/**
 * 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 = 64) in;

layout (binding = 0) readonly buffer tensorInA { float16_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;
    uint nb01;
    uint nb02;
    uint nb11;
    uint nb12;
    int ne0;
    int ne1;
} pcs;

shared float sum[gl_WorkGroupSize.x];

void main() {
    const uint r0 = gl_WorkGroupID.x;
    const uint r1 = gl_WorkGroupID.y;
    const uint im = gl_WorkGroupID.z;

    const uint x = (r0*pcs.nb01 + im*pcs.nb02) / 2 + pcs.inAOff; // Based from inA
    const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // based from inB

    sum[gl_LocalInvocationID.x] = 0.0;

    for (uint i = gl_LocalInvocationID.x; i < pcs.ne00; i += gl_WorkGroupSize.x) {
        sum[gl_LocalInvocationID.x] += float(inA[x+i]) * float(inB[y+i]);
    }

    // accumulate the sum from all threads in the threadgroup
    barrier();
    memoryBarrierShared();
    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
        if (gl_LocalInvocationID.x < i) {
            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
        }
        barrier();
        memoryBarrierShared();
    }

    if (gl_LocalInvocationID.x == 0) {
        out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = sum[0];
    }
}