whispercpp 1.3.3 → 1.3.5

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

@@ -1,162 +0,0 @@
-
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
-layout (constant_id = 1) const uint32_t Br = 1;
-layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-layout (constant_id = 4) const uint32_t Clamp = 0;
-layout (constant_id = 5) const uint32_t D_split = 16;
-
-
-layout (push_constant) uniform parameter {
-    uint32_t N;
-    uint32_t KV;
-
-    uint32_t ne1;
-    uint32_t ne2;
-    uint32_t ne3;
-
-    uint32_t neq2;
-    uint32_t neq3;
-    uint32_t nek2;
-    uint32_t nek3;
-    uint32_t nev2;
-    uint32_t nev3;
-    uint32_t nem1;
-
-    uint32_t nb01;
-    uint32_t nb02;
-    uint32_t nb03;
-    uint32_t nb11;
-    uint32_t nb12;
-    uint32_t nb13;
-    uint32_t nb21;
-    uint32_t nb22;
-    uint32_t nb23;
-    uint32_t nb31;
-
-    float scale;
-    float max_bias;
-    float logit_softcap;
-
-    uint32_t mask;
-    uint32_t n_head_log2;
-    float m0;
-    float m1;
-
-    uint32_t gqa_ratio;
-    uint32_t split_kv;
-    uint32_t k_num;
-} p;
-
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
-
-#if defined(A_TYPE_PACKED16)
-#define BINDING_IDX_K 0
-#define BINDING_IDX_V 1
-layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
-#endif
-
-#if defined(DATA_A_Q4_0)
-#define BLOCK_BYTE_SIZE 18
-
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
-    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
-    uint shift = (iqs & 0x10) >> 2;
-    vui_lo >>= shift;
-    vui_hi >>= shift;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
-}
-#endif
-
-#if defined(DATA_A_Q8_0)
-#define BLOCK_BYTE_SIZE 34
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
-    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
-}
-#endif
-
-#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
-
-
-// Store column zero. This is used to save per-row m and L values for split_k.
-ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    if (r < N && c == 0) {
-        uint32_t offset = iq2 + r;
-        data_o[o_offset + offset] = D_TYPE(elem);
-    }
-    return elem;
-}
-
-// Load the slope matrix, indexed by Q's dimension 2.
-ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
-{
-    const uint32_t h = iq2 + (r % p.gqa_ratio);
-
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
-
-    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
-}
-
-uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
-         iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
-         q_stride, k_stride, v_stride, m_stride;
-
-void init_indices()
-{
-    N = p.N;
-    KV = p.KV;
-
-    i = gl_WorkGroupID.x;
-    split_k_index = 0;
-
-    if (p.k_num > 1) {
-        i = 0;
-        split_k_index = gl_WorkGroupID.x;
-    }
-
-    Tr = CEIL_DIV(N, Br);
-
-    start_j = split_k_index * p.split_kv / Bc;
-    end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
-
-    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
-    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
-    iq2 = gl_WorkGroupID.y * p.gqa_ratio;
-    iq3 = gl_WorkGroupID.z;
-
-    // broadcast factors
-    rk2 = p.neq2/p.nek2;
-    rk3 = p.neq3/p.nek3;
-
-    rv2 = p.neq2/p.nev2;
-    rv3 = p.neq3/p.nev3;
-
-    // k indices
-    ik3 = iq3 / rk3;
-    ik2 = iq2 / rk2;
-
-    // v indices
-    iv3 = iq3 / rv3;
-    iv2 = iq2 / rv2;
-
-    // nb?1 are already divided by the type size and are in units of elements.
-    // When using grouped query attention, Q is indexed by iq2, so the stride
-    // should be nb02 (which is in bytes).
-    q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
-    k_stride = p.nb11;
-    v_stride = p.nb21;
-    // When using grouped query attention, all rows use the same mask (stride 0).
-    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
-    // that prevents the compiler from folding the "&" through the select
-    // and breaking the alignment detection.
-    m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
-}

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

@@ -1,118 +0,0 @@
-#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_mmq_funcs.comp

@@ -1,99 +0,0 @@
-#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.comp"
-
-// Each iqs value maps to a 32-bit integer
-
-#if defined(DATA_A_Q4_0)
-i32vec2 repack(uint ib, uint iqs) {
-    // Use 2-byte loads since a q4_0 block (18 bytes) is not divisible by 4
-    const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2    ],
-                                   data_a[ib].qs[iqs * 2 + 1]);
-    const uint32_t vui = pack32(quants);
-    return i32vec2( vui       & 0x0F0F0F0F,
-                   (vui >> 4) & 0x0F0F0F0F);
-}
-
-ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y));
-}
-#endif
-
-#if defined(DATA_A_Q4_1)
-i32vec2 repack(uint ib, uint iqs) {
-    // Use 4-byte loads since a q4_1 block (20 bytes) is divisible by 4
-    const uint32_t vui = data_a_packed32[ib].qs[iqs];
-    return i32vec2( vui       & 0x0F0F0F0F,
-                   (vui >> 4) & 0x0F0F0F0F);
-}
-
-ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
-}
-#endif
-
-#if defined(DATA_A_Q5_0)
-i32vec2 repack(uint ib, uint iqs) {
-    // Use 2-byte loads since a q5_0 block (22 bytes) is not divisible by 4
-    const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2    ],
-                                   data_a[ib].qs[iqs * 2 + 1]);
-    const uint32_t vui = pack32(quants);
-    const int32_t qh = int32_t((uint32_t(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]) >> (4 * iqs));
-    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
-                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
-
-    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
-                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
-
-    return i32vec2(v0, v1);
-}
-
-ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y));
-}
-#endif
-
-#if defined(DATA_A_Q5_1)
-i32vec2 repack(uint ib, uint iqs) {
-    // Use 4-byte loads since a q5_1 block (24 bytes) is divisible by 4
-    const uint32_t vui = data_a_packed32[ib].qs[iqs];
-    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
-    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
-                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
-
-    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
-                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
-
-    return i32vec2(v0, v1);
-}
-
-ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
-}
-#endif
-
-#if defined(DATA_A_Q8_0)
-int32_t repack(uint ib, uint iqs) {
-    // Use 2-byte loads since a q8_0 block (34 bytes) is not divisible by 4
-    return pack32(i16vec2(data_a[ib].qs[iqs * 2    ],
-                          data_a[ib].qs[iqs * 2 + 1]));
-}
-
-ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
-    return ACC_TYPE(float(q_sum) * da * dsb.x);
-}
-#endif
-
-#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
-FLOAT_TYPE get_d(uint ib) {
-    return FLOAT_TYPE(data_a[ib].d);
-}
-#endif
-
-#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
-FLOAT_TYPE_VEC2 get_dm(uint ib) {
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
-}
-#endif

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

@@ -1,58 +0,0 @@
-#include "types.comp"
-
-#extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_spirv_intrinsics: enable
-
-#if RTE16
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif
-
-layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer Y {int data_pos[];};
-layout (binding = 2) readonly buffer Z {float data_ff[];};
-layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
-
-layout (push_constant) uniform parameter {
-    uint ncols;
-    uint n_dims;
-    float freq_scale;
-    uint p_delta_rows;
-    float freq_base;
-    float ext_factor;
-    float attn_factor;
-    float corr_dims[2];
-    float theta_scale;
-    uint has_ff;
-    uint ne02;
-    uint s1;
-    uint s2;
-    int sections[4];
-    uint is_back;
-} p;
-
-float rope_yarn_ramp(const float low, const float high, const uint i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta) {
-    float mscale = p.attn_factor;
-    // Get n-d rotational scaling corrected for extrapolation
-    float theta_interp = p.freq_scale * theta_extrap;
-    float theta = theta_interp;
-    if (p.ext_factor != 0.0f) {
-        float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
-        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-
-        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
-    }
-    // Backprogagation uses inverted rotation
-    if (p.is_back != 0) {
-        theta = -theta;
-    }
-    cos_theta = cos(theta) * mscale;
-    sin_theta = sin(theta) * mscale;
-}