whispercpp 1.3.4 → 1.3.6

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl

@@ -0,0 +1,454 @@
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+
+#include "types.glsl"
+
+// Each iqs value maps to a 32-bit integer
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q4_1)
+// 2-byte loads for Q4_0 blocks (18 bytes)
+// 4-byte loads for Q4_1 blocks (20 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+#ifdef DATA_A_Q4_0
+    buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
+                                           data_a_packed16[ib].qs[iqs * 2 + 1]));
+
+    if (iqs == 0) {
+        buf_a[buf_ib].dm = FLOAT_TYPE(data_a_packed16[ib].d);
+    }
+#else // DATA_A_Q4_1
+    buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
+
+    if (iqs == 0) {
+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+    }
+#endif
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
+
+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    int32_t q_sum = 0;
+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
+        const uint32_t vui = cache_a[ib_a].qs[iqs];
+        const i32vec2 qs_a = i32vec2( vui       & 0x0F0F0F0F,
+                                     (vui >> 4) & 0x0F0F0F0F);
+
+        const int32_t qs_b0 = cache_b.qs[iqs];
+        const int32_t qs_b1 = cache_b.qs[iqs + 4];
+
+        q_sum += dotPacked4x8EXT(qs_a.x, qs_b0);
+        q_sum += dotPacked4x8EXT(qs_a.y, qs_b1);
+    }
+
+#ifdef DATA_A_Q4_0
+    return ACC_TYPE(float(cache_a[ib_a].dm) * (float(q_sum) * float(cache_b.ds.x) - 8.0 * float(cache_b.ds.y)));
+#else // DATA_A_Q4_1
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm.x) * float(cache_b.ds.x) + float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
+#endif
+}
+#endif
+
+#if defined(DATA_A_Q5_0) || defined(DATA_A_Q5_1)
+// 2-byte loads for Q5_0 blocks (22 bytes)
+// 4-byte loads for Q5_1 blocks (24 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+#ifdef DATA_A_Q5_0
+    buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
+                                           data_a_packed16[ib].qs[iqs * 2 + 1]));
+
+    if (iqs == 0) {
+        buf_a[buf_ib].dm = FLOAT_TYPE(data_a_packed16[ib].d);
+        buf_a[buf_ib].qh = pack32(u16vec2(data_a_packed16[ib].qh[0], data_a_packed16[ib].qh[1]));
+    }
+#else // DATA_A_Q5_1
+    buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
+
+    if (iqs == 0) {
+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+        buf_a[buf_ib].qh = data_a_packed32[ib].qh;
+    }
+#endif
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
+    cache_a[reg_ib].qh = buf_a[buf_ib].qh;
+
+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    int32_t q_sum = 0;
+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
+        const uint32_t vui = cache_a[ib_a].qs[iqs];
+        const int32_t qh = int32_t(cache_a[ib_a].qh >> (4 * iqs));
+        const int32_t qs_a0 = int32_t(vui & 0x0F0F0F0F)
+                         | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
+        const int32_t qs_a1 = int32_t((vui >> 4) & 0x0F0F0F0F)
+                         | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
+
+        const int32_t qs_b0 = cache_b.qs[iqs];
+        const int32_t qs_b1 = cache_b.qs[iqs + 4];
+
+        q_sum += dotPacked4x8EXT(qs_a0, qs_b0);
+        q_sum += dotPacked4x8EXT(qs_a1, qs_b1);
+    }
+
+#ifdef DATA_A_Q5_0
+    return ACC_TYPE(float(cache_a[ib_a].dm) * (float(q_sum) * float(cache_b.ds.x) - 16.0 * float(cache_b.ds.y)));
+#else // DATA_A_Q5_1
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm.x) * float(cache_b.ds.x) + float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
+#endif
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+// 2-byte loads for Q8_0 blocks (34 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+    buf_a[buf_ib].qs[iqs] = pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2],
+                                           data_a_packed16[ib].qs[iqs * 2 + 1]));
+
+    if (iqs == 0) {
+        buf_a[buf_ib].dm = FLOAT_TYPE(data_a_packed16[ib].d);
+    }
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
+
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    int32_t q_sum = 0;
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
+        const int32_t qs_b = cache_b.qs[iqs];
+
+        q_sum += dotPacked4x8EXT(qs_a, qs_b);
+    }
+
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm) * float(cache_b.ds.x));
+}
+#endif
+
+#if defined(DATA_A_MXFP4)
+// 1-byte loads for mxfp4 blocks (17 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+    const uint32_t qs = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
+                                      data_a[ib].qs[iqs * 4 + 1],
+                                      data_a[ib].qs[iqs * 4 + 2],
+                                      data_a[ib].qs[iqs * 4 + 3]));
+
+    const u8vec4 i_a0 = unpack8( qs       & 0x0F0F0F0F);
+    const u8vec4 i_a1 = unpack8((qs >> 4) & 0x0F0F0F0F);
+
+    buf_a[buf_ib].qs[iqs    ] = pack32(i8vec4(kvalues_mxfp4[i_a0.x], kvalues_mxfp4[i_a0.y], kvalues_mxfp4[i_a0.z], kvalues_mxfp4[i_a0.w]));
+    buf_a[buf_ib].qs[iqs + 4] = pack32(i8vec4(kvalues_mxfp4[i_a1.x], kvalues_mxfp4[i_a1.y], kvalues_mxfp4[i_a1.z], kvalues_mxfp4[i_a1.w]));
+
+    if (iqs == 0) {
+        buf_a[buf_ib].d = FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e) * 0.5);
+    }
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].d = buf_a[buf_ib].d;
+
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    int32_t q_sum = 0;
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
+
+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
+    }
+
+    return ACC_TYPE(float(cache_a[ib_a].d) * float(cache_b.ds.x) * float(q_sum));
+}
+#endif
+
+// For k-quants, ib and iqs still assume 32-wide blocks, but k-quants are 256-wide
+// iqs still refers to a 32-bit integer, meaning 0..7 for 32-wide quants
+#if defined(DATA_A_Q2_K)
+// 4-byte loads for Q2_K blocks (84 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
+
+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
+
+    // Repack 4x4 quants into one int
+    const uint32_t vals0 = (data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x03030303;
+    const uint32_t vals1 = (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x03030303;
+    const uint32_t vals2 = (data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x03030303;
+    const uint32_t vals3 = (data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x03030303;
+
+    buf_a[buf_ib].qs[iqs] = vals0 | (vals1 << 2) | (vals2 << 4) | (vals3 << 6);
+
+    if (iqs == 0) {
+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
+        buf_a[buf_ib].scales = unpack8(uint32_t(data_a_packed16[ib_k].scales[iqs_k / 8])).xy; // vec4 used due to #12147
+    }
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
+    cache_a[reg_ib].scales = buf_a[buf_ib].scales;
+
+    [[unroll]] for (uint iqs = 0; iqs < 2; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    int32_t sum_d = 0;
+    int32_t sum_m = 0;
+
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        const uint8_t scale = cache_a[ib_a].scales[iqs / 4];
+        const int32_t scale_m = int32_t(scale >> 4) * 0x01010101; // Duplicate 8-bit value across 32-bits.
+        const int32_t qs_a = int32_t((cache_a[ib_a].qs[iqs / 4] >> ((iqs % 4) * 2)) & 0x03030303);
+
+        sum_d += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]) * (scale & 0xF);
+        sum_m += dotPacked4x8EXT(scale_m, cache_b.qs[iqs]);
+    }
+
+    return ACC_TYPE(float(cache_b.ds.x) * (float(cache_a[ib_a].dm.x) * float(sum_d) - float(cache_a[ib_a].dm.y) * float(sum_m)));
+}
+#endif
+
+#if defined(DATA_A_Q3_K)
+// 2-byte loads for Q3_K blocks (110 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint hm_idx = iqs * QUANT_R_MMQ;
+    const uint iqs_k = (ib % 8) * 8 + hm_idx;
+
+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
+    const uint hm_shift = iqs_k / 8;
+
+    // Repack 2x4 quants into one int
+    // Add the 3rd bit instead of subtracting it to allow packing the quants
+    // vec4 for unpack8 used due to #12147
+    const i8vec2 vals00 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2        ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2    ] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals01 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 1    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 1] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals10 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 2    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 2] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals11 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 3    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 3] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    buf_a[buf_ib].qs[iqs] = pack32(u8vec4(vals00.x, vals00.y, vals01.x, vals01.y)) |
+                           (pack32(u8vec4(vals10.x, vals10.y, vals11.x, vals11.y)) << 4);
+
+    if (iqs == 0) {
+        const uint is = iqs_k / 4;
+        const i8vec2 scales = i8vec2(unpack8(uint32_t(((data_a_packed16[ib_k].scales[(is % 8      ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
+                                                     (((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4))).xy); // vec4 used due to #12147
+
+        buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales - 32));
+    }
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].d_scales = buf_a[buf_ib].d_scales;
+
+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    float result = 0.0;
+    int32_t q_sum = 0;
+
+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
+        // Subtract 4 from the quants to correct the 3rd bit offset
+        const int32_t qs_a = pack32(unpack8(int32_t((cache_a[ib_a].qs[iqs / 2] >> ((iqs % 2) * 4)) & 0x0F0F0F0F)) - int8_t(4));
+
+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
+    }
+    result += float(cache_a[ib_a].d_scales[0]) * float(q_sum);
+    q_sum = 0;
+
+    [[unroll]] for (uint iqs = 4; iqs < 8; iqs++) {
+        const int32_t qs_a = pack32(unpack8(int32_t((cache_a[ib_a].qs[iqs / 2] >> ((iqs % 2) * 4)) & 0x0F0F0F0F)) - int8_t(4));
+
+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
+    }
+    result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
+
+    return ACC_TYPE(float(cache_b.ds.x) * result);
+}
+#endif
+
+#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
+// 4-byte loads for Q4_K blocks (144 bytes) and Q5_K blocks (176 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
+
+    const uint qs_idx = (iqs_k / 16) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 16) / 8) * 4;
+
+    // Repack 2x4 quants into one int
+#if defined(DATA_A_Q4_K)
+    const uint32_t vals0 = (data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x0F0F0F0F;
+    const uint32_t vals1 = (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x0F0F0F0F;
+
+    buf_a[buf_ib].qs[iqs] = vals0 | (vals1 << 4);
+#else // defined(DATA_A_Q5_K)
+    const uint qh_idx = iqs * QUANT_R_MMQ;
+    const uint qh_shift = iqs_k / 8;
+
+    buf_a[buf_ib].qs[iqs] = int32_t(((data_a_packed32[ib_k].qs[qs_idx] >> qs_shift) & 0x0F0F0F0F) |
+                                   (((data_a_packed32[ib_k].qh[qh_idx] >> qh_shift) & 0x01010101) << 4));
+#endif
+
+    if (iqs == 0) {
+        // Scale index
+        const uint is = iqs_k / 8;
+        u8vec2 scale_dm;
+        if (is < 4) {
+            scale_dm = u8vec2(data_a[ib_k].scales[is] & 0x3F, data_a[ib_k].scales[is + 4] & 0x3F);
+        } else {
+            scale_dm = u8vec2((data_a[ib_k].scales[is+4] & 0xF) | ((data_a[ib_k].scales[is-4] & 0xC0) >> 2),
+                              (data_a[ib_k].scales[is+4] >>  4) | ((data_a[ib_k].scales[is  ] & 0xC0) >> 2));
+        }
+
+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(vec2(data_a_packed32[ib_k].dm) * vec2(scale_dm));
+    }
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
+
+    [[unroll]] for (uint iqs = 0; iqs < 8 / QUANT_R_MMQ; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    int32_t q_sum = 0;
+
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+#if defined(DATA_A_Q4_K)
+        const int32_t qs_a = int32_t((cache_a[ib_a].qs[iqs / 2] >> ((iqs % 2) * 4)) & 0x0F0F0F0F);
+#else // defined(DATA_A_Q5_K)
+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
+#endif
+
+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
+    }
+
+    return ACC_TYPE(float(cache_b.ds.x) * float(cache_a[ib_a].dm.x) * float(q_sum) - float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
+}
+#endif
+
+#if defined(DATA_A_Q6_K)
+// 2-byte loads for Q6_K blocks (210 bytes)
+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint ql_idx = (iqs_k / 32) * 16 + iqs_k % 16;
+    const uint ql_shift = ((iqs_k % 32) / 16) * 4;
+
+    const uint qh_idx = (iqs_k / 32) * 8 + iqs;
+    const uint qh_shift = ((iqs_k % 32) / 8) * 2;
+
+    const i8vec2 vals00 = (unpack8(int32_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))).xy |
+                          unpack8(int32_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4)).xy) - int8_t(32);
+    const i8vec2 vals01 = (unpack8(int32_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))).xy |
+                          unpack8(int32_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4)).xy) - int8_t(32);
+    buf_a[buf_ib].qs[iqs] = pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y));
+
+    if (iqs == 0) {
+        const uint is = iqs_k / 4;
+        const i8vec2 scales = unpack8(int32_t(data_a_packed16[ib_k].scales[is / 2])).xy;
+
+        buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales));
+    }
+}
+
+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
+    cache_a[reg_ib].d_scales = buf_a[buf_ib].d_scales;
+
+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
+    }
+}
+
+ACC_TYPE mmq_dot_product(const uint ib_a) {
+    float result = 0.0;
+    int32_t q_sum = 0;
+
+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
+
+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
+    }
+    result += float(cache_a[ib_a].d_scales[0]) * float(q_sum);
+    q_sum = 0;
+
+    [[unroll]] for (uint iqs = 4; iqs < 8; iqs++) {
+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
+
+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
+    }
+    result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
+
+    return ACC_TYPE(float(cache_b.ds.x) * result);
+}
+#endif
+
+void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bool is_in_bounds) {
+    if (is_in_bounds) {
+        const uint ib_outer = ib / 4;
+        const uint ib_inner = ib % 4;
+
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
+        }
+
+        const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
+        buf_b[buf_ib].qs[iqs * 4    ] = values.x;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
+    } else {
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
+        }
+
+        buf_b[buf_ib].qs[iqs * 4    ] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = 0;
+    }
+}
+
+void block_b_to_registers(const uint ib) {
+    cache_b.ds = buf_b[ib].ds;
+    [[unroll]] for (uint iqs = 0; iqs < BK / 4; iqs++) {
+        cache_b.qs[iqs] = buf_b[ib].qs[iqs];
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl

@@ -0,0 +1,78 @@
+#if defined(DATA_A_Q4_0)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    FLOAT_TYPE dm;
+};
+#elif defined(DATA_A_Q4_1)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q5_0)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    uint32_t qh;
+    FLOAT_TYPE dm;
+};
+#elif defined(DATA_A_Q5_1)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    uint32_t qh;
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q8_0)
+#define QUANT_R_MMQ 1
+// AMD likes 4, Intel likes 1 and Nvidia likes 2
+// #define BK_STEP 1
+struct block_a_cache {
+    int32_t qs[32/4];
+    FLOAT_TYPE dm;
+};
+#elif defined(DATA_A_MXFP4)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE d;
+};
+#elif defined(DATA_A_Q2_K)
+#define QUANT_R_MMQ 4
+struct block_a_cache {
+    uint32_t qs[2];
+    u8vec2 scales;
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q3_K)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[4];
+    FLOAT_TYPE_VEC2 d_scales;
+};
+#elif defined(DATA_A_Q4_K)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[4];
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q5_K)
+#define QUANT_R_MMQ 1
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q6_K)
+#define QUANT_R_MMQ 1
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE_VEC2 d_scales;
+};
+#endif
+
+struct block_b_cache
+{
+    int32_t qs[8];
+    FLOAT_TYPE_VEC2 ds;
+};

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

@@ -8,9 +8,9 @@
 #extension GL_KHR_shader_subgroup_basic : enable
 #endif
 
-#include "rte.comp"
-#include "types.comp"
-#include "utils.comp"
+#include "rte.glsl"
+#include "types.glsl"
+#include "utils.glsl"
 
 layout (push_constant) uniform parameter2
 {
@@ -23,16 +23,100 @@ layout (push_constant) uniform parameter2
     uint rms_partials;
 } p;
 
-// Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
-// layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
-// layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
-layout (binding = 0) buffer A {A_TYPE data_a[];} a[];
-layout (binding = 0) buffer D {D_TYPE data_d[];} d[];
-
-layout (binding = 0, std430) buffer PartialBuf {float partial_sums[];} partials[];
+// No readonly/writeonly decorations. Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
+layout (binding = 0)  buffer A0 {A_TYPE data_a[];} a0;
+layout (binding = 1)  buffer A1 {A_TYPE data_a[];} a1;
+layout (binding = 2)  buffer A2 {A_TYPE data_a[];} a2;
+layout (binding = 3)  buffer A3 {A_TYPE data_a[];} a3;
+layout (binding = 4)  buffer A4 {A_TYPE data_a[];} a4;
+layout (binding = 5)  buffer A5 {A_TYPE data_a[];} a5;
+layout (binding = 6)  buffer A6 {A_TYPE data_a[];} a6;
+layout (binding = 7)  buffer A7 {A_TYPE data_a[];} a7;
+layout (binding = 8)  buffer A8 {A_TYPE data_a[];} a8;
+layout (binding = 9)  buffer A9 {A_TYPE data_a[];} a9;
+layout (binding = 10) buffer A10 {A_TYPE data_a[];} a10;
+layout (binding = 11) buffer A11 {A_TYPE data_a[];} a11;
+layout (binding = 0)  buffer D0 {D_TYPE data_d[];} d0;
+layout (binding = 1)  buffer D1 {D_TYPE data_d[];} d1;
+layout (binding = 2)  buffer D2 {D_TYPE data_d[];} d2;
+layout (binding = 3)  buffer D3 {D_TYPE data_d[];} d3;
+layout (binding = 4)  buffer D4 {D_TYPE data_d[];} d4;
+layout (binding = 5)  buffer D5 {D_TYPE data_d[];} d5;
+layout (binding = 6)  buffer D6 {D_TYPE data_d[];} d6;
+layout (binding = 7)  buffer D7 {D_TYPE data_d[];} d7;
+layout (binding = 8)  buffer D8 {D_TYPE data_d[];} d8;
+layout (binding = 9)  buffer D9 {D_TYPE data_d[];} d9;
+layout (binding = 10) buffer D10 {D_TYPE data_d[];} d10;
+layout (binding = 11) buffer D11 {D_TYPE data_d[];} d11;
+layout (binding = 0, std430)  buffer PartialBuf0 {float partial_sums[];} partials0;
+layout (binding = 1, std430)  buffer PartialBuf1 {float partial_sums[];} partials1;
+layout (binding = 2, std430)  buffer PartialBuf2 {float partial_sums[];} partials2;
+layout (binding = 3, std430)  buffer PartialBuf3 {float partial_sums[];} partials3;
+layout (binding = 4, std430)  buffer PartialBuf4 {float partial_sums[];} partials4;
+layout (binding = 5, std430)  buffer PartialBuf5 {float partial_sums[];} partials5;
+layout (binding = 6, std430)  buffer PartialBuf6 {float partial_sums[];} partials6;
+layout (binding = 7, std430)  buffer PartialBuf7 {float partial_sums[];} partials7;
+layout (binding = 8, std430)  buffer PartialBuf8 {float partial_sums[];} partials8;
+layout (binding = 9, std430)  buffer PartialBuf9 {float partial_sums[];} partials9;
+layout (binding = 10, std430) buffer PartialBuf10 {float partial_sums[];} partials10;
+layout (binding = 11, std430) buffer PartialBuf11 {float partial_sums[];} partials11;
 
 layout(constant_id = 0) const uint num_srcs = 2;
 
+FLOAT_TYPE load_a(uint b, uint i) {
+    switch (b) {
+    case 0:  return FLOAT_TYPE(a0.data_a[i]);
+    case 1:  return FLOAT_TYPE(a1.data_a[i]);
+    case 2:  return FLOAT_TYPE(a2.data_a[i]);
+    case 3:  return FLOAT_TYPE(a3.data_a[i]);
+    case 4:  return FLOAT_TYPE(a4.data_a[i]);
+    case 5:  return FLOAT_TYPE(a5.data_a[i]);
+    case 6:  return FLOAT_TYPE(a6.data_a[i]);
+    case 7:  return FLOAT_TYPE(a7.data_a[i]);
+    case 8:  return FLOAT_TYPE(a8.data_a[i]);
+    case 9:  return FLOAT_TYPE(a9.data_a[i]);
+    case 10: return FLOAT_TYPE(a10.data_a[i]);
+    case 11: return FLOAT_TYPE(a11.data_a[i]);
+    default: return FLOAT_TYPE(0);
+    }
+}
+
+void store_d(uint b, uint i, FLOAT_TYPE v) {
+    switch (b) {
+    case 0:  d0.data_d[i] = D_TYPE(v); break;
+    case 1:  d1.data_d[i] = D_TYPE(v); break;
+    case 2:  d2.data_d[i] = D_TYPE(v); break;
+    case 3:  d3.data_d[i] = D_TYPE(v); break;
+    case 4:  d4.data_d[i] = D_TYPE(v); break;
+    case 5:  d5.data_d[i] = D_TYPE(v); break;
+    case 6:  d6.data_d[i] = D_TYPE(v); break;
+    case 7:  d7.data_d[i] = D_TYPE(v); break;
+    case 8:  d8.data_d[i] = D_TYPE(v); break;
+    case 9:  d9.data_d[i] = D_TYPE(v); break;
+    case 10: d10.data_d[i] = D_TYPE(v); break;
+    case 11: d11.data_d[i] = D_TYPE(v); break;
+    default: break;
+    }
+}
+
+void store_partial(uint b, uint i, float v) {
+    switch (b) {
+    case 0:  partials0.partial_sums[i] = v; break;
+    case 1:  partials1.partial_sums[i] = v; break;
+    case 2:  partials2.partial_sums[i] = v; break;
+    case 3:  partials3.partial_sums[i] = v; break;
+    case 4:  partials4.partial_sums[i] = v; break;
+    case 5:  partials5.partial_sums[i] = v; break;
+    case 6:  partials6.partial_sums[i] = v; break;
+    case 7:  partials7.partial_sums[i] = v; break;
+    case 8:  partials8.partial_sums[i] = v; break;
+    case 9:  partials9.partial_sums[i] = v; break;
+    case 10: partials10.partial_sums[i] = v; break;
+    case 11: partials11.partial_sums[i] = v; break;
+    default: break;
+    }
+}
+
 uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
     return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
 }
@@ -78,10 +162,10 @@ void main() {
 
         FLOAT_TYPE sum = FLOAT_TYPE(0);
         [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
-            sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
+            sum += load_a(s, src_idx(s, i00, i01, i02, i03));
         }
         sum_sq += sum*sum;
-        d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
+        store_d(num_srcs, dst_idx(i00, i01, i02, i03), sum);
 
         idx += num_threads;
     }
@@ -104,7 +188,7 @@ void main() {
         }
 
         if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
-            partials[num_srcs + 1].partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+            store_partial(num_srcs + 1, orig_idx / (num_iter * num_threads), sum_sq);
         }
     }
 #endif

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

@@ -0,0 +1,20 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+    data_d[i] = D_TYPE(-float(data_a[i]));
+}