/**
 * 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 = 8, local_size_y = 8) in;

layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };

layout (push_constant) uniform parameter {
    uint inAOff;
    uint inBOff;
    uint outOff;
    int ne00;
    int ne10;
    int ne0;
} pcs;

shared float sum[64];

void main() {
    const uint nb = uint(pcs.ne00/QK4_0);

    const uint r0 = gl_WorkGroupID.x;
    const uint r1 = gl_WorkGroupID.y;

    const uint x = r0*nb; // Based from inA without base offset
    const uint y = r1*uint(pcs.ne10) + pcs.inBOff; // Based from inB

    const uint nth = gl_WorkGroupSize.x*gl_WorkGroupSize.y;
    const uint ith = gl_WorkGroupSize.y*gl_LocalInvocationID.x + gl_LocalInvocationID.y;

    const uint ix = gl_LocalInvocationID.y/4;           // 0 or 1
    const uint iy = gl_LocalInvocationID.y - 4*ix;      // 0...3

    const uint first = 4 * iy;

    float sumf = 0.0;

    for (uint i = 2*gl_LocalInvocationID.x + ix; i < nb; i += 2*gl_WorkGroupSize.x) {
        const uint index = (x+i)*sizeof_block_q4_0+pcs.inAOff;
        const float d = float(u8BufToFloat16(inA, index));

        const uint xl = first; // Based from bl->qs
        const uint yl = y + i * QK4_0 + first; // Based from inB

        vec2 acc = vec2(0.0, 0.0);

        for (int j = 0; j < 4; ++j) {
            const uint8_t b = inA[index+2+xl+j];
            acc.x += inB[yl+j] * (b & 0xF) + inB[yl+j+16] * (b >> 4);
            acc.y += inB[yl+j] + inB[yl+j+16];
        }

        sumf += d * (acc.x - 8.*acc.y);
    }

    sum[ith] = sumf;

    //
    // Accumulate the sum from all threads in the threadgroup
    //
    barrier();
    if (ith == 0) {
        float sumTotal = 0.0;
        for (uint i = 0; i < nth; ++i) {
            sumTotal += sum[i];
        }
        out_[r1*uint(pcs.ne0) + r0 + pcs.outOff] = sumTotal;
    }
}