whispercpp 1.3.4 → 1.3.5

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

@@ -1,105 +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(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * 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(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-}
-#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(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * 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(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-}
-#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(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    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_MXFP4)
-FLOAT_TYPE get_d(uint ib) {
-    return FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e));
-}
-#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

@@ -1,55 +0,0 @@
-#include "types.comp"
-
-#extension GL_EXT_shader_16bit_storage : require
-
-#include "rte.comp"
-
-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;
-}

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add.tmpl.wgsl

@@ -1,44 +0,0 @@
-#define(VARIANTS)
-
-[
-  {
-    "REPLS": {
-      "TYPE" : "f32",
-    }
-  },
-  {
-    "REPLS": {
-      "TYPE" : "f16",
-    }
-  }
-]
-
-#end(VARIANTS)
-
-#define(SHADER)
-
-enable f16;
-
-#include "binary_head.tmpl"
-
-@group(0) @binding(0)
-var<storage, read_write> src0: array<{{TYPE}}>;
-
-@group(0) @binding(1)
-var<storage, read_write> src1: array<{{TYPE}}>;
-
-@group(0) @binding(2)
-var<storage, read_write> dst: array<{{TYPE}}>;
-
-@group(0) @binding(3)
-var<uniform> params: Params;
-
-override wg_size: u32;
-@compute @workgroup_size(wg_size)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x < params.ne) {
-        dst[params.offset_dst + gid.x] = src0[params.offset_src0 + gid.x] + src1[params.offset_src1 + src1_index(gid.x)];
-    }
-}
-
-#end(SHADER)

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_in_place.tmpl.wgsl

@@ -1,41 +0,0 @@
-#define(VARIANTS)
-
-[
-  {
-    "REPLS": {
-      "TYPE" : "f32",
-    }
-  },
-  {
-    "REPLS": {
-      "TYPE" : "f16",
-    }
-  }
-]
-
-#end(VARIANTS)
-
-#define(SHADER)
-
-enable f16;
-
-#include "binary_head.tmpl"
-
-@group(0) @binding(0)
-var<storage, read_write> src0: array<{{TYPE}}>;
-
-@group(0) @binding(1)
-var<storage, read_write> src1: array<{{TYPE}}>;
-
-@group(0) @binding(2)
-var<uniform> params: Params;
-
-override wg_size: u32;
-@compute @workgroup_size(wg_size)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x < params.ne) {
-        src0[params.offset_dst + gid.x] = src0[params.offset_src0 + gid.x] + src1[params.offset_src1 + src1_index(gid.x)];
-    }
-}
-
-#end(SHADER)

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl

@@ -1,60 +0,0 @@
-enable f16;
-
-@group(0) @binding(0)
-var<storage, read_write> src: array<f32>;
-
-@group(0) @binding(1)
-var<storage, read_write> dst: array<f16>;
-
-struct Params {
-    ne: u32,            // total number of elements
-    offset_src: u32,    // in elements
-    offset_dst: u32,    // in elements
-
-    // Strides (in elements) — may be permuted
-    stride_src0: u32,
-    stride_src1: u32,
-    stride_src2: u32,
-    stride_src3: u32,
-
-    stride_dst0: u32,
-    stride_dst1: u32,
-    stride_dst2: u32,
-    stride_dst3: u32,
-
-    // Logical shape (same for both tensors)
-    ne0: u32,
-    ne1: u32,
-    ne2: u32,
-    ne3: u32,
-};
-
-@group(0) @binding(2)
-var<uniform> params: Params;
-
-override wg_size: u32;
-@compute @workgroup_size(wg_size)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x >= params.ne) {
-        return;
-    }
-
-    var i = gid.x;
-
-    let i3 = i / (params.ne2 * params.ne1 * params.ne0);
-    i = i % (params.ne2 * params.ne1 * params.ne0);
-
-    let i2 = i / (params.ne1 * params.ne0);
-    i = i % (params.ne1 * params.ne0);
-
-    let i1 = i / params.ne0;
-    let i0 = i % params.ne0;
-
-    let src_idx = i0 * params.stride_src0 + i1 * params.stride_src1 +
-                  i2 * params.stride_src2 + i3 * params.stride_src3;
-
-    let dst_idx = i0 * params.stride_dst0 + i1 * params.stride_dst1 +
-                  i2 * params.stride_dst2 + i3 * params.stride_dst3;
-
-    dst[params.offset_dst + dst_idx] = f16(src[params.offset_src + src_idx]);
-}

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul.tmpl.wgsl

@@ -1,44 +0,0 @@
-#define(VARIANTS)
-
-[
-  {
-    "REPLS": {
-      "TYPE" : "f32",
-    }
-  },
-  {
-    "REPLS": {
-      "TYPE" : "f16",
-    }
-  }
-]
-
-#end(VARIANTS)
-
-#define(SHADER)
-
-enable f16;
-
-#include "binary_head.tmpl"
-
-@group(0) @binding(0)
-var<storage, read_write> src0: array<{{TYPE}}>;
-
-@group(0) @binding(1)
-var<storage, read_write> src1: array<{{TYPE}}>;
-
-@group(0) @binding(2)
-var<storage, read_write> dst: array<{{TYPE}}>;
-
-@group(0) @binding(3)
-var<uniform> params: Params;
-
-override wg_size: u32;
-@compute @workgroup_size(wg_size)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x < params.ne) {
-        dst[params.offset_dst + gid.x] = src0[params.offset_src0 + gid.x] * src1[params.offset_src1 + src1_index(gid.x)];
-    }
-}
-
-#end(SHADER)

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_in_place.tmpl.wgsl

@@ -1,41 +0,0 @@
-#define(VARIANTS)
-
-[
-  {
-    "REPLS": {
-      "TYPE" : "f32",
-    }
-  },
-  {
-    "REPLS": {
-      "TYPE" : "f16",
-    }
-  }
-]
-
-#end(VARIANTS)
-
-#define(SHADER)
-
-enable f16;
-
-#include "binary_head.tmpl"
-
-@group(0) @binding(0)
-var<storage, read_write> src0: array<{{TYPE}}>;
-
-@group(0) @binding(1)
-var<storage, read_write> src1: array<{{TYPE}}>;
-
-@group(0) @binding(2)
-var<uniform> params: Params;
-
-override wg_size: u32;
-@compute @workgroup_size(wg_size)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x < params.ne) {
-        src0[params.offset_dst + gid.x] = src0[params.offset_src0 + gid.x] * src1[params.offset_src1 + src1_index(gid.x)];
-    }
-}
-
-#end(SHADER)

data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_in_place.wgsl

@@ -1,48 +0,0 @@
-@group(0) @binding(0)
-var<storage, read_write> a: array<f32>;
-
-struct Params {
-    offset: u32, // in elements
-
-    // Strides (in elements)
-    stride1: u32,
-    stride2: u32,
-    stride3: u32,
-
-    // Shape
-    ne0: u32,
-    ne1: u32,
-    ne2: u32,
-    ne3: u32,
-
-    eps: u32
-};
-
-@group(0) @binding(1)
-var<uniform> params: Params;
-
-override wg_size: u32;
-@compute @workgroup_size(wg_size)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x >= params.ne1 * params.ne2 * params.ne3) {
-        return;
-    }
-
-    // one thread per row
-    var i = gid.x;
-    let i3 = i / (params.ne2 * params.ne1);
-    i = i % (params.ne2 * params.ne1);
-    let i2 = i / params.ne1;
-    let i1 = i % params.ne1;
-    let i_row = params.offset + i3 * params.stride3 + i2 * params.stride2 + i1 * params.stride1;
-
-    var sum = 0.0f;
-    for (var j: u32 = 0; j < params.ne0; j++) {
-        sum += a[i_row + j] * a[i_row + j];
-    }
-    let eps = bitcast<f32>(params.eps);
-    let scale = 1.0/sqrt(sum/f32(params.ne0) + eps);
-    for (var j: u32 = 0; j < params.ne0; j++) {
-        a[i_row + j] = scale * a[i_row + j];
-    }
-}