vulkan: roll bk loop in matmul for asahi linux (#24663)

* vulkan: roll bk loop in matmul for asahi linux

* vulkan: fix inline comment

* vulkan: revert BK-loop unroll change

* vulkan: edit spirv directly for asahi roll bk loop

* vulkan: remove trailing whitespace at the end of comments

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

@@ -2475,6 +2475,85 @@ static bool ggml_vk_strip_decode_vector(const uint32_t * code, size_t word_count
     return true;
 }
 
+// Remove the loop unrolling hint of the matmul shader's BK loop
+// and replace it with the dont_unroll hint for better performance on
+// hardware like Apple M1/M2.
+// Assumes 1. code comes from mul_mm.comp 2. the K-tile loop has no loop
+// control hint and 3. the BK loop is the last loop nested directly inside
+// the K-tile loop.
+// Returns true when the input was modified; returns false otherwise
+// without touching `out`.
+static bool ggml_vk_roll_bk_loop(const uint32_t * code, size_t word_count, std::vector<uint32_t> & out) {
+    if (word_count < 5) {
+        return false;
+    }
+
+    struct vk_spv_loop {
+        size_t   header;
+        size_t   end;
+        uint32_t control;
+    };
+
+    std::vector<vk_spv_loop> loops;
+
+    // Collect a list of all loops in the module.
+    for (size_t pos = 5; pos < word_count; ) {
+        const uint32_t wc = code[pos] >> spv::WordCountShift;
+        const uint32_t op = code[pos] & spv::OpCodeMask;
+        if (wc == 0 || pos + wc > word_count) {
+            return false;
+        }
+
+        if (op == spv::OpLoopMerge && wc >= 4) { loops.push_back({ pos, 0, code[pos + 3] }); }
+
+        if (op == spv::OpLabel && wc >= 2) {
+            for (auto & l : loops) {
+                if (l.end == 0 && code[l.header + 1] == code[pos + 1]) { l.end = pos; }
+            }
+        }
+
+        pos += wc;
+    }
+
+    auto encloses = [](const vk_spv_loop & a, const vk_spv_loop & b) {
+        return a.header < b.header && b.header < a.end;
+    };
+
+    // Find the BK loop.
+    const vk_spv_loop * bk = nullptr;
+    for (const auto & h : loops) {
+        if (h.control != spv::LoopControlUnrollMask) {
+            continue;
+        }
+        const vk_spv_loop * parent = nullptr;
+        bool has_child = false;
+        for (const auto & g : loops) {
+            if (encloses(g, h) && (!parent || g.header > parent->header)) {
+                parent = &g;
+            }
+            if (encloses(h, g)) {
+                has_child = true;
+            }
+        }
+        // BK loop should be the last loop nested inside the loop with no hint
+        // and have at least one child loop.
+        if (parent &&
+            parent->control == spv::LoopControlMaskNone &&
+            has_child &&
+            (!bk || h.header > bk->header)) {
+            bk = &h;
+        }
+    }
+    if (!bk) {
+        return false;
+    }
+
+    // set DontUnroll instead of Unroll
+    out.assign(code, code + word_count);
+    out[bk->header + 3] = spv::LoopControlDontUnrollMask;
+    return true;
+}
+
 static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, size_t spv_size, const void* spv_data, const std::string entrypoint,
                                          uint32_t parameter_count, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants,
                                          bool disable_robustness, bool require_full_subgroups, uint32_t required_subgroup_size) {
@@ -2558,6 +2637,22 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
     }
 #endif
 
+#if VK_HEADER_VERSION >= 287
+    // Roll the mul_mm BK loop on Asahi Linux. Skip bf16 and the mul_mmq pipelines.
+    if (device->driver_id == vk::DriverId::eMesaHoneykrisp &&
+        pipeline->name.rfind("matmul", 0) == 0 &&
+        pipeline->name.find("bf16") == std::string::npos &&
+        pipeline->name.find("q8_1") == std::string::npos) {
+        const uint32_t * src   = spirv.empty() ? reinterpret_cast<const uint32_t *>(spv_data) : spirv.data();
+        size_t           src_n = spirv.empty() ? spv_size / sizeof(uint32_t) : spirv.size();
+        std::vector<uint32_t> rolled;
+        if (ggml_vk_roll_bk_loop(src, src_n, rolled)) {
+            spirv = std::move(rolled);
+            shader_module_create_info = vk::ShaderModuleCreateInfo({}, spirv.size() * sizeof(uint32_t), spirv.data());
+        }
+    }
+#endif
+
     pipeline->shader_module = device->device.createShaderModule(shader_module_create_info);
 
     vk::PushConstantRange pcr(