@novastera-oss/llamarn 0.4.1 → 0.4.3-beta4

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp

@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    data_d[i] = D_TYPE(abs(float(data_a[i])));
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "types.comp"
-#include "generic_binary_head.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/add.comp

@@ -6,8 +6,8 @@
 #extension GL_KHR_shader_subgroup_basic : enable
 #endif
 
-#include "types.comp"
-#include "generic_binary_head.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
 
 const uint num_threads = 256;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp

@@ -0,0 +1,28 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= p.ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03);
+
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset()]));
+
+        idx += num_threads;
+    }
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp

@@ -2,7 +2,7 @@
 
 #extension GL_EXT_control_flow_attributes : require
 
-#include "types.comp"
+#include "types.glsl"
 
 layout (push_constant) uniform parameter
 {

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp

@@ -0,0 +1,20 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    // p.param1 = start, p.param2 = step
+    float value = p.param1 + p.param2 * float(i);
+    data_d[i] = D_TYPE(value);
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp

@@ -1,44 +1,42 @@
 #version 450
 #extension GL_EXT_control_flow_attributes : enable
 
-#include "types.comp"
+#include "types.glsl"
 
 layout(constant_id = 0) const int BLOCK_SIZE = 1024;
-layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
+layout(constant_id = 1) const int NCOLS_PADDED_LOG2 = 10;
 #define ASC 0
 
 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[];};
-layout (binding = 1)          buffer D {int data_d[];};
+layout (binding = 2) writeonly buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
+    uint ncols_padded;
+    uint ncols_padded_log2;
+    uint nrows;
     uint order;
+    uint outer_start;
+    uint outer_end;
+    uint inner_start;
+    uint inner_end;
 } p;
 
-shared int dst_row[BLOCK_SIZE];
-shared A_TYPE a_sh[BLOCK_SIZE];
+shared ivec2 dst_row[BLOCK_SIZE];
 
-void swap(uint idx0, uint idx1) {
-    int tmp = dst_row[idx0];
-    dst_row[idx0] = dst_row[idx1];
-    dst_row[idx1] = tmp;
-}
-
-void argsort(bool needs_bounds_check) {
+void argsort(bool needs_bounds_check, const uint row) {
     // bitonic sort
     const int col = int(gl_LocalInvocationID.x);
-    const uint row = gl_WorkGroupID.y;
 
     const uint row_offset = row * p.ncols;
 
     // initialize indices
-    dst_row[col] = col;
-    a_sh[col] = data_a[row_offset + col];
+    dst_row[col] = ivec2(col, floatBitsToInt(data_a[row_offset + col]));
     barrier();
 
-    uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
+    uint num_outer_loop_iters = NCOLS_PADDED_LOG2;
     [[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
         uint num_inner_loop_iters = outer_idx + 1;
         [[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
@@ -47,14 +45,15 @@ void argsort(bool needs_bounds_check) {
             int idx_0 = (col & k) == 0 ? col : ixj;
             int idx_1 = (col & k) == 0 ? ixj : col;
 
-            int sh_idx_0 = dst_row[idx_0];
-            int sh_idx_1 = dst_row[idx_1];
-            bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
-            bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
+            ivec2 sh_idx_0 = dst_row[idx_0];
+            ivec2 sh_idx_1 = dst_row[idx_1];
+            bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.ncols : false;
+            bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= p.ncols : false;
 
             if ((idx_0_oob ||
-                (!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
-                swap(idx_0, idx_1);
+                (!idx_1_oob && intBitsToFloat(sh_idx_0.y) > intBitsToFloat(sh_idx_1.y))) && (ixj > col)) {
+                dst_row[idx_0] = sh_idx_1;
+                dst_row[idx_1] = sh_idx_0;
             }
 
             barrier();
@@ -63,17 +62,25 @@ void argsort(bool needs_bounds_check) {
 
     if (col < p.ncols) {
         if (p.order == ASC) {
-            data_d[row_offset + col] = dst_row[col];
+            data_d[row_offset + col] = dst_row[col].x;
         } else {
-            data_d[row_offset + p.ncols - col - 1] = dst_row[col];
+            data_d[row_offset + p.ncols - col - 1] = dst_row[col].x;
         }
     }
 }
 
 void main() {
     if (p.ncols == BLOCK_SIZE) {
-        argsort(false);
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
     } else {
-        argsort(true);
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp

@@ -0,0 +1,114 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_memory_scope_semantics : enable
+#pragma use_vulkan_memory_model
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int WG_UNROLL_FACTOR = 2;
+#define ASC 0
+
+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[];};
+layout (binding = 1) workgroupcoherent buffer B {ivec2 tmp_idx[];};
+layout (binding = 2) workgroupcoherent buffer D {int data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint ncols;
+    uint ncols_padded;
+    uint ncols_padded_log2;
+    uint nrows;
+    uint order;
+    uint outer_start;
+    uint outer_end;
+    uint inner_start;
+    uint inner_end;
+} p;
+
+void argsort(bool needs_bounds_check, const uint row) {
+    // bitonic sort
+    int col = int(gl_GlobalInvocationID.x);
+    col = (col % BLOCK_SIZE) + (col / BLOCK_SIZE) * BLOCK_SIZE * WG_UNROLL_FACTOR;
+
+    const uint row_offset = row * p.ncols;
+    uint idx_offset = row * p.ncols_padded;
+
+    bool need_barrier = false;
+
+    // initialize indices
+    if (p.outer_start == 0 && p.inner_start == 0) {
+        [[unroll]] for (int u = 0; u < WG_UNROLL_FACTOR; ++u) {
+            uint c = u*BLOCK_SIZE + col;
+            if (c < p.ncols_padded) {
+                ivec2 v = ivec2(c, floatBitsToInt(data_a[row_offset + c]));
+                tmp_idx[idx_offset + c] = v;
+            }
+        }
+        need_barrier = true;
+    }
+
+    [[unroll]] for (uint outer_idx = p.outer_start, k = (2 << outer_idx); outer_idx < p.outer_end; k *= 2, outer_idx++) {
+        uint inner_end = min(p.inner_end, outer_idx + 1);
+        for (uint j = k >> (p.inner_start + 1), inner_idx = p.inner_start; inner_idx < inner_end; j /= 2, inner_idx++) {
+            if (need_barrier) {
+                controlBarrier(gl_ScopeWorkgroup, gl_ScopeWorkgroup, gl_StorageSemanticsBuffer, gl_SemanticsAcquireRelease);
+            }
+            need_barrier = true;
+            [[unroll]] for (int u = 0; u < WG_UNROLL_FACTOR; ++u) {
+                int c = u*BLOCK_SIZE + col;
+                const int ixj = int(c ^ j);
+
+                if (ixj < c) {
+                    continue;
+                }
+
+                int idx_0 = (c & k) == 0 ? c : ixj;
+                int idx_1 = (c & k) == 0 ? ixj : c;
+
+                ivec2 sh_idx_0 = tmp_idx[idx_offset + idx_0];
+                ivec2 sh_idx_1 = tmp_idx[idx_offset + idx_1];
+                bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.ncols : false;
+                bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= p.ncols : false;
+
+                if ((idx_0_oob ||
+                    (!idx_1_oob && intBitsToFloat(sh_idx_0.y) > intBitsToFloat(sh_idx_1.y)))) {
+                    tmp_idx[idx_offset + idx_0] = sh_idx_1;
+                    tmp_idx[idx_offset + idx_1] = sh_idx_0;
+                }
+            }
+        }
+    }
+
+    if (p.outer_end == p.ncols_padded_log2 &&
+        p.inner_end >= p.ncols_padded_log2 + 1) {
+        controlBarrier(gl_ScopeWorkgroup, gl_ScopeWorkgroup, gl_StorageSemanticsBuffer, gl_SemanticsAcquireRelease);
+        [[unroll]] for (int u = 0; u < WG_UNROLL_FACTOR; ++u) {
+            uint c = u*BLOCK_SIZE + col;
+            if (c < p.ncols) {
+                if (p.order == ASC) {
+                    data_d[row_offset + c] = tmp_idx[idx_offset + c].x;
+                } else {
+                    data_d[row_offset + p.ncols - c - 1] = tmp_idx[idx_offset + c].x;
+                }
+            }
+        }
+    }
+}
+
+void main() {
+    if (p.ncols == p.ncols_padded) {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    } else {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    }
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(ceil(x));
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "types.comp"
-#include "generic_unary_head.comp"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "types.comp"
-#include "generic_binary_head.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "types.comp"
-#include "generic_unary_head.comp"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
 
 #extension GL_EXT_control_flow_attributes : require
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "types.comp"
+#include "types.glsl"
 
 layout (push_constant) uniform parameter
 {

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -11,12 +11,12 @@
 #    extension GL_KHR_shader_subgroup_shuffle : enable
 #endif
 
-#include "types.comp"
+#include "types.glsl"
 
 // shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
 layout(binding = 0) readonly buffer A {
     A_TYPE knl_data[];
-};  // src0 - kernel:   [KW, KH, Cin, Cout]
+};  // src0 - kernel:   [KW, KH, Cin, Cout] for conv_2d, [KW, KH, Cout, Cin] for conv_transposed_2d
 
 layout(binding = 1) readonly buffer B {
     B_TYPE src_data[];
@@ -62,8 +62,6 @@ layout(push_constant) uniform parameter {
     uint32_t nb3;
 
     // fastdiv helper values
-    uint32_t KWmp;   uint32_t KWL;
-    uint32_t KWKHmp; uint32_t KWKHL;
     uint32_t OWmp;   uint32_t OWL;
     uint32_t OWOHmp; uint32_t OWOHL;
 }
@@ -80,6 +78,15 @@ layout(constant_id = 4) const uint TS_K            = 8;
 layout(constant_id = 5) const uint use_collectives = 1;
 layout(constant_id = 6) const uint SHMEM_PAD       = 4;
 
+layout(constant_id = 7)  const uint s0             = 1;
+layout(constant_id = 8)  const uint s1             = 1;
+layout(constant_id = 9)  const uint p0             = 0;
+layout(constant_id = 10) const uint p1             = 0;
+layout(constant_id = 11) const uint d0             = 1;
+layout(constant_id = 12) const uint d1             = 1;
+layout(constant_id = 13) const uint KW             = 1;
+layout(constant_id = 14) const uint KH             = 1;
+
 uint32_t       tid     = gl_LocalInvocationID.x;
 const uint32_t WG_SIZE = gl_WorkGroupSize.x;
 
@@ -88,7 +95,7 @@ uint splitWork(uint work_size, uint block_size) {
 }
 
 uint32_t K   = p.Cout;
-uint32_t CRS = p.Cin * p.KH * p.KW;
+uint32_t CRS = p.Cin * KH * KW;
 uint32_t NPQ = p.N * p.OH * p.OW;
 
 uint32_t n_elems_out = K * NPQ;
@@ -183,7 +190,7 @@ void main() {
     }
 #endif
     /* Advance block in CRS dim */
-    for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
+    [[dont_unroll]] for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
         uint32_t CRS_idx_a;
         uint32_t Cin_idx_a;
         uint32_t KH_idx_a;
@@ -196,10 +203,10 @@ void main() {
         uint32_t cached_KW_idx;
         if (use_collectives == 1) {
             cached_CRS_idx                = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
-            cached_Cin_idx                = fastdiv(cached_CRS_idx, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-            uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
-            cached_KH_idx                 = fastdiv(cached_CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-            cached_KW_idx                 = cached_CRS_remainder - cached_KH_idx * p.KW;
+            cached_Cin_idx                = cached_CRS_idx / (KW * KH);
+            uint32_t cached_CRS_remainder = cached_CRS_idx % (KW * KH);
+            cached_KH_idx                 = cached_CRS_remainder / KW;
+            cached_KW_idx                 = cached_CRS_remainder % KW;
 
             CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
             Cin_idx_a = subgroupShuffle(cached_Cin_idx, Ac);
@@ -207,25 +214,29 @@ void main() {
             KW_idx_a  = subgroupShuffle(cached_KW_idx, Ac);
         } else {
             CRS_idx_a              = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-            Cin_idx_a              = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-            uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-            KH_idx_a               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-            KW_idx_a               = CRS_remainder - KH_idx_a * p.KW;
+            Cin_idx_a              = CRS_idx_a / (KW * KH);
+            uint32_t CRS_remainder = CRS_idx_a % (KW * KH);
+            KH_idx_a               = CRS_remainder / KW;
+            KW_idx_a               = CRS_remainder % KW;
         }
 #else
         CRS_idx_a     = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-        Cin_idx_a     = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH); / (p.KW * p.KH);
-        CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-        KH_idx_a      = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-        KW_idx_a      = CRS_remainder - KH_idx_a * p.KW;
+        Cin_idx_a     = CRS_idx_a / (KW * KH);
+        CRS_remainder = CRS_idx_a % (KW * KH);
+        KH_idx_a      = CRS_remainder / KW;
+        KW_idx_a      = CRS_remainder % KW;
 #endif
 
         /* Load kernel to A_block: (BS_K x BS_CRS)*/
-        for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
             uint32_t B_ly    = r_offset + Ar;
             uint32_t B_lx    = Ac;
             uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
+#ifdef TRANSPOSE
+            uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + K_idx * p.nb02 + Cin_idx_a * p.nb03, K * CRS - 1);
+#else
             uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + Cin_idx_a * p.nb02 + K_idx * p.nb03, K * CRS - 1);
+#endif
             float    val     = knl_data[knl_idx];
             if (K_idx >= K || CRS_idx_a >= CRS) {
                 val = 0.0;
@@ -254,25 +265,37 @@ void main() {
                 KW_idx_b  = subgroupShuffle(cached_KW_idx, r_offset + Br);
             } else {
                 CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-                Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-                uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-                KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-                KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+                Cin_idx_b              = CRS_idx_b / (KW * KH);
+                uint32_t CRS_remainder = CRS_idx_b % (KW * KH);
+                KH_idx_b               = CRS_remainder / KW;
+                KW_idx_b               = CRS_remainder % KW;
             }
 #else
             CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-            Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-            uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-            KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-            KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+            Cin_idx_b              = CRS_idx_b / (KW * KH);
+            uint32_t CRS_remainder = CRS_idx_b % (KW * KH);
+            KH_idx_b               = CRS_remainder / KW;
+            KW_idx_b               = CRS_remainder % KW;
 #endif
 
-            uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
-            uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
+#ifdef TRANSPOSE
+            uint32_t H_idx_x_s1 = OH_idx - KH_idx_b * d1 + p1;
+            uint32_t W_idx_x_s0 = OW_idx - KW_idx_b * d0 + p0;
+            uint32_t H_idx = H_idx_x_s1 / s1;
+            uint32_t W_idx = W_idx_x_s0 / s0;
+#else
+            uint32_t H_idx = OH_idx * s1 + KH_idx_b * d1 - p1;
+            uint32_t W_idx = OW_idx * s0 + KW_idx_b * d0 - p0;
+#endif
             uint32_t src_idx =
                 min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
             float val = src_data[src_idx];
-            if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
+            if (CRS_idx_b >= CRS || NPQ_idx >= NPQ
+                || H_idx >= p.H || W_idx >= p.W // Lower bound checks aren't necessary. (idx >= 0x80000000 for such case)
+#ifdef TRANSPOSE
+                || (H_idx_x_s1 - H_idx * s1 != 0) || (W_idx_x_s0 - W_idx * s0 != 0)
+#endif
+                ) {
                 val = 0.0;
             }
             Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "types.comp"
+#include "types.glsl"
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};   // src0 - kernel:    [K, Cout, Cin]
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};   // src1 - input:     [L, Cin]

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "types.comp"
-#include "generic_unary_head.comp"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp

@@ -1,8 +1,8 @@
 #version 450
 
-#include "types.comp"
-#include "generic_unary_head.comp"
-#include "dequant_funcs.comp"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+#include "dequant_funcs.glsl"
 
 #if defined(DATA_A_IQ4_NL) || defined(DATA_A_MXFP4)
 // 16 invocations needed for init_iq_shmem

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "rte.comp"
-#include "types.comp"
+#include "rte.glsl"
+#include "types.glsl"
 
 #if defined(SET_ROWS) && QUANT_K == 1
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
@@ -14,11 +14,18 @@ const uint BLOCK_SIZE = 32;
 layout (binding = 0) readonly buffer S {float data_s[];};
 
 #if defined(SET_ROWS)
-#include "generic_binary_head.comp"
-layout (binding = 1) readonly buffer C {uvec2 data_i[];};
+#include "generic_binary_head.glsl"
+layout (binding = 1) readonly buffer C {B_TYPE data_i[];};
 layout (binding = 2) writeonly buffer Q {A_TYPE data_q[];};
+
+#if B_SIZE == 64
+#define DATA_I_SWIZZLE .x
+#else
+#define DATA_I_SWIZZLE
+#endif
+
 #else
-#include "generic_unary_head.comp"
+#include "generic_unary_head.glsl"
 layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
 #endif
 
@@ -259,7 +266,7 @@ void main() {
     uint i11 = fastmod(i02, p.ne11);
     uint i10 = i01;
 
-    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()].x;
+    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()] DATA_I_SWIZZLE;
 
     uint src0_idx = src0_idx(i00, i01, i02, i03) + get_aoffset();
     uint dst_idx = dst_idx(i00 / QUANT_K, i1, i02, i03) + get_doffset();