whispercpp 1.3.1 → 1.3.2

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp

@@ -0,0 +1,31 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { float inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
+    for (int j = 0; j < k; j++) {
+        out_[y + j] = inA[x + j];
+    }
+}
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp

@@ -0,0 +1,38 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 2
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q4_0
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q4_0 get_unaligned_block_q4_0(uint index) {
+    block_q4_0 fres;
+    fres.d = u8BufToFloat16(inA, index);
+    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
+        fres.qs[it] = inA[index+2+it];
+    }
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q4_0 block = get_unaligned_block_q4_0(index);
+    return dequantize_q4_0(block, il);
+}
+
+#include "op_getrows.comp"

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp

@@ -0,0 +1,39 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 2
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q4_1
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q4_1 get_unaligned_block_q4_1(uint index) {
+    block_q4_1 fres;
+    fres.d = u8BufToFloat16(inA, index);
+    fres.m = u8BufToFloat16(inA, index+2);
+    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
+        fres.qs[it] = inA[index+4+it];
+    }
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q4_1 block = get_unaligned_block_q4_1(index);
+    return dequantize_q4_1(block, il);
+}
+
+#include "op_getrows.comp"

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp

@@ -0,0 +1,44 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 16
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q6_k
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q6_k get_unaligned_block_q6_k(uint index) {
+    block_q6_k fres;
+    [[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
+        fres.ql[it] = inA[index + it];
+    }
+    [[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
+        fres.qh[it] = inA[index + QK_K/2 + it];
+    }
+    [[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
+        fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
+    }
+    fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q6_k block = get_unaligned_block_q6_k(index);
+    return dequantize_q6_k(block, il);
+}
+
+#include "op_getrows.comp"

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

@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1024) 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 nb00;
+    int nb01;
+    int nb02;
+    int nb03;
+    int ne10;
+    int ne11;
+    int ne12;
+    int ne13;
+    int nb10;
+    int nb11;
+    int nb12;
+    int nb13;
+    int ne0;
+    int nb0;
+    int nb1;
+    int nb2;
+    int nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint i13 = i03 % pcs.ne13;
+    const uint i12 = i02 % pcs.ne12;
+    const uint i11 = i01 % pcs.ne11;
+
+    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
+    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
+    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1)  / 4);
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
+        const uint i10 = i0 % pcs.ne10;
+        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp

@@ -0,0 +1,69 @@
+#version 450
+
+#include "common.comp"
+
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x_id = 0) in;
+
+layout (binding = 0) readonly buffer tensorInA { float16_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;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne10;
+    int ne11;
+    int ne12;
+    uint nb10;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    int ne0;
+    int ne1;
+    uint r2;
+    uint r3;
+} pcs;
+
+#define N_F16_F32 4
+
+void main() {
+    const uint r0 = gl_WorkGroupID.x;
+    const uint rb = gl_WorkGroupID.y*N_F16_F32;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb03;
+
+    const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
+
+    for (uint row = 0; row < N_F16_F32; ++row) {
+        uint r1 = rb + row;
+        if (r1 >= pcs.ne11) {
+            break;
+        }
+
+        const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
+
+        float sumf = 0;
+        for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
+            sumf += float(inA[x+i]) * float(inB[y+i]);
+        }
+
+        const float all_sum = subgroupAdd(sumf);
+        if (subgroupElect()) {
+            out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp

@@ -0,0 +1,51 @@
+#version 450
+
+#include "common.comp"
+
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#extension GL_EXT_debug_printf : enable
+
+// device subgroup size
+layout (local_size_x_id = 0) in;
+
+layout(binding = 0) readonly buffer tensorInA { float 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 ne11;
+  int ne12;
+  uint nb01;
+  uint nb02;
+  uint nb11;
+  uint nb12;
+  uint nb1;
+  uint nb2;
+}
+pcs;
+
+
+void main() {
+  uvec3 gid = gl_WorkGroupID;
+
+  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
+  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
+
+  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
+  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
+  float sum = 0.0f;
+  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
+      sum += float(inA[x+i]) * float(inB[y+i]);
+  }
+
+  const float all_sum = subgroupAdd(sum);
+  if (subgroupElect()) {
+    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
+  }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp

@@ -0,0 +1,33 @@
+#version 450
+
+#include "common.comp"
+
+#define BLOCKS_IN_QUANT QK4_0
+#define SIZE_OF_BLOCK sizeof_block_q4_0
+#define N_ROWS 4
+
+#include "op_mul_mv_q_n_pre.comp"
+
+// The q4_0 version of this function
+float block_q_n_dot_y(uint block_index, uint yb, uint il) {
+    vec2 acc = vec2(0.0, 0.0);
+    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
+    float d = float(u8BufToFloat16(inA, index));
+    float sumy = 0.0f;
+    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
+        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
+
+        const float yl0 = inB[yb + i];
+        const float yl1 = inB[yb + i + 1];
+        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
+        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
+
+        sumy += yl0 + yl1 + yl8 + yl9;
+
+        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
+        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
+    }
+    return d * (sumy * -8.f + acc[0] + acc[1]);
+}
+
+#include "op_mul_mv_q_n.comp"

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp

@@ -0,0 +1,35 @@
+#version 450
+
+#include "common.comp"
+
+#define BLOCKS_IN_QUANT QK4_1
+#define SIZE_OF_BLOCK sizeof_block_q4_1
+#define N_ROWS 4
+
+#include "op_mul_mv_q_n_pre.comp"
+
+// The q4_1 version of this function
+float block_q_n_dot_y(uint block_index, uint yb, uint il) {
+    vec2 acc = vec2(0.0, 0.0);
+    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
+    float d = float(u8BufToFloat16(inA, index));
+    float m = float(u8BufToFloat16(inA, index+2));
+
+    float sumy = 0.0f;
+    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
+        const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
+
+        const float yl0 = inB[yb + i];
+        const float yl1 = inB[yb + i + 1];
+        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
+        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
+
+        sumy += yl0 + yl1 + yl8 + yl9;
+
+        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
+        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
+#include "op_mul_mv_q_n.comp"

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp

@@ -0,0 +1,140 @@
+#version 450
+
+#include "common.comp"
+
+#define N_DST 4
+#define SIZE_OF_BLOCK sizeof_block_q4_k
+
+layout(local_size_x = 4) in;
+layout(local_size_y = 8) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { block_q4_k 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 ne10;
+    int ne0;
+    int ne1;
+    int ne01;
+    int ne02;
+    int ne12;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    uint r2;
+    uint r3;
+} pcs;
+
+void main() {
+    const uint16_t kmask1 = uint16_t(0x3f3f);
+    const uint16_t kmask2 = uint16_t(0x0f0f);
+    const uint16_t kmask3 = uint16_t(0xc0c0);
+
+    const uint ix = gl_SubgroupInvocationID/8;  // 0...3
+    const uint it = gl_SubgroupInvocationID%8;  // 0...7
+    const uint iq = it/4;     // 0 or 1
+    const uint ir = it%4;     // 0...3
+
+    const uint nb = pcs.ne00/QK_K;
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = r0 * N_DST;
+    const uint ib_row = first_row * nb;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
+    const uint offset1 =        r1*pcs.nb11 + (i12       )*pcs.nb12 + (i13       )*pcs.nb13;
+
+    const uint xblk = offset0 + pcs.inAOff;
+    const uint y = (offset1 / 4) + pcs.inBOff;
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
+    float all_sum = 0.f;
+
+    uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    for (uint ib = ix; ib < nb; ib += 4) {
+        const uint blk_idx = ib + xblk;
+
+        float sumy[4] = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = inB[y4+i+  0]; sumy[0] += yl[i+0];
+            yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
+            yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
+            yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK);
+
+            uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
+            uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
+            uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
+            uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
+            uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
+
+            uint16_t sc16[4];
+            sc16[0] = sc_0 & kmask1;
+            sc16[1] = sc_2 & kmask1;
+            sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
+            sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
+
+            float acc1[4] = {0.f, 0.f, 0.f, 0.f};
+            float acc2[4] = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
+                uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
+                acc1[0] += yl[i+0] * (q1 & 0x000F);
+                acc1[1] += yl[i+1] * (q1 & 0x0F00);
+                acc1[2] += yl[i+8] * (q1 & 0x00F0);
+                acc1[3] += yl[i+9] * (q1 & 0xF000);
+                acc2[0] += yh[i+0] * (q2 & 0x000F);
+                acc2[1] += yh[i+1] * (q2 & 0x0F00);
+                acc2[2] += yh[i+8] * (q2 & 0x00F0);
+                acc2[3] += yh[i+9] * (q2 & 0xF000);
+            }
+
+            uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
+            uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
+            uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
+            uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
+            uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
+            uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
+            uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
+            uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
+
+            float dall = float(inA[blk_idx + row_idx].d);
+            float dmin = float(inA[blk_idx + row_idx].dmin);
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
+                               (acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
+                               (acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
+                               (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
+                dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
+        }
+
+        y4 += 4 * QK_K;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = subgroupAdd(sumf[row]);
+        if (subgroupElect()) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp

@@ -0,0 +1,106 @@
+#version 450
+
+#include "common.comp"
+
+#define SIZE_OF_BLOCK sizeof_block_q6_k
+
+layout(local_size_x_id = 0) in;
+layout(local_size_y_id = 1) 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 ne10;
+    int ne0;
+    int ne1;
+    int ne01;
+    int ne02;
+    int ne12;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    uint r2;
+    uint r3;
+} pcs;
+
+void main() {
+    const uint8_t kmask1 = uint8_t(0x03);
+    const uint8_t kmask2 = uint8_t(0x0C);
+    const uint8_t kmask3 = uint8_t(0x30);
+    const uint8_t kmask4 = uint8_t(0xC0);
+
+    const uint nb = pcs.ne00/QK_K;
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint x = row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
+    const uint yy = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
+
+    float sumf = 0;
+
+    // bits of invocation ID for gl_SubgroupSize=32:
+    //  x   x   x   x   x
+    //  4   3   2   1   0
+    // (     tid     ) ix
+    //  ip (   il    )
+
+    const uint block_stride = gl_SubgroupSize / 16;         // number of blocks each subgroup processes
+    const uint tid  = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
+    const uint ix   = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
+    const uint ip   = tid/8;        // first or second half of block (0 or 1)
+    const uint il   = tid%8;        // each half has 8 parts, one per scale
+    const uint n    = 4;            // 4 scales at a time (and 4 sums)
+    const uint l0   = n*il;         // offset into half-block, 0..28
+    const uint is   = 8*ip + l0/16; // 0, 1, 8, 9
+
+    const uint y_offset = 128*ip + l0;
+    const uint q_offset_l = 64*ip + l0;
+    const uint q_offset_h = 32*ip + l0;
+
+    for (uint i = ix; i < nb; i += block_stride) {
+
+        const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
+
+        const uint qlIndex = q_offset_l;
+        const uint q2Index = qlIndex + QK_K/8;
+        const uint qhIndex = q_offset_h;
+        const uint y = yy + i * QK_K + y_offset;
+
+        float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+        for (uint l = 0; l < n; ++l) {
+            const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
+            const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
+            const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
+
+            sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
+            sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
+            sums[2] += inB[y+l+64] * (int8_t((currentQ1  >> 4) | ((currentQh & kmask3) << 0)) - 32);
+            sums[3] += inB[y+l+96] * (int8_t((currentQ2  >> 4) | ((currentQh & kmask4) >> 2)) - 32);
+        }
+
+        float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
+        sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
+    }
+
+    const float tot = subgroupAdd(sumf);
+    if (subgroupElect()) {
+        out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
+    }
+}