Skip to content

Commit 6e84b0a

Browse files
qnixsynapseqnixsynapse
authored
SYCL : SOFTMAX F16 mask support and other fixes (#11261)
Implemented ggml_sycl_op_soft_max() F16 src1(mask) support for which a pragma deprecation warning was added during #5021. To do this, had to decouple it from ggml_sycl_op_flatten which always considered src1 to be of fp32 type(many OP functions are dependent on it). * SYCL: SOFTMAX F16 mask support and other fixes * test-backend-ops: Add F16 mask test cases
1 parent 2b8525d commit 6e84b0a

File tree

4 files changed

+64
-49
lines changed

4 files changed

+64
-49
lines changed

‎ggml/src/ggml-sycl/ggml-sycl.cpp

+1-5
Original file line numberDiff line numberDiff line change
@@ -3878,10 +3878,6 @@ static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor
38783878
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_diag_mask_inf);
38793879
}
38803880

3881-
static void ggml_sycl_soft_max(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
3882-
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_soft_max);
3883-
}
3884-
38853881
static void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
38863882
GGML_ASSERT(ggml_is_contiguous(dst->src[0])); // TODO: this restriction is temporary until non-cont support is implemented
38873883
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_rope);
@@ -4090,7 +4086,7 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
40904086
ggml_sycl_diag_mask_inf(ctx, dst);
40914087
break;
40924088
case GGML_OP_SOFT_MAX:
4093-
ggml_sycl_soft_max(ctx, dst);
4089+
ggml_sycl_op_soft_max(ctx, dst);
40944090
break;
40954091
case GGML_OP_ROPE:
40964092
ggml_sycl_rope(ctx, dst);

‎ggml/src/ggml-sycl/softmax.cpp

+33-23
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
#include "norm.hpp"
1+
#include "softmax.hpp"
22

3-
template <bool vals_smem, int ncols_template, int block_size_template>
4-
static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par,
3+
template <bool vals_smem, int ncols_template, int block_size_template, typename T>
4+
static void soft_max_f32(const float * x, const T * mask, float * dst, const int ncols_par,
55
const int nrows_y, const float scale, const float max_bias, const float m0,
66
const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) {
77
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
@@ -29,7 +29,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
2929
slope = sycl::pow(base, float(exp));
3030
}
3131

32-
float *vals = vals_smem ? buf + std::max(nwarps, WARP_SIZE) : dst + rowx * ncols;
32+
float *vals = vals_smem ? buf + sycl::max(nwarps, WARP_SIZE) : dst + rowx * ncols;
3333
float max_val = -INFINITY;
3434

3535
for (int col0 = 0; col0 < ncols; col0 += block_size) {
@@ -42,7 +42,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
4242
const int ix = rowx*ncols + col;
4343
const int iy = rowy*ncols + col;
4444

45-
const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f);
45+
const float val = x[ix]*scale + (mask ? slope*static_cast<float>(mask[iy]) : 0.0f);
4646

4747
vals[col] = val;
4848
max_val = sycl::max(max_val, val);
@@ -65,7 +65,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
6565
item_ct1.barrier(sycl::access::fence_space::local_space);
6666
max_val = buf[lane_id];
6767
for (size_t i = 1; i < nreduce; i += 1) {
68-
max_val = std::max(max_val, buf[lane_id + i * WARP_SIZE]);
68+
max_val = sycl::max(max_val, buf[lane_id + i * WARP_SIZE]);
6969
}
7070
max_val = warp_reduce_max(max_val, item_ct1);
7171
}
@@ -122,8 +122,8 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
122122
}
123123
}
124124

125-
template <bool vals_smem, int ncols_template, int block_size_template>
126-
static void soft_max_f32_submitter(const float * x, const float * mask, float * dst, const int ncols_par,
125+
template <bool vals_smem, int ncols_template, int block_size_template, typename T>
126+
static void soft_max_f32_submitter(const float * x, const T * mask, float * dst, const int ncols_par,
127127
const int nrows_y, const float scale, const float max_bias, const float m0,
128128
const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
129129
const size_t n_local_scratch, queue_ptr stream) {
@@ -141,7 +141,8 @@ static void soft_max_f32_submitter(const float * x, const float * mask, float *
141141
});
142142
}
143143

144-
static void soft_max_f32_sycl(const float * x, const float * mask,
144+
template<typename T>
145+
static void soft_max_f32_sycl(const float * x, const T * mask,
145146
float * dst, const int ncols_x, const int nrows_x,
146147
const int nrows_y, const float scale, const float max_bias,
147148
queue_ptr stream, int device) {
@@ -223,29 +224,38 @@ static void soft_max_f32_sycl(const float * x, const float * mask,
223224
}
224225
}
225226

226-
void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
227-
const ggml_tensor *src1, ggml_tensor *dst,
228-
const float *src0_dd, const float *src1_dd,
229-
float *dst_dd,
230-
const queue_ptr &main_stream) {
227+
void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
231228

232-
GGML_ASSERT(src0->type == GGML_TYPE_F32);
229+
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
233230
GGML_ASSERT( dst->type == GGML_TYPE_F32);
234231

235-
#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 support")
236-
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021")
237-
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
232+
GGML_ASSERT(!dst->src[1] || dst->src[1]->type == GGML_TYPE_F16 || dst->src[1]->type == GGML_TYPE_F32); // src1 contains mask and it is optional
238233

239-
const int64_t ne00 = src0->ne[0];
240-
const int64_t nrows_x = ggml_nrows(src0);
241-
const int64_t nrows_y = src0->ne[1];
234+
const int64_t ne00 = dst->src[0]->ne[0];
235+
const int64_t nrows_x = ggml_nrows(dst->src[0]);
236+
const int64_t nrows_y = dst->src[0]->ne[1];
242237

243238
float scale = 1.0f;
244239
float max_bias = 0.0f;
245240

246241
memcpy(&scale, dst->op_params + 0, sizeof(float));
247242
memcpy(&max_bias, dst->op_params + 1, sizeof(float));
248243

249-
soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00,
250-
nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
244+
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
245+
float * dst_dd = static_cast<float *>(dst->data);
246+
247+
ggml_sycl_set_device(ctx.device);
248+
dpct::queue_ptr main_stream = ctx.stream();
249+
250+
if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F16) {
251+
const sycl::half * src1_dd = static_cast<sycl::half *>(dst->src[1]->data);
252+
soft_max_f32_sycl<sycl::half>(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias,
253+
main_stream, ctx.device);
254+
} else if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F32) {
255+
const float * src1_dd = static_cast<const float *>(dst->src[1]->data);
256+
soft_max_f32_sycl<float>(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
257+
} else {
258+
/* mask unavailable */
259+
soft_max_f32_sycl<float>(src0_dd, nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
260+
}
251261
}

‎ggml/src/ggml-sycl/softmax.hpp

+1-5
Original file line numberDiff line numberDiff line change
@@ -15,10 +15,6 @@
1515

1616
#include "common.hpp"
1717

18-
void ggml_sycl_op_soft_max(ggml_backend_sycl_context &ctx, const ggml_tensor *src0,
19-
const ggml_tensor *src1, ggml_tensor *dst,
20-
const float *src0_dd, const float *src1_dd,
21-
float *dst_dd,
22-
const queue_ptr &main_stream);
18+
void ggml_sycl_op_soft_max(ggml_backend_sycl_context &ctx, ggml_tensor *dst);
2319

2420
#endif // GGML_SYCL_SOFTMAX_HPP

‎tests/test-backend-ops.cpp

+29-16
Original file line numberDiff line numberDiff line change
@@ -2347,11 +2347,12 @@ struct test_soft_max : public test_case {
23472347
const ggml_type type;
23482348
const std::array<int64_t, 4> ne;
23492349
const bool mask;
2350+
const ggml_type m_prec;
23502351
const float scale;
23512352
const float max_bias;
23522353

23532354
std::string vars() override {
2354-
return VARS_TO_STR5(type, ne, mask, scale, max_bias);
2355+
return VARS_TO_STR6(type, ne, mask, m_prec, scale, max_bias);
23552356
}
23562357

23572358
// the 1024 test with bias occasionally fails:
@@ -2363,9 +2364,10 @@ struct test_soft_max : public test_case {
23632364
test_soft_max(ggml_type type = GGML_TYPE_F32,
23642365
std::array<int64_t, 4> ne = {10, 5, 4, 3},
23652366
bool mask = false,
2367+
ggml_type m_prec = GGML_TYPE_F32,
23662368
float scale = 1.0f,
23672369
float max_bias = 0.0f)
2368-
: type(type), ne(ne), mask(mask), scale(scale), max_bias(max_bias) {}
2370+
: type(type), ne(ne), mask(mask), m_prec(m_prec), scale(scale), max_bias(max_bias) {}
23692371

23702372
ggml_tensor * build_graph(ggml_context * ctx) override {
23712373
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
@@ -2374,7 +2376,7 @@ struct test_soft_max : public test_case {
23742376

23752377
ggml_tensor * mask = nullptr;
23762378
if (this->mask) {
2377-
mask = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ne[0], ne[1]);
2379+
mask = ggml_new_tensor_2d(ctx, m_prec, ne[0], ne[1]);
23782380
ggml_set_name(mask, "mask");
23792381
}
23802382

@@ -4150,17 +4152,28 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
41504152
for (float scale : {1.0f, 0.1f}) {
41514153
for (int64_t ne0 : {16, 1024}) {
41524154
for (int64_t ne1 : {16, 1024}) {
4153-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, scale, max_bias));
4154-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, scale, max_bias));
4155+
if (mask) {
4156+
for (ggml_type m_prec : {GGML_TYPE_F32, GGML_TYPE_F16}) {
4157+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, m_prec, scale, max_bias));
4158+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, scale, max_bias));
4159+
}
4160+
} else {
4161+
/* The precision of mask here doesn't matter as boolean mask is false */
4162+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, GGML_TYPE_F32, scale, max_bias));
4163+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, GGML_TYPE_F32, scale, max_bias));
4164+
}
41554165
}
41564166
}
41574167
}
41584168
}
41594169
}
4160-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, 0.1f, 0.0f));
4161-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 0.0f));
4162-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 0.0f));
4163-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 8.0f));
4170+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F32, 0.1f, 0.0f));
4171+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F16, 0.1f, 0.0f));
4172+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, GGML_TYPE_F32, 0.1f, 0.0f));
4173+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32, 0.1f, 0.0f));
4174+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16, 0.1f, 0.0f));
4175+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32, 0.1f, 8.0f));
4176+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16, 0.1f, 8.0f));
41644177

41654178
for (float max_bias : {0.0f, 8.0f}) {
41664179
for (float scale : {1.0f, 0.1f}) {
@@ -4296,13 +4309,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
42964309
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {8192, 512, 2, 1}, {0, 2, 1, 3}));
42974310
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {3072, 512, 2, 1}, {0, 2, 1, 3}));
42984311

4299-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, 1.0f, 0.0f));
4300-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, 1.0f, 0.0f));
4301-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, 1.0f, 0.0f));
4302-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, 1.0f, 0.0f));
4303-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {256, 256, 20, 1}, false, 1.0f, 0.0f));
4304-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, 1.0f, 0.0f));
4305-
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, 1.0f, 0.0f));
4312+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
4313+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
4314+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
4315+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
4316+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {256, 256, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
4317+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
4318+
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
43064319

43074320
test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32, 10, 1, 1}));
43084321
test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));

0 commit comments

Comments
 (0)