whispercpp 1.3.1 → 1.3.2

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp

@@ -0,0 +1,73 @@
+#version 450
+
+#include "common.comp"
+
+#include "op_mul_mv_q_n_pre.comp"
+
+#define SIZE_OF_D 2
+
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+#define NB_Q8_0 8
+
+void main() {
+    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const int nr  = N_DST;
+    const int nsg = N_SIMDGROUP;
+    const int nw  = N_SIMDWIDTH;
+
+    const int nb = pcs.ne00/QK8_0;
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+
+    const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
+    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
+
+    float yl[NB_Q8_0];
+    float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
+
+    const uint ix = gl_SubgroupInvocationID.x/4;
+    const uint il = gl_SubgroupInvocationID.x%4;
+
+    uint yb = y + ix * QK8_0 + NB_Q8_0*il;
+
+    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
+    for (uint ib = ix; ib < nb; ib += nw/4) {
+        for (int i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = inB[yb + i];
+        }
+
+        for (int row = 0; row < nr; row++) {
+            const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
+            float sumq = 0.f;
+            for (int iq = 0; iq < NB_Q8_0; ++iq) {
+                const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
+                sumq += qs_iq * yl[iq];
+            }
+            const float16_t d = u8BufToFloat16(inA, x + block_offset);
+            sumf[row] += sumq*d;
+        }
+
+        yb += NB_Q8_0 * nw;
+    }
+
+    for (int row = 0; row < nr; ++row) {
+        const float tot = subgroupAdd(sumf[row]);
+        if (subgroupElect() && first_row + row < pcs.ne01) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp

@@ -0,0 +1,52 @@
+void main() {
+    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    // pointers to src0 rows
+    uint ax[N_ROWS];
+    for (int row = 0; row < N_ROWS; ++row) {
+        const uint offset0 = (first_row + row)*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
+
+        ax[row] = offset0 + pcs.inAOff;
+    }
+
+    const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
+
+    float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
+
+    const uint ix = gl_SubgroupInvocationID/2;
+    const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
+
+    uint yb = y + ix * BLOCKS_IN_QUANT + il;
+
+    //debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
+    //    gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
+    //    gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
+
+    for (uint ib = ix; ib < nb; ib += 16) {
+        for (int row = 0; row < N_ROWS; row++) {
+            sumf[row] += block_q_n_dot_y(ax[row] + ib, yb, il);
+        }
+
+        yb += BLOCKS_IN_QUANT * 16;
+    }
+
+    for (int row = 0; row < N_ROWS; ++row) {
+        const float tot = subgroupAdd(sumf[row]);
+        if (first_row + row < pcs.ne01 && subgroupElect()) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp

@@ -0,0 +1,28 @@
+layout(local_size_x_id = 0) in;
+layout(local_size_y = 8) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int  ne00;
+    int  ne01;
+    int  ne02;
+    int  ne10;
+    int  ne12;
+    int  ne0;
+    int  ne1;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    uint r2;
+    uint r3;
+} pcs;

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp

@@ -0,0 +1,84 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 256) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+    // MEAN
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float mean = sum[0];
+
+    // recenter
+    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] = in_[x+i00] - mean;
+    }
+
+    // VARIANCE
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float variance = sum[0];
+
+    const float scale = 1.0f/sqrt(variance + pcs.eps);
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] *= scale;
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp

@@ -0,0 +1,21 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} 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] = max(0.0, in_[i + pcs.inOff]);
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp

@@ -0,0 +1,53 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 512) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+
+    const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
+
+    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] = in_[x+i00] * scale;
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+pcs.n_dims/2]);
+
+            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+pcs.n_dims/2];
+
+            out_[dst_data]              = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+1]);
+
+            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+1];
+
+            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp

@@ -0,0 +1,19 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    float scale;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp

@@ -0,0 +1,23 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    float scale;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 8;
+
+    for (uint x = 0; x < 8; x++) {
+        const uint i = baseIndex + x;
+        out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 4;
+
+    for (uint x = 0; x < 4; x++) {
+        const uint i = baseIndex + x;
+        const float y = in_[i + pcs.inOff];
+        out_[i + pcs.outOff] = y / (1.0 + exp(-y));
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp

@@ -0,0 +1,72 @@
+// TODO: implement multi-simd softmax (llama.cpp commit e16b9fa4)
+
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x_id = 0) 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;
+    int ne00;
+    int ne01;
+    int ne02;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+    uint n_head_log2;
+    int mask;
+} pcs;
+
+void main() {
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
+    const uint psrc0 = extra_off + pcs.inAOff; // Based from inA
+    const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
+    const uint pdst = extra_off + pcs.outOff; // Based from out_
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (pcs.max_bias > 0.0f) {
+        int64_t h = i02;
+
+        float base = h < pcs.n_head_log2 ? pcs.m0 : pcs.m1;
+        int64_t exp = h < pcs.n_head_log2 ? h + 1 : 2*(h - pcs.n_head_log2) + 1;
+
+        slope = pow(base, float(exp));
+    }
+
+    // parallel max
+    float localMax = uintBitsToFloat(0xFF800000);
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f));
+    }
+    float max_ = subgroupMax(localMax);
+
+    // parallel sum
+    float localSum = 0.0f;
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f) - max_);
+        localSum += exp_psrc0;
+        out_[pdst + i00] = exp_psrc0;
+    }
+
+    const float sum = subgroupAdd(localSum);
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        out_[pdst + i00] /= sum;
+    }
+}