simple_v_extension/specification/mv.x.rst

   1 MV.X and MV.swizzle
   2 ===================
   3
   4 swizzle needs a MV.  see below for a potential way to use the funct7 to do a swizzle in rs2.
   5
   6 +---------------+-------------+-------+----------+----------+--------+----------+--------+--------+
   7 | Encoding      | 31:27       | 26:25 | 24:20    | 19:15    | 14:12  | 11:7     | 6:2    | 1:0    |
   8 +---------------+-------------+-------+----------+----------+--------+----------+--------+--------+
   9 | RV32-I-type   + imm[11:0]                      + rs1[4:0] + funct3 | rd[4:0]  + opcode + 0b11   |
  10 +---------------+-------------+-------+----------+----------+--------+----------+--------+--------+
  11 | RV32-I-type   + fn4[3:0]    + swizzle[7:0]     + rs1[4:0] + 0b000  | rd[4:0]  + OP-V   + 0b11   |
  12 +---------------+-------------+-------+----------+----------+--------+----------+--------+--------+
  13
  14 * funct3 = MV: 0b000 for FP, 0b001 for INT
  15 * OP-V = 0b1010111
  16 * fn4 = 4 bit function.
  17 * fn4 = 0b0000 - MV-SWIZZLE
  18 * fn4 = 0bNN01 - MV-X, NN=elwidth (default/8/16/32)
  19 * fn4 = 0bNN11 - MV-X.SUBVL NN=elwidth (default/8/16/32)
  20
  21 swizzle (only active on SV or P48/P64 when SUBVL!=0):
  22
  23 +-----+-----+-----+-----+
  24 | 7:6 | 5:4 | 3:2 | 1:0 |
  25 +-----+-----+-----+-----+
  26 |   w |   z |   y |   x |
  27 +-----+-----+-----+-----+
  28
  29 MV.X has two modes: SUBVL mode applies the element offsets only within a SUBVL inner loop. This can be used for transposition.
  30
  31 ::
  32
  33   for i in range(VL):
  34      for j in range(SUBVL):
  35         regs[rd] = regs[rd+regs[rs+j]]
  36
  37 Normal mode will apply the element offsets incrementally:
  38
  39 ::
  40
  41   for i in range(VL):
  42      for j in range(SUBVL):
  43         regs[rd] = regs[rd+regs[rs+k]]
  44           k++
  45
  46
  47 Pseudocode for element width part of MV.X:
  48
  49 ::
  50
  51   def mv_x(rd, rs1, funct4):
  52       elwidth = (funct4>>2) & 0x3
  53       bitwidth = {0:XLEN, 1:8, 2:16, 3:32}[elwidth] # get bits per el
  54       bytewidth = bitwidth / 8 # get bytes per el
  55       for i in range(VL):
  56           addr = (unsigned char *)&regs[rs1]
  57           offset = addr + bytewidth # get offset within regfile as SRAM
  58           # TODO, actually, needs to respect rd and rs1 element width,
  59           # here, as well.  this pseudocode just illustrates that the
  60           # MV.X operation contains a way to compact the indices into
  61           # less space.
  62           regs[rd] = (unsigned char*)(regs)[offset]
  63
  64 The idea here is to allow 8-bit indices to be stored inside XLEN-sized
  65 registers, such that rather than doing this:
  66
  67 .. parsed-literal::
  68     ldimm x8, 1
  69     ldimm x9, 3
  70     ldimm x10, 2
  71     ldimm x11, 0
  72     {SVP.VL=4} MV.X x3, x8, elwidth=default
  73
  74 The alternative is this:
  75
  76 .. parsed-literal::
  77     ldimm x8, 0x00020301
  78     {SVP.VL=4} MV.X x3, x8, elwidth=8
  79
  80 Thus compacting four indices into the one register.  x3 and x8's element
  81 width are *independent* of the MV.X elwidth, thus allowing both source
  82 and element element widths of the *elements* to be moved to be over-ridden,
  83 whilst *at the same time* allowing the *indices* to be compacted, as well.
  84
  85 ----
  86
  87 potential MV.X?  register-version of MV-swizzle?
  88
  89 +-------------+-------+-------+----------+----------+--------+----------+--------+--------+
  90 | Encoding    | 31:27 | 26:25 | 24:20    | 19:15    | 14:12  | 11:7     | 6:2    | 1:0    |
  91 +-------------+-------+-------+----------+----------+--------+----------+--------+--------+
  92 | RV32-R-type + funct7        + rs2[4:0] + rs1[4:0] + funct3 | rd[4:0]  + opcode + 0b11   |
  93 +-------------+-------+-------+----------+----------+--------+----------+--------+--------+
  94 | RV32-R-type + 0b0000000     + rs2[4:0] + rs1[4:0] + 0b001  | rd[4:0]  + OP-V   + 0b11   |
  95 +-------------+-------+-------+----------+----------+--------+----------+--------+--------+
  96
  97 * funct3 = MV.X
  98 * OP-V = 0b1010111
  99 * funct7 = 0b000NN00 - INT MV.X, elwidth=NN (default/8/16/32)
 100 * funct7 = 0b000NN10 - FP MV.X, elwidth=NN (default/8/16/32)
 101 * funct7 = 0b0000001 - INT MV.swizzle to say that rs2 is a swizzle argument?
 102 * funct7 = 0b0000011 - FP MV.swizzle to say that rs2 is a swizzle argument?
 103
 104 question: do we need a swizzle MV.X as well?
 105
 106 macro-op fusion
 107 ===============
 108
 109 there is the potential for macro-op fusion of mv-swizzle with the following instruction and/or preceding instruction.
 110 <http://lists.libre-riscv.org/pipermail/libre-riscv-dev/2019-August/002486.html>
 111
 112 VBLOCK context?
 113 ===============
 114
 115 additional idea: a VBLOCK context that says that if a given register is used, it indicates that the
 116 register is to be "swizzled", and the VBLOCK swizzle context contains the swizzling to be carried out.
 117
 118 mm_shuffle_ps?
 119 ==============
 120
 121 __m128 _mm_shuffle_ps(__m128 lo,__m128 hi,
 122        _MM_SHUFFLE(hi3,hi2,lo1,lo0))
 123 Interleave inputs into low 2 floats and high 2 floats of output. Basically
 124    out[0]=lo[lo0];
 125    out[1]=lo[lo1];
 126    out[2]=hi[hi2];
 127    out[3]=hi[hi3];
 128
 129 For example, _mm_shuffle_ps(a,a,_MM_SHUFFLE(i,i,i,i)) copies the float
 130 a[i] into all 4 output floats.
 131
 132 Transpose
 133 =========
 134
 135 assuming a vector of 4x4 matrixes is stored as 4 separate vectors with subvl=4 in struct-of-array-of-struct form (the form I've been planning on using):
 136 using standard (4+4) -> 4 swizzle instructions with 2 input vectors with subvl=4 and 1 output vector with subvl, a vectorized matrix transpose operation can be done in 2 steps with 4 instructions per step to give 8 instructions in total:
 137
 138 input:
 139 | m00 m10 m20 m30 |
 140 | m01 m11 m21 m31 |
 141 | m02 m12 m22 m32 |
 142 | m03 m13 m23 m33 |
 143
 144 transpose 4 corner 2x2 matrices
 145
 146 intermediate:
 147 | m00 m01 m20 m21 |
 148 | m10 m11 m30 m31 |
 149 | m02 m03 m22 m23 |
 150 | m12 m13 m32 m33 |
 151
 152 finish transpose
 153
 154 output:
 155 | m00 m01 m02 m03 |
 156 | m10 m11 m12 m13 |
 157 | m20 m21 m22 m23 |
 158 | m30 m31 m32 m33 |
 159
 160 <http://web.archive.org/web/20100111104515/http://www.randombit.net:80/bitbashing/programming/integer_matrix_transpose_in_sse2.html>
 161
 162
 163 ::
 164
 165    __m128i T0 = _mm_unpacklo_epi32(I0, I1);
 166    __m128i T1 = _mm_unpacklo_epi32(I2, I3);
 167    __m128i T2 = _mm_unpackhi_epi32(I0, I1);
 168    __m128i T3 = _mm_unpackhi_epi32(I2, I3);
 169
 170    /* Assigning transposed values back into I[0-3] */
 171    I0 = _mm_unpacklo_epi64(T0, T1);
 172    I1 = _mm_unpackhi_epi64(T0, T1);
 173    I2 = _mm_unpacklo_epi64(T2, T3);
 174    I3 = _mm_unpackhi_epi64(T2, T3);
 175
 176 Transforms for DCT
 177 ==================
 178
 179 <https://opencores.org/websvn/filedetails?repname=mpeg2fpga&path=%2Fmpeg2fpga%2Ftrunk%2Frtl%2Fmpeg2%2Fidct.v>
 180