whispercpp 1.3.1 → 1.3.2

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -0,0 +1,169 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16)
+#define K_PER_ITER 8
+#else
+#define K_PER_ITER 2
+#endif
+
+
+uint a_offset, b_offset, d_offset, y_offset;
+
+void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
+{
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
+        const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
+        const uint iybs = col - col%QUANT_K; // y block start index
+
+#if K_PER_ITER == 8
+#if QUANT_R == 2
+        const vec4 bv02 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]);
+        const vec4 bv13 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4]);
+        const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
+        const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
+#else
+        const vec4 bv0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]);
+        const vec4 bv1 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4 + 1]);
+#endif
+#else
+        // Check if the second of the pair of elements is OOB, and don't fetch B or
+        // accumulate it. We still fetch a pair of elements for A, which is fine for
+        // quantized formats since they'll be within the same block. We should
+        // probably skip fetching the second element for F16/F32, but as of now we
+        // still do.
+        const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
+
+        FLOAT_TYPE b0 = 0, b1 = 0;
+        b0 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]);
+        if (!OOB) {
+            b1 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]);
+        }
+#endif
+        uint ibi = first_row*p.ncols;
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib = (ibi + col)/QUANT_K; // block index
+            ibi += p.ncols;
+
+#if K_PER_ITER == 8
+            vec4 v = dequantize4(ib, iqs, a_offset);
+            vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+
+            const vec2 dm = get_dm(ib, a_offset);
+            if (dm.y != 0) { // quant has min component
+                v = v * dm.x + dm.y;
+                v2 = v2 * dm.x + dm.y;
+            }
+
+            // matrix multiplication
+            FLOAT_TYPE rowtmp = dot(bv0, v);
+            rowtmp += dot(bv1, v2);
+
+            if (dm.y == 0)
+                rowtmp *= dm.x;
+
+            temp[j][n] += rowtmp;
+#else
+            const vec2 v = dequantize(ib, iqs, a_offset);
+
+            // matrix multiplication
+            temp[j][n] = fma(FLOAT_TYPE(v.x), b0, temp[j][n]);
+            if (!OOB) {
+                temp[j][n] = fma(FLOAT_TYPE(v.y), b1, temp[j][n]);
+            }
+#endif
+        }
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    const uint tid = gl_LocalInvocationID.x;
+
+    get_offsets(a_offset, b_offset, d_offset);
+    a_offset /= QUANT_K;
+
+    y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
+    if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
+        num_iters++;
+    }
+    int unroll_count = 4;
+    uint unrolled_iters = num_iters & ~(unroll_count - 1);
+
+#if K_PER_ITER == 2
+    // If the K dimension is odd, we need lastiter==true on the last iteration
+    // so OOB is computed correctly. Skip some unrolling to make that happen.
+    if ((p.ncols & 1) != 0 &&
+        unrolled_iters == num_iters &&
+        unrolled_iters > 0) {
+        unrolled_iters -= unroll_count;
+    }
+#endif
+
+    uint i = 0;
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
+        }
+    }
+
+    unroll_count = 2;
+    unrolled_iters = num_iters & ~(unroll_count - 1);
+
+#if K_PER_ITER == 2
+    if ((p.ncols & 1) != 0 &&
+        unrolled_iters == num_iters &&
+        unrolled_iters > 0) {
+        unrolled_iters -= unroll_count;
+    }
+#endif
+
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
+        }
+    }
+    while (i < num_iters) {
+        iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true);
+        i++;
+    }
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp

@@ -0,0 +1,118 @@
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_8bit_storage : require
+
+#ifdef MUL_MAT_ID
+#define EXPERT_COUNT 8
+#endif
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+#ifdef MUL_MAT_ID
+layout (binding = 3) readonly buffer IDS {int data_ids[];};
+#endif
+
+#include "dequant_funcs.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ncols;
+    uint stride_a;
+    uint stride_b;
+    uint stride_d;
+
+    uint batch_stride_a;
+    uint batch_stride_b;
+    uint batch_stride_d;
+
+#ifdef MUL_MAT_ID
+    uint nei0;
+    uint ne11;
+#else
+    uint ne02;
+    uint ne12;
+    uint broadcast2;
+    uint broadcast3;
+#endif
+} p;
+
+void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.y;
+#else
+    const uint batch_idx = gl_GlobalInvocationID.y;
+#endif
+
+#ifndef MUL_MAT_ID
+    uint batch_idx_a = 0;
+    if (batch_idx != 0) {
+        const uint i13 = batch_idx / p.ne12;
+        const uint i12 = batch_idx % p.ne12;
+
+        const uint i03 = i13 / p.broadcast3;
+        const uint i02 = i12 / p.broadcast2;
+
+        batch_idx_a = i03 * p.ne02 + i02;
+    }
+#else
+    const uint expert_id = data_ids[expert_idx];
+#endif
+
+    a_offset =
+#ifdef MUL_MAT_ID
+            expert_id * p.batch_stride_a;
+#else
+            batch_idx_a * p.batch_stride_a;
+#endif
+    b_offset =
+#ifdef MUL_MAT_ID
+            (expert_idx % p.ne11) * p.stride_b;
+#else
+            batch_idx * p.batch_stride_b;
+#endif
+    d_offset =
+#ifdef MUL_MAT_ID
+            expert_idx * p.stride_d;
+#else
+            batch_idx * p.batch_stride_d;
+#endif
+}
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
+layout (constant_id = 2) const uint NUM_COLS = 1;
+
+shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
+
+void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    // sum up partial sums and write back result
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            tmpsh[j][n][tid] = temp[j][n];
+        }
+    }
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                    tmpsh[j][n][tid] += tmpsh[j][n][tid + s];
+                }
+            }
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
+            }
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp

@@ -0,0 +1,82 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 32 * ib32;
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint16_t[4] scales = data_a[ibi].scales;
+        const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+        const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
+
+        const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1));
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1));
+            const uint qs = data_a[ibi].qs[4 * ib32 + l];
+            const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+            const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1);
+
+            const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+                [[unroll]] for (int k = 0; k < 4; ++k) {
+                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
+                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
+                }
+                temp[j][n] = fma(dl, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

@@ -0,0 +1,79 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 32 * ib32;
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint qh = data_a[ibi].qh[ib32];
+        const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+        const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const uint qs = data_a[ibi].qs[4 * ib32 + l];
+            const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
+            const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+                [[unroll]] for (int k = 0; k < 4; ++k) {
+                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
+                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
+                }
+                temp[j][n] = fma(dl, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp

@@ -0,0 +1,90 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint nibble_shift = 4 * (itid & 1);
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
+        const float db = d * (0.5 + scale) * 0.25;
+
+        const uint qh = data_a[ibi].qh[ib32];
+        const u8vec2 qs16 = unpack8(uint32_t(data_a_packed16[ibi].qs[itid])).xy; // vec4 used due to #12147
+        const u8vec2 sign16 = unpack8(uint32_t(data_a_packed16[ibi].qs[QUANT_K / 16 + itid])).xy;
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint8_t sign = sign16[l];
+            const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300);
+            const uvec2 grid = iq2s_grid[qs];
+            const vec4 grid0 = vec4(unpack8(grid.x));
+            const vec4 grid1 = vec4(unpack8(grid.y));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp

@@ -0,0 +1,87 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint nibble_shift = 4 * (itid & 1);
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
+        const float db = d * (0.5 + scale) * 0.25;
+
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint qs = data_a[ibi].qs[2 * itid + l];
+            const uint sign = qs >> 9;
+            const uint sign7 = bitCount(sign);
+            const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x));
+            const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}