From 0787b80db8b78398061b7256e9cbf045cbcccb72 Mon Sep 17 00:00:00 2001
From: Jared Van Bortel <jared@nomic.ai>
Date: Mon, 22 Jan 2024 17:42:05 -0500
Subject: [PATCH] kompute : remove broken mulrow kernel -> 1 less test failure

---
 CMakeLists.txt                 |  2 --
 ggml-kompute.cpp               | 55 ++++++----------------------------
 kompute-shaders/op_mulrow.comp | 25 ----------------
 3 files changed, 9 insertions(+), 73 deletions(-)
 delete mode 100644 kompute-shaders/op_mulrow.comp

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 01f01bfee..c3533b969 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -520,7 +520,6 @@ if (LLAMA_KOMPUTE)
           kompute-shaders/op_add.comp
           kompute-shaders/op_addrow.comp
           kompute-shaders/op_mul.comp
-          kompute-shaders/op_mulrow.comp
           kompute-shaders/op_silu.comp
           kompute-shaders/op_relu.comp
           kompute-shaders/op_gelu.comp
@@ -553,7 +552,6 @@ if (LLAMA_KOMPUTE)
           shaderop_add.h
           shaderop_addrow.h
           shaderop_mul.h
-          shaderop_mulrow.h
           shaderop_silu.h
           shaderop_relu.h
           shaderop_gelu.h
diff --git a/ggml-kompute.cpp b/ggml-kompute.cpp
index 163b0a29a..45a579b3b 100644
--- a/ggml-kompute.cpp
+++ b/ggml-kompute.cpp
@@ -9,7 +9,6 @@
 #include "shaderop_add.h"
 #include "shaderop_addrow.h"
 #include "shaderop_mul.h"
-#include "shaderop_mulrow.h"
 #include "shaderop_silu.h"
 #include "shaderop_relu.h"
 #include "shaderop_gelu.h"
@@ -671,37 +670,6 @@ static void ggml_vk_mul(
     seq.record<kp::OpAlgoDispatch>(s_algo);
 }
 
-static void ggml_vk_mulrow(kp::Sequence& seq,
-                 const std::shared_ptr<kp::Tensor>& inA,
-                 const std::shared_ptr<kp::Tensor>& inB,
-                 const std::shared_ptr<kp::Tensor>& out,
-                 uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-                 uint32_t size, uint32_t row = 0) {
-
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mulrow_comp_spv,
-        kp::shader_data::op_mulrow_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        uint32_t row;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        row
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
 static void ggml_vk_scale(kp::Sequence& seq,
                    const std::shared_ptr<kp::Tensor>& in,
                    const std::shared_ptr<kp::Tensor>& out,
@@ -1516,20 +1484,15 @@ void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph
                     } break;
                 case GGML_OP_MUL:
                     {
-                        if (ggml_nelements(src1) == ne10) {
-                            // src1 is a row
-                            ggml_vk_mulrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
-                        } else {
-                            ggml_vk_mul(
-                                seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                ne00, ne01, ne02, ne03,
-                                nb00, nb01, nb02, nb03,
-                                ne10, ne11, ne12, ne13,
-                                nb10, nb11, nb12, nb13,
-                                ne0,
-                                nb0, nb1, nb2, nb3
-                            );
-                        }
+                        ggml_vk_mul(
+                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                            ne00, ne01, ne02, ne03,
+                            nb00, nb01, nb02, nb03,
+                            ne10, ne11, ne12, ne13,
+                            nb10, nb11, nb12, nb13,
+                            ne0,
+                            nb0, nb1, nb2, nb3
+                        );
                     } break;
                 case GGML_OP_SCALE:
                     {
diff --git a/kompute-shaders/op_mulrow.comp b/kompute-shaders/op_mulrow.comp
deleted file mode 100644
index ae7106320..000000000
--- a/kompute-shaders/op_mulrow.comp
+++ /dev/null
@@ -1,25 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    uint row;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = inA[i + pcs.inAOff] * inB[(i % pcs.row) + pcs.inBOff];
-    }
-}
\ No newline at end of file