vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

port failing dequant callbacks from mul_mm

Browse source
This commit is contained in:
Jeff Bolz 2025-01-23 13:58:21 -06:00 • committed by Rémy O
parent 097befd022
commit 10cc151af5
1 changed files with 48 additions and 45 deletions
Download patch file Download diff file Expand all files Collapse all files

93
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
View file

@ -368,23 +368,25 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2
float16_t dequantFuncIQ2_S(const in decodeBufIQ2_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
uint idx = coordInBlock[1];
uint lsb = idx & 1;
idx /= 2;
const uint ib32 = (idx & 0xE0) >> 5; // 0..7
const uint ib8 = (idx & 0xF8) >> 3; // 0..31
const uint qhshift = 2 * (ib8 % 4);
const uint ib8 = (idx % 128) / 4; // 0..31
const uint ib32 = ib8 / 4; // 0..7
const uint scale = (bl.block.scales[ib32] >> ((idx & 0x10) >> 4)) & 0xf;
const uint scale = (bl.block.scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
const uint qs = bl.block.qs[ib8];
const uint qh = bl.block.qh[ib32];
const uint sign = bl.block.qs[QUANT_K / 8 + ib8];
const uint qhshift = 2 * (ib8 % 4);
const uint sign = bl.block.qs[QUANT_K / 8 + ib8] >> (2 * (idx % 4));
const float16_t dscale = bl.block.d * 0.25hf * (0.5hf + float16_t(scale));
const uint8_t g = unpack8(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 4) >> 2])[idx & 3];
float16_t ret = dscale * float16_t(g) * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
return ret;
const float d = float(bl.block.d);
const float db = d * 0.25 * (0.5 + scale);
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
return float16_t(v[lsb]);
}
#endif
@ -399,25 +401,28 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3
float16_t dequantFuncIQ3_XXS(const in decodeBufIQ3_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufIQ3_XXS_packed16 bl16 = decodeBufIQ3_XXS_packed16(bl);
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
uint idx = coordInBlock[1];
uint lsb = idx & 1;
idx /= 2;
const uint ib32 = (idx & 0xE0) >> 5; // 0..7
const uint ib4 = (idx & 0xFC) >> 4; // 0..63
const uint is16 = QUANT_K / 8 + 2 * ib32; // index in packed16
const uint iqs = (idx % 128) / 2; // 0..63
const uint is = QUANT_K / 4 + 4 * (iqs / 8); // 8 values
const uint8_t qs = bl.block.qs[ib4];
const uint signscale = pack32(u16vec2(bl16.block.qs[is16], bl16.block.qs[is16+1]));
const float16_t dscale = bl.block.d * 0.5hf * (0.5hf + float16_t(signscale >> 28));
uint sign = bitfieldExtract(signscale, 7 * int(ib4 & 3), 7);
sign |= bitCount(sign) << 7;
const uint8_t g = unpack8(iq3xxs_grid[qs])[idx & 3];
float16_t ret = dscale * float16_t(g) * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
return ret;
const float d = float(bl.block.d);
const uint qs = bl.block.qs[iqs];
const uint signs = pack32(u8vec4(
bl.block.qs[is+0],
bl.block.qs[is+1],
bl.block.qs[is+2],
bl.block.qs[is+3]
));
const float db = d * 0.5 * (0.5 + (signs >> 28));
const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
return float16_t(v[lsb]);
}
#endif
@ -428,26 +433,24 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3
float16_t dequantFuncIQ3_S(const in decodeBufIQ3_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
uint idx = coordInBlock[1];
uint lsb = idx & 1;
idx /= 2;
const uint iqs = (idx & 0xFC) >> 2; // 0..63
const uint iqh = (idx & 0xE0) >> 5; // 0..7
const uint qhbit = iqs & 7;
const uint isgn = (idx & 0xF8) >> 3; // 0..31
const uint is = (idx & 0xC0) >> 6; // 0..3
const uint8_t scale = (bl.block.scales[is] >> ((idx & 0x20) >> 3)) & uint8_t(0xF);
const float16_t dscale = d * (1.0hf + float16_t(2 * scale));
const uint iqs = (idx % 128) / 2; // 0..63
const uint iqh = iqs / 8;
const float d = float(bl.block.d);
const uint qs = bl.block.qs[iqs];
const uint qh = (bl.block.qh[iqh] << (8 - qhbit)) & 0x100;
const uint8_t sign = bl.block.signs[isgn];
const uint qh = bl.block.qh[iqh];
const int8_t sign = int8_t(bl.block.signs[iqs / 2] >> (2 * (idx % 4)));
const uint scale = bl.block.scales[iqs / 16];
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
const uint g = unpack8(iq3s_grid[qs | qh])[idx & 3];
float16_t ret = dscale * float16_t(g) * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
return ret;
return float16_t(v[lsb]);
}
#endif