@novastera-oss/llamarn 0.4.1 → 0.4.3-beta4

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

@@ -0,0 +1,67 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+// workgroup does 32x32 tile, but uses 32x8 threads
+#define TILE_DIM 32
+layout(local_size_x = 32, local_size_y = 8, local_size_z = 1) in;
+
+shared uint sh[TILE_DIM][TILE_DIM + 1];
+
+void iter(uvec3 wg_id) {
+    const uint tile_col = wg_id.x;
+    const uint tile_row = wg_id.y;
+
+    const uint tid_col = gl_LocalInvocationID.x;
+    const uint tid_row = gl_LocalInvocationID.y;
+
+    const uint i2 = wg_id.z % p.ne12;
+    const uint i3 = wg_id.z / p.ne12;
+    const uint i02 = i2;
+    const uint i03 = i3;
+
+    // The workgroup does TILE_DIM x TILE_DIM, but swaps the LSBs of the
+    // src coords to make memory accesses contiguous, dst has tid.x in i0,
+    // src has tid.x in i01
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i00 = tile_col * TILE_DIM + tid_row + 8 * y;
+        const uint i01 = tile_row * TILE_DIM + tid_col;
+        if (i00 < p.ne00 && i01 < p.ne01 && i02 < p.ne02 && i03 < p.ne03) {
+            const uint src_idx = i00 * p.nb00 + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+            sh[tid_row + 8 * y][tid_col] = uint(data_a[get_aoffset() + src_idx]);
+        }
+    }
+
+    barrier();
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i0 = tile_col * TILE_DIM + tid_col;
+        const uint i1 = tile_row * TILE_DIM + tid_row + 8 * y;
+        if (i0 < p.ne10 && i1 < p.ne11 && i2 < p.ne12 && i3 < p.ne13) {
+            const uint dst_idx = i0 * p.nb10 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+            // load transposed
+            data_d[get_doffset() + dst_idx] = D_TYPE(sh[tid_col][tid_row + 8 * y]);
+        }
+    }
+}
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+void main() {
+    uint z = gl_WorkGroupID.z;
+    uint y = gl_WorkGroupID.y;
+    bool need_barrier = false;
+    for (uint z = gl_WorkGroupID.z; z < p.ne12 * p.ne13; z += gl_NumWorkGroups.z) {
+        for (uint y = gl_WorkGroupID.y; y < CEIL_DIV(p.ne11, TILE_DIM); y += gl_NumWorkGroups.y) {
+            for (uint x = gl_WorkGroupID.x; x < CEIL_DIV(p.ne10, TILE_DIM); x += gl_NumWorkGroups.x) {
+                if (need_barrier) {
+                    barrier();
+                }
+                need_barrier = true;
+                iter(uvec3(x, y, z));
+            }
+        }
+    }
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/cos.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/count_equal.comp

@@ -2,8 +2,8 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
-#include "types.comp"
-#include "generic_head.comp"
+#include "types.glsl"
+#include "generic_head.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/{dequant_funcs.comp → dequant_funcs.glsl}

@@ -2,7 +2,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
 #endif
 
-#include "types.comp"
+#include "types.glsl"
 
 #if defined(A_TYPE_PACKED16)
 layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
@@ -437,7 +437,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 #if defined(DATA_A_MXFP4)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]);
+    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]) * 0.5;
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     vec2 v0 = dequantize(ib, iqs, a_offset);
@@ -478,3 +478,139 @@ vec2 get_dm(uint ib, uint a_offset) {
     return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
 }
 #endif
+
+#if defined(DATA_A_Q2_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
+    const uint scalesi = iqs / 8;                      // 0..15
+    const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
+
+    const uvec2 qs = uvec2(data_a[a_offset + ib].qs[qsi], data_a[a_offset + ib].qs[qsi + 1]);
+    const uint scales = data_a[a_offset + ib].scales[scalesi];
+    const vec2 dm = vec2(data_a[a_offset + ib].dm);
+
+    return dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q3_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 64;                     // 0,1
+    const uint qsi = n * 32 + (iqs % 16) * 2;    // 0,2,4..62
+    const uint hmi =          (iqs % 16) * 2;    // 0,2,4..30
+    const uint j = (iqs % 64) / 4;               // 0..3
+    const uint is = iqs / 8;                     // 0..15
+    const uint halfsplit = ((iqs % 64) / 16);    // 0,1,2,3
+    const uint qsshift = halfsplit * 2;          // 0,2,4,6
+    const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
+
+    const int8_t us = int8_t(((data_a[a_offset + ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF)
+                          | (((data_a[a_offset + ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
+    const float dl = float(data_a[a_offset + ib].d) * float(us - 32);
+
+    return vec2(dl * float(int8_t((data_a[a_offset + ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[a_offset + ib].hmask[hmi    ] & m) != 0) ? 0 : 4)),
+                dl * float(int8_t((data_a[a_offset + ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[a_offset + ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 32;                   // 0,1,2,3
+    const uint b = (iqs % 32) / 16;            // 0,1
+    const uint is = 2 * n + b;                 // 0..7
+    const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+
+    const vec2 loadd = vec2(data_a[a_offset + ib].dm);
+
+    const uint scidx0 = (is < 4) ? is : (is + 4);
+    const uint scidx1 = (is < 4) ? is : (is - 4);
+    const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint scidxshift1 = (is < 4) ? 0 : 2;
+    const uint mbidx0 = is + 4;
+    const uint mbidx1 = (is < 4) ? is + 4 : is;
+    const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+    const uint mbidxshift0 = (is < 4) ? 0 : 4;
+    const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+    const uint8_t sc = uint8_t((data_a[a_offset + ib].scales[scidx0] & 0xF) | ((data_a[a_offset + ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+    const uint8_t mbyte = uint8_t((data_a[a_offset + ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[a_offset + ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+    const float d = loadd.x * sc;
+    const float m = -loadd.y * mbyte;
+
+    return vec2(fma(d, float((data_a[a_offset + ib].qs[qsi    ] >> (b * 4)) & 0xF), m),
+                fma(d, float((data_a[a_offset + ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q5_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 32;                   // 0,1,2,3
+    const uint b = (iqs % 32) / 16;            // 0,1
+    const uint is = 2 * n + b;                 // 0..7
+    const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+    const uint qhi = (iqs % 16) * 2;           // 0,2,4..30
+
+    const uint8_t hm = uint8_t(1 << (iqs / 16));
+
+    const vec2 loadd = vec2(data_a[a_offset + ib].dm);
+
+    const uint scidx0 = (is < 4) ? is : (is + 4);
+    const uint scidx1 = (is < 4) ? is : (is - 4);
+    const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint scidxshift1 = (is < 4) ? 0 : 2;
+    const uint mbidx0 = is + 4;
+    const uint mbidx1 = (is < 4) ? is + 4 : is;
+    const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+    const uint mbidxshift0 = (is < 4) ? 0 : 4;
+    const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    const uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+    const uint8_t sc    = uint8_t((data_a[a_offset + ib].scales[scidx0] & 0xF)                         | ((data_a[a_offset + ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+    const uint8_t mbyte = uint8_t(((data_a[a_offset + ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[a_offset + ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+    const float d = loadd.x * sc;
+    const float m = -loadd.y * mbyte;
+
+    return vec2(fma(d, float((data_a[a_offset + ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[a_offset + ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m),
+                fma(d, float((data_a[a_offset + ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[a_offset + ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif
+
+#if defined(DATA_A_Q6_K)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    iqs /= 2;
+    const uint n = iqs / 64;                    // 0,1
+    const uint b = (iqs % 64) / 32;             // 0,1
+    const uint is_b = (iqs % 16) / 8;           // 0,1
+    const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
+    const uint is = 8 * n + qhshift + is_b;     // 0..15
+    const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
+    const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+
+    const float dscale = float(data_a[a_offset + ib].d) * float(data_a[a_offset + ib].scales[is]);
+
+    return vec2(dscale * float(int8_t(((data_a[a_offset + ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[a_offset + ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32),
+                dscale * float(int8_t(((data_a[a_offset + ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[a_offset + ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(1, 0);
+}
+#endif

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/{dequant_funcs_cm2.comp → dequant_funcs_cm2.glsl}

@@ -1,5 +1,17 @@
 
-#include "types.comp"
+#include "types.glsl"
+
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufF32 {
+   vec4 block;
+};
+
+float16_t dequantFuncF32(const in decodeBufF32 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const vec4 v = bl.block;
+    const uint idx = coordInBlock[1];
+    const f16vec4 vf16 = f16vec4(v);
+    return vf16[idx];
+}
 
 layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ4_0 {
    block_q4_0_packed16 block;
@@ -108,7 +120,7 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2
 float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
     decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
-    const f16vec2 d = bl.block.d;
+    const f16vec2 dm = bl.block.dm;
     const uint idx = coordInBlock[1];
 
     const uint scalesi = (idx & 0xF0) >> 4;             // 0..15
@@ -119,7 +131,7 @@ float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2
     qs = unpack8(qs)[idx & 1];
 
     const uint scales = bl.block.scales[scalesi];
-    float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
+    float16_t ret = dm.x * float16_t(scales & 0xF) * float16_t(qs) - dm.y * float16_t(scales >> 4);
     return ret;
 }
 
@@ -668,7 +680,7 @@ float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords
     uint32_t qs = bl.block.qs[iqs];
     qs >>= shift;
     qs &= 0xF;
-    float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
+    float16_t ret = float16_t(kvalues_mxfp4[qs] * d * 0.5);
     return ret;
 }
 #endif
@@ -717,4 +729,6 @@ float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords
 #define dequantFuncA dequantFuncIQ4_NL
 #elif defined(DATA_A_MXFP4)
 #define dequantFuncA dequantFuncMXFP4
+#elif defined(DATA_A_F32)
+#define dequantFuncA dequantFuncF32
 #endif

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/{dequant_head.comp → dequant_head.glsl}

@@ -10,4 +10,4 @@ layout (push_constant) uniform parameter
     uint nel;
 } p;
 
-#include "types.comp"
+#include "types.glsl"

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp

@@ -2,7 +2,7 @@
 
 #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 
@@ -29,7 +29,7 @@ void main() {
         uint qs = data_a[ib].qs[4 * ib32 + l];
         const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l];
         qs |= (qh << (8 - 2 * l)) & 0x300;
-        const uvec2 grid = iq2s_grid[qs & 511];
+        const uvec2 grid = iq2s_grid[qs];
         const u8vec4 grid0 = unpack8(grid.x);
         const u8vec4 grid1 = unpack8(grid.y);
         data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0));

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 
@@ -33,7 +33,8 @@ void main() {
     [[unroll]] for (uint l = 0; l < 4; ++l) {
         const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
         const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit
-        const uvec2 grid = iq2xxs_grid[data_a[ib].qs[8 * is + l]];
+        const uint qs = data_a[ib].qs[8 * is + l];
+        const uvec2 grid = iq2xxs_grid[qs];
         const u8vec4 grid0 = unpack8(grid.x);
         const u8vec4 grid1 = unpack8(grid.y);
         data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 
@@ -22,15 +22,16 @@ void main() {
     const uint b_idx = 256 * ib + 32 * is;
 
     const float d = float(data_a[ib].d);
-    const float db = d * (1 + 2 * ((data_a[ib].scales[is] >> (4 * (is % 2))) & 0xf));
+    const float db = d * (1 + 2 * ((data_a[ib].scales[is / 2] >> (4 * (is % 2))) & 0xf));
 
     // We must produce 32 values using 4 sign bytes, 1 qh byte, 8 qs bytes.
     uint qh = data_a[ib].qh[is];
     [[unroll]] for (uint l = 0; l < 8; ++l) {
-        uint qs = data_a[ib].qs[8 * is + l];
-        uint gidx = qs | ((qh << (8 - l)) & 256);
-        uint8_t signs = data_a[ib].signs[8 * is + l / 2] >> (4 * (l & 1));
-        u8vec4 grid = unpack8(iq3s_grid[gidx]);
+        const uint iqs = 8 * is + l;
+        const uint qs = data_a[ib].qs[iqs];
+        const uint gidx = qs | ((qh << (8 - l)) & 256);
+        const uint8_t signs = data_a[ib].signs[iqs / 2] >> (4 * (l & 1));
+        const u8vec4 grid = unpack8(iq3s_grid[gidx]);
         data_b[b_idx + 4 * l + 0] = D_TYPE(db * grid.x * ((signs & 1) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 4 * l + 1] = D_TYPE(db * grid.y * ((signs & 2) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 4 * l + 2] = D_TYPE(db * grid.z * ((signs & 4) != 0 ? -1.0 : 1.0));

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 
@@ -35,8 +35,10 @@ void main() {
         const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
         // Restore parity bit.
         const uint sign8 = sign7 | (bitCount(sign7) << 7);
-        const u8vec4 grid0 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l]]);
-        const u8vec4 grid1 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l + 1]]);
+        const uint qs0 = data_a[ib].qs[8 * is + 2 * l];
+        const uint qs1 = data_a[ib].qs[8 * is + 2 * l + 1];
+        const u8vec4 grid0 = unpack8(iq3xxs_grid[qs0]);
+        const u8vec4 grid1 = unpack8(iq3xxs_grid[qs1]);
         data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0));
         data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0));

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 
@@ -26,7 +26,7 @@ void main() {
     const float d = e8m0_to_fp32(data_a[ib].e);
 
     [[unroll]] for (uint l = 0; l < 8; ++l) {
-        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]);
-        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]);
+        data_b[b_idx + l +  0] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]));
+        data_b[b_idx + l + 16] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]));
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 
@@ -24,8 +24,8 @@ void main() {
         const uint ql_idx = 32 * ip + il;
         const uint8_t qs = data_a[i].qs[32 * ip + il];
 
-        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
-        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].dm.x);
+        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].dm.y);
         data_b[y_idx +  0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
         data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
         data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 
@@ -20,8 +20,8 @@ void main() {
         const uint is = 2 * il;
         const uint n = 4;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].dm.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].dm.y);
 
         const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
         const uint qs_idx = 32*il + n * ir;

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 
@@ -19,8 +19,8 @@ void main() {
         const uint ir = tid % 16;
         const uint is = 2 * il;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].dm.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].dm.y);
 
         const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir;
         const uint qs_idx = 32*il + 2 * ir;

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp

@@ -1,6 +1,6 @@
 #version 450
 
-#include "dequant_head.comp"
+#include "dequant_head.glsl"
 
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 

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

@@ -10,7 +10,7 @@ layout (push_constant) uniform parameter
     uint n_past;
 } p;
 
-#include "types.comp"
+#include "types.glsl"
 
 layout(local_size_x = 1, local_size_y = 512, local_size_z = 1) in;
 

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

@@ -1,7 +1,7 @@
 #version 450
 
-#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/exp.comp

@@ -1,7 +1,8 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "rte.glsl"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

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

@@ -0,0 +1,19 @@
+#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 = fill value
+    data_d[i] = D_TYPE(p.param1);
+}