/**
 * Copyright (c) 2023 Nomic, Inc. All rights reserved.
 *
 * This software is licensed under the terms of the Software for Open Models License (SOM),
 * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
 * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
 */

#version 450

#extension GL_EXT_shader_16bit_storage: require
#extension GL_EXT_shader_8bit_storage: require
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
#extension GL_EXT_control_flow_attributes: enable

#define QK4_0 32
#define QR4_0 2
#define QK4_1 32

#define GELU_COEF_A 0.044715
#define SQRT_2_OVER_PI 0.79788456080286535587989211986876

#ifndef QK_K
#define QK_K 256
#endif

#if QK_K == 256
#define K_SCALE_SIZE 12
#else
#define K_SCALE_SIZE 4
#endif

#define BM 128
#define BN 128
#define BK 8
#define TM 8
#define TN 8

#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)

#define sizeof_block_q4_0 0x12
#define sizeof_block_q4_1 0x14
struct block_q4_0 {
    float16_t d;
    uint8_t qs[QK4_0 / 2];
};
struct block_q4_1 {
    float16_t d;
    float16_t m;
    uint8_t qs[QK4_1 / 2];
};

#ifndef QK_K
#define QK_K 256
#endif

#if QK_K == 256
#define K_SCALE_SIZE 12
#else
#define K_SCALE_SIZE 4
#endif

struct block_q2_K {
    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
    uint8_t qs[QK_K/4];      // quants
    float16_t d;           // super-block scale for quantized scales
    float16_t dmin;        // super-block scale for quantized mins
};
// 84 bytes / block

struct block_q3_K {
    uint8_t hmask[QK_K/8];     // quants - high bit
    uint8_t qs[QK_K/4];        // quants - low 2 bits
#if QK_K == 64
    uint8_t scales[2];
#else
    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
    float16_t d;             // super-block scale
};

#if QK_K == 64
typedef struct {
    float16_t    d[2];          // super-block scales/mins
    uint8_t scales[2];
    uint8_t qs[QK_K/2];    // 4-bit quants
} block_q4_K;
#else
struct block_q4_K {
    float16_t d;             // super-block scale for quantized scales
    float16_t dmin;          // super-block scale for quantized mins
    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
    uint8_t qs[QK_K/2];        // 4--bit quants
};
#endif

#if QK_K == 64
struct block_q5_K {
    float16_t  d;                     // super-block scales/mins
    int8_t  scales[QK_K/16];     // 8-bit block scales
    uint8_t qh[QK_K/8];          // quants, high bit
    uint8_t qs[QK_K/2];          // quants, low 4 bits
};
#else
struct block_q5_K {
    float16_t d;                      // super-block scale for quantized scales
    float16_t dmin;                   // super-block scale for quantized mins
    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
    uint8_t qh[QK_K/8];          // quants, high bit
    uint8_t qs[QK_K/2];          // quants, low 4 bits
};
// 176 bytes / block
#endif

struct block_q6_K {
    uint8_t ql[QK_K/2];      // quants, lower 4 bits
    uint8_t qh[QK_K/4];      // quants, upper 2 bits
    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
    float16_t d;                  // super-block scale
};
// 210 bytes / block

layout(local_size_x = 1) in;

layout (binding = 0) readonly buffer tensorIn { float in_[]; };
layout (binding = 1) writeonly buffer tensorOut { float out_[]; };

layout (push_constant) uniform parameter {
    uint inOff;
    uint outOff;
    uint n_past;
    int n_dims;
    int mode;
    float freq_base;
    float freq_scale;
    uint nb00;
    uint nb01;
    uint nb02;
    uint nb03;
    int ne0;
    uint nb0;
    uint nb1;
    uint nb2;
    uint nb3;
} pcs;

void main() {
    const uint i3 = gl_WorkGroupID.z;
    const uint i2 = gl_WorkGroupID.y;
    const uint i1 = gl_WorkGroupID.x;

    const bool is_neox = (pcs.mode & 2) != 0;
    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);

    const uint p = ((pcs.mode & 1) == 0 ? pcs.n_past + i2 : i2);

    float theta = pcs.freq_scale * float(p);

    if (!is_neox) {
        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
            const float cos_theta = cos(theta);
            const float sin_theta = sin(theta);

            theta *= theta_scale;

            const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inOff; // Based from in
            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_

            const float x0 = in_[src];
            const float x1 = in_[src+1];

            out_[dst_data] = x0*cos_theta - x1*sin_theta;
            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
        }
    } else {
        // TODO: implement
    }
}