 @@
 #version 450
 
-#include "generic_head.comp"
 #include "types.comp"
 
 #extension GL_EXT_control_flow_attributes : enable
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -11,16 +11,49 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 
+layout (push_constant) uniform parameter
+{
+    uint n_cols;
+    uint ne01, ne02;
+    uint nb01, nb02, nb03;
+    uint nb11, nb12, nb13;
+    float weight;
+    uint misalign_offsets;
+    uint ne0_12mp, ne0_12L;
+    uint ne0_1mp, ne0_1L;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
+
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
+    const float weight = p.weight;
+
+    const uint i03 = fastdiv(row, p.ne0_12mp, p.ne0_12L);
+    const uint i03_offset = i03 * p.ne01*p.ne02;
+    const uint i02 = fastdiv(row - i03_offset, p.ne0_1mp, p.ne0_1L);
+    const uint i01 = row - i03_offset - i02*p.ne01;
+
+    const uint src_idx = get_aoffset() + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+    const uint dst_idx = get_doffset() + i01 * p.nb11 + i02 * p.nb12 + i03 * p.nb13;
 
-    tmp[col] = FLOAT_TYPE(0.0f);
+    tmp[col] = FLOAT_TYPE(0.0);
 
-    for (uint i = col; i < p.KX; i += BLOCK_SIZE) {
-        tmp[col] += FLOAT_TYPE(data_a[row*p.KX + i]);
+    for (uint i = col; i < p.n_cols; i += BLOCK_SIZE) {
+        tmp[col] += FLOAT_TYPE(data_a[src_idx + i]);
     }
 
     barrier();
@@ -32,6 +65,6 @@ void main() {
     }
 
     if (col == 0) {
-        data_d[row] = D_TYPE(tmp[0]);
+        data_d[dst_idx] = D_TYPE(tmp[0] * weight);
     }
 }

diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index a51527ca55c2359eb0dc3a14385a88e4e8e8cbff..2e53f8e21a5a29b5f18805e34418702063b56a20 100644 (file)
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -4300,20 +4300,32 @@ struct test_sum : public test_case {
 struct test_sum_rows : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
+    const bool permute;
+    const bool slice;
 
     std::string vars() override {
-        return VARS_TO_STR2(type, ne);
+        return VARS_TO_STR4(type, ne, permute, slice);
     }
 
     test_sum_rows(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 5, 4, 3})
-        : type(type), ne(ne) {}
+            std::array<int64_t, 4> ne = {10, 5, 4, 3},
+            bool permute = false, bool slice = false)
+        : type(type), ne(ne), permute(permute), slice(slice) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_param(a);
         ggml_set_name(a, "a");
 
+        if (slice) {
+            a = ggml_view_4d(ctx, a,
+                             ne[0], ne[1], ne[2] / 2, ne[3] - 1,
+                             a->nb[1], a->nb[2] * 2, a->nb[3], /*offset=*/a->nb[3]);
+        }
+        if (permute) {
+            a = ggml_permute(ctx, a, 0, 2, 3, 1);
+        }
+
         ggml_tensor * out = ggml_sum_rows(ctx, a);
         ggml_set_name(out, "out");
 
@@ -6195,6 +6207,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, true, false));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, false, true));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, true, true));
     test_cases.emplace_back(new test_mean());
     test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 1, 1, 1 }));
     test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 33, 1, 1, 1 }));